content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
After receiving paper reviews, authors may optionally submit a rebuttal to address the reviewers' comments, which will be limited to a {\bf one page} PDF file. Please follow the steps and style guidelines outlined below for submitting your author response.
Note that the author rebuttal is optional and, following similar guidelines to previous CVPR conferences, it is meant to provide you with an opportunity to rebut factual errors or to supply additional information requested by the reviewers. It is NOT intended to add new contributions (theorems, algorithms, experiments) that were not included in the original submission. You may optionally add a figure, graph or proof to your rebuttal to better illustrate your answer to the reviewers' comments.
Per a passed 2018 PAMI-TC motion, reviewers should not request additional experiments for the rebuttal, or penalize authors for lack of additional experiments. This includes any experiments that involve running code, e.g., to create tables or figures with new results. \textbf{Authors should not include new experimental results in the rebuttal}, and reviewers should discount any such results when making their final recommendation. Authors may include figures with illustrations or comparison tables of results reported in the submission/supplemental material or in other papers.
The rebuttal must adhere to the same blind-submission as the original submission and must comply with this rebuttal-formatted template.
\subsection{Response length}
Author responses must be no longer than 1 page in length including any references and figures. Overlength responses will simply not be reviewed. This includes responses where the margins and formatting are deemed to have been significantly altered from those laid down by this style guide. Note that this \LaTeX\ guide already sets figure captions and references in a smaller font.
\section{Formatting your Response}
{\bf Make sure to update the paper title and paper ID in the appropriate place in the tex file.}
All text must be in a two-column format. The total allowable width of the text
area is $6\frac78$ inches (17.5 cm) wide by $8\frac78$ inches (22.54 cm) high.
Columns are to be $3\frac14$ inches (8.25 cm) wide, with a $\frac{5}{16}$ inch
(0.8 cm) space between them. The top margin should begin
1.0 inch (2.54 cm) from the top edge of the page. The bottom margin should be
1-1/8 inches (2.86 cm) from the bottom edge of the page for $8.5 \times
11$-inch paper; for A4 paper, approximately 1-5/8 inches (4.13 cm) from the
bottom edge of the page.
Please number all of your sections and any displayed equations. It is important
for readers to be able to refer to any particular equation.
Wherever Times is specified, Times Roman may also be used. Main text should be
in 10-point Times, single-spaced. Section headings should be in 10 or 12 point
Times. All paragraphs should be indented 1 pica (approx. 1/6 inch or 0.422
cm). Figure and table captions should be 9-point Roman type as in
Figure~\ref{fig:onecol}.
List and number all bibliographical references in 9-point Times, single-spaced,
at the end of your response. When referenced in the text, enclose the citation
number in square brackets, for example~\cite{Authors14}. Where appropriate,
include the name(s) of editors of referenced books.
\begin{figure}[t]
\begin{center}
\fbox{\rule{0pt}{1in} \rule{0.9\linewidth}{0pt}}
\end{center}
\caption{Example of caption. It is set in Roman so that mathematics
(always set in Roman: $B \sin A = A \sin B$) may be included without an
ugly clash.}
\label{fig:long}
\label{fig:onecol}
\end{figure}
\subsection{Illustrations, graphs, and photographs}
All graphics should be centered. Please ensure that any point you wish to make is resolvable in a printed copy of the response. Resize fonts in figures to match the font in the body text, and choose line widths which render effectively in print. Many readers (and reviewers), even of an electronic copy, will choose to print your response in order to read it. You cannot insist that they do otherwise, and therefore must not assume that they can zoom in to see tiny details on a graphic.
When placing figures in \LaTeX, it's almost always best to use \verb+\includegraphics+, and to specify the figure width as a multiple of the line width as in the example below
{\small\begin{verbatim}
\usepackage[dvips]{graphicx} ...
\includegraphics[width=0.8\linewidth]
{myfile.eps}
\end{verbatim}
}
{\small
\bibliographystyle{ieee_fullname}
\section{Introduction}
Recognizing objects in video is an important yet costly task. Many real-world applications, such as autonomous driving and video surveillance, require efficient processing of video data. Deep convolutional neural network (CNN) models are a powerful tool for recognition tasks. However, CNN models are computationally demanding, making real-time processing infeasible on embedded devices.
As a result, many methods have been proposed to exploit the short-range correspondence that exists in many videos~\cite{zhu2017deep,mao2018catdet,luo2019detect,wang2019fast,liu2019looking}.
These methods extract explicit or implicit correspondence between frames, such as optical flow~\cite{zhu2017deep}, object tracking~\cite{mao2018catdet,luo2019detect}, motion fields in compressed video learned with LSTM~\cite{wang2019fast}. Discovering cheap and reliable correspondence in video greatly reduces processing cost while preserving prediction quality.
\begin{figure}[h!]
\centering
\includegraphics[width=0.47\textwidth]{figures/Figure1.pdf}
\caption{For short-range template matching, PatchNet combines the simplicity of correlation filter methods and the learnability of Siamese networks. PatchNet works by fitting a very efficient CNN on a patch-wise correlation map instead of image pixels.
}
\label{fig:comparison}
\vspace{-5pt}
\end{figure}
On the other hand, the recent progress on compact CNN models~\cite{howard2017mobilenets,tan2019efficientnet} has placed stricter computation budget for temporal correspondence. The budget for a CNN model has been pushed from $\sim$10GFLOPs of VGG~\cite{simonyan2014very} and ResNet~\cite{he2016deep} to $\sim$600MFLOPs in mobile regime~\cite{radosavovic2020designing}. One may find the existing techniques no longer yield significant speed-ups, as the cost to compute correspondence is no longer negligible. As an example, the recent EfficientDet-D0~\cite{tan2019efficientdet} costs only 2.5GFLOPs, while optical flow by FlowNet-Inception~\cite{zhu2017deep} alone costs about 1.8GFLOPs, 72\% of the backbone itself.
Intuitively, matching a previously seen object should be cheaper than searching for a new object. The proposed PatchNet is designed for this low-cost \textit{short-range template matching} problem. It has a low computation footprint, like correlation filter methods~\cite{henriques2014high,bolme2010visual}, but with robustness similar to deep CNN based Siamese Networks~\cite{li2018high}. As shown in Figure~\ref{fig:comparison}, instead of using a single correlation filter, PatchNet splits a template into patches and generates a multi-channel correlation map. This correlation map, treated as a feature map, is then fed into a shallow CNN model to predict object locations. In addition, we propose a hierarchical bounding box regression scheme that enables bounding box regression with just a few layers.
The contributions of this work are listed as follows:
\begin{itemize}
\item This work describes the short-range template matching problem, and demonstrates that patch-wise correlation enables very efficient model design for this problem.
\item For video object detection, patchNet achieves 4.9x speed-up over R-FCN ResNet-101 and 3.4x speed-up over EfficientDet-D0, with less than 1\% mAP loss.
\item For visual object tracking, PatchNet achieves 2.6x speed-up compared to SiamRPN without loss of accuracy, or 3.9x with 1\% success score loss.
\item Experiments on Jetson Nano demonstrates near-linear speed-ups associated with FLOP reduction by PatchNet.
\end{itemize}
\section{Related Work}
\textbf{Visual object tracking} (VOT) considers the problem of tracking an arbitrary object in video solely by its template in the first frame. In the past, Correlation Filter based methods such as MOSSE~\cite{bolme2010visual} and KCF~\cite{henriques2014high} have shown great success on both accuracy and performance. Recent progress of Deep Convolutional Neural Network has enabled substantial advances in this field. SiameseFC~\cite{koch2015siamese} first formulates the VOT problem in the Siamese Network paradigm. SiamRPN~\cite{li2018high} incorporates Region Proposal Networks (RPN) from object detection, which boosts accuracy while reducing model run time. Many works follow this paradigm to further improve tracking results, such as DaSiamRPN~\cite{zhu2018distractor}, SiamMask~\cite{wang2019fast} and SiamRCNN~\cite{voigtlaender2019siam}.
Siamese Networks work well on a wide range of problems apart from VOT, such as one-shot recognition~\cite{koch2015siamese}, re-identification~\cite{chung2017two}, and similarity measuring~\cite{rao2016deep}. The foundation of all these problems is template matching, for which Siamese Network is a suitable solution.
\textbf{Video object detection} (VOD) aims to reduce the computational cost or improve accuracy of object detection with temporal information. Most VOD models are based on image object detection models, e.g., Faster R-CNN~\cite{ren2015faster}, R-FCN~\cite{dai2016r} and SSD Multi-Box~\cite{liu2016ssd}.
Temporal information extraction, the key for efficient and accurate video recognition, can be classified as feature level or object level. Feature-level temporal information discovers correspondence across feature maps. Deep Feature Flow~\cite{zhu2017deep} employs a reduced FlowNet to estimate optical flow and extrapolates high-level CNN features. LSTM-aided SSD~\cite{liu2018mobile} adds an LSTM layer to learn implicit correspondence between multi-frame features.
Object-level temporal information leverages previous object detections for future detections. These methods are typically associated with object tracking. Detect to Track~\cite{feichtenhofer2017detect} adds additional regression head and training loss to encourage temporally coherent detection. Detect or Track~\cite{luo2019detect} employs a SiamFC model to locate detected objects and skip detection. CaTDet~\cite{mao2018catdet} performs selected-region detection based on a multi-object tracker's predictions.
\section{Short-range Template Matching}
\subsection{Applications}
Our goal is to match a seen object in a near future frame at a very low computation cost. By matching existing objects across frames, it is possible to skip the high-cost detection computation in most frames. We list two applications that benefit from efficient short-range template matching.
\textbf{Efficient video object detection} methods exploit the redundancy in consecutive frames to reduce computation cost. For example, one can avoid detection in inter-frames via template matching~\cite{luo2019detect,yao2020deep} or feature extrapolation~\cite{zhu2017deep}. In this work, we run the base detector only on sparse keyframes and running short-range template matching on inter-frames to reduce overall workload.
\textbf{Efficient visual object tracking} can also be achieved using short-range template matching. With a low-cost model, like PatchNet, we run expensive Siamese models on sparse key frames only. Again, short-range template matching is run on the inter-frames. Key frames can be selected in an online manner, when the number of skipped frames exceeds a limit or the similarity between consecutive frames falls below a threshold.
\subsection{Template matching with patches}
Short-range template matching is very similar to visual object tracking, which is also solved as a template matching problem.
In contrast to the visual object tracking problem where object appearance can drastically change, low-level features of an object are well preserved over a short period, as shown in Figure~\ref{fig:patch_aggregate}. Although similar features exist across multiple frames, their spatial alignment can vary due to movement, distortion, or scale change.
\begin{figure}[t!]
\centering
\includegraphics[width=0.43\textwidth]{figures/lowlevel.pdf}
\caption{Similar low-level features widely exist in consecutive frames, although they may move to different relative positions. This type of redundancy inspires a new approach to reduce computation -- learning a CNN on correlation features instead of directly on pixels.
}
\label{fig:patch_aggregate}
\end{figure}
Correlation filter methods~\cite{henriques2014high,bolme2010visual} have been widely applied to visual object tracking.
A filter is created from the template and correlated with subsequent frames.
The peak correlation score locates the object. Correlation filters are lightweight and robust to certain distortions, but especially prone to spatial variation~\cite{li2018robust,danelljan2016discriminative}.
On the other hand, modern CNN models usually have small filters learned from data and many layers to aggregate features from low-level to high-level, and demonstrate good invariance to spatial variation.
Such robustness is attributed not only to large amount of training data and learnable parameters, but also to the great representation power of the CNN architecture.
\begin{figure}[t!]
\centering
\centering
\includegraphics[width=0.39\textwidth]{figures/Patch-spread.pdf}
\caption{The learned CNN model relaxes the spatial adjacency of patches and creates more robust predictions. Left: correlation filter suffers from scale change. Right: aggregated patch score from the CNN permits limited movement of patches which is more robust to scale change. }
\vspace{-10pt}
\label{fig:toycase}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.45\textwidth]{figures/offset_with_scale.pdf}
\caption{Localization accuracy versus scale for 1000 randomly selected ImageNet-VID objects. PatchNet works on a larger range of scales than full template correlation.}
\label{fig:patch_dist}
\end{figure}
\begin{figure*}[h!]
\centering
\includegraphics[width=0.85\textwidth]{figures/PatchNet3.pdf}
\caption{Overview of PatchNet. The patch correlation layer resembles a 2-D convolutional layer but with replaceable weights, which are evenly split patches from the template. It outputs a multi-channel feature map, each channel corresponding to a patch's correlation scores. The aggregation subnet, which is a 3-stage CNN, aggregates the correlation map into a single-channel feature map that localizes the object center by the peak. In addition, it has a hierarchical bounding box regression branch based on correlation information.
}
\label{fig:patch_aggregate}
\end{figure*}
As short-range template matching requires low-cost methods, we look for an approach with the economy of a correlation filter but with the robustness of a deep neural network.
Inspired by the small filters and hierarchical feature extraction of a CNN, we improve upon the correlation filters by splitting a template into small patches and spatially aggregating their correlation scores. The aggregation is learned by a shallow CNN model, which combines patch-wise correlation scores into object-wise scores. This \textit{Patch Aggregation} has two main advantages. First, low level features in patches are robust to scale changes~\cite{lowe1999object}. Second, aggregation relaxes the adjacency constraint of patches using convolution and max pooling layers, creating more robust responses to spatial distortion.
The example in Figure~\ref{fig:toycase} illustrates the scale invariance of patch aggregation, where we backtrace and identify the contributing patches to the peak response score.
We further sample 1000 objects from ImageNet VID, resize them to multiple scales and plot the center mismatch of predictions and groundtruth in Figure~\ref{fig:patch_dist}.
Full template correlation is very sensitive to scale change. Patch aggregation, on the other hand, has smaller center mismatch across a large range of scales.
In addition to more accurate centers, we will show that patch aggregation can identify new bounding boxes at low cost, which correlation filters find difficult.
\section{PatchNet Architecture}
Figure~\ref{fig:patch_aggregate} shows the PatchNet architecture.
PatchNet is divided into a \textit{patch correlation layer} and an \textit{aggregation subnet}. The patch correlation layer splits the template into patches and generates a 3D patch-wise correlation map. The aggregation subnet performs localization and bounding box estimation using the correlation map. Together they form a very efficient template matching model.
\subsection{Patch Correlation Layer}
In the patch correlation layer, a fixed-size template is split into a fixed number of patches. Each patch serves as a convolution filter, and together all patches form a 4D tensor as convolution weight. After that, patchwise correlation is performed by a normal 2D convolution function to generate a 3D correlation map.
This 3D correlation map contains rich patch similarity information per channel, which is then treated as a feature map and fed into the aggregation subnet to localize the object center and regress the bounding box.
Plain image correlation often falsely responds to background~\cite{bolme2010visual}, so does plain patch correlation.
We propose a simple yet effective learning-based method to mitigate this issue. PatchNet selects weighted Fourier-domain features rather than plain image features. It has been reported that low-frequency components are more robust to perturbations~\cite{rahaman2018spectral}. With learnable Fourier coefficients and end-to-end training, the patch correlation layer learns frequency components that are robust for recognition. In addition, Fourier transformation can be efficiently executed on both CPU and GPU.
Figure~\ref{fig:feat_sub} illustrates patch splitting and Fourier re-weighting. The template image, first warped to a fixed size ($KN \times KN)$, is then divided into $N^2$ patches of size $K\times K$. After they are evenly cropped out from the template, the $N^2$ patches are then transformed to the Fourier domain, multiplied by learned weights, and then converted back to the spatial domain. When $K$ has a base of 2, the Fourier transformation steps can be accelerated using an FFT with a time complexity of $O(N^2KlogK)$, which is almost negligible compared with the following convolution step's $O(HWN^2K^2)$.
The learned Fourier coefficients are visualized in Figure~\ref{fig:fcoef}. Due to the symmetry of Fourier transformation, frequency components are mirrored across axes $x=4$ and $y=4$. Therefore, we omit the mirrored weights. It is interesting that PatchNet learns to suppress both high frequencies and the constant (DC) base, implying that low-frequency information is more reliable for object recognition.
\begin{figure}[h!]
\centering
\includegraphics[width=0.45\textwidth]{figures/FourierImportance.pdf}
\vspace{-5pt}
\caption{In the patch correlation layer, the template is split into patches, reweighted in the Fourier domain and converted to convolutional filters.}
\label{fig:feat_sub}
\end{figure}
\begin{figure}[h!]
\centering
\begin{subfigure}{0.25\textwidth}
\includegraphics[width=\textwidth]{figures/ff_vis.pdf}
\end{subfigure}
\hfill
\begin{subfigure}{0.22\textwidth}
\includegraphics[width=0.5\textwidth]{figures/patch_kernel.pdf}
\end{subfigure}
\caption{Left: learned Fourier coefficients. Right: patches before and after Fourier reweighting. The constant (DC) component and high-frequency components are both suppressed by the training process.}
\label{fig:fcoef}
\end{figure}
\subsection{Aggregation Subnet}
The Aggregation Subnet performs target localization and hierarchical bounding-box regression.
In our design, each convolutional layer aggregates four channels that correspond to four adjacent patches, reducing the channel dimension by $4\times$. Each pooling layer has a kernel size of 2 and stride of 2, reducing spatial dimensions by $2\times$. The output feature map has a single channel that indicates similarity to the template, similar to correlation filters.
While there have been past works on localizing objects with just a few convolutional layers~\cite{valmadre2017end}, no existing shallow model, to our knowledge, is capable of bounding box regression. PatchNet presents a bounding box estimation strategy that works with just a few layers.
It is based on a simple observation that the patch distribution information can inform bounding box estimation, as shown in Figure~\ref{fig:toycase}. This information is hierarchically aggregated from smaller patches to larger patches, by convolution and modified max pooling.
Hierarchical bounding box regression predicts offsets of the bounding box boundary, represented by $(\Delta x_{min}, \Delta y_{min}, \Delta x_{max}, \Delta y_{max})$.
It is a three-stage procedure coupled with score aggregation. In each stage, it combines the previous stage's offset by a convolutional layer, plus the current stage's new incremental offset by the selection signal from the score pooling layer. Following this procedure, we hierarchically aggregate the per-patch offsets into box offsets.
\begin{figure}[b!]
\centering
\includegraphics[width=0.48\textwidth]{figures/AggregateBlock3.pdf}
\caption{The building block of aggregation subnet. The lower localization path is normal conv-pool layers, while the upper bounding box regression path actively adds the partial offset from the pooling stage.
Input offset to the first block is set to zero.
}
\label{fig:aggregate}
\end{figure}
Figure~\ref{fig:aggregate} shows the aggregation subnet's basic building block.
The aggregation block takes two tensors as input: the score tensor represents patch-wise similarity while the offset tensor represents patch-wise boundary offsets. As each offset consists of 4 scalars, the offset tensor has four times the number of channels as the score tensor.
Here the selection signal is the pooling indices. The intuition is that the pooling choice represents a patch shift, which contributes a learned offset to the bounding box.
This operation first selects the offset according to the pooled score tensor, and then adds the offset bias based on its relative location inside the pooling kernel. In practice, the selection is ``soft'', i.e., a weighted combination by the magnitude of scores, instead of ``hard'' selection by indices. Equation~\ref{eq:poolselect} describes the output offset of a pooling kernel $\mathcal{K}$, where $S$, $F$, $b$ are scores, offsets and bias, respectively.
\begin{align}
\begin{split}
\mathbf{f}_{out} = & \sum_{\mathbf{x} \in \mathcal{K}} \hat{S} [\mathbf{x}](F[\mathbf{x}] + b[\mathbf{x}]), \\
\quad & \text{where} \quad \hat{S} = softmax(S)
\end{split}
\label{eq:poolselect}
\end{align}
Learned aggregation is performed by convolutional layers. In the two parallel Conv2D layers, offsets and scores of neighboring patches are aggregated into larger patches, along with reduction in the channel dimension. The convolutional layer for scores is initialized with weights from the patch aggregation model. The layer for offsets is initialized so that the offset of a patch is averaged from four adjacent smaller patches.
The input offset to the first aggregation block is always set to zero.
The final output is a single-channel response map and a 4-channel offset map. In the inference stage, we select the peak in the response map and the corresponding offset to estimate the new bounding box.
\subsection{Training}
\textbf{Objective functions}.
We train PatchNet with a sum of localization loss and a bounding box regression loss. We adopt a simple L1 localization loss that penalizes large response scores at non-center locations. In SiamRPN, the localization task is trained as a binary classification problem with cross entropy. However, in our experiments, cross entropy loss leads to inferior localization accuracy for the shallow CNN model. Equation~\ref{eq:centerness} describes this localization loss, where the response map is denoted by $S$ and groundtruth center location by $\textbf{x}_{gt}$. $\mathcal{D}$ is the Manhattan distance. This loss function relaxes the penalty for locations close to the center, as neighboring regions typically respond stronger than backgrounds.
\begin{align}
\label{eq:centerness}
\mathcal{L}_{c}(\mathbf{S}, \mathbf{x}_{gt})= \sum_\mathbf{x} max (S[\mathbf{x}] - S[\mathbf{x}_{gt}] + \alpha \mathcal{D}(\mathbf{x}, \mathbf{x}_{gt}), 0)
\end{align}
For bounding box regression, the widely used smooth L1 loss is adopted, although it should be noted that output and groundtruth are both boundary offsets, different from typical bounding box regression.
\textbf{Structural sparsity constraints}.
$N\times N$ patches are stored along the same channel dimension for efficient implementation. However, in this approach we lose the patch adjacency information. Ideally, a 4D score tensor with 4D convolution would preserve the patch adjacency, but that is beyond the capability of most deep learning libraries.
An alternative is to enforce structural sparsity in the convolution filters. We set zeros for most of the channel-wise connections except those from adjacent patches. Our experiment shows this approach reduces the gap between the training and validation losses and improves tracking robustness.
\section{Experiments}
PatchNet solves the short-range template matching problem, which is exploited to speed up video processing.
In this section, we demonstrate the experimental results of PatchNet on video object detection (VID) and visual object tracking (VOT).
The PatchNet model is trained on the GOT-10k dataset~\cite{huang2019got}, following the training process in the SiamFC repo\footnote{ \href{https://github.com/huanglianghua/siamfc-pytorch}{https://github.com/huanglianghua/siamfc-pytorch}}. We use the same PatchNet weights for both VID and VOT benchmarks without retraining, showing that it is easy to generalize to different tasks and integrate with other models.
\subsection{Video object detection}
We select two object detection models, R-FCN ResNet-101 and EfficientDet-D0 as our baselines. R-FCN ResNet-101 is a high-precision detection model that is the baseline in many VID works~\cite{feichtenhofer2017detect,zhu2017deep}. For the ImageNet VID dataset, it consumes an average of 169GFLOPs per frame, which is computationally prohibitive on many edge devices. In contrast, EfficientDet-D0 is a state-of-the-art cost-effective detection model, consuming only 2.5GFLOPs per frame. The average number of objects per frame is 7 for R-FCN and 4 for EfficientDet on ImageNet VID.
We adopt a computation-reduction approach similar to Detect or Track~\cite{luo2019detect}. In this approach, the base detector runs on key frames and PatchNet tracks detected objects on non-key frames. The objects from key frames, together with the detection confidence scores, are preserved for a fixed number of frames unless they exceed the frame boundary.
Our results are compared against Deep Feature Flow (DFF) and other template matching methods, including full template correlation and SiamFC. DFF tries to alleviate expensive deep CNN feature extraction by propagating CNN features according to optical flow. DFF has extra overhead mostly from FlowNet, while PatchNet is much more compact. In addition, DFF requires joint training, while PatchNet does not need to be coupled with the detection model.
As shown in Figure~\ref{fig:dffvspatch}, our method generally achieves close or better speed-accuracy trade-off than other methods, especially at high amounts of computation reduction. Key results are reported in Table~\ref{tab:vid}.
\begin{figure}[h!]
\includegraphics[width=0.46\textwidth]{figures/dff_patch.pdf}
\vspace{-10pt}
\caption{Complexity/accuracy trade-off with R-FCN ResNet-101 as backbone on VID dataset. Unlike DFF, PatchNet does not require retraining on video dataset, while still achieving significant speed-ups.}
\label{fig:dffvspatch}
\end{figure}
\begin{figure}[h!]
\includegraphics[width=0.46\textwidth]{figures/interval_speed.pdf}
\vspace{-10pt}
\caption{Projected relationship of key frame interval and model complexity reduction. PatchNet enables significant FLOP reduction for both models.}
\label{fig:interval_flops}
\end{figure}
\begin{table}[h!]
\centering
\begin{tabular}{cccccc}
\toprule
Model & Method & mAP & FLOPs & Reduction\\
\midrule
\multirow{2}{*}{R-FCN }& Baseline & 0.738 & 169G & \textbackslash \\
\multirow{2}{*}{ResNet-101} & DFF & 0.725 & 34.9G & 4.8$\times$ \\
& PatchNet & 0.731 & 34.2G & \textbf{4.9$\times$} \\
\midrule
EfficientDet & Baseline & 0.597 & 2.5G & \textbackslash \\
-D0 & PatchNet & 0.589 & 0.73G & \textbf{3.4$\times$} \\
\bottomrule
\end{tabular}
\caption{PatchNet achieves better FLOPs/mAP trade-off than DFF. DFF's results are obtained with model checkpoints provided in the official GitHub repo.}
\label{tab:vid}
\vspace{-5pt}
\end{table}
\begin{table*}[h!]
\centering
\begin{tabular}{cccccc}
\toprule
Baseline model & Keyframe interval & Success score & Precision score & FLOPs & Reduction\\
\midrule
& 1 & 0.575 & 0.766 & 2.7G & \textbackslash \\
SiamFC & 2.7 & 0.573 & 0.784 & 1.1G & 2.5$\times$\\
& 3.6 & 0.562 & 0.776 & 0.8G & 3.2$\times$\\
\midrule
& 1 & 0.630 & 0.837 & 4.1G & \textbackslash\\
SiamRPN-2x & 2.7 & 0.629 & 0.839 & 1.6G & 2.6$\times$ \\
& 4.2 & 0.620 & 0.841 & 1.1G & 3.9$\times$\\
\bottomrule
\end{tabular}
\vspace{5pt}
\caption{Results on OTB2015 dataset, with SiamRPN and SiamFC as baseline models. The SiamRPN-2x model has 2x channels compared to the original paper. When keyframe interval is 1, we run the default baseline model.}
\label{tab:vot}
\vspace{-5pt}
\end{table*}
We implemented EfficientDet for ImageNet VID and show that PatchNet can still yield meaningful speed-ups.
DFF for EfficientDet is not reported in Table~\ref{tab:vid}, because even the smallest FlowNet-Inception is about the same complexity as EfficientDet-D0. On contrary, PatchNet can speedup the very compact EfficientDet-D0 almost linearly, as indicated in Figure~\ref{fig:interval_flops}.
\subsection{Visual object tracking}
We select SiamFC and SiamRPN as our baseline models. SiamFC is the predecessor of recent advances of Siamese models for VOT. It has an AlexNet backbone for feature extraction and computes similarity by cross-correlation. Due to lack of bounding box regression, multi-scale evaluation is necessary to handle scale change. SiamRPN adds a RPN head for bounding box regression, eliminating multi-scale evaluation while improving tracking robustness. The SiamRPN model \footnote{\href{https://github.com/huanglianghua/siamrpn-pytorch}{https://github.com/huanglianghua/siamrpn-pytorch}} uses double-channel AlexNet as the backbone, therefore we denote it by SiamRPN-2x throughout this section.
We test PatchNet with the skip-frame scheme. Key frames are selected in an online manner: when PatchNet predicts a confidence score lower than a threshold or when the number of inter-frames exceeds a limit.
Our results are reported in Table~\ref{tab:vot}. The skip-frame method reduces average workload of both SiamFC and SiamRPN by about 3x. For both models, the success score drops by about 1 percent while the precision score is slightly improved.
We also evaluate our methods on UAV123 ~\cite{mueller2016benchmark}, a high-resolution aerial-view dataset that is larger and more realistic than the previous OTB2015 dataset. As shown in Table~\ref{tab:uav_exp}, our skip-frame method maintains the success score and slightly improves the precision score when we run the base model every 3 frames. With a tolerance of 1 percent accuracy drop, we can further increase the keyframe interval to 5. The results show that PatchNet can generalize to different datasets without additional training.
Table~\ref{tab:vot_comparison} shows PatchNet's advantage as an inter-frame tracker compared with other methods. All methods have key-frame intervals of roughly 4 frames. The dummy tracker does not predict any new location in the inter frames, therefore suffers from significant accuracy drop. Kernelized Correlation Filter (KCF)~\cite{henriques2014high} is a very fast correlation filter method, but it has inferior accuracy compared with PatchNet. In our experiment we use OpenCV's KCF implementation. Patch Aggregation is PatchNet without Fourier reweighting and bounding box regression. It achieves performance intermediate between KCF and the full PatchNet.
\begin{table}[h!]
\begin{tabular}{cccc}
\toprule
Keyframe & \quad Success \quad \quad & Precision \quad & Reduction\\
interval & score & score & \\
\midrule
1 & 0.599 & 0.770 & \textbackslash \\
3 & 0.600 & 0.780 & 2.8$\times$ \\
5 & 0.589 & 0.777 & 3.6$\times$ \\
\bottomrule
\end{tabular}
\caption{Results on UAV dataset with SiamRPN-2x as the baseline model.}
\label{tab:uav_exp}
\end{table}
\begin{table}[h!]
\begin{tabular}{cccc}
\toprule
Skip-frame & Keyframe & Success & Precision \\
method & interval & score & score \\
\midrule
Original & 1 & 0.630 & 0.837 \\
Dummy & 4.0 & 0.204 & 0.152 \\
KCF & 3.7 & 0.584 & 0.770 \\
Patch Aggregation & 3.7 & 0.609 & 0.815 \\
PatchNet & \textbf{4.2} & \textbf{0.620} & \textbf{0.841} \\
\bottomrule
\end{tabular}
\caption{Comparison with other choices for the skip-frame method. Results obtained on OTB2015, with SiamRPN-2x as the keyframe model.}
\label{tab:vot_comparison}
\end{table}
\subsection{Ablation study}
We analyze the design choices that affect PatchNet's ability to capture short-range correspondence, as measured by the miss rate of groundtruth objects that are 10 frames apart on the ImageNet VID dataset. An IoU score below 0.7 is decided as a miss, which is stricter than the object detection problem. We name this experimental setting as short-range ImageNet VID benchmark.
\textbf{Patch size} affects the representation and generalization power of PatchNet. A patch size of 1 is unlikely to work, as pixel correlation does not yield meaningful information; on the other hand, large patch sizes will incur extra computation overhead, which is unwanted for edge applications. A patch size as large as the object will convert PatchNet to a correlation filter method.
Figure~\ref{fig:patchsize} shows the impact of patch size on accuracy and complexity. The number of patches is fixed to keep the aggregation subnet unchanged. Larger patch sizes yields diminishing return on accuracy, while greatly increasing computation cost.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figures/ablation_patchsize.pdf}
\vspace{-10pt}
\caption{Accuracy and complexity of different patch sizes on the short-range ImageNet VID benchmark.}
\label{fig:patchsize}
\end{figure}
\textbf{Fourier reweighting} and \textbf{hierarchical bounding box regression} are two novel components in PatchNet. Table~\ref{tab:ablation_component} shows their impact on accuracy and overall computation cost.
Starting from Patch Aggregation, Fourier basis weighting gives a 1.7\% miss rate improvement while increasing computation cost by only 1 MFLOPs. Such a small overhead is attributed to Fast Fourier Transformation for base-2 patches.
Hierarchical bounding box regression, on the other hand, gives a 3.2\% miss rate improvement while increasing computation footprint by 12.4 MFLOPs.
\begin{table}[t!]
\centering
\begin{tabular}{ccc}
\toprule
Method & Miss rate & FLOPs\\
\midrule
Patch Aggregation & 0.208 & 44.2M \\
w/ Fourier reweighting & 0.199 & 45.2M \\
w/ BB regression & 0.184 & 56.6M \\
PatchNet & 0.167 & 57.6M \\
\bottomrule
\end{tabular}
\caption{Analysis of Fourier reweighting and hierarchical bounding box regression on the short-range ImageNet VID benchmark.}
\label{tab:ablation_component}
\end{table}
\subsection{Speed on edge device}
We evaluate PatchNet on an NVIDIA Jetson Nano to measure its performance on a typical edge device.
Efficient model execution on mobile platforms has become a focus in many recent works~\cite{howard2017mobilenets,tan2019efficientnet}. In this section, we demonstrate that PatchNet is able to accelerate all aforementioned models on a mobile platform.
All models listed in Figure~\ref{fig:speed_flops} are tested with PyTorch-1.4 on Jetson Nano with a quad-core ARM Cortex-A57 CPU and an 128-core Maxwell GPU, with a batch size of 1.
As shown in the figure, PatchNet has orders of magnitude lower on-device latency and consumes orders of magnitude fewer FLOPs than the other models.
In Table~\ref{tab:nano}, we list the speed-ups in accordance to Table~\ref{tab:vid} \& \ref{tab:vot} in previous sections. PatchNet achieves 2.8 - 4.3$\times$ speedup on four listed models (excluding data IO time). Model optimization with TensorRT may further improve the speed, as already shown on many edge device applications~\cite{hadidi2019characterizing,jo2020benchmarking}.
\begin{figure}[t!]
\centering
\includegraphics[width=0.47\textwidth]{figures/speed_vs_flops.pdf}
\vspace{-5pt}
\caption{Single-frame inference speed on Jetson Nano. PatchNet is orders of magnitude faster than other models, making it an ideal inter-frame model. }
\label{fig:speed_flops}
\end{figure}
\begin{table}[t!]
\centering
\begin{tabular}{cccc}
\toprule
\multirow{ 2}{*}{Model} & Baseline & Skip-frame & \multirow{ 2}{*}{Speed-up} \\
& (ms) & (ms) & \\
\midrule
R-FCN & \multirow{ 2}{*}{1094} & \multirow{ 2}{*}{254} & \multirow{ 2}{*}{4.3$\times$} \\
ResNet101 & & & \\
EfficientDet-D0 & 227 & 56 & 4.1$\times$ \\
\midrule
SiamFC & 114 & 52 & $2.8\times$\\
SiamRPN-2x & 232 & 65 & 3.6$\times$ \\
\bottomrule
\end{tabular}
\vspace{5pt}
\caption{Model time and speed-ups achieved by PatchNet on Jetson Nano. Keyframe interval is 5 for video object detection, and 3.6 \& 4.2 for SiamFC \& SiamRPN-2x, following previous experimental settings.
}
\label{tab:nano}
\end{table}
\section{Conclusion}
This paper describes PatchNet, a very small neural network model for template matching in short-range video frames, and demonstrates its effectiveness in video object detection and visual object tracking. On both tasks, PatchNet achieves 3-5x speedup with little accuracy drop without retraining the model on the target dataset. As recent advances drive the need for fast video understanding on mobile platforms and compact models like EfficientDet-D0 are developed, previous heavyweight temporal acceleration methods find difficulty achieving speedup. PatchNet provides a viable solution that offers significant and effortless speed-ups on video tasks.
{\small
\bibliographystyle{ieee_fullname}
|
\chapter{Introduction}
\pagenumbering{arabic}
\hspace{0.5cm} Material processing by laser interactions has attracted a wide interest both in scientific and engineering communities. The possibilities in micro- and nano-structuring utilizing laser energy density to make material changes or damages \cite{nolte1997ablation} enable a tremendous number of applications, such as optical data storage\cite{glezer1996three,watanabe1998three,juodkazis2003recording,canioni2008three}, waveguides \cite{szameit2010discrete,davis1996writing,miura1997photowritten,nolte2003femtosecond,eaton2011high}, Bragg gratings \cite{gattass2008femtosecond,marshall2006direct,zhang2007single,marshall2010point, zhang2018efficient}, volume gratings \cite{klyukin2017volume,klyukin2018thermal}, photonic quantum circuits \cite{marshall2009laser,matthews2009manipulation,aspuru2012photonic}, micro-welding \cite{tamaki2005welding,horn2008investigations}, metasurfaces for light field manipulation \cite{drevinskas2017ultrafast, wang2016optically}, micromechanics \cite{maruo2003submicron}, microfludic devices \cite{bellouard2004fabrication,wu2009femtosecond,amato2012integrated}, and biological channels \cite{wu2015channel,kim2009single,sugioka2012femtosecond}. These expanded applications stem from the fundamental progresses and improvements in high power lasers. In 1987, the finding of laser ablation without heat-affected zone by ultrafast lasers whose pulse duration was less than a few picoseconds had an important impact on the researches working in the field of laser interactions with materials \cite{srinivasan1987ablation,sugioka2014ultrafast,kuper1987femtosecond}. Since then, and, importantly, because of the invention of the chirped-pulse amplification in 1985 by D. Strickland and G. Mourou \cite{strickland1985compression} and further development of this technique \cite{rudd1993chirped}, high quality micro- and nano-fabrication attracted lots of attentions for the applications in semiconductors and dielectrics. The pioneering work for producing optical waveguides inside a glass via ultrafast lasers was reported in 1996 by K. Davis et al. \cite{davis1996writing}. In fact, at attempt to fabricate such optical elements by ultra-violet(UV) light was first performed much earlier, in the 1970s \cite{hill1978photosensitivity}. However, it was until the development of the high energy density pulsed lasers that a plausibility to greatly increase the refractive-index changes appeared due to multiphoton absorption. Other improvements were proposed to increase the spatial resolutions in the three-dimensional micro-fabrications. Nowadays, the two-photon polymerization enables the three-dimensional fabrications of objects with dimensions down to sub-100 nm \cite{serbin2003femtosecond,emons2012two,sun1999three}.
Furthermore, the laser-induced periodic surface structures (LIPSS) emerged as another useful surface nano-structuring technique. The effect was firstly reported in 1965 by M. Birnbaum during examining the surface damage on semiconductors (germanium, silicon, gallium arsenide, gallium antimonide, and indium antimonide) by a focused Ruby laser \cite{birnbaum1965semiconductor}. The parallel straight lines (or called grooves) were observed at laser powers lower than the formation of crack patterns. Later, this effect was also observed for pulsed lasers and it turned out that the laser wavelength, fluence, polarization, and number of irradiated pulses greatly affected the LIPSS generation\cite{van1982laser,bonse2012femtosecond,bonse2016laser}. Almost in the same time, a number of theoretical studies were focused on the understanding of the LIPSS formation \cite{emmony1973laser,keilmann1982periodic,sipe1983laser,dufft2009femtosecond,bonse2009role,rudenko2016random,rudenko2019amplification}. The applications of these structures were widely examined by both scientific and engineering communities. One of the direct results was that the laser generated grooves served as diffractive optical elements manipulating the optical spectra. These nano-gratings were controlled by laser polarization, wavelength, fluence, and scan speed, which was useful for color marking, encryption, and optical data storage \cite{dusser2010controlled,vorobyev2008colorizing,ahsan2011colorizing,ou2014colorizing,yao2012selective}. Moreover, the contact angle was increased after the formation of LIPSS, so that superhydrophobic surfaces were produced \cite{barberoglou2009bio,kietzig2009patterned,kietzig2011laser}. In addition, the LIPSS were capable to increase the tribological performances \cite{yu1999laser,yasumaru2011frictional,wang2015reduction}.
Apart from the laser processing of solid or bulk materials, their porous counterparts have attracted more and more interest in the past decades. Comparing to the bulk glasses, such as fused silica and BK7, porous glasses (typical pore sizes range from 2 to 20 nm \cite{toquer2011effect,veiko2016femtosecond,burchianti2006reversible}) were easier to be densified by lasers and impregnated of nanoparticles or dyes, so that they presented promising applications in the field of optical filters \cite{alexeev1993porous,sendova2012plasmonic}, microfluidic channels \cite{liao2010three,ju2011fabrication,sugioka2012femtosecond}, and optical waveguides \cite{antropova2012peculiarities,simmons1979optical}. For example, the refractive index changes in a bulk glass were normally in the range of ($\Delta n \approx 10^{-3}$ to $10^{-2}$ \cite{eaton2011high}), which were realized because of the local mass-density or material changes after ultra-short laser induced heating or shock-waves. The solidification and decompaction in bulk materials were restricted owing to their intrinsic physical and chemical limitations; in contrast, the pores and their random connected networks inside porous glasses allowed larger changes in local mass density and in refractive index ($\Delta n \approx 0.12$) \cite{veiko2016femtosecond,simmons1979optical}. To date, most microfluidic devices were two-dimensional that possessed limited applications for integrated devices \cite{liao2014femtosecond}. A recent developed technique has enabled the fabrication of 3D microfluidic structures in porous glass with arbitrary sizes and geometries \cite{liao2010three,ju2011fabrication,liao2012rapid,ju2013direct}. The process was based on an ultra-fast laser writing of porous glass immersed in water followed by a post-annealing to collapse nanopores. Based on this technique, a square wave-shape channel\cite{liao2010three}, a large volume microchamber\cite{ju2011fabrication}, and a 3D microfluidic mixer inside a glass chip\cite{liao2012rapid} were successfully fabricated. Furthermore, V. P. Veiko et al. used ultrafast lasers to process porous glass and examined densification, decompatction, void formation regimes as a function of pulse energy and scanning speed \cite{veiko2016femtosecond}. These processes accompanied by heat diffusion and multi-pulse accumulation, as well as nonlinear light propagation, absorption and plasma scattering. Thus, it was hard to explain the obtained results without a detailed modeling.
In addition, mesoporous (typical pore size ranges between 2 and 50 nm) glasses loaded with nanoparticles served as emerging platforms for quantum-dots applications \cite{litvin2013pbs,letant2006semiconductor}, optical recording \cite{royon2010silver},nano-wire synthesizing \cite{huang2000ag}, Q-switched laser \cite{zhang2019method}, light-controlled atomic dispensers \cite{burchianti2004light}, and drug delivering \cite{li2012mesoporous,slowing2007mesoporous} as their strong capabilities in absorbing nanocomposites. The kinetics of the embedded nanoparticles enabled additional possibilities in laser processing of porous glasses and, particularly, in a better control over their properties. It was shown by L. Marmugi and E. Mariotti et al. \cite{marmugi2014laser} that the adsorption/desorption and nanoparticle formation in nanoporous glasses were strongly affected by the laser irradiation. A properly chosen laser wavelength and illumination time was able to control the concentration of the formed NPs. In those materials loaded with nanopartilces, the heat transfer at nanometric scales differed from macroscopic scales was of great interest. Using the time-resolved optical pump-probe spectroscopy, N. D. Fatti et al. studied the thermal interface resistances of metal nanoparticles embedded in a homogeneous matrix \cite{gandolfi2018ultrafast, banfi2012temperature, stoll2015time}. Furthermore, the presence of silver nanoparticles in soda-lime glasses allowed the possibility of generating self-arranged periodic Ag nanoparticles by an intense UV laser\cite{tite2017ag,goutaland2013laser}. In this case, the LIPSS were formed under laser exposure before any surface damages appeared. Recently, a two-beam interference of a picosecond laser shining on the surface of a silver-nanocomposite porous glass was shown to cause grating formation and its transformation to grooves at higher laser fluences \cite{andreeva2020picosecond}. At lower fluences, the structures were parallel to the interference pattern, while they were converted to the grooves orientated parallel to laser polarization as soon as distribution were generated at high fluences.
Despite the laser processing of solid-state materials, laser ablation in liquids (LAL) and LAL-based nanoparticle synthesis in solutions were extensively studied during the past decades \cite{geohegan1998time,kabashin2003synthesis,amendola2009laser,hashimoto2012studies}.
To better control nanoparticle size distribution, laser-induced nanofabrication in liquids with and without following laser-induced colloidal fragmentation was proposed. For example, in 2003, A. V. Kabashin et al. \cite{kabashin2003synthesis} reported synthesis of colloidal nanoparticles during femtosecond laser ablation of gold in water. It was found that the nanoparticle size was small (3-10 nm) at low laser fluences (<400$J/cm^2$) while large nanoparticles with a broad size
as soon as the fluence exceeded the identified threshold.
Unlike porous glasses that only acted as the serving platforms, porous semiconductors interacted with the encapsulated nanocomposites providing additional controlling degrees of freedom in transparent conductors \cite{granqvist2007transparent,han2008improved,liu2010transparent}, photoconduction \cite{hoyer1995photoconduction,zaban1998photosensitization,kamat2010beyond}, and photocatalytics \cite{chen2011new,wang2012plasmonic,awazu2008plasmonic,chen2010plasmonic,zheng2007ag,georgekutty2008highly}. In those hosting media, the most studied semiconductors were TiO$_2$ and ZnO as they presented significant photoactive behaviors, and were naturally abundant and chemically stable. The ability to generate conducting electrons and holes after absorbing photons could activate the oxidation and reduction of several chemical compounds. However, the bulk semiconductors were inefficient in the photocatalytic process as plenty of conducting electrons or holes were reduced by recombination during their migrations to the material surfaces before interacting with chemicals in the vicinity \cite{hashimoto2005tio2,zhang2013plasmonic,wang2012plasmonic,awazu2008plasmonic,chen2010plasmonic}. In contrast, the porous state of materials having larger contacting surfaces with the environment and smaller recombination probability could boost the photoactivated process. The dramatic boosting was observed by a recently proposed idea of the so-called "plasmonic photocatalysis" \cite{awazu2008plasmonic,kubacka2011advanced,chandrasekharan2000improving,kamat2002photophysical,hirakawa2005charge,zhao1996sol}, which encapsulated nanoparticles of noble metals such as Au and Ag into these semiconductors. The photocatalysis benefited from the light absorption in visible and the nearfield enhancement due to the surface plasmon resonances of Au and Ag nanoparticles. Typically, TiO$_2$ and ZnO had relatively large band gaps ($E_g = 3.2 eV$), so that they only absorbed the UV light. The plasmonic absorption by metallic nanoparticles occurs across the visible to infrared and the induced free electrons were found effectively injected into the nearby semiconductors of suitable band gaps \cite{mongin2012ultrafast,tian2005mechanisms,hirakawa2005charge, yu2006size,takahashi2011solid}. In addition, the contacting of a noble metal with a semiconductor formed the Schottky junction that efficiently separated the electrons and holes to avoid the recombination by taking advantages of the internal electric field in a certain region. After the process, silver nanoparticles were positively charged and tended to dissolve into Ag$^+$ ions to keep the neutral state. On the other hand, electrons diffused from TiO$_2$ films and accepted by the Ag$^+$ ions thanks to the photocatalytic process when illuminated by the UV light. The nucleation increased the quantity of monomers in the matrix (hosting medium), which resulted into the growth of nanoparticles by Ostwald ripening or coalescence. Because of the size-dependent optical responses of Ag nanoparticles, the spectra or colors of these samples were reversible. This effect was recognized as "photochromism" \cite{ohko2003multicolour,naoi2004tio2,naoi2005switchable,diop2015magnetron}, which has numerous applications in optical data storage \cite{wang2011optical,rakuljic1995optical}, and, particularly, in the re-writable colored images \cite{yokoyama2000fulgides}.
In the past decades, metamaterials composed of artificially designed structures had attracted tremendous interests, due to their abilities of engineering the optical amplitude, phase, and polarization at will that not found in nature. The applications were wide and not limited to electromagnetic/optical cloak \cite{schurig2006metamaterial,cai2007optical,valentine2009optical,ergin2010three}, aberration-free lens \cite{aieta2013aberrations,loo2012broadband,ma2010broadband}, ultra-thin lens \cite{chen2012dual,pors2013broadband}, metasurface based holography \cite{yu2011light,huang2015aluminum,zheng2015metasurface,ni2013metasurface,huang2013three}, polarization imaging \cite{arbabi2018full,li2015circularly}, and chirality-dependent pulse modulation \cite{yan2018chirality,yan2019electromagnetically}. However, these structures were normally in sub-wavelength scale and required high-precision technologies such as electron beam lithography, focused ion beam, and nanoimprinting that were either expensive or unsuitable for large-scale productions. In recent years, the "NANOPARTICULE" team of Laboratoire Hubert Curien developed a cost-effective way of producing nanoparticles and nano-gratings by using the continuous-wave and pulsed-lasers to direct write mesoporous TiO$_2$ films loaded with Ag nanoparticles \cite{crespo2010reversible,crespo2012one,destouches2014self,baraldi2016polarization,destouches2013photo}. The applications in color printing \cite{diop2017spectral} and image multiplexing \cite{sharma2019tailoring,sharmalaserdriven2019} were demonstrated.
The laser irradiation of the semiconductor loaded with Ag nanoparticles was identified as a complex process consisting of light absorption at plasmonic resonances, photo-oxidation, reduction, and nanoparticles growth. Based on a simplified model without heat diffusion, the writing-speed threshold of dramatic growth of Ag nanoparticles was understood \cite{liu2015understanding,liu2016selfthesis}. Below the certain writing-speed, only small nanoparticles were observed. Nevertheless, the formation of Ag nanoparticles and nanogratings as a function of laser writing-speed was not well understood.
In addition to numerous experimental studies, a series of theoretical and numerical models were developed to better understand the physical and chemical process involved. First part of the proposed models considered lasers just as heat sources, which can be valid for rather long laser pulses (nanosecond and longer) and for a continuous-wave laser irradiation \cite{tian2004temperature,tao2009thermal}. Several studies also examined ultra-short laser interactions \cite{couairon2002light,sprangle2002propagation,miyamoto2011evaluation,sun2013role,couairon2007femtosecond, rudenko2016random,gulley2012interaction,wu2005femtosecond, chekalin2015light}. In particular, T. Brabec and F. Krausz (1997) \cite{brabec1997nonlinear}, A. Couairon (2002) \cite{couairon2002light}, P. Sprangle (2002) \cite{sprangle2002propagation} and A. L. Gaeta (2000) et al. \cite{gaeta2000catastrophic} developed an envelope equation modeling the propagation of ultra-short pulses by assuming that the pules envelope is slowly changing along the direction of light transportation. For multiple pulses (repetition rate of 100 kHz), I. Miyamoto et al. proposed a simplified thermal conduction model and simulated the picosecond laser modifications of bulk glasses. By comparing with experiments at different laser energy and repetition rate, the nonlinear absorptivity was obtained and used for the thermal diffusion model. It was concluded that the laser energy was absorbed through avalanche ionization seeded by thermally excited free-electrons \cite{miyamoto2011evaluation}. In addition, for inertial confinement fusion \cite{duchateau2009simple}, G. Duchateau et al. proposed a model considering the solid-to-plasma transition to better simulate the laser-plasma coupling. It turned out that around 100 picoseconds were required to induce a full plasma state, suggesting that the process of solid-to-plasma transition could not be ignored for inertial confinement fusion applications.
To date, very limited numerical studies have been performed in understanding the laser interactions with porous materials \cite{liao2015high,antropova2012peculiarities,li2004comparison, rudenko2016random}.
In which, most studies \cite{antropova2012peculiarities,li2004comparison} were concentrated on a continuous wave laser irradiation.
For ultra-fast lasers, Y. Liao et al. analyzed the morphological features of nanogratings in porous glass and concluded that the formed nanovoids were a manifestation of spherical nanoplasma produced by multiphoton ionization \cite{liao2015high}. By assuming the nanoplasma with a size of tens of nanometers, they performed a finite difference time domain simulation to explain the formation of subwavelength-width cracks. In 2016, A. Rudenko et al. proposed a more rigours model of ultrafast laser-plasma coupling and simulated the evolution from small spherical nanometric inhomogeneities to periodical nanoplanes \cite{rudenko2016random}.
Furthermore, laser-induced nanoparticle formation was modeled in several studies \cite{volkov2007numerical,boyer2012modeling,delfour2015mechanisms,liu2015understanding, liu2017three}.
Many of the performed investigations considered that nanoparticles were generated in gases or in liquid solutions. In these studies, thermal effects, nucleation and growth of nanopartilces, laser absorption and hydrodynamics were discussed. There is, however, still a lack of coupled multi-physical models and simulations suitable for prediction of the nanoparticle size or their formation regimes.
Recently, experiments were performed by shining a scanning continues wave laser on porous TiO$_2$ films containing Ag ions and nanoclusters. Interestingly, the resulting nanoparticles are found to be well-organized into nanograting-like periodic structures enabling unique optical properties and several promising applications, namely for imaging and security. To explain nanoparticle formation, a numerical approach was also proposed explaining only a part of the observed effects \cite{destouches2014self,baraldi2016polarization}. In 2015, Z. Liu et al. \cite{liu2015understanding} developed a model considering size shrinkage by photo-oxidation, growth via Ostwald ripening, and Ag$^+$ reduction into free Ag$^0$ to explain the laser-writing speed related growth. This model was based on Kumar's nanopaticle kinetics model \cite{kumar2008efficient} and was inspired by the nanoclusters formation modeling in glass performed by Kaganovskii et al. \cite{kaganovskii2007formation} In these previous models, however, temperature field was not well calculated and coupled to nanoparticle growth taking place in time and space. For instance, such important process as heat diffusion was completely disregarded and only single nanoparticles were taken into account for in the laser-induced heating.
Despite the limitations, previously studies clearly confirmed that laser interactions with materials containing nano-clusters was a set of complicated process involving many electromagnetic, chemical and physical mechanisms taking place at very different time and space scales. In particular, light absorption and thermal diffusion were important in the process of nanocluster evolution. Therefore, much more rigours models were required for well understanding the involved processes and providing guidelines for better controlling laser-induced micro- and nano-structures. There were many open questions, namely concerning the role of heat diffusion, the correct description of nanoparticle sizes and the related absorption maps, the explanation of several anti-intuitive effects, such as temperature rise with increasing laser scanning speed, etc.
This thesis was performed in the "LASERMODE" team at Hubert Curien Laboratory, UMR CNRS 5516/ Univ. Lyon/ Univ. Jean Monnet in Saint-Etienne, France (LabHC) in a close cooperation with the experimental "NANOPARTICULE" group in the same Laboratory. A part of the experiments was also carried out in The Laser Micro- and nano- structuring Laboratory of the ITMO University (Saint-Petersburg, Russian Federation) in the frame of the PHC KOLMOGOROV Formalas project operated by Campus France. The first motivation was to better understand the laser micro-processing of porous glass by pulsed-lasers. Secondly, since the laser nano-processing was important and interesting as it acted as an emerging promising and cost-efficient technique for constructing nanoscale building blocks for light manipulation, models for laser processing of mesoporous TiO$_2$ films load with Ag nanoparticles were developed. Then, simulations were carried out to better understand the involved processes and, thus, better control laser treatment, and, thus, to get the desired sub-wavelength structures.
The thesis was financed by the French Ministry of Science and Education and Jean Monnet University. Several missions were supported by the Laboratoire Hubert Curien and by the FormaLas project. Numerical simulations were performed on the local cluster at Hubert Curien Laboratory and the CINES supercomputer in France.
After the introduction presented in this Chapter, the thesis is organized as follows:
Chapter 2 discusses the results of a femtosecond laser inscription of periodic void arrays in porous glass. The periodic voids and high aspect ratio densification were presented. The formation of the periodic void structures were shown to rely on laser parameters such as deposited energy and scanning speed. A thermodynamic analysis was performed to explain the linear relation of period and number of applied laser pulses per spot.
Chapter 3 is focused on the continuous-wave laser writing of silver nanoparticles inside mesoporous TiO$_2$ thin films. A coupled model of nanoparticles growth, photo-oxidation, reduction, and heat diffusion was proposed to understand the writing-speed dependent phenomena. The spatial size distribution was revealed to be non-uniform that resulted into the transmission inhomogeneity during laser scanning the sample, which was shown in accordance with in-situ experiments. The performed study also depicted a novel view that Ag NPs grew ahead of the beam center due to an expanded temperature field. The reduction process was insignificant due to the low concentration of the reducing agents. Exhaust simulations were performed on the study of parameters such as activation energy of Ag diffusion, photo-ionization rate, and initial concentration of Ag$^+$ and Ag$^0$ in the mesoporous TiO$_2$. The phenomenon of temperature increasing with writing speed was discussed. Simulations showed the same trend with the analysis of the phase transition of TiO$_2$. The multi-scale model was simulated in three-dimensional and the results were compared with the in-situ transmission maps.
Chapter 4 considers femtosecond laser writing of Ag nanoparticles in mesoporous TiO$_2$ thin films. Compared to the continuous-wave cases, the ultra-fast lasers provided additional possibilities in generating surface grooves on TiO$_2$. The laser propagation was simulated based on the nonlinear Schr\"{o}dinger equation coupled with plasma equation. Due to the low pulsed-energy and loosely focusing conditions, the laser deposited energy was shown to be negligible for the studied laser setups. Then, the heat accumulation models were developed that were coupled with silver nanoparticles growth were simulated, which showed similar trends with the experiments.
Chapter 5 presents general conclusions and perspectives of this work.
\afterpage{\null \newpage}
\chapter{Femtosecond laser inscription of periodic void structures and high aspect ratio densification in porous glass}
\hspace{0.5cm} This chapter reports a thermodynamic analysis of ultra-fast laser induced periodic voids and high aspect ratio densification in porous glass. Previously, the femtosecond laser irradiation of porous glass, leading to stress-free densification, decompaction and void formation were demonstrated as a function of pulse energy and laser writing speed \cite{veiko2016femtosecond}. It lacked, however, well understanding of the multi-physical processes usually accompanied by nonlinear propagation of pulsed beams, light absorption, heat diffusion, and hydrodynamics. Furthermore, the formation of perfectly controlled periodic voids structure was shown to rely on laser energy per pulse and writing speed \cite{ma2017well}, which remained unclear. For this sake, a model based on the nonlinear Schr\"{o}dinger equation coupled with electron plasma generation is proposed to estimate the absorption of each pulse, which was then used for the multi-pulses heat accumulation. The estimated dimensions by temperature simulations are well correlated with the obtained shapes of the densified regions. The formations of the periodic voids can be understood by a superposition of two energy sources: the moving laser and the created static heat source. The model can by used by coupling hydrodynamics in the future to have a better understanding of the involved process.
\newpage
\section{Introduction}
The capacity of femtosecond laser systems to locally modify transparent materials has been used for many promising applications in different areas ranging from photonics to microfluidics \cite{gattass2008femtosecond}. The main advantage of the ultra-short laser pulses is in an extremely high precision of the energy deposition. Therefore, laser machining can be performed in volume of different glasses enabling three-dimensional inscription of numerous structures, such as photonics crystals, optical memories, waveguides, gratings, couplers, chemical and biological membranes and other devises \cite{graf2007pearl,bellouard2011femtosecond}. Previous studies have already revealed major mechanisms of femtosecond laser irradiation of typical glasses, such as fused silica, BK7 and calcium fluoride \cite{schaffer2001laser, stuart1995laser, rethfeld2006free, taylor2008applications, richter2013formation, bulgakova2015modification}. Such phenomena as photo-ionization and avalanche ionization leading to laser-induced breakdown were in focus of both experimental and theoretical studies starting from the invention of high-power systems \cite{bloembergen1997brief,schaffer2001laser}. It was shown that very small regions could be efficiently treated in a well-determined way by using only a central part of the Gaussian radial distribution of the laser beam \cite{joglekar2004optics}.
The results of laser machining can considerably differ if instead of a single laser pulse, trains of high repetition rate ultra-short laser pulses are used \cite{schaffer2001micromachining, miyamoto2011evaluation}. In fact, when time between laser pulses is shorter than that required for energy diffusion outside laser-treated volume, a considerable heat accumulation may lead to a pronounced enhancement in the size of the laser-affected zone. This effect was observed in the experiments with nanojoule laser energy at laser repetition rate as high as $25$MHz. For such multi-pulse laser irradiation regime Miyamoto et al. \cite{miyamoto2011evaluation} calculated thermal fields revealing stationary regimes for several optical materials and laser parameters. Moreover, it was shown that even upon a single ultra-short laser interaction with glass, the induced thermal effects could last for a significant time up to nanoseconds in dielectric materials \cite{veiko2014effective}.
Additionally, femtosecond laser-induced formation of nanopores and voids was investigated for the initially non-porous optical materials \cite{watanabe2000optical, juodkazis2006laser, bellouard2011femtosecond, lancry2013ultrafast}. Such effects as strong explosion and shock wave formation were shown to play a role in a single void formation for tightly focusing conditions \cite{gamaly2006laser, hallo2007model, mezel2008formation, bulgakova2015modification, rapp2016high}, whereas cavitation and nanopores were found to play a role in volume leading to nanograting formation when laser focusing was not so strong \cite{lancry2013ultrafast, rudenko2016random, cerkauskaite2017ultrafast}.
It was demonstrated, furthermore, that silica glasses with different densities including fused silica, porous glass, and aerogel could be also successfully processed \cite{veiko2016femtosecond, cerkauskaite2017ultrafast}. For porous glasses, several regimes were identified \cite{veiko2016femtosecond} with different corresponding types of material modifications ranging from densification and microvoid formation to wider channels appearance. In fact, laser-induced densification is expected in the regime of lower energy deposition, whereas the formation of channels was reported for larger laser absorbed energy. For intermediate absorbed energy, the so-called "decompaction" regime was identified accompanying by microscopic void formation during transverse laser scanning of porous glass in volume. We note that local control of porosity also allows the formation of waveguides and other embedded elements. Additionally, porous materials can be used for laser-assisted nanoparticle formation with promising photonic, plasmonic or photochromic properties. These investigations also open up a way toward an easier fabrication of optical sensors based on porous materials to monitor changes in the environment.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.6\textwidth]{./figuresChapter4/figure_1.pdf}
\caption{Illustration of transverse laser scanning geometry (a) and the final structure (b).}
\label{fgr:a}
\end{figure}
Previously, ultra-short lasers were shown to induce voids both in the direction of laser propagation and in the laser scanning direction. The former structures were attributed to such effects as self-focusing and filamentation \cite{kanehira2005periodic}, standing electron plasma wave \cite{sun2007standing} or spherical aberration \cite{song2008formation, wang2011fabrication, ahsan2014formation}. The transverse laser scanning of non-porous glass (Fig. \ref{fgr:a}) was performed only in a few studies and mostly for non-porous glasses \cite{bellouard2011femtosecond, richter2013formation, cvecek2014gas, matsuo2015spontaneous}. In particular, femtosecond laser pulses were shown to generate self-organized bubble patterns \cite{bellouard2011femtosecond} in fused silica at repetition rate as high as $9.4 MHz$ and scanning speed up to $31 mm/s$. A transition between chaotic and self-organized patterns at high scanning rate (above $10 mm/s$) was revealed and attributed to similarities with rather complicated non-linear dynamical systems.
Thus, an efficient control over laser inscription is still puzzling, particularly for porous glasses. In fact, no systematized control possibilities over threshold, size or period have been proposed so far. That is why the present study is aimed at the identification of the major physical mechanisms involved the ultra-short laser-induced void structure formation in volume of a porous glass. Such parameters as structure periodicity, size and the required pulse energy will be explained.
\section{Results and discussion}
\subsection{Experimental study of void formation}
A series of experiments have been performed for porous silica glass plates. Their chemical composition is (mass fraction,\%): 94.73SiO$_2$-4.97B$_2$O$_3$-0.30Na$_2$O and the expected trace content ($\leqslant 0.1$ mass \%) of Al$_2$O$_3$. The average pore radius is $\approx 2 nm$, the porosity $ 0.26 cm^3 / cm^3$, and the specific surface area of the pores is $\approx 210 m^2/g$. The transmission of the used porous glass is high in the visible and near-IR range of wavelengths ($0.2-2.5 \mu m$). The experimental set-up is shown in Fig. \ref{fgr:b}. Here, Satsuma Yb-doped fiber laser with wavelength $\lambda = 515 nm$ at second-harmonic, temporal pulse width $\tau = 200 fs$ passes through a nonlinear optical crystal (3) and a plate (4) placed at $45^{o}$ angle to the laser beam direction (1). Material modifications are obtained by varying sample moving speed, $V_{s}$, from $0.0125 - 3.75 mm/s$ with respect to the focused laser beam and by changing laser pulse energy $E_p$ from 1.5 to 2.34 $\mu J$ at a constant repetition rate ($\nu = 500 kHz$). The experiments were performed by Dr. Roman Zakoldaev (ITMO University) and Dr. Maksim Sergeev (ITMO University) at Lebedev Physical Institute, Moscow.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.8\textwidth]{./figuresChapter4/figure_2.pdf}
\caption{Illustration of the experimental setup, where tightly focused
(20$\times$, NA $= 0.4$) ultra-short laser pulses ($200 fs$, $500 kHz$, $515 nm$)
create decompaction region inside of the porous glass.}
\label{fgr:b}
\end{figure}
In this case, around 5 percents of laser energy is reflected from the plate (4) and arrives at the power measuring device (11). Then, the laser beam is focused by the objective (5) (LOMO, $20 \times$, $NA = 0.4$) in volume of the glass film/plate (6) at the plane of the laser modified region formation (MR). When laser (1) is turned on, the optical table (7) (Standa 8SMC1-USBhF) starts moving along one of the axis X or Y. At the same time, a part of the transmitted laser radiation is registered by the power-meter (9) located behind the collective lens (8) placed right after the glass plate (6). Another part of the laser energy that is reflected by the place (4) is used to control laser power. The control over the optical table (7) movement and its synchronization with laser power supply (2) is provided by the personal computer (10).
At rather low laser energy and the number of laser shots per focusing spot, a moderate densification, or, compaction is observed in the porous glass. In the intermediate range of laser energy per pulse and of the number of laser pulses per spot, the decompaction structure is observed (Fig. \ref{fgr:c}) and is typically characterized by a series of voids with a size up to ten micrometers, which changes only slightly with the number of laser pulses per focusing spot. Importantly, the observed microvoids appear with a periodicity from a few microns to several tens of micrometers (Figs. \ref{fgr:c} and \ref{fgr:d}(a)) in the considered laser irradiation regime. The distance between the voids can be varied by the total laser energy delivered per focusing diameter, $E_p$, or by the number of laser pulses per spot, $N$ , determined by the scanning speed. Here, a couple of interesting observations can be done.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.42\textwidth]{./figuresChapter4/figure_3_1.pdf}
\includegraphics[width=0.35\textwidth]{./figuresChapter4/figure_3_2.pdf}
\caption{(a-c) Optical transmission microscopy of decompaction regions obtained for different writing speeds; (d) image of periodic lines of voids formed inside of the porous glass at a constant pulse energy ($Ep = 2.2\mu J$) and with different number of pulses per focusing spot: (1) 2600, (2) 3200, (3) 5330, (4) 32000 and (5) 80000. Experimental data from Ref. \cite{ma2017well}.}
\label{fgr:c}
\end{figure}
Firstly, the period $\Lambda \sim \xi N$, increases almost linearly with $N$ (or, decreasing scanning speed), where $\xi$ is a constant length defined by the slope of the line in Fig. \ref{fgr:d}(a). Here, $\xi \approx 0.62 nm$. Secondly, the performed experiments clearly demonstrate that the laser energy required for the formation of the decompaction structures (energy threshold) also linearly increases with the number of the applied laser pulses per laser irradiated spot (Fig. \ref{fgr:d}(b)). Here, $E_a = E_0 \times N$, where $E_0$ depends on the focusing conditions and is on the order of 2.5 $\mu$J.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.45\textwidth]{./figuresChapter4/figure_4_1.pdf}
\includegraphics[width=0.45\textwidth]{./figuresChapter4/figure_4_2.pdf}
\caption{(a) The period of the decompaction regions as a function of the number of laser pulses per spot at a constant pulse energy ($E_p = 2.2 \mu J$); (b) the total threshold energy required to obtain the decompaction of PG using mentioned experimental setup as a function of the number of pulses per focusing spot. Experimental data from Ref. \cite{ma2017well}.}
\label{fgr:d}
\end{figure}
\subsection{Experimental study of high aspect ratio densification}
For high repetition rate multi-pulse laser irradiation, the obtained results depend not only on laser pulse energy, $E_p$, but also on the number of laser pulses per laser spot, $N$. In this case, it is convenient to use the total incident laser energy, which is defined as follows $E_L = E_p N$. Thus, in the present study a series of ultra-short laser PG scanning were performed for two different focusing conditions by varying $E_L$. When $E_L$ is small, porous glass densification is proven to take place \cite{veiko2016femtosecond,ma2017well, veiko2018direct} (Fig. \ref{fgr:Chapt4Exp2all}(a). In the intermediate range of $E_L$, a series of voids are typically formed with a size up to ten micrometers (decompaction) (Fig. \ref{fgr:Chapt4Exp2all}(b)). Further increase of $E_L$ leads to the channel formation.
\begin{figure}[htbp]
\centering\includegraphics[width=0.9\textwidth]{./figuresChapter4/exp2/Fig_2.pdf}
\caption{Top view and cross-section microphotos of fabricated structures in PG: densification region (a,b) obtained by $E_L$ = 2 mJ and 10X, NA = 0.25 lens, and voids (c,d) by $E_L$ = 2 mJ and 20X, NA = 0.4 lens. Measured ratio of $w$ to the laser beam waist diameter (e) for two lenses and $h$ to the Rayleigh length ($2 z_0$) (f) as a function of $E_L$. Experimental dependencies obtained for different NA: width (g) and height (h) of the modified region as a function of $E_L$. The scale bar is 10 $\mu$m. Experimental data from Ref. \cite{itina2019ultra}.}
\label{fgr:Chapt4Exp2all}
\end{figure}
Here, we focus attention at the possibilities to obtain the deepest (the highest) possible PG densification. The focusing position was the same for all the cases, $300 \mu$m below the surface. For this, we firstly examine the shapes of the laser-modified zones, their heights ($h$) and widths ($w$), as presented as a function of $E_L$ (Fig. \ref{fgr:Chapt4Exp2all}). The obtained results show that when the lens with $\mathrm{NA} = 0.4$ is used, only decompaction regime is observed (Fig. \ref{fgr:Chapt4Exp2all}(a) and Fig. \ref{fgr:Chapt4Exp2all}(b)). This regime is characterized by a big void formation at all the considered laser energies \cite{veiko2016femtosecond, ma2017well}. In this case, an increase in $E_L$ leads to a more spherical laser-affected zone.
On the contrary, PG densification without void formation is possible if the total incident laser energy is below 7.5 mJ at $\mathrm{NA} = 0.25$. Figs. \ref{fgr:Chapt4Exp2all}(c) and (d) demonstrate how the dimensions of the laser-densified volume evolve at $\mathrm{NA} = 0.25$ as a function of laser energy. We note a considerable difference in the shape and energy-dependent behavior of the laser-affected volume: while the corresponding ($w$) increases both for $\mathrm{NA} = 0.4$ and $\mathrm{NA} = 0.25$, an increase in $h$ is much more pronounced when $\mathrm{NA}$ is smaller.
Based on the obtained results for $\mathrm{NA} = 0.25$, three different sub-regimes can be distinguished as follows:
(i) densification regime favorable for waveguide recording at $E_L$ smaller than 2 mJ;
(ii) densification regime with a considerably increase the depth-to-width aspect ratio at $E_L$ from 2 to 7.5 mJ;
(iii) decompaction regime at $E_L$ higher than 7.5 mJ.
It is noted that the second regime is more appropriate for barrier formation, where care should be taken to elongate the densified region in depth. The use of lenses with even smaller numerical aperture leads to an unacceptable broadening of the laser-affected zone.
\section{Modeling and discussion}
To explain the obtained experimental results, a thermodynamic analysis of the laser heating is performed. Every applied femtosecond laser pulses generates and heats free carries inside the glass. The subsequent electron-ion/matrix relaxation takes place on a picosecond time scale. In the considered multi-pulse regime, however, material temperature, or "base temperature" \cite{miyamoto2011evaluation,wu2005femtosecond} rises during longer time, $t_h$, corresponding to the focal spot irradiation time and creating an almost spherical region of the energy accumulation (Fig. \ref{fgr:e}(a)). This region can be then considered as a static heat source which appears consecutively in addition to the moving laser. Note, that the involved physical processes occur on very different time scales, where the energy relaxation stage is much longer than the laser absorption and heating ones.
The thermodynamic analysis is based on a modification of the procedure proposed by Miyamoto et al. \cite{miyamoto2011evaluation,sun2007standing} for multi-pulse laser irradiation regime. In our model, both light propagation and energy absorption per shot are calculated by using nonlinear Schr\"{o}dinger equation (NLSE) coupled with material ionization including photo-ionization, avalanche and electron trapping\cite{wu2005femtosecond}. The effective optical index of the studied porous glass \cite{veiko2016femtosecond}, $n_0$, is estimated $\approx$1.342 according to the effective medium theory, e.g. Maxwell-Garnett theory. Moreover, the photo-ionization rate is calculated by using Keldysh theory \cite{keldysh1965diagram}. The NLSE equation used here is written as follows
\begin{equation}
j \frac{\partial}{\partial z} \varphi = - \frac{1}{2k} {\nabla}^2_{\tau} \varphi + \frac{\upsilon_g}{2} \frac{{\partial}^2 \varphi}{\partial t^2} - k \Delta n \varphi
\label{eqn:Eq01}
\end{equation}
\noindent where $\varphi(x,y,z,t)$, ${\nabla}_{\tau}$, $k$ and $\upsilon_g \approx 361 {fs}^2/cm$ are the envelop function of the electric field, the Laplacian operator in the XY plane, the wave number and the group velocity dispersion coefficient, respectively. The term $\Delta n$ includes the optical index change and photo-ionization,
\begin{equation}
\Delta n = j \frac{W_{PI} U_g } {n \varepsilon_0 c_0 {|\varphi|}^2 k} + n_2 \frac{n \varepsilon_0 c_0 {|\varphi|}^2}{2} + j \frac{\sigma_B \rho}{2 k} - \frac{\sigma_B \omega \tau_s \rho}{2 k}
\label{eqn:Eq02}
\end{equation}
\noindent where $W_{PI}$ is the photo-ionization rate and the related to material properties, such as band gap $U_g = 9 eV$, electromagnetic wave angular frequency $\omega$, effective mass of electron $m \approx 0.86 m_e$, and the light intensity \cite{keldysh1965diagram}. $m_e \approx 9.1 \times 10^{-31} kg$ is the free electron mass. The temporal optical index during pulsed laser, $n$, is estimated by Drude model, and $n_2 \approx 3.5 \times 10^{-16} cm^2/W$, $\varepsilon_0 \approx 8.854 \times 10^{-12} F/m$, $c_0 \approx 2.9979 \times 10^8 m/s$, $\tau_s \approx 1.0 fs$ are the Kerr-effect coefficient, vacuum permittivity, light speed in vacuum and electron collision time, respectively. $\sigma_B$ is the cross section for inverse Bremsstrahlung absorption, $\sigma_B = (e^2\tau_s / m(\omega^2 \tau_s^2 + 1) \cdot (1/n c_0 \varepsilon_0))$, where $e \approx 1.6022 \times 10^{-19} C$ is the elementary charge. The free electron density is described by the following single rate equation during photo-ionization \cite{wu2005femtosecond}
\begin{equation}
d \rho / d t = (W_{PI} + \eta I \rho) (1-\rho / \rho_m) - \rho / \tau_p
\label{eqn:Eq03}
\end{equation}
\noindent which takes into account the multi-photon ionization, avalanche $\eta = \sigma / U_{eff}$ and electron relaxation time $\tau_p \approx 150 fs$. Here, $\omega_0$, $\lambda$, $\rho_m \approx 2.2 \times 10^{22} cm^{-3}$, $I = n c_0 \varepsilon_0 |\varphi|^2 / 2$ and $U_{eff} = (1+m/m_e) \cdot (U_g + e^2 |\varphi|^2 / 4m \omega^2)$ \cite{kaiser2000microscopic} are the waist radius, maximum plasma density, laser wavelength in vacuum, laser intensity and effective band gap, respectively.The solution of these equations results in the following distribution of the total absorbed energy
\begin{equation}
q(z) = \pi \omega_0^2 [1+(\frac{\lambda z}{\pi \omega_0^2 n_0})^2] \int_t \frac{4\pi \cdot imag(n)}{\lambda} I(z,t) dt
\label{eqn:Eq04}
\end{equation}
\noindent Finally, thermal behavior at time $t$ is considered as a heating by consecutive laser sources separated along the scanning direction with distances of $Vs / \nu$. The analytical solution of the thermal equation is derived by Green's function method as follows \cite{miyamoto2011evaluation}
\begin{equation}
\Delta T(x,y,z,t) = \frac{1}{\pi \rho_g c_g} \sum_{i=0}^{N} \frac{1}{\sqrt[]{\pi \alpha (t-i/\nu)}} Q_i
\label{eqn:Eq05}
\end{equation}
\begin{equation}
Q_i = \int \frac{q(z)}{\omega_z^2 + 8\alpha(t-i/\nu)} exp[-\frac{2\{[x+V_s(t-i/\nu)]^2 + y^2\} }{\omega_z^2+8\alpha(t-i/\nu)} - \frac{(z-z')^2}{4\alpha t}] dz'
\label{eqn:Eq06}
\end{equation}
\noindent where $\omega_z = \omega_0 \sqrt[]{1+(\frac{\lambda z}{\pi \omega_0^2 n_0})^2}$ is the estimation of laser beam radius along $z$ axis. Thermal parameters such as mass density $\rho_g$, heat capacity $c_g$ and thermal diffusity $\alpha$ are shown in Table \ref{table:a}. With all these equations at hand, we estimate the temperature distribution and evolution during multi-pulse laser scanning. Fig. \ref{fgr:e} shows the 2D base temperature distribution and evolution at various scanning speed.
\subsection{Periodical void formation}
The formation of the periodic patterns in multi-pulse regime is often attributed to the heat accumulation \cite{graf2007pearl, matsuo2015spontaneous}. In ultra-short laser processing, heat accumulation is typically considered as non-negligible when time between successive laser pulses, $\delta t = 1 / \nu$, becomes shorter than the characteristic cooling time of the focal spot. When these times are comparable as in the present study (laser repetition rate is $500 kHz$), heat can escape from the focal region avoiding strong thermo-mechanical effects.
\begin{table}[ht!]
\centering
\caption{Parameters summary used in simulation}
\begin{tabular}{llll}
\hline
\hline
Parameter & Description & Value or definition &Reference\\ \hline
$ \lambda $ & Laser wavelength in vacuum & $515$ nm &\\
$n_0$ & Effective optical index of porous glass & 1.342 &\\
$\upsilon_g$ & Group velocity dispersion coefficient & $361$ {fs}$^2$/cm & Ref(\cite{wu2005femtosecond,gulley2012interaction})\\
$ m_e $ & Free electron mass & $9.1\times 10^{-31}$kg &\\
$ m $ & Effective electron mass & $0.86 m_e$ &Ref(\cite{wu2005femtosecond})\\
$ n_2 $ & Nonlinear refractive index & $3.5 \times 10^{-16}$ cm$^2$/W &Ref(\cite{wu2005femtosecond})\\
$ \varepsilon_0 $ & Vacuum permittivity & $ 9.954 \times 10 ^{-12}$ F/m &\\
$ c_0 $ & Light velocity & $2.9979\times 10^8$ m/s &\\
$ \tau_s $ & Electron collision time & $1.0$ fs &Ref(\cite{wu2005femtosecond})\\
$ \tau_p $ & Electron relaxation time & $150$ fs &Ref(\cite{wu2005femtosecond})\\
$ e $ & Elementary charge & $1.6022\times 10^{-19}$ C &\\
$ \rho_m $ & Maximum plasma density & $2.2\times 10^{22}$ cm$^{-3}$ &Ref(\cite{wu2005femtosecond})\\
$ U_g $ & Material band gap & $9$ eV &\\
$ \tau $ & Pulse duration & $200$ fs &\\
$ \nu $ & Repetition rate & $500$ kHz &\\
$ V_s $ & Scanning speed & $0.125 - 100$ mm/s &\\
$ \omega_0 $ & Laser beam waist radius & $2.45 - 2.6\mu $m &\\
$ E_p $ & Pulse energy & $1.5 - 2.3$ $\mu$J &\\
$ \rho_g $ & Mass density of porous glass & $2.1 - 2.2$ g/cm$^3$&\\
$ c_g $ & Heat capacity & $0.8 - 1.6$ J/g/K&\\
$ \alpha $ & Thermal diffusity & $2.7 \times 10^{-7}$ m$^2$/s&\\
\hline
\hline
\label{table:a}
\end{tabular}
\end{table}
In our case, laser source moves continuously in the direction transverse to the propagation one. As a result, laser beam passes the focal volume with lateral diameter $d_f=2R_f$ during the heating time $t_h=d_f/V_{s}$ by applying $N=\nu\frac{2R}{V_{s}}$ laser pulses. Thus, the linear dependencies revealed in Fig. \ref{fgr:c} can be presented as follows
\begin{equation}
\Lambda (N) = \xi N = \xi N_0 + \xi (N - N_0) = d_v + \xi N_1
\label{eqn:Eq1}
\end{equation}
\noindent for the period of the void structure, and
\begin{equation}
E_a(N) = E_0 N = E_0 N_0 + E_0 (N - N_0) = E_v + E_0 N_1
\label{eqn:Eq2}
\end{equation}
\noindent for the threshold laser energy respectively, where $N_0$ is the minimum number of laser pulses per focusing spot in the decompaction regime; $N_1 =( N - N_0)$; $d_v$ is the typical void diameter, and $E_v = E_0 N_0$ is the minimum laser energy per focusing spot required for decompaction. Thus, these dependencies indicate that only a small fraction of energy goes to the void formation. When the number of laser pulses overcomes $N_0$, then a large fraction of energy is lost, and this amount rises linearly with the number of laser pulses as well as structure period.
\begin{figure*}[ht!]
\centering
\includegraphics[width=0.6\textwidth]{./figuresChapter4/figure_5_1.pdf}\\
\includegraphics[width=0.4\textwidth]{./figuresChapter4/figure_5_2.pdf}
\includegraphics[width=0.4\textwidth]{./figuresChapter4/figure_5_3.pdf}
\caption{(a) XZ view of base temperature distribution ($E_p = 2.2 \mu J$, $V_s=1 mm/s$, $\nu=500 kHz$, $\omega_0=2.45 \mu m$) after 10,000 pulses irradiated on the porous glass; (b) temporal profiles of base temperature increase in positions of (1,0,0)$\mu m$ and (10,0,0)$\mu m$ at various laser scanning speed; (c) 1D temperature increase along X axis at various scanning speed after 10,000 pulses illuminated on the material, as one can see, at speed lower than $10$ mm/s the heat front propagates faster than the laser scan; Note, in this calculation, the coordinate is moving at speed of $V_s$; for simplicity, heat parameters are set as constants with $c_g = 1.6J/g/K$, $\rho_g = 2.2 g/cm^3$ and $\alpha = 2.7 \times 10^{-7} m^2/s$.}
\label{fgr:e}
\end{figure*}
Material decompaction occurs when the absorbed energy is high enough to heat material in the focal region (at $d_f / 2$) up to a temperature exceeding the glass softening temperature and when the energy source supplies enough energy for the void formation. It was estimated that $\approx 2.5 \mu J$ per pulse is required for the tensile stress in the localized molten region of fused silica to exceed several MPa, which is enough for a void formation (see Ref. \cite{bulgakova2015modification}). As one can see in Fig. \ref{fgr:d}(b), in our case $\approx 2.2 \mu J$ is the laser energy per pulse, but at least $E_v$ is required to create a void structure and this energy rises with $N$. The energy losses include energy spent for light reflection, scattering, and transmission. Additionally, laser energy is required for such effects as material ionization (both field and avalanche ionization processes), plasma heating, phase transitions (here, softening, densification, plasma formation), and it is also dissipated via pressure waves, heat conductivity, convection and viscous flow motion, and even radiation. As a result, only a very small fraction of the absorbed energy actually leads to the void formation.
Apparently, energy losses increase linearly with $N$, so that the most effective regime for void structure formation is when $N$ is rather small, or at the highest laser writing speed required for the void structure formation (see Fig. \ref{fgr:d}(b)) . In this case, voids are closely packed and $\Lambda = d_v$. The typical void size is $\approx 10 \mu m$ corresponds to the radius of the molten and superheated region and exceeding the diameter of the focal region (around $4 \mu m$), where gas/plasma phase is formed.
To estimate void size, we define the temperature required for void formation for the corresponding laser irradiation conditions by applying the Grady's criterion for spall in liquid \cite{grady1988spall}. According to the viscoelastic energy conservation law, the elastic energy of the deformation is defined as $\approx Y\sigma{t^2}$, where $t$ is the deformation time limited by the time between two pulses $2 {\mu}s$ and $\sigma =\frac{\Delta{\rho}}{\rho t}$ is the laser-induced strain rate, should be sufficient to do the work against the dissipation forces defined by ${\eta}\sigma$. The viscosity of porous glass is lower than that of the pristine fused silica and its dependencies on the glass porosity have been previously reported \cite{sura1990viscosity, boccaccini1995viscosity}. For instance, for porosity of $26\%$, the viscosity is approximately ten times lower than it for the pristine silica. This way, it is possible to derive the maximum viscosity required for cavitation as follows $\eta_{max} = Y\sigma{t}^2 \approx 10^6 Pa \cdot s$. These viscosities correspond to the temperature $T_d$ of order $2500 K$, exceeding boiling temperature of the material. The results of the performed calculations presented in Fig. \ref{fgr:e} indicate that the thermo-affected zone with the temperatures exceeding $2500 K$ is around $10 {\mu}m$, which agrees well with the experimental results (Fig. \ref{fgr:c}).
We note that for the solid material, another equation \cite{gamaly2006laser, cerkauskaite2017ultrafast} can be derived based on the energy conservation law as follows: assuming that the internal energy of the material with volume of $\frac{4}{3}\pi r^3$ is around the absorbed laser pulse energy, one gets $\frac{4}{3}\pi Y r_{0}^3 \approx E_{abs}$. The fact that voids grow only slightly with laser energy
means that upon a certain laser energy above the densification threshold, the absorbed fraction is almost constant, whereas energy losses should grow linearly. Then, for the diameter of a spherical void can be estimated as follows $d = C\sqrt[3]{\frac{6 \alpha E_p}{\pi Y}}$, where $\alpha$ is the fraction of laser energy $E_p$ absorbed by the material; $Y$ is Young modulus; $C = \sqrt[3]{\frac{\Delta{\rho}}{\rho}}$ is the compression ratio. For fused silica, Young modulus $Y = 74.5 GPa$, whereas for porous glass it can be smaller, for instance $Y = 17.5 GPa$ was reported by Cerkauskaite et al. \cite{cerkauskaite2017ultrafast}. For the absorbed energy on the order of $2 \mu J$, the estimation gives void diameter $\approx 10 \mu m$, which is more than twice larger than the focusing diameter $d_f$.
\begin{figure}[ht!]
\centering
\includegraphics[angle=0, width=0.5\textwidth]{./figuresChapter4/figure_6.pdf}
\caption{Schematics illustrating the void structure formation}
\label{fgr:f}
\end{figure}
It remains to explain structure periodicity, and, namely, the observed almost linear period dependency on the number of laser pulses per focusing spot. Considering that cooling time is rather long and a scanning heating source supplies energy all the time, a stationary regime is expected in the absence of the void formation. In this regime, once void is formed, the temperature field outside the void is induced by two heat sources: (i) a moving laser source; and (ii) a static spherical heat source formed at the location of the previously formed void (Fig. \ref{fgr:f}). The solution of a static thermodynamic problem is straightforward and gives $T = Q / 4 \pi k r$ outside of the void, where $k$ is thermal conductivity and $Q$ can be considered to be proportional to $E_0 N$, indicated by the revealed linear dependency of the decompaction threshold. Heat diffusion induces densification and softening. The fronts of these phase transitions propagate longer distances when more energy is absorbed, therefore $\Lambda=r_d = Q / 4 \pi k T_d$. Once laser overcomes the modified region, again enough energy can be spent for the next void. When $r_d$ is smaller than $r_f/2$, we get a closely packed void structure, otherwise the voids are spaced with a distance $\Lambda$ that rises linearly with $N$.
\subsection{Densification}
The observed results rely on the physical processes involved in the formation of laser-induced modifications. When femtosecond laser is used, non-linear ionization generates electron plasma (ELP). Both defect formation and ELP relaxation lead to the local modifications in the refractive index. As a result, optical properties, such as scattering and absorption, are modified.
The generated ELP strongly affects laser propagation. Our calculations \cite{ma2017well} show that the maximum ELP density depends on both laser focusing conditions and laser pulse energy (Fig. \ref{fgr:Chapter4_exp_4all}(a)). Importantly, optical breakdown takes place at laser energy of $~1.9 \mu$J for $\mathrm{NA} = 0.25$, whereas for $\mathrm{NA}=0.4$ it is expected at at a much smaller laser energy ($~0.7 \mu$ J). The transmission also drops at much smaller energy for $\mathrm{NA}=0.4$ than for $\mathrm{NA} = 0.25$ ((Fig. \ref{fgr:Chapter4_exp_4all}(b)). These results explain the difference in the material modifications regimes observed for these two different laser focusing conditions.
\begin{figure}[!ht]
\centering\includegraphics[width=0.9\textwidth]{./figuresChapter4/exp2/figure4.pdf}
\caption{ Calculated maximum electron density (a) and light transmission (b) for single ultra-short laser pulse. The model parameters are given in \cite{ma2017well}. }
\label{fgr:Chapter4_exp_4all}
\end{figure}
Importantly, if laser power is above the critical one $P_{cr} = 3.77\lambda ^2 / 8 \pi n_0 n_1$, where $n_0$ and $n_1$ stand for the linear (1.34) and nonlinear refractive ($\approx 0.4\times 10^{-19} m^2/W$) indexes of the material \cite{cimek2017experimental}, $P_{cr} = 7.4\times 10^5 W$), the laser propagation of a Gaussian laser pulse is accompanied by such effects as self-focusing and filamentation \cite{couairon2007femtosecond}. Further increase in laser power becomes inefficient (a so-called "intensity clamping effect" \cite{xu2008effect}).
In all the experiments, laser power exceeded the critical one. In fact, $P_0 = E_p/\tau$, it ranges from $5.0\times 10^5 W$ to $1.2 \times 10^7 W$ for $E_p$ at $0.1\mu J$ and $2.3 \mu J$ respectively. Despite the fact that the critical power does not depend on the laser focusing, both the characteristic length of the filament appearance $z_f$ (1) and filament dimension strongly depends on the laser beam waist \cite{couairon2007femtosecond}, where
\begin{equation}
z_f = \frac{0.367 k \omega^2_0}{\sqrt{[(P_0 / P_{cr})^{1/2} - 0.852]^2 - 0.0219}}
\label{eqn:eq1}
\end{equation}
\noindent where, $k$ is a coefficient. Note that the equation would be invalid for incident power $P_0 > 100 P_{cr}$ \cite{couairon2007femtosecond}. Because $z_f$ is proportional to $\omega_0^2$, even rather small difference in $\omega_0$ affects the filamentation process. In our case, when $\mathrm{NA} = 0.25$ ($2 \omega _0 = 5.7 \mu m$), and if $\mathrm{NA}=0.4$ ($2 \omega_0 = 3.6 \mu m$), so that the ratio of $z_f$ parameter in our case is $\approx 2.5$. This means that for $\mathrm{NA} = 0.25$ laser propagates $\approx 2.5$ longer distance before the possible appearance of self-focusing than at $\mathrm{NA}=0.4$. During the propagation path, the absorption and heating of ELP typically lead to laser energy losses. In fact, denser plasma formation and heating lead to the decompaction (void formation) rather than to the formation of a densified region. Thus, at smaller values of $E_L$, mostly densification is observed for $\mathrm{NA} = 0.25$ (Fig. \ref{fgr:Chapt4Exp2all}).
\begin{figure}[!ht]
\centering\includegraphics[width=0.8\textwidth]{./figuresChapter4/exp2/figure5a.pdf}\\
\centering\includegraphics[width=0.8\textwidth]{./figuresChapter4/exp2/figure5c.pdf}
\caption{Heat accumulation by an ultra-fast laser of high repetition-rate (500 kHz $\mathrm{NA} = 0.25$) at various pulse energies and scanning speeds (a). Laser moves in the positive X direction (from left to right). Dimension of the densified area: width (b) and elongation (c). Here, these dimensions are estimated based on the temperature of $1000^\circ\text{C}$ \cite{veiko2016femtosecond}. }
\label{fgr:Chapter4_exp_5all}
\end{figure}
As we have noted in the Introduction, additional effects, such as defect, electronic and/or thermal accumulation, play a role in a multi-pulse regime. Particularly, heat accumulation depends on laser pulse width, repetition rate, laser spot dimensions and scanning speed, as well as on the material properties.
For this case, a combined model was developed \cite{ma2017well} that carefully accounts for ultra-short laser light propagation, material ionization, and heating. A series of calculations are performed for the case of $\mathrm{NA} = 0.25$ (Fig. \ref{fgr:Chapter4_exp_5all}). Based on the calculated temperature distributions (Fig. \ref{fgr:Chapter4_exp_5all}(a)), we evaluate the sizes of the laser-modified region (Fig. \ref{fgr:Chapter4_exp_5all}(b) and (c)). In the presented experiments discussed above, the system is designed to control the total incident laser energy $E_L$. To better compare the multi-pulse simulations with experiments, the results are shown as a function of $E_L$, which is calculated as $E_L = \frac{2\omega_0 f_p E_p}{Vs}$, where $f_p = 500 kHz$ is the repetition rate, $Vs$ is the laser scanning speed. The performed simulations reveal how the corresponding temperature profiles change with the increase in laser pulse energy for two laser scan speeds (1 mm/s and 45 mm/s). At high speed, a tail appears in the temperature distribution if laser pulse energy is high enough. This tail is commonly observed behind powerful moving heat sources. The higher the speed of the heat source, the longer is the tail. At lower laser energies, this tail is, however, less pronounced.
Importantly, Figure \ref{fgr:Chapter4_exp_5all} confirms the experimentally observed rise and saturation in the dimensions of the laser-modified area. Here, they were estimated based on the dimension of the region where temperature exceeds softening one. The observed saturation in the dimensions of laser-densified region is mainly caused by the reflection and scattering effects of the ELP.
We note, furthermore, that laser-induced modifications in local optical properties are known to cause local field enhancement\cite{pereira2008laser, rasedujjaman2018polarization} and pulse-to-pulse elongation of the laser-affected region. Upon many pulse, depending on laser energy, repetition rate, induced refractive index change, and the size of the modified area, such accumulation effects either lead to a further increase in the aspect ratio of the laser-modified region\cite{rudenko2016random, mcleod2008subwavelength}, or to the appearance of a big void in its center if laser energy is high enough \cite{watanabe2000optical, veiko2016femtosecond, bulgakova2015modification}. It is estimated that about 2.5$ \mu J$ per pulse is required the void formation \cite{bulgakova2015modification, ma2017well}.
\section{Conclusion}
The periodic void structures produced in a well-controlled manner inside porous glasses by femtosecond laser pulses. The trains of ultrashort laser pulses with a properly chosen energy per pulse and writing speed not only provide the required energy deposition, but also allow an efficient inscription of the periodic arrays of voids in these materials. Contrary to powerful single laser pulses that are known to produce explosion-like effects, pulse trains can be used to considerably reduce non-desirable effects, such as material cracking.
The performed thermodynamic analysis have shown that the possibility of an efficient control over laser micromachining in volume relies on a better understanding of the physical mechanisms involved. For the given pulse repetition rate, the main controlling parameters being the laser scanning speed and the mean energy per pulse, while the former defines both the structure period and the required laser energy per focusing spot. The period has been found to depend on the size of the densified region induced by the created almost static spherical heat source located at the void position. The final structure is induced by a superposition of two energy sources: the moving laser and the created static heat source.
Additionally, the performed analysis has indicated that the void formation is more efficient when the minimum number of laser pulses per focusing spot corresponds to the writing speed as high as about $0.47mm/s$. Under these conditions, a compact void structure has been written in the porous glass, so that the structure period is equal to the void diameter. For lower scanning speed, both threshold and period increase linearly as a result of the rising energy losses. The lost energy goes mostly for material heating and densification thus leading to a reduced absorption at the passages between the voids.
The performed study thus provides a way of a well-controlled laser inscription of void arrays in the bulk of porous materials. The obtained results can be useful for advancing not only a wide range of photonics applications, but also for many applications in microfluidics, biology and other fields.
Furthermore, the ultra-short laser interaction with PG in a high repetition rate multi-pulse regime was investigated. Conditions favorable for the inscription of either highly symmetric wave-guides or high aspect-ratio membranes and barriers were determined, which was explained based on a series of numerical calculations. Laser focusing affects filament appearance distance. In the considered case, this effect is minor compared to the effect of the laser-induced ionization since the created ELP strongly affects thermal field as well as the shape of the laser-affected zone. The characteristic dimensions of the calculated temperature field are well correlated with the obtained shapes of the densified regions. The maximum density of the created electron plasma was found, furthermore, to strongly depend on the focusing conditions.
Laser-based high aspect ratio densification is promising for numerous applications in microfluidics and for the fabrication of the integrated multi-purpose sensors. The identified laser irradiation conditions required for the efficient void-free material densification thus provide a way toward a better control over direct laser writing of membranes, barriers and various integrated devices in porous glasses.
\afterpage{\null \newpage}
\chapter{Continuous-Wave laser writing of silver NPs in mesoporous TiO$_2$ thin films}
\hspace{0.5cm} Comparing to porous dielectrics, the semiconducting counterparts such as TiO$_2$ and ZnO exhibit additional possibilities in photoactive applications. The photosensitive abilities were activated by the process of porous semiconductors interact with nanoscale dopants through transferring free electrons. In recent years, the interaction of semiconducting media with metallic nanoparticles demonstrated laser controlling behaviors, which was promising for optical data storage, color printing, and multiplexing \cite{destouches2014self,liu2015understanding,baraldi2016polarization,sharma2019laser}.
In this chapter, we present numerical insights in continuous-wave laser writing silver nanoparticles (NPs) inside mesoporous TiO$_2$ thin films. Controlling the formation of NPs by direct laser writing ensures an efficient way of engineering the optical responses of nanocomposite materials through adjusting laser parameters such as power, focusing and writing speed. The process of a NP's growth, photo-oxidation, and reduction affect the light absorption by an ensemble of laser-activated NPs. However, those self-consistent process are not well understood. Previous studies \cite{liu2015understanding, liu2016laser, liu2016selfthesis} explained the speed threshold switching the growth or shrinkage of silver NPs by neglecting heat diffusion, it remained unclear, however, how the NP size can be controlled by the laser writing speed. In this chapter, based on the coupled calculations of size-variation, light absorption, and heat diffusion by an ensemble of NPs, we propose a two-dimensional model taking account of spatial information of size and temperature to have a further investigation into the process. The spatial size distribution is revealed to be non-uniform leading to the transmission inhomogeneity along the laser written lines. This fact is confirmed by the in-situ transmission experiments. The performed study also depicts a novel view in which NPs grow ahead of the laser beam center due to the heat diffusion. The nonlinear growth never stops until it exhausts the majority of the free Ag$^0$ in the matrix, while the amount of Ag$^0$ by reduction cannot compensates the consumption. After that, the photo-oxidation dominates the process and finally acts as the controlling role in NP size depending on the writing speed. The simulations also show that it is not the activation energy of Ag$^0$ diffusion but the ionization efficiency affect the final NP size, which helps understand in how to improve the laser processing of differently prepared samples.
Reproduced in part with permission from [H. Ma et al., J. Phys. Chem. C 123.42 (2019): 25898-25907.] Copyright [2019] American Chemical Society.
\section{Introduction}
Plasmonic nanoparticles (NPs) and nanostructures have great potential in photovoltaics \cite{atwater2010plasmonics, nakayama2008plasmonic}, nano-imaging \cite{kawata2009plasmonics}, bio-sensing \cite{anker2008biosensing}, photoelectric sensors \cite{brongersma2015plasmon, clavero2014plasmon} and light engineering metamaterials \cite{yin2013photonic,luk2010fano, ergin2010three}. A lot of plasmonic metasurfaces have been created during the last decade to design new optical functionalities. Most of them resulted from perfectly arranged metallic oligomers and were produced by focused ion beam and electron beam lithography, two techniques that are still expensive, time consuming and not well suited to large surfaces. Nanoimprint lithography of metallic nanostructures overcomes these limitations and is expected to become a way to spread the use of plasmonic metasurfaces at the industrial level. This technique is, however, limited to reproduction of masks. Laser direct writing is another technology that provides more flexibility, is well suited for large scale processing and allows a control of plasmonic properties at sub-wavelength level.
Among the metasurfaces that have great promises for applications is titanium dioxide (TiO$_2$) loaded with metallic NPs. TiO$_2$ is one of the most studied and widely applied host-materials due to its unique characteristics, such as excellent optical transmittance in visible wavelength, high refractive index, and stable chemical properties \cite{matsumoto2001room, clavero2014plasmon}. Additionally, TiO$_2$ is a good n-type semiconductor material and it allows fast electron acceptance because of the high density of states in the conduction band\cite{clavero2014plasmon}. However, only ultraviolet light can activate the photovoltaic process due to the intrinsic wide-bandgap limitations of TiO$_2$. Embedding silver NPs in such a material forms a Schottky barrier, which boosts the photovoltaic efficiency in the visible region. The plasmonic resonance induces a high density of free electrons whose energies exceed the Schottky barrier and makes an increase in the quantum efficiency possible. This so-called "plasmonic hot-carrier" phenomenon has attracted wide interest in research communities during the past decade \cite{clavero2014plasmon, brongersma2015plasmon, tatsuma2017plasmon}. By injecting hot-electrons from silver NPs to the surrounding semiconductor matrix, NPs become unstable and tend to dissolve into Ag$^+$ ions, a mechanism also named photo-oxidation \cite{naoi2004tio2}. In the meantime, hot electrons also relax through electron-phonon coupling that heats the Ag NPs. This heating leads to other mechanisms, such as chemical reduction of silver ions and NP growth, mechanisms compete with the aforementioned photo-oxidation tending to shrink the NPs \cite{liu2015understanding, baraldi2016polarization, liu2017three}.
Using scanning laser writing for such TiO$_2$:Ag film provides an elegant way of controlling the NP size, organization and anisotropy, and thus control the optical properties of such plasmonic metasurfaces. Previously, the growth of Ag NPs in mesoporous films of amorphous TiO$_2$ was investigated as a function of scanning speed for different laser powers and focusing. The first developed model \cite{liu2015understanding} was able to simulate the experimentally observed phenomenon showing that NPs grow only above a speed threshold and tend to oxidize below this threshold. In this model, a set of differential equations was solved numerically to demonstrate the competition between shrinkage and growth during laser writing. Later, the model was improved to simulate the size distribution of the grown NPs as a function of time and speed \cite{liu2017three}. These previous numerical studies suffered, however, from several limitations. Namely, they modeled the evolution of the system in a point without considering the effect of heating by the NPs grown few micrometers apart and still under the laser beam. The temperature was then calculated by assuming a homogeneous spatial size distribution of NPs under the laser spot. This led to a good description of the final NP size distributions as a function of speed, but also to a rather questionable growth kinetics and temperature history.
Here, we propose a model, which accounts for both thermal and spatial size distribution during laser writing to better investigate the growth kinetics during the laser writing and examine the relative role of the underlying physico-chemical mechanisms involved in the laser-induced nanoparticle size evolution. The Galerkin method and adaptive time stepping techniques are used to solve the nonlinear diffusion-controlled growth model based on the open-source finite element method library (Deal.II 8.5.0) \cite{arndt2017deal}. A set of 2D and 3D simulations are carried out and compared to new experimental results obtained based on in-situ transmission measurements. The modeling part summarizes the major physico-chemical processes involved and presents how the thermal equation is coupled. The last part analyzes the obtained results and provides new insights into the dynamics of the system under continuous wave laser scanning. In particular, we examine the importance and interplay of the different mechanisms involved.
\section{Experimental}
\subsection{Sample preparation and in-situ characterization setup}
Mesoporous films of amorphous TiO$_2$ are prepared according to a previously published procedure based on a sol-gel process \cite{crespo2014changes}. The obtained films have pore sizes from 5 to 20 nm and a thickness estimated to 230$\pm$50nm by scanning electron microscopy (SEM - FEI Nova nanoSEM 200) on the cross-section. Silver is introduced in the pores by soaking the mesoporous sample into an ammoniacal silver nitrate solution and small silver nanoparticles are created in the film by exposing it to UV light (400 $\mathrm{\mu W cm^{-2}}$ at 254 nm wavelength). Ag NPs are observed after this process with an averaged size around 3nm \cite{liu2017haadf}. The detailed descriptions of the process were reported previously \cite{liu2015understanding,crespo2010reversible,crespo2012one,crespo2014changes, babonneau2018real}. The experiments were carried out by Dr. Said Bakhti at Lawrence Berkeley National Laboratory under the supersion of Nathalie Destouches (Laboratoire Hubert Curien) and Daniel S Slaughter (Lawrence Berkeley National Laboratory).
\begin{figure*}[ht!]
\centering
\includegraphics[width=0.8\textwidth]{./figuresChapter5/figure_1a_v2.pdf}\\
\includegraphics[width=0.8\textwidth]{./figuresChapter5/figure_1b_v3.pdf}
\caption{In-situ transmission experiments. Schematic of the laser setup (a), transmission images of the sample under writing at various translating speeds (b), map showing the transmission along a line parallel to the scanning direction and passing through the beam center (dashed line in the left-handed image) (c), transmission profile along the dashed line of (c) at 0.3 mm/s (d), and the minimum value of transmission along the scanning direction versus scanning speed (e). The green dashed line stands for the writing laser of spot size $20 \mu m$ at $1 / e^{2}$ ($17 \mu m$ FWHM). The collimated beam before focusing is 1.7 mm full width at $1 / e^{2}$ (1 mm FWHM). Experimental results are taken from Ref. \cite{ma2019laser}.}
\label{fgr:1all}
\end{figure*}
Figure \ref{fgr:1all}(a) shows the setup used for in situ measurements of the sample transmission during laser writing. The writing laser is a 1W continuous-wave laser emitting at 532 nm and linearly polarized along the y-direction on the sample plane. It is focused on the sample surface by a 5 cm focal length lens. Imaging is performed thanks to a probe nanosecond laser emitting at 527 nm, a wavelength absorbed by growing Ag NPs. The latter is collimated on the sample and then imaged on a CCD camera through a couple of lenses, the first one being the same as the one used to focus the writing beam. Without writing beam, the sample transmittance is fully stable over time, meaning that the nanosecond laser does not modify the nanocomposite film. The imaging beam is linearly polarized (x-axis) with a polarization perpendicular to the writing beam. A linear polarizer just before the camera and a band-pass filter stop all photons from the writing beam before the camera. The writing beam is thus not seen directly on the images where only the transmission changes of the samples are observed. The writing beam position on the images is determined before carrying out in situ experiments. During the experiments, the sample is translated at a constant speed along the x-axis, in the positive direction (according to the sketch in Figure \ref{fgr:1all}(a)).
\subsection{Results of the in-situ transmission}
Transmission images are calculated by normalizing the raw images by the averaged intensity recorded in the laser front area. Images are recorded for various scanning speeds (0.1-40 mm/s) and a selection is shown in Figure \ref{fgr:1all}(b). In all cases, transmission varies in the laser written line and our attention is focused on the area located around the laser beam corresponding to the distance "0" in Figure \ref{fgr:1all}(d). The profiles along the line center show the evolution of transmission from a position 100 $\mu m$ in front of the laser beam center to a position 100 $\mu m$ after the beam center. The measured transmission strongly decreases in the front (leading) edge of the laser beam before increasing back near the beam center and reaching a constant value in the back edge, as shown in Figure \ref{fgr:1all}(c) and (d). Unexpectedly, the lowest transmission is before the laser center. Moreover, its value decreases with increasing speed until saturation (Figure \ref{fgr:1all}(e)). The transmission inhomogeneity indicates the variation of the NPs' characteristics (such as size, density, shape and order) along the translation direction. To better understand the mechanisms involved in the NP growth and their kinetics during the laser scan, we perform numerical simulations of the measured behaviors accounting for photo-oxidation, thermal diffusion, diffusion-limited growth, reduction, and light scattering.
\section{Modeling}
In this model, we are going to present how the isolated Ag NPs grow and shrinkage affecting the heat absorption inside the laser spot and vice versa. A complicated interplay during the dynamical process is revealed. Figure \ref{fgr:2all}(a) illustrates the laser writing of a colorful line by controlling writing speed. We also emphasize that Ag NPs start to grow ahead of the laser beam center as shown in Figure \ref{fgr:2all}(b). The average temperature indicating the heat absorption is a phenomenon relying on a group of NPs. The variation of a single NP size could barely explain the results shown in (Figure \ref{fgr:2all}(c)), which makes great difference with the results of the previous model \cite{liu2015understanding}.
\begin{figure}[!hb]
\centering
\includegraphics[width=0.8\textwidth]{./figuresChapter5/figure_2_v4.pdf}
\caption{Illustration of the laser writing of a colorful line (a) and the mechanisms involved during laser excitation of Ag NPs inside the TiO$_2$. Ag NPs grow ahead of the laser beam and then shrink due to photo-oxidation (b). NPs inside the laser spot act as the heat sources and contribute to the temperature field due to the thermal diffusion (c). Injection and tunneling of plasmon-induced hot-electrons (d), process of photo-oxidation (e), heating and diffusion (f), reduction (g), and diffusion-limited growth (h) and (i).}
\label{fgr:2all}
\end{figure}
\subsection{Photo-oxidation}
This mechanism is the one that leads to electron emission from NPs upon photon absorption. As sketched in Figure \ref{fgr:2all}(d), when Ag NPs are in contact with TiO$_2$, the electron energy required to allow electrons to leave the NPs is lowered to the value of the Schottky barrier that is created at the interface between the metal and the n-type semiconductor. Electrons excited through the localized surface plasmon resonance of NPs at the laser wavelength (2.4 eV) can cross the Schottky barrier by injection or tunneling. The band bending prevents the injected electrons from escaping back to the metallic crystals. As a result, the positively charged NPs are unstable and tend to dissolve into Ag$^+$ ions to keep the neutral state (Figure \ref{fgr:2all}(e)) \cite{liu2015understanding}. The number of silver atoms leaving a nanoparticle per unit time due to the above processes was calculated as \cite{liu2015understanding}:
\begin{equation}
n_{oxi}(t) =\eta_0 \frac{ \sigma_{abs}(R,\lambda) I(x, y, z, t)} {h \nu}
\label{eqn:oxidationRate}
\end{equation}
\noindent where $\eta_0$ is the ionization efficiency, $\sigma_{abs}$ the absorption cross-section of a nanoparticle of radius $R$ at laser wavelength $\lambda$, $I$ the laser intensity, $h$ the Planck constant and $\nu$ the light frequency. The ionization efficiency was estimated experimentally in reference \cite{liu2015understanding}. $\sigma_{abs}$ is calculated by using the classical Mie theory.
\subsection{Reduction}
Figure \ref{fgr:2all}(g) depicts the reduction process where Ag$^+$ ions can diffuse and recombine with electrons from TiO$_2$ conduction band or from compounds absorbed in the film such as H$_2$, CO from the atmosphere or NO$^-_3$ from the initial silver nitrate solution \cite{crespo2011irradiance, kazuma2011nanoimaging, kazuma2012photoelectrochemical, clavero2014plasmon, kawahara2005electron}. Silver ions can come from the initial non-reduced silver species present in the film or from the photo-oxidation process. The reduction process leads to Ag$^0$ monomers that can contribute to the NP growth according to the next mechanism. We assume a homogeneous Ag$^+$ ion and Ag$^0$ monomer distribution for a small cuboid, in which matter conservation is presumed. The decrease of Ag$^+$ concentration due to the reduction per unit time has the following form \cite{liu2015understanding, kaganovskii2007formation}:
\begin{equation}
\frac{d C_{Ag^+}}{d t} |_{red} = \exp \Big( -\frac{E_p}{N_A k_B T} \Big) \cdot D^{red}_{0} \exp \Big( -\frac{E_D}{N_A k_B T} \Big)
\cdot C_{red} C^{2/3}_{Ag^+}
\label{eqn:reduction}
\end{equation}
\noindent where $E_p$ and $E_D$ are the activation energies, $D^{red}_{0}$ is the diffusion coefficient of the reducer at infinite temperature, $C_{red}$ the concentration of reducing agent, and $C_{Ag^+}$ the concentration of Ag$^+$. $N_A$ is the Avogadro constant, $k_B$ the Boltzman constant and $T$ the absolute temperature.
The first exponential function describes the probability of reduction, while the second term stands for the temperature dependent diffusion of the reducing agents.
\subsection{Diffusion-controlled growth}
Based on previous experiments \cite{liu2015understanding,crespo2010reversible,crespo2012one,crespo2014changes, babonneau2018real} we suppose that primary tiny NPs, nanoclusters, or "monomers" are formed in the porous matrix due to ultraviolet illumination. A uniform distribution of small and monodisperse NPs is considered in the initial state. Thermal diffusion is considered to be the main initiator of the Ag NP growth in the TiO$_2$ matrix under laser irradiation (Figure \ref{fgr:2all}(h) and (i)), as was shown by Liu et al. \cite{liu2015understanding,liu2016laser}.
To determine the total number of monomers absorbed by an Ag NP of radius $R$ per unit time, the Fick's first law is applied:
\begin{equation}
n_{abs}(t) = 4 \pi R^2 D_{Ag^0}(T) \frac{d C_{Ag^0}}{d r}|_{r=R}
\label{eqn:fick}
\end{equation}
\noindent where the diffusion coefficient $D_{Ag^0}(T)$ is described by the Arrhenius law as follows
\cite{mcbrayer1986diffusion, butrymowicz1974diffusion}:
\begin{equation}
D_{Ag^0}(T) = D_{0Ag^0}\exp \Big( -\frac{E_{Ag^0}}{N_A k_B T} \Big)
\label{eqn:DAg0defi}
\end{equation}
The homogeneous distribution $C_{Ag^0}(r)$ is calculated by solving the steady state diffusion equation \cite{landau1981course}:
\begin{equation}
\frac{1}{r} \frac{\partial ^2}{\partial r^2}rC_{Ag^0}(r) = 0
\label{eqn:densityMonomer}
\end{equation}
\noindent with boundary conditions $C_{Ag^0}(r\to \infty) = C_{Ag^0}(\infty)$ and $C_{Ag^0}(r=R) = C_{Ag^0}(R)$. The analytic solution is obvious:
\begin{equation}
C_{Ag^0}(r) = C_{Ag^0}(\infty) - [C_{Ag^0}(\infty) - C_{Ag^0}(R)] \frac{R}{r}
\label{eqn:densityMonomerSolution}
\end{equation}
\noindent The equilibrium concentration at the surface of the NP is defined by the following thermodynamic equation \cite{liu2015understanding, kaganovskii2007formation, landau1981course, thanh2014mechanisms,kwon2011formation}:
\begin{equation}
C_{Ag^0}(R) = S_{Ag^0} ( 1 + \frac{2 \gamma \omega}{R k_B T} )
\label{eqn:densityMonomerSurface}
\end{equation}
\noindent where, $S_{Ag^0}$ is the Ag$^0$ solubility in mesoporous TiO$_2$ film, $\gamma$ the nanoparticle-matrix interfacial tension, and $\omega$ the atomic volume of Ag in crystal.
Substituting Eq. \ref{eqn:densityMonomerSolution} and Eq. \ref{eqn:densityMonomerSurface} into Eq. \ref{eqn:fick}, the total amount of monomers absorbed by the Ag NP per unit time is given by:
\begin{equation}
n_{abs}(t) = 4 \pi R D_{Ag^0}(T) [C_{Ag^0}^{BS}- S_{Ag^0} ( 1 + \frac{2 \gamma \omega}{R k_B T} )]
\label{eqn:fick2}
\end{equation}
\noindent where $E_{Ag^0}$ the activation energy of diffusion, and $C_{Ag^0}^{BS} \equiv C_{Ag^0}(\infty)$ is the concentration of neutral silver monomers in the matrix.
\subsection{Light absorption by Ag NPs and thermal diffusion}
In addition to re-radiation and injection, the rest of the absorbed laser energy decays via the electron-phonon and phonon-phonon interactions that heat the irradiated region (Figure \ref{fgr:2all}(f,c)). The electron relaxations and electron-phonon coupling occur on the time scale of few picoseconds, while phonon-phonon interactions take up to a few nanoseconds \cite{delfour2015mechanisms, metwally2015fluence, pustovalov2016light, plech2004laser}. Since Ag NPs grow more slowly than the electron-phonon interaction speed, the classical thermal diffusion equation is introduced:
\begin{equation}
C_m \frac{\partial T}{\partial t} =\nabla \cdot ( k_{m} \nabla T) + \alpha_{abs} \cdot I(x,y,z,t)
\label{eqn:thermal}
\end{equation}
\noindent where $k_m$ is the heat conductivity, and $C_m$ the volumetric heat capacity.
Although the absorbed laser energy by a Ag NP is partially relaxed via injection or tunneling, the majority of the energy is finally converted into heat \cite{brongersma2015plasmon, clavero2014plasmon,kawahara2005electron}. It is reasonable to estimate the heating source by taking into account the total absorbed energy. Normally, a laser beam attenuates exponentially in a composite having the mixture of nano-inclusions. Many efforts \cite{malasi2014mie, doyle1989optical, rysselberghe1932remarks} have been devoted to effective medium theories (EMTs) aiming to bridge the Fresnel medium together. Their key common ideas are that, as the concentration of inclusions goes to zero, the effective dielectric function should behave asymptotically as the host medium, and follow the Kramers-Kronig relation \cite{malasi2014mie}. In such media, the Beer-Lambert law describes both light absorption and transmission. Herein, the laser intensity at any position is given by:
\begin{equation}
I(x,y,z,t) = [1 - \Lambda(x,y) ] I_0(x, y, t) \exp (-z \alpha_{abs})
\label{eqn:intensity02}
\end{equation}
\noindent where $\alpha_{abs} = \sigma_{abs} \cdot C_{NP}$ is the absorption coefficient, $\Lambda(x,y)$ the surface reflection, and $I_0(x,y,t) = \frac{P_0}{\pi w_0^2 / 2} \cdot \exp\{-\frac{-2[(x-V_s t)^2+y^2]}{w_0^2}\}$ the Gaussian shape laser intensity at time $t$ and point $(x, y)$, where $x$ is the coordinate along the translation direction on the top surface of the TiO$_2$ film. Here, $C_{NP}$ is the NP concentration, $w_0$ the beam radius at $1/e^2$, $P_0$ the laser power, and $V_s$ the laser scanning speed.
For simplicity, we use the Maxwell-Garnett (MG) theory to estimate the effective index of the thin film. The surface reflection $\Lambda(x,y)$ is then estimated using the Fresnel equation.
\subsection{The coupled equations}
A link between NP concentration $C_{NP}$ and radius $R$ was previously revealed \cite{liu2015understanding} by analysis of \textit{post mortem} HAADF-STEM micrographs of samples produced at different speeds. The dependency is obtained by fitting the experiments according to an empirical law. It is given as follows \cite{liu2015understanding}:
\begin{equation}
C_{NP} = \frac{1}{(a_1 R + a_2)^3}
\label{eqn:CNP}
\end{equation}
\noindent where $a_1 \approx 2.6$ and $a_2 \approx 8.7 \mathrm{nm}$.
The amount of crystallized silver changes in two ways: the variation of size and of the total number of NPs in the volume. Therefore, the differential of concentration is:
\begin{equation}
\frac{d C_{Ag_{cst}} }{d t} = N_{Ag_{cst}} \frac{d C_{NP}}{d t} + C_{NP} [n_{abs}(t) - n_{oxi}(t)]
\label{eqn:ConcenAgcrystal}
\end{equation}
\noindent where $N_{Ag_{cst}} = 4 \pi R^3 / 3 \omega$ is the number of crystallized Ag atoms in a nanoparticle. For each nanoparticle, the variation of total number of atoms is due to growth and oxidation. Therefore, the radius of a nanoparticle changes as:
\begin{equation}
\frac{d R}{d t} = [n_{abs}(t) - n_{oxi}(t)] \cdot \frac{\omega}{4 \pi R^2}
\label{eqn:dRnp}
\end{equation}
\noindent Thus, the amount of crystallized Ag atoms in the studied unit cell varies due to the changes of NP concentration as \cite{liu2015understanding}:
\begin{equation}
\begin{aligned}
N_{Ag_{cst}} \frac{d C_{NP}}{d t} &= N_{Ag_{cst}} \frac{d C_{NP}}{d R} \frac{d R}{d t} \\
& = \frac{R}{3} [n_{abs}(t) - n_{oxi}(t)] \frac{d C_{NP}}{d R}
\label{eqn:dNcst}
\end{aligned}
\end{equation}
\noindent where, $C_{NP}$ is the concentration of NPs in the unit cell. The right part of Eq. \ref{eqn:dNcst} implies its contribution to oxidation. The mass conservation of Ag in an unit cell leads to the following equations \cite{liu2015understanding}:
\begin{equation}
\frac{d C_{Ag^+}}{d t} =n_{oxi}(t) [ C_{NP}(t) + \frac{R(t)}{3} \frac{d C_{NP}} {d R} ]
- \frac{d C_{Ag^+}}{d t} |_{red}
\label{eqn:ConcenAg1}
\end{equation}
\begin{equation}
\frac{d C_{Ag^0}^{BS}}{d t} = - n_{abs}(t) [ C_{NP}(t) + \frac{R(t)}{3} \frac{d C_{NP}} {d R} ]
+ \frac{d C_{Ag^+}}{d t} |_{red}
\label{eqn:ConcenAg0}
\end{equation}
\subsection{The extended 2D model}
Eq. \ref{eqn:dRnp} to Eq. \ref{eqn:ConcenAg0} describe the full coupled optical, thermal, photo-oxidizing and growing field. The Arrhenius description of diffusion coefficients implies a nonlinear system that makes the 3D calculation extremely time-consuming.
The extended 2D model provides an appropriate approach with less computational efforts to simulate the multi-scale problem and access the spatial information of NP size and temperature in a sample. We simulate the process in cross-section, which consists of two coupled steps. In the first step, the temperature field is computed by FEM. With the knowledge of the heat distribution, the growth, oxidation, and reduction are calculated adaptively by using the Bulirsch-Stoer algorithm \cite{press2007section}. The approach provides second-order accuracy with comparatively few calculation steps. The forecasting time step at each grid is obtained by the Bulirsch-Stoer algorithm. The minimum time step is used for the calculation of temperature and growth in the next step. In this way, few computational efforts are required and the accuracy is guaranteed. The studied structure consists of a TiO$_2$ film 200nm thick on top of a glass (Figure \ref{fgr:3all}(a)). The effective refractive index of the mesoporous titania film is estimated to $n = 1.7$ and the initial diameter of Ag NPs homogeneously distributed in the film is assumed to be 3 nm \cite{liu2015understanding}. The width of the structure is set to $200 \mu m$ and the laser scans the sample from $x = -50 \mu m$ to $x = 50 \mu m$ to avoid the boundary absorption of heat (Figure \ref{fgr:3all}(b)). Figure \ref{fgr:3all} shows the simulated results on the top surface (as illustrated in Figure \ref{fgr:3all} (c)) of the TiO$_2$ film at a laser writing speed of 50 $\mathrm{\mu m/s}$. Table \ref{table:Chapter5a} summarizes the parameters used in simulations.
\begin{table*}[ht!]
\centering
\caption{A summary of the parameters in simulation}
\begin{tabular}{llll}
\hline
\hline
Parameter & Description & Value or definition &Reference\\
\hline
$n$ & optical index of TiO$_2$ film & 1.7 & Ref. \cite{liu2015understanding}\\
$P_0$ & laser power & 180mW & \\
$w_0$ & waist radius & $\mathrm{10 \mu m}$&\\
$h_f $ & thickness of TiO$_2$ film & 200 nm & Ref. \cite{baraldi2016polarization}\\
$\omega$ & atomic volume of crystallized Ag & 10.27 $\mathrm{cm^3/mol}$ &Ref. \cite{singman1984atomic}\\
$E_{Ag^0} $ & activation energy of Ag$^0$ diffusion & 0.95eV &Ref. \cite{mcbrayer1986diffusion, butrymowicz1974diffusion, liu2015understanding}\\
$D_{0Ag^0}(500K) $ & Ag$^0$ diffusion coefficient at 500K &0.6$\times 10^{-14} \mathrm{m^2/s}$ &Ref. \cite{mcbrayer1986diffusion, butrymowicz1974diffusion, liu2015understanding}\\
$\eta _0$ & ionization efficiency & 2.3 $\times 10^{-5}$ &Ref. \cite{liu2015understanding, crespo2010reversible}\\
$E_p$ & activation energy of reduction & 0.037eV &Ref. \cite{liu2015understanding, kaganovskii2007formation}\\
$C_{red} $ & concentration of reducing agent (RA) & 2.258 $\mathrm{\times 10^{-4} mol/m^3}$ &\\
$E_D $ & activation energy of RA diffusion & 0.4eV &Ref. \cite{haran1998determination, torresi1987hydrogen}\\
$D_0^{red}(500K) $ & RA diffusion coefficient at 500K & 2.5$\times 10^{-14} \mathrm{m^2/s}$ &Ref. \cite{haran1998determination, torresi1987hydrogen}\\
$S_{Ag^0} $ & solubility of Ag$^0$ & $1.66 \mathrm{mol/m^3}$ &Ref. \cite{liu2015understanding, kaganovskii2007formation, prudenziati2004dissolution}\\
$\gamma$ &Ag NP - TiO$_2$ interfacial tension & $0.506 \mathrm{J/m^2}$ &Ref. \cite{liu2015understanding, kaganovskii2007formation}\\
$C_{Ag^0}^{BS}(t=0 s)$ & initial Ag$^0$ concentration in matrix & $1.495 \mathrm{\times 10^{4} mol/m^3}$ &Ref. \cite{liu2015understanding, liu2016laser}\\
$C_{Ag^+}(t=0 s)$ & initial Ag$^+$ concentration in matrix & $0.996 \mathrm{\times 10^{4} mol/m^3}$ &Ref. \cite{liu2015understanding, liu2016laser}\\
$T(t=0s)$ & initial temperature & 300K &\\
$R(t=0s)$ & initial radius & 1.5nm &Ref. \cite{liu2015understanding}\\
\hline
\hline
\label{table:Chapter5a}
\end{tabular}
\end{table*}
\section{Simulation results}
The 2D model clearly demonstrates the spatial inhomogeneity during the kinetic process (Figure \ref{fgr:3all}). For instance, as one can see in Figure \ref{fgr:3all}(j), the NP's spatial-size-distribution (SSD) varies from uniform (3 nm in diameter) at the initial stage to non-uniform upon laser writing with big NPs formed at the beginning of the line (43 nm) and under the laser beam (20 nm), whereas the final NP size at the middle of the laser line is 6.8 nm after laser irradiation. Simultaneously, the temperature profile (Figure \ref{fgr:3all}(e)) changes with its maximum value rapidly increasing to 870 $\mathrm{^{\circ} C}$ at the beginning and more slowly decreasing to a steady one at 528 $\mathrm{^{\circ} C}$ after translating the sample by about 25 $\mu m$. Figure \ref{fgr:3all}(f) shows the laser during the writing.
\begin{figure*}[!ht]
\centering
\includegraphics[width=0.9\textwidth]{./figuresChapter5/figure_3_v2}
\caption{Simulations of the extended 2D model. The structure used in calculations (a), and an example of the temperature distributions at different times (b). Time variation maps of NPs size (d), temperature (e), normalized light intensity (f) along the translation of the laser beam at the top surface (c) of the film. (g), (h), (i) and (j) stand for the different states of the sample at four different times. All simulations are carried out at a scan speed of 50 $\mu m/s$.}
\label{fgr:3all}
\end{figure*}
To clarify the whole process from the beginning, the NP's SSD (solid green line), temperature (solid red line) and laser profiles (dashed blue line) are plotted in the same figures at various times (Figure \ref{fgr:3all}(g) to (j)). The curves at different times clearly show the main stages during laser writing. At the first step, NPs inside the laser spot grow rapidly due to the positive feedback (Figure \ref{fgr:3all}(g) and (h)). Accordingly, the temperature increases dramatically to 870 $\mathrm{^{\circ} C}$ as illustrated in Figure \ref{fgr:3all}(d) and (h)). After then, the shrinkage of NPs by oxidation dominates the process. The positive correlation between NPs size and absorption leads to the temperature decrease. As a matter of fact, the decreasing of temperature slows down the reaction speed, which in turn decreases the maximum size in the laser front edge (Figure \ref{fgr:3all}(i)). The negative feedback continues until a steady state sets in. Afterward, the stationary stage persists till the end as plotted in Figure \ref{fgr:3all}(i) to (j). Indeed, the above processes are valid for all scanning speeds that are lower than the maximum speed for NP growth. It is obvious that for high writing speeds that are comparable with the reaction rate, there is less time for NPs to grow. In this case, the laser-induced NP growth is not expected. The discussion will be presented in the next section.
Transforming the coupled equations into a moving coordinate system (MCS) \cite{jensen1980solution}($x=x'-\int_{0}^{t} Vs(t')dt'$ and $t=t'$) implies the steady-state equations (see Eq. \ref{eqn:Chapter5supeq1}). The stationary solutions depend on moving boundary conditions that are computationally difficult by FEM. The transient results of the extended 2D model reveal the existence of these kinds of solutions for various writing speeds (as shown in Figure \ref{fgr:3all} and Figure \ref{fgr:4all}).
\begin{equation}
\left\{
\begin{aligned}
\boldsymbol{V_s} \cdot \nabla R + \frac{(n_{abs} - n_{oxi})\omega}{4 \pi R^2} &= 0 \\
\boldsymbol{V_s} \cdot \nabla C_{Ag^0}^{BS} - n_{abs}[C_{NP} + \frac{1}{3} R \frac{d C_{NP}} {d R}] + \frac{d C_{Ag^+}}{d t} |_{red} &= 0 \\
\boldsymbol{V_s} \cdot \nabla C_{Ag^+} + n_{oxi} [C_{NP} + \frac{1}{3} R \frac{d C_{NP}} {d R}] - \frac{d C_{Ag^+}}{d t} |_{red} &=0 \\
C_m \boldsymbol{V_s} \cdot \nabla T+ \nabla \cdot ( k_{m} \nabla T) + \alpha_{abs} \cdot I(x, y, z) &= 0
\label{eqn:Chapter5supeq1}
\end{aligned}
\right.
\end{equation}
\noindent where $\boldsymbol{V_s}$ is the moving speed of the laser at laboratory coordinate.
The photo-oxidation process is particularly important in the understanding of the speed-dependent final size. For all writing speeds, as discussed above, the NPs size profiles are nearly the same in the first stage owing to their rapid growth. In fact, the positive feedback stops as soon as the free Ag$^0$ atoms in matrix are exhausted as implied by Eq. \ref{eqn:fick2}. After size-shrinkage and temperature decrease, the stationary states are set with the total amount of photo-oxidized Ag$^0$ depending on the integration time of oxidation. In other words, there is more time at low scanning speeds to oxidize the NPs, so that lower temperatures are reached. The SSD width is broader at the beginning and thinner at stationary state as shown in Figure \ref{fgr:4all} (a) to (d). Note that reduction has little influence on the growth due to the low concentration of reducing agent, as reported by Liu et al.\cite{liu2015understanding}.
To get rid of the rapid growth at the beginning of the laser writing and to consider only the steady state behavior of the sample, the studied position is set at $x = 30 \mu m$ (shown in Figure \ref{fgr:4all}(e)) with its distance 80$\mu m$ away from the laser center at $ t = 0 s$. Figure \ref{fgr:4all}(f) plots the time variations of the NP size in this location for various writing speeds. The growth process lasts from a few tenths of a second to seconds, and the growth rate depends on the writing speed. According to Eq. \ref{eqn:fick2}, the temperature strongly affects the growth rate. This means that the sample experiences different temperatures at different speeds, which is clearly shown in Figure \ref{fgr:4all}(g). In all cases, the Ag NP growth is comparatively much slower than what was shown in reference \cite{liu2015understanding} that ignored the collective behavior of NPs and the thermal diffusion.
\begin{figure*}[!htb]
\centering
\includegraphics[width=0.95\textwidth]{./figuresChapter5/figure_4_v3}
\caption{Time variation of the NP size along the translation direction at various writing speeds (a) to (d). To compare results at various speeds, the time variation of NP size (f) at the studied position (e) are plotted. Dashed lines show when the laser beam reaches the studied location.}
\label{fgr:4all}
\end{figure*}
Heat diffusion is strongly contributes to the inhomogeneous SSD of NPs. According to Figure \ref{fgr:4all}(f), for all writing speeds, NPs start to grow even before the laser arrives and the size increase is stopped before the maximum intensity is reached. It is clear that the heat diffusion due to phonon-phonon interaction is faster than the growth by atoms diffusion. Thus, the temperature has broader profiles than NP's SSD. That is why the NPs in the laser vicinity are pre-heated and once the temperature exceeds the activation threshold, NPs start growing. Nevertheless, the temperature-dependent growth rate is faster than the studied writing speeds. Otherwise, the end of the growth process would be expected after reaching the maximum of temperature, or laser intensity.
\begin{figure*}[ht!]
\centering
\includegraphics[width=0.9\textwidth]{./figuresChapter5/figure_5_v5}
\caption{Simulated NP size and transmission along the writing line at the end of the calculation for various writing speeds. NP size (a). Sample transmission calculated by Maxwell-Garnett (b) or Mie coupled with Clausius-Mossotti model (c). Calculated transmission maps by MG (d) and CM-Mie (e) model. Minimum transmission versus writing speed for the two models (f). Comparison of transmission by simulation (MG and CM-Mie) and in-situ experiment (g). The dashed black lines stand for the laser profiles.}
\label{fgr:5all}
\end{figure*}
\subsection{Comparison with experiments}
Experimentally, the transmissions are measured in the steady state far from the starting positions. Herein, the discussion is then limited to the comparison with simulated results at steady-state. For each writing speed, we obtain the NPs' size profile after the laser moves $100 \mu m$. The NP's SSDs are shown in Figure \ref{fgr:5all}(a). Here, laser moves toward the positive x-axis with its intensity profile shown by the dashed black line. As discussed above, the maximum sizes appear before the laser-center arrives at all the writing speeds. To compare with the \textit{in-situ} experiments, transmission profiles are retrieved based on two different effective-medium models. The Maxwell-Garnett approximation is commonly used in literature \cite{pedrueza2011novel}. However, it is only valid for the NPs whose sizes are much smaller than laser wavelength \cite{malasi2014mie}. For larger NPs, compared with MG, the Mie theory associated with Clausius-Mossotti equation \cite{rysselberghe1932remarks} provides a size-dependent correction by accounting for the multipole contributions (CM-Mie) \cite{malasi2014mie, doyle1989optical}. The transmission profiles at wavelength 527 nm are calculated according to the Fresnel equation under normal incidence:
\begin{equation}
\begin{aligned}
&\mathrm{Trans} = \frac{4 Z_{air} Re(Z_{sub}) }{|Z_{air} B + C|^2} \\
\mathrm{where} \\
\left[
\begin{array}{c}
B\\
C
\end{array}
\right]
&=
\left[
\begin{array}{cc}
cos \delta _m & j sin \delta _m / Z_m \\
j Z_m sin \delta _m & cos \delta _m
\end{array}
\right]
\left[
\begin{array}{c}
1\\
Z_{sub}
\end{array}
\right] \\
\delta _m &= 2 \pi n_m h_f / \lambda
\label{eqn:fresnel}
\end{aligned}
\end{equation}
\noindent where $Z_i$ is the media impedance (i = air, effective medium or glass substrate), $h_f = 200 nm$ the film thickness, and $n_m$ the effective index.
Figure \ref{fgr:5all}(b) and (c) show the retrieved transmission profiles at various writing speeds based on MG and CM-Mie respectively. Though the value of transmission by MG and CM-Mie differ, the trends are quite similar: the transmission decreases in the laser front edge before growing back inside the irradiated area. No matter which EMT model is used, the position at transmission minimum is always located at the laser front edge for every writing speed. The calculations show that the transmission minimum is close to the location of the largest NPs in the laser front edge.
Furthermore, the transmission maps calculated using MG (Figure \ref{fgr:5all}(d)) and CM-Mie model (Figure \ref{fgr:5all}(e)) have similar trends with the results of the in-situ experiment shown by Figure \ref{fgr:1all}(c). Though the writing speed in our experiments and simulations can differ, the transmission dip (shown by green or blue color in the maps) is becoming broader and moves farther toward the laser front edge when the writing speed is increased.
In addition, the transmission minimum decreases and saturates while increasing the writing speed (Figure \ref{fgr:5all}(f)). These results are consistent with the in-situ experiments as shown in Figure \ref{fgr:5all}(g). The location-dependent transmissions are due to the NPs formation along the laser scanning direction. Nevertheless, we note that the value of the transmission in the simulation and experiment are not exactly the same. The differences are probably due to several modeling assumptions. For example, a single-size model was used, while the size of Ag NPs is dispersed. Secondly, the effective medium theory ignores the NP-NP interactions, which can also affect the electromagnetic response if nanoparticle density is getting high. Interestingly, the transmission obtained by using the CM-Mie model is underestimated. In principle, CM-Mie is an improvement that should work better than the MG model by taking into account the contributions of multipoles, such as quadrupole, octupole and etc. These differences are shown in Figure \ref{fgr:5all}(d), (e), and (g) indicating that the effective medium theory could only approximately describe the true composition of the porous thin films with a few layers of NPs inside.
\begin{figure}[ht]
\centering
\includegraphics[width=0.4\textwidth]{./figuresChapter5/figure_6a}
\includegraphics[width=0.4\textwidth]{./figuresChapter5/figure_6b}\\
\includegraphics[width=0.4\textwidth]{./figuresChapter5/figure_6c}
\caption{Influences of the activation energy of free Ag$^0$ diffusion ($\mathrm{\eta _0 = 2.3 \times 10^{-5}}$). The final NP size after laser scanning (a) and the maximum size in the laser front (b) for various activation energy, and (c) the spatial distribution of NP size with $E_{Ag^0} = \mathrm{0.95eV}$.}
\label{fgr:6all}
\end{figure}
\subsection{Activation energy of diffusion}
The activation energy of metal ion diffusion in the studied system affects NP's growth threshold. The exact activation energy in the mesoporous TiO$_2$ film is not known to the best of our knowledge. The activation energy is estimated from the Ag diffusion in silicon dioxide and copper, which gives the value around 1.0 eV \cite{liu2015understanding,mcbrayer1986diffusion, butrymowicz1974diffusion}. Thus, a reasonable value for Ag in mesoporous TiO$_2$ ranges from 0.5eV to 1.5 eV \cite{liu2015understanding}.
To understand how the activation energy influences the NP growth, the NP final size and maximum size in the laser front edge are calculated at various scanning speeds with different activation energies. They are shown in Figure \ref{fgr:6all}(a) and (b). Interestingly, the maximum size never exceeds 4 nm for activation energy $E_{Ag^0} =\mathrm{1.15eV}$. In this case, the activation energy is too high to allow the rapid positive feedback to appear. The size increase and saturation occur for lower energies. Moreover, the activation energy never alters the final size and maximum size profiles except for the high writing speeds. It means that the activation energy has a little impact on the saturation process. Instead, an increase of the activation energy decreases the growing speed that has a great influence on the up-threshold of writing speed. To understand what happens at high speed, NP size profiles are plotted in Figure \ref{fgr:6all}(c) at various scanning speeds along x-axis. At high speed the growth never starts since there is not enough time for NPs in a rapid moving laser spot (e.g. $\mathrm{600 \mu m/s}$) to grow large enough to activate the feedback promptly.
\subsection{Oxidation rates}
The photo-oxidation process strongly affects the steady-state parameters. Figure \ref{fgr:7all} shows the dependence of the NP final size and maximum size on the writing speed at various oxidation rates. The two size-profiles saturate at lower speeds while decreasing the oxidation rate. The result again confirms the fact that the oxidation process contributes greatly to the scanning speed dependent phenomenon. Once the rapid feedback is activated, the fast growth of NPs never stops until the concentration of Ag$^0$ decreases to the value close to Ag$^0$ solubility. This is also the reason why the activation energy ($E_{Ag^0}$) never alters the saturation profile. On the contrary, the oxidation leads to the size-shrinkage with an amount inversely proportional to the scanning speed. The suggested NP size increases while decreasing the oxidation rate for the same scanning speed. Owing to the positive correlation between absorption and size, lower oxidation rates saturate at lower speeds. Furthermore, oxidation affects the positive feedback. For a given activation energy, an increase of the oxidation rate means more time is required for a NP to grow above a threshold size. This is the case for oxidation rate $\eta_0 = 1.6 \times 10^{-5}$ and $\eta_0 = 2.8 \times 10^{-5}$ at $\mathrm{600\mu m/s}$. For the high oxidation rate $\eta_0 = 9.2 \times 10^{-5}$ as shown in Figure \ref{fgr:7all}, the rapid positive feedback never appears.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.43\textwidth]{./figuresChapter5/figure_7a}
\includegraphics[width=0.43\textwidth]{./figuresChapter5/figure_7b}
\caption{Influence of ionization efficiency ($\mathrm{E_{Ag^0} = 0.95 eV}$). Final NP size after laser writing (a) and Maximum size in the laser front edge (b).}
\label{fgr:7all}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[width=0.55\textwidth]{./figuresChapter5/figure_DEAg0_p1}
\caption{Influence of Ag$^0$ diffusion coefficient ($\mathrm{E_{Ag^0} = 0.95 eV}$, and $\mathrm{\eta_0 = 2.3 \times 10 ^{-5}}$). }
\label{fgr:chapter5_DEAg0_p1}
\end{figure}
\subsection{Diffusion coefficient at 500K}
The Ag $^0$ diffusion coefficient is important to silver NP's growth speed. In the previous discussions we have seen the abrupt descending of the NP size (as shown in Figures \ref{fgr:6all} and \ref{fgr:7all}) at fast writing speeds. The mechanism has been attributed to the low growth speed. In porous material, especially in mesoporous TiO$_2$ thin films, the diffusion coefficient of silver can be larger than it is inside the bulk material, such as the silicon, and fused silica. Based on this assumption, we perform simulations by increasing the diffusion coefficient of silver at 500K. Figure \ref{fgr:chapter5_DEAg0_p1} shows the final NP size over different writing speeds. The abrupt descending tends to a right move to the high-speed end while increasing the diffusion coefficient, which disappears for $\mathrm{D_{0Ag^0}(500K) = 1.2 \times 10 ^{-14} m^2/s}$ and $\mathrm{D_{0Ag^0}(500K) = 4.8 \times 10 ^{-14} m^2/s}$ even at the highest writing speed studied (namely, $\mathrm{1000\mu m /s}$). Moreover, the curve of the final NP size with speed does not depend on the silver diffusion coefficient. The mechanism lies in the fact as we discussed previously that the photo-oxidation is the key role in laser writing-speed controlled phenomena. Due to the plasmonic size-dependent absorption, once the NPs inside the laser spot start to grow, the increment in size leads to a higher power absorption. The increased temperature further speed up the growth process. As a result, the growth is quite non-linear and it happens before the oxidation be a role. The abrupt decrements in the Figure \ref{fgr:chapter5_DEAg0_p1} are because of the insufficient time for the size-accumulation to activate the rapid growth as shown by Figure \ref{fgr:6all} (c).
\begin{figure}[ht!]
\centering
\includegraphics[width=\textwidth]{./figuresChapter5/figure_CAg0Ag1}
\caption{Concentrations of silver in the forms of ions and free atoms in matrix ($\mathrm{E_{Ag^0} = 0.95 eV}$, $\mathrm{\eta_0 = 2.3 \times 10 ^{-5}}$, and $\mathrm{D_{0Ag^0}(500K) = 0.8 \times 10 ^{-14} m^2/s}$). Laser scans from left to the right. The initial concentrations of silver ions and atoms are shown in Table \ref{table:Chapter5a}.}
\label{fgr:chapter5_CAg0Ag1}
\end{figure}
\subsection{Concentrations of free Ag$^0$ and Ag$^+$}
The previous discussions have shown that silver NPs starts to grow even before the laser center arrives and it never stops until the majority of the Ag$^0$ atoms in the matrix are consumed. To better illustrate the underlying process, here we plot spatial distributions of Ag$^+$ ions, Ag$^0$ atoms and NP size alongside with laser profile in Figure \ref{fgr:chapter5_CAg0Ag1}. In the regime of laser-activated rapid growth of NPs (namely, Figure \ref{fgr:chapter5_CAg0Ag1} (a) and (b)), the thermal diffusion allowed NPs' growth appears at positions more than 30$\mu m$ ahead of the laser beam center. This, on the other hand, can easily be seen from the drop-down positions of the Ag$^0$ concentration curves. In the region from $x = 60 \mu m$ to $x = 80 \mu m$, no changes are observed for the Ag$^+$ concentration. Thus, this region only exists the process of thermal-activated growth of silver NPs. The size of Ag NPs reaches the maximum value until it exhausts the majority number of the Ag$^0$ atoms due to the fast growth speed. Thereafter, the growth is nearly stopped when the laser arrives and the photo-oxidation tends to dissolve the NPs into Ag$^+$ ions. The green curves represent the ratio of the real time quantity and the initial number of Ag$^+$ ions. The maximum sizes of the Ag NP ahead of the laser beam are close for the two scanning speeds ($\mathrm{V_S} = 70 \mu m/s, 500 \mu m/s$). After passing-through of the laser beam, the quantity of Ag$^+$ is increased by a factor of 1.26 and 2.05 for $\mathrm{V_S} = 500 \mu m/s$ and $\mathrm{V_S} = 70 \mu m/s$, respectively. The slower writing speed has longer time for the oxidation to shrink the NPs, therefore, larger amount of ions are observed. in addition, no abrupt transformations of Ag$^+$ into Ag$^0$ are presented in the simulations. This is because of the low efficiency of the reduction process studied here (Table \ref{table:Chapter5a}, Eq. \ref{eqn:reduction}). Therefore, the process is not the traditional photochromism \cite{naoi2004tio2,crespo2010reversible}. The samples are changed and are not reversible unless efficient methods are introduced to pump-up the Ag$^0$ quantity in the matrix after the laser processing.
Figure \ref{fgr:chapter5_CAg0Ag1} (c) explains the abrupt decrements of the NP size curve (green dashed line in Figure \ref{fgr:chapter5_DEAg0_p1}) at $\mathrm{V_S} = 800 \mu m/s$. The process is not stable and can not be discussed using the stationary-state descriptions used above. However, it is clear that the size of Ag NP is increased when the laser scans from left to the right. The decreasing concentration of Ag$^0$ along the X-axis shows the accelerating of the growth process, which is thermal activated as we discussed previously. According to Mie absorption, for the studied Ag size range (3 nm to 50 nm in diameter) at 532 nm wavelength, the larger the size the higher the absorption is. The higher temperature results to the faster diffusion of Ag$^0$ to be absorbed by a NP. Due to the limited computational resources, the simulated distances for the laser scanning is set to 100 $\mu m$. The distances required for the laser to scan before activating the rapid growth at $\mathrm{V_S} = 800 \mu m/s$ is unknown.
\begin{figure}[ht!]
\centering
\includegraphics[width=\textwidth]{./figuresChapter5/figure_CAg0CAg1_p2}
\caption{Simulated NP size at maximum in the laser front (a) and at final stage after laser scans (b) for different initial concentrations of Ag$^+$ and Ag$^0$ in TiO$_2$ thin films. (c) the influence of initial $C_{Ag^0}$ by increasing the value from $1.495 \times 10^4 mol/m^3$ to $1.528 \times 10^4 mol/m^3$, as a comparison to Figure \ref{fgr:chapter5_CAg0Ag1}(c). (Note that $\mathrm{E_{Ag^0} = 0.95 eV}$, $\mathrm{\eta_0 = 2.3 \times 10 ^{-5}}$, and $\mathrm{D_{0Ag^0}(500K) = 0.8 \times 10 ^{-14} m^2/s}$).}
\label{fgr:chapter5_CAg0Ag1_p2}
\end{figure}
The influences by initial concentrations of Ag$^+$ and Ag$^0$ are shown in Figure \ref{fgr:chapter5_CAg0Ag1_p2}. From the previous discussions we find that once the process is activated, the nonlinear growth of NPs never stops until the initial amount of Ag$^0$ are dramatically consumed. This means the maximum size in the laser front should be larger while increasing the concentration of Ag$^0$, which is confirmed by the simulation as shown in Figure \ref{fgr:chapter5_CAg0Ag1_p2}(a). Moreover, since the size-shrinkage is monotonically decreasing with writing speed of the laser, the final NP is also increased (Figure \ref{fgr:chapter5_CAg0Ag1_p2}(b)). The initial Ag$^+$ ion concentration has negligible impacts on the studied model (Figure \ref{fgr:chapter5_CAg0Ag1_p2}(a-b)). This illustrates the inefficient reduction to covert the Ag$^+$ to Ag$^0$. In addition, the increments of Ag$^0$ concentration accelerates the growth speed in average. Comparing the dashed blue curve in Figure \ref{fgr:chapter5_CAg0Ag1_p2}(a) to the dashed green curve in Figure \ref{fgr:chapter5_DEAg0_p1} at point $V_S = 800 \mu m/s$, it is obvious that the maximum size in the laser front increases from 5.0 nm ($C_{Ag^0} = 1.495 \times 10^4 mol/m^3$) to 50.2 nm ($C_{Ag^0} = 1.528 \times 10^4 mol/m^3$). The spatial distribution of Ag$^+$ and Ag$^0$ concentration, and Ag NP size that are plotted in Figure \ref{fgr:chapter5_CAg0Ag1_p2} (c) clearly demonstrate the nonlinear activation of the growth process. Because the positive feedback of size increase with light absorption, it results to the rising of the maximum size in laser front during laser scanning. At first, the size increases very slowly as shown both in Figure \ref{fgr:chapter5_CAg0Ag1} (c) and Figure \ref{fgr:chapter5_CAg0Ag1_p2} (c). However, the averaged growth speed is larger for $C_{Ag^0} = 1.528 \times 10^4 mol/m^3$ than the case of $C_{Ag^0} = 1.495 \times 10^4 mol/m^3$, so that the Ag size in the laser front can be larger and once it exceeds the specific threshold size, the rapid growth is observed as shown in Figure \ref{fgr:chapter5_CAg0Ag1_p2} (c).
\section{The relation of temperature-rise and writing speed}
The first empirical relation of the temperature-rise and the laser writing speed \cite{liu2016selfthesis} comes from the observations of TiO$_2$ phase changes. As reported \cite{kholmanov2003influence, ocana1992low, yuan1995characterization, mosaddeq2000refractive}, a thermal annealing of initial amorphous TiO$_2$ prepared by processing of sol-gel \cite{yuan1995characterization, bersani1998raman, vinogradov2014low}, physical and chemical vapour deposition\cite{goossens1998gas}, and colloidal chemistry methods \cite{ma1998investigation} results to the phase transition of amorphous-to-anatase at 400 - 600 $^{\circ} C$ and anatase-to-rutile at 800 - 1000$^{\circ} C$. Indeed, the temperature at phase transitions differs from different processes and conditions \cite{bersani1998raman, kholmanov2003influence, sharma2019crystal, yanagisawa1999crystallization}. The phase transition from anatase to rutile normally appears at higher temperature than it from amorphous to anatase.
\subsection{Raman spectrum acts as the signature of temperature history}
In the studied laser writing Ag embedded TiO$_2$ thin films, the phase of TiO$_2$ can be estimated by the Raman spectroscopy. Figure \ref{fgr:chapter5_tempRise_raman} shows the Raman spectra of the laser irradiated samples over different writing speeds. The experimental results are from Ref. \cite{liu2016selfthesis}, which were carried out by Zeming Liu at Laboratoire Hubert Curien under the supervision of Nathalie Destouches. The characteristic peaks of anatase phase are 144 , 197 and 639 cm$^{-1}$ at E$_g$ modes, 399 and 519 cm$^{-1}$ at B$_{1g}$ modes, and 513 cm$^{-1}$ at A$_{1g}$ mode \cite{chen2007titanium, swamy2005finite, ohsaka1978raman}. The fundamental peaks of rutile phase are E$_g$ mode (447 cm$^{-1}$), B$_{1g}$ (143 cm$^{-1}$), A$_{1g}$ mode (612 cm$^{-1}$), and B$_{2g}$ (826 cm$^{-1}$) \cite{porto1967raman, swamy2006size}. It is clear that the initial sample before laser treatment is amorphous. The appearance of the peaks at anatase phase starts from writing speed of Vt + 5 $\mu m/s$, where Vt stands for the threshold speed of NP growth \cite{liu2016selfthesis}. The peaks are progressively distinct while increasing the writing speed. The rutile phase appears from Vt + 1,700 $\mu m/s$. The anatase peaks are completely disappear at Vt + 2,300 $\mu m/s$. Based on the above analyses, the temperature-rise inside the Ag:TiO$_2$ film shows a positive correlation with the writing speed.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.85\textwidth]{./figuresChapter5/figure_raman_01}
\caption{Raman spectra of samples after different laser writing speeds. The excitation wavelength for Raman characterization is 633 nm. The experimental data are taken from Ref. \cite{liu2016selfthesis}. }
\label{fgr:chapter5_tempRise_raman}
\end{figure}
A fine illustration of the temperature-Vs relation shown in the range from Vt + 5 $\mu m/s$ to Vt + 2,100 $\mu m/s$ can be revealed by further exploitation of the size effect of anatase TiO$_2$ on Raman spectra \cite{chen2007titanium, zhang2000raman, li2005raman, bersani1998raman, ivanda1999effects}. The size influence on Raman lineshape comes from the phonon confinement that leads to the breakdown of the phonon momentum selection rule at $q \approx 0$\cite{zhang2000raman}. This is the case for nanocrystals due to the spatially confined phonons all over the Brillouin zone, which will contribute to the first-order Raman scattering \cite{zhang2000raman, chen2007titanium, li2005raman}. For spherical nanocrystals at first-order scattering, the Raman intensity is estimated as \cite{tiong1984effects, shen1984raman, zhang2000raman, li2005raman, bersani1998raman, ivanda1999effects}:
\begin{equation}
I(\omega) \propto \int_{BZ}\frac{|C(0,\mathbf{q})|^2}{(\omega-\omega(\mathbf{q}))^2 + (\Gamma_0/2)^2} d^3 \mathbf{q}
\label{eqn:phonon-confinement_p1}
\end{equation}
\noindent where $\omega$ is the Raman shift expressed in wavenumber, $\omega(\mathbf{q}) = \omega_0 + \Delta [1 - cos(qa)]$ is the phonon dispersion function, $a = 0.3768$ nm is the lattice parameter, $\Delta = 20 cm^{-1}$ is the width of the phonon dispersion curve, $\omega_0 = 144 cm^{-1}$ is the zone center phonon frequency, $|C(0,\mathbf{q})|^2 = \exp(-q^2 D^2 / 16 \pi^2)$ is the confinement function, $D$ is the size of the nanocrystals, $\Gamma_0 = 7 cm^{-1}$ is the Raman linewidth at room temperature (the constant value of 7 $cm^{-1}$ corresponds to bulk anatase), and $d^3 \mathbf{q} \propto q^2 dq$ for a three-dimensional confinement that is suitable for power, and quantum dots; $d^3 \mathbf{q} \propto q dq$ for quantum wires (two-dimensional confinement); and $d^3 \mathbf{q} \propto dq$ for quantum wells (one-dimensional confinement) \cite{li2005raman, chen2007titanium}. For nanocrystals, the three-dimensional confinement is the most convincing model \cite{li2005raman, chen2007titanium, lei2001fabrication, kelly1997raman, swamy2005finite, barborini2002engineering}.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.9\textwidth]{./figuresChapter5/figure_raman_03}
\caption{Raman spectra by phonon confinement model of various nanocrystal sizes (a). Experimental Raman spectra of amorphous TiO$_2$ samples after annealing at various temperature, the data are taken from Ref. \cite{sharma2019crystal}, and are normalized in the range of 120-180 cm$^{-1}$ as $I_{norm}(\omega) = \frac{I(\omega) - min[I(\omega)]}{max\{I(\omega) - min[I(\omega)]\}}$ (b). The relation of TiO$_2$ nanocrystal size and annealing temperature (c). The crystallized size are obtained by fitting the experimental data from (b) to phonon confinement model. The insert-map of (c) shows the fitting curve (blue) along with the experimental data (red dot).}
\label{fgr:chapter5_tempRise_raman_p1}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[width=0.85\textwidth]{./figuresChapter5/figure_raman_02}
\caption{Raman spectra in the range of 100-200 cm$^{-1}$ of samples after different laser writing speeds (a). The excitation wavelength for Raman characterization is 633 nm. The experimental data are taken from Ref. \cite{liu2016selfthesis}, and are normalized in the range of 100-200 cm$^{-1}$ as $I_{norm}(\omega) = \frac{I(\omega) - min[I(\omega)]}{max\{I(\omega) - min[I(\omega)]\}}$ (b).}
\label{fgr:chapter5_tempRise_raman_p2}
\end{figure}
The typical Raman spectra calculated by the three-dimensional phonon confinement model with different size of TiO$_2$ nanocrystals are shown as Figure \ref{fgr:chapter5_tempRise_raman_p1} (a). As the nanocrystal size decreases, the spectrum is broadened asymmetrically and shifted towards the higher frequency end. These features can be used for estimating the TiO$_2$ nanocrystal size by fitting the theoretical Raman intensity with the experiments \cite{sharma2019crystal} of annealing initial amorphous samples at various temperatures (see Figure \ref{fgr:chapter5_tempRise_raman_p1}(b-c)). The experiments were carried out by Nipun Sharma at Laboratoire Hubert Curien under the supervision of Nathalie Destouches \cite{sharma2019crystal}. The samples calcined at temperature between 100 and 300 $^{\circ} C$ are still amorphous. The distinct peaks of the Raman curves appear at temperature from 400 $^{\circ} C$. The lineshape tends to be more symmetric and narrowing as the temperature increases. The fitting between the phonon confinement and the experiments reveals a relation of nanocrystals' growth and annealing temperature. It is reasonable that at higher temperature the crystals are larger.
The temperature history of Ag:TiO$_2$ thin films under CW laser irradiation can then be estimated by the method discussed above. To show this, the Raman spectra in the range of 100 - 200 $cm^{-1}$ at different writing speeds are shown in Figure \ref{fgr:chapter5_tempRise_raman_p2}(a). The experimental data are taken from Ref. \cite{liu2016selfthesis} and are then normalized for the comparisons of line characteristics (see Figure \ref{fgr:chapter5_tempRise_raman_p2}(b)). To compare with the TiO$_2$ calcining experiments discussed above \cite{sharma2019crystal}, the data showing strong anatase peaks are selected (speed from Vt+50$\mu m/s$ to Vt+2,100$\mu m/s$). Though the resolution of the curves are limited, it is clear that the Raman spectrum are becoming more symmetric and narrowing while the writing speed is increased. The TiO$_2$ are more crystallized into larger nanocrystals for higher writing speed. The conclusion is that the maximum temperature inside Ag:TiO$_2$ films increases with laser writing speed in the range of Vt+50$\mu m/s$ to Vt+2,100$\mu m/s$).
In short summary of this subsection, the annealing experiments of TiO$_2$ that are reported in many researches \cite{chen2007titanium, swamy2006size, swamy2005finite} showing the TiO$_2$ phase-transitions, which roughly provides the temperature history. Moreover, the phonon confinement model in accordance with the calcined mesoporous experiments performed by N. Sharma et al. \cite{sharma2019crystal} provides fine information of the temperature. Based on such analyses, the post mortem Raman spectra \cite{liu2016selfthesis} of Ag:TiO$_2$ reveal a progressively increasing nature of the temperature with the writing speed of a CW laser. However, this phenomenon looks counter-intuitive since it seems less deposited laser energy leads to higher temperature rise. Is there something missing while mentioning this behavior? In the next subsection, we will go into the depth of the temperature discussion by the assistance of the multi-physical model.
\subsection{Multi-physical simulation of the temperature during laser writing}
In the previous sections, we have discussed the importance of temperature in the growth of Ag NPs inside TiO$_2$, but rarely mentioned the relation of temperature and laser writing speed shown by the model. To clarify the temperature behavior, we are going to discuss the transient maximum-temperature inside the sample during the laser scanning and then concentrates on the steady-state. The reason why we can split the process into a few stages including rapid growth and steady-state actually arises from, for one thing, variations of the spatial size-distribution of Ag NPs, for another, mostly the temperature evolutions.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.85\textwidth]{./figuresChapter5/figure_temp_rise_p3}
\caption{Results of the two-dimensional model. Illustration of the laser scanning direction (a). The top green line represents the locations for retrieving laser profile (b) and temperature distribution (c) from the model. The dynamical temperature history is taken at position $x = 20.1 \mu m$ (c). The maximum temperature in the TiO$_2$ as a function of time during laser writing (e-f). The activation energy is $\mathrm{E_{Ag^0} = 0.95 eV}$, and the ionization rate is $\mathrm{\eta_0 = 2.3 \times 10 ^{-5}}$. Other parameters are shown in Table \ref{table:Chapter5a}.}
\label{fgr:chapter5_tempRise_p1}
\end{figure}
The whole history of the temperature distribution (on the top surface of TiO$_2$) is simulated and plotted in Figure \ref{fgr:chapter5_tempRise_p1}. To better compare the temperature field, the maximum temperature of the structure versus the time is plotted in Figure \ref{fgr:chapter5_tempRise_p1}(e-f). As it is previously shown but rarely discussed in Figure \ref{fgr:3all}(e), the temperature increases from room temperature to a rather high value dramatically before decaying slowly till a steady distribution. This trend is continuously repeated for all the performed simulations at different writing speed. For curves standing for writing speed from 30 $\mu m/s$ to 260 $\mu m/s$, there are time ranges where the maximum temperature inside the structure is constant. Since NPs' growth and reduction process are greatly affected by the temperature, the maximum temperature provides the information of the highest reaction rate inside the TiO$_2$ thin films. Thus, the reactions during the time ranges with constant temperature are steady. Due to this reason, these states can be regraded as steady states.
The temperature at steady state increases with the writing speed (Figure \ref{fgr:chapter5_tempRise_p1}(e) and (f)). This is also shown in Figure \ref{fgr:chapter5_tempRise_p1} (d) by studying the dynamical temperature at position $x = 20.1 \mu m$. The position can be anywhere locating at the steady state shown by Figure \ref{fgr:chapter5_tempRise_p1}(e) and (f). Expect the maximum value has a positive correlation with the writing speed, the temperature evolves differently. The faster the laser writing speed, the faster the temperature changes. This can be useful for calcined or biological applications where the speed of the annealing and the temperature are required to be well controlled. To investigate the relation of maximum temperature with the laser writing speed, simulations at writing speed from 30 $\mu m/s$ to 500 $\mu m/s$ are performed and shown in Figure \ref{fgr:chapter5_tempRise_p2}. The two-dimensional maps in Figure \ref{fgr:chapter5_tempRise_p2}(a) show the time-evolution of the spatial-distribution of temperature on top of the TiO$_2$ thin film. The tilted the angle is, the higher the temperature it has in the TiO$_2$. The maximum temperature is found to increase at first following by a saturation as the writing speed increases (Figure \ref{fgr:chapter5_tempRise_p2}(b)).
\begin{figure}[ht!]
\centering
\includegraphics[width=0.9\textwidth]{./figuresChapter5/figure_temp_rise_p2}
\caption{Simulation and comparisons with experiments. The x-time map of the temperature (a). The maximum temperature at steady state versus laser writing speed (b). The experimental data are taken from Ref. \cite{liu2016selfthesis}. The activation energy is $\mathrm{E_{Ag^0} = 0.95 eV}$, and the ionization rate is $\mathrm{\eta_0 = 2.3 \times 10 ^{-5}}$. Other parameters are shown in Table \ref{table:Chapter5a}.}
\label{fgr:chapter5_tempRise_p2}
\end{figure}
The comparisons of temperature by experiments can be performed by the in-situ Raman spectroscopy. In fact, the Raman scattering or inelastic scattering has two possible outcomes: the photons having lower energy than the incident (known as the Stokes shift); and the photons is of higher energy than the incident (known as the anti-Stokes shift). Though the peak-locations of the Stokes and anti-Stokes in the spectra are symmetric around $\omega = 0$ $cm^{-1}$, the intensities differs because of the different populations of initial states. The initial states of Stokes modes has larger population than that of the ant-Stokes modes for a thermodynamic equilibrium system \cite{maher2008raman,cui1998noncontact,nemanich1984raman,santoro2004situ}. The anti-Stokes modes obeys the Boltzmann distribution $\exp[-\hbar \omega_{\nu} / (k_B T)]$, which as a result gives the ratio of the anti-Stokes and Stokes intensities \cite{maher2008raman,cui1998noncontact,nemanich1984raman,santoro2004situ}:
\begin{equation}
\frac{I_{AS} (\nu)}{I_S (\nu)} = B_0 \frac{(\omega_L + \omega_{\nu})^4}{(\omega_L - \omega_{\nu})^4} \exp[-\hbar \omega_{\nu} / (k_B T)]
\label{eqn:Raman_temp_p2}
\end{equation}
\noindent where $\hbar$ is the reduced Planck's constant, $\omega_{\nu}$ and $\omega_{L}$ are circular frequencies of the vibration modes and the probing laser, $k_B$ is the Boltzmann constant, $T$ is the absolute temperature, and $B_0$ is a constant. The temperature can be retrieved from in-situ Raman experiments based on the Eq. \ref{eqn:Raman_temp_p2}. Here, the experimental data are taken from Ref. \cite{liu2016selfthesis}. The writing speed by the experiments was limited to 500 $\mu m/s$ in order to get a sufficient acquisition time ($\lambda = 533 nm$). The experiments show an increasing tendency in the temperature as the writing speed increases, which confirms the simulations. Noting that only anatase nanocrystals can be observed for the cases at writing speed below 500 $\mu m/s$ \cite{liu2016selfthesis}, the temperature by simulations saturates at speeds of 200-300 $\mu m/s$ is thus inconsistent. Before going further to discuss how the simulations can be fitted with the experiments, the following discussions will be how to understand the temperature behavior.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.85\textwidth]{./figuresChapter5/figure_temp_rise_p1}
\caption{The instant light absorption during laser scanning. The Ag NP size (a), temperature (b) are plotted as the laser-center locates at $x=50 \mu m$. The absorbed laser power per unit volume (c) and the total laser power absorbed (d) by the film are calculated. ($\mathrm{E_{Ag^0} = 0.95 eV}$, $\mathrm{\eta_0 = 2.3 \times 10 ^{-5}}$).}
\label{fgr:chapter5_tempRise_p3}
\end{figure}
In order to understand why the less deposited-energy (higher writing speed) leads to higher temperature rise in the studied system, the instant absorption along x-axis are calculated and shown in Figure \ref{fgr:chapter5_tempRise_p3}. It is noted that the laser-center locations at $x=50 \mu m$ representing the steady state. The starting position of the laser locates at $x= - 50 \mu m$. The second term on the right-hand side of Eq. \ref{eqn:thermal} is calculated for different writing speeds and shown in Figure \ref{fgr:chapter5_tempRise_p3}(c). The corresponding NP size and temperature distributions are shown in Figure \ref{fgr:chapter5_tempRise_p3}(a) and (b). It is obvious that the area below the absorbing curves increases with the writing speed. The total power absorbed by the Ag:TiO$_2$ thin film are calculated by integrating the heat source term of Eq. \ref{eqn:thermal} over the whole thin film (Figure \ref{fgr:chapter5_tempRise_p3}(d)). Clearly, the absorption at-peak and in-average at steady state are both increased at higher writing speed. Consequently, as shown by Figure \ref{fgr:chapter5_tempRise_p3}(b), the maximum temperature has a positive correlation with the writing speed.
Further investigations of the heat equation (Eq. \ref{eqn:thermal}) and the NP size distribution (Figure \ref{fgr:chapter5_tempRise_p3}(a)) reveal that the Ag NP's intrinsic plasmonic absorption is responsible for the boosted power absorption. In the modeling, we used the Mie theory to calculate the NP's absorption by assuming that NPs are randomly distributed and the interactions between NP-NP can be ignored. Figure \ref{fgr:chapter5_mieCal}(a) shows the map of the absorption efficiency factor ($Q_{abs} = \sigma_{abs} / \pi R^2$), where $\sigma_{abs}$ is the cross-section of absorption and $R$ is the radius of Ag NP. In heat equation (Eq. \ref{eqn:thermal}), recall that $\alpha_{abs} = C_{NP} \times \sigma_{abs}$, the NP's abundance or concentration ($C_{NP}$) determines the averaged absorption at any positions. The absorption coefficient $\alpha_{abs}$ is plotted in Figure \ref{fgr:chapter5_mieCal}(b). At wavelength $\lambda = 532 nm$, the absorption efficiency factor ($Q_{abs}$) and the absorption coefficient ($\alpha_{abs}$) peaks at 67.6 nm and 65.7 nm in diameter, respectively (see Figure \ref{fgr:chapter5_mieCal}(d)). At plasmonic resonance, the near-field enhancement of electric field is shown in Figure \ref{fgr:chapter5_mieCal}(b). The maximum NP size by the multiphysical modeling is around 46 nm, which is below the peak-absorption size. For the possible sizes of Ag NPs studied by the model are in the left range where the absorption increases while the NP is larger. We have known that, during the laser irradiation, the process forms large NP ahead of laser beam center following by the shrinkage due to photo-oxidation. It results to comparatively larger size at higher writing speed at any positions behind the locations of the maximum size (Figure \ref{fgr:chapter5_tempRise_p3}(a)). Thus, according to the size-dependent absorption shown in Figure \ref{fgr:chapter5_mieCal}, both the maximum and averaged temperature-rise in the laser spot increases as the writing speed is faster. The limitation of the modeling is obvious that the single-size may give negative response of temperature as the NP size exceeds the peak-absorption size (65.7 nm or 67.7 nm). In fact, the NP size is dispersed according to experiments \cite{liu2017haadf, liu2016selfthesis, liu2015understanding}, and the temperature never decreases as the writing speed increases to the upper limit (around $30 mm/s$) of the setups \cite{liu2016selfthesis, destouches2014self, liu2016laser}.
\begin{figure}[ht!]
\centering
\includegraphics[width=\textwidth]{./figuresChapter5/figure_Mie_cal}
\caption{Maps of absorption efficiency factor (a) and absorption coefficient (b). Electromagnetic simulations (COMSOL) of plane wave ($\lambda = 532 nm$) scattering by a Ag NP (c). The absorption efficiency factor ($Q_{abs}$) and absorption coefficient ($\alpha_{abs}$) as functions of Ag NP size at wavelength 532nm (d).}
\label{fgr:chapter5_mieCal}
\end{figure}
Due to lack of the direct experimental proof by in-situ TEM characterization, the in-direct diagnostics by in-situ transmission and or in-situ temperature can be used for interpolating the model parameters. To do so, one needs to know how do Ag diffusion and ionization rate affect the measurable quantities and what are the trends by these coefficients. For this purpose, here, this discussion will focus on the temperature-rise. We have seen from the previous sections that the Ag diffusion activation energy $E_{Ag^0}$, the diffusion ability $D_{0Ag^0}(500K)$, and the ionization rate $\eta_0$ have great influence on the NPs' final size after laser treatment, and the maximum size in the laser front. The resulted temperature-rise is also affected. Figure \ref{fgr:chapter5_DEAg0_maxT_p1} shows the influences of $E_{Ag^0}$, $D_{0Ag^0}(500K)$, and $\eta_0$ to the maximum temperature-rise at steady states. Hereafter in this subsection, for simplicity, the mentioned maximum temperature-rise stands for the steady states.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.9\textwidth]{./figuresChapter5/figure_DEAg0_maxT_p1}
\caption{The maximum temperature at steady state as a function of the writing speed influenced by activation energy of Ag$^0$ (a), ionization efficiency (b), and the diffusion ability $\mathrm{D_{0Ag^0}(500K)}$ (c).}
\label{fgr:chapter5_DEAg0_maxT_p1}
\end{figure}
\noindent Firstly, an increase in activation energy ($E_{Ag^0}$) slow-down the NP's growth, which can be seen from Figure \ref{fgr:chapter5_DEAg0_maxT_p1}(a) and Figure \ref{fgr:6all}(c) with $E_{Ag^0}$ ranges from 0.75 ev to 1.15 eV. The break firstly appears at $Vs = 600 \mu m/s$ for $E_{Ag^0} = 0.95 eV$, which decreases to $V_S = 300 \mu m/s$ for $E_{Ag^0} = 1.05 eV$. Secondly, the diffusion ability ($D_{0Ag^0}(500K)$) has positive consequences to the NP's growth. The influence, for instance, as shown in Figure \ref{fgr:chapter5_DEAg0_maxT_p1}(c) speed-up the NP's growth. The speed at break position increases from $600 \mu m/s$ to $800 \mu m/s$ before completely disappearing as increasing the diffusion ability ($D_{0Ag^0}(500K)$) from 0.6 to 1.2 $\times 10^{-14} m^2/s$. A further increase of the diffusion ability does not change the curve. Moreover, the Ag diffusion does not affect the maximum temperature-rise (Figure \ref{fgr:chapter5_DEAg0_maxT_p1}(a,c)), on the contrary, the photo-oxidation rate ($\eta_0$) shift the saturating position of the curve (Figure \ref{fgr:chapter5_DEAg0_maxT_p1}(b)). The essential lies in the growth nature of Ag NPs in the studied system which the Ag NPs' growth is ahead of the laser beam due to the heat-diffusion; the growth stops by consuming the majority of the $Ag^0$ in the matrix while the quantity of $Ag^0$ by reduction is too small to feed back into the growth, so that the photo-oxidation dominates the process. As consequences, the oxidation rate affect the spatial-distribution of NP size along the x-axis, which according to the absorption relation shown by Figure \ref{fgr:chapter5_mieCal} affect the maximum temperature-rise. It is obvious that increasing the ionization rate $\eta_0$ shall decrease the NP's size and the temperature. As what is found previously, too high ionization rate shall affect the first rapid-growth (e.g. Figure \ref{fgr:6all}(c)), as a result the temperature at $\eta_0 = 9.2 \times 10^{-5}$ never exceeds $200 ^{\circ}C$.
Indeed, the temperature by experiments increases from speed at a few tens of $\mu m/s$ to 30 mm/s (Figure \ref{fgr:chapter5_tempRise_raman}). The simulated temperature till now seems to saturate and even breaks at speed much lower than the experiments. Now, the question is whether the model can be interpolated close to that experimental findings by adjusting the diffusion parameters and the ionization rate. Before coming back the question, a referring to the parameter discussion would help to understand why one can do this. First of all, the ionization rate is the only one alter the curve shape and the saturation; it, however, suffers from temperature break by simply increasing the ionization rate. Furthermore, a decrease of the activation energy $E_{Ag^0}$ removes the break points. The last but not the least, the increase of diffusion ability ($D_{0Ag^0}(500K)$) accelerates the NP's growth and thus removes the break points in the temperature curve. Figure \ref{fgr:chapter5_tempRise_p4}(a) shows the temperature vs laser writing speed at $E_{Ag^0} = 0.75 eV$, $\eta_0 = 6.9 \times 10 ^{-5}$, and $D_{0Ag^0}(500K) = 4.0 \times 10 ^{-14} m^/2$. The break in the temperature curve disappears at this choice of Ag diffusion and ionization parameters. The absorbance in Figure \ref{fgr:chapter5_tempRise_p4}(a) is calculated as follows:
\begin{equation}
\mathrm{Absorbance} =\int_{x = -100 \mu m}^{x = 100 \mu m} \int_z \alpha_{abs} \cdot I(x,z) dxdz
\label{eqn:thermal_absorbance}
\end{equation}
\noindent where, the term $\alpha_{abs} \cdot I(x,y,z)$ stand is the second term of the heat equation \ref{eqn:thermal}, the distributions on top surface of the structure are shown in Figure \ref{fgr:chapter5_tempRise_p4}(b). The saturation of the temperature-rise vs writing speed is due to the small size change by photo-oxidation in the laser beam (Figure \ref{fgr:chapter5_tempRise_p4}(c)). The temperature profiles at studied writing speed along with the laser intensity profile (black dashed line) are shown in Figure \ref{fgr:chapter5_tempRise_p4}(d). Obviously, both the maximum temperature and the averaged temperature look increase with the writing speed.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.95\textwidth]{./figuresChapter5/figure_temp_rise_p4_v2}
\caption{The temperature as a function of the laser writing speed and the corresponding absorbance. ($\mathrm{E_{Ag^0} = 0.75 eV}$, $\mathrm{\eta_0 = 6.9 \times 10 ^{-5}}$, and $\mathrm{D_{0Ag^0}(500K) = 4.0 \times 10 ^{-14} m^2/s}$). Laser intensity profiles are shown in black dashed lines.}
\label{fgr:chapter5_tempRise_p4}
\end{figure}
\section{3D simulation}
From the previous calculations, we find that solving the coupled equations is non-trivial due to the intrinsic multi-scale problem. Firstly, the growth speed of NPs strongly depends on the temperature, thus making time-steps range from 0.1 ns to 1 $\mu s$ to keep second-order accuracy according to the maximum temperature it experiences. At the same time, the scanning speed of laser is comparatively small that varies from 100 $\mu m/s$ to 50 mm/s. For instance, to study the laser scans 100 $\mu m$ on the sample, in the
case of a typical value to blind computing, the time step is set to 1 ns, it requires more than 2,000,000 steps to finish the calculation. Moreover, the calculation of heat diffusion based on finite element method (FEM) is time consuming.
Comparing to the 2D model, 3D simulations provide a much more complete information. Albeit calculating the temperature and NP size is more difficult in 3D due to the matrix assembling in FEM algorithm, a good design of the data structure in C/C++ enables a scalable and parallel computing on a supercomputer. To
accelerate the calculation, the code is developed based on the distributed-memory parallel programming algorithm. The computational domain consists of 200 nm thick TiO$_2$ film on top of 50 $\mu m$ glass substrate, with width in X and Y of 100 $\mu m$ and 200 $\mu m$, respectively. Figure \ref{fgr:chapter5_3D_p1} shows the meshes used in the FEM calculation. The number of degrees of freedom is 96,865. Thank to the CINES project, we used 32-cores of Intel(R) Xeon(R) CPU-E5-2690 v3 @ 2.6GHz for three weeks to perform the 3D calculations.
Table \ref{table:Chapter5a_2} summarizes the parameters used in 3D simulations.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.75\textwidth]{./figuresChapter5/figure_chapter5_3D_p1.pdf}
\caption{3D view of the meshes used in FEM calculations.}
\label{fgr:chapter5_3D_p1}
\end{figure}
\begin{table*}[ht!]
\centering
\caption{A summary of the parameters in 3D simulation}
\begin{tabular}{llll}
\hline
\hline
Parameter & Description & Value\\
\hline
$P_0$ & laser power& 400mW \\
$\lambda$ & laser wavelength& 532nm \\
$w_0$ & waist radius & $\mathrm{9 \mu m}$\\
$E_{Ag^0} $ & activation energy of Ag$^0$ diffusion & 0.75eV \\
$D_{0Ag^0}(500K) $ & Ag$^0$ diffusion coefficient at 500K &$16 \times 10^{-14} \mathrm{m^2/s}$\\
$\eta _0$ & ionization efficiency & 6.9 $\times 10^{-5}$ \\
$C_{Ag^0}^{BS}(t=0 s)$ & initial Ag$^0$ concentration in matrix & $1.495 \mathrm{\times 10^{4} mol/m^3}$\\
$C_{Ag^+}(t=0 s)$ & initial Ag$^+$ concentration in matrix & $0.996 \mathrm{\times 10^{4} mol/m^3}$ \\
\hline
\hline
\label{table:Chapter5a_2}
\end{tabular}
\end{table*}
Figure \ref{fgr:chapter5_3D_p2} shows top views of laser intensity, temperature, and NP size at different time. In the presented results in Figure \ref{fgr:chapter5_3D_p2}, the laser writing speed is 500 $\mu m/s$; the beam waist size is $2w_0 = 18 \mu m$, at $e^{-2}$ of the peak intensity; the laser power is 400 mW; and the laser wavelength is 532 nm. In the calculation, the laser center goes from (-50,0,0) to (50,0,0) $\mu m$ in +x direction. The results present a general behavior for all scanning speeds. In particular, one can observe the fast growth of NPs at the stating position followed by the size shrinkage and by a transition to the steady state after laser leaves the starting position. At the starting position, due to the positive feedback of plasmonic absorption and NP size, a fast or explosive growth is observed accompanying by the dramatic temperature increase. The following photo-ionization shrinks the NP size in the laser beam, making asymmetry size from laser front to the tail. This process continues until the steady-state is set. The asymmetric NP size profile leads to the light absorption efficiency decrease from the laser front to the tail, which shifts the temperature distribution with its maximum value ahead of laser beam center. In addition, because of heat diffusion, NPs are found to grow even before the center of the laser beam arrives.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.82\textwidth]{./figuresChapter5/figure_chapter5_3D_p2.pdf}
\caption{Top views of temperature, laser intensity and Ag NP size distributions at different time. The laser writing speed is $500 \mu m/s$.}
\label{fgr:chapter5_3D_p2}
\end{figure}
To compare the results of our 3D simulations with the in-situ experiments of transmission during laser writing, maps of laser intensity, NP size, temperature and the corresponding transmission by MG and CM-Mie models are presented in Figure \ref{fgr:chapter5_3D_p3}. The presented results correspond to the final step (the laser center stops at $x = 50 \mu m$) for each writing speed. The Figures show that, in contrast to the 2D model, here we can analyze NPs' size-distribution along the Y direction, which is perpendicular to the writing direction. The NP's size inside the laser spot is smaller than that at the periphery. It can be understood by the growth and shrinkage process. As indicated by Eq. \ref{eqn:fick2} and Eq. \ref{eqn:DAg0defi}, the NPs start the nonlinear growth as soon as the temperature exceed a specific value. The consequence is that, in all the studied writing speeds, the NP size in the laser front forms "\rotatebox{90}{U}" shapes. The photo-oxidation only affects the region inside the laser spot, which results to the small size of Ag NP inside the spot while remaining the outside region to be unaffected. The trend of size increase with writing speed revealed by the 2D model is valid inside the spot region. Since the larger NPs outside the laser spot have little contributions to the absorption and the heat conduction are assumed to be constant in this study, the relation of maximum temperature increase with writing speed is not affected.
\begin{figure}[ht!]
\centering
\includegraphics[width=\textwidth]{./figuresChapter5/figure_chapter5_3D_p3}
\caption{Top views of laser intensity, NP size, Temperature and the corresponding transmission maps for different writing speeds. The results are from the final step in each simulations.}
\label{fgr:chapter5_3D_p3}
\end{figure}
Better comparisons with in-situ transmission experiments are done by directly calculating the transmission maps. The shape of the transmission maps by MG and CM-Mie are close to the in-situ experiments shown in Figure \ref{fgr:1all}(b). For both MG and CM-Mie mode, the transmission inside the laser spot affected line are higher than the vicinity, which is similar to the experiments. Moreover, the higher the writing speed is the lower the transmission in both the center of edge of the written line. Based on the NP size analyses above by the 3D simulations, one can deduce that the size of Ag NP inside the written lines are larger as the writing speed increasing in experiments. In addition, as the temperature is resulted from absorptions by the ensemble of NPs inside the laser spot and diffusion around the vicinity, getting the relation of maximum temperature verses single NP size is impossible. However, the in-situ experiments show better transparency inside the written lines than the model, which may result from particles movement, thickness reduction of the thin film under high temperature, imperfect considerations of size-variations (or size dispersion), a changing in heat conductance, and other changes inside the mesoporous material in reality.
\section{Conclusions}
\hspace{0.5cm} We have proposed an extended and self-consistent model of laser-induced Ag NP growth in TiO$_2$ film. The multi-scale problem has been theoretically investigated by introducing the Bulirsch-Stoer algorithm coupled with Finite Element Method, so that the extended 2D model was solved numerically in an affordable time. The extended 2D model takes into consideration spatial size-distribution of NPs and provides new insights into the understanding of the writing speed-dependent phenomena.
Firstly, the silver NPs have been shown to grow in the laser front edge due to pre-heating by thermal diffusion. Secondly, photo-oxidation has been found to be responsible for the writing speed-dependent size. The spatial size-distribution and temperature profile vary dynamically until a steady-state appears. The implied stationary solutions of the moving coordinate system are found to exist for various speeds by the transient analyses.
Furthermore, the activation energy of Ag$^0$ diffusion has been shown to affect the growing speed, but to only slightly influence the size-saturation. On the contrary, the ionization efficiency is found to have a great impact on the saturation, which again proves the important role of oxidation in the speed controlled phenomena. In the meantime, high oxidation prohibits the rapid growth of NPs that levels up the threshold size of the positive feedback.
The proposed study provides an overview of the influences of Ag$^0$ activation energy and ionization efficiency in mesoporous TiO$_2$ film on the NP size, which gives a new route for comparing the ionization efficiency in different samples.
We note, finally, that the proposed model can be also applied for many similar systems composed of metallic nanoparticles embedded in semiconductors such as, for instance, Ag:ZnO or Au:TiO$_2$.
\afterpage{\null\newpage}
\chapter{Multi-pulsed-laser writing of Ag NPs in TiO$_2$ thin films}
\hspace{0.5cm} This chapter is focused on multi-physical modeling of ultrashort laser-induced Ag nanoparticle formation in TiO$_2$ films. Comparing with continuous-wave laser processing, pulsed-lasers having higher laser intensities and short pulses that are more likely in producing surface grooves (LIPSS), which provides additional possibilities in nano-structuring of porous semiconductors encapsulated with metallic NPs. It was recently shown that under specific fluence and writing speed, the surfaces grooves and Ag nano-gratings closed to the substrate were both generated \cite{liu2017three,liu2016selfthesis,sharmaTailoring2019,sharma2019laser}. The optical anisotropy was demonstrated as results of these subwavelength structures. It was convenient to control laser parameters for switching different structures. However, the multi-physical process consisting of Ag NPs growth, Ag$^+$ reduction, light absorption and heat accumulations were not well understood. To this end, this chapter will discuss the corresponding phenomena and propose simulation models to understand the underlying mechanisms. The performed simulations demonstrated an experimentally similar region of laser fluence and writing speed where Ag NPs grew dramatically. The proposed models paved the way for simulating Ag NPs and Ag nanograting formation by pulsed lasers in future.
\newpage
\section{Introduction}
Controlling nanoscale metallic structures is of great interest by research communities and in industrial applications. The laser based nano-structuring technique is emerging great potential as it is cost-efficient and able for large-scale production. Comparing with continuous-wave laser, the pulsed-laser has higher energy density and shorter time in processing that is more likely to generate nano-grooves on the material surface. For the hosting material, the presence of metallic nanoparticles (NPs) such as Ag, Al and Au NPs has two effects: on one hand, the optical response is tunable by the size-dependent plasmonic resonance; on the other, the plasmon-induced near-field enhancement decreases the laser fluence for damaging the material surface. As consequences, the NP shape is altered anisotropically by high intensity lasers \cite{stalmashonak2009intensity,voss2019situ,catone2018plasmon}. On the other hand, self-organizations of periodic grooves are more promising during laser processing. Based on this idea, several researches were done on the formation of regular gratings \cite{eurenius2008grating,yadavali2014dc}.
Recently, it was demonstrated that two kinds of gratings could be formed by the femtosecond laser irradiation of mesoporous TiO$_2$ films loaded with Ag NPs \cite{liu2017three,sharmalaserdriven2019, sharmaTailoring2019}. The surface grooves of period 490 - 500 nm (perpendicular to the laser polarization) closing to the laser wavelength (515 nm) and silver nanogratings of period 310 nm (parallel to the laser polarization) inside the TiO$_2$ film were formed at certain fluences and writing speeds. Table \ref{table:6b} summarized the experimental results. The fluence and speed dependent phenomena allowed potential applications in optical data storage and multiplexing \cite{liu2017three,sharmalaserdriven2019}.
\begin{table*}[ht!]
\centering
\caption{Femtosecond laser writing of TiO$_2$ doped with Ag NPs. Experimental data are taken from Ref. \cite{liu2016selfthesis,liu2017three}.}
\begin{tabular}{lllll}
\hline
\hline
Name&Fluence ($mJ/cm^{2}$)& Speed ($mm/s$) & Surface grooves & Silver Nanogratings \\
\hline
G1&42 - 62 & 50 -120 & 490 nm, $\perp \mathbf{E}$ & NPs within LIPSS valleys \\
G2&46 - 54 & 5 - 40 & 500 nm, $\perp \mathbf{E}$ & 310 nm, $\parallel\mathbf{E}$ \\
G3&37 - 46 & 5 - 30 & 500 nm, $\perp \mathbf{E}$ & homogeneous NPs (50 nm) \\
G4&29 - 37 & 5 - 30 & None & Anisotropic NPs (<15 nm) \\
\hline
\hline
\hline
\label{table:6b}
\end{tabular}
\end{table*}
However, it lacks well understanding of the multi-pulsed laser processing the material, which is accompanied with light absorption, thermal diffusion, heat accumulation, Ag NP growth, and Ag$^+$ reduction. To this end, this chapter focuses on the modeling of the involved processes. Following the discussion of nonlinear propagation and absorption by the glass substrate, two models of heat accumulation are proposed and coupled with the Ag NP growth and Ag$^+$ reduction. Results are discussed and compared in the following section.
\section{Pulsed-laser propagation}
Substrate absorption and heating strongly depend on laser parameters. Particularly, a set of model modifications are required if a pulsed laser is used instead of a CW laser. In this Chapter, attention is focused on multi-pulsed laser processing of mesoporous TiO$_2$ films with embedded Ag NPs.
We first consider the experimental conditions reported in Ref. \cite{liu2016selfthesis,liu2017three,sharmaTailoring2019}. The fluence of the pulsed laser ranges from 20 to 62 $mJ/cm^{2}$, the spot size (at $1/e^2$ intensity criterion) is around 35 $\mu m$, the repetition rate is 500 kHz, the pulse duration is 300 fs, the wavelength in vacuum is 515 nm, and the writing speed ranges from 5 to 120 mm/s. The corresponding pulse energy varies from 0.1 to 0.3 $\mu J$ by their studies. To check whether the substrate absorption should be considered, 2D+1 (time) simulations based on the nonlinear Schr\"{o}dinger equation (Eq. \ref{eqn:Eq01}) coupled with the plasma equation (Eq. \ref{eqn:Eq03}) are performed. Different pulse energies and focusing conditions are studied.
\begin{figure}[ht!]
\centering
\includegraphics[width=\textwidth]{./figuresChapter6/chap6_nlse_v2_1}
\caption{Simulated propagation of the time-averaged laser intensity (a), time-space distribution of laser intensity at $z=12 \mu m$ and $z = 3.4 mm$, maximum plasma density along z-axis (c), and laser induced temperature-rise (d). The incident pulse energy is 0.3 $\mu J$; and the spot size is $35 \mu m$.}
\label{fgr:chapter6_nlse_fig1}
\end{figure}
Figure \ref{fgr:chapter6_nlse_fig1} shows the evolution of the beam shape at different positions. The laser incidence is in the positive z-direction. For the conditions reported in Ref. \cite{liu2016selfthesis,liu2017three}, the beam shape changes very slowly as shown in Figure \ref{fgr:chapter6_nlse_fig1} (a-b) from z = 36 $\mu m$ to z = 3.4 $mm$. This result is in agreement with the self-focusing and filamentation conditions \cite{couairon2007femtosecond} $P_{cr} = 3.77 \lambda ^2 / 8\pi n_0 n_1$, where $n_0 = 1.52$ is the linear refractive index, and $n_1 = 3.5 \times 10^{-16} cm^2/W$ is the nonlinear refractive index. The calculated laser power approaches the critical power $P_{cr} \approx 7.5 \times 10 ^{5} W$.
The calculated time-averaged intensity is shown in Fig. \ref{fgr:chapter6_nlse_fig1} (a), which proves the self-focusing effect. The corresponding maximum plasma density increases from $0.75 \times 10^{12} cm^{-3}$ at the incident depth to around $2.0 \times 10^{12} cm^{-3}$ at z = 2.1 mm along the propagation direction before decreasing due to the self-focusing (Fig. \ref{fgr:chapter6_nlse_fig1} (c)). Nevertheless, the absorption in the glass is small in this case as shown in Fig. \ref{fgr:chapter6_nlse_fig1} (d).
\begin{figure}[ht!]
\centering
\includegraphics[width=\textwidth]{./figuresChapter6/chap6_nlse_v2_2}
\caption{Simulated propagation of the time-averaged laser intensity (a), time-space distribution of laser intensity at $z = 36 \mu m$ and $z = 672 \mu m$, maximum plasma density along z-axis (c), and laser induced temperature-rise (d). The incident pulse energy is 1.5 $\mu J$; and the spot size is $35 \mu m$.}
\label{fgr:chapter6_nlse_fig2}
\end{figure}
To calculate sample temperature, two-temperature model is commonly used for ultra-short laser interaction with metals and semiconductors. Such approach is required to account for the absence of equilibrium electron-phonon/lattice/matrix during the laser pulse. Here, we are interested by much longer time scales. We consider only matrix temperature and the laser induced temperature rises in the calculations as follows:
\begin{equation}
\begin{aligned}
\delta T &= \frac{1}{C_V} \int \sigma \mathbf{E} \cdot \mathbf{E} \\
& \approx \frac{1}{C_V} \int_t \frac{4\pi \cdot imag(n(x,z,t))}{\lambda} I(x,z,t) dt
\end{aligned}
\label{eqn:chapter6_eq_glass_abs_1}
\end{equation}
\noindent where $C_V$ is the volumetric thermal capacity; $I(z,t)$ is the laser intensity; and $n(x,z,t)$ is the refractive index.
\begin{figure}[ht!]
\centering
\includegraphics[width=\textwidth]{./figuresChapter6/chap6_nlse_v2_3}
\caption{Simulated propagation of the time-averaged laser intensity (a), time-space distribution of laser intensity at $z = 2.4 \mu m$ and $z = 122.4 \mu m$, maximum plasma density along z-axis (c), and laser induced temperature-rise (d). The incident pulse energy is 1.5 $\mu J$; and the spot size is $6 \mu m$.}
\label{fgr:chapter6_nlse_fig3}
\end{figure}
To study the propagation of more powerful and tightly focused laser pulses inside the glass substrate, laser pulse energy of 1.5 $\mu J$, and spot size of 35 $\mu m$ are set in the simulations. The time-averaged intensity presented in Fig.\ref{fgr:chapter6_nlse_fig2} (a) shows a more tightly self-focused beam whose peak position is at around 0.76 mm after the incidence. Figure \ref{fgr:chapter6_nlse_fig2}(b) shows the temporal variations of intensities at z = $12 \mu m$ and $672 \mu m$. In this case, the beam shapes change greatly along z-axis. As a result, the maximum plasma density (Fig. \ref{fgr:chapter6_nlse_fig2} (c)) reaches $6\times 10^{18} cm^{-3}$. The laser induced temperature rise is smaller than 1.6 K as shown in Fig. \ref{fgr:chapter6_nlse_fig2} (d).
Further focusing the beam into the spot size of 6 $\mu m$ (pulse energy of 1.5 $\mu J$) increases the calculated temperature rise. In this case, the plasma absorption and scattering are prominent, so that the beam shapes changes dramatically as shown in Fig. \ref{fgr:chapter6_nlse_fig3} (a-b). The time-averaged intensity clearly shows the scatterings in Fig. \ref{fgr:chapter6_nlse_fig3} (a). The maximum plasma promptly decreases from $6.2 \times 10 ^{20} cm^{-3}$ to $0.6 \times 10 ^{20} cm^{-3}$ in the first 20 $\mu m$ range (Fig. \ref{fgr:chapter6_nlse_fig3} (c)). The maximum temperature-rise reaches 250 K on the surface and decreases along the z-axis with a line shape whose radius is smaller than the beam radius (Fig. \ref{fgr:chapter6_nlse_fig3} (d)).
Because relatively low fluences are used in the discussed experiments, light absorption takes place inside the TiO$_2$ films, so that the absorption by the glass substrate can be ignored. Laser propagation and nonlinear effects are neglected in this study due to the lack of information about nonlinear effects in TiO$_2$ and since rather low laser intensity is favorable for the considered applications. In addition, the near-field enhancement by the localized plasmonic resonances is also disregarded here. In the previous chapter, we have treated the film as an effective medium when considering the absorption; likewise, here, we introduce the same idea and try to simulate the growth of Ag NPs pulse-by-pulse.
\section{Heat accumulation}
The direct attempt to find the accumulated laser energy is to solve the general heat equation using the Green's function method as following:
\begin{equation}
\left\{
\begin{aligned}
C_V \frac{\partial T}{\partial t} - \nabla \cdot k_T \nabla T &= 0\\
T(x,y,z, t=0) &= \delta (x, y, z)
\label{eqn:heatEQChap6_eq1}
\end{aligned}
\right.
\end{equation}
\noindent where $C_V$ is the specific heat capacity, $k_T$ is the thermal conductivity, and $\delta(x,y,z)$ is the Dirac delta function. The Green's function is found in the form \cite{beck1992heat}:
\begin{equation}
\begin{aligned}
G_T(x,y,z,t) = \big(\frac{1}{\sqrt{4 \pi \beta_T t}}\big)^3 \exp{\big(-\frac{x^2 + y^2 + z^2}{4 \beta_T t}\big)}
\label{eqn:heatEQChap6_eq1_2}
\end{aligned}
\end{equation}
\noindent Thus, the general solution of the heat transfer equation is given by:
\begin{equation}
\begin{aligned}
T(x,y,z,t) = \int_{t_s} \int_{x_s,y_s,z_s} G_T(x-x_s,y-y_s,z-z_s,t-t_s) f(x_s,y_s,z_s,t_s) dt_s dx_s dy_s dz_s
\label{eqn:heatEQChap6_eq1_3}
\end{aligned}
\end{equation}
\noindent where $T(x,y,z,t)$ is the temperature difference induced by the heat source, $\beta_T = k_T / C_V$ is the thermal diffusivity and $f(x_s,y_s,z_s,t_s)$ is the heat source. For pulsed-laser studied by Z. Liu et al. \cite{liu2017three, liu2016selfthesis}, the absorption inside the glass substrate can be ignored as discussed in the previous section. Thus, only thin film absorption is considered in this case. We remind that the laser pulse width is around 300 fs, while the growth of Ag NPs and thermal diffusion take place on nanosecond and longer time scales. For integration, the Green's function is close to the Dirac delta function. Therefore, the time integration term is written as:
\begin{equation}
\begin{aligned}
\int_{t_s} G_T(x,y,z,t-t_s) f (x,y,z,t_s) dt_s = G_T(x,y,z,t-t_s^i) T_0(x,y,z,t=t_s^i)
\label{eqn:heatEQChap6_eq1_4}
\end{aligned}
\end{equation}
\noindent where $t_s^i$ is time corresponding to the peak intensity of the pulse.
In addition, the initial temperature after absorbing a single pulse energy is calculated by neglecting heat diffusion during such short time and is written as:
\begin{equation}
\begin{aligned}
T_0(x,y,z,t=t_s) &= \frac{1}{C_V} \int_{\tau_p}\alpha_{abs}(x,y,z,t) I(x,y,z,t) dt\\
&= \frac{1}{C_V} \int_{\tau_p} C_{NP} Q_{abs}(\omega) \pi R_{NP}^2 I(x,y,z,t) dt
\label{eqn:heatEQChap6_eq1_4_2}
\end{aligned}
\end{equation}
\noindent where $I(x,y,z,t)$ is the laser intensity, $\tau_p$ is the pulse duration, $\alpha_{abs} = C_{NP} \pi R_{NP}^2 Q_{abs}$ is the absorption coefficient inside the film, $C_{NP}$ is the concentration of the Ag NP, $Q_{abs}(\omega)$ is the absorption efficiency factor, $\omega$ is the circular frequency of light, and $R_{NP}$ is the radius of an Ag NP. Therefore, the Eq. \ref{eqn:heatEQChap6_eq1_3} is written as:
\begin{equation}
\begin{aligned}
T(x,y,z,t) &= \int_{x_s,y_s,z_s} G_T(x-x_s,y-y_s,z-z_s,t-t_s^i) \\
& \times \frac{1}{C_V}\int_{\tau_p} \alpha_{abs} I(x_s,y_s,z_s,t) dt dx_s dy_s dz_s, \\
&(x,\: y,\: z,\: t \in \textbf{R}, \text{and} \: t > t_s^i)
\label{eqn:heatEQChap6_eq1_5}
\end{aligned}
\end{equation}
\noindent where $t_s^i$ is the time of the incident pulse. For simplicity, laser intensity inside the TiO$_2$ thin film $I(x,y,z,t)$ can be written as:
\begin{equation}
\begin{aligned}
I(x,y,z,t) = \frac{E_p}{\tau_p \pi w_0^2 /2} \exp{\big(-\frac{2(x^2+y^2)}{w_0^2}\big)} \exp{\big(-\frac{8(t-t_s^i)^2}{\tau_p ^2}\big)} \exp{\big(-\alpha_{abs} z \big)}
\label{eqn:heatEQChap6_eq1_6}
\end{aligned}
\end{equation}
\noindent where $E_p$ is the pulse energy, and $w_0$ is the beam waist (at $e^{-2}$ of the peak intensity). Recalling that $\int_{-\infty}^{\infty} \exp{-\frac{(x-x_0)^2}{2 a^2}}dx = a \sqrt{2 \pi}$, the one variable integration is as follows:
\begin{equation}
\begin{aligned}
&\int_{x_s} G_T(x-x_s, t-t_s^i) \times T_0(x_s, t_s) dx_s\\
&= \frac{1}{C_V} \int_{x_s} \frac{1}{\sqrt{4\pi \beta_T(t-t_s^i)}} \exp{\big(-\frac{(x-x_s)^2}{4 \beta_T (t-t_s^i)} \big)} \exp{\big(-\frac{2x_s^2}{w_0^2} \big)} dx_s\\
&\times \alpha_{abs} I_0 \frac{\tau_p \sqrt{2 \pi}}{4} \exp{\big(-\alpha_{abs} z \big)} \\
&=\frac{1}{C_V} \sqrt{\frac{w_0^2}{w_0^2+8\beta_T(t-t_s^i)}} \exp{\big(- \frac{2x^2}{w_0^2+8\beta_T(t-t_s^i)}\big)} \\
&\times \alpha_{abs} I_0 \frac{\tau_p \sqrt{2 \pi}}{4} \exp{\big(-\alpha_{abs} z \big)}
\label{eqn:heatEQChap6_eq1_7}
\end{aligned}
\end{equation}
\noindent where $ I_0 = \frac{E_p}{\tau_p \pi w_0^2 /2}$ is the peak intensity. It should be noted that the Eq. \ref{eqn:heatEQChap6_eq1_7} is only valid as the absorption coefficient $\alpha_{abs}$ is independent of $x$.
\subsection{Homogeneous absorption in x and y}
If the absorption coefficient $\alpha_{abs}(x,y,z,t)$ is homogeneous in x and y, then $\alpha_{abs}(x,y,z,t) = \alpha_{abs}(z,t)$. The temperature difference induced is expressed as:
\begin{equation}
\begin{aligned}
T^i(x,y,z,t;t_s^i) &=\frac{w_0^2}{C_V[w_0^2+8\beta_T(t-t_s^i)]} \exp{\big(- \frac{2(x^2+y^2)}{w_0^2+8\beta_T(t-t_s^i)}\big)} I_0 \frac{\tau_p \sqrt{2 \pi}}{4} \\
& \times \int_{0}^{h_0} \frac{1}{\sqrt{4\pi \beta_T(t-t_s^i)}} \exp{\big(-\frac{(z-z_s)^2}{4 \beta_T (t-t_s^i)} \big)} \alpha_{abs} \exp{\big(-\alpha_{abs} z_s \big)} dz_s
\label{eqn:heatEQChap6_eq1_8}
\end{aligned}
\end{equation}
\noindent where $h_0$ is the thickness of the TiO$_2$ film, the rest of the integration along z-axis is zero because of the non-absorption in the glass substrate at the studied laser energies and focusing conditions. It should be noted that the variables $x$, $y$, and $z$ are relative to the pulse; in a global coordinate system, the transformation is $x_g = x + Vs \cdot t_s^i$, $y_g = y$, $z_g = z$. This is because the Green's function is written in such a coordinate system (see Eq. \ref{eqn:heatEQChap6_eq1} and Eq. \ref{eqn:heatEQChap6_eq1_2}). Note that this equation is valid when the substrate does not absorb laser energy.
In the case of multi-pulse process, the temperature difference $T(x,y,z,t)$ is written as:
\begin{equation}
\begin{aligned}
T(x,y,z,t) = \sum_i^N T^i(x - Vs \cdot t_s^i,y,z,t;t_s^i), \text{for} \: t \geq t_s^N
\label{eqn:heatEQChap6_eq1_9}
\end{aligned}
\end{equation}
\noindent where $N$ is the total number of pulses irradiated on the sample, and the coordinate is in the global coordinate system.
For a scanning laser beam, the laser-affected region is in the vicinity of the laser spot. We assume that laser scans in the positive x-axis direction at the speed of $V_s$. Then, we transform the Eq. \ref{eqn:heatEQChap6_eq1_8} into a moving coordinate system ($x'=x_g + V_s t, y'=y_g, z'=z_g, t'=t$). The temperature difference induced by a single pulse can be written as:
\begin{equation}
\begin{aligned}
T_m^i(x,y,z,t;t_s^i) &=\frac{w_0^2}{C_V[w_0^2+8\beta_T(t-t_s^i)]} \exp{\big(- \frac{2[(x+V_s (t-t_s^i))^2+y^2]}{w_0^2+8\beta_T(t-t_s^i)}\big)} I_0 \frac{\tau_p \sqrt{2 \pi}}{4} \\
& \times \int_{0}^{h_0} \frac{1}{\sqrt{4\pi \beta_T(t-t_s^i)}} \exp{\big(-\frac{(z-z_s)^2}{4 \beta_T (t-t_s^i)} \big)} \alpha_{abs} \exp{\big(-\alpha_{abs} z_s \big)} dz_s
\label{eqn:heatEQChap6_eq1_10}
\end{aligned}
\end{equation}
\noindent whereas for the multi-pulse regime, one obtains the following expression
\begin{equation}
\begin{aligned}
T_m(x,y,z,t) = \sum_i^N T_m^i(x,y,z,t;t_s^i), \text{for} \: t \geq t_s^N
\label{eqn:heatEQChap6_eq1_11}
\end{aligned}
\end{equation}
In general, Eq. \ref{eqn:heatEQChap6_eq1_8}, Eq. \ref{eqn:heatEQChap6_eq1_9}, Eq. \ref{eqn:heatEQChap6_eq1_10}, and Eq. \ref{eqn:heatEQChap6_eq1_11} have no analytical solutions because of the asymmetric function of $\alpha_{abs}(x,y,z,t)$, which dynamically changes due to the growth of Ag NPs while heating by laser pulses. It is noted that the deduced Eq. \ref{eqn:heatEQChap6_eq1_8}, Eq. \ref{eqn:heatEQChap6_eq1_9}, Eq. \ref{eqn:heatEQChap6_eq1_10} and Eq. \ref{eqn:heatEQChap6_eq1_11} are only valid for lasers whose pulse duration is much shorter than both the growth of Ag NPs and the heat diffusion time (normally in nanosecond scale). Otherwise, the dynamic changes of $\alpha_{abs}$ should be considered due to the size variation of Ag NPs inside the time integration term (Eq. \ref{eqn:heatEQChap6_eq1_4}) . Moreover, if the NP size changes, the obtained formulas from Eq. \ref{eqn:heatEQChap6_eq1_7} to Eq. \ref{eqn:heatEQChap6_eq1_11} are invalid. In this case, the convolution in Eq. \ref{eqn:heatEQChap6_eq1_7} has no semi-analytic solutions. Nevertheless, these equations can be used for the estimation of base temperature before growth of Ag NPs. This is correct, since the threshold temperature required for the growth of NPs is high comparing to the room temperature. As a result, multiple pulses to accumulate the base temperature so as to exceed the threshold for a dramatic growth of Ag NPs.
\subsection{Homogeneous absorption in y and z}
As in the Chapter focused on CW laser writing of Ag NPs inside the TiO$_2$ film, here the NP size changes dramatically along the laser scanning direction $x$. Furthermore, if the temperature variation along z-axis (direction of laser incidence) inside the film can be ignored, the single-pulse induced temperature difference $T^i(x,y,z,t)$ only has one integration of $x$. In this way, the resulted semi-anatylic equation can be largely simplified accelerating the temperature calculation. In this case, $\alpha_{abs}(x,y,z,t) = \alpha_{abs}(x,t)$ and the temperature difference is written as follows:
\begin{equation}
\begin{aligned}
&T^i(x,y,z,t;t_s^i) =\\
&\frac{1}{C_V}\sqrt{\frac{w_0^2}{[w_0^2+8\beta_T(t-t_s^i)]}} \exp{\big(- \frac{2(y^2)}{w_0^2+8\beta_T(t-t_s^i)}\big)} I_0 \frac{\tau_p \sqrt{2 \pi}}{4} \\
& \times \int_{x_s} \alpha_{abs}(x_s,t_s^i) \frac{1}{\sqrt{4\pi \beta_T(t-t_s^i)}} \exp{\big(-\frac{(x-x_s)^2}{4 \beta_T (t-t_s^i)} \big)} \exp{\big(-\frac{2x_s^2}{w_0^2} \big)} \\
& \times \int_{0}^{h_0} \frac{1}{\sqrt{4\pi \beta_T(t-t_s^i)}} \exp{\big(-\frac{(z-z_s)^2}{4 \beta_T (t-t_s^i)} \big)} \exp{\big(-\alpha_{abs}(x_s,t_s^i) \cdot z_s \big)} dz_s dx_s
\label{eqn:heatEQChap6_eq1_13}
\end{aligned}
\end{equation}
\noindent Since $\int_{x_0}^{x_1} \exp{-\frac{(x+c)^2}{2a^2}}dx = \sqrt{\frac{\pi}{2}} a \big[ \erf{\frac{x_1+c}{a\sqrt{2}} - \erf{\frac{x_0+c}{a\sqrt{2}}}} \big]$, the Eq. \ref{eqn:heatEQChap6_eq1_13} is expressed as follows:
\begin{equation}
\begin{aligned}
&T^i(x,y,z,t;t_s^i) =\\
&\frac{1}{C_V}\sqrt{\frac{w_0^2}{[w_0^2+8\beta_T(t-t_s^i)]}} \exp{\big(- \frac{2(y^2)}{w_0^2+8\beta_T(t-t_s^i)}\big)} I_0 \frac{\tau_p \sqrt{2 \pi}}{4} \\
& \times \int_{x_s} \alpha_{abs}(x_s,t_s^i) \frac{1}{\sqrt{4\pi \beta_T(t-t_s^i)}} \exp{\big(-\frac{(x-x_s)^2}{4 \beta_T (t-t_s^i)} \big)} \exp{\big(-\frac{2x_s^2}{w_0^2} \big)} \\
& \times \frac{1}{2} \bigg( \erf{\frac{h0+2\beta_T(t-t_s^i)\alpha_{abs}(x_s,t_s^i) - z}{\sqrt{4\beta_T(t-t_s^i)}}} - \erf{\frac{h_1|_{h_1 \to 0^-} + 2\beta_T(t-t_s^i)\alpha_{abs}(x_s,t_s^i) - z}{\sqrt{4\beta_T(t-t_s^i)}}} \bigg) \\
& \times \exp{\bigg(-\alpha_{abs}(x_s,t_s^i) \big(z - \beta_T (t-t_s^i) \alpha_{abs}(x_s,t_s^i) \big) \bigg) } dx_s
\label{eqn:heatEQChap6_eq1_14}
\end{aligned}
\end{equation}
\noindent In most cases, we consider only the top surface $z = 0$ and the line along the laser scanning direction $y = 0$. Thus, the Eq. \ref{eqn:heatEQChap6_eq1_13} can be simplified as follows:
\begin{equation}
\begin{aligned}
&T^i(x,t;t_s^i)|_{y=0,z=0} =\\ &\frac{1}{C_V}\sqrt{\frac{w_0^2}{[w_0^2+8\beta_T(t-t_s^i)]}} I_0 \frac{\tau_p \sqrt{2 \pi}}{4} \\
& \times \int_{x_s} \alpha_{abs}(x_s,t_s^i) \frac{1}{\sqrt{4\pi \beta_T(t-t_s^i)}} \exp{\big(-\frac{(x-x_s)^2}{4 \beta_T (t-t_s^i)} \big)} \exp{\big(-\frac{2x_s^2}{w_0^2} \big)} \\
& \times \frac{1}{2} \bigg( \erf{\frac{h0+2\beta_T(t-t_s^i)\alpha_{abs}(x_s,t_s^i)}{\sqrt{4\beta_T(t-t_s^i)}}} - \erf{\frac{h_1|_{h_1 \to 0^-} + 2\beta_T(t-t_s^i)\alpha_{abs}(x_s,t_s^i)}{\sqrt{4\beta_T(t-t_s^i)}}} \bigg) \\
& \times \exp{\bigg( \beta_T (t-t_s^i) \big(\alpha_{abs}(x_s,t_s^i) \big)^2 \bigg) } dx_s
\label{eqn:heatEQChap6_eq1_15}
\end{aligned}
\end{equation}
\noindent The numerical introduction of the above equation is quite tricky and normally requires that the condition $2\beta_T(t-t_s^i)\alpha_{abs}(x_s,t_s^i) < 100 h_0$ is fulfilled. Otherwise, the last exponential term $\exp{\bigg( \beta_T (t-t_s^i) \big(\alpha_{abs}(x_s,t_s^i) \big)^2 \bigg)}$ is divergent, so that the numerical errors can grow, so that the calculations can fail. This condition can be very useful in accelerating simulations. In fact, because of heat diffusion, for all pulses whose irradiation time $t_s^i$ smaller than $100h_0/(2\beta_T \alpha_{abs}(x_s,t_s^i))$ are less than 10\%. In the case of the studied Ag NPs inside TiO$_2$ films, the considered time is around 10 $\mu s$. A numerical solution of this issue should be included and will be discussed in the next section. However, this approximations can lead to an underestimated temperature. Furthermore, the Green's function should be carefully considered. For this, numerical introduction of the Dirac delta function can be of the form $\delta_{\epsilon}(x) = \frac{1}{\epsilon} \varphi(x/\epsilon)$ \cite{zahedi2010delta,engquist2005discretization,smereka2006numerical,beale2008proof,towers2007two,towers2008convergence}, where one can choose $\varphi(x) = \frac{1}{2}(1+cos(\pi x))$ for $x \leqslant 1$ and $\varphi(x) = 0 $ for $x > 1$.
In the multi-pulses regime, the accumulated energy in the form of temperature difference is written as follows:
\begin{equation}
\begin{aligned}
T(x,t)|_{y=0,z=h_0} = \sum_i^N T^i(x - Vs \cdot t_s^i,t;t_s^i)|_{y=0,z=h_0}, \text{for} \: t \geq t_s^N
\label{eqn:heatEQChap6_eq1_16}
\end{aligned}
\end{equation}
\noindent where the coordinates $x$, $y$, $z$ and $t$ are in the global coordinate system.
The deduced Eq. \ref{eqn:heatEQChap6_eq1_15} and Eq. \ref{eqn:heatEQChap6_eq1_16} take into considerations spatial (along x-axis) and kinetic variations of absorption, which are the general descriptions of the temperature field. Because there is only one integration left in the calculation, the simulations can be drastically accelerated by coupling Eq. \ref{eqn:heatEQChap6_eq1_15} and Eq. \ref{eqn:heatEQChap6_eq1_16} with the NPs growth. In contrast, the three-dimensional calculations by finite element method are inefficient and time-consuming, requiring huge computational resources. However, the finite element method (FEM) is more accurate, since it considers diffusion inside thin films. In contrast, the Eq. \ref{eqn:heatEQChap6_eq1_15} and Eq. \ref{eqn:heatEQChap6_eq1_16} assume that the film is very thin and that the heated layer can be regarded as a boundary condition.
\section{The coupled kinetic equations}
The modeling of Ag NPs formation is discussed in the previous Chapter. The coupled processes of growth, photo-oxidation and reduction are described at any position as follows \cite{liu2015understanding,liu2016selfthesis}:
\begin{equation}
\left\{
\begin{aligned}
&\frac{d R_{NP}}{d t} = [n_{abs}(t) - n_{oxi}(t)] \cdot \frac{\omega_{Ag}}{4 \pi R_{NP}^2}\\
&\frac{d C_{Ag^+}}{d t} =n_{oxi}(t) [ C_{NP}(t) + \frac{R_{NP}(t)}{3} \frac{d C_{NP}} {d R_{NP}} ] - \frac{d C_{Ag^+}}{d t} |_{red}\\
&\frac{d C_{Ag^0}^{BS}}{d t} = - n_{abs}(t) [ C_{NP}(t) + \frac{R_{NP}(t)}{3} \frac{d C_{NP}} {d R_{NP}} ] + \frac{d C_{Ag^+}}{d t} |_{red}\\
\end{aligned}
\right.
\label{eqn:heatChap6_sec_couple_eq1}
\end{equation}
\noindent where $n_{oxi}$ is the number of silver atoms leaving a nanoparticle per unit time due to photo-oxidation, $\frac{d C_{Ag^+}}{d t} |_{red}$ is the decrease of Ag$^+$ concentration due to the reduction per unit time, and $n_{abs}$ is the total amount of monomers absorbed by the Ag NP per unit time, which are described in chapter 5 and in Ref. \cite{liu2015understanding, liu2016selfthesis, liu2016laser}. These variables can be written as:
\begin{equation}
\left\{
\begin{aligned}
&n_{abs}(t) = 4 \pi R_{NP} D_{0Ag^0}\exp \Big( -\frac{E_{Ag^0}}{N_A k_B T} \Big) [C_{Ag^0}^{BS}- S_{Ag^0} ( 1 + \frac{2 \gamma \omega_{Ag}}{R_{NP} k_B T} )]\\
&\frac{d C_{Ag^+}}{d t} |_{red} = \exp \Big( -\frac{E_p}{N_A k_B T} \Big) \cdot D^{red}_{0} \exp \Big( -\frac{E_D}{N_A k_B T} \Big) \cdot C_{red} C^{2/3}_{Ag^+}\\
&n_{oxi}(t) =\eta_0 \frac{ \sigma_{abs}(R,\lambda) I(x, y, z, t)} {h \nu}
\end{aligned}
\right.
\label{eqn:heatChap6_sec_couple_eq2}
\end{equation}
The above coupled equations are in the simplified forms since the processes of Ag$^+$, Ag$^0$, and Ag NPs diffusion are neglected. In addition, the diversity in NP's size is not considered. The photo-oxidation process is regarded to be temperature independent \cite{liu2015understanding,liu2016selfthesis}. The parameters of the coupled equations are summarised in Table \ref{table:6a}.
\newpage
\begin{table*}[ht!]
\centering
\caption{The summary of parameters}
\begin{tabular}{ll}
\hline
\hline
Parameter & Description \\
\hline
& \\
Ag NP growth, oxidation, and Ag$^+$ reduction&\\
\hline
$\eta _0$ & ionization efficiency\\
$\omega_{Ag}$ & atomic volume of crystallized Ag\\
$E_{Ag^0} $ & activation energy of Ag$^0$ diffusion\\
$D_{0Ag^0}(500K)=D_{0Ag^0}\exp \Big( -\frac{E_{Ag^0}}{N_A k_B 500K} \Big)$ & Ag$^0$ diffusion coefficient at 500K\\
$E_p$ & activation energy of reduction \\
$C_{red} $ & concentration of reducing agent (RA) \\
$E_D $ & activation energy of RA diffusion\\
$D_0^{red}(500K) = D^{red}_{0} \exp \Big( -\frac{E_D}{N_A k_B 500K} \Big)$ & RA diffusion coefficient at 500K \\
$S_{Ag^0} $ & solubility of Ag$^0$ \\
$\gamma$ &Ag NP - TiO$_2$ interfacial tension \\
$k_B$ & the Boltzmann constant\\
$N_A$ & the Avogadro constant\\
$T$ & temperature \\
$R_{NP}$ & radius of Ag NP \\
$C_{Ag^0}^{BS}(t=0 s)$ & initial Ag$^0$ concentration in matrix \\
$C_{Ag^+}(t=0 s)$ & initial Ag$^+$ concentration in matrix \\
& \\
Laser pulse and thermal diffusion&\\
\hline
$E_p$ & pulse energy \\
$\tau_p$ & pulse duration \\
$k_r$ & repetition rate\\
$w_0$ & waist radius (at $e^{-2}$)\\
$C_V$ & volumetric heat capacity \\
$k_T$ & thermal conductivity\\
$\beta_T$ & thermal diffusivity\\
$t_s^i$ & time at peak intensity of the $i^{\text{th}}$ pulse\\
$\alpha_{abs}(x_s,y_s,z_s,t_s^i)$ & absorption coefficient\\
$x_s,y_s,z_s$ & the coordinates relative to the $i^{\text{th}}$ pulse\\
$h_0 $ & thickness of TiO$_2$ film\\
&\\
\hline
\hline
\label{table:6a}
\end{tabular}
\end{table*}
\newpage
\subsection{Coupling algorithm for the semi-analytic function}
Because of the simplified form of Eqs. (\ref{eqn:heatEQChap6_eq1_15} - \ref{eqn:heatEQChap6_eq1_16}), the calculation of temperature field can be paralleled and become scalable for GPU. Nowadays, GPU has more than a few hundreds of cores, and is cheaper than CPU on a personal computer. To take advantages of such parallel schemes for accelerating the multi-scale problem, the developed model is shown in Fig. \ref{fgr:chapter6_coupledAna_1}. The simulations of Ag NPs growth, photo-oxidation, reduction, and heating are performed in 1D, which stands for the top surface of the TiO$_2$ thin film. The glass thermal diffusivity $\beta_T \approx 3.4 \times 10 ^{-7} m^{2} s^{-1}$ is considered in Eqs. (\ref{eqn:heatEQChap6_eq1_15} - \ref{eqn:heatEQChap6_eq1_16}). The Bulirsch-Stoer algorithm \cite{press2007section} is used for simulating the variations of Ag$^0$ atoms, Ag$^+$ ions, and Ag NPs inside the TiO$_2$ film. The discretization is of the second accuracy and can largely accelerate the simulation by estimating the time steps.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.72\textwidth]{./figuresChapter6/chap6_coupledAna_1}
\caption{Schematics of the developed semi-analytic model. The temperature are calculated on the GPU and by pulse-by-pulse accumulations, which is then coupled with the Ag NP's growth, photo-oxidation, and reduction.}
\label{fgr:chapter6_coupledAna_1}
\end{figure}
\subsection{Coupling algorithm for the finite element method}
The the implementation of coupling is direct in a finite element method (FEM). For an ultrafast pulse whose duration is much shorter the heat diffusion, the laser induced temperature difference, as given by Eq. \ref{eqn:heatEQChap6_eq1_4_2}. The heating by each pulse and the cooling between two pulses can be described as follows:
\begin{equation}
\left\{
\begin{aligned}
& C_V \frac{\partial T}{\partial t} - \nabla \cdot k_T \nabla T = 0\\
& \delta T^i(x,y,z;t_s^i) = \frac{1}{C_V} \int_{\tau_p} \alpha_{abs}(x,y,z,t_s^i) I(x,y,z,t) dt, \: i \in [1, N]
\label{eqn:heatEQChap6_femCoupled_eq1}
\end{aligned}
\right.
\end{equation}
\noindent where, $N$ is the number of incident pulses. The Crank-Nicolson discreatization \cite{kays1967convective} is used for the heat transfer equations. The developed FEM model is depicted in Fig. \ref{fgr:chap6_coupledFEM_1} (a-c). In contrast to the semi-analytic model, the thermal properties of 200 nm TiO$_2$ film are considered in the FEM.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.8\textwidth]{./figuresChapter6/chap6_coupledFEM_1}
\caption{Schematics of the developed FEM-based model. }
\label{fgr:chap6_coupledFEM_1}
\end{figure}
\noindent For simplicity, the specific heat capacity of anatase TiO$_2$ is taken from Ref. \cite{saeedian2012specific}, thermal conductivity is from Ref. \cite{batmunkh2014thermal}, the effective volumetric heat capacity and effective thermal conductivity of the mesoporous TiO$_2$ films are calculated by $C_{V_{eff}} = C_{V_{TiO_2}} (1-v_p) + C_{V_{air}} v_p $, and $k_{T_{eff}} = k_{T_{TiO_2}} (1-v_p) + k_{T_{air}} v_p$ \cite{smith-01115971}, where $v_p$ is the volume fraction of pores. The porosity is considered from 45\% to 60\% as reported in Ref. \cite{sharma2019crystal}. The calculated effective specific heat capacity is $C_{V_{eff}} \approx 1.008 \times 10 ^6 J/m^3/K$, and effective thermal conductivity is $k_{T_{eff}} \approx 0.2 W/m/K$. For the glass, we consider a bulk material with specific heat of $C_V \approx 2.352 \times 10 ^6 J/m^3/K$, and thermal conductivity from $k_T \approx 0.5 W/m/K$ to $k_T \approx 1.8 W/m/K$.
\section{Results and Discussions}
This section deals with the results of the proposed models. Then, comparisons with the experiments are also discussed. In the first part, we present the base temperature by assuming that the Ag NPs do not grow during few first pulses. The results are compared with more rigorous calculations performed by using the finite element method model. In the second subsection, we report the semi-analytic results after laser scans 100 $\mu m$ length in the x-direction. Furthermore, the results of the FEM model are shown in the following subsection. The differences between the two models are compared and explained. Finally, the comparisons of the results obtained in simulations and experiments are presented.
\subsection{Base temperature before NP's growth}
In the previous Chapter, we have discussed the continuous-wave laser interactions of Ag NPs inside TiO$_2$ thin films. It was found that the Ag NPs grow very slowly in the beginning before the so-called "dramatic growth" takes place. This effect arises as the NP's size exceeds a certain value. In the region before the dramatic growth, one can estimate the temperature rise by using a homogeneous size assumption. In the case of pulsed laser writing of Ag NPs inside TiO$_2$ films, this can be done by looking at the base temperature and checking whether the Ag NPs can grow dramatically or not under the particular laser-irradiation conditions. In this case, the semi-analytic formula (Eq. \ref{eqn:heatEQChap6_eq1_10}) has an analytic solution as follows:
\begin{equation}
\begin{aligned}
&T_m^i(x,y,z,t;t_s^i) =\\
&\frac{w_0^2}{C_V[w_0^2+8\beta_T(t-t_s^i)]} \exp{\big(- \frac{2[(x+V_s (t-t_s^i))^2+y^2]}{w_0^2+8\beta_T(t-t_s^i)}\big)} \alpha_{abs} I_0 \frac{\tau_p \sqrt{2 \pi}}{4} \\
& \times \frac{1}{2} \bigg( \erf{\frac{h0+2\beta_T(t-t_s^i)\alpha_{abs}}{\sqrt{4\beta_T(t-t_s^i)}}} - \erf{\frac{h_1|_{h_1 \to 0^-} + 2\beta_T(t-t_s^i)\alpha_{abs}}{\sqrt{4\beta_T(t-t_s^i)}}} \bigg) \\
& \times \exp{\bigg( \beta_T (t-t_s^i) \big(\alpha_{abs} \big)^2 \bigg) }
\label{eqn:heatEQChap6_eq2_21}
\end{aligned}
\end{equation}
\begin{figure}[ht!]
\centering
\includegraphics[width=0.9\textwidth]{./figuresChapter6/chap6_alpha_homo_fig1}
\caption {Temperature-rise pulse-by-pulse (a). The dashed red line stands for the base-temperature. The writing speed has little influence on the temperature profiles (for the first 30 pulses) in the range of 5 mm/s to 50 mm/s. Base temperature rise vs number of pulses for $V_s = 5$ mm/s (b) and $V_s = 50$ mm/s. The position is taken as x=0, y=0, and z=0.}
\label{fgr:chap6_alpha_homo_fig1}
\end{figure}
\noindent where $\alpha_{abs} = 5.0 \times 10^4 m^{-1}$ as we take the radius $R_{NP} = 1.5 nm$ for the initial Ag NPs \cite{liu2016selfthesis,liu2015understanding, liu2017haadf}. The cross-section of Ag absorption is calculated by the Mie theory at wavelength 515 nm, which corresponds to the experimental results obtained by using a femtosecond laser system with fundamental wavelength of 1030 nm, whose frequency is then doubled by a BBO crystal \cite{liu2017three,liu2016selfthesis, sharmaTailoring2019,sharmalaserdriven2019}. The laser spot size is $2w_0 = 35 \mu m$, the pulse duration is $\tau_p = 300 fs$, the laser energy is $E_p = 0.3 \mu J$, and the repetition rate $k_r$ is 500 kHz.
The analytical equation (Eq. \ref{eqn:heatEQChap6_eq2_21}) clearly demonstrates the pulse-by-pulse accumulation (see Fig. \ref{fgr:chap6_alpha_homo_fig1} (a)). For each pulse, the temperature-rise at laser center is around $19.3 ^{\circ} C$ and that decades quickly before the next pulse comes in. Because of the high repetition rate, the accumulation is observed looking at the dashed red line in Fig. \ref{fgr:chap6_alpha_homo_fig1} (a). To find out the influences of the laser writing speed, the base-temperature rises at $Vs = 5$ mm/s (Fig. \ref{fgr:chap6_alpha_homo_fig1} (b)) and $Vs = 50$ mm/s (Fig. \ref{fgr:chap6_alpha_homo_fig1} (c)) are calculated. In the case of $Vs = 5$ mm/s, the base-temperature rise increases quickly to 20 $^{\circ} C$ by the first 200 pulses and gradually to 40 $^{\circ} C$ after 1600 pulses. In comparison, for the writing speed of $Vs = 50$ mm/s, the temperature tends to saturate after 300 pulses. This result can be understood by the effective number of pulses ($N_{eff} = w_0 k_r / Vs$ ) inside the laser spot. The corresponding effective number of pulses are 1750 and 175 for $Vs = 5$ mm/s and $Vs = 50$ mm/s, respectively. For the writing speed of $Vs = 50$ mm/s, it is then clear that pulses whose number larger than 175 has gradually less impact on the accumulation because of the increasing distances for the heat to diffuse. As for $Vs = 5$ mm/s, this effect is relatively insignificant. Intuitively, if the writing speed is large enough that the diffusion can not catch up with, the contributions in accumulation by the effective number of pulses should decrease quickly. Here, set a limit at the speed of around 50 mm/s, which are the typical experimental values \cite{liu2017three, liu2016selfthesis, sharmaTailoring2019, sharmalaserdriven2019}.
\begin{figure}[ht!]
\centering
\includegraphics[width=\textwidth]{./figuresChapter6/chap6_alpha_homo_fig2}
\caption{Base-temperature rise obtained by using the effective number of pulses $N_{eff} = 1750$ (a) and (b), and $N_{eff} = 175$ (c) and (d). 2D maps (a) and (c) stand for the top surface of TiO$_2$ films. Fig. (b) and (d) are temperature profiles along the line at y=0 on the top surface.}
\label{fgr:chap6_alpha_homo_fig2}
\end{figure}
The spatial distributions of base-temperature rise at two different writing speeds are shown in Fig. \ref{fgr:chap6_alpha_homo_fig2}. Because the Eq. \ref{eqn:heatEQChap6_eq2_21} is deduced from the 3D heat equation, the solution is accurate without ignoring the diffusion in the y-axis (as we did in the 2D finite-element-model). To show this, the 2D maps of base-temperature rise at $Vs = 5$ mm/s and $Vs = 50$ mm/s are calculated and plotted in Fig. \ref{fgr:chap6_alpha_homo_fig2} (a) and (c), respectively. The circle-shapes by the two figures clearly show the y-axis diffusion. The maximum temperature is the same as Fig. \ref{fgr:chap6_alpha_homo_fig1} due to the same equation. The temperature shapes become comma-alike as the writhing speed increases, which are the results of diffusion effects. The diffusion lags can be seen from the Fig. \ref{fgr:chap6_alpha_homo_fig2} (b). In the case of Fig. \ref{fgr:chap6_alpha_homo_fig2} (d), the temperature difference is small and only small number of pulses are calculated, so that the lag is not clearly seen. However, this can be revealed by performing the simulation for more number of pulses.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.82\textwidth]{./figuresChapter6/chap6_fem_fewPulses_fig1}
\caption{Results of the finite element method model for the first 30 pulses. The temporal temperature at laser center (a). The spatial diffusion shown in pulse-by-pulse (b).}
\label{fgr:chap6_fem_fewPulses_fig1}
\end{figure}
To check the validity of the homogeneous assumption, simulations are performed based on the FEM model. The results by the first 30 pulses ($Vs = 5$ mm/s) are shown in Fig. \ref{fgr:chap6_fem_fewPulses_fig1}. The temporal temperature shown in Fig. \ref{fgr:chap6_fem_fewPulses_fig1} (a) is identical to the Fig. \ref{fgr:chap6_alpha_homo_fig1} (a). Figure \ref{fgr:chap6_fem_fewPulses_fig1} (b) shows the process of pulse absorption in the TiO$_2$ film and heat diffusion in x- and z-axis. The accumulation effects are clearly depicted. After each pulse, the absorbed laser energy is confined inside the TiO$_2$ layer, which is, however, quickly damped by diffusing the energy into the media in the vicinity. Since the thermal diffusivity of glass is larger than the mesoporous TiO$_2$, the heat diffusion (into the glass substrate) is faster along the laser incident direction than the horizontal writing directions (see Fig. \ref{fgr:chap6_fem_fewPulses_fig1} (b)). This is neglected in the semi-analytic model. Within the first 30 pulses, the laser only moves 0.3 $\mu m$ for $Vs = 5$ mm/s. Though the temperature damped dramatically before the next pulse comes in, the residual temperature is accumulated pulse-by-pulse that can in some how greatly affect the growing speed of Ag NPs as soon as the base exceeds a certain value. However, this is not observed in the first 30 pulses. It is noted that the critical temperature depends on the diffusion properties of Ag$^0$ as shown in the previous chapter. For the case after hundreds of pulses irradiation, it will be shown in the next discussions that the homogeneous absorption is inaccurate. Therefore, the coupled models are required to understand the writing-speed influences on the pulsed-laser controlling Ag NPs growth inside TiO$_2$ films.
\subsection{Results of coupled semi-analytic equations}
The diffusion in z-axis (last erf and exponential terms in Eq. \ref{eqn:heatEQChap6_eq1_15}) is significant in the beginning after each laser pulse. The z-diffusion term is written as:
\begin{equation}
\begin{aligned}
&\varphi_z(t,t_s^i) = \\
&\frac{1}{2} \bigg( \erf{\frac{h0+2\beta_T(t-t_s^i)\alpha_{abs}(x_s,t_s^i)}{\sqrt{4\beta_T(t-t_s^i)}}} - \erf{\frac{h_1|_{h_1 \to 0^-} + 2\beta_T(t-t_s^i)\alpha_{abs}(x_s,t_s^i)}{\sqrt{4\beta_T(t-t_s^i)}}} \bigg) \\
& \times \exp{\bigg( \beta_T (t-t_s^i) \big(\alpha_{abs}(x_s,t_s^i) \big)^2 \bigg) }
\label{eqn:heatEQChap6_eq1_22}
\end{aligned}
\end{equation}
In simulations of the heat accumulation by multiple pulses, however, the z-diffusion term is inefficient at large number of pulses. Because heat diffusion becomes much slower after a certain time, we try to fit the diffusion function by using a very simple one. Another reason why we introduce this approximation is that the numerical errors become significant after a certain time, as shown in Fig. \ref{fgr:chap6_coupledFit_fig1}(a) and (b). It is found that an abrupt changes in the z-diffusion curves depends on the absorption coefficient. In addition, we found that the breakdown does not disappear even if one uses smaller time steps. This turns out be due to the numerical errors in multiplication of a divergent function with a infinitesimal. To accelerate the parallel calculation on a GPU, the for-loop should be avoided. In addition, we found that if the $t_s^i$ is smaller than $100h_0/(2\beta_T \alpha_{abs}(x_s,t_s^i)) \approx 10 \mu s$, the z diffusion part tends to be very small. Here, we use $\beta_T = 3.4 \times 10 ^{-7} m^{2} s^{-1}$ for a glass substrate, the maximum absorption coefficient $\alpha_{abs} \approx 3.4 \times 10 ^{6} m^{-1}$, and the film thickness is 200 nm. Figure \ref{fgr:chap6_coupledFit_fig1} shows that the diffusion parts become identical and asymptotically approaching to zero, which are less than 4\% after 20 $\mu s$ for all the studied absorption cases. However, the small portion can become significant in the accumulation regime, so that it cannot be neglected in the multi-pulses regime. In order to reduce the calculation of exponentials on a GPU (which is quite inefficient), we fit the z-diffusion part after $20 \mu s$ for the studied system. The resulted equation can be written as follows:
\begin{equation}
\varphi_z(t) = m/(t)^n
\label{eqn:heatEQChap6_eq1_17_fit}
\end{equation}
\noindent where $m = 96.3\times 10^{-6} (s^{1/2})$, $n = 1/2$, and $t = t_g - t_s^i$, $t_g$ is the global time.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.6\textwidth]{./figuresChapter6/chap6_coupledFit_fig1}
\caption{Illustrations of numerical errors of the z-diffusion part (Eq. \ref{eqn:heatEQChap6_eq1_22}).}
\label{fgr:chap6_coupledFit_fig1}
\end{figure}
\noindent Figure \ref{fgr:chap6_coupledFit_fig2} shows the comparisons of the fitting to the Eq. \ref{eqn:heatEQChap6_eq1_22}. The good fitting is in the range from 20 $\mu s$ to 10 $ms$ indicating that we can accelerate the simulation by replacing exponential calculations by such a simple function. Furthermore, the numerical error is avoided in multiplying infinitesimal with a divergent quantity.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.5\textwidth]{./figuresChapter6/chap6_coupledFit_fig2}
\caption{Calculated fitting function of the z-diffusion part.}
\label{fgr:chap6_coupledFit_fig2}
\end{figure}
To further decrease the simulation time for this multi-scale problem, unlike the procedure used in the step marching method, long loops should be simplified in Eq. \ref{eqn:heatEQChap6_eq1_16}. To do so, the accounted number of pulses is defined as $N_c$, which is the nearest to $N_c$ pulses at time $t$. In this case, the temperature contribution by the pulses before the time $t - Nc/k_r$ is ignored. The z-diffusion part is, as illustrated in Fig. \ref{fgr:chap6_coupledFit_fig1} (a) and Fig. \ref{fgr:chap6_coupledFit_fig2}, less than 4\% after $3 \mu s$ and 1\% after $100 \mu s$. Thus, a typical choice of the accounted number of pulses can be $N_c = 100$ and the corresponding time of the heat transportation is $200 \mu s$. The temperature error is estimated to be less than 1\% of each pulses. The numerical solution of the coupled semi-analytic equations is described in this way. The photo-oxidation is not considered because of the femtosecond time scale.
The results of the carried out simulations are shown in Fig. \ref{fgr:chap6_coupled_ana_fig1} for laser fluences of 103 $m J / cm^{2}$ (a) and 268 $m J / cm^{2}$ (b). The size of NPs by higher fluence is bigger than the lower fluence one. Though the considered fluences are around 2 and 4 times larger than the experimental ones (62 $m J / cm^{2}$), a so-called 'fast-growth' due to a strong coupling of size-increment and light absorption is never observed in the two cases. The temperature is not high enough to activate the nonlinear growth of Ag NPs so as to make the size exceeding a certain value for a fast-growth.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.85\textwidth]{./figuresChapter6/chap6_coupled_ana_fig1}
\caption{Calculated profiles of Ag NP size, temperature and laser intensity. The laser scans from $x=-50 \mu m$ to $ 50 \mu m$. Here, laser fluences are 103 $m J / cm^{2}$ for (a) and 268 $m J /cm^{2}$ for (b). $N_c$ stands for the number of accounted laser pulses at time $t$. The green lines show the NP size, red lines cprrespond to temperature, and the dashed black lines demonstrate normalized laser intensities.}
\label{fgr:chap6_coupled_ana_fig1}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[width=0.85\textwidth]{./figuresChapter6/chap6_coupled_ana_fig2}
\caption{Simulation results by different number of accounted pulses: 200 (a), 300 (b), and 400 (c).}
\label{fgr:chap6_coupled_ana_fig2}
\end{figure}
To understand the reason why the temperature is underestimated, simulations are performed for different numbers of laser pulses ($N_c$), as shown in Fig. \ref{fgr:chap6_coupled_ana_fig2}. Laser fluence is set to be 103 $m J / cm^{2}$, and the writing speed is $5 mm/s$. For the cases with the number of accounted pulses $N_c = 100$, 200, 300, and 400, the maximum temperature are 57.2, 61.7, 64.4 and 65.4 $^{\circ} C$, respectively. The maximum size of Ag NPs are the same and of 3.08 nm for the studied number of accounted pulses. Furthermore, the heat diffusion is overestimated since the calculations by Eq. \ref{eqn:heatEQChap6_eq2_21} only considers the TiO$_2$ as the boundary condition. As a result, the estimated threshold in laser fluence is overestimated.
\begin{figure}[ht!]
\centering
\includegraphics[width=\textwidth]{./figuresChapter6/chap6_coupled_ana_fig3}
\caption{Simulations results of incident laser fluence at 309 $m J/cm^{2}$. The relation of the maximum NP size and temperature with number of pulses (a). The NP size, temperature and laser profiles at 2064 pulses (b), and 2481 pulses (c).}
\label{fgr:chap6_coupled_ana_fig3}
\end{figure}
As an an illustration of the fast growth, simulation are then performed for the laser fluence of 309 $m J/cm^{2}$ as shown in Fig. \ref{fgr:chap6_coupled_ana_fig3}. In this case, the reactions are much faster than in the case of 268 $m J/cm^{2}$. As a result, when the size of Ag NP exceeds 4.5 nm, the so-called "fast-growth" is clearly observed (Fig. \ref{fgr:chap6_coupled_ana_fig3}(a)) due to the strong nonlinear relation of size-absorption at wavelength 515 nm. The spatial profiles of NP size, temperature, and laser intensity shown in Fig. \ref{fgr:chap6_coupled_ana_fig3}(b) depicts the state before the "fast-growth". Similar to the continuous-wave writing of Ag NPs in TiO$_2$ films, the peaks of temperature, NP size, and laser intensity are misaligned. However, the maximum temperature is observed behind the beam center because of the slow growth-speed before this time. It is anticipated that the temperature is higher after the fast-growth so that the growth of NPs stops before the laser center arrives. Figure \ref{fgr:chap6_coupled_ana_fig3}(c) shows the case after the fast-growth. It is found that the estimated temperature exceeds 5000 $^{\circ} C$ after the fast growth, which looks impossible by experiments \cite{liu2017three,sharmalaserdriven2019}. Previous experiments \cite{destouches2014self} have shown that the big NPs after growth tend to reside into a single layer toward the glass side. In this case, the effective medium assumption is unreasonable so that the laser absorption should be calculated based on more rigorous methods such as FDTD, FEM and coupled waves.
\subsection{Results of coupled equations by FEM}
In the previous section we have developed a semi-analytical model to simulate the pulsed laser writing of Ag NPs. Because of the large loops in numerical calculations, simplifications are made to accelerate the simulations by only considering the contributions of the nearest $N_c$ pulses to the temperature. Furthermore, the heat transportation inside TiO$_2$ is regarded as the one like glass. All these assumptions lead to the overestimation of the fluence threshold for the growth of Ag NPs. A more accurate method can be developed by directly solving the heat equation using the step-marching algorithm in the framework of finite element method. In this way, the heat diffusion inside the TiO$_2$ film is considered.
Figure \ref{fgr:chap6_fem_all_fig1} shows the finite element method simulations at different number of irradiated pulses. The writing speed is 5 mm/s, the fluence is 103 $m J/cm^{2}$, , the spot size 35 $\mu m$, and repetition rate is 500 kHz. The effective conductivity, capacity of porous TiO$_2$ are considered. The film thickness is set to 200 nm. The initial temperature is low so that the size of the Ag NP increases slowly before the laser moves a distance of 25 $\mu m$ in the positive x-axis direction (before 2520 pulses as illustrated in the Fig. \ref{fgr:chap6_fem_all_fig1} (a)). For the next 360 pulses, the growth speed is accelerated so that the maximum NP size increase from 4.5 nm to 7.1 nm. Since then, the nonlinear relation of size-absorption leads to the dramatic growth of Ag NPs in the next short 115 pulses. In contrast, this fast-growth is never observed by the semi-analytic model with the same laser conditions. The threshold, as discussed in the previous section, is around $309 m J/cm^{2}$ that corresponds to pulse energy of $1.5 \mu J$, which is 4-5 times larger than the experiments of $60 m J/cm^{2}$ \cite{liu2017three, liu2016selfthesis}. The finite element method model, however, gives the threshold around $103 mJ/cm^{2}$ is much closer.
\newpage
\begin{figure}[ht!]
\centering
\includegraphics[width=0.95\textwidth]{./figuresChapter6/chap6_fem_all_fig1}
\caption{FEM simulations of Ag NPs growth under the pulsed-laser irradiation. The temperature distribution at different number of pulses (a), and the corresponding NPs, temperature and laser intensity on the top surface (b). The laser fluence is 103 $m J/cm^{2}$, the writing speed is 5 $mm/s$, and the spot size is 35 $\mu m$ at $1/e^2$.}
\label{fgr:chap6_fem_all_fig1}
\end{figure}
\newpage
\begin{figure}[ht!]
\centering
\includegraphics[width=0.5\textwidth]{./figuresChapter6/chap6_fem_all_fig2}
\caption{NP size, temperature, and laser profile after 8180 pulses irradiated on the sample. The dashed black line stands for the laser profile. Here, laser fluence is 103 $m J/cm^{2}$, the writing speed is 5 $mm/s$, and the spot size is 35 $\mu m$ at $1/e^2$.}
\label{fgr:chap6_fem_all_fig2}
\end{figure}
Figure \ref{fgr:chap6_fem_all_fig2} shows the size of NP, temperature, and laser profile after the laser moves a distance of 81.1 $\mu m$ from $x = -50 \mu m$, which corresponds to 8180 pulses. In reality, the maximum time-step is limited to less than 6 ns so as to precisely simulate the heat diffusion in TiO$_2$. The maximum size of Ag NP is around 48 nm, which is determined by the initial quantity of Ag$^0$ atoms inside the TiO$_2$ matrix. It it obvious that the NPs can grow rapidly due to the rather high temperature (around 3400 $^{\circ} C$). However, whether the temperature is overestimated is not known due to the lack of experiments. Intuitively, the glass is melt at this temperature. Recalling that the size of the formed NPs are comparable to the film thickness, the estimation of absorption obtained by using the classical effective medium approximation is no longer valid. Therefore, as indicated by the previous discussions, the model should be further modified to make a better prediction of the temperature.
\begin{figure}[ht!]
\centering
\includegraphics[width=\textwidth]{./figuresChapter6/chap6_fem_all_fig3}
\caption{Calculated NP size, temperature, and laser profiles by laser writing speed of 20 $mm/s$ and $50 mm/s$. The dashed black line stands for the laser profile. The laser fluence is 103 $m J/cm^{2}$, and the spot size is 35 $\mu m$ at $1/e^2$.}
\label{fgr:chap6_fem_all_fig3}
\end{figure}
Most of the experiments \cite{liu2017three, liu2016selfthesis, sharmaTailoring2019, sharmalaserdriven2019} have shown that NPs are smaller than 10 or 20 nm for speed range from 50 mm/s to 120 mm/s. To understand the high speed writing of Ag NPs, simulations are performed at speed of 20 mm/s and 50 mm/s of laser fluence of $103 mJ/cm^{2}$ (Fig. \ref{fgr:chap6_fem_all_fig3}). For both cases, the fast-growth of NPs are not observed due to the lower accumulated temperature. The speed-threshold by experiments are actually higher than the simulation, which is due to the overestimation of the Ag$^0$ activation energy. As revealed in the study for continuous-wave laser, the model covers the spectrum range by experiments if model coefficients are properly chosen. To better fit the results, more efforts are required, particularly in the parameter optimization methodology.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.65\textwidth]{./figuresChapter6/chap6_fem_all_fig4}
\caption{NP size, temperature, and laser profiles obtained for laser writing speed of 5 $mm/s$. The dashed black line stands for the laser profile. The laser fluence is 62 $m J/cm^{2}$, and the spot size is 35 $\mu m$ at $1/e^2$.}
\label{fgr:chap6_fem_all_fig4}
\end{figure}
Simulations performed for the same laser conditions ($62 mJ/cm^{2}$) with experiments \cite{liu2017three,liu2016selfthesis} are studied, which are shown in Fig. \ref{fgr:chap6_fem_all_fig4}. For the lowest speed that is around 5 mm/s, the size of Ag NPs never exceeds 3.5 nm, so that the fast-growth is not observed. Though the simulation is different from the experiments at the same condition, the results show that the existence of fluence threshold and the trend is in accordance with the experiments.
The simulated relation of Ag NP size with various fluences and writing speeds are summarized in Fig. \ref{fgr:chapter6_exp_02}. The experimental data are taken from Ref. \cite{liu2016selfthesis, liu2017three}. For G1 structure, LIPSS appeared on the TiO$_2$ surface which accompanied with Ag NPs (around 10-20 nm in diameter) within the grooves valleys and smaller NPs on the peak positions (Fig. \ref{fgr:chapter6_exp_02}(a)). As for G2, apart from the LIPSS on the material surface, Ag nano-gratings were formed closing to the glass substrate (Fig. \ref{fgr:chapter6_exp_02}(b-c)).
\begin{figure}[ht!]
\centering
\includegraphics[width=0.8\textwidth]{./figuresChapter6/chapter6_exp_02}
\caption{Comparisons of simulations with experiments by Ref. \cite{liu2016selfthesis}. The HAADF STEM image of G1 film cross-section (a); SEM images of G2 (b) and HAADF STEM top view of G2 (insert map), cross-section (c); EM images of G3 (d), the insert map is HAADF STEM image; AFM image of G4; the plot of Ag NP size by FEM and semi-analytic models ($Nc = 100$); the colored squred regions are from experiments by Ref. \cite{liu2016selfthesis}. White arrows represent the laser polarizations.}
\label{fgr:chapter6_exp_02}
\end{figure}
For the case of G3, the NPs were formed homogeneously on the top surface (Fig. \ref{fgr:chapter6_exp_02}(d)). As soon as the fluence decreased, LIPSS and Ag nano-gratings were indistinguishable and NPs were small (5 - 15 nm) for G4. Though the proposed models do not consider the formation of LIPSS and Ag nano-gratings, the performed simulations show the same trend in terms of the Ag NP growth (Fig. \ref{fgr:chapter6_exp_02}(f)). For the semi-analytic model taking effective pulses as Nc = 100, the NP size do not increase until the fluence increases to $309 mJ/cm^2$. The simulations by the FEM model is much closer to the experiments where large NPs are observed. The differences of these two models come from two factors as explained in the previous section: firstly, not enough effective pulses are considered for the semi-analytic model; secondly, the diffusion in TiO$_2$ films are ignored by the semi-analytic model. Thus, the FEM results are more precise in this case. Nevertheless, the fluence of generating large NPs is $103 mJ/cm^2$ by simulation and $37 mJ/cm^2$ by experiments \cite{liu2016selfthesis, liu2017three}. Probably, the plasmonic induced near-field enhancement boosts the light absorption by nonlinear effects, which is not considered by the present models. Furthermore, the estimated activation energy and coefficient of Ag diffusion by simulations are far away from the practical cases. The presented threshold for NP growth only exists in a specific region of fluences and writing speed that is observed by the experiments. The corresponding temperature rises shown by simulations provide reasonable explanations of the temperature decrease as the writing speed increases or the fluence decreases.
\section{Conclusions}
We proposed models describing the ultra-fast laser processing of mesoporous TiO$_2$ films loaded with Ag NPs. The femtosecond laser propagation in glass substrate was simulated by the nonlinear Schr\"{o}dinger equation coupled with the plasma equation and concluded that the absorption by glass can be ignored under the studied laser fluences and focusing conditions. Based on this analysis, a semi-analytic heat accumulation equation by a multi-pulse laser was deduced and coupled with the growth of Ag NPs and Ag$^+$ reduction processes. To largely accelerate the simulation, a GPU based model was developed. Different effective number of pulses were compared. The threshold fluence of Ag growth was shown to be around $309 mJ/cm^2$. For comparisons, the heat accumulation was simulated by a finite element method and coupled with the Ag growth related equations. As a result, the threshold fluence of Ag growth was around $103 mJ/cm^2$, which was much closer to the experiments by Ref. \cite{liu2017three, liu2016selfthesis}. The differences by the two models were explained and it concluded that the finite element method model was more precise comparing to the semi-analytic model. In the finite element method model, the heat diffusion in the mesoporous TiO$_2$ was considered. The simulations showed a similar region of fluence and writing speed where Ag NPs grew dramatically; outside the region, only small NPs were obtained. Because of the complexity of the practical process, the formation of surface grooves (LIPSS) and NP gratings were not considered by the simulation. Nevertheless, the proposed models paved the way for the modeling of Ag NPs formation and nano-gratings by pulsed-lasers in the future.
\afterpage{\null\newpage}
\chapter{General conclusions and outlook}
\section{Conclusions}
\hspace{0.5cm} This thesis provided the descriptions of several recently developed models and the results of numerical simulations of laser interaction with porous glass and mesoporous TiO$_2$ loaded with Ag nanoparticles. Part of the simulations were performed for porous materials without nanoclusters. In the another part, porous films were impregnated with metallic ions, irradiated by a weak UV light forming primary nanoparticles affecting material absorption and enabling their further growth or shrinking by the considered following laser irradiation.
Firstly, the periodic micro-void structures produced inside porous glass by femtosecond laser pulses were explained by a thermodynamic analysis. The laser propagation model was based on the nonlinear Schr\"{o}dinger equation coupled with electron plasma equation. The simulations were conducted in 1D+1 (time) taking advantage of the GPU parallelization. The performed thermodynamic analysis showed the possibility to efficiently control laser micro-machining in volume. For the given pulse repetition rate, the laser writing speed had great influences on the size and period of the formed voids. Furthermore, the densifications at low pulse energies were studied by simulation at different focusing conditions. The characteristic dimensions of the calculated temperature field were well correlated with experiments. The performed study provided a way of understanding the laser inscription of void arrays and densifications in porous glasses, which are of interest for porous glass based applications in optical waveguides, filters, and other integrated devices for microfluidics.
Comparing to the porous glass, laser writing of mesoporous TiO$_2$ films loaded with Ag nanoparticles enabled additional possibilities in the generation of different NP sizes as a function of laser power, focusing or writing speed. To elucidate the optimum laser irradiation conditions, a model taking into considerations Ag nanoparticle shrinkage via photo-oxidation, Ag$^+$ ion reduction, growth by Ostwald ripening, and plasmonic (size-dependent) light absorption was developed. The performed simulations showed dynamical changes in the variables such as Ag nanoparticles size, temperature field, and Ag$^+$ concentration. It was shown that the laser writing speed controlled the Ag nanoparticles size. As a result, the calculated transmission profile was inhomogeneous along the writing direction, which was demonstrated to be writing speed dependent. In order to accelerate the multi-scale simulation, the nanoparticle related processes such as growth, reduction, and photo-oxidation were solved using the Bulirsch-Stoer algorithm. The performed calculations depicted a novel view that Ag nanoparticles grow ahead of the laser spot center due to the heat diffusion. The mechanisms were attributed to the following processes
(\rom{1}) thermal activated nonlinear growth of nanoparticles that never stopped until the majority of free Ag$^0$ in the matrix were consumed;
(\rom{2}) the amount of Ag$^0$ created by reduction failed to compensate the Ag$^0$ consumption process (\rom{1});
(\rom{3}) as the growth process stopped, the photo-oxidation dominated the size-variation and finally played the controlling role;
As a result, the maximum size of Ag nanoparticles appeared in the laser front edge. The final size of Ag nanoparticle depended on the laser writing speed due to the different time for photo-oxidation. The simulated transmission profiles showed similar trend with the results of the in-situ transmission experiments. The differences were attributed to the remaining model limitations. The single size model, effective medium theory for thin film doped with Ag nanoparticles were considered as the origins of these differences.
Both spatial size distribution of nanoparticles and their plasmonic size-dependent absorption resulted in the temperature variations observed at different laser writing speeds. In particular, the temperature increase with writing speed was shown by the simulation, which seemed to be abnormal as less deposited energy led to higher temperature. This effect was explained by that fact that the Ag nanoparticle size was larger at a higher writing speed. Thus, it was the same for the absorption. In other words, the light absorption was more efficient at higher writing speeds. This result was confirmed by the TiO$_2$ phase-transition experiments. By studying the influences of the activation energy of Ag diffusion, diffusion coefficient of Ag, photo-oxidation rate, and initial concentration of Ag$^0$ and Ag$^+$, the model was shown to be robust in explaining the results of Raman experiments. Furthermore, a three-dimensional model was developed and simulated thanked to the high performances of the parallelization and scalable finite element method. The three-dimensional simulations reproduced the laser written lines in agreement with the experimental results.
In addition, the multi-physical modeling of Ag nanoparticle growth by a multi-pulsed laser was studied. This topic is interesting because the pulsed laser provided additional abilities in generating two kinds of nanostructures: the laser induced periodic surface grooves (LIPSS) and Ag nanogratings inside the TiO$_2$ film. To estimate the deposited laser energy, femtosecond laser propagation inside the glass substrate was studied. It was concluded that the absorptions inside the glass volume could be ignored at the considered wavelength, fluences and focusing conditions. Based on these results, a semi-analytic model was developed to estimate the heat accumulation by multiple pulses, which was coupled with the Ag nanoparticle growth equations. The simulations were accelerated by using a GPU. The threshold fluence of Ag growth was shown to be around $309 mJ/cm^2$. For comparisons, a finite element method based model was proposed for simulating the heat diffusion coupled with Ag nanoparticle growth. The threshold fluence was estimated to be around $103
mJ/cm^2$, which was closer to the reported experimental values. The finite element method was more precise comparing to the semi-analytic model since the heat diffusion inside the TiO$_2$ thin films was considered. The simulations were carried out for different fluences and writing speeds showing a region where Ag nanoparticles grew dramatically, similar to the reported experiments.
It should be noted that the developed models can be further extended to account for the formation of LIPSS and Ag nano-gratings in such media by coupling with nanoparticle migrations, surface melting and hydrodynamics. Thus, the range of the potential applications is rather wide. This work is, however, out of the scope of the present manuscript.
\section{Perspectives}
In laser processing of mesoporous TiO$_2$ films loaded with Ag NPs, self-organized Ag nanogratings with period in subwavelength scale are particularly attractive for optical applications, such as optical data storage \cite{rakuljic1995optical}, color printing\cite{diop2017spectral}, and image multiplexing \cite{sharmalaserdriven2019,sharmaTailoring2019}. In such films, standing waves are known to be formed\cite{destouches2014self}. The number of supported modes depends on film thickness, refractive index and wavelength. The presence of the guiding modes allows optical modulations of spatial distribution of Ag NPs in a sub-wavelength scale. Previously, a model based on the coupled mode theory \cite{destouches2014self} was proposed to predict the period of such nanostructures. It was concluded that guided waves enhanced due to the growth of Ag NPs. Among the involved mechanisms, several researches discussed possible contributions of the electromagnetic forces \cite{eurenius2008grating}, diffusion due to temperature gradient (Soret effect) \cite{braibanti2008does}, and diffusion due to density gradients \cite{smetanina2016modeling}. In future, the possible roles of ion and cluster diffusion should be better accounted for. Particularly, one of the possible effects is based on the so-called Soret-Dufour equation involving thermophoresis.
\subsection{Gratings induced by photo-oxidation}
In Chapter 3, where CW laser writing of Ag NPs in TiO$_2$ films was considered, the roles of photo-oxidation was proportional to the number of the absorbed photons by each nanoparticle. In this case, mostly laser amplitude played a role. Nevertheless, the scattered photons have also a probability to flow to other nanoparticles to be scattered and partly absorbed in the considered nanocoposite films. Figure \ref{fgr:chapter5_selfOrg_fig0} (a) shows the studied structure. The thickness of the TiO$_2$ film is 200 nm and its effective refractive index equals to 1.7 \cite{liu2016selfthesis,liu2015understanding}. Under laser illumination, Ag NPs can be considered as absorbing and scattering centers. Under dipole approximation conditions, these NPs act, in fact, as dipole sources (Fig. \ref{fgr:chapter5_selfOrg_fig0} (b)). For the laser polarization perpendicular to the studied plane (Fig. \ref{fgr:chapter5_selfOrg_fig0} (a,c)), waves with the incidence angle larger than the critical angle of the total reflection are reflected back into the film. As a result, the standing waves are formed, as the FDTD results have shown \cite{oskooi2010meep} in Fig. \ref{fgr:chapter5_selfOrg_fig0} (d).
\begin{figure}[ht!]
\centering
\includegraphics[width=0.95\textwidth]{./figuresChapter5Self/section1_fig0}
\caption{Thin film wave-guiding activated by a dipole source. The diagram of the structure (a); the dipole radiation pattern (b); illustrations of the propagation of waves at different angles (c); FDTD simulations (d).}
\label{fgr:chapter5_selfOrg_fig0}
\end{figure}
Only limited numbers of guiding modes are supported by the film, depending on the light wavelength, film thickness, and the refractive index. The guiding modes are obtained by the mode expansion method \cite{lohmeyer2002mode}. The period of standing waves are shown in Fig. \ref{fgr:chapter5_selfOrg_fig01} (a) for different TiO$_2$ thickness. For the laser wavelength of 532 nm, the 200 nm TiO$_2$ film supports only a single mode for TE and TM modes.
In the case of Ag NPs, or dipole sources, scattered light polarization corresponds to the TE mode. The activated guiding waves are parallel to the laser polarization, which is the same orientation as for Ag nanogratings after laser processing \cite{liu2016selfthesis, baraldi2016polarization,destouches2014self}.
The guiding-wave period obtained by the FDTD simulations is close to the mode expansion results as shown in Fig. \ref{fgr:chapter5_selfOrg_fig01} (b). As the thickness of the TiO$_2$ increases to around 420 nm and 810 nm, two and three modes appear for both TE and TM.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.7\textwidth]{./figuresChapter5Self/section1_fig0_2}
\caption{Standing wave period obtained by the mode expansion method for different thickness of TiO$_2$ films (a). Results of the FDTD simulation for a dipole-source activated standing wave.}
\label{fgr:chapter5_selfOrg_fig01}
\end{figure}
Considering the scattering centers (Ag NPs) as the TE oscillating sources, the standing waves contribute to the photo-oxidation. Therefore, the number of silver atoms leaving a nanoparticle per unit time due to the above-discussed processes is modified and can be written as:
\begin{equation}
n_{oxi}(t) =\eta_0 \frac{I(x, y, z, t)} {h \nu} \bigg( \sigma_{abs}(R,\lambda) + \xi \sigma_{sca}(R,\lambda) sin^2(\frac{\pi x}{d}) \bigg)
\label{eqn:conclu_eq1}
\end{equation}
\noindent where $I(x,y,z,t)$ is the laser intensity, $\eta_0$ is the ionization rate, $h \nu$ is the photon energy, $\sigma_{abs}(R,\lambda)$ is the absorption cross-section of the Ag NP with radius $R$, $\xi$ is the absorption ratio of scattered photons and ranges within (0, 1), $\sigma_{abs}(R,\lambda)$ is the cross-section of scattering, $d$ is the period of standing wave inside thin films, and $x$ is the coordinate. The squared sin function is a simplification of the guiding waves, which is invariant when laser moves.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.95\textwidth]{./figuresChapter5Self/section1_fig1}
\caption{Ag NP size distribution obtained due to the photo-oxidation modulation at writing speed $Vs = 150 \mu m/s$. (a) the time-x map of Ag NP size; (b) the magnification of the black dashed region shown in (a); (c) the spatial distribution of Ag NP when the laser center locates at $x = 36.6\mu m$; (d) the magnified view of the black dashed region in (c). The read dashed line in (c) represents the location where NP has the maximum size.}
\label{fgr:chapter5_selfOrg_fig1}
\end{figure}
In the presented Thesis, it is found that Ag NPs start growing at the laser front edge due to heat diffusion. The nonlinear growth was fast and it never stopped until the initial Ag$^0$ atoms were exhausted. The maximum size was ahead of the laser beam at a distance typically in the range of a few ten micrometers. Based on this idea and to largely decrease the computation efforts, simulations are performed by supposing that the distance at the maximum size is $10 \mu m$ before the laser center; the maximum size is set to be 50 nm; the absorption ratio of scattered photons is taken as 0.3; the ionization rate is $\eta_0 = 6.9 \times 10^{-5}$; laser wavelength $\lambda = 532 nm$. Because the reduction of Ag$^+$ into Ag$^0$ only slightly affects the growth, the simulations can be regarded as the process after the fast growth. In this way, the Ag NP spatial size distribution is modulated by the photo-oxidation. To accelerate the simulation, the code is written in "CUDA C" taking advantages of the GPU parallelization. The studied distance is $100 \mu m$, the spatial discretization is $\delta x = 17 nm$, and the period of the standing wave is $d = 340 nm$.
Figure \ref{fgr:chapter5_selfOrg_fig1}(a-d) shows the simulated Ag NP size at writing speed $Vs = 150 \mu m/s$. The laser scans from $x = 0\mu m$ to $x = 100\mu m$. Figure \ref{fgr:chapter5_selfOrg_fig1} (a-b) clearly shows the periodic profiles after photo-oxidation. While the laser moves from left side of $x = 0 \mu m$ to the right at $x = 100 \mu m$, the homogeneous spatial size distribution is modified and results into the formation of the nanoscale grooves. The further analysis as demonstrated in Fig. \ref{fgr:chapter5_selfOrg_fig1} (c-d) shows the period around 340 nm, which is exactly the period of the standing wave. The size at valley is around 20 nm and at peak is 25 nm.
For the laser writing speed of $Vs = 20 \mu m/s$ and smaller, the periodic grooves disappear. Figure \ref{fgr:chapter5_selfOrg_fig2} (a-b) shows the temporal variations of the Ag NP size along x-axis. It is found that at this writing speed, the grooves only exist inside the laser beam. Outside the region, the Ag NP size decreases to zero. In this case, the standing wave modulations are cancelled as the NP disappears by oxidation.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.95\textwidth]{./figuresChapter5Self/section1_fig2}
\caption{Ag NP size distribution by the photo-oxidation modulation at writing speed $Vs = 20 \mu m/s$. (a) the time-x map of Ag NP size; (b) the spatial distribution of Ag NP when the laser center locates at $x = 36.6\mu m$. The read dashed line in (b) represents the location where NP has the maximum size.}
\label{fgr:chapter5_selfOrg_fig2}
\end{figure}
In another opposite regime of the modulation at high writing speed, the resulted grooves are insignificant. Figure \ref{fgr:chapter5_selfOrg_fig3} (a) show the simulated Ag NP size along x-axis at different time. The difference in Ag NP size at peak and valley is small (less than 2 nm) as shown by Fig. \ref{fgr:chapter5_selfOrg_fig3}(b). In this case, due to the limited time for the photo-oxidation, the standing wave modulation is small.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.95\textwidth]{./figuresChapter5Self/section1_fig3}
\caption{Ag NP size distribution by the photo-oxidation modulation at writing speed $Vs = 20 mm/s$. (a) the time-x map of Ag NP size; (b) the spatial distribution of Ag NP when the laser center locates at $x = 36.6\mu m$. The read dashed line in (b) represents the location where NP has the maximum size.}
\label{fgr:chapter5_selfOrg_fig3}
\end{figure}
The size of Ag NP after photo-oxidation modulation is calculated and shown in Fig. \ref{fgr:chapter5_selfOrg_fig5} (a) for various speeds. It is clear that the high contrast grooves only exist if the writing speed is neither too slow (<40 $\mu m/s$) nor too fast (>30 $mm/s$). At low writing speed, the Ag NP tend to dissolve completely because of the strong size-reduction by photo-oxidation. For laser writing at a high speed, the standing wave modulation is insignificant because the oxidation time is limited and is too short. The maximum and minimum sizes and their difference for the grooves obtained at various writing speeds are shown in Fig. \ref{fgr:chapter5_selfOrg_fig5} (b-c). The contrast is defined as $\frac{D_{max}-D_{min}}{D_{max}+D_{min}}$. The abrupt rising followed by gradual decrements of the contrast value show the corresponding behavior of the standing wave modulation, which is writing speed dependent. The self-organization of Ag nano-gratings by continuous-wave laser was shown to take places only in a specific speed region ranging from $150 \pm 100 \mu m/s$ to $3 \pm 2 mm/s$ \cite{destouches2014self,liu2016selfthesis}. Here, the simulations present similar speed range. Nevertheless, the modeling is different from the experiments by Ref. \cite{destouches2014self}. For example, large Ag NPs (> 35 nm in diameter) were everywhere after the standing wave modulation by simulations; in contrast, these large NPs were aligned periodically and spaced in chains that formed 1D nanoparticle gratings \cite{destouches2014self}. Indeed, NPs moved by external forces to diffuse and finally formed the 1D pattern was not considered by this model. To the best of our knowledge, the self-organization process remains unclear till now. Possible origins may due to the electomagnetic force \cite{eurenius2008grating}, the Soret effect \cite{braibanti2008does}, and the electrostatic force. Though the optical tweezers \cite{ashkin1970acceleration} are commonly used for trapping particles in a liquid, whether the electromagnetic force can move Ag NPs in a mesoporous material by a low intensity (e.g. continuous-wave) laser is, however, questionable and should be accounted for only in for small easily polarised species.
Based on a simplified model, photo-oxidation modulation can lead to the formation of Ag nano-gratings. In addition, it is possible that NP motion and trapping can also play a role in the self-assembled grating formation.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.9\textwidth]{./figuresChapter5Self/section1_fig5}
\caption{Calculated map of Ag NP size vs writing speed and x-axis (a); the maximum, minimum and contrast of NP size for various speed (b); the NP size difference at the peak and in the valley (c).}
\label{fgr:chapter5_selfOrg_fig5}
\end{figure}
\subsection{Species migration}
As far as species migration is concerned, metallic species generally tend to move in mesoporous matrices subjected to laser radiation. The laser-triggered gradients of concentration, $C$, temperature, $T$, and electrostatic force, $\boldsymbol{F}$, are able to induce ion and cluster migration as follows:
\begin{equation}
\boldsymbol{J} = -D \nabla C - C D_T \nabla T + m C \boldsymbol{F}+C\boldsymbol{U}
\label{eqn:chap6_soret_eq2}
\end{equation}
\noindent where D is diffusion coefficient of the Ag species; $D_T$ is the thermal diffusion coefficient; $m =D/kT$; $k$ is Boltzmann constant; $\boldsymbol{U}$ is the velocity of a convective flow; and $T(r)$ is the thermal field induced by the laser. For different size of NPs, Soret coefficient is calculated as a function of temperature as follows: \cite{braibanti2008does}
\begin{equation}
\begin{aligned}
S_T (T,R) &= D_T / D \\
&= S^{\infty}_{T}(R) [1 - \exp(\frac{T_{sc}(R) - T}{T_0(R)})]
\label{eqn:chap6_soret_eq3}
\end{aligned}
\end{equation}
\noindent where $T_0$ is a fitting parameter, $R$ is the nanoparticle radius, and $T_{sc}(R)$ is the specific temperature, at which $S_T$ changes its sign.
When thermal gradient prevails, ion and small cluster drift velocity depends on their concentration, $C$, and on the size-dependent thermophoretic mobility, or thermal diffusion coefficient ($D_T$). Depending on the sign of $D_T$, the drift can be realized toward either the coldest or the hottest side, leading to a steady state concentration.
Thus, the presented model can be extended by including additional effects affecting Ag NPs growth and thermal diffusion to simulate the self-assembly of nanogratings. Optical properties of these structures is also a subject of additional studies that should be performed in future. This topic is too large is out of scope of the present thesis.
\newpage
\chapter{Appendix}
\section{Appendix A: deep neural networks for predicting nanocrystal size of anatase TiO$_2$}
Deducing the TiO$_2$ nanocrystal size based on Raman spectra is presented in this section. Traditionally, the inverse problem without human's handful adjustments of parameters can be done by the optimization algorithms. However, fitting of the Raman spectra requires selection method, patience and time. Alternatively, the problem can be solved without fitting algorithm through deep neural networks. Deep learning has attracted widespread attentions and increasing interests due to its great success in computer vision \cite{krizhevsky2012imagenet}, speech recognition \cite{hinton2012deep}, and gaming \cite{silver2016mastering}. Very recently, because of its capabilities in multi-parameters fitting on a high-dimensional space, deep learning algorithms have started to be applied in the field of researches of material design \cite{sanchez2018inverse, carrasquilla2017machine,raccuglia2016machine, yao2019intelligent}, microscopy and spectroscopy \cite{chen2014deep, wu2017air,esteva2017dermatologist, rivenson2017deep, ota2018ghost, rivenson2018phase}, and nano- photonic design \cite{liu2018training, ma2018deep,wiecha2019pushing, xu2019deep, malkiel2018plasmonic}. The algorithms of deep neural networks are also explored to be applied on photonic platforms \cite{lin2018all, shen2017deep, mengu2019analysis} to boost the calculations aiming at near light speed computing.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.8\textwidth]{./figuresChapter5/figure_Raman_ml_01}
\caption{Mean absolute error during training (a). Fully connected structure obtained by using Raman spectra as the input, while the output is nanocrystal size (b).}
\label{fgr:raman_ml_01}
\end{figure}
Among the deep neural networks for inverse design problem, the significant challenges in training the network based on non-unique mapped data sets are inevitable. This is fundamental for many physical problems in photonic design, and curve fittings. As a matter of fact, the non-unique spectrum-to-crystal size by the phonon confinement model is unfortunately the case. Certainly, one can solve this problem by pealing off the training data into a new data set, so that each data inside it is distinct. However, it requires time and patience, and turned out be inefficient \cite{liu2018training,kabir2008neural}.
To solve this problem, D. Liu et al. proposed a method based on a tandem network structure\cite{liu2018training}. The cascaded network consists of an inverse training network and a forward training network. The idea is to construct a network that can generate data in a distinct space based on the separated parts of networks. The process can mainly be divided as the following steps: firstly, forward training the second part network; the forward process is efficient because the mapping is unique in this direction; secondly, construct the first part of the network and feed the results into the previously trained network; the data set for this step can be very large, because the input are the same as the output; during the training, the weights inside the pre-trained network by the first step are set to be constant; after the training, the network automatically selects the weights so that the generated data locates in a distinct space.
Inspired by this idea, the nanocrystal of anatase TiO$_2$ are estimated using a deep neural network in the fitting Raman spectrum. The data set is generated based on the phonon confinement model (Eq. \ref{eqn:phonon-confinement_p1}). Before diving into the cascading network, the first attempt is to train an inverse-fitting network directly. Figure \ref{fgr:raman_ml_01}(a) shows the mean absolute error during training. The corresponding network structure is shown in Figure \ref{fgr:raman_ml_01}(b). It is obvious that the network is very hard to be trained as the mean absolute error oscillates even after 1000 epoches. The network used here has two hidden layers of 100 unit cells inside each layer. The input layer has 200 cells and the output is one cell. The training is still very hard even for deeper networks of 6 layers of 500 cells inside each layer.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.95\textwidth]{./figuresChapter5/figure_Raman_ml_02}
\caption{Training efficiency for the structure using size as the input and Raman spectrum as the output (a).The generating network (the dashed red rectangle) using discrete experimental data as input and the output is fed into the pre-trained network by step (a) (dashed green rectangle). Mean absolute error vs epoch for network shown figure (b).}
\label{fgr:raman_ml_02}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[width=\textwidth]{./figuresChapter5/figure_Raman_ml_03}
\caption{The network after training is then used for prediction (a). The predicted Raman spectra for experiments calcined at 400, 500, 600 and 700 $^{\circ} C$ are shown in figure (b), (c), (d), and (e) respectively. Experimental data (samples without seeds) are taken from Ref. \cite{sharma2019crystal}. The yellow dashed rectangle shows the network part for predicting the nanocrystal size (f).}
\label{fgr:raman_ml_03}
\end{figure}
On the contrary, the forward training is very efficient. Figure \ref{fgr:raman_ml_02}(a) shows the mean absolute error decrease rapidly and the vibrations is very small. The forward training network has identical structure to the inverse network used above, which only reverse the input and output layer. In the next step, a new network is designed and shown in Figure \ref{fgr:raman_ml_02}(b) (red dashed rectangle) having 6 hidden layers. The convolution layers are introduced for identifying Raman curve features. For example, the first hidden layer has 64 convolution cores of 3 pixels each. They are used to identify different features. The maxpooling layer of core of 3 pixels is to minimize the data size. The dropout layer is used to avoid over-fitting. The input layer has 27 cells which is the size of the experimental Raman spectrum. The output of this part network is then fed into the pre-trained forward network (shown by green dashed rectangle in Figure \ref{fgr:raman_ml_02}(b)).
After being successfully connected, the total network becomes a new deep network whose output layer has 200 cells representing the Raman spectrum in the range of 100 to 200 $cm^{-1}$. As discussed above, the data set used for training can be very large since the input and output are Raman spectra. For this sake, the Raman spectra are generated by the phonon confinement model. The input spectrum are selected according to the experimental Raman shift. The training by this network is then very efficient as shown by Figure \ref{fgr:raman_ml_02}(c).
The whole trained network (Figure \ref{fgr:raman_ml_03}(a)) can be used for Raman spectrum generation and fitting. Splitting the first part of that network, it can be used for predicting the nano crystal size of TiO$_2$. Figure \ref{fgr:raman_ml_03}((b)-(e)) shows the Raman curve comparisons of prediction and Experiment. The nano crystal size is predicted by the network shown by Figure \ref{fgr:raman_ml_03}(f). The good fittings indicates the inverse training is successful as indicated by the mean absolute error in Figure \ref{fgr:raman_ml_02}(c). Nevertheless, in contrast to the theoretical model, the experimental spectra are more expanded in the left part of the curve (for instance, in \ref{fgr:raman_ml_03}(b)). In fact, the differences comes from the phonon confinement model itself: the uncertainty of the dispersion relation of the anatase TiO$_2$, the assumption of isotropic dispersion relations, and the absence of consideration of surface effects for very small nanocrystals \cite{sharma2019crystal,zhang2000raman}.
\section{Appendix B: the timestep variations during laser writing in simulations}
\begin{figure}[ht!]
\centering
\includegraphics[width=0.95\textwidth]{./figuresChapter5/chapter5_appendix_p1}
\caption{Adaptive timestep (a) and the location of laser center (b) as functions of step number during the calculation. The relation of timestep and location of laser center is plotted in (c). The initial laser spot size is also depicted. The dramatic variations of timestep is obviously seen to happen inside the initial laser spot, which corresponds to the fast temperature change before a steady state.}
\label{fgr:chapter5_appd}
\end{figure}
The Bulirsch-Stoer algorithm provides the forecasting timestep that is very useful in calculations. As discussed in the previous sections, the time step varies according to the reaction speed. Thus, the timestep acts as the benchmark of the whole system. Figure \ref{fgr:chapter5_appd}(a) shows the relation of adaptive timestep with step number in calculations. The initial timestep is set to 0.1 ns. The timestep changes dramatically in the first few hundreds of steps, which slowly changes as the step marches forward before the steady value is reached. The same trend are observed throughout simulations at different writing speeds. The timestep at steady-state is smaller and the calculation stops earlier as the writing speed is higher. To identify the major role to timestep, the laser center locations are calculated as $V_S \times t$ and shown in Figure \ref{fgr:chapter5_appd}(b). The variations of timestep only appear before the laser moves 20 $\mu m$ in the writing direction. To find out the locations of non- steady-states, the relation of timestep and laser center locations are plotted in Figure \ref{fgr:chapter5_appd}(c). The insert map shows the laser spot. It is obvious that the timestep changes very quickly inside the region of size less than the laser spot(e.g. 20$\mu m$). As one may record from Figure \ref{fgr:chapter5_tempRise_p1} that the temperature changes at the same time, the temperature seems to be the major role in affecting the timestep. This is possible because temperature has great influences on the physicochemical process in Ag NP's growth.
\bibliographystyle{unsrt}
|
\section{Introduction}
\label{sec:intro}
\input{introduction.tex}
\section{Preliminaries and Problem Statement}
\label{Sec:Prelims}
\input{preliminaries.tex}
\section{Distributed Soft Covering using Algebraic Structured Random Codes}
\label{Sec:DistribuetdSoftCovering}
\subsection{Change of Measure Soft Covering}\label{lem:changeMeasureCovering}
\input{covering.tex}
\subsection{Main Result}
\input{distributed.tex}
\section{Proof of Distributed Soft Covering using Algebraic Structured Random Codes}
\input{proof.tex}
\appendices
\input{proofLemmas}
\input{proofProp}
\newpage
\bibliographystyle{IEEEtran}
\subsection{Binning of Random Encoders}
\label{sec:PMF binning}
We next proceed to binning the above constructed collection of random encoders.
Since, UCC is already a union of several cosets, we associate a bin to each coset, and place all the codewords of a coset in the same bin. For each $i\in \mathbb{F}_p^{l_1}$ and $j\in \mathbb{F}_p^{l_2}$, let $\mathcal{B}^{(\mu_1)}_1(i) \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \mathbb{C}(G^{},h_1^{(\mu_1)}(i))$ and $\mathcal{B}^{(\mu_2)}_2(j) \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \mathbb{C}(G^{},h_2^{(\mu_2)}(j))$ denote the $i^{th}$ and the $j^{th}$ bins, respectively. Formally, we define the following PMFs.
\begin{eqnarray}\label{Eqn:dist_PMFInducedByEncoder1}
p_{M_{1}|X_{1}^{n}}^{(\mu_1)}(m_{1}|x_{1}^{n}) =
\begin{cases}
\mathbbm{1}_{\{m_i=0\}}& \mbox{if } s_{i}^{(\mu_i)}(x_{i}^{n}) > 1,\\
1-s_{i}^{(\mu_i)}(x_{i}^{n})&\mbox{if }m_{i}=0 \mbox{ and } s_{i}^{(\mu_i)}(x_{i}^{n}) \!\in [0, 1],\\
\displaystyle \sum_{a_i\in\mathbb{F}_p^{k}}E^{(\mu_i)}_{L_i|X_i^n}(a_i,\!m_i|x_i^n) &\mbox{if }m_{i}\neq0 \mbox{ and
}s_{i}^{(\mu_i)}(x_{i}^{n}) \!\in [0, 1],
\end{cases}\nonumber
\end{eqnarray}
for all $x_i^n \in T_{\delta}(X_i)$, $s_i^{(\mu_i)}(x_{i}^{n})$ defined as $s_i^{(\mu_i)}(x_{i}^{n}) \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{a_i\in\mathbb{F}_p^{k}}\sum_{m_i \in \mathbb{F}_p^{l_i}}E^{(\mu_i)}_{L_i|X_i^n}(a_i,m_i|x_i^n)$ and $i\in[2]$.
For $x_1^n \notin T_{\delta}(X_1)$, we let $ p_{M_{1}|X_{1}^{n}}^{(\mu_1)}(m_{1}|x_{1}^{n}) = \mathbbm{1}_{\{m_1=0\}}.$
With this definition note that, $ \sum_{m_1=0}^{2^{nR_1}}p_{M_1|X_1^n}^{(\mu_1)}(m_1|x_1^n) = 1$ for all $\mu_1 \in [2^{nC_1}]$ and $x^n_1 \in \mathcal{X}_1^n$ and similarly, $ \sum_{m_2=0}^{2^{nR_2}}p_{M_2|X_2^n}^{(\mu_2)}(m_2|x_2^n) = 1$ for all $\mu_2 \in [2^{nC_2}]$ and $ x^n_2 \in \mathcal{X}_2^n$.
Also, note that the effect of introducing binning (by defining the above PMFs) is in reducing the communication rates from $(S_1, S_2)$ to $(R_1,R_2)$, where $ R_i = \frac{l_i}{n}\log{p}, i\in\{1,2\}$. Now, we move on to describing the decoder.
\subsection{Decoder mapping }
We create a decoder that takes as an input a pair of bin numbers and produces a sequence $W^n \in \mathbb{F}^n_p$. More precisely, we define a mapping $f^{(\mu)}$ for $\mu \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} (\mu_1,\mu_2)$, acting on the messages $(m_1,m_2)$ as follows.
On observing $ \mu $ and the classical indices $ (m_1,m_2) \in \mathbb{F}_p^{l_1}\times \mathbb{F}_p^{l_2} $ communicated by the encoder, the decoder constructs $D^{(\mu)}_{i,j} \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \{\tilde{a} \in \mathbb{F}_p^{k}: \tilde{a}G^{} + h_1^{(\mu_1)}(i)+h_2^{(\mu_2)}(j) \in \mathcal{T}_{\hat{\delta}}^{(n)}(Z) \},$ and $f^{(\mu)}(m_1,m_2)$
\begin{align}
\mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}}
\begin{cases}
\tilde{a}G^{} + h_1^{(\mu_1)}(i)+h_2^{(\mu_2)}(j) &\quad \text{ if } D^{(\mu_1,\mu_2)}_{i,j} \equiv \{\tilde{a}\} \\
w^n_0 &\quad \text{ otherwise },
\end{cases}
\label{eq:POVM-17}
\end{align}
where $\hat{\delta} = p\delta$ and
{$w_0^n$ is an additional sequence added to $ \mathbb{F}_p^n $}.
Further, $f^{(\mu)}(m_1,m_2)=w_0^n$ for $i=0$ or $j=0$.
The decoder then performs a stochastic processing of the output and chooses $y^n$ according to PMF $p^n_{Y|Z}(y^n|f^{(\mu)}(m_1,m_2))$. This
implies the PMF $p^{(\mu_1)}_{Y^n|M_1 M_2}(\cdot|\cdot)$ is given by
\begin{eqnarray}
\label{Eqn:dist_PMFInducedByDecoder}
p^{(\mu)}_{Y^n|M_1M_2}(\cdot|m_1,m_2) = p^n_{Y|Z}(y^{n}|f^{(\mu)}(m_1,m_2)).
\end{eqnarray}
We now begin our analysis of the total variation term given in (\ref{eq:dist_main_lemma}).
\subsection{Analysis of Total Variation}
Our goal is to prove the existence of a collections $c_{1},c_{2}$ for which (\ref{eq:dist_main_lemma}) holds. We do this via random coding. Specifically, we prove that $\mathbb{E}[{K}] \leq \epsilon$, where the expectation is over the ensemble of codebooks. The PMF induced on the ensemble of codebooks is as specified below. The codewords of the random codebook
$C_{i}^{(\mu_i)} = (\mathtt{W}_i(a_i,m_i,\mu_i): a_i \in \mathbb{F}_p^k, m_i \in \mathbb{F}_p^{l_i})$ for each $\mu_i \in [2^{nC_i}]$
are only pairwise independent \cite{pradhanalgebraic} and distributed with PMF $ \mathbb{P}(\mathtt{W}_{i}(a_i,m_i,\mu_i ) = w_{i}^{n}) = \frac{1}{p^n}$
for each $i \in [2].$
\noindent \textbf{Step 1: Error caused by not covering}\\
We begin by splitting $K$ into two terms using the triangle inequality as $K \leq S + \tilde{S}$, where
\begin{align}
S \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} &\sum_{\underline{x}^n,y^n}\Bigg|p^n_{\underline{X}Y}(\underline{x}^n,y^n) - \sum_{\substack{ \mu_1,\mu_2 }}\sum_{\substack{m_1 >0,\\m_2 >0}}
\frac{p^{n}_{\underline{X}}(\underline{x}^n)p^{(\mu_1)}_{M_1|X_1^{n}}(m_1|x_1^{n})}{2^{n(C_1+C_2)}} p^{(\mu_2)}_{M_2|X_2^{n}}(m_2|x_2^{n}
)p^{(\mu) } _ { Y^ { n }
|\underline{M}}(y^{n}|\underline{m})\Bigg|, \nonumber\\
\tilde{S} \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} &\sum_{\underline{x}^n,y^n}\Bigg|\sum_{\substack{ \mu_1,\mu_2 }}\sum_{\substack{m_1=0 \medcup m_2=0}}
\frac{p^{n}_{\underline{X}}(\underline{x}^n)}{2^{n(C_1+C_2)}}p^{(\mu_1)}_{M_1|X_1^{n}}(m_1|x_1^{n})p^{(\mu_2)}_{M_2|X_2^{n}}(m_2|x_2^{n}
)p^{(\mu) } _ { Y^ { n }
|\underline{M}}(y^{n}|\underline{m})\Bigg|. \nonumber
\end{align}
\noindent Note that $\widetilde{S}$ captures the error induced by not covering $p^n_{\underline{X}Y}$.
For the term corresponding to $\widetilde{S}$, we
prove the following result by developing the following lemma below followed by a proposition.
\begin{lem}\label{lem:NotaPMFBound}
For the above defined notations, for $i \in \{1,2\}$, if $S_i \geq I(X_i;W_i) + \delta_{c_i}$, then the following holds true
\begin{align}\label{eq:NotaPMF}
\frac{1}{2^{nC_i}}\sum_{\substack{ \mu_i }}\sum_{{x}_i^n}{p^{n}_{{X}_i}({x}_i^n)} \mathbb{P}\Bigg(\bigg[\sum_{a_i\in\mathbb{F}_p^{k}}\sum_{m_i \in \mathbb{F}_p^{l_i}}E^{(\mu_i)}_{L_i|X_i^n}(a_i,m_i|x_i^n)\bigg] > 1\Bigg) \leq \epsilon_{c_i}
\end{align}
\end{lem}
\begin{proof}
The proof is provided in Appendix \ref{proof_lem:NotaPMFBound}.
\end{proof}
\begin{proposition}\label{prop:Lemma for Tilde_S}
There exist functions $\epsilon_{\widetilde{S}}(\delta), $ and $\delta_{\widetilde{S}}(\delta)$,
such that for all sufficiently small $\delta$ and sufficiently large $n$, we have $\mathbb{E}[\widetilde{S}] \leq\epsilon_{\widetilde{S}}(\delta) $, if $S_1 > I(X_1;W_1) - H(W_1) +\log{p} + \delta_{\widetilde{S}} $ and $S_2 > I(X_2;W_2) - H(W_2) + \log{p} +\delta_{\widetilde{S}},$ where $\epsilon_{\widetilde{S}}, \delta_{\widetilde{S}} \searrow 0$ as $\delta \searrow 0$.
\end{proposition}
\begin{proof}
The proof is provided in Appendix \ref{proof_prop:Lemma for Tilde_S}
\end{proof}
Now we move on to removing from $S$ the error that is induced due to binning.
\noindent{\bf Step 2: Error caused by binning}\\
Note that $S$ can be simplified using the definitions of $P_{M_1|X_1^n}^{(\mu_1)}(\cdot|\cdot)$, $P_{M_2|X_2^n}^{(\mu_2)}(\cdot|\cdot)$, and $p^{(\mu) } _ { Y^ { n }
|\underline{M}}(y^{n}|\underline{m}) $ as
\begin{align}
S \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} & \sum_{\underline{x}^n,y^n}\Bigg|p^n_{\underline{X}Y}(\underline{x}^n,y^n) -
\sum_{\substack{ \mu_1,\mu_2 }}\sum_{\substack{ m_1 > 0, \\m_2>0 }}\sum_{\substack{a_1 \in \mathbb{F}_p^k\\a_2 \in \mathbb{F}_p^k}}\sum_{\substack{w_1,w_2 \in \mathbb{F}_p^n}} \frac{p^{n}_{\underline{X}}(\underline{x}^n)}{2^{n(C_1+C_2)}}E^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n})E^{(\mu_2)}_{W^n_2|X_2^{n}}\!(w^n_2|x_2^{n}) \nonumber \\ & \hspace{150pt}
\mathbbm{1}_{\left\{ \mathtt{w}_1(a_1,m_1,\mu_1) = w_1^n\right\}}\mathbbm{1}_{\left\{ \mathtt{w}_2(a_2,m_2,\mu_2) = w_1^n\right\}}
p_{Y|Z}^n(y^{n}|f^{(\mu)}(m_1,m_2))\Bigg|,\nonumber
\end{align}
where $E^{(\mu_i)}_{W_i^n|X_i^n}(w_i^n|x_i^n)$ is defined as
\begin{align}
E^{(\mu_i)}_{W_i^n|X_i^n}(w_i^n|x_i^n) & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \frac{p^n}{2^{nS_i}(1+\eta)}{p_{W_i|X_i}^n(w_i^n|x_i^n)} \mathbbm{1}_{\left\{w_i^n \in T_{\delta}(W_i|x_i^n)\right\}} \mathbbm{1}_{\{s_i^{(\mu_i)}(x_i^n) \leq 1\}}, \quad \text{ for i $\in \{1,2\}$ } .
\nonumber
\end{align}
Further by defining $\gammaCoeff$ and $\zetaCoeff$ as
\begin{align}
\gammaCoeff & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} |\{\mathtt{w}_1(a_1,m_1,\mu_1) = w_1^n\} | = \sum_{m_1 > 0}\sum_{a_1 \in \mathbb{F}_{q}^k} \mathbbm{1}_{\{\mathtt{w}_1(a_1,m_1,\mu_1) = w_1^n\}} \quad \text{ and } \nonumber \\
\zetaCoeff & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} |\{\mathtt{w}_2(a_2,m_2,\mu_2) = w_2^n\} | = \sum_{m_2 > 0}\sum_{a_2 \in \mathbb{F}_{q}^k} \mathbbm{1}_{\{\mathtt{w}_2(a_2,m_2,\mu_2) = w_2^n\}}.
\end{align}we bound $S$ using triangle inequality as $S \leq S_1 + S_2$, where
\begin{align}
S_1 &\mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{\underline{x}^n,y^n}\Bigg|p^n_{\underline{X}Y}(\underline{x}^n,y^n) - \sum_{\substack{ \mu_1,\mu_2 }}\sum_{\substack{w_1^n,w_2^n \in \mathbb{F}_p^n}} \frac{p^{n}_{\underline{X}}(\underline{x}^n)}{2^{n(C_1+C_2)}} \gammaCoeff \zetaCoeff
\nonumber \\
&\hspace{170pt}E^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n})E^{(\mu_2)}_{W^n_2|X_2^{n}}(w^n_2|x_2^{n}
)p^{n } _ {Y|Z}(y^{n}|w_1^n\oplus w_2^n)\Bigg|,\nonumber \\
S_2 &\mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{\underline{x}^n,y^n}
\sum_{\substack{ \mu_1,\mu_2 }}\sum_{\substack{ m_1 > 0, \\m_2>0 }}\sum_{\substack{a_1 \in \mathbb{F}_p^k\\a_2 \in \mathbb{F}_p^k}} \sum_{\substack{w_1^n,w_2^n \in \mathbb{F}_p^n}} \!\frac{p^{n}_{\underline{X}}(\underline{x}^n)}{2^{n(C_1+C_2)}} E^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n})E^{(\mu_2)}_{W^n_2|X_2^{n}}(w^n_2|x_2^{n}
) \nonumber \\ &\hspace{80pt} \mathbbm{1}_{\left\{ \mathtt{w}_1(a_1,m_1,\mu_1) = w_1^n\right\}}\mathbbm{1}_{\left\{ \mathtt{w}_2(a_2,m_2,\mu_2) = w_1^n\right\}} \Bigg|p^n_{Y|Z}(y^{n}|w_1^n\oplus w_2^n)\! -\! p^n_{Y|Z}(y^{n}|f^{(\mu)}(m_1,\!m_2))\Bigg|.\nonumber
\end{align}
To bound the term corresponding to $S_2$, we provide the following proposition.
\begin{proposition}[Mutual Packing]\label{prop:Lemma for S_2}
There exist $\epsilon_{S_{2}}(\delta),$ such that for all sufficiently small $\delta$ and sufficiently large $n$, we have $\mathbb{E}\left[{S}_2\right] \leq \epsilon_{{S_2}}(\delta) $, if $ S_{1} - R_1 < \log{p} - H(Z)$, or equivalently, $ S_{2} - R_2 < \log{p} - H(Z)$, where $\epsilon_{{S}_2} \searrow 0$ as $\delta \searrow 0$.
\end{proposition}
\begin{proof}
The proof is provided in Appendix \ref{proof_prop:Lemma for S_2}.
\end{proof}
\noindent Now, we move on to analyzing the term corresponding to $S_1$.\\
\noindent\textbf{Step 3: Term concerning Alice's encoding}\\
\noindent In this step, we separately analyze the action of the two encoders in approximating the product distribution $p_{\underline{X}Y}^n(\cdot)$. For that, we split $S_{1}$ as $ S_{1} \leq Q_1 + Q_2 $, where
\begin{align}
Q_1 \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} & \sum_{\underline{x}^n,y^n}\Bigg|p^n_{\underline{X}Y}(\underline{x}^n,y^n) - \frac{1}{2^{nC_1}}\sum_{\substack{ \mu_1 }}\sum_{\substack{w_1^n,w_2^n \in \mathbb{F}_p^n}}{p^{n}_{\underline{X}}(\underline{x}^n)}\gammaCoeff E^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n})\nonumber \\ & \hspace{3in} p^n_{W^n_2|X_2^{n}}(w^n_2|x_2^{n}
)p^n_ { Y|{Z}}(y^{n}|w_1^n+w_2^n)\Bigg|,\nonumber \\
Q_2 \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}}& \!\! \sum_{\underline{x}^n,y^n}\Bigg|\frac{1}{2^{nC_1}}\sum_{\substack{ \mu_1 }}\sum_{\substack{w_1^n,w_2^n \in \mathbb{F}_p^n}} {p^{n}_{\underline{X}}(\underline{x}^n)}\gammaCoeff E^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n}) \nonumber \\ & \hspace{80pt}\left(p^n_{W^n_2|X_2^{n}}(w^n_2|x_2^{n})\! - \zetaCoeff E^{(\mu_2)}_{W^n_2|X_2^{n}}(w^n_2|x_2^{n})\right)
p^n_ { Y|{Z}}(y^{n}|w_1^n\!\!+\!\!w_2^n) \Bigg|.\nonumber
\end{align}
With this partition, the terms within the trace norm of $ Q_1 $ differ only in the action of Alice's encoder. And similarly, the terms within the norm of $ Q_2 $ differ only in the action of Bob's encoder. Showing that these two terms are small forms a major portion of the achievability proof. \\
\noindent{\textbf{Analysis of $ Q_1$}:}
To prove $ Q_1 $ is small, we characterize the rate constraints which ensure that an upper bound to $ Q_1 $ can be made to vanish in an expected sense. In addition, this upper bound becomes useful in obtaining a single-letter characterization for the rate needed to make the term corresponding to $ Q_2 $ vanish. For this, we define $ J $ as
\begin{align}
J \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} &\sum_{\underline{x}^n,w_2^n,y^n}\Bigg|p^n_{\underline{X}W_2Y}(\underline{x}^n,y^n) - \frac{1}{2^{nC_1}}\sum_{\substack{ \mu_1}}\sum_{\substack{w_1^n}} {p^{n}_{\underline{X}}(\underline{x}^n)}\gammaCoeff E^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n})\nonumber \\
&\hspace{3in}p^n_{W^n_2|X_2^{n}}(w^n_2|x_2^{n})
p^n_ { Y|{Z}}(y^{n}|w_1^n+w_2^n)\Bigg|.\nonumber
\end{align}
By again using triangle inequality we obtain $J \leq J_1 + J_2$, where
\begin{align*}
J_1 & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{\underline{x}^n,w_2^n,y^n}\Bigg|p^n_{\underline{X}W_2Y}(\underline{x}^n,y^n) - \frac{1}{2^{nC_1}}\sum_{\substack{ \mu_1}}\sum_{\substack{w_1^n}} {p^{n}_{\underline{X}}(\underline{x}^n)}\gammaCoeff \bar{E}^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n})\nonumber \\
&\hspace{3in}p^n_{W^n_2|X_2^{n}}(w^n_2|x_2^{n})
p^n_ { Y|{Z}}(y^{n}|w_1^n+w_2^n)\Bigg| \\
J_2 & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{\underline{x}^n,w_2^n,y^n}\Bigg|\frac{1}{2^{nC_1}}\sum_{\substack{ \mu_1}}\sum_{\substack{w_1^n}} {p^{n}_{\underline{X}}(\underline{x}^n)}\gammaCoeff \left(\bar{E}^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n}) - E^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n})\right)\nonumber \\
&\hspace{3in}p^n_{W^n_2|X_2^{n}}(w^n_2|x_2^{n})
p^n_ { Y|{Z}}(y^{n}|w_1^n+w_2^n)\Bigg|
\end{align*}
where $\bar{E}^{(\mu_1)}_{W^n_1|X_1^{n}}(\cdot|\cdot)$ is defined as
\begin{align}
E^{(\mu_1)}_{W_1^n|X_1^n}(w_1^n|x_1^n) & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \frac{p^n}{2^{nS_1}(1+\eta)}{p_{W_1|X_1}^n(w_1^n|x_1^n)} \mathbbm{1}_{\left\{w_1^n \in T_{\delta}(W_1|x_1^n)\right\}}.
\end{align}
To prove the term corresponding to $J_1$ is small, consider the following proposition.
\begin{proposition}
\label{prop:Lemma for J1}
There exist $\epsilon_{J_1}(\delta), \delta_{J_1}(\delta)$ such that for all sufficiently small $\delta$ and sufficiently large $n$, we have $\mathbb{E}\left[J_1\right] \leq \epsilon_{J_1}$ if $S_1 + C_1 \geq I(W_1;X_1X_2ZW_2) +\log{p} - H(W_1) + \delta_{J_1}$, where $\epsilon_{{J_1}}, \delta_{{J_1}} \searrow 0$ as $\delta \searrow 0$.
\end{proposition}
\begin{proof}
The proof is provided in Appendix \ref{proof_prop:Lemma for J1}.
\end{proof}
\noindent Now, consider the term corresponding to $J_2$. This can be simplified as
\begin{align*}
J_2 &= \frac{1}{2^{nC_1}}\sum_{{x}_1^n} \sum_{\substack{ \mu_1}} {p^{n}_{\underline{X}}(\underline{x}^n)}\left(\sum_{\substack{w_1^n}}\gammaCoeff \bar{E}^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n})\right)\mathbbm{1}_{\{s_1^{(\mu_1)}(x^n_1) > 1\}} \\
& \leq \frac{1}{2^{nC_1}}\sum_{{x}_1^n} \sum_{\substack{ \mu_1}} {p^{n}_{\underline{X}}(\underline{x}^n)} \mathbb{E}\left[\sum_{\substack{w_1^n}}\gammaCoeff \bar{E}^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n})\right]\mathbbm{1}_{\{s_1^{(\mu_1)}(x^n_1) > 1\}} \\
& \hspace{10pt} + \frac{1}{2^{nC_1}}\sum_{{x}_1^n} \sum_{\substack{ \mu_1}} {p^{n}_{\underline{X}}(\underline{x}^n)} \left|\sum_{\substack{w_1^n}}\gammaCoeff \bar{E}^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n}) - \mathbb{E}\left[\sum_{\substack{w_1^n}}\gammaCoeff \bar{E}^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n})\right]\right| \\
& \leq \frac{1}{2^{nC_1}}\sum_{{x}_1^n} \sum_{\substack{ \mu_1}} {p^{n}_{\underline{X}}(\underline{x}^n)}\mathbb{E}\left[\sum_{\substack{w_1^n}}\gammaCoeff \bar{E}^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n})\right] \mathbbm{1}_{\{s_1^{(\mu_1)}(x^n_1) > 1\}} \\
& \hspace{10pt} + \frac{1}{2^{nC_1}}\sum_{{x}_1^n} \sum_{\substack{ \mu_1}} {p^{n}_{\underline{X}}(\underline{x}^n)} \left|\sum_{\substack{w_1^n}}\gammaCoeff \bar{E}^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n}) - \mathbb{E}\left[\sum_{\substack{w_1^n}}\gammaCoeff \bar{E}^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n})\right]\right| \\
& \leq H_0 + H_1 + \epsilon'',
\end{align*}
for
\begin{align*}
H_0 & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \frac{1}{2^{nC_1}}\sum_{{x}_1^n} \sum_{\substack{ \mu_1}} {p^{n}_{\underline{X}}(\underline{x}^n)} \mathbbm{1}_{\{s_1^{(\mu_1)}(x^n_1) > 1\}} \\
H_1 & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \frac{1}{2^{nC_1}}\sum_{{x}_1^n} \sum_{\substack{ \mu_1}} \left| {p^{n}_{\underline{X}}(\underline{x}^n)} - \frac{p^n}{2^{nS_1}(1+\eta)} \sum_{\substack{w_1^n}}\sum_{m_1 > 0}\sum_{a_1 \in \mathbb{F}_{q}^k} {p_{X_1W_1}^n(x_1^n,w_1^n)}\mathbbm{1}_{\{\mathtt{w}_1(a_1,m_1,\mu_1) = w_1^n\}}\right|
\end{align*}
where we use $ \mathbb{E}\left[\sum_{\substack{w_1^n}}\gammaCoeff \bar{E}^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n})\right] \leq 1$ in defining $H_0$ and $H_1$ is obtained by adding the sequences $w_1^n \notin \mathcal{T}_{\delta}^{(n)}(W_1)$ within the summation.
Now, we can provide an upper bound on $H_0$ and $H_1$ using the Lemmas \ref{lem:NotaPMFBound} and \ref{lem:changeMeasureCoveringLemma}, respectively, as $\mathbb{E}[H_0 + H_1 ] \leq \epsilon_{H} $ if $S_1 \geq I(X_1;W_1) + \delta_{H}$. Therefore, since $Q_1 \leq J$, hence $ Q_1$, can be made arbitrarily small for sufficiently large n, if $S_1+C_1 > I(W_1;X_1X_2YW_2) - H(W_1) + \log{p} + \delta_{J}$ . Now we move on to bounding $ Q_2 $.\\
\noindent{\bf Step 4: Analysis of Bob's encoding}\\Step 3 ensured that the random variables $X_1X_2YW_2$ are close to a product PMF in total variation. In this step, we approximate the PMF of random variables $X_1X_2Y$ using the Bob's encoding rule and bound the theorem corresponding to $Q_2$. We proceed with the following proposition.
\begin{proposition} \label{prop:Lemma for Q2}
There exist functions $\epsilon_{Q_{2}}(\delta) $ and $\delta_{{Q}_{2}}(\delta)$, such that for all sufficiently small $\delta$ and sufficiently large $n$, we have $\mathbb{E}[{Q}_2] \leq\epsilon_{{Q_2}}$, if $ S_1+C_1\geq I(W_1;X_1X_2YW_2) - H(W_1)_{} + \log{p} + \delta_{Q_2}$ and $ S_2+C_2 \geq I(W_2;X_1X_2Y) - H(W_2)_{} + \log{p} + \delta_{Q_2}$, where $\epsilon_{Q_2},\delta_{{Q}_2} \searrow 0$ as $\delta \searrow 0$.
\end{proposition}
\begin{proof}
The proof is provided in Appendix \ref{proof_prop:Lemma for Q_2}.
\end{proof}
Hence, in bounding the terms corresponding to $Q_1$ and $Q_2$, we have obtained the following constraints:
\begin{align}
S_1 + C_1 &\geq I(W_1;X_1X_2YW_2) - H(W_1)_{} + \log{p}, \nonumber \\
S_2+C_2 &\geq I(W_2;X_1X_2Y) - H(W_2)_{} + \log{p}. \label{SCrate1}
\end{align}
By doing an exact symmetric analysis, but by replacing the first encoder by a product distribution instead of the second encoder in $S_1$, we obtain the following constraints
\begin{align}
S_1 + C_1 &\geq I(W_1;X_1X_2Y) - H(W_1)_{} + \log{p}, \nonumber \\
S_2+C_2 &\geq I(W_2;X_1X_2YW_1) - H(W_2)_{} + \log{p}. \label{SCrate2}
\end{align}
By time sharing between the above rates \eqref{SCrate1} and \eqref{SCrate2}, one can obtain the following rate constraints
\begin{align}
S_1\! +\! C_1 & \!\geq\! I(W_1;X_1X_2Y) - H(W_1)_{} + \log{p}, \nonumber\\
S_2\!+\!C_2 &\!\geq \! I(W_2;X_1X_2Y) - H(W_2)_{} + \log{p}, \nonumber\\
S_1 \!+\!S_2\! +\!C_1\! +\!C_2 &\!\geq\! I(W_1W_2;X_1X_2Y)\! -\! H(W_1,W_2)\! + \!2\log{p}. \nonumber
\end{align}
\subsection{Rate Constraints}
To sum-up, we showed that the \eqref{eq:dist_main_lemma} holds for sufficiently large $n$ and with probability sufficiently close to 1, if the following bounds holds while incorporating the time sharing random variable $Q$ taking values over the finite set $\mathcal{Q}$\footnote{Since $Q$, the time sharing random variable is employed in the standard way we omit its discussion here.}:
\begin{align}\label{eq:rate-region_long}
S_1 & \geq I(X_1;W_1|Q) - H(W_1|Q) + \log{p}, \nonumber \\
S_2 & \geq I(X_2;W_2|Q) - H(W_2|Q) + \log{p}, \nonumber \\
S_1+ C_1 & \geq I(X_1X_2Y;W_1|Q) - H(W_1|Q) + \log{p}, \nonumber \\
S_2+ C_2 & \geq I(X_1X_2Y;W_2|Q) - H(W_2|Q) \!+\! \log{p}, \nonumber \\
S_1 +S_2 +C_1 +C_2 &\geq I(W_1W_2;X_1X_2Y|Q) - H(W_1,\!W_2|Q) + 2\log{p}, \nonumber \\
S_1 - R_1 &= S_2 - R_2\leq \log{p} - H(W_1\oplus W_2|Q), \nonumber \\
0 \leq R_1 &\leq S_1, \quad 0 \leq R_2 \leq S_2,\nonumber \\
C_1+C_2&\leq C, \quad C \geq 0
\end{align}
Lastly, we complete the proof of the theorem using the following lemma.
\begin{lem}
Let $\mathcal{R}_1$ denote the set of all $(R_1,R_2,C)$ for which there exists $(S_1,S_2)$ such that the septuple $(R_1,R_2, C, S_1, S_2, {C}_1, {C}_2)$ satisfies the inequalities in \eqref{eq:rate-region_long}. Let, $\mathcal{R}_2$ denote the set of all triples $(R_1, R_2, C)$ that satisfies the inequalities in \eqref{eq:rate-region} given in the statement of the theorem. Then, $
\mathcal{R}_1=\mathcal{R}_2$.
\end{lem}
\begin{proof}
This follows from Fourier-Motzkin elimination \cite{fourier-motzkin}.
\end{proof}
\section{Proof of Lemmas}
\subsection{Proof of Lemma \ref{lem:NotaPMFBound}}\label{proof_lem:NotaPMFBound}
Let $K$ denote the left hand side of \eqref{eq:NotaPMF}. Further, for the purpose of this proof, we skip the subscript $i$
Bounding the a-typical sequences of $x^n$ from the summation gives $K = K_1 + \epsilon_{X},$ where
\begin{align*}
K_1 \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \frac{1}{2^{nC}}\sum_{\substack{ \mu }}\sum_{{x}^n \in \mathcal{T}_{\delta}^{(n)}(X)}{p^{n}_{{X}}({x}^n)} \mathbb{P}\Bigg(\bigg[\sum_{a\in\mathbb{F}_p^{k}}\sum_{m \in \mathbb{F}_p^{l}}E^{(\mu)}_{L|X^n}(a,m|x^n)\bigg] > 1\Bigg),
\end{align*}
and $\epsilon_{X} \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{x^n \notin \mathcal{T}_{\delta}^{(n)}(X)}{p^{n}_{{X}}({x}^n)}$. Note that $\epsilon_{X}(\delta) \searrow 0 $ as $\delta \searrow 0.$ With that, it remains to show the $K_1$ can be made arbitrarily small in expected sense.
Toward that, define
\begin{align*}
Z^{(\mu)}_{x^n}(a,m) & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \frac{p^n}{(1+\eta)}\!\sum_{w^n }p_{XW}^n(x^n\!\!,w^n)\mathbbm{1}_{\{\mathtt{w}^n(a,m,\mu) = w^n\}}\mathbbm{1}_{\{w^n \in \mathcal{T}_{\delta}^{(n)}(W|x^n)\}},\quad \!\!\!
Z^{(\mu)}_{x^n} \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \frac{1}{2^{nS}}\!\!\sum_{m>0}\sum_{a \in \mathbb{F}_q^k}\!\!Z^{(\mu)}_{x^n}(a,m)
\end{align*}
Observe the following upper and lower bounds on $\mathbb{E}[Z^{(\mu)}_{x^n}]$.
\begin{align}
\mathbb{E}[Z^{(\mu)}_{x^n}] & = \frac{p^n}{(1+\eta)}\sum_{w^n \in \mathcal{T}_{\delta}^{(n)}(W|x^n)}p_{XW}^n(x^n,w^n)\frac{1}{p^n} \leq \frac{p_X^n(x^n)}{(1+\eta)} \label{eq:upperbound}\\
\mathbb{E}[Z^{(\mu)}_{x^n}] & = \frac{1}{(1+\eta)}\sum_{w^n \in \mathcal{T}_{\delta}^{(n)}(W|x^n)}p_{XW}^n(x^n,w^n) \geq \frac{p_X^n(x^n) 2^{n\delta_{w}}}{(1+\eta)} , \label{eq:lowerbound}
\end{align}
where the inequalities above uses the typicality arguments and $ \delta_{w}(\delta) \searrow 0$ as $\delta \searrow 0.$
Using these bounds, we perform the following simplification.
\begin{align}\label{eq:ineq1}
\mathbb{P}\Bigg(\bigg[\sum_{a\in\mathbb{F}_p^{k}}\sum_{m \in \mathbb{F}_p^{l}}E^{(\mu)}_{L|X^n}(a,m|x^n)\bigg] > 1\Bigg) & = \mathbb{P}\Bigg( Z^{(\mu)}_{x^n} > p_{X}^n(x^n) \Bigg) \leq \mathbb{P}\Bigg( Z^{(\mu)}_{x^n} > (1+\eta)\mathbb{E}[Z^{(\mu)}_{x^n} ] \Bigg)
\end{align}
where the inequality above uses the inequality from \eqref{eq:upperbound}. Further, we have
\begin{align}\label{eq:ineq2}
\mathbb{P}\Bigg( \bigg|Z^{(\mu)}_{x^n} - \mathbb{E}[Z^{(\mu)}_{x^n}] \bigg| > \eta\mathbb{E}[Z^{(\mu)}_{x^n} ] \Bigg)
\leq \frac{\mathbb{E}\bigg|Z^{(\mu)}_{x^n} - \mathbb{E}[Z^{(\mu)}_{x^n}] \bigg|}{\eta\mathbb{E}[Z^{(\mu)}_{x^n}]} \leq \frac{(1+\eta)\sqrt{Var\left(Z^{(\mu)}_{x^n}\right)}}{\eta 2^{-n\delta_w}p^n_X(x^n)}
\end{align}
where the first inequality follows from Markov Inequality, and the second uses (i) the Jensen's inequality for square-root function and (ii) the bound from \eqref{eq:lowerbound}.
Combining the inequalities \eqref{eq:ineq1} and \eqref{eq:ineq2} using union bound, we obtain
\begin{align}
\mathbb{P}\Bigg(\bigg[\sum_{a\in\mathbb{F}_p^{k}}\sum_{m \in \mathbb{F}_p^{l}}E^{(\mu)}_{L|X^n}(a,m|x^n)\bigg] > 1\Bigg) \leq \frac{\sqrt{Var\left(Z^{(\mu)}_{x^n}\right)}}{\eta 2^{-n\delta_w}p^n_X(x^n)} \leq \frac{(1+\eta)\sqrt{ 2^{-n(S + H(X|W) + H(W) -\delta'')}}}{\eta 2^{-n\delta_w}p^n_X(x^n)} \label{eq:finalexp}
\end{align}
where the last inequality follows by simplifying $Var\left(Z^{(\mu)}_{x^n}\right) $ similar to the one in \cite[Lemma 19]{cuff2009communication} to obtain
\begin{align}
Var\left(Z^{(\mu)}_{x^n}\right) \leq 2^{-n(S + H(X|W) + H(W) -\delta'')}\frac{1}{(1+\eta)^2} \nonumber.
\end{align}
Substituting the simplification of \eqref{eq:finalexp} in $K_1$ completes the proof.
\section{Proof of Propositions}
\subsection{Proof of Proposition \ref{prop:Lemma for Tilde_S}}
\label{proof_prop:Lemma for Tilde_S}
We bound $\Tilde{S}$ as $\Tilde{S} \leq \Tilde{S}_1 + \Tilde{S}_2 + \Tilde{S}_3, $ where
\begin{align}
\Tilde{S}_1 & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{\underline{x}^n,y^n}\sum_{\substack{ \mu_1,\mu_2 }}\sum_{\substack{m_2 >0}}
\frac{p^{n}_{\underline{X}}(\underline{x}^n)}{2^{n(C_1+C_2)}} p^{(\mu_1)}_{M_1|X_1^{n}}(0|x_1^{n}) p^{(\mu_2)}_{M_2|X_2^{n}}(m_2|x_2^{n})p^{(\mu)}_{Y^{n}|M_1M_2}(y^{n}|0,m_2) \nonumber \\
\Tilde{S}_2 & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{\underline{x}^n,y^n}\sum_{\substack{ \mu_1,\mu_2 }}\sum_{\substack{m_2 >0}}
\frac{p^{n}_{\underline{X}}(\underline{x}^n)}{2^{n(C_1+C_2)}}p^{(\mu_1)}_{M_1|X_1^{n}}(m_1|x_1^{n}) p^{(\mu_2)}_{M_2|X_2^{n}}(0|x_2^{n})p^{(\mu)}_{Y^{n}|M_1M_2}(y^{n}|m_1,0) \nonumber \\
\Tilde{S}_3 & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{\underline{x}^n,y^n}\sum_{\substack{ \mu_1,\mu_2 }}\sum_{\substack{m_2 >0}}
\frac{p^{n}_{\underline{X}}(\underline{x}^n)}{2^{n(C_1+C_2)}}p^{(\mu_1)}_{M_1|X_1^{n}}(0|x_1^{n}) p^{(\mu_2)}_{M_2|X_2^{n}}(0|x_2^{n})p^{(\mu)}_{Y^{n}|M_1M_2}(y^{n}|0,0) \nonumber
\end{align}
\textbf{Analysis of $\tilde{S}_{1}$: } Consider the following simplification with regards to $\tilde{S}_1.$
\begin{align}
\Tilde{S}_1
&= \sum_{\underline{x}^n}\sum_{\substack{ \mu_1,\mu_2 }}
\frac{p^{n}_{\underline{X}}(\underline{x}^n)p^{(\mu_1)}_{M_1|X_1^{n}}(0|x_1^{n})}{2^{n(C_1+C_2)}}\left(\sum_{\substack{m_2 >0}} p^{(\mu_2)}_{M_2|X_2^{n}}(m_2|x_2^{n})\left(\sum_{y^n}p^{(\mu)}_{Y^{n}|M_1M_2}(y^{n}|0,m_2)\right)\right) \nonumber \\
& = \sum_{{x}^n_1}\sum_{\substack{ \mu_1 }}
\frac{p^{n}_{{X}_1}({x}_1^n)p^{(\mu_1)}_{M_1|X_1^{n}}(0|x_1^{n})}{2^{nC_1}} \nonumber \\
&\leq \sum_{{x}_1^n \in \mathcal{T}_{\delta}^{(n)}(X_1)}\sum_{\substack{ \mu_1 }}
\frac{p^{n}_{{X}_1}({x}_1^n)p^{(\mu_1)}_{M_1|X_1^{n}}(0|x_1^{n})}{2^{nC_1}} \left(\mathbbm{1}_{\{s_1^{(\mu_1)}(x^n_1) \leq 1\}} + \mathbbm{1}_{\{s_1^{(\mu_1)}(x^n_1) > 1\}} \right) + \epsilon_{X_1} \nonumber \\
& = \Tilde{S}_{11} +\Tilde{S}_{12} + \epsilon_{X_1} \nonumber
\end{align}
where we define $\epsilon_{X_1}(\delta) \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} 1- \sum_{x^n \in \mathcal{T}_{\delta}^{(n)}(X_1)}p^{n}_{{X}_1}({x}_1^n)$ and
\begin{align}
\Tilde{S}_{11} & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{{x}_1^n \in \mathcal{T}_{\delta}^{(n)}(X_1)}\sum_{\substack{ \mu_1 }}\frac{p^{n}_{{X}_1}({x}_1^n)} {2^{nC_1}}(1-s_1^{(\mu_1)}(x^n_1))\mathbbm{1}_{\{s_1^{(\mu_1)}(x^n_1) \leq 1\}}, \nonumber \\
\Tilde{S}_{12} & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{{x}_1^n\in \mathcal{T}_{\delta}^{(n)}(X_1)}\sum_{\substack{ \mu_1 }}\frac{p^{n}_{{X}_1}({x}_1^n)} {2^{nC_1}}\mathbbm{1}_{\{s_1^{(\mu_1)}(x^n_1) > 1\}}. \nonumber
\end{align}
Now, we bound each of the above terms.
For the term corresponding to $\tilde{S}_{11}$, consider the following.
\begin{align}
\tilde{S}_{11} & \labelrel\leq{subeq:1_S11} \sum_{{x}_1^n \in \mathcal{T}_{\delta}^{(n)}(X_1)}\sum_{\substack{ \mu_1 }}\frac{p^{n}_{{X}_1}({x}_1^n)} {2^{nC_1}}\left|1-\sum_{a_1\in\mathbb{F}_p^{k}}\sum_{m_1 \in \mathbb{F}_p^{l_1}}E^{(\mu_1)}_{L_1|X_1^n}(a_1,m_1|x_1^n)\right| \nonumber \\
& \labelrel={subeq:2_S11} \sum_{{x}_1^n \in \mathcal{T}_{\delta}^{(n)}(X_1)}\sum_{\substack{ \mu_1 }}\frac{1} {2^{nC_1}}\left|p^{n}_{{X}_1}({x}_1^n)-\frac{p^n}{2^{nS_1}}\sum_{a_1\in\mathbb{F}_p^{k}}\sum_{m_1 \in \mathbb{F}_p^{l_1}}\sum_{\substack{w_1^n \in T_{\delta}(W_1|x_1^n)}}\frac{p^n_{W_1X_1}\!\!(w_1^n,x_1^n)}{(1+\eta)}\!\mathbbm{1}_{\{\mathtt{w}_1(a_1,m_1,\mu_1) = w_1^n\}}\right| \nonumber \\
& \labelrel\leq{subeq:3_S11} \sum_{{x}_1^n \in \mathcal{T}_{\delta}^{(n)}(X_1)}\sum_{\substack{ \mu_1 }}\frac{1} {2^{nC_1}}\left|p^{n}_{{X}_1}({x}_1^n)-\frac{p^n}{2^{nS_1}}\sum_{a_1\in\mathbb{F}_p^{k}}\sum_{m_1 \in \mathbb{F}_p^{l_1}}\sum_{\substack{w_1^n}}\frac{p^n_{W_1X_1}\!\!(w_1^n,x_1^n)}{(1+\eta)}\!\mathbbm{1}_{\{\mathtt{w}_1(a_1,m_1,\mu_1) = w_1^n\}}\right| \nonumber \\
& \hspace{30pt} + \label{eq:simplify_S11_tilde}
\sum_{{x}_1^n \in \mathcal{T}_{\delta}^{(n)}(X_1)}\sum_{\substack{ \mu_1 }}\frac{p^n}{2^{n(S_1+C_1)}}\sum_{a_1\in\mathbb{F}_p^{k}}\sum_{m_1 \in \mathbb{F}_p^{l_1}}\sum_{\substack{w_1^n \notin T_{\delta}(W_1|x_1^n)}}\frac{p^n_{W_1X_1}\!\!(w_1^n,x_1^n)}{(1+\eta)}\!\mathbbm{1}_{\{\mathtt{w}_1(a_1,m_1,\mu_1) = w_1^n\}}
\end{align}
where \eqref{subeq:1_S11} follows by bounding the indicator by $1$ and using the definition of $s_1^{(\mu_1)}(\cdot)$, \eqref{subeq:2_S11} uses the definition of $E^{(\mu_1)}_{L_1|X_1^n}(a_1,\!m_1|x_1^n)$ as defined in Definition \eqref{def:E_L|X}, the inequality \eqref{subeq:3_S11} follows from triangle inequality.
Taking expectation on \eqref{eq:simplify_S11_tilde} over the first encoders codebook generation, we obtain
\begin{align}
\mathbb{E}_{\mathbbm{C}_1}[\tilde{S}_{11}] & \leq \frac{1} {2^{nC_1}}\sum_{\substack{ \mu_1 }}\mathbb{E}_{\mathbbm{C}_1}\left[\sum_{{x}_1^n \in\mathcal{T}_{\delta}^{(n)}(X_1)} \left|p^{n}_{{X}_1}({x}_1^n)-\frac{p^n}{2^{nS_1}}\sum_{a_1\in\mathbb{F}_p^{k}}\sum_{m_1 \in \mathbb{F}_p^{l_1}}\sum_{\substack{w_1^n}}\frac{p^n_{W_1X_1}\!\!(w_1^n,x_1^n)}{(1+\eta)}\!\mathbbm{1}_{\{\mathtt{w}_1(a_1,m_1,\mu_1) = w_1^n\}}\right|\right] \nonumber \\
& \hspace{30pt} + \label{eq:simplify_S11_tilde1}
\sum_{{x}_1^n \in \mathcal{T}_{\delta}^{(n)}(X_1)}\sum_{\substack{ \mu_1 }}\frac{p^n}{2^{n(S_1+C_1)}}\sum_{a_1\in\mathbb{F}_p^{k}}\sum_{m_1 \in \mathbb{F}_p^{l_1}}\sum_{\substack{w_1^n \notin T_{\delta}(W_1|x_1^n)}}\frac{p^n_{W_1X_1}\!\!(w_1^n,x_1^n)}{(1+\eta)}\frac{1}{p^n}
\end{align}
For the first term in \eqref{eq:simplify_S11_tilde1}, we use Lemma \eqref{lem:changeMeasureCoveringLemma} and obtain $\mathbb{E}[\tilde{S}_{11}] \leq \epsilon_{\tilde{S}_{11}}$ if $S_1 \leq I(X_1;W_1) - H(W_1) + \log{p} + \delta_{\tilde{S}_{11}}$. As for the second term we can use typicality arguments and bound it as
\begin{align}
\sum_{{x}_1^n \in \mathcal{T}_{\delta}^{(n)}(X_1)}& \sum_{\substack{ \mu_1 }}\frac{p^n}{2^{n(S_1+C_1)}}\sum_{a_1\in\mathbb{F}_p^{k}}\sum_{m_1 \in \mathbb{F}_p^{l_1}}\sum_{\substack{w_1^n \notin T_{\delta}(W_1|x_1^n)}}\frac{p^n_{W_1X_1}\!\!(w_1^n,x_1^n)}{(1+\eta)}\frac{1}{p^n} \nonumber \\
& \leq \sum_{{x}_1^n \in \mathcal{T}_{\delta}^{(n)}(X_1)}\sum_{\substack{w_1^n \notin T_{\delta}(W_1|x_1^n)}}\frac{p^n_{W_1X_1}\!\!(w_1^n,x_1^n)}{(1+\eta)} \leq {\epsilon'}
\end{align}
where $\epsilon' \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{{x}_1^n \in \mathcal{T}_{\delta}^{(n)}(X_1)}\sum_{\substack{w_1^n \notin T_{\delta}(W_1|x_1^n)}}{p^n_{W_1X_1}\!\!(w_1^n,x_1^n)} $
Finally, we have the term corresponding to $\tilde{S}_{12}$. For this, we use Lemma \ref{lem:NotaPMFBound} and hence obtain $\mathbb{E}[\tilde{S}_{12}] \leq \epsilon_{\tilde{S}_{12}}$ if $S_2 \leq I(X_1;W_1) - H(W_1) + \log{p} + \delta_{\tilde{S}_{12}}$.
\textbf{Analysis of $\tilde{S}_2$: } Due to the symmetry in $\tilde{S}_1$ and $\tilde{S}_2, $ the analysis of $\tilde{S}_2$ follows very similar arguments at that of $\tilde{S}_1$ and therefore we obtain $\mathbb{E}[\tilde{S}_2] \leq \epsilon_{\tilde{S}_2}$ if $S_2 \geq I(X_2;W_2) - H(W_2) +\log{p} + \delta_{\tilde{S}_2}.$
\textbf{Analysis of $\tilde{S}_3: $} Follows by merging the above analysis of $\tilde{S}_1$ and $\tilde{S}_2.$
\subsection{Proof of Proposition \ref{prop:Lemma for S_2}} \label{proof_prop:Lemma for S_2}
Recalling $S_2$, we have
\begin{align}
S_2 & = \sum_{\underline{x}^n}
\sum_{\substack{ \mu_1,\mu_2 }}\sum_{\substack{ m_1 > 0, \\m_2>0 }}\sum_{\substack{a_1 \in \mathbb{F}_p^k\\a_2 \in \mathbb{F}_p^k}} \sum_{\substack{w_1^n,w_2^n \in \mathbb{F}_p^n}} \!\frac{p^{n}_{\underline{X}}(\underline{x}^n)}{2^{n(C_1+C_2)}} E^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n})E^{(\mu_2)}_{W^n_2|X_2^{n}}(w^n_2|x_2^{n}
) \nonumber \\ &\hspace{50pt} \mathbbm{1}_{\left\{ \mathtt{w}_1(a_1,m_1,\mu_1) = w_1^n\right\}}\mathbbm{1}_{\left\{ \mathtt{w}_2(a_2,m_2,\mu_2) = w_1^n\right\}}
\sum_{y^n}\Bigg|p^n_{Y|Z}(y^{n}|w_1^n\oplus w_2^n)\! -\! p^n_{Y|Z}(y^{n}|f^{(\mu)}(m_1,\!m_2))\Bigg|\nonumber \\
& \leq 2 \sum_{\underline{x}^n}
\sum_{\substack{ \mu_1,\mu_2 }}\sum_{\substack{ m_1 > 0, \\m_2>0 }}\sum_{\substack{a_1 \in \mathbb{F}_p^k\\a_2 \in \mathbb{F}_p^k}} \sum_{\substack{w_1^n,w_2^n \in \mathbb{F}_p^n}} \!\frac{p^{n}_{\underline{X}}(\underline{x}^n)}{2^{n(C_1+C_2)}} \frac{p^np^n}{2^{n(S_1+S_2)}(1+\eta)^2}{p_{W_1|X_1}^n(w_1^n|x_1^n)} {p_{W_2|X_2}^n(w_2^n|x_2^n)} \nonumber \\ &\hspace{55pt} \mathbbm{1}_{\left\{w_i^n \in T_{\delta}(W_i|x_i^n)\right\}}\mathbbm{1}_{\left\{w_i^n \in T_{\delta}(W_i|x_i^n)\right\}} \mathbbm{1}_{\left\{ \mathtt{w}_1(a_1,m_1,\mu_1) = w_1^n\right\}}\mathbbm{1}_{\left\{ \mathtt{w}_2(a_2,m_2,\mu_2) = w_1^n\right\}}\mathbbm{1}_{\{w_1^n \oplus w_2^n,m_1,m_2\}} \nonumber
\end{align}
where we define $\mathbbm{1}_{\{w_1^n \oplus w_2^n,m_1,m_2\}} $ as
\begin{align}
\mathbbm{1}_{\{w^n,m_1,m_2\}} \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \mathbbm{1} \left\{ \exists (\tilde{w}^n, \tilde{a}): \tilde{w}^n G + h_1^{(\mu_1)}(m_1) + h_2^{(\mu_2)}(m_2), \tilde{w}^n \in \mathcal{T}_{\delta}^{(n)}(W_1 \oplus W_2), \tilde{w}^n \neq w^n \right\}. \nonumber
\end{align}
Using this we obtain,
\begin{align}
\mathbb{E}[{S}_2] & \leq 2 \sum_{\underline{x}^n}
\sum_{\substack{ \mu_1,\mu_2 }}\sum_{\substack{ m_1 > 0, \\m_2>0 }}\sum_{\substack{a_1 \in \mathbb{F}_p^k\\a_2 \in \mathbb{F}_p^k}} \sum_{\substack{w_1^n,w_2^n \in \mathbb{F}_p^n}} \!\frac{p^{n}_{\underline{X}}(\underline{x}^n)}{2^{n(C_1+C_2)}} \frac{p^np^n}{2^{n(S_1+S_2)}(1+\eta)^2}{p_{W_1|X_1}^n(w_1^n|x_1^n)} {p_{W_2|X_2}^n(w_2^n|x_2^n)} \nonumber \\ &\hspace{55pt} \mathbbm{1}_{\left\{w_i^n \in T_{\delta}(W_i|x_i^n)\right\}}\mathbbm{1}_{\left\{w_i^n \in T_{\delta}(W_i|x_i^n)\right\}} \mathbbm{1}_{\left\{ \mathtt{w}_1(a_1,m_1,\mu_1) = w_1^n\right\}}\mathbbm{1}_{\left\{ \mathtt{w}_2(a_2,m_2,\mu_2) = w_1^n\right\}}\mathbbm{1}_{\{w_1^n \oplus w_2^n,m_1,m_2\}} \nonumber
\end{align}
Note that, we have
\begin{align}
\mathbb{E}&\left[\mathbbm{1}_{\{w_1^n \oplus w_2^n,m_1,m_2\}} \mathbbm{1}_{\left\{ \mathtt{w}_1(a_1,m_1,\mu_1) = w_1^n\right\}}\mathbbm{1}_{\left\{ \mathtt{w}_2(a_2,m_2,\mu_2) = w_1^n\right\}} \right] \nonumber \\
& \leq \sum_{\tilde{a} \neq a}\sum_{\substack{\tilde{w}^n \in \mathcal{T}_{\delta}^{(n)}(W_1 \oplus W_2) \\ \tilde{w}^n \neq w_1^n \oplus w^n_2}} \frac{1}{p^n}\frac{1}{p^n}\frac{1}{p^n} \leq 2^{n(H(W_1 \oplus W_2) + \delta_z)}p^{3n -k},
\end{align}
where $\delta_z(\delta) \searrow 0 $ as $\delta \searrow 0.$ This gives
\begin{align*}
\mathbb{E}[{S}_2] & \leq 2 \sum_{\underline{x}^n}
\sum_{\substack{ m_1 > 0, \\m_2>0 }}\sum_{\substack{a_1 \in \mathbb{F}_p^k\\a_2 \in \mathbb{F}_p^k}} \sum_{\substack{w_1^n,w_2^n \in \mathbb{F}_p^n}} {p^{n}_{\underline{X}}(\underline{x}^n)}\frac{2^{n(H(W_1 \oplus W_2) + \delta_z)}p^{n-k}}{2^{n(S_1+S_2)}(1+\eta)^2}{p_{W_1|X_1}^n(w_1^n|x_1^n)} {p_{W_2|X_2}^n(w_2^n|x_2^n)} \nonumber \\
& \leq 2^{n\left[(S_1 - R_1 ) - (\log{p} H(W_1 \oplus W_2) - \delta_{S_2})\right] },
\end{align*}
where $\delta_{S_2}(\delta) \searrow 0 $ as $\delta \searrow 0.$
\subsection{Proof of Proposition \ref{prop:Lemma for J1}} \label{proof_prop:Lemma for J1}
Substituting the definition of $\bar{E}^{(\mu_1)}_{W^n_1|X_1^{n}}(\cdot|\cdot)$ and $\gammaCoeff$ in $J_1$, we obtain
\begin{align*}
J_1 & = \sum_{\underline{x}^n,w_2^n,y^n}\Bigg|p^n_{\underline{X}W_2Y}(\underline{x}^n,y^n) - \frac{1}{2^{nC_1}}\sum_{\substack{ \mu_1}}\sum_{\substack{w_1^n}} \sum_{m_1 > 0}\sum_{a_1 \in \mathbb{F}_{q}^k} {p^{n}_{\underline{X}}(\underline{x}^n)}
\frac{p^n}{2^{nS_1}(1+\eta)}{p_{W_1|X_1}^n(w_1^n|x_1^n)} \nonumber \\
&\hspace{1.5in}\mathbbm{1}_{\left\{w_1^n \in T_{\delta}(W_1|x_1^n)\right\}}\mathbbm{1}_{\{\mathtt{w}_1(a_1,m_1,\mu_1) = w_1^n\}} p^n_{W_2|X_2}(w^n_2|x_2^{n})
p^n_ { Y|{Z}}(y^{n}|w_1^n+w_2^n)\Bigg| \\
& = \sum_{\underline{x}^n,w_2^n,y^n}\Bigg|p^n_{\underline{X}W_2Y}(\underline{x}^n,y^n) \nonumber \\
& \hspace{0.5in} - \frac{1}{(1+\eta)}\frac{p^n}{2^{n(S_1+C_1)}}\sum_{\substack{ \mu_1}}\sum_{\substack{w_1^n \in T_{\delta}(W_1|x_1^n)}} \sum_{m_1 > 0}\sum_{a_1 \in \mathbb{F}_{q}^k} {p^{n}_{\underline{X}\underline{W}Y}(\underline{x}^n , \underline{w}^n,y^n)} \mathbbm{1}_{\{\mathtt{w}_1(a_1,m_1,\mu_1) = w_1^n\}} \Bigg| \\
& \leq J_{11} + J_{12},
\end{align*}
where
\begin{align*}
J_{11} & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{\underline{x}^n,w_2^n,y^n} \Bigg|p^n_{\underline{X}W_2Y}(\underline{x}^n,y^n) - \frac{1}{(1+\eta)}\frac{p^n}{2^{n(S_1+C_1)}} \sum_{\substack{\mu_1}} \sum_{\substack{w_1^n}} \sum_{\substack{m_1 > 0\\a_1 \in \mathbb{F}_{q}^k}} {p^{n}_{\underline{X}\underline{W}Y}(\underline{x}^n, \underline{w}^n, y^n)} \mathbbm{1}_{\{\mathtt{w}_1(a_1,m_1,\mu_1) = w_1^n\}} \Bigg| \\
J_{12} & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{\underline{x}^n,w_2^n,y^n}\Bigg| \frac{1}{(1+\eta)}\frac{p^n}{2^{n(S_1+C_1)}}\sum_{\substack{ \mu_1}}\sum_{\substack{w_1^n \notin \mathcal{T}_{\delta}^{(n)}(W_1)}} \sum_{m_1 > 0}\sum_{a_1 \in \mathbb{F}_{q}^k} {p^{n}_{\underline{X}\underline{W}Y}(\underline{x}^n , \underline{w}^n,y^n)} \mathbbm{1}_{\{\mathtt{w}_1(a_1,m_1,\mu_1) = w_1^n\}} \Bigg|
\end{align*}
As for the term $J_{11}$, we use Lemma \ref{lem:changeMeasureCoveringLemma} and obtain the following bound on $\mathbb{E}[J_{11}]$ as, $\mathbb{E}[J_{11}] \leq \epsilon_{J_{11}}$ if $S_1 + C_1 \geq I(W_1;X_1X_2ZW_2) + \log{p} - H(W_1) + \delta_{J_{11}}.$
For the term $J_{12},$ applying expectation gives
\begin{align*}
\mathbb{E}[J_{12}] \leq \frac{1}{(1+\eta)}\sum_{\underline{x}^n,w_2^n,y^n} \sum_{w_1^n \notin \mathcal{T}_{\delta}^{(n)}(W_1)}{p^{n}_{\underline{X}\underline{W}Y}(\underline{x}^n , \underline{w}^n,y^n)} \leq \epsilon'.
\end{align*}
where $\epsilon'(\delta) \searrow 0$ as $\delta \searrow 0.$
\subsection{Proof of Proposition \ref{prop:Lemma for Q2}}
\label{proof_prop:Lemma for Q_2}
We begin by
applying triangle inequality on $Q_2$ to obtain $Q_2 \leq F_1 + F_2$, where
\begin{align*}
F_1 & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{\underline{x}^n,y^n}\Bigg|\frac{1}{2^{n(C_1+C_2)}}\sum_{\substack{ \mu_1,\mu_2 }}\sum_{\substack{w_1^n,w_2^n \in \mathbb{F}_p^n}} {p^{n}_{\underline{X}}(\underline{x}^n)}\gammaCoeff E^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n}) \nonumber \\ & \hspace{80pt}\left(p^n_{W_2|X_2}(w^n_2|x_2^{n})\! - \zetaCoeff \bar{E}^{(\mu_2)}_{W^n_2|X_2^{n}}(w^n_2|x_2^{n})\right)
p^n_ { Y|{Z}}(y^{n}|w_1^n\oplus w_2^n) \Bigg|, \\
F_2 & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{\underline{x}^n,y^n}\Bigg|\frac{1}{2^{n(C_1+C_2)}}\sum_{\substack{ \mu_1,\mu_2 }}\sum_{\substack{w_1^n,w_2^n \in \mathbb{F}_p^n}} {p^{n}_{\underline{X}}(\underline{x}^n)}\gammaCoeff E^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n}) \nonumber \\ & \hspace{70pt}\zetaCoeff\left( \bar{E}^{(\mu_2)}_{W^n_2|X_2^{n}}(w^n_2|x_2^{n}) - {E}^{(\mu_2)}_{W^n_2|X_2^{n}}(w^n_2|x_2^{n})\right)
p^n_ { Y|{Z}}(y^{n}|w_1^n\oplus w_2^n) \Bigg|,
\end{align*}
where $ \bar{E}^{(\mu_2)}_{W^n_2|X_2^{n}}(\cdot|\cdot)$ is defined as
\begin{align*}
\bar{E}^{(\mu_2)}_{W^n_2|X_2^{n}}(w^n_2|x_2^{n}) & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \frac{p^n}{2^{nS_2}(1+\eta)}{p_{W_2|X_2}^n(w_2^n|x_2^n)} \mathbbm{1}_{\left\{w_2^n \in T_{\delta}(W_2|x_2^n)\right\}}
\end{align*}
Considering the term corresponding to $F_1$, we bound it using triangle inequality applied by adding and subtracting the following terms within its modulus:
\begin{align*}
(i) \;& \sum_{w_1^n,w_2^n \in \mathbb{F}_p^n} {p^{n}_{\underline{X}}(\underline{x}^n)}p^n_{W_1|X_1}(w^n_1|x_1^{n}) p^n_{W_2|X_2}(w^n_2|x_2^{n})
p^n_ { Y|{Z}}(y^{n}|w_1^n\oplus w_2^n) \\
(ii) \; & \frac{1}{2^{nC_2}} \sum_{\mu_2} \sum_{w_1^n,w_2^n \in \mathbb{F}_p^n} {p^{n}_{\underline{X}}(\underline{x}^n)}p^n_{W_1|X_1}(w^n_1|x_1^{n}) \frac{p^n\zetaCoeff}{2^{nS_2} (1+\eta)} p^n_{W_2|X_2}(w^n_2|x_2^{n})
p^n_ { Y|{Z}}(y^{n}|w_1^n\oplus w_2^n) \\
(iii) \; & \frac{1}{2^{n(C_1+C_2)}} \sum_{\mu_1, \mu_2} \sum_{w_1^n,w_2^n \in \mathbb{F}_p^n}\!\!\!\!\!\! {p^{n}_{\underline{X}}(\underline{x}^n)}\gammaCoeff E^{(\mu_1)}_{W^n_1|X^n_1}(w^n_1|x_1^{n}) \frac{p^n\zetaCoeff}{2^{nS_2} (1+\eta)}\ p^n_{W_2|X_2}(w^n_2|x_2^{n})
p^n_ { Y|{Z}}(y^{n}|w_1^n\oplus w_2^n)
\end{align*}
This gives the following bound $F_1 \leq F_{11} + F_{12} + F_{13} + F_{14},$ where
\begin{align*}
F_{11} & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{\underline{x}^n,y^n}\Bigg|\frac{1}{2^{nC_1}}\sum_{\substack{ \mu_1 }}\sum_{\substack{w_1^n,w_2^n \in \mathbb{F}_p^n}} {p^{n}_{\underline{X}}(\underline{x}^n)}\left(\gammaCoeff E^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n}) - p^n_{W^n_1|X_1^{n}}(w^n_1|x_1^{n}) \right)p^n_ { Y|{Z}}(y^{n}|w_1^n\oplus w_2^n) \Bigg| \\
F_{12} & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{\underline{x}^n,y^n}\Bigg| \sum_{w_1^n,w_2^n \in \mathbb{F}_p^n} {p^{n}_{\underline{X}}(\underline{x}^n)}p^n_{W_1|X_1}(w^n_1|x_1^{n})\left( p^n_{W_2|X_2}(w^n_2|x_2^{n})\right. \\ & \hspace{2.7in} \left. - \frac{p^n\zetaCoeff}{2^{nS_2} (1+\eta)} p^n_{W_2|X_2}(w^n_2|x_2^{n}) \right)
p^n_ { Y|{Z}}(y^{n}|w_1^n\oplus w_2^n)\Bigg| \\
F_{13} & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{\underline{x}^n,y^n}\Bigg| \frac{1}{2^{nC_2}} \sum_{\mu_2} \sum_{w_1^n,w_2^n \in \mathbb{F}_p^n} {p^{n}_{\underline{X}}(\underline{x}^n)}\left(p^n_{W_1|X_1}(w^n_1|x_1^{n}) \right. \\ & \hspace{1in} \left. - \frac{1}{2^{nC_1}} \sum_{\mu_1} \gammaCoeff E^{(\mu_1)}_{W^n_1|X^n_1}(w^n_1|x_1^{n})\right) \frac{p^n\zetaCoeff}{2^{nS_2} (1+\eta)} p^n_{W_2|X_2}(w^n_2|x_2^{n})
p^n_ { Y|{Z}}(y^{n}|w_1^n\oplus w_2^n) \Bigg| \\
F_{14} & \mathrel{\ensurestackMath{\stackon[1pt]{=}{\scriptstyle\Delta}}} \sum_{\underline{x}^n,y^n}\Bigg|\frac{1}{2^{n(C_1+C_2)}} \sum_{\mu_1, \mu_2} \sum_{\substack{w_1^n\in \mathbb{F}_p^n}}\sum_{w_2^n \notin \mathcal{T}_{\delta}^{(n)}(W_2)}{p^{n}_{\underline{X}}(\underline{x}^n)}\gammaCoeff \bar{E}^{(\mu_1)}_{W^n_1|X^n_1}(w^n_1|x_1^{n}) \frac{p^n\zetaCoeff}{2^{nS_2} (1+\eta)}\\ & \hspace{3.5in} p^n_{W_2|X_2}(w^n_2|x_2^{n})
p^n_ { Y|{Z}}(y^{n}|w_1^n\oplus w_2^n) \Bigg| \\
\end{align*}
We start by analyzing $F_{11}.$ Note that $F_{11}$ is exactly similar to the term $Q_1$ and hence using the same rate constraints as $Q_1$, this term can be bounded. Next consider the term corresponding to $F_{12}.$ Substituting the definition of $\zetaCoeff$ gives
\begin{align*}
F_{12} & = \sum_{\underline{x}^n,y^n}\Bigg| {p^{n}_{\underline{X}Y}(\underline{x}^n,y^n)} - \frac{p^n}{2^{nS_2} (1+\eta)} \sum_{w_2^n \in \mathbb{F}_p^n} \sum_{\substack{m_2 > 0\\a_2 \in \mathbb{F}_{q}^k}} \mathbbm{1}_{\{\mathtt{w}_2(a_2,m_2,\mu_2) = w_2^n\}} p^{n}_{\underline{X}W_2Y}(\underline{x}^n,w_2^n,y^n) \Bigg|
\end{align*}
Lemma \ref{lem:changeMeasureCoveringLemma} gives us functions $\epsilon_{F_{12}}(\delta), \delta_{F_{12}}(\delta) $ such that if $$S_1 \geq I(W_2;X_1X_2Y) - H(W_2) + \log{p} + \delta_{F_{12}},$$ then $\mathbb{E}[F_{12}] \leq \epsilon_{F_{12}}$, where $\epsilon_{F_{12}}(\delta), \delta_{F_{12}}(\delta) \searrow 0 $ as $\delta \searrow 0.$
\noindent Now, we move on to considering in the term corresponding to $F_{13}.$ Taking expectation with respect to $G, h_1^{(\mu_1)}$ and $h_2^{(\mu_2)}$ gives
\begin{align*}
\mathbb{E}[F_{13}] & = \mathbb{E}_{G,h_1}\left[ \sum_{\underline{x}^n,y^n}\Bigg| \frac{1}{2^{nC_2}} \sum_{\mu_2} \sum_{w_1^n,w_2^n \in \mathbb{F}_p^n} {p^{n}_{\underline{X}}(\underline{x}^n)}\bigg(p^n_{W_1|X_1}(w^n_1|x_1^{n}) \right.
\\ & \hspace{0.3in} \left. - \frac{1}{2^{nC_1}} \sum_{\mu_1} \gammaCoeff E^{(\mu_1)}_{W^n_1|X^n_1}(w^n_1|x_1^{n})\bigg) \mathbb{E}_{h_2|G}\left[\frac{p^n\zetaCoeff}{2^{nS_2} (1+\eta)} \right] p^n_{W_2|X_2}(w^n_2|x_2^{n})
p^n_ { Y|{Z}}(y^{n}|w_1^n\oplus w_2^n)\right]\\
& = \frac{1}{(1+\eta)}\mathbb{E}_{G,h_1}\left[ \sum_{\underline{x}^n,y^n}\Bigg| \frac{1}{2^{nC_2}} \sum_{\mu_2} \sum_{w_1^n,w_2^n \in \mathbb{F}_p^n} {p^{n}_{\underline{X}}(\underline{x}^n)}\bigg(p^n_{W_1|X_1}(w^n_1|x_1^{n}) \right.
\\ & \hspace{1.5in} \left. - \frac{1}{2^{nC_1}} \sum_{\mu_1} \gammaCoeff E^{(\mu_1)}_{W^n_1|X^n_1}(w^n_1|x_1^{n})\bigg) p^n_{W_2|X_2}(w^n_2|x_2^{n})
p^n_ { Y|{Z}}(y^{n}|w_1^n\oplus w_2^n)\right]\\
& = \mathbb{E}\left[\frac{J}{(1+\eta)}\right],
\end{align*}
where the above equalities follows from the fact that $h_1^{(\mu_1)}$ and $h_1^{(\mu_1)}$ were generated independently and from using the definition of $J$ as stated earlier. Therefore, using the same analysis and rate constraints as $J$, we can bound the term $F_{13}$. Finally, we remain with the term $F_{14}$. Applying expectation on $F_{14}$ gives
\begin{align*}
\mathbb{E}\left[F_{14}\right] & \leq \mathbb{E}_{G,h_2}\left[\sum_{\underline{x}^n}\Bigg| \sum_{\mu_1, \mu_2} \!\!\!\!\!\sum_{\substack{w_1^n\in \mathbb{F}_p^n\\w_2^n \in \mathcal{T}_{\delta}^{(n)}(W_2)}}\!\!\!\! {p^{n}_{\underline{X}}(\underline{x}^n)}\frac{E_{h_1}\left[\gammaCoeff\right]}{2^{n(C_1+C_2)}} E^{(\mu_1)}_{W^n_1|X^n_1}(w^n_1|x_1^{n}) \frac{p^n\zetaCoeff}{2^{nS_2} (1+\eta)} p^n_{W_2|X_2}(w^n_2|x_2^{n})\Bigg| \right]\\
& \leq \sum_{\underline{x}^n} \frac{1}{2^{nC_2}} \sum_{\mu_2} \sum_{\substack{w_1^n\in \mathcal{T}_{\delta}^{(n)}(W_1)\\w_2^n \notin \mathcal{T}_{\delta}^{(n)}(W_2)}}\!\!\!\! {p^{n}_{\underline{X}}(\underline{x}^n)} \frac{p^n_{W_1|X_1}(w^n_1|x_1^{n})}{(1+\eta)} \mathbb{E}\left[\frac{p^n\zetaCoeff}{2^{nS_2} (1+\eta)}\right] p^n_{W_2|X_2}(w^n_2|x_2^{n}) \\
& \leq \frac{1}{(1+\eta)^2}\sum_{\underline{x}^n} \sum_{\substack{w_2^n \notin \mathcal{T}_{\delta}^{(n)}(W_2)}}\!\!\!\! {p^{n}_{\underline{X}}(\underline{x}^n)} p^n_{W_2|X_2}(w^n_2|x_2^{n}) \leq \epsilon_w',
\end{align*}
where $\epsilon_w'(\delta) \searrow 0$ as $\delta \searrow 0.$ This completes the analysis for the term corresponding to $F_1$. Finally we remain with the analysis of the term $F_{2}.$ Simplifying $F_2$ gives
\begin{align}
F_2 & \leq \frac{1}{2^{n(C_1+C_2)}} \sum_{\substack{ \mu_1,\mu_2 }} \sum_{\underline{x}^n} {p^{n}_{\underline{X}}(\underline{x}^n)}
\Bigg(\sum_{\substack{w_1^n\in \mathbb{F}_p^n}} \gammaCoeff E^{(\mu_1)}_{W^n_1|X_1^{n}}(w^n_1|x_1^{n})\Bigg) \nonumber \\ & \hspace{70pt}\Bigg|\sum_{\substack{w_2^n \in \mathbb{F}_p^n}}\zetaCoeff\left( \bar{E}^{(\mu_2)}_{W^n_2|X_2^{n}}(w^n_2|x_2^{n}) - {E}^{(\mu_2)}_{W^n_2|X_2^{n}}(w^n_2|x_2^{n})\right) \Bigg|, \nonumber \\
& = \frac{1}{2^{nC_2}} \sum_{\substack{ \mu_1,\mu_2 }} \sum_{\underline{x}^n} {p^{n}_{\underline{X}}(\underline{x}^n)}
\bigg(\sum_{m_1 > 0}p_{M_1|X^n_1}(m_1|x_1^n) \bigg) \nonumber \\ & \hspace{70pt}\Bigg|\sum_{\substack{w_2^n \in \mathbb{F}_p^n}}\zetaCoeff\left( \bar{E}^{(\mu_2)}_{W^n_2|X_2^{n}}(w^n_2|x_2^{n}) - {E}^{(\mu_2)}_{W^n_2|X_2^{n}}(w^n_2|x_2^{n})\right) \Bigg|, \nonumber \\
& \leq \frac{1}{2^{nC_2}} \sum_{\substack{ \mu_1,\mu_2 }} \sum_{\underline{x}^n} {p^{n}_{\underline{X}}(\underline{x}^n)}
\Bigg|\sum_{\substack{w_2^n \in \mathbb{F}_p^n}}\zetaCoeff\left( \bar{E}^{(\mu_2)}_{W^n_2|X_2^{n}}(w^n_2|x_2^{n}) - {E}^{(\mu_2)}_{W^n_2|X_2^{n}}(w^n_2|x_2^{n})\right) \Bigg|, \nonumber \\
= \tilde{S}_{2}, \nonumber
\end{align}
where the last inequality above follows by using $\bigg(\sum_{m_1 > 0}p_{M_1|X^n_1}(m_1|x_1^n) \bigg) \leq 1$ and the last equality follows by recalling the definition of $\tilde{S}_2$. Therefore, using the constraints obtained in the analysis of $\tilde{S}_2$, we complete the proof of the proposition.
\subsection{Achievability}
\label{SubSec:p2pAchievability}
\subsection{Converse}
\label{SubSec:SideInfoConverse}
The proof for the converse is provided in \cite{CuffPhDThesis}.
\end{proof} |
\section{\textbf{\ Introduction}}
Let's assign, once for all, a quantum physical system (described through the
Hilbert space $\mathbb{\mathcal{H}}$) prepared in the (pure) state $[\psi ]$
in $\mathbb{P(\mathcal{H})}$ and a measuring apparatus for the observable
corresponding to the self-adjoint operator $A$ on $\mathbb{\mathcal{H}}$.
Let's suppose to perform infinite times the measurement of $A$ on $[\psi ]$
giving to each value a different label $z$ (where $z$ varies in a set $Z$)
and to collect all the values obtained in a function $\widetilde{A
:Z\rightarrow \mathbb{R}$; for example the label space $Z$ could be the
Minkowski spacetime and each label $z$ be the event where/when the measured
value begins to be available.
Proceeding with the measurements, from the thickening and the thinning of
the labels in $Z$, let's suppose that a probability measure $\mu $ is
emerging informing us with the number $\mu (X)$ how probable it is to find a
label in the subset $X$ \ (with $X$ in a suitable $\sigma -$algebra
\mathcal{X}$ of subsets in $Z$).
The measure space $(Z,\mathcal{X,}\mu )$ so obtained allows us to compute
through the number $\mu \left[ \left( \widetilde{A}\right) ^{-1}(B)\right]
=\left( \widetilde{A}_{\ast }\mu \right) (B)$ the probability that the
measured values fall in a given borel subset $B$ of $\mathbb{R}$ and to
check if in our laboratory the probabilities furnished by the Quantum
Mechanics are verified: it is so if $\left( \widetilde{A}_{\ast }\mu \right)
(B)=\left\langle E_{B}^{A}\right\rangle _{\psi }$ for every borel subset $B$
of $\mathbb{R}$ (where $\left\{ E_{B}^{A}\right\} _{B}$ is the spectral
measure of the operator $A$).
If we introduce the cumulative monotone function $F:\mathbb{R\rightarrow
\left[ 0,1\right] $ defined by $F(r)=\left\langle E_{(-\infty
,r]}^{A}\right\rangle _{\psi }$ and its borel measure $\nu
_{F}(B)=\left\langle E_{B}^{A}\right\rangle _{\psi }$this condition can be
written, more synthetically, as: $\widetilde{A}_{\ast }\mu =\nu _{F}$; this
will imply also: $\left\langle A\right\rangle _{\psi }=\int_{Z}\widetilde{A
\cdot d\mu $, that is the attended value of QM for $A$ on $\psi $ is the
natural mean value of $\widetilde{A}$ with respect to $\mu $.
Given $F$ the condition $\widetilde{A}_{\ast }\mu =\nu _{F}$ is not enough
to reobtain a single values function $\widetilde{A}\mathbb{.}$ By experience
we know that a different laboratory will provide a different function still
verifying the probabilities of QM. Therefore we ask how many solutions the
following mathematical problem has: assigned the cumulative function $F$ and
the measure space $(Z,\mathcal{X,}\mu )$ find all borel functions
\widetilde{A}:Z\rightarrow \mathbb{R}$ verifying $\widetilde{A}_{\ast }\mu
=\nu _{F}$.
From a physical viewpoint to solve this problem it means to find all the
values function $\widetilde{A}$, relative to $A$ and $[\psi ]$, in all
laboratories where our scheme is appliable, starting from the function $F$
that summarizes only the attended probabilities.
If we suppose that $(Z,\mathcal{X})$ is a standard uncountable borel space
and that the probability measure $\mu $ is without atoms (that is $\mu
(\left\{ z\right\} )=0$ for every $z$ in $Z$) our problem has the following
answer: for each borel function $\alpha :Z\rightarrow ]0,1[$ (a barrier
function) such that $\alpha _{\ast }\mu =\lambda _{]0,1[}$ we get the
solution: $\widehat{A}_{\alpha }(z)=\min \left\{ r\in \mathbb{R;}\text{
F(r)\geq \alpha (z)\right\} $and every solution to the problem coincides
(essentially) with one of these functions $\widehat{A}_{\alpha }$ (that we
will call observable functions of $A$ on $[\psi ]$).
We observe that , fixed the observable $A$ and the state $[\psi ]$, the
measured value $\widehat{A}_{\alpha }(z)$ depends also on the label $z$ and
on the function $\alpha $. The variable $z$ plays the role of a hidden
random variable since its variations give raise to the attended
probabilities and (in agreement with the Kochen and Specker theorem) the
measured value depends also on the function $\alpha $ that tells us how the
operator $A$ representing the measurement apparatus and the physical system
in the state $[\psi ]$ are to be connected to give the measured value.
The value $\alpha (z)$ is not hidden, in general: in fact we will see that
the function $\alpha (z)$ is worth generally: $\ \alpha (z)=F(\widehat{A
_{\alpha }(z))$. Therefore, fixed the space of labels and performing the
various measurements, we can build the structure $(Z,\mathcal{X,}\mu )$ with
the values function $\widehat{A}_{\alpha }$ as well as\ the basic behaviour
of the function $\alpha $.
In this context it is evident why $A$ and $[\psi ]$ are not enough to
determine a univocal measured value: you don't know the value of the
function $\alpha $ on the label $z$. That is you don't know how the
measurement apparatus and the physical system are matematically correlated.
For example in the case of the famous Schr\oe dinger's cat if $\left[ \psi
\right] $ is the state of the system inside the box and $A$ is a test
telling if your cat is awake or not, the value $\widehat{A}_{\alpha }(z)$ is
either $1$ or $0$ according to $\alpha (z)>1-\left\langle A\right\rangle
_{\psi }$ or $\alpha (z)\leq 1-\left\langle A\right\rangle _{\psi }$.
Conversely if you know $\alpha (z)$ there is no room for uncertainty: your
measurement apparatus will show on your display the outcome:$\ \min \left\{
r\in \mathbb{R;}F(r)\geq \alpha (z)\right\} $ because the function $\alpha $
is exactly what you need to pass from $F$ to $\widehat{A}_{\alpha }$.
The insufficiency of $A$ and $[\psi ]$ induces us to propose a revision of
the measurement process: when you perform the measurement of the observable
A$ on a physical system prepared in the state $[\psi ]$ the outcome depends
also on the value $\alpha (z)$ taken by the ''scalar field'' $\alpha $ on
the ''hidden'' label $z$.
The condition $\alpha _{\ast }\mu =\lambda _{]0,1[}$ joined with the
hypothesis that the measure $\mu $ depends only on the state $[\psi ]$ makes
it natural to see $\alpha $ as an enrichment of the quantum state and
suggests to pass from considering the ''incomplete'' state $[\psi ]$ to
consider the ''complete'' state $([\psi ],\alpha ,z).$
Called $\mathcal{C}$ the set of all these possible ''complete states'' every
self-adjoint operator $A$ defines a function $\widehat{A}:\mathcal
C\rightarrow }\mathbb{R}$ by $\widehat{A}([\psi ],\alpha ,z)=\min \left\{
r\in \mathbb{R;}F(r)\geq \alpha (z)\right\} $, that is, on complete states
each observable \textbf{has a well defined value}. The theory we are
describing here is \textbf{realistic}: given the proper ingredients, the
measurement process can only produce an outcome that you can know a priori.
QM is logically compatible with a realistic description.
Moreover we show how it is possible to enrich each unitary map $\mathcal
U:H\rightarrow H}$ to an automorphism $\widehat{\mathcal{U}}:\mathcal
C\rightarrow C}$ in such a way to have a group homomorphism and moreover
\widehat{\mathcal{U}^{-1}A\mathcal{U}}=\widehat{A}\circ \widehat{\mathcal{U}}
$. Therefore each unitary evolution in $\mathcal{H}$ can be seen as the
apparent part of a deterministic hidden evolution in $\mathcal{C}$.
All this works for a general borel standard (uncountable) space $Z$ but if
we choose the phase space as space of labels there is a natural choice for
the measures $\mu _{\lbrack \psi ]}$ making the functions of the positions
and the functions of the momentum observable functions. Note that you need
different barriers functions for position and momentum observables.
In the end it is worth mentioning that in this context a phenomenon of
dependence of the ''state'' on the way you get the outcomes appears : there
are operators $A$ such that the operation of squaring the value $\widehat{A
([\psi ],\alpha ,z)$ is not given by $A^{2}$ on the ''state'' $([\psi
],\alpha ,z)$ but on a changed ''state'' $([\psi ],\beta ,z)$.
An observer reading the value $v=\widehat{A}(\left[ \psi \right] ,\alpha ,z)$
and then computing the value $v^{2}$ can continue to refer it to the
inaltered ''state'' $(\left[ \psi \right] ,\alpha ,z)$ but another observer
looking at the entire process as a measurement of the observable $A^{2}$ is
compelled to refer the outcome to the ''state'' $(\left[ \psi \right] ,\beta
,z)$ generally different from the previous one.
\section{\textbf{\ Symbols}}
In the real line we will consider the family $\mathcal{B}\left( \mathbb{R
\right) $ of all its borel subsets; the symbol $\chi _{B}$ will denote the
characteristic function of the subset $B$. A borel function on $\mathbb{R}$
is called here compact (bornological) if it sends bounded sets in bounded
sets: all bounded real functions and all continuous real functions are
compact.
Given a monotone, non-decreasing function $F:\mathbb{R\rightarrow \lbrack }
\mathbb{,}1\mathbb{]}$ if it is right-continuous with $\inf F=0$ and $\sup
F=1$ it is well defined its quasi-inverse (or quantile) function $F^{(-1)}
\mathbb{]}0\mathbb{,}1\mathbb{[\rightarrow }\mathbb{R}$ given by:
F^{(-1)}(s)=\min \left\{ r;\text{ }F(r)\geq s\right\} $. For every monotone,
non-decreasing, right-continuous function $F$ on $\mathbb{R}$ we will denote
by$\ \nu _{F}$ the borel measure characterized by $\nu _{F}]a,b]=F(b)-F(a)$.
We will denote by $\lambda _{(a,b)}=\nu _{id}$ the borel-lebesgue measure on
the interval $(a,b)$.
Given a self-adjoint operator $A$ on a Hilbert space $\mathbb{\mathcal{H}}$
with spectral family $\left\{ E_{(-\infty ,r]}^{A}\right\} _{r\in \mathbb{R
} $ and a unitary vector $\psi $ we will denote by $F_{\psi }^{A}$ the
monotone function defined by $F_{\psi }^{A}(r)=\left\langle E_{(-\infty
,r]}^{A}\right\rangle _{\psi }$. Note that $\nu _{F_{\psi
}^{A}}(B)=\left\langle E_{B}^{A}\right\rangle _{\psi }$ for every $B$ in
\mathcal{B}\left( \mathbb{R}\right) $, moreover $F_{\psi }^{A}$ is
non-decreasing, right-continuous with $\inf F_{\psi }^{A}=0$ and $\sup
F_{\psi }^{A}=1$, so it is well defined $\left( F_{\psi }^{A}\right) ^{(-1)}
\mathbb{]}0\mathbb{,}1\mathbb{[\rightarrow }\mathbb{R}$. Note that the
function $F_{\psi }^{A}$ is continuous in $r$ if and only if $\left\langle
E_{\left\{ r\right\} }^{A}\right\rangle _{\psi }=0$.
We will denote by $\Omega $ the family of all linear operators (essentially)
self-adjoint on $\mathbb{\mathcal{H}}$ (those associated to quantum
observables) and by $\Omega _{b}$ its subfamily of bounded operators. For
every $A$ in $\Omega $ and every borel real function $b$ the composition
b\circ A$ is still in $\Omega $, for every $A$ in $\Omega _{b}$ and every
compact borel real function $b$ the composition $b\circ A$ is still in
\Omega _{b}$.
The function $F:\Omega \times \mathbb{P(\mathcal{H})\times R\rightarrow }$
\mathbb{[}0\mathbb{,}1\mathbb{]}$ given by $F(A,[\psi ],r)=F_{\psi }^{A}(r)$
resumes all the probabilistic content of a quantum system represented by the
Hilbert space $\mathbb{\mathcal{H}}$.
Given two borel spaces $(X,\mathcal{A})$ and $(Y,\mathcal{B})$ a measure
\mu $ on $(X,\mathcal{A})$ and a borel map $\varphi :X\rightarrow Y$ the
image measure of $\mu $ by $\varphi $ is the measure $\varphi _{\ast }\mu $
on $Y$ defined by $\left( \varphi _{\ast }\mu \right) (B)=\mu (\varphi
^{-1}(B))$. Let's remember that $\left( F_{\psi }^{A}\right) _{\ast
}^{(-1)}\lambda _{\mathbb{]}0\mathbb{,}1\mathbb{[}}=\nu _{F_{\psi }^{A}}$
(cfr. \textbf{[KS]}). In general, for any operator $A$ and any borel
function $b$ we have: $\nu _{F_{\psi }^{b\circ A}}=b_{\ast }\nu _{F_{\psi
}^{A}}$.
\section{\textbf{\ Quantum mechanics on a standard space}}
We fix a separable Hilbert complex space $\mathbb{\mathcal{H}}$ of infinite
dimension, a borel standard (uncountable) space $(Z,\mathcal{X})$ and a
probability measure without atoms $\mu _{\lbrack \psi ]}$ on $Z$ for each
[\psi ]$ in the projective space $\mathbb{P(\mathcal{H})}$.
What follows works for any such structure $\mathcal{(}Z\mathcal{,X,H,\mu
_{\cdot })$.
First of all let's guarantee that our future definitions will not be empty.
\begin{theorem}
Let $(Z,\mathcal{X})$ be a borel standard (uncountable) space
\begin{enumerate}
\item There exists a borel equivalence between $(Z,\mathcal{X})$ and
(]0,1[,\mathcal{B(]}0\mathcal{,}1\mathcal{[)})$
\item Each borel equivalence $\sigma :Z\diagdown M\rightarrow
]0,1[\diagdown N$, where $M\in \mathcal{X}$ and $N\in \mathcal{B(]}0\mathcal
,}1\mathcal{[)})$ with $\lambda (N)=0$, defines a probability measure $\mu
=\left( \sigma ^{-1}\right) _{\ast }\lambda _{]0,1[\diagdown N}$ without
atoms on $Z$
\item Conversely given a probabily measure $\mu $ without atoms on $Z$
there exists a borel equivalence $\sigma :Z\rightarrow ]0,1[$ (defined out
of borel null subsets) such that $\sigma _{\ast }\mu =\lambda _{]0,1[}$.
\end{enumerate}
\end{theorem}
\begin{proof}
1. Cfr. \textbf{[KS]} Supplement to ch. 5
2. Obvious
3. Cfr \textbf{[Ry]} Thm. 9 p. 327
\end{proof}
\begin{definition}
Let's fix a self-adjoint linear operator $A$ on $\mathcal{H}$ and a quantum
state $[\psi ]$ in $\mathbb{P(\mathcal{H})}$. A \textbf{probability barrier
associated to\ }$[\psi ]$ \textbf{\ }is a borel function $\alpha
:Z\rightarrow ]0,1[$ such that $\alpha _{\ast }\mu _{\lbrack \psi ]}=\lambda
_{]0,1[}$.
\end{definition}
\begin{definition}
A \textbf{complete (pure) state} on $\mathcal{(}Z\mathcal{,X,H,\mu }_{\cdot
})$ is a triple $(\left[ \psi \right] ,\alpha ,z)$ of a pure state $\left[
\psi \right] $ in $\mathbb{P(\mathcal{H})}$, a barrier $\alpha _{\left[ \psi
\right] }$ associated to $\left[ \psi \right] $ and a point $z$ in $Z$.
\end{definition}
\begin{notation}
Let's denote by $\mathcal{C}$ the set of all complete states on $\mathcal{(}
\mathcal{,H,\mu }_{\cdot })$.
\end{notation}
\begin{definition}
The \textbf{values function} $v:\Omega \times \mathcal{C\rightarrow }\mathbb
R}$ is defined by:
\begin{equation*}
v(A;\left[ \psi \right] ,\alpha _{\lbrack \psi ]},z)=\min \left\{ r\in
\mathbb{R:}\text{ }\left\langle E_{(-\infty ,r]}^{A}\right\rangle _{\psi
}\geqq \alpha _{\lbrack \psi ]}(z)\right\} =\left( F_{\psi }^{A}\right)
^{(-1)}\left( \alpha _{\lbrack \psi ]}(z)\right)
\end{equation*}
\end{definition}
\begin{notation}
Fixed a self-adjoint linear operator $A$ on $\mathcal{H}$ its \textbf{values
function} is the function $\widehat{A}:\mathcal{C\rightarrow }\mathbb{R}$
defined by: $\widehat{A}(\left[ \psi \right] ,\alpha _{\lbrack \psi
]},z)=v(A;\left[ \psi \right] ,\alpha _{\lbrack \psi ]},z)$. Fixed also a
pure state $\left[ \psi \right] $ and a barrier $\alpha _{\lbrack \psi ]}
associated to $\left[ \psi \right] $ we will denote by $\widehat{A}_{\alpha
\left[ \psi \right] }:Z\rightarrow \mathbb{R}$ the function $\widehat{A}$
with $\left[ \psi \right] $ and $\alpha _{\lbrack \psi ]}$ fixed.
\end{notation}
\begin{theorem}
\begin{enumerate}
\item $\widehat{A}_{\alpha \left[ \psi \right] }=\left( F_{\psi
}^{A}\right) ^{(-1)}\circ \alpha _{\lbrack \psi ]}$ is a borel function
\item $\widehat{A}_{\alpha \left[ \psi \right] \ast }\mu _{\left[ \psi
\right] }=\nu _{F_{\psi }^{A}}$.
\end{enumerate}
\end{theorem}
\begin{proof}
\begin{enumerate}
\item A quasi-inverse function is monotonic.
\item $\widehat{A}_{\alpha \left[ \psi \right] \ast }\mu _{\left[ \psi
\right] }=\left( \left( F_{\psi }^{A}\right) ^{(-1)}\circ \alpha _{\lbrack
\psi ]}\right) _{\ast }\mu _{\left[ \psi \right] }=\left( F_{\psi
}^{A}\right) _{\ast }^{(-1)}\lambda _{]0,1[}=\nu _{F_{\psi }^{A}}$.
\end{enumerate}
\end{proof}
\begin{remark}
The property $\widehat{A}_{\alpha \left[ \psi \right] \ast }\mu _{\left[
\psi \right] }=\nu _{F_{\psi }^{A}}$ means that for each borel subset $B$ in
$\mathbb{R}$ the probability $\mu _{\left[ \psi \right] }\left( \left(
\widehat{A}_{\alpha \left[ \psi \right] }\right) ^{-1}(B)\right) $ that the
value of the function $\widehat{A}_{\alpha \left[ \psi \right] }$ falls in
B $ is the probability $\nu _{F_{\psi }^{A}}(B)=\left\langle
E_{B}^{A}\right\rangle _{\psi }$ assigned by the Quantum Mechanics for $A$,
\left[ \psi \right] $ and $B$.
\end{remark}
\begin{exercise}
\begin{itemize}
\item For each projector $E$ on $\mathcal{H}$ we have: $\widehat{\left(
E\right) }_{\alpha \left[ \psi \right] }(z)=\chi _{L\text{ }_{\left[ \psi
\right] }}(z)$ with $L_{\left[ \psi \right] }=\left\{ z\in Z;1-\left\langle
E\right\rangle _{\psi }<\alpha _{\left[ \psi \right] }(z)<1\right\} $
\item Let $\left\{ E_{0},...,E_{m,}...,E_{\infty }\right\} $ be a sequence
of pairwise orthogonal projectors in $\mathcal{H}$ such that
I=\sum_{k}E_{k} $\ \ and let $\lambda _{0},...,\lambda _{m},...,\lambda
_{\infty }$ be a sequence of real numbers with $\lambda _{\infty }=0$,
\left\{ \lambda _{m}\right\} _{m\geqslant 0}$ strictly increasing to $0$ and
defining the bounded operator $A=\sum_{k}\lambda _{k}\cdot E_{k}$, its
function $\widehat{A}_{\alpha \left[ \psi \right] }$ is the $\sigma $-simple
function $\widehat{A}_{\alpha \left[ \psi \right] }=\sum_{k}\lambda
_{k}\cdot \chi _{L_{k\text{ }\left[ \psi \right] }}$where:
\begin{eqnarray*}
L_{k\left[ \psi \right] } &=&\left\{ z\in Z;\sum_{j=0,...,k-1}\left\langle
E_{j}\right\rangle _{\psi }<\alpha _{\left[ \psi \right] }(z)\leq
\sum_{j=0,...,k}\left\langle E_{j}\right\rangle _{\psi }\right\} \\
L_{\infty \left[ \psi \right] } &=&\left\{ z\in
Z;\sum_{j=0,...,k}\left\langle E_{j}\right\rangle _{\psi }<\alpha _{\left[
\psi \right] }(z)\text{ for each }k<\infty \right\}
\end{eqnarray*}
Note that\ the family $\left\{ L_{k\left[ \psi \right] }\right\} $ makes a
partition of $Z$.
\end{itemize}
\end{exercise}
\begin{theorem}
When $A$ is a bounded self-adjoint operator, a function $f:Z\rightarrow
\mathbb{R}$ borel and essentially bounded for $\mu _{\left[ \psi \right] }
verifies $f_{\ast }\mu _{\left[ \psi \right] }=\nu _{F_{\psi }^{A}}$ if and
only if for each borel compact function $b$ and each unitary vector $\psi $,
we have:
\begin{equation*}
\int_{Z}\left( b\circ f\right) (z)\cdot d\mu _{\left[ \psi \right]
}=\left\langle \psi ,\left( b\circ A\right) \psi \right\rangle
\end{equation*}
\end{theorem}
\begin{proof}
$\left[ \Longrightarrow \right] $ Since $f$ is essentially bounded for $\mu
_{\left[ \psi \right] }$ and $b$ is compact the function $b\circ f$ is
essentially bounded for $\mu _{\left[ \psi \right] }$ and absolutely
integrable, moreover: $\int_{Z}\left( b\circ f\right) (z)\cdot d\mu _{\left[
\psi \right] }=\int_{\mathbb{R}}b\cdot d(f_{\ast }\mu _{\left[ \psi \right]
})=\int_{\mathbb{R}}b\cdot d(\nu _{F_{\psi }^{A}})=\left\langle b\circ
A\right\rangle _{\psi }$
$\left[ \Longleftarrow \right] $For each borel subset $B$ in $\mathbb{R}$ we
have $\left( f_{\ast }\mu _{\left[ \psi \right] }\right) (B)=\int_{\mathbb{R
}\chi _{B}\cdot d(f_{\ast }\mu _{\left[ \psi \right] })=\int_{Z}\left( \chi
_{B}\circ f\right) \cdot d\mu _{\left[ \psi \right] }=\left\langle \chi
_{B}\circ A\right\rangle _{\psi }=\nu _{F_{\psi }^{A}}(B)$
\end{proof}
\begin{notation}
Introducing the symbol $\left\langle \left\langle b\circ f\right\rangle
\right\rangle _{\psi }=\int_{Z}b\circ f\cdot d\mu _{\left[ \psi \right] }$
we can write $\left\langle \left\langle b\circ f\right\rangle \right\rangle
_{\psi }=\left\langle b\circ A\right\rangle _{\psi }$. In particular
\left\langle \left\langle f^{k}\right\rangle \right\rangle _{\psi
}=\left\langle A^{k}\right\rangle _{\psi }$ for every $k\geq 0$.
\end{notation}
\begin{remark}
The next property even though it appears quite often in synthetic data
generation it deserves to be full proved in our context.
\end{remark}
\begin{lemma}
Let $F:\mathbb{R\rightarrow \lbrack }0\mathbb{,}1\mathbb{]}$ be a monotone,
non-decreasing and right continuous function with $\inf F(\mathbb{R})=0$ and
$\sup F(\mathbb{R})=1$. Let $f:]0,1[\rightarrow \mathbb{R}$ a borel function
such that $f_{\ast }\lambda _{]0,1[}=\nu _{F}$, there exist two borel null
subsets $M$ and $N$ in $]0,1[$ and a borel function $\alpha :]0,1[\setminus
M\rightarrow ]0,1[\setminus N$ such that $\alpha _{\ast }\lambda
_{]0,1[\setminus M}=\lambda _{]]0,1[\setminus N}$ and $f\mid
_{]0,1[\setminus M}=F^{(-1)}\circ \alpha $.
\end{lemma}
\begin{proof}
The function $F$ can have a countable set $\left\{ x_{k}\right\} _{k\geq 1}$
of discontinuity points; let's write $a_{k}=F^{-}(x_{k})<F(x_{k})=b_{k}$.
Let's denote by $][a_{k},b_{k}[$ the interval $[a_{k},b_{k}[\setminus
(\left\{ a_{k}\right\} \cap F(\mathbb{R}))$.
Moreover the function $F$ can be constant on a countable family of disjoint
proper maximal intervals $\left\{ J_{n}\right\} _{n\geq 1}$. Let's write
F(J_{n})=\left\{ c_{n}\right\} $. We have $F^{-1}(F(r))=J_{n}$ if $r\in
J_{n} $ and instead $F^{-1}(F(r))=\left\{ r\right\} $ if $r\notin \bigcup
J_{n}$. If $a_{k}\in F(\mathbb{R})$ there exists $J_{n}$ such that $\left\{
a_{k}\right\} =F(J_{n})=\left\{ c_{n}\right\} $.
The restricted function $F\mid :\left( \mathbb{R\setminus }\bigcup
J_{n}\right) \rightarrow \left( ]0,1[\setminus \left[ \bigcup
][a_{k},b_{k}[\cup \left\{ c_{n}\right\} \right] \right) $ is a bijective
borel equivalence with $\left( F\mid \right) ^{-1}=F^{(-1)}\mid $(observe
that $]0,1[\setminus \bigcup ][a_{k},b_{k}[\subset F(\mathbb{R})$). \ Taken
M_{0}=f^{-1}(\bigcup J_{n}\cup \left\{ x_{k}\right\} )$ and $N_{0}=\bigcup
][a_{k},b_{k}]\cup \left\{ c_{n}\right\} $ we have $\lambda (M_{0})=\lambda
(N_{0})=0$ and the function $\alpha _{0}=F\circ f:]0,1[\setminus
M_{0}\rightarrow ]0,1[\setminus N_{0}$ is a borel function with
F^{(-1)}\circ \alpha _{0}=f$ and $\alpha _{0\ast }\lambda =\lambda $.
Since $\lambda \left[ f^{-1}(\left\{ x_{k}\right\} )\right]
=b_{k}-a_{k}=\lambda ][a_{k},b_{k}[$ for each $k\geq 1$ there exists a borel
null subset $M_{k}$ of $f^{-1}(\left\{ x_{k}\right\} )$ and a borel null
subset $N_{k}$ of $][a_{k},b_{k}[$ and a borel and measure equivalence
\alpha _{k}:f^{-1}(\left\{ x_{k}\right\} )\setminus M_{k}\rightarrow
][a_{k},b_{k}[\setminus N_{k}$.
Let's take $M=\bigcup_{n\geq 0}M_{n}$ and $N=\bigcup_{n\geq 0}N_{n}$, the
map $\alpha :]0,1[\setminus M\rightarrow ]0,1[\setminus N$ defined by
\alpha _{0}=F\circ f$ on $]0,1[\setminus f^{-1}(\bigcup J_{n}\cup \left\{
x_{k}\right\} )$ and $a_{k}$ on $f^{-1}(\left\{ x_{k}\right\} )\setminus
M_{k}$ verifies the thesis of the lemma.
\end{proof}
\begin{theorem}
A borel function $f_{_{\left[ \psi \right] }}:Z\rightarrow \mathbb{R}$
verifies $f_{_{\left[ \psi \right] }\ast }\mu _{\left[ \psi \right] }=\nu
_{F_{\psi }^{A}}$ if and only if there exists a barrier $\alpha _{\lbrack
\psi ]}$ associated to $\left[ \psi \right] $ such that $f_{_{\left[ \psi
\right] }}=\left( F_{\psi }^{A}\right) ^{(-1)}\circ \alpha _{\lbrack \psi ]}$
(out of a borel $\mu _{\left[ \psi \right] }-$null subset).
\end{theorem}
\begin{proof}
$\left[ \Longrightarrow \right] $ We already proved that $\left[ \left(
F_{\psi }^{A}\right) ^{(-1)}\circ \alpha _{\lbrack \psi ]}\right] _{\ast
}\mu _{\left[ \psi \right] }=\nu _{F_{\psi }^{A}}$.
$\left[ \Longleftarrow \right] $Since $Z$ is an uncountable borel standard
space we can find a borel $\mu _{\left[ \psi \right] }-$null subspace $M$ in
$Z$, a borel null subspace $N$ in $]0,1[$ and a borel equivalence $\sigma
:Z\diagdown M\rightarrow ]0,1[\diagdown N$ with $\sigma _{\ast }\mu _{\left[
\psi \right] }=\lambda _{]0,1[}$. Therefore the function $f\circ \sigma
^{-1}:]0,1[\diagdown N\rightarrow \mathbb{R}$ verifies: $\left( f\circ
\sigma ^{-1}\right) _{\ast }\lambda _{]0,1[}=\nu _{F_{\psi }^{A}}$.
For the previous Lemma there exists a borel function $\beta
:]0,1[\rightarrow ]0,1[$ with $\beta _{\ast }\lambda _{]0,1[}=\lambda
_{]0,1[}$ and $f\circ \sigma ^{-1}=\left( F_{\psi }^{A}\right) ^{(-1)}\circ
\beta $ out of a borel null subset, therefore $f=\left( F_{\psi }^{A}\right)
^{(-1)}\circ \alpha _{\left[ \psi \right] }^{A}$ out of a borel null subset
where $\alpha _{\left[ \psi \right] }^{A}=\beta \circ \sigma $ is a barrier
associated to $\left[ \psi \right] $.
\end{proof}
\begin{remark}
From the definition of quasi-inverse function for $f_{_{\left[ \psi \right]
}}=\left( F_{\psi }^{A}\right) ^{(-1)}\circ \alpha _{\lbrack \psi ]}$ we
have: \ \ $F_{\psi }^{A}(f_{_{\left[ \psi \right] }}(z)^{-})\leq \alpha
_{\lbrack \psi ]}(z)\leq F_{\psi }^{A}(f_{_{\left[ \psi \right] }}(z))$.
Therefore if $f_{_{\left[ \psi \right] }}(z)$ is not a discontinuity point
of $F_{\psi }^{A}$ we have: $\alpha _{\lbrack \psi ]}(z)=F_{\psi }^{A}(f_{_
\left[ \psi \right] }}(z))$.
\end{remark}
\section{\textbf{\ Representations of bounded operators by functions}}
\begin{definition}
A \textbf{complex of barriers} is a family $\alpha =$ $\left\{ \alpha
_{\lbrack \psi ]}^{A}\right\} _{A\text{ }\in \Omega \text{,}\left[ \psi
\right] \in \mathbb{P(\mathcal{H})}}$ where each $\alpha _{\lbrack \psi
]}^{A}$ is a barrier for $\left[ \psi \right] $.
\end{definition}
\begin{notation}
Fixed a complex of barriers $\alpha $ for each self-adjoint operator $A$ in
\Omega $ it is defined the function $\widehat{A}_{\alpha }:\mathbb{P
\mathcal{H})\times }Z\rightarrow \mathbb{R}$ by $\widehat{A}_{\alpha }([\psi
],z)=v(A;[\psi ],\alpha _{\lbrack \psi ]}^{A},z)=\left( \left( F_{\psi
}^{A}\right) ^{(-1)}\circ \alpha _{\lbrack \psi ]}^{A}\right) (z)$ . Two
such functions $\widehat{H}_{\alpha }$ and $\widehat{K}_{\beta }$will be
considered \textbf{equivalent} if, for each $\left[ \psi \right] $ in
\mathbb{P(\mathcal{H})}$ the functions $\widehat{H}_{\alpha \left[ \psi
\right] }$ and $\widehat{K}_{\beta \left[ \psi \right] }$ are equal out of a
borel subset $\mu _{\left[ \psi \right] }$-null.
\end{notation}
\begin{notation}
Given two self-adjoint linear operators $A_{1}$ and $A_{2}$ and a complex of
barriers $\alpha $ if $\widehat{A}_{1\alpha }\equiv \widehat{A}_{2\alpha }$
then $A_{1}=A_{2}$.
\end{notation}
\begin{proof}
For each pure state $\left[ \psi \right] $ from $\widehat{A}_{1\alpha \left[
\psi \right] }\equiv \widehat{A}_{2\alpha \left[ \psi \right] }$ it follows
\nu _{F_{\psi }^{A_{1}}}=\nu _{F_{\psi }^{A_{2}}}$ and therefore $F_{\psi
}^{A_{1}}=F_{\psi }^{A_{2}}$ since both are right-continuous functions with
\inf =0$ and $\sup =1$.
Then $\left\langle E_{(-\infty ,r]}^{A_{1}}\right\rangle _{\psi
}=\left\langle E_{(-\infty ,r]}^{A_{2}}\right\rangle _{\psi }$ for each $r$
in $\mathbb{R}$ and each unitary vector $\psi $. That is $A_{1}=A_{2}$.
\end{proof}
\begin{lemma}
For each $A$ in $\Omega _{b}$ and each $\left[ \psi \right] $ in $\mathbb{P
\mathcal{H})}$: $\mu _{\left[ \psi \right] }\left[ (\widehat{A}_{\alpha
\left[ \psi \right] })^{-1}]a,b]\right] =0$ if and only if $(\widehat{A
_{\alpha \left[ \psi \right] })^{-1}]a,b]=\varnothing $.
\end{lemma}
\begin{proof}
Since $\widehat{A}_{\alpha \left[ \psi \right] \ast }\mu _{\left[ \psi
\right] }=\nu _{F_{\psi }^{A}}$ we have $\mu _{\left[ \psi \right] }\left[
\widehat{A}_{\alpha \left[ \psi \right] })^{-1}]a,b]\right] =0$ if and only
if $F_{\psi }^{A}(a)=F_{\psi }^{A}(b)$ that is equivalent to $\left[ \left(
F_{\psi }^{A}\right) ^{(-1)}\right] ^{-1}]a,b]=\varnothing $ that implies $
\widehat{A}_{\alpha \left[ \psi \right] })]a,b]=\varnothing $.
\end{proof}
\begin{theorem}
For each $A$ in $\Omega _{b}$ we have $\overline{\widehat{A}_{\alpha }
\mathbb{P(\mathcal{H})\times }Z)}=spec(A)$.
\end{theorem}
\begin{proof}
For a number $r$ in $\mathbb{R}$ it holds $r\notin spec(A)$ if and only if
there exists $\sigma >0$ with $E_{]r-\sigma ,r+\sigma ]}^{A}=0$ (cfr.
\textbf{[W]} Thm. 7.22) that is equivalent to $\mu _{\left[ \psi \right]
\left[ \widehat{A}_{\alpha \left[ \psi \right] }^{-1}]r-\sigma ,r+\sigma
\right] =0$ and to $\widehat{A}_{\alpha \left[ \psi \right] }^{-1}]r-\sigma
,r+\sigma ]=\varnothing $ for each $\left[ \psi \right] $ in $\mathbb{P
\mathcal{H})}$ that is $r\notin \overline{\widehat{A}_{\alpha }(\mathbb{P
\mathcal{H})\times }Z)}$.
\end{proof}
\begin{definition}
A\textbf{\ representation of }$\Omega _{b}$\textbf{\ with functions of }
\mathcal{(}Z\mathcal{,X,\mu }_{\cdot })$ is a map $\widetilde{(\cdot )
:\Omega _{b}\rightarrow \mathbb{R}^{\mathbb{P(\mathcal{H})\times }Z}$ such
that for each $A$ in $\Omega _{b}$ and for each $\left[ \psi \right] $ in
\mathbb{P(\mathcal{H})}$:
\begin{enumerate}
\item the function $\widetilde{A}_{\left[ \psi \right] }:Z\rightarrow
\mathbb{R}$ is borel and essentially bounded for $\mu _{\left[ \psi \right]
} $
\item $\widetilde{A}_{\left[ \psi \right] \ast }\mu _{\left[ \psi \right]
}=\nu _{F_{\psi }^{A}}$
\end{enumerate}
\end{definition}
\begin{theorem}
Given a representation $\widetilde{(\cdot )}:\Omega _{b}\rightarrow \mathbb{
}^{\mathbb{P(\mathcal{H})\times }Z}$\textbf{\ }of\textbf{\ }$\Omega _{b}
\textbf{\ }on\textbf{\ }$\mathcal{(}Z\mathcal{,X,\mu }_{\cdot })$ there
exists a complex of barriers $\alpha $ such that $\widetilde{A}\equiv
\widehat{A}_{\alpha }$.
\end{theorem}
\begin{proof}
Given a representation $\widetilde{(\cdot )}:\Omega _{b}\rightarrow \mathbb{
}^{\mathbb{P(\mathcal{H})\times }Z}$ for each $A$ and $\left[ \psi \right] $
we have: $\widetilde{A}_{\left[ \psi \right] \ast }\mu _{\left[ \psi \right]
}=\nu _{F_{\psi }^{A}}$. This implies there exists a barrier $\alpha _{\left[
\psi \right] }^{A}$ associated to $\left[ \psi \right] $such that
\widetilde{A}_{\left[ \psi \right] }\equiv \left( F_{\psi }^{A}\right)
^{(-1)}\circ \alpha _{\lbrack \psi ]}^{A}=\widehat{A}_{\alpha \left[ \psi
\right] }$.
\end{proof}
\begin{theorem}
A map $\widetilde{(\cdot )}:\Omega _{b}\rightarrow \mathbb{R}^{\mathbb{P
\mathcal{H})\times }Z}$ with each function $\widetilde{A}_{\left[ \psi
\right] }:Z\rightarrow \mathbb{R}$ borel and $\mu _{\left[ \psi \right] }-
essentially bounded for each $\left[ \psi \right] $ in $\mathbb{P(\mathcal{H
)}$, is a representation of $\Omega _{b}$ if and only if for each bounded
self-adjoint operator $A$ on $\mathcal{H}$, for each $\left[ \psi \right] $
in $\mathbb{P(\mathcal{H})}$ and for each borel compact function $b$ we
have:
\begin{equation*}
\int_{Z}\left( b\circ \widetilde{A}_{\left[ \psi \right] }\right) (z)\cdot
d\mu _{\left[ \psi \right] }=\left\langle \psi ,\left( b\circ A\right) \psi
\right\rangle \text{ (when }\left\| \psi \right\| =1\text{)}
\end{equation*}
\end{theorem}
\begin{remark}
Given a representation $\widetilde{(\cdot )}$ of $\Omega _{b}$ for every
self-adjoint bounded operator $A$ and every borel compact function $b
\mathbb{R\rightarrow R}$ the attended value $\left\langle b\circ
A\right\rangle _{\psi }$ is equal to the mean value $\int_{Z}\left( b\circ
A_{\left[ \psi \right] }\right) (z)\cdot d\mu _{\left[ \psi \right] }$ and
also to the mean value $\int_{Z}\widetilde{(b\circ A)}_{\left[ \psi \right]
}(z)\cdot d\mu _{\left[ \psi \right] }$suggesting that it could be
\widetilde{(b\circ A)}_{\left[ \psi \right] }=b\circ \widetilde{A}_{\left[
\psi \right] }$. This is true, for example, if $b:\mathbb{R\rightarrow R}$
is a strictly increasing homeomorphism: in fact in this case we have (cfr.
\textbf{[W]}, 7.3): $F_{\psi }^{b\circ A}=F_{\psi }^{A}\circ b^{-1}$,
\left( F_{\psi }^{b\circ A}\right) ^{(-1)}=b\circ \left( F_{\psi
}^{A}\right) ^{(-1)}$ and then $\widehat{\left( b\circ A\right) }_{\alpha
}=b\circ \widehat{A}_{\alpha }$ for every self-adjoint operator $A$ and
every complex of barriers $\alpha $. However the next theorem tells us that
this is not universally true whatever representation you consider.
\end{remark}
\begin{theorem}
\begin{enumerate}
\item It does not exist a complex of barriers $\alpha $ such that $b\circ
\widehat{A}_{\alpha }=$ $\widehat{\left( b\circ A\right) }_{\alpha }$ for
every self-adjoint bounded operator $A$ and every compact borel function $b$.
\item For every self-adjoint bounded operator $A$ and every compact borel
function $b$ there exists a suitable complex $\beta $ of barriers such that
b\circ \widehat{A}_{\alpha }\equiv \widehat{\left( b\circ A\right) }_{\beta
} $
\end{enumerate}
\end{theorem}
\begin{proof}
\begin{enumerate}
\item For every projector $E$ the function $\widehat{E}_{\alpha }$ is a
characteristic function. Fixed $\left[ \psi _{0}\right] $ in $\mathbb{P
\mathcal{H})}$ with $\left\| \psi _{0}\right\| =1$ and $z_{0}$ in $Z$ it is
well defined the function $G:\left\{ \mathbb{\psi ;}\text{ }\psi \text{ is
unitary}\right\} \mathbb{\rightarrow }\left\{ 0\mathbb{,}1\right\} $ by
G(\psi )=\widehat{\left( E_{\psi }\right) }_{\alpha }(\left[ \psi _{0}\right]
,z_{0})=\chi _{\left\{ 1-\alpha _{\lbrack \psi _{0}]}(z_{0})<\left|
\left\langle \psi ,\psi _{0}\right\rangle \right| ^{2}\right\} }(\psi )$.
Let's make the extra hypothesis that such a complex of barriers $\alpha $
would esist, it will imply that for the function $G$ we have: $\sum_{_{k\geq
1}}G(\psi _{k})=1$ for every orthonormal basis $\left\{ \psi _{k}\right\}
_{k\geq 1}$. We can choose a suitable sequence $\left\{ c_{k}\right\}
_{k\geq 1}$ of distinct real numbers in such a way that the self-adjoint
operator $A=\sum_{_{k\geq 1}}c_{k}\cdot E_{\psi _{k}}$ is bounded. We have
for the extra hypothesis: $1=\widehat{I}_{\alpha }=\widehat{\left(
\sum_{_{k\geq 1}}E_{\psi _{k}}\right) }_{\alpha }=\widehat{\left[ \left(
\sum_{_{k\geq 1}}\chi _{c_{k}}\right) \circ A\right] }_{\alpha }=\left(
\sum_{_{k\geq 1}}\chi _{c_{k}}\right) \circ \widehat{A}_{\alpha }$ therefore
$\sum_{_{k\geq 1}}\widehat{\left( E_{\psi _{k}}\right) }_{\alpha
}=\sum_{_{k\geq 1}}\widehat{\left( \chi _{c_{k}}\circ A\right) }_{\alpha
}=\sum_{_{k\geq 1}}\left( \chi _{c_{k}}\circ \widehat{A}_{\alpha }\right)
=\left( \sum_{_{k\geq 1}}\chi _{c_{k}}\right) \circ \widehat{A}_{\alpha }=1
. Then $G$ is a Gleason frame function of weight 1(cfr. \textbf{[G]}) and
there exists a bounded self-adjoint operator $D$ such that $G(\psi
)=\left\langle \psi ,D\psi \right\rangle $ for every unitary $\psi $. This
implies that $G$ is continuous and constantly $0$ or constantly $1$,
contradicting, in both cases, $\sum_{_{k\geq 1}}G(\psi _{k})=1$ for any
orthonormal basis $\left\{ \psi _{k}\right\} _{k\geq 1}$.
\item For the function $\left( b\circ \widehat{A}_{\alpha }\right) _{\left[
\psi \right] }$we have $\left( b\circ \widehat{A}_{\alpha }\right) _{\left[
\psi \right] \ast }\mu _{\left[ \psi \right] }=b_{\ast }\nu _{F_{\psi
}^{A}}=\nu _{F_{\psi }^{b\circ A}}$. In fact: $\nu _{F_{\psi }^{b\circ
A}}(B)=\left\langle E_{B}^{b\circ A}\right\rangle _{\psi }=\left\langle \chi
_{B}^{{}}(b\circ A)\right\rangle _{\psi }=\left\langle \chi
_{b^{-1}(B)}^{{}}(A)\right\rangle _{\psi }=\nu _{F_{\psi }^{A}}(b^{-1}(B))$
(cfr. \textbf{[W] }Ch. 7.3).
\end{enumerate}
\end{proof}
\begin{remark}
. When $b\circ \widehat{A}_{\alpha }\equiv \widehat{\left( b\circ A\right)
_{\beta }\neq \widehat{\left( b\circ A\right) }_{\alpha }$ the composition
with the function $b$ compels us to consider a new complex of barriers
\beta $, different from the previous $\alpha $ used to define $\widehat{A
_{\alpha }$. An observer reading the value $v=\widehat{A}(\left[ \psi \right]
,\alpha _{\left[ \psi \right] }^{A},z)$ and then computing the value $b(v)$
can continue to refere it to the inaltered ''state'' $(\left[ \psi \right]
,\alpha _{\left[ \psi \right] }^{A},z)$ but another observer looking at the
entire process as a measurement of the observable $b\circ A$ seems compelled
to refere the outcome to the ''state'' $(\left[ \psi \right] ,\beta _{\left[
\psi \right] }^{A},z)$ generally different from the previous one. A possible
expanation could be that the ''state'' of a physical system can depend on
the way the outcomes emerge (cfr. \textbf{[Ro]), }however other
erxplanations are possible.
\end{remark}
\begin{example}
Let's take $Z=]0,1[$ and $\mu _{\lbrack \psi ]}=\lambda $ for every $\left[
\psi \right] $ in $\mathbb{P(\mathcal{H})}$. Let's decompose the space\
\mathcal{H}$ as $\mathcal{H}\mathbb{\mathcal{=}}\mathcal{I}\mathbb{\mathcal
\oplus }}\mathcal{J}\mathbb{\mathcal{\oplus }}\mathcal{K}$ with $\mathcal{I}
, $\mathcal{J}$ and $\mathcal{K}$ orthogonal two by two and let's denote by
E$, $F$ and $G$ the orthogonal projections, respectively, on $\mathcal{I}$,
\mathcal{J}$ and $\mathcal{K}$ . Let's take a unitary vector $\psi $ such
that $\left\langle E\right\rangle _{\psi }=1/8$, $\left\langle
F\right\rangle _{\psi }=1/4$ and $\left\langle G\right\rangle _{\psi }=5/8$
and let's consider $A=E-G$ (with $A^{2}=E+G$). It is not difficult to
compute $\left[ \left( F_{\psi }^{A}\right) ^{(-1)}\right] ^{2}(s)=\chi
_{]0,5/8[\cup ]7/8,1[}(s)$ and $\left[ \left( F_{\psi }^{A^{2}}\right)
^{(-1)}\right] (s)=\chi _{]1/4,1[}(s)$. This means $\left( \widehat{A^{2}
\right) _{\beta }=\chi _{]1/4,1[}(\beta (s))$ and $\left( \widehat{A
_{\alpha }\right) ^{2}=\chi _{]0,5/8]\cup ]7/8,1[}(\alpha (s))$ so $\left(
\widehat{A}_{\alpha }\right) ^{2}\equiv \left( \widehat{A^{2}}\right)
_{\beta }$if we take $\beta (s)\equiv \sigma (\alpha (s))$ where $\sigma
(u)=(u+3/8)$ $MOD$ $1$.
\end{example}
\section{\textbf{\ Observable functions}}
\begin{definition}
A function $f:\mathbb{P(\mathcal{H})\times }Z\rightarrow \mathbb{R}$ will be
called an \textbf{observable function} if each $f_{\left[ \psi \right]
}:Z\rightarrow \mathbb{R}$ is a borel function and there exists a
self-adjoint operator $A$ on $\mathbb{\mathcal{H}}$ such that $f_{\left[
\psi \right] \ast }\mu _{\left[ \psi \right] }=\nu _{F_{\psi }^{A}}$ for
every $\left[ \psi \right] $ in $\mathbb{P(\mathcal{H})}$.
\end{definition}
\begin{remark}
The observable functions associated to an operator $A$ emulate on each
\left[ \psi \right] $ the values taken by the concrete apparatuses measuring
the observable $A$ on $\left[ \psi \right] $.
\end{remark}
\begin{remark}
For every self-adjoint operator $A$ and every complex of barriers $\alpha $
the function $\widehat{A}_{\alpha }$ is an observable function. Conversely
for every observable function $f$ there exist a self-adjoint operator $A$
and a complex of barriers $\alpha $ such that $f_{\left[ \psi \right]
}\equiv \widehat{A}_{\alpha \left[ \psi \right] }$ for every $\left[ \psi
\right] $ in $\mathbb{P(\mathcal{H})}$.
\end{remark}
\begin{notation}
Let's denote by $\mathcal{O}$ the set of all the observable functions and by
$Op:\mathcal{O\rightarrow }\Omega $ the well defined map sending an
observable function to its associated operator; we will denote by $\mathcal{
}_{b}$ the subset of $\mathcal{O}$ of all observables whose operator is
bounded. Fixed a complex of barriers $\alpha $ we will denote by $\mathcal{O
_{\alpha }$ the family of all the functions $\widehat{A}_{\alpha }$ with $A$
a self-adjoint operator and by $\mathcal{O}_{\alpha b}$ the family of
functions $\widehat{A}_{\alpha }$ with $A$ a bounded self-adjoint operator.
We have $\mathcal{O=}\bigcup_{\alpha }\mathcal{O}_{\alpha }$ and $\mathcal{O
_{b}\mathcal{=}\bigcup_{\alpha }\mathcal{O}_{\alpha b}$.
\end{notation}
\begin{theorem}
For each observable function $f$ and each borel function $b$ the composition
$b\circ f$ is an observable function with $Op(b\circ f)=b\circ Op(f)$.
\end{theorem}
\begin{proof}
If $f$ is in $\mathcal{O}$ and $b$ is borel then each $\left( b\circ
f\right) _{\left[ \psi \right] }=b\circ f_{\left[ \psi \right] }$ is borel
and: $(b\circ f)_{\left[ \psi \right] \ast }\mu _{\left[ \psi \right]
}=b_{\ast }\nu _{F_{\psi }^{Op(f)}}=\nu _{F_{\psi }^{b\circ Op(f)}}$.
\end{proof}
\begin{theorem}
Let's consider the Hilbert space $\mathcal{H=}\bigoplus_{s=-\sigma
,...,\sigma }L^{2}(\mathcal{\mathbb{R}}_{s}^{3},\lambda )$ of a particle of
spin $\sigma $, $\ $the standard borel space $Z\mathcal{=}\bigcup_{s
\mathcal{\mathbb{R}}_{s}^{6}$ and the probability measures: $\mu _{\lbrack
\psi ]}=\sum_{\left\| \psi _{s}\right\| \neq 0}\frac{1}{\left\| \psi
_{s}\right\| ^{2}}\cdot \left| \psi _{s}\right| ^{2}\otimes \left| \mathcal{
}\psi _{s}\right| ^{2}\cdot \lambda _{\mathcal{\mathbb{R}}_{s}^{6}}$ (when
\left\| \psi \right\| =1$).
\begin{enumerate}
\item For every borel function $f$ the function $f(p_{1},p_{2},p_{3})$ is
an observable function with operator $f(P_{1},P_{2},P_{3})$
\item For every borel function $g$ the function $g(q_{1},q_{2},q_{3})$ is
an observable function with operator $g(Q_{1},Q_{2},Q_{3})$
\item The function $s$ is an observable function with operator $S$ (defined
by $S((\psi _{s})_{s})=\left( s\cdot \psi _{s}\right) _{s}$)
\end{enumerate}
\end{theorem}
\begin{proof}
\begin{enumerate}
\item Applying the Thm. 7.17 of \textbf{[W]} to the operator
g(Q_{1},Q_{2},Q_{3})$ it is not difficult to prove that $E_{(-\infty
,r]}^{g(Q)}\psi =\chi _{\left\{ g(q)\leq r\right\} }\cdot \psi $ and
F_{\psi }^{g(Q)}(r)=\sum_{\left\| \psi _{s}\right\| \neq 0}\iint_{\mathcal
\mathbb{R}}^{6}}\chi _{\left\{ g(q)\leq r\right\} }\cdot \left| \psi
_{s}\right| ^{2}\cdot dqdp$. Therefore $\left( g(q)_{\ast }\mu _{\lbrack
\psi ]}\right) (-\infty ,r]=\nu _{F_{\psi }^{g(Q)}}(-\infty ,r]$.
\item Using $f(P)=\mathcal{F}^{-1}\cdot f(Q)\cdot \mathcal{F}$ we have
F_{\psi }^{f(P)}=\sum_{\left\| \psi _{s}\right\| \neq 0}F_{\psi
_{s}}^{f(P)}=\sum_{\left\| \psi _{s}\right\| \neq 0}F_{\mathcal{F}\psi
_{s}}^{f(Q)}$. Moreover $\left[ f(p)_{\ast }\mu _{\lbrack \psi ]}\right]
(-\infty ,r]=\sum_{\left\| \psi _{s}\right\| \neq 0}\iint_{\mathcal{\mathbb{
}}^{6}}\chi _{\left\{ f(p)\leq r\right\} }\cdot \frac{\left| \psi
_{s}\right| ^{2}\cdot \otimes \left| \mathcal{F}\psi _{s}\right| ^{2}}
\left\| \psi _{s}\right\| ^{2}}dqdp=$ $=\sum_{\left\| \psi _{s}\right\| \neq
0}\int_{\mathcal{\mathbb{R}}^{3}}\chi _{\left\{ f(p)\leq r\right\} }\cdot
\left| \mathcal{F}\psi _{s}\right| ^{2}\cdot dq=\left( \sum_{\left\| \psi
_{s}\right\| \neq 0}F_{\mathcal{F}\psi _{s}}^{f(Q)}\right) (r)=\nu _{F_{\psi
}^{f(P)}}(-\infty ,r]$.
\item Applying the Thm. 7.17 of \textbf{[W]} to the operator $S$ it is not
difficult to prove that $\left( E_{(-\infty ,r]}^{S}\psi \right) (x,s)=\chi
_{(-\infty ,r]}(s)\cdot \psi _{s}(x)$ and $F_{\psi }^{S}(r)=\sum_{s\leq
r}\left\| \psi _{s}\right\| ^{2}$. Therefore $s_{\ast }\mu _{\lbrack \psi
]}=\nu _{F_{\psi }^{S}}$.
\end{enumerate}
\end{proof}
\begin{remark}
So for a suitable choice of a complex of barriers the function
g(q_{1},q_{2},q_{3})$ emulates the values of the operator
g(Q_{1},Q_{2},Q_{3})$. Then taken as $g$ the projection on the component
q_{j}$ there exists a quantum observable function giving the $j-th$
component of the position for $j=1,2,3$. If you have the skill to produce
concrete apparatuses with assigned barriers the components of the position
are outcomes for the label $(q,p)$! Analogously for the components of the
momentum (with different barriers).
\end{remark}
\begin{remark}
Therefore on the phase space each complex of barriers $\alpha $ with the map
$\widehat{(\cdot )}_{\alpha }:\Omega \rightarrow \mathcal{O}_{\alpha }$
(togeter with $\mu _{\cdot }$) defines a QMPS in the sense that
\left\langle \psi ,A\psi \right\rangle =\int_{Z}\widehat{A}_{\alpha \left[
\psi \right] }\cdot d\mu _{\left[ \psi \right] }$ (for each unitary vector
\psi $ and each self-adjoint operator $A$ such that both terms of the
equality are defined). The main differences with the usual QMPS is the
dependence of $\widehat{A}_{\alpha \left[ \psi \right] }$ on $\left[ \psi
\right] $ and the lack of any explicit hypothesis of linearity on $A$ (cfr.
\textbf{[CZ], [Gr], [M], [P]}).
\end{remark}
\section{\textbf{\ Dynamics}}
\begin{definition}
For each $f$ in $\mathcal{O}_{\alpha b}$ the \textbf{quadratic function
associated to} $f$ is the function
\begin{equation*}
Q_{f}:\mathcal{H\diagdown }\left\{ 0\right\} \mathcal{\rightarrow }\mathbb{R
\text{ defined by }\mathcal{Q}_{f}(\psi )=\left\| \psi \right\| ^{2}\cdot
\int_{Z}f_{\left[ \psi \right] }(z)\cdot d\mu _{\left[ \psi \right] }
\end{equation*}
\end{definition}
\begin{remark}
The function $\mathcal{Q}_{f}$ can be considered defined and differentiable
on all $\mathcal{H}$ (even in $0$) since it is equal to $\left\langle \psi
,Op(f)\psi \right\rangle $.
\end{remark}
\begin{theorem}
For each $f$ in $\mathcal{O}_{\alpha b}$ we have:
\begin{equation*}
Op(f)\psi =\frac{1}{2}Grad_{\psi }\mathcal{Q}_{f}
\end{equation*}
\end{theorem}
\begin{proof}
Let $A=Op(f)$ for every $\varphi $ in $\mathcal{H}$ we have: $RE\left\langle
\varphi ,\frac{1}{2}Grad_{\psi }\mathcal{Q}_{f}\right\rangle =\frac{1}{2}R
\left[ \varphi _{\psi }(\mathcal{Q}_{f})\right] =\frac{1}{2}RE\left[
\lim_{t\rightarrow 0}\frac{1}{t}\left( \left\langle \psi +t\varphi ,A\psi
+tA\varphi \right\rangle -\left\langle \psi ,A\psi \right\rangle \right)
\right] =$ $RE\left\langle \varphi ,A\psi \right\rangle $ therefore $A\psi
\frac{1}{2}Grad_{\psi }\mathcal{Q}_{f}$.
\end{proof}
\begin{definition}
Given $f$ and $g$ in $\mathcal{O}_{\alpha b}$ we define $\left\{ f,g\right\}
_{\alpha }=\widehat{\left( \left\{ Op(f),Op(g)\right\} \right) }_{\alpha }$
(where $\left\{ A,B\right\} =-\frac{i}{2}\left[ A,B\right] $) and $f\circ
_{\alpha }g=\widehat{\left( Op(f)\circ Op(g)\right) }_{\alpha }$ (where
A\circ B=\frac{1}{2}\left( AB+BA\right) $).
\end{definition}
\begin{remark}
Note that here: $A\cdot B=A\circ B+i\left\{ A,B\right\} $. The space
\mathcal{O}_{\alpha b}$ with $\left\{ \cdot ,\cdot \right\} _{\alpha }$ is a
Lie-algebra and $\mathcal{O}_{\alpha b}$ with $\circ _{\alpha }$ is a Jordan
algebra. In the complexification of $\mathcal{O}_{\alpha b}$ we can consider
the product $f\times _{\alpha }g=\widehat{\left( Op(f)\cdot Op(g)\right)
_{\alpha }=f\circ _{\alpha }g+i\cdot \left\{ f,g\right\} _{\alpha }$ that
makes it an associative algebra. We have: $\left\langle \left\langle \left\{
f,g\right\} _{\alpha }\right\rangle \right\rangle _{\psi }=\left\langle
\left\{ Op(f),Op(g)\right\} \right\rangle _{\psi }$, $\left\langle
\left\langle f\circ _{\alpha }g\right\rangle \right\rangle _{\psi
}=\left\langle Op(f)\circ Op(g)\right\rangle _{\psi }$ and $\left\langle
f\times _{\alpha }g\right\rangle _{\psi }=\left\langle Op(f)\cdot
Op(g)\right\rangle _{\psi }$. Moreover: $Q_{f\times _{\alpha }g}(\psi )
\frac{1}{4}\left\langle Grad_{\psi }Q_{f},Grad_{\psi }Q_{g}\right\rangle $,
Q_{f\circ _{\alpha }g}(\psi )=$ $RE\left[ Q_{f\times _{\alpha }g}(\psi
\right] $\ and $Q_{\left\{ f,g\right\} _{\alpha }}(\psi )=IM\left[
Q_{f\times _{\alpha }g}(\psi )\right] $.
\end{remark}
\begin{theorem}
Let $H$ be a bounded self-adjoint operator on $\mathcal{H}$ , let $\alpha
_{\cdot \text{ }}$be a complex of barriers, let $h=\widehat{H}_{\alpha }$
and let $\left\{ \psi _{t}\right\} _{t\in \mathbb{R}}$ be a differentiable
unitary path in $\mathcal{H}$.
Are equivalent:
\begin{enumerate}
\item $i\frac{d}{dt}|_{t=t_{0}}\psi _{t}=H\psi _{t_{0}}$
\item $\frac{d}{dt}|_{t=t_{0}}\int_{Z}f_{\left[ \psi _{t}\right] }\cdot
d\mu _{\left[ \psi _{t}\right] }=2\cdot \int_{Z}\left\{ f,h\right\} _{\alpha
\left[ \psi _{t_{0}}\right] }\cdot d\mu _{_{\left[ \psi _{t_{0}}\right] }}$
for every $f$ in $\mathcal{O}_{\alpha b}$
\end{enumerate}
\end{theorem}
\begin{proof}
$\left[ 1)\Longrightarrow 2)\right] $ $\frac{d}{dt}|_{t=t_{0}}\int_{Z}f_
\left[ \psi _{t}\right] }\cdot d\mu _{\left[ \psi _{t}\right] }=\frac{d}{dt
|_{t=t_{0}}\left\langle \psi _{t},Op(f)\psi _{t}\right\rangle =\left\langle
\psi _{t_{0}},-i\left[ Op(f),H\right] \psi _{t_{0}}\right\rangle =$
=\int_{Z}\widehat{\left( -i\left[ Op(f),H\right] \right) }_{\alpha _{\left[
\psi _{t_{0}}\right] }}\cdot d\mu _{_{\left[ \psi _{t_{0}}\right]
}}=2\int_{Z}\left\{ f.h\right\} _{\alpha \left[ \psi _{t_{0}}\right] }\cdot
d\mu _{_{\left[ \psi _{t_{0}}\right] }}$.
$\left[ 2)\Longrightarrow 1)\right] $ Let's fix $t_{0}$ in $\mathbb{R}$ and
a bounded self-adjoint operator $A$ such that $A\psi _{t_{0}}=\frac{d}{dt
|_{t=t_{0}}\psi _{t}+iH\psi _{t_{0}}$. Then taken $f=\widehat{A}_{\alpha }$:
$2\left\| \frac{d}{dt}|_{t=t_{0}}\psi _{t}+iH\psi _{t_{0}}\right\| ^{2}
\frac{d}{dt}|_{t=t_{0}}\left\langle \psi _{t},A\psi _{t}\right\rangle
-\left\langle \psi _{t_{0}},-i\left[ A,H\right] \psi _{t_{0}}\right\rangle =$
$=\frac{d}{dt}|_{t=t_{0}}\int_{Z}f_{\left[ \psi _{t}\right] }\cdot d\mu _
\left[ \psi _{t}\right] }-2\int_{Z}\left\{ f,h\right\} _{\alpha \left[ \psi
_{t_{0}}\right] }\cdot d\mu _{_{\left[ \psi _{t_{0}}\right] }}=0$.
\end{proof}
\begin{definition}
For every observable function $f$ and every $\left[ \psi \right] $ in
\mathbb{P(\mathcal{H})}$ the\textbf{\ dispersion of }$f$\textbf{\ on }$\left[
\psi \right] $ is the non-negative number $\Delta _{\psi }^{f}=\sqrt
\left\langle \left\langle \left[ f-\left\langle f\right\rangle _{\psi
\right] ^{2}\right\rangle \right\rangle _{\psi }}=\sqrt{\left\langle
\left\langle f^{2}\right\rangle \right\rangle _{\psi }-\left\langle
\left\langle f^{{}}\right\rangle \right\rangle _{\psi }^{2}}$.
\end{definition}
\begin{remark}
$\Delta _{\psi }^{f}=\Delta _{\psi }^{Op(f)}$ in fact: $\left( \Delta _{\psi
}^{f}\right) ^{2}=\left\langle \left\langle \left[ f-\left\langle
f\right\rangle _{\psi }\cdot 1\right] ^{2}\right\rangle \right\rangle _{\psi
}=\left\langle \left[ Op(f)-\left\langle f\right\rangle _{\psi }\cdot
\right] ^{2}\right\rangle _{\psi }=\left( \Delta _{\psi }^{Op(f)}\right)
^{2} $
\end{remark}
\begin{theorem}
(Heisenberg's uncertainty principle) For every $f$ and $g$ in $\mathcal{O
_{\alpha b}$ we have $\Delta _{\psi }^{f}\cdot \Delta _{\psi }^{g}\geq
\left| \left\langle \left\langle \left\{ f,g\right\} _{\alpha }\right\rangle
\right\rangle _{\psi }\right| $
\end{theorem}
\begin{proof}
$\Delta _{\psi }^{f}\cdot \Delta _{\psi }^{g}=\Delta _{\psi }^{Op(f)}\cdot
\Delta _{\psi }^{Op(g)}\geq \left| \left\langle \left\{ Op(f),Op(g)\right\}
_{\alpha }\right\rangle _{\psi }\right| =\left| \left\langle \left\langle
\left\{ f,g\right\} _{\alpha }\right\rangle \right\rangle _{\psi }\right| $
\end{proof}
\begin{definition}
A \textbf{complex of equivalences} is a family $\sigma =\left\{ \sigma
_{\lbrack \psi ]}\right\} _{[\psi ]\in \mathbb{P(\mathcal{H})}}$ of borel
equivalences $\sigma _{\lbrack \psi ]}:Z\rightarrow ]0,1[$ (defined out of
null borel subsets) with $\sigma _{\lbrack \psi ]\ast }\mu _{\lbrack \psi
]}=\lambda _{]0,1[}$ for each $[\psi ]$ in $\mathbb{P(\mathcal{H})}$.
\end{definition}
\begin{definition}
A unitary operator $\mathcal{U}$ together with a complex of equivalences
\sigma $ defines on the space of complete states an \textbf{automorphism }
\widehat{\mathcal{U}}_{\sigma }:\mathcal{C\rightarrow C}$ by:
\begin{equation*}
\widehat{\mathcal{U}}_{\sigma }([\psi ],\alpha _{\lbrack \psi ]},z)=\left(
\mathcal{U}\psi ],\left( \alpha _{\lbrack \psi ]}\circ \sigma _{\lbrack \psi
]}^{-1}\circ \sigma _{\lbrack \mathcal{U}\psi ]}\right) _{[\mathcal{U}\psi
]},\left( \sigma _{\lbrack \mathcal{U}\psi ]}^{-1}\circ \sigma _{\lbrack
\psi ]}\right) (z)\right)
\end{equation*}
\end{definition}
\begin{remark}
Given $\mathcal{U}$ and $\sigma $ the automorphism $\widehat{\mathcal{U}
_{\sigma }$ can be given equivalently by: $\widehat{\mathcal{U}}_{\sigma
}([\psi ],\alpha _{\lbrack \psi ]},z)=\left( [\mathcal{U}\psi ],\left(
\alpha _{\lbrack \psi ]}\circ \tau _{\lbrack \psi ]}^{-1}\right) _{[\mathcal
U}\psi ]},\tau _{\lbrack \psi ]}(z)\right) $ where $\tau _{\lbrack \psi
]}=\sigma _{\lbrack \mathcal{U}\psi ]}^{-1}\circ \sigma _{\lbrack \psi
]}:Z_{[\psi ]}\rightarrow Z_{[\mathcal{U}\psi ]}$ is a borel equivalence
(defined out of $\mu -$null borel subsets) with $\tau _{\lbrack \psi ]\ast
}\mu _{\lbrack \psi ]}=\mu _{\lbrack U\psi ]}$.
\end{remark}
\begin{remark}
Written $\widehat{\mathcal{U}}_{\sigma }([\psi ],\alpha ,z)=\left( [\mathcal
U}\psi ],\alpha ^{\prime },z^{\prime })\right) $ note that $\alpha ^{\prime
}(z^{\prime })=\alpha (z)$.
\end{remark}
\begin{remark}
Each automorphism $\widehat{\mathcal{U}}_{\sigma }$ induces the projective
transformation $\ \left[ \mathcal{U}\right] $ $\mathcal{:}\mathbb{P(\mathcal
H})\rightarrow P(\mathcal{H})}\mathcal{\ \ }$ defined by $\left[ \mathcal{U
\right] [\psi ]=$ $[\mathcal{U}\psi ]$. Fixed a complex of equivalences
\sigma $ we have $\widehat{\mathcal{U}}_{\sigma }=\widehat{\mathcal{V}
_{\sigma }$ if and only if $\left[ \mathcal{U}\right] =\left[ \mathcal{V
\right] $. The automorphisms with a fived $\sigma $ make a group isomorphic
to the group of projective transformations of $\mathbb{P(\mathcal{H})}$.
\end{remark}
\begin{theorem}
Given a 1-parameter group $\left\{ \mathcal{U}_{t}\right\} _{t\in \mathbb{R
} $ of unitary transformations of $\mathcal{H}$ and a complex of
equivalences $\sigma $, the family $\left\{ \widehat{\mathcal{U}}_{t\sigma
}\right\} _{t\in \mathbb{R}}$ is a 1-parameter group of automorphisms of
\mathcal{C}$ inducing the 1-parameter group $\left\{ \left[ \mathcal{U}_{t
\right] \right\} _{t\in \mathbb{R}}$ of projective transformations of
\mathbb{P(\mathcal{H})}$.
\end{theorem}
\begin{proof}
Obvious.
\begin{remark}
A 1-parameter group of automorphisms of $\mathcal{C}$ gives a deterministic
evolution for every initial complete state $([\psi _{0}],\alpha _{0[\psi
_{0}]},z_{0})$. Every evolution $\left\{ \mathcal{U}_{t}\psi _{0}\right\}
_{t\in \mathbb{R}}$ in $\mathcal{H}$ of a unitary vector $\psi _{0}$ with
\left\{ \mathcal{U}_{t}\right\} _{t\in \mathbb{R}}$ 1-parameter group of
unitary transformations, can be seen, in several ways, as the apparent part
of an evolution of complete states.
\begin{theorem}
For every self-adjoint operator $A$, every complex of equivalences $\sigma $
and every unitary operator $\mathcal{U}$ we have:
\begin{equation*}
\widehat{\mathcal{U}^{-1}A\mathcal{U}}=\widehat{A}\circ \widehat{\mathcal{U}
_{\sigma }
\end{equation*}
\end{theorem}
\end{remark}
\end{proof}
\begin{proof}
$\widehat{\left( \mathcal{U}^{-1}A\mathcal{U}\right) }([\psi ],\alpha
_{\lbrack \psi ]},z)=\left( F_{\mathcal{U}\psi }^{A}\right) ^{(-1)}(\alpha
_{\lbrack \psi ]}(z))$ and $\left( \widehat{A}\circ \widehat{\mathcal{U}
_{\sigma }\right) ([\psi ],\alpha _{\lbrack \psi ]},z)=$
$=\widehat{A}([\mathcal{U}\psi ],\left( \alpha _{\lbrack \psi ]}\circ \sigma
_{\lbrack \psi ]}^{-1}\circ \sigma _{\lbrack \mathcal{U}\psi ]}\right) _{
\mathcal{U}\psi ]},\left( \sigma _{\lbrack \mathcal{U}\psi ]}^{-1}\circ
\sigma _{\lbrack \psi ]}\right) (z))=$
$=\left( F_{\mathcal{U}\psi }^{A}\right) ^{(-1)}\left( \alpha _{\lbrack \psi
]}\circ \sigma _{\lbrack \psi ]}^{-1}\circ \sigma _{\lbrack \mathcal{U}\psi
]}\circ \sigma _{\lbrack \mathcal{U}\psi ]}^{-1}\circ \sigma _{\lbrack \psi
]}\right) =\left( F_{\mathcal{U}\psi }^{A}\right) ^{(-1)}(\alpha _{\lbrack
\psi ]}(z))$.
\end{proof}
\begin{theorem}
Given a bounded self-adjoint operator $A$, a 1-parameter group $\left\{
\mathcal{U}_{t}\right\} _{t\in \mathbb{R}}$ of unitary transformations of
\mathcal{H}$, a unitary vector $\psi _{0\text{ }}$, a complex of barriers
\alpha _{0}$ \ we have:
\begin{equation*}
\left\langle \psi _{t},A\psi _{t}\right\rangle =\int_{Z}\widehat{A}_{\alpha
_{0}}([\psi _{t}],z)\cdot d\mu _{_{\lbrack \psi _{0}]}}(z)
\end{equation*}
where $\psi _{t}=\mathcal{U}_{t}\psi _{0}$.
\end{theorem}
\begin{proof}
Fixed a complex of equivalences $\sigma $ we have: $\left\langle \psi
_{t},A\psi _{t}\right\rangle =\left\langle \psi _{0},\left( \mathcal{U
_{t}^{-1}A\mathcal{U}_{t}\right) \psi _{0}\right\rangle =\int_{Z}\widehat
\left( \mathcal{U}_{t}^{-1}A\mathcal{U}_{t}\right) }_{\alpha _{0}[\psi
_{0}]}(z)\cdot d\mu _{_{\lbrack \psi _{0}]}}=$ $=\int_{Z}\widehat{\left(
A\right) }(\widehat{\mathcal{U}}_{t\sigma })([\psi _{0}],\alpha _{0[\psi
_{0}]},z)\cdot d\mu _{_{\lbrack \psi _{0}]}}(z)=$ $=\int_{Z}\left( F_{\psi
_{t}}^{A}\right) ^{(-1)}\left( \alpha _{0[\psi _{0}]}\circ \sigma _{\lbrack
\psi _{0}]}^{-1}\circ \sigma _{\lbrack \psi _{0}]}\circ \sigma _{\lbrack
\psi _{t}]}^{-1}\circ \sigma _{\lbrack \psi _{t}]}(z)\right) \cdot d\mu
_{_{\lbrack \psi _{0}]}}(z)=$ $=\int_{Z}\widehat{\left( A\right) }_{\alpha
_{0}[\psi _{t}]}(z)\cdot d\mu _{_{\lbrack \psi _{0}]}}(z)$.
\end{proof}
\begin{remark}
The attended value of $A$ on $\psi _{t}$ (evolution of $\psi _{0\text{ }}$
at the time $t$) is equal to the mean value of $\widehat{A}_{\alpha _{0}}$
on the points $([\psi _{t}],z)$ (''evolutions'' of the points $([\psi
_{0}],z)$ at the time $t$).
\end{remark}
\section{\textbf{BIBLIOGRAPHY}}
\
\textbf{[CZ]\ }\ T.L. Curtright and C.K. Zachos: Quantum mechanics in phase
space.World Scientific - Singapore (2005)
\textbf{[G] }\ \ \ \ A. M. \ Gleason: Measures on the closed subspaces of a
Hilbert space. Jour. of Math. and Mech. Vol. 6 (1957) \ 885-89
\textbf{[Gr]} \ \ H.J. Groenewold: On the principles of elementary quantum
mechanics. Physica 12 (1946) 405-460
\textbf{[KS]} \ \ J.L. Kelley and T.P. Srinivasan: Measure and integral. \
Springer-Verlag - NY 1988
\textbf{[M] }\ \ \ J.E.Moyal: Quantum mechanics as a statistical theory. \
Proc. Cambridge Phil. Soc. 45 (1949) 99-124
\textbf{[P]} \ \ \ \ J. C. T. Pool: Mathematical aspects of the Weyl
correspondence. Jour. of Math. Phys. Vol. 7 N. 1 Jan. 1966
\textbf{[Ro] \ }C. Rovelli: Relational Quantum Mechanics. Int. Jour. of
Teor. Phys., 35 (1996) 1637-1678
\textbf{[Ry]} \ \ \ H. L. Royden: Real analysis - The Macmillan company -
London 1968
\textbf{[W]} \ \ J. Weidmann: Linear Operators in Hilbert Spaces.
Springer-Verlag, NY 1980
\end{document}
|
\section{Introduction}
\documentclass[11pt,a4paper,leqno]{article}
\usepackage{amsmath}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsthm}
\usepackage{hyperref}
\usepackage[margin=2.5cm]{geometry}
\newtheorem{theorem}{Theorem}[section]
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{lemma}[theorem]{Lemma}
\numberwithin{equation}{section}
\title{Convergence of linear combinations of iterates of an inner function}
\author{Artur Nicolau\thanks{Supported in part by the Generalitat de Catalunya (grant 2017 SGR 395) and the Spanish Ministerio de Ciencia e Innovaci\'on (project MTM2017-85666-P)}\\
{\small
\begin{tabular}{@{}c}
Departament de Matem\`atiques\\
Universitat Aut\`onoma de Barcelona\\
and Centre de Recerca Matem\`atica\\
08193 Barcelona\\
{\tt <EMAIL>}
\end{tabular}}}
\date{}
\begin{document}
\maketitle
\begin{abstract}
\noindent Let $f$ be an inner function with $f(0)=0$ which is not a rotation and let $f^{n}$ be its $n$-th iterate. Let $\{a_{n}\}$ be a sequence of complex numbers. We prove that the series $\sum a_{n}f^{n}(\xi)$ converges at almost every point~$\xi$ of the unit circle if and only if $\sum |a_n|^2 < \infty$. The main step in the proof is to show that under this assumption, the function $F= \sum a_n f^n$ has bounded mean oscillation. We also prove that $F$ is bounded on the unit disc if and only if $\sum |a_n| < \infty$. Finally we describe the sequences of coefficients $\{a_n \}$ such that $F$ belongs to other classical function spaces, as the disc algebra and the Dirichlet class.
\end{abstract}
\section{Introduction}
Inner functions are analytic mappings from the unit disc~$\mathbb{D}$ of the complex plane into itself whose radial limits are of modulus one at almost every point of the unit circle~$\partial\mathbb{D}$. Inner functions are a central notion in Complex and Functional Analysis. See for instance \cite{Ga}. Any inner function~$f$ induces a map from~$\partial\mathbb{D}$ into itself defined at almost every point~$\xi\in \partial\mathbb{D}$ by $f(\xi)=\lim\limits_{r\to 1}f(r\xi)$. Let $f^{n}$ denote the $n$-th iterate of~$f$, which is again an inner function. The main purpose of this paper is to study the convergence of linear combinations of iterates of an inner function and to show that in a certain sense, the iterates~$f^{n}\colon \partial\mathbb{D}\to \partial\mathbb{D}$ behave as independent random variables. This phenomena has been extensively studied in the context of random series of functions (\cite{Ka}) but it also occurs in other settings where independence is not present, as in the theory of lacunary series (see Section 6 of Chapter V and Section 5 of Chapter XVI of \cite{Zy}).
We start discussing pointwise convergence. Let $\{X_{n}\}$ be a sequence of independent random variables with mean~$0$ and finite variances~$V(X_{n})$. The classical Khintchine--Kolmogorov Convergence Theorem asserts that $\sum X_{n}$ converges almost surely if and only if $\sum V(X_{n})<\infty$. In our context we have the following analogous result.
\begin{theorem}\label{theo1.1}
Let $f$ be an inner function with $f(0)=0$ which is not a rotation and let $\{a_{n}\}$ be a sequence of complex numbers. Then the following conditions are equivalent:
\begin{enumerate}
\item[(a)] The series $\sum\limits_{n=1}^{\infty} a_{n}f^{n}(\xi)$ converges at almost every point~$\xi\in \partial\mathbb{D}$.
\item[(b)] The set $\Bigl\{\xi\in \partial\mathbb{D}:\sup\limits_{N}\Bigl|\sum\limits_{n=1}^{N} a_{n}f^{n}(\xi)\Bigr|<\infty\Bigr\}$ has positive Lebesgue measure.
\item[(c)] The complex numbers $\{a_{n}\}$ satisfy $\sum\limits_{n=1}^{\infty} |a_{n}|^{2}<\infty$.
\end{enumerate}
\end{theorem}
Let $m$ denote normalized Lebesgue measure in $\partial\mathbb{D}$. For $0<p<\infty$ let $\mathbb{H}^{p}(\mathbb{D})$ be the classical Hardy space of analytic functions~$F$ in~$\mathbb{D}$ such that
$$
\|F\|_{p}^{p}=\sup_{r<1}\int_{\partial\mathbb{D}}|F(r\xi)|^{p}\,dm(\xi)<\infty.
$$
Let $\mathbb{H}^{\infty}(\mathbb{D})$ be the algebra of bounded analytic functions~$F$ in~$\mathbb{D}$ and $\|F\|_{\infty}=\sup \{|F(z)|: z\in \mathbb{D}\}$. Let~$\operatorname{BMOA}$ be the space of analytic functions~$F$ in~$\mathbb{D}$ such that
$$
\|F\|^{2}_{\operatorname{BMOA}(\mathbb{D})}=\sup_{z\in \mathbb{D}} \int_{\partial\mathbb{D}} |F(\xi)-F(z)|^{2} P(z,\xi)\,dm(\xi)<\infty.
$$
Here $P(z,\xi)=(1-|z|^{2})|\xi-z|^{-2}$ is the Poisson kernel. The subspace~$\operatorname{VMOA}$ is formed by the functions~$F\in\operatorname{BMOA}$ such that the integral above tends to $0$ as $|z|$ tends to~$1$.
Any function~$F\in \mathbb{H}^{p}(\mathbb{D})$, $0 < p\le\infty$, has non-tangential limit, denoted by~$F(\xi)$, at almost every point~$\xi\in \partial\mathbb{D}$. Moreover if $p\ge 1$, then
$$
F(z)=\int_{\partial\mathbb{D}}F(\xi)P(z,\xi)\,dm(\xi),\quad z\in \mathbb{D}.
$$
We have $\mathbb{H}^{\infty}(\mathbb{D})\subseteq \operatorname{BMOA}(\mathbb{D})\subseteq \mathbb{H}^{p}(\mathbb{D})$ for any $0<p<\infty$. Actually $\operatorname{BMOA}(\mathbb{D})$ is the natural substitute of~$\mathbb{H}^{\infty}(\mathbb{D})$ in several results of the theory of Hardy spaces. See \cite{Ga} for all these well known results.
The proof of Theorem~\ref{theo1.1} is based in the following $\operatorname{BMOA}$ type estimate.
\begin{theorem}\label{theo1.2}
Let $f$ be an inner function with $f(0)=0$ which is not a rotation. Then there exists a constant $C=C(f)>0$ such that for any positive integer~$N$, any set $\{a_{n}:n=1,\dots,N\}$ of complex numbers and any point $z\in\mathbb{D}$, we have
$$
\int_{\partial\mathbb{D}}\left| \sum_{n=1}^{N} a_{n}(f^{n}(\xi)-f^{n}(z))\right|^{2} P(z,\xi)\,dm(\xi)\le C\sum^{N}_{n=1}|a_{n}|^{2} (1-|f^{n}(z)|^{2}).
$$
\end{theorem}
Let $f$ be an analytic selfmapping of $\mathbb{D}$ with $f(0)=0$ which is not a rotation. The classical Denjoy--Wolff Theorem asserts that $f^{n}$ tends to~$0$ uniformly on compact sets of~$\mathbb{D}$. See Chapter V of \cite{Sh}. The proof of Theorem~\ref{theo1.2} uses the interplay between those dynamical properties of the inner function~$f$ as a self mapping of $\partial\mathbb{D}$ and those as a self mapping of~$\mathbb{D}$. In particular we use a result due to Pommerenke (\cite{Po}) which provides an exponential decay in the Denjoy--Wolff Theorem.
It was proved in~\cite{NS2} that $\sum a_{n}f^{n}$ converges in $\mathbb{H}^{2}(\mathbb{D})$ if and only if $\sum |a_{n}|^{2}<\infty$. As a consequence of Theorem~\ref{theo1.2} we will show that under this last condition, we actually have $\sum a_{n}f^{n}\in\operatorname{BMOA}(\mathbb{D})$.
\begin{corollary}\label{coro1.3}
Let $f$ be an inner function with $f(0)=0$ which is not a rotation and let $\{a_{n}\}$ be a sequence of complex numbers. Assume $\sum\limits^{\infty}_{n=1}|a_{n}|^{2}<\infty$.
\begin{enumerate}
\item[(a)] Then $F=\sum\limits_{n=1}^{\infty} a_{n}f^{n}\in \operatorname{BMOA}(\mathbb{D})$ and for almost every $\xi\in\partial\mathbb{D}$ we have
\begin{equation}\label{eq1.1}
\lim_{r\to 1}F(r\xi)=\sum^{\infty}_{n=1}a_{n}f^{n}(\xi).
\end{equation}
Moreover there exists a constant~$C=C(f)>0$ only depending on~$f$, such that
$$
C^{-1} \sum^{\infty}_{n=1}|a_{n}|^{2} \le \|F\|^{2}_{\operatorname{BMOA}(\mathbb{D})}\le C\sum^{\infty}_{n=1}|a_{n}|^{2}.
$$
\item[(b)] If $f$ is a finite Blaschke product, then $F=\sum\limits^{\infty}_{n=1}a_{n}f^{n}\in\operatorname{VMOA}$.
\end{enumerate}
\end{corollary}
The following Khintchine type estimate follows easily from the previous result.
\begin{corollary}\label{coro1.4}
Let $f$ be an inner function with $f(0)=0$ which is not a rotation and let $\{a_{n}\}$ be a sequence of complex numbers. Assume $\sum\limits^{\infty}_{n=1}|a_{n}|^{2}<\infty$. Then $\sum\limits^{\infty}_{n=1}a_{n} f^{n}\in \mathbb{H}^{p}(\mathbb{D})$ for any~$0<p<\infty$. Moreover for any $0<p<\infty$, there exists a constant~$C(p,f)>0$ such that for any sequence of complex numbers~$\{a_{n}\}$ we have
$$
C(p,f)^{-1}\left( \sum^{\infty}_{n=1}|a_{n}|^{2}\right)^{1/2} \le\left\|\sum_{n=1}^{\infty}a_{n}f^{n}\right\|_{p}\le C(p,f)\left(\sum^{\infty}_{n=1}|a_{n}|^{2}\right)^{1/2}.
$$
\end{corollary}
By Theorem~\ref{theo1.1}, if $\sum |a_{n}|^{2}<\infty$, the series~$\sum a_{n}f^{n}(\xi)$ converges at almost every $\xi\in\partial\mathbb{D}$, while if $\sum |a_{n}|^{2}=\infty$, the series~$\sum a_{n}f^{n}(\xi)$ converges at almost no point~$\xi\in\partial\mathbb{D}$.~If the coefficients satisfy the stronger condition~$\sum |a_{n}|<\infty$, it is clear that $\sum a_{n}f^{n}\in\mathbb{H}^{\infty}(\mathbb{D})$. Our next result says that this condition is also necessary.
\begin{theorem}\label{theo1.5}
Let $f$ be an inner function with $f(0)=0$ which is not a rotation. Let $\{a_{n}\}$ be a sequence of complex numbers with $\sum\limits_{n=1}^\infty |a_{n}|^{2}<\infty$ and $F=\sum\limits_{n=1}^\infty a_{n}f^{n}$. Assume there exists an arc~$I\subset \partial\mathbb{D}$ such that $\sup\{|F(\xi)|:\xi\in I\}<\infty$. Then $\sum\limits_{n=1}^\infty |a_{n}|<\infty$.
\end{theorem}
The proof of Theorem~\ref{theo1.5} is the most technical part of the paper.~The main idea is to construct by induction a sequence of arcs~$I_{k}\subset I$ and a sequence~$M_{k}$ of positive integers tending to infinity such that
\begin{equation}\label{eq1.2}
\frac{1}{m(I_{k})} \int_{I_{k}}\operatorname{Re}F\,dm\ge A\sum^{M_{k}}_{j=1}|a_{j}|-B,\quad k=1,2,\dotsc,
\end{equation}
where $A$ and $B$ are positive constants independent of~$k$.~We start showing that if $|f(z)|$ is sufficiently small, the two terms in the conclusion of Theorem~\ref{theo1.2} are comparable. This provides a way of finding arcs~$I_{k}\subset I$ and of splitting $F=\sum\limits_{j}F_{j}$ into blocks~$F_{k}$ such that
$$
\frac{1}{m(I_{k})} \int_{I_{k}}\sum_{j=1}^{k}
\operatorname{Re}F_{j}\,dm\ge C\sum^{k}_{j=1}\|\operatorname{Re}F_{j}\|_{2},
\quad k=1,2,\dotsc
$$
Since the blocks~$F_{j}$ have a uniformly bounded number of terms, $\|\operatorname{Re}F_{j}\|_{2}$ can be estimated from below by the sum of the modulus of the corresponding coefficients and \eqref{eq1.2} follows easily.
Let $\overline{\mathbb{H}^{\infty}(\mathbb{D})}$ denote the closure of $\mathbb{H}^{\infty}(\mathbb{D})$ in $\operatorname{BMOA}(\mathbb{D})$. This space was studied in~\cite{GJ}. See also~\cite{NS1} and~\cite{SS}. Theorems~\ref{theo1.1}, \ref{theo1.5} and Corollary~\ref{coro1.3} lead to the following result.
\begin{corollary}\label{coro1.6}
Let $f$ be an inner function with $f(0)=0$ which is not a rotation. Let $\{a_{n}\}$ be a sequence of complex numbers with $\sum\limits_{n=1}^\infty |a_{n}|^{2}<\infty$ but $\sum\limits_{n=1}^\infty |a_{n}|=\infty$. Then $F=\sum\limits^{\infty}_{n=1}a_{n}f^{n}\in \overline{\mathbb{H}^{\infty}(\mathbb{D})}$. Moreover
$$
F(\xi)=\lim_{r\to 1}F(r\xi)=\sum^{\infty}_{n=1}a_{n}f^{n}(\xi)
$$
at almost every $\xi\in \partial\mathbb{D}$ but $\sup\{|F(\xi)|:\xi\in I\}=\infty$ for any arc~$I\subset \partial\mathbb{D}$.
\end{corollary}
Theorems~\ref{theo1.1} and \ref{theo1.5} and Corollaries~\ref{coro1.3} and \ref{coro1.4} can be understood as analogues to classical results in the theory of lacunary series (see Section 6 of Chapters V and VI of~\cite{Zy}). However it should be noted that no lacunarity assumption is needed in our results.
The paper is organized as follows. Next section is devoted to the proof of Theorems~\ref{theo1.1}, \ref{theo1.2} and Corollaries~\ref{coro1.3} and \ref{coro1.4}. In Section~\ref{sec3} we start collecting several auxiliary results which are used in the proof of Theorem~\ref{theo1.5}. In Section~\ref{sec4} we describe in terms of the coefficients, those linear combinations of iterates which belong to certain classical function spaces, such as the disc algebra, the Dirichlet space or the Bloch space. Finally last section is devoted to state some related open questions we have not explored.
It is a pleasure to thank Od\'i Soler i Gibert for some very helpful comments.
\section{BMO estimates and pointwise convergence}\label{sec2}
Let $\rho(z,w)$ denote the pseudohyperbolic distance between the points~$z,w\in\mathbb{D}$ defined as $\rho(z,w)=|z-w||1-\overline wz|^{-1}$. Schwarz's Lemma asserts that any analytic selfmapping~$f$ of~$\mathbb{D}$ contracts hyperbolic distances, that is, $\rho(f(z),f(w))\le \rho(z,w)$ for any $z,w\in\mathbb{D}$. Equivalently ${D}_{h}(f)(z)\le 1$, $z\in \mathbb{D}$, where
\begin{equation}\label{eq2.1}
D_{h}(f)(z)=\frac{(1-|z|^{2})|f'(z)|}{1-|f(z)|^{2}},\quad z\in \mathbb{D},
\end{equation}
is the hyperbolic derivative of~$f$ at the point~$z$. Our first auxiliary result is a quantitative version of the Denjoy--Wolff Theorem which is essentially due to Pommerenke (\cite{Po}). Its short proof is included for the sake of completeness.
\begin{lemma}\label{lem2.1}
Let $f\in H^{\infty}(\mathbb{D})$, $\|f\|_{\infty}\le 1$ with $f(0)=0$ which is not a rotation.
\begin{enumerate}
\item[(a)] Then, for any $0<r<1$, there exists $0<c=c(r,f)<1$ such that $1-|z|\le c(1-|f(z)|)$ if $|z|\ge r$.
\item[(b)] Assume $f'(0) \neq 0$. Then there exists $0<r_{0}=r_{0}(f)<1$ such that
$$
|f^{n}(z)|\le r_{0}^{-1} |f'(0)|^{n}|z|,\quad n\ge 1,\quad\text{if}\quad |z|\le r_{0}.
$$
\item[(c)] Assume $f'(0)=0$. Then $|f^{n}(z)|\le |z|^{2^{n}}$, $z\in \mathbb{D}$.
\end{enumerate}
\end{lemma}
\begin{proof}
Since $f(0)=0$, Schwarz's Lemma gives
$$
\rho\left(\frac{f(z)}{z},f'(0)\right)\le |z|,\quad z\in \mathbb{D}.
$$
Denote $\psi(z)=z(z+|f'(0)|)(1+|f'(0)|z)^{-1}$, $z\in \mathbb{D}$, to obtain
\begin{equation}\label{eq2.2}
|f(z)|\le \psi (|z|),\quad z\in \mathbb{D}.
\end{equation}
Then
$$
1-|f(z)|\ge 1-\psi (|z|)=(1-|z|)\frac{1+|z|}{1+|f'(0)||z|}.
$$
Given $0<r<1$, taking $c=(1+r)^{-1}(1+|f'(0)|r)$ the estimate in (a) follows.
We now prove (b). For $n=1,2,\dotsc$, consider the function~$g_{n}(z)=\psi^{n}(z)|f'(0)|^{-n}$. It is known that $\{g_{n}\}$ converges uniformly on compact sets of~$\mathbb{D}$ to an analytic function~$g$ on~$\mathbb{D}$, known as the K\"onigs function of~$\psi$, which satisfies $g(\psi(z))=|f'(0)|g(z)$, $z\in \mathbb{D}$. See \cite[pp.~89--93]{Sh}. Moreover for $0\le x\le 1$ we have
$$
g_{n+1}(x)=\frac{\psi^{n+1}(x)}{|f'(0)|^{n+1}} =g_{n}(x) \frac{1+|f'(0)|^{n-1}g_{n}(x)} {1+|f'(0)|^{n+1}g_{n}(x)}\ge g_{n}(x).
$$
Hence $g_{n}(x)\le g(x)$, $n=1,2,\dotsc,0<x<1$. Note that there exists a constant $\delta=\delta (|f'(0)|)>0$ such that $\psi$ is univalent in $\{z\in \mathbb{D}:|z|<\delta\}$. Hence $g_{n}$ is also univalent in $\{z\in \mathbb{D}: |z|<\delta\}$. By Koebe Distortion Theorem, there exists $0<r_{0}=r_{0}(f)<1$ such that $|g(w)|<1$ if $|w|<r_{0}$. By Schwarz's Lemma we deduce $|g(w)| \leq r_{0}^{-1}|w|$ if $|w|<r_{0}$. Applying \eqref{eq2.2}, for any $|z| < r_{0}$ we have
$$
|f^{n}(z)|\le \psi^{n} (|z|)\le |f'(0)|^{n} g(|z|)\le r_{0}^{-1} |f'(0)|^{n}|z|,\quad n=1,2,\dotsc
$$
This proves (b). Assume now $f'(0)=0$. Note that \eqref{eq2.2} gives $|f(z)|\le |z|^{2}$, $z\in \mathbb{D}$. Iterating we obtain (c).
\end{proof}
For future reference we state the following easy consequence of Lemma~\ref{lem2.1}.
\begin{corollary}\label{coro2.2}
Let $f\in \mathbb{H}^{\infty}(\mathbb{D})$, $\|f\|_{\infty}\le 1$ with $f(0)=0$ which is not a rotation. Then there exist constants $0<r_{0}=r_{0}(f)<1$ and $0<c_{0}=c_{0}(f)<1$ such that
$$
|f^{n}(z)|\le r_{0}^{-1} c_{0}^{n}|z|\quad \text{if}\quad |z|\le r_{0}.
$$
\end{corollary}
\begin{proof}
If $f'(0)\ne 0$ let $r_{0}=r_{0}(f)$ be the constant given by part~(b) of Lemma~\ref{lem2.1} and $c_{0}=|f'(0)|$. If $f'(0)=0$, pick any $0<r_{0}<1$. Part~(c) of Lemma~\ref{lem2.1} gives $|f^{n}(z)|\le r_{0}^{2^{n}-1}|z|$, $|z|\le r_{0}$. Since $2^{n}-1\ge n$, $n\ge 1$, we can take $c_{0}=r_{0}$.
\end{proof}
For future reference we also state the following auxiliary result whose proof can be found in \cite[Theorem~9]{NS2}.
\begin{lemma}\label{lem2.3}
Let $f$ be an inner function with $f(0)=0$ which is not a rotation. Let $\{a_{n}\}$ be a sequence of complex numbers. Then
$$
\frac{1-|f'(0)|}{1+|f'(0)|} \sum^{N}_{n=1}|a_{n}|^{2} \le \left\|\sum^{N}_{n=1}a_{n}f^{n}\right\|_2^2 \le \frac{1+|f'(0)|}{1-|f'(0)|} \sum^{N}_{n=1}|a_{n}|^{2},\quad N=1,2,\dotsc
$$
\end{lemma}
We are now ready to prove Theorem~\ref{theo1.2}.
\begin{proof}[Proof of Theorem~\ref{theo1.2}]
Without loss of generality we can assume $f^{n}(z)\ne 0$ for any $n\ge 1$. Denote $F=\sum\limits^{N}_{n=1}a_{n}f^{n}$. Then
$$
\int_{\partial \mathbb{D}}|F(\xi)-F(z)|^{2}P(z,\xi)\,dm (\xi)=\sum^{N}_{n=1}|a_{n}|^{2}c_{n,n}+2\operatorname{Re}\sum^{N-1}_{n=1}\sum^{N}_{k>n}\overline{a}_{n}a_{k}c_{k,n},
$$
where
$$
c_{k,n}=\int_{\partial \mathbb{D}}\overline{(f^{n}(\xi)-f^{n}(z))} (f^{k}(\xi)-f^{k}(z))P(z,\xi)\,dm(\xi),\quad k,n=1,\dotsc,N.
$$
For $k\ge n$, the function~$\overline{f^{n}(\xi)} f^{k}(\xi)=f^{k}(\xi)/f^{n}(\xi)$, $\xi\in\partial \mathbb{D}$, has an analytic extension to~$\mathbb{D}$. Then
$$
c_{k,n}=\frac{f^{k}(z)}{f^{n}(z)}-f^{k}(z) \overline{f^{n}(z)}=\frac{f^{k}(z)(1-|f^{n}(z)|^{2})}{f^{n}(z)},\quad k\ge n.
$$
Hence
\begin{equation}\label{eq2.3}
\int_{\partial \mathbb{D}}|F(\xi)-F(z)|^{2}
P(z,\xi)=A+2\operatorname{Re}B,
\end{equation}
where
\begin{align*}
A&=\sum^{N}_{n=1} |a_{n}|^{2} (1-|f^{n}(z)|^{2}),\\*[5pt]
B&=\sum^{N-1}_{n=1}\sum^{N}_{k>n}\overline{a}_{n} a_{k}\frac{f^{k}(z)(1-|f^{n}(z)|^{2})}{f^{n}(z)}.
\end{align*}
The idea is that the cross term~$B$ can be estimated by the diagonal term~$A$. Let $0<r_{0}=r_{0}(f)<1$ and $0<c_{0}=c_{0}(f)<1$ be the constants appearing in Corollary~\ref{coro2.2}. The Denjoy-Wolff Theorem says that the iterates $f^n$ converge to $0$ uniformly on compacts of $\mathbb{D}$. So we can consider the smallest positive integer $\ell=\ell(z)$ such that $|f^{\ell}(z)|\le r_{0}$. Corollary~\ref{coro2.2} gives that $|f^{k}(z)|\le r_{0}^{-1} c_{0}^{k-n} |f^{n}(z)|$ if $k\ge n >\ell$. Assume $\ell <N$. We have
$$
\sum^{N-1}_{n\ge \ell}|a_{n}|\frac{1-|f^{n}(z)|^{2}}{|f^{n}(z)|} \sum^{N}_{k>n} |a_{k}| |f^{k}(z)|\le r_{0}^{-1} \sum^{N-1}_{n\ge \ell}|a_{n}| \sum^{N}_{k>n}|a_{k}|c_{0}^{k-n}.
$$
Writting $j=k-n$ and applying Cauchy--Schwarz's inequality, last double sum can be bounded by
$$
\sum^{N- \ell}_{j=1}c_{0}^{j} \sum^{N-j}_{n\ge \ell} |a_{n}| |a_{n+j}| \le (1-c_{0})^{-1} \sum^{N}_{n\ge \ell}|a_{n}|^{2}.
$$
Since $|f^{n}(z)|\le |f^{\ell}(z)|\le r_{0}$ for $n\ge \ell$, last expression above is bounded by
$$
(1-r_{0}^{2})^{-1} (1-c_{0})^{-1}\sum^{N}_{n\ge \ell} |a_{n}|^{2} (1-|f^{n}(z)|^{2}).
$$
Hence we only need to estimate $B_{1}+B_{2}$ where
\begin{align*}
B_{1}&= \sum^{\ell-1}_{n=1} |a_{n}| \frac{1-|f^{n}(z)|^{2}}{|f^{n}(z)|} \sum^{\ell}_{k>n}|a_{k}| |f^{k}(z)|,\\*[5pt]
B_{2}&= \sum^{\ell-1}_{n=1} |a_{n}| \frac{1-|f^{n}(z)|^{2}}{|f^{n}(z)|} \sum^{N}_{k>\ell}|a_{k}| |f^{k}(z)|.
\end{align*}
Since $|f^{k}(z)|\le |f^{n}(z)|$ for $k\ge n$, writting $j=k-n$ we obtain
$$
B_{1}\le \sum^{\ell-1}_{n=1} |a_{n}| (1-|f^{n}(z)|^{2}) \sum^{\ell}_{k>n}|a_{k}| \le
\sum^{\ell}_{j=1}\sum^{\ell-j}_{n=1}|a_{n}| |a_{n+j}| (1-|f^{n}(z)|^{2}).
$$
Since $|f^{n}(z)|\ge r_{0}$ for $n<\ell$, part~(a) of Lemma~\ref{lem2.1} provides a constant $0<c_{1}=c_{1}(f)<1$ such that $1-|f^{n}(z)|\le c_{1}^{j} (1-|f^{n+j}(z)|)$ for any $n<n+j\le \ell$. Hence
$$
B_{1}\le 2\sum^{\ell}_{j=1}c_{1}^{j/2} \sum^{\ell-j}_{n=1} |a_{n}| (1-|f^{n}(z)|)^{1/2} |a_{n+j}|(1-|f^{n+j}(z)|)^{1/2}.
$$
Applying Cauchy--Schwarz's inequality, we deduce
$$
B_{1}\le 2 (1-c_{1}^{1/2})^{-1} \sum^{\ell}_{n=1}|a_{n}|^{2} (1-|f^{n}(z)|).
$$
Finally we estimate $B_{2}$. Note that Corollary~\ref{coro2.2} gives $|f^{k}(z)|\le r_{0}^{-1} c_{0}^{k-\ell}|f^{\ell}(z)|$ if $k\ge \ell$. Since $|f^{n}(z)|\ge |f^{\ell}(z)|$ for $n\le\ell$, we have
$$
B_{2}\le r_{0}^{-1} \sum^{\ell-1}_{n=1}|a_{n}| (1-|f^{n}(z)|^{2})\sum^{N}_{k>\ell }|a_{k}|c_{0}^{k-\ell}.
$$
Cauchy--Schwarz's inequality gives
\begin{equation}\label{eq2.4}
\sum^{N}_{k>\ell}|a_{k}|c_{0}^{k-\ell} \le (1-c_{0}^2)^{-1/2} \left(\sum^{N}_{k>\ell}|a_{k}|^{2}\right)^{1/2}
\end{equation}
and
$$
\sum^{\ell-1}_{n=1}|a_{n}| (1-|f^{n}(z)|)\le \left(\sum^{\ell-1}_{n=1}|a_{n}|^{2} (1-|f^{n}(z)|)\right)^{1/2} \left(\sum^{\ell-1}_{n=1}(1-|f^{n}(z)|)\right)^{1/2}.
$$
Since $1-|f^{n}(z)|\le c_{1}^{\ell-n}$ for $1\le n<\ell$, last sum is bounded by~$(1-c_{1})^{-1}$ and we deduce
\begin{equation}\label{eq2.51}
\sum^{\ell-1}_{n=1}|a_{n}| (1-|f^{n}(z)|)\le (1-c_{1})^{-1/2} \left(\sum^{\ell-1}_{n=1}|a_{n}|^{2} (1-|f^{n}(z)|)\right)^{1/2}.
\end{equation}
Now applying \eqref{eq2.4} and \eqref{eq2.51} we deduce
\begin{equation*}
\begin{split}
B_{2}&\le c_{2} \left(\sum_{n=1}^{\ell-1} |a_{n}|^{2} (1-|f^{n}(z)|^{2})\right)^{1/2} \left(\sum^{N}_{k>\ell}|a_{k}|^{2}\right)^{1/2}\\*[5pt]
&\le \frac{c_{2}}{2}\left(\sum^{\ell-1}_{n=1}|a_{n}|^{2} (1-|f^{n}(z)|^{2})+\sum^{N}_{k>\ell}|a_{k}|^{2}\right),
\end{split}
\end{equation*}
where $c_{2}=2 r_0^{-1} (1-c_{0}^2)^{-1/2}(1-c_{1})^{-1/2}$. Finally observe that $|f^{k}(z)|\le r_{0}$ for $k>\ell$ to deduce
$$
\sum^{N}_{k>\ell}|a_{k}|^{2}\le (1-r_{0}^{2})^{-1} \sum^{N}_{k>\ell}|a_{k}|^{2} (1-|f^{k}(z)|^{2}).
$$
This finishes the proof if $\ell <N$. Assume now $\ell \geq N$, that is $|f^k (z)| > r_0$ for any $k<N$. Then we argue as in the estimate of $B_1$ replacing $\ell$ by $N$, to obtain
$$
\sum^{N-1}_{n=1} |a_{n}| \frac{1-|f^{n}(z)|^{2}}{|f^{n}(z)|} \sum^{N}_{k>n}|a_{k}| |f^{k}(z)| \leq 2 (1- c_1^{1/2})^{-1} \sum_{n=1}^{N} |a_{n}|^{2} (1-|f^{n}(z)| ).
$$
This finishes the proof.
\end{proof}
The following well known auxiliary result plays a fundamental role in the classical work of Paley and Zygmund, as well as in the study of pointwise convergence of random series of functions (\cite{Ka}). Its short proof is included for the sake of completeness.
\begin{lemma}[Paley--Zygmund inequality]\label{lem2.4}
Let $(X,\Omega,d\mu)$ be a probability space and $Z\colon X\to [0,\infty)$ be a positive square integrable random variable. Then for any $0<\lambda <1$, we have
$$
\mu\left\{x\in X:Z(x)>\lambda\int_{X}Z\,d\mu\right\}\ge (1-\lambda)^{2} \frac{\left(\displaystyle\int_{X}Z\,d\mu\right)^{2}}{\displaystyle\int_{X}Z^{2}\,d\mu}.
$$
\end{lemma}
\begin{proof}
We can assume $\displaystyle\int_{X}Z\,d\mu=1$. Let $W_{\lambda}$ denote the indicator function of the set~$\{x\in X: Z(x) \leq \lambda\}$. Cauchy--Schwarz's inequality gives
$$
1=\int_{X}ZW_{\lambda}\,d\mu +\int_{X}Z(1-W_{\lambda})\,d\mu\le \lambda +\left( \int_{X}Z^{2}\,d\mu\right)^{1/2} \mu \{x\in X:Z(x)>\lambda\}^{1/2}.
$$
\end{proof}
We are now ready to prove Theorem \ref{theo1.1} stated in the Introduction.
\begin{proof}[Proof of Theorem~\ref{theo1.1}]
It is obvious that (a) implies (b). We start proving that (c) implies (a). By Lemma \ref{lem2.3}, the series $\sum a_{n}f^{n}$ converges in $\mathbb{H}^{2}(\mathbb{D})$. Let $F(z)=\sum\limits_{n=1}^{\infty} a_{n}f^{n}(z)$, $z\in\mathbb{D}$. Hence the non-tangential limit
$$
F(\xi) =\lim_{{z {\to}_{\nless} \xi} }F(z)
$$
exists for almost every $\xi\in\partial\mathbb{D}$. For almost every $\xi\in \partial\mathbb{D}$ we will construct a sequence of points~$\{z_{N}=z_{N}(\xi): N=1,2,\dotsc\}$ tending non-tangentially to~$\xi$, such that for any $\varepsilon>0$ there exists an integer $N_{0}=N_{0}(\varepsilon,\xi)>0$ satisfying
\begin{equation}\label{eq2.5}
\left|\sum^{N}_{n=1}a_{n} f^{n}(\xi) - F(z_{N})\right|<\varepsilon,\quad N \geq N_{0}.
\end{equation}
It is clear that \eqref{eq2.5} implies the statement in~(a).~By Corollary~\ref{coro2.2}, there exist constants $0<r_{0}=r_{0}(f)<1$, $0<c_{0}=c_{0}(f)<1$ such that
\begin{equation}\label{eq2.6}
|f^{n}(z)|\le r_{0}^{-1}c_{0}^{n} |z|,\quad |z|\le r_{0},\quad n=1,2,\dotsc
\end{equation}
Fix $N\ge 1$.~Write $F_{N}=\sum\limits^{N}_{n=1}a_{n}f^{n}$.~Since $f(0)=0$ and $f^{N}$ is inner, there exists a subset $S=S_{N}\subset\partial\mathbb{D}$ with $m(\partial\mathbb{D}\backslash S)=0$ such that for any $\xi\in S$ there exists $0<r_{N}=r_{N}(\xi)< 1$ such that $|f^{N} (r_{N}\xi)|=r_{0}$. Since $f^{N}$ tends to~$0$ uniformly on compact sets of~$\mathbb{D}$ as~$N\to\infty$, we deduce that $r_{N}\to 1$ as $N\to\infty$. Note also that $|f^{j}(r_{N}\xi)|\ge |f^{N}(r_{N}\xi)|=r_{0}$ for $1\le j\le N$. For $\xi\in S$ consider the arc~$I(N,\xi)=\{z\in\partial\mathbb{D}:|z-\xi|<1-r_{N}\}$. Apply Vitali's Covering Lemma (see~\cite[p.~27]{EG}) to obtain a subcollection~$\{I(N,\xi_{k}):k=1,2,\dotsc\}$ of pairwise disjoint arcs such that $m(\partial\mathbb{D}\backslash\cup 5I(N,\xi_{k}))=0$. Here $5I$ denotes the arc in the unit circle having the same center than $I$ and with $m(5I) = 5 m(I)$. Given $\varepsilon>0$ consider the set
$$
E(N)=\bigcup_{k}\{ \xi\in 5 I(N,\xi_{k}): |F_{N}(\xi)-F_{N}(\xi_{k})|\ge\varepsilon\}
$$
and
$$
E=\bigcap_{\ell\ge 1}\bigcup_{N\ge\ell}E(N).
$$
We first show that $m(E)=0$. By the classical Borel--Cantelli Lemma, it is sufficient to show
\begin{equation}\label{eq2.7}
\sum^{\infty}_{N=1}m(E(N))<\infty.
\end{equation}
Observe that Theorem~\ref{theo1.2} gives that $F\in\operatorname{BMOA}$. Since $P(r_{N} \xi_{k},\xi)$ is comparable to $m(I(N,\xi_{k}))^{-1}$ for $\xi\in 5I(N,\xi_{k})$, there exists a constant~$c_{1}=c_{1}(f)>0$ such that for any $k,N\ge 1$, we have
$$
\frac{1}{m(I(N,\xi_{k}))}\int_{5I(N,\xi_{k})} |F_{N}(\xi)-F_{N}(r_{N}\xi_{k})|^{2}\,dm(\xi)\le
c_{1}\sum^{N}_{n=1}|a_{n}|^{2} (1-|f^{n}(r_{N}\xi_{k})|).
$$
By part~(a) of Lemma~\ref{lem2.1} and the choice of~$r_{N}$, there exists a constant $0<c<1$ such that $1-|f^{n}(r_{N}\xi_{k})|\le c^{N-n}(1-r_{0})$, $n=1,\dotsc,N$. Hence
$$
\frac{1}{m(I(N,\xi_{k}))}\int_{5I(N,\xi_{k})}|F_{N}(\xi)-F_{N}(r_{N}\xi_{k})|^{2}\,dm(\xi)\le c_{1}(1-r_{0})\sum^{N}_{n=1}|a_{n}|^{2}c^{N-n}, \quad k,N\ge 1.
$$
Given $\varepsilon >0$, we deduce that for any $k,N\ge 1$, we have
$$
\frac{1}{m(I(N,\xi_{k}))} m(\{\xi\in 5I(N,\xi_{k}):|F_{N}(\xi)-F_{N}(r_{N}\xi_{k})|>\varepsilon\})\le \frac{c_{1}(1-r_{0})}{\varepsilon^{2}} \sum^{N}_{n=1}|a_{n}|^{2}c^{N-n}.
$$
Since $\{I(N,\xi_{k}): k=1,2, \dotsc\}$ are pairwise disjoint, we deduce
$$
m(E(N))\le \frac{c_{1}(1-r_{0})}{\varepsilon^{2}} \sum^{N}_{n=1}|a_{n}|^{2} c^{N-n}.
$$
Then
$$
\sum^{\infty}_{N=1}m(E(N)) \le \frac{c_{1}(1-r_{0})}{\varepsilon^{2}} \sum^{\infty}_{N=1} \sum^{N}_{n=1}|a_{n}|^{2}c^{N-n} = \frac{c_{1}(1-r_{0})}{\varepsilon^{2}(1-c)} \sum^{\infty}_{n=1}|a_{n}|^{2}.
$$
This proves \eqref{eq2.7} and shows that $m(E)=0$. Finally for almost every $\xi\in \partial\mathbb{D}\backslash E$ we will construct the sequence $\{z_{N}=z_{N}(\xi) \}$ for which \eqref{eq2.5} holds. Since $m(\partial\mathbb{D}\backslash\cup 5I(N,\xi_{k}))=0$, for almost every $\xi\in\partial\mathbb{D}\backslash E$ there exists $N_{0}=N_{0}(\xi)>0$ such that for any~$N>N_{0}$ there exists $\xi_{k}\in\partial\mathbb{D}$ with $\xi\in 5I(N,\xi_{k})$ and
\begin{equation}\label{eq2.8}
|F_{N}(\xi)-F_{N}(r_{N}\xi_{k} )|<\varepsilon.
\end{equation}
The choice of $r_{N}$ and \eqref{eq2.6} gives that
$$
|f^{n}(r_{N}\xi_{k})|\le c_{0}^{n-N},\quad n\ge N.
$$
Hence
\begin{equation}\label{eq2.9}
\sum_{n>N}|a_{n}| |f^{n}(r_{N}\xi_{k})|\le \sum_{n>N}|a_{n}|c_{0}^{n-N}\le \left(1-c_{0}^2 \right)^{-1/2} \left(\sum_{n>N}|a_{n}|^{2}\right)^{1/2}.
\end{equation}
Pick $z_N = z_N (\xi) = r_N \xi_k$, where $\xi_k$ is chosen such that $\xi\in 5I(N,\xi_{k})$. Now \eqref{eq2.8} and \eqref{eq2.9} imply \eqref{eq2.5}. Note that since $\xi\in 5I(N, \xi_{k})$ and $r_{N}\to 1$, the points~$z_{N}$ converge non-tangentially to~$\xi$ as $N$ tends to $\infty$.
We now prove that (b) implies (c). Given two sets $A,B\subset\partial\mathbb{D}$ we use the notation $A\overset{\text{a.e.}}{=}B$ if $A$ and $B$ differ at most on a set of Lebesgue measure zero. Consider the set
$$
A=\left\{\xi\in\partial\mathbb{D}:\sup_{N}\left|\sum^{N}_{n=1}a_{n}f^{n}(\xi)\right|<\infty\right\}.
$$
Note that $f^{-k} (f^{k}(A))\overset{\text{a.e.}}{=} A$ for any $k=1,2, \dotsc$ Since the mapping $f\colon \partial\mathbb{D}\to \partial\mathbb{D}$ is exact (\cite{Ne}) we deduce that $m(A)=0$ or $m(A)=1$. Our assumption gives $m(A)=1$. Write
$$
F_{N}=\sum^{N}_{n=1}a_{n}f^{n}
$$
and $s_{N}=\|F_{N}\|_{2}$. By Lemma~\ref{lem2.3} there exists a constant~$c=c(f)>1$ such that
\begin{equation}\label{eq2.10}
c^{-1} \sum^{N}_{n=1}|a_{n}|^{2}\le s^{2}_{N}\le c\sum^{N}_{n=1}|a_{n}|^{2},\quad N=1,2,\dotsc
\end{equation}
Now the Paley--Zygmund inequality stated in Lemma~\ref{lem2.4} gives that for any $0<\lambda<1$ we have
$$
m\{\xi\in\partial\mathbb{D}:|F_{N}(\xi)|^{2}\ge \lambda s_{N}^{2}\}\ge (1-\lambda)^{2}\frac{s_{N}^{4}}{\displaystyle\int_{\partial\mathbb{D}}|F_{N}|^{4}\,dm}.
$$
It was proved in \cite[part (c) of Theorem~9]{NS2} that there exists a constant~$c_{1}=c_{1}(f)>0$ such that
$$
\int_{\partial\mathbb{D}}|F_{N}|^{4}\,dm \le c_{1} s_{N}^{4}.
$$
Hence
$$
m\{\xi\in\partial\mathbb{D}: |F_{N}(\xi)|^{2}\ge \lambda s_{N}^{2}\}\ge c_{1}^{-1} (1-\lambda)^{2},\quad 0<\lambda <1.
$$
Since $m(A)=1$, we deduce that $s_{N}$ is bounded. Applying \eqref{eq2.10} we deduce $\sum\limits_{n}|a_{n}|^{2}<\infty$.
\end{proof}
We are now ready to prove Corollaries~\ref{coro1.3} and~\ref{coro1.4}.
\begin{proof}[Proof of Corollary~\ref{coro1.3}]
By Lemma \ref{lem2.3}, the series $\sum a_{n}f^{n}$ converges in $\mathbb{H}^{2}(\mathbb{D})$. Let $F=\sum\limits_{n\ge 1} a_{n}f^{n}$. Theorem~\ref{theo1.2} gives that $F\in\operatorname{BMOA}(\mathbb{D})$ and $\|F\|^{2}_{\operatorname{BMOA}(\mathbb{D})}\le C \sum\limits_{n\ge 1}|a_{n}|^{2}$. The converse estimate follows easily from Lemma \ref{lem2.3} which gives
$$
\|F\|_{\operatorname{BMOA}(\mathbb{D})}^{2}\ge \|F\|_{2}^{2}\ge \frac{1-|f'(0)|}{1+|f'(0)|} \sum\limits_{n=1}^{\infty} |a_{n}|^{2} .
$$
The identity~\eqref{eq1.1} follows from \eqref{eq2.5}. This finishes the proof of (a). Assume now that $f$ is a finite Blaschke product. Fix a number $0<r<1$ and let $0<c=c(r,f)<1$ be the constant appearing in part~(a) of Lemma~\ref{lem2.1}. For $z\in\mathbb{D}$ let $N=N(z)$ be the smallest positive integer such that $|f^{N}(z)|\le r$. Note that $|f^{n}(z)|\le r$ for $n\ge N$ and $1-|f^{n}(z)|\le c^{N-n}$ for $1\le n\le N$. Since $f$ is a finite Blaschke product, we have
\begin{equation}\label{eq2.11}
\lim_{|z|\to 1}N(z)=+\infty.
\end{equation}
Now
$$
\sum^{\infty}_{n=1}|a_{n}|^{2} (1-|f^{n}(z)|)\le \sum^{N}_{n=1}|a_{n}|^{2}c^{N-n}+\sum^{\infty}_{n=N+1}|a_{n}|^{2}
$$
and \eqref{eq2.11} gives
$$
\lim_{|z|\to 1}\sum^{\infty}_{n=1}|a_{n}|^{2} (1-|f^{n}(z)|^{2})=0.
$$
Theorem~\ref{theo1.2} finishes the proof.
\end{proof}
\begin{proof}[Proof of Corollary~\ref{coro1.4}]
Let $F=\sum\limits_{n=1}^{\infty}a_{n}f^{n}$ and $s^{2}=\sum\limits_{n=1}^{\infty}|a_{n}|^{2}$. We first prove the upper estimate.
Consider the distribution function $\Phi(\lambda)=m\{\xi\in\partial\mathbb{D}: |F(\xi)|>\lambda\}$, $\lambda>0$. Corollary~\ref{coro1.3} and the John--Nirenberg Theorem give that there exist universal constants $A,B>0$ such that
$$
\Phi(\lambda)\le A e^{-B\lambda/s},\quad \lambda>0.
$$
Then
$$
\|F\|_{p}^{p}=\int_{0}^{\infty} p\lambda^{p-1} \Phi (\lambda)\,d\lambda\le C(p)s^{p},
$$
where $C(p)$ is a constant depending on $A$, $B$ and $p$. The lower estimate follows from the following standard duality argument. We can assume $p<2$. By H\"older's inequality
$$
\int_{\partial\mathbb{D}}|F|^{2}\,dm\le \left(\int_{\partial\mathbb{D}} |F|^{p}\,dm\right)^{1/p}\left(
\int_{\partial\mathbb{D}} |F|^{q}\,dm\right)^{1/q},
$$
where $p^{-1}+q^{-1}=1$. By Lemma~\ref{lem2.3} and the upper estimate we have already proved, there exists a constant~$c=c(f,p)>0$ such that
$$
\sum^{\infty}_{n=1}|a_{n}|^{2}\le c\left( \int_{\partial\mathbb{D}} |F|^{p}\,dm\right)^{1/p} \left(\sum^{\infty}_{n=1}|a_{n}|^{2}\right)^{1/2}.
$$
This finishes the proof.
\end{proof}
\section{Uniform estimates}\label{sec3}
Given $F\in L^{1}(\partial \mathbb{D})$ and $z\in \mathbb{D}$, let $F(z)$ denote the value of its harmonic extension at~$z$, that is,
$$
F(z)=\int_{\partial \mathbb{D}}F(\xi)P(z,\xi)\,dm (\xi).
$$
If $F$ is the characteristic function of a measurable set~$E\subset\partial \mathbb{D}$, the corresponding function is called the harmonic measure of the set~$E$ from the point~$z\in \mathbb{D}$ and will be denoted by $F(z) = w (z, E)$. A function $F\in L^{1}(\partial \mathbb{D})$ is in the space~$\operatorname{BMO}(\partial \mathbb{D})$ if
$$
\|F\|_{\operatorname{BMO}(\partial \mathbb{D})}^{2}=\sup_{z\in \mathbb{D}}\int_{\partial \mathbb{D}}|F(\xi)-F(z)|^{2}P(z,\xi)\,dm(\xi)<\infty.
$$
Equivalently, $F\in L^{1}(\partial \mathbb{D})$ is in $\operatorname{BMO}(\partial \mathbb{D})$ if and only if there exists a constant~$C=C(F)>0$ such that for any arc~$I\subset \partial \mathbb{D}$, we have
$$
\frac{1}{m(I)}\int_{I}|F-F_{I}|^{2}\,dm\le C.
$$
Here $F_{I}=m(I)^{-1}\int_{I}F\,dm$ denotes the mean of~$F$ on~$I$. See Chapter VI of~\cite{Ga}. For $z\in \mathbb{D}$ let $ \tau_{z}$ be the automorphism of~$\mathbb{D}$ given by $\tau_{z}(w)=(w-z)(1-\overline{z}w)^{-1}
$, $w\in \mathbb{D}$. The John--Nirenberg Theorem applied to $F\circ \tau_{z}-F(z)$ provides two universal constants~$A,B>0$ such that
\begin{equation}\label{eqJN}
w(z,\{\xi\in\partial \mathbb{D}: |F(\xi)-F(z)|>\lambda\})\le A e^{-B\lambda/\|F\|_{\operatorname{BMO}(\partial \mathbb{D})}},\quad \lambda>0.
\end{equation}
See \cite{Ba}. The proof of Theorem \ref{theo1.5} is based on a careful study of the oscillation of the partial sums of $\sum a_n f^n$. We start with an auxiliary result which holds for any $BMO$ function.
\begin{lemma}\label{lem3.1}
For any $0<c<1$ there exists a constant~$0< \delta=\delta(c) <1$ such that the following statement holds. Let $F$ be a real valued function defined on~$\partial \mathbb{D}$ with $\|F\|_{\operatorname{BMO}(\partial \mathbb{D})}=1$. Let $z\in\mathbb{D}$ such that
\begin{equation}\label{eq3.1}
\int_{\partial\mathbb{D}}|F(\xi)-F(z)|^{2}P(z,\xi)\,dm(\xi)\ge c^{2}.
\end{equation}
Then $w(z,\{\xi\in\partial\mathbb{D}: F(\xi)-F(z)\ge \delta\})\ge \delta$.
\end{lemma}
\begin{proof}
Given $0\le a<b$, let $E(a,b)$ denote the set of points~$\xi\in\partial\mathbb{D}$ such that $a\le |F(\xi)-F(z)|\le b$. Apply the John--Nirenberg estimate~\eqref{eqJN} to find a constant~$c_{1}=c_{1}(c)>0$, such that
$$
\int_{\partial\mathbb{D}\backslash E(0,c_{1})} |F(\xi)-F(z)|^{2}P(z,\xi)\,dm(\xi)\le \frac{c^{2}}{4}.
$$
Since
$$
\int_{E(0,c/2)}|F(\xi)-F(z)|^{2}P(z,\xi)\,dm(\xi)\le \frac{c^{2}}{4},
$$
we deduce that
$$
\int_{E(c/2,c_{1})}|F(\xi)-F(z)|^{2}P(z,\xi)\,dm(\xi)\ge\frac{c^{2}}{2}.
$$
Hence
\begin{equation}\label{eq3.2}
w\left(z,\left\{\xi\in\partial \mathbb{D}: |F(\xi)-F(z)|\ge \frac{c}{2}\right\}\right)\ge \frac{c^{2}}{2c_{1}^{2}}.
\end{equation}
Apply the John--Nirenberg estimate~\eqref{eqJN} to find a constant~$c_{2}=c_{2}(c)>0$ such that
\begin{equation}\label{eq3.3}
\int_{\partial\mathbb{D}\backslash E(0,c_{2})} |F(\xi)-F(z)|P(z,\xi)\,dm(\xi)\le \frac{c^{3}}{16c_{1}^2}.
\end{equation}
Let $0<\delta=\delta(c)<c/2$ be a small constant to be fixed later.~We will show that if $\delta >0$ is conveniently chosen, then
\begin{equation}\label{eq3.4}
w(z,\{\xi:F(\xi)-F(z)>\delta\})\ge\min \left\{\frac{c^{3}}{16c_{1}^{2}c_{2}},\frac{c^{2}}{16c_{1}^{2}}\right\}.
\end{equation}
We argue by contradiction. Assume
\begin{equation}\label{eq3.5}
w(z,\{\xi:F(\xi)-F(z)>\delta\})\le\min \left\{\frac{c^{3}}{16c_{1}^{2}c_{2}},\frac{c^{2}}{16c_{1}^{2}}\right\}.
\end{equation}
Apply \eqref{eq3.2} to deduce that $w(z,\{\xi:F(\xi)-F(z)<-c/2\})\ge 7c^{2}/16c_{1}^{2}$. Hence
\begin{equation}\label{eq3.6}
\int_{\{\xi:F(\xi)-F(z)\le 0\}} (F(\xi)-F(z))P(z,\xi)\,dm(\xi)\le \frac{-7c^{3}}{32c_{1}^{2}}.
\end{equation}
For $0< \gamma < \tau$ consider the set $G(\gamma, \tau) = \{\xi \in \partial \mathbb{D} : \gamma \leq F(\xi )- F(z) \leq \tau\}$. We have
\begin{equation*}
\begin{split}
\int_{\{\xi:F(\xi)-F(z)>0\}} (F(\xi)-F(z))P(z,\xi)\,dm(\xi)&\le\delta +\int_{G(\delta , c_2)}
(F(\xi)-F(z))P(z,\xi)\,dm(\xi) \\
&\quad+\int_{\{\xi:F(\xi)-F(z)>c_{2}\}}(F(\xi)-F(z))P(z,\xi)\,dm(\xi).
\end{split}
\end{equation*}
The choice~\eqref{eq3.3} of $c_{2}$ gives that last integral is bounded by $c^{3}/16c_{1}^{2}$. Moreover by~\eqref{eq3.5}, we have
$$
\int_{G(\delta , c_2)} (F(\xi)-F(z))P(z,\xi)\,dm(\xi) \le c_{2}w(z,\{\xi:F(\xi)-F(z)>\delta\})\le \frac{c^{3}}{16c_{1}^{2}}.
$$
We deduce
\begin{equation}\label{eq3.7}
\int_{\{\xi: F(\xi)-F(z)>0\}} (F(\xi)-F(z))P(z,\xi)\,dm(\xi) \le\delta +\frac{c^{3}}{8c_{1}^{2}}.
\end{equation}
Choosing $0<\delta <c^{3}/16c_{1}^{2}$ we observe that \eqref{eq3.6} and \eqref{eq3.7} contradict the identity
$$
\int_{\partial\mathbb{D}} (F(\xi)-F(z))P(z,\xi)\,dm(\xi) =0.
$$
Hence choosing $0<\delta<\min\{c/2,c^{3}/16c_{1}^{2}\}$, estimate~\eqref{eq3.4} holds. This finishes the proof.
\end{proof}
Our next auxiliary result says that $F=\sum\limits_{n=M}^{N}a_{n}f^{n}$ satisfies condition~\eqref{eq3.1} in Lemma~\ref{lem3.1} if $|f^{M}(z)|$ is sufficiently small.
\begin{lemma}\label{lem3.2}
Let $f$ be an inner function with $f(0)=0$ which is not a rotation. Then there exists a constant~$ 0< \varepsilon =\varepsilon (f)< 1$ such that if $M<N$ are positive integers, $\{a_{n}:M \le n\le N\}$ are complex numbers and $z\in\mathbb{D}$ satisfies $|f^{M}(z)|\le\varepsilon$, we have
$$
\int_{\partial\mathbb{D}}\left|\sum^{N}_{n=M}a_{n}(f^{n}(\xi)-f^{n}(z))\right|^{2} P(z,\xi)\,dm(\xi)\ge \frac{1}{2}\, \frac{1-|f'(0)|}{1+|f'(0)|} \sum^{N}_{n=M}|a_{n}|^{2}.
$$
\end{lemma}
\begin{proof}
By continuity we can assume that $f^{n}(z)\ne 0$ for any~$n$. Write $F=\sum\limits_{n=M}^{N}a_{n}f^{n}$. In \eqref{eq2.3} we already noted that
\begin{equation}\label{eq3.8}
\begin{split}
\int_{\partial \mathbb{D}}|F(\xi)-F(z)|^{2}
P(z,\xi)\,dm(\xi)&=\sum^{N}_{n=M} |a_{n}|^{2}(1-|f^{n}(z)|^{2})\\
&\quad+2\operatorname{Re}\sum^{N-1}_{n=M}
\sum^{N}_{k>n}\overline{a}_{n}a_{k}(1-|f^{n}(z)|^{2}) \frac{f^{k}(z)}{f^{n}(z)}.
\end{split}
\end{equation}
The idea is to compare the expression above with
\begin{equation}\label{eq3.9}
\int_{\partial\mathbb{D}}|F(\xi)|^{2}\,dm(\xi)=\sum^{N}_{n=M}|a_{n}|^{2} +2\operatorname{Re}\sum^{N-1}_{n=M}\sum^{N}_{k>n}\overline{a}_{n}a_{k}f'(0)^{k-n}.
\end{equation}
Assume $f'(0)\ne 0$. For $k>n\ge M$ consider the function~$g=g_{k,n}$ defined by $g(w)=f^{k-n}(w)$, $w\in \mathbb{D}$. By part~(b) of Lemma~\ref{lem2.1}, there exists a constant~$0<r_{0}=r_{0}(f)<1$ such that $|g(w)|\le r_{0}^{-1} |f'(0)|^{k-n}|w|$ if $|w|<r_{0}$. Hence there exists a constant~$c(r_{0})>0$ such that
\begin{equation}\label{eq3.10}
\left|\frac{g(w)}{w}-g'(0)\right|\le c(r_{0})|f'(0)|^{k-n}|w|,\quad |w|\le \frac{r_{0}}{2}.
\end{equation}
Let $0<\varepsilon < r_{0}/2$ be a constant to be fixed later and assume $|f^{M}(z)|\le \varepsilon $. Taking $w=f^{n}(z)$ in \eqref{eq3.10} we obtain
$$
\left|\frac{f^{k}(z)}{f^{n}(z)}-f'(0)^{k-n}\right| \le c(r_{0})|f'(0)|^{k-n} |f^{n}(z)|\le \varepsilon c(r_{0})|f'(0)|^{k-n}, \quad k>n \geq M.
$$
By part (b) of Lemma 2.1, we also have $|f^{k}(z)|\le r_{0}^{-1}|f'(0)|^{k-M}\varepsilon $ and $|f^{n}(z)|\le r_{0}^{-1}|f'(0)|^{n-M}\varepsilon$ if $k \ge M$ and $n\ge M$. Then
\begin{equation*}
\begin{split}
&\left|\sum^{N-1}_{n=M} \sum^{N}_{k>n}\overline{a}_{n}a_{k}\left( (1-|f^{n}(z)|^{2})\frac{f^{k}(z)}{f^{n}(z)}-f'(0)^{k-n}\right)\right|\\*[5pt]
\le
&\sum^{N-1}_{n=M}\sum^{N}
_{k>n}|a_{n}| |a_{k}| (\varepsilon c(r_{0})|f'(0)|^{k-n}+r_{0}^{-2}\varepsilon^{2}|f'(0)|^{k+n-2M}).
\end{split}
\end{equation*}
As in the proof of Theorem \ref{theo1.2}, writting $j =k-n$ and applying Cauchy-Schwarz inequality, we find a constant $c(f)>0$ such that
\begin{align*}
&\sum^{N-1}_{n=M} \sum^{N}_{k>n}|a_{n}| |a_{k}| |f'(0)|^{k-n}\le
c(f)\sum^{N}_{n=M}|a_{n}|^{2},\\*[5pt]
&\sum^{N-1}_{n=M} \sum^{N}_{k>n}|a_{n}| |a_{k}| |f'(0)|^{k+n-2M}\le
c(f)\sum^{N}_{n=M}|a_{n}|^{2}.
\end{align*}
We deduce that there exists a constant $c(f,r_{0})>0$ such that
$$
\left|\sum^{N-1}_{n=M}\sum^{N}_{k>n}\overline{a}_{n}a_{k}\left( (1-|f^{n}(z)|^{2}) \frac{f^{k}(z)}{f^{n}(z)}-f'(0)^{k-n}\right)\right|\le c(f,r_{0})\varepsilon \sum^{N}_{n=M}|a_{n}|^{2}.
$$
Applying \eqref{eq3.8} and \eqref{eq3.9}, we deduce
$$
\left|\int_{\partial\mathbb{D}}|F(\xi)-F(z)|^{2}P(z,\xi)\,dm(\xi)-\int_{\partial\mathbb{D}} |F(\xi)|^{2}\,dm(\xi)\right|\le \varepsilon^{2} \sum^{N}_{n=M}|a_{n}|^{2}+2c(f,r_{0})\varepsilon \sum^{N}_{n=M}|a_{n}|^{2}.
$$
Recall that
$$
\int_{\partial\mathbb{D}}|F(\xi)|^{2}\,dm(\xi)\ge \frac{1-|f'(0)|}{1+|f'(0)|}\sum^{N}_{n=M}|a_{n}|^{2}.
$$
Choose $0<\varepsilon <r_{0} / 2 $ small enough so that $\varepsilon^{2}+2c(f,r_{0})\varepsilon <(1-|f'(0)|) (1+|f'(0)|)^{-1}/2$ and deduce
$$
\int_{\partial\mathbb{D}}|F(\xi)-F(z)|^{2} P(z,\xi)\,dm(\xi)\ge \frac{1}{2}\, \frac{1-|f'(0)|}{1+|f'(0)|}\sum^{N}_{n=M}|a_{n}|^{2}.
$$
This finishes the proof if $f'(0) \neq 0$. Assume now $f'(0)=0$.~Apply part~(c) of Lemma~\ref{lem2.1} to the inner function~$f^{k-n}$ to obtain $|f^{k}(z)|\le |f^{n}(z)|^{2^{k-n}}$, $k>n$. Let $ 0 < \varepsilon < 1$ be a (small) constant to be fixed later and assume $|f^M (z)| \leq \varepsilon$. If $k > n\ge M$, we have
$$
\frac{|f^k (z)|}{|f^n (z)|} \leq |f^n (z)|^{2^{k-n} - 1} \leq \varepsilon^{2^{k-n} - 1} \leq \varepsilon^{k-n}.
$$
Writing $j=k-n$ and applying Cauchy-Schwarz's inequality, we obtain
$$
\left|\sum^{N-1}_{n=M}\overline{a}_{n}\sum^{N}_{k>n}a_{k}(1-|f^{n}(z)|^{2})\frac{f^{k}(z)}{f^{n}(z)}\right|\le
\sum^{N-1}_{n=M}\sum^{N}_{k>n}|a_{n}| |a_{k}|\varepsilon^{k-n}\le \varepsilon (1-\varepsilon)^{-1} \sum^{N}_{n=M} |a_{n}|^{2}.
$$
As before, applying \eqref{eq3.8} and \eqref{eq3.9} we deduce
$$
\left|\int_{\partial \mathbb{D}}|F(\xi)-F(z)|^{2}P(z,\xi)\,dm(\xi)-\int_{\partial\mathbb{D}}|F(\xi)|^{2}\,dm(\xi)\right|\le (\varepsilon^{2}+2\varepsilon (1-\varepsilon)^{-1})\sum^{N}_{n=M}|a_{n}|^{2}.
$$
Since
$$
\int_{\partial\mathbb{D}}|F(\xi)|^{2}\,dm(\xi)=\sum^{N}_{n=M}|a_{n}|^{2}
$$
we only need to pick $\varepsilon >0$ small enough so that $\varepsilon^{2}+ 2 \varepsilon (1- \varepsilon)^{-1}<1/2$.
\end{proof}
Next auxiliary result is an easy consequence of Lemmas~\ref{lem3.1} and \ref{lem3.2}.
\begin{lemma}\label{lem3.3}
Let $f$ be an inner function with $f(0)=0$ which is not a rotation.~Then there exist two constants~$0 < \varepsilon =\varepsilon (f)< 1$ and $0 < c =c(f) < 1$ such that the following statement holds. Let $M<N$ be positive integers, $z\in\mathbb{D}$ satisfying $|f^{M}(z)|<\varepsilon$ and $\{a_{n}:M\le n\le N\}$ a set of complex numbers. Then there exists a set~$E=E(z,\{a_{n}\}, f)\subset \partial\mathbb{D}$ with $w(z,E)\ge c$ such that
$$
\operatorname{Re} \sum^{N}_{n=M}a_{n}f^{n}(\xi)\ge c \left(\sum_{n=M}^{N}|a_{n}|^{2}\right)^{1/2},\quad \xi\in E.
$$
\end{lemma}
\begin{proof}
Note that for any analytic function~$G\in \mathbb{H}^{2}(\mathbb{D})$ with $G(z)=0$, we have
\begin{equation}\label{eq3.11}
\int_{\partial\mathbb{D}}(\operatorname{Re}G(\xi))^{2} P(z,\xi)\,dm(\xi)=\frac{1}{2}\int_{\partial\mathbb{D}}|G(\xi)|^{2}P(z,\xi)\,dm(\xi).
\end{equation}
Let $0<r_{0}=r_{0}(f)<1$ and $0<c_{0}=c_{0}(f)<1$ be the constants given in Corollary~\ref{coro2.2} and let $\varepsilon_{0}=\varepsilon_{0}(f)>0$ be the constant appearing in Lemma~\ref{lem3.2}. Let $0<\varepsilon<\min \{\varepsilon_{0},r_{0}\}$ be a (small) constant to be fixed later. Lemma~\ref{lem3.2} and the identity \eqref{eq3.11} applied to the function~$G=F-F(z)$, where $F=\sum\limits_{n=M}^{N} a_{n}f^{n}$, give that
$$
\int_{\partial\mathbb{D}}(\operatorname{Re}F(\xi)-\operatorname{Re}F(z))^{2}P(z,\xi)\,dm(\xi)\ge c(f)\sum^{N}_{n=M}|a_{n}|^{2},
$$
where $c(f)=4^{-1}(1-|f'(0)|)(1+|f'(0)|)^{-1}$. Applying Lemma~\ref{lem3.1} to the function
$$
(\operatorname{Re}F)\|\mathrm{Re}\,F\|^{-1}_{\operatorname{BMO}(\partial\mathbb{D})} ,
$$
we find a constant~$\delta=\delta(f)>0$ and a set $E=E(z,\{a_{n}\}, f)\subset\partial\mathbb{D}$ with $w(z,E)\ge\delta$ such that
$$
\operatorname{Re}F(\xi)-\operatorname{Re}F(z)\ge\delta \|\mathrm{Re}\,F\|_{\operatorname{BMO}(\partial\mathbb{D})},\quad \xi\in E.
$$
By Corollary~\ref{coro1.3}, $\|\mathrm{Re}\, F\|^{2}_{\operatorname{BMO}(\partial\mathbb{D})}$ is comparable to $\sum\limits_{n=M}^{N}|a_{n}|^{2}$. Reducing $\delta>0$ if necessary, we can assume
\begin{equation}\label{eq3.12}
\operatorname{Re}F(\xi)-\operatorname{Re}F(z)\ge \delta\left(\sum^{N}_{n=M}|a_{n}|^{2}\right)^{1/2},\quad \xi\in E.
\end{equation}
We have $|f^{M}(z)|<\varepsilon$. Note that by Corollary~\ref{coro2.2}, we obtain $|f^{n}(z)|\le r_{0}^{-1}c_{0}^{n-M}\varepsilon$, $n\ge M$. Cauchy-Schwarz inequality yields
$$
|F(z)|\le \varepsilon c(c_{0},r_{0})\left(\sum^{N}_{n=M}|a_{n}|^{2}\right)^{1/2},
$$
where $c(c_{0},r_{0})$ is a constant depending on~$c_{0}$ and~$r_{0}$. Pick $0 < \varepsilon < \min \{ \varepsilon_0 , r_0\}$ sufficiently small such that $\varepsilon c(c_{0},r_{0}) \le \delta/2$. Taking $c =\delta/2$ the proof is completed.
\end{proof}
The next easy auxiliary result is stated for future reference.
\begin{lemma}\label{lem3.4}
Let $F\in\operatorname{BMO}(\partial\mathbb{D})$ and $0<c<1$. Then there exists a constant $\varepsilon=\varepsilon(c)>0$ such that if $I\subset \partial\mathbb{D}$ is an arc with $m ( \{ \xi\in I: \operatorname{Re}F(\xi)\ge c\|F\|_{\operatorname{BMO}(\partial\mathbb{D})} \} )\ge cm(I)$ and
\begin{equation}\label{3.99}
\frac{1}{m(I)}\int_{I}|F-F_{I}|^{2}\,dm\le \varepsilon\|F\|^{2}_{\operatorname{BMO}(\partial\mathbb{D})},
\end{equation}
then $\operatorname{Re}(F_{I})\ge 2^{-1}c\|F\|_{\operatorname{BMO}(\partial\mathbb{D})}$.
\end{lemma}
\begin{proof}
Denote $E=\{\xi\in I:\operatorname{Re}F(\xi)\ge c\|F\|_{\operatorname{BMO}(\partial\mathbb{D})}\}$.~We argue by contradiction.~Assume $\operatorname{Re}(F_{I})\le 2^{-1}c\|F\|_{\operatorname{BMO}(\partial\mathbb{D})}$. Then the estimate \eqref{3.99} gives
$$
\frac{c^{2}}{4}\|F\|^{2}_{\operatorname{BMO}(\partial\mathbb{D})}\frac{m(E)}{m(I)} \le \varepsilon \|F\|^{2}_{\operatorname{BMO}(\partial\mathbb{D})}
$$
and we deduce $m(E)\le 4\varepsilon c^{-2}m(I)$. Since $m(E)\ge cm(I)$, picking $0<\varepsilon < c^{3}/4$ one gets a contradiction.
\end{proof}
Given an arc~$I$ centered at $\xi\in\partial\mathbb{D}$ we denote by~$z(I)$ the point~$z(I)=(1-m(I))\xi$. Conversely, given a point~$z\in\mathbb{D}\backslash\{0\}$ let $I(z)$ be the arc in the unit circle such that $z(I(z))=z$. Given an arc~$I\subset\partial\mathbb{D}$ and a number~$0<c<1/m(I)$, we denote by~$cI$ the arc which has the same center as~$I$ and with $m(cI)=cm(I)$. Given an arc~$I\subset\partial\mathbb{D}$ we consider its dyadic decomposition~$\mathcal{D}(I)=\bigcup\limits_{n\ge 0}\mathcal{D}_{n}(I)$ where $\mathcal{D}_{n}(I)$ is the set of the $2^{n}$~pairwise disjoint subarcs of~$I$ of Lebesgue measure~$2^{-n}m(I)$.
\begin{lemma}\label{lem3.5}
Let $f$ be an inner function with $f(0)=0$ which is not a rotation. Then there exist constants~$0< \varepsilon =\varepsilon (f)<1$ and $0< c=c(f) < 1$ such that the following statement holds. Let $0<\gamma<1$ be a constant with $1-\gamma\le \varepsilon$, let $M<N$ be positive integers, $\{ a_{n}:M\le n\le N\}$ be complex numbers, and $z\in\mathbb{D}$ with $|f^{M}(z)|<\varepsilon$. Then there exist pairwise disjoint arcs~$\{I_{k}\}$ with $c^{-1}I_{k}\subset c^{-1}I(z)$ for any~$k$, such that
\begin{enumerate}
\item[(a)] For any $k=1,2,\dotsc$, we have $ \tau (1-\gamma)\le 1-|f^{N}(z(I_{k}))|\le 1-\gamma,$
where $\tau >0$ is a universal constant.
\item[(b)] $\sum m(I_{k})\ge c m(I(z))$.
\item[(c)] For any $k=1,2,\dotsc$, we have
$$
\frac{1}{|I_{k}|} \int_{I_{k}}\operatorname{Re}\sum^{N}_{n=M}a_{n}f^{n}\,dm\ge c \left(\sum^{N}_{n=M}|a_{n}|^{2}\right)^{1/2}.
$$
\end{enumerate}
\end{lemma}
\begin{proof}
Denote $F=\sum\limits_{n=M}^{N}a_{n}f^{n}$. Let $0<\varepsilon_{0}=\varepsilon_{0}(f)<1$, $0<c_{0}=c_{0}(f)<1$ be the constants given by Lemma~\ref{lem3.3}. Let $0<\varepsilon<\varepsilon_{0}(f)/2$ be a (small) constant to be fixed later. Assume $|f^{M}(z)|<\varepsilon$. Apply Lemma~\ref{lem3.3} to find a set~$E\subset\partial\mathbb{D}$ with $w(z,E)\ge c_{0}$ such that
\begin{equation}\label{eq3.13}
\operatorname{Re}F(\xi)\ge c_{0}\left(\sum^{N}_{n=M}|a_{n}|^{2}\right)^{1/2},\quad \xi\in E.
\end{equation}
Reducing $E$ if necessary, we can assume that there exists a constant~$c_{1}=c_{1}(c_{0})>0$ such that $E\subset c_{1}^{-1}I(z)$ and $m(E) \ge c_{1} m(I(z))$. Let $I_{0}=c_{1}^{-1}I(z)$ and consider its dyadic decomposition~$\mathcal{D}(I_{0})$. Fix $0<\gamma<1$ with $1-\gamma<\varepsilon$. Let $\mathcal{G}_{0}=\{I_{k}: k\ge 1\}$ be the collection of maximal dyadic arcs in~$\mathcal{D}(I_{0})$
such that
\begin{equation}\label{eq3.14}
|f^{N}(z_{k})| \geq \gamma.
\end{equation}
Here $z_{k}=z(I_{k})$, $k\ge 1$. By maximality, the arcs in the collection $\mathcal{G}_{0}$ are pairwise disjoint. Note also that by maximality and Schwarz's Lemma, there exists a universal constant~$\tau >0$ such that
\begin{equation}\label{eq3.15}
\tau (1- \gamma)\le 1-|f^{N}(z_{k})|\le 1-\gamma,\quad k\ge 1.
\end{equation}
This is the statement in (a). Since $f^{N}$ is inner, the union of $\{I_{k}:k\ge 1\}$ covers almost every point of~$I_{0}$. Consider the subcollection~$\mathcal{G}_{1} \subset\mathcal{G}_{0}$ of those dyadic arcs~$I_{k}\in\mathcal{G}_{0}$ such that
\begin{equation}\label{eq3.16}
m(E\cap I_{k})\ge c_{2} m(I_{k}).
\end{equation}
Here $c_{2}>0$ is a small constant to be fixed later. We will show that the subcollection~$\mathcal{G}_{1} $ satisfies the conditions (b) and (c) in the statement. Let $\mathcal{L}$ be the collection of indices~$k$ such that $I_{k}\in\mathcal{G}_{1}$. Note that
$$
c_{1}m(I(z))\le m(E)\le \sum_{k\notin\mathcal{L}} m(E\cap I_{k}) +\sum_{k\in\mathcal{L}}m(I_{k}).
$$
If $k\notin \mathcal{L}$, we have $m(E\cap I_{k})\le c_{2}m(I_{k})$. Hence the first sum is bounded by~$c_{2}c_{1}^{-1} m(I(z))$. Pick $c_{2}>0$ sufficiently small so that $c_{1}-c_{2}c_{1}^{-1}>c_{1}/2$ to deduce
\begin{equation}\label{eq3.17}
\sum_{k\in \mathcal{L}} m( I_{k} ) \ge \frac{c_{1}}{2} m(I(z)).
\end{equation}
Choosing $0<c<c_1 / 2$, the estimate~\eqref{eq3.17} gives the statement in (b). Note that $|f^{N}(z)|\le |f^{M}(z)|<\varepsilon$ and $|f^{N}(z_{k})|\ge \gamma$, for any positive integer $k$. By Schwarz's Lemma $\rho(z_{k},z)\ge \rho(\gamma,\varepsilon)$. Since $0<1-\gamma<\varepsilon$, taking $\varepsilon>0$ sufficiently small we deduce that $\rho (\gamma,\varepsilon)$ is close to~$1$ and then $m(I_{k}) \le c_3(\varepsilon) m(I_{0})$ for any $k \geq 1$, where $c_3 (\varepsilon)$ is tends to~$0$ as $\varepsilon$ tends to~$0$. If $c>0$ and $\varepsilon >0$ are taken sufficiently small we have $c^{-1} I_k \subset c^{-1} I(z)$ for any $k\geq 1$.
By Theorem~\ref{theo1.2}, there exists a constant~$c_4 = c_4 (f)>0$ such that for any~$k\ge 1$ we have
$$
\frac{1}{m(I_{k})} \int_{I_{k}}\left| F(\xi)- F_{I_k} \right|^{2}\,dm(\xi)\le c_4 \sum^{N}_{n=M}|a_{n}|^{2}(1-|f^{n}(z_{k})|^{2}).
$$
By part~(a) of Lemma~\ref{lem2.1}, there exists a constant~$0<c_{5}=c_{5}(\gamma,f)<1$ such that $1-|f^{n}(z_{k})|\le c_{5}^{N-n} (1-|f^{N}(z_{k})|)$, $n\le N$. Since $1-|f^{N}(z_{k})|\le 1-\gamma$, we obtain
$$
\frac{1}{m(I_{k})} \int_{I_{k}}\left|F(\xi)- F_{I_k}\right|^{2}\,dm(\xi)\le
2c_4 (1-\gamma)\sum^{N}_{n=M}|a_{n}|^{2}c_{5}^{N-n}.
$$
Recall that $m(E\cap I_{k})\ge c_{2} m(I_{k})$ if $k\in\mathcal{L}$. Then if $1-\gamma$ is sufficiently small, by Lemma~\ref{lem3.4}, the estimate~\eqref{eq3.13} implies that
$$
\frac{1}{m(I_{k})} \int_{I_{k}}\operatorname{Re}F\,dm \ge \frac{c_{0}}{2}\left(\sum^{N}_{n=M}|a_{n}|^{2}\right)^{1/2}.
$$
Choosing $0<c<c_0 / 2$, last estimate gives the statement in (c).This finishes the proof.
\end{proof}
Our next auxiliary result is the building block of the main construction in the proof of Theorem~\ref{theo1.5}.
\begin{lemma}\label{lem3.7}
Let $f$ be an inner function with $f(0)=0$ which is not a rotation. Then there exist constants~$0<\varepsilon=\varepsilon(f)<1$ and $0<c=c(f)<1$ such that the following statement holds. Let $0<\gamma<1$ be a constant with $1-\gamma\le \varepsilon$, let $\{a_{n}\}$ be a sequence of complex numbers, let $M_{1}<N_{1}<M_{2}<N_{2}$ be positive integers and
$$
F_{i}=\sum^{N_{i}}_{n=M_{i}}a_{n}f^{n},\quad i=1,2.
$$
Let $I\subset \partial\mathbb{D}$ be an arc such that $c(1-\gamma)< 1-|f^{N_{1}}(z(I))|\le 1-\gamma$ and $|f^{M_{2}}(z(I))|\le \varepsilon$. Then there exists an arc~$J$ with $c^{-1}J\subset c^{-1}I$, such that
\begin{gather}
c (1-\gamma)\le 1-|f^{N_{2}}(z(J))|\le 1-\gamma,\label{eq3.18}\\
\frac{1}{|J|} \int_{J}\operatorname{Re}F_{2}\,dm \ge c\left(\sum^{N_{2}}_{n=M_{2}}|a_{n}|^{2}\right)^{1/2},\label{eq3.19}\\
\frac{1}{|J|} \int_{J}\operatorname{Re}F_{1}\,dm \ge \frac{1}{|I|}\int_{I}\operatorname{Re}F_{1}\,dm-
c^{-1}(1-\gamma)^{1/2}
\left(\sum^{N_{1}}_{n=M_{1}}|a_{n}|^{2}t^{N_{1}-n}\right)^{1/2},\label{eq3.20}
\end{gather}
where $t=(1+\gamma)^{-1} (1+|f'(0)|\gamma)<1$.
\end{lemma}
\begin{proof}
Let $\varepsilon_0>0$ and $c_0 >0$ be the constants appearing in Lemma~\ref{lem3.5}. Let $0 < \varepsilon < \varepsilon_0$ be a constant to be fixed later. Let $0<\gamma<1$ with $1-\gamma<\varepsilon$. Apply Lemma~\ref{lem3.5} to the function $F_2$ to find arcs~$J_{k}$ with $c_{0}^{-1}J_{k}\subset c_{0}^{-1}I$, such that $\tau (1-\gamma)\le 1-|f^{N_{2}}(z(J_{k}))|\le 1-\gamma$, $\sum m(J_{k})\ge c_{0} m(I)$ and
$$
\frac{1}{|J_{k}|}\int_{J_{k}}\operatorname{Re} F_{2}\,dm\ge c_{0}\left( \sum^{N_{2}}_{n=M_{2}}|a_{n}|^{2}\right)^{1/2}.
$$
Here $\tau>0$ is the universal constant given by part (a) of Lemma~\ref{lem3.5}. The constant $c$ will satisfy $0<c<c_0$ and the arc $J$ will be one of the arcs of the family $\{J_k \}$. Hence estimates \eqref{eq3.18} and \eqref{eq3.19} will follow from the construction. Theorem~\ref{theo1.2} gives that there exists a constant $c_1=c_1(f)>0$ such that
\begin{equation}\label{3.100}
\frac{1}{c_{0}^{-1}m(I)}\int_{c_{0}^{-1}I}|\mathrm{Re}\,F_{1}- \mathrm{Re}\,F_{1} (z(c_{0}^{-1} I))|^{2}\,dm\le c_1 \sum^{N_{1}}_{n=M_{1}}|a_{n}|^{2} (1-|f^{n}(z(c_{0}^{-1} I))|^{2}).
\end{equation}
Note that
\begin{equation*}
\begin{split}
| (\operatorname{Re}F_{1})_{I} - \mathrm{Re}\,F_{1} (z(c_{0}^{-1} I)) | &\leq \frac{1}{m(I)} \int_I |\mathrm{Re}\,F_{1}- \mathrm{Re}\,F_{1} (z(c_{0}^{-1} I))| dm \\
&\leq \left( \frac{1}{ m(I)} \int_{ I} |\mathrm{Re}\,F_{1}- \mathrm{Re}\,F_{1} (z(c_{0}^{-1} I))|^2 dm
\right)^{1/2}
\end{split}
\end{equation*}
Applying \eqref{3.100} we find a constant $c_2 = c_2 (f, c_0) >0$ such that
\begin{equation}\label{3.101}
\frac{1}{m(I)}\int_{c_{0}^{-1}I}|\mathrm{Re}\,F_{1}- (\operatorname{Re}F_{1})_{I} |^{2}\,dm\le c_2 \sum^{N_{1}}_{n=M_{1}}|a_{n}|^{2} (1-|f^{n}(z(c_{0}^{-1} I))|^{2}).
\end{equation}
By Schwarz's Lemma, there exists a constant $c_3 = c_3 (c_0) >0$ such that $1-|f^{n}(z(c_{0}^{-1} I ))|^{2} \leq c_3 (1-|f^{n}(z(I))|^{2})$ for any $n \geq 1$. Since $1-|f^{N_{1}}(z(I))|\le 1-\gamma$, part~(a) of Lemma~\ref{lem2.1} provides a constant~$0<c_4<1$ such that $1-|f^{n}(z(I))|\le c_4^{N_{1}-n}(1-\gamma)$, $n\le N_{1}$. Actually, according to Lemma~\ref{lem2.1} one can take $c_4=(1+\gamma)^{-1}(1+|f'(0)|\gamma)$. Applying \eqref{3.101} we obtain a constant $c_5 = c_5 (c_0 , f) >0$ such that
\begin{equation}\label{3.102}
\frac{1}{m(I)}\int_{c_{0}^{-1}I}|\mathrm{Re}\,F_{1}- (\operatorname{Re}F_{1})_{I} |^{2}\,dm\le c_5 (1- \gamma) \sum^{N_{1}}_{n=M_{1}}|a_{n}|^{2} c_4^{N_{1}-n} .
\end{equation}
Let $K>0$ be a large constant to be fixed later. Let $\mathcal{G}$ be the subcollection of those arcs~$J_{k}$ such that
$$
(\operatorname{Re}F_{1})_{J_{k}}\le (\operatorname{Re}F_{1})_{I}-K(1-\gamma)^{1/2} \left(\sum^{N_{1}}_{n=M_{1}}|a_{n}|^{2} c_4^{N_{1}-n}\right)^{1/2}.
$$
Note that for any $J_k \in \mathcal{G}$ we have
$$
\frac{1}{m(J_k)} \int_{J_k} | \operatorname{Re}F_{1} - (\operatorname{Re}F_{1})_{I}| dm \geq |(\operatorname{Re}F_{1})_{J_k} - (\operatorname{Re}F_{1})_{I}| \geq K(1-\gamma)^{1/2} \left(\sum^{N_{1}}_{n=M_{1}}|a_{n}|^{2} c_4^{N_{1}-n}\right)^{1/2} .
$$
Adding over $k$, we obtain
$$
\sum_{J_{k}\in\mathcal{G}} \int_{J_k} | \operatorname{Re}F_{1} - (\operatorname{Re}F_{1})_{I}| dm \geq K(1-\gamma)^{1/2} \left(\sum^{N_{1}}_{n=M_{1}}|a_{n}|^{2} c_4^{N_{1}-n}\right)^{1/2} \sum_{J_{k}\in\mathcal{G}}m(J_{k}) .
$$
Applying \eqref{3.102} we deduce
$$
\sum_{J_{k}\in\mathcal{G}}m(J_{k})\le \frac{c_5^{1/2}}{K} m(I).
$$
Since $\sum m(J_{k})\ge c_{0}m(I)$, taking $K>0$ large enough so that $c_5^{1/2} K^{-1} < c_0 / 2$, we deduce
$$
\sum_{J_{k}\notin\mathcal{G}}m(J_{k})\ge \frac{c_{0}}{2}m(I)
$$
and one can take as $J$ in the statement, any of the arcs~$J_{k}\notin\mathcal{G}$.
\end{proof}
We are now ready to prove Theorem\ref{theo1.5}.
\begin{proof}[Proof of Theorem~\ref{theo1.5}]
Since $\sum a_{n}f^{n}$ converges at a set of positive measure, Theorem~\ref{theo1.1} gives that $\sum |a_{n}|^{2}<\infty$. Hence the function~$F=\sum a_{n}f^{n}$ is in~$\operatorname{BMOA}$. The plan of the proof is to find constants~$A,B>0$, a sequence of arcs~$\{I_{k}\}$ contained in~$I$ and a sequence of positive integers~$\{M_{k}\}$ tending to infinity such that
\begin{equation}\label{eq3.22}
\frac{1}{|I_{k}|} \int_{I_{k}}\operatorname{Re}F\,dm\ge A\sum^{M_{k}}_{n=1}|a_{n}|-B.
\end{equation}
It is clear that \eqref{eq3.22} finishes the proof. The construction of the arcs~$\{I_{k}\}$ and the sequence~$\{M_{k}\}$ is performed inductively. We start using an idea of Weiss (\cite{We}). Let $T'<T$ be two (large) positive integers to be fixed later such that $T/T'$ is integer. Split the sum~$F=\sum a_{n}f^{n}$ into blocks of length~$T$, that is, $F=\sum\limits_{k\ge 0}G_{k}$, where
$$
G_{k}=\sum^{T-1}_{n=0}a_{kT+n}f^{kT+n},\quad k=0,1,2,\dotsc
$$
Next split $G_{2k}$ into successive blocks of length~$T'$ and pick the subblock for which the sum of the modulus of the coefficients is the least. In other words, pick $\mathcal{G}_{k}$ a subset of $T'$ consecutive integers in the interval~$[2kT,(2k+1)T)$ such that
$$
\sum_{n\in \mathcal{G}_{k}}|a_{n}|\le \sum_{\ell \in \mathcal{G}}|a_{\ell}|
$$
for any subset~$\mathcal{G}$ of $T'$ consecutive integers in~$[2kT,(2k+1)T)$. Since the number of disjoint subblocks of length $T'$ in~$[2kT,(2k+1)T)$ is $T/T'$ we deduce
\begin{equation}\label{eq3.23}
\sum_{n\in \mathcal{G}_{k}}|a_{n}| \le \frac{T'}{T}\sum^{T-1}_{n=0}|a_{2kT+n}|.
\end{equation}
Each set of indices~$\mathcal{G}_{k}$, $k\ge 1$, has $T'$~elements and the corresponding block
$$
S_{k}=\sum_{n\in\mathcal{G}_{k}}a_{n}f^{n}
$$
will be called a short block. The long blocks are defined as the blocks between two short blocks. More concretely
$$
L_{1}=\sum^{N_{1}}_{n=0}a_{n}f^{n},
$$
where $N_{1}+1$ is the first index in $\mathcal{G}_{1}$ and
$$
L_{k}=\sum^{N_{k}}_{n=M_{k}}a_{n}f^{n},\quad k>1,
$$
where $M_{k}-1$ is the largest index in~$\mathcal{G}_{k-1}$ and $N_{k}+1$ is the smallest index in~$\mathcal{G}_{k}$. Note that each short block has $T'$~terms while the number of terms in a long block is between $T$ and $3T$. Note also that if $\mathcal{L}_{k}=\{n\in\mathbb{Z}: M_{k}\le n\le N_{k}\}$ denotes the set of indices appearing in the long block~$L_{k}$, then \eqref{eq3.23} implies
\begin{equation}\label{eq3.24}
\sum^{L}_{k=1}\sum_{n\in \mathcal{L}_{k}}|a_{n}|\ge \left(1-\frac{T'}{T}\right) \sum^{N_{L}}_{n=1}|a_{n}|,\quad L=1,2,\dotsc
\end{equation}
The idea is that \eqref{eq3.24} will imply that the short blocks are irrelevant and the construction of the arcs~$\{I_{k}\}$ and the sequence of integers~$\{M_{k}\}$ verifying \eqref{eq3.22} will depend on the long blocks. Moreover, the fact that two long blocks are separated by a short block will provide a sort of independence between the long blocks.
Let $s_{k}$ denote the $\ell^{2}$-norm of the coefficients in the long block~$L_{k}$, that is,
$$
s_{k}^{2}=\sum_{n\in \mathcal{L}_{k}} |a_{n}|^{2},\quad k=1,2,\dotsc
$$
Let $\varepsilon=\varepsilon(f)>0$ and $c=c(f)>0$ be the constants given by Lemma~\ref{lem3.7}. Let $0<\gamma=\gamma(f)<1$ be a constant to be fixed later satisfying $1-\gamma<\varepsilon$. Let $D(0,R)$ denote the disc centered at the origin of radius $R$. By the Denjoy--Wolff Theorem, $f^{n}$ tends to~$0$ uniformly on compact sets of~$\mathbb{D}$. The integer~$T'$ will be taken large enough so that
\begin{equation}\label{eq3.25}
f^{n} ( D(0, 1- c (1- \gamma) )) \subset D(0, \varepsilon), \quad n \geq T' .
\end{equation}
Fix an arc $I^{*}\subset I$ such that $c^{-1}I^{*}\subset I$. The construction starts with the first long block~$L_{k_{0}}$ such that $|f^{M_{k_{0}}}(z(I^{*}))|<\varepsilon$. Without loss of generality we can assume that the constant $c$ is smaller than the constants appearing in Lemma~\ref{lem3.5}. Apply Lemma~\ref{lem3.5} to find an arc~$I_{k_{0}} $ with $c^{-1}I_{k_{0}}\subset c^{-1}I^{*}\subset I$ such that
\begin{gather*}
c(1-\gamma)\le 1-|f^{N_{k_{0}}} (z(I_{k_{0}}))|\le 1-\gamma,\\*[5pt]
\frac{1}{m(I_{k_{0}})}\int_{I_{k_{0}}}\operatorname{Re} L_{k_{0}}\,dm\ge c s_{k_{0}}.
\end{gather*}
Assume by induction that we have constructed arcs~$I_{k_{0}},I_{k_{0}+1},\dotsc, I_{k}$, $k \geq k_{0}$, with $c^{-1}I_{j}\subset c^{-1}I_{j-1}$, $j=k_{0}+1,\dotsc, k$, such that the following two conditions hold
\begin{gather}
c(1-\gamma)\le 1-|f^{N_{k}}(z(I_{k}))|\le 1-\gamma,\label{eq3.26}\\*[5pt]
\frac{1}{m(I_{k})}\int_{I_{k}}\operatorname{Re}\left(\sum^{k}_{j=k_{0}}L_{j}\right)\,dm\ge c\sum^{k}_{j=k_{0}}s_{j}
-c^{-1}(1-\gamma)^{1/2}\sum^{k-1}_{\ell=k_{0}}\left(\sum^{\ell}_{j=k_{0}}t^{N_{\ell}-N_{j}}s_{j}^{2}\right)^{1/2}.\label{eq3.27}
\end{gather}
Here $0<t<1$ is the constant appearing in Lemma~\ref{lem3.7} and when $k=k_0$, we replace the right hand side term of \eqref{eq3.27} by $c s_{k_0}$. Recall that two different long blocks are separated by a short block which has $T'$ terms. Hence the estimates~\eqref{eq3.25} and \eqref{eq3.26} give that $|f^{M_{k+1}}(z(I_{k}))|\le \varepsilon$. Apply Lemma~\ref{lem3.7} to $F_{1}=\sum\limits_{j=k_{0}}^{k}L_{j}$ and $F_{2}=L_{k+1}$, to find an arc~$I_{k+1}$ with $c^{-1}I_{k+1}\subset c^{-1}I_{k}$ such that
\begin{gather}
c (1-\gamma)\le 1-|f^{N_{k+1}}(z(I_{k+1}))|\le 1-\gamma,\nonumber\\*[5pt]
\frac{1}{m(I_{k+1})}\int_{I_{k+1}}\operatorname{Re}L_{k+1}\,dm\ge c s_{k+1},\label{eq3.28}\\*[5pt]
\frac{1}{m(I_{k+1})}\int_{I_{k+1}}\operatorname{Re}\sum^{k}_{j=k_{0}}L_{j}\,dm\ge \frac{1}{m(I_{k})}\int_{I_{k}}
\operatorname{Re}\sum^{k}_{j=k_0}L_{j}\,dm-
c^{-1}(1-\gamma)^{1/2}\left(\sum^{k}_{j=k_{0}}t^{N_{k}-N_{j}}s_{j}^{2}\right)^{1/2}.\label{eq3.29}
\end{gather}
Then \eqref{eq3.29} and the induction assumption~\eqref{eq3.27} give
$$
\frac{1}{m(I_{k+1})}\int_{I_{k+1}}\operatorname{Re}\left(\sum^{k}_{j=k_{0}}L_{j}\right)\,dm\ge c \sum^{k}_{j=k_{0}}s_{j}-c^{-1}(1-\gamma)^{1/2}\sum^{k}_{\ell=k_{0}}\left(\sum^{\ell}_{j=k_{0}}t^{N_{\ell}-N_{j}}s_{j}^{2}\right)^{1/2}.
$$
Applying \eqref{eq3.28} we deduce
$$
\frac{1}{m(I_{k+1})}\int_{I_{k+1}}\operatorname{Re}\left(\sum^{k+1}_{j=k_{0}}L_{j}\right)\,dm\ge c \sum^{k+1}_{j=k_{0}}s_{j}- c^{-1}(1-\gamma)^{1/2}\sum^{k}_{\ell=k_{0}}\left(\sum^{\ell}_{j=k_{0}}t^{N_{\ell}-N_{j}}s_{j}^{2}\right)^{1/2}.
$$
This concludes the inductive step.
Note that for any $k>k_{0}$, we have
$$
\sum^{k-1}_{\ell=k_{0}}\left(\sum^{\ell}_{j=k_{0}}t^{N_{\ell}-N_{j}}s_{j}^{2}\right)^{1/2}\le \sum^{k-1}_{\ell=k_{0}}\sum^{\ell}_{j=k_{0}}t^{(N_{\ell}-N_{j})/2}s_{j}=
\sum^{k-1}_{j=k_{0}}s_{j }\sum^{k-1}_{\ell=j}t^{(N_{\ell}-N_{j})/2}\le (1-t^{1/2})^{-1}
\sum^{k-1}_{j=k_{0}}s_{j}.
$$
Pick $0< \gamma< 1$ sufficiently close to~$1$ so that $c^{-1}(1-\gamma)^{1/2}(1-t^{1/2})^{-1}\le c/2$. The estimate~\eqref{eq3.27} gives
\begin{equation}\label{eq3.30}
\frac{1}{m(I_{k})}\int_{I_{k}}\operatorname{Re}\left(\sum^{k}_{j=k_{0}}L_{j}\right)\,dm\ge \frac{c}{2}\sum^{k}_{j=k_{0}}s_{j}.
\end{equation}
Recall that each long block has at most $3T$~terms. Hence Cauchy-Schwarz's inequality gives
$$
\sum_{n\in \mathcal{L}_{j}}|a_{n}|\le s_{j}(3T)^{1/2},\quad j=1,2,\dotsc
$$
Then apply \eqref{eq3.30} to deduce
\begin{equation}\label{eq3.31}
\frac{1}{m(I_{k})} \int_{I_{k}}\operatorname{Re} \sum^{k}_{j=k_{0}} (L_{j}+S_{j})\,dm\ge \frac{c}{2} (3T)^{-1/2}\sum^{k}_{j=k_{0}}\sum_{n\in\mathcal{L}_{j}}|a_{n}|-\sum^{k}_{j=k_{0}}\sum_{n\in\mathcal{G}_{j}}|a_{n}|,\quad k\ge k_{0}.
\end{equation}
Let $\mathcal{A}_{k}$ be the set of indices appearing in $\sum\limits_{j=k_{0}}^{k}(L_{j}+S_{j})$, that is, $\mathcal{A}_{k}=\bigcup\limits_{j=k_{0}}^{k}(\mathcal{L}_{j}\cup\mathcal{G}_{j})$, $k\ge k_{0}$. Applying \eqref{eq3.31}, \eqref{eq3.23} and \eqref{eq3.24} we obtain
$$
\frac{1}{m(I_{k})}\int_{I_{k}}\operatorname{Re} \sum^{k}_{j=k_{0}} (L_{j}+S_{j})\,dm \ge \frac{c}{2} (3T)^{-1/2}\left(1-\frac{2T'}{T}\right)\sum_{n\in\mathcal{A}_{k}}|a_{n}|-\frac{2T'}{T}\sum_{n\in\mathcal{A}_{k}}|a_{n}|.
$$
We now choose $T$ so that $T'T^{-1/2}$ is sufficiently small so that
$$
\frac{c}{2} (3T)^{-1/2}\left(1-\frac{2T'}{T}\right)-\frac{2T'}{T}\ge \frac{c}{4}T^{-1/2}.
$$
Then
\begin{equation}\label{eq3.32}
\frac{1}{m(I_{k})}\int_{I_{k}}\operatorname{Re} \sum^{k}_{j=k_{0}}(L_{j}+S_{j})\,dm\ge \frac{c}{4} T^{-1/2}\sum_{n\in\mathcal{A}_{k}}|a_{n}|.
\end{equation}
As observed previously, $|f^{M_{k+1}}(z(I_{k}))|\le \varepsilon$. By Corollary~\ref{coro2.2} there exists a constant~$0<c_{1}<1$ such that $|f^{n}(z(I_{k}))|\le c_{1}^{n-M_{k+1}}$, $n\ge M_{k+1}$. Then
$$
\sum_{n\ge M_{k+1}}|a_{n}| |f^{n}(z(I_{k}))|\le (1-c_{1}^2)^{-1/2}\left( \sum_{n\ge M_{k+1}}|a_{n}|^{2}\right)^{1/2}.
$$
Since there exists a constant~$c_2 = c_2(f)>0$ such that
$$
\left|\frac{1}{m(I_{k})}\int_{I_{k}}\sum_{n\ge M_{k+1}}a_{n}f^{n}\,dm-\sum_{n\ge M_{k+1}}a_{n}f^{n}(z(I_{k}))\right|\le c_2 \left(\sum_{n\ge M_{k+1}}|a_{n}|^{2}\right)^{1/2},
$$
the estimate \eqref{eq3.32} gives
$$
\frac{1}{m(I_{k})}\int_{I_{k}}\operatorname{Re}\sum^{\infty}_{n=1}a_{n}f^{n}\,dm\ge \frac{c}{4} T^{-1/2}\sum_{n\in\mathcal{A}_{k}}|a_{n}|-(c_2 +(1-c_{1}^2)^{-1/2})\left(\sum_{n\ge 1}|a_{n}|^{2}\right)^{1/2}.
$$
This gives \eqref{eq3.22} and finishes the proof.
\end{proof}
\section{Other function spaces}\label{sec4}
Let $A(\mathbb{D})$ denote the disc algebra, that is, the space of analytic functions in~$\mathbb{D}$ which extend continuously to $\overline{\mathbb{D}}=\{z\in\mathbb{C}: |z|\le 1\}$. The only inner functions which belong to~$A(\mathbb{D})$ are the finite Blaschke products. See ~\cite[p.~72]{Ga}. Our next result is an easy consequence of Theorem~\ref{theo1.5}.
\begin{theorem}\label{theo4.1}
Let $f$ be an inner function with $f(0)=0$ which is not a rotation and let $\{a_{n}\}$ be a sequence of complex numbers not identically zero. Then $\sum\limits_{n=1}^{\infty} a_{n}f^{n}\in A(\mathbb{D})$ if and only if $f$ is a finite Blaschke product and $\sum\limits_{n=1}^{\infty} |a_{n}|<\infty$.
\end{theorem}
\begin{proof}
Write $F=\sum\limits_{n=1}^{\infty} a_{n}f^{n}$. The sufficiency follows from the uniform convergence of the partial sums~$\sum\limits_{n=1}^{N}a_{n}f^{n}$. We now prove the necessity. Assume $F\in A(\mathbb{D})$. Since $\sup \{|F(\xi)|: \xi\in \partial \mathbb{D}\}<\infty$, Theorem~\ref{theo1.5} gives that $\sum\limits_{n=1}^{\infty} |a_{n}|<\infty$. We now show that $f$ is a finite Blaschke product arguing by contradiction. Assume that $f$ does not extend analytically at any neighbourhood of the point~$\xi\in\partial \mathbb{D}$. By Frostman's Theorem (see~\cite[p.~77]{Ga}) there exists a set~$E\subset\mathbb{D}$ of logarithmic capacity zero such that for any $\alpha\in \mathbb{D}\backslash E$ one can find a sequence~$\{z_{k} \} = \{ z_{k}(\alpha) \}$, of points in~$\mathbb{D}$ converging to~$\xi$, such that $f(z_{k})=\alpha$ for any~$k$.
Then
$$
F(z_{k})=\sum^{\infty}_{n=1}a_{n}f^{n-1}(\alpha),\quad k=1,2,\dotsc
$$
Here $f^0 (\alpha) = \alpha , \alpha \in \mathbb{D}$. Since $\{z_{k}\}$ converges to~$\xi$ and $F \in A(\mathbb{D})$, for any $\alpha\in\mathbb{D}\backslash E$ we have
$$
\sum_{n=1}^{\infty} a_{n}f^{n-1}(\alpha)=F(\xi).
$$
Hence the function $\sum\limits^{\infty}_{n=1}a_{n}f^{n-1}$ is constant in $\mathbb{D}$. Since it vanishes at the origin, we deduce\linebreak $\sum\limits_{n=1}^{\infty} a_{n}f^{n-1}\equiv 0$. Hence $\sum\limits_{n=1}^{\infty} a_{n}f^{n}\equiv 0$. By Lemma \ref{lem2.3}, $a_n=0$ for any $n \geq 1$, which gives the contradiction.
\end{proof}
Let $\mathcal{D}$ denote the Dirichlet space of analytic functions~$f$ in~$\mathbb{D}$ such that
$$
\|f\|^{2}_{\mathcal{D}}=\int_{\mathbb{D}}|f'(z)|^{2}\,dA(z)<\infty,
$$
where $dA(z)$ denotes the area measure. Note that $\|f\|^{2}_{\mathcal{D}}$ can be understood as the area of the image domain~$f(\mathbb{D})$, counting multiplicities. Hence the only inner functions which belong to the Dirichlet space are finite Blaschke products. Our next result is a description of linear combinations of iterates of an inner function which belong to the Dirichlet space.
\begin{theorem}\label{theo4.2}
Let $f$ be a finite Blaschke product with $N>1$~zeros and with $f(0)=0$. Let $\{a_{n}\}$ be a sequence of complex numbers with $\sum\limits_{n=1}^{\infty}|a_{n}|^{2}<\infty$ and $F=\sum\limits_{n=1}^{\infty}a_{n}f^{n}$. Then $F \in \mathcal{D}$ if and only if $\sum\limits^{\infty}_{n=1}|a_{n}|^{2}N^{n} < \infty$. Moreover there exists a universal constant $C=C(f)>0$ such that
$$
C^{-1}\sum^{\infty}_{n=1}|a_{n}|^{2}N^{n}\le \int_{\mathbb{D}}|F'(z)|^{2}\,dA(z)\le C\sum^{\infty}_{n=1}|a_{n}|^{2}N^{n}.
$$
\end{theorem}
\begin{proof}
Note that
$$
\int_{\mathbb{D}}|F'(z)|^{2}\,dA(z)=\sum^{\infty}_{n=1}|a_{n}|^{2}b_{n,n}+2\operatorname{Re} \sum^{\infty}_{n=1}\overline{a}_{n}\sum_{k>n}a_{k}b_{k,n},
$$
where
$$
b_{k,n}=\int_{\mathbb{D}}\overline{(f^{n})'(z)} (f^{k})'(z)\,dA(z),\quad k,n=1,2,\dotsc
$$
Since $f^{n}$ is a Blaschke product with $N^{n}$~zeros, the area counting multiplicities, of~$f^{n}(\mathbb{D})$ is~$\pi N^{n}$. Then
$$
\int_{\mathbb{D}}|(f^{n})'(z)|^{2}\,dA(z)=\pi N^{n}.
$$
On the other hand, the change of variables formula with multiplicities (see~\cite[p.~122]{EG}) gives
$$
b_{k,n}=\int_{\mathbb{D}}|(f^{n})'(z)|^{2} (f^{k-n})' (f^{n}(z))\,dA(z)=
\int_{\mathbb{D}} N^{n}(f^{k-n})'(w)\,dA(w),\quad k>n.
$$
The mean value property gives $b_{k,n}=\pi N^{n} f'(0)^{k-n}$, $k>n$. Hence
$$
\frac{1}{\pi}\int_{\mathbb{D}} |F'(z)|^{2}\,dA(z)=\sum^{\infty}_{n=1}|a_{n}|^{2}N^{n}+2\operatorname{Re} \sum^{\infty}_{n=1} \overline{a}_{n}N^{n}\sum_{k>n}a_{k} f'(0)^{k-n} =\vec{a}^{\,t} T \vec{a},
$$
where $\vec{a}$ denotes the vector~$(a_{n} N^{n/2})_{n\ge 1}$ and $T$ is the Toeplitz matrix whose entries are
$$
t_{n,k}=(f'(0) N^{-1/2})^{k-n}, \quad k \geq n; \qquad t_{n,k}=(\overline{f'(0)} N^{-1/2})^{n-k}, \quad n \geq k.
$$
Consider the symbol
$$
t(\xi)=\sum^{\infty}_{n=-\infty} t_{n,0} \xi^{n}= \frac{1-|f'(0)|^{2}N^{-1}} {|1- f'(0) N^{-1/2}\xi|^{2}},\quad \xi\in\partial\mathbb{D}.
$$
It is well known that $T$ diagonalizes and its eigenvalues are between the minimum and the maximum of~$t$. See~\cite{BG}. Since $C(f,N)^{-1}\le t(\xi)\le C(f,N)$, a.e.\ $\xi\in\partial\mathbb{D}$, where
$$
C(f,N)=\frac{1+|f'(0)|N^{-1/2}}{1-|f'(0)|N^{-1/2}},\quad \text{a.e.\ } \xi\in\partial\mathbb{D},
$$
we deduce that
$$
C(f,N)^{-1}\sum^{\infty}_{n=1}|a_{n}|^{2} N^{n}\le \frac{1}{\pi}\int_{\mathbb{D}} |F'(z)|^{2}\,dA(z) \le C(f,N) \sum^{\infty}_{n=1} |a_{n}|^{2} N^{n}.
$$
\end{proof}
Let $\mathcal{B}$ denote the Bloch space of analytic functions~$f$ in~$\mathbb{D}$ such that
$$
\|f\|_{\mathcal{B}}=\sup_{z\in\mathbb{D}} (1-|z|^{2}) |f'(z)|<\infty.
$$
It is well known that a lacunary series is in the Bloch space if and only if its coefficients are uniformly bounded. See~\cite{ACP}. In our setting this condition is still sufficient but we will see that it is not necessary.
\begin{theorem}\label{theo4.3}
Let $f$ be an inner function with $f(0)=0$ which is not a rotation and let $\{a_{n}\}$ be a bounded sequence of complex numbers. Then $F=\sum\limits_{n=1}^{\infty} a_{n}f^{n}\in\mathcal{B}$ and there exists a constant $C=C(f)>0$ such that $\|F\|_{\mathcal{B}}\le C \sup\limits_{n \geq 1}|a_{n}|$.
\end{theorem}
\begin{proof}
Let $0<r_{0}=r_{0}(f)<1$ and $0<c_{0}=c_{0}(f)<1$ be the constants given by Corollary~\ref{coro2.2}. Since $f^{n}$ tends to~$0$ uniformly on compacts of~$\mathbb{D}$, for any $z\in \mathbb{D}$ we can pick $N(z)$ to be the minimum positive integer $n$ such that $|f^{n}(w)|\le r_{0}$ for any $w\in \mathbb{D}$ with $\rho(w,z)\le 1/2$. Corollary~\ref{coro2.2} gives
\begin{equation}\label{eq4.1}
\sup\{ |f^{n}(w)|: \rho(w,z)\le r_{0}\}\le c_{0}^{n-N(z)},\quad n\ge N(z), \quad z \in \mathbb{D}.
\end{equation}
By Cauchy's estimate, there exist a universal constant $c_{1}>0$ such that
$$
(1-|z|)| (f^{n})'(z)|\le c_{1}c_{0}^{n-N(z)},\quad n\ge N(z), \quad z \in \mathbb{D}.
$$
Hence
\begin{equation}\label{eq4.2}
(1-|z|)\sum_{n\ge N(z)}|(f^{n})'(z)|\le c_{1} (1-c_{0})^{-1}, \quad z \in \mathbb{D}.
\end{equation}
Let $M= N(z) - 1$. Note that there exists a point $w$ with $\rho (w, z) \leq 1/2 $ such that $|f^M (w)| \geq r_0$. According to part~(a) of Lemma~\ref{lem2.1}, there exists a constant~$0<c_{2}<1$ such that $1-|f^{n}(w)|\le c_{2}^{M-n}$ for any $n < N(z)$. By Schwarz's Lemma there exists a constant $c_3 >0$ such that $1-|f^{n}(z)|\le c_3 c_{2}^{M-n}$ for any $n < N(z)$. Since $(1-|z|^{2})|(f^{n})'(z)|\le 1 - |f^{n}(z)|^{2}$, $z\in \mathbb{D}$, we deduce
\begin{equation}\label{eq4.3}
(1-|z|^{2})\sum_{n=1}^{M} |(f^{n})'(z)|\le c_3 (1-c_{2})^{-1}, \quad z \in \mathbb{D}.
\end{equation}
The estimates \eqref{eq4.2} and \eqref{eq4.3} give that $\|F\|_{\mathcal{B}}\le c\sup\limits_{n}|a_{n}|$ with $c=c_{1}(1-c_{0})^{-1}+ c_3 (1-c_{2})^{-1}$.
\end{proof}
Next we will show that the converse estimate in Theorem~\ref{theo4.3} does not hold, that is, there exist an inner function~$f$ with $f(0)=0$ and unbounded sequence~$\{a_{n}\}$ of complex numbers such that $\sum a_{n}f^{n}\in\mathcal{B}$. We start with an auxiliary result on the hyperbolic derivative~$D_{h}f$ defined in \eqref{eq2.1}. Denote $f^{0}(z)=z, z \in \mathbb{D}$.
\begin{lemma}\label{lem4.4}
Let $f\in H^{\infty}(\mathbb{D})$, $\|f\|_{\infty}\le 1$. Then for any $n=1,2,\dotsc$, we have
$$
D_{h}(f^{n})(z)=\prod^{n-1}_{k=0}D_{h}(f) (f^{k}(z)),\quad z\in\mathbb{D}.
$$
\end{lemma}
\begin{proof}
Note that $(f^{n})'(z)=\prod\limits_{k=0}^{n-1} f'(f^{k}(z))$, $n\ge 1$. Then
$$
(1-|z|^{2})\frac{(f^{n})'(z)}{1-|f^{n}(z)|^{2}}=\prod^{n-1}_{k=0}
\frac{(1-|f^{k}(z)|^{2}f'(f^{k}(z))}{1-|f^{k+1}(z)|^{2}}.
$$
\end{proof}
Let $f\in H^{\infty}$ with $\|f\|_{\infty}\le 1$. Schwarz's Lemma gives that $D_{h}f(z)\le 1$, $z\in\mathbb{D}$. Given any positive gauge function $w\colon [0,1]\to (0,+\infty)$ satisfying a mild regularity condition, such that
$$
\int^{1}_{0}\frac{w^{2}(t)}{t}\,dt = \infty,
$$
there exists an inner function~$f$ with $f(0)=0$ such that $D_{h}f(z)\le w(1-|z|)$, $z\in\mathbb{D}$. See \cite{AAN} or \cite{Sm}. In particular, given any $0< \tau <1$, there exists an inner function~$f$ with $f(0)=0$ such that $D_{h}f(z)\le \tau$. Let $\{a_{n}\}$ be any sequence of complex numbers such that
\begin{equation}\label{eq4.4}
\sum^{\infty}_{n=1}|a_{n}| \tau^{n}<\infty.
\end{equation}
We next show that the function~$F=\sum\limits^{\infty}_{n=1}a_{n}f^{n}$ is in the Bloch space. Note that
$$
(1-|z|^{2})|F'(z)| \leq \sum^{\infty}_{n=1} |a_{n} | |(f^{n})' (z) | (1-|z|^{2})=\sum^{\infty}_{n=1} |a_{n}|
(1-|f^{n}(z)|^{2})D_{h}(f^{n})(z) , \quad z \in \mathbb{D}.
$$
Apply Lemma~\ref{lem4.4} to deduce $D_{h}(f^{n})(z)\le \tau^{n}$, $n\ge 1$, $z \in \mathbb{D}$. Then
$$
(1-|z|^{2})|F'(z)|\le \sum^{\infty}_{n=1}|a_{n}| \tau^{n}, \quad z \in \mathbb{D}.
$$
Hence if condition~\eqref{eq4.4} holds, we have $F\in\mathcal{B}$.
We finish this section showing the following converse of part (b) of Corollary 1.3.
\begin{corollary}\label{coro1.10}
Let $f$ be an inner function with $f(0)=0$ which is not a rotation and let $\{a_{n}\}$ be a non-identically zero sequence of complex numbers with $\sum\limits^{\infty}_{n=1}|a_{n}|^{2}<\infty$. Assume
$F=\sum\limits_{n=1}^{\infty} a_{n}f^{n}\in \operatorname{VMOA}(\mathbb{D})$. Then $f$ is a finite Blaschke product.
\begin{proof}
Assume $f$ is not a finite Blaschke product. then there exists a sequence of points $\{z_k \}$ in $\mathbb{D}$ with $|z_k| \to 1$ such that $f(z_k) \to 0$, as $k \to \infty$. Lemma 3.2 gives that
$$
\int_{\partial\mathbb{D}} \left| F(\xi)- F(z_k) \right|^{2} P(z_k,\xi)\,dm(\xi)\ge \frac{1}{2}\, \frac{1-|f'(0)|}{1+|f'(0)|} \sum^{\infty}_{n=1}|a_{n}|^{2},
$$
if $k$ is sufficiently large. Since the integral above tends to $0$ as $k$ tends to $\infty$, we deduce that $a_n =0$ for any $n$. This finishes the proof.
\end{proof}
\end{corollary}
\section{Open problems}\label{sec5}
We close the paper mentioning some open problems we have not explored.
\paragraph{\ Analytic continuation and smoothness.}
Hadamard's Theorem says that if a lacunary power series can be extended analytically across an arc of the unit circle, it can actually be extended analytically to a neighbourhood of the closed unit disc (see \cite[p.~208]{Zy}). It is natural to ask if the following analogous result holds.
\paragraph{Problem 1.} Let $f$ be an inner function with $f(0)=0$ which is not a rotation. Let $\{a_{n}\}$ be a sequence of complex numbers whose modulus are square summable and consider $F=\sum\limits^{\infty}_{n=1}a_{n}f^{n}$. Assume that $F$ extends analytically across an arc~$I\subset \partial\mathbb{D}$. Is it true that there exist $n_{0}>0$ such that $a_{n}=0$ for any $n>n_{0}$ and that $f^n$ extends analytically across $I$ for any $n \leq n_0$?
\vspace{0.2cm}
For $0<\alpha<1$, let $\operatorname{Lip}_{\alpha}(\overline{\mathbb{D}})$ denote the space of continuous functions $g$ in $\overline{\mathbb{D}}$ for which there exists a constant $C=C(g)>0$ such that $|g(z) - g(w)| \leq C |z-w|^{\alpha}$, for any pair of points $z, w \in \overline{\mathbb{D}}$. Lacunary series which belong to $\operatorname{Lip}_{\alpha}(\overline{\mathbb{D}})$ can be described in terms of their coefficients. It is then natural to ask for the corresponding result in our context.
\paragraph{Problem 2.} Let $f$ be a finite Blaschke product with $f(0)=0$ which is not a rotation. Describe the sequences of complex numbers $\{a_{n}\}$ with square summable modulus, such that $F=\sum\limits^{\infty}_{n=1}a_{n}f^{n}\in\operatorname{Lip}_{\alpha}(\overline{\mathbb{D}})$.
\paragraph{\ Peano curves.}
Let $\sum a_{k}z^{n_{k}}$ be a lacunary series with $\sum |a_{k}|=\infty$ and $\lim a_{k}=0$. Then for any $w\in\mathbb{C}$, there exists $\xi\in\partial\mathbb{D}$ such that $\sum a_{k}\xi^{n_{k}}$ converges to~$w$ (\cite{We}). It is then natural to ask for the following analogous result.
\paragraph{Problem 3.} Let $f$ be an inner function with $f(0)=0$ which is not a rotation. Let $\{a_{n}\}$ be a sequence of complex numbers with $\sum\limits_{n=1}^\infty |a_{n}|=\infty$, but $\lim_{n \to \infty} a_{n}=0$. Is it true that for any $w\in \mathbb{C}$, there exists $\xi\in \partial\mathbb{D}$ such that $\sum\limits_{n=1}^\infty a_{n}f^{n}(\xi)$ converges to~$w$?
There is also a version of the question above in the interior of the unit disc. Murai proved that any lacunary series $\sum a_{k}z^{n_{k}}$ convergent in $\mathbb{D}$ with $\sum |a_{k}|=\infty$, takes any complex value infinitely often. See \cite{Mu}. It is natural to ask for the following analogous result.
\paragraph{Problem 4.} Let $f$ be an inner function with $f(0)=0$ which is not a rotation. Let $\{a_{n}\}$ be any sequence of complex numbers such that $F(z)=\sum\limits_{n=1}^{\infty}a_{n} f^{n}(z)$ is analytic in~$\mathbb{D}$.
Assume $\sum\limits_{n=1}^{\infty} |a_{n}|=\infty$. Is true that for any $w\in\mathbb{C}$ there exists $z\in\mathbb{D}$ such that $F(z)=w$?
\paragraph{\ Abel's Theorem.}
Let $F(z) = \sum a_n z^n$ be a power series with radius of convergence $1$. Let $\xi \in \partial \mathbb{D}$ such that $ \sum a_n {\xi}^n$ converges. The classical Abel's Theorem says that $F(z)$ tends to $ \sum a_n {\xi}^n$, as $z$ approaches non-tangentially to $\xi$. The classical Hardy-Littlewood High Indices Theorem asserts that for lacunary series, the converse holds. In other words, if $F(z) = \sum a_k z^{n_k}$ is a lacunary series which has limit $L$ when $z$ approaches non-tangentially a point $\xi \in \partial \mathbb{D}$, then $ \sum a_k {\xi}^{n_k}$ converges to $L$. It is then natural to ask for the following analogous results.
\paragraph{Problem 5.} Let $f$ be an inner function with $f(0)=0$ which is not a rotation. Let $\{a_{n}\}$ be any sequence of complex numbers such that $F(z)=\sum\limits_{n=1}^{\infty}a_{n} f^{n}(z)$ is analytic in~$\mathbb{D}$. Let $\xi \in \partial \mathbb{D}$ such that $f^n (\xi)= \lim\limits_{r \to 1} f^n (r \xi) $ exists for any positive integer $n$. Assume $ \sum\limits_{n=1}^\infty a_n f^n (\xi)$ converges. Is it true that
$\lim\limits_{r \to 1} F(r \xi)$ exists?
\paragraph{Problem 6.} Let $f$ be an inner function with $f(0)=0$ which is not a rotation. Let $\{a_{n}\}$ be any sequence of complex numbers such that $F(z)=\sum\limits_{n=1}^{\infty}a_{n} f^{n}(z)$ is analytic in~$\mathbb{D}$. Assume $F(z)$ has limit when $z$ approaches non-tangentially a point $\xi \in \partial \mathbb{D}$. Is it true that $ \sum\limits_{n=1}^\infty a_n f^n (\xi)$ converges?
|
\section{Introduction}
\label{sec:introduction}
\addspan{In building artificial intelligence (AI) agents, referring to how brains function in real environments can accelerate development by reducing the design space.
}
Hippocampal formation (HF) supports crucial neural capabilities, such as spatial cognition, self-localization for navigation, mapping, and episodic memory.
In neuroscience, HF and its functions have attracted increasing attention in recent years.
The hippocampus has long been considered the brain region responsible for configuring the cognitive map \citep{Tolman1948, Okeefe1978placecells}.
To this end, designated neurons, such as place cells in the hippocampus~\citep{Okeefe1978placecells} and grid cells in the medial entorhinal cortex (MEC), exist to execute these functions~\citep{Fyhn2004,Hafting2005gridcells}.
From the perspective of computational neuroscience, numerous computational model-based studies have focused on functions involving the hippocampus~\citep{milford2004ratslam, Madl2015, Schapiro2017, Banino2018, Kowadlo2019, Scleidorovich2020}.
Alongside these computational studies, the use of brain-inspired AI and intelligent robotics is crucial to the implementation of these spatial functions.
From an engineering perspective, simultaneous localization and mapping (SLAM)~\citep{thrun2005probabilistic} represents a typical approach in computational geometry and robotics.
Spatial cognition and place understanding are important challenges that must be overcome to facilitate the advance of robotics~\citep{taniguchi2018TCDSsurvey}.
However, despite the abundancy of neuroscience knowledge related to HF and the progress in AI technology, combining knowledge from both fields and applying it to robotics remains a major challenge.
\textbf{Purposes}:
This study aims to bridge the gap between neuroanatomical/biological findings of the HF and engineering technologies of probabilistic generative models (PGMs), particularly in AI and robotics.
{This paper is a feasibility study on the methodology proposed by \citet{Yamakawa2021-yy}.}
We establish a correspondence between the function/structure of the HF in neuroscience and spatial cognitive methods in robotics.
The main objectives of this study are as follows.
\begin{itemize}
\item To provide suggestions for the construction of a computational model with functions of HF by surveying the association between SLAM in robotics and HF in neuroscience.
\item To construct a \textit{brain reference architecture} (BRA) that operates with biologically valid and consistent functions, as a specification for implementing a brain-inspired model.
\end{itemize}
\textbf{Type of paper}:
This report is a hypothesis-suggestion paper that presents a novel argument, interpretation, or model intended to introduce a hypothesis/theory, based on a literature review, and provides the direction for its verification.
We construct a \textit{hippocampal formation-inspired probabilistic generative model} (HF-PGM) as a highly adequate hypothesis by referencing neuroanatomical/biological findings.
{The adequacy of the constructed model can be evaluated} by adopting the \textit{structure-constrained interface decomposition} (SCID) method~\citep{Yamakawa2021-yy} for hypothesis generation.
A \textit{generation-inference process allocation} task is used to solve particular problems related to the mapping of PGMs to brain circuits.
This task involves allocating all anatomical connections to either the generative or the inference process.
Here, we learn the connections between modules from the brain and utilize engineering technologies for the parts having insufficient findings of the brain structure and function.
The available findings on HF have been primarily gathered from studies on rodents and only partially from the human brain.
Engineering technologies refer to methods related to spatial awareness (e.g., SLAM, navigation, place recognition, spatial concept formation, and semantic mapping)~\citep{thrun2005probabilistic,kostavelis2015semantic}.
Thus, we present computational HF models to support a feasible hypothesis from an engineering perspective.
Furthermore, we provide suggestions for the construction of methods to study spatial cognitive functions using robotics and {AI} technology from the perspective of neuroscience.
The proposed PGM is designed using an extension of the development method proposed by \citet{Yamakawa2021-yy}.
{Therefore, the adequacy of the architecture is based on the evaluation criteria proposed in the same work.}
These evaluation criteria require (a) the brain information flow (BIF) to be consistent with the anatomical findings, (b) the hypothetical component diagram (HCD) to be consistent with the structure of the BIF, and (c) the HCD to be able to achieve the expected computational functions.
The proposed PGM corresponding to the neural circuitry of the HF is obtained through a design that is a natural extension of SLAM based on engineering practice, and the design procedure is detailed in Section~\ref{sec:hpf-pgm:model}.
Hence, the feasibility (c) of HCD is computationally {supported} without any new implementations or simulations.
\textbf{Contributions}:
The main contributions of this study are as follows.
\begin{itemize}
\item We proffer and apply the generation-inference process allocation, an approach that allows neural circuits to be interpreted as PGM, for the first time.
\item We clarify that the function and structure of HF can be consistently represented as an extension of the previously reported SLAM models by considering relevant findings on HF and SLAM.
\item We show that the phase precession of brain activity in the HF can be formulated as a \textit{discrete-event queue} by repeatedly performing sequential Bayesian inference on the joint probability distribution of the variables.
\item We illuminate the direction of future challenges that are important for the development of HF-inspired models through discussions on spatial movement, hierarchies, and physical constraints.
\end{itemize}
This paper is structured as follows.
Section~\ref{sec:background} describes the background and motivation.
Section~{\ref{sec:bra}} describes the methodology for constructing a brain reference architecture and PGM.
Section~\ref{sec:neuroscience} summarizes the neuroscientific findings on HF.
Section~\ref{sec:hpf-slam} describes relevant topics, including brain-inspired SLAM, computational models of the HF, concept formation and semantic understanding of location and space, and deep generative models.
Section~\ref{sec:hpf-pgm} describes the proposed HF-PGM from the perspective of the model structure.
We construct a novel PGM that is highly consistently with the neuroanatomical/biological findings of HF by integrating allocentric/egocentric visual information.
{Section~\ref{sec:hpf-pgm:queue} describes the formulation of a discrete-event queue based on the phase precession queue assumption in HF as one of the inference algorithms for PGMs.}
Section~\ref{sec:conclusion} summarizes the findings of this study and discusses future directions.
\section{Backgrounds and motivation}
\label{sec:background}
\subsection{Bridging the gap between two fields}
\label{sec:background:gap}
Models facilitated by engineering approaches, such as robotics and AI, are capable of supporting HF modeling.
Numerous aspects of the hippocampus remain unknown.
Hence, there are limitations to creating an HF model solely based on existing hippocampus-related knowledge in neuroscience and biology.
Models for problem-solving based on engineering do not encounter these types of restrictions.
It is possible to build a model with functions similar to those of the hippocampus without actively utilizing its knowledge.
Such engineering models constructed for solving actual tasks may have implications for neuroscience and biology.
Furthermore, the knowledge on neuroscience is useful for engineering applications in robotics and AI.
Robotics generally depend on a particular task and involves practical applications (e.g., accurate 3-dimensional (3D) modeling of the environment.
Brain-inspired {AI} emphasizes learning from the brain and mapping brain circuitry and functions, followed by searching for possible practical applications.
Building a system that refers to the brain that actually operates with functions for various tasks opens engineering possibilities from novel perspectives.
Brain reference architecture, which links neuroscience and robotics while enabling their application in robots, is currently an important field of research.
\subsection{PGM-based cognitive architecture}
\label{sec:background:pgm}
Probabilistic models that have succeeded as neocortical computational models are also useful for modeling HF.
In particular, PGMs represent the process of generating observational stimuli by assuming dependencies between random variables.
Probabilistic inference (e.g., Bayesian) may be employed by adopting PGMs, which infer states behind sensory stimuli as latent variables, enabling the construction of internal representations.
The observed variables correspond to the stimuli.
Adopting PGMs is valid from the perspective of the Bayesian brain hypothesis, which states, ``the brain represents sensory signals probabilistically in the form of probability distributions''~\citep{Knill2004,doya2007bayesian}.
Furthermore, a reason for modeling brain-inspired architecture using PGMs is that the Neuro-symbol emergence in the robotics tool kit (SERKET) architecture can be used both theoretically and practically~\citep{Taniguchi2020neuro}.
Neuro-SERKET enables PGMs having several functions to have distributed development, which can be used to develop integration modules.
This architecture enables the integration of models that imitate other regions in the brain into a whole-brain model.
Therefore, the PGMs for HF proposed in this study may be used as a module to represent the whole-brain's integrated cognitive architecture~\citep{Taniguchi2021wb-pgm}.
\begin{figure}[!tb]
\centering
\includegraphics[width=1.0\textwidth]{SCID_GIPA.pdf}
\caption{
Overview of modeling processes for brain reference architecture \addspan{(BRA)}.
The brain information flow \addspan{(BIF)} and hypothetical component diagrams \addspan{(HCD)} are constructed using the structure-constrained interface decomposition \addspan{(SCID)} method.
The probabilistic generative model (PGM) is one of the expression formats of the hypothetical component diagrams constructed by generation-inference process allocation.
}
\label{fig:scid_gipa}
\end{figure}
\section{Brain reference architecture ({BRA}) construction}
\label{sec:bra}
This section describes the methodology for constructing the PGM as a BRA.
Figure~\ref{fig:scid_gipa} shows the relationship of components for the modeling processes.
Section~\ref{sec:background:wba} introduces an overview on BRA.
Section~\ref{sec:hpf-pgm:scid} {describes construction process and adequacy evaluation for BIF and HCD.}
Section~\ref{sec:hpf-pgm:gipa} describes a problem-solving approach for associating anatomical structures in the brain with PGMs (i.e., generation-inference process allocation).
The generation-inference process allocation is devised and demonstrated for the first time in this study.
\subsection{Brain reference architecture}
\label{sec:background:wba}
This study adopts an approach for building parts of modules in a whole-brain architecture, which is a brain-inspired artificial general intelligence development approach that emphasizes the architecture.
Design methodologies for combining machine learning that imitates the mesoscopic-level brain structures of rodents and humans are being studied to gather knowledge on the architecture of the brain.
Specifically, the whole-brain architecture initiative standardizes information corresponding to the fulfilment of requirement specifications for brain-inspired software as BRA data\footnote{{\url{https://wba-initiative.org/wiki/en/brain_reference_architecture}}} and promotes role sharing in the BRA design and utilization.
The whole-brain architecture initiative includes a manual\footnote{{BRA Data Preparation Manual: \url{https://docs.google.com/document/d/1_t3W_dkFmGjfBhz3_EEZ2FCYyzrJi_1ZOtoPlOru8dc/edit\#heading=h.6766fia4kgtf}}} for preparing brain reference architecture (BRA) as the standard notation for guiding the development of brain-inspired software.
BRA, which is the reference architecture at the mesoscopic level, represents a consistent description of (i) the brain information flow (BIF) related to anatomical structures and (ii) the hypothetical component diagram (HCD) related to its computational functions \citep{Yamakawa2021-yy}.
BIF is a directed graph comprising partial circuits and connections that represent the anatomical structure of neural circuits in the brain.
HCD is a directed graph describing component dependencies, which can be associated with any BIF sub-graph.
The functional mechanisms described in HCD are hypothetical, as the name implies, and different hypotheses may be presented by neuroscientists from different perspectives.
In many cases, brain-inspired software is implemented by engineers who do not necessarily have a deep understanding of neuroscience.
Therefore, it is useful to examine HCD candidates that cannot be clearly dismissed based on current neuroscience knowledge, even if they cannot be confirmed as ground truth.
\subsection{{Construction process and adequacy evaluation}}
\label{sec:hpf-pgm:scid}
In this study, we use a \textit{structure-constrained interface decomposition} (SCID) method developed by the whole-brain architecture initiative~\citep{Yamakawa2021-yy} to design an HCD consistent with the anatomical structure of the HF.
Using the SCID method, \citet{Fukawa2020} applied the concept to the neural circuit of HF for the first time and identified a circuit that performs a path integration function on MEC.
\begin{figure}[!tb]
\centering
\includegraphics[width=0.98\textwidth]{Modeling_processes.pdf}
\caption{
{Modeling processes and section structure by structure-constrained interface decomposition (SCID) method and generation-inference process allocation}
}
\label{fig:model_process}
\end{figure}
We created one plausible HCD by applying steps 1 and 2 {(see \ref{apdx:sec:bra} for the detail)} mainly to the HF.
{In this study, PGM is positioned as a form of expression for HCD.}
Figure~\ref{fig:model_process} shows a table of PGM modeling processes in the SCID method.
The execution result of step 1-A is mainly explained in Section~\ref{sec:neuroscience}.
BIF is shown in Fig.~\ref{fig:hpf_circuit}.
The execution result of step 1-B is explained in Section~\ref{sec:hpf-slam}.
The execution result of step 1-C is explained in Section~\ref{sec:hpf-pgm:roi}.
The execution result of Step 2 is explained in Section~\ref{sec:hpf-pgm:model}.
The contents partially examined in Step 3 are described in Section~\ref{sec:hpf-pgm:step3}.
The authenticity of BIF was established by \citet{Yamakawa2021-yy} in the section ``Adequacy evaluation of BIF''.
We further elaborate on this adequacy in Section~\ref{sec:neuroscience}.
Similarly, the consistency of HCD with BIF has been described in the section ``Adequacy evaluation of HCD'' in \citet{Yamakawa2021-yy}.
The functionality of HCD is ensured by the presence of SLAM models (details are presented in Section~\ref{sec:hpf-pgm:model}.
The model was created to be consistent with BIF, thus ensuring structure-consistency (discussed in Section~\ref{sec:hpf-pgm:step3}).
\subsection{Generation-inference process allocation}
\label{sec:hpf-pgm:gipa}
We describe the task developed in this study to solve particular problems for mapping brain circuits when modeling PGMs.
PGMs have the restriction that they must be directed acyclic graphs; hence, loops cannot be represented.
(See \ref{apdx:pgm_slam:PGM} for a basic description of PGMs.)
However, some brain circuits have loop structures.
In most cases, it is difficult to assign acyclic PGMs to a brain circuit.
Furthermore, in ordinary PGMs, signal propagation in the inference process causes signals to propagate in the direction opposite to that of the links used during the generative process.
In contrast, in brain neural circuits, signal propagations between regions propagated by electrical spikes that propagate terminally on axons are essentially unidirectional.
Generally, modeling a PGM of any existing interarea connection in the brain is a difficult task.
To eliminate this restriction, we adopt a combination of generative/inference models\footnote{As other solutions, we can use probabilistic models based on factor/undirected graphs instead of generative models.}.
In other words, we assume that an \textit{amortized variational inference} is introduced, as described in \ref{apdx:pgm_slam:PGM}.
The amortized variational inference can define the link structure of the inference process without depending on the link structure of the generative process.
The model used for amortized variational inference can be designed with a high degree of freedom, as long as it is consistent with the link structure of the generative process.
Even if loops are present, there is a time delay in the signal transduction in the neural circuits of the brain.
Therefore, we introduce \textit{next-time generation}\footnote{It means assuming state-space models for time-series data.}, which expresses the progress of time for the generation process, as shown in Fig.~\ref{fig:slam_gm} (b).
Therefore, it becomes easy to relate the link structure of PGMs to the structure of actual brain neural circuits.
This allocating task is the generation-inference process allocation.
This task consistently allocates all links into a generative or inference model in a dependency network.
In neural circuits adjacent to the cortex, attention must be paid to avoid inconsistencies at the cortex interface.
If the neuroscientific findings are uncertain, a reasonable and feasible engineering model is chosen.
As a prerequisite for performing this task, each node of these variables\footnote{Nodes are basically random variables; furthermore, they can be represented as temporary variables during the calculation process of deterministic variables.} is principally associated with a uniform circuit (see \cite{Yamakawa2021-yy}), which is the minimum descriptive unit of the BIF.
Generally, applying generation-inference process allocation to any existing interarea connections in the brain is not an easy task.
However, the major interarea connections of the neocortex can be allocated to either the generative or inference process.
In the neocortex, a feedforward pathway transmits signals from lower to higher areas while processing signals received by sensors, and a feedback pathway transmits signals in the opposite direction \citep{Markov2013-zd,Markov2014-ez}.
In computational neuroscience theories (e.g., Bayesian brain \citep{Knill2004,doya2007bayesian} and predictive coding \citep{Rao1999}), inference and generation are assumed to be processed by the feedforward and feedback pathways, respectively.
Therefore, the flow of the inference model, represented by the dotted arrow in the graphical model, is associated with the pathway of the feedforward signal, and the flow of the generative model represented by the solid arrow is associated with the pathway of the feedback signal.
\section{Neuroscientific findings of hippocampal formation}
\label{sec:neuroscience}
In this section, we survey the connections and functions of HF and its surrounding areas.
The core region of interest of the HF focused in this study includes the cornu ammonis-1 and -3 areas (CA1 and CA3) and dentate gyrus (DG) in the hippocampus, lateral/medial entorhinal cortex (LEC/MEC), subiculum (Sb), and parasubiculum (ParaSb).
The presubiculum (PreSb), perirhinal cortex (PER), postrhinal cortex (POR), retrosplenial cortex (RSC), and medial septum are also investigated as areas adjacent to the HF, that is, as regions for sending and receiving the signal.
\subsection{Anatomical and physiological findings}
\label{sec:neuroscience:connections}
\begin{figure}[!tb]
\centering
\includegraphics[width=1.0\textwidth]{HPF_BIF.pdf}
\caption{
Brain information flow (BIF) represents the connection of neural circuits between HF and its surrounding areas.
It is based on anatomical studies and reviews \citep{Amaral1989,Witter2000,Fukawa2020}.
{Blue, lateral entorhinal cortex (LEC) and perirhinal cortex (PER); red, medial entorhinal cortex (MEC), presubiculum (PreSb), parasubiculum (ParaSb) and postrhinal cortex (POR); yellow, the hippocampus and subiculum (Sb); green, the medial septum; and purple, retrosplenial cortex (RSC).}
{The hippocampus includes the cornu ammonis-1 and -3 areas (CA1 and CA3) and dentate gyrus (DG).}
Colored arrows indicate that the signal is sent from the area indicated by that color to the other area at the tip of the arrow.
}
\label{fig:hpf_circuit}
\end{figure}
With respect to neuroscientific findings, we conducted a survey by adding detailed findings on LEC and hippocampus to the findings in the review of \citet{Fukawa2020} centered on the hippocampus and MEC.
These findings are also based on an original anatomical review of the structure of HF~\citep{Amaral1989,Witter2000}.
Figure~\ref{fig:hpf_circuit} shows the connection relationship of the HF circuit, which includes the hippocampus, Sb, PreSb, ParaSb, and entorhinal cortex.
The hippocampus comprises DG, CA1, and CA3.
The entorhinal cortex is divided into MEC and LEC.
CA1 and Sb are further separated distally and proximally~\citep{Knierim2014}.
In this study, MEC is divided into six layers: I, II, III, Va, Vb, and VI, and LEC is similarly divided into six layers: I, II, III, IV, V, and VI~\citep{Shepherd2013, Fukawa2020}.
MEC IV is excluded because it has few neurons.
As input to LEC, a connection exists from the POR and PER to LEC II and III~\citep{Nilssen2019}.
The output from LEC has a connection from LEC II to DG and CA3 and from LEC III to CA1 and Sb.
From CA3, there are two connections, Sb proximal through CA1 distal and Sb distal through CA1 proximal~\citep{Knierim2014}.
The nonspatial signal is conveyed from LEC to the hippocampus~\citep{Hargreaves2005}.
The MEC and LEC signals are integrated by the DG and CA3 in the hippocampus~\citep{Chen2013, Knierim2014}.
Neural cells with various expressions have been observed in the hippocampus and its surrounding areas.
Place cells exist in CA3 and CA1~\citep{Okeefe1978placecells}; they are active only when an animal enters a specific place in the environment; they do not fire elsewhere.
It is thought that an environmental cognitive map is stored as a neural circuit according to the place cells~\citep{Okeefe1978placecells}.
Grid cells exist in MEC, ParaSb, and PreSb~\citep{Hafting2005gridcells}.
Although the activity of place cells represents one place in the environment, grid cells are activated in a hexagonal grid at multiple places in the environment.
Head-direction cells exist in MEC, PreSb, ParaSb, and RSC~\citep{Taube1990, Taube2007, Grieves2017}; they are active only when the head faces a specific direction, regardless of the location of the viewer.
Border cells and boundary vector cells are cells in Sb and MEC that selectively fire near the border, irrespective of whether there are objects or obstacles~\citep{lever2009}.
Speed cells exist in MEC and exhibit activity that depends on the speed of movement of the subject~\citep{Kropff2015, Hinman2016}.
Spatial view cells that respond to the current landscape are known to exist on the primate's hippocampus~\citep{Rolls2013}.
Furthermore, event cells have been discovered that represented the content and order of the events experienced~\citep{Terada2017}.
Events that are not evenly spaced in time are processed centrally by the LEC and stored only when the state changes.
\subsection{Functions}
\label{sec:neuroscience:functions}
HF is a brain region that controls short-term and episodic memory in vertebrates. It is deeply involved in spatial memory functions, such as spatial learning and exploration.
The behavior of an animal when traveling to a destination by selecting an appropriate route while simultaneously acquiring environmental information is called navigation.
Cognitive mapping, where animals form a map of the spatial positional relationship of various items in the environment by exploration and act accordingly using this map, is proposed as a psychological concept deeply related to navigation~\citep{Tolman1948,Giocomo2011}.
The hippocampus is thought to be crucial to the formation of cognitive maps in mammals.
Self-localization, where animals consistently recognize their current position, is indispensable for the navigation and formation of cognitive maps.
One of the functions that support the self-localization ability of animals is path integration~\citep{McNaughton2006}.
Path integration is a function that outputs self-position after movement upon input of the initial position, head-direction, and movement signals, including speed and movement direction \citep{Raudies2015a}.
These researchers argued that the region responsible for path integration was either MEC II stellate cells or MEC III.
Path integration is calculated using grid cells~\citep{Gil2018}.
Their firing pattern represents metric properties, and the firing pattern of place cells represents the position index information without metric properties~\citep{Buzsaki2013,Fukawa2020}.
The prospective speed signal is calculated by MEC III or MEC Vb.
Furthermore, in the hippocampus, there is a loop structure wherein the movement signal by path integration in MEC and signal from the LEC are integrated to obtain accurate position information.
DG and CA3 execute the functions of pattern separation and pattern completion, respectively \citep{Bakker2008}.
Pattern separation is the ability to identify the difference between two perceptual patterns, and pattern completion is the ability to generalize and complement using similar signals from partial observation and noisy environments.
\begin{figure}[!tb]
\centering
\includegraphics[width=0.55\textwidth]{map_and_view.pdf}
\caption{
Egocentric and allocentric representations of space.
(a) Egocentric visual information. First-person perspective and self-centered coordinate system.
(b) Allocentric visual information. Objective perspective and object center coordinate system .
}
\label{fig:map_and_view}
\end{figure}
LECs and MECs have been reported to perform different processing functions owing to the difference in the signals sent from PER and POR.
LEC and MEC process proximal and distal landmarks, respectively~\citep{Kuruvilla2020}.
In addition, LEC processes signals from an egocentric perspective, whereas MEC processes signals from an allocentric perspective~\citep{Hargreaves2005,Byrne2007,Deshmukh2011,Bicanski2018a,Alexander2020,Wang2020}.
Figure~\ref{fig:map_and_view} shows the difference between egocentric and allocentric perspectives.
Considering these factors, two types of cognitive maps are generated: survey and route maps~\citep{Shemyakin1962OrientationIS}.
Therefore, LEC and MEC are expected to be responsible for spatial cognition of the route and survey maps, respectively.
These two types of views are used in car navigation systems and online map apps (e.g., Google Street View).
\section{Relevant topics and correspondence between HF and SLAM}
\label{sec:hpf-slam}
We introduce the research on computational models related to spatial cognition methods, including SLAM.
Furthermore, we discuss the association between functions of HF and computational models based on the content covered in \ref{apdx:pgm_slam} and Section~\ref{sec:neuroscience}.
First, we introduce the HF computational models and brain-inspired SLAM in Section~\ref{sec:hpf-slam:models}.
Second, we introduce the models relating to place category formation and investigate their relationship with HF in Section~\ref{sec:spatial_concept}.
Finally, the neural-network models and their associations with PGMs are described in Section~\ref{sec:hpf-slam:neuralnet}.
\subsection{Computational models for HF and brain-inspired SLAM}
\label{sec:hpf-slam:models}
Herein, we describe the HF computational model and the relationship between HF and SLAM.
Several studies have discussed the relationship between SLAM, partially observable Markov decision process graphical models, their probabilistic inferences, and HF circuits and functions~\citep{Penny2013,Madl2018,Gaussier2019,Bermudez-Contreras2020}.
SLAM functions mainly correspond to the loop circuits of the MEC and hippocampus~\citep{Fukawa2020}.
\citet{Bermudez-Contreras2020} discussed the association with artificial neural networks and reinforcement learning in spatial navigation.
Furthermore, several studies were conducted on computational models of functions relating to the hippocampus from the perspective of computational neuroscience~\citep{Schapiro2017,Banino2018,Kowadlo2019,Scleidorovich2020}.
Several hippocampus-inspired SLAM methods have been proposed~\citep{milford2004ratslam,Tang2018,Yu2019,Zou2020}.
\citet{milford2004ratslam} implemented a biologically inspired mapping system, RatSLAM, which is related to place cells in the hippocampus of a rodent.
Some of the existing SLAM functions can be partially associated with each functional cell in the HF.
Metric representation in the MEC is approximated by the map coordinate system in SLAM.
Grid cells represent spatial metrics to determine coordinate axes~\citep{Hafting2005gridcells}.
Models of grid-cell spatial firing have been proposed in computational neuroscience studies \citep{Zilli2012}.
In contrast, the map representation in numerous SLAM models is either a 2D or 3D Cartesian coordinate system.
Particularly, the occupancy-grid map of SLAM can be regarded as a combined representation of grid and border cells.
Additionally, in the self-localization model designed for engineering~\citep{ishibushi2015statistical}, Gaussian distributions are arranged at equal intervals, resulting in a representation resembling grid cells.
The state variables in SLAM that represent posture are defined by position coordinates and orientation.
Therefore, posture is associated with the firing of grid and head-direction cells.
In landmark-based SLAM, which is a time-series state-space model, the landmark position is estimated using a Kalman filter~\citep{montemerlo2002fastslam}.
This corresponds to the MEC function for processing distal landmark signals obtained from POR~\citep{Kuruvilla2020}.
In general SLAM, it is assumed that the landmarks are static.
Expanding this to dynamic objects is expected to enable the estimation of the position of the movement of other individuals, for example, as a dynamic event that is part of the external context in the environment.
This type of expanded SLAM model may be a candidate model for a research report on the activity pattern of place cells in other rats~\citep{Danjo2018}.
In the brain, self-localization is performed by theta-phase precession~\citep{Hafting2008}.
Phase precession is explicitly modeled as a spatial cell model in \citet{Zou2020}.
However, in the SLAM studies, there are a few models that explicitly implement phase precession, even in brain-inspired models.
Hence, it is implemented by different processes that are functionally similar in engineering.
In this study, we propose a discrete-event queue, as shown in Section~\ref{sec:hpf-pgm:queue}.
Episodic memory is an important function both in the hippocampus and robotics.
In the computational neuroscience field, models of the hippocampus in episodic memory have been proposed~\citep{Mcnaughton1987,Treves1994,ME1997}.
Several methods incorporating episodic memory have been proposed for robotics~\citep{Tang2017,Ueda2018,ref:furuta2018everyday,Zou2020}.
\citet{Ueda2018} proposed a brain-inspired method (i.e., a particle filter on episode) for agent decision-making.
An episode that replays in CA3 may be modeled with a generative process.
There are various ways to express episodic memory; however, robots can retain temporal transitions with observed events and refer to them subsequently.
\subsection{Semantic mapping and spatial concept formation in robotics}
\label{sec:spatial_concept}
In mobile robots, it is essential to appropriately generalize and form place categories while dealing with the uncertainty of observations.
Hence, a semantic mapping approach, including semantics of places and objects, has been actively developed~\citep{kostavelis2015semantic,Garg2020}.
To address these issues, PGMs for spatial concept formations have been constructed \citep{isobe2017learning,ataniguchi_IROS2017,hagiwara2018hierarchical,ataniguchi2020spcoslam2,Katsumata2020SpCoMapGAN}.
\citet{ataniguchi_IROS2017,ataniguchi2020spcoslam2} realized a PGM for online spatial concept acquisition with simultaneous localization and map-ping (SpCoSLAM), which conducts place categorization and mapping through unsupervised online learning from multimodal observations.
Visual information is used as landscape features reminiscent of spatial view cells~\citep{Rolls2013}.
HF is also centrally involved in the formation of place categories, semantic memory, and understanding of the meaning of places by integrating signals from each sensory organ~\citep{Buzsaki2013}.
We consider the aforementioned models as candidates for a cognitive module with functions similar to those of the HF.
SpCoSLAM is a model that arguably imitates some functions of the hippocampus and cerebral cortex.
From the viewpoint of computational efficiency and estimation accuracy, \citet{ataniguchi2020spcoslam2} proposed an inference algorithm that sequentially re-estimates some recent events and accordingly updates global parameters in older observations.
Assuming that the training data (i.e., the event-based robotic experience) represent episodic memory and spatial concepts represent semantic memory, their algorithms can sequentially extract concepts from short-term episodic memory to form a semantic memory.
\citet{isobe2017learning} proposed a model for place categorization using self-position and object recognition results.
This model shows that the categorization accuracy is higher when weighing is performed while considering only the objects that are close to the robot, rather than using all objects that are evident from the robot's viewpoint for place categorization.
This result is also consistent with the neuroscientific findings of proximal landmarks~\citep{Kuruvilla2020}.
Furthermore, the hierarchical multimodal latent Dirichlet allocation~\citep{Ando2013hMLDA} provides a categorical representation of hierarchical locations~\citep{hagiwara2018hierarchical}.
The multi-layered k-means was adopted to extract the hierarchical positional features of a space.
Arguably, this corresponds to the hierarchical representation of grid cells in an MEC~\citep{Hafting2005gridcells} and is considered a valid model based on neuroscientific findings.
Although the aforementioned algorithms and models were not originally inspired by biology or neuroscience, such research is highly suggestive.
\subsection{Neural-network models}
\label{sec:hpf-slam:neuralnet}
Deep neural-network models, such as world models~\citep{Ha2018}, represent various types of information in the latent state space.
In ordinary SLAM, the functional shapes of the distributions (i.e., global parameters) for motion models, which represent state transitions and the measurement models that match observations to a map, are designed and set by humans.
Therefore, SLAM problem means the estimation of local latent parameters.
In contrast, world models do not have explicit models embedded with prior knowledge.
Deep neural-network models learn elements corresponding to global and local parameters in PGMs simultaneously.
The estimation of motion and measurement models, including map representations on neural networks, is related to predictive coding.
As computational models for HF, grid cells and their similar spatial representations have been reproduced by deep neural-networks, including long short-term memory (LSTM) or multi-layered recurrent neural networks (RNNs)~\citep{noguchi2018navigation,Grid2018,Banino2018}.
A vector representation similar to that of grid cells is acquired as an internal state by LSTMs, which predicts the current posture from the past velocity and angular velocity with dropouts~\citep{Banino2018}.
In contrast, from an engineering point of view, models for place representation and spatial concepts using deep neural-networks have been proposed.
The room space is learned from the visual-motor experience using two sub-networks comprising a deep auto-encoder and an LSTM~\citep{yamada2017learning}.
By integrating the spatial concept formation model introduced in Section~\ref{sec:spatial_concept} with generative adversarial networks, \citet{Katsumata2020SpCoMapGAN} transferred global spatial knowledge from multiple environments to a new environment.
These approaches have some suggestive elements that can be interpreted as an HF model.
With the advent of Bayesian deep learning and deep PGMs, it has become possible to discuss neural networks within the framework of PGMs.
In particular, by implementing it in the framework of deep PGMs, it is possible to naturally incorporate hierarchical RNNs into a generative model, such as variational auto-encoder and generative adversarial networks.
For example, the predictive-coding-inspired variational RNN~\citep{Ahmadi2019} is not a model for the hippocampus or navigation task, but a hierarchical RNN-based deep PGM.
Deep PGMs are achieved by amortized variational inference, a type of variational inference that introduces functions to transform observation data into parameters of the approximate posterior distribution.
Therefore, it is possible to formulate arbitrary variable transformations and nonlinear functions as probability distributions.
The framework for maximizing the evidence lower bound by variational inference is equivalent to free energy minimization.
Therefore, if the HF models are described by PGMs, they can be naturally connected to the free energy principle~\citep{friston2019} and world models~\citep{Ha2018}.
\section{HF-PGM: Hippocampal formation-inspired probabilistic generative models}
\label{sec:hpf-pgm}
We constructed a graphical model of HF and its surrounding regions: HF-PGM.
We describe the HF-PGM in association with the HF by interpreting and modeling it using a graphical model.
Section~\ref{sec:hpf-pgm:roi} describes the region of interest, top-level function, and input/output signals.
Section~\ref{sec:hpf-pgm:executing_gipa} describes the execution for allocation of generative or inference process in connections.
Section~\ref{sec:hpf-pgm:model} describes the time-series representation, encoder-decoder representation, and representation associated with BIF in the proposed HF-PGM.
Section~\ref{sec:hpf-pgm:step3} discusses the consistency of the model with scientific knowledge.
\subsection{Region of interest, top-level function, and input/output}
\label{sec:hpf-pgm:roi}
The region of interest targeted in this study to realize the proposed model includes LEC, MEC, CA1, CA3, DG, Sb, and ParaSb.
Additionally, adjacent areas of the region of interest (i.e., areas where the signal can be obtained or a signal is sent) are PreSb, POR, PER, and RSC.
The connection relationship of each region is based on Fig.~\ref{fig:hpf_circuit}.
\citet{Fukawa2020} mainly focused on the MEC and hippocampal circuits.
In addition, LEC is considered in this study.
Hypotheses on LEC functions were derived from several papers and used for model construction.
This study assumes that the model works when the agent is walking in an awake state.
The top-level function for the assumed region of interest integrates allocentric and egocentric information once and outputs each prediction.
The following activities describe the processing inside the model used to realize the top-level function:
\\
{\textbf{(i) Self-localization (path integration and observational correction)}:
Self-localization is performed by path integration and observation-based prediction correction (as explained in \ref{apdx:pgm_slam:SLAM} and Section~\ref{sec:neuroscience:functions}).
The difference from conventional SLAM is that the prediction is corrected using integrated information from the LEC-variables, and the prediction of the future times is output to other regions.
For details, refer to Section~\ref{sec:hpf-pgm:model:pgm-time}.
\\
\textbf{(ii) Place categorization by integrating allocentric and egocentric information}:
By integrating the allocentric visual information processed by the MEC and the egocentric visual information processed by the LEC, a place category representing semantic memory about a place is formed~\citep{Buzsaki2013}.
As a PGM, it can be modeled as a multimodal categorization.
For details, refer to ``integration of information'' in Section~\ref{sec:hpf-pgm:model:pgm-bif}.}
The inputs to the model include variables in POR, PER, and RSC.
The internal representations at these parts are treated as observed variables in PGM.
The outputs of the model are the predicted values in POR and PER at the next time-step.
In PGM, the latent variables and parameters of the conditional distribution are obtained as the estimated values.
The definition of each corresponding variable is described in Table~\ref{tab:element_of_graphical_model}.
\subsection{Executing generation-inference process allocation}
\label{sec:hpf-pgm:executing_gipa}
We constructed HF-PGM to be consistent with SLAM's PGM based on the generation-inference process allocation procedure (see Section~\ref{sec:hpf-pgm:gipa}).
The connection between POR and MEC II superficial on the BIF (Fig.~\ref{fig:hpf_circuit}) can be regarded as a feedforward pathway.
An inference process can be allocated to this connection in the PGM.
In contrast, the connection between MEC V and POR on the BIF can be regarded as a feedback pathway.
A generative process can be allocated to this connection in the PGM.
The connections between LEC and PER are assigned in the same way.
There exist limitations to performing generation-inference process allocation for the inside of the hippocampus while considering only the connectivity with the neocortex.
Therefore, the engineering formulation of the SLAM modeled as PGM is used as a reference.
Partially following the generative process in the PGM of SLAM (see \ref{apdx:pgm_slam:SLAM}), the links related to the state transition, motion models $p(x_{t} \mid x_{t-1}, u_{t})$, were assigned as the generative processes.
\subsection{Models and operating principle}
\label{sec:hpf-pgm:model}
\begin{figure}[!tb]
\centering
\includegraphics[width=1.0\textwidth]{HPF_PGM.pdf}
\caption{
Graphical model representation of HF-PGM {as the final form}.
It is drawn on the BIF, representing the circuit diagram of the HF shown in Fig.~\ref{fig:hpf_circuit}.
The area surrounded by a gray frame is the region of interest targeted in this study.
Gray nodes represent observed variables, whereas white nodes represent unobserved latent variables.
Black arrows represent the generative process, and dotted arrows represent the inference process.
The flat arrow with $\Delta t $ indicates the generation of the variable in the next time step.
Nodes surrounded by gray circles are assumed to be functionally similar and may be treated as the same variable.
}
\label{fig:hpf_pgm}
\end{figure}
In this section, we describe the model structure and function of the proposed HF-PGM.
We associate variables with each region of the HF.
Some anatomical connections are omitted in the PGM for engineering feasibility.
Descriptions of global parameters are omitted from the graphical model.
In this study, a specific shape is not particularly limited in probability distributions.
Only random variables, their functional definitions, and dependencies between the random variables are assumed.
Here, the node does not necessarily have to be a random variable.
Signals from outside the HF-PGM module are treated as observed variables, and it is assumed that the observation of each input is converted from raw data into individual high-level features through each processing module; additionally, it enters the HF-PGM.
Models may be considered from raw data that can be observed by robots; however, because we intend focus on the top-level function, for the time being, we consider them as extracted features.
The {final form of }HF-PGM is shown in Fig.~\ref{fig:hpf_pgm}.
HF-PGM was expressed using inference and generative processes based on the generation-inference process allocation.
We describe the variables of HF-PGM in Table~\ref{tab:element_of_graphical_model}.
Next, two types of descriptions in the table are described.
Physiological findings include a description of a physiological phenomenon observed at a specific site on the BIF or the function inferred from it.
Because there are many uncertain elements in physiological findings related to LECs, the description is omitted here.
The function of the components is a description of the computational functions assumed for each component included in the HCD.
As a detail of the aforementioned contents with references, we released a pre-screening version of the BRA data\footnote{Hippocampal formation BRA data (pre-screening version): \url{https://docs.google.com/spreadsheets/d/1xf5tIj2qzHh9a52a2p9K5b8ggra825bvBdXKhghF2W4/edit\#gid=0}} \addspan{as supplementary material (\texttt{hf\_bra.xlsx})}, which describes the navigation functions of the HF.
Sections~\ref{sec:hpf-pgm:model:pgm-time} -- \ref{sec:hpf-pgm:model:pgm-bif} explain the flow of changes from the general SLAM models in a step by step manner, as shown in Fig.~\ref{fig:hpf_pgm}.
This implies that the functions of SLAM's PGM can be decomposed and associated based on the anatomical structure of HF.
Meanwhile, the functions of the HF-PGM can include the functions of the conventional SLAM (discussed in Section \ref{sec:hpf-slam:models}) and the spatial concept formation models (introduced in Section \ref{sec:spatial_concept}).
The operating principle, as an internal process for realizing top-level function in the HF-PGM, is shown below.
These also serve as sub-functions subordinate to the top-level function.
{\scriptsize
\begin{tabularx}{\linewidth}{p{11mm}X||p{9mm}p{45mm}}
\caption{
Description of variables in the HF-PGM.
Grey backgrounds are associated with PGM for SLAM, as shown in Fig.~\ref{fig:slam_gm}.
The corresponding variables are listed together in symbol.
}
\\ \hline
\textbf{Region} & \textbf{Physiological findings} & \textbf{Symbol} & \textbf{Function of components on HCD} \\
\hline
CA1\, distal & Non-spatial semantic memory & $C$ & Place category (internal representation of visual spatial information) \\
CA1 proximal & Place cells & $r$ & Position distribution (cluster information regarding positions) \\
Sb\, proximal & State at CA1 distal with time delay & $C^{\prime}$ & Place category at the previous time \\
Sb distal & State at CA1 proximal with time delay & $r^{\prime}$ & Position distribution at the previous time \\
CA3 & Pattern completion, information integration & $H^{\rm{I\hspace{-.1em}I\hspace{-.1em}I }}$ & Integrated semantic memory and episodic memory of information from $X$ and $Z$ \\
DG & Pattern separation & $H^{\rm DG}$ & Integrated semantic memory \\
\rowcolor[gray]{0.9}
MEC \rm{I\hspace{-.1em}I } & Grid cells, Border cells, path integration & $X^{\rm{I\hspace{-.1em}I }}$ ($\{x_{t}\}$) & Self-position information, predictive distribution \\
\rowcolor[gray]{0.9}
MEC \rm{I\hspace{-.1em}I\hspace{-.1em}I } & Grid cells, head-direction cells, border cells & $X^{\rm{I\hspace{-.1em}I\hspace{-.1em}I }}$ ($\{x_{t}\}$) & Self-posture information (position and orientation), observation likelihood \\
MEC Va,Vb,VI & Prospective speed calculation, feedback to POR & $g$ & Predictive representation of $X$ (Prediction at future time regarding movement/speed amount or posture) \\
LEC \rm{I\hspace{-.1em}I } & --- & $Z^{\rm{I\hspace{-.1em}I }}$ & Abstraction of information from $y^{\rm{PER}}$ (transmission of prediction, generation of prediction signal) \\
LEC \rm{I\hspace{-.1em}I\hspace{-.1em}I } & --- & $Z^{\rm{I\hspace{-.1em}I\hspace{-.1em}I }}$ & Abstraction of information from $y^{\rm{PER}}$ (Observation transmission) \\
LEC V,VI & Feedback to PER & $l$ & Predictive representation of $Z$ (Prediction at future time from the difference between $C^{\prime}$ and $C$) \\
\rowcolor[gray]{0.9}
ParaSb & --- & $X^{\prime}$ ($\{x_{t-1}\}$) & Self-posture information \\
\rowcolor[gray]{0.9}
POR & --- & $y^{\rm{POR}}$ ($\{y_{t}\}$) & Allocentric visual information (distal distance/landmarks, absolute object positions) \\
PER & --- & $y^{\rm{PER}}$ & Egocentric visual information (proximal distance/landmarks, relative object positions, object category, landscape information) \\
\rowcolor[gray]{0.9}
RSC & Head direction signal & $u^{w}$ ($\{u_{t}\}$) & Rotational speed movement \\
\hline
\label{tab:element_of_graphical_model}
\end{tabularx}
}
\subsubsection{PGM representation according to hippocampus and MEC structures}
\label{sec:hpf-pgm:model:pgm-time}
\begin{figure}[!tb]
\centering
\includegraphics[width=0.7\textwidth]{HPF_PGM_time.pdf}
\caption{
Time-series version of graphical model representation of only the MEC side in HF-PGM.
The generation and inference processes are drawn simultaneously.
The subscript representing time reflects the time of the outside world, in which the observation is obtained.
Notably, it does not represent the time of internal processing.
Observations up to the current time, $ t $, are obtained.
}
\label{fig:hpf_pgm_time}
\end{figure}
The graphical model is shown in Fig.~\ref{fig:hpf_pgm_time}, which is a variant of the PGM of SLAM (see Fig.~\ref{fig:slam_gm}) and is mapped with reference to the hippocampal and MEC loop structures.
This model is an extension of a partially observable Markov decision process.
In this model representation, we omit the connections on the LEC side. However, the signal obtained from the LEC side was originally integrated via $H_{t}$, which corresponds to the integrated higher-level internal representation.
Corresponding to this graphical model of the neural circuits in the hippocampus and MEC, we obtain the following:
CA1/Sb ($r_{t-1}$)
$\rightarrow$
ParaSb ($X^{\prime}_{t}$)
\&
MEC Vb, VI ($g_{t}$)
+
POR/RSC ($y_{t}^{\text{POR}}, u_{t}$)
$\rightarrow$
MEC \rm{I\hspace{-.1em}I } (${X}_{t}$)
$\rightarrow$
DG/CA3 ($H_{t}$)
\&
CA1/Sb ($r_{t}$)
$\rightarrow$
MEC Vb, VI ($g _{t+1}$)
$\rightarrow$
POR ($\hat{y}_{t+1}^{\text{POR}}$).
Some are defined as the same variable, assuming that they have similar functions; however, the regions are different.
HF-PGM differs from the traditional PGM of SLAM when generating predictions of observation for the next cycle.
SLAM directly estimates the current self-posture, $x_{t}$, from the previous self-posture, $x_{t-1}$.
Furthermore, the generative process of SLAM is the arrow from $x_{t}$ to $y_{t}$ at time $t$.
In contrast, HF-PGM generates the self-posture, $X_{t+1}$, from $X_{t}$ via variables, such as $H_{t}$, $r_{t}$, $X^{\prime}_{t+1}$, and $g_{t+1}$.
The higher-level representation of place is denoted as $r_{t}$.
$g_{t+1}$ is responsible for the prediction of the next time.
It is assumed that $r_{t}$ comprises differential information about the time delay between CA1 and Sb.
Hence, the prediction, $g_{t+1}$, is generated from $r_{t}$, and the time-differential information is used.
This prediction by the difference calculation is not performed in conventional SLAM.
Additionally, $r_{t-1}$ at the previous time, $t-1$, generates the self-posture, $X^{\prime}_{t}$.
$X^{\prime}_{t}$ is responsible for conveying the position information of the previous time.
The self-position, $X_{t}$, is predicted from $X^{\prime}_{t}$ and the movement amount, $u_{t}$.
Subsequently, the self-position, $X_{t}$, is corrected by the predicted value, $g_{t}$, and the observation, $y_{t}^{\text{POR}}$.
It is assumed that the predicted value, $g_{t}$, is determined by reducing the error so that $X_{t}$ can be generated consistently with other variables.
It is also possible to minimize the prediction error of $\hat{y}_{t}^{\rm POR}$, which is generated from the prediction, $g_{t}$, and the actual observation, $y_{t}^{\rm POR}$: predictive coding~\citep{Rao1999}.
Related to the theta-phase precession phenomena in the HF circuit, signals that circulate in the looping circuit in HF require units of discrete-event queues in a current state, which includes both near-past and near-future predictions (see also Section~\ref{sec:hpf-pgm:queue}).
Additionally, given the nature of theta-phase precession, variables that form a large loop may not have self-transition~\citep{Butler2018}.
Therefore, it is assumed that $H_{t}$, $r_{t}$, and $X_{t}$ do not have a direct time self-transition despite having a direct neural connection.
\subsubsection{Integrated with LEC}
\label{sec:hpf-pgm:model:pgm-mae}
\begin{figure}[!tb]
\centering
\includegraphics[width=0.8\textwidth]{HPF_PGM_simple_MAE.pdf}
\caption{
Encoder--decoder version of graphical model representation in HF-PGM.
}
\label{fig:hpf_pgm_MAE}
\end{figure}
Figure~\ref{fig:hpf_pgm_MAE} is a time-omitted representation of the graphical model in Fig.~\ref{fig:hpf_pgm_time} with additional connections on the LEC side.
The projection from LEC V, VI $l$ to LEC II, III $Z$ assumes a connection structure similar to MEC.
This graphical model representation clearly shows the result of generation-inference process allocation.
Because the HF-PGM in Fig.~\ref{fig:hpf_pgm_MAE} is compressed in time, a circulation occurs in the generative process.
Therefore, the notation for the next time generation process introduced in Fig.~\ref{fig:slam_gm} is used.
Notably, there exists an arbitrariness of position with which the time progress can be allocated in the PGM loop.
This model has an encoder--decoder structure and can be seen as an extended form of variational encoders--decoders~\citep{Bahuleyan2018} with two modalities and a condition.
Because the model structure is similar to encoder--decoder models, the input and output have the same variables with the next time generation process for the inference and generation sides.
The signals (${y}^{\text{PER}}$ and ${y}^{\text{POR}}$) from PER and POR-parts are processed into $Z$ and $X$ by LEC and MEC, respectively, integrated into $H$ by DC/CA3-part and divided into $C$ and $r$.
Subsequently, $\hat{y}^{\text{PER}}$ and $\hat{y}^{\text{POR}}$ are returned from the predictive information, $l$ and $g$.
During training, the loss function can be designed to match the input to the generated output.
The structures of $Z$ to $C$ and $X$ to $r$ are similar to those of the skip connection (i.e., the contracting path) and the U-net~\citep{Ronneberger2015}.
Furthermore, a loop structure has a recurrent time delay on the MEC side, which is consistent with the anatomy of the MEC and hippocampus.
It is suggested that this loop structure is crucial to self-localization~\citep{Fukawa2020}.
\subsubsection{HF-PGM associated with BIF}
\label{sec:hpf-pgm:model:pgm-bif}
Finally, corresponding to the BIF of the HF, Fig.~\ref{fig:hpf_pgm} is obtained.
The following is a detailed description of each part of the HF-PGM based on Fig.~\ref{fig:hpf_pgm}.
\textbf{Input and output}:
The amount of rotational movement, $u_{t}^{w}$, is transmitted from RSC to MEC \rm{I\hspace{-.1em}I\hspace{-.1em}I } ($X^{\rm{I\hspace{-.1em}I\hspace{-.1em}I }}$) via PreSb.
The translational speed is assumed to be calculated inside the MEC, where the speed cells exist~\citep{Kropff2015, Hinman2016}, because the corresponding region cannot be clearly identified.
Hence, the amount of movement, $u_{t}$, is obtained by integrating the difference information sent from $r$ and $r^{\prime}$ with $u_{t}^{w}$.
In engineering, it is possible to calculate the speed from the difference information. For example, there are methods such as optical flow.
A discussion related to the aforementioned sentences is provided in \ref{apdx:future:speed}.
Additionally, this time-difference information, as mentioned in Section~\ref{sec:hpf-pgm:model:pgm-time}, is useful for predicting the internal state and input signal at the next time\footnote{As the variables in CA3 and Sb are separated in Fig.~\ref{fig:hpf_pgm}, it is considered that C and r hold time-difference information as an internal representation.
Because $r^{\prime}$ and $C^{\prime}$ are said to represent the state having a time delay of $r$ and $C$ in BIF, further studies are required to assess the assumptions on the separability and generation of these variables.}.
POR $y^{\rm POR}$ mainly deals with distal landmarks and distance signals~\citep{Kuruvilla2020}.
Distant information is useful for self-localization because it can be obtained more robustly than proximal information while moving.
PER $y^{\rm PER}$ mainly deals with proximal landmarks and distance signals.
This signal is useful to avoid proximal obstacles.
Additionally, proximal objects may be useful in forming a place category for the current location~\citep{isobe2017learning}.
Landscape information can also be used to roughly identify the current location~\citep{Rolls2013}.
These are similar to the treatments of observations in spatial concept formation, as described in Section~\ref{sec:spatial_concept}.
\textbf{Role of LEC/MEC}:
Generally, in the neocortex, layer \rm{I\hspace{-.1em}I\hspace{-.1em}I } is thought to be responsible for the transmission of observations, and layer \rm{I\hspace{-.1em}I } for predictions~\citep{Yamakawa2020}.
Given that the LEC/MEC is also part of the neocortex, its role is likely to be preserved.
LEC/MEC \rm{I\hspace{-.1em}I\hspace{-.1em}I } receives observation signals from the POR and PER and projects them mainly to CA1 and other areas.
LEC/MEC \rm{I\hspace{-.1em}I } generates predictive signals and projects them to DG and CA3.
Hence, it is considered that LEC/MEC \rm{I\hspace{-.1em}I } mainly calculates the predictive distribution, and LEC/MEC \rm{I\hspace{-.1em}I\hspace{-.1em}I } calculates the observation likelihood.
Because the MEC has a coordinate system with grid cells, $X$ represents the robot's posture in the environment.
MEC \rm{I\hspace{-.1em}I }, \rm{I\hspace{-.1em}I\hspace{-.1em}I } receives the observation, $y^{\rm POR}$, from the POR and the rotational movement, $u_{t}^{w}$, from the RSC.
Egocentric visual information $y^{\rm PER}$ (e.g., proximal object signal~\citep{Kuruvilla2020}) is sent to $Z^{\rm{I\hspace{-.1em}I }}$, $Z^{\rm{I\hspace{-.1em}I\hspace{-.1em}I }}$ in LEC.
The variable $l$ of LEC V, VI is expected to be latent variables that serve as intermediates to send a generative signal to the PER.
\textbf{Integration of information}:
Allocentric and egocentric information is integrated into the hippocampus.
Hence, DG $H^{\rm DG}$ and CA3 $H^{\rm III}$ form an abstract internal representation of a place that integrates information from the visual information, $Z$, in LEC and the positional information, $X$, in MEC.
We believe that this internal representation corresponds to the spatial concepts.
Furthermore, DG and CA3 are said to have functions of pattern separation and completion, respectively ~\citep{Bakker2008}.
{These functions are crucial in the integration of multimodal information.}
From the PGM perspective, pattern separation may be modeled by parameters that determine the Bayesian prior distribution~\citep{Sanders2020}.
Specifically, the concentration parameter in Dirichlet process clustering is involved in the automatic determination of the number of clusters~\citep{Neal2000a}.
{In short, allocentric and egocentric information are expected to form a cluster within a unified latent space.}
Pattern completion can be viewed as the process of regenerating information from DG as defect completion by resampling from a probability distribution.
Additionally, error correction of self-position is believed to occur in the hippocampus~\citep{Fukawa2020}.
By this process, simultaneously integrated information $H^{\rm DG}$ and $H^{\rm{I\hspace{-.1em}I\hspace{-.1em}I }}$ can be used.
We assume that place-category formation or information processing occurs in the CA1 distal region.
Further, we assume that location-dependent category formation corresponding to place cells occurs in the CA1 proximal region.
\subsection{Consistency of model with scientific knowledge}
\label{sec:hpf-pgm:step3}
This section discusses the consistency of the model using scientific knowledge{ and how the outcome can be tested}.
We believe that the HF-PGM is highly feasible because it is consistent with the anatomical findings of HF, although there may be more detailed variables and dependencies.
The main reason is that the original PGM of SLAM already exists.
We also follow the path integration of MEC and hippocampus, as discussed in \citet{Fukawa2020}.
Furthermore, the agreement with BIF and engineering operating principles is described in Sections~\ref{sec:hpf-pgm:model}.
Consequently, we successfully map the HF as a PGM by introducing the generation-inference process allocation.
In summary, HF-PGM is highly consistent with the brain structure of HF.
The effectiveness of HF-PGM can be verified by solving tasks that can be achieved by integrating LEC (egocentric signal) and MEC (allocentric signal). For example, in situations where localization is difficult in SLAM, we can investigate whether the place category information can distinguish the position. In addition, HF-PGM can be realized as a concrete model by integrating the neural network-based world model and SLAM. Then, further experiments can be conducted to determine whether the latent space representation of the world model, which tends to be unstable in learning, can be complemented by the geometric information of SLAM.
\addspan{
We have also provided a pseudo-source code in the form of Neuro-SERKET architecture~\citep{Taniguchi2020neuro} for HF-PGM as supplementary material (\texttt{HF-PGM\_serket.py}).
}
The implementation of HF-PGM on a robotic platform is a hot research topic for future studies.
Hence, we plan to explore the detailed structure of each model element in subsequent studies.
The HF-PGM proposed in this study is a specimen, and the actual implementation may require more concrete discovery in terms of engineering.
Therefore, the following issues must be mitigated, including
(i) selection of a type for each probability distribution during the generative process,
(ii) when performing amortized variational inference, selection of the function shape of the inference model and that of the architecture of the neural networks,
and
(iii) ensuring a real-time algorithm that includes the learning of global parameters.
The aforementioned issues can be solved by model selection and architecture search/optimization in a framework similar to neural architecture search.
\section{Abstraction as discrete-event queue}
\label{sec:hpf-pgm:queue}
We provide the interpretation of the \textit{phase precession queue assumption}, which is one of the functionalities of the HF as an estimation of the probability distribution in PGM.
As explained in Section~\ref{sec:hpf-pgm:model:pgm-time}, a discrete-event queue based on this assumption was constructed to explain the parts that cannot be expressed by the graphical model structure of the PGM alone.
The phase precession queue assumption plays an important role in the development of an HCD from BIF. Thus, its relationship with the PGM framework is indirect.
In the HF, the theta-phase precession is known to process the experience by discretizing it and compressing it within a time step, as shown in Figure~\ref{fig:hpf_queue}~(a).
Herein, stimuli from the external world are sampled at the period of theta waves (8--13 Hz), and it is believed that the present, past, and future events are encoded in phase \citep{Terada2017}.
By abstracting the information within one phase as a queue, as shown in this subsection, the calculation of the queue can be interpreted as filtering the current state, smoothing the past, and predicting the future.
\subsection{Introducing discrete-event queue}
\label{sec:hpf-pgm:queue:intro}
To comprehensibly model the compressed time process handled by the theta-phase precession, the time granularity of the entire model must be detailed.
To avoid too much complexity, we introduce the phase precession queue assumption, which states that the signal held by the theta-phase precession can be regarded as a time queue containing the past, present, and future events at the current time.
The assumption is as follows:
\begin{quote}
\textbf{Phase precession queue assumption}:
The observed state compressed within one cycle of theta waves circulating on a pentasynaptic loop circuit can be regarded as a discrete-event queue representing the state sampled at discrete time intervals. Here, the pentasynaptic loop circuit is formed in the hippocampus and MEC by the projection sequence MEC\,II-DG-CA3-CA1-Sb-ParaSb-MEC\,II.
\end{quote}
There are several reasons why the aforementioned assumption would be considered reasonable.
As a neuroscientific finding, HF is expected to be modeled as a discrete-event queue with a finite buffer capacity according to the analysis of large-scale network communication in macaque monkeys~\citep{Misic2014}.
Further, in intelligent systems that deal with a world with hidden Markov properties, it is useful to have a discrete-event queue capability to retain the entire observed signal for a short time.
From an engineering perspective, a discrete-event queue can be easily realized as a memory array for the number of time steps to be stored. Here, the memory elements on the same array are considered to have the same meaning, but at different times. However, the neural circuitry of the brain has a restriction that it cannot have more than one representation with the same meaning. Therefore, it seems that we have no choice but to use a method like the theta-phase precession, which compresses information in the time dimension on the same neuron.
The fact that PraSb and MEC II on the pentasynaptic loop circuit receives direct projections from the medial septum, which generates theta rhythms, is consistent with the assumption that this circuit is involved in the theta-phase precession.
\begin{figure}[!tb]
\centering
\includegraphics[width=1.00\textwidth]{queue.pdf}
\caption{
(a) Phase precession queue assumptions and (b) discrete-event queue processing.
(a) shows the information representation in the theta-phase precession of MEC \rm{I\hspace{-.1em}I } when the rat moves from left to right.
$x_{3} | y_{1:3}$ represents an estimate of the state, given the observations up to time $t = 3$.
Fields 2--4 at the bottom of the rat represent the spatial regions where the three grid cells fire in place fields.
The horizontal axis of the bottom table in (b) represents the time of the estimated state, and the vertical axis represents the time of observation.
The red box represents the queue.
The queue has the estimated/predicted states five times from $t-2$ to $t + 2$ at the current time, $ t $.
The vertical axis of the table in (b), which is the observation time, corresponds to the horizontal axis in (a), which is the actual time.
}
\label{fig:hpf_queue}
\end{figure}
\subsection{Processing for discrete-event queue}
\label{sec:hpf-pgm:queue:formulation}
From the aforementioned discussion, the information held in the phase precession can be interpreted as a discrete-event queue.
If this is the case, the process shown in Fig.~\ref{fig:hpf_queue} (b) is performed.
Theoretically, the discrete-event queue can be regarded as a sequential estimation problem for the joint posterior distribution in multiple states.
The variables are the same as those of the PGM of SLAM, as shown in Fig.~\ref{fig:slam_gm}.
Table~\ref{tab:element_of_graphical_model} lists the correspondence of variables with HF-PGM.
Notably, $X_{t}$ in the HF-PGM is a variable with an internal representation equivalent to $x_{t-2:t+2}$.
Figure~\ref{fig:hpf_queue} (b) shows a table presenting a simplified notation for the discrete-event queue.
Each element in the table is a conditional marginal probability distribution in the state, $x_{j}$, at time $j$ under the condition of observations up to time $i$.
Here, the control value, $u_{t}$, and the integrated information from the LEC are omitted.
The probability distribution of the discrete-event queue is shown in Equation (\ref{eq:queue_prob}).
\begin{eqnarray}
\text{Queue}(t) &=& p(x_{t-2:t+2} \mid y_{1:t}), \qquad j \in \{ i-2 \leqq i \leqq i+2 \}, \quad i=t, \nonumber
\\ &=& \eta \underbrace{p(x_{t+2} \mid x_{t+1})}_{\text{Prediction}} \underbrace{p(y_{t} \mid x_{t})}_{\text{Filtering}} \underbrace{p(x_{t-2:t+1} \mid y_{1:t-1})}_{\int \text{Queue}(t-1) dx_{t-3}}, \quad t \geqq 3.
\label{eq:queue_prob}
\end{eqnarray}
Here, we assume that $\text{Queue}(t-1)$ is calculated at the previous time, $t-1$.
$i$ and $ j$ in the formula correspond to those in Fig.~\ref{fig:hpf_queue}.
${\int \text{Queue}(t-1) dx_{t-3}}$ denotes the operation of the integrated-out (i.e., marginalizing)~\citep{murphy2012machine} of $x_{t-3}$ from $\text{Queue}(t-1)$.
As shown in Eq.~(\ref{eq:queue_prob}), $\text{Queue}(t)$ is a recurrence formula expressed by $\text{Queue}(t-1)$, and a sequential calculation similar to the Bayes filter is possible.
The discrete-event queue can be interpreted as an algorithm that combines the filtering with the smoother and the prediction~\citep{kitagawa2014computational}.
This queue calculation is applicable to PGMs of any partially observable Markov decision processes, not just to simple PGMs for SLAM.
Refer to \ref{apdx:pgm_slam:PGM} for the formulae of the predictive/smoothing distributions.
In general online self-localization,
the belief, $ bel (x_{t}) $, shown in \ref{apdx:pgm_slam:SLAM}, which is the distribution when $i=j$, is estimated without using the queue.
{This is called filtering~\citep{kitagawa2014computational,thrun2005probabilistic}.}
The smoothing distribution can correct past self-position estimates from later observations.
SpCoSLAM 2.0~\citep{ataniguchi2020spcoslam2} introduced the fixed-lag smoother for sequential and accurate state estimation.
The predictive distribution predicts the future self-position from the current state and the learning result of the environment, providing a trajectory that avoids obstacles.
Additionally, long-term predictions can be made by repeating predictions ahead by one term.
For example, such predictions have already been realized in a map-based motion model~\citep{thrun2005probabilistic} and a stochastic model predictive control~\citep{Li2019c} for autonomous vehicles.
In addition, predictions related to the generation of spatial behavior are discussed in \ref{apdx:future:navigation}.
\section{Conclusion}
\label{sec:conclusion}
We sought to bridge the findings of HFs in neuroscience and SLAM methods in AI and robotics.
We summarized the SLAM methods in PGMs and the neuroscientific findings of the HFs and investigated their associations.
{This paper presents a case study on the framework reported by \citet{Yamakawa2021-yy}.}
The main contribution of this study is the construction of a PGM for the HF that satisfies the {evaluation} criteria for the BRA design.
Specifically, the BIF was designed to be consistent with the anatomy of the HF, and the HCD was then designed by extending the existing SLAM model to be consistent with the structure of BIF.
{We intend to evaluate the effectiveness of this framework and HF-PGM in future studies.}
In addition, the generation-inference process allocation solved particular problems regarding the mapping of PGMs to brain circuits.
The HF-PGM is significantly different from the previous SLAM models; it integrates LEC and MEC and introduces a discrete-event queue.
Such structures were not found in most SLAMs and are very suggestive in engineering modeling.
This study operates as part of the grand challenge of realizing the whole-brain architecture using PGMs~\citep{Taniguchi2021wb-pgm}.
A whole-brain PGM will be expected to integrate submodules of PGMs for multiple brain regions by using the Neuro-SERKET architecture~\citep{Taniguchi2020neuro}.
In the case of HF-PGM, visual information obtained from other PGM modules corresponding to its surrounding areas, including the visual cortex, could be connected as observed variables.
Other areas of HF connection, such as the prefrontal cortex, were not targeted by HF-PGM in this study.
Integration with a PGM module that mimics such areas can be considered in future studies.
{Other detailed discussions related to open questions and future perspectives are presented in \ref{apdx:future}.}
\section*{Acknowledgement}
This work was partially supported by the Japan Society for the Promotion of Science (JSPS) KAKENHI under Grant JP20K19900 and by the Ministry of Education, Culture, Sports, Science and Technology (MEXT)/JSPS KAKENHI, under Grant JP16H06569 in \#4805 (Correspondence and Fusion of Artificial Intelligence and Brain Science) and JP17H06315 in \#4905 (Brain information dynamics underlying multi-area interconnectivity and parallel processing).
We would like to thank Editage (www.editage.com) for English-language editing.
\bibliographystyle{model5-names}\biboptions{authoryear}
|
\section*{Conflicts of Interest}
\vspace*{-0.1in}
\section*{Acknowledgement}
This work was partially supported by the Natural Sciences and Engineering Research Council (NSERC) of Canada.
\vspace*{-0.1in}
\bibliographystyle{asrl}
\section{Introduction}
Describing rigid-body motion is a fundamental concept in robotics, computer vision, computer graphics, augmented reality, classical mechanics, and beyond. From \citet{mozzi1763} and \citet{chasles1830} we have the {\em Mozzi-Chasles theorem}, which says that the most general rigid-body displacement can be produced by a translation along a line (the {\em screw axis}) followed by (or preceeded by) a rotation about that line. A complete screw theory was later developed by \citet{ball1900}. \citet{murray94} popularized the modern approach to describing rigid-body motion (in the context of manipulator robotics) using the {\em special Euclidean group}, where screw motions are elegantly described by mapping elements of the Lie algebra onto the Lie group through the matrix exponential. It is through this sequence of historical events that multiple communities have come to think of the matrix exponential as the {\em canonical} way to map a six-degree-of-freedom {\em vector} to a rigid-body {\em pose}. Optimization and uncertainty representation for pose quantities now also commonly make use of matrix Lie group tools and the exponential map \citep{chirikjian09,barfoot_tro14,barfoot_ser17}.
Yet, there are many representations of three-dimensional {\em rotations} in active use today: Euler angles, unit-length quaternions (a.k.a., Euler parameters), rotation vectors, Cayley-Gibbs-Rodrigues parameters, modified Rodrigues parameters, and many more \citep{hughes86}. Of these, only the {\em rotation vector} (i.e., angle multiplied with a unit-length axis) maps to an element of the {\em special orthogonal group} (i.e., a rotation matrix) through the matrix exponential.
However, \citet{bauchau03} synthesized the class of {\em vectorial} parameterizations of rotation. In particular, they elegantly describe a general mapping from a three-degree-of-freedom vector to a rotation matrix that encompasses many commonly employed representations of rotation, including Cayley-Gibbs-Rodrigues and modified Rodrigues parameters.
\citet{bauchau03b} and later \citet{bauchau11b} expanded the idea of a general vector mapping to the special Euclidean group representing poses, where they focus on the $6 \times 6$ {\em adjoint} representation of pose, $\mbox{Ad}(SE(3))$. \citet{bauchau11} provides an excellent summary of both the rotation and adjoint pose vectorial mappings. Interestingly, these works do not present the general vector mappings for the common $4 \times 4$ representation of pose, $SE(3)$, although some works do explore special cases \citep{borri00,selig07}. We revisit the idea of a {\em vectorial parameterization of pose}, extending what is known in a few specific ways. First, we explore the more common $4 \times 4$ representation, $SE(3)$, showing that for every parameterization of rotation, there are infinite possibilities for the parameterization of pose having different properties from which one might choose; each has a different coupling between the translational and rotational variables. Second, we show that our general vector mapping provides the special cases of the exponential map and the {\em Cayley transformation} \citep{cayley1846} for $SE(3)$, which has been previously discussed in the literature \citep{borri00, selig07}. Finally, we hope that this paper serves to further shine a light on the important long-term efforts of \citet{bauchau11} on the vectorial parameterizations of rotations and poses, which we believe is of great interest to the robotics community.
Further to our extensions of the vectorial parameterization of pose, which is a theoretical contribution, we also demonstrate how to use the vectorial parameterization of pose in various canonical robotics applications. In particular, the classic problems of pose interpolation, pose control, and pose estimation in a pointcloud alignment framework, are each addressed using different vectorial parameterizations of pose. We show using an alternative pose parameterization over the classic matrix exponential can lead to different results in terms of performance and robustness.
The paper is organized as follows. Section~\ref{sec:math} provides some brief mathematical background. Section~\ref{sec:vecmap} introduces our vectorial parameterization of pose and discusses several properties thereof. While our main contribution is the theory, Section~\ref{sec:apps} provides three applications highlighting that different pose mappings may be beneficial in different situations. Section~\ref{sec:confu} concludes the paper.
\section{Mathematical Preliminaries}\label{sec:math}
We briefly review some key concepts and notation \citep{barfoot_ser17} to prepare for what follows.
\subsection{Matrix Lie Groups}
The {\em special orthogonal group}, representing rotations, is the set of valid rotation matrices:
\begin{equation}
\label{eq:SO3}
SO(3) = \left\{ \mathbf{C} \in \mathbb{R}^{3\times3} \; | \; \mathbf{C} \mathbf{C}^T = \mathbf{1}, \mbox{det} \,\mathbf{C} = 1 \right\},
\end{equation}
where $\mathbf{1}$ is the identity matrix.
It is common to map a vector, $\mbs{\phi} \in \mathbb{R}^3$, to a rotation matrix, $\mathbf{C}$, through the matrix exponential,
\begin{equation}
\mathbf{C}(\mbs{\phi}) = \exp\left( \mbs{\phi}^\wedge \right),
\end{equation}
where $(\cdot)^\wedge$ is the skew-symmetric operator,
\begin{equation}
\mbs{\phi}^\wedge = \begin{bmatrix} \phi_1 \\ \phi_2 \\ \phi_3 \end{bmatrix}^\wedge = \begin{bmatrix} 0 & -\phi_3 & \phi_2 \\ \phi_3 & 0 & -\phi_1 \\ -\phi_2 & \phi_1 & 0 \end{bmatrix}.
\end{equation}
The mapping is surjective-only, meaning every $\mathbf{C}$ can be produced by many different values for $\mbs{\phi}$.
The {\em special Euclidean group}, representing \index{poses} poses (i.e., translation and rotation), is the set of valid transformation matrices:
\begin{equation}
\label{eq:se3}
SE(3) = \left\{ \mathbf{T} = \begin{bmatrix} \mathbf{C} & \mathbf{r} \\ \;\,\mathbf{0}^T & 1 \end{bmatrix} \in \mathbb{R}^{4\times4} \; \Biggl| \; \mathbf{C} \in SO(3), \, \mathbf{r} \in \mathbb{R}^3 \right\}.
\end{equation}
It is again common to map a vector, $\mbs{\xi} \in \mathbb{R}^6$, to a transformation matrix, $\mathbf{T}$, through the matrix exponential,
\begin{equation}
\mathbf{T}(\mbs{\xi}) = \exp\left( \mbs{\xi}^\wedge \right),
\end{equation}
where
\begin{equation}
\mbs{\xi}^\wedge = \begin{bmatrix} \mbs{\rho} \\ \mbs{\phi} \end{bmatrix}^\wedge = \begin{bmatrix} \mbs{\phi}^\wedge & \mbs{\rho} \\ \mathbf{0}^T & 0 \end{bmatrix}.
\end{equation}
As is common practice \citep{barfoot_ser17}, we have broken the pose vector, $\mbs{\xi}$, into a translational component, $\mbs{\rho}$, and a rotational component, $\mbs{\phi}$.
The mapping is also surjective-only, meaning every $\mathbf{T}$ can be produced by many different values for $\mbs{\xi}$.
Finally, the {\em adjoint} of pose is given by
\begin{equation}\label{eq:se3adjointmap1}
\;\mbox{\boldmath ${\cal T}$\unboldmath}(\mbs{\xi}) = \mbox{Ad}\left( \mathbf{T} \right) = \begin{bmatrix} \mathbf{C}(\mbs{\phi}) & \mathbf{r}^\wedge \mathbf{C}(\mbs{\phi}) \\ \mathbf{0} & \mathbf{C}(\mbs{\phi}) \end{bmatrix},
\end{equation}
which is now $6 \times 6$. We will refer to the set of adjoints as $\mbox{Ad}(SE(3))$. We can map a vector, $\mbs{\xi} \in \mathbb{R}^6$, to an adjoint transformation matrix again through the matrix exponential map:
\begin{equation}
\;\mbox{\boldmath ${\cal T}$\unboldmath}(\mbs{\xi}) = \exp\left( \mbs{\xi}^\curlywedge \right),
\end{equation}
where
\begin{equation}
\mbs{\xi}^\curlywedge = \begin{bmatrix} \mbs{\rho} \\ \mbs{\phi} \end{bmatrix}^\curlywedge = \begin{bmatrix} \mbs{\phi}^\wedge & \mbs{\rho}^\wedge \\ \mathbf{0} & \mbs{\phi}^\wedge \end{bmatrix}.
\end{equation}
The mapping is again surjective-only, meaning every $\;\mbox{\boldmath ${\cal T}$\unboldmath}$ can be produced by many different values for $\mbs{\xi}$.
\subsection{Key Identities}
We will be exploring mappings (including the exponential map) from vectors to Lie groups and will have occasion to work with series expressions of rotation and transformation matrices. Each group has associated with it an identity that can be used to limit the number of terms in such series, related to the Cayley-Hamilton theorem.
For rotations, we have that the characteristic equation of $\mbs{\phi}^\wedge$ is
\begin{equation}
\left| \lambda \mathbf{1} - \mbs{\phi}^\wedge \right| = \lambda ( \lambda^2 + \underbrace{\phi_1^2 + \phi_2^2 + \phi_3^2}_{\phi^2} ) = \lambda^3 + \phi^2 \lambda = 0,
\end{equation}
where $\lambda$ are the eigenvalues of $\mbs{\phi}^\wedge$.
From the Cayley-Hamilton theorem we can claim that
\begin{equation} \label{eq:ch1}
\mbs{\phi}^{\wedge^3} + \phi^2 \mbs{\phi}^\wedge \equiv \mathbf{0},
\end{equation}
since $\mbs{\phi}^\wedge$ must satisfy its own characteristic equation.
For poses, we have that the characteristic equation of $\mbs{\xi}^\wedge$ is
\begin{equation}\label{eq:se3char}
\left| \lambda \mathbf{1} - \mbs{\xi}^\wedge \right| = \left| \begin{matrix} \lambda \mathbf{1} - \mbs{\phi}^\wedge & -\mbs{\rho} \\ \mathbf{0}^T & \lambda \end{matrix} \right| = \lambda^4 + \phi^2 \lambda^2 = 0.
\end{equation}
Again, from the Cayley-Hamilton theorem we can claim that
\begin{equation}\label{eq:ch2}
\mbs{\xi}^{\wedge^4} + \phi^2 \mbs{\xi}^{\wedge^2} \equiv \mathbf{0},
\end{equation}
since $\mbs{\xi}^\wedge$ must satisfy its own characteristic equation.
Finally, for adjoint poses we have that the characteristic equation of $\mbs{\xi}^\curlywedge$ is
\begin{equation}\label{eq:adse3char}
\left| \lambda \mathbf{1} - \mbs{\xi}^\curlywedge \right| = \left| \begin{matrix} \lambda \mathbf{1} - \mbs{\phi}^\wedge & -\mbs{\rho}^\wedge \\ \mathbf{0} & \lambda \mathbf{1} - \mbs{\phi}^\wedge \end{matrix} \right| = \left( \lambda^3 + \phi^2 \lambda \right)^2 = \lambda^6 + 2 \phi^2 \lambda^4 + \phi^4 \lambda^2 = 0.
\end{equation}
We could again employ the Cayley-Hamilton theorem to create a similar identity to the other two cases, but it turns out that the {\em minimal polynomial} of $\mbs{\xi}^\curlywedge$ is actually \citep{barfoot_ser17, deleuterio_tro21},
\begin{equation}\label{eq:adid}
\mbs{\xi}^{\curlywedge^5} + 2\phi^2 \mbs{\xi}^{\curlywedge^3} + \phi^4 \mbs{\xi}^\curlywedge \equiv \mathbf{0},
\end{equation}
which is one order lower than the characteristic equation in this case. This is because both the algebraic and geometric multiplicities of eigenvalue $\lambda=0$ in~\eqref{eq:adse3char} are two (two Jordan blocks of size one), allowing us to drop one copy when constructing the minimal polynomial; in~\eqref{eq:se3char} the algebraic multiplicity of $\lambda = 0$ is two, while the geometric multiplicity is one (one Jordan block of size two).
\section{Vector Mappings}\label{sec:vecmap}
In this section, we will show that there exist several options for mapping vectors onto the two most common matrix Lie groups used in robotics and computer vision. Our approach builds on the work of \citet{bauchau03}, who worked this out for $SO(3)$, the group of rotations and \citet{bauchau03b,bauchau11} who extended this work to $\mbox{Ad}(SE(3))$, the adjoint representation of pose. We show how to extend the rotation work to $SE(3)$ directly.
\subsection{Vector Mappings for Rotations}
It is well known that there exist many different ways to parameterize the group of three-dimensional rotations, $SO(3)$. Naturally, such rotations have three degrees of freedom. Representations with three parameters are known to always have singularities while representations with more than three parameters have constraints to keep the degrees of freedom at three \citep{stuelpnagel1964, hughes86}. \citet{bauchau03} synthesized previous work on three-parameter representations that can be considered {\em vectors}; they show that all such parameterizations can be written as
\begin{equation}\label{eq:so3def}
\mbs{\phi} = \phi(\varphi) \, \mathbf{a} \in \mathbb{R}^3,
\end{equation}
where $\mathbf{a}$ is the unit-length axis of rotation, $\varphi$ the angle of rotation, and $\phi(\varphi)$ a {\em generating function}. The scalar generating function is an odd function of the rotation angle that satisfies
\begin{equation}\label{eq:genprop}
\lim_{\varphi \rightarrow 0} \frac{\phi(\varphi)}{\varphi} = \kappa,
\end{equation}
with $\kappa$ a real normalization constant. We will assume $\kappa = 1$ in what follows as this makes it easier to compare different generating functions directly. The generating function is essentially a nonlinear warping of one angle, $\varphi$, to new angle, $\phi$.
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{genfunc.pdf}
\caption{Example generating functions, $\phi(\varphi)$, that map rotation angle, $\varphi$ to different vector representations of rotation and pose. The independent variable is shown on the vertical axis to save space. We see that the $2\tan\frac{\varphi}{2}$ function is quite a bit more extreme than the others}
\label{fig:genfunc}
\end{figure*}
Some common examples of vector parameterizations of rotation are,
\begin{equation}
\underbrace{\mbs{\phi} = \varphi \mathbf{a}}_{\mbox{rotation vector}}, \quad \underbrace{\mbs{\phi} = 2 \tan \frac{\varphi}{2} \mathbf{a}}_{\mbox{Cayley-Gibbs-Rodrigues}}, \quad \underbrace{\mbs{\phi} = 4 \tan \frac{\varphi}{4} \mathbf{a}}_{\mbox{modified Rodrigues}},
\end{equation}
where we have ensured $\kappa = 1$ in~\eqref{eq:genprop} by including normalization constants where necessary.
Notably, the Euler angle sequences, a common three-parameter representation, do not fit the vector definition since they are not parallel to the axis of rotaton. Vector parameterizations come along with singularities but these will not be a roadblock for the way in which we intend to use them. Figure~\ref{fig:genfunc} depicts several generating functions.
\citet{bauchau03} show that the corresponding rotation matrix, $\mathbf{C} \in SO(3)$, can be constructed from the general vector parameterization as follows:
\begin{equation}\label{eq:so3map}
\mathbf{C}(\mbs{\phi}) = \mathbf{1} + \frac{\nu^2}{\epsilon}\mbs{\phi}^\wedge + \frac{\nu^2}{2}\mbs{\phi}^{\wedge^2},
\end{equation}
where
\begin{equation}
\nu(\varphi) = \frac{2 \sin(\varphi/2)}{\phi(\varphi)}, \quad \epsilon(\varphi) = \frac{2 \tan(\varphi/2)}{\phi(\varphi)}.
\end{equation}
The group-mapped-from-vector function is a surjective-only mapping from $\mathbb{R}^3$ to $SO(3)$, which means every rotation matrix can be generated by many vectors (owing to the rotation angle wrapping around).
The expression for the $SO(3)$ Jacobian, $\mathbf{J}(\mbs{\phi})$, that maps the derivative of the vector parameters to angular velocity,
\begin{equation}
\mbs{\omega} = \mathbf{J}(\mbs{\phi}) \dot{\mbs{\phi}},
\end{equation}
is given by \citep{bauchau03}
\begin{equation}\label{eq:so3jac}
\mathbf{J}(\mbs{\phi}) = \mu \mathbf{1} + \frac{\nu^2}{2} \mbs{\phi}^\wedge + \frac{1}{\phi^2}\left( \mu - \frac{\nu^2}{\epsilon}\right) \mbs{\phi}^{\wedge^2},
\end{equation}
where
\begin{equation}
\mu(\varphi) = \left(\frac{d\phi(\varphi)}{d\varphi}\right)^{-1}.
\end{equation}
The inverse of this Jacobian is given by \citep{bauchau03}
\begin{equation}\label{eq:so3jacinv}
\mathbf{J}(\mbs{\phi})^{-1} = \frac{1}{\mu} \mathbf{1} - \frac{1}{2} \mbs{\phi}^\wedge - \frac{1}{\phi^2} \left( \frac{1}{\epsilon} - \frac{1}{\mu} \right)\mbs{\phi}^{\wedge^2},
\end{equation}
which we will require when dealing with $SE(3)$.
As the rotation angle, $\varphi$, becomes small, we have the following infinitesimal expressions:
\begin{equation}
\mathbf{C}(\mbs{\phi}) \approx \mathbf{1} + \mbs{\phi}^\wedge, \quad \mathbf{J}(\mbs{\phi}) \approx \mathbf{1} + \frac{1}{2} \mbs{\phi}^\wedge.
\end{equation}
These will become important when considering perturbations.
The result from this section is general, but it will be instructive to consider some specific examples. Table~\ref{tab:so3vecparams} provides some common vector parameterizations and their associated mappings to $SO(3)$. The first three mappings are appealing in that they do not require calculating the angle of rotation, $\varphi$, and instead can be built directly from $\mbs{\phi}$.
\begin{table*}[t]
\centering
\caption{Example vector parameterizations of rotation and the associated mappings to $SO(3)$. Note the connection to the Cayley transformation in the $\tan$ parameterizations. The entries in the table are ordered from steepest to shallowest generating function (right to left in Figure~\ref{fig:genfunc})}
\begin{tabular}{lcccc}
\hline
& $\phi(\varphi)$ & $\mathbf{C}(\mbs{\phi})$ & validity range \\ \hline
Cayley-Gibbs-Rodrigues & $2 \tan \frac{\varphi}{2}$ & $\left( \mathbf{1} - \frac{1}{2}\mbs{\phi}^\wedge \right)^{-1} \left( \mathbf{1} + \frac{1}{2}\mbs{\phi}^\wedge \right)$ & $| \varphi |< \pi$ \\
modified Rodrigues & $4 \tan \frac{\varphi}{4}$ & $\left( \mathbf{1} - \frac{1}{4}\mbs{\phi}^\wedge \right)^{-2} \left( \mathbf{1} + \frac{1}{4}\mbs{\phi}^\wedge \right)^2$ & $| \varphi | < 2\pi$ \\
rotation vector & $\varphi $ & $\exp\left(\mbs{\phi}^\wedge \right)$ & $| \varphi | < 2\pi$ \\
Bauchau-Trainelli & $4\sin\frac{\varphi}{4}$ & $\mathbf{1} + \cos\frac{\varphi}{2} \cos\frac{\varphi}{4}\mbs{\phi}^\wedge + \frac{1}{2} \cos^2 \frac{\varphi}{4}\mbs{\phi}^{\wedge^2}$ & $| \varphi | < 2\pi$ \\
Euler-Rodrigues & $2 \sin\frac{\varphi}{2}$ & $\mathbf{1} + \cos\frac{\varphi}{2} \mbs{\phi}^\wedge + \frac{1}{2} \mbs{\phi}^{\wedge^2}$ & $| \varphi | < \pi$ \\ \hline
\end{tabular}
\label{tab:so3vecparams}
\end{table*}
\subsection{Vector Mappings for Poses}
We next extend the idea of generalized vector mappings to $SE(3)$. As pose changes have six degrees of freedom, our vector parameterization will take the form
\begin{equation}
\mbs{\xi} = \begin{bmatrix} \mbs{\rho} \\ \mbs{\phi} \end{bmatrix} \in \mathbb{R}^6,
\end{equation}
where the rotational parameters, $\mbs{\phi} \in \mathbb{R}^3$, are the same as for $SO(3)$ and $\mbs{\rho} \in \mathbb{R}^3$ are the translational parameters (not necessarily equivalent to the actual translation in Euclidean space, $\mathbf{r}$). We would like to generalize the $SO(3)$ result such that we have a surjective-only mapping from $\mathbb{R}^6$ to $SE(3)$.
Motivated by the Rodrigues-type form of~\eqref{eq:so3map} above, we define a general vector pose mapping to be of the form
\begin{equation}\label{eq:posemap}
\mathbf{T}(\mbs{\xi}) = \mathbf{1} + a\, \mbs{\xi}^{\wedge} + b\, \mbs{\xi}^{\wedge^2} + c\, \mbs{\xi}^{\wedge^3},
\end{equation}
for some $a$, $b$, and $c$ such that the result is an element of $SE(3)$. We know that higher-order powers of $\mbs{\xi}^\wedge$ are unnecessary owing to~\eqref{eq:ch2}, which comes from the Cayley-Hamilton theorem.
It is not immediately obvious that this definition of a pose vector parameterization for $SE(3)$ is consistent with \citet[\S 14.4.1]{bauchau11}, who notes that we can define such a mapping for $\mbox{Ad}(SE(3))$ as one that is {\em parallel to an eigenvector of the resulting adjoint pose matrix associated with the unit eigenvalue}. It turns out that the two definitions are equivalent and we will demonstrate this shortly after first examining the conditions on $a$, $b$, and $c$ to ensure $\mathbf{T} \in SE(3)$.
Using the definition of $\mbs{\xi}^\wedge$, we can manipulate our pose mapping into the form,
\begin{equation}\label{eq:posemap2}
\mathbf{T}(\mbs{\xi}) = \begin{bmatrix} \mathbf{C}(\mbs{\phi}) & \mathbf{D}(\mbs{\phi}) \mbs{\rho} \\ \mathbf{0}^T & 1 \end{bmatrix} = \begin{bmatrix} \mathbf{1} + (a - \phi^2 c ) \,\mbs{\phi}^\wedge + b \,\mbs{\phi}^{\wedge^2} & \left( a \mathbf{1} + b \,\mbs{\phi}^\wedge + c \,\mbs{\phi}^{\wedge^2} \right) \mbs{\rho} \\ \mathbf{0}^T & 1 \end{bmatrix},
\end{equation}
where we used~\eqref{eq:ch1} to simplify the top-left entry.
Looking to~\eqref{eq:se3}, the top-left entry in~\eqref{eq:posemap2} must be the rotation matrix:
\begin{equation}\label{eq:Cabc}
\mathbf{C}(\mbs{\phi}) = \mathbf{1} + (a - \phi^2 c) \,\mbs{\phi}^\wedge + b \,\mbs{\phi}^{\wedge^2}.
\end{equation}
Comparing to~\eqref{eq:so3map}, the following conditions must hold:
\begin{equation}\label{eq:se3mapconditions}
a - \phi^2 c = \frac{\nu^2}{\epsilon}, \quad b = \frac{\nu^2}{2}.
\end{equation}
The second condition leaves no choice, while the first provides some flexibility.
Turning to the top-right entry of~\eqref{eq:posemap2}, the matrix
\begin{equation}\label{eq:Dabc}
\mathbf{D}(\mbs{\phi}) = a \mathbf{1} + b \,\mbs{\phi}^\wedge + c \,\mbs{\phi}^{\wedge^2},
\end{equation}
is the {\em coupling matrix} between the translational and rotational degrees of freedom. Because there is flexibility in the choice of $a$ and $c$, this means there are an infinite number of possibilities for $\mathbf{D}(\mbs{\phi})$ that satisfy our pose mapping definition in~\eqref{eq:posemap}.
Under the conditions in~\eqref{eq:se3mapconditions}, the defined properties of the generating function, and the further condition that $\mathbf{D}(\mbs{\phi})$ must be invertible (except at some singularities), we can claim that our chosen mapping in~\eqref{eq:posemap} is surjective-only from $\mathbb{R}^6$ to $SE(3)$. This is not too hard to see since every $\mathbf{C}(\mbs{\phi})$ can be generated by many $\mbs{\phi}$ and every $\mathbf{r}$ can be generated by some $\mbs{\rho}$ provided $\mathbf{D}(\mbs{\phi})$ invertible. The inverse map is discussed further in Section~\ref{sec:invmap}.
Returning to the question of consistency of our definition of pose vector parameterizations with that of \citet{bauchau11}, we want to show that $\mbs{\xi}$ {\em is} an eigenvector of $\!\!\;\mbox{\boldmath ${\cal T}$\unboldmath} = \mbox{Ad}(\mathbf{T})$ associated with a unit eigenvalue, i.e.,
\begin{equation}
\;\mbox{\boldmath ${\cal T}$\unboldmath} \mbs{\xi} = \mbs{\xi}.
\end{equation}
Filling in the details of $\!\!\;\mbox{\boldmath ${\cal T}$\unboldmath}$ from~\eqref{eq:se3adjointmap1} we have
\begin{equation}
\;\mbox{\boldmath ${\cal T}$\unboldmath} \mbs{\xi} = \begin{bmatrix} \mathbf{C} & \left( \mathbf{D} \mbs{\rho} \right)^\wedge \mathbf{C} \\ \mathbf{0} & \mathbf{C} \end{bmatrix} \begin{bmatrix} \mbs{\rho} \\ \mbs{\phi} \end{bmatrix} = \begin{bmatrix} \left( \mathbf{C} - \mbs{\phi}^\wedge \mathbf{D} \right) \mbs{\rho} \\ \mbs{\phi} \end{bmatrix},
\end{equation}
where we have used $\mathbf{C} \mbs{\phi} = \mbs{\phi}$. Using~\eqref{eq:Cabc},~\eqref{eq:Dabc}, and~\eqref{eq:ch1} we see that $\mathbf{C} - \mbs{\phi}^\wedge \mathbf{D} = \mathbf{1}$ so that $\!\!\;\mbox{\boldmath ${\cal T}$\unboldmath} \mbs{\xi} = \mbs{\xi}$. Thus, our $SE(3)$ definition of pose vector parameterizations is equivalent to that of \citet{bauchau11} for $\mbox{Ad}(SE(3))$.
The nonuniqueness of the pose parameterization that we discussed earlier can also be explained from the eigenstructure, since there are two linearly independent eigenvectors for $\!\!\;\mbox{\boldmath ${\cal T}$\unboldmath}$ associated with its two unit eigenvalues \citep{bauchau11,deleuterio_tro21}; the $SE(3)$ pose vector parameterization therefore has a free parameter, as we showed by another means.
\subsection{Screws, Pose Adjoints, and Commutative Mappings}
As discussed in the previous section, we must choose values for the parameters, $a$, $b$, and $c$ to construct the rotation matrix in~\eqref{eq:Cabc}. However, we have a free parameter in making this choice, as long as the parameters satisfy~\eqref{eq:se3mapconditions}. One possibility is to choose
\begin{equation}\label{eq:abc}
a = \mu, \quad b = \frac{\nu^2}{2}, \quad c = \frac{1}{\phi^2}\left( \mu - \frac{\nu^2}{\epsilon}\right),
\end{equation}
which satisfy the conditions in~\eqref{eq:se3mapconditions} and importantly this particular choice means that we can write our vector pose parameter as \citep[eq. 14.34]{bauchau11}
\begin{equation}\label{eq:screw}
\mbs{\xi} = \phi(\varphi) \, \mathbf{s},
\end{equation}
where $\mathbf{s}$ is the constant {\em screw axis} of the Mozzi-Chasles theorem. The screw can be written in the form
\begin{equation}
\mathbf{s} = \begin{bmatrix} p \,\mathbf{a} + \mathbf{m}^\wedge \mathbf{a} \\ \mathbf{a} \end{bmatrix},
\end{equation}
where $p$ is the (scalar) {\em pitch}, $\mathbf{a}$ is the (unit) axis of rotation, and $\mathbf{m}$ is the moment arm; $(\mathbf{m}^\wedge \mathbf{a}, \mathbf{a})$ are the {\em Pl\"{u}cker coordinates}. The relationship in~\eqref{eq:screw} parallels the $SO(3)$ definition of \citep{bauchau03} captured in~\eqref{eq:so3def}; we have the product of the generating function and a (screw) axis. In the case that $\phi(\varphi) = \varphi$, then $\mbs{\xi} = \varphi \mathbf{s}$ is the familiar {\em twist}. We underscore that it is the specific selection of $a$ and $c$ in~\eqref{eq:abc} that has made this connection; other choices for $a$ and $c$ will not result in a constant screw.
The selection of $a$ and $c$ in~\eqref{eq:abc} means that the pose mapping becomes
\begin{equation}\label{eq:se3map2}
\mathbf{T}(\mbs{\xi}) = \mathbf{1} + \mu \, \mbs{\xi}^{\wedge} + \frac{\nu^2}{2} \mbs{\xi}^{\wedge^2} + \frac{1}{\phi^2}\left( \mu - \frac{\nu^2}{\epsilon}\right) \mbs{\xi}^{\wedge^3}.
\end{equation}
As the rotation angle, $\varphi$, becomes small, we have the infinitesimal expression,
\begin{equation}
\mathbf{T}(\mbs{\xi}) \approx \mathbf{1} + \mbs{\xi}^\wedge,
\end{equation}
which will become important when considering perturbations.
Choosing $a$ and $c$ as in~\eqref{eq:abc} also has the effect of selecting
\begin{equation}
\mathbf{D}(\mbs{\phi}) = \mathbf{J}(\mbs{\phi}),
\end{equation}
the Jacobian of $SO(3)$ for the chosen parameterization \citep{bauchau11}, which can be verified by substituting into~\eqref{eq:so3jac}.
Choosing $\mathbf{D}(\mbs{\phi}) = \mathbf{J}(\mbs{\phi})$ gives our $SE(3)$ mapping an important property related to pose {\em adjoints}.
With these definitions, we can establish the following desirable commutative relationship:
\begin{equation}
\begin{CD}
4 \times 4@. @. \mbs{\xi}^\wedge \in \frak{se}(3) @>\mbox{mapping}>> \mathbf{T} \in SE(3) \\
@. \qquad @. @VV\mbox{ad}V @VV\mbox{Ad}V \\
6 \times 6 @. @. \mbs{\xi}^\curlywedge \in \mbox{ad}(\frak{se}(3)) @>\mbox{mapping}>> \;\mbox{\boldmath ${\cal T}$\unboldmath} \in \mbox{Ad}(SE(3))
\end{CD}
\end{equation}
This means we can transform our vector parameterization of pose to its adjoint in two equivalent ways.
To verify this commutative property, we can write the series form for the adjoint mapping as
\begin{equation}\label{eq:se3adjointmap2}
\;\mbox{\boldmath ${\cal T}$\unboldmath}(\mbs{\xi}) = \mathbf{1} + d \, \mbs{\xi}^\curlywedge + e \, \mbs{\xi}^{\curlywedge^2} + f \, \mbs{\xi}^{\curlywedge^3} + g \, \mbs{\xi}^{\curlywedge^4},
\end{equation}
for some unknowns, $d$, $e$, $f$, $g$. We do not need higher-order terms in this Rodrigues-type series owing to~\eqref{eq:adid}, which can be used to reduce quintic and higher terms. Multiplying out and comparing terms between~\eqref{eq:se3adjointmap1} and~\eqref{eq:se3adjointmap2}, we find that
\begin{equation}\label{eq:defg}
d = a + \phi^2 (f - c), \quad e = \phi^2 g + b, \quad f = \frac{1}{2}\left( ab-c \right), \quad g = \frac{1}{2} \left( b^2 - c(a-\phi^2 c) \right),
\end{equation}
in terms of the earlier choices for $a$, $b$, and $c$. There may also be other choices for $a$ and $c$ (recall these are not unique) that impose the commutative relationship for $SE(3)$, but this was not explored further.
Owing to smoothness of the generating function and its convergence to $\varphi$ in the limit of an infinitesimally small rotation, one might be able to argue that our vector mappings offer alternatives (to the exponential map) for the local diffeomorphism from the Lie algebra, $\frak{se}(3)$, to the Lie group, $SE(3)$. In such a case, by changing the map we are merely `warping' how the Lie algebra maps onto the Lie group.
\subsection{Cayley Transformations}
It is known \citep{borri00,bauchau03} that for the Cayley-Gibbs-Rodrigues parameterization of rotation, $\phi(\varphi) = 2 \tan \frac{\varphi}{2}$, we can write the rotation matrix in terms of the Cayley transformation:
\begin{equation}
\mathbf{C}(\mbs{\phi}) = \left( \mathbf{1} - \frac{1}{2}\mbs{\phi}^\wedge \right)^{-1} \left( \mathbf{1} + \frac{1}{2}\mbs{\phi}^\wedge \right).
\end{equation}
In fact, this can be generalized to the case where $\phi(\varphi) = 2m \tan \frac{\varphi}{2m}$ with $m$ a positive integer using higher-order Cayley transformations \citep{bauchau03}:
\begin{equation}
\mathbf{C}(\mbs{\phi}) = \left( \mathbf{1} - \frac{1}{2m}\mbs{\phi}^\wedge \right)^{-m} \left( \mathbf{1} + \frac{1}{2m}\mbs{\phi}^\wedge \right)^m.
\end{equation}
However, things become somewhat more complicated for poses.
\citet{borri00} and later \citet{selig07} demonstrated that the Cayley transformation can be used to map pose vectors to $SE(3)$ according to
\begin{equation}\label{eq:se3cay}
\mathbf{T}(\mbs{\xi}) = \left( \mathbf{1} - \frac{1}{2m}\mbs{\xi}^\wedge \right)^{-m} \left( \mathbf{1} + \frac{1}{2m}\mbs{\xi}^\wedge \right)^m,
\end{equation}
where we have generalized to higher-order transformations.
\citet{borri00} and \citet{selig07} also show that the Cayley transformation can be used to map pose vectors to $\mbox{Ad}(SE(3))$ according to
\begin{equation}\label{eq:se3Adcay}
\;\mbox{\boldmath ${\cal T}$\unboldmath}(\mbs{\xi}) = \left( \mathbf{1} - \frac{1}{2m}\mbs{\xi}^\curlywedge \right)^{-m} \left( \mathbf{1} + \frac{1}{2m}\mbs{\xi}^\curlywedge \right)^m,
\end{equation}
where we have generalized again to higher-order transformations. Disappointingly, \citet{selig07} demonstrates that starting from the same $\mbs{\xi}$ and applying~\eqref{eq:se3cay} and~\eqref{eq:se3Adcay} does not result in an equivalent transformation, i.e., $\;\mbox{\boldmath ${\cal T}$\unboldmath}(\mbs{\xi}) \neq \mbox{Ad}(\mathbf{T}(\mbs{\xi}))$; the commutative property for adjoints does not hold.
We can actually say something more than \citet{selig07}. It turns out that~\eqref{eq:se3cay} is {\em not} equivalent to~\eqref{eq:se3map2}, but~\eqref{eq:se3Adcay} {\em is} equivalent to~\eqref{eq:se3adjointmap2}, for the family of rotation parameterizations, $\phi(\varphi) = 2m \tan \frac{\varphi}{2m}$. The explanation for this lies in the choice of $\mathbf{D}(\mbs{\phi})$, discussed earlier. For the $6 \times 6$ case, choosing $\mathbf{D}(\mbs{\phi}) = \mathbf{J}(\mbs{\phi})$ results in the Cayley transformation, while for the $4 \times 4$ case it does not; to connect to the Cayley transformation in the $4 \times 4$ case for $m=1$, we must pick $\mathbf{D}(\mbs{\phi}) = \frac{1}{2}(\mathbf{C}(\mbs{\phi}) + \mathbf{1}) \neq \mathbf{J}(\mbs{\phi})$, as shown by \citet{selig07}.
To show that~\eqref{eq:se3Adcay} holds, consider the $m=1$ case. If we left-multiply by the inverse of the first factor we should have
\begin{equation}
\left( \mathbf{1} - \frac{1}{2}\mbs{\xi}^\curlywedge \right) \;\mbox{\boldmath ${\cal T}$\unboldmath}(\mbs{\xi}) = \mathbf{1} + \frac{1}{2}\mbs{\xi}^\curlywedge.
\end{equation}
Multiplying out the left side we have
\begin{equation}
\mathbf{1} + \left(d-\frac{1}{2}(1-\phi^4 g)\right) \mbs{\xi}^\curlywedge + \left(e-\frac{1}{2}d\right) \mbs{\xi}^{\curlywedge^2} + \left(f-\frac{1}{2}(e-2\phi^2 g)\right) \mbs{\xi}^{\curlywedge^3} + \left(g-\frac{1}{2}f \right) \mbs{\xi}^{\curlywedge^4} = \mathbf{1} + \frac{1}{2}\mbs{\xi}^\curlywedge,
\end{equation}
where we have made use of~\eqref{eq:adid} once to reduce a quintic term. Comparing the coefficients of each term it must be that
\begin{equation}
d-\frac{1}{2}(1-\phi^4 g) = \frac{1}{2}, \quad e-\frac{1}{2}d = 0, \quad f-\frac{1}{2}(e-2\phi^2 g) = 0, \quad g-\frac{1}{2}f = 0.
\end{equation}
Substituting~\eqref{eq:defg} and then~\eqref{eq:abc} confirms these conditions for $\phi(\varphi) = 2 \tan \frac{\varphi}{2}$.
\subsection{Cayley-Type Factorizations}
Following the previous section that focussed on the $\phi(\varphi) = 2m \tan \frac{\varphi}{2m}$ family, we next consider some novel Cayley-type mappings for different generating functions. To our knowledge these have not been presented before.
It turns out that we can factor {\em any} of the vectorial rotation parameterizations according to
\begin{equation}
\mathbf{C}(\mbs{\phi}) = \left( \mathbf{1} - \lambda \mbs{\phi}^\wedge \right)^{-1} \left( \mathbf{1} + \gamma \mbs{\phi}^\wedge \right) = \left( \mathbf{1} - \gamma \mbs{\phi}^\wedge \right)^{-1} \left( \mathbf{1} + \lambda \mbs{\phi}^\wedge \right),
\end{equation}
where
\beqn{lamgam}
\gamma & = & \frac{1}{2} \epsilon, \\
\lambda & = & \frac{\nu^2}{\epsilon} - \frac{1}{2} \epsilon \left( 1 - \frac{1}{2} \phi^2 \nu^2 \right).
\end{eqnarray}\end{subequations}
In the case that $\phi(\varphi) = 2 \tan \frac{\varphi}{2}$ we recover $\gamma = \lambda = \frac{1}{2}$. The inverse of $\mathbf{C}(\mbs{\phi})$ is obtained by flipping the signs in the factorization:
\begin{equation}
\mathbf{C}(\mbs{\phi})^{-1} = \mathbf{C}(\mbs{\phi})^T = \left( \mathbf{1} + \lambda \mbs{\phi}^\wedge \right)^{-1} \left( \mathbf{1} - \gamma \mbs{\phi}^\wedge \right) = \left( \mathbf{1} + \gamma \mbs{\phi}^\wedge \right)^{-1} \left( \mathbf{1} - \lambda \mbs{\phi}^\wedge \right).
\end{equation}
The relationships can be verified by simply multiplying out and comparing to~\eqref{eq:so3map}.
We can also factor some $SE(3)$ parameterizations in an analogous way:
\begin{subequations}
\begin{eqnarray}
\mathbf{T}(\mbs{\xi}) & = & \left( \mathbf{1} - \lambda \mbs{\xi}^\wedge \right)^{-1} \left( \mathbf{1} + \gamma \mbs{\xi}^\wedge \right) = \left( \mathbf{1} - \gamma \mbs{\xi}^\wedge \right)^{-1} \left( \mathbf{1} + \lambda \mbs{\xi}^\wedge \right), \\
\mathbf{T}(\mbs{\xi})^{-1} & = & \left( \mathbf{1} + \lambda \mbs{\xi}^\wedge \right)^{-1} \left( \mathbf{1} - \gamma \mbs{\xi}^\wedge \right) = \left( \mathbf{1} + \gamma \mbs{\xi}^\wedge \right)^{-1} \left( \mathbf{1} - \lambda \mbs{\xi}^\wedge \right),
\end{eqnarray}
\end{subequations}
where $\gamma$ and $\lambda$ are again given by~\eqref{eq:lamgam}. To achieve this factorization, we must select $c = \frac{1}{4} \nu^2 \epsilon$ in~\eqref{eq:se3mapconditions}, which consequently selects a $\mathbf{D}(\mbs{\phi}) \neq \mathbf{J}(\mbs{\phi})$.
Finally, the pattern holds for $\mbox{Ad}(SE(3))$ where we have
\beqn{}
\;\mbox{\boldmath ${\cal T}$\unboldmath}(\mbs{\xi}) & = & \left( \mathbf{1} - \lambda \mbs{\xi}^\curlywedge \right)^{-1} \left( \mathbf{1} + \gamma \mbs{\xi}^\curlywedge \right) = \left( \mathbf{1} - \gamma \mbs{\xi}^\curlywedge \right)^{-1} \left( \mathbf{1} + \lambda \mbs{\xi}^\curlywedge \right), \\
\;\mbox{\boldmath ${\cal T}$\unboldmath}(\mbs{\xi})^{-1} & = & \left( \mathbf{1} + \lambda \mbs{\xi}^\curlywedge \right)^{-1} \left( \mathbf{1} - \gamma \mbs{\xi}^\curlywedge \right) = \left( \mathbf{1} + \gamma \mbs{\xi}^\curlywedge \right)^{-1} \left( \mathbf{1} - \lambda \mbs{\xi}^\curlywedge \right),
\end{eqnarray}\end{subequations}
where $\gamma$ and $\lambda$ are yet again given by~\eqref{eq:lamgam}. To achieve this, we must select
\begin{equation}
d = \phi^2 f + \frac{\nu^2}{\epsilon}, \quad e = \frac{1}{2} \epsilon \phi^2 f + \frac{\nu^2}{2}, \quad f = \frac{\nu^2 \epsilon}{4 + \phi^2 \epsilon^2}, \quad g = \frac{1}{2}\epsilon f,
\end{equation}
in~\eqref{eq:se3adjointmap2}. As discussed in the previous section, when $\phi(\varphi) = 2 \tan \frac{\varphi}{2}$ this happens to select $\mathbf{D}(\mbs{\phi}) = \mathbf{J}(\mbs{\phi})$, but this is not true for the other generating functions.
\subsection{Inverse Mappings}\label{sec:invmap}
We briefly discuss how to go the other way from $\mathbf{T}$, or $\;\mbox{\boldmath ${\cal T}$\unboldmath}$, back to $\mbs{\xi}$ in the general case. The first step is to extract $\mathbf{C}(\mbs{\phi})$ and $\mathbf{r}$ from $\mathbf{T}$, or $\;\mbox{\boldmath ${\cal T}$\unboldmath}$. From $\mathbf{C}(\mbs{\phi})$, we can exploit the facts that
\begin{equation}
\mbox{tr}(\mathbf{C}(\mbs{\phi})) = 2 \cos\varphi + 1, \quad \mathbf{C}(\mbs{\phi}) \mathbf{a} = \mathbf{a},
\end{equation}
which are true independent of the chosen vector parameterization. The first gives us the rotation angle, $\varphi$, while the second is an eigenproblem that can be solved for axis $\mathbf{a}$ (unit-length eigenvector corresponding to eigenvalue of $1$); see \citet{hughes86} for an explicit formula. Owing to the singularity of the vector parameterization, we choose $\varphi \in (-\pi,\pi]$. We can then calculate $\mbs{\phi} = \phi(\varphi) \, \mathbf{a}$.
Once we have $\mbs{\phi}$, we can calculate $\mbs{\rho}$ according to
\begin{equation}
\mbs{\rho} = \mathbf{D}(\mbs{\phi})^{-1} \mathbf{r},
\end{equation}
and then assemble $\mbs{\rho}$ and $\mbs{\phi}$ into $\mbs{\xi}$. In the case of $ \mathbf{D}(\mbs{\phi}) = \mathbf{J}(\mbs{\phi})$, the inverse, $\mathbf{J}(\mbs{\phi})^{-1}$, is given by~\eqref{eq:so3jacinv} and naturally has singularities of which to be aware.
\subsection{Compounding Poses}
For $SO(3)$, we seek to compound two rotations such that
\begin{equation}
\mathbf{C}(\mbs{\phi}) = \mathbf{C}(\mbs{\phi}_2) \mathbf{C}(\mbs{\phi}_1).
\end{equation}
\citet{bauchau03} show that the vector parameterization can be directly compounded according to
\beqn{}
\varphi & = & 2 \cos^{-1} \left( \nu(\varphi_1)\nu(\varphi_2) \left( \frac{1}{\epsilon(\varphi_1)\epsilon(\varphi_2)} - \frac{1}{4} \, \mbs{\phi}_1^T \mbs{\phi}_2 \right)\right), \\
\mbs{\phi} & = & \frac{\nu(\varphi_1) \nu(\varphi_2)}{\nu(\varphi)} \left( \frac{1}{\epsilon(\varphi_2)}\mbs{\phi}_1 + \frac{1}{\epsilon(\varphi_1)}\mbs{\phi}_2 - \frac{1}{2} \mbs{\phi}_1^\wedge \mbs{\phi}_2 \right) .
\end{eqnarray}\end{subequations}
The first of these is used to calculate the compound rotation angle, $\varphi$, while the second provides the complete vectorial parameterization, $\mbs{\phi}$.
For $SE(3)$, \citet{condurache20} recently discussed a closed-form solution for compounding two pose vectors in the case that $\mbs{\phi} = \varphi \mathbf{a}$. \citet{bauchau03b} showed that we can easily generalize this result to any of our vectorial pose parameterizations. The compounding of two poses is
\begin{multline}
\mbf{T}(\mbs{\xi}) = \begin{bmatrix} \mathbf{C}(\mbs{\phi}) & \mathbf{D}(\mbs{\phi}) \mbs{\rho} \\ \mathbf{0}^T & 1 \end{bmatrix} = \mbf{T}(\mbs{\xi}_2) \mbf{T}(\mbs{\xi}_1) = \begin{bmatrix} \mathbf{C}(\mbs{\phi}_2) & \mathbf{D}(\mbs{\phi}_2) \mbs{\rho}_2 \\ \mathbf{0}^T & 1 \end{bmatrix} \begin{bmatrix} \mathbf{C}(\mbs{\phi}_1) & \mathbf{D}(\mbs{\phi}_1) \mbs{\rho}_1 \\ \mathbf{0}^T & 1 \end{bmatrix} \\ = \begin{bmatrix} \mathbf{C}(\mbs{\phi}_2) \mathbf{C}(\mbs{\phi}_1) & \mathbf{C}(\mbs{\phi}_2) \mathbf{D}(\mbs{\phi}_1)\mbs{\rho}_1 + \mathbf{D}(\mbs{\phi}_2) \mbs{\rho}_2 \\ \mathbf{0}^T & 1 \end{bmatrix} .
\end{multline}
Comparing entries we see that
\beqn{}
\mathbf{C}(\mbs{\phi}) & =& \mathbf{C}(\mbs{\phi}_2) \mathbf{C}(\mbs{\phi}_1), \\
\mbs{\rho} & = & \mathbf{D}(\mbs{\phi})^{-1} \mathbf{C}(\mbs{\phi}_2) \mathbf{D}(\mbs{\phi}_1)\mbs{\rho}_1 + \mathbf{D}(\mbs{\phi})^{-1} \mathbf{D}(\mbs{\phi}_2) \mbs{\rho}_2 .
\end{eqnarray}\end{subequations}
We can use the approach of \citet{bauchau03} to calculate the compound rotation, $\mbs{\phi}$, and then calculate $\mbs{\rho}$ afterwards. We require an expression for $ \mathbf{D}(\mbs{\phi})^{-1}$, which in the case of $ \mathbf{D}(\mbs{\phi}) = \mathbf{J}(\mbs{\phi})$ is given by~\eqref{eq:so3jacinv}.
\subsection{Additional Useful Properties of Vector Mappings}
There are a few other important properties of our vector mappings of which we frequently make use when manipulating expressions in state-estimation problems. Owing to the well-known identity \citep{barfoot_ser17},
\begin{equation}
\left( \mathbf{C} \mathbf{v} \right)^\wedge \equiv \mathbf{C} \mathbf{v}^\wedge \mathbf{C}^T,
\end{equation}
where $\mathbf{C} \in SO(3)$ and $\mathbf{v} \in \mathbb{R}^3$, we have that
\begin{equation}
\mathbf{C}_1 \mathbf{C}_2(\mbs{\phi}_2) \mathbf{C}_1^T \equiv \mathbf{C}_2\left( \mathbf{C}_1 \mbs{\phi}_2 \right).
\end{equation}
The proof is as follows:
\begin{multline}\label{eq:Cv}
\mathbf{C}_1 \mathbf{C}_2(\mbs{\phi}_2) \mathbf{C}_1^T = \mathbf{C}_1 \left( \mathbf{1} + \alpha \, \mbs{\phi}_2^\wedge + \beta \mbs{\phi}_2^{\wedge^2} \right) \mathbf{C}_1^T = \mathbf{1} + \alpha \, \mathbf{C}_1 \mbs{\phi}_2^\wedge \mathbf{C}_1^T + \beta \, \mathbf{C}_1 \mbs{\phi}_2^\wedge \mathbf{C}_1 \, \mathbf{C}_1^T \mbs{\phi}_2^\wedge \mathbf{C}_1^T \\
= \mathbf{1} + \alpha \left( \mathbf{C}_1 \mbs{\phi}_2 \right)^\wedge + \beta \left( \mathbf{C}_1 \mbs{\phi}_2 \right)^{\wedge^2} = \mathbf{C}_2\left( \mathbf{C}_1 \mbs{\phi}_2 \right),
\end{multline}
where $\alpha$ and $\beta$ are standing in for the coefficients in~\eqref{eq:so3map}.
For $SE(3)$, we can use the identities \citep{barfoot_ser17},
\begin{gather}
\left( \!\! \;\mbox{\boldmath ${\cal T}$\unboldmath} \mathbf{x} \right)^\wedge \equiv \mathbf{T} \mathbf{x}^\wedge \mathbf{T}^{-1}, \\ \left( \!\!\;\mbox{\boldmath ${\cal T}$\unboldmath} \mathbf{x} \right)^\curlywedge \equiv \;\mbox{\boldmath ${\cal T}$\unboldmath} \mathbf{x}^\curlywedge \;\mbox{\boldmath ${\cal T}$\unboldmath}^{-1}
\end{gather}
with $\mathbf{T} \in SE(3)$, $\;\mbox{\boldmath ${\cal T}$\unboldmath} = \mbox{Ad}(\mathbf{T})$, $\mathbf{x} \in \mathbb{R}^6$ to establish
\begin{gather}
\mathbf{T}_1 \mathbf{T}_2(\mbs{\xi}_2) \mathbf{T}_1^{-1} \equiv \mathbf{T}_2 \left( \!\!\;\mbox{\boldmath ${\cal T}$\unboldmath}_1 \mbs{\xi}_2 \right), \\ \;\mbox{\boldmath ${\cal T}$\unboldmath}_1 \;\mbox{\boldmath ${\cal T}$\unboldmath}_2(\mbs{\xi}_2) \;\mbox{\boldmath ${\cal T}$\unboldmath}_1^{-1} \equiv \;\mbox{\boldmath ${\cal T}$\unboldmath}_2 \left( \!\!\;\mbox{\boldmath ${\cal T}$\unboldmath}_1 \mbs{\xi}_2 \right),
\end{gather}
where the proofs are similar to~\eqref{eq:Cv}.
\section{Applications}\label{sec:apps}
In this section, we touch on three classic applications of pose vectors: (i) interpolation between two poses (useful in computer graphics, for example), (ii) a servoing control problem, and (iii) the alignment of two pointclouds. In each, we show how different choices for the vector parameterization can produce different results.
\subsection{Interpolation}
\begin{figure*}[t]
\centering
\includegraphics[width=0.95\textwidth]{example1.pdf}
\caption{An example of `linear' interpolation between an initial pose (green) and final pose (red) for different vector parameterizations of pose. The paths are straight lines in $\mathbb{R}^6$ but when mapped to $SE(3)$ produce different curves. Each parameterization takes a unique path with the $\phi(\varphi) = 2\tan\frac{\varphi}{2}$ generating function being the most distinct. The figure compares the path taken by each of the generating functions from Figure~\ref{fig:genfunc} (same colour scheme) as well as displays the full pose trajectory individually for three generating functions}
\label{fig:example1}
\end{figure*}
Pose interpolation has been discussed extensively in the literature; \citet{han16,han18} provide some representative examples. Here our goal is not to claim a new method of pose interpolation but rather to show the effect of simply changing the generating function on a basic interpolation scheme defined in the vectorspace.
We can define a `linear' interpolation scheme for poses using any of our vector parameterizations:
\begin{equation}
\mbs{\xi} = (1-s) \, \mbs{\xi}_{\rm final} + s \, \mbs{\xi}_{\rm initial},
\end{equation}
where $\mbs{\xi}_{\rm initial}$ is an initial pose, $\mbs{\xi}_{\rm final}$ a final pose, and $s \in [0,1]$ the interpolation variable. As we vary $s$ from $0$ to $1$, the vector, $\mbs{\xi}$ will vary and so too will $\mathbf{T}(\mbs{\xi})$. This allows the smooth transition from one pose to another.
Figure~\ref{fig:example1} provides a qualitative example of this scheme for all five generating functions, $\phi(\varphi)$, listed in Table~\ref{tab:so3vecparams} with $\mathbf{D}(\mbs{\phi}) = \mathbf{J}(\mbs{\phi})$. The paths taken each correspond to a {\em geodesic} in $\mathbb{R}^6$, but when mapped to $SE(3)$ are quite different. Notably, none of the paths are straight lines in Euclidean space owing to the coupling between translation and rotation. As the rotational motion becomes smaller, the paths followed by the different generating functions look more similar owing to~\eqref{eq:genprop}. Looking more closely at Figure~\ref{fig:example1}, we can also observe that the spacing of the steps are quite different for the different generating functions, even though we stepped $s$ uniformly. The $\tan$ generating functions take larger steps at the beginning but smaller ones at the end, while the $\sin$ ones do the opposite, and the $\phi(\varphi) = \varphi$ function keeps the steps even throughout. Naturally, we can make the steps in the interpolation variable, $s$, as small as we like to obtain smooth interpolation; we chose fairly large steps to keep the plot in Figure~\ref{fig:example1} from becoming too cluttered.
\subsection{Pose Servoing Control}
\begin{figure*}[t]
\centering
\includegraphics[width=0.75\textwidth]{example2.pdf}
\caption{An example of pose servoing control using a `linear' controller in pose vector space. Once mapped to $SE(3)$ we see different behaviours from the different pose mappings. In this example, the reference pose is moving along the straight black line and there is an initial error. The figure compares the path taken to join the reference by each of the generating functions from Figure~\ref{fig:genfunc} (same colour scheme) as well as displays the full pose trajectory individually for three generating functions}
\label{fig:example2}
\end{figure*}
Pose servoing is also a well studied topic. Here we simply want to show the effect of changing the pose mapping on a simple control law defined in the vectorspace. We define a `linear' control law to servo onto a reference path. Consider a reference pose, $\mathbf{T}_{\rm ref}$, that we want to track. We can compare our current pose, $\mathbf{T}(\mbs{\xi})$, to this reference and put it through the inverse mapping to create an `error vector':
\begin{equation}
\mbs{\xi}_{\rm error} \leftarrow \mathbf{T}(\mbs{\xi}) \mathbf{T}_{\rm ref}^{-1},
\end{equation}
where the left arrow stands in for the inverse of any of our pose mappings. We can then adjust our pose to be slightly closer to the reference pose (which is moving) using a simple control law:
\begin{equation}
\dot{\mbs{\xi}} = -\kappa \, \mbs{\xi}_{\rm error}.
\end{equation}
with $\kappa >0$ a gain. As the error $\mbs{\xi}_{\rm error}$ approaches zero, the moving pose, $\mathbf{T}(\mbs{\xi})$ will approach the reference pose, $\mathbf{T}_{\rm ref}$.
Figure~\ref{fig:example2} provides an example of servoing to a reference pose that is moving straight along the $x$-axis, starting from a poor initial pose that is translated and rotated from the reference. We kept the vector control law (including the gain) identical and varied only the choice of pose mapping. As expected, we see different transient behaviour resulting from the different pose mappings as they seek the reference.
\subsection{Pointcloud Alignment}
We use this section to discuss aligning two clouds of three-dimensional points. Our problem is similar to that of \citet{barfoot_aa11}, but formulated directly on $SE(3)$.
\subsubsection{Problem Setup}
Let us consider that we have two sets of observations of some three-dimensional points, $\mathbf{p}_j$ and $\mathbf{q}_j$, expressed in two different reference frames. We assume the points are expressed in homogeneous format,
\begin{equation}
\mathbf{p} = \begin{bmatrix} \mathbf{v} \\ 1 \end{bmatrix}, \quad \mathbf{v} = \begin{bmatrix} x \\ y \\ z \end{bmatrix} .
\end{equation}
Our goal is to find the rigid-body transformation, $\mathbf{T}$, that minimizes the following sum-of-squares cost function,
\begin{equation}
J(\mathbf{T}) = \frac{1}{2} \sum_j \mathbf{e}_j^T \mathbf{W}_j \mathbf{e}_j,
\end{equation}
where
\begin{equation}
\mathbf{e}_j = \mathbf{q}_j - \mathbf{T} \, \mathbf{p}_j,
\end{equation}
and $\mathbf{W}_j > 0$ is a matrix weight for each associated point pair.
\subsubsection{General Perturbation Solution}\label{sec:genper}
Owing to the matrix weights, there is no closed-form solution to this optimization problem and we must turn to an iterative scheme \citep{barfoot_ser17}. If we have some initial guess (or operating point) for our pose solution, $\mathbf{T}_{\rm op}$, we can seek an optimal perturbation, $\delta\mathbf{T}(\mbs{\xi})$, that iteratively reduces our cost function. Any of the pose mappings discussed earlier could be used in this regard:
\begin{equation}
\mathbf{T} = \delta\mathbf{T}(\mbs{\xi}) \, \mathbf{T}_{\rm op} \approx \left( \mathbf{1} + \mbs{\xi}^\wedge \right) \mathbf{T}_{\rm op},
\end{equation}
where notably the infinitesimal expression is the same for all our pose mappings.
Inserting this into the error we have
\begin{equation}
\mathbf{e}_j \approx \mathbf{q}_j - \left( \mathbf{1} + \mbs{\xi}^\wedge \right) \mathbf{T}_{\rm op} \mathbf{p}_j = \underbrace{\mathbf{q}_j - \mathbf{T}_{\rm op} \mathbf{p}_j}_{\mathbf{e}_{j,{\rm op}}} - \left( \mathbf{T}_{\rm op} \mathbf{p}_j \right)^\odot \mbs{\xi} = \mathbf{e}_{j,{\rm op}} - \left( \mathbf{T}_{\rm op} \mathbf{p}_j \right)^\odot \mbs{\xi},
\end{equation}
where we have made use of $\mbs{\xi}^\wedge \mathbf{p} \equiv \mathbf{p}^\odot \mbs{\xi}$ \citep{barfoot_tro14, barfoot_ser17} with
\begin{equation}
\mathbf{p}^\odot = \begin{bmatrix} \mathbf{1} & -\mathbf{v}^\wedge \\ \mathbf{0}^T & \mathbf{0}^T \end{bmatrix}.
\end{equation}
Our approximate error expression is now linear and inserting it into the cost function we have
\begin{equation}\label{eq:cost1}
J(\mbs{\xi}) \approx \frac{1}{2} \sum_j \left( \mathbf{e}_{j,{\rm op}} - \left( \mathbf{T}_{\rm op} \mathbf{p}_j \right)^\odot \mbs{\xi} \right)^T \mathbf{W}_j \left( \mathbf{e}_{j,{\rm op}} - \left( \mathbf{T}_{\rm op} \mathbf{p}_j \right)^\odot \mbs{\xi} \right),
\end{equation}
which is exactly quadratic in $\mbs{\xi}$, and can be minimized by solving the following linear system of equations for $\mbs{\xi}$:
\begin{equation}
\left(\sum_j \left( \mathbf{T}_{\rm op} \mathbf{p}_j \right)^{\odot^T} \mathbf{W}_j \left( \mathbf{T}_{\rm op} \mathbf{p}_j \right)^\odot \right) \, \mbs{\xi} = \sum_j \left( \mathbf{T}_{\rm op} \mathbf{p}_j \right)^{\odot^T} \mathbf{W}_j \mathbf{e}_{j,{\rm op}}.
\end{equation}
After solving for the optimal $\mbs{\xi}$, we can update our initial guess using our pose mapping of choice,
\begin{equation}
\mathbf{T}_{\rm op} \leftarrow \delta\mathbf{T}(\mbs{\xi}) \, \mathbf{T}_{\rm op},
\end{equation}
and iterate to convergence. This approach is discussed in detail by \citet{barfoot_ser17} for the $\phi(\varphi) = \varphi$ case, while here we show it generalizes to any vectorial mapping of pose.
\subsubsection{Alternate Cayley Perturbation Solution}\label{sec:cayper}
The previous section relied on using the infinitesimal expression for pose at each iteration. Here we take a slightly different strategy based on the Cayley transformation that is similar in spirit to that of \citep{mortari07,majji11,junkins11,wong16,wong18,qian20}. We will refer to the resulting algorithm as {\em CayPer}. If we choose the $\phi(\varphi) = 2 \tan \frac{\varphi}{2}$ generating function we can then let our perturbation to the operating point be
\begin{equation}
\delta \mathbf{T}(\mbs{\xi}) = \left( \mathbf{1} - \frac{1}{2} \mbs{\xi}^\wedge \right)^{-1} \left( \mathbf{1} + \frac{1}{2} \mbs{\xi}^\wedge \right).
\end{equation}
Our point error is then
\begin{equation}
\mathbf{e}_j = \mathbf{q}_j - \left( \mathbf{1} - \frac{1}{2} \mbs{\xi}^\wedge \right)^{-1} \left( \mathbf{1} + \frac{1}{2} \mbs{\xi}^\wedge \right) \mathbf{T}_{\rm op} \, \mathbf{p}_j.
\end{equation}
We can then define a new error, $\mathbf{e}_j^\prime$, by premultiplying by $ \left( \mathbf{1} - \frac{1}{2} \mbs{\xi}^\wedge \right)$:
\begin{equation}
\mathbf{e}_j^\prime = \left( \mathbf{1} - \frac{1}{2} \mbs{\xi}^\wedge \right) \mathbf{e}_j = \left( \mathbf{1} - \frac{1}{2} \mbs{\xi}^\wedge \right) \mathbf{q}_j - \left( \mathbf{1} + \frac{1}{2} \mbs{\xi}^\wedge \right) \mathbf{T}_{\rm op} \mathbf{p}_j = \mathbf{e}_{j,{\rm op}} - \frac{1}{2}\left( \mathbf{q}_j + \mathbf{T}_{\rm op} \mathbf{p}_j \right)^\odot \, \mbs{\xi}.
\end{equation}
Importantly, we have that
\begin{equation}
\lim_{\mbs{\xi} \rightarrow \mathbf{0}} \mathbf{e}^\prime_j = \mathbf{e}_j.
\end{equation}
This suggests that we can define a new cost function,
\begin{multline}
J^\prime(\mbs{\xi}) = \frac{1}{2} \sum_j \mathbf{e}_j^{\prime^T} \mathbf{W}_j \mathbf{e}_j^\prime \\ \approx \frac{1}{2} \sum_j \left( \mathbf{e}_{j,{\rm op}} - \frac{1}{2}\left( \mathbf{q}_j + \mathbf{T}_{\rm op} \mathbf{p}_j \right)^\odot \, \mbs{\xi} \right)^T \mathbf{W}_j \left( \mathbf{e}_{j,{\rm op}} - \frac{1}{2}\left( \mathbf{q}_j + \mathbf{T}_{\rm op} \mathbf{p}_j \right)^\odot \, \mbs{\xi} \right),
\end{multline}
which we minimize at each iteration instead of~\eqref{eq:cost1}. The optimal perturbation, $\frac{1}{2}\mbs{\xi}$, is this time the solution to the linear system,
\begin{equation}
\left(\sum_j \left( \mathbf{q}_j + \mathbf{T}_{\rm op} \mathbf{p}_j \right)^{\odot^T} \mathbf{W}_j \left( \mathbf{q}_j + \mathbf{T}_{\rm op} \mathbf{p}_j \right)^\odot \right) \, \frac{1}{2} \mbs{\xi} = \sum_j \left( \mathbf{q}_j + \mathbf{T}_{\rm op} \mathbf{p}_j \right)^{\odot^T} \mathbf{W}_j \mathbf{e}_{j,{\rm op}}.
\end{equation}
The update to the operating point is then
\begin{equation}
\mathbf{T}_{\rm op} \leftarrow \left( \mathbf{1} - \frac{1}{2} \mbs{\xi}^\wedge \right)^{-1} \left( \mathbf{1} + \frac{1}{2} \mbs{\xi}^\wedge \right) \, \mathbf{T}_{\rm op},
\end{equation}
which requires only linear algebraic operations (e.g., no trigonometric functions). Notably, on the first iteration with $\mathbf{T}_{\rm op} = \mathbf{1}$, the CayPer approach becomes a one-shot Weighted Optimal Linear Attitude and Translation Estimator (WOLATE) \citep{mortari07,majji11,junkins11,wong16,wong18,qian20} directly on $SE(3)$. However, we can also continue to iterate to convergence of the original problem.
\subsubsection{Experiment}
We conducted a simulated pointcloud alignment experiment motivated by aligning two sets of matched features from a stereo camera. Our approach is very similar in nature to that of \citet{maimone07} and \citet{barfoot_aa11} in that we align matched stereo points in Euclidean space rather than image space.
We generated feature locations in image space, $\mathbf{y}$, by observing a set of static landmarks using a stereo camera model:
\begin{equation}
\mathbf{y} = \begin{bmatrix} u_\ell \\ v_\ell \\ u_r \\ v_r \end{bmatrix} = \underbrace{\begin{bmatrix} f & 0 & c_u & f \frac{b}{2} \\ 0 & f & c_v & 0 \\ f & 0 & c_u & -f\frac{b}{2} \\ 0 & f & c_v & 0 \end{bmatrix}}_{\mathbf{M}} \, \frac{1}{z} \begin{bmatrix} x \\ y \\ z \\ 1 \end{bmatrix} + \mathbf{n},
\end{equation}
where $\mathbf{n} \sim \mathcal{N}(\mathbf{0}, \mathbf{R})$ and $\mathbf{M}$ is a combined parameter matrix for the stereo rig. We used focal length $f = 200$ [pixels], stereo baseline $b = 0.25$ [m], optical center $(c_u, c_v) = (0,0)$ [pixels], and measurement covariance $\mathbf{R} = 0.25^2 \mathbf{1}$ [pixels$^2$]. This model was used to generate two noisy pointclouds where the second frame is transformed by groundtruth pose, $\mathbf{T}_{\rm gt}$, from the first. Once the noisy feature locations were generated in the two different stereo image frames, we transformed these back to two Euclidean pointclouds, $\mathbf{p}_j$ and $\mathbf{q}_j$, using the inverse stereo camera model. The alignment problem is then to recover $\mathbf{T}_{\rm gt}$ using $\mathbf{p}_j$ and $\mathbf{q}_j$. We assume the points have already been correlated correctly and do not treat the outlier rejection problem here.
\begin{figure*}[t]
\centering
\includegraphics[width=0.49\textwidth]{example3_fig2.pdf}
\includegraphics[width=0.49\textwidth]{example3_fig1.pdf}
\caption{Results of pointloud alignment problem for both the exponential map and CayPer algorithm. The left plot shows the convergence history of the cost function $J$ for $100$ trials where the initial pose was randomly selected. We see that the exponential map regularly takes a long time to converge and often becomes stuck in a local minimum when the initial pose is far from the groundtruth. CayPer always converged in $20$ iterations or less and always ended at the global minimum. The right plot shows a histogram of the costs, $J$, at the final iteration for $1000$ trials where the initial pose was randomly selected. Again, we see that in approximately $20$\% of the trials the exponential map failed to converge to the global minimum while CayPer succeeded in all trials}
\label{fig:example3}
\end{figure*}
To generate the weight matrices, $\mathbf{W}_j$, for our cost function, we mapped the isotropic covariance in image space, $\mathbf{R}$, through the inverse stereo camera model to produce an (inverse) covariance of our error in Euclidean space:
\begin{equation}
\mathbf{W}_j^{-1} = \mathbb{E} [ \mathbf{e}_j \mathbf{e}_j^T ].
\end{equation}
The resulting $\mathbf{W}_j^{-1}$ is a long, skinny uncertainty ellipsoid in the depth direction of the camera as established by \citet{matthies87}. This anisotropic measurement covariance makes this problem particularly challenging. As far we know, there is no practical algorithm that can guarantee finding the global minimum of the cost function, $J$, in this anisotropic situation. \citet{yang20} show how to certify finding the global minimum in the case that $\mathbf{W}_j^{-1}$ is isotropic (i.e., a scalar times the identity matrix) even with significant outliers, but this does not apply in the chosen situation.
We selected a pyramid distribution of $12$ landmarks in the field of view from $5$ to $15$ m away from the camera. The camera was translated $1$ m in the depth, $z$, direction to generate the two noisy pointclouds. We then ran many trials where we varied (i) the random draw of the measurement noise, and (ii) the random draw of the initial guess for the alignment algorithms. Our goal was to see if choosing different vector parameterizations of pose would result in different paths to get from the initial guess to the groundtruth pose.
\subsubsection{Results}
\begin{figure*}[p]
\centering
\includegraphics[width=0.75\textwidth]{example3good.pdf}
\rule{0.7\textwidth}{0.4pt}
\includegraphics[width=0.75\textwidth]{example3bad.pdf}
\caption{Two trials of the pointcloud alignment problem. In both cases, CayPer successfully finds the correct minimum while the exponential map succeeds in the top case but not the bottom where it becomes stuck in an undesirable local minimum. Red box is the initial guess, green box is the groundtruth pose, black dots are the groundtruth landmark locations, green dots are the first noisy pointcloud (static), magenta box (hidden by green box) is the CayPer final pose, blue box (when not hidden by green box) is the exponential map final pose, magenta dots are the second noisy pointcloud (transformed) at the final CayPer pose, blue dots are the second noisy pointlcoud (transformed) at the final exponential map pose}
\label{fig:example3b}
\end{figure*}
We found that using different vector parameterizations of pose in the general perturbation approach of Section~\ref{sec:genper} made little difference and so selected the `exponential map' as a representative method. However, the CayPer approach of Section~\ref{sec:cayper} turned out to be a lot better at not getting stuck in poor local minima.
Figure~\ref{fig:example3} shows quantitative convergence results for several trials. We see that the CayPer approach does not get stuck in local minima as easily as the exponential map approach. In fact, in $1000$ trials CayPer always arrived at the global minimum, but we are not prepared to claim this will always be true at this point. The exponential map frequently became trapped in local minima when the initial guess was particularly bad. We also found that even when both algorithms did find the global minimum, CayPer took fewer iterations to get there; both algorithms were iterated until $\mbs{\xi}^T \mbs{\xi} < 10^{-10}$ or hit $100$ iterations.
Figure~\ref{fig:example3b} shows two of the $1000$ trials we ran. The top case shows the final alignment for a trial where both algorithms successfully converged to the global minimum. The bottom shows a trial where CayPer succeeded but the exponential map failed (catastrophically).
\section{Conclusion and Future Work}\label{sec:confu}
We presented a family of different vector mappings of pose that may find application in computer vision, robotics, and even graphics. We showed that given a vector parameterization of rotation, there is actually an infinite number of possible vector pose mappings with different couplings between the rotational and translational components of the pose. We explored commutative mappings as well as Cayley and Cayley-type mappings. We discussed inverse mappings as well as provided a general formula for the compounding of two general pose mappings. Finally, we showed several of these pose mappings in action on three different applications: pose interpolation, pose servoing control, and pointcloud alignment estimation. The main conclusion is that while we might think of the exponential map as the {\em canonical} way to map vectors to poses, there are many other possibilities that might be preferable, depending on the application. In particular, we found that the Cayley transformation was particularly successful in the CayPer pointcloud alignment method we presented.
Moving forward, there are certainly other theoretical results and connections that might be found related to vector pose mappings. Moreover, several of these ideas might extend to analysis of other useful matrix Lie groups that involve $SO(3)$ as a building block. On the practical side, further analysis of the CayPer algorithm is worthwhile. We also believe that using the Cayley transformation to iteratively optimize poses in more involved problems such as pose-graph optimization and simultaneous localization and mapping will be worth pursuing.
|
\section{Introductory remarks}
\label{Sec: Introduction}
The ability to influence and orchestrate the group behaviour of large ensembles of people is a central theme that arises in many of the challenges facing humanity today. For example, the aggregate effects of our behaviours, coupled with accelerating population growth, is one of the principal drivers behind climate change as well as human induced societal pollution. Furthermore, the inability of humans to design, and then to respect, fair and equitable societal contracts (rules and laws that guide the interactions between people, as well as societal infrastructures) directly impacts the quality-of-life of our citizens. This inability permeates all aspects of society affecting not only our ability to embrace new disruptive business models arising from, for example, the Sharing Economy and Servitization Economy \cite{Shorten_book}, but also other more fundamental aspects of societies, ranging from healthcare outcomes and life expectancy, to congestion in cities. These failures are directly related to our inability to orchestrate human behaviours in some desired manner. To place this latter observation in context, the issue of compliance with social contracts (social distancing, respecting quarantines, etc.) in the context of the current Covid19 pandemic, is just one, albeit extremely sobering illustration of the devastating impact of our current inability to design, manage and enforce desired behaviours in a wider societal context.\newline
Our objective is this paper is to develop a control theoretic framework which will underpin and empower the design and realization of new forms of social contracts to guide the interaction of citizens, devices, and urban infrastructures. Our principal goal is to develop methods that can enforce compliance with social contracts in a range of Sharing Economy applications. Sharing Economy applications often require the good behaviour of humans; in particular, that people should return shared assets in a timely manner as promised, and in good condition. Our basic proposal is to deploy digital tokens as a bond or deposit to ensure {\em compliance}: if the agent remains in compliance then the token is returned, otherwise the token is lost. This issue of compliance has often been approached indirectly using for example a {\em game theoretic} framework, or using some privacy averse voting or recommendation platform. The game theoretic approach is based on creating an equilibrium that incentivizes good behaviour; self-regulatory recommendation platforms simply seek, in principle, to {\em name and shame} miscreants. Both strategies are flawed and are subject to attack from resource-rich nefarious actors (for example by making false recommendations). Our approach is fundamentally different. Building on ideas from stochastic approximations, hyperbolic discounting, and notions of digital identity, we use distributed ledger (DLT) technology coupled with a suite of control algorithms to create a personalised economic commitment mechanism which will enforce compliance. We also design {\em privacy preserving} distributed personalized interventions which encourage good behaviour for parties which share and trade assets in cities \cite{accesspaper}. It is worth also noting that this approach is well grounded in behavioural economics by the theory of hyperbolic discounting \cite{HD}, and can be implemented using DLT's that support Smart Contracts and which do not require transaction fees. Furthermore our work in this arena addresses the thorny problem of how to price compliance individually, by using a control theoretic approach.cThis is not a new idea per se as the use of control theory to model pricing signals have been applied in many different fields. As a first example, link pricing concepts are very common in networking \cite{srikant}, and many stochastic signalling strategies can be interpreted as a price \cite{book,arieh,cdc}. Moreover, a large body of literature has been published on the topic of dynamic pricing to increase the quality of service provided to customers in various domains. The term, commonly used to refer to this field, is \emph{Transactive control}, i.e., the use financial transactions, as a feedback loop, to improve quality of service in various domains: some examples of work in this area can be found in \cite{Phan}-\cite{Junjie}. Other specific instances of dynamic pricing are\cite{Li} (incentivizing users to schedule electricity-consuming applications more prudently), \cite{Chekired} (managing EVs charging and discharging in order to reduce the peak loads), \cite{Bejestani} (combining the classical hierarchical control in the power grid with market transactions) and \cite{Hao} (where the authors propose a transactive control system of commercial building heating, ventilation, and air-conditioning for demand response).\newline
While at first glance this may resemble traditional approaches to pricing, the methodology is actually quite different to that discussed elsewhere, both from a philosophical and technical perspective. In all the aforementioned works, in fact, the control mechanism is obtained using economic transactions in the form of \emph{dynamic toll pricing}; what we propose is a form of \emph{dynamic deposit pricing}: we explicitly price the {\em risk of not using an asset correctly}, rather than {\em access to the asset}, and we provide a rigorous methodology to design personalised interventions supporting aggregated regulation constraints. In this context the specific contributions of this paper are to: (i) present a control system for stochastic agents, enabled by the use of DLTs, to enforce a desired level of compliance to social contracts in a fair manner; and (ii) present a theoretical analysis that establishes the stochastic convergence of the proposed control system.\newline
\subsection{Paper structure} The remainder of this paper is organised as follows: In Section II we discuss why the use of Distributed Ledger Technologies is desirable when designing a compliance scheme. Sections III and IV describe the proposed control strategy and provide theoretical guarantees on its convergence. In Section V we show the effectiveness of the proposed approach through extended Monte Carlo simulations and, finally, Section VI summarizes the presented results and outlines future lines of research.\newline
\section{Preamble: Distributed Ledger Technologies, Social Contracts, and Compliance}
\textcolor{black}{As discussed in the introduction, the basic idea advocated in this paper is to exploit the opportunities afforded by distributed ledgers (often referred to as DLTs; one example is Blockchain) when designing and enforcing social contracts. In particular, we wish to exploit and combine the notions of digital identity and smart contracts coming from DLTs with the rigorous design methods afforded by control theory, in order to design personalised interventions which can orchestrate the behaviour of an entire ensemble of human-like actors.}\newline
\textcolor{black}{In our context a {\bf Social Contract} is a set of rules, or policies, designed to govern the interaction between humans, other humans, and societal infrastructures. For example, in the context of shared assets such as a pool of vehicles, a social contract might require that vehicles are returned to a specified location at a contracted time. Another example of a social contract is the requirement that plastic bottles are returned to point-of-sale after use. For example, in Germany and other countries, bottle deposits, or {\em pfands}, are used to encourage consumers to adhere to the social contract. We argue that in many cases such schemes do not lead to the desired resultss. Often, the minimal marginal cost of the {\em pfand} is not large enough to encourage good behaviour; instead, in many cases the plastic bottles end up in waste bins. This effectively creates a bounty for anybody who is willing to sorting through public rubbish bins in order to find these discarded bottles. The sight of people searching through rubbish bins is common in cities in the western world. So rather than incentivising responsible behaviour, the enforcement of this social contract has incentivised unfortunate people (who often belong to the most vulnerable in society) to sort rubbish on others' behalf. Often they must collect many bottles just to make minimum wage and the sorting ecosystem is in fact akin to a modern form of exploitation. In the case of bottle deposits, the enforcement mechanism is fragile in many ways: through the creation of a de-facto currency that can be redeemed by anyone; by not placing a high enough price on the
deposit to incentivise compliance; and by not applying differential penalties to miscreants depending on levels of personalised compliance. A further issue in such systems is that they can be attacked by resource-rich nefarious actors. Anyone with enough resources and bad intent could attack the entire system by simply dumping many plastic bottles in the ocean, and thereby result in an increased {\em pfand} for everyone. }\newline
\textcolor{black}{The major impediment in solving the problems alluded to above are related to the inability to assign deposits to individuals. If we could do this in a privacy preserving manner we would be able to create systems where only those purchasing goods could redeem deposits. We could also create personalised interventions based on differential penalties, and be able to price risk of misbehaviour accordingly. However enforcing social contracts and nudging people to comply with rules for the greater good is a delicate and complex task. In particular the use of privacy-invasive mechanisms to micromanage behaviour is not acceptable and might lead to dangerous scenarios where a centralised authority violates the rights of individuals. How is it then possible to mitigate the risks while still maintaining the advantages for the community? One technology that offers great potential in this context is {\bf Distributed Ledger Technology}. We believe that DLTs represent a significant step towards a democratization process for these compliance mechanisms. In particular, DLTs possess a number of properties that make them desirable in a context where the presence of a central authority might be detrimental:\newline}
A DLT is nothing more than a shared database. By its very nature it is decentralised and therefore no central authority is required in order to achieve consensus amongst users. In DLTs, due to the cryptographic nature of the private address\protect\footnote{https://laurencetennant.com/papers/anonymity-iota.pdf}, transactions are pseudo-anonymous. This ensures that users are able to protect their privacy behind a large number of transactions \cite{cryptoprivacy}. Of course, it is still feasible to trace back the original transaction and the true owner but, through randomization of the address, it is possible to make it expensive for malicious entities to trace the transactions. This makes DLTs robust from a privacy perspective, while at the same time allowing for the issuing of digital identities. Transactions in DLTs can be encrypted. This allows every agent to manage access to the data present in their own transactions. In our setting the only requirement is that the ownership of the tokens needs to remain accessible to the compliance control algorithms, whereas other information (e.g., user quality of service, statistics on the usage of the system) can be encrypted. This allows each user to maintain ownership of their data and to use them as they please (e.g., to monetize them at a later stage). Moreover, due to their intrinsic consensus mechanism, DLTs are also immutable (meaning that it is not possible for single users to alter the content of previous transactions) and resistant to doublespending (a potential flaw in a digital cash scheme in which the same single digital token can be spent more than once). {\color{black}All these properties make DLTs the ideal infrastructure for a distributed compliance system.} Accordingly, the basic idea is to use digital tokens pegged at a stable fixed value, in the form of a cryptocurrency, to nudge users to comply with a social contract $E$. Token balances, and records of compliance on the ledger, associated with digital identities are the basis of the compliance system. To be more specific, let $E$ represents a general statement, such as \emph{people should wear a face mask} or \emph{each car needs to obey traffic signals}, and its generality allows us to model a very large class of problems. The digital token is thus used as a bond, or digital deposit, to ensure
that various agents comply with rule $E$: if the agent behaves according in compliance with the contract then tokens are returned in their entirety to the owner, whereas if the agent misbehaves the some tokens are lost. Repeat offenders may in fact pay more and lose more tokens. Thus, the risk of losing a token is the mechanism that encourages agents to comply with these social contracts. Note that this system is based on the loss of tokens and is not based on the willingness of agents to report misbehaviour to a central authority (such as the police). Rather than endangering civic rights, such systems should be regarded as protecting the rights of citizens by allowing a more nuanced and fairer form of punishment for indiscretions; {\em fair} in the sense that miscreants might pay more depending on the level of their compliance over time. Finally, in order to address the compliance issue, we will focus on the use of DLTs built around Directed Acyclic Graphs (DAGs) such as IOTA\cite{Popov}\cite{Tangle}. This is due to the fact that such ledgers faciliate large transaction speeds, and are fee-less. In contrast, many standard payment systems (e.g., VISA, Mastercard) and classical Blockchain architectures (e.g., Bitcoin, Ethereum) require users to pay a fee for each transaction. This makes such systems inadequate to serve as the backbone for the proposed compliance scheme. It is essential that the deposited token is returned in its entirety to the owner in the event of full compliance with rule $E$: the proposed social compliance mechanism woudl break down if the agent were required to pay a fee every time a token is deposited or returned, as this would effectively erode the value of her tokens over time.\newline
The proposed architecture is shown in Figure \ref{fig: architecture}. The scheme is divided into three main components: the Distributed Ledger discussed in this section, which acts as the communication backbone for the whole infrastructure, the Physical Layer, in which agents interact with their environment in the setting of the social contract $E$, and finally the Controller, whose task is to regulate the price of the token bonds in order to achieve the desired level of compliance. The latter two components of the architecture will be the focus of the next two Sections. A complete discussion of DAG-based DLTs and their comparison to classical Blockchain is beyond the scope of this paper; the interested reader can refer to \cite{accesspaper} \cite{Popov} \cite{IoTPaper1} \cite{IoTPaper2} \cite{TaCPaper} for a thorough discussion of their properties. For the purpose of this paper, all we need to assume is that there is a fast and secure way to execute the deposit and retrieval of these token bonds. \newline
Before proceeding to the analysis and the modelling of the proposed framework it is worth stressing that the issue of compliance is often not incorporated into algorithms which are designed to regulate, control, and optimise city infrastructures. Many studies addressing human behaviour in this context assume full compliance with policies that have been engineered to optimally organise city infrastructures. As an example, consider traffic flow optimization: a crucial element that is often left out is that humans break rules, and the effect of this rule-breaking profoundly affects how cities operates and how well the engineered algorithms actually perform.
\section{Policy choices for compliance with social contracts}
\label{sec: Policy}
As explained before we are interested in using DLT's to create a type of {\em digital bond} which will encourage compliance with social contracts. Figure \ref{fig: Token} provides a visual representation of this basic idea. In order to engage with this social scheme each agent stakes an amount of tokens (to which some monetary cost is associated) that acts as a bond. The whole process can be made automatic by using smart contracts (these are distributed computer programs which execute as soon as certain conditions are met) \cite{ethereum} to carry out the operations of deposit and return of the tokens. All these operations are recorded on a DLT that is shared amongst all agents (anonymously). \newline
A basic question that arises is how to price this bond: namely, {\em how many tokens should be required as a bond in order to assure compliance with a social contract?} Clearly, if this number is too low, one can expect low levels of compliance (as in the case of the aforementioned plastic bottle deposits), and if it is too high, activity will cease and the social contract will be meaningless. In what follows, we shall develop a method for personalised pricing of the bond based on a feedback signal. The feedback signal will be designed so that aggregate levels of compliance satisfy some constraint. Before proceeding we present two examples of social contracts and show how they lead to different policy choices.
The first example is akin to the traffic signal situation mentioned in the previous section. In this kind of application, a desirable policy might be the following: if you behave, your token is returned, otherwise you lose your token. The second example concerns an agent who enters and moves within a public building (such as an airport or a train station) where the social contract might represent a rule such as {\em keep your mask on}. Clearly, in this type of contract, if the agent behaves then all tokens are returned. But what should happen if the agent misbehaves? One policy might be to issue and redeem tokens at discrete intervals of time. This would punish participants who misbehave multiple times, and would also incentivise those individuals who remove their mask to wear it again so as to avoid further penalty. An alternative policy would be for a miscreant to lose their tokens, but for the pricing algorithm to operate at discrete intervals without further loss of tokens. In this case the miscreant is incentivised to wear the mask again so as to keep their personalised price as low as possible. Below we summarise some policies that are of interest to us.\newline
\begin{itemize}
\item {\bf Fixed penalty policy: } Before participating in the social scheme each agent deposits a certain amount of tokens, the amount being set by the controller. When the action is completed or when the agent exits the scheme, all tokens are returned in the event that they complied with rule $E$; otherwise no tokens are returned to the agent. In the latter case the pricing algorithm continues to adjust the price based on both the agents' level of compliance and that of the network.\newline
\item{\bf Adaptive penalty policies:} Initially each agent deposits a certain amount of tokens, the amount being set by the controller. The contract is reissued at every time-step.
At each time step, compliant agents retrieve their tokens, and stake new ones to continue the activity. Non-compliant agents lose all their tokens every time they do not comply. At all time steps, the pricing algorithm continues to adjust the price based on both the agents' level of compliance and that of the network.\newline
\item {\bf Event driven policies: } Initially each agent deposits a certain amount of tokens. Whenever the agent fails to comply with rule $E$ the tokens are lost; in order to keep participating in the scheme the agent needs to deposit more tokens. In this version of the scheme the amount of tokens that are required as a bond changes over time (again a smart contract can easily take care of the update process).\newline
\end{itemize}
Clearly, these are just three possible policies that might be adopted by the issuer of a social contract, and many others are possible. Our main contribution in this paper is to develop a modelling and feedback control strategy to describe and enable a wide class of policies that include the three aforementioned ones.
\begin{figure}
\includegraphics[width=1\columnwidth]{Architecture.eps}
\caption{The three components of the proposed compliance architecture are: the Distributed Ledger that acts as the communication backbone of the infrastructure, the compliance policy and the feedback controller.}
\label{fig: architecture}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.7\columnwidth]{Token.eps}
\caption{Compliance Policy: each agent deposits a certain amount of tokens that are returned at the end of the activity. $T_C$ is a transaction recorded in a DAG-based DLT that certifies that the tokens have been deposited. Similarly if the agent complies, another transaction is recorded to return the tokens.}
\label{fig: Token}
\end{figure}
\section{Mathematical Framework}
As previously explained, we are interested in designing a feedback mechanism to avoid scenarios in which the value of the bond is either too low (leading to non-compliance) or too high (meaning that agents would not engage in the scheme for fear of losing their tokens). The issue of finding this value is the subject of this section. We will use typical elements of control theory in a stochastic environment where a large number of agents interact with one another and are subject to rule $E$.\newline
Accordingly we consider $n$ agents and, for each of them, we define independent binary random variables $\{ M_i(k) \}_{i=1}^n$, for discrete values of $k$, such that
\begin{equation}
\mathbb{P}(\text{$i$ complies with rule $E$ at time $k$}) = \mathbb{P}(M_i(k) = 1)
\label{eq: prob}
\end{equation}
Moreover, we assume that the probability of these events is entirely dependant on a constant $q_i$, which represents the proclivity of each agent to comply with rules, and two control variables, $C(k)$, $c_i(k)$. The variable $C(k)+c_i(k)$ represents the value of the token bond staked by agent $i$ at time-step $k$. The combination $q_i + C(k)+c_i(k)$ determines the likelihood that agent $i$ will comply with the rule at time-step $k+1$. Then, (\ref{eq: prob}) can be expressed as
\begin{equation}
\mathbb{P}(M_i(k+1) = 1) = p\left(q_i + C(k)+c_i(k)\right)
\end{equation}
with $p: \mathbb{R} \longrightarrow [0,1]$ being a monotone increasing function (which is used to bind the probability between 0 and 1). $C(k)$ and $c_i(k)$ represent, respectively, a global and an individual feedback signal whose purpose is to regulate the behaviour of each agent so as to achieve the desired level of compliance. Accordingly, we consider the following control laws, $\forall k \in \mathbb{N}$ and $\forall
i \in \{1,\dots, n\}$,
\begin{eqnarray}\label{recur1}
C(k+1) &=& C(k) + \alpha \, \left(Q^* - n^{-1} \sum_{i=1}^n M_i(k)\right)\nonumber \\
c_i(k+1) &=& c_i(k) + \beta \, \left(Q^* - \overline{M_i(k)}\right)
\end{eqnarray}
with $\alpha > 0$ and $\beta > 0$ being two constants, $Q^* \in [0,1]$ being the desired level of compliance and $\overline{M_i(k)}$ representing a windowed time average of the compliance of agent $i$, defined as
\begin{eqnarray}\label{def:M-bar}
\overline{M_i(k)} = (1 - \gamma) \, \sum_{j=1}^k \gamma^{k-j} \, M_i(j).
\end{eqnarray}
In this last expression,
the factor $(1 - \gamma)^{-1}$ plays the role of the length of the window for the average, with $\gamma < 1$.
Notice that the proposed framework is very flexible and it would be possible to employ more sophisticated control laws. In this paper, however, we limit ourselves to the study of a proportional action and the extension to more complex feedback loops will be the subject of a future work. \newline
The reason to use both a global and an individual control signal, as opposed to just an individual or a global one, is that these two feedback signals achieve different complementary goals:\newline
\begin{itemize}
\item \emph{Fairness:} Due to differences in individual behaviour, some agents are going to comply with rules less than others. This means that if only a global shared signal were used to control the behaviour of multiple agents, the signal would be driven up by the behaviour of the least complying users, and this would result in an unfair price for the most \emph{virtuous} agents. On the other hand, the introduction of a personalised cost ensures that individuals are going to be priced according to their own behaviour (e.g., the less you comply the more you are going to pay, and vice versa).\newline
\item \emph{Distributed trading of compliance levels:} While the presence of a global cost is not necessary to guarantee the desired level of average compliance in the presented framework (because if every agent's compliance signal were equal to the target value $Q^*$ then the overall average compliance would be $Q^*$), its absence would make the system vulnerable to the repeated misbehaviour of malicious agents who purposely attempt to drive down the compliance level. The introduction of a global signal ensures that the system is able to achieve the desired level of compliance even in the presence of this kind of disturbance. In effect, the global cost allows compliant agents to compensate for non-compliant ones. This aspect is further explored in Section \ref{sec: Simulations}.\newline
\item {\em Pricing attacks:} In traditional pricing models, even one nefarious agent could, in principle, drive up the cost for all agents simply by misbehaving. In view of the previous comment, a natural concern is that similar effects might be possible in our schemes. Fortunately, in our scheme such attacks are not possible. Even though non-compliant agents may drive up the control signal $C(k)$ and hence drive up the cost of the bond, compliant agents will always receive their full deposit after expiration of a contract, leaving them unaffected by the increased price. On the other hand miscreants would continue to lose tokens while they drive $C(k)$ to a high value. Furthermore, in the event that miscreants prevent the desired levels of compliance being reached, $c_i(k)$ may in fact tend to zero, further rewarding compliant agents. \newline
\end{itemize}
{\em Remark:} Before moving on, we want to point out that, since we are considering a DLT as the communication backbone of the whole architecture, the loss of a token is recorded on each agent's copy of the ledger. This means that everyone is aware at all times of the behaviour $M_i(k)$ of every other actor in the scheme (as the loss of the token implies non compliance at time $k$). Therefore, equations (\ref{recur1}) can be validated by each user individually. Also, notice that equations (\ref{recur1}) are consistent with the scenarios described in Section \ref{sec: Policy}.
\section{Theoretical Analysis}
In this section we provide a theoretical analysis of the convergence properties of the stochastic processes $\{C(k), c_i(k)\}$.
We begin by noticing that $\overline{M_i(k)} $ satisfies the following recursion formula
\begin{eqnarray}
\overline{M_i(k+1)} = \gamma \, \overline{M_i(k)} + (1 - \gamma) \, M_i(k+1).
\end{eqnarray}
We also define the function
\begin{eqnarray*}
p(x) = {\rm mid} \{0, 1, x\} = \begin{cases}
x & \text{if } x \in (0,1)\\
0 & \text{if } x \leq 0\\
1 & \text{if } x \geq 1
\end{cases}.
\end{eqnarray*}
Let ${\cal F}(k)$ be the $\sigma$-algebra generated by the random variables
$\{M_i(j) : 1 \le i \le n, \, 1 \le j \le k\}$. Conditioned on ${\cal F}(k)$, the Bernoulli random variables
$\{M_i(k+1)\}$ ($i=1,\dots,n$) are independent and the
distribution of $M_i(k+1)$ is
\begin{eqnarray}\label{def:M}
\mathbb{P}(M_i(k+1)=1 \,|\, {\cal F}(k)) = p(q_i+ C(k) + c_i(k)).
\end{eqnarray}
Before proceeding we note from this preamble that we chose a specific form for the probability function $p$. However, we emphasize that
this choice of a single defined function is merely to aid exposition and that our results can be extended to allow personalized
probability functions $\{p_i\}$ for each agent, where in each these functions are non-decreasing uniformly Lipschitz function.\newline
\medskip
Our main result is a concentration of probability for the time averaged compliance variables
$\overline{M_i(k)}$ around the target value $Q^*$ in the regime
\begin{eqnarray*}
\alpha << \begin{cases} \hskip0.1in \beta \cr 1 - \gamma \end{cases} \hskip-0.18in\Bigg\} << 1
\end{eqnarray*}
Accordingly we introduce a small parameter $\epsilon$ and a window size parameter $w$,
and we allow $\alpha, \beta, \gamma$ to scale with $\epsilon, w$ as follows:
\begin{eqnarray}\label{def:scaling}
\alpha = \epsilon^{3/2} \, \alpha_0, \quad
\beta = \epsilon \, \beta_0 \, w, \quad
1 - \gamma = \epsilon \, w
\end{eqnarray}
where $\alpha_0, \beta_0$ are fixed constants.
We will also assume that the parameter $\beta_0$ satisfies the following condition:
\begin{eqnarray}\label{beta0}
0 < \beta_0 \le 1
\end{eqnarray}
\medskip
\begin{thm}\label{thm1}
Suppose that the processes $\{C(k), c_i(k), M_i(k)\}$ satisfy the recursion equations (\ref{recur1}) and (\ref{def:M})
and the parameters $\alpha, \beta, \gamma$ satisfy the relations (\ref{def:scaling}) and (\ref{beta0}).
Then there are positive constants $B, \epsilon_0$ such that for all $\epsilon < \epsilon_0$,
for all initial values $\{c_i(0),C(0)\}$ satisfying $0 \le q_i + C(0) + c_i(0) \le 1$, and for all $\delta > 0$ and $i \in \{1,\dots,n\}$,
\begin{eqnarray}\label{thm1:eq1}
\limsup_{k \rightarrow \infty} \mathbb{P}\left(|\overline{M_i(k)} - Q^*| > \delta\right) \le B \, \delta^{-2} \, \epsilon \qquad
\end{eqnarray}
\end{thm}
\medskip
\par\noindent{\em Remark 1}:
The result of Theorem \ref{thm1} says that the long-run average compliance level $\overline{M_i(k)}$ ultimately stays close to the target value $Q^*$, for every agent $i$. The crucial term is the factor $\epsilon$ on the right side of (\ref{thm1:eq1}).
By choosing $\epsilon$ sufficiently small the deviation between average compliance and $Q^*$ can be reduced to any desired level
for all agents.
\medskip
\par\noindent{\em Remark 2}:
Although we chose a specific form for the probability function $p$, our results can be extended to allow personalized
probability functions $\{p_i\}$ for each agent, where in each case
$p_i : \mathbb{R} \rightarrow [0,1]$ is a non-decreasing uniformly Lipschitz function.
\subsection{Proof of Theorem \ref{thm1}}
We will establish a bound on the difference $|\overline{M_i(k)} - Q^*|$ for each agent $i=1,\dots,n$.
Accordingly we fix the index $i$ and introduce the variables
\begin{eqnarray}
X_i(k) &=& q_i + C(k+1) + c_i(k+1), \\
\xi_i(k) &=& M_i(k) - p(X_i(k-1)),\nonumber
\end{eqnarray}
and we define the $\mathbb{R}^{2}$-valued process
\begin{eqnarray}
Y(k) = (Y_1(k), Y_2(k))^T = \left(\overline{M_i(k)}, X_i(k)\right)^T.
\end{eqnarray}
Then the equations (\ref{recur1}) and (\ref{def:M}) imply that $Y(k)$ is measurable with respect to ${\cal F}(k)$ and satisfies the recursion relation
\begin{eqnarray}\label{eqn-Y}
Y(k+1) = Y(k) + \epsilon h(Y(k)) + \epsilon \theta(k+1) + \epsilon^{3/2} G(k+1)
\end{eqnarray}
where $h =(h_1,h_2): \mathbb{R}^{2} \rightarrow \mathbb{R}^{2}$,
$\theta = (\theta_1,\theta_2)$ is a $\mathbb{R}^{2}$-valued
martingale difference, and $G(k+1) = (G_1,G_2) \in \mathbb{R}^{2}$ is bounded and measurable with respect to ${\cal F}(k+1)$.
Specifically we have
\begin{eqnarray}
h_1(y) &=& w \, \left(p(y_{2}) - y_1\right), \\
h_{2}(y) &=& \beta_0 \, w \, \left(Q^* - y_1 \right), \\
G_1 &=& 0, \\
G_2 &=& \alpha_0 \, \left(Q^* - \frac{1}{n} \, \sum_{j=1}^n M_{j}(k+1)\right) + \nonumber\\
& & \epsilon^{1/2} \, \beta_0 \, w^2 \, \left(\overline{M_i(k)} - M_{i}(k+1)\right), \\
\theta_1(k+1) &=& w \, \xi_i(k+1), \\
\theta_{2}(k+1) &=& 0.
\end{eqnarray}
Clearly $G(k)$ is uniformly bounded and the function $h$ is uniformly Lipschitz on $\mathbb{R}^{2}$,
so there are constants $K_1, L$ such that
\begin{eqnarray*}
\| G(k+1) \| &\le & K_1 \quad \text{for all $k \ge 0$} \\
\| h(y) - h(z) \| & \le & L \, \| y - z \| \quad \text{for all $y,z \in \mathbb{R}^{2}$}.
\end{eqnarray*}
We also define the associated deterministic dynamical system $y(k) = (y_1(k),y_2(k))$ on $\mathbb{R}^2$ by
\begin{eqnarray}\label{def:class}
y(k+1) = y(k) + \epsilon \, h(y(k))
\end{eqnarray}
and observe that the system (\ref{def:class}) has the unique fixed point $y^* = (Q^*,Q^*)^T$.
\begin{lemma}\label{lem1}
\noindent{\bf a)} There is $K_2' > 1$ such that for all $\epsilon$ sufficiently small and
all $y(0) \in [0,1]^{2}$, the sequence $y(k)$ defined by
(\ref{def:class}) satisfies
\begin{eqnarray}\label{lem1:eq1}
- K_2' \le y_2(k) \le K_2' \quad \text{for all $k \ge 0$}.
\end{eqnarray}
\noindent{\bf b)}
There are constants $\tau, B_1 < \infty$ such that for all $\epsilon$ sufficiently small and
all $y(0)$ satisfying $y_1(0) \in [0,1]$ and $|y_2(0)| \le 2 K_2'$, the sequence $y(k)$ defined by
(\ref{def:class}) satisfies
\begin{eqnarray}\label{lem1:eq2}
\| y(\tau \epsilon^{-1}) - y^* \| \le \frac{1}{2} \, \| y(0) - y^* \| + B_1 \, \epsilon.
\end{eqnarray}
\end{lemma}
\medskip
Lemma \ref{lem1} will be proved in the Appendix.
First we note that $\xi_i(k+1)=0$ if $X_i(k) \le 0$ or $X_i(k) \ge 1$, so that
$Y=y$ whenever $Y$ is outside the square $[0,1]^2$.
Now suppose that $0 \le Y_2(k-1) \le 1$, $Y_2(k) \le 0$ and define
$y(j)$ by the system (\ref{def:class}) with $y(0) = Y(k)$.
It follows that
\begin{eqnarray*}
Y(k+j)=y(j) + \epsilon^{3/2} \, \sum_{i=1}^j G(k+i),
\end{eqnarray*}
for all $j$ such that $y_2(l) \le 0$, all $0 \le l \le j$.
Equation (\ref{lem1:eq1}) implies that
\begin{eqnarray*}
Y_2(k+j) \ge - K_2' - O(\epsilon^{3/2}) \, j \quad
\end{eqnarray*}
for all $j$ such that $y_2(l) \le 0$, all $0 \le l \le j$, while (\ref{lem1:eq2}) implies that $y_2(j) \ge 0$ for some $j = O(\epsilon^{-1})$. Therefore
\begin{eqnarray*}
Y_2(k+j) \ge - K_2' - O(\epsilon^{1/2}) \quad \text{for all $j$ such that $y_2(j) \le 0$}.
\end{eqnarray*}
This bound applies to any excursion of the process $Y(k)$ into the region $Y_2 \le 0$, and therefore
we deduce that $Y_2(k) \ge - K_2' - O(\epsilon^{1/2})$ for all $k$.
Similar reasoning shows that $Y_2(k) \le K_2' + O(\epsilon^{1/2})$ for all $k$.
Since $K_2' > 1$, for $\epsilon$ sufficiently small it follows that
\begin{eqnarray}\label{bound-Y}
Y_2(0) \in [0,1] \Rightarrow |Y_2(k)| \le 2 K_2' \quad \text{for all $k \ge 0$}.
\end{eqnarray}
Note that $0 \le Y_1(k) \le 1$ for all $k$, and therefore $\|Y(k)\| \le \left(1 + 4 K_2'^2\right)^{1/2}$ for all $k$.
\medskip
We now complete the proof of Theorem \ref{thm1} using ideas and techniques from \cite{Borkar}, Chapter 9.
We will use the norm
\begin{eqnarray}
\| X \|_2 = \left(\mathbb{E}[\|X\|^2]\right)^{1/2}
\end{eqnarray}
The next Lemma addresses the martingale part of (\ref{eqn-Y}); its proof appear in the Appendix.
\vskip0.1in
\begin{lemma}\label{lem2}
For all $k \ge 1$,
\begin{eqnarray}\label{increm1}
\|\sum_{j=1}^{k} \theta(j)\|_2 \le \frac{1}{2} \, k^{1/2} \, w
\end{eqnarray}
\end{lemma}
We now use Lemma \ref{lem2} to bound the difference between the process $Y(k)$ and the
sequence $y(k)$. From (\ref{eqn-Y}) we deduce that
\begin{eqnarray}
Y(k) - Y(0) &=& \epsilon \, \sum_{j=0}^{k-1} h(Y(j)) + \epsilon \, \sum_{j=0}^{k-1} \theta(j+1) +\nonumber \\
& & + \epsilon^{3/2} \, \sum_{j=0}^{k-1} G(j+1) \\
y(k) - y(0) &=& \epsilon \, \sum_{j=0}^{k-1} h(y(j))
\end{eqnarray}
Assuming $Y(0)=y(0)$ we get
\begin{eqnarray}
Y(k) - y(k)& =& \epsilon \, \sum_{j=0}^{k-1} h(Y(j)) - h(y(j)) + \epsilon \, \sum_{j=0}^{k-1} \theta(j+1) + \nonumber \\
& & + \epsilon^{3/2} \, \sum_{j=0}^{k-1} G(j+1)
\end{eqnarray}
and therefore
\begin{eqnarray}
\| Y(k) - y(k) \|_2 &\le & \epsilon \, L \, \sum_{j=1}^{k-1} \| Y(j) - y(j) \|_2 + \epsilon \, \frac{1}{2} \, k^{1/2} \, w \nonumber \\
& & + \epsilon^{3/2} \, K_1 \, k
\end{eqnarray}
Accordingly
\begin{eqnarray*}
\sup_{0 \le k \le T} \| Y(k) - y(k) \|_2 &\le & \epsilon \, L \, \sum_{j=1}^{T-1} \sup_{0 \le l \le j} \| Y(l) - y(l) \|_2 + \nonumber \\
& & + \epsilon \, \frac{1}{2} \, T^{1/2} \, w + \epsilon^{3/2} \, K_1 \, T
\end{eqnarray*}
We now define $T = \epsilon^{-1} \tau$ (where $\tau$ was introduced in Lemma \ref{lem1}(b)), and by applying the discrete Gr\"onwall inequality we deduce that
\begin{eqnarray}
\sup_{0 \le k \le T} \| Y(k) - y(k) \|_2 \le \epsilon^{1/2} \, \left(\frac{\tau^{1/2} \, w}{2} + \tau K_1\right) \, e^{L \tau}
\end{eqnarray}
\medskip
For any $k \ge 1$ we write $k = m T + j$ where $m,j$ are integers and $0 \le j \le T-1$.
We define $y(l)$ as the solution of (\ref{def:class}) with initial condition $y(0) = Y(k)$.
Combining the bound (\ref{bound-Y}) with (\ref{lem1:eq2}) we deduce that
\begin{eqnarray*}
\| Y(k+T) - y^* \|_2 & \le & \| Y(k+T) - y(T) \|_2 + \| y(T) - y^* \|_2 \\
& \le & \epsilon^{1/2} \, \left(\frac{\tau^{1/2} \, w}{2} + \tau K_1\right) \, e^{L \tau} + \\
& & +\frac{1}{2} \, \| y(0) - y^* \|_2 + B_1 \epsilon \\
& < & \epsilon^{1/2} \, K_3 + \frac{1}{2} \, \| Y(k) - y^* \|_2
\end{eqnarray*}
where
\begin{eqnarray}
K_3 = \left(\frac{\tau^{1/2} \, w}{2} + \tau K_1\right) \, e^{L \tau} + B_1
\end{eqnarray}
By iterating this bound and using (\ref{bound-Y}) we deduce that
\begin{eqnarray*}
\| Y(k+T) - y^* \|_2 &\le & 2 \epsilon^{1/2} K_3 + 2^{-(m+1)} \| Y(j) - y^* \|_2 \\
&\le & 2 \epsilon^{1/2} K_3 + 2^{-(m+1)} \left[2^{1/2} + K_4 \right]
\end{eqnarray*}
where
\begin{eqnarray}
K_4 = \left(1 + 4 K_2'^2\right)^{1/2}
\end{eqnarray}
Since $m \rightarrow \infty$ as $k \rightarrow \infty$ it follows that
\begin{eqnarray}\label{limsup-E}
\limsup_{k \rightarrow \infty} \| Y(k) - y^* \|_2 \le 2 \, \epsilon^{1/2} \, K_3
\end{eqnarray}
Also using Markov's inequality we get
\begin{eqnarray}\label{limsup-P}
\mathbb{P}\left(|\overline{M_i(k)} - Q^*| > \delta\right) &=& P\left(|Y_1(k) - Q^*| > \delta\right) \nonumber \\
& \le & \mathbb{P}\left(\| Y(k) - y^* \| > \delta\right) \nonumber \\
& \le & \delta^{-2} \, \mathbb{E}\left[\| Y(k) - y^* \|^2\right]
\end{eqnarray}
Therefore (\ref{thm1:eq1}) follows by combining (\ref{limsup-E}) and (\ref{limsup-P}), with
\begin{eqnarray}
B = 4 \, K_3^2\newline
\end{eqnarray}
\section{Simulations}
\label{sec: Simulations}
In this section we provide simulations for the proposed control scheme, to show the effectiveness of our approach and to highlight the need for both a global and individual signal.\newline
Specifically, we consider the following four scenarios:\newline
\begin{itemize}
\item[I] A scenario where only the global signal $C(k)$ is used to regulate the behaviour of each agent (i.e., $c_i(k) = 0, \, \forall k, \, \forall i$);
\item[II] A scenario where both global and individual signals are used to regulate the behaviour of each agent;
\item[III] A scenario where only the individual signal is used to regulate the behaviour of each agent and a subset of individuals $D \subset \{1,\dots,n\}$ refuses to comply with rule $E$ (i.e., $\mathbb{P} (M_i(k) = 1) = 0, \forall k, \forall i \in D$);
\item[IV] A scenario where both the global and the individual signals are used to regulate the behaviour of each agent and a subset of individuals $D \subset \{1,\dots,n\}$ refuses to comply with rule $E$.
\end{itemize}
In all four scenarios we set $n=1000$, $\alpha =0.025$, $\beta =0.1$, $Q^*=0.85$, $\gamma = 0.95$, $C(0) = 0$ and $c_i(0) = 0, \, \forall i \in \{1,\dots, n\}$ and we sample the base compliance levels $q_i$ from the uniform distribution in $[0.1, 0.35]$. Moreover, due to the stochastic nature of the system, for each scenario we perform 150 Monte Carlo simulations and we average over the obtained realizations in order to obtain statistically meaningful results. All simulations are performed on Matlab 2020a.\newline
\subsection{Scenarios I and II}
Figures \ref{fig: global} and \ref{fig: individual} show the results of the simulations in scenario I and II. Even by visual inspection it is clear that the lack of an individual signal to control the agents' behaviour leads to unfair results: while the overall compliance converges to $Q^*$, this is achieved at the expense of the users that would behave better under normal circumstances (i.e., the users with larger $q_i$), that are forced to comply with a higher probability than $Q^*$ in order to compensate for the behaviour of the less compliant agents. Of course, this is undesirable and the use of the personalised cost, as shown in Figure \ref{fig: individual}, tackles this problem by adjusting the individual price depending on the past behaviour of each agent.
\subsection{Scenarios III and IV}
While in Scenarios I and II we explored how the lack of an individual cost leads to unfair results, it is less clear why a global cost is needed at all. In fact, equations (\ref{recur1}) show that, when $C(k)$ is set to zero, the personalised control signals would be sufficient to drive the average behaviour to the desired level of compliance. Nevertheless, without the global cost, the system might fail to achieve the desired target for compliance in scenarios when for some reason a certain number of agents fail to comply repeatedly with rule $E$. This could be due to malfunctions or malicious behaviour. This is highlighted in Figures \ref{fig: indOnly10}, related to scenario III, where 10 \% of the agents, for $k \leq 100$ does not comply with rule $E$ and the system is not able to achieve the desired level of compliance $Q^*$. On the other hand, in scenario IV, shown in Figures \ref{fig: Glob10}, it is possible to see that the presence of the global signal corrects this disturbance, thus making the system more robust to malfunctions and malicious behaviour (of course, the drawback is the the honest agents will have to comply more in order to compensate for the misbehaviour of the non compliant users).
\begin{figure}
\includegraphics[width=1\columnwidth]{Glob.eps}
\caption{Compliance control using only the global signal. Compliance is enforced but at the expense of some agents that have to comply more than others.}
\label{fig: global}
\end{figure}
\begin{figure}
\includegraphics[width=1\columnwidth]{GlobInd.eps}
\caption{Compliance control using both the global and the individual signals. Compliance is achieved and the use of personalised signals make it so that every agent contributes in a fair way.}
\label{fig: individual}
\end{figure}
\begin{figure*}
\includegraphics[width=1\textwidth]{OnlyInd10.eps}
\caption{Compliance control using only the individual signals. For $k \leq 100$ 10\% of the agents does not comply with rule $E$.}
\label{fig: indOnly10}
\end{figure*}
\begin{figure*}
\includegraphics[width=1\textwidth]{Glob10.eps}
\caption{Compliance control using both the global and the individual signals. For $k \leq 100$ 10\% of the agents does not comply with rule $E$.}
\label{fig: Glob10}
\end{figure*}
\section{Conclusions}
\label{sec: Conclusions and Future Research}
In this paper we explored the use of a feedback control system to regulate the behaviour of stochastic agents and to enforce the desired level of compliance, both globally and individually. The use of personalised feedback signals takes into account the behaviour of each agent and leads to fair regulation, with respect to each individual base compliance $q_i$, whereas the global signal increases the robustness of the control system to malfunctions and malicious behaviour. We proved a theorem that establishes that the averaged compliance of each agent, under the proposed regulation scheme, will accumulate around the target compliance $Q^*$ and finally we validated our results through extensive Monte Carlo simulations. As per future lines of research we intend to provide theoretical results for the robustness of the proposed compliance control against malicious actors, explore different formulations of fairness to include, as an example, the economic status of each agent. Finally, we intend to extend our framework by using elements of game theory to take into account more complex scenarios.
\section{Acknowledgements}
Ferraro, Zhao and Shorten are funded in part by the IOTA Foundation, by EPSRC project EP/V018450/1, and by Science Foundation Ireland grant 16/IA/4610 respectively.
|
\section{Introduction}
\qquad SPD’s have many applications, such as QKD \cite{Kiktenko_2018}, 3D imaging systems (LIDAR) \cite{Yu:17}, optical time domain reflectometry (OTDR) \cite{OTD_article}, fluorescence microscopy \cite{Lee_2016}, biomarker tomography \cite{8740927}, astronomical telescopy \cite{10.1117/12.2265785} etc. SPD’s structure is not universal, it’s designed for a specific application and operating parameters. There are several kinds of devices can be used as a single photon sensor, for example, avalanche photodiode (APD) \cite{7747499}, single photon avalanche diode (SPAD) \cite{7283534}, negative feedback avalanche diode (NFAD) \cite{NFAD_article_2}, semiconductor matrices \cite{10.1117/12.2540769}, superconducting nanowires \cite{Li:16}.
In this paper, we have used an SPD, based on InGaAs/InP SPAD, as an investigation object. This SPD is designed for use in compact (standard server rack compatible) QKD device \cite{Kiktenko_2017}, This device works on 1550 nm optical wavelength. APD structure is not optimized for receive optical pulses with average energy less than one photon per pulse, so the choice of SPAD for this problem is optimal. Because of an integrated resistor, using NFAD increases the speed of the electrical circuit, but these diodes have high noise characteristics \cite{NFAD_article} due to many problems while producing. Superconducting nanowire single-photon detector (SNSPD) are large and expensive, which makes it difficult to create a compact QKD device.
SPAD control electronics \cite{Stipcevic, Jong} has a set of important for effective diode’s work parameters, such as DC bias voltage, AC bias voltage (gating signal), the shape of gating signal, and dead time. It’s quite important but a laborious task to find the optimal shape of gating signal to make $QE$ high enough ($10$ -- $20$ \%) and $DCR$ low enough ($<100$ Hz). In this paper, we explore the influence of gating parameters to $QE$ and $DCR$ to make some recommendations about how to tune an SPD to increase the efficiency of a whole QKD device.
There are SPD’s functional scheme and the influence analysis of the gating parameters on $QE$, $DCR$ and afterpulsing probability (AP) in section 2. There are SPD’s measuring stand scheme and methodology for characterizing SPD's parameters in section 3. There are experimental data analysis and general recommendations for improving SPD’s performance in section 4. There is summary in section 5.
\section{Electrical circuit and the influence of gating parameters on the physical processes in SPAD}
\qquad The development of electrical control circuitry of SPAD to provide certain operations with diode is an important task when design SPD. In this paper, the electrical circuit, which provides a periodical crossing of the SPAD to Geiger mode, passive avalanche quenching with active reset and dead time is used \cite{10.1117/12.685808}. We consider the dead time as the time interval when SPD is not sensitive to optical radiation.
The functional diagram of the developed SPD is shown in Figure \ref{fig:el_sch}. SPAD is reverse biased with the direct bias voltage $V_{b1}$ and alternate bias voltage $V_g$ from the cathode side. The alternate bias voltage (gates) is a $312.5$ MHz frequency sine electrical wave. Terminating resistor $R_L$ and quenching resistor $R_{q}$ are used to quench the avalanche and reset the SPAD after a breakdown. There are amplifiers and suppression filters to eliminate the gating signal influence on the avalanche signal processing on the anode side of the SPAD. Comparator converts analog avalanche signal to a certain logic level digital one. This signal arrives at the output logic block and the quenching driver simultaneously. Output logic block makes a signal with certain amplitude and duration that goes to the receiver. Quenching driver makes a dead time pulse. It decreases the direct bias voltage by a fixed amount $V_q$ within a $\tau_{dt}$ time.
\begin{figure}[h]
\centering
\includegraphics[width=0.8\linewidth]{Gen_Scheme_1_Eng.pdf}
\caption{Functional circuit of SPD. In the circuit: Bias Gener. -- DC bias generator; Sine Gener. -- alternating sinusoidal bias voltage (gate) generator with a frequency of $312.5$ MHz; BSF -- the barrage filters to eliminate the influence of gates on the processing of avalanche signals.}
\label{fig:el_sch}
\end{figure}
The most important and determining SPAD’s operating mode output parameters of the presented electrical circuit are the parameters of the AC and DC bias voltages, the shape of the gating signal, and the dead time. The maintaining temperature of the SPAD has also particular importance. These parameters must be selected for each specific used SPAD since its performance can vary significantly from diode to diode even in the same batch \cite{compstudy}. This approach allows us to achieve the best performance of SPD.
\qquad AC bias voltage is a $312.5$ MHz frequency sinusoidal signal with amplitude $V_g$ peak-peak. It is applied to SPAD’s cathode in addition to the DC bias voltage $V_{b1(b2)}$. Gating voltage has to exceed SPAD’s breakdown voltage $V_{br}$ to make single photon detection possible on a value $V_{ex}$ within a $t_g$ time (Geiger mode time). Direct bias voltage $V_{b1}$ falls to $V_{b2}$ after detection to exclude the probability of following avalanche breakdown. The dead time pulse shape is made as shown in Figure \ref{fig:pic_time} to reduce the effect of transient processes on single photon detection. The time when the bias voltage does not exceed the breakdown voltage of the SPAD is called dead time $\tau_{dt}$. The operation mode of SPAD described above is called gated mode with dead time.
\begin{figure}[h]
\centering
\includegraphics[width=1\linewidth]{pic_time_5.jpg}
\caption{Bias voltage on SPAD.}
\label{fig:pic_time}
\end{figure}
It is Geiger mode when $V_a> V_{br}$. Caught in SPAD’s multiplication region free carriers generate a positive feedback avalanche process. Free carriers in SPAD’s heterostructure can appear due to photogeneration effect, direct and indirect tunneling, relaxation of traps, and other less probable effects, like Poole-Frenkel effect and photon-assisted generation through defects \cite{6504460}. Photogeneration and thermal generation prevail in the absorption region because of the low energy gap of the InGaAs material, as is shown in \cite{6504460, 6247522}. The tunneling generation effect is determined with the energy gap and the value of electric field in material. It takes place in absorption and multiplication regions.
Traps for minority charge carriers, electrons, prevail in InGaAs/InP SPAD’s heterostructure. Relaxation of these traps can trigger avalanches. This effect is called the afterpulsing. It imposes some serious restrictions on key distribution speed in QKD systems. We don’t focus on the afterpulses in this work. We have minimized contribution to SPD’s noise by afterpulses due to the fixed value of dead time equal to $\approx 3.5 \ \mu s$, so we have not considered afterpulsing further.
Such SPD parameters like direct bias voltage $V_{b1}$, gate signal shape, and amplitude $V_g$ define maximum SPAD’s excess voltage $V_{ex}$ and diode’s time in Geiger mode. These values can be estimated from equations \ref{eq:1} and \ref{eq:2}.
\begin{equation} \label{eq:1}
V_{ex} = V_{b1} + V_g - V_{br},
\end{equation}
\begin{equation} \label{eq:2}
t_g = \frac{1}{\omega} \left(\pi - 2 \arcsin \left(\frac{V_{br} - V_{b1}}{V_g}\right)\right),
\end{equation}
where $\omega$ -- is circular frequency of sinusoidal gating signal.
There are analytical graphs for the shape of SPAD’s excess voltage $V_{ex}$ under different gate amplitude shown in Figure \ref{fig:gate1}a. There is a laser pulse waveform from the measuring stand (see Fig. \ref{fig:meas_sch}) shown in Figure \ref{fig:gate1}b. There are $t_g$ versus $V_g$ under different $V_{ex}$ diagrams shown in Figure \ref{fig:gate1}c. Also, we have placed $t_g$ versus $V_{b1}$ under different excess voltage values $V_{ex}$ diagrams in Figure \ref{fig:gate1}d.
\begin{figure}[h]
\begin{minipage}[h]{0.47\linewidth}
\center{\includegraphics[width=1\linewidth]{gsbv.jpg}} a) \\
\end{minipage}
\hfill
\begin{minipage}[h]{0.47\linewidth}
\center{\includegraphics[width=1\linewidth]{las_pulse_en_2.jpg}} \\b)
\end{minipage}
\vfill
\begin{minipage}[h]{0.47\linewidth}
\center{\includegraphics[width=1\linewidth]{gtsv.jpg}} c) \\
\end{minipage}
\hfill
\begin{minipage}[h]{0.47\linewidth}
\center{\includegraphics[width=1\linewidth]{gtbv.jpg}} d) \\
\end{minipage}
\caption{a) The shape of excess voltage $V_{ex}$ for various values of gate amplitude: $V_{g} = \{2, 5, 8\}$ V and excess voltage $V_{ex} = 1$ V; b) the waveform of the laser pulse obtained at the measuring stand (see Fig. \ref{fig:meas_sch}), with FWHM equal to $40$ ps; c) the dependence of the geiger mode time $t_g$ on the gate amplitude $V_g$ for various values of the excess voltage $V_{ex} = \{0.8, 1.2, 1.6\}$ V; d) the dependence of the geiger time $t_g$ on the direct bias voltage $V_{b1}$ for various values of the excess voltage $V_{ex} = \{0.8, 1.2, 1.6\}$ V. General parameters: $V_{br} = 70$ V, $T = 3.2$ ns.}
\label{fig:gate1}
\end{figure}
The rise of excess voltage increases the probability of an avalanche excitation with free charge carriers in the multiplication region. Increase of Geiger mode time $t_g$ causes an increase of dark count probability per gate, including afterpulsing, and $QE$. Timing jitter of the laser pulse and the SPAD determine the distribution parameters of a random variable – the moment of avalanche generation. An increase in gating time increases the probability of photogeneration charge carrier to excite the avalanche process in SPAD. So, a variation of gating parameters leads to a change of $QE$ and $DCR$, which ratio we have to optimize to achieve an effective operating mode of SPD.
\section{SPD’s performance parameters measurement}
\qquad We have used the setup shown in Figure \ref{fig:meas_sch} to measure the SPD’s parameters. This setup includes a synchronization system (1), laser source (2), an optical assembly with optical adapters and beam splitters, a frequency counter (3), a system of controlled optical attenuators with output power control (5), multiphoton detector (6) and oscilloscope (7).
\begin{figure}[h]
\centering
\includegraphics[width=0.8\linewidth]{Meas_Sch_1.pdf}
\caption{Functional scheme of SPD’s measurement setup: 1 -- synchronization system, 2 –- laser pulses source 1550 nm wavelength, 3 -- frequency counter Keysight 53230A, 4 –- tested SPD, 5 -– a system of controlled optical attenuators with output power control, 6 –- multiphoton detector NewFocus Model \#1014, 7 -– oscilloscope Lecroy WaveMaster 830Zi-B-R, OA –- optical adapter, BS –- optical beam splitter.}
\label{fig:meas_sch}
\end{figure}
Laser source emits optical pulses at $1550$ nm wavelength. These pulses are generated in the same frequency grid as the gating signal. In the experimental data we present below the repetition rate of laser pulses is $100$ kHz. The delay between gating signal and optical pulses is controlled with the synchronization system. Optical pulses arrive at the input of 90/10 beam splitter, then they are split into two beams with different intensities. A beam with a higher intensity arrives at the input of a wideband optical detector, in which electrical output is connected to an oscilloscope. This detector is used to control the shape, duration, and repetition rate of optical pulses incidented on the SPD (see Fig. \ref{fig:gate1}b)).
A beam with a lower intensity arrives at the two series-connected controlled attenuators system input. Due to the possibility of measuring the power of light radiation after the first attenuator and a fixed attenuation coefficient of the second attenuator, it is possible to adjust and maintain the power of laser pulses at the output of about $0.2$ photons per pulse. Then optical pulses arrive at the 50/50 beam splitter, and after it, two detectors can be connected to the measuring system simultaneously. There is 0.1 photon per pulse arrives at the input of each detector after beam splitter.
The signal, attenuated to a level of $0.1$ photon per pulse, arrives at the tested SPD, where the photon is detected. The number of one-photon states significantly exceeds the number of two-photon states at given energy of light radiation \cite{Gisin}, which ensures the maximum approximation of the results of stand measurements of the SPD performance to the real SPD performance in the QKD setup. The output signal from the detector is incidents simultaneously to a frequency counter and an oscilloscope using an electric power divider. $QE$ and $DCR$ are determined by indications of the frequency counter. The oscilloscope displays a time histogram of triggers, which determines the dead time and the afterpulsing probability.
\section{Results}
\qquad In this work, we determine the effective mode of SPD’s operation as ensuring the minimum $DCR$ value at a given $QE$ value. To identify the effective operating mode of the SPD, the $DCR$ dependence on the gate amplitude has been measured at constant values of the quantum efficiency, such as $10\%$, $15\%$, $20\%$, and at a maintained temperature on the SPAD equal to $228$ K. The measurement results for the three detectors are shown in Figure \ref{fig:dcr1}.
\begin{figure}[h]
\begin{minipage}[h]{0.47\linewidth}
\center{\includegraphics[width=1\linewidth]{dcr_1.jpg}} a) \\
\end{minipage}
\hfill
\begin{minipage}[h]{0.47\linewidth}
\center{\includegraphics[width=1\linewidth]{dcr_2.jpg}} \\b)
\end{minipage}
\begin{center}
\begin{minipage}[h]{0.47\linewidth}
\center{\includegraphics[width=1\linewidth]{dcr_3.jpg}} c) \\
\end{minipage}
\end{center}
\caption{The $DCR$ dependence on the gate amplitude $V_g$ at different values of $QE$ for SPDs with different SPADs: a) SPD №1; b) SPD №2; c) SPD №3.}
\label{fig:dcr1}
\end{figure}
For SPD №1 and SPD №3 with an increase of gating amplitude $DCR$ level decreases (Fig. \ref{fig:dcr1} a), c)) due to the decrease of Geiger time $t_g$ (see Fig. \ref{fig:gate1}). For SPD №2, the $DCR$ level does not decrease as much as for the other tested SPDs (Fig. \ref{fig:dcr1} b)). We can see $DCR$ decrease by more than two times with an increase of gating voltage from $2.1$ V to $4.1$ V at $QE = 10\%$ for SPD №1. $DCR$ decreases more than three times with an increase of gating signal amplitude from $1$ V to $3.7$ V at $QE = 10\%$ for SPD №3.
There is a sharp increase in the $DCR$ value for SPD №1 and SPD №2 starting from a certain value of the gating signal amplitude. This effect is conditioned by the physical processes, occurring in the SPAD of the corresponding SPD. We suggest that these processes may occur because of the reasons presented below.
The size of the depletion region of SPAD constantly oscillates because the bias voltage on the SPAD is sinusoidal and the redistributed charges have some inertia. These processes determine the electrical capacity of the heterostructure. We consider that with high-amplitude voltage oscillations, a free charge carrier remains near the absorption region. This charge subsequently leads to the generation of an avalanche process and contributes to the $DCR$.
We also consider two other hypotheses of the origin of the effect of a sharp increase in $DCR$. It should be noted that these hypotheses have been experimentally falsified. However, the studies we have carried out make it possible to more fully understand the processes occurring in the SPD and to build regularities between the main parameters. It can be used to optimize the SPD’s operation.
\subsection{The increase in the probability of tunneling generation hypothesis}
\qquad The first hypothesis for a sharp increase in $DCR$ is an increase in the probability of tunnel charge generation, which may be due to an extremely sharp increase in the excess bias voltage $V_{ex}$.
The dependence of $DCR$ on excess voltage $V_{ex}$ at different values of the gating signal amplitude is shown in Figure \ref{fig:dcr2}. For SPD №1 we can see that with the same value of excess voltage $V_{ex}$, $DCR$ with gating amplitude values $V_g = 4.6$ V and $V_g = 4.8$ V significantly exceeds $DCR$ with gating amplitude values $V_g = 2.35$ V and $3.9$ V (Fig. \ref{fig:dcr2}a). A similar situation is observed for SPD №2: at the same value of excess voltage $V_{ex}$, $DCR$ with gating amplitude value $V_g = 5.1$ V significantly exceeds $DCR$ with gating amplitude values $V_g = 3$ V and $V_g = 4.1$ V (Fig. \ref{fig:dcr2}b).
\begin{figure}[h]
\begin{minipage}[h]{0.47\linewidth}
\center{\includegraphics[width=1\linewidth]{dcr_11.jpg}} a) \\
\end{minipage}
\hfill
\begin{minipage}[h]{0.47\linewidth}
\center{\includegraphics[width=1\linewidth]{dcr_22.jpg}} \\b)
\end{minipage}
\begin{center}
\begin{minipage}[h]{0.47\linewidth}
\center{\includegraphics[width=1\linewidth]{dcr_33.jpg}} c) \\
\end{minipage}
\end{center}
\caption{The $DCR$ dependence on excess voltage $V_{ex}$ at different values of the gating signal amplitude $V_g$ for SPDs with different SPADs: a) SPD №1; b) SPB №2; c) SPD №3. }
\label{fig:dcr2}
\end{figure}
An increase in the $DCR$ value with increasing $V_{ex}$ and a constant gating signal amplitude is shown in the graphs in Figure \ref{fig:dcr2}. It is mainly due to an increase in Geiger time. A change in the excess voltage $V_{ex}$ within $[0, \ 2]$ V about the breakdown voltage $V_{br} \approx 70$ V does not have such a strong effect on the probability of tunneling generation.
It should be noted that the graph for SPD №1 with gate amplitude values $V_g = 4.6$ V and $V_g = 4.8$ V has an effect associated with a decrease in the $DCR$ value with an increase in excess voltage $V_{ex}$. We find it difficult to put forward an assumption about the cause of this effect. We will consider it in future works.
\subsection{The decrease in the photon detection probability hypothesis}
\qquad The second hypothesis for a sharp increase in $DCR$ is a sharp decrease in the $QE$ value with an increase in the gate amplitude and a constant excess voltage $V_{ex}$.
We show the dependence of $QE$ on the excess voltage $V_{ex}$ at different values of the gating amplitude in Figure \ref{fig:qe1}. With an increase in $V_{ex}$ and a constant gating amplitude, an increase in $QE$ is observed in all graphs due to an increase in the probability of transporting the primary electric charge from the absorption region to the multiplication region. These processes also cause an increase in the probability of an avalanche process.
With an increase in the gating amplitude and a constant value of $V_{ex}$, a decrease in the value of $QE$ due to a decrease in the Geiger mode time $t_g$ is observed. Firstly, this effect is due to a decrease in the probability of a photon arriving in a time window $t_g$ due to the presence of a timing jitter of the photon emission time. Even in the case of a photon arriving at SPAD at the specified time window, the avalanche generation may not occur, since the primary electric charge may appear in the multiplication region at the moment when the bias voltage on SPAD is less than the breakdown voltage. Secondly, the most favorable time interval for detection, characterized by high values of the excess voltage $V_{ex}$, is reduced.
We discovered that this hypothesis is not the reason for the observed effect of a sharp increase in $DCR$. As we can see in the Figure \ref{fig:qe1}a, the decrease in $QE$ with a constant value of $V_{ex}$ and an increase in the gating signal amplitude occurs by approximately equal values for SPD №1. It is necessary to increase $V_{ex}$ also by approximately equal values to maintain $QE$ at the same level. Thus, with gate amplitude values $V_g = \{4.8, 4.6, 2.35, 3.9\}$ V, there are similar trends in the change in the probability of photon detection.
Similar statements can be made for SPD №2. The described effect was also observed on this detector. However, for the gate amplitude value $V_g = 4.1$ V, at certain values of $V_{ex}$, the probability of photon detection is higher than at $V_g = 5.1$ V (Fig. \ref{fig:qe1}b).
We also can see all the described tendencies of $QE$ change for the SPD №3 (Fig. \ref{fig:qe1}c). However, there is one more interesting effect. For $V_g = 1.2$ V, $QE$ grows much faster than for $V_g = 2$ V and $V_g = 3.7$ V.
\begin{figure}[h]
\begin{minipage}[h]{0.47\linewidth}
\center{\includegraphics[width=1\linewidth]{qe_1.jpg}} a) \\
\end{minipage}
\hfill
\begin{minipage}[h]{0.47\linewidth}
\center{\includegraphics[width=1\linewidth]{qe_2.jpg}} \\b)
\end{minipage}
\vfill
\begin{center}
\begin{minipage}[h]{0.47\linewidth}
\center{\includegraphics[width=1\linewidth]{qe_3.jpg}} c) \\
\end{minipage}
\end{center}
\caption{The $QE$ dependence on excess voltage $V_{ex}$ at different values of the gating signal amplitude $V_g$ for SPDs with different SPADs: a) SPD №1; b) SPB №2; c) SPD №3.}
\label{fig:qe1}
\end{figure}
Three SPDs with identical structures and different SPAD diodes were measured. We can conclude that define a universal empirical formula that could relate the main operational parameters of SPDs with the main parameters of the control voltage, such as $QE$ and $DCR$ with $V_{b1}$ and $V_g$, is not possible at this stage of research. This is due to errors that arise during the production of the SPAD heterostructure. It is observed even in one batch of devices and makes it difficult to formalize the dependence for the main SPAD parameters, as shown, for example, in \cite{compstudy}.
Based on the results obtained, it is possible to formulate general recommendations for setting up an SPD with a similar control electrical circuit. To decrease the $DCR$, the gating signal amplitude should be increased, starting from the lowest possible value, with the $QE$ unchanged, until the moment when a strong increase of $DCR$ begins.
\section{Conclusion}
\qquad In this work, we have studied the influence of the gating signal parameters on the main operational parameters of the SPD: $QE$ and $DCR$. The first result of the work is the proposed universal recommendations for increasing the performance of the SPD with a sinusoidal gate signal, passive quenching, and active or passive reset. Recommendations are based on the established relationship of $DCR$ decrease with increasing gating signal amplitude, starting from the lowest possible value, up to a certain boundary value. However, when reaching a certain amplitude of the gate, a sharp increase in the $DCR$ value begins. Using the described dependence has made it possible to reduce the $DCR$ by up to three times. The presented method for reducing the noise characteristics is based on reducing the Geiger mode time $t_g$. The results obtained make it possible to optimize the operational parameters of the sine-wave gated SPD. It improves the operational parameters of the whole QKD device.
The second result is the observation of the effect of a sharp increase in the $DCR$ value with increasing gating signal amplitude and constant $QE$. We have assumed the possible interpretation of this effect. A free charge carrier remains near the absorption zone with high-amplitude oscillations, which subsequently leads to the generation of an avalanche process. We perceive such avalanches as noise. It is necessary to exclude the hypothesis of a sharp increase in $DCR$ because of tunneling charge generation, or a sharp decrease in $QE$ with an increase in the gating signal amplitude and a constant excess voltage $V_{ex}$. It has been shown experimentally. A more detailed study of the nature of this effect based on numerical modeling will be presented in a separate work.
The third result is confirmation of the conclusion of the article \cite{compstudy}. It consists in the difficulty of drawing up a general empirical dependence for individual operational parameters of the SPD, that’s why we need to study each device with its mathematical model. However, in \cite{compstudy} only the afterpulse effect was considered. In this work, we have concluded that similar difficulties arise when trying to empirically relate the main operational parameters of the SPD with the main parameters of the control voltage, namely $QE$ and $DCR$ with $V_{b1}$ and $V_g$.
|
\section{Introduction}\label{sec:introduction}}
\IEEEPARstart{A}{}utonomous driving motivates much of contemporary visual deep learning research.
However, many commercially successful approaches to autonomous driving control rely on a wide array of views and sensors, reconstructing 3D point clouds of the surroundings before inferring object instance trajectories in 3D.
In contrast, human observers have no difficulty perceiving the 3D world in space and time from simple sequences of 2D images rather than 3D point clouds, even though human stereo vision only reaches several meters.
Recent progress in monocular object detection and scene segmentation offers the promise to make low-cost mobility widely available.
In this paper, we focus on monocular 3D detection and tracking so that this low-cost system can be robust and reason in 3D even without 3D sensors.
\begin{figure}[t]
\begin{center}
\includegraphics[width=1.0\linewidth]{fig/Fig_teaser.pdf}
\end{center}
\figcaption{Monocular quasi-dense detection and tracking in 3D.}{
Our dynamic 3D tracking pipeline predicts 3D bounding box association of observed target from quasi-dense object proposals in image sequences captured by a monocular camera with an ego-motion sensor.}
\label{fig:teaser}
\end{figure}
Monocular 3D detection and tracking are inherently intertwined. 3D detection is challenging by itself in the absence of depth measurements or strong priors given a single image. However, current 3D object detectors based on deep learning show promise in capturing geometric priors. This detection basis makes short-term 3D tracking more robust and makes long-term tracking possible. At the same time, 3D tracking information across multiple frames can assist 3D detection in future frames. Note that 3D tracking in the world coordinate can be achieved by projecting object 3D extents in each frame to the world coordinate so that a still object stays still although it moves in 2D observations. This 3D transformation makes both 3D detection and tracking robust as ego-motion of the sensor platform is factored out.
We propose an online 3D tracking framework to detect and track 3D objects in world coordinates from a series of monocular images. Figure~\ref{fig:teaser} provides an overview of our 3D detection and tracking task. After generating object proposals from a monocular image, we estimate their 3D properties, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, 3D bounding box center, depth, dimension, and orientation with a multi-head network. Our model also learns an instance descriptor represented as a vector in a latent space via quasi-dense similarity learning. The quasi-dense mechanism leverages densely populated object proposals, instead of scarce ground truth bounding boxes, to learn instance similarity mappings on a high-dimensional latent space and decide if two object detections are the same object. We utilize the instance feature embedding, estimated 3D information, and object motion from lifted 3D object instance poses to associate instances over time. Notably, we leverage novel motion-aware association and depth-ordering matching algorithms to overcome the occlusion and reappearance problems in tracking. Finally, our model captures the movement of instances in a world coordinate system and updates their 3D poses using velocity-based LSTM (VeloLSTM) to estimate motion along a trajectory, integrating single-frame observations associated with an object instance over time.
Like any deep network, our model is data-hungry.
The more data we feed it, the better it performs.
However, early public datasets are either limited to static scenes~\cite{wen2015ua}, lack the required ground truth trajectories~\cite{milan2016mot16}, or are too small to train contemporary deep models~\cite{geiger2012kitti}.
To bridge this gap, we resort to realistic video games.
We use a new pipeline to collect large-scale 3D trajectories from a realistic synthetic driving environment, augmented with dynamic meta-data associated with each observed scene and object instance.
Thanks to the hype of autonomous driving, more and more large-scale datasets, and benchmarks have surged onto the desk.
However, most of the datasets, collected in days to months and with costly human resources to reach a reasonable scale for a data-hungry network, are licensed for only non-commercial use.
Our 3D vehicle tracking simulation dataset democratizes the autonomous driving field that our simulation dataset is open to both the research community and industrial side. As a result, we believe the presented synthetic driving dataset would foster the community in discovering the power of a deep network as an entry point, especially in combination with sim-to-real approaches, for the large dataset and benchmarks.
In summary, we propose a new pipeline while revising each module building upon the findings of our initial work~\cite{Hu2019Mono3DT}. The key contributions and extended results are highlighted below.
\minisection{Quasi-Dense 3D Object Tracking Pipeline.}
We integrate Quasi-Dense Similarity Learning~\cite{pang2020quasidense} into our tracking framework. It learns to match objects from dense object proposals, instead of only a few annotated foreground regions.
Our quasi-dense 3D tracking pipeline outperforms the camera-based state of the art by near $500\%$ while bridging the gap to LiDAR-based methods on the nuScenes 3D tracking benchmark.
Please refer to~Section~\ref{sec:quasi_dense_similarity_feature} for the method details.
\minisection{3D Confidence, Motion-based Similarity and VeloLSTM.}
In addition to the 2D tracking upgrade, we revised our detection and tracking method from all aspects, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, detector, tracker, and motion model.
We introduce a 3D confidence in~Section~\ref{sec:3d_estimation}, aggregating 2D bounding box confidence and depth confidence, that benefits the bounding box filtering.
In~Section~\ref{sec:3d_tracking}, we improve our centroid-based data association scheme with a motion-based similarity so that the tracker is encouraged to match reappeared candidates in trajectories.
Lastly, in~Section~\ref{sec:lstm_motion}, we enhance our VeloLSTM module to model not only object velocity, but heading angle and dimension.
With all the new techniques, our 3D tracking method has improved the rank on the KITTI 2D tracking benchmarks outperforming existing online approaches.
\minisection{Detailed Experiments on large-scale Datasets.}
Apart from the KITTI and our simulated dataset, we applied our extended model to experiment on urban-driving situations with recently collected large-scale datasets, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, nuScenes and Waymo Open.
On nuScenes, we focus on detailed module design comparison and hyper-parameter tuning, while on Waymo Open, we discuss the advantages and limitations of a monocular-based method.
\section{Related Works}
\minisection{2D Object Detection.}
Object detection reaped many of the benefits from the success of convolutional representation.
There are two mainstream deep detection frameworks:
1) two-step detectors: R-CNN~\cite{girshick2014rich}, Fast R-CNN~\cite{girshick2015fast}, and Faster R-CNN~\cite{ren2015faster}.
2) one-step detectors: SSD~\cite{liu2016ssd}, YOLO9000~\cite{redmon2017yolo9000} and RetinaNet~\cite{lin2017focal}. While these methods rely on so-called anchor boxes to predict offsets for bounding box estimation which are usually derived from dataset depended statistics, more recent works study anchor-free object detection by predicting keypoints such as centers or corners~\cite{zhou2019objects, law2018cornernet, tian2019fcos}. We utilize Faster R-CNN, one of the most popular object detectors, as our detection model basis.
\minisection{Image-based 3D Object Detection.}
A large body of research \cite{chen20153d, mono3d, mousavian2017deep3dbox, m3drpn, monodis, pseudolidar, am3d, 3drcnn, deepmanta} has focused on 3D object detection rather than detecting objects on the image plane since understanding the environment in 3D is of fundamental importance for many vision applications such as robotic navigation and autonomous driving.
Over the last years, these methods predominantly regressed the 3D bounding box parameters directly from the image domain using 2D CNNs \cite{chen20153d, mono3d, mousavian2017deep3dbox, m3drpn, monodis}. More recent works \cite{pseudolidar, am3d} propose to transform the data representation to detect objects by utilizing recent advances in point cloud based 3D object detection \cite{zhou2018voxelnet, shi2018pointrcnn, lang2019pointpillars, chen2017multi} where methods showed impressive performance on point clouds retrieved from LiDAR sensors. Other works leverage shape prior cues to enhance the performance of image-based 3D object detection \cite{3drcnn, deepmanta}.
In contrast, our work extends beyond the single-frame domain, leveraging temporal information to enhance monocular 3D detection by utilizing instance-level uncertainty estimation in combination with our LSTM-based motion module.
\minisection{2D Object Tracking.}
Object tracking in the image-domain has been explored extensively in the last decade~\cite{yilmaz2006object,salti2012adaptive,smeulders2014visual}.
Early methods~\cite{bolme2010visual,gaidon2016virtual,kristan2015visual} track objects based on correlation filters. Recent ConvNet-based methods typically leverage pre-trained object recognition networks. Some generic object trackers are trained entirely online, starting from the first frame of a given video~\cite{hare2016struck,babenko2009visual,kalal2012tracking}.
While object tracking has been tackled by many different paradigms~\cite{luo2017motreview}, including tracking by object verification~\cite{tao2016siamese}, tracking by correlation~\cite{bertinetto2016fully} and tracking by detection~\cite{feichtenhofer2017detectandtrack}, we build upon on the recent successes of tracking by detection methods~\cite{bergmann2019tracktor, mykheievskyi2020odesa, zhang2019mmMOT}. These are typically distinguished by their data association mechanism, for which many different algorithms have been explored, such as network flow~\cite{zhang2008global}, conditional random fields~\cite{choi2015near}, multi-hypothesis tracking~\cite{kim2015multiple} and quadratic pseudo boolean optimization~\cite{ess2008mobile}. These association mechanisms are fueled by different appearance and location based cues such as consistency of 2D and 3D motion~\cite{choi2015near, bergmann2019tracktor, Scheidegger2018pmbm} as well as visual appearance similarity~\cite{voigtlaender2019mots, bergmann2019tracktor}. Our method combines the strengths of different similarity cues for object tracking.
\minisection{3D Object Tracking.}
The previously discussed methods for object tracking in the image-domain usually only take 2D visual features into consideration, where the search space is restricted near the original position of the object. This works well for a static observer, but fails in a dynamic 3D environment. Therefore, various research works have proposed to further leverage 3D information to narrow down the search space and stabilize the trajectory of target objects \cite{MOTBeyondPixels, luiten2020track, Scheidegger2018pmbm, Li_2018_ECCV, Osep17ICRAciwt}. Scheidegger~\emph{et al}\onedot.~\cite{Scheidegger2018pmbm} estimate 3D object positions from single images and add a 3D Kalman filter on the 3D positions to get more consistent localization results and thus improve association. Another approach by Luiten~\emph{et al}\onedot.~\cite{luiten2020track} proposes to use dynamic 3D reconstruction for tracklet association in 3D space to improve long-term tracking. Osep ~\emph{et al}\onedot.~\cite{Osep17ICRAciwt} and Li~\emph{et al}\onedot.~\cite{Li_2018_ECCV} study the extension of this paradigm to 3D bounding box tracking using 3D information obtained from stereo cameras. Recent work by Weng~\emph{et al}\onedot.~\cite{weng2020ab3dmot} proposes specific evaluation metrics for this task and provides a baseline relying on 3D detections from LiDAR. Our work is in line with the recent developments in this field and aims to improve data association by leveraging 3D information, but goes beyond this by integrating both visual appearance and 3D localization information. Moreover, we utilize only a monocular camera and GPS information to track objects in 3D.
\minisection{Joint Detection and Tracking.}
While accurate object detection is a crucial component for tracking, i.e. to track an object throughout a video we must first detect it in every frame it is present, information from tracking can also provide strong priors for object detection. Therefore, many works have studied ways of integrating techniques from object detection with tracking algorithms. Feichtenhofer~\emph{et al}\onedot.~\cite{feichtenhofer2017detectandtrack} compute correlation maps between a pair of frames which are subsequently used to predict the 2D bounding box deformation of each instance between the two frames. In \cite{bergmann2019tracktor}, the authors utilize a single-frame Faster-RCNN model to predict the 2D bounding box deformation from the second stage refinement module alone. Recently, Lu~\emph{et al}\onedot.~\cite{Lu2020RetinaTrack} extend RetinaNet \cite{lin2017focal} to learn track-level instance embeddings via a triplet loss in a joint detection and tracking model. In contrast to the popular triplet loss training scheme, Pang~\emph{et al}\onedot.~\cite{pang2020quasidense} propose to exploit given object proposals within the Faster-RCNN framework to learn instance embedding similarity using densely-connected contrastive pairs and show impressive improvement to instance embedding quality. Our work leverages these findings within our 3D detection and tracking framework, extending the work of Pang~\emph{et al}\onedot. beyond the image domain.
In the 3D tracking domain, the paradigm of joint detection and tracking also caught researcher's attention recently, with Yin~\emph{et al}\onedot.~\cite{yin2020center} combining 3D LiDAR detection with the center-based association paradigm from CenterTrack \cite{zhou2020centertrack} to perform joint 3D detection and tracking. However, because the depth can be perceived directly in LiDAR data, the task is much easier. Our work in contrast is able to perform joint 3D detection and tracking solely from 2D image information, estimating the 3D properties of the visible objects from the image information as well as associating these objects over time.
\minisection{Autonomous Driving Datasets.}
Driving datasets have comprised some of the most popular benchmarks for computer vision algorithms in the last decade. Benchmarks like KITTI~\cite{geiger2012kitti}, UA-DETRAC~\cite{wen2015ua}, Cityscapes~\cite{cordts2016cityscapes} and Oxford RobotCar~\cite{maddern2017robotcar} provide well annotated ground truth for visual odometry, stereo reconstruction, optical flow, scene flow, object detection and tracking as well as semantic segmentation. However, due to the high effort that the annotation of these datasets requires, these benchmarks have been limited in scale. In recent years, the topic of autonomous driving catched on more and more in the industry, providing the resources for new, large-scale driving benchmarks for computer vision, providing annotations for 3D computer vision tasks like 3D object detection and tracking at an unprecedented scale. Therefore, benchmarks like BDD100K~\cite{yu2018bdd100k}, NuScenes~\cite{caesar2019nuscenes}, Argoverse~\cite{Chang2019argoverse} and Waymo Open \cite{waymo} have attracted a lot of attention by the research community.
Still, accurate 3D annotations are challenging to obtain and expensive to measure with 3D sensors like LiDAR. To overcome this difficulty, there has been significant work on virtual driving datasets: virtual KITTI~\cite{gaidon2016virtual}, SYNTHIA~\cite{ros2016synthia}, GTA5~\cite{richter2016playing}, VIPER~\cite{richter2017playing}, CARLA~\cite{dosovitskiy2017carla}, and Free Supervision from Video Games (FSV)~\cite{pk-fsvg-2018}. These datasets have the potential to lower the costs of training accurate deep learning models for applications like autonomous driving tremendously, thus opening up the capabilities of these to a broader audience.
The closest dataset to ours is VIPER~\cite{richter2017playing}, which provides a suite of videos and annotations for various computer vision problems while we focus on object tracking.
We extend FSV~\cite{pk-fsvg-2018} to include object tracking in both 2D and 3D, as well as fine-grained object attributes, control signals from driver actions.
In the next section, we describe how to generate 3D object trajectories from 2D dash-cam videos.
Considering the practical requirement of autonomous driving, we primarily focus on online tracking systems, where only the current and constant number of past frames are accessible to a tracker.
\section{Joint 3D Detection and Tracking}
\label{sec:joint_3d_detection_tracking}
Our goal is to jointly track objects across frames and infer the full 3D information of all tracks from a single monocular video stream and a GPS sensor. The 3D information includes the position, orientation, and dimensions of each object instance.
Figure~\ref{fig:overview} shows an overview of our system.
Images are first passed through a backbone network and Region Proposal Network (RPN) to generate 2D object proposals (Section~\ref{sec:object_detection}).
These 2D proposals are then fed into two lightweight multi-head networks to infer per-instance similarity feature embedding (Section~\ref{sec:quasi_dense_similarity_feature}) and 3D information (Section~\ref{sec:3d_estimation}).
Using both feature embedding and 3D information to generate similarity metrics between all trajectories and detected proposals, we leverage estimated 3D object instance of current trajectories to track them through time (Section~\ref{sec:3d_tracking}).
We introduce motion-aware data association and depth-ordering matching to solve the occlusion problem in tracking.
Finally, we refine the 3D location of objects using the motion model through the newly matched trajectory (Section~\ref{sec:lstm_motion}).
\subsection{Problem Formulation}
\label{sec:problem_formulation}
We phrase the 3D tracking problem as a supervised learning problem.
The goal is to find $N$ trajectories $\mathbb{T} = \{\mathbf{\tau}^1,\ldots,\mathbf{\tau}^N\}$ that match the ground truth trajectories in a video. Each trajectory $\mathbf{\tau}^i_{a,b}$ links to a sequence of detected object states $\langle \mathbf{s}^{(i)}_a,\mathbf{s}^{(i)}_{a+1},\ldots,\mathbf{s}^{(i)}_b \rangle$ starting at the first visible frame $a$ and ending at the last visible frame $b$.
The state of an object $i$ at frame $a$ is given by $\mathbf{s}^i_a = [P, O, D, F, \Delta P] \in \mathbb{R}^{10}$
where $P$ defines the 3D world location $(x, y, z)$ of the object center, $O$ for object orientation $\theta$, $D$ for dimensions ($l, w, h$), $F$ for appearance feature $f_\subfix{app}$, and $\Delta P$ stands for its velocity $(\dot{x}, \dot{y}, \dot{z})$.
The 3D object states in the world coordinates enable the use of our depth-ordering matching and motion-aware association.
On a moving platform with a pose $[\mathbf{R}|\mathbf{t}]$, the mounted camera captures objects found in the 3D world.
We describe an object in the form of a 3D bounding box $\mathbf{X} \in \mathbb{R}^{3\times8}$ in the world coordinates to avoid the changing reference axes introduced by ego-motion.
Each 3D bounding box $\mathbf{X}$ can be viewed as the composition of object location $P$, dimensions $D$, and orientation $O$.
The location $P$ can be projected onto the camera plane using camera pose $[\mathbf{R}|\mathbf{t}]$ as a point $P_\subfix{cam}=(x_\subfix{cam}, y_\subfix{cam}, d)$ where $d$ is the depth from camera.
The $P_\subfix{cam}$ can be further projected to image plane using camera intrinsics $\mathbf{K}$ as the projection of the 3D bounding box center $C=({u_c}, {v_c})$.
The intrinsic parameter $\mathbf{K}$ can be obtained from camera calibration.
Therefore, we can ``lift'' an object center $C$ in the image plane with a depth $d$ using camera parameters $\mathbf{M}=\mathbf{K}[\mathbf{R}|\mathbf{t}]$ to a object location $P$ in the world.
Given the mapping from 3D points in the 3D world and 2D points in the image is known, we enclose a 2D bounding box $B=(u_\subfix{min},v_\subfix{min},u_\subfix{max},v_\subfix{max})$ over the projected 3D bounding box $\mathbf{M} \times \mathbf{X}$ as the 2D annotation for object detectors.
The extrinsic parameter $[\mathbf{R}|\mathbf{t}]$ can be retrieved from a commonly equipped GPS or IMU sensor.
The recorded extrinsic parameter will be used later in the 3D tracking phase to cancel out the ego-motion of the moving platform.
The whole system is powered by an end-to-end multi-head convolutional network trained on a considerable amount of ground truth supervision.
Next, we discuss each component in more detail.
\subsection{Candidate Box Detection}
\label{sec:object_detection}
We employ Faster R-CNN~\cite{ren2015faster} trained on our dataset to provide object candidates in the form of bounding boxes.
As a two-stage detector, Faster R-CNN first generates regions of interest (RoIs) using a multiple-scale Region Proposal Network (RPN) and then utilizes a region convolutional neural network (R-CNN) to classify the RoI as well as to refine its localization.
The multiple-scale RPN generates RoIs from the image feature and obtains its regional feature maps at different levels as the input of our multi-head detector, 3D estimator, and quasi-dense feature extractor networks.
RoI align~\cite{he2017mask} is used instead of RoI pool to obtain the regional representation, which reduces the misalignment of two-step quantization.
Each object proposal (Figure~\ref{fig:overview}(a)) corresponds to a 2D bounding box $B$ as well as an estimated projection of the 3D bounding box center $\hat{C}$.
The proposals are used to locate the candidate objects and extract their appearance features.
However, the center of the objects' 3D bounding box usually does not project directly to the center of its 2D counterparts.
Hence, we also provide an estimation of the 3D bounding box center for better accuracy.
\minisection{Projection of 3D bounding box center.}
To estimate the 3D layout from single images more accurately, we extend the regression process to predict a projected 2D point of the 3D bounding box center from an RoI aligned feature $F$ using Smooth L1 loss~\cite{huber1964loss}.
Estimating a projection of the 3D center is crucial since a small gap in the image coordinate can cause a tremendous shift in 3D.
With the extended multi-head detector, the model simultaneously regresses a bounding box $B$ and an estimated center $\hat{C}$ from an object proposal.
We discuss how an estimated center $\hat{C}$ is used in~Section~\ref{sec:3d_estimation} for lifting predicted 3D bounding boxes.
\begin{figure*}[htp]
\begin{center}
\includegraphics[width=1.0\linewidth]{fig/Fig_quasi_dense.pdf}
\end{center}
\figcaption{The illustration of the quasi-dense similarity learning.}{We leverage quasi-dense object proposals (all the circles), instead of traditional sparse ground truth (solid circles), to train a discriminative feature space by comparing the region proposal pairs between the key frame and the reference frame. The quasi-dense instance similarity loss pulls the feature embedding of different object away from its paired target proposal and draws the embedding of same object pairs together in a high dimensional space.}
\label{fig:quasi_dense}
\end{figure*}
\subsection{Quasi-dense Similarity Learning}
\label{sec:quasi_dense_similarity_feature}
Data association, one challenging problem in the multiple object tracking field, requires distinguishable feature embeddings to match detections and tracklets~\cite{hermans2017defense, valmadre2017siamesefc}.
We leverage quasi-dense proposals to train a discriminative feature space using contrastive loss.
Figure~\ref{fig:quasi_dense} depicts the difference of a quasi-dense approach comparing to conventional sparse ground-truth based approaches.
Unlike \textit{sparse} feature learning, which is commonly used in multiple object tracking that learns only from ground-truth bounding boxes, quasi-dense similarity learning utilizes all object proposals to discriminate positive proposals from negative ones.
Unlike \textit{dense} approaches, which is commonly used in single object tracking that compare tracklets with the whole image, our approach focuses on potential regions of interest to reduce the computational redundancy.
To obtain the proposal-to-proposal quasi-dense instance embedding correspondences, we establish pair-wise instance similarity loss between two neighboring frames, namely, a key frame and a reference frame.
The multiple-scale RPN generates RoIs from the two images and obtains their regional feature maps at different levels.
We assign RoIs with IoU scores to an object over $0.7$ as \textit{positive} proposals, and those lower than $0.3$ as \textit{negative} proposals.
Using the regional feature maps, we propose a quasi-dense embedding head to extract feature embeddings from positive and negative proposals.
Adapting contrastive characteristics to the tracking problem, we defined region proposal pairs from two frames matched to the same object identity as positive pairs, and those matched to different identities are negative ones.
Given a key frame at time $a$, we sample a reference frame within a temporal interval $n$, where $n\in [-3, 3]$ throughout all the experiments.
For each target proposal $\mathbf{s}_{a}$, we optimize its feature embedding $F_{\mathbf{s}_{a}}$ using cross entropy loss combining a non-parametric softmax activation~\cite{wu2018nonparamsoftmax}.
\begin{align}
\mathcal{L}_\subfix{embed} & = -\log
\frac{\text{exp}(F_{\mathbf{s}_{a}} \cdot F_{\mathbf{s}_{a+n}}^{+})}
{\text{exp}(F_{\mathbf{s}_{a}} \cdot F_{\mathbf{s}_{a+n}}^{+}) + \sum_{F_{\mathbf{s}_{a+n}}^{-}}\text{exp}(F_{\mathbf{s}_{a}} \cdot F_{\mathbf{s}_{a+n}}^{-})},
\label{eqn:quasi_dense_loss}
\end{align}
where feature embedding $F_{\mathbf{s}_{a}}$ of the target proposal is learned to associate its positive referenced embedding $F_{\mathbf{s}_{a+n}}^{+}$, and distinguish all its negative referenced ones $F_{\mathbf{s}_{a+n}}^{-}$.
However, the formula above considers only one positive object proposal.
To fully leverage the quasi-dense characteristics, we can balance the number of positive and negative examples by comparing the target proposal to all positive proposals.
Since we accumulate the whole positive targets, Equation~\ref{eqn:quasi_dense_loss} can be reformulated as follows
\begin{align}
\mathcal{L}_\subfix{embed} & = \log [1 + \sum_{F_{\mathbf{s}_{a+n}}^{+}} \sum_{F_{\mathbf{s}_{a+n}}^{-}} \text{exp}(F_{\mathbf{s}_{a}} \cdot F_{\mathbf{s}_{a+n}}^{-} - F_{\mathbf{s}_{a}} \cdot F_{\mathbf{s}_{a+n}}^{+})]
\label{eqn:quasi_dense_loss_update}
\end{align}
The loss term $\mathcal{L}_\text{embed}$ of the above formulation minimizes the cosine distance of the target proposal to all positive referenced examples while maximizing the cosine distance to all negative examples.
By balancing positive and negative examples, we encourage the network to learn an embedding space that discriminates between instances most effectively while being invariant to input perturbations like change in viewpoint or lighting.
In addition, we apply an auxiliary loss encouraging the learning of cosine similarity of target proposal $F_{\mathbf{s}_{a}}$ and references $F_{\mathbf{s}_{a+n}}$
\begin{align}
\mathcal{L}_\subfix{aux} & = (\frac{F_{\mathbf{s}_{a}} \cdot F_{\mathbf{s}_{a+n}}}{||F_{\mathbf{s}_{a}}|| \cdot ||F_{\mathbf{s}_{a+n}}||} - \mathds{1}({\mathbf{s}_{a}, \mathbf{s}_{a+n}}))^2
\label{eqn:quasi_dense_loss_aux}
\end{align}
where $\mathds{1}({\mathbf{s}_{a}, \mathbf{s}_{a+n}})$ is the matching pair indicator function producing $1$ if $\mathbf{s}_{a}$ and $\mathbf{s}_{a+n}$ match to the same object and $0$ otherwise.
We define the full quasi-dense instance similarity loss as
\begin{align}
\mathcal{L}_\subfix{similarity} & = \lambda_\subfix{embed}\,\mathcal{L}_\subfix{embed}+\mathcal{L}_\subfix{aux}
\label{eqn:quasi_dense_loss_all}
\end{align}
In short, the embedding head optimizes the quasi-dense feature embedding pairs to discriminate different instances using instance similarity loss $\mathcal{L}_\subfix{similarity}$. Next, we discuss how we estimate the 3D extent of an object.
\subsection{3D Bounding Box Estimation}
\label{sec:3d_estimation}
We estimate complete 3D bounding box information (Figure~\ref{fig:overview}(b)) from an RoI in the image via a feature representation of the pixels in the 2D bounding box.
The RoI feature vector $F$ is extracted from a deep convolutional backbone network.
Each of the 3D information is estimated by passing the RoI features through a $2$-layer convolutional sub-network with shared parameters and subsequently a $4$-layer 3x3 convolution sub-network, which extends the stacked linear layers design of Mousavian~\emph{et al}\onedot~\cite{mousavian2017deep3dbox}.
We estimate all necessary parameters for inferring 3D bounding boxes from images, whereas Mousavian~\emph{et al}\onedot~\cite{mousavian2017deep3dbox} focus on object orientation and dimensions from 2D boxes.
Besides, our approach integrates 2D detection with 3D estimation, 3D tracking and motion refinement, while Mousavian~\emph{et al}\onedot~\cite{ mousavian2017deep3dbox} crops the input image with pre-computed boxes.
This network is trained using ground truth depth $d$, 3D bounding box center projection $C$, dimensions $D$, and orientation $O$.
We explain how we estimate each component in detail in the following paragraphs.
\minisection{3D World Location.}
Unlike previous approaches~\cite{mono3d, mousavian2017deep3dbox, Hu2019Mono3DT} that lift objects in camera coordinates, we infer 3D location $P$ in the world coordinates from monocular images.
The network learns to regress a logarithmic depth value $\log(\hat{d})$ scaled with a constant $r$ with Smooth L1 loss~\cite{huber1964loss}.
Compared with the characteristics of inverse depth value encoding $1/d$ in~\cite{Hu2019Mono3DT} which compresses the target range ($[5, 100]$) drastically to ($[0.80, 0.99]$), our scaled logarithmic target provides a more realistic scaling inside our target range of the encoded depth value ($[1.39, 4]$).
The scaled logarithmic approach remedies the curse of regression that small encoded depth value change results in a large depth estimation variation.
A projected 3D location $P$ is calculated using an estimated 2D projection of the 3D object center $C$ as well as the depth $d$ and camera transformation $\mathbf{M}$.
\minisection{Initial Projection of 3D Bounding Box Center.}
The 3D object location in the current camera coordinate system can be obtained by estimating the depth of the detected 2D bounding box $B$ as well as the 3D object center projection $C$ and the camera calibration.
The 3D world position $P$ can be inferred by transforming the 3D object location based on the observer's pose.
We find that accurately estimating the 3D object center and its projection is critical for accurate 3D bounding box localization.
However, the projection of the 3D center can be far away from the 2D bounding box center.
Naturally, a center shift from the 2D box center since the 2D bounding box encloses the 3D object from the observer's perspective.
After occlusion and truncation, the estimated 2D bounding box can hardly cover the correct dimension with only the visible area of the objects.
In contrast, the 3D bounding box is defined by the full physical dimensions of an object.
The projected 3D center can lie even outside the detected 2D boxes.
For instance, the 3D bounding box of a truncated object can be out of the camera view. These situations are illustrated in Figure~\ref{fig:3d_center}.
\begin{figure}[htp]
\begin{center}
\includegraphics[width=1.0\linewidth]{fig/Fig_2D_3D.pdf}
\end{center}
\figcaption{3D projected centers are perferred over 2D bounding boxes.}{
A visible 2D center projection point may wrongly locate the object away from the ground plane in the 3D coordinates and would inevitably suffer from more severe scenarios, occluded or truncated.
The center shift causes object misalignment for GT and predicted tracks and harms 3D IoU AP performance.}
\label{fig:3d_center}
\end{figure}
\minisection{Single-frame 3D Confidence.}
Given the high variance in estimating the object-observer relationship in the real world, we find a single-frame depth estimation is prone to predict an erroneous guess if the target object is only partially observable or in unforeseen situations.
While the uncertainty of the model estimation has been well-studied~\cite{kendall2017uncertainty, brazil2020kinematic}, we aim to solve the uncertainty problem in a spatial-temporal approach, focusing on the confidence of the 3D estimation.
During training, the network is supervised by the target confidence score $c_\subfix{depth} = \exp({-|\hat{d} - d|/r})$, which we formulated as an exponential of negative depth distance of prediction and ground truth scaled by a constant $r$.
The learned confidence $\hat{c}_\subfix{depth}$ is used during inference as a 3D objectness score in the detection phase, and as a weighting score in the tracking phase.
The confidence helps to handle the ambiguity of the single-frame depth estimation by balancing single-frame depth estimation and multiple-frame motion model location prediction.
\minisection{Object Orientation.}
Given the image coordinate distance $\hat{u} = \hat{u}_{c} - w_\subfix{img} / 2$ from an object center $\hat{u}_{c}$ to the horizontal center of an image $w_\subfix{img} / 2$ and the focal length $f$, the object rotation $\theta_\subfix{cam}$ in the camera coordinates transformed from an observed vehicle heading $\theta_\subfix{obs}$ with simple geometry, $\theta_\subfix{cam} = [\theta_\subfix{obs} + \arctan (\hat{u} / f)] \mod 2\pi$.
Later we can restore the global rotation $\theta$ in the world coordinates using ego-vehicle heading obtained in $\mathbf{R}$.
Following Mousavian~\emph{et al}\onedot~\cite{mousavian2017deep3dbox} for $\theta_\subfix{obs}$ estimation, we first classify the angle into two bins and then regress the residual relative to the bin center using Smooth L1 loss~\cite{huber1964loss}.
\minisection{Object Dimension.}
In driving scenarios, the high variance of the distribution of the dimensions of different categories of vehicles (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, car, bus) results in difficulty classifying various vehicles using unimodal object proposals.
Therefore, we regress the dimensions $(\hat{l}, \hat{w}, \hat{h})$ to the ground truth dimensions $(l, w, h)$ using Smooth L1 loss~\cite{huber1964loss}.
To sum up, estimating an object's 3D properties provides us with an observation of its location $P$ with orientation $O$, dimensions $D$ and 2D projection of its 3D center $C$. Next, we discuss how we associate the estimation across frames.
\subsection{Data Association and Tracking}
\label{sec:3d_tracking}
Given a set of candidate detections $\mathbb{S}_a = \{\mathbf{s}^1_a,\ldots,\mathbf{s}^M_a\}$ and a set of tracks $\mathbb{T}_a = \{\mathbf{\tau}^J,\ldots,\mathbf{\tau}^K\}$ at frame $a$ where $1 \leq J \leq K \leq N$ from $N$ trajectories, our goal is to associate each track $\mathbf{\tau}$ with a candidate detection $\mathbf{s}$, spawn new tracks, or end a track (Figure~\ref{fig:overview}(c)) in an online fashion.
We solve the data association problem by using a weighted bipartite matching algorithm.
Affinities between tracks and new detections are calculated from three criteria,
(1) the deep representation similarity $\mathbf{A}_\subfix{deep}(\mathbf{\tau}_a, \mathbf{s}_a)$ of the appearance embeddings $f_\subfix{app}$ between an accumulated feature $F_{\mathbf{\tau}_a}$ from existing tracks and a learned feature $F_{\mathbf{s}_a}$ from object detections;
(2) the overlap of bounding boxes $\mathbf{A}_\subfix{iou}(\mathbf{\tau}_a, \mathbf{s}_a)$ between the projection of a 3D bounding box in current trajectories forward in time $B_{\mathbf{\tau}_a}$ and one of the bounding box candidates $B_{\mathbf{s}_a}$;
and (3) the motion similarity $\mathbf{A}_\subfix{motion}(\mathbf{\tau}_a, \mathbf{s}_a)$ of a pseudo object motion vector $V_{\mathbf{s}_a} = \overrightarrow{P_{\mathbf{\tau}_a} P_{\mathbf{s}_a}}$ and the accumulated motion vector $V_{\mathbf{\tau}_a}$ from current trajectories.
Each trajectory is projected forward in time using the estimated velocity of an object and camera ego-motion.
Here, we assume that ego-motion is given by a sensor, like GPS, an accelerometer, gyro, or IMU.
We define an affinity matrix $\mathbf{A}(\mathbb{T}_a, \mathbb{S}_a) \in \mathbb{R}^{||\mathbb{T}_a|| \cdot ||\mathbb{S}_a||}$ between the information of existing tracks $\mathbb{T}_a$ and candidate detections $\mathbb{S}_a$ as a joint probability of appearance, location, and velocity correlation.
\begin{align}
\mathbf{A}_\subfix{deep}(\mathbf{\tau}_a, \mathbf{s}_a) &= \frac{\sigma_x(F_{\mathbf{\tau}_a} F_{\mathbf{s}_a}^T) + \sigma_y(F_{\mathbf{\tau}_a} F_{\mathbf{s}_a}^T)}{2}
\label{eqn:affinity_deep} \\
\mathbf{A}_\subfix{iou}(\mathbf{\tau}_a, \mathbf{s}_a) &= \exp(\frac{-|B_{\mathbf{\tau}_a} - B_{\mathbf{s}_a}|}{c}),
\label{eqn:affinity_iou} \\
\mathbf{A}_\subfix{motion}(\mathbf{\tau}_a, \mathbf{s}_a) &= \exp(\frac{-|V_{\mathbf{\tau}_a} - V_{\mathbf{s}_a}|}{c}))
\label{eqn:affinity_motion}
\end{align}
where $\sigma_x$, $\sigma_y$ are sigmoid functions applied along $x$, $y$ direction.
The affinity matrix, $\mathbf{A}(\mathbb{T}_a, \mathbb{S}_a)$, aggregates the semantic, spatial and temporal similarities.
\begin{equation}
\begin{split}
\mathbf{A}(\mathbf{\tau}_a, \mathbf{s}_a)
&= w_\subfix{deep} \mathbf{A}_\subfix{deep}(\mathbf{\tau}_a, \mathbf{s}_a) \\
&+ (1 - w_\subfix{deep}) \mathbf{A}_\subfix{motion}(\mathbf{\tau}_a, \mathbf{s}_a) \cdot \mathbf{A}_\subfix{iou}(\mathbf{\tau}_a, \mathbf{s}_a)
\label{eqn:affinity_sum}
\end{split}
\end{equation}
The weight $w_{deep}$ balances appearance similarity and a hybrid of motion and 3D location overlap.
We utilize a mixture of discriminative quasi-dense feature embeddings, bounding box overlaps, and location reasoning as similarity measures across frames, similar to the design of POI~\cite{yu2016poi}.
We solve data association greedily instead of using Kuhn-Munkres algorithm~\cite{Kuhn1955hungarian} in Mono3DT~\cite{Hu2019Mono3DT}.
Comparing to 2D tracking, 3D-oriented tracking is more robust to ego-motion, visual occlusion, overlaps, and re-appearances.
When a target is temporarily occluded, the corresponding 3D motion estimator can roll-out for a period of time and relocate the 2D location at each new point in time via the camera coordinate transformation.
\minisection{Data Association Scheme.}
\label{sec:data_association}
Similar to previous methods~\cite{Wojke2017deepsort,xiang2015mdptrack,Sadeghian2017untrackable}, we model the lifespan of a tracker using four major subspaces in MDP state space: $\{\mathtt{birth}, \mathtt{tracked}, \mathtt{lost}, \mathtt{death}\}$.
For each new set of detections, the tracker is updated ($\mathtt{tracked}$) using pairs with the highest affinity score (Equation~\ref{eqn:affinity_sum}).
Each unmatched detection spawns a new tracklet ($\mathtt{birth}$); however, an unmatched tracklet is not immediately terminated ($\mathtt{death}$), as tracklets can naturally disappear ($\mathtt{lost}$) into an occluded region and reappear later.
We address the dynamic object inter-occlusion problem by introducing a motion vector propagation.
A $\mathtt{lost}$ tracklet will not update its feature representation until it reappears, but we still predict its 3D location using the estimated motion.
We continue to predict the 3D location of unmatched tracklets until they disappear from our tracking range (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot $0.15m$ to $100m$) or die out after their lifespan (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot $10$ time-steps).
With the sweeping scheme above, we keep disappeared object tracks inside the tracker memory to be able to recover from object occlusions while still keeping the computational requirements for storing all object tracks at a feasible level.
\begin{figure}[t]
\begin{center}
\includegraphics[width=1.0\linewidth]{fig/Fig_ordering.pdf}
\end{center}
\figcaption{Illustration of depth-ordering matching.}{
Given the tracklets and detections, we sort them into a list by depth order.
For each detection of interest (DoI), we calculate similarities of the object state between DoI and each tracklet.
The location of the object in the world coordinates naturally provides higher probabilities to tracklets near the DoI.
}
\label{fig:ordering}
\end{figure}
\minisection{Depth-Ordering Matching.}
\label{sec:object_3d_matching}
We introduce instance depth ordering for assigning a detection to neighbor tracklets, which models the strong prior of relative depth ordering found in human perception.
Extending the formulation in Hu~\emph{et al}\onedot~\cite{Hu2019Mono3DT}, we consider probabilities of matching each detection of interest (DoI) to potential tracklets using centroid distance of their object state $\mathbf{s}_a$.
Additionally, we introduce object orientation and dimensions to discover different neighbor objects with similar centroid distances.
From the view of each DoI, we compare the pairwise distance of object state $\mathbf{s}_a$ with all the tracklets $\mathbf{\tau}_a$ in the world coordinates.
We directly enforce a smooth distance transition of an exponential function to replace the discrete depth order relationship.
To cancel out the ordering ambiguity of a distant tracklet, we leverage the nature of negative exponential to cut off the matching probability of a distant tracklet smoothly.
So Equation~\ref{eqn:affinity_iou} becomes
\begin{equation}
\begin{split}
\mathbf{A}_\subfix{iou}(\mathbf{\tau}_a, \mathbf{s}_a) &= \exp(\frac{-|\mathbf{\tau}_a - \mathbf{s}_a|}{r}),
\label{eqn:affinity_iou_update}
\end{split}
\end{equation}
where the exponential function with a constant $r$ captures the concept that distant objects contribute less to the affinities.
It naturally provides higher probabilities of linking similar neighbor tracklets than those far away.
In this way, we solve the data association problem of moving objects with the help of 3D trajectories in world coordinates.
Figure~\ref{fig:ordering} depicts the pipeline of depth ordering.
\begin{figure}[t]
\begin{center}
\includegraphics[width=1.0\linewidth]{fig/Fig_motion.pdf}
\end{center}
\figcaption{Demonstration of motion-aware data association.}{
The yellow tracklet is visible all the time, while the red tracklet is occluded by the blue tracklet at frame $T-1$.
Given pseudo motion vectors (dot arrows from a tracklet to all detection candidates) and the temporal-aggregated motion vector (dashed arrows) by the tracklet, our motion-based data association scheme is encouraged match reappeared candidates in trajectories.
}
\label{fig:data_association_scheme}
\end{figure}
\minisection{Motion-aware Data Association.}
\label{sec:motion_aware}
Assuming a linear movement of an object, we expect a $\mathtt{lost}$ object to reappear at a predictable location based on its velocity and the previously seen locations.
Equation~\ref{eqn:affinity_motion} encodes both the motion magnitude $||V||$ and heading angle $\angle V$ into the similarity.
Specifically, $\mathbf{A}_\subfix{motion}$ considers accumulated motion vectors $V_{\mathbf{\tau}_a}$ from tracklet trajectories and pseudo motion vectors $V_{\mathbf{s}_a}$ from single-frame depth estimation.
Since the network can not estimate a velocity directly from a single frame, we leverage the bipartite affinity that aggregates the pseudo motion vectors $\overrightarrow{P_{\mathbf{\tau}_a} P_{\mathbf{s}_a}}$ from all $P_{\mathbf{\tau}_a}$ to $P_{\mathbf{s}_a}$.
However, a motion-only similarity in~Equation~\ref{eqn:affinity_motion} does not distinguish similar motion vectors in different locations.
We propose a new $\mathbf{A}_\subfix{motion}(\mathbf{\tau}_a, \mathbf{s}_a)$ that counts in the motion vectors as well as the centroid location of the objects.
\begin{equation}
\begin{split}
\mathbf{A}_\subfix{motion}(\mathbf{\tau}_a, \mathbf{s}_a) &= \mathbf{w_\subfix{cos}}\,\mathbf{A}_\subfix{centroid} + (1-\mathbf{w_\subfix{cos}})\,\mathbf{A}_{pseudo}
\label{eqn:affinity_motion_update}
\end{split}
\end{equation}
which composed of the following terms
\begin{equation}
\begin{split}
\mathbf{A}_\subfix{centroid} &= \exp(\frac{-|P_{\mathbf{\tau}_a} - P_{\mathbf{s}_a}|}{r}) \\
\mathbf{A}_\subfix{pseudo} &= \exp(\frac{-|V_{\mathbf{\tau}_a} - V_{\mathbf{s}_a}|}{r}) \\
\mathbf{w}_\subfix{cos} &= \angle(V_{\mathbf{\tau}_a}, V_{\mathbf{s}_a})
\label{eqn:affinity_motion_update_decompose}
\end{split}
\end{equation}
where $\mathbf{w_\subfix{cos}}$ is the normalized cosine similarity of two motion vectors $\angle(V_{\mathbf{\tau}_a}, V_{\mathbf{s}_a}) \in [0, 1]$.
The $\mathbf{w_\subfix{cos}}$ favors centroid distance if two motion vectors within a hemisphere of the same direction and weights more on pseudo motion otherwise.
Therefore, we can find the most aligned motion vector that assigns the detected object $\mathbf{s}_a$ to the target tracklet $\mathbf{\tau}_a$.
Figure~\ref{fig:data_association_scheme} illustrates how our motion-aware data association scheme works.
In the next subsection, we show how a motion model helps refine the estimated 3D object state.
\subsection{Motion Model Refinement}
\label{sec:lstm_motion}
For any new tracklet, the detection network is trained to predict the object state $s$ by leveraging RoI features.
For any previously tracked object, the motion model network learns a mixture of multi-view monocular 3D estimates by merging the object state from previous visible frames and the current frame.
Given the learned motion and updated object state, the motion model network predicts the next possible object state to assist the data association procedure.
The update and predict cycles refine the final object state in tandem.
\minisection{VeloLSTM: Deep Motion Estimation and Update.}
To exploit the temporal consistency of objects, we associate the information across frames using two LSTMs for object state prediction and updating (Figure~\ref{fig:overview}(c)).
We use two-layer LSTMs with $128$-dim hidden state size to keep track of a 3D object state in the world coordinates $\mathbf{s}^{(i)}_a = \{P, O, D\} = \{x, y, z, \theta, l, w, h\}$ from the $64$-dim output feature.
Prediction LSTM (P-LSTM) models the dynamic object state $\mathbf{s}_{a}$ in 3D world coordinates based on the previous refined object state $\overline{\mathbf{s}}_{a-1}$ with predicted object velocity $\dot{\mathbf{s}}_{a} = \{\Delta P, \Delta O, \Delta D\}$ from previously updated velocities $\dot{\mathbf{s}}_{a-n:a-1}$.
We encode previous $n=5$ frames of object velocity into a $64$-dim velocity feature to model object motion and acceleration from the trajectory.
P-LSTM predicts the next possible object state $\tilde{\mathbf{s}}_{a} = \overline{\mathbf{s}}_{a-1} + \dot{\mathbf{s}}_{a}$.
Updating LSTM (U-LSTM) balances current observed object state $\hat{\mathbf{s}}_{a}$ from 3D estimation module and previously predicted state $\tilde{\mathbf{s}}_{a-1}$ considering current 3D confidence of depth estimation $\hat{\mathbf{c}}_\subfix{depth}$.
We encode observed, predicted object states into $64$-dim state features and 3D confidence scores into a $64$-dim confidence feature.
By concatenating observation, prediction state features and confidence feature together, we obtained a $192$-dim feature embedding as the input of U-LSTM.
Given the fusion information, U-LSTM refines the object state $\bar{\mathbf{s}}_{a}$ and updates velocities $\dot{\mathbf{s}}_{a-n:a-1}$.
Modeling motion in 3D world coordinates naturally cancels out adverse effects of ego-motion, allowing our model to handle missed and occluded objects.
The LSTMs keep updating the predicted object state $\tilde{\mathbf{s}}_{a-1}$ while assuming a linear velocity model if there is no matched bounding box.
Therefore, we model 3D motion (Figure~\ref{fig:overview}(d)) in world coordinates allowing occluded tracklets to move along motion-plausible paths while managing the birth and death of moving objects.
Both LSTM modules are trained with predicted object states $\tilde{\mathbf{s}}_{a}$, refined object states $\overline{\mathbf{s}}_{a}$ from detector prediction, and ground truth object state $\mathbf{s}_{a}$ trajectories.
The refinement loss $\mathtt{L}_\subfix{refine}(\overline{\mathbf{s}}_{a}, \mathbf{s}_{a})$ and prediction loss $\mathtt{L}_\subfix{predict}(\tilde{\mathbf{s}}_{a}, \mathbf{s}_{a})$ aim to reduce the distance of estimated and ground-truth object state using Smooth L1 loss~\cite{huber1964loss}.
The linear motion loss $\mathtt{L}_\subfix{linear}(y_{a-n:a}) = \sum_{a-n:a}|(y_{a} - y_{a-1}) - (y_{a-1} - y_{a-2})| / \Delta t$ focuses on the smooth transition over time $\Delta t$ of object state refinement $\mathtt{L}_\subfix{linear}(\overline{\mathbf{s}}_{a-n:a})$ and prediction $\mathtt{L}_\subfix{linear}(\tilde{\mathbf{s}}_{a-n:a})$.
Conclusively, our pipeline consists of a single-frame monocular 3D object detection model for object-level pose inference and recurrent neural networks for inter-frame object association and matching.
We extend the region processing to include 3D estimation by employing multi-head modules for each object instance.
We introduce a motion-aware association to solve the inter-object occlusion problem.
For tracklet matching, depth ordering lowers the mismatch rate by filtering out distant candidates from a target.
The LSTM motion estimator updates the velocity and states of each object independent of camera movement or interactions with other objects.
The final pipeline produces accurate and smooth object trajectories in the 3D world coordinate system.
\section{3D Vehicle Tracking Simulation Dataset}
\label{sec:gta_vehicle_dataset}
It is laborious and expensive to annotate a large-scale 3D bounding box image dataset even in the presence of LiDAR data, although it is much easier to label 2D bounding boxes on tens of thousands of videos~\cite{yu2018bdd100k}.
Therefore, no such dataset collected from real sensors is available to the research community.
To resolve the data problem, we turn to driving simulation to obtain accurate 3D bounding box annotations at no cost of human efforts.
Our data collection and annotation pipeline extend the previous works like VIPER~\cite{richter2017playing} and FSV~\cite{pk-fsvg-2018}, especially in terms of linking identities across frames.
Our simulation is based on \textit{Grand Theft Auto V}, a modern game that simulates a functioning city and its surroundings in a photo-realistic three-dimensional world.
To associate object instances across frames, we utilize in-game API to capture global instance id and corresponding 3D annotations directly.
In contrast, VIPER leverages a weighted matching algorithm based on a heuristic distance function, which can lead to inconsistencies.
It should be noted that our pipeline is real-time, providing the potential of large-scale data collection, while VIPER requires expensive off-line processings.
\minisection{Dataset Statistics.}
Compared to the others, our dataset has more diversity regarding instance scales (Figure~\ref{fig:statistics_scale}) and closer instance distribution to real scenes (Figure~\ref{fig:statistics_ins}).
To help understand our dataset and its difference, we show more statistics in the appendix.
\begin{figure}[htpb]
\begin{subfigure}{0.45\linewidth}
\includegraphics[width=\linewidth]{fig/Fig_statistics/Scale.pdf}
\caption{Instance scale}\label{fig:statistics_scale}
\end{subfigure}
\begin{subfigure}{0.45\linewidth}
\includegraphics[width=\linewidth]{fig/Fig_statistics/ins_per_img2.pdf}
\caption{Instances per image}\label{fig:statistics_ins}
\end{subfigure}
\begin{subfigure}{0.45\linewidth}
\includegraphics[width=\linewidth]{fig/Fig_statistics/durations.pdf}
\caption{Frames of tracking}\label{fig:statistics_duration}
\end{subfigure}
\begin{subfigure}{0.45\linewidth}
\includegraphics[width=\linewidth]{fig/Fig_statistics/distance.pdf}
\caption{Vehicle distance}\label{fig:statistics_distance}
\end{subfigure}
\figcaption{Statistical summary of our dataset}{in comparison of KITTI~\cite{geiger2012kitti}, VKITTI~\cite{gaidon2016virtual}, VIPER~\cite{richter2017playing}, and Cityscapes\cite{cordts2016cityscapes}}
\label{fig:statistics}
\end{figure}
\section{Experiments}
\label{sec:experiments}
We evaluate our 3D detection and tracking pipeline on the real-world driving scenes, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, Waymo Open dataset, nuScenes Tracking, KITTI MOT benchmarks, and our large-scale simulation dataset featuring a wide variety of road conditions in a diverse virtual environment.
To follow the submission policy on the above public testing benchmarks, we only evaluate the final performance of our best result on the public testing benchmark.
\subsection{Dataset Details}
\label{sec:dataset_details}
\minisection{GTA 3D Vehicle Tracking dataset} recorded raw data at $12$ FPS, which is helpful for temporal aggregation.
With the goal of autonomous driving in mind, we focus on vehicles closer than $150m$, and also filtered out the bounding boxes whose areas are smaller than $256$ pixels.
The dataset is then split into train, validation and test set with a ratio $10:1:4$.
\minisection{KITTI MOT benchmark~\cite{geiger2012kitti}} provides real-world driving scenarios.
Since the KITTI benchmark is relatively small-scale, we additionally train on the whole KITTI detection training set to enhance our 3D estimation module in the connection of lifting 3D bounding boxes from a single image.
We train a full model with the whole detection and tracking training set for the public benchmark submission.
For bounding box center comparison, we train on the whole detection training set and half of the tracking training set, and then we evaluate the performance difference on the other half of the tracking training set.
\minisection{nuScenes dataset~\cite{caesar2019nuscenes}} is another real-world benchmark containing street scenes in Boston and Singapore captured from a moving vehicle equipped with different sensors, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, $6$ cameras, $1$ LiDAR, and $4$ radars.
It provides 3D annotations for LiDAR data with $10$ object classes for detection task, and $7$ object classes for tracking task.
There are $700$ training sequences, $150$ validation sequences and $150$ test sequences in the nuScenes dataset.
Every sequence collects images at $12$ FPS, denoted as \textit{full frames}, and only those sampled keyframes, annotated at $2$ FPS, are used for evaluation.
Each sequence consists of about $40$ keyframes per camera.
We apply our motion-based pipeline on the frames in a higher frame rate (\textit{full frames}) and produce refined tracking results on a keyframe basis.
\minisection{Waymo Open Dataset~\cite{waymo}} collects images at $10Hz$ from $5$ different directions with partially overlapped regions: front, front left, front right, side left, and side right.
There are $1,150$ segments of $20s$ each, $798$ segments for training, $202$ segments for validation, and $150$ segments for testing.
Every segment contains about $198$ frames per camera.
Waymo dataset provides $3$ types of annotations: 2D annotations, 3D LiDAR annotations, and projected 3D LiDAR annotations.
The annotated object classes are vehicles, pedestrians, cyclists, and signs. However, signs are not used in the 3D detection and 3D tracking benchmark.
\minisection{Cross Camera Aggregation.}
For both nuScenes and Waymo, we align all camera results together in an online fashion because their official evaluations only support the LiDAR-based perspective.
The cross-camera aggregation criteria focus on removing duplicated objects across different cameras whose 3D euclidean distance is within $2$ meters. For an object with smaller width or height, such as a pedestrian, the threshold is set to $1$ meter.
Once duplicates are found, we keep the bounding box with a higher 3D score and discard the lower ones to handle the overlapping problem.
On the nuScenes dataset, we found using only camera data achieving competitive results compared to LiDAR baselines since all $6$ cameras can be stitched together to reconstruct the 360-degree scene.
On the other hand, the five camera setting on Waymo Open Dataset lacks information at the rear camera and covers only $5/8$ azimuth angle of the horizon.
But objects are annotated in LiDAR sweeps.
The limitation above refrains our method from reaching meaningful evaluation scores on Waymo Open 3D detection or tracking benchmark.
Therefore, we conduct a view angle ablation study in~Table~\ref{tab:waymo_validation_evaluation} to validate the power of our data-driven method.
\subsection{Training and Evaluation}
\label{sec:training_and_evaluation}
\minisection{Network Specification.}
Our 2D detection and 3D center estimation are based (Section~\ref{sec:object_detection}) on Faster RCNN~\cite{ren2015faster} with image classification pre-trained weights on ImageNet~\cite{russakovsky2015imagenet} from TorchVision~\cite{adam2019pytorch}.
For the KITTI dataset with a smaller amount of data, we choose a light-weight yet powerful DLA-34-up~\cite{yu2018dla} architecture, while for the large driving datasets, \emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, nuScenes, GTA, Waymo, we resort to a $101$-layer ResNet~\cite{resnet}.
The anchors of the RPN in Faster R-CNN span $5$ scales and $5$ ratios, in order to detect small objects at distance as well as larger, close objects.
We use a maximum of $2000$ candidate objects before NMS.
The 3D estimation network (Section~\ref{sec:3d_estimation}) estimate 3D bounding box parameters from features extracted from RoI align~\cite{he2017mask}.
For the quasi-dense similarity learning (Section~\ref{sec:quasi_dense_similarity_feature}), we use $\lambda_{embed} = 0.25$ throughout the experiments.
The LSTM motion module (Section~\ref{sec:lstm_motion}) is trained for $100$ epochs on $10$ sample frames per object trajectory with $128$ sequences per batch for large-scale datasets and $8$ for KITTI.
\minisection{Training Procedure.}
The general training procedure is conducted for $24$ epochs using SGD~\cite{robbins1951sgd} with a momentum of $0.9$ and a weight decay of $10^{-4}$.
The learning rate linearly increases from $5\times10^{-4}$ to $5\times10^{-3}$ over the initial $1000$ warm-up training steps.
We use a batch size of $8$ images per GPU and train on $4$ GPUs, resulting in an effective mini-batch size of $32$.
We apply image horizontal flipping as data augmentation in the training phase, and there is no augmentation performed for test/validation phase.
We optimize our model on each dataset with different training procedures according to its amount of data and the GPU memory limit.
The image resolution is resized and padded to fit the minimum down sample scale $32$.
We use $1600\times900$ resolution for nuScenes, $1280\times854$ for Waymo, $1485\times448$ for KITTI, and $1920\times1080$ for GTA.
Following CenterTrack~\cite{zhou2020centertrack}, we fine-tune our KITTI model from a GTA 3D detection model.
Given the multiple cameras setup in nuScenes and Waymo, we treat each camera video as a different sequence during both training and inference.
Since nuScenes doesn't provide 2D annotations, we project the annotated 3D bounding boxes to the image plane to train the 2D branch of our model.
On the Waymo Open dataset, we use only the 3D LiDAR annotations and project them to the image coordinates as our 2D annotations since the 2D and 3D bounding boxes are annotated independently.
For more details on training strategies, please refer to the appendix.
\minisection{3D Detection Evaluation.}
We evaluate the performance of 3D detection with refined object state estimation of different tracking methods.
For KITTI and GTA, we use the formal evaluation metrics of detection 3D mAP from KITTI~\cite{geiger2012kitti} in the 3D Object Detection Evaluation.
To inspect the precision-recall curve deeper, we also use detection 3D mAP from COCO~\cite{lin2014mscoco} with a range of IoU thresholds from $50$ to $95$ at a step of $5$.
As stated in~\cite{monodis}, we use $40$ recall positions in 3D object detection evaluation instead of the $11$ recall positions proposed in the original Pascal VOC benchmark~\cite{everinham2010voc}.
For nuScenes, we use the benchmark evaluation metrics, called nuScenes detection score (NDS), which is the weighted sum of mean Average Precision (mAP) and several True Positive (TP) metrics.
For Waymo, we also use the 3D detection benchmark evaluation metrics where their primary metric is mean Average Precision weighted by Heading (mAPH) with different 3D IoU thresholds for each class.
\minisection{Multiple Object Tracking Evaluation.}
For KITTI and GTA, we use the official KITTI 2D tracking benchmark metrics following CLEAR~\cite{bernardin2008evaluating} and Li et al.~\cite{li2009learning}, including Multiple Object Tracking Accuracy (MOTA), Multiple Object Tracking Precision (MOTP), Miss-Match (MM), False Positive (FP), False Negative (FN), Mostly-Tracked (MT), Partly-Tracked (PT), and Mostly-Lost (ML), etc.
MOTA indicates the tracking performance of a tracker in dealing with three common errors, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, false positives, missed targets and identity switches.
\begin{equation}
MOTA = 1 - \frac{IDS + FP + FN}{GT} \label{eqn:MOTA}
\end{equation}
where $GT$ stands for number of ground-truth positives.
For nuScenes, we follow the 3D tracking benchmark evaluation metrics using Average Multi-Object Tracking Accuracy (AMOTA) as the main metric. They average the MOTA metric over different recall thresholds
\begin{align}
\begin{split}
AMOTA &= \frac{1}{n-1} \sum_{r \in \{ \frac{1}{n-1}, \frac{2}{n-1}, ..., 1 \} } {MOTA}_{\subfix{r}}, \label{eqn:AMOTA} \\
{MOTA}_{\subfix{r}} &= \text{max} (0, \\
& 1 - \mathbf{\alpha} \frac{IDS_{\subfix{r}} + FP_{\subfix{r}} + FN_{\subfix{r}} - (1 - r){GT}}{r {GT}})
\end{split}
\end{align}
with recall threshold $\mathbf{r}$ is calculated at each of the interpolation $\mathbf{n}$-points for $n = 40$, tolerance ratio $\mathbf{\alpha}$ aims to retain the non-zero values of MOTA$_{\subfix{r}}$.
We report $\mathbf{\alpha} = 1.0$, denoted as AMOTA@1, for the official benchmark and $\mathbf{\alpha} = 0.2$, denoted as [email protected], for backward comparison with previous arts.
Instead of using 3D IoU threshold, the nuScenes tracking benchmark choose a centroid distance threshold of $2$ meters on the ground plane as the matching criterion.
For Waymo, we report the official 3D tracking evaluation metrics, which follow the metrics of CLEAR~\cite{bernardin2008evaluating} with different 3D IoU thresholds for each class.
However, the limitations above (Section~\ref{sec:dataset_details}) refrain our method from reaching meaningful evaluation scores on benchmarks using 3D IoU based metrics.
\minisection{Discussion of Evaluation Metrics}
The analysis from~\cite{caesar2019nuscenes} has shown that objects, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, pedestrians, bicycles, are hard for camera-based methods to have meaningful IoU scores, while those objects are difficult to be detected in LiDAR.
Similarly, we observe from quantitative results on Waymo that the IoU threshold at $0.7$ prohibits the camera-based method from evaluating meaningful tracking performance.
However, the score does not cohere to our qualitative results.
We argue that high IoU thresholds become impractical with uncertain object state $\mathbf{s}_{a}$ measurements as the metric does not reward a tracker for associating all matches right if it makes a moderate error on the object state.
Therefore, we report extended 3D tracking evaluation~\cite{Chang2019argoverse} for a better understanding of the association measure in the 3D extent. The MOTA$_\subfix{3D}$ is based on the object centroid distance, denoted as MOTP$_\subfix{C}$, from tracked objects to the ground truth within a detection association range, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, $2$ meters. MOTP$_\subfix{O}$ calculates orientation angle difference to the vertical axis, and MOTP$_\subfix{I}$ stands for the amodal shape error, using $1 - \text{IoU}$ with aligned orientation and centroid.
We conduct ablation studies in~Table~\ref{tab:cross_dataset_centroid_based_evaluation} with detailed metrics on translation, scale, orientation as accompanying metrics to validate the power of our data-driven method.
\subsection{Ablation Experiments}
\label{sec:ablation_experiments}
In the following, we discuss the results of our ablation experiments on the design choices of each component (3D box estimation, data association, motion modeling) in our quasi-dense 3D tracking pipeline.
We report the 3D tracking performance comparison on nuScenes validation set for dropping sub-affinity scores in Table~\ref{tab:nusc_ablation_drop_affinity}, using full images in Table~\ref{tab:nusc_ablation_full_frame}, and swapping different modules in Table~\ref{tab:nusc_ablation_swap_module}.
We also investigate the importance of using a 3D object center using the KITTI sub-val set in~Table~\ref{tab:kitti_projection_3d_center}.
\minisection{Importance of each sub-affinity matrix.}
From~Table~\ref{tab:nusc_ablation_drop_affinity}, we observe that adding a deep feature affinity ($\mathbf{A}_{\subfix{deep}}$) distinguishes two near-overlapping objects and thus increasing AMOTA@1 with an observable margin. Adding 3D object state affinity ($\mathbf{A}_{\subfix{iou}}$) also contributes to AMOTA@1 effectively. Lastly, we found in~Table~\ref{tab:nusc_ablation_full_frame} that applying our model on \textit{full frames} during inference helps the motion model estimate an objects' velocity more accurately. Motion-based methods benefit from using a higher frame rate setting, \textit{full frames}, with a $(0.2421-0.2329)/0.2421 = 3.8\%$ improvement on LSTM and $0.6\%$ on KF3D.
\begin{table}[thpb]
\figcaption{Ablation study of tracking performance with different methods on nuScenes validation set.}{Results in (a) suggest that the deep feature affinity matrix contributes to data association. From (b), we observe a robust trend that a motion model with a higher frame rate image sequence encourages tracking performance.}
\begin{subtable}[t]{\linewidth}
\centering
\small
\caption{Ablation study of dropping affinity matrix.}
\adjustbox{width=\linewidth}{
\begin{tabular}{c|cccc|rr}
\toprule
{Method} & $\mathbf{A}_{\subfix{deep}}$ & $\mathbf{A}_{\subfix{iou}}$ & $\mathbf{A}_{\subfix{depth}}$ & $\hat{\mathbf{c}}_{depth}$ & {AMOTA@1} $\uparrow$ & {AMOTP} $\downarrow$ \\
\midrule
\multirow{5}{*}{KF3D}
& - & \checkmark & \checkmark & \checkmark & 0.1779 & 1.596 \\
& \checkmark & - & \checkmark & \checkmark & 0.2013 & 1.540 \\
& \checkmark & \checkmark & - & \checkmark & 0.2213 & \textbf{1.535} \\
& \checkmark & \checkmark & \checkmark & - & 0.2301 & \textbf{1.535} \\
& \checkmark & \checkmark & \checkmark & \checkmark & \textbf{0.2306} & \textbf{1.535} \\
\bottomrule
\end{tabular}
}
\label{tab:nusc_ablation_drop_affinity}
\end{subtable}
\begin{subtable}[t]{\linewidth}
\centering
\small
\medskip
\caption{Ablation study of \textit{full frames}.}
\adjustbox{width=0.9\linewidth}{
\begin{tabular}{c|c|rr}
\toprule
{Method} & {full frames} & {AMOTA@1} $\uparrow$ & {AMOTP} $\downarrow$ \\
\midrule
\multirow{2}{*}{KF3D}
& - & 0.2306 & 1.535 \\
& \checkmark & \textbf{0.2321} & \textbf{1.530} \\
\midrule
\multirow{2}{*}{VeloLSTM}
& - & 0.2329 & 1.528 \\
& \checkmark & \textbf{0.2421} & \textbf{1.518} \\
\bottomrule
\end{tabular}
}
\label{tab:nusc_ablation_full_frame}
\end{subtable}
\end{table}
\minisection{Importance of sub-affinity module design.}
Comparing boxes IoU sub-affinity ($\mathbf{A}_{\subfix{iou}}$), adding depth-order matching, denoted \textit{3D} in~Table~\ref{tab:nusc_ablation_swap_module}, as overlapping threshold helps identify falsely rejected objects and increases overall AMOTA@1 scores by a relative $(0.2306-0.1381)/0.2306 = 40.1\%$ AMOTA@1.
With motion-aware association, denoted \textit{motion} in~Table~\ref{tab:nusc_ablation_swap_module}, our full model filtering out objects with wrong motion velocity reaches $0.2306$ AMOTA@1 comparing to $0.2234$ using only \textit{cosine} angle difference module or $0.2300$ using \textit{centroid} distance module.
\begin{table}[thpb]
\centering
\small
\figcaption{Tracking performance comparison with different designs of affinity matrix on nuScenes validation set.}{Results suggest that using 3D IoU and motion-based sub-affinity matrix yielding the best validation performance of our final pipeline.}
\adjustbox{width=0.80\linewidth}{
\begin{tabular}{ll|rr}
\toprule
$\mathbf{A}_{\subfix{iou}}$ & $\mathbf{A}_{\subfix{depth}}$ & {AMOTA@1} $\uparrow$ & {AMOTP} $\downarrow$ \\
\midrule
2D & \multirow{3}{*}{motion} & 0.1381 & 1.674 \\
BEV & & 0.1929 & 1.599 \\
3D & & \textbf{0.2306} & \textbf{1.535} \\
\midrule
\multirow{3}{*}{3D} & cosine & 0.2234 & \textbf{1.534} \\
& centroid & 0.2300 & 1.536 \\
& motion & \textbf{0.2306} & 1.535 \\
\bottomrule
\end{tabular}
}
\label{tab:nusc_ablation_swap_module}
\end{table}
\minisection{Importance of 3D center projection estimation.}
We estimate the 3D location of a bounding box by predicting the projection of its center and depth, while Mousavian~\emph{et al}\onedot~\cite{mousavian2017deep3dbox} uses the center of detected 2D boxes directly.
Table~\ref{tab:kitti_projection_3d_center} shows the comparison of these two methods on the KITTI dataset.
The result indicates that the correct 3D projections provide a higher tracking capacity for a motion module to associate candidates and significantly reduce the centroid misalignment (MOTP$_\subfix{C}$).
\begin{table}[htpb]
\figcaption{Importance of using projection of 3D bounding box center estimation on KITTI sub-validation set.}{
We evaluate our KF3D model using different center inputs $C$ on the Car category to reveal the importance of estimating the projection of a 3D center.
The increase of MOTA and higher 3D IoU AP with COCO $(50:5:95)$ suggest that the projection of a 3D center benefits our tracking pipeline over the 2D center.
}
\centering
\adjustbox{width=\linewidth}{
\begin{tabular}{l|lrr||lrr}
\toprule
Method & Range & MOTA $\uparrow$ & MOTP$_\subfix{C}$ $\downarrow$ & Difficulty & $AP_\subfix{bev}^\subfix{COCO}$ $\uparrow$ & $AP_\subfix{3d}^\subfix{COCO}$ $\uparrow$ \\
\midrule
\multirow{3}{*}{2D Cen}
& 0-30m & 77.28 & 0.56 & Easy & 24.91 & 15.11 \\
& 0-50m & 70.78 & 0.60 & Medium & 23.53 & 15.87 \\
& 0-100m & 65.15 & 0.62 & Hard & 21.99 & 14.99 \\
\midrule
\multirow{3}{*}{3D Cen}
& 0-30m & \textbf{79.50} & \textbf{0.44} & Easy & \textbf{41.71} & \textbf{36.74} \\
& 0-50m & \textbf{72.53} & \textbf{0.51} & Medium & \textbf{33.73} & \textbf{29.30} \\
& 0-100m & \textbf{66.77} & \textbf{0.54} & Hard & \textbf{31.05} & \textbf{26.67} \\
\bottomrule
\end{tabular}
}
\label{tab:kitti_projection_3d_center}
\end{table}
\subsection{Motion Modeling Comparison.}
\label{sec:motion_modeling_comparison}
For motion-based models, update-and-predict cycles help single-frame 3D estimates.
To examine the effectiveness of our VeloLSTM, we compare variants of motion models to pure detection in refining the object locations along their trajectories.
\minisection{Pure Detection (Detection).}
A single-frame monocular 3D estimate provides a noisy measurement of an object learned from the data distribution.
\minisection{Dummy Motion Model (Momentum).}
Comparing with the non-linear motion model, we experimented with a dummy motion model that directly updated the object state with a fixed momentum on observed measurements and predicted the next possible state with zero motion.
By applying motion momentum, we expect the dummy model to output linearly smoothed trajectories.
\minisection{Kalman Filter 3D baseline (KF3D).}
We take a Kalman filter-based motion model as our baseline which models $$\{x, y, z, \theta, l, w, h, \Delta x, \Delta y, \Delta z\}$$ in 3D world coordinates.
The well-tuned Kalman filter~\cite{kalman1960new} models a sequence of observed measurements with Gaussian noise assumptions to find a prediction location that minimize a mean of square error.
\minisection{Deep Motion Estimation and Update (VeloLSTM).}
As mentioned in~Section~\ref{sec:lstm_motion}, we propose a VeloLSTM network to model the vehicle motion.
To analyze the effectiveness of our deep motion model, we compare our LSTM model with traditional 3D Kalman filter (KF3D) and single-frame 3D estimation (Detect) using 3D detection and tracking evaluation metrics from nuScenes.
Table~\ref{tab:nusc_motion_helps_3dap} shows that KF3D provides a small improvement in NDS via trajectory smoothing within prediction and observation.
On the other hand, our VeloLSTM module provides a learned estimation based on past $n=5$ velocity predictions and current frame observation, which may compensate for the observation error.
Our VeloLSTM module achieves the highest AMOTA@1, AMOTP, and NDS, among the other methods.
\begin{table}[htpb]
\centering
\figcaption{Comparison of location refinement with different motion model on nuScenes validation set.}{Result suggests the VeloLSTM achieves the highest performance among the other methods. Besides, employing a motion model benefits 3D detection and tracking. We use \textit{full frames} and final affinity matrices.}
\small
\adjustbox{width=0.9\linewidth}{
\begin{tabular}{l|rrr}
\toprule
{Method} & {NDS} & {AMOTA@1} $\uparrow$ & {AMOTP} $\downarrow$ \\
\midrule
Detection & 0.3622 & 0.222 & 1.538 \\
Momentum & 0.3620 & 0.227 & 1.532 \\
KF3D & 0.3634 & 0.232 & 1.530 \\
VeloLSTM & \textbf{0.3666} & \textbf{0.242} & \textbf{1.518} \\
\bottomrule
\end{tabular}
}
\label{tab:nusc_motion_helps_3dap}
\end{table}
\subsection{Real-world Evaluation.}
\label{sec:real_world_evaluation}
Besides evaluating on large-scale synthetic data, we utilize nuScenes~\cite{caesar2019nuscenes}, Waymo Open~\cite{waymo} and KITTI~\cite{geiger2012kitti} benchmarks to compare our model abilities with other state of the arts.
Major results are listed in Table~\ref{tab:nusc_tracking_test} for nuScenes dataset and Table~\ref{tab:waymo_tracking_test} and Table~\ref{tab:waymo_detection_test} for Waymo Open dataset.
Qualitative results on all three datasets are shown in~Figure~\ref{fig:qualitative_result}.
\begin{figure*}[htpb]
\adjustbox{width=1.0\linewidth}{
\setlength{\tabcolsep}{2pt}
\begin{tabular}{ccc}
\includegraphics[width=0.33\textwidth]{fig/Fig_qualitative_vis/nusc_test_71_tracking-0002.jpg} &
\includegraphics[width=0.33\textwidth]{fig/Fig_qualitative_vis/nusc_test_71_tracking-0003.jpg} &
\includegraphics[width=0.33\textwidth]{fig/Fig_qualitative_vis/nusc_test_71_tracking-0004.jpg} \\
\includegraphics[width=0.33\textwidth]{fig/Fig_qualitative_vis/waymo_test_11_tracking-0002.jpg} &
\includegraphics[width=0.33\textwidth]{fig/Fig_qualitative_vis/waymo_test_11_tracking-0003.jpg} &
\includegraphics[width=0.33\textwidth]{fig/Fig_qualitative_vis/waymo_test_11_tracking-0004.jpg}
\end{tabular}
}
\figcaption{Qualitative results on testing set of nuScenes (First row), Waymo Open (Second row) datasets.}{Our proposed quasi-dense 3D tracking pipeline estimates accurate 3D extent and robustly associates tracking trajectories from a monocular image. We show predicted 3D bounding boxes and trajectories colored with tracking IDs.
}
\label{fig:qualitative_result}
\end{figure*}
\begin{figure}[htpb]
\adjustbox{width=1.0\linewidth}{
\setlength{\tabcolsep}{2pt}
\begin{tabular}{c}
\includegraphics[width=\textwidth]{fig/Fig_qualitative_vis/kitti_test_13_tracking-0001.jpg} \\
\includegraphics[width=\textwidth]{fig/Fig_qualitative_vis/kitti_test_13_tracking-0002.jpg} \\
\includegraphics[width=\textwidth]{fig/Fig_qualitative_vis/kitti_test_13_tracking-0003.jpg} \\
\includegraphics[width=\textwidth]{fig/Fig_qualitative_vis/kitti_test_13_tracking-0004.jpg}
\end{tabular}
}
\figcaption{Qualitative results on testing set of KITTI datasets.}{Our VeloLSTM continues to predict the 3D object state of unmatched tracklets after they disappear. With motion-based data association, our tracker successfully recovers tracked vehicle (cyan-colored box) from object occlusions after the vehicle re-appeared. We show predicted 3D bounding boxes and trajectories colored with tracking IDs.
}
\label{fig:qualitative_result_kitti}
\end{figure}
\minisection{nuScenes tracking challenge.}
Our monocular 3D tracking method outperforms all the published methods with a large margin.
Center-Track-Vision~\cite{zhou2020centertrack} uses two consecutive frames to generate inter-frame motion for object detection and 3D tracking.
Center-Track-Open~\cite{zhou2020centertrack} fuses LiDAR information in the Center-Track-Vision pipeline with Megvii-detector~\cite{zhu2019cbgs} to generate 3D detection.
The LiDAR-based baselines~\cite{caesar2019nuscenes} uses state-of-the-art LiDAR-based detectors~\cite{shi2018pointrcnn, lang2019pointpillars, zhu2019cbgs} to estimate accurate bounding boxes and feed into a Kalman-Filter-based 3D tracker, AB3DMOT~\cite{weng2020ab3dmot}.
Our method takes only one image for 3D object detection and tracking and leverages our deep motion model to refine object trajectories compared to the prior arts.
The nuScenes tracking dataset linearly interpolates GT tracks to avoid track fragments from LiDAR point filtering and removes GT objects without LiDAR points. Both invisible objects with annotation and visible objects without annotation prohibit the camera-based methods from optimizing bounding box estimation.
Nevertheless, our quasi-dense 3D tracking approach reaches $0.217$ AMOTA with near five times tracking accuracy of the best vision-only submission among all published methods.
\begin{table}[htpb]
\centering
\figcaption{Tracking performance on the testing set of nuScenes tracking benchmark~\cite{caesar2019nuscenes}.}{We report the average AMOTA@1 and AMOTP over $7$ categories on the benchmark with only published methods shown. Our quasi-dense 3D tracking pipeline outperforms the best camera-based submission (underlined) by near $500\%$ while bridging the gap to LiDAR-based methods on the nuScenes 3D tracking benchmark.
}
\adjustbox{width=\linewidth}{
\begin{tabular}{ll|rr}
\toprule
Method & Modality & AMOTA@1 $\uparrow$ & AMOTP $\downarrow$ \\
\midrule
\textbf{QD3DT (Ours)} & Camera & \textbf{0.217} & 1.550 \\
Megvii-AB3DMOT~\cite{caesar2019nuscenes} & LiDAR & 0.151 & 1.501 \\
CenterTrack-Open~\cite{zhou2020centertrack} & LiDAR + Camera & 0.108 & \textbf{0.989} \\
CenterTrack-Vision~\cite{zhou2020centertrack} & Camera & \underline{0.046} & 1.543 \\
PointPillars-AB3DMOT~\cite{caesar2019nuscenes} & LiDAR & 0.029 & 1.703 \\
Mapillary-AB3DMOT~\cite{caesar2019nuscenes} & LiDAR & 0.018 & 1.790 \\
\bottomrule
\end{tabular}
}
\label{tab:nusc_tracking_test}
\end{table}
\minisection{Waymo Open benchmark.}
We provide a strong baseline being the first entry in camera-only modality for both 3D tracking and detection challenges shown in~Table~\ref{tab:waymo_tracking_test} and~\ref{tab:waymo_detection_test}.
\begin{table}[htpb]
\begin{subtable}[t]{\textwidth}
\centering
\caption{3D Tracking Performance}
\adjustbox{width=0.9\linewidth}{
\begin{tabular}{l|l|rr}
\toprule
{Method} & {Modality} & {MOTA} $\uparrow$ & {MOTP} $\downarrow$ \\
\midrule
HorizonMOT3D~\cite{wang20201st} & Lidar+Camera & 0.6407 & 0.1577 \\
CenterPoint~\cite{yin2020center} & Lidar & 0.5938 & 0.1637 \\
pillars\_kf\_baseline~\cite{waymo} & Lidar & 0.4008 & 0.1856 \\
\midrule
QD3DT (Ours) & Camera & 0.0001 & 0.0658 \\
\bottomrule
\end{tabular}
}
\label{tab:waymo_tracking_test}
\end{subtable}
\begin{subtable}[t]{\textwidth}
\centering
\medskip
\caption{3D Detection Performance.}
\adjustbox{width=0.9\linewidth}{
\begin{tabular}{l|l|rr}
\toprule
{Method} & {Modality} & {mAPH} $\uparrow$ & {mAP} $\uparrow$ \\
\midrule
HorizonMOT3D~\cite{wang20201st} & Lidar+Camera & 0.7783 & 0.7823 \\
PV-RCNN~\cite{openpcdet2020} & Lidar & 0.7323 & 0.7369 \\
CenterPoint~\cite{yin2020center} & Lidar & 0.7299 & 0.7342 \\
\midrule
QD3DT (Ours) & Camera & 0.0233 & 0.0242 \\
\bottomrule
\end{tabular}
}
\label{tab:waymo_detection_test}
\end{subtable}
\figcaption{3D detection and tracking performance in Vehicle Level 2 difficulty on the testing set of Waymo dataset.}{Our pipeline serves as a strong baseline with non-zero scores using only camera modality in the 3D detection and tracking challenge on LiDAR-dominated benchmark.}
\end{table}
\begin{table}[htpb]
\centering
\figcaption{3D detection and tracking performance in Vehicle Level 2 difficulty on the validation set of Waymo dataset.}{We compare 3D detection and tracking results on different extent of the observable area. LiDAR-based GT covers a 360-degree area, while camera-based GT filters out annotations outside of camera view. MOTA and mAPH increase greatly while lowering IoU thresholds slightly.}
\small
\adjustbox{width=0.9\linewidth}{
\begin{tabular}{l|l|c|rr}
\toprule
{Modality} & {Range} & {IoU} & {MOTA} $\uparrow$ & {mAPH} $\uparrow$ \\
\midrule
\multirow{3}{*}{LiDAR-based GT} &
\multirow{3}{*}{Overall} & 0.3 & 0.1183 & 0.2242 \\
& & 0.5 & 0.0195 & 0.1161 \\
& & 0.7 & 1.8e-06 & 0.0286 \\
\midrule
\multirow{3}{*}{Camera-based GT} &
\multirow{3}{*}{Overall} & 0.3 & 0.1867 & 0.3401 \\
& & 0.5 & 0.0308 & 0.1743 \\
& & 0.7 & 2.8e-06 & 0.0408 \\
\bottomrule
\end{tabular}
}
\label{tab:waymo_validation_evaluation}
\end{table}
\subsection{Evaluation Metrics.}
\label{sec:evaluation_metrics}
We observe from~Table~\ref{tab:waymo_tracking_test} that the IoU threshold at $0.7$ prohibits the camera-based method from evaluating meaningful tracking performance.
However, the score does not cohere to our qualitative results in~Figure~\ref{fig:qualitative_result}.
Therefore, we conduct ablation studies to validate the power of our data-driven method.
We first compare 3D detection and tracking results on different extent of observable area,~\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, for the annotations from the whole field of view, denoted \textit{LiDAR-based GT}, and ones from full camera field of view, denoted \textit{Camera-based GT}.
The \textit{LiDAR-based GT} setting constantly yields inferior performance with approximate a $5/8$ to the \textit{Camera-based GT} ones in different IoU thresholds in~Table~\ref{tab:waymo_validation_evaluation}.
The difference matches the ratio of an observable area by cameras and area by a LiDAR.
We turn to centroid-based metrics to validate our pipeline's performance since the IoU-based metrics set up an unreachable bar for the camera-based method.
With centroid-based metrics, the tracking scores of our final model on nuScenes and Waymo Open datasets list in~Table~\ref{tab:cross_dataset_centroid_based_evaluation}.
The similarity in MOTA and MOTP performance suggests that our pipeline has enough capacity to handle different datasets with minimal change of training procedure, underlining the robustness of our tracking pipeline.
From~Table~\ref{tab:cross_dataset_centroid_based_evaluation} and~Table~\ref{tab:waymo_validation_evaluation}, our method acts comparable in IoU-based metrics with a threshold at $0.3$ to centroid-based metrics with a threshold at $2$ meters on the camera-based GT setting.
We argue that the IoU threshold at $0.7$ is overly biased toward a precise 3D object attributes estimation, like depth estimation, 3D center, orientation, 3D center or dimensions.
A trajectory with perfect matching but a moderate error in one of the object attributes, \emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, $1$m off in depth, is not counted as a true positive. However, it is treated equally as those with a completely wrong estimation \emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, $10$m away.
Besides, the experiment in~Table~\ref{tab:cross_dataset_centroid_based_evaluation} and~Table~\ref{tab:waymo_validation_evaluation} show an interesting discrepancy in reported performance by different types of evaluation metrics, which supports analysis in~\cite{caesar2019nuscenes} to choose centroid-based evaluation metrics as a fair comparison to both camera-based and LiDAR-based methods.
\begin{table}[htpb]
\figcaption{Centroid-based evaluation across datasets.}{
This experiment uses our final model and the \textit{full frames} setting for nuScenes.
We evaluate the results in different datasets to reveal the importance of choosing evaluation metrics. The similarity of MOTA and MOTP performance suggests that our pipeline has been well trained and performs equally well on the Waymo Open dataset.
}
\centering
\adjustbox{width=\linewidth}{
\begin{tabular}{l|lrrrr}
\toprule
Method & Range & MOTA $\uparrow$ & MOTP$_\subfix{C}$ $\downarrow$ & MOTP$_\subfix{O}$ $\downarrow$ & MOTP$_\subfix{I}$ $\downarrow$ \\
\midrule
\multirow{3}{*}{Waymo}
& 0-30m & 57.33 & 0.61 & 14.66 & 0.18 \\
& 0-50m & 36.87 & 0.74 & 16.10 & 0.18 \\
& 0-100m & 18.66 & 0.79 & 17.63 & 0.19 \\
\midrule
\multirow{3}{*}{nuScenes}
& 0-30m & 44.47 & 0.64 & 14.18 & 0.19 \\
& 0-50m & 25.93 & 0.73 & 14.84 & 0.19 \\
& 0-100m & 12.14 & 0.75 & 15.05 & 0.19 \\
\bottomrule
\end{tabular}
}
\label{tab:cross_dataset_centroid_based_evaluation}
\end{table}
\subsection{Amount of Data Matters.}
\label{sec:amount_of_data_matters}
We train the depth estimation module with $1\%$, $10\%$, and $100\%$ training data on the GTA dataset.
The results show how we can benefit from more data in Table~\ref{tab:data_amount}, where a consistent trend of performance improvement emerges as the number of data increases.
The trend of our results with a different amount of training data indicates that large-scale 3D annotation is helpful, especially with the ground truth of distant and small objects.
\begin{table}[htpb]
\figcaption{Performance of 3D estimation on object detection IoU mAP.}{
The evaluation demonstrates the effectiveness of our model from each separate metric.
With different amounts of training data in our GTA dataset, the results suggest that large data capacity benefits the performance of a data-hungry network.}
\centering
\adjustbox{width=0.9\linewidth}{
\begin{tabular}{lr|cccc}
\toprule
\multirow{2}{*}{Dataset} & \multirow{2}{*}{Amount} & \multicolumn{4}{c}{Medium} \\
& & $AP_\subfix{bev}^{70}$ & $AP_\subfix{3d}^{70}$ & $AP_\subfix{bbox}^{70}$ & $AP_\subfix{aos}^{70}$ \\
\midrule
\multirow{3}{*}{GTA}
& {1\%} & 4.10 & 1.31 & 78.18 & 75.59 \\
& {10\%} & 14.61 & 8.45 & 91.66 & 90.83 \\
& {100\%} & \textbf{21.61} & \textbf{15.63} & \textbf{94.62} & \textbf{94.23} \\
\bottomrule
\end{tabular}
}
\label{tab:data_amount}
\end{table}
\subsection{Comparison of matching algorithms.}
\label{sec:comparison_of_matching_algorithm}
A tracker solves the assignment problem often using combinatorial optimization algorithms, \emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, Hungarian algorithm.
However, from the comparison in~Table~\ref{tab:nusc_ablation_matching_algo}, we found the well-trained instance embeddings and 3D object states are robust in matching possible pairs without needing an optimizing algorithm.
Besides, our method benefits from the lower computational complexity compared to using polynomial-time optimization algorithms.
Therefore, we choose to greedily match bipartite quasi-dense pairs of detections and tracklets throughout the experiments.
\begin{table}[htpb]
\centering
\figcaption{Comparison of the different matching algorithms on nuScenes validation set.}{We use only keyframes and all affinity matrices as our setting. We found that using greedy matching yields similar results to the Hungarian matching but with less computation complexity on well-trained quasi-dense embedding pairs.}
\adjustbox{width=0.9\linewidth}{
\begin{tabular}{l|rrrrr}
\toprule
Matching Algorithm & {AMOTA@1} $\uparrow$ & {[email protected]} $\uparrow$ \\
\midrule
Hungarian~\cite{Kuhn1955hungarian} & 0.230 & 0.3479 \\
Greedy & 0.230 & 0.3479 \\
\bottomrule
\end{tabular}
}
\label{tab:nusc_ablation_matching_algo}
\end{table}
\section{Conclusion}
We propose an online 3D object detection and tracking framework, combining quasi-dense similarity learning and 3D instance dynamics, to track moving objects in a 3D world.
Our updated pipeline consists of four parts: a single-frame monocular 3D object inference model, cross-frames contrastive feature learning network, multimodal affinity matching scheme for inter-frame object association, and an LSTM-based motion model for refining the 3D extents from location trajectories.
Besides, we introduce a 3D detection confidence to provide a balancing cue of single-frame depth estimation and multi-frame motion model refinement.
We present 3D bounding boxes depth-ordering matching for robust instance association and utilize motion-based 3D trajectory prediction for re-identification of occluded vehicles.
We design an object movement learning module, termed VeloLSTM, to update each object's location independent of camera movement.
Moreover, our ablation study and experiments show that our quasi-dense 3D tracking pipeline takes advantage of dynamic 3D trajectories and offers robust data association on urban-driving situations.
On the Waymo Open benchmark, we established the first strong baseline in camera modality with positive scores in both the 3D detection and tracking challenge.
Our quasi-dense 3D tracking pipeline outperforms the camera-based state of the art by near $500\%$ while bridging the gap to LiDAR-based methods on nuScenes 3D tracking benchmark.
\section*{Foreword}
This appendix provides technical details about our monocular quasi-dense 3D object tracking network, our training procedures for different datasets, and more qualitative and quantitative results.
Section~\ref{apn:data_stats} offers frame- and object-based statistical summaries of our proposed simulation dataset.
Section~\ref{apn:training_detail} describes our training procedure, and network setting of each dataset.
Section~\ref{apn:experiments} illustrates various comparisons, including inference time, network settings and performance results, of our method on nuScenes, Waymo, GTA, and KITTI dataset.
\section{Dataset Statistics}
\label{apn:data_stats}
To help understand our dataset and its difference, we show more statistics.
\minisection{Dataset Comparison.}
Table~\ref{tab:dataset} demonstrates a detailed comparison with related datasets, including detection, tracking, and driving benchmarks.
KITTI-D~\cite{geiger2012kitti} and KAIST~\cite{hwang2013multispectral} are mainly detection datasets, while KITTI-T~\cite{geiger2012kitti}, MOT15~\cite{leal2015motchallenge}, MOT16~\cite{milan2016mot16}, and UA-DETRAC~\cite{wen2015ua} are primarily 2D tracking benchmarks.
The common drawback could be the limited scale, which cannot meet the growing demand for training data.
Compared to the related synthetic dataset, Virtual KITTI~\cite{gaidon2016virtual} and VIPER~\cite{richter2017playing}, we additionally provide fine-grained attributes of object instances, such as color, model, maker attributes of a vehicle, motion and control signals, which leaves the space for the imitation learning system.
\begin{table*}[htpb]
\figcaption{Comparison to related dataset for detection and tracking (Upper half: real-world, Lower half: synthetic).}{We only count the size and types of annotations for training and validation (D=detection, T=tracking, C=car, P=pedestrian). To our knowledge, our dataset is the largest 3D tracking benchmark for dynamic scene understanding, with control signals of driving, sub-categories of object.}
\label{tab:dataset}
\centering
\adjustbox{width=0.9\linewidth}{
\begin{tabular}{cccccccccc}
\toprule
Dataset & Task & Object & Frame & Track & Boxes & 3D & Weather & Occlusion & Ego-Motion\\ \midrule
KITTI-D~\cite{geiger2012kitti} & D & C,P & 7k & - & 41k & \checkmark & - & \checkmark & - \\
KAIST~\cite{hwang2013multispectral} & D & P & 50k & - & 42k & - & \checkmark & \checkmark & - \\
KITTI-T~\cite{geiger2012kitti} & T & C & 8k & 938 & 65k & \checkmark & - & \checkmark & \checkmark \\
MOT15-3D~\cite{leal2015motchallenge} & T & P & 974 & 29 & 5632 & \checkmark & \checkmark & - & - \\
MOT15-2D~\cite{leal2015motchallenge} & T & P & 6k & 500 & 40k & - & \checkmark & - & - \\
MOT16~\cite{milan2016mot16} & T & C,P & 5k & 467 & 110k & - & \checkmark & \checkmark & - \\
UA-DETRAC~\cite{wen2015ua} & D,T & C & 84k & 6k & 578k & - & \checkmark & \checkmark & - \\
\midrule
Virtual KITTI~\cite{gaidon2016virtual} & D,T & C & 21k & 2610 & 181k & \checkmark & \checkmark & \checkmark & \checkmark \\
VIPER~\cite{richter2017playing} & D,T & C,P & 184k & 8272 & 5203k & \checkmark & \checkmark & \checkmark & \checkmark \\
Ours & D,T & C,P & 688k & 325k & 10068k & \checkmark & \checkmark & \checkmark & \checkmark \\
\bottomrule
\end{tabular}
}
\end{table*}
\minisection{Number of Instances in Each Category.}
The vehicle diversity is also very large in the GTA world, featuring $15$ fine-grained subcategories. We analyzed the distribution of the $15$ subcategories in Figure~\ref{fig:statistics_category}.
Besides instance categories, we also show the distribution of occluded (Figure~\ref{fig:statistics_occlusion}) and truncated (Figure~\ref{fig:statistics_truncated}) instances to support the problem of partially visible in the 3D coordinates.
\begin{figure}[htpb]
\minipage{1.0\linewidth}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_statistics_supp/Types.pdf}
\subcaption{Category}\label{fig:statistics_category}
\endminipage
\hfill
\minipage{0.49\linewidth}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_statistics_supp/Truns.pdf}
\subcaption{Truncation}\label{fig:statistics_truncated}
\endminipage
\minipage{0.49\linewidth}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_statistics_supp/Occ.pdf}
\subcaption{Occlusion}\label{fig:statistics_occlusion}
\endminipage
\figcaption{The statistics of object in our dataset.}{}
\label{fig:statistics_object}
\end{figure}
\minisection{Weather and Time of Day.}
Figure~\ref{fig:statistics_scene} shows the distribution of weather, hours of our dataset.
It features a full weather cycle and time of a day in a diverse virtual environment.
By collecting various weather cycles (Figure~\ref{fig:statistics_weather}), our model learns to track with a higher understanding of environments.
With different times of a day (Figure~\ref{fig:statistics_hours}), the network handles changeable perceptual variation.
\begin{figure}[htpb]
\minipage{0.49\linewidth}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_statistics_supp/Weather.pdf}
\subcaption{Weather}\label{fig:statistics_weather}
\endminipage
\minipage{0.49\linewidth}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_statistics_supp/Hours.pdf}
\subcaption{Hours}\label{fig:statistics_hours}
\endminipage
\figcaption{The statistics of scene in our dataset.}{}
\label{fig:statistics_scene}
\end{figure}
\minisection{Examples of Our Dataset.}
Figure~\ref{fig:dataset_example} shows some visual examples in different time, weather and location.
\begin{figure*}[htpb]
\adjustbox{width=1.0\linewidth}{
\setlength{\tabcolsep}{2pt}
\begin{tabular}{rl}
\includegraphics[width=0.49\linewidth, keepaspectratio]{fig/Fig_dataset_example/1540502872975_final.jpg} &
\includegraphics[width=0.49\linewidth, keepaspectratio]{fig/Fig_dataset_example/1540517865792_final.jpg} \\
\includegraphics[width=0.49\linewidth, keepaspectratio]{fig/Fig_dataset_example/1540520555979_final.jpg} &
\includegraphics[width=0.49\linewidth, keepaspectratio]{fig/Fig_dataset_example/1540587579898_final.jpg} \\
\includegraphics[width=0.49\linewidth, keepaspectratio]{fig/Fig_dataset_example/1540592284296_final.jpg} &
\includegraphics[width=0.49\linewidth, keepaspectratio]{fig/Fig_dataset_example/1539042300186_final.jpg} \\
\includegraphics[width=0.49\linewidth, keepaspectratio]{fig/Fig_dataset_example/1539050273914_final.jpg} &
\includegraphics[width=0.49\linewidth, keepaspectratio]{fig/Fig_dataset_example/1540486835782_final.jpg}
\end{tabular}
}
\figcaption{Examples of our GTA dataset.}{We collected a diverse set of driving scenes in different weather, location and time of a day.}
\label{fig:dataset_example}
\end{figure*}
\section{Training Details}
\label{apn:training_detail}
We optimize our model on each dataset with different training procedures according to its amount of data and the limit of the GPU memory.
\minisection{Training Procedure.}
For nuScenes, we train our model on the training set with all $10$ object classes and evaluate the performance on the validation set with $10$ object classes for detection task and $7$ object classes for tracking task.
We trained the model on $8$ GPUs with a total batch size of $32$ for $24$ epochs. The learning rate was linearly increased from $1\times10^{-3}$, to $1\times10^{-2}$ over the initial $1000$ warm-up training steps decreased by $0.5$ after $16$ and $22$ epochs.
For Waymo, we trained the model on $8$ GPUs with a total batch size of $16$ for $24$ epochs. The learning rate was linearly increased from $5\times10^{-4}$, to $5\times10^{-3}$ over the initial $1000$ warm-up training steps. Also, due to the large amount of the training data, we decrease the learning rate by $0.5$ after $8$, $12$, $16$, $20$, and $22$ epochs.
For GTA, we trained the model on $8$ GPUs with a total batch size of $24$ for $24$ epochs. The learning rate was linearly increased from $1\times10^{-3}$, to $1\times10^{-2}$ over the initial $1000$ warm-up training steps. Due to the large amount of the training data, we decreased the learning rate by $0.5$ after $8$, $12$, $16$, $20$, and $22$ epochs.
For KITTI, we fine-tuned the model from a GTA 3D detection model on $4$ GPUs with a total batch size of $8$ for $24$ epochs. The learning rate was linearly increased from $5\times10^{-4}$, to $5\times10^{-3}$ over the initial $1000$ warm-up training steps and decreased by $0.5$ after $16$ and $22$ epochs.
\section{Experiments}
\label{apn:experiments}
\minisection{Inference Time.}
The average inference time is $123.3+2.3+12.3+1.9+22.2=162.0$ ms with $11095$ frames on a single RTX 2080Ti GPU (see Table~\ref{tab:inference_time} for details).
Note that the KITTI benchmark focuses only on the non-detection part inference time ($22.2$ ms).
\begin{table}[htpb]
\adjustbox{width=0.7\linewidth}{
\begin{tabular}{l|c}
\toprule
Phase & Second \\
\midrule
\textbf{Detection} & \textbf{0.1256} \\
\midrule
2D Box, 3D Estimation & 0.1233 \\
Contrastive Feature & 0.0023 \\
\midrule
\textbf{3D transform} & \textbf{0.0142} \\
\midrule
3D Lifting (img to world) & 0.0123 \\
Reprojection (world to img) & 0.0019 \\
\midrule
\textbf{Tracking} & \textbf{0.0222} \\
\midrule
LSTM Predict & 0.0076 \\
Greedy Matching & 0.0019 \\
LSTM Update & 0.0009 \\
Misc. & 0.0118 \\
\midrule
\textbf{Total} & \textbf{0.1620} \\
\bottomrule
\end{tabular}
}
\figcaption{Inference Time (second) of our proposed framework on KITTI tracking benchmark.}{We recorded the wall clock of the execution time consumption of each module, which might have slight overhead, \emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, functions with result dumping. Bold fonts are the summation of each sub-module and misc. time. Note that elapsed time for object detection is not included in the specified runtime of the KITTI benchmark.}
\label{tab:inference_time}
\end{table}
\minisection{Data Association Weights.}
We use different weights of appearance during our experiment, 3D IoU overlap, and motion overlap for corresponding methods.
We select weight ratios based on the results of our validation set.
For deep appearance only methods, we give $100\%$ weighting to $w_{\subfix{deep}}$.
For 3D related methods, we set $w_{\subfix{deep}}$ with $0.5$ to balance appearance and 3D extents.
\minisection{Tracking Performance on nuScenes dataset.}
We report the full table of nuScenes tracking benchmark in~Table~\ref{tab:nusc_tracking_test_full}.
\begin{table*}[tpb]
\centering
\figcaption{Tracking performance on the testing set of nuScenes tracking benchmark~\cite{caesar2019nuscenes}.}{We report the average AMOTA@1 and AMOTP over $7$ categories on the benchmark. Camera-based and published methods are shown. Our quasi-dense 3D tracking achieves the top ranking with about $5$ times the 3D tracking performance of the best camera-based submission, CenterTrack Vision~\cite{zhou2020centertrack} (score underlined), at the time of submission. We mark the best result in \textbf{bold} and runner-up in \textit{italic}.
}
\adjustbox{width=0.8\linewidth}{
\begin{tabular}{lll|rr}
\toprule
Method & Published & Modality & [email protected] $\uparrow$ & AMOTP $\downarrow$ \\
\midrule
\textbf{Ours} & & Camera & \textbf{0.217} & 1.550 \\
DEFT & & Camera & \textit{0.177} & 1.564 \\
Megvii-AB3DMOT & \cite{caesar2019nuscenes} & LiDAR & 0.151 & 1.501 \\
CenterTrack-Open & \cite{zhou2020centertrack} & LiDAR + Camera & 0.108 & \textbf{0.989} \\
DBNet & & Camera & 0.072 & \textit{1.489} \\
ProTracker & & Camera & 0.072 & 1.628 \\
Buffalo\_Vision & & Camera & 0.059 & 1.490 \\
CaTracker & & Camera & 0.053 & 1.611 \\
CenterTrack-Vision & \cite{zhou2020centertrack} & Camera & \underline{0.046} & 1.543 \\
GDLG & & Camera & 0.045 & 1.819 \\
PointPillars-AB3DMOT & \cite{caesar2019nuscenes} & LiDAR & 0.029 & 1.703 \\
Mapillary-AB3DMOT & \cite{caesar2019nuscenes} & LiDAR & 0.018 & 1.790 \\
\bottomrule
\end{tabular}
}
\label{tab:nusc_tracking_test_full}
\end{table*}
\minisection{Tracking Performance on KITTI dataset.}
As mentioned in the main paper, we resort to KITTI~\cite{geiger2012kitti} tracking benchmarks to compare our model abilities in the real-world scenario.
We have improved our previous results~\cite{Hu2019Mono3DT} to third place with five ranks improvement and have increased from $84.52$ to $86.41$ in MOTA.
Our quasi-dense 3D tracking method ranked in the top tier among all the published methods upon the time of submission.
Please note that we focus on tracking module improvements and not over-design the detector-specific architecture to fit on a small dataset.
Results are listed in Table~\ref{tab:kitti_full_tracking}.
\begin{table*}[htpb]
\figcaption{Tracking performance on the testing set of KITTI tracking benchmark.}{Improving from mono3DT~\cite{Hu2019Mono3DT}, we have greatly increased the ranking position by leveraging the quasi-dense contrastive feature in the data association part. However, the goal objective of our training aims for 3D object tracking, and we are not optimizing the network for the 2D tracking task. Only published methods are reported.
We mark the best result in \textbf{bold} and runner-up in \textit{italic}.}
\centering
\adjustbox{width=0.8\linewidth}{
\begin{tabular}{lcc|rrrrrr}
\toprule
{Benchmark} & LiDAR & Stereo & {MOTA} $\uparrow$ & {MOTP} $\uparrow$ & {MT} $\uparrow$ & {ML} $\downarrow$ & {FP} $\downarrow$ & {FN} $\downarrow$ \\
\midrule
SRK\_ODESA~\cite{mykheievskyi2020odesa} & & & \textbf{90.03} & 84.32 & \textbf{82.62} & \textbf{2.31} & \textit{451} & \textit{2887} \\
CenterTrack~\cite{zhou2020centertrack} & & & \textit{89.44} & 85.05 & \textit{82.31} & \textbf{2.31} & 849 & \textbf{2666} \\
Ours & & & 86.41 & \textbf{85.82} & 75.38 & \textit{2.46} & 804 & \textit{3761} \\
Quasi-Dense~\cite{pang2020quasidense} & & & 85.76 & 85.01 & 69.08 & 3.08 & 517 & 4288 \\
JRMOT~\cite{shenoi2020jrmot} & \checkmark & & 85.70 & 85.48 & 71.85 & 4.00 & 772 & 4049 \\
MASS~\cite{karunasekera2019mass} & & & 85.04 & 85.53 & 74.31 & 2.77 & 742 & 4101 \\
MOSTFusion~\cite{voigtlaender2019mots} & & \checkmark & 84.83 & 85.21 & 73.08 & 2.77 & 681 & 4260 \\
mmMOT~\cite{zhang2019mmMOT} & \checkmark & & 84.77 & 85.21 & 73.23 & 2.77 & 711 & 4243 \\
Mono3DT~\cite{Hu2019Mono3DT} & & & 84.52 & 85.64 & 73.38 & 2.77 & 705 & 4242 \\
BeyondPixels~\cite{MOTBeyondPixels} & & & 84.24 & \textit{85.73} & 73.23 & 2.77 & 705 & 4247 \\
AB3DMOT~\cite{weng2020ab3dmot} & \checkmark & & 83.84 & 85.24 & 66.92 & 11.38 & 1059 & 4491 \\
PMBM~\cite{Scheidegger2018pmbm} & \checkmark & & 80.39 & 81.26 & 62.77 & 6.15 & 1007 & 5616 \\
extraCK~\cite{gunduz2018lightweight} & & & 79.99 & 82.46 & 62.15 & 5.54 & 642 & 5896 \\
MCMOT-CPD~\cite{Lee2016MCMOTCPD} & & & 78.90 & 82.13 & 52.31 & 11.69 & \textbf{316} & 6713 \\
NOMT~\cite{choi2015nomt} & & \checkmark & 78.15 & 79.46 & 57.23 & 13.23 & 1061 & 6421 \\
MDP~\cite{xiang2015mdptrack} & & & 76.59 & 82.10 & 52.15 & 13.38 & 606 & 7315 \\
DSM~\cite{frossard2018dsm} & & & 76.15 & 83.42 & 60.00 & 8.31 & 578 & 7328 \\
SCEA~\cite{hong2016scea} & & & 75.58 & 79.39 & 53.08 & 11.54 & 1306 & 6989 \\
CIWT~\cite{Osep17ICRAciwt} & & & 75.39 & 79.25 & 49.85 & 10.31 & 954 & 7345 \\
NOMT-HM~\cite{choi2015nomt} & & \checkmark & 75.20 & 80.02 & 50.00 & 13.54 & 1143 & 7280 \\
mbodSSP~\cite{lenz2015followme} & & & 72.69 & 78.75 & 48.77 & 8.77 & 1918 & 7360 \\
\bottomrule
\end{tabular}
}
\label{tab:kitti_full_tracking}
\end{table*}
\minisection{Qualitative results.}
We show our evaluation results in Figure~\ref{fig:qualitative_nusc} on the test set of nuScenes dataset and Figure~\ref{fig:qualitative_waymo} on the test set of Waymo Open Dataset.
The trajectories in~Figure~\ref{fig:qualitative_nusc} demonstrate our long-term tracking ability in the nuScenes dataset.
The results of Waymo show that our proposed 3D tracking pipeline establishes a strong baseline using monocular images.
The solid rectangular in bird's eye view stands for the predicted vehicle.
The 3D bounding boxes and the rectangular are colored with their tracking ID.
The figures are best viewed in color.
\begin{figure*}[htpb]
\minipage{0.50\textwidth}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_qualitative_nuscenes/nusc_test_56_tracking-0001.jpg}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_qualitative_nuscenes/nusc_test_56_tracking-0002.jpg}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_qualitative_nuscenes/nusc_test_56_tracking-0003.jpg}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_qualitative_nuscenes/nusc_test_56_tracking-0004.jpg}
\endminipage
\minipage{0.28\textwidth}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_qualitative_nuscenes/nusc_test_56_birdsview-0001.jpg}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_qualitative_nuscenes/nusc_test_56_birdsview-0002.jpg}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_qualitative_nuscenes/nusc_test_56_birdsview-0003.jpg}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_qualitative_nuscenes/nusc_test_56_birdsview-0004.jpg}
\endminipage
\figcaption{Qualitative results on the testing set of nuScenes dataset.}{Our QD-3DT robustly tracks all observed objects and locates them in 3D. We show predicted 3D bounding boxes and trajectories colored with tracking IDs. Better visualization with color.}
\label{fig:qualitative_nusc}
\end{figure*}
\begin{figure*}[htpb]
\minipage{0.46\textwidth}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_qualitative_waymo/waymo_test_3_tracking-0001.jpg}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_qualitative_waymo/waymo_test_3_tracking-0002.jpg}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_qualitative_waymo/waymo_test_3_tracking-0003.jpg}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_qualitative_waymo/waymo_test_3_tracking-0004.jpg}
\endminipage
\minipage{0.31\textwidth}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_qualitative_waymo/waymo_test_3_birdsview-0001.jpg}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_qualitative_waymo/waymo_test_3_birdsview-0002.jpg}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_qualitative_waymo/waymo_test_3_birdsview-0003.jpg}
\includegraphics[width=1.0\linewidth, keepaspectratio]{fig/Fig_qualitative_waymo/waymo_test_3_birdsview-0004.jpg}
\endminipage
\figcaption{Qualitative results on the testing set of Waymo Open dataset.}{Our QD-3DT accurately tracks all observed objects and locates them in 3D. We show predicted 3D bounding boxes and trajectories colored with tracking IDs. Better visualization with color.}
\label{fig:qualitative_waymo}
\end{figure*}
\minisection{Evaluation Video.}
We have uploaded a showcase video that demonstrates video inputs with estimated 3D bounding boxes in the camera view and tracked trajectories in bird's eye view on nuScenes, KITTI, and Waymo Open dataset. Please refer to the showcase video for more qualitative examples.
\section*{Acknowledgments}
\else
\section*{Acknowledgment}
\fi
Hu would like to thank the National Center for High-performance Computing for computer time and facilities, and Google PhD Fellowship Program for their support.
The authors gratefully acknowledge the support of Berkeley AI Research, Berkeley DeepDrive and Ministry of Science and Technology of Taiwan (MOST-110-2634-F-007-016), MOST Joint Research Center for AI Technology and All Vista Healthcare.
|
\section{Introduction}
Circular data are collected when the topic of interest is a direction or a time of day. These particular data appear in many
applications: earth sciences (e.g. wind directions), medicine (e.g. circadian rhythm), ecology (e.g. animal movements), forensics (crime incidence).
Different surveys on statistical methods for circular data can be found: \cite{Mardia},
\cite{JammaSengup}, \cite{LeyVerdebout} or more recently \cite{PewseyGarcia}. In the present work, we consider a mixture model with two components equal up to a rotation. We observe $X_1, \dots, X_n$ a sample of data on $\S^1$ with probability distribution function:
\begin{equation}\label{model}
g(x)=p_0 f(R_0^{-1} x)+(1-p_0) f(Q_0^{-1} x)=p_0 f( x-\alpha_0)+(1-p_0) f(x-\beta_0).
\end{equation}
In the right hand side we have identified $f:\S^1\to \mathbb{R}$ and its periodized version on $\mathbb{R}$.
Here $R_0$ and $Q_0$ are two unknown rotations of the circle.
$R_0$ is a rotation with angle $\alpha_0$ and $Q_0$ is a rotation with angle $\beta_0$. The aim is to estimate both $\theta_0=(p_0,\alpha_0,\beta_0)$ and the nonparametric part $f$.
Bimodal circular data are commonly encountered in many scientific fields, for instance in climatology, animal orientations or in earth sciences.
For the analysis of wind directions, see \cite{HernandezScarpa} and for animal orientations, the dragonflies data set presented in \cite{Batschelet}. %
In geosciences, one can cite the
cross-bed orientations data set obtained in the middle Mississipian Salem Limestone of central Indiana and which was presented by the Seminar Sedimentation (\cite{Sedimentation}). Last but not least, the paper of \cite{Lark} analyzes some geological data sets and clearly favours for some of them a two component mixture of von Mises distributions.
\medskip
Mixture models for describing multimodal circular data date back to \cite{Pearson} and have been largely used since then.
An important case in the literature is the mixture of two von Mises distributions which has been explored in numerous works. Let us cite among others papers by \cite{Bartels}, \cite{Spurr} or \cite{ChenLiFu}. From a practical point of view, algorithms have also been proposed to deal with mixture of two von Mises distributions, including maximum likelihood algorithms by \cite{JonesJames} or a characteristic function based procedure by \cite{SpurrKout}. Note that on the unit hypersphere, \cite{Banerjee} investigated clustering methods for mixtures of von Mises Fisher distributions. In our framework, we shall not assume any parametric form of the density and hence the model is said to be semiparametric. {To the best of our knowledge, this is the first work devoted to the study of the semiparametric mixture model for circular data}. {This semiparametric model is more complex and intricate than the usual parametric one encountered in the circular literature}. In the spherical case, \cite{KimKoo} studied the general mixture framework for a location parameter but assuming that the nonparametric part $f$ is known. On the real line, this semiparametric model has been studied by \cite{bordes2006}, \cite{Hunterwang07}, \cite{butucea2014} or \cite{GassiatRousseau} for dependent latent variables. For the multivariate case, see for instance \cite{HallZhou}, \cite{HallNeeman}, \cite{GassiatRousseauVernet}, \cite{Hohmann}. When dealing with the specific case of one of the two components being parametric, one refers to work by \cite{MaYao} and references therein.
Note that we can rewrite model (\ref{model}) as
\begin{equation}\label{modelprime}
X_i = Y_i+\varepsilon_i \: \pmod{2\pi},\qquad i=1,\dots,n,
\end{equation}
where $Y_i$ has density $f$ and $\varepsilon_i$ is a Bernoulli angle, which is equal to $\alpha_0$ with probability $p_0$ and $\beta_0$ otherwise. Accordingly, model (\ref{model}) can be viewed as a circular convolution model with unknown noise operator $\varepsilon$. The circular convolution model has been studied by \cite{Golden} in the case of known noise operator whereas \cite{JohannesSchwarz} dealt with unknown error distribution but have at their disposal an independent sample of the noise to estimate this latter. It is worth pointing out that
\cite{Golden} and \cite{JohannesSchwarz} made the usual assumptions on the decay of the Fourier coefficients of the density of $\varepsilon$, whereas in model \eqref{model} the Fourier coefficients are not decreasing.
\medskip
Identifiability questions are at the heart of the theory of mixture models and the circular context is no exception. Thus, our first task is to study the identifiability of the model. From a mathematical point of view, the topology of the circle makes the problem very different from the linear case. In the circular parametric case, \cite{Fraser} obtained identifiability results for the von Mises distributions, which were extended in \cite{Kent83} to generalized von Mises distributions while \cite{holzmann2004}) focused on wrapped distributions, basing their analysis on the Fourier coefficients. Here, the Fourier coefficients turn out to be very useful as well but the nonparametric paradigm makes the study quite different and intricate. Our identifiability results are obtained under mild assumptions on the Fourier coefficients. We require that the coefficients are real which can be related to the usual symmetry assumption in mixture models (see for instance \cite{Hunterwang07}) and we impose that only the first 4 coefficients do not vanish. Interestingly enough, some not intuitive phenomena appear. {A striking case} occurs when the angles $\alpha_0$ and $\beta_0$ are distant from $2\pi/3$, model \eqref{model} is then nonidentifiable which is quite surprising at first sight.
Once the identifiability of the model is obtained, we resort to a contrast function in the line of \cite{butucea2014} to estimate the Euclidian parameter $\theta_0$. In that regard, we prove the consistency of our estimator and an asymptotic normality result. Thereafter, for the estimation of the nonparametric part, a penalized empirical risk estimation method is used. The estimator of the density turns out to be adaptive (meaning that it does not require the specification of the unknown smoothness parameter), {a property which was not reached so far for this semiparametric model even in the linear case}. The procedure devised is hence relevant for practical purposes. We prove an oracle inequality and minimax rates are achieved by our estimator for Sobolev regularity classes. Eventually, a numerical section shows the good performances of the whole estimation procedure.
The paper is organized as follows. Section 2 is devoted to the identifiability of the model. Section 3 tackles the estimation of the parameter $\theta_0$ whereas Section 4 focuses on the estimation of the nonparametric part. Finally Section 5 presents numerical implementations of our procedure. Proofs are gathered in Section 6.
\section{Identifiability}
In this section, to keep the notation as light and clear as possible, we drop the subscript $0$ in the parameters. For any function $g$ and any angle $\alpha$, denote $g_\alpha(x):=g(x-\alpha)$. For any complex number $a$, $\overline a$ is the complex conjugate of $a$. For any integrable function $\phi: \S^1\to \mathbb{R}$, we denote for any $l\in \mathbb{Z}$,
$\phi^{\star l}=\int_{\S^1} \phi(x) e^{-i l x} \frac{dx}{2\pi}$, the Fourier coefficients. {Note also that we use notation $f$ and $f'$ for two densities, where $f'$ is not the derivative of $f$.}
\medskip
Let us now study the identifiability of our model (\ref{model})
where the data have density $ p f( x-\alpha)+(1-p) f( x-\beta)$. First, it is obvious that if $p=0$, $\alpha$ is not identifiable, and if $p=1$, $\beta$ is not identifiable. In the same way, $p$ is not identifiable if $\alpha=\beta$.
Moreover, as explained in \cite{Hunterwang07} for a translation mixture on the real line, the case $p=1/2$ has to be avoided. Indeed, denoting $g$ a
density and for instance
$f=\frac12 g_1+\frac12 g_{-1}$
and
$f'=\frac12 g_2+\frac12 g_{-2}$
we have
$$f_1+f_5=f'_{2}+f'_{4}.$$
In addition, it is well known that, in such a mixture model, $(p, \alpha, \beta)$ cannot be distinguished from
$(1-p,\beta,\alpha)$: it is the so-called \textit{label switching} problem.
So we will assume that $p\in (0,1/2)$ (for mixtures on $\mathbb{R}$ it is assumed alternatively that $\alpha<\beta$ but ordering angles is less relevant).
Now let us study the specific problems of identifiability on the circle, that do not appear on $\mathbb{R}$.
First, if $f$ is the uniform probability, the model is not identifiable, so we have to exclude this case. Another case to exclude is the case of $\delta$-periodic functions.
Indeed in this case
$f_{\alpha}=f_{\alpha+\delta}$.
These functions have the property that $f^{\star l}=0$ for all $l\notin (2\pi/\delta)\mathbb{Z}$. So we will require that the Fourier coefficients of $f$ do not cancel out too much. Here we will assume
$$\text{ for all } l\in \{1,2,3,4\}, \qquad f^{\star l}\neq 0,\text{ and} \quad f^{\star l}=\overline{f^{\star l}} .$$
This last assumption can be related to the symmetry of $f$. Indeed if $f$ is zero-symmetric then all its Fourier coefficients are real. Symmetry is a usual assumption in this mixture context, to distinguish between the translations of $f$:
for any $\delta\in \mathbb{R}$, $$p f( x-\alpha)+(1-p) f(x-\beta)=p f_\delta( x-\alpha+\delta)+(1-p) f_\delta(x-\beta+\delta)$$
More precisely, \cite{Hunterwang07} show that symmetry is a sufficient and necessary condition for identifiability of the model mixture on $\mathbb{R}$.
In the circle framework, it is natural to work with Fourier coefficients rather than Fourier transform as on $\mathbb{R}$. A lot of circular densities have their Fourier coefficients real, provided that their location parameter is $\mu=0$: for example the Jones-Pewsey density, which includes the cardioid, the wrapped Cauchy density, and the von Mises density.
Here we require the assumption only for the first 4 Fourier coefficients of $f$ (due to our proof), which is milder than symmetry.
Let us now state our identifiability result under these assumptions. Note that \cite{holzmann2004} have studied the identifiability of this model when $f$ belongs to a parametric scale-family of densities, but here we face a nonparametric problem concerning $f$.
\begin{theorem} \label{identifiability}
Assume that $\theta=(p, \alpha, \beta)$ and $\theta'=(p', \alpha', \beta')$ belong to
$$\left\{(p, \alpha, \beta)\in (0,1/2)\times \S^1\times \S^1, \quad \alpha\neq \beta\pmod{2\pi}\right\}$$
and that
$f,f'$ belongs
to
$$ \left\{f: \S^1\to \mathbb{R} \text{ density such that, for all } l\in \{1,2,3,4\}, \; f^{\star l}\in \mathbb{R}\backslash \{0\}\right\}.$$
Suppose
$pf_{\alpha}+(1-p)f_{\beta}=p'f'_{\alpha'}+(1-p')f'_{\beta'}$. Then
\begin{enumerate}
\item either $(p',\alpha', \beta')$=$(p,\alpha, \beta)$ and $f'=f$,
\item or $(p', \alpha', \beta')=(p, \alpha+\pi, \beta+\pi)$ and
$f'=f_{\pi}$
\item or if $\beta-\alpha=\pi \pmod{2\pi}$, then $f'$ is a linear combination of $f$ and $f_{\pi}$,
and
either
$(\alpha',\beta')=(\alpha, \beta)$,
or $(\alpha',\beta')=(\beta, \alpha)$.
\item or if $\beta-\alpha=\pm 2\pi/3 \pmod{2\pi}$, then $f'$ is a linear combination of
$f_{\pi/3}, f_{-\pi/3}, f_{\pi}$
and $p'=(1-2p)/(2-3p)$ and
(a) if $\beta-\alpha=2\pi/3 $,
$(\alpha',\beta')=(\alpha+\pi,\beta-\pi/3)$
{or $(\alpha',\beta')=(\alpha,\beta+2\pi/3)$},
(b) if $\beta-\alpha=-2\pi/3 $,
$(\alpha',\beta')=(\alpha+\pi,\beta+\pi/3)$
{or $(\alpha',\beta')=(\alpha,\beta-2\pi/3)$.}
\end{enumerate}
\end{theorem}
Case 2. arises from a specific feature of circular distributions: if $f$ is symmetric with respect to 0 then it is symmetric with respect to $\pi$. Unlike the real case, a symmetry assumption does not exclude the case $f'(x)=f(x-\pi)$. To bypass this we could assume for instance
$ f^{\star 1}> 0.$
Indeed %
for each $l\in \mathbb{Z}$, $(f_{\pi})^{\star l}=f^{\star l}(-1)^l$, so the Fourier coefficients of $f$ and $ f_\pi$ have opposite sign for any odd $l$.
With our assumption, we recover among $f$ and $ f_\pi$ the one with positive first Fourier coefficient, i.e. with positive mean resultant length.
Nevertheless our estimation procedure begins with the parametric part so that this assumption concerning only the nonparametric part will not allow us to distinguish $\alpha$ from $\alpha+\pi$ in this first parametric estimation step. That is why we rather choose to assume that $\alpha$ and $\beta$ belong to $[0,\pi) \pmod{\pi}$.
Case 3. concerns bipolar data since $\alpha$ and $\beta$ are diametrically opposed (separated by $\pi$ radians). In this case $\alpha$ and $\beta$ are identifiable, but $p$ and $f$ not.
{
Indeed, for any density $f$ and any $0<p'\leq p <1/2$, we can find $q\in (0,1]$ such that $f'=qf+(1-q)f_{\pi}$ verifies $pf_{\alpha}+(1-p)f_{\beta}=p'f'_{\alpha'}+(1-p')f'_{\beta'}$.
}
{
Thus our result demonstrates that bimodal data sets with opposite modes lead to non-identifiability issues, and this highlights a fundamental issue in considering a too large class of possible densities.
}
{Let us now discuss the case 4., which is the most curious
({we shall only comment the first case (a)}, the other is similar).
{Let us set
$$f'(x)=(1-p)f\left(x-\frac{\pi}3\right)+ (1-p) f\left(x+\frac{\pi}3\right)+ (2p-1) f(x-\pi).$$
This function is symmetric if $f$ is symmetric, verifies $\int_{\S^1} f'=1$ and may be positive for some values of $p$ (depending on $f$): see Figure~\ref{figf'}.}
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.6]{Rplotffprime.pdf}
\end{center}
\caption{Plot of a circular density $f$ (dashed blue), and of
$f'=(1-p)f_{\frac{\pi}3}+ (1-p) f_{-\frac{\pi}3}+ (2p-1) f_\pi$
(solid red).
Here $f$ is the von Mises density with mean 0 and concentration 1.
{In this case, $f'$ is positive as soon as $p\geq 0.36$, here $p=0.4$.}
}
\label{figf'}
\end{figure}
Then we can write $f'_{\pi/3}$:
$$f'\left(x-\frac\pi{3}\right)=(1-p)f\left(x-\frac{2\pi}3\right)+ (1-p) f\left(x\right)+ (2p-1) f\left(x-\frac{4\pi}3\right),$$
as well as $f'_{\pi}$:
$$f'(x-\pi)=(1-p)f\left(x-\frac{4\pi}3\right)+ (1-p) f\left(x-\frac{2\pi}3\right)+ (2p-1) f(x).$$
Hence a mixture of $f'_{\pi}$ and $f'_{\pi/3}$ gives a mixture of $f(x), f(x-\frac{2\pi}3), f(x-\frac{4\pi}3)$:
\begin{eqnarray*}p'f'(x-\pi)+(1-p')f'\left(x-\frac\pi{3}\right)&=&
[p'(2p-1)+(1-p')(1-p)]f(x)\\
&&+[p'(1-p)+(1-p')(1-p)] f\left(x-\frac{2\pi}3\right)\\
&&+[p'(1-p)+(1-p')(2p-1)]f\left(x-\frac{4\pi}3\right)
\end{eqnarray*}
If now $p'=(1-2p)/(2-3p)$, then $p'(1-p)+(1-p')(2p-1)=0$ and the third component $f(x-\frac{4\pi}3)$ vanishes. Thus
$$p' f'(x-\pi)+(1-p')f'\left(x-\frac{\pi}3\right) =p f( x )+(1-p)f\left(x-\frac{2\pi}3\right).$$}
In such a particular case, we cannot identify $\theta$ nor $f$. However this happens only when $\beta-\alpha=\pm 2\pi/3$. So,
to exclude these cases, we will now assume $\beta \neq \alpha \pmod{2\pi/3}$.
\bigskip
Finally, we shall assume
that $f\in \mathcal{F}$ with some assumptions for $\mathcal{F}$:
\begin{assumption}\label{hypf}
$$\mathcal F\subset \left\{f: \S^1\to \mathbb{R} \text{ density s.t. for all } l\in \{1,2,3,4\}, \quad f^{\star l}\in \mathbb{R}\backslash \{0\}\right\}$$
\end{assumption}
or
\begin{assumption}\label{hypf2}
$$\mathcal F\subset \left\{f: \S^1\to \mathbb{R} \text{ density s.t. for all } l\in \{1,2,3,4\}, \quad f^{\star l}\in \mathbb{R}\backslash \{0\},\quad f^{\star 1}>0\right\}$$
\end{assumption}
and we shall assume that $\theta\in \Theta$ with some assumptions for $\Theta$:
\begin{assumption}\label{hyptheta2}
$$\Theta \subset \left\{(p, \alpha, \beta)\in \left(0, \frac12\right)\times \S^1\times \S^1, \quad \alpha\neq \beta \pmod{\pi, 2\pi/3}\right\}$$
\end{assumption}
where $\alpha\neq \beta \pmod{2\pi/3,\pi}$ means
$\beta-\alpha\notin\{-\frac{2\pi}{3},0,\frac{2\pi}{3},\pi\}+ 2\pi\mathbb{Z}$, or
\begin{assumption}\label{hyptheta}
$$\Theta \subset \left\{(p, \alpha, \beta)\in \left(0, \frac12\right)\times [0,\pi)\times [0,\pi), \quad \alpha\neq \beta \pmod{2\pi/3}\right\}$$
\end{assumption}
Note that Assumption \ref{hyptheta} implies Assumption \ref{hyptheta2}, and Assumption \ref{hypf2} implies Assumption \ref{hypf}. We can write the following result.
\begin{corollary}
Under Assumptions \ref{hypf} and \ref{hyptheta}, or under Assumptions \ref{hypf2} and \ref{hyptheta2}, model \eqref{model} is identifiable.
Under Assumptions \ref{hypf} and \ref{hyptheta2}, model \eqref{model} is identifiable modulo $\pi$, that is to say that if
$pf_{\alpha}+(1-p)f_{\beta}=p'f'_{\alpha'}+(1-p')f'_{\beta'}$
then $p'=p$ and
either $(\alpha', \beta')$=$(\alpha, \beta)$ and $f'=f$,
or $(\alpha', \beta')=(\alpha+\pi, \beta+\pi)$ and
$f'=f_{\pi}$.
\end{corollary}
Moreover, the proof of Theorem~\ref{identifiability} provides the following statement.
\begin{lemma} \label{EqMtheta}
Under Assumption~\ref{hyptheta2}, denoting
$
M^l(\theta):= pe^{-i \alpha l }+ (1-p)e^{-i\beta l}
$, for all $\theta,\theta' \in \Theta$,
$$\forall 1 \leq l \leq 4, \; \Im\left(M^l(\theta')\overline{M^l(\theta)}\right) =0 \Leftrightarrow \theta'=\theta\text{ or } \theta'{=}\theta+\pi.$$
where $\theta'{=}\theta+\pi$ means $(p', \alpha', \beta')=(p, \alpha+\pi, \beta+\pi)$.
\end{lemma}
\section{Estimation for the parametric part}
Now, let us denote for all $l \in \mathbb{Z}$
\[
M^l(\theta):= pe^{-i \alpha l }+ (1-p)e^{-i\beta l}.
\]
In model (\ref{model}) the Fourier coefficients of $g$ satisfy for any $l$:
\begin{equation*}\label{coeff-g}
g^{\star l} = (p_0e^{-i \alpha_0 l} + (1-p_0) e^{-i \beta_0 l} )f^{\star l }.
\end{equation*}
Thus $g^{\star l} = M^l(\theta_0)f^{\star l } $ and the previous lemma gives that $\theta=\theta_0$ (or $\theta_0+\pi$) if and only if, for each $l\in \{1,\dots, 4\}$,
$$ \Im\left(M^l(\theta_0)\overline{M^l(\theta)}\right) =0 \Leftrightarrow \Im\left(g^{\star l} \overline{M^l(\theta)}\right) =0 $$
using that $f^{\star l }$ are non-zero real numbers.
This invites us to consider
\[
S(\theta):=
\sum_{l =-4}^{4} \left ( \Im \left ( g^{\star l}\overline{{M}^l(\theta)} \right ) \right )^2 =
\sum_{l =-4}^{4} \left ( \Im \left ( g^{\star l}\{pe^{i \alpha l} + (1-p) e^{i \beta l} \} \right ) \right )^2.
\]
Note that $g^{\star0}\overline{{M}^{0}(\theta)}=1/(2\pi)$ and that $ \Im \left ( g^{\star (-l)}\overline{{M}^{-l}(\theta)} \right ) =
\Im \left (\overline{ g^{\star l}}{{M}^l(\theta)} \right ) =-\Im \left ( g^{\star l}\overline{{M}^l(\theta)} \right ) $ so that
we can also write \[
S(\theta)=
2\sum_{l =1}^{4} \left ( \Im \left ( g^{\star l}\overline{{M}^l(\theta)} \right ) \right )^2 .
\]
The empirical counterpart of $S(\theta)$ is
\begin{eqnarray*}
\tilde S_n(\theta) &=& \sum_{l=-4}^4 \left ( \Im \left ( \widehat {g^{\star l}}\overline{ M^l(\theta)} \right ) \right )^2\\
&=& \sum_{l=-4}^4 \left (\Im \left ( \frac {1}{2 \pi n} \sum_{k=1}^n e^{-i l X_k}\overline{ M^l(\theta)} \right ) \right)^2 \\
&=&\frac{1}{4 \pi^2n^2} \sum_{l=-4}^4\sum_{1\leq k,j\leq n } \Im \left ( e^{i lX_k}{ M^l(\theta)}\right ) \Im \left ( e^{i lX_j}{ M^l(\theta)} \right ) .
\end{eqnarray*}
Next, we consider a slightly modified version of $ \tilde S_n(\theta)$ by removing the diagonal terms
\begin{equation}\label{Sn}
S_n(\theta) = \frac{1}{ 4 \pi^2n(n-1) } \sum_{l=-4}^{4} \sum_{1\leq k \neq j\leq n} \Im \left ( e^{i lX_k}{ M^l(\theta)}\right ) \Im \left ( e^{i lX_j}{ M^l(\theta)} \right ) .
\end{equation}
Let us denote
\[
Z_{ k}^l(\theta):= \Im \left ( \frac{e^{i lX_k}}{2 \pi} M^l(\theta) \right )\quad
\text{ and }\quad
J^l(\theta) := \Im \left (\overline{g^{\star l}}{ M^l(\theta)}\right ).
\]
Hence
\[
S_n(\theta)= \frac{1}{n(n-1)} \sum_{l=-4}^{4} \sum_{1\leq k \neq j\leq n} Z_{k}^l(\theta) { Z_{j}^l(\theta) }.
\]
Note that we have
$\mathbb{E}( Z_{k}^l(\theta) )= J^l(\theta),$
and $S_n(\theta)$ is an unbiased estimator of $S(\theta)$.
Let the estimator of $\theta_0$ be
\begin{equation}\label{estimator_theta}
\hat \theta_n = \argmin_{\theta \in \Theta} S_n(\theta).
\end{equation}
For this estimator we can prove the following consistency result.
\begin{theorem}\label{consistance}
Consider $\Theta$ a compact set included in
$$\left\{(p, \alpha, \beta)\in \left(0, 1/2\right)\times \S^1\times \S^1, \quad \alpha\neq \beta \pmod{2\pi/3,\pi}\right\}$$
and the estimator $\hat \theta_n = \argmin_{\theta \in \Theta} S_n(\theta). $ We have
$\hat \theta_n \rightarrow \theta_0 \pmod{\pi}$ in probability.
\end{theorem}
The last convergence means that for all $\epsilon>0$, the probability $\P(\|\hat \theta_n-\theta_0\|\leq \epsilon \text{ or } \|\hat \theta_n-\theta_0-\pi\|\leq \epsilon)$ tends to 1 when $n$ goes to $+\infty$, where $\|.\|$ denotes the Euclidean norm.
\begin{proof}
$\Theta$ is a compact set and $S$ is continuous. Lemma~\ref{lip} ensures that $S_n$ is Lipschitz hence uniformly continuous, and
Proposition~\ref{variance} ensures that for all $\theta$,
$| S_n(\theta)-S(\theta)|$ tends to 0 in probability.
Then
it is sufficient to apply a classical Lemma to conclude. See the details in Section~\ref{proof:consistence}
\end{proof}
From now on, we assume that $\Theta$ is a compact set included in
$\left(0, \frac12\right)\times [0,\pi)\times [0,\pi)$, as in Assumption \ref{hyptheta}.
Then, $\theta_0+\pi$ is excluded and under Assumption \ref{hyptheta}, $\hat \theta_n \rightarrow \theta_0 $ in probability.
Moreover this estimator is asymptotically normal. We denote $\dot \phi(\theta)$ the gradient of any function $\phi$ with respect to $\theta=(p,\alpha,\beta)$, $\ddot \phi(\theta)$ the Hessian matrix and for any matrix $A$, we denote $A^\top$ its transpose.
\begin{theorem}\label{normasymp}
Consider $\Theta$ a compact set included in $$\left\{(p, \alpha, \beta)\in \left(0, 1/2\right)\times [0,\pi)\times [0,\pi), \quad \alpha\neq \beta \pmod{2\pi/3}\right\}$$
and the estimator $\hat \theta_n = \argmin_{\theta \in \Theta} S_n(\theta) .$ Assume that $\theta_0\in \Theta$.
{ Let $\mathcal{A}$ be the Hessian matrix of $S$ in $\theta_0$: $\mathcal{A}=\ddot S(\theta_0)= 2\sum_{l=-4}^4 \dot J^l(\theta_0) \dot J^l(\theta_0)^\top$.
Then, if $\mathcal{A}$ is invertible,
$$
\sqrt{n}(\hat \theta_n -\theta_0) \overset{d}{\longrightarrow} \mathcal{N}(0, \Sigma),
$$
where $\Sigma= \mathcal{A}^{-1}V\mathcal{A}^{-1} $,
$V=4\mathbb{E}(UU^\top)$ and $U= \sum_{l=-4}^4 \dot J^l(\theta_0) Z^l_{1}(\theta_0) $. }
\end{theorem}
The proof can be found in Section~\ref{proof:normasymp}.
Note that $\mathcal{A}$ can be estimated by $\ddot S(\hat \theta_n)$ and $V$ by
$$\frac{4}{n^3}\sum_{1\leq k, j,j'\leq n}\sum_{-4\leq l,l'\leq 4} Z_k^{l}(\hat\theta_n)Z_k^{l'}(\hat\theta_n)\dot Z_j^{l}(\hat\theta_n)(\dot Z_{j'}^{l'}(\hat\theta_n))^\top$$
(see details in Section~\ref{sec:estimV}).
Thus we can estimate the covariance matrix $\Sigma$ and deduce an asymptotic confidence region.
{
We also prove the following result on the quadratic risk of the estimator $\hat\theta_n $, which is useful for the sequel (see Section~\ref{sec:proof1varthetachap} for a proof).
\begin{proposition}\label{varthetachap}
Under the assumptions of Theorem~\ref{normasymp}, there exists a numerical constant $K$ such that, for all $\theta_0\in \Theta$ and for all $n\geq 1$
$$\mathbb{E}\|\hat\theta_n-\theta_0\|^2\leq Kn^{-1},$$
where the norm is the Euclidean norm in $\mathbb{R}^3$.
\end{proposition}
}
\section{Nonparametric part}
Let us now estimate the nonparametric part. We shall use the following norm: for any function $\phi$, we denote $\|\phi\|_2=\left(\frac{1}{2\pi}\int_{\S^1} \phi^2(x)dx\right)^{1/2}.$
Recall that for all $l\in \mathbb{Z}$, $g^{\star l }=M^l(\theta_0)f^{\star l}$
where $g$ is the density of the observations $X_k$ and $g^{\star l }$ its Fourier coefficient.
Then $f^{\star l}=g^{\star l }/M^l(\theta_0)$. We can verify that $M^{l}(\theta_0)\neq 0$. Indeed, for any $\theta\in \Theta$,
$$|M^l(\theta)|^2=p^2+(1-p)^2+2p(1-p)\cos[l(\beta-\alpha)]\geq (1-2p)^2>0.$$
Nevertheless this division by $M^l(\theta_0)$ requires us to impose a new assumption. We assume that there exists $P\in (0,1/2)$ such that $0<p<P$ for any $p$, i. e.
\begin{assumption}\label{hypthetabis} $\Theta$ is a compact set included in
$$\left\{(p, \alpha, \beta)\in \left(0, P\right)\times [0,\pi)\times [0,\pi), \quad \alpha\neq \beta \pmod{2\pi/3}\right\}.$$
\end{assumption}
Under this assumption, $|M^l(\theta)|$ is always bounded from below by $1-2P$.
Now, to estimate $g^{\star l }=\int_{\S^1}e^{-il x}g(x)dx/(2
\pi)$, it is natural to define
$$\widehat{g^{\star l}}=\frac1{2\pi n}\sum_{k=1}^n e^{-ilX_k}.$$
If $\hat \theta=\hat \theta_n$ is the previous estimator of the parametric part, we set the plugin estimator of the Fourier coefficient:
$$\widehat{f^{\star l}}=\frac1{2\pi n}\sum_{k=1}^n M^{l}(\hat \theta)^{-1}e^{-ilX_k}.$$
Finally, for
{$L$ an integer}, set
$$\hat f_L(x)=\sum_{l=-L}^L\widehat{f^{\star l}}e^{ilx}. $$
To measure the performance of this estimator, we use Parseval equality to write
$$\|f-\hat f_L\|_2^2=\sum_{|l|>L} |f^{\star l}|^2+\sum_{l=-L}^L |f^{\star l}-\widehat{f^{\star l}}|^2$$
which is the classical bias variance decomposition.
Moreover it is possible to prove that the variance term satisfy $\sum_{l=-L}^L \mathbb{E}|f^{\star l}-\widehat{f^{\star l}}|^2=O(\frac{2L+1}{n})$ (see Lemma~\ref{majnun} below).
To control the bias term we recall %
{the definition of the
Sobolev ellipsoid:
$$W(s, R)=\{ f: \S^1\to \mathbb{R},\quad \sum_{l\in \mathbb{Z}} (1+l^2)^{s} |f^{\star l}|^2 \leq R^2\}.$$ }
For such a smooth $f$,
the risk of estimator $\hat f_L$ is then bounded in the following way:
$$\mathbb{E}\|f-\hat f_L\|_2^2\leq R^2\left(1+L^2\right)^{-s}+ C\frac{2L+1}{n}. $$
It is clear that an optimal value for $L$ is of order $n^{1/(2s+1)}$ but this value is unknown. We rather choose a data-driven method to select $L$.
We introduce a classical minimization of a penalized empirical risk. Set
\begin{equation}\label{resolution_level}
\widehat{L}=\underset{L\in \mathcal{L}}{\argmin} \left\{-\sum_{l=-L}^L | \widehat{f^{\star l}}|^2+\lambda \frac{2L+1}{n}\right\}
\end{equation}
where $\mathcal{L}$ is a finite set of resolution level, and $\lambda$ a constant to be specified later.
The next theorem states an oracle inequality which highlights the bias variance decomposition of the quadratic risk and justifies our estimation procedure.
\begin{theorem}\label{oracle}
\label{oracleineq}
Assume Assumption \ref{hypf} and Assumption \ref{hypthetabis} .
Assume that $f$ belongs to the Sobolev ellipsoid $W(s, R)$ with
{ $s \geq 1$. Let $\hat L$ defined in \eqref{resolution_level} with
$\mathcal{L}=\{{0},1, \dots, \lfloor cn^{\frac1{2s_0+1}} \rfloor \}$ for some $s_0>1$ and some positive constant $c$.}
Let $\epsilon>0$.
If the penalty constant verifies
$\lambda > {(3/\pi^2)}(1+\epsilon^{-1})(1-2P)^{-2}$
then,
$$\mathbb{E}\|\hat f_{\widehat{L}}-f\|_2^2\leq (1+2\epsilon)\mathbb{E}\min_{L\in \mathcal{L}} \left\{\|\hat f_{ L}-f\|_2^2+2\lambda \frac{2L+1}{n}\right\}+\frac{C {(1+R^2)}}{n}$$
where $C$ is a positive constant depending on
{
$P,s_0,c,\epsilon,\lambda$.
Moreover, if $s\geq s_0$,
$$\sup_{f\in W(s, R)}\sup_{\theta_0\in \Theta}\mathbb{E}_{\theta_0,f}\|\hat f_{\widehat{L}}-f\|_2^2=O\left(R^2
n ^{-2s/(2s+1)}\right).$$
}
\end{theorem}
In consequence our estimator has a quadratic risk in $n ^{-2s/(2s+1)}$.
{
Regarding the lower bound note that for any estimator $\tilde f_n$
$$\sup_{f\in W(s, R)}\sup_{\theta_0\in \Theta}\mathbb{E}_{\theta_0,f}\| \tilde f_n-f\|_2^2\geq
\sup_{f\in W(s, R)}\mathbb{E}_{\theta,f}\|\tilde f_n-f\|_2^2$$
for some arbitrary $\theta\in \Theta$,
so that the problem is reduced to a purely nonparametric lower bound.
In the case of direct observations this quantity is lower bounded by $\underline{C}n ^{-2s/(2s+1)}$, see Theorem 11 and its proof in \cite{baldikerk2009} (case $d=1$ for the circle $\S^1$).
We can use this proof to prove the lower bound in our mixture case. Indeed, for any densities $f_1$ and $f_2$, if
$g_i(x)=pf_i(x-\alpha)+(1-p)f_i(x-\beta)$ is the associated density of our observations, then the Kullback-Leibler divergence verifies
$$K(g_1dx,g_2dx)\leq \int \frac{(g_1-g_2)^2}{g_2}\leq 2 \int \frac{(f_1-f_2)^2}{f_2}$$
and the rest of the proof is identical. Thus
$$\sup_{f\in W(s, R)}\sup_{\theta_0\in \Theta}\mathbb{E}_{\theta_0,f}\|\hat f_{\widehat{L}}-f\|_2^2\geq \underline{C}
n ^{-2s/(2s+1)}$$
and our estimator is optimal minimax.
}
\begin{remark}
Note that the penalty only depends on $P$ which is some safety margin around 1/2, that can be chosen by the statistician. For the practical choice of the penalty, see Section \ref{section:simus}.
\end{remark}
Eventually, note that some densities may be \textit{supersmooth}, in the following sense:
$$\sum_{l\in \mathbb{Z}} \exp(2b|l|^r)|f^{\star l}|^2 \leq R^2. $$
In this case, the quadratic bias is bounded by $R^2\exp(-2bL^r)$ which gives the following fast rate of convergence:
$$\mathbb{E}\|\hat f_{\widehat{L}}-f\|_2^2=O\left(\frac{(\log n)^{1/r}}{n}\right).$$
\section{Numerical results}
\label{section:simus}
All computations are performed with Matlab software and the Optimization Toolbox.
We shall implement our statistical procedure to both estimate the parameter $\theta_0$ and the density $f$. We consider three popular circular densities, namely the von Mises density, the wrapped Cauchy and the wrapped normal densities. We remind their expression (see \cite{LeyVerdebout}).
The von Mises density is given by:
$$
f_{VM}(x)=\frac{1}{2\pi I_{0}(\kappa) } e^{\kappa \cos(x-\mu)},
$$
with $\kappa \geq 0$, $I_0(\kappa)$ the modified Bessel function of the first kind and of order $0$.
The wrapped Cauchy distribution has density:
$$
f_{WC}(x)= \frac{1}{2\pi} \frac{1-\gamma^2}{1+ \gamma^2 -2 \gamma\cos(x-\mu)},
$$
with $0 \leq \gamma \leq 1.$
The wrapped normal density expression is:
$$
f_{WN}(x)=\frac{1}{\sigma\sqrt{2\pi}}\sum_{k=-\infty}^\infty e^{- \frac{(x-\mu+2k\pi)^2}{2\sigma^2}},
$$
$\sigma>0$. For more clarity, we set $\sigma^2=:-2 \log (\rho).$ Hence, we have $0\leq \rho\leq 1$.
All these densities are characterized by a concentration parameter $\kappa$, $\gamma$ or $\rho$ and a location parameter $\mu$. Remind that values $\kappa=0$, $\gamma=0$ and $\rho=0$ correspond to the uniform density on the circle. To meet symmetry assumptions of Theorem \ref{identifiability}, we consider in the sequel that the location parameter is set to $\mu=0$.
\medskip
First, let us focus on the parametric part. We set $\theta_0= (p_0, \alpha_0, \beta_0)=(\frac 1 4, \frac \pi 8, \frac{2\pi}{3})$. Obtaining the estimate $\hat \theta_n$ of $\theta_0$ (see (\ref{estimator_theta})) requires to solve a nonlinear minimization problem. To this end, we resort to the function \textit{fmincon} of the Matlab Optimization toolbox. The function \textit{fmincon} finds a constrained minimum of a function of several variables. Two parameters are to be specified: the domain over which the minimum is searched and an initial value. We consider the domain $\{ (0,\frac 1 2) \times [0, \pi) \times [0, \pi) \}$. For more stability and to avoid possible local minimums, we perform the procedure over $10$ initials values uniformly drawn on $\{ (0,\frac 1 2) \times [0, \pi) \times [0, \pi) \}$. The final estimator $\hat \theta_n$ corresponds to the minimum value of the empirical contrast $S_n(\theta)$ given in (\ref{Sn}) over the $10$ trials.
Table \ref{MISE-theta} gathers mean squared errors for our estimation procedure. When analyzing Table \ref{MISE-theta}, one clearly sees that increasing the number of observations improves noticeably the performances. As expected, von Mises densities with smaller concentration parameter are more difficult to estimate. Nonetheless, the overall performances are satisfying. Table \ref{MISE-theta-Spurr} displays the performances of the method-of-moments estimation procedure developed by \cite{SpurrKout} to handle the problem of estimating the parameters in mixtures of von Mises distributions.
To fairly compare the two methods, Table \ref{MISE-theta-Spurr-2} gives the \cite{SpurrKout} performances but this time when estimating on the same domain than ours e.g $\{ (0,\frac 1 2) \times [0, \pi) \times [0, \pi) \}$. At closer inspection, the \cite{SpurrKout} method seems to behave better to estimate angles $\alpha_0$ and $\beta_0$ while our method may appear more competitive for estimating $p_0$. It is worth noticing that the method by \cite{SpurrKout} is completely parametric and takes advantage of the knowledge of the distributions. In this regard, our procedure which is semiparametric is competitive with a parametric method.
\medskip
Figure \ref{histogrammes} illustrates the asymptotic normality of our estimator $\hat \theta_n$ stated in Theorem \ref{normasymp}.
\bigskip
\begin{table}[!h] \begin{center} \begin{tabular}{c |c c c| c c c}
density & \multicolumn{3}{c}{$n=100$ } & \multicolumn{3}{c}{$n=1000$}\\
& $p$ & $\alpha$ & $\beta$ & $p$ & $\alpha$ & $\beta$ \\
\hline
$f_{VM}$, $\kappa=2$ & 0.0121 & 0.6848 & 0.1131 & 0.0017 & 0.1919 & 0.0238 \\
$f_{VM}$, $\kappa=5$ & 0.0030& 0.0285 & 0.0049 & 1.4632e-04 & 0.0017 & 4.4861e-04 \\
$f_{VM}$, $\kappa=7$ & 0.0033 & 0.0133 & 0.0031 & 1.6721e-04 & 0.0013 & 3.0102e-04 \\
$f_{WC}$, $\rho=0.8$ & 0.0029 & 0.0124 &0.0024 & 2.0788e-04 & 8.5435e-04 & 1.8942e-04 \\
$f_{WN}$, $\rho=0.8$ &0.0077 & 0.1679 & 0.0457& 0.0020 & 0.0238& 0.0037
\end{tabular}
\caption{Mean squarred errors for estimating parameter $\theta_0$ over $50$ Monte Carlo replications.}
\label{MISE-theta}
\end{center}\end{table}
\begin{table}[!h] \begin{center} \begin{tabular}{c |c c c| c c c}
density & \multicolumn{3}{c}{$n=100$ } & \multicolumn{3}{c}{$n=1000$}\\
& $p$ & $\alpha$ & $\beta$ & $p$ & $\alpha$ & $\beta$ \\
\hline
$f_{VM}$, $\kappa=2$ & 0.0938 & 0.4212 & 0.1171 & 0.0116 & 0.0685 & 0.0062 \\
$f_{VM}$, $\kappa=5$ & 0.0031 & 0.0360 & 0.0049 & 2.9965e-04 & 0.0025 & 6.6273e-04 \\
$f_{VM}$, $\kappa=7$ & 0.0031 &0.0084 & 0.0029 & 2.4553e-04 & 0.0014 &3.5541e-04 \\
\end{tabular}
\caption{Spurr and Koutbeiy procedure: mean squared errors for estimating parameter $\theta_0$ over $50$ Monte Carlo replications on $\{ (0,1) \times [0, 2\pi) \times [0, 2\pi) \}$}
\label{MISE-theta-Spurr}
\end{center}\end{table}
\begin{table}[!h] \begin{center} \begin{tabular}{c |c c c| c c c}
density & \multicolumn{3}{c}{$n=100$ } & \multicolumn{3}{c}{$n=1000$}\\
& $p$ & $\alpha$ & $\beta$ & $p$ & $\alpha$ & $\beta$ \\
\hline
$f_{VM}$, $\kappa=2$ & 0.0231 & 0.2117 & 0.0351 & 0.0112 &0.0635 & 0.0081 \\
$f_{VM}$, $\kappa=5$ & 0.0032 & 0.0409 & 0.0042 & 4.1489e-04 & 0.0022 & 6.3122e-04 \\
$f_{VM}$, $\kappa=7$ & 0.0026 & 0.0094 & 0.0029 & 2.3197e-04 & 0.0010 & 2.8350e-04 \\
\end{tabular}
\caption{Spurr and Koutbeiy procedure: mean squared errors for estimating parameter $\theta_0$ over $50$ Monte Carlo replications on $\{ (0,\frac 1 2) \times [0, \pi) \times [0, \pi) \}$}
\label{MISE-theta-Spurr-2}
\end{center}\end{table}
\begin{figure}[!h]\begin{center}
\begin{tabular}{cc}
$\alpha$ & $\beta$ \\
\includegraphics[scale=0.23]{histo_alpha_1000obs_100MC.pdf} & \includegraphics[scale=0.23]{histo_beta_1000obs_100MC.pdf}
\end{tabular}
$p$\\
\includegraphics[scale=0.23]{histo_p_1000obs_100MC.pdf}
\caption{Histograms of the centered and standardized statistics $\hat \theta_n$ for the von Mises density $f_{VM}$ with $\kappa=5$, $n=1000$ observations and $100$ Monte Carlo replications}
\label{histogrammes}
\end{center}\end{figure}
Now, let us turn to the nonparametric estimation part namely the estimation of the density $f$. The estimator of $f$ is given by $\hat f_{\widehat{L}}$ (see Theorem \ref{oracleineq}). It requires the computation of a data-driven resolution level choice $\widehat{L}$ (given in (\ref{resolution_level})) which implies a tuning parameter $\lambda$. To select the proper $\lambda$, we follow the data-driven slope estimation approach due to Birg\'e and Massart (see \cite{BirgeMassart2001} and \cite{BirgeMassart2007}). An overview in practice is presented in \cite{Baudry}. To implement the slope heuristics method, one has to plot for ${L=0}$ to $L_{\max}$ the couples of points $(\frac{2L+1}{n} , \sum_{l=-L}^L | \widehat{f^{\star l}}|^2)$. For $L \geq L_0$, one should observe a linear behaviour (see Figure \ref{slope}). Then, once the slope is estimated, say $a$, by a linear regression method, one eventually takes $\hat \lambda = 2a$ and the final resolution level is:
$$\widehat{L}=\underset{L\in \mathcal{L}}{\argmin} \left\{-\sum_{l=-L}^L | \widehat{f^{\star l}}|^2+\hat \lambda \frac{2L+1}{n}\right\}.$$
\begin{figure}[!h]\begin{center}
\begin{tabular}{c}
\includegraphics[scale=0.35]{slope.pdf}
\end{tabular}
\caption{For the wrapped Cauchy density $f_{WC}$ with $\gamma=0.8$ and $n=1000$: plot of couples $(\frac{2L+1}{n}, \sum_{l=-L}^L | \widehat{f^{\star l}}|^2)$ for $L=\{1,\dots, 50 \} $.}
\label{slope}
\end{center}\end{figure}
\begin{figure}[!h]\begin{center}
\begin{tabular}{cc}
$f$ & $g$ \\
\includegraphics[scale=0.23]{graphfFisher51000-2.pdf}&\includegraphics[scale=0.23]{graphgFisher51000-2.pdf} \\
\includegraphics[scale=0.23]{graphfCauchy081000-2.pdf}&\includegraphics[scale=0.23]{graphgCauchy081000-2.pdf} \\
\includegraphics[scale=0.23]{graphfNormale081000-2.pdf}&\includegraphics[scale=0.23]{graphgNormale081000-2.pdf}
\end{tabular}
\caption{Estimation of the density $f$ and the mixture density $g$ for $n=1000$. In red, the density, in dotted lines its estimate. From top to bottom: the von Mises density with $\kappa=5$, the wrapped Cauchy with $\gamma=0.8$ and the wrapped normal density with $\rho=0.8$. }
\label{graphiques}
\end{center}\end{figure}
Finally, Figure \ref{graphiques} shows reconstructions of the density $f$ and the mixture density $g$ as well. The estimates are good.
\begin{remark}
Note that for the two exceptional cases, when $p_0=0$ or $f$ is the uniform density, our procedure performs well. Indeed, if $p_0=0$, our method yields that $\alpha=\beta$ and retrieves that there is only one component in the mixture. When $f$ is the uniform density,
our algorithm selects $\hat L=0$ which yields the uniform distribution.
\end{remark}
\section{Proofs}
\subsection{Proof of Theorem~\ref{identifiability} (identifiability)}
Denote
\[
M^l(\theta):= pe^{-i \alpha l }+ (1-p)e^{-i\beta l}.
\]
Suppose
$p f( x-\alpha)+(1-p) f(x-\beta)=p' f'( x-\alpha')+(1-p') f'(x-\beta')$.
The calculation of the Fourier coefficients gives, for all $l\in \mathbb{Z}$,
$ f^{\star l}M^l(\theta)=(f')^{\star l} M^l(\theta')$
which implies $$ f^{\star l}|M^l(\theta)|^2=(f')^{\star l} M^l(\theta')\overline{M^l(\theta)}.$$ Then, our assumptions on $f$ and $f'$ entail
$$M^l(\theta')\overline{M^l(\theta)}\text{ is real}\quad \forall l\in \{1,2,3,4\}.$$
Let us now study the consequence of this fact.
Denote
$$\gamma_1=\alpha'-\beta,\; \gamma_2=\alpha'-\alpha,\;\gamma_3=\beta'-\beta,\;\gamma_4=\beta'-\alpha$$
the 4 angles.
Denote also the associated weights in $(0,1)$:
$$\lambda_1=p'(1-p),\; \lambda_2=p'p, \;\lambda_3=(1-p')(1-p),\; \lambda_4=(1-p')p.$$
With this notation
$
M^l(\theta')\overline{M^l(\theta)}=
\lambda_1e^{-i\gamma_1l}+\lambda_2e^{-i \gamma_2 l}+\lambda_3e^{-i \gamma_3 l} +\lambda_4e^{-i\gamma_4l}.$$
Then
$M^l(\theta')\overline{M^l(\theta)}$ is real if and only if $\sum_{k=1}^4 \lambda_k \sin (l\gamma_k)=0$
and we have to solve the equations
\begin{equation}
\label{equasin}
\forall l=1,2,3,4,\qquad \sum_{k=1}^4 \lambda_k \sin (l\gamma_k)=0.
\end{equation}
This system of equations is studied in Lemmas~\ref{detA} and \ref{CasAnnul}
below.
Let us now reason with the representatives of the $\gamma_k$ in $(-\pi,\pi]$. Lemma~\ref{CasAnnul} says that the possible values for the $\gamma_k$'s are $0, \pi, \gamma, -\gamma$, for some $\gamma \in (0,\pi)$.
Note that here
\begin{equation}
\label{eqdif}\gamma_1-\gamma_2=\gamma_3-\gamma_4=\alpha-\beta\neq 0\;\;
\text{ and }\;\;
\gamma_1-\gamma_3=\gamma_2-\gamma_4=\alpha'-\beta'\neq 0
\end{equation}
and then
the $\gamma_k$'s take at least 2 different values: either 4 different values; or $\gamma_2=\gamma_3$ and the other distinct; or $\gamma_1=\gamma_4$ and the other distinct; or $\gamma_2=\gamma_3$ and $\gamma_1=\gamma_4$.
$\bullet$ Let us first study the case where all the $\gamma_k$'s are distinct. There are 4!=24 ways of having
$(\gamma_{i_1}, \gamma_{i_2}, \gamma_{i_3}, \gamma_{i_4})=(-\gamma, 0, \gamma,\pi)$. But 16 combinations lead to $p=1/2$ or $p'=1/2$.
For example, if
$(\gamma_{1}, \gamma_{2}, \gamma_{3}, \gamma_{4})=(-\gamma, 0, \gamma, \pi)$ then \eqref{equasin} becomes
$$\lambda_1 \sin (-l\gamma)+\lambda_2 \sin (0)+\lambda_3 \sin (l\gamma)+\lambda_4 \sin (l\pi)=0.$$
Thus $\lambda_1=\lambda_3$, which gives $p'=1/2$. In the same way, there are 4 possibilities giving $\lambda_1=\lambda_3$, 4 possibilities giving $\lambda_1=\lambda_2$, 4 possibilities giving $\lambda_2=\lambda_4$, 4 possibilities giving $\lambda_3=\lambda_4$. All of this is impossible, since $p,p'\in (0,1)\backslash\{ 1/2\}$.
In addition, in the 4 cases where
$\gamma_1=-\gamma_4$, we obtain via \eqref{eqdif} $\gamma_3=-\gamma_2$ which is impossible if $\{\gamma_2,\gamma_3\}=\{0,\pi\}$.
Idem if $\gamma_2=-\gamma_3$ and $\{\gamma_1,\gamma_4\}=\{0,\pi\}$. Thus it is finally impossible that all the $\gamma_k$'s are distinct.
$\bullet$ Let us now study the case where the $\gamma_k$'s take 3 distinct values ($\gamma_2=\gamma_3$ or $\gamma_1=\gamma_4$) and belong to $\{0, \pi, \gamma\}$ or $\{0, \pi, -\gamma\}$. In the case where $\gamma_2=\gamma_3$, coming back to equation $\eqref{equasin}$, we understand that all the rearrangements lead to $\lambda_4=0$ or $\lambda_1=0$ or $\lambda_2+\lambda_3=0$, which is impossible. In the same way, if $\gamma_1=\gamma_4$, equation \eqref{equasin} leads to
$\lambda_2=0$ or $\lambda_3=0$ or $\lambda_1+\lambda_4=0$, which is impossible.
$\bullet$ The next case is when the $\gamma_k$'s take 3 distinct values and belong to $\{0, \gamma, -\gamma\}$ or $\{\pi, \gamma, -\gamma\}$. If $\gamma_2=\gamma_3$, we can then list the 6 cases:
$$\begin{array}{ccc|c}
\gamma_1 & \gamma_2=\gamma_3 & \gamma_4 & \text{consequence}\\
\hline
-\gamma & 0/\pi & \gamma & p=p',
{ \alpha'-\alpha=\beta'-\beta =0\pmod{\pi}} \\
\gamma & 0/\pi &- \gamma & p=p',{ \alpha'-\alpha=\beta'-\beta =0\pmod{\pi}} \\
-\gamma & \gamma & 0/\pi &\lambda_1=\lambda_2+\lambda_3
\\
\gamma & -\gamma & 0/\pi &\lambda_1=\lambda_2+\lambda_3
\\
0/\pi & \gamma & -\gamma& \lambda_4=\lambda_2+\lambda_3
\\
0/\pi & -\gamma & \gamma &\lambda_4=\lambda_2+\lambda_3
\\
\end{array}
$$
Note that $\lambda_1=\lambda_2+\lambda_3\Leftrightarrow p'(2-3p)=1-p$,
which is possible only if $p<1/2$ and $p'>1/2$ (recall that we suppose $p<1/2$ and $p'<1/2$). In the same way $\lambda_4=\lambda_2+\lambda_3\Leftrightarrow p'(1-3p)=1-2p$, which is possible only if $p>1/2$ and $p'<1/2$.\\
Finally, if $\gamma_1=\gamma_4$, we have the 6 last cases:
$$\begin{array}{ccc|c}
\gamma_2 & \gamma_1=\gamma_4 & \gamma_3 & \text{consequence}\\
\hline
-\gamma & 0/\pi & \gamma & p'=1-p\\%,\alpha'=\beta,\beta'= \alpha \pmod{\pi} \\
\gamma & 0/\pi &- \gamma & p'=1-p,\\%\alpha'=\beta,\beta'= \alpha \pmod{\pi} \\
-\gamma & \gamma & 0/\pi & p'=\frac{p}{3p-1}
\\
\gamma & -\gamma & 0/\pi & p'=\frac{p}{3p-1}
\\
0/\pi & \gamma & -\gamma&
{p'=\frac{1-2p}{2-3p}, \beta- \alpha=\pm 2\pi/3}\\
0/\pi & -\gamma & \gamma&
{p'=\frac{1-2p}{2-3p}, \beta- \alpha=\pm 2\pi/3}
\end{array}
$$
Note that the 4 first lines of this table are impossible since $p,p'\in(0,1/2)$ and $p'=p/(3p-1)\notin (0,1)$ if $0<p<1/2$. Let us detail the lines 5 and 6. In these cases, $\lambda_1+\lambda_4-\lambda_3=0$ which provides $p'=(1-2p)/(3-2p)$.
Moreover \eqref{eqdif} implies that $3\gamma_1=\gamma_2=0\pmod{\pi}$ and $2\gamma_1=\beta-\alpha=\alpha'-\beta'$.
According to the values of $\gamma_1$ and $\gamma_2$, there are 4 possibilities
$\diamond$ $\beta-\alpha= 2\pi/3$ and $(\alpha',\beta')=(\alpha, \beta +2\pi/3)$,
$\diamond$ $\beta-\alpha= 2\pi/3$ and $(\alpha',\beta')=(\alpha+\pi, \beta -\pi/3)$
$\diamond$ $\beta-\alpha= -2\pi/3$ and $(\alpha',\beta')=(\alpha, \beta -2\pi/3)$
$\diamond$ $\beta-\alpha=- 2\pi/3$ and $(\alpha',\beta')=(\alpha+\pi, \beta +\pi/3)$
\color{black}
$\bullet$ The last case occurs when the $\gamma_k$'s take 2 distinct values.
If the $\gamma_k$'s take exactly 2 different values, using \eqref{eqdif}, necessarily
$$\gamma_1=\gamma_4 \text{ and }\gamma_2=\gamma_3 \pmod{2\pi}
\Rightarrow 0=\gamma_1-\gamma_4 +\gamma_3-\gamma_2=2(\alpha-\beta) \pmod{2\pi}$$
which is possible only if $\alpha-\beta= \pi\pmod{2\pi}$ (recall that $\alpha-\beta$ is always assumed $\neq 0$). And in the same way $\alpha'-\beta'= \pi\pmod{2\pi}$.
Then $\gamma_1-\gamma_2=\alpha-\beta=\pi\pmod{2\pi}$. Thus the two different values of the $\gamma_k$'s are at a distant of $\pi$.
The first possibility is that these two values are $0$ and $\pi$, which corresponds to the first case of Lemma~\ref{CasAnnul}. There are two subcases:
1a. $(\gamma_1,\gamma_2,\gamma_3,\gamma_4)=(\pi,0,0,\pi)$ or 1b.
$(\gamma_1,\gamma_2,\gamma_3,\gamma_4)=(0,\pi,\pi,0)$.
In the subcase 1a. $(\alpha',\beta')=(\alpha, \beta)$.
Equations
$$\begin{cases}
pf+(1-p)f_{\pi}=p'f'+(1-p')f'_{\pi}\\
pf_{\pi}+(1-p)f=p'f'_{\pi}+(1-p')f'
\end{cases} $$
entails that $f'$ is a linear combination of $f$ and $f_{\pi}$.
In the subcase $1b$. $(\alpha',\beta')=(\alpha+\pi, \beta+\pi)=(\beta,\alpha)$.
The second possibility is that the two distinct values $\gamma_1=\gamma_4$ and $\gamma_2=\gamma_3$ are not multiples of $\pi$, which
corresponds to the fourth case of Lemma~\ref{CasAnnul}. Then
$(\gamma_1,\gamma_2,\gamma_3,\gamma_4)=(\gamma_1,-\gamma_1,-\gamma_1,\gamma_1)$ and $$\gamma_1-(-\gamma_1)=\gamma_1-\gamma_2=\pi \pmod{2\pi}$$ which entails $\gamma_1=\pi/2 \pmod{\pi}$.
Equation \eqref{equasin} becomes
$$(\lambda_1-\lambda_2-\lambda_3+\lambda_4)\sin (l\pi/2)=0$$
so that $\lambda_1+\lambda_4=\lambda_2+\lambda_3$, which gives
$$p'(1-p)+p(1-p')=p'p+(1-p')(1-p)\Rightarrow p'+p -2pp'=1/2\Rightarrow p'=1/2$$
which is impossible.
$\bullet$ Let us recap the only possible cases that we have obtained:
$\triangleright$ $p=p', \alpha'-\alpha=\beta'-\beta =0\pmod{\pi} $,
$\triangleright$ $p'=\frac{1-2p}{2-3p}, \beta- \alpha =\pm 2\pi/3$,
with the four possibilities described above,
$\triangleright$ $\beta-\alpha= \pi$, $(\alpha',\beta')=(\alpha, \beta)$ or $(\alpha',\beta')=( \beta, \alpha)$.
This completes the proof of the theorem.
\begin{lemma}\label{detA}
Let $\gamma_1, \dots, \gamma_4$ be four reals. Let $A$ be the matrix $(\sin(i \gamma_j))_{1\leq i,j\leq 4}$. Then
$$\det A=64\prod_{k=1}^4\sin (\gamma_k) \prod_{1\leq i<j \leq 4} (\cos(\gamma_i)-\cos(\gamma_j)).
$$
\end{lemma}
\begin{proof}
From matrix $A$, doing line modification $L_3 \leftarrow L_3-L_1$, and $L_4 \leftarrow L_4-L_2$, we obtain (recall that
$\sin(2p)=2\sin(p)\cos(p)$ and $\sin(p)-\sin(q)=2 \sin(\frac{p-q}{2})\cos(\frac{p+q}{2})$)
$$\det A= \begin{vmatrix}
\sin(\gamma_1) & \sin(\gamma_2) & \sin(\gamma_3) & \sin(\gamma_4) \\
2\sin(\gamma_1)\cos(\gamma_1) & 2\sin(\gamma_2)\cos(\gamma_2)& 2\sin(\gamma_3)\cos(\gamma_3)&2\sin(\gamma_4)\cos(\gamma_4) \\
2\sin(\gamma_1)\cos(2\gamma_1) & 2\sin(\gamma_2)\cos(2\gamma_2)& 2\sin(\gamma_3)\cos(2\gamma_3)&2\sin(\gamma_4)\cos(2\gamma_4) \\
2\sin(\gamma_1)\cos(3\gamma_1) & 2\sin(\gamma_2)\cos(3\gamma_2)& 2\sin(\gamma_3)\cos(3\gamma_3)&2\sin(\gamma_4)\cos(3\gamma_4)
\end{vmatrix}. $$
Using $4$-linearity of the determinant:
$$\det A=8\left(\prod_{j=1}^4\sin (\gamma_j)\right)
\begin{vmatrix}
1& 1& 1 &1\\
\cos(\gamma_1) & \cos(\gamma_2)& \cos(\gamma_3)&\cos(\gamma_4) \\
\cos(2\gamma_1) & \cos(2\gamma_2)& \cos(2\gamma_3)&\cos(2\gamma_4) \\
\cos(3\gamma_1) & \cos(3\gamma_2)& \cos(3\gamma_3)&\cos(3\gamma_4)
\end{vmatrix} .$$
Now, denote $x_k=\cos(\gamma_k)$ and remark that $\cos(i\gamma_k)=T_i(\cos \gamma_k)=T_i(x_k)$ where $T_i$ is the $i$th Chebyshev polynomial:
$T_0=1, T_1=X, T_2= 2X^2-1, T_3=4X^3-3X$.
We have $T_2+T_0=2X^2$ and $T_3+3T_1=4X^3$.
Then, doing $L_3 \leftarrow L_3+L_1$, and $L_4 \leftarrow L_4+3L_2$:
$$\det A=8\left(\prod_{j=1}^4\sin (\gamma_j)\right)
\begin{vmatrix}
1& 1& 1 &1\\
x_1 & x_2 & x_3 & x_4\\
2x_1^2 &2 x_2^2 & 2x_3^2 & 2x_4^2\\
4x_1^3 &4 x_2^3 & 4x_3^3 & 4x_4^3
\end{vmatrix}
=64\left(\prod_{j=1}^4\sin (\gamma_j)\right)
\begin{vmatrix}
1& 1& 1 &1\\
x_1 & x_2 & x_3 & x_4\\
x_1^2 & x_2^2 & x_3^2 & x_4^2\\
x_1^3 & x_2^3 & x_3^3 & x_4^3
\end{vmatrix}$$
This is a Vandermonde matrix, hence
$$\det A
=64\left(\prod_{j=1}^4\sin (\gamma_j)\right)
\prod_{1\leq i<j\leq 4} (x_i-x_j)=64\prod_{k=1}^4\sin (\gamma_k) \prod_{1\leq i<j \leq 4} (\cos(\gamma_i)-\cos(\gamma_j)).
$$
\end{proof}
\begin{lemma}\label{CasAnnul}
Let $\gamma_1, \dots, \gamma_4$ be four reals.
Let $\lambda_1, \dots, \lambda_4 \in \mathbb{R}\backslash\{0\}$ such that
\begin{equation}
\label{eqsin}
\sum_{k=1}^4 \lambda_k \sin (l\gamma_k)=0,\qquad l=1,\dots,4.
\end{equation}
Then, one of the following cases holds:
\begin{enumerate}
\item All $\gamma_k$ are multiples of $\pi$.
\item Exactly two $\gamma_k$ are multiples of $\pi$: $\gamma_{i_1}=\gamma_{i_2}=0 \pmod{\pi}$
and $\gamma_{i_3}=\pm\gamma_{i_4}\pmod{2\pi}$.
\item Only one $\gamma_k$ is multiple of $\pi$: $\gamma_{i_1}=0 \pmod{\pi}$ and
$\gamma_{i_2}=\pm \gamma_{i_3}=\pm\gamma_{i_4} \pmod{2\pi}$.
\item No $\gamma_k$ is multiple of $\pi$ and
$\gamma_1=\pm\gamma_{2}=\pm \gamma_{3}=\pm\gamma_{4} \pmod{2\pi}$.
\end{enumerate}
\end{lemma}
\begin{proof} First observe that,
since $\sum_{k=1}^4 \lambda_k \sin (l\gamma_k)=0$ with $\lambda\neq 0_{\mathbb{R}^4}$, necessarily
det($A$)=0 where $A=(\sin(i \gamma_j))_{1\leq i,j\leq 4}$. Using Lemma~\ref{detA}
\begin{equation}\label{eqdetA}
\prod_{k=1}^4\sin (\gamma_k) \prod_{1\leq i<j \leq 4} (\cos(\gamma_i)-\cos(\gamma_j))=0.
\end{equation}
Now, let us study the various cases that make this quantity vanish. \\
For the first case, note that if three $\gamma_k$ are multiples of $\pi$: $\gamma_{i_1}=\gamma_{i_2}=\gamma_{i_3}=0 \pmod{\pi}$ then
equation \eqref{eqsin} becomes
$ \lambda_{i_4}\sin (l\gamma_{i_4})=0$ and the last angle is also null modulo $\pi$.\\
In case 2., equation \eqref{eqsin} entails
$$ \lambda_{i_3}\sin (l\gamma_{i_3})+ \lambda_{i_4}\sin (l\gamma_{i_4})=0,\qquad l=1,2$$
with $\gamma_{i_3}\neq 0 \pmod{\pi}, \gamma_{i_4}\neq 0 \pmod{\pi}.$
Then, since $( \lambda_{i_3}, \lambda_{i_4})\neq(0,0)$,
$$0= \begin{vmatrix}
\sin(\gamma_{i_3}) & \sin(\gamma_{i_4}) \\
\sin(2\gamma_{i_3})\ & \sin(2\gamma_{i_4})
\end{vmatrix}
=2\sin(\gamma_{i_3})\sin(\gamma_{i_4})(\cos(\gamma_{i_4})-\cos(\gamma_{i_3})).$$
Then $\cos(\gamma_{i_3})=\cos(\gamma_{i_4})$.
Either $\gamma_{i_3}=\gamma_{i_4} \pmod{2\pi}$,
or $\gamma_{i_3}=-\gamma_{i_4} \pmod{2\pi}$.\\
Let us now study case 3. For the sake of simplicity we assume that
$\gamma_{4}=0 \pmod{\pi}$ and $\gamma_{k}\neq 0 \pmod{\pi}$ for $k=1,2,3$.
Equation \eqref{eqsin} gives
$$ \lambda_{1}\sin (l\gamma_{1})+ \lambda_{2}\sin (l\gamma_{2})+\lambda_3\sin(l\gamma_3)=0,\qquad l=1,2,3.$$
With the same proof as Lemma~\ref{detA}, we obtain
$$\prod_{k=1}^3\sin (\gamma_k) \prod_{1\leq i<j \leq 3} (\cos(\gamma_i)-\cos(\gamma_j))=0.$$
Then $\gamma_1=\pm \gamma_2 \pmod{2\pi}$ or $\gamma_1=\pm \gamma_3 \pmod{2\pi}$ or $\gamma_2=\pm \gamma_3 \pmod{2\pi}$. Moreover, if, for example, $\gamma_1=\pm \gamma_2 \pmod{2\pi}$ then
$$ (\lambda_{1}\pm \lambda_2)\sin (l\gamma_{1})+\lambda_3\sin(l\gamma_3)=0,\qquad l=1,2$$
We are reduced to the previous case, then $\gamma_1=\pm\gamma_3 \pmod{2\pi}$.\\
In the case 4., equation \eqref{eqdetA} becomes $ \prod_{1\leq i<j \leq 4} (\cos(\gamma_i)-\cos(\gamma_j))=0$, which provides 6 possible equalities.
Assume, for example, $\cos(\gamma_1)-\cos(\gamma_2)=0$ and consequently $\gamma_1=\pm \gamma_2 \pmod{2\pi}$. Then
$$( \lambda_{1}\pm \lambda_2)\sin (l\gamma_{1})+\lambda_3\sin(l\gamma_3)+ \lambda_{4}\sin (l\gamma_{4})=0,\qquad l=1,2,3.$$
Reasoning as in previous case, $\gamma_{1}=\pm \gamma_{3}=\pm\gamma_{4} \pmod{2\pi}$.
\end{proof}
\subsection{Proof of Theorem~\ref{consistance} (consistency)}
\label{proof:consistence}
This proof and the following are inspired from \cite{butucea2014}.
Let us denote $\tilde \Theta =(0,1/2) \times \S^1\times \S^1$. Denote $\dot \phi(\theta)$ the gradient of any function $\phi$ with respect to $\theta=(p,\alpha,\beta)$, and $\ddot \phi(\theta)$ the Hessian matrix.
The proof of Theorem~\ref{consistance} relies on some preliminary results, given in the sequel.
\begin{proposition}\label{contrast} Under Assumption~\ref{hyptheta2}
the contrast function $S$ verifies the following properties:
$S(\theta) \geq 0$, and $S(\theta)=0$ if and only if $\theta=\theta_0$ or $\theta=\theta_0+\pi$.
\end{proposition}
\begin{proof}
It is clear that $S(\theta)\geq 0$ and that
$$S(\theta_0)=\sum_{l =-4}^{4} \left ( \Im \left ( g^{\star l } \overline{{M}^l(\theta_0)} \right ) \right )^2=\sum_{l =-4}^{4} \left ( \Im \left ( f^{\star l } |M^l(\theta_0)|^2\right ) \right )^2=0.$$
By Lemma~\ref{EqMtheta}, if $\theta \neq \theta_0 \pmod{\pi}$,
there exists $l_1\in \{1,\dots, 4\}$ such that
$ \Im\left(M^{l_1}(\theta_0)\overline{M^{l_1}(\theta)}\right) \neq 0$ %
so that
$
S(\theta)\geq \left ( \Im \left ({ g^{\star l_1}}\overline{{M}^{l_1}(\theta)} \right ) \right )^2 >0.
$
\end{proof}
\begin{lemma}\label{bornes}
\begin{enumerate}
\item For all $\theta$ in $\tilde\Theta$, $|M^l(\theta)| \leq 1$.
\item For all $1\leq k\leq n$, for all $l$ in $\mathbb{Z}$, \[
\sup_{\theta\in \tilde\Theta} | Z^l_{k}(\theta) |\leq \frac{1}{2\pi}, \qquad \sup_{\theta\in \tilde\Theta} | J^l(\theta)| \leq \frac{1}{2\pi}.
\]
\item For all $1\leq k\leq n$, for all $l$ in $\mathbb{Z}$,
{
\[
\sup_{\theta\in \tilde\Theta} \| \dot Z^l_{k}(\theta) \|\leq \frac{ 2 +| l| }{ \sqrt{2}\pi}, \qquad \sup_{\theta\in \tilde\Theta} \| \dot J^l(\theta)\| \leq \frac{ 2 +| l| }{ \sqrt{2}\pi} .
\]
}
where $\|.\|$ is the Euclidean norm.
\item For all $1\leq k\leq n$, for all $l$ in $\mathbb{Z}$, \[
\sup_{\theta\in \tilde\Theta} \| \ddot Z^l_{k}(\theta) \|_F\leq \frac{|l| + l^2}{\pi} , \qquad \sup_{\theta\in \tilde\Theta} \| \ddot J^l(\theta)\|_F \leq \frac{|l| + l^2}{\pi}.
\]
where $\|.\|_F$ is the Frobenius norm.
\end{enumerate}
\end{lemma}
\begin{proof}
Point 1 is straightforward.
2. Let us start with $ Z^l_{ k}(\theta)$. We recall that
$
Z_{ k}^l(\theta)= \Im \left ( \frac{ e^{i lX_k}}{2 \pi} M^l(\theta) \right ). $
Then
$$
|Z_{k}^l(\theta)| \leq \frac{1}{2 \pi} |M^l(\theta)| \leq \frac{ 1}{2 \pi}.
$$
Furthermore
\[
| J^l(\theta) | \leq {|g^{\star l}| }{|M^l(\theta)|} \leq \frac{ 1}{2\pi} \int_{\S_1} g \leq \frac{1 }{2 \pi}.
\]
3. We have
\[
\dot Z_{k}^l(\theta) =
\frac{1}{2\pi}\Im \left( e^{i lX_k} \dot M^{l }(\theta)\right)= \frac{1}{2\pi}\Im \left ( e^{i lX_k}
\begin{pmatrix} e^{-il\alpha} - e^{-il \beta}\\
-il pe^{-i \alpha l}\\
- i l(1-p) e^{-i \beta l}\\
\end{pmatrix}\right)
\]
and
\[
\dot J^l(\theta) = \Im \left ( \overline{g^{\star l}} \dot M^{l }(\theta) \right)= \Im \left ( \overline{g^{\star l}}\begin{pmatrix} e^{-il\alpha} - e^{-il \beta}\\
-il pe^{-i \alpha l}\\
- i l(1-p) e^{-i \beta l}\\
\end{pmatrix}\right).
\]
We get
{
$$\|\dot Z_{k}^l(\theta) \| \leq \frac{1}{2 \pi} \left(2^2 + p^2l^2 + (1-p)^2 l^2\right)^{1/2} \leq \frac{ 2 +| l| }{ \sqrt{2}\pi}$$
}
and we have the same bound for $\|\dot J^l(\theta) \| $.
4. We have
\begin{eqnarray*}
\ddot Z_{ k}^l(\theta) &=& \Im \left ( \frac{ e^{i lX_k}}{ 2\pi } \ddot M^{l }_{}(\theta) \right ) \\
&=& \Im \left ( \frac{ e^{i lX_k}}{ 2\pi} \begin{pmatrix} 0 & -ile^{-il\alpha} & il e^{-il\beta}\\
- il e^{-il\alpha } & -l^2 pe^{-il \alpha} & 0\\
ile^{-il \beta} & 0 & -l^2(1-p) e^{-i \beta l}
\end{pmatrix} \right ).
\end{eqnarray*}
Thus
$$
\| \ddot Z^l_{k}(\theta) \|_F \leq \frac{1}{2\pi}\left(4l^2+l^4p^2+l^4(1-p)^2\right)^{1/2}\leq \frac{|l| + l^2}{\pi}.
$$
We bound $\| \ddot J^l (\theta) \|_F$ in the same way. This ends the proof of the lemma.
\end{proof}
\begin{lemma}\label{bornes2} There exists a numerical positive constant $C$ such that the following inequalities hold.
1. For all $1\leq k\leq n$, for all $l$ in $\mathbb{Z}$
\[ \forall \theta,\theta'\in \tilde\Theta \qquad
\| \dot Z^l_{k}(\theta) -\dot Z^l_{k}(\theta') \| \leq C \| \theta- \theta'\| (1+|l|+l^2).
\]
2. We also have
\[
\| \ddot Z^l_{k}(\theta) -\ddot Z^l_{k}(\theta') \|_F \leq C \| \theta - \theta' \| (1 +| l| +l^2+| l|^3).
\]
\end{lemma}
\begin{proof}
We use Taylor expansions at first order and then apply same bounding techniques as in Lemma \ref{bornes}.
\end{proof}
\begin{lemma} \label{lip}
\begin{enumerate}
\item The function $S$ is Lipschitz continuous over $\tilde \Theta$.
\item The function $S_n(\theta)$ is Lipschitz continuous over $\tilde \Theta$.
\item {The function $\ddot S_n(\theta)$ is Lipschitz continuous over $\tilde \Theta$ with respect to Frobenius norm},
{with Lipschitz constant not depending on $n$}.
\end{enumerate}
\end{lemma}
\begin{proof}
We will write $C$ for a numerical constant that may change from line to line but is numerical.
Let us start with point 1. We recall that $S(\theta)= \sum_l J^{l}(\theta)^2$. Let $\theta$ and $\theta'$ in $\tilde \Theta$. As $\tilde \Theta$ is a convex set, we get, thanks to the mean value theorem
\begin{eqnarray*}
|S(\theta)-S(\theta')| &= &\left| \sum_{l=-4}^4 J^{l}(\theta)^2-J^{l}(\theta')^2\right|
=\left |2(\theta -\theta')^\top \sum_{l=-4}^4 J^l(\theta_u)\dot J^{l}(\theta_u) \right|\\
&\leq & C \| \theta -\theta' \|\sum_{l=-4}^4 (1+ |l|)
\leq C\| \theta -\theta' \|
\end{eqnarray*}
with $\theta_u$ lying on the line connecting $\theta$ to $\theta'$, and using Lemma \ref{bornes}.
Let us shift to point 2. Due to the mean value theorem, we have
\begin{eqnarray*}
|S_n(\theta) - S_n(\theta') |&=& \left| \frac{1}{n(n-1)} \sum_{k \neq j} \sum_{l=-4}^4 \left ( Z^l_{k}(\theta) Z^l_{j}(\theta) - Z^l_{k}(\theta') Z^l_{j}(\theta') \right ) \right |\\
&=& \left| \frac{1}{n(n-1)} \sum_{k \neq j} \sum_{l=-4}^4 \left ( (\theta- \theta')^\top \nabla [ Z^l_{k}(\theta) Z^l_{j}(\theta)] |_{\theta= \theta_u} \right ) \right |\\
&=&\left| \frac{2(\theta -\theta')^\top}{n(n-1)} \sum_{k \neq j} \sum_{l=-4}^4 \dot Z^{l}_{k}(\theta_u)Z^{l}_{j}(\theta_u) \right |,
\end{eqnarray*}
with $\theta_u$ lying on the line connecting $\theta$ to $\theta'$. Then using 1. and 2. of Lemma \ref{bornes} we get
\begin{eqnarray*}
| S_n(\theta) - S_n(\theta') | &\leq& \frac{C \| \theta- \theta' \|}{n(n-1)} \sum_{k \neq j} \sum_{l=-4}^4 (1+|l |) \leq C \| \theta- \theta' \|
\end{eqnarray*}
which ends the proof of the second point.
Concerning point 3. we have that
$$
\ddot S_n(\theta)= \frac{2}{n(n-1)} \sum_{k \neq j} \sum_{l=4}^4 (\ddot Z^l_{k}(\theta) Z^l_{j}(\theta)+ \dot Z^l_{k}(\theta) \dot Z^l_{j}(\theta)^\top).
$$
Hence
\begin{eqnarray*}
\| \ddot S_n(\theta) - \ddot S_n(\theta') \|_F
&\leq &\frac{2}{n(n-1)} \sum_{k \neq j} \sum_{l=-4}^4 \bigg ( \| (\ddot Z^l_{k}(\theta) - \ddot Z^l_{k}(\theta')) Z^l_{j}(\theta) \|_F \\
&&+ \| \ddot Z^l_{k}(\theta' )( Z^l_{j}(\theta)- Z^l_{j}(\theta') ) \|_F
+ \| \dot Z^l_{k}(\theta' )( \dot Z^l_{j}(\theta)- \dot Z^l_{j}(\theta')^\top ) \|_F \\
& & +
\| ( \dot Z^l_{k}(\theta')- \dot Z^l_{k}(\theta) ) \dot Z^l_{j}(\theta)^\top \|_F \bigg )
\end{eqnarray*}
Using Taylor expansions and Lemma \ref{bornes} and \ref{bornes2}, we get that
$$
\| \ddot S_n(\theta) - \ddot S_n(\theta') \|_F \leq C\| \theta -\theta' \| \sum_{l=-4}^4 (1+ |l|+ l^2 +|l|^3) .
$$
\end{proof}
\begin{proposition}\label{variance} There exist a positive constant $C$
such that
\[
\sup_{\theta\in \tilde \Theta} \mathbb{E} [(S_n(\theta)- S(\theta))^2] \leq \frac{C}{n}.
\]
\end{proposition}
\begin{proof} The definitions of $S_n$ and $S$ provide
\[
S_n(\theta)- S(\theta) =\frac{1}{n(n-1)} \sum_{l=-4}^4\sum_{k \neq j} \left( Z^l_{k}(\theta) { Z^l_{ j}(\theta) } - J^{l}(\theta)^2 \right)
=T_n + V_n
\]
where
\[
T_n= \frac{2}{n(n-1)} \sum_{l=-4}^4 \sum_{k < j}( Z^l_{k}(\theta) - J^l(\theta)) ( {Z^l_{ j}(\theta) }- {J^l(\theta)})
\]
and
\[V_n =
\frac{2}{n} \sum_{l=-4}^4 \sum_{k=1 }^n (Z^l_{k}(\theta)- {J^l(\theta)}) J^l(\theta).
\]
Note that $\mathbb{E}(Z^l_{k}(\theta) - J^l(\theta))=0$ which entails $\mathbb{E}[T_nV_n]=0$.
Then
\[
\mathbb{E}\left[\left(S_n(\theta)- S(\theta) \right)^2\right]=\mathbb{E}\left[\left(T_n+V_n\right)^2\right]=\mathbb{E}\left[T_n^2\right]+\mathbb{E}\left[V_n^2\right].
\]
Now,
since the variables $ \left(\sum_{l=-4}^4 ( Z_{k}^l(\theta) - J^l(\theta)) ( {Z_{j}^l(\theta) } - J^l(\theta))\right)_{k<j}$ are uncorrelated,
\begin{eqnarray*}
\mathbb{E}[T_n^2] &=& \frac{2}{n(n-1)} \mathbb{E} \left [ \left ( \sum_{l=-4}^4 ( Z_{1}^l(\theta) - J^l(\theta)) ( {Z_{2}^l(\theta) } - { J^l(\theta))} \right ) ^2 \right ] \\
& \leq & \frac{2}{n(n-1)} \mathbb{E} \left [ \left ( \sum_{l=-4}^4 \frac{2}{2\pi} \cdot \frac{2}{2\pi} \right )^2 \right ]
\leq \frac{ C}{2n}
\end{eqnarray*}
using Lemma \ref{bornes}.
We focus now on $V_n$: in the same way
\begin{eqnarray*}
\mathbb{E}[V_n^2] &=& \frac 4 n \mathbb{E} \left [ \left ( \sum_{l=-4}^4 (Z_{1}^l(\theta)-J^l(\theta)) {J^l(\theta) } \right )^2 \right ] \\
& \leq & \frac 4 n \mathbb{E} \left [ \left ( \sum_{l=-4}^4 \frac{2}{2\pi} \cdot \frac{1}{2\pi} \right )^2 \right ] \leq \frac{ C}{2n} ,
\end{eqnarray*}
using Lemma \ref{bornes} again.
\color{black}
\end{proof}
{ Theorem~\ref{consistance} is finally proved using the following lemma, its assumptions
being ensured by Proposition~\ref{contrast}, Lemma~\ref{lip} and Proposition~\ref{variance}.
}
\begin{lemma}\label{lem:convmin}
Assume that $\Theta$ is a compact set and let $S: \Theta \to \mathbb{R}$ be a continuous function. Assume that
$$S(\theta)=\min_{\Theta} S \Leftrightarrow \theta=\theta_0 \text{ or }\theta=\theta_0'$$ where $\theta_0,\theta_0'\in \Theta$. Let $S_n:\Theta \to \mathbb{R}$ be a function which is uniformly continuous and such that for all $\theta$
$| S_n(\theta)-S(\theta)|$ tends to 0 in probability.
Let $\hat\theta_n$ be a point such that $S_n(\hat\theta_n)=\inf_{\Theta} S_n$. Then $\hat \theta_n \rightarrow \theta_0$ or $\theta_0'$ in probability.
\end{lemma}
This is a classical result in the theory of minimum contrast estimators, when $\theta_0=\theta'_0$ (see \cite{vandervaart} or \cite{DCD}).
We reproduce the proof since it is slightly adapted to the case of two argmins.
\begin{proof} Let $\epsilon>0$ and $B$ be the union of the open ball with center $\theta_0$ and radius $\epsilon$
and the open ball with center $\theta_0'$ and radius $\epsilon$.
Since $S$ is continuous and $B^c\subset\Theta$ is a compact set, there exists $\theta_\epsilon \in B^c$ such that
$S(\theta_\epsilon)=\inf_{B^c} S$.
Using the assumption, since $\theta_\epsilon\neq \theta_0$, and $\theta_\epsilon\neq \theta_0'$
$$\delta:=S(\theta_\epsilon)-S(\theta_0)>0.$$
Since $S_n$ is uniformly continuous, there exists $\alpha>0$ such that
$$\forall \theta, \theta' \qquad \|\theta-\theta'\|< \alpha \Rightarrow
|S_n(\theta)-S_n(\theta')|\leq \delta/2.$$
Moreover $B^c$ is a compact set then there exists a finite set $(\theta_i)$ such that $B^c\subset \cup_{i=1}^I B(\theta_i,\alpha)$.
Denote $\Delta_n:=\max_{0\leq i \leq I}| S_n(\theta_i)-S(\theta_i)|$. The assumption ensures that $\Delta_n$ tends to 0 in probability.
Let $\theta\in B^c$. There exists $1\leq i\leq I$ such that $\|\theta-\theta_i\|<\alpha$, and then $|S_n(\theta)-S_n(\theta_i)|\leq \delta/2$.
Thus
\begin{align*}
S_n(\theta)-S_n(\theta_0)&=(S_n(\theta)-S_n(\theta_i))+(S_n(\theta_i)-S(\theta_i))\\
&+(S(\theta_i)-S(\theta_0))+(S(\theta_0)-S_n(\theta_0))\\
& \geq -\delta/2-\Delta_n+\delta-\Delta_n
\end{align*}
using that $S(\theta_i)-S(\theta_0)\geq S(\theta_\epsilon)-S(\theta_0)=\delta$.
Then \begin{align*}
\inf_{\theta\in B^c} S_n(\theta)-S_n(\theta_0) \geq \delta/2-2\Delta_n.
\end{align*}
Now, if $\|\hat \theta_n-\theta_0\|\geq \epsilon$ and $\|\hat \theta_n-\theta_0'\|\geq \epsilon$
then $\hat\theta_n \in B^c$ and
$$\inf_{\theta \in \Theta} S_n(\theta)=S_n(\hat\theta_n)=\inf_{\theta \in B^c} S_n(\theta).$$
In particular $\inf_{\theta\in B^c} S_n(\theta)\leq S_n(\theta_0)$ so that
\begin{eqnarray*}
\P(\|\hat \theta_n-\theta_0\|\geq \epsilon \text{ and }\|\hat \theta_n-\theta_0'\|\geq \epsilon)&\leq &\P(0\geq \inf_{\theta \in B^c} S_n(\theta)- S_n(\theta_0) \geq \delta/2- 2\Delta_n)\\
&\leq &
\P(\Delta_n\geq \delta/4) \longrightarrow 0
\end{eqnarray*}
since $\Delta_n$ tends to 0 in probability.
\end{proof}
\color{black}
\subsection{Proof of Theorem~\ref{normasymp} (asymptotic normality)}
\label{proof:normasymp}
The Taylor's theorem and the definition of $\hat\theta_n$ give
$$
\dot S_n(\hat \theta_n) = \dot{S_n}(\theta_0) + \ddot{S_n}(\theta_n^*)(\hat \theta_n -\theta_0)=0,
$$
where $\theta_n^*$ lies in the line segment with extremities $\theta_0$ and $\hat \theta_n$. Equivalently we have,
$$
\ddot{S_n}(\theta_n^*)(\hat \theta_n -\theta_0) = - \dot{S_n}(\theta_0).
$$
We recall that
$$
S_n(\theta_0)= \frac{1}{n(n-1)} \sum_{k \neq j} \sum_{l=-4}^4 Z^l_{k}(\theta_0) Z^l_{j}(\theta_0)
$$
and
$$
\dot S_n(\theta_0)= \frac{2}{n(n-1)} \sum_{k \neq j} \sum_{l=-4}^4 \dot Z^l_{k}(\theta_0) Z^l_{j}(\theta_0)
$$
and
$$
\ddot S_n(\theta_0)= \frac{2}{n(n-1)} \sum_{k \neq j} \sum_{l=-4}^4\ddot Z^l_{k}(\theta_0) Z^l_{j}(\theta_0)+ \dot Z^l_{k}(\theta_0) \dot Z^l_{j}(\theta_0)^\top.
$$
\textbf{Step 1}- Let us prove that
$$
\sqrt{n} \dot S_n(\theta_0) \overset{d}{\longrightarrow} \mathcal{N}(0,V).
$$
We remind by Lemma~\ref{EqMtheta} that $
J^l(\theta_0)=0.
$
Hence
$$
\mathbb{E}(\dot S_n(\theta_0))= 2 \sum_{l=-4}^4 \dot J^l(\theta_0) J^l(\theta_0) =0.
$$
We can break down $\dot S_n(\theta_0)$ in the following way:
\begin{eqnarray*}
\dot S_n(\theta_0) &=& \frac{2}{n(n-1)} \sum_{k \neq j} \sum_{l=-4}^4 (\dot Z^l_{k}(\theta_0) - \dot J^l(\theta_0) + \dot J^l(\theta_0)) Z^l_{j}(\theta_0) \\
&=& \frac{4}{n(n-1)} \sum_{k< j} \sum_{l=-4}^4( \dot Z^l_{k}(\theta_0) - \dot J^l(\theta_0)) Z^l_{j}(\theta_0)
+ \frac{2}{n} \sum_{j=1}^n \sum_{l=-4}^4 \dot J^l(\theta_0) Z^l_{j}(\theta_0) \\
&=:& A_n + B_n.
\end{eqnarray*}
Note that $A_n$ and $B_n$ are centered variables. Let us show that $\sqrt{n}A_n= o_P(1)$.
Note that the variables $W_{jk}:= \left(\sum_{l=-4}^4 ( \dot Z_{k}^l(\theta_0) - \dot J^l(\theta_0)) Z_{j}^l(\theta_0) \right)_{k<j}$ are centered and uncorrelated. Then
$$\mathbb{E}(\|A_n\|^2)=\mathbb{E}\left(\left\|\frac{4}{n(n-1)} \sum_{k< j}W_{jk}\right\|^2\right)
=\frac{8}{n(n-1)} \mathbb{E}\|W_{12}\|^2.$$
Using Lemma \ref{bornes}, there exists $C>0$ such that
$$\|W_{12}\|\leq \sum_{l=-4}^4 \frac{2(1+|l|)}{\sqrt{2}\pi}\frac{1}{2\pi}\leq C$$ so that $\mathbb{E}(\|\sqrt{n}A_n\|^2)\leq {8C^2}/{(n-1)} $.
Finally, invoking Markov inequality we have that $\sqrt{n}A_n= o_P(1)$.
We can write $\sqrt{n}B_n$ in the following way:
\[
\sqrt{n} B_n = \frac{2}{\sqrt{n}} \sum_{k=1}^n U_k(\theta_0) ,
\]
where we set $U_{k}(\theta_0):= \sum_{l =-4}^4 \dot J^l(\theta_0) Z^l_{k}(\theta_0) $.
Note that the $U_{k}(\theta_0)$'s are i.i.d and centered. Invoking the central limit theorem, we have that
$$
\frac{1}{\sqrt{n}} \sum_{k=1}^n U_k(\theta_0) \overset{d}{\longrightarrow} \mathcal{N}(0,V/4),
$$
where $V/4$ is the covariance matrix of $U_1(\theta_0)$, equal to $\mathbb{E}(U_1(\theta_0)U_1(\theta_0)^\top)$.
\medskip
\textbf{Step 2}- Let us prove that $\ddot S_n(\theta_n^*) \overset{P}{\longrightarrow} \mathcal{A}(\theta_0)$
where $\mathcal{A}(\theta_0)= 2 \sum_{l=-4}^4 \dot J^l(\theta_0) \dot J^l(\theta_0)^\top $.
First, we have
\begin{eqnarray*}
\mathbb{E}(\ddot S_n(\theta_0)) &=& \ddot S(\theta_0)=2\sum_{l=-4}^4 (\ddot J^l(\theta_0) \underbrace{J^l(\theta_0)}_{=0}+ \dot J^l(\theta_0) \dot J^l(\theta_0)^\top) \\
&=& 2 \sum_{l=-4}^4 \dot J^l(\theta_0) \dot J^l(\theta_0)^\top=\mathcal{A}(\theta_0) .
\end{eqnarray*}
Next we write the decomposition
$$
\| \ddot S_n(\theta_n^*) - \mathcal{A}(\theta_0) \|_F \leq \| \ddot S_n(\theta_n^*) - \ddot S_n(\theta_0) \|_F + \| \ddot S_n(\theta_0) - \mathbb{E} \ddot S_n(\theta_0) \|_F
$$
We get due to the Lipschitz property of $\ddot S_n$ stated in Lemma \ref{lip} that
\[
\P \left ( \| \ddot S_n(\theta_0) -\ddot S_n(\theta_n^*) \|_F \geq \varepsilon \right ) \leq \P \left ( K\| \theta_n^* -\theta_0 \| \geq \varepsilon \right ) \rightarrow 0,
\]
because $\hat \theta_n \overset{P}{\longrightarrow} \theta_0.$
%
Last, let us focus on the term $ \| \ddot S_n(\theta_0) - \mathbb{E} \ddot S_n(\theta_0) \|_F$. We remind that
$$
\ddot S_n(\theta_0)- \mathbb{E} \ddot S_n(\theta_0) = \frac{2}{n(n-1)} \sum_{k \neq j} \sum_{l=-4}^4 \left(\ddot Z^l_{k}(\theta_0) Z^l_{j}(\theta_0)+ \dot Z^l_{k}(\theta_0) \dot Z^l_{j}(\theta_0)^\top
-\dot J^l(\theta_0) \dot J^l(\theta_0)^\top\right).
$$
From now on, we drop indices $l$ and $\theta_0$ to simplify the notation. We center the variables in order to find uncorrelatedness:
\begin{eqnarray*}
\ddot Z_{k} Z_{j}+\dot Z_{k} \dot Z_{j}^\top-\dot J\dot J^\top=
\underbrace{\left(\ddot Z_{k}-\ddot J\right) Z_{j}}_{A}+
\underbrace{\ddot J Z_j}_{B}+
\underbrace{(\dot Z_{k} -\dot J)(\dot Z_{j}-\dot J)^\top}_{C}
\\+\underbrace{\dot J(\dot Z_{j}-\dot J)^\top}_{D}
+\underbrace{( \dot Z_{k} -\dot J)(\dot J)^\top}_{E}
\end{eqnarray*}
(remind that $\mathbb{E}(Z_j)=J^l(\theta_0)=0$). Then
$
\ddot S_n(\theta_0)- \mathbb{E} \ddot S_n(\theta_0) = 2\sum_{l=-4}^4 (A+B +C+D+E)
$
where
\begin{align*}
A&=\frac{2}{n(n-1)} \sum_{k < j}\left(\ddot Z_{k}-\ddot J\right) Z_{j}\\
B&=\frac{1}{n} \sum_{ j=1}^n\ddot JZ_j\\
C&=\frac{2}{n(n-1)} \sum_{k < j}(\dot Z_{k} -\dot J)(\dot Z_{j}-\dot J)^\top\\
D&=\frac{1}{n} \sum_{j=1}^n\dot J(\dot Z_{j}-\dot J)^\top\\
E&=\frac{1}{n} \sum_{k =1}^n ( \dot Z_{k} -\dot J)\dot J^\top=D^\top
\end{align*}
Using the weak law of large numbers for uncorrelated centered variables, we obtain that
$\| \ddot S_n(\theta_0) - \mathbb{E} \ddot S_n(\theta_0) \|_F \overset{P}{\to} 0$ which completes the step 2.
Finally it is sufficient to apply Slutsky's Lemma to obtain
the theorem.
\subsection{Estimation of the covariance}\label{sec:estimV}
\begin{proposition}\label{estimV} Consider notation and assumptions of Theorem~\ref{normasymp}.
Let $V=4\mathbb{E}(U_1U_1^\top)$ where $U_1= \sum_{l=-4}^4 Z^l_{1}(\theta_0)\dot J^l(\theta_0) .$ Then $$\frac{4}{n^3}\sum_{1\leq k, j,j'\leq n}\sum_{-4\leq l,l'\leq 4} Z_k^{l}(\hat\theta_n)Z_k^{l'}(\hat\theta_n)\dot Z_j^{l}(\hat\theta_n)(\dot Z_{j'}^{l'}(\hat\theta_n))^\top.$$ tends almost surely toward $V$ when $n$ tends to $+\infty$.
\end{proposition}
Thus we obtain a consistent estimator for $V$ (that allows to estimate the covariance $\Sigma$). Nevertheless this estimator is biased. Notice that the quantity
$$ \frac{4}{n(n-1)(n-2)}\sum_{k=1}^n\sum_{j\neq k}\sum_{j'\notin\{k,j\}}\sum_{-4\leq l,l'\leq 4} Z_k^{l}(\theta_0)Z_k^{l'}(\theta_0)\dot Z_j^{l}(\theta_0)(\dot Z_{j'}^{l'}(\theta_0))^\top$$
has expectation
$$ 4\sum_{-4\leq l,l'\leq 4} \mathbb{E}[Z_1^{l}(\theta_0)Z_1^{l'}(\theta_0)]\dot J^{l}(\theta_0)(\dot J^{l'}(\theta_0))^\top=V$$
and we could also prove (with some additional technicalities in the following proof about the uniform convergence in $k$) that it tends almost surely toward $V$.
However, we lose the "unbiased" property when replacing $\theta_0$ by $\hat \theta_n$.
\subsubsection*{Proof of Proposition~\ref{estimV}}
Let $U_k=\sum_{l=-4}^4 Z_k^l(\theta_0)\dot J^l(\theta_0)$. The law of large numbers gives
$$V=\mathbb{E}(4U_1U_1^\top)=\lim_{n\to \infty} \frac{4}{n}\sum_{k=1}^n U_kU_k^\top$$
where the convergence is almost sure.
Moreover
$$U_kU_k^\top=\sum_{-4\leq l,l'\leq 4} Z_k^{l}Z_k^{l'}\dot J^{l}(\dot J^{l'})^\top=\lim_{n\to \infty}\frac1{n^2}
\sum_{-4\leq l,l'\leq 4} Z_k^{l}Z_k^{l'}\sum_{1\leq j,j'\leq n}\dot Z_j^{l}(\dot Z_{j'}^{l'})^\top$$
where the convergence is almost sure and we have dropped the $\theta_0$ for the sake of simplicity. This convergence is uniform in $k$ in the following sense: there exists a set with probability 1 for which for any $\varepsilon>0$, there exists $N\geq 1$ such that for all $n\geq N$ and for all $1\leq k\leq n$
$$\left\| \frac1{n^2}
\sum_{-4\leq l,l'\leq 4} Z_k^{l}Z_k^{l'}\sum_{1\leq j,j'\leq n}\dot Z_j^{l}(\dot Z_{j'}^{l'})^\top-U_kU_k^\top\right\|\leq \varepsilon$$
Indeed
\begin{align*}
&\left\| \frac1{n^2}
\sum_{-4\leq l,l'\leq 4} Z_k^{l}Z_k^{l'}\sum_{1\leq j,j'\leq n}\dot Z_j^{l}(\dot Z_{j'}^{l'})^\top-U_kU_k^\top\right\|\\
&=
\left\|
\sum_{-4\leq l,l'\leq 4} Z_k^{l}Z_k^{l'}\left(\frac1{n^2}\sum_{1\leq j,j'\leq n}\dot Z_j^{l}(\dot Z_{j'}^{l'})^\top-\dot J^{l}(\dot J^{l'})^\top\right)\right\|\\
&\leq \frac{1}{4\pi^2}\sum_{-4\leq l,l'\leq 4} \left\|\left(\frac1{n^2}\sum_{1\leq j,j'\leq n}\dot Z_j^{l}(\dot Z_{j'}^{l'})^\top-\dot J^{l}(\dot J^{l'})^\top\right)\right\|.
\end{align*}
Then we use the following lemma:
"If $v_{nk} \to v_k$ uniformly, with $(v_{nk})$ and $(v_k)$ bounded, and if $n^{-1}\sum_{k=1}^n v_k \to v$ then $n^{-1}\sum_{k=1}^n v_{nk} \to v$."
To prove this lemma, notice that, for a given positive $\varepsilon$, for $n$ large enough
\begin{eqnarray*}
\left|\frac{1}{n}\sum_{k=1}^n v_{nk}-v\right|
&\leq & \frac{1}{n}\sum_{k=1}^{N} \left|v_{nk}-v_k\right|+
\frac{1}{n}\sum_{k=N+1}^n \left| v_{nk}-v_k\right|+\left|\frac{1}{n}\sum_{k=1}^n v_{k}-v\right| \\
&\leq & \frac{N}{n}(\sup_{kn}|v_{nk}|+\sup_k |v_k|)+\frac{n-N}{n}\varepsilon+ \varepsilon \leq 3\varepsilon.
\end{eqnarray*}
That provides
$$V=\lim_{n\to \infty} \frac{4}{n^3}\sum_{1\leq k, j,j'\leq n}\sum_{-4\leq l,l'\leq 4} Z_k^{l}Z_k^{l'}\dot Z_j^{l}(\dot Z_{j'}^{l'})^\top$$
where the convergence is almost sure. Here all the $Z_k$ are depending on $\theta_0$, but we can use the consistency of $\hat\theta_n$ to finally assert
$$V=\lim_{n\to \infty} \frac{4}{n^3}\sum_{1\leq k, j,j'\leq n}\sum_{-4\leq l,l'\leq 4} Z_k^{l}(\hat\theta_n)Z_k^{l'}(\hat\theta_n)\dot Z_j^{l}(\hat\theta_n)(\dot Z_{j'}^{l'}(\hat\theta_n))^\top.$$
{
\subsection{Proof of Proposition~\ref{varthetachap}}
\label{sec:proof1varthetachap}}
We use the proof of Theorem~\ref{normasymp}. We have seen that
$$
\ddot{S_n}(\theta_n^*)(\hat \theta_n -\theta_0) = - \dot{S_n}(\theta_0),
$$
with $\theta_n^*$ in the line segment with extremities $\theta_0$ and $\hat \theta_n$.
Recall that $\dot S_n(\theta_0) =A_n+B_n$ with
\begin{eqnarray*}
A_n &= & \frac{4}{n(n-1)} \sum_{k< j} \sum_{l=-4}^4( \dot Z^l_{k}(\theta_0) - \dot J^l(\theta_0)) Z^l_{j}(\theta_0) \\
B_n &=&
=\frac{2}{n} \sum_{k=1}^n U_k(\theta_0)
\end{eqnarray*}
where $U_{k}(\theta_0):= \sum_{l =-4}^4 \dot J^l(\theta_0) Z^l_{k}(\theta_0).$
%
Note that the $U_{k}(\theta_0)$'s are i.i.d and centered so that
$$ \mathbb{E}\left\|\sum_k U_k(\theta_0)\right\|^2
=\sum_{j=1}^3\mbox{Var}\left(\sum_k U_{kj}(\theta_0)\right)
=\sum_{j=1}^3n\mbox{Var}\left(U_{1j}(\theta_0)\right)
\leq n c_1$$
using Lemma~\ref{bornes}. Here $c_1$ is a numerical constant.
Thus $\mathbb{E}\|B_n\|_1^2\leq 4c_1/n$.
In the same way the variables $W_{jk}:= \left(\sum_{l=-4}^4 ( \dot Z_{k}^l(\theta_0) - \dot J^l(\theta_0)) Z_{j}^l(\theta_0) \right)_{k<j}$ are centered and uncorrelated, and also bounded.
Then
$$\mathbb{E}(\|n(n-1)A_n\|^{2})=\mathbb{E}\left(\left\| \sum_{k< j}W_{jk}\right\|^{2}\right)
=n(n-1)\mathbb{E}\left(\left\| W_{12}\right\|^{2}\right)\leq n(n-1)c_2.$$
Then $\mathbb{E}\left( \|A_n\|^{2}\right)\leq c_2/n$ and
$\sup_n\mathbb{E}\left( n\|\dot S_n(\theta_0)\|^{2}\right)\leq 8c_1+2c_2<\infty$.
In the proof of Theorem~\ref{normasymp}, we noted that $\ddot S_n(\theta_n^*)
$ tends to $\ddot S(\theta_0)$ in probability. Actually we can prove that the convergence is almost sure.
Indeed the strong law of large numbers is true for uncorrelated variables if their second moments have a common bound (see e.g. \cite{chung2001}) so that $$\ddot S_n(\theta_0) -\ddot S(\theta_0)=\ddot S_n(\theta_0) - \mathbb{E} \ddot S_n(\theta_0) \overset{a.s.}{\longrightarrow} 0.$$
Since $\ddot S_n$ is continuous, it is sufficient to show that the convergence of $\hat\theta_n$ towards $\theta$ is almost sure and this will imply that $\ddot S_n(\theta^*_n)$ converges almost surely towards $\ddot S_n(\theta_0)$ (recall that $\theta_n^*$ in the line segment with extremities $\theta_0$ and $\hat \theta_n$).
To do this, remark first that $ S_n(\theta) - S(\theta) \overset{a.s.}{\longrightarrow} 0$ by the strong law of large numbers for uncorrelated variables again (see the decomposition of $S_n-S$ in the proof of Proposition~\ref{variance}). Now, we come back to the proof of Lemma~\ref{lem:convmin} (in the case of a unique minimum $\theta_0$), with this new assumption that $S_n(\theta)$ tends to $S(\theta)$ almost surely. The proof shows that for any $\epsilon>0$ there exist $\delta(\epsilon)>0$ and $\Delta_n(\epsilon)$ which tends to 0 almost surely such that
$$\|\hat\theta_n-\theta_0\|\geq \epsilon\Rightarrow \Delta_n(\epsilon)\geq \delta(\epsilon)/4.$$
Let $\Gamma=\cap_{p\geq 1}\{\Delta_n(1/p)\to 0\}$. This set has probability 1 and on this set, for any $\varepsilon>0$, taking $p \geq 1/\varepsilon$, there exists $N\geq 1$ such that for any $n\geq N$
$$\Delta_n(1/p)< \delta(1/p)/4 \text{ and then } \|\hat\theta_n-\theta_0\|< (1/p)\leq \varepsilon.$$
This ensures that on the set $\Gamma$, $\hat\theta_n$ tends to $\theta_0$, and finally $\ddot S_n(\theta_n^*)$ tends to $\ddot S(\theta_0)$ almost surely.
Now, since $\ddot S(\theta_0)$ is assumed invertible, there exists $n_1$ such that for all $n\geq n_1$, $\ddot S_n(\theta_n^*) $ is invertible and
$\|\ddot S_n(\theta_n^*)^{-1}\|_{op}\leq 2\|\ddot S_n(\theta_0)^{-1}\|_{op}:=C(\theta_0)$ a.s.
Then
$$n\|\hat \theta_n-\theta_0\|^{2}\leq
C(\theta_0)^2 n\|\dot S_n(\theta_0)\|^{2}\quad \text{ a.s.}$$
and
$$\mathbb{E}(n\|\hat \theta_n-\theta_0\|^{2})\leq
C(\theta_0)^2\mathbb{E}( n\|\dot S_n(\theta_0)\|^{2})\leq C(\theta_0)^2(8c_1+2c_2).$$
Moreover $\sup_{\theta\in \Theta}C(\theta)<\infty $ because
$\Theta$ is a compact set and $\theta \mapsto \|\ddot S_n(\theta)^{-1}\|_{op}$ is continuous.
\subsection{Proof of Theorem~\ref{oracleineq} (nonparametric estimation)}
The proof of the oracle inequality is based on Lemma~\ref{lemadap} and Lemma~\ref{majnun} below.
The conclusion follows, choosing $2\gamma=\epsilon/(1+\epsilon)$
and $\lambda=\gamma^{-1}\kappa(1-2P)^{-2}=2\kappa(1+\epsilon^{-1})(1-2P)^{-2}$,
{ and $q=(2s_0+1)/3$}.
Let us derive the rate of convergence, which is the second result of Theorem~\ref{oracleineq}. We use the notation of Lemma~\ref{lemadap} and the notation $\|.\|_{\ell}$ for the natural norm of $\ell^2(\mathbb{C}^\mathbb{Z})$. Let $L\in \mathcal L$.
{Since
$\nu_n(t)= \sum_{l\in \mathbb{Z}} \overline{t_{l}} (\widehat{f^{\star l}}- f^{\star l})$,
$$\sum_{l=-L}^L |\widehat{f^{\star l}}- f^{\star l}|^2=\nu_n({\widehat{f^{\star}_L}-f^{\star}_L})\leq \sup_{t\in B_L}\nu_n(t)\|{\widehat{f^{\star}_L}-f^{\star}_L}\|_{\ell}$$
where we denote $ f_L^\star$ the sequence in $\mathbb{C}^\mathbb{Z}$ such that
$( f_L^\star)_l={f^{\star l}}$ if $-L\leq l \leq L$ and 0 otherwise.
Hence $\|{\widehat{f^{\star}_L}-f^{\star}_L}\|_{\ell}^2\leq \sup_{t\in B_L}\nu_n(t)\|{\widehat{f^{\star}_L}-f^{\star}_L}\|_{\ell}$ so that
$\|{\widehat{f^{\star}_L}-f^{\star}_L}\|_{\ell}\leq \sup_{t\in B_L}\nu_n(t)$.
Then, using Lemma~\ref{majnun}
\begin{eqnarray*}
\mathbb{E}\sum_{l=-L}^L |\widehat{f^{\star l}}- f^{\star l}|^2
&=&\mathbb{E}\|{\widehat{f^{\star}_L}-f^{\star}_L}\|_{\ell}^2
\leq \frac{\kappa}{(1-2P)^2} \frac{2L+1}{n}+\frac{C {(1+R^2)}}{n}\\
&\leq & C' {(1+R^2)}\frac{2L+1}{n}.
\end{eqnarray*}
Using Parseval's identity,
$$\mathbb{E}\|f-\hat f_L\|_2^2=\sum_{|l|>L} |f^{\star l}|^2+C' {(1+R^2)}\frac{2L+1}{n}\leq
{R^2(1+L^2)^{-s}+C' {(1+R^2)}\frac{2L+1}{n}.}$$
Thus, the oracle inequality gives
\begin{eqnarray*}
\mathbb{E}\|\hat f_{\widehat{L}}-f\|_2^2&\leq & (1+2\epsilon)\min_{L\in \mathcal{L}} \left\{
{R^2(1+L^2)^{-s}}+(C' {(1+R^2)}+2\lambda)\frac{2L+1}{n}\right\}\\
&&+\frac{C {(1+R^2)}}{n}\\
&\leq& C'' {R^2} n^{-2s/(2s+1)}
\end{eqnarray*}
{
choosing $L=L_0=\lfloor C n^{1/(2s+1)}\rfloor $.
This choice is possible since $s\geq s_0$ and then $L_0 $ belongs to $\mathcal{L}$.
}\\
\begin{lemma}
\label{lemadap}
Let $\lambda>0$ and $\mathcal{L}$ be a finite set of resolution level and define
$$\widehat{L}=\underset{L\in \mathcal{L}}{\argmin} \left\{-\sum_{l=-L}^L | \widehat{f^{\star l}}|^2+\lambda \frac{2L+1}{n}\right\}.$$
Then, for all $0<\gamma<1/2$,
\begin{eqnarray*}
(1-2\gamma)\|\hat f_{\widehat{L}}-f\|_2^2\leq \min_{L\in \mathcal{L}} \left\{(1+2\gamma)\|\hat f_{ L}-f\|_2^2+2\lambda \frac{2L+1}{n}\right\}\\+
{
\frac{1}{\gamma}\max_{L\in\mathcal{L}}\left( \sup_{t\in B_{L}}\nu_n^2(t)-\lambda \gamma\frac{2L+1}{n}\right)
}
\end{eqnarray*}
where
$B_L=\{t\in \mathbb{C}^{\mathbb{Z}}, \, \sum_{l\in \mathbb{Z}} |t_{l}|^2=1, \,t_l=0 \text{ if }|l|>L \}$
and
$\nu_n(t)= \sum_{l\in \mathbb{Z}} \overline{t_{l}} (\widehat{f^{\star l}}- f^{\star l})$.
\end{lemma}
\begin{proof}
We recall that the dot product $\langle f,g\rangle $ means $\frac{1}{2\pi}\int \overline{f(x)}g(x)dx$ and that $\|.\|_2$ is the associated norm. Usual Fourier analysis gives for any $L$:
\begin{align*}
\|\hat f_L-f\|_2^2&=-\|\hat f_L\|_2^2+2(\|\hat f_L\|_2^2-\langle \hat f_L,f\rangle )+\|f\|_2^2\\
&=-\sum_{l=-L}^{ L} | \widehat{f^{\star l}}|^2+2\sum_{l=-L}^L\overline{\widehat{f^{\star l}}}(\widehat{f^{\star l}}-f^{\star l})+\|f\|_2^2\\
&=-\sum_{l=-L}^{L} | \widehat{f^{\star l}}|^2+2\nu_n(\widehat f_L^\star)+\|f\|_2^2
\end{align*}
where we denote $\widehat f_L^\star$ the sequence in $\mathbb{C}^\mathbb{Z}$ such that
$(\widehat f_L^\star)_l=\widehat{f^{\star l}}$ if $-L\leq l \leq L$ and 0 otherwise.
Now let $L$ be an arbitrary resolution level in $\mathcal{L}$. Using the definition of $\widehat{L}$,
$$-\sum_{l=-\widehat{L} }^{\widehat{L}} | \widehat{f^{\star l}}|^2+\lambda \frac{2\widehat{L}+1}{n}\leq -\sum_{l=-L}^L | \widehat{f^{\star l}}|^2+\lambda \frac{2L+1}{n}.$$
Thus
$$\|\hat f_{\widehat{L}}-f\|_2^2-2\nu_n(\hat f_{\widehat{L}}^\star)+\lambda \frac{2\widehat{L}+1}{n}\leq \|\hat f_L-f\|_2^2-2\nu_n(\hat f_L^\star)+\lambda \frac{2L+1}{n}$$
which leads to
$$\|\hat f_{\widehat{L}}-f\|_2^2\leq \|\hat f_L-f\|_2^2+2\nu_n(\widehat f_{\widehat{L}}^\star-\widehat f_L^\star)-\lambda \frac{2\widehat{L}+1}{n}+\lambda \frac{2L+1}{n}.$$
But, denoting by $\|.\|_{\ell}$ the natural norm of $\ell^2(\mathbb{C}^\mathbb{Z})$
\begin{align*}
2\nu_n(\widehat f_{\widehat{L}}^\star-\widehat f_L^\star)&=
2\nu_n\left(\frac{\widehat f_{\widehat{L}}^\star-\widehat f_L^\star}{\|\widehat f_{\widehat{L}}^\star-\widehat f_L^\star\|_{\ell}}\right)\|\widehat f_{\widehat{L}}^\star-\widehat f_L^\star\|_{\ell}\\
2\left|\nu_n(\widehat f_{\widehat{L}}^\star-\widehat f_L^\star)\right|&\leq \gamma\|\widehat f_{\widehat{L}}^\star-\widehat f_L^\star\|_{\ell}^2 + \frac{1}{\gamma}\left|\nu_n\left(\frac{\widehat f_{\widehat{L}}^\star-\widehat f_L^\star}{\|\widehat f_{\widehat{L}}^\star-\widehat f_L^\star\|_{\ell}}\right)\right|^2\\
&\leq 2\gamma (\|\hat f_{\widehat{L}}-f\|_2^2 +
\| f-\hat f_L\|_2^2 )
+ \frac{1}{\gamma}\sup_{t\in B_{L\vee \widehat{L}}}|\nu_n(t)|^2
\end{align*}
where $L\vee \widehat{L}=\max(L, \widehat{L})$.
Thus
\begin{align*}
\|\hat f_{\widehat{L}}-f\|_2^2(1-2\gamma) \leq \|\hat f_L-f\|_2^2(1+2\gamma)+ \frac{1}{\gamma}\sup_{t\in B_{L\vee \widehat{L}}}|\nu_n(t)|^2-\lambda \frac{2\widehat{L}+1}{n}+\lambda \frac{2L+1}{n}\\
\leq \|\hat f_L-f\|_2^2(1+2\gamma)+ \frac{1}{\gamma}\left(\sup_{t\in B_{L\vee \widehat{L}}}|\nu_n(t)|^2-\lambda \gamma \frac{2\widehat{L}+2L+2}{n}\right)+2\lambda \frac{2L+1}{n}\\
\leq \|\hat f_L-f\|_2^2(1+2\gamma)+2\lambda \frac{2L+1}{n}+\frac{1}{\gamma}\max_{L'\in\mathcal{L}}\left(\sup_{t\in B_{L'}}|\nu_n(t)|^2-\lambda \gamma \frac{2L'+1}{n}\right).
\end{align*}
\end{proof}
\begin{lemma} \label{majnun}
Assume Assumption Assumption \ref{hypf} and \ref{hypthetabis}.
Assume that $f$ belongs to the Sobolev ellipsoid $W(s,R)$
{
with $s\geq 1$.
Assume that
$\mathcal{L}=\{0, \dots, L_n\}$ with ${L}_n$ such that $L_n^3\leq C_{\mathcal L}n^{1/q}$ for some $q>1$.
Then, with the notation of Lemma~\ref{lemadap},
for all $\kappa> 3/(2\pi^2)$,
}
$$\mathbb{E}\max_{L\in \mathcal{L}}\left( \sup_{t\in B_L}|\nu_n(t)|^2-\frac{\kappa}{(1-2P)^2} \frac{2L+1}{n}\right)\leq \frac{C {(1+R^2)}}{n},$$
where $C$ is a positive constant depending on
{$P,q, C_{\mathcal L}, \kappa$}.
\end{lemma}
\begin{proof}
Denote $R^l=\frac{1}{M^l(\hat \theta)}-\frac{1}{M^l( \theta_0)}$.
First note that
$$\nu_n(t)=\frac1{2\pi n}\sum_{k=1}^n\sum_{l\in \mathbb{Z}} \overline{t_{l}} \left(\frac{e^{-ilX_k}}{M^l(\hat \theta)}-
\frac{2\pi g^{\star l}}{M^l( \theta_0)}\right)
=\nu_{n,1}(t)+\nu_{n,2}(t)+\nu_{n,3}(t)$$
where
\begin{align*}
\nu_{n,1}(t)&=\frac1{2\pi n}\sum_{k=1}^n\sum_{l\in \mathbb{Z}} \overline{t_{l}} \left(\frac{e^{-ilX_k}-2\pi g^{\star l}}{M^l( \theta_0)}\right)\\
\nu_{n,2}(t)&=\frac1{2\pi n}\sum_{k=1}^n\sum_{l\in \mathbb{Z}} \overline{t_{l} }\left(e^{-ilX_k}-2\pi g^{\star l}\right)R^l\\
\nu_{n,3}(t)&=\frac1{n}\sum_{k=1}^n\sum_{l\in \mathbb{Z}} \overline{ t_{l}} g^{\star l}R^l=\sum_{l\in \mathbb{Z}} \overline{ t_{l}} g^{\star l}R^l.
\end{align*}
{
Thus $|\nu_n|^2\leq 3|\nu_{n,1}|^2+3|\nu_{n,2}|^2+3|\nu_{n,3}|^2$, and,
if $\kappa_1=\kappa/3$,
\begin{align*}
&\mathbb{E}\max_L\left( \sup_{ B_L}|\nu_n|^2-\frac{\kappa}{(1-2P)^2} \frac{2L+1}{n}\right)\leq
3\mathbb{E}\sum_L\left( \sup_{ B_L}|\nu_{n,1}|^2-\frac{\kappa_1}{(1-2P)^2} \frac{2L+1}{n}\right)_+\\
&+
3\mathbb{E}\max_L\left( \sup_{B_L}|\nu_{n,2}|^2\right)
+3\mathbb{E}\max_L\left( \sup_{B_L}|\nu_{n,3}|^2\right)
\end{align*}
where $a_+=\max(a,0)$ denotes the positive part of $a$.\\
}
\textit{\underline{Control of $\nu_{n,3}$}}
First note that
$$\left|g^{\star l}R^l\right|=\left| f^{\star l}\frac{M^l( \theta_0)-M^l(\hat \theta)}{M^l( \hat\theta)}\right| \leq \frac{|f^{\star l}|}{1-2P}
\left|M^l( \theta_0)-M^l(\hat \theta)\right|.$$
Thus, using Schwarz inequality
$$\sup_{t\in B_L} |\nu_{n,3}(t)|^2 \leq \sum_{l=-L}^L \frac{|f^{\star l}|^2}{(1-2P)^2}
\left|M^l( \theta_0)-M^l(\hat \theta)\right|^2.$$
Moreover
\begin{eqnarray*}
|M^l( \theta_0)-M^l(\hat \theta)|
&\leq & \left|(p_0-\hat p)e^{-i\alpha_0 l}+\hat p (e^{-i\alpha_0 l}-e^{-i\hat \alpha l})
+(1-p_0-1+\hat p)e^{-i\beta_0 l}\right.
\\&&+ \left. (1-\hat p) (e^{-i\beta_0 l}-e^{-i\hat \beta l})\right|\\
&\leq & |p_0-\hat p|+|e^{-i\alpha_0 l}-e^{-i\hat \alpha l}|
+|p_0-\hat p|+|e^{-i\beta_0 l}-e^{-i\hat \beta l}|\\
&\leq & 2 |p_0-\hat p|+|l||\alpha_0-\hat \alpha |+|l||\beta_0 -\hat \beta|\leq 2 |l| \|\theta_0-\hat \theta\|_1
\end{eqnarray*}
(note that it is also true for $l=0$ since $M^0( \theta_0)=M^0(\hat \theta)=1$).
Thus, for any $L\in\mathcal{L}$
$$\sup_{t\in B_L} |\nu_{n,3}(t)|^2\leq \sum_{l=-L_n}^{L_n} \frac{|f^{\star l}|^2}{(1-2P)^2} 4 |l|^2 \|\theta_0-\hat \theta\|_1^2
$$
{
Since $s\geq 1$
$$\max_{L\in\mathcal{L}}\left(\sup_{ B_L} |\nu_{n,3}|^2\right) \leq \frac4{(1-2P)^2}\sum_{l=-L_n}^{L_n}{|f^{\star l}|^2} |l|^{2s} \|\theta_0-\hat \theta\|_1^2
\leq \frac{4R^2}{(1-2P)^2}\|\theta_0-\hat \theta\|_1^2.
$$
According to Proposition~\ref{varthetachap} and inequality $\|x\|_1^2\leq 3 \|x\|^2$,
\label{Ktheta}
there exists a constant $K>0$ such that
$\mathbb{E}(n\|\hat \theta - \theta_0\|_1^2)\leq K$. Then
$$\mathbb{E}\max_{L\in\mathcal{L}}\left(\sup_{ B_L} |\nu_{n,3}|^2\right)\leq \frac{C_3R^2}{n}$$
with $C_3=4K/(1-2P)^2$.
\medskip
\textit{\underline{Control of $\nu_{n,2}$}}
Note that
$$|R^l |\leq
\frac{1}{(1-2P)^2}
\left|M^l( \theta_0)-M^l(\hat \theta)\right|
\leq
\frac{2}{(1-2P)^2}
|l|\|\hat \theta-\theta_0\|_1,$$ so for $t\in B_L$,
\begin{align*}
| \nu_{n,2}(t)|^2 \leq \left( \sum_{l=-L}^{L} |\overline{t_l}(\widehat{g^{\star l}} - g^{\star l})R^l|\right)^2
\leq \frac{4}{(1-2P)^4}\sum_{l=-L}^{L} |\widehat{g^{\star l}} - g^{\star l}|^2 l^2\|\hat \theta-\theta_0\|_1^2.
\end{align*}
Then, for any $L\in \mathcal{L}$
\begin{align*}
\sup_{t\in B_L} |\nu_{n,2}(t)|^2
&\leq \frac{4}{(1-2P)^2}\sum_{l=-L}^{L} |\widehat{g^{\star l}} - g^{\star l}|^2 l^2\|\hat \theta-\theta_0\|_1^2
\end{align*}
and
\begin{align*}
\max_{L\in \mathcal{L} }\sup_{t\in B_L} |\nu_{n,2}(t)|^2
&\leq \frac{4}{(1-2P)^2}\sum_{l=-L_n}^{L_n} |\widehat{g^{\star l}} - g^{\star l}|^2 l^2\|\hat \theta-\theta_0\|_1^2.
\end{align*}
Using H\"{o}lder's inequality, for any $p,q\geq 1$ such that $\frac1p+\frac1q=1$,
\begin{align*}
\mathbb{E}\max_{L\in \mathcal{L} }\sup_{t\in B_L} |\nu_{n,2}(t)|^2
& \leq\frac{4}{(1-2P)^2} \sum_{l=-L_n}^{L_n} l^2 \mathbb{E}^{1/p}(|\widehat{g^{\star l}} - g^{\star l}|^{2p})\mathbb{E}^{1/q}\|\hat \theta-\theta_0\|_1^{2q}
\end{align*}
%
But Proposition~\ref{varthetachap} gives us
$$\mathbb{E}\|\hat \theta-\theta_0\|_1^{2q}\leq (1+2\pi+2\pi)^{2q-2}\mathbb{E}\left(3\|\hat \theta-\theta_0\|^{2}\right)\leq
K'(q)n^{-1}.$$
Moreover, we can apply the Rosenthal inequality to the variables $Y_k=e^{ilX_k}-\mathbb{E}(e^{ilX_k})$: there exists $C(2p)>0$ such that
\begin{eqnarray*}
\mathbb{E} \left|\sum_{k=1}^n Y_k\right|^{2p}&\leq & C(2p)\left(
\sum_{k=1}^n \mathbb{E}|Y_k|^{2p}+\left(\sum_{k=1}^n \mathbb{E}|Y_k|^2\right)^{p}
\right)\\
&\leq & C(2p)\left(
n 2^{2p}+\left(4n\right)^{p}
\right)\leq C'(p)n^{p}
\end{eqnarray*}
that provides
$$\mathbb{E}(|\widehat{g^{\star l}} - g^{\star l}|^{2p})=\mathbb{E}\left((2\pi)^{-2p}
\left |\frac1n\sum_{k=1}^n Y_k\right|^{2p}\right)
\leq (2\pi)^{-2p}C'(p)n^{-p}.$$
Thus
\begin{align*}
\mathbb{E}\max_{L\in \mathcal{L} }\sup_{t\in B_L} |\nu_{n,2}(t)|^2
&\leq \frac{4}{(1-2P)^2} \sum_{l=-L_n}^{L_n} l^2
(2\pi)^{-2}C'(p)^{1/p}n^{-1}
K'(q)^{1/q}n^{-1/q}\\
& \leq \frac{C''(q)}{(1-2P)^2}n^{-1-1/q}L_n^3.
\end{align*}
Since $L_n^3\leq C_{\mathcal L} n^{1/q}$, we obtain
\begin{align*}
\mathbb{E}\max_{L\in \mathcal{L} }\sup_{t\in B_L} |\nu_{n,2}(t)|^2
& \leq \frac{C_2}{n}
\end{align*}
with $C_2=C''(q)C_{\mathcal L}/(1-2P)^2.$
}
\medskip
\textit{\underline{Control of $\nu_{n,1}$}}
To control $\nu_{n,1}$ , we need Talagrand's inequality.
\begin{lemma}\label{InegTalagrand}
Let $X_1,\ldots,X_n$ be i.i.d. random variables, and define $\nu_n(t)=\frac{1}{n}\sum_{k=1}^n\psi_t(X_k)-\mathbb{E}[\psi_t(X_k)]$, for $t$ belonging to a countable class $\mathcal{B}$ of real-valued measurable functions. Then, for $\delta>0$, there exist three constants $c_l$, $l=1,2,3$, such that
\begin{eqnarray}\label{TalagrandIntegre}\displaystyle\mathbb{E}\left[ \left(\sup_{t\in\mathcal{B}}\left|\nu_n\left(t\right)\right|^2-c(\delta)H^2\right)_{+}\right]&\leq& c_1\left\{\frac{v}{n}\exp\left(-c_2\delta\frac{nH^2}{v}\right)\right.\\\nonumber &&\displaystyle\left.+\frac{M_1^2}{C^2(\delta)n^2}\exp\left(-c_3C(\delta)\sqrt{\delta}\frac{nH}{M_1}\right)\right\},\end{eqnarray}
\noindent with $C(\delta)=(\sqrt{1+\delta}-1)\wedge 1$, $c(\delta)=2(1+2\delta)$ and
$$\displaystyle \sup_{t\in\mathcal{B}} \|\psi_t\|_{\infty} \leq M_1\mbox{, }\mathbb{E}\left[ \sup_{t\in\mathcal{B}}\left|\nu_n(\psi_t)\right|\right]\leq H\mbox{, and }\sup_{t\in\mathcal{B}}\mbox{Var}\left(\psi_t\left(X_1\right)\right)\leq v.$$
\end{lemma}
Inequality (\ref{TalagrandIntegre}) is a classical consequence of Talagrand's inequality given in \citet{KR2005}: see for example Lemma 5 (page 812) in \cite{lacour08}. Using density arguments, we can apply it to the unit sphere of a finite dimensional linear space.
Here $\nu_{n,1}(t)=\frac{1}{n}\sum_{k=1}^n\psi_t(X_k)-\mathbb{E}[\psi_t(X_k)]$ with
$$\psi_t(X)=\frac1{2\pi }\sum_{l\in \mathbb{Z}} \overline{t_{l}} \frac{e^{-ilX}}{M^l( \theta_0)}, \qquad \mathbb{E}(\psi_t(X))=\sum_{l\in \mathbb{Z}} \overline{t_{l}} \frac{g^{\star l}}{M^l( \theta_0)}$$
Let us compute $M_1, H$ and $v$.
\begin{itemize}
\item Using Cauchy Schwarz inequality, for $t\in B_L$,
\begin{align*}
|\psi_t(u)|^2&=\left|\frac1{2\pi }\sum_{l=-L}^L \overline{t_{l}} \frac{e^{-ilu}}{M^l( \theta_0)}\right|^2 \leq \frac1{4\pi ^2}\sum_{l=-L}^L|t_l|^2\sum_{l=-L}^L \left| \frac{e^{-ilu}}{M^l( \theta_0)}\right| ^2\\
&\leq \frac{1}{4\pi^2(1-2p_0)^2}(2L+1),
\end{align*}
thus $M_1=\frac{1}{2\pi(1-2p_0)}\sqrt{2L+1}.$
\item
Using Cauchy Schwarz inequality, for $t\in B_L$,
\begin{align*}
\sup_{t\in {B}_L}\left|\frac{1}{2\pi n}\sum_{k=1}^n\sum_{l\in \mathbb{Z}} \overline{ t_{l}} \left(\frac{e^{-ilX_k}}{M^l(\theta_0)}-
\mathbb{E}\left(\frac{e^{-ilX_k}}{M^l( \theta_0)}\right)\right)\right|^2\\
\leq \sum_{l=-L}^L\left|\frac{1}{2\pi n}\sum_{k=1}^n\left(\frac{e^{-ilX_k}}{M^l(\theta_0)}-
\mathbb{E}\left(\frac{e^{-ilX_k}}{M^l( \theta_0)}\right)\right)\right|^2,
\end{align*}
then
\begin{align*}
\mathbb{E}\left(\sup_{t\in {B}_L}\left|\nu_{n,1}(\psi_t)\right|^2\right)
& \leq \sum_{l=-L}^L \mbox{Var} \left(
\frac{1}{2\pi n}\sum_{k=1}^n\frac{e^{-ilX_k}}{M^l(\theta_0)}\right)
\leq \sum_{l=-L}^L \frac{1}{4\pi^2 n} \mbox{Var} \left(
\frac{e^{-ilX_1}}{M^l(\theta_0)}\right)\\
& \leq \frac{1}{4\pi^2 n} \sum_{l=-L}^L\mathbb{E}\left|
\frac{e^{-ilX_1}}{M^l(\theta_0)}\right|^2
\leq \frac{1}{4\pi^2(1-2p_0)^2} \frac{2L+1}{ n},
\end{align*}
thus by Jensen’s inequality $H^2=\frac{1}{4\pi^2(1-2p_0)^2} \frac{2L+1}{ n}$.
\item It remains to control the variance. If $t\in B_L$
\begin{align*}
\mbox{Var}(\psi_t(X))\leq \mathbb{E}\left| \frac1{2\pi }\sum_{l=-L}^L \overline{ t_{l}} \frac{e^{-ilX}}{M^l( \theta_0)}\right|^2
=\frac1{4\pi^2 }\sum_{l,l'} t_{l}\overline{t_{l'} }\frac{\mathbb{E}(e^{-ilX}\overline{e^{-il'X}})}{M^l( \theta_0)\overline{M^{l'}( \theta_0)}} \\
=\frac1{2\pi }\sum_{l,l'} t_{l}\overline{t_{l'} }\frac{g^{\star(l-l')}}{M^l( \theta_0)\overline{M^{l'}( \theta_0)}}
\end{align*}
Using twice Schwarz inequality
\begin{align*}
\mbox{Var}(\psi_t(X))&\leq \frac1{2\pi }\sqrt{\sum_{l}\left|\frac{t_{l}}{M^l( \theta_0)}\right|^2\sum_l\left|\sum_{l'} \frac{\overline{t_{l'} }}{\overline{M^{l'}( \theta_0)}}g^{\star(l-l')}\right|^2} \\
&\leq \frac1{2\pi (1-2p_0)}\sqrt{\sum_l\left|\sum_{l'} \frac{\overline{t_{l'} }}{\overline{M^{l'}( \theta_0)}}g^{\star(l-l')}\right|^2} \\
&\leq \frac1{2\pi (1-2p_0)}\sqrt{\sum_l\sum_{l'} \left|\frac{\overline{t_{l'} }}{\overline{M^{l'}( \theta_0)}}\right|^2\sum_{l'}\left|g^{\star(l-l')}\right|^2} \\
&\leq \frac1{2\pi (1-2p_0)}\sqrt{\sum_l\frac{1}{|1-2p_0|^2}\sum_{j\in \mathbb{Z}}\left|g^{\star j}\right|^2} \\
&\leq \frac{ { R}}{2\pi (1-2p_0)^2}\sqrt{2L+1}, \\
\end{align*}
since $\sum_{j\in \mathbb{Z}}\left|g^{\star j}\right|^2\leq \sum_{j\in \mathbb{Z}}\left|f^{\star j}\right|^2 {\leq R^2}$.
Thus $v=\frac{ { R}}{2\pi (1-2p_0)^2}\sqrt{2L+1} $.
\end{itemize}
Inequality (\ref{TalagrandIntegre}) becomes
\begin{eqnarray*}
& &\displaystyle\mathbb{E}\left[ \left(\sup_{t\in {B}_L}\left|\nu_{n,1}\left(t\right)\right|^2-\frac{c(\delta)}{4\pi^2(1-2p_0)^2} \frac{2L+1}{ n}\right)_{+}\right]\\
&& \leq
c_1\left\{\frac{R\sqrt{2L+1} }{2\pi (1-2p_0)^2n}\exp\left(-c_2\delta\frac{\sqrt{2L+1}}{2\pi R}\right)\right.\\
&&\qquad \displaystyle\left.+\frac{2L+1}{4\pi^2(1-2p_0)^2C^2(\delta)n^2}\exp\left(-c_3C(\delta)\sqrt{\delta n}\right)\right\}\\
&& \leq \frac{K\max(R,1)}{n}
\left\{\sqrt{2L+1} \exp\left(-c\sqrt{2L+1}\right)+\frac{2L+1}{n}\exp\left(-c\sqrt{ n}\right)\right\}
\end{eqnarray*}
with $K$
and $c$ positive constants depending on $P,c_1,c_2,c_3, \delta$.
%
This ends the control of $\nu_{n,1}$ with $\kappa_1=\frac{c(\delta)}{4\pi^2}$ since
$$\sum_{L\in \mathcal{L}}\left\{\sqrt{2L+1} e^{-c\sqrt{2L+1}}+\frac{2L+1}{n}e^{-c\sqrt{ n}}\right\}\leq \sum_{L=0}^{\infty}\sqrt{2L+1} e^{-c\sqrt{2L+1}}+\sharp\mathcal{L} e^{-c\sqrt{ n}}=
O(1).$$
{
Finally it is sufficient to take
$$\kappa\geq 3\kappa_1=
\frac{3(2+4\delta)}{4\pi^2}=
\frac{3}{2\pi^2}+\frac{3\delta}{\pi^2}$$
to conclude the proof. Since $\delta$ can be chosen arbitrary small, and we have assumed $\kappa>3/(2\pi^2)$, this condition is satisfied.
}
\end{proof}
\section*{Acknowledgement}
The authors would like to thank the Editors and one anonymous referee for valuable comments and suggestions leading to corrections and improvements of the article.
\bibliographystyle{apalike}
|
\section{Introduction}
Project-based learning~\cite{doi:10.1177/1365480216659733} has been used worldwide as an active methodology~\cite{Freeman8410} to provide the students with a more realistic and professional experience. In the software engineering undergraduate program at PUC Minas in Brazil, we make extensive use of project-based courses. During each semester of the program, students have to develop a software application \cite{Nelson:2017}. In general the projects are a way to motivate the students to take responsibility for their learning process. They challenge themselves through the process and go deeper and further than in traditional courses. Furthermore, since this generation is more socially aware they are keen to solve problems brought by the surrounding community.
In each semester, the projects have different goals and are always related to the other courses that students are taking in the same semester. These other courses are the anchors for the projects. Students work in groups that vary in size from 3 to 6 depending on the project and on the specific course. Releases occur every week or every other week depending on the course. Usually they like to collaborate and improve these skills. Because of this team work, we have always motivated the students to find repository tools where they could share their work, collaborate and keep track of the versions of the system under development. During the project, students would share their work with their supervising professor so that they can provide feedback and directions on how to continue, as well as suggest improvements.
From the point of view of the professors, the shared repositories were only a way to visualize student work. Each group used a different repository according to the familiarity that they had with such tools. Sometimes, even inside a single team it was hard to achieve consensus about which tool to use, and we have awkward decisions such as two repositories in different tools, one for the backend and another for the frontend. By observing these not-so-good practices, the supervising professors started to discuss what would be a better way to improve this situation.
Some of these projects have real clients. Two project courses are officially considered extension courses, meaning the students work directly with the community surrounding the university to provide this community with a concrete return of the application of the students' knowledge. At the end of the semester, some of these clients and stakeholders want to further develop and elaborate these projects. They also need to have access to these repositories. Furthermore, some of these projects span more than one semester, done by different teams. The need to have more control over these repositories to guarantee their longevity became apparent.
Considering the needs of the program and the five professors who coordinate it, it is necessary to keep track of the projects to register what students are doing, to evaluate how the project-based courses are evolving, and to showcase the best projects. Every semester the best projects are published in the program's website and in a magazine describing the clients' problems and software solutions the teams came up with.
All those different needs and perspectives result in a challenge to manage projects every semester and their evolution overtime. Currently, we have approximately 60 projects per semester, with 287 enrolled students, and 13 supervising professors. Seeking a software approach that allows us to manage learning, participation and meet stakeholders needs, we developed an approach based on the GitHub Classroom tool. In the next sections, we describe the proposed approach (Section~\ref{sec:approach}), results of experience in its use (Section~\ref{sec:experience}), and the road ahead, analyzing the new insights on metrics for monitoring their learning, on student work portfolios, and on tooling support for project-based courses (in Section~\ref{sec:ahead}).
\section{A GitHub Classroom Approach}
\label{sec:approach}
GitHub Classroom\footnote{Available at: https://classroom.github.com/ (Accessed: 01 January 2021)} is a tool that provides a dashboard for standardizing, organizing, and managing student' code repositories and grading work automatically. The GitHub Classroom is integrated with the GitHub platform\footnote{Available at: http://github.com/ (Accessed: 01 January 2021)}. Students and professors must have an account on this platform. In the GitHub Classroom terminology, a {\it classroom} is composed of a set of students, a set of professors and teaching assistants. When created, a classroom is linked to an {\it organization} on the GitHub platform. Classrooms of all courses in the undergraduate program are associated with the same organization. The inclusion of students in the classroom can be done automatically through an integration with the Learning Management Systems.
Professors and students act on GitHub Classroom as follows. Professors can create classrooms and, within those classrooms, create {\it assignments} for students. An assignment is usually a coding task, which can be performed by a student individually or by a group of students. For each student/group who accepts an assignment, GitHub Classroom creates a {\it code repository} automatically. Depending on the configuration made by the professor, the repository can be public or private, and can also have a standard structure of folders and files based in a predefined template. The student receives the link to the repository created by GitHub Classroom. All student's repositories associated with the classroom are maintained in the organization. Students can act in their repositories the same way they do in any other repository on the GitHub platform, by performing typical operations such as: clone, commit, pull, and push. They can also document and manage their projects by creating {\it wiki pages} and {\it project boards} made up of issues, pull requests, and notes.
The novelty of this approach is using GitHub classroom not just on a course, but in all project-based courses within the scope of the undergraduate program in a shared, structured, and institutionalized way. The GitHub organization belonging to the program is central. An organization is a namespace where all student's repositories exist. It provides tools to manage subgroups of people with shared ownership of repositories. While the classroom is specific to a course and is archived at the end of the semester, the repositories stored in the organization persist active overtime. They can be maintained by other students, teachers and provided to clients or other stakeholders. Besides each professor having the possibility to analyze the learning outcomes of his/her students over the semester, an analysis of all students enrolled in the undergraduate program is also possible. From the organization, coordinators of the program can take a snapshot of classwork produced by students in all courses in each semester.
\section{Report of the first experience}
\label{sec:experience}
On the first semester of 2020 we decided to run a pilot of our approach based on GitHub Classroom in one of the eight project-based courses. We chose the course on the fourth semester of the curriculum because of a few reasons: 1) students are mature enough and really understand the good practice of sharing a common repository among the team members; 2) in this course there are real clients for the projects and at the end students have to hand in the complete project to the client; and, 3) in this course a project in one semester is usually continued by another team in the following semester.
Considering all these reasons, starting with the fourth semester course was a good choice. The classrooms were created by the owner of our own organization in GitHub, one of the professors who also coordinates the program. Besides creating the classrooms, he also gave owner permissions to the supervisors. Following the approach discussed in Section~\ref{sec:approach}, students accepted the assignment and received a link to enter the repository automatically generated by GitHub classroom. Each team used its repository throughout the semester. This pilot experience was perceived as being successful by the supervising professors.
In the second semester of 2020 we decided to adopt our GitHub Classroom approach for all eight project-based courses. We had a meeting in the beginning of the semester where we explained to all supervising professors the dynamics of the tool. We also created a repository template to be used for all courses. Supervising professors could add to the template structure for each of their courses but they had to follow a primitive structure. In this structure we guarantee folders and files to be organized in such a way as to ensure that at the end of the semester we would find every information we may need on these projects, and that students understand the organization of the different types of artifacts produced in the course.
In this section, we detail the results of the experience by describing professor' perceptions, student' perceptions and lessons learned.
\subsection{Professor Perceptions}
We had a meeting at the end of the second semester of 2020 to hear how the faculty felt about using GitHub classroom. All 13 supervising professors who participated in the experience and used the proposed GitHub classroom approach attended the meeting. In general they reported that they want to learn more about the tool to be able to use more features in the projects, and to help them with marking, and providing feedback to the students. A recurring demand in the faculty is the definition of a set of metrics measured directly from the student's activity log in the repositories and that supports the assessment process.
\subsection{Student Perceptions}
We also conducted a survey with students to see what was their impression of using the repository environment set up in the course. We had 287 students enrolled in project-based courses. They all received the survey to answer and we received 58 answers back (approximately $20\%$).
From the 58 respondents, 44 ($75.86\%$) thought the experience was positive, 3 ($5.17\%$) thought the experience was negative, and 11 (18.97\%) had a neutral opinion. About the work in teams using GitHub as a repository, 42 ($72.41\%$) students said that was satisfactory, 12 ($20.69\%$) said it brought more complexity to the team and project dynamics, and 4 ($6.90\%$) said it was indifferent.
Asked to provide a general score to the experience, the averaged score was \textit{4.3 out of 5}, which we think is very good for a first experience. We also evaluate more deeply the resources used by the teams (Fig.~\ref{fig:resources}), students commitment and interest in using our GitHub-based approach (Fig.~\ref{fig:commitmentinterest}), students' perceived benefits, difficulties, and recommendations (Fig.~\ref{fig:perceptions}).
Figure~\ref{fig:resources} shows a histogram of the resources used by the teams during the project-based course with our GitHub Classroom approach. Most used resources were code repository and version control, branch control, and evolution of parallel versions. This is expected because our infrastructure has just been implemented and there is still a lot to be explored, in terms of the resources provided by GitHub classroom, GitHub organization and repositories, as we detail in Section~\ref{sec:ahead}.
\begin{figure*}[bht]
\centering
\includegraphics[width=11.1cm]{GH73.png}
\caption{Resources used by the teams. Total of 58 respondents, each being able to select more than one option.}
\label{fig:resources}
\end{figure*}
The students assessed their level of interest and commitment on a 5-points semantic differential scale ranging from not {\it committed/interested at all} to {\it fully committed/interested} (Figure~\ref{fig:commitmentinterest}). Overall, the results show that students are more likely to be very committed than not committed, with 76.9\% indicating well or fully committed. The equivalent tendency was observed for the level of interest, with $54,2\%$ indicating well or fully interested.
\begin{figure*}[htb]
\centering
\subfigure[Level of Commitment ]{\includegraphics[width=8.3cm]{GH33.png}}
\subfigure[Interest in Learning about GitHub]{ \includegraphics[width=8.3cm]{GH43.png}}
\caption{Students' self-reported commitment and interest. Total of 58 respondents, each being able to select only one option.}
\label{fig:commitmentinterest}
\end{figure*}
Finally, Figure~\ref{fig:perceptions} shows students' perceived benefits, difficulties, and overall recommendations. The major difficulty was lack of technical knowledge about the tool ($39.65\%$ of students). The major benefit was the shared repository used by the team in the organization ($79\%$ of the students). More than three quarters of students ($79\%$ of the students) recommended us to stick to the approach in the following semesters.
\begin{figure*}[hbt]
\centering
\subfigure[Difficulties, being each student able to select more than one option.]{\includegraphics[width=8.3cm]{GH53.png}}
\subfigure[Benefits, being each student able to select more than one option.]{\includegraphics[width=8.3cm]{GH63.png}}
\subfigure[Recommendation. Each student is able to select only one option.]{\includegraphics[width=8.2cm]{GH93.png}}
\caption{Students' perceptions and recommendations. Total of 58 respondents.}
\label{fig:perceptions}
\end{figure*}
In addition to these questions, students also had a space where they could write about the experience in general and provide suggestions. Some suggestions include: ``we would like to use Markdown to prepare the project documentation instead of Word''; ``I would rather use the repository tool that my team finds most appropriate''; ``My team was not able to integrate our repository directly with Heroku and that made it much harder to deploy in Heroku.''
\subsection{Lessons Learned}
From professors, to students, to coordinators and clients in the community, everyone saw an improvement from previous experiences. In general the majority found that the experience was good. From theses results we could observe the following factors to improve: 1) students felt the need for more training with the tool; 2) students need to understand more about configuration management to take full advantage of this proposal; and 3) students need to understand the benefits in using the tool as soon as they start their projects.
From these students answers we have decided to: 1) allow and motivate students to use Markdown for the project documentation which will allow professors to understand the contribution of the team to the project documentation in the same manner as we can do with code; 2) motivate the students to ensure they understand the benefits of using the tool, and to prepare more training; and 3) to make the repositories public to allow for the integration with other tools and environments.
\section{The Road Ahead: New Requirements and Improvements}
\label{sec:ahead}
Using code versioning platforms and repository management in the teaching-learning context has become increasingly popular in teaching software engineering~\cite{Feliciano:2016,Parizi2018,Angulo2019,Hsing:2019}, but also in teaching other skills~\cite{Fiksel2019}. In this work, we {\it use GitHub classroom not just on a course, but in all project-based courses in the undergraduate program in a shared, structured, and institutionalized way.} This approach allows us to provide students with project-based learning closer to the software development approach practiced in the industry~\cite{garousi:2020}.
The first experiences provide us with insights on how this can evolve to enable new innovations in the teaching-learning process of software engineering. In this section, we focus on discussing these insights.
When applying a project-based learning methodology to courses with many students, monitoring the gain in knowledge and the development of skills and attitudes by each student is challenging. For example, a student may hide behind his team and it is hard to monitor the contributions of each student. This monitoring is carried out by professors to conduct formative assessment, provide individual feedback to each student on each team, and foster students' self-reflection~\cite{Baltes2018}. A standardization and centralization based on repositories, following project management practices, may {\it enable the automatic extraction of learning metrics and indicators to support analysis of knowledge, skills, and attitudes}. There are some efforts in this direction, focusing on measuring engagement, collaboration and quality of contribution to the team~\cite{Makiaho2015,Parizi2018}, mitigating possible inaccuracies and attempts at cheating~\cite{Tushev2020,Mohammad2020,Kun:2020}. However, there is a lack of convergence and standardization for metrics and learning indicators suitable for each stage of student training. On an undergraduate program in which each semester the student works on a project, changing the scope and increasing the complexity of the project, the definition of metrics and learning indicators suitable for each stage is relevant. To this end, it is necessary to {\it establish a theoretical framework covering knowledge, skills and attitudes associated with the students' maturity in each stage of a sequence of project-based courses}.
In addition to the other skills that are developed in the projects, but that are not directly associated with coding tasks, our approach of integrating GitHub Classroom, GitHub organizations and the students' GitHub in the different projects-based courses allows the student to obtain, in an iterative and incremental way, experiences with software licensing, collaboration in software development teams, configuration management, and continuous integration. {\it An educational tooling support yet to be fully explored and exploited is the integration with other apps/bots/extensions that can be used with GitHub repositories}, such as automating quality assurance procedures and software deployment activities. The use of this sort of tool is increasingly common in the industry ~\cite{lebeuf2017software,Lebeuf:2019} and is a demand we have received from students. They discover apps/bots developed through frameworks like ProBot\footnote{Available at: https://probot.github.io/apps/ (Accessed: 07 January 2021)} that they would like to use in their projects and request our authorization to use them in the organization's repositories. We believe there is room for using, testing, and developing tools to support software engineering activities within this context, directly as a tool for the GitHub platform or as an external tool integrated via the application programming interface (API) provided by that platform.
Our approach for managing student's repositories in project-based courses allows the {\it creation of the students' portfolio verified by professors during the undergraduate program}.
Such a portfolio is of academic and professional importance to students~\cite{cheng2019innovative,olaniyi2020survey}. Nowadays, GitHub profiles act as a portfolio signaling technical expertise and experience of developers~\cite{Feliciano:2016,Hsing:2019,Wang:2020}. It is a relevant factor in the hiring process at many companies. In this way, the student's skills are demonstrated in the projects in which she/he worked. Of course, as the learning process is continuous and incremental, a portfolio that describes the skills already acquired can also point out which new skills and experiences are yet to be developed. This can become a new tool and parameter of analysis to be considered by both students and professors when defining the next steps on the learning path.
To summarize, we advocate that: (1) using tooling support is essential in the teaching-learning process in Software Engineering when students, professors, program coordinators, and real clients are involved; and (2) building a shared, structured, and persistent repository in project-based courses makes it possible to integrate these stakeholders and support education by employing metrics to monitor skill development, conducting formative assessment, providing individual feedback in teamwork, and verifying student work portfolios.
\section{Author Profiles}
{\bf Maria Augusta Nelson} received her Ph.D. degree in Computer Science from the University of Waterloo, Canada, in 2003. She is an Associate Professor at the Pontifical Catholic University of Minas Gerais (PUC Minas), in Brazil, where she has been teaching Software Engineering, Software Testing, and Software Design courses since 2004. She was part of the team that conceived the Software Engineering Undergraduate Program and is now in charge of the program. She leads the Structuring Teaching Core, which is a group of professors responsible for conceiving, checking, questioning, and improving the pedagogical structure of the program. Her research interests are in computer education, focusing in software engineering education, requirements engineering and curriculum design involving also distance and continuing education. She is a member of the Brazilian Computer Society, participating in the Computer Education Interest Group (GIEC). She has served in the program committee of the IEEE/ACM 1ST International Workshop On Software Engineering Curricula for Millennials (SECM), and the second International Workshop On Software Engineering Education for Millennials (SEEM). Further information is available on her ORCID at https://orcid.org/0000-0002-1151-1362.
{\bf Lesandro Ponciano} received a Ph.D. degree in Computer Science from the Federal University of Campina Grande, Brazil, in 2015. He is an Associate Professor at the Pontifical Catholic University of Minas Gerais (PUC Minas), in Brazil, where he has been teaching Requirements Engineering, Software Testing, and Human-Computer Interaction courses since 2016. In the Software Engineering program of PUC Minas, he is also a member of the Structuring Teaching Core, which is a group of professors who are responsible for conceiving, checking, questioning, and improving the pedagogical structure of the program. His research interests are in computer education, human-computer interaction, and software engineering, focusing on engagement, credibility, coordination, and learning analysis. He is a member of the Brazilian Computer Society, participating in the Computer Education Interest Group (GIEC). He is certified as a GitHub Campus Advisor\footnote{Available at: https://education.github.com/teachers/advisors Accessed: 07 January 2021)}, which is a program part of the GitHub Education. Further information is available on his ORCID at https://orcid.org/0000-0002-5724-0094.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
The notion of vacuum state in quantum field theory is observer dependent \cite{Davies:1984rk}. The importance of this basic fact was only started to be appreciated with the discovery of black hole quantum radiance \cite{Hawking:1974sw,Unruh:1976db}, which is one of its main consequences. The existence of a temperature and thermal features associated to the horizon of a black hole is directly related to two of the major open issues in quantum gravity: the fate of unitarity in the quantum evolution of black holes \cite{Hawking:1976ra,Mathur:2009hf,Giddings:2019vvj} and the nature of their Bekenstein-Hawking entropy \cite{Bekenstein:1973ur,Carlip:2008rk}, and it is at the basis of the interpretation of gravity as a ``thermodynamics of space-time" \cite{Jacobson:1995ab}.
The ambiguity in the choice of the vacuum state in curved space-time relies on the possibility of having different choices of time-like Killing vectors that can be used to decompose the space of solutions of the classical field equation into positive and negative energy subspaces \cite{Ashtekar:1975zn}. The former will define, together with a positive definite inner product, the one-particle Hilbert space for static observers with respect to the Killing vector which generates evolution in their proper time. From such Hilbert space the standard Fock space construction follows with the associated vacuum and multi-particle states \cite{Wald:1995yp}. The Hawking effect boils down to the fact that the Hartle-Hawking vacuum state associated to freely falling observer is a thermal state at the Hawking temperature for static observers with respect to Schwarzschild time whose associated Killing vector becomes null at the horizon.
In maximally symmetric spaces the quantization ambiguity associated to different possible choices of time evolution is eliminated by the requirement of invariance of the vacuum state under the full group of isometries. In Minkowski space-time this singles out the standard vacuum state associated to inertial time evolution under the ordinary time translation Killing vector. However, as it is well known, one can also consider time evolution generated by a boost Killing vector which is time-like in the Rindler wedges and whose orbits correspond to worldlines of uniformly accelerating observers. For such observers the inertial vacuum is populated by a thermal distribution of particles at a temperature proportional to the magnitude of their four-acceleration: the so-called Unruh temperature \cite{Unruh:1976db}.
Quite interestingly, if one restricts the attention to scale-invariant fields the range of possible candidates for generators of time evolution can be extended to include {\it conformal Killing vectors}. In Minkowski space-time this has recently been used \cite{Higuchi:2017gcd,Wald:2019ygd} (see also \cite{Crispino:2007eb,Olson:2010jy}) to consider ``Milne quantization" of a two-dimensional massless scalar field in the future and past cones, analogous to Rindler quantization in the Rindler wedges, where time evolution is generated by the dilation conformal Killing vector. Observers along the radial orbits of dilations are comoving observers in a Milne universe (which is just Minkowski space-time described in expanding coordinates covering only the future cone) which perceive the inertial vacuum as populated by a thermal distribution of Milne particles.
These considerations suggest that in higher dimensions, and here we refer for simplicity to the $3+1$ dimensional case, one could define time evolution in terms of conformal Killing vectors which are time-like in certain regions of Minkowski space-time and similar thermal effects should manifest. Of particular interest are, to this extent, {\it radial conformal symmetries}, whose full classification for Minkowski space-time has been given in \cite{RCM}. Radial conformal Killing vectors divide Minkowski space-time into causal domains separated by light-like surfaces which are {\it conformal Killing horizons}, they are analogous to stationary spherically symmetric Killing horizons of black hole geometries and have similar thermodynamic properties \cite{DeLorenzo:2017tgx, DeLorenzo:2018ghq}. Among such radial conformal Killing vectors are the ones which map a causal diamond into itself and whose orbits describe worldlines of observers confined within the diamond. The fact that these diamond observers should perceive the inertial Minkowski vacuum as a thermal state was first suggested in \cite{Martinetti:2002sz} where the claim was supported with arguments which used the so-called ``thermal time hypothesis''. Another attempt to provide evidence of such diamond temperature was given in \cite{Su:2016wop} using a model of Unruh-DeWitt detector moving along the diamond observer trajectory.
In \cite{Arzano:2020thh} the present author pointed out that radial conformal symmetries are in correspondence with generators of time evolution in conformal quantum mechanics, a $0+1$ dimensional conformal field theory \cite{deAlfaro:1976vlx}. Using this correspondence, for which the ``wordline quantum mechanics" of diamond observers at the origin\footnote{This result has analogies with the worldline quantum mechanics description of the temperature perceived by a static patch observer in de Sitter space \cite{Anninos:2011af,Nakayama:2011qh}} is viewed as a conformal quantum mechanics, it was shown that the Minkowski vacuum is perceived by diamond observers as a thermal state at a temperature inversely proportional to the size of the diamond. The results of \cite{Arzano:2020thh}, besides providing a simple but powerful tool for studying the thermodynamics of causal regions in Minkowski space-time, also show that the possibility of selecting different generators of time evolution in conformal quantum mechanics leads to thermal effects in complete analogy with higher dimensional quantum field theory. Since the focus of \cite{Arzano:2020thh} was on causal diamonds, and the corresponding hyperbolic time evolution restricted to a finite time domain in conformal quantum mechanics, the natural question is whether similar thermal effects can be associated to other choices of time evolution like, for example, the one determined by the generator of dilations, and what is the physical interpretation of the associated temperature in Minkowski space-time.
In this work we show that, besides the ``diamond temperature" associated to time evolution restricted to a finite time domain of conformal quantum mechanics, there is a natural temperature associated to time evolution generated by dilations and restricted to positive (or negative) times. These two thermal effects have their origin in the {\it thermofield double} structure of the vacuum state associated to the ``inertial" time evolution generator which is here evidenced using purely group theoretical arguments. The thermal effect associated to time evolution restricted to half of the time line indicates that in Minkowski space-time static Milne observers, i.e. observers at the origin whose proper time evolution is determined by dilations, perceive the Minkowski vacuum as a thermal state. Such result provides evidence for the existence of a ``Milne temperature" for space-time dimensions higher than two and exhibits its intimate connection with diamond temperature given that, as it shown below, the two pictures are connected on the conformal quantum mechanics side by a conformal transformation. Finally, since both the diamond and Milne temperatures are non-vanishing along the worldlines of static observers sitting at the origin with vanishing acceleration, our result demonstrates that the fundamental ingredient for quantum vacuum thermal effects is the presence of a horizon, in this case associated to the non-eternal nature of the lifetime of the observer, rather than a non-vanishing acceleration.
\section{Radial conformal Killing vectors in Minkowski space-time and time evolution in conformal quantum mechanics}
Let us start from the line element of Minkowski space-time written in spherical coordinates
\begin{equation}\label{minksph}
d s^2 = -d t^2 + d r^2 + r^2 d \Omega^2\,,
\end{equation}
where $d \Omega^2 = d \theta^2 + \sin^2 \theta\, d\phi^2$. The most general radial conformal Killing vector $\xi$ for which $\mathcal{L}_{\xi} \eta_{\mu\nu} \propto \eta_{\mu\nu}$ where $\mathcal{L}_{\xi}$ is the Lie derivative and $\eta_{\mu\nu}$ is the Minkowski, metric is given by \cite{RCM}
\begin{equation}\label{xi}
\xi = a K_0 + b D_0 + c P_0\,,
\end{equation}
where $K_0$, $D_0$ and $P_0$ generate, respectively, special conformal transformations, dilations and time translations and close the $\mathfrak{sl}(2,\mathbb{R})$ Lie algebra
\begin{equation}\label{sl1}
[P_0,D_0]= P_0\,,\qquad [K_0,D_0]= - K_0\,,\qquad [P_0,K_0]= 2 D_0\,.
\end{equation}
These generators can be written in terms of radial and time derivatives as
\begin{eqnarray}\label{}
P_0 & = & \partial_t\\
D_0 & = & r\, \partial_r + t\, \partial_t\\
K_0 & = & 2 t r\, \partial_r + (t^2+r^2)\, \partial_t \,,
\end{eqnarray}
so that \eqref{xi} becomes
\begin{equation}\label{xi2}
\xi = \left(a(t^2+r^2)+b t +c\right)\, \partial_t + r (2 a t + b)\, \partial_r\,,
\end{equation}
and in light-cone coordinate $u=t+r$ and $v=t-r$ we have
\begin{equation}
\xi = (a u^2 + b u + c)\, \partial_u + (a v^2 + b v + c)\, \partial_v\,.
\end{equation}
In conformal quantum mechanics the generators $K_0$, $D_0$ and $P_0$ play the role of different time evolution operators. In terms of the {\it inertial} time variable $\tau$ associated to the generator $P_0$ they can be written as differential operators
\begin{eqnarray}\label{getau}
P_0 &=& \partial_{\tau}\\
D_0 &=& \tau\, \partial_{\tau}\label{dzee}\\
K_0 &=& \tau^2\, \partial_{\tau}\,.
\end{eqnarray}
The first important observation is that radial conformal Killing vectors for static observers at the origin $r=0$ and on the light-cones $u=const.$ or $v=const.$ coincide with the most general time evolution operator in conformal quantum mechanics
\begin{equation}
G = a K_0 + b D_0 + c P_0 = (a \tau^2 + b \tau + c)\, \partial_{\tau}\,.
\end{equation}
The causal structure of the Killing vector \eqref{xi2} in Minkowski space-time has been studied in \cite{RCM}. The quantity $\Delta = b^2 - 4 ac$ characterizes the norm of the Killing vector and is used to classify different types of generators. The generator $P_0$ for which $\Delta < 0$ is everywhere time-like and in conformal quantum mechanics generates time evolution in terms of ordinary translation spanning the whole real (time) line. The generator $D_0$ with $\Delta > 0$ and $a=0$ is time-like inside the light-cone emanating from the origin, null on the light-cone and space-like outside. In conformal quantum mechanics it describes time evolution in terms of dilations restricted to the positive or negative real line. In Minkowski space-time it describes {\it conformal time} evolution in a Milne universe\footnote{The Milne universe being flat can be indeed mapped to the future (or past, inverting the time direction) inner light cone of Minkowski space-time. It corresponds to a slicing of the latter in terms constant time hyper-surfaces given by Euclidean hyperboloids, orbits of boosts.} \cite{RCM,Olson:2010jy,Wald:2019ygd}. The other relevant generator is the radial conformal Killing vector which maps a causal diamond, i.e. the region $|t|+|r|<\alpha$, into itself given by
\begin{equation}\label{zeta}
S_0 = \frac{1}{2} \left(\alpha P_0 - \frac{K_0}{\alpha}\right) = \frac{1}{2 \alpha} \left((\alpha^2-t^2-r^2) \partial_t - 2 t r \partial_r \right)\,.
\end{equation}
The causal structure of this Killing vector for which $\Delta > 1$ and $a \neq 0$ is more articulated. It is null on the light-cones emanating from $r=0$ at times $\alpha$ and $-\alpha$. It is time-like inside or outside both light-cones, i.e. within the {\it causal diamond} of radius $\alpha$ and its causal complement, and space-like everywhere else. In conformal quantum mechanics the combination $S_0$ generates time evolution restricted to the finite domain $t\in (-\alpha, \alpha)$, the analogue of the {\it diamond} on the real line.
Notice that the quantity $\Delta = b^2 - 4 ac$ is proportional to the determinant of $G$ seen as a $2\times 2$ matrix and thus it is invariant under the adjoint action of $SL(2,\mathbb{R})$ on its Lie algebra. Such determinant can be used to classify different types of transformations generated by $G$. In particular while $P_0$, for which $\Delta < 0$, generates {\it parabolic} transformations of the Euclidean plane (null rotations if we think of the relation between $SL(2,\mathbb{R})$ and the three dimensional Lorentz group $SO(2,1)$), $D_0$ and $S_0$, for which $\Delta > 0$, generate $SL(2,\mathbb{R})$ elements belonging to the class of {\it hyperbolic} transformations whose representatives on the Lorentz group side are boost transformations. Thus we can map one into another using an $SL(2,\mathbb{R})$ transformation.
In order to do so let us first introduce time coordinates adapted to the evolution defined by the generators $D_0$ and $S_0$ i.e. such that
\begin{eqnarray}
\frac{D_0}{\alpha} &= &\partial_{\nu}\label{adnusig}\\
\frac{S_0}{\alpha} &=& \partial_{\sigma}\label{adnusig2}\,,
\end{eqnarray}
defining the Milne and diamond time for static observers with semi-infinite and finite lifetimes respectively. We can easily derive the transformation between the inertial time $\tau$ and the Milne and diamond time coordinates $\nu$ and $\sigma$ comparing \eqref{adnusig} and \eqref{adnusig2} with the expressions \eqref{dzee} and \eqref{zeta}. One obtains
\begin{eqnarray}
\nu & = & \alpha \log{|\tau|} + C_1\\
\sigma & = & \alpha \log{\frac{|\tau+\alpha|}{|\tau-\alpha|}} + C_2\,.
\end{eqnarray}
where the integration constants $C_1$ and $C_2$ can be determined by requiring dimensional consistency. We set
\begin{eqnarray}
C_1 & = & - \log{2\alpha}\\
C_2 & = & 0\,,
\end{eqnarray}
the reason for the choice $2 \alpha$ for the argument of the $\log$ will be clear later. We thus have for the Milne time
\begin{eqnarray}
\tau = &2\alpha\, \exp{\frac{\nu}{\alpha}}& \qquad \text{for} \,\, \tau>0\label{ttaup}\\
\tau = & - 2\alpha\, \exp{\frac{\nu}{\alpha}}& \qquad \text{for} \,\, t<0\,.
\end{eqnarray}
We see that on the positive half-line $\tau>0$ as $\nu$ ranges from $-\infty$ to $\infty$ we go from $\tau=0$ to $\tau=\infty$. On the negative half line the time flow is inverted and we go from $\tau=0$ to $\tau=-\infty$ as $\nu$ ranges from $-\infty$ to $\infty$. Notice that, even though there is no natural time scale associated with the future cone, a constant with dimensions of time is needed in order to define a time evolution operator in terms of the generator of dilations $\tau\, \partial_\tau$. This time scale is the point on the real line measured in time $\tau$ corresponding to the origin $\nu=0$ of Milne time.\\
For the $S$-time variable $\sigma$ we have that in the region $|\tau|<\alpha$, i.e. within the diamond, i.e.
\begin{equation}
\tau = \alpha\, \tanh{\sigma/2\alpha}
\end{equation}
while for $|\tau| > \alpha$ one has
\begin{equation}
\tau = \alpha \coth{\sigma/2\alpha}\,.
\end{equation}
We see that within the diamond as the diamond time ranges from $-\infty$ to $\infty$ we cover the region from $t=\alpha$ to $t=-\alpha$.
Let us now consider a new time coordinate $\tau'(\tau)$ in terms of which the dilation generator $D = \tau\, \partial_\tau$ acts as the generator $S$ of conformal transformations which preserve the diamond, i.e.
\begin{equation}
\tau\, \partial_\tau = \frac{1}{2 \alpha} \left(\alpha^2- \tau'^2\right) \partial_{\tau'}\,.
\end{equation}
Such equation can be easily solved leading to
\begin{equation}
\log{\frac{|\tau'+\alpha|}{|\tau'- \alpha|}} = \log{|\tau|} + C \,.
\end{equation}
For dimensional consistency we choose for the integration constant $C= -\log{A}$, where $A$ has dimensions of time and thus must be proportional to $\alpha$. Restricting to the positive half-line $\tau > 0$ we have
\begin{equation}\label{ttpa}
\tau = A\,\, \frac{\tau'+\alpha}{\tau'-\alpha}
\end{equation}
In order to determine the constant $A$ we first notice that the point $\tau'=-\alpha$ gets mapped to $\tau=0$ while $\tau=\alpha$ to $\tau = \infty$. The origin of the diamond $\tau'=0$ gets mapped to $\tau= -A$ so we know that $A$ must be negative. Let us further notice that the origin of the diamond proper time $\sigma=0$ corresponds to $\tau'=0$. Under the map \eqref{ttpa} this is mapped to $\tau=-A$. Now we require that the origins of the diamond and Milne proper times are ``syncronized" i.e. the point corresponding to $\sigma=0$ in $\tau'$ coordinates is mapped to the point corresponding to $\nu=0$ in $\tau$-coordinates and thus, from \eqref{ttaup},
\begin{equation}
2 \alpha = -A\,.
\end{equation}
We thus have
\begin{equation}\label{DtoS}
\tau = - 2 \alpha\, \frac{\tau'+\alpha}{\tau'-\alpha}
\end{equation}
which can be inverted to give
\begin{equation}\label{StoD}
\tau' = \alpha\, \frac{\tau - 2\alpha}{\tau + 2 \alpha}\,.
\end{equation}
Notice how this transformations correspond to the map from the causal diamond to the Rindler wedge and viceversa \cite{Casini:2011kv, Jacobson:2018ahi} when written in radial light-cone coordinates, indeed the Rindler wedge and the future cone meet on the light cone.
\section{The thermofield double of conformal quantum mechanics}
As shown in \cite{Chamon:2011xk,Jackiw:2012ur} conformal quantum mechanics can be seen as a one-dimensional conformal field theory in which correlation functions are built from eigenstates of the time translation operator. Let us recall the main features of this construction. In conformal quantum mechanics we work with the ``quantum mechanical" counterparts of the generators introduced in the previous section i.e. $H = iP_0$, $D= iD_0$ and $K=i K_0$ closing the algebra
\begin{equation}\label{sl2}
[H,D]=i H\,,\qquad [K,D]= - i K\,,\qquad [H,K] = 2 i D\,.
\end{equation}
Let us look for states $|\tau \rangle$ on which the $SL(2, \mathbb{R})$ generator $H$ acts as a $\tau$-derivative
\begin{equation}
H |\tau \rangle = -i\, \partial_\tau\, |\tau \rangle\,.
\end{equation}
These states were first constructed in \cite{deAlfaro:1976vlx} starting from an irreducible representation of the $\mathfrak{sl}(2,\mathbb{R})$ Lie algebra. Upon defining the operators
\begin{equation}
L_\pm \equiv \frac{1}{2}\, \left(\frac{K}{\alpha} - \alpha\, H \right)\, \pm i\, D,
\end{equation}
and
\begin{equation}
L_0 \equiv \frac{1}{2}\, \left(\frac{K}{\alpha} + \alpha\, H \right)\,,
\end{equation}
with commutation relations
\begin{equation}\label{eq:su11Alg}
[L_-,L_+] = 2 L_0\,, \quad [L_0, L_{\pm}] = \pm L_{\pm}\,,
\end{equation}
such representation is realized in terms of the eigenstates $\ket{n}$ of $L_0$
\begin{subequations}\label{rjAdS6}
\begin{gather}
L_0 \ket{n} = r_n\, \ket{n}\label{rjAds6-a}\\
r_n = r_0 + n,\ \ r_0 > 0, \ \ n = 0,1 \ldots \nonumber\\
\braket{n}{n^\prime} = \delta_{n\, n^\prime} \nonumber \\
L_\pm \ket{n} = \sqrt{r_n \, (r_n \pm 1) - r_0 \, (r_0 -1)} \, \ket{n \pm 1} \label{rjAds6-b}
\end{gather}
\end{subequations}
where $r_0$ is related to the eigenvalue of the Casimir operator of the $\mathfrak{sl}(2,\mathbb{R})$ algebra
\begin{equation}
\mathcal{C}= \frac{1}{2}\left(K H + H K \right) - D^2 = r_0 (r_0-1)\,.
\end{equation}
Notice that dimensional consistency requires the introduction of a scale $\alpha$ with the dimension of time in the definition of the operators $L_+$ and $L_-$ and, in light of the correspondence with conformal symmetries in Minkowski space-time, we can give a geometric interpretation to such scale, following the discussion in the previous section, in terms of the radius of a causal diamond whose boundaries are invariant under the flow of the conformal Killing vector $S_0$.
In \cite{Chamon:2011xk,Jackiw:2012ur} it was found that the $|\tau \rangle$ states can be obtained from the $ \ket{n=0} $ ``vacuum state" via
\begin{equation}\label{opcft}
|\tau \rangle = \mathcal{O}(\tau) \ket{n=0} = N(\tau)\exp(-\omega(\tau)L_+) \ket{n=0}\,,
\end{equation}
where
\begin{eqnarray}
N(\tau) &=& [\Gamma(2r_0)]^{1/2}\left(\frac{\omega(\tau)+1}{2}\right)^{2r_0} \ , \nonumber \\
\omega(\tau) &=& \frac{a+i\tau}{a-i\tau} \ .
\end{eqnarray}
The action of the other generators on such $\tau$-states are given by
\begin{eqnarray}
D \ket{\tau} &=& - i\, \left(\tau\, \frac{d}{d \tau} + r_0 \right) \ket{\tau},\\
K \ket{\tau} &=& -i \left(\tau^2\, \frac{d}{d \tau} + 2\, r_0\, \tau \right) \ket{\tau}\,.
\end{eqnarray}
In \cite{Chamon:2011xk} the two point functions of the theory are identified with the inner product between the $\tau$-states
\begin{equation}\label{2pf}
G_2 (\tau_1, \tau_2) \equiv \braket{\tau_1}{\tau_2} = \frac{\Gamma\, (2 r_0)\, \alpha^{2 r_0}}{[2 i\, (\tau_1 - \tau_2)]^{2 r_0}}\,,
\end{equation}
where $r_0$ plays the role of the conformal weight. General $n$-point functions with $n>2$ are obtained by inserting generic primary operators between $\langle \tau_1|$ and $|\tau_2\rangle$. When $\tau=0$ one has $\omega(\tau) = 1$ and thus
\begin{equation}\label{tze1}
|\tau =0 \rangle = \Gamma(2r_0)^{1/2}\, \exp(-L_+) \ket{n=0}
\end{equation}
One immediately sees from \eqref{2pf} that such $\tau$-vacuum is not normalizable due to the divergence of the two-point function for coincident points.
In \cite{Arzano:2020thh} it was noted that for $r_0 = 1$ the two-point function \eqref{2pf} is proportional to the two-point function of a free massless scalar field in Minkowski space-time evaluated along the trajectory of a static inertial observer at the origin. This fact was used to derive the diamond temperature\footnote{It should be noted that $H$ and $D$ form a sub-algebra of $\mathfrak{sl}(2,\mathbb{R})$ which generates the group of affine transformations of the real line, also known as the $ax+b$ group. In \cite{Arzano:2018oby} it was shown, using one-dimensional fields carrying a representation of the $ax+b$ group and Bogoliubov transformations between $H$-modes and $D$-modes, that the $H$-vacuum is seen as a thermal distribution of $D$-particles when restricted to the positive half line.} upon re-writing \eqref{2pf} in terms of the time variable associated to the diamond time evolution generator $S=iS_0$. Here we provide an alternative derivation of the same temperature by looking directly at the structure of the $\tau$-vacuum. This will lead to a unified description of the diamond and Milne temperatures.
Let us start from a simple observation which, nonetheless, is a key passage of our derivation. It is known from the quantum optics literature (see e.g. \cite{Ban:1992zz,Chaturvedi:1991zzb}) that the operators $L_{\pm}$ and $L_0$ with commutation rules \eqref{eq:su11Alg} can be realized in terms of {\it two sets} of creation and annihilation operators $a_L, a_L^\dagger, a_R, a_R^\dagger$ as
\begin{equation}\label{casl}
L_+=a_L^\dagger a_R^\dagger\,, \quad L_- = a_L a_R\,,
\quad L_0=\frac{1}{2}\left( a_L^\dagger a_L + a_R^\dagger a_R+1\right)\,.
\end{equation}
It immediately follows from \eqref{tze1} that the $\tau$-vacuum can be written as\footnote{From this point on we set $r_0=1$ as in \cite{Arzano:2020thh}.}
\begin{equation}\label{lasl}
|\tau =0 \rangle = \exp\left[ - a^\dagger_L a^\dagger_R\right]|0\rangle_L \otimes |0\rangle_R~,
\end{equation}
where we made explicit the bipartite structure of the vacuum state $\ket{n=0}$ emerging from the decomposition \eqref{lasl}
\begin{equation}
\ket{n=0} = |0\rangle_L \otimes |0\rangle_R\,.
\end{equation}
With simple manipulations we have that
\begin{equation}\label{ttens}
\begin{aligned}
|t =0 \rangle = & \sum_{n=0}^\infty \frac{(-1)^n}{n!}\,\left(a^\dagger_L a^\dagger_R\right)^n|0\rangle_L \otimes |0\rangle_R\\
= & \, \sum_{n=0}^\infty (-1)^n |n\rangle_L \otimes |n\rangle_R\,,
\end{aligned}
\end{equation}
where $|n\rangle_L$ and $|n\rangle_R$ are eigenstates of the ``left" and ``right" number operators $N_L= a^\dagger_L a_L $ and $N_R= a^\dagger_R a_R$. Our second key observation is that the Lie algebra
\begin{equation}\label{lzeropm}
[L_-,L_+] = 2 L_0\,, \quad [L_0, L_{\pm}] = \pm L_{\pm}\,,
\end{equation}
can be realized via another combination of $H$, $D$ and $K$, namely
\begin{equation}\label{lzeros}
L_0= i S\,,\qquad L_+= \frac{1}{2}\left(D-R\right)\,,\qquad L_-=2 \left(D+R\right)\,,
\end{equation}
where $R=\frac{1}{2}\, \left(\frac{K}{\alpha} + \alpha\, H \right)$ is generator of elliptic transformations corresponding to rotations in the three-dimensional Lorentz group. Under the identification \eqref{lzeros} we can now interpret the $|n\rangle$ states as eigenstates of the diamond time evolution generator $S$ and, in particular, the bipartite vacuum $\ket{n=0} = |0\rangle_L \otimes |0\rangle_R$ as the ``ground state'' of such Hamiltonian. In light of the correspondence with radial conformal symmetries in Minkowski space-time, we can think of the vacuum state $\ket{n=0}$ as the conformal quantum mechanics counterpart of the the vacuum state associated the generator of time evolution for static diamond observers, i.e. as the diamond analogue of the Boulware vacuum for Rindler observers.
Along the same lines we can think of the $\tau$-vacuum $|\tau =0 \rangle$ as the conformal quantum mechanics counterpart of the Hartle-Hawking vacuum of a massless scalar field in Minkowski space-time. Such connection is a bit more subtle and should be understood in terms of a state-operator correspondence illustrated in \cite{Chamon:2011xk}. As observed in \cite{Arzano:2020thh} the Minkowski light-cone two-point function maps to the $CFT_1$ two-point function constructed from the $|\tau \rangle$ states \eqref{2pf}. The latter can be written as
\begin{equation}
\langle \tau_1|\tau_2\rangle = \langle n=0 |\, \mathcal{O}(\tau_1)\, \mathcal{O}(\tau_2)\,| n=0 \rangle\,,
\end{equation}
with $\mathcal{O}(\tau)$ defined in \eqref{opcft}. The vacuum sate $| n=0 \rangle$ is not conformally invariant and the operators $\mathcal{O}(\tau)$ do not transform as primary operators under conformal transformations \cite{Chamon:2011xk}. However, as shown in \cite{Chamon:2011xk}, the combination
\begin{equation}
|\tau\rangle = \mathcal{O}(\tau)| n=0 \rangle
\end{equation}
satisfies a constraint which combines both the requests of the invariance of the vacuum with that of the operators $\mathcal{O}_{r_0}(t)$ to be primary. In other words the two point functions $\langle t_1|t_2\rangle$ behave as if they were constructed from a conformal invariant vacuum in analogy with the two-point function for massless scalar field in Minkowski space-time constructed from the Hartle-Hawking vacuum. Indeed it turns out \cite{Chamon:2011xk} that such two-point functions can be expressed as
\begin{equation}
\langle \tau_1|\tau_2\rangle = \langle \tau=0 | e^{-i (\tau_1 - \tau_2)H}| \tau=0 \rangle\,,
\end{equation}
i.e. a quantum mechanical counterpart of the Minkowski space-time two-point function with $| \tau=0 \rangle$ playing the role of the Hartle-Hawking state.
The analogy between the $\tau$-vacuum and the Hartle-Hawking state is actually deeper than it what might appear at first glance. Indeed the state $| \tau=0 \rangle$, like the Hartle-Hawking vacuum, exhibits the structure of a {\it thermofield double}. To see this let us recall the expression of $L_0$ in terms of creation and annihilation operators \eqref{casl}. Using the identification \eqref{lzeros}, from \eqref{ttens} we can write the $\tau$-vacuum as
\begin{equation}\label{tfdc1}
|\tau =0 \rangle = -i \, \sum_{n=0}^\infty e^{i\pi L_0}\, |n\rangle_L|n\rangle_R\, = -i \, \sum_{n=0}^\infty e^{-\pi S}\, |n\rangle_L|n\rangle_R\,.
\end{equation}
Such superposition of states is the same as the one appearing in the thermofield double state of a bosonic oscillator \cite{Khanna:2009zz,Chapman:2018hou}. Indeed, starting from the $\tau$-vacuum density matrix $\rho = |\tau =0 \rangle \langle \tau=0|$ and tracing over one set of $L$ or $R$ degrees of freedom one obtains a thermal density matrix at temperature $1/2\pi$ with respect to the modular Hamiltonian $S$. Now, of course, $S$ does not have the dimensions of a Hamiltonian, rather one should see the operator $\frac{S}{\alpha}$ as the modular Hamiltonian of the system so that the temperature associated to the thermofield double \eqref{tfdc1} is given by
\begin{equation}
T_{S} = \frac{1}{2\pi \alpha}\,,
\end{equation}
i.e. precisely the diamond temperature obtained in \cite{Arzano:2020thh}.
An interesting upshot of the analysis above is that it can be used to also give a group theoretical derivation of the Milne temperature. Indeed, under the conformal map \eqref{StoD} the identification \eqref{lzeros} leads to the following realization of the algebra \eqref{lzeropm}
\begin{equation}\label{lzerod}
L_0= i D\,,\qquad L_+= -\alpha H\,,\qquad L_-=\frac{K}{\alpha}\,.
\end{equation}
With such identification the vacuum state $\ket{n=0}$ is now seen as the conformal quantum mechanics analogue of the vacuum state associated to $D$, the generator of conformal time translations in the future cone, and the temperature
\begin{equation}
T_{D} = \frac{1}{2\pi \alpha}\,,
\end{equation}
is simply the temperature perceived by a Milne observer sitting at the origin of Minkowski space-time in the inertial vacuum. In two space-time dimensions, where null planes and null cones coincide and Milne and Rindler particles can de identified on the the light cone \cite{Wald:2019ygd}, our arguments give a unified, group theoretical derivation of the diamond, Milne and Unruh temperatures.
\section{Conclusions}
The considerations above show that the structure of conformal quantum mechanics is rich enough to reproduce, in the extremely simplified context of a one-dimensional model, the basic features which characterize vacuum thermal effects in quantum field theory, i.e. the appearance of the global vacuum as a thermal state populated of excitations associated to Hamiltonians whose evolution does not cover the entire time domain of the theory. Remarkably, using the correspondence between time evolution in conformal quantum mechanics and radial conformal symmetries in Minkowski space-time seen as generators of proper time, this result provides direct evidence for the fact that the inertial vacuum appears as a thermal state to observer whose time evolution is not {\it eternal}. The results presented suggest that the thermodynamic properties of causal diamonds and the Milne ``patch" of Minkowski space are intimately connected. It is tempting to speculate that the tools of conformal quantum mechanics which are used to provide evidence for such connection might provide an alternative and perhaps simpler way to describe entanglement features of quantum fields across the causal boundaries of such regions. A thorough exploration of such hypothesis is left to future studies.
|
\section{Introduction}
\label{sec:introduction}
Adversarial robustness of deep neural networks (DNNs) has received increasing attention in recent years.
Related studies include adversarial defense and attacks~\cite{szegedy2013intriguing,goodfellow2014explaining}.
In terms of defense, adversarial training is an effective and the most widely-used method~\cite{madry2018towards,zhang2019theoretically,wang2019dynamic,wang2020improving,wu2020adversarial}.
In spite of their fast development, the essential mechanism of the adversarial robustness is still unclear.
Thus, the understanding of adversarial attacks and defense is an emerging direction in recent years.
\citet{ilyas2019adversarial} demonstrated adversarial examples could be attributed to the presence of non-robust yet discriminative features.
Some methods~\cite{gilmer2018relationship,weng2018evaluating} explored the mathematical bound for the model robustness.
\citet{zhang2019interpreting,tsipras2018robustness} found adversarial training helped DNNs learn a more interpretable (more shape-biased) representation.
Besides the feature interpretability,
\citet{tsipras2018robustness} further showed an inherent tension between the adversarial robustness and the generalization power.
Unlike above perspectives for explanations,
we aim to propose a unified view to explain the essential reason why and how adversarial examples emerge, as well as essential mechanisms of various adversarial defense methods.
We rethink the adversarial robustness from the novel perspective of interactions between input variables of a DNN.
It is because in an adversarial example, adversarial perturbations on different pixels do not attack the DNN independently.
Instead, perturbation pixels usually interact with each other to form a specific pattern for attacking.
Surprisingly, we find that such interactions can explain various aspects of adversarial robustness.
Specifically, this study aims to answer the following three questions.
$\bullet$ \textbf{How to disentangle feature representations that are sensitive to adversarial perturbations.}
Based on the game-theoretic interactions, we aim to summarize the distinct property of feature representations, which are sensitive to adversarial perturbations, among overall feature representations.
$\bullet$ \textbf{How to explain the effectiveness of the adversarial training.}
The above summarized property of sensitive feature representations also provides a new perspective to explain the utility of the adversarial training, \emph{i.e.} why and how the adversarial training penalizes such sensitive feature representations.
$\bullet$ \textbf{How to unify various adversarial defense methods in a single theoretic system.}
Our research provides a unified understanding for the success of adversarial defense methods~\cite{DBLP:journals/corr/abs-1906-03499,devries2017improved,jere2020singular}.
\begin{figure}[t]
\centering
\begin{minipage}[t]{0.67\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/fig1-n.pdf}
\end{minipage}
\hfill
\begin{minipage}[t]{0.28\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/vis.pdf}
\end{minipage}
\caption{(Left) Adversarial attacks mainly affect high-order interactions in input samples. High-order interactions in adversarially trained DNNs are more robust than those in standard DNNs. (Right) Regions with strong low-order and high-order interactions, which are visualized by the method extended from \cite{zhang2020interpreting-dropout}.}
\label{fig:fig1}
\vspace{-10pt}
\end{figure}
As a prerequisite of analyzing the adversarial robustness, let us first revisit the interaction between input variables in DNNs.
Let a set of input variables collaborate with each other to form an inference pattern,
the Shapley interaction index~\cite{grabisch1999axiomatic} is a standard metric to measure the numerical benefits to the inference from their collaborations.
This metric can be further extended to the multi-order interaction~{\cite{zhang2020interpreting-dropout}}.
For the interaction between two input variables ($i,j$), the interaction order measures the number of contextual variables that influence the significance of the interaction between $i$ and $j$.
In other words, low-order interactions represent simple collaborations between input variables with small contexts, while high-order interactions indicate complex collaborations over large contexts (see Figure~\ref{fig:fig1} (right)).
We further prove that the network output can be decomposed into the sum of multi-order interactions between different pairs of input variables.
Thus, the overall effects of adversarial perturbations on the network output can be decomposed into elementary effects on different interaction components.
Therefore, we can explain adversarial robustness using such elementary interaction components.
(1)
\textbf{We discover and partially prove that adversarial perturbations mainly affect high-order interactions, rather than low-order interactions} (see Figure~\ref{fig:fig1} (left)). In comparison, low-order interactions are naturally robust to attacks.
In other words, adversarial attacks mainly affect the complex and large-scale collaborations among most pixels in the image.
Based on this, we can successfully disentangle sensitive feature representations, \emph{i.e.} high-order interactions.
\textit{Interaction vs. frequency \& rank.}
Some studies explained adversarial examples as high-frequency features~\cite{yin2019fourier,wang2020high,harder2021spectraldefense} and high-rank features~\cite{jere2020singular}.
We have conducted experiments to show that high-order interactions can better explain the essential property of attacking-sensitive representations, \emph{i.e.} the complex and large-scale visual concepts.
(2)
A clear difference between standard DNNs and adversarially trained DNNs is as follows.
\textbf{Adversarial training significantly increases the robustness of high-order interactions.}
In other words, attacks mainly affect complex collaborations in standard DNNs, while for adversarially trained DNNs, complex collaborations are not so vulnerable \emph{w.r.t.} simple collaborations.
Then, we further explain the reason for the high robustness of adversarially trained DNNs.
For interactions of each order, we define the disentanglement metric to identify whether interactions of this order are discriminative for the classification of a specific category, or represent common knowledge shared by different categories.
For example, in Figure~\ref{fig:fig1} (left), interactions representing the blue water may be shared by different categories, while interactions corresponding to the head of the red-breasted merganser are discriminative for this category.
Based on the disentanglement metric, we discover that \textbf{compared with standard DNNs, adversarially trained DNNs usually encode more discriminative low-order interactions.
Discriminative low-order interactions make high-order interactions of adversarially trained DNNs robust to attacks,} because contexts of high-order interactions are composed of many small contexts of low-order interactions.
For example, if the simple (low-order) interactions for the red-breasted merganser are learned to be discriminative, instead of being shared by the bicycle category, then it is difficult to attack this image towards the bicycle category.
It is because it is difficult to use the low-order interactions of the red-breasted merganser's head to construct high-order interactions of bicycles.
(3)
\textbf{Our research provides a unified understanding for the success of several existing adversarial defense methods}, including the attribution-based detection of adversarial examples~\cite{DBLP:journals/corr/abs-1906-03499}, the recoverability of adversarial examples to normal samples, the cutout method~\cite{devries2017improved}, and
the rank-based method~\cite{jere2020singular} (which is proved by \cite{jere2020singular} to be related to frequency-based methods~\cite{yin2019fourier,wang2020high,wu20143d}).
\textbf{Above findings also slightly revise the previous explanation of adversarially learned features}~\cite{goodfellow2014explaining,tsipras2018robustness,dong2017towards,zhang2019interpreting}.
They claimed that adversarial training learned more information about foreground shapes. We discover that these adversarially learned features are actually low-order interactions (usually local shapes), instead of modeling the global shape of the foreground.
{\textbf{Explainable AI system based on game-theoretic interactions.}
In fact, our research group led by Dr. Quanshi Zhang have proposed game-theoretic interactions, including interactions of different orders~\cite{zhang2020game} and multivariate interactions~\cite{zhang2021interpreting}.
The interaction can be used as an typical metric to explain signal processing in DNNs from different perspectives.
For example, {the game-theoretic metric} can be used to guide the learning of baseline values of Shapley values~\citep{ren2021learning}.
Furthermore, we have built up a tree structure to explain the hierarchical interactions between words encoded by NLP models~\cite{zhang2021building}.
We have also used interactions to explain the generalization power of DNNs~\cite{zhang2020interpreting-dropout}.
The interaction can also explain how adversarial perturbations contribute to the attacking task~\cite{wang2021interpreting}, and explain the transferability of adversarial perturbations~\cite{wang2020unified}.
Furthermore, we have also used the interaction to formulate the visual aesthetics~\cite{cheng2021hypothesis} and signal-processing properties of different types of visual concepts~\cite{cheng2021game} in DNNs.
As an extension of the system of game-theoretic interactions, in this study, we explain the adversarial robustness based on interactions.}
\section{Related work}
\label{related_work}
Attacking methods can be roughly summarized into
white-box attacks~\cite{szegedy2013intriguing,goodfellow2014explaining,kurakin2016adversarial,papernot2016limitations,carlini2017towards} and black-box attacks~\cite{liu2016delving,papernot2017practical,chen2017zoo,bhagoji2018practical,ilyas2018black,wu2020skip,bai2020improving}.
For defense, adversarial training is one of the most effective and widely-used defense methods~\cite{madry2018towards,zhang2019theoretically,wang2019dynamic,wang2020improving,wu2020adversarial}.
Other defense methods include masking gradients~\cite{papernot2017practical,nayebi2017biologically}, modifying networks~\cite{cisse2017houdini,gao2017deepcloak}, and applying pre-processing on input images for testing~\cite{das2017keeping,meng2017magnet,xie2019feature,bai2019hilbert}.
\textit{Explanations for adversarial examples.}
Some previous studies focused on the reason for the existence of adversarial examples.
\citet{goodfellow2014explaining} explained adversarial examples as a result of the high linearity of feature representations.
\citet{gilmer2018relationship,ma2018characterizing} proved that the existence of adversarial examples was due to the geometry of the high-dimensional manifold.
\citet{xie2019feature,xu2018structured,xu2019interpreting,bai2021improving} discovered that adversarial perturbations usually activated substantial ``noise'' and semantically irrelevant features.
\citet{engstrom2019exploring} investigated the vulnerability of DNNs to rotations and translations.
\citet{tsipras2018robustness,ilyas2019adversarial} demonstrated that adversarial examples were attributed to non-robust yet discriminative features.
\textit{Understandings of adversarial training.}
\citet{athalye2018obfuscated} suggested that adversarial training did not cause the obfuscated gradients phenomenon, which boosted the robustness.
\citet{goodfellow2014explaining,dong2017towards,tsipras2018robustness,zhang2019interpreting} found that adversarially trained DNNs learned more shape-biased features than standard DNNs.
\citet{chalasani2020concise} proved that adversarially trained DNNs \emph{w.r.t.} the $\ell_\infty$ attack exhibited more sparse attributions.
\citet{tian2021analysis} showed that there is robustness imbalance among classes in adversarial training.
\citet{song2018constructing} considered the adversarial training as the enumeration of all potential adversarial perturbations.
\citet{wang2019dynamic} explained adversarial training from the perspective of min-max optimization.
\citet{yin2019fourier,wang2018analyzing,harder2021spectraldefense} discovered adversarial training pushed DNNs to utilize low-frequency components in inputs.
\textit{Understanding of the robustness.}
\citet{szegedy2013intriguing,hein2017formal} computed Lipschitz constant to explain the robustness.
\citet{ignatiev2019relating,boopathy2020proper} investigated the connection between network interpretability and adversarial robustness.
\citet{tsipras2018robustness} proved the inherent tension between adversarial robustness and standard generalization power.
\citet{fawzi2018analysis,gilmer2018relationship,weng2018evaluating} proved lower/upper bounds on the robustness.
\citet{pal2020game} proposed a game-theoretic framework to understand the adversarial robustness, by formulating the game between attackers and defenders \emph{w.r.t.} a trained DNN. Under the game-theoretic framework, \citet{pal2020game} theoretically proved that the FGM attack and the random smoothing defense formed a Nash Equilibrium under some assumptions.
Unlike previous explanations of the existence of adversarial examples and adversarial robustness, we explain adversarial examples and adversarial training from a new perspective, \emph{i.e.} the complexity of visual concepts that are learned by a DNN.
\section{Decomposing attacking utility into interactions of multiple orders}
\label{sec:interaction_shapley}
\textbf{Preliminaries: using Shapley values~\cite{shapley1953value} and the Shapley interaction index~\cite{grabisch1999axiomatic} to explain a DNN.}
\label{sec:shapley}
The Shapley value~\cite{shapley1953value} in game theory is widely considered as an unbiased estimation for the importance or contribution of each player in a game.
Given a trained DNN and the input with $n$ variables (\emph{e.g.} an image with $n$ pixels, a sentence with $n$ words) $N=\{1,\cdots,n\}$, we can take input variables as players, and consider the network output as the reward.
Shapley values can fairly divide and assign numerical effects on the network output to each input variable.
More specifically, let $2^N\overset{\text{def}}{=} \{S|S\subseteq N\}$ denote all potential subsets of $N$, and each subset $S$ represents a specific context (\emph{e.g.} a set of pixels in an image).
$v(S)\in\mathbb{R}$ represents the scalar network output when we keep variables in $S$ unchanged and mask variables in $N\setminus S$ by following settings in \cite{ancona2019explaining} (\emph{i.e.} setting a masked variable to the average value over different input samples).\footnote{For the DNN trained for multi-category classification, we set ${v(S)=\log p(y=c|}\text{given an input with variables in}~N\setminus S~\text{masked})$, where $c$ is selected as either the ground-truth category of the input, or the incorrectly classified category after attacks.}
In particular, $v(N)$ refers to the network output \emph{w.r.t.} the entire input $N$ (\emph{e.g.} the whole image), and $v(\emptyset)$ denotes the output when we mask all variables.
In this way, the Shapley value $\phi(i)$ unbiasedly measures the importance of the variable $i$ to the network output.
\vspace{-5pt}
\begin{equation}
\begin{small}
\phi(i) ={\sum}_{S\subseteq N\setminus \{i\}} p(S) \left[v(S\cup \{i\})-v(S)\right],\quad p(S)\overset{\text{def}}{=}\frac{(n-|S|-1)! |S|!}{n!}
\end{small}
\label{eq:shapley_value}
\end{equation}
In this way, the network output can be considered as the sum of elementary importances of input variables, \emph{i.e.} $v(N)=\sum_{i\in N}\phi(i)+v(\emptyset)$.
The Shapley value has been proved to satisfy four desirable properties, \emph{i.e.} \textbf{\textit{linearity, nullity, symmetry}} and \textbf{\textit{efficiency}} properties, thereby being regarded as a fair method to allocate the network output to each input variable~\cite{weber1988probabilistic}.
Input variables of a DNN do not contribute to the network output independently.
Instead, different variables collaborate with each other to affect the network output, and we use the interaction to quantify the numerical utility of such collaborations.
To this end, the Shapley interaction index~\cite{grabisch1999axiomatic} measures the utility of the collaboration between two input variables ($i,j$) by examining \textbf{whether the absence/presence of an input variable $j$ will change the importance of the other variable $i$.}
Their interaction is defined as
$I(i,j)=\tilde{\phi}(i)_{\text{$j$ always present}} - \tilde{\phi}(i)_{\text{$j$ always absent}}$,
where $\tilde{\phi}(i)_{\text{$j$ always present}}$ and $\tilde{\phi}(i)_{\text{$j$ always absent}}$ denote the importance of the variable $i$ when $j$ is always present and when $j$ is always absent, respectively (please see \cite{grabisch1999axiomatic} for more details about $I(i,j)$).
\begin{figure}[t]
\centering
\begin{minipage}[b]{0.67\linewidth}
\includegraphics[width=\linewidth]{figs/vis-interaction.pdf}
\end{minipage}
\hfill
\begin{minipage}[b]{0.32\linewidth}
\includegraphics[width=\linewidth]{figs/vis-n.pdf}
\end{minipage}
\vspace{-15pt}
\caption{(Left) Computation of the $m$-order interaction between two input variables ($m=2$). (Right) Contexts of the $m$-order interaction {\small $I^{(m)}_{ij}$}. Variables $(i,j)$ are indicated by red boxes. We extend the visualization method in \cite{zhang2021interpreting}, which visualized contexts of the Shapley interaction index, to show contexts of multi-order interactions (details in the supplementary material).
Red (blue) colors indicate pixels that frequently (infrequently) appear in contexts which cause significant interactions.}
\label{fig:interaction}
\vspace{-5pt}
\end{figure}
\textbf{Decomposition of the existing Shapley interaction index into multiple orders.}
{\citet{zhang2020interpreting-dropout}} further decomposes the Shapley interaction index $I(i,j)$~\cite{grabisch1999axiomatic} into interactions of different orders.
\begin{equation}
\begin{small}
I(i,j) = \frac{1}{n-1}{\sum}_{m=0}^{n-2} I^{(m)}_{ij},\quad
I^{(m)}_{ij} = {\mathbb{E}}_{S\subseteq N\setminus\{i,j\},|S|=m} \left[\Delta v(i,j,S) \right]
\end{small}
\end{equation}
where {\small $\Delta v(i,j,S)= v(S\cup\{i,j\})-v(S\cup\{i\})-v(S\cup\{j\})+v(S)$}.
\textbf{Proof of decomposing network output into multi-order interactions to explain adversarial attacks.}
In this study, we successfully prove that the network output can be decomposed as the weighted sum of multi-order interactions.
This enables us to use such interactions to explain adversarial attacks.
Before the decomposition, let us first use interactions to understand the attack.
As Figure~\ref{fig:interaction} (left) shows, the interaction component of each order {\small $I^{(m)}_{ij}$} represents collaborations between two input variables {\small $(i,j)$} with a specific contextual complexity.
Let us consider input variables {\small $i,j$} and an arbitrary set of $m$ contextual variables.
If {\small$I^{(m)}_{ij} > 0$}, it indicates that the presence of the variable $j$ will increase the importance of the variable $i$. Thus, we consider variables {\small $(i,j)$} have a positive interaction.
If {\small$I^{(m)}_{ij}<0$}, it indicates a negative interaction.
The $m$-order interaction between two input variables $(i,j)$ measures the average interaction between $(i,j)$ when we consider their collaborations with $m$ contextual variables ($m$ pixels).
From another perspective, the order $m$ measures the number of contextual variables that influence the interaction between $i$ and $j$.
For a low order $m$, {\small $I^{(m)}_{ij}$} reflects the interaction between $i$ and $j$ \emph{w.r.t.} simple contexts of a few variables.
For a high order $m$, {\small $I^{(m)}_{ij}$} corresponds to the interaction \emph{w.r.t.} complex contexts with massive variables.
Figure~\ref{fig:interaction} (right) visualizes the contexts $S$ corresponding to strong low-order interactions and strong high-order interactions, respectively.
\citet{cheng2021game} has also proven that low-order interactions (local collaborations) mainly reflect simple and common concepts (features), and high-order interactions (global collaborations) usually represent complex and global features.
\textit{$\bullet$ Properties of multi-order interactions.}
{We have proven that {\small $I^{(m)}_{ij}$} satisfies \textbf{\textit{linearity, nullity, commutativity, symmetry}}, and \textbf{\textit{efficiency}}
properties (please see the supplementary material for details).}
In particular, the \textit{efficiency property} is given as follows.
\textit{Efficiency property:}
The output of the DNN given the entire input $v(N)$ can be decomposed into interactions
of different orders, \emph{i.e.} {\small$v(N) = v(\emptyset)+\sum_{i\in N}\phi^{(0)}(i)+\sum_{i,j\in N,i\ne j}\sum_{m=0}^{n-2} J^{(m)}_{ij}$, where $J^{(m)}_{ij} \overset{\text{def}}{=} \frac{n-1-m}{n(n-1)} I^{(m)}_{ij}$, and $\phi^{(0)}(i) \overset{\text{def}}{=} v(\{i\})-v(\emptyset)$}.
\textit{$\bullet$ Decomposing the overall attacking utility into interactions of different orders.}
According to the \textit{efficiency property} of interactions, the output of the DNN \emph{w.r.t.} the image $x$ can be decomposed into the sum of multi-order interactions between different pairs of input variables.
Thus, we can decompose the overall utility of adversarial perturbations on the network output into elementary effects on interactions.
\vspace{-5pt}
\begin{equation}
\begin{small}
\Delta v(N|x) \overset{\text{def}}{=} v(N|x)-v(N|x^{\text{adv}}) = \underbrace{\Delta v(\emptyset|x) + \sum_{i\in N} \Delta \phi^{(0)}(i|x)}_{\text{usually can be ignored}} + \sum_{i\ne j\in N}\sum_{m=0}^{n-2} \Delta J^{(m)}_{ij}(x)
\end{small}
\label{eq:v(n)}
\end{equation}
where {\small $x\in \mathbb{R}^n$} denotes the normal sample, and {\small $x^{\text{adv}}=x+\Delta x \in \mathbb{R}^n$} denotes the adversarial example.
{\small $v(N|x^{\text{adv}})$} denotes the network output given all variables in {\small $N$} of the adversarial example {\small $x^{\text{adv}}$}.
{\small $v(N|x)$} corresponds to the normal sample.
The first term {\small $\Delta v(\emptyset)=v(\emptyset|x)-v(\emptyset|x^{\text{adv}})=0$}.
In the second term, {\small $\Delta \phi^{(0)}(i|x) \overset{\text{def}}{=} (v(\{i\}|x)-v(\emptyset|x))-(v(\{i\}|x^{\text{adv}})-v(\emptyset|x^{\text{adv}}))=v(\{i\}|x)-v(\{i\}|x^{\text{adv}})$}.
Because in most applications, the importance of a single variable (\emph{e.g.} a pixel) is usually small, we can ignore this term.
In the third term, {\small $\Delta J^{(m)}_{ij}(x)$} denotes the attacking utility of the $m$-order interaction between variables {\small $(i,j)$} in {\small $x$}.
{\small$\Delta J^{(m)}_{ij}(x)\overset{\text{def}}{=}\frac{n-1-m}{n(n-1)} \Delta I^{(m)}_{ij}(x)$}, where {\small$\Delta I^{(m)}_{ij}(x)\overset{\text{def}}{=}I^{(m)}_{ij}(x)-I^{(m)}_{ij}(x^{\text{adv}})$} measures the elementary effects of adversarial perturbations on the $m$-order interaction.
\section{Explaining adversarial attacks and defense using interactions}
\label{sec:explain}
To simplify the story, we only study the simplest and widely-used untargeted and targeted $\ell_\infty$ PGD attacks~\cite{madry2018towards}.
\subsection{Attacks mainly affect high-order interactions}
\label{sec:explain_attack}
\textbf{Attacking utility of $m$-order interactions.}
According to Eq.~\eqref{eq:v(n)}, effects of adversarial attacks mainly depend on changes of interactions {\small $\Delta I^{(m)}_{ij}(x)$} (or {\small $\Delta J^{(m)}_{ij}(x)$}).
Thus, in this section, we further conduct experiments to show that adversarial attacks mainly affect {\small $\Delta I^{(m)}_{ij}(x)$} of high orders.
In order to compare the multi-order interactions in normal samples and those in adversarial examples, we measure the average interaction of a specific order $m$ among different pairs of variables in different input images, \emph{i.e.} {\small $I^{(m)}=\mathbb{E}_{x\in\Omega} \mathbb{E}_{i,j}[I^{(m)}_{ij}(x)]$},
where {\small $\Omega\subseteq \mathbb{R}^n$} denotes the set of all samples.
We use the normal samples {\small $\Omega_{\text{nor}}$} to compute {\small $I^{(m)}_{\text{nor}}$} and use adversarial examples {\small $\Omega_\text{adv}$} to compute {\small $I^{(m)}_{\text{adv}}$}.
In this way, {\small $\Delta I^{(m)}\overset{\text{def}}{=}I^{(m)}_{\text{nor}}-I^{(m)}_{\text{adv}}$} represents effects of attacks on the $m$-order interactions.
Similarly, {\small $\Delta J^{(m)}\overset{\text{def}}{=} \mathbb{E}_{x\in\Omega}\mathbb{E}_{i,j}[\Delta J^{(m)}_{ij}(x)]=\frac{n-1-m}{n(n-1)}\Delta I^{(m)}$} measures the attacking utility of the $m$-order interactions according to Eq.~\eqref{eq:v(n)}.
\begin{figure}[t]
\centering
\begin{minipage}[b]{0.43\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/untarget-interaction-n.pdf}
\end{minipage}
\begin{minipage}[b]{0.56\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/target-interaction-n.pdf}
\end{minipage}
\vspace{-15pt}
\caption{The multi-order interaction $I^{(m)}_\text{nor}$ and $I^{(m)}_\text{adv}$ of standard DNNs and adversarially trained DNNs. Adversarial perturbations mainly affected high-order interactions.}
\label{fig:explain_attack}
\vspace{-10pt}
\end{figure}
\textbf{Comparing attacking utilities between interactions of different orders (in terms of {\small $I^{(m)}$}).} In order to measure {\small $I^{(m)}_\text{nor}$} and {\small $I^{(m)}_\text{adv}$}, we conducted experiments on ResNet-18/34/50~\cite{he2016deep}, Wide-ResNet50-4~\citep{zagoruyko2016wide}, DenseNet-161~\cite{huang2017densely}, and VGG-16~\cite{simonyan2015very}.
For each DNN, we obtained both the standardly trained version and the adversarially trained version on the ImageNet dataset~\cite{imagenet2015}\footnote{We used pretrained models released by \citet{salman2020adversarially}.}.
We also conducted experiments on PointNet++~\cite{qi2017pointnet} learned on the ModelNet10 dataset~\cite{wu20143d}, which is a 3D dataset.
We measured the interaction {\small $I^{(m)}$} by setting {\small $v(S|x)=\log p(y=y^\text{truth}|\text{given variables in}~S~\text{in the input}~x~\text{and mask variables in}~N\setminus S)$}, and {\small $x\in \Omega$} was sampled from the validation set.
By following settings in \cite{ancona2019explaining}, the masked variables in {\small $N\setminus S$} were set to the average value over different input samples.
However, the computational cost of {\small $ I^{(m)}$} was intolerable.
To reduce the computational cost, we applied the sampling-based approximation method in \cite{zhang2020interpreting-dropout} and
computed interactions at the grid level, rather than the pixel level. Please see the supplementary material for more discussions.
We considered both untargeted and targeted\footnote{In the untargeted attack, we considered the misclassified category as the target category. In the targeted attack, the target label was set as the misclassified category in the untarget attack.\label{fn:target} When we consider interaction \emph{w.r.t.} the target category, we set {\small $v(S) = \log p(y=y^{\text{target}}|\text{given variables in}~S~\text{and mask variables in}~N\setminus S)$}.} PGD attacks~\cite{madry2018towards} with the {\small $\ell_{\infty}$} constraint {\small $\Vert \Delta x \Vert_{\ell_\infty} \le \epsilon$} to generate adversarial examples. We set $\epsilon=32/255$ by following the setting in~\cite{xie2019feature}.
The step size was set to $2/255$.
For fair comparisons, we controlled the perturbation generated for each image to have similar attacking utility of 8.
The attacking utility in untargeted attacks was defined as {\small $U_\text{untarget}(x) = h_{y^\text{truth}}(N|x)-h_{y^\text{truth}}(N|x^{\text{adv}})$}, where
{\small $h_{y^\text{truth}}(N|x)$} and {\small $h_{y^\text{truth}}(N|x^{\text{adv}})$} denote network outputs of the ground-truth category {\small $y^\text{truth}$} before the softmax layer, when taking the normal sample and the adversarial example as input, respectively.
The attacking utility in targeted attacks was defined as {\small $U_\text{target}(x) = (h_{y^\text{target}}(N|x^{\text{adv}})-h_{y^\text{truth}}(N|x^{\text{adv}}))-(h_{y^\text{target}}(N|x)-h_{y^\text{truth}}(N|x))$}, where {\small $h_{y^\text{target}}(\cdot)$} denote the network output of the target category\footref{fn:target} {\small $y^\text{target}$}.
\textit{Results.} Figure~\ref{fig:explain_attack} shows the multi-order interactions within normal samples and adversarial examples.
We found that adversarial perturbations significantly decreased high-order interactions, in both adversarially trained DNNs and standard DNNs.
For example, in Figure~\ref{fig:explain_attack} (a), low-order interactions ($m<0.5 n$) in standard ResNet-18 decreased a little, while the interaction of order $m=0.95n$ dropped from $0.01$ to $-0.059$.
This phenomenon was consistent with the heuristic findings of \citet{dong2017towards} that neurons corresponding to high-level semantics were ambiguous.
\textit{Additional experiments on the target category and other categories.} Beyond above analysis about interactions \emph{w.r.t.} the ground-truth category, we also conducted experiments and found that high-order interactions \emph{w.r.t.} the target category\footref{fn:target} increased significantly.
Please see the supplementary material for details.
In sum, high-order interactions were much more sensitive than low-order interactions, which verified our conclusions.
\textbf{Theoretic explanation of the sensitivity of high-order interactions.}
Besides the above empirical observation, we can also theoretically prove that high-order interactions are more sensitive than low-order interactions.
\textit{\textbf{Proposition 1} (equivalence between the multi-order interaction and the mutual information):
Given an input sample $x\in X \subseteq \mathbb{R}^n$ and the network output $Y$, we define $X_S=\{x_S|x\in X\}$ where $S\subseteq N$; each $x_S$ represents the sample, where variables not in $S$ are masked.
If $v(S)$ is set as the entropy of classification {\small$v(S) = H(Y|X_S)=\sum_{x_S} p(x_S) H(Y|X_S= x_S)$},} then {\small $I^{(m)}_{ij} =\mathbb{E}_{S\subseteq N\setminus\{i,j\},|S|=m} MI(X_i;X_j;Y|X_S)$}.
The conditional mutual information $MI(X_i;X_j;Y|X_S)$ measures the remaining mutual information\footnote{Note that unlike the bivariate mutual information, $MI(X_i;X_j;Y|X_S)$ can be negative.} between $X_i,X_j$ and $Y$, when $X_S$ is given.
When $X_S$ (with $m$ pixels) is given, $MI(X_i;X_j;Y|X_S)$ can be considered as the benefit to the inference from the interaction between $X_i$ and $X_j$.
Please see the supplementary material for proofs.
Proposition 1 indicates that compared to low-order interactions, high-order interactions are conditioned on larger contexts $S$, \emph{i.e.} conditioned on more contextual perturbations, thereby suffering more from adversarial perturbations.
Note that in Proposition 1, we set {\small $\hat{v}(S)=H(Y|X_S)$}, which is slightly different from setting {\small $v(S)=\log p(y=y^\text{truth}|S,x)$} in Section~\ref{sec:explain_attack}. Nevertheless, the trend of {\small $v(S)$} can roughly reflect the negative trend of {\small $\hat{v}(S)$}, {which is discussed in the supplementary material}.
\textbf{Comparison with frequency-based methods.}
Previous studies~\cite{yin2019fourier,wang2020high,harder2021spectraldefense} explained adversarial perturbations as high-frequency features.
Figure~\ref{fig:frequency} (left) shows the difference in the frequency between normal samples and adversarial samples.
We find that high-order interactions can better distinguish adversarial examples and normal samples, than the frequency of features, which indicates that the interaction metric can better explain the essential property of adversarial perturbations.
Please see the supplementary material for more details.
\subsection{Explaining specific interactions encoded by adversarially trained DNNs}
\label{sec:explain_training}
\textbf{Discovery.} In order to understand the difference in signal-processing behaviors between adversarially trained DNNs and standard DNNs,
we have proven that the overall attacking utility {\small $\mathbb{E}_{x\in\Omega} [\Delta v(N|x)]$} can be approximately decomposed as the sum of attacking utilities on multi-order interactions {\small $\Delta J^{(m)}$}, according to Eq.~\eqref{eq:v(n)}.
Figure~\ref{fig:diff} shows the attacking utility of multi-order interactions \emph{i.e.} {\small $\Delta J^{(m)}=\frac{n-1-m}{n(n-1)} \Delta I^{(m)}$}.
\begin{figure}[t]
\centering
\begin{minipage}[t]{0.48\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/diff.pdf}
\vspace{-15pt}
\caption{Distribution of compositional attacking utilities caused by interactions of different orders {\small $\{\Delta J^{(m)}\}$}.}
\label{fig:diff}
\end{minipage}
\hfill
\begin{minipage}[t]{0.48\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/disentanglement-n.pdf}
\vspace{-15pt}
\caption{The interaction disentanglement {\small $D^{(m)}$}. Interactions of adversarially trained DNNs exhibited higher disentanglement than those of standard DNNs.}
\label{fig:disentanglement}
\end{minipage}
\vspace{-5pt}
\end{figure}
We noticed that for standard DNNs, adversarial attacks mainly affected high-order interactions.
For adversarially trained DNNs, although high-order interactions were usually sensitive to attacks, low/middle-order interactions took up more attacking utilities.
This phenomenon actually can also be observed in Figure~\ref{fig:explain_attack}, which is discussed in the supplementary material.
Considering Eq.~\eqref{eq:v(n)}, this indicated that adversarial perturbations towards adversarially trained DNNs penalized both complex features of global collaborations and simple features of local collaborations.
\textbf{Explanation.} We further explain the above phenomena, \emph{i.e.} why high-order interactions in adversarially trained DNNs are more robust to attacks than those in standard DNNs.
To this end, we propose the following disentanglement metric for the interaction of a specific order.
\vspace{-3pt}
\begin{equation}
\begin{small}
\begin{aligned}
\!\!\!\! D^{(m)}\!\! =\! \mathbb{E}_{x\in \Omega}\mathbb{E}_{\substack{i,j\in N\\ i\ne j}}\frac{ |I^{(m)}_{ij}(x)|}{\mathbb{E}_{S\subseteq N\setminus\{i,j\}, |S|=m} |\Delta v(i,j,S|x)|}
\!=\! \mathbb{E}_{x\in \Omega}\mathbb{E}_{\substack{i,j\in N\\ i\ne j}}\frac{ |\mathbb{E}_{S\subseteq N\setminus\{i,j\}, |S|=m} \Delta v(i,j,S|x)|}{ \mathbb{E}_{S\subseteq N\setminus\{i,j\}, |S|=m} |\Delta v(i,j,S|x)|}
\end{aligned}
\end{small}
\end{equation}
The above disentanglement metric examines whether or not the $m$-order interactions represent discriminative information for a specific category.
The high value of $D^{(m)}$ indicates that the $m$-order interactions relatively purely describe specific categories. In other words, the interactions between $(i,j)$ under different contexts $S$ consistently have the same effects (either positive or negative) on the inference of a specific category.
Let us consider the following toy example.
When the pair of $(i,j)$ consistently have positive interactions towards a specific category under different contexts, \emph{i.e.} $\forall S\subseteq N\setminus\{i,j\},|S|=m, \Delta v(i,j,S|x) >0$, then we have $D^{(m)}=1$.
This means that the $m$-th interactions between $i$ and $j$ stably promote the output probability of this category.
In contrast, a low value of $D^{(m)}$ indicates that interactions between $(i,j)$ represent diverse categories.
\emph{I.e.} given different contexts $S$, interactions between $(i,j)$ sometimes have positive effects on a specific category, and sometimes have negative effects.
Please see the supplementary material for more discussions.
\textit{Experiments.} Figure~\ref{fig:disentanglement} compares the interaction disentanglement $D^{(m)}$ of standard DNNs and the disentanglement of adversarially trained DNNs.
Interactions of adversarially trained DNNs were more disentangled than those of standard DNNs, especially for low-order interactions.
This indicated that low-order interactions in adversarially trained DNNs encoded more category-specific information for inference than low-order interactions in standard DNNs.
Based on this observation, we could explain the robustness of high-order interactions to attacks in adversarially trained DNNs.
Adversarial training learned more category-specific low-order interactions, which boosted the difficulty of attacking high-order interactions.
This was because high-order interactions (usually global features) were usually constructed by low-order interactions (usually local features).
Let us take the peacock in Figure~\ref{fig:interaction} (right) for example. In this image, low-order interactions (simple features) represented the body of the peacock and the green garden. In this case, it was difficult to attack this image to other categories (\emph{e.g.} the bicycle category) by constructing high-order interactions of the bicycle category using these peacock low-order features.
Due to the difficulty of attacking high-order interactions, the adversarial examples towards the adversarially trained DNN had to attack low-order interactions, instead.
\subsection{Unifying four existing adversarial defense methods}
\label{sec:unifying}
\textbf{$\bullet$ Explaining the attribution-based method of detecting adversarial examples.}
\citet{DBLP:journals/corr/abs-1906-03499} proposed a method to use the attribution score of input variables to detect adversarial examples.
In order to prove the effectiveness of this method, we define the multi-order Shapley value just like the multi-order interaction, as follows. {\small $\phi^{(m)}(i|x)\overset{\text{def}}{=} \mathbb{E}_{S\subseteq N\setminus\{i\},|S|=m}[v(S\cup \{i\}|x)-v(S|x)]$}, where $m$ denotes the order of the Shapley value.
Then, we can prove that the attribution score which was used in~\cite{DBLP:journals/corr/abs-1906-03499} to detect adversarial exmaples, can be writtern as {\small $\phi^{(n-1)}(i|x)=v(N|x)-v({N\setminus\{i\}|x})$}.
We further prove that $\phi^{(m)}(i|x)$ can be decomposed into interactions of different orders.
\begin{equation}
\begin{small}
v(N|x)\! -\! v(\emptyset|x)\! =\! \frac{1}{n}\sum_{i\in N}\sum_{m=0}^{n-1}\phi^{(m)}(i|x),~~\phi^{(m)}(i|x) \!=\! \mathbb{E}_{j\in N\setminus\{i\}} \!\! \left[\sum_{k=0}^{m-1} I^{(k)}_{ij}(x)\right] \!\! + \phi^{(0)}(i|x)
\label{eq:multi-order shapley}
\end{small}
\end{equation}
According to the above equation, the overall adversarial effects can also be decomposed into elementary effects on $\phi^{(m)}(i|x)$.
$\phi^{(0)}(i|x)=v(i|x)-v(\emptyset|x)$ is usually small and can be ignored, as discussed below Eq.~\eqref{eq:v(n)}.
Among Shapley values of all orders, only the {\small$(n-1)$}-order component $\phi^{(n-1)}(i|x)$ contains the interactions of the highest order, which are the most sensitive interactions,
according to Section~\ref{sec:explain_attack}.
Therefore, the Shapley value component $\phi^{(n-1)}(i|x)$ is supposed to be the most sensitive to adversarial perturbations, which proves the effectiveness of the detection of adversarial examples.
Please see the supplementary material for related proofs and discussions.
\textbf{$\bullet$ Explaining high recoverability of adversarial examples towards adversarially trained DNNs.}
The adversarial recoverability of a DNN is referred to as whether the DNN's adversarial examples can be inverted back to the normal sample by minimizing the classification loss.
Specifically, given an adversarial example $x^{\text{adv}}$ generated by the untargeted attack~\cite{madry2018towards} via $\max_{\Vert x^{\text{adv}}-x \Vert _p\le \epsilon} \ell(x^{\text{adv}},y^{\text{truth}})$, we conduct a targeted attack~\cite{madry2018towards} on the adversarial sample to invert the classification result back to its ground-truth label, \emph{i.e.} $\min_{\delta^{\prime}}\ell(x^{\text{adv}}+\delta^{\prime},y^{\text{truth}})$. Let $\hat{x} = x^{\text{adv}}+\delta^{\prime}$ denote the recovered sample.
If $\Vert \hat{x}-x \Vert_2 \le \Vert x^{\text{adv}}-x \Vert_2$, it indicates a high recoverability; otherwise, a low recoverability.
We find that \textbf{\textit{adversarial examples generated by adversarially trained DNNs usually exhibit higher recoverability than those generated by standard DNNs.}}
This phenomenon can be explained by Proposition 1 in Section~\ref{sec:explain_attack},
\emph{i.e.} low-order interactions correspond to the mutual information conditioned on smaller adversarial perturbations, thereby suffering less from attacks.
Therefore, we can partially explain the high recoverability of adversarial examples towards adversarially trained DNNs.
It is because adversarially trained DNNs mainly focus on low-order interactions (see Section~\ref{sec:explain_training}), which leads to high recoverability.
Besides, we conducted experiments on ResNet-18/50 and DenseNet-161 trained on the ImageNet dataset, which also validated the high recoverability of adversarial examples generated on adversarially trained DNNs.
Please see the supplementary material for details.
\begin{figure}[t]
\centering
\begin{minipage}[b]{0.49\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/frequency.pdf}
\end{minipage}
\hfill
\begin{minipage}[b]{0.49\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/rank.pdf}
\end{minipage}
\vspace{-5pt}
\caption{(Left) The game-theoretic interaction is a more significant metric to distinguish normal samples and adversarial examples. (Right) Multi-order interactions in images of different ranks.}
\vspace{-8pt}
\label{fig:frequency}
\end{figure}
\textbf{$\bullet$ Explaining the rank-based method.}
\citet{jere2020singular} discovered that standard DNNs paid certain attention to high-rank (\emph{w.r.t.} the Singular Value Decomposition) features, but adversarial training put more attention on low-rank features.
This implied that adversarial examples mainly contained high-rank features.
We used \cite{jere2020singular} to reduce the SVD rank of an image (either a normal sample or an adversarial example) to {\small $\frac{1}{4}k$}-rank image and {\small $\frac{1}{16}k$}-rank image, where $k$ denoted the full rank of the image.
Figure~\ref{fig:frequency} (right) shows that
(1) compared with normal samples, the rank-reducing operation on adversarial examples significantly increased high-order interactions.
This indicated our conclusions in Section~\ref{sec:explain_attack} could explain that adversarial perturbations were mainly high-rank components in images.
(2) When the rank of images was further reduced from {\small $\frac{1}{4}k$} to {\small $\frac{1}{16}k$}, the rank-reducing operation affected both low-order and high-order interactions. However, when we reduce the rank of either normal samples or adversarial examples, they exhibited similar effects on low-order interactions, but exhibited dramatically different effects on high-order interactions.
This indicated that \textbf{the most discriminative factor of distinguishing adversarial examples from normal samples was high-order interactions, rather than the rank-reducing operation.}
In other words, the game-theoretic interaction presents a more essential property of adversarial attacks.
\begin{table}[t]
\caption{Ratio of adversarial examples whose classification results were corrected.
\label{tab:dropout}}
\centering
\resizebox{0.46\linewidth}{!}{
\begin{tabular}{c|c|c|c|c}
\hline
\multirow{2}*{Standard DNNs} & \multirow{2}*{cutout} & \multicolumn{3}{|c}{ours} \\
\cline{3-5}
{} & {} & $\alpha=0.1$ & $\alpha=0.2$ & $\alpha=0.3$ \\
\hline
ResNet-18 & 7.12\% & 19.62\% & 35.55\% & \textbf{44.14\%} \\
ResNet-50 & 10.63\% & 28.53\% & 45.53\% & \textbf{52.33\%} \\
DenseNet-161 & 11.32\% & 27.16\% & 46.15\% & \textbf{56.35\%} \\
\hline
\end{tabular}}
\hspace{3pt}
\resizebox{0.46\linewidth}{!}{
\begin{tabular}{c|c|c|c|c}
\hline
{$~$ Adversarially$~~$} & \multirow{2}*{cutout} & \multicolumn{3}{|c}{ours} \\
\cline{3-5}
{$~$ trained DNNs$~~$} & {} & $\alpha=0.1$ & $\alpha=0.2$ & $\alpha=0.3$ \\
\hline
ResNet-18 & 23.92\% & 22.46\% & 33.78\% & \textbf{39.18\%} \\
ResNet-50 & 24.27\% & 23.23\% & 35.23\% & \textbf{41.54\%} \\
DenseNet-161 & 35.65\% & 24.65\% & 39.25\% & \textbf{45.74\%} \\
\hline
\end{tabular}}
\vspace{-12pt}
\end{table}
\textbf{$\bullet$ Explaining the cutout method.}
\citet{devries2017improved} proposed the cutout method to boost the robustness of DNNs by masking out a square region of the input during training.
This operation can be roughly considered as applying a random dropout operation~\cite{srivastava2014dropout} to input images.
The dropout operation on input images has been proven to destroy interactions between input variables~\cite{zhang2020interpreting-dropout}.
To this end, we have further proven that the dropout operation mainly hurts high-order interactions, rather than low-order interactions (please see the proof in the supplementary material).
In this way, the cutout method can be considered as removing high-order interactions and pushing DNNs to learn more insensitive low-order interactions.
Therefore, the learned DNN is robust to adversarial attacks.
In experiments, we applied the dropout operation to the misclassified adversarial examples, in order to verify the attacking-alleviating influence on the network output of removing high-order interactions.
If the removal of high-order interactions could alleviate the effects of adversarial attacks, it would prove the attacking-alleviating influence of removing high-order interactions.
Given ResNet-18/50 and DenseNet-161 trained on the ImageNet dataset,
we selected normal samples that are correctly classified in the validation set of the ImageNet dataset, and generated adversarial examples with the $\ell_\infty$ untargeted PGD attack based on~\cite{xie2019feature}.
The attack was stopped once it succeeded.
We applied the dropout operation with different dropout rates $\alpha$ at the pixel level, and the dropped pixels were filled with the average value of surrounding pixels.
For the cutout method, we set the length of the side of the masked square regions as $112$, which was half of the side length of the input sample (dropping $\alpha=25\%$ pixels of the input), following settings in \cite{devries2017improved}.
Table~\ref{tab:dropout} reports the ratio of adversarial examples whose classification results were corrected.
When we applied dropout to adversarial examples, the classification accuracy of adversarial examples significantly increased,
and even performed better than the cutout method.
This indicated that the removal of high-order interactions presented a more essential defense mechanism than the cutout method.
\section{Conclusion}
\label{sec:conclusion}
In this paper, we have used the multi-order interaction to provide a unified understanding for the success of adversarial attacks and defense.
Based on the multi-order interaction, we have explained adversarial attacking as mainly affecting high-order interactions between input variables.
Furthermore, we have also explained the success of adversarial training, \emph{i.e.} learning more category-specific low-order interactions to boost the robustness.
Besides, we have further provided a unified explanation for some existing adversarial defense methods.
\begin{ack}
{This work is partially supported by the National Nature Science Foundation of China (No. 61906120, U19B2043), Shanghai Natural Science Foundation (21JC1403800,21ZR1434600), Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102).
This work is also partially supported by Huawei Technologies Inc.
Prof. Yisen Wang is partially supported by the National Natural Science Foundation of China under Grant 62006153, and Project 2020BD006 supported by PKU-Baidu Fund.
Xin Wang is supported by Wu Wen Jun Honorary Doctoral Scholarship, AI Institute, Shanghai Jiao Tong University.}
\end{ack}
\newpage
{
\small
\bibliographystyle{plainnat}
|
\section{Introduction}
In the past 20 years, most of the physical layer technologies, i.e., modulations, multiple access waveforms, coding techniques and time/frequency multiplexing, have flourished over the evolution of cellular systems. However, up to 4G, time-frequency domain technology has been explored to increase overall system capacity \cite{9040431}, \cite{9321326}. The recent developments in 5G and beyond technologies support emerging applications such as smart homes, vehicular networks, augmented reality (AR), virtual reality (VR) with unprecedented rates enabled by recent advances in massive multiple-input multiple-output (MIMO), millimeter-wave (mmWave) communications, network slicing, small cells, and Internet of things (IoT). These complex structures of 5G and beyond technologies can be captured by using data-driven approach machine learning (ML) algorithm \cite{7792374,8666641, 8972389}. The strong learning, reasoning, intelligent recognition abilities of ML allow the network structure to train and adapt itself to support the diverse demands of the systems without human intervention \cite{8360430}.
The extraordinary growth of data traffic on wireless communication has driven the need to examine the highest frequency spectrum to meet the requirements by using mmWave communications \cite{6736750}. The frequency range of the mmWave communication system is between 30 and 300 GHz the available bandwidth is about 250 GHz. Enabling mmWave communication faces mainly three critical challenges \cite{8395149, 7400949} i) the sensitivity for atmospheric attenuation obligates it to propagate solely by line-of-sight paths, ii) hand over problem between base stations (BSs), iii) adjustment of the large numbers of beamforming arrays. In addition, due to the use of large antenna arrays and low complexity, transceiver demands are captured by using ML algorithms for mmWave communication.
mmWave communication systems require the pointing of the narrow beams. The goal is to choose the best beams for the analogue beamforming with both receiver and transmitter having multi-antenna arrays. A beam codeword is a set of analogue phase-shifted values applied to the antenna elements due to an analogue beam \cite{8767936}\cite{9048614}. In \cite{9034044}, deep learning base beam selection is proposed that exploits channel state information for the sub-6 GHz links. In addition to beam prediction, information about the locations and sizes of vehicles in the communication environment are used in \cite{8644288} to predict the optimal beam pair. Locational based beamforming solutions are more suitable for line-of-sight (LOS) communication. The same locations for the non-line-of-sight (NLOS)transmission need different beamforming solution.
The integration of the ML for the 6G beyond technologies lead to potential security concerns \cite{9229146}, \cite{sagduyu2021adversarial}. Especially, wireless communication systems have security vulnerabilities due to their nature. The studies 6G beyond technologies with ML methods should be evaluated in terms of security. Current research is just building the AI models for the 6G communication problems. On the other hand, security concerns are ignored in previous studies. Based on the shortcomings of the literature's security concepts, we deal with the security problem of machine learning application for beamforming prediction. Alkhateeb et al. \cite{8395149} proposed a feed-forward deep learning model for RF beamforming codeword prediction with several base stations (BSs) with multiple users. The BSs beamforming vectors are predicted from the received signals using the omni and quasi-omni beam patterns to enable both LOS and NLOS transmissions. While the proposed method in \cite{8395149} showed promising results for the beamforming problem, the deep learning algorithm itself security was not investigated. Based on ML-based beamforming prediction's shortcomings, in this study, we focus on adversarial attack strategies based on loss maximisation-based attacks against proposed AI models for 6G mmWave communication. We consider the adversarial machine learning attacks to poisoning the beamforming prediction model \cite{8395149}. Thus our main contributions for this paper are as follows:
\begin{itemize}
\item We show that an undefended RF beamforming codeword deep learning model's prediction performance will decrease with the craftily designed adversarial noise.
\item We demonstrated that the adversarial training based robustness approach is one of the mitigation methods for this domain.
\end{itemize}
The rest of the paper is organized as follows: Section \ref{sec:preliminaries} describes preliminary information about model uncertainty and uncertainty quantification. Section \ref{sec:system_overview} shows our system overview. Section \ref{sec:experiments} evaluates the proposed uncertainty quantification method. Section \ref{sec:conclusion} concludes this paper.
\vspace{2pt}
\emph{Notations}
In this paper, we employ the following notations:
\begin{itemize}
\item Vectors denote in lowercase bold font and matrices in uppercase bold font i.e., $ \textbf{a} $ and $ \textbf{A} $ respectively.
\item For a given vector $ \textbf{a} $, $ \textbf{a}_{i,j} $ and $ \textbf{a}_{k} $ denote $ (i,j) $-th component of $ \textbf{a} $ and $ k $-th component of $ \textbf{a} $ respectively. For a integer $ d $, $ d_{i,j} $ means $ (i,j) $-th components of matrix $ \textbf{D} $.
\end{itemize}
\section{Preliminary Information} \label{sec:preliminaries}
\subsection{Downlink Transmission}
Let consider a mmWave communication system as in Figure \ref{fig:downlink_comm} where $N$ is the number of the BSs with equipped $M$ antennas are serving for one mobile user who has single antenna. All the BSs are connected with a cloud processing unit. The transmitted signal $ \textbf{s} = [s_1, s_2,\dots, s_K]$ with $K$ subcarriers is firstly precoded by using code vector $ \textbf{c}_{k}= [c_{k,1}, c_{k,2},\dots, c_{k,N}]^T$ and then is transformed into time domain with using $K$-point IFFT operation. Thus, the baseband signal from the $n$-th BS and $k$-th subcarrier is
\begin{equation}\label{Eqn:b1}
\textbf{x}_{k,n} = \textbf{f}_{n}c_{k,n}s_k
\end{equation}
where $\textbf{f}_{n}$ is the beam steering vector is defined for each BS antennas as [$\textbf{f}_{n}]_m = \frac{1}{\sqrt{M}}e^{j\theta_{n,m}}$ where $\theta_{n,m}$ is a quantized angle. RF precoding matrix $\textbf{F}^RF= blkdiag(\textbf{c}_{1}, \textbf{c}_{2\dots,\textbf{c}_{N}}) \in \mathbf{C}^{NM\times N}$ The received signal at the $k$-th subcarrier is expressed as
\begin{equation}\label{Eqn:b2}
\textbf{y}_{k} =\sum_{n=1}^{N} \textbf{h}_{k,n}^T \textbf{x}_{k,n} + v_k
\end{equation}
where $v_k$ is additive white Gaussian noise (AWGN) with variance $ \sigma ^{2} $, i.e., $N(0, \sigma^2)$.
\subsection{Effective Achievable Rate}
Perfect channel information satisfies optimum achievable rate, however, the channel state information requires large training overhead due to the large number of the antennas. On the other hand, the channel information and beamforming vector need to be updated as the user moves. This issues can be captured as with the channel coherence time $T_{C}$, and channel beam coherence time $T_{B}$, which are examined in detail in \cite{7742901}. The multi-path channel and beams stay aligned on the $T_{C}$ and $T_{B}$ duration respectively. The channel training and beamforming design take place in the first $T_{tr}$, the rest of it is used to the data transmission. To develop a model with efficient channel training and beamforming design, the effective achievable rate need to be maximized. The final problem formulations \cite{8395149} are
\begin{equation}\label{Eqn:b3}
\begin{aligned}
& \prod \left(T_{tr},\lbrace \textbf{c}_{k}\rbrace_{k=1}^{K},\textbf{F}^R, \mathcal{F} \right)= \\
& {\mathrm{argmax}} \left( 1- \frac{T_{TR}}{T_{B}} \right) \sum_{k=1}^{K} \log_2 \left( 1+ SNR\vert \sum_{n=1}^{N} \textbf{h}_{k,n}^T \textbf{f}_{n}c_{k,n} \vert^2 \right)
\end{aligned}
\end{equation}
\begin{equation}\label{Eqn:b4}
s.t.\, \, \, \textbf{f}_{n} \in \mathcal{F}, \, \, \forall n
\end{equation}
\begin{equation}\label{Eqn:b5}
\| c_{k}\| ^2 = 1 \ \ \forall k
\end{equation}
where $\mathcal{F}$ is the quantized codebook for the BSs RF beamforming vectors. Solving these equations determine a solution for a low channel training ahead and realize the beamforming vector to satisfy the maximum achievable rate, $R$.
\subsection{Using Deep Learning Algorithms to estimate RF beamforming vectors}
Using the benefits of machine learning algorithms is a novel solution for a massive amount of MIMO channel training and scanning a large number of narrow beams. The beams depend on the environmental conditions like user and BSs locations, furniture, trees, building e.t.c. It is too difficult to define these environment conditions as a closed-form equation. The solution is to use omni and quasi-omni beam pattern to predict the best RF beamforming vectors. Using these beam patterns benefits to take into account the reflection and diffraction of the pilot signal.
The deep learning solution consists of two states: training and prediction. Firstly, the deep learning model learns the beams according to the omni-received pilots. Secondly, the model uses the trained data to predict the RF beamforming vector for the current condition.
\subsubsection{Training Steps}The user sends uplink training pilot sequences for each beam coherence time $T_{B}$. BSs combine received pilot sequences on RF beamforming vector $B$ and fed back to the cloud. The cloud uses the received sequences from all the BSs as the input of the deep learning algorithm to find the achievable rate in (\ref{Eqn:b6}) for every RF beamforming vector to represent the desired outputs, i.e. $\textbf{g}_{p}$ is the channel coefficient from ommi beam.
\begin{equation}\label{Eqn:b6}
{R}_{n} ^{(p)}=\frac{1}{K} \sum_{n=1}^{N} \log_2 \left( 1+ SNR\vert \textbf{h}_{k,n}^T \textbf{g}_{p} \vert^2 \right)
\end{equation}
\subsubsection{Learning Steps} In this stage, the trained deep learning model is used to prediction the RF beamforming vectors. Firstly, the user sends an uplink pilot sequence. The BSs combine these sequences and send them to the cloud. Then, the cloud uses the trained deep learning model to predict the best RF beamforming vectors to maximize the achievable rate for each BS. Finally, BSs use the predicted RF beamforming vector to estimate the effective channel $\textbf{h}_{k,n}$.
To sum up, machine learning algorithms find diverse applications in a wireless communication system that we consider the RF beamforming vector prediction \cite{8395149}. On the other hand, security concerns in wireless communication are also a problem for the ML algorithm.
\begin{figure}[t!]
\centering
\includegraphics[width=1.0\linewidth]{img/drawMIMO.png}
\caption{Block diagram of the mmWave beamforming system.}
\label{fig:downlink_comm}
\end{figure}
In the following section, we will briefly describe adversarial machine learning, attack environments, and adversarial training that we have used in this study.
\subsection{Adversarial Machine Learning}
Adversarial machine learning is an attack technique that attempts to fool neural network models by supplying craftily manipulated input with a small difference. The obvious intention is to produce a failure in a neural network model\cite{2016arXiv161101236K}.
Attackers apply model evasion attacks for phishing attacks, spams, and executing malware code in an analysis environment \cite{8965459}. There are also some advantages to attackers in misclassification and misdirection of models. In such attacks, the attacker does not change training instances. Instead, he tries to make some small perturbations in input instances in the model's inference time to make this new input instance seem safe (i.e. normal behaviour) \cite{2021arXiv210204150F}. We mainly concentrate on this kind of adversarial attacks in this study. There are many attacking methods for deep learning models, and the Fast-Gradient Sign Method (FGSM) is the most straightforward and powerful attack type. We only focus on the FGSM attack, but our solution to prevent this attack can be applied to other adversarial machine learning attacks.
FGSM works by utilizing the gradients of the neural network to create an adversarial example to evade the model. For an input instance $\mathbf{x}$, the FGSM utilizes the gradients $\nabla_x$ of the loss value $\ell$ for the input instance to build a new instance $\mathbf{x}^{adv}$ that maximizes the loss value of the classifier hypothesis $h$. This new instance is named the adversarial instance. We can summarize the FGSM using the following explanation:
\begin{equation}
\mathbf{x}^{adv} = \mathbf{x} + \epsilon \cdot sign(\nabla_x \ell(\mathbf{\theta},\mathbf{x},y))
\end{equation}
By adding a slowly modest noise vector $\eta \in \mathbb{R}^n$ whose elements are equal to the sign of the features of the gradient of the cost function $\ell$ for the input $\mathbf{x} \in \mathbb{R}^n$, the attacker can easily manipulate the output of a deep learning model.
Figure \ref{fig:fgsm_detail} shows the details of the FGSM attack.
\begin{figure}[h]
\centering
\includegraphics[width=1.0\linewidth]{img/fgsm_detail.pdf}
\caption{FGSM attack steps. The input vector $\mathbf{x} \in \mathbb{R}^n$ is poisoned with loss maximization direction. }
\label{fig:fgsm_detail}
\end{figure}
\subsection{Adversarial Training}
Adversarial training is a widely recommended defense that implies generating adversarial instances using the gradient of the victim classifier, and then re-training the model with the adversarial instances and their respective labels. This technique has demonstrated to be efficient in defending models from adversarial attacks.
Let us first think a common classification problem with a training instances $X \in \mathbb{R}^{m \times n}$ of dimension $d$, a label space $Y$ We assume the classifier $h_\theta$ has been trained to minimize a loss function $\ell$as follows:
\begin{equation}
\label{eq:cost-func}
\underset{\theta}{min}\frac{1}{m} \sum_{i=1}^{m} \ell(h_\theta(\mathbf{x}_i,y_i))
\end{equation}
Given a classifier model $h_\theta(\cdot)$ and an input instance $x$, whose responding output is $y$, an adversarial instance $x^*$ is an input such that:
\begin{equation}
\label{eq:adv_ex}
h_\theta(x^*) \neq y \,\,\,\,\, \wedge \, d(x,x^*) < \epsilon
\end{equation}
where $d(\cdot,\cdot)$ is the distance metric between two input instances original input $x$ and adversarial version $x^*$. Most actual adversarial model attacks transform Equation \ref{eq:adv_ex} into the following optimization problem:
\begin{equation}
\underset{x}{\mathrm{\textbf{argmax}}} \, \ell \left(h_\theta(x^*),y\right)
\end{equation}
\begin{equation}
s.t. d(x,x^*) < \epsilon
\end{equation}
where $\ell$ is loss function between predicted output $h(\cdot)$ and correct label $y$.
In order to mitigate such attacks, at per training step, the conventional training procedure from Equation \ref{eq:cost-func} is replaced with a \texttt{min-max} objective function to minimize the expected value of the maximum loss, as follows:
\begin{equation}
\underset{\theta}{min} \, \underset{(x,y)}{\mathbb{E}} \left(\underset{d(x,x^*)<\epsilon}{max} \ell(h(x^*),y) \right)
\end{equation}
\section{System Model}\label{sec:system_overview}
\subsection{Adversarial Training}
Figure \ref{fig:deepmimo_adv_learning} shows the adversarial training process. After the model is trained, adversarial inputs are created using the model itself, combined with legitimate users information and added to the training. When the model reaches the steady-state state, the training process is completed. In this way, the model will both predict RF beamforming codeword for legitimate users while at the same time being immune to the craftily designed noise attack that will be added as input.
\begin{figure}[htbp!]
\centering
\includegraphics[width=1.0\linewidth]{img/deepmimo_adv_learning.pdf}
\caption{The diagram of RF beamforming codeword adversarial training.}
\label{fig:deepmimo_adv_learning}
\end{figure}
\subsection{Capability of the Attacker}
We assumed that the attacker's primary purpose is to manipulate the RF model by applying a carefully crafted noise to the input data. In a real-world scenario, this white-box setting is the most desired choice for an attacker that does not take the risks of being caught in a trap. The problem is that it requires the attacker to access the model from outside to generate adversarial examples. After manipulating the input data, the attacker can exploit the RF beamforming codeword prediction model's vulnerabilities in the same manner as in an adversary's sandbox environment. The prediction model predicts the adversarial instances when the attacker can convert some model's outputs as other outputs (i.e. wrong prediction).
However, to prevent this noise addition from being easily noticed, the attacker must answer an optimization problem to determine which regions in the input data (i.e. beamforming) must be modified. By solving this optimization problem using one of the available attack methods \cite{8965459}, the attacker aims to reduce the prediction performance on the manipulated data as much as possible. In this study, to limit the maximum allowed perturbation allowed for the attacker, we used $l_\infty$ norm, which is the maximum difference limit between original and adversarial instance.
Figure \ref{fig:adv-ml} shows the attack scenario. The attacker gets an legitimate input, $\mathbf{x}$, creates a noise vector with an $\epsilon$ budget $\eta = \epsilon \cdot sign(\nabla_x \ell(\mathbf{\theta},\mathbf{x},y))$, sums the input instance and the craftily designed noise to create adversarial input $\mathbf{x}^{adv} = \mathbf{x} + \eta$.
\begin{figure}[htbp!]
\centering
\includegraphics[width=1.0\linewidth]{img/adv-ml.pdf}
\caption{Original milimeter-wave beam prediction model results}
\label{fig:adv-ml}
\end{figure}
\section{Experiments}\label{sec:experiments}
In the experiments, we tested three different scenarios
\begin{itemize}
\item \textbf{SC1:} Undefended beamforming codework prediction model without any adversary
\item \textbf{SC2:} Undefended beamforming codework prediction model with FGSM attack
\item \textbf{SC3:} Adversarial trained beamforming codework prediction model with FGSM attack
\end{itemize}
The experiments were performed using the Python scripts and machine learning libraries: Keras, Tensorflow, and Scikit-learn, on the following machine: 2.8 GHz Quad-Core Intel Core i7 with 16GB of RAM. For all scenarios, two models, undefended and adversarial trained, were built to obtain prediction results. In the first model, the model is trained without any input poisoning. The first model (i.e. undefended model) was used with legitimate users (for SC1) and adversaries (for SC2). The second model (i.e. the adversarially trained model) was used under the FGSM attack. The hyper-parameters such as the number of hidden layers and the number of neurons in the hidden layers, the activation function, the loss function, and the optimization method are the same for both models.
The model architectures are given in Table \ref{tab:dl_arch} and the hyper-parameters selected in Table \ref{tab:dl_arch_params}.
\begin{table}[htbp!]
\centering
\caption{Model architecture}
\label{tab:dl_arch}
\begin{tabular}{c|c}
\hline \textbf{Layer type} & \textbf{Layer information} \\ \hline \hline
Fully Connected + ReLU & 100 \\
Fully Connected + ReLU & 100 \\
Fully Connected + ReLU & 100 \\
Fully Connected + TanH & 1 \\ \hline
\end{tabular}
\end{table}
\begin{table}[htbp!]
\centering
\caption{Milimater-wave beam prediction model parameters}
\label{tab:dl_arch_params}
\begin{tabular}{c|c}
\hline
\textbf{Parameter} & \textbf{Value} \\
\hline \hline
Optimizer & Adam \\
Learning rate & 0.01 \\
Batch Size & 100 \\
Dropout Ratio & 0.25 \\
Epochs & 10 \\
\hline
\end{tabular}
\end{table}
\subsection{Research Questions}\label{RQ}
We consider the following two research questions (RQs):
\begin{itemize}
\item \textbf{RQ1}: Is the deep learning based RF beamforming codeword predictor vulnerable for adversarial machine learning attacks?
\item \textbf{RQ2}: Is the iterative adversarial training approach a mitigation method for the adversarial attacks in beamforming prediction?
\end{itemize}
\subsection{RF Beamforming Data Generator}\label{sec:datagenerator}
We employed the generic deep learning dataset for millimeter-wave and massive MIMO applications (DeepMIMO) data generator in our experiments \cite{alkhateeb2019deepmimo}. Figure \ref{fig:scenario} shows the bird's-eye view of a section of the O1' ray-tracing scenario, showing the two streets' intersection.
\begin{figure}[htbp!]
\centering
\includegraphics[width=1.0\linewidth]{img/scenario_o1.jpg}
\caption{Original scenario \cite{alkhateeb2019deepmimo}. }
\label{fig:scenario}
\end{figure}
In this section, we conduct experiments on the mmWave communication and massive MIMO applications dataset from the publicly available data set repository. We implemented the proposed mitigation method using Keras and TensorFlow libraries in the Python environment.
\begin{table}[htbp!]
\centering
\caption{Adversarial settings of our experiments perturbation budget $\epsilon$}
\label{tab:adversarial_settings}
\begin{tabular}{c|c|c}
\hline
\textbf{Attack} & \textbf{Parameters} & \textbf{$l_p$ norm}\\
\hline \hline
FGSM & $\epsilon \in [0.01, \cdots 0.1]$ & $l_\infty$\\
\hline
\end{tabular}
\end{table}
\subsection{Results for RQ1}\label{sec:res_rq1}
Figure \ref{fig:org} shows the original undefended deep learning model results without any attack. According to the figure, the deep learning model's predictions are very close the original value.
\begin{figure}[h]
\centering
\includegraphics[width=0.8\linewidth]{img/org.pdf}
\caption{Original (Undefended) RF beamforming codeword deep learning model results. }
\label{fig:org}
\end{figure}
Figure \ref{fig:history_org} shows the training history of the beamforming prediction model with 35.000 training instances. The model is trained with clean (non-perturbated) instances.
\begin{figure}[h]
\centering
\includegraphics[width=0.8\linewidth]{img/history_org.pkl.pdf}
\caption{The beamforming prediction model history.}
\label{fig:history_org}
\end{figure}
Figure \ref{fig:attacked} shows the performance results of the beamforming prediction model's evaluation results under the FGSM attack. We have used $l_{inf}$ norm as the distance metric, which shows the maximum allowable perturbation amount for each item in the input vector $\mathbf{x}$.
\begin{figure}[h]
\centering
\includegraphics[width=0.8\linewidth]{img/attacked.pdf}
\caption{Attacked (Undefended) RF beamforming codeword deep learning model results. }
\label{fig:attacked}
\end{figure}
Figure \ref{fig:rq1} shows the mean squared error (MSE) of the performance results with normal and attacked beamforming input.
\begin{figure}
\centering
\includegraphics[width=0.7\linewidth]{img/rq1.pdf}
\caption{The performance results for SC1 and SC2.}
\label{fig:rq1}
\end{figure}
According to the figures, the undefended RF beamforming codeword prediction model is vulnerable for the FGSM attack. The MSE performance result of the model under attack is approximately 40 (i.e. $\frac{0.00843 (Normal)}{0.00021 (Attacked)} \approx 40.14$) times higher.
\subsection{Results for RQ2}\label{sec:res_rq2}
Adversarial training is a popularly advised defense mechanism that proposes generating adversarial instances using the victim model's loss function, and then re-training the model with the newly generated adversarial instances and their respective outputs. This approach has proved to be effective in protecting deep learning models from adversarial machine learning attacks.
Figure \ref{fig:adv_training} shows the adversarial trained deep learning model results with FGSM attack. According to the figure, the deep learning model's predictions are very close to the original (i.e. undefended and non-attacked) value in Figure \ref{fig:org}.
\begin{figure}[h]
\centering
\includegraphics[width=0.8\linewidth]{img/adv_training.pdf}
\caption{Attacked (Undefended) RF beamforming codeword deep learning model results with adversarial training.}
\label{fig:adv_training}
\end{figure}
Figure \ref{fig:rq1} shows the MSE of the performance results for all scenarios.
\begin{figure}[h]
\centering
\includegraphics[width=0.8\linewidth]{img/rq2.pdf}
\caption{The performance results for all scenarios.}
\label{fig:my_label}
\end{figure}
\subsection{Threats to Validity}\label{sec:threats_to_val}
A key \textit{external validity} threat is related to the generalization of results \cite{a06430d231c644d8a1278057a2dd6956}. We used only the RF beamforming dataset in our experiments, and we need more case studies to generalize the results. Moreover, the dataset reflects different types of milimeter-wave beams.
Our key \textit{construct validity} threat is related to the selection of attack type FGSM. Nevertheless, note that this attack is from the literature \cite{a06430d231c644d8a1278057a2dd6956} and applied to several deep learning usage domains. In the future, we will conduct dedicated empirical studies to investigate more adversarial machine learning attacks systematically.
Our main \textit{conclusion validity} threat is due to finding the best attack budget $\epsilon$ that is responsible for manipulating the legitimate user's signal for poisoning the beamforming prediction model. To mitigate this threat, we repeated each experiment 20 times to reduce the probability that the results were obtained by chance. In a standard neural network training, all weights are initialized uniformly at random. In the second stage, using optimization, these weights are updated to fit the classification problem. Since the training started with a probabilistic approach, there is a possibility of facing optimization's local minimum problem. To eliminate the local minimum problem, we repeat the training 20 times to find the $\epsilon$ value that gives the best attack result. In each repetition, the weights were initialized uniformly at random but with different values. If the optimization function failed to find the global minimum in the next experiment, it is likely to see it as the weights have been initialized with different values.
\section{Conclusions and Future Works}\label{sec:conclusion}
This research discussed one of the security issues related to RF beamforming codeword prediction models' vulnerabilities and solutions: (1) Is the deep learning-based RF beamforming codeword predictor vulnerable for adversarial machine learning attacks? (2) Is the iterative adversarial training approach a mitigation method for the adversarial attacks in beamforming prediction? We conducted experiments with the DeepMIMO ray tracing scenario to answer these questions. Our results confirm that the original model is vulnerable to a modified FGSM attack. One of the mitigation methods is the iterative adversarial training approach. Our empirical results also show that iterative adversarial training successfully increases the RF beamforming prediction performance and creates a more accurate predictor, suggesting that the strategy can improve the predictor's performance.
\bibliographystyle{ieeetr}
|
\section{Introduction}
It is time-consuming and difficult to manually design neural architectures under specific hardware constraints. Neural architecture search (NAS) \cite{evolutionnas}\cite{oneshotnas}\cite{rlnas} aiming at automatically searching the best neural architecture is thus highly demanded. However, how to efficiently and flexibly determine the architectures conforming to various constraints is still very challenging \cite{onceforall}.
The earliest NAS methods were developed based on reinforcement learning (RL) \cite{mnasnet}\cite{rlnas} or the evolution algorithm \cite{evolutionnas}. However, extremely expensive computation is needed. For example, 2,000 GPU days are needed by an RL method \cite{rlnas}, and 3,150 GPU days are needed by the evolution algorithm \cite{evolutionnas}.
\begin{comment}
For \emph{efficiency}, we mean the time required to find the best architecture for a specific hardware constraint. For \emph{flexibility}, we mean that, for various hardware constraints, a new best architecture should be efficiently derived from the existing architecture. The earliest NAS methods were developed based on reinforcement learning \cite{mnasnet}\cite{rlnas} or the evolution algorithm \cite{evolutionnas} which is extremely inefficient and inflexible. For example, in \cite{mnasnet}, it is inefficient that searching the best neural architecture for a specific hardware constraint takes 3800 GPU days. In addition, 3800N GPU days is required for N different hardware constraints, which makes NAS difficult to support various hardware resource flexibly.
When N different hardware constraints is required, 3800N GPU days
However, extremely expensive computation is needed. For example, 2,000 GPU days are needed by a reinforcement learning method \cite{rlnas}, and 3,150 GPU days are needed by the evolution algorithm \cite{evolutionnas}. The huge search cost makes NAS very \emph{inefficient}.
Although a number of elegant NAS methods have been proposed to automate architecture design, searching a neural architecture for a specific hardware constraint is still time-consuming and inefficient. Besides, It is inflexible to re-search N times for N different hardware constraints.
The earliest NAS methods were developed based on reinforcement learning \textcolor{blue}{(RL)} \cite{mnasnet}\cite{rlnas} or the evolution algorithm \cite{evolutionnas}. However, extremely expensive computation is needed. For example, 2,000 GPU days are needed by a \textcolor{red}{reinforcement learning method}\textcolor{blue}{RL} \cite{rlnas}, and 3,150 GPU days are needed by the evolution algorithm \cite{evolutionnas}. The huge search cost makes NAS very inefficient. Moreover, N times re-searching are required for N different hardware constraints, which is very inflexibility for NAS.
\end{comment}
\begin{figure}[t]
\begin{center}
\includegraphics[width=1\linewidth]{image/SGNAS_overview_new.png}
\end{center}
\caption{Overview of SGNAS. Given the target hardware constraint as the input, the architecture generator can generate architecture parameters instantly within the inference time of one forward pass. With the generated parameters, the specific architectures can be sampled from the unified supernet.}
\label{fig:SGNAS_overview}
\end{figure}
To improve searching efficiency, one-shot NAS methods \cite{oneshotnas}\cite{darts}\cite{proxylessnas}\cite{fbnet} were proposed to encode the entire search space into an over-parameterized neural network, called a \emph{supernet}. Once the supernet is trained, all sub-networks in the supernet can be evaluated by inheriting the weights of the supernet without additional training. One-shot NAS methods can be divided into two categories: differentiable NAS (DNAS) and single-path NAS.
In addition to optimizing the supernet, DNAS \cite{gdas}\cite{darts}\cite{fbnet}\cite{pcdarts:}\cite{atomnas} utilizes additional differentiable parameters, called \emph{architecture parameters}, to indicate the architecture distribution in the search space. Because DNAS couples architecture parameters optimization with supernet optimization, for $N$ different hardware constraints, the supernet and the architecture parameters should be trained jointly for $N$ times to find $N$ different best architectures. This makes DNAS methods inflexible.
\begin{comment}
DNAS searches the best neural architecture under the specific hardware constraint by jointly optimizing the supernet and the architecture parameters. After the optimizing, the best neural architecture can be sampled from the trained architecture parameters. However, the best neural architecture sampled from the trained architecture parameter is only suitable to a specific hardware constraint. With this scheme, N time re-searching is needed for N different neural architectures specific to N different hardware constraints. This make Differentiable method is less scable.
\end{comment}
In contrast, single-path methods \cite{uniform_sampling}\cite{fairnas}\cite{scarlet}\cite{greedynas} decouple supernet training from architecture searching. For supernet training, only a single path consisting of one block in each layer is activated and is optimized in one iteration. The main idea is to simulate discrete neural architectures in the search space and save GPU memory consumption. Once the supernet is trained, different search strategies, like the evolution algorithm \cite{greedynas}\cite{uniform_sampling}, can be used to search the architecture under different constraints without retraining the supernet. Single-path methods are thus more flexible than DNAS. However, re-executing the search strategy $N$ times for $N$ different constraints is costly and not flexible enough.
\begin{comment}
For supernet training, only a single path consisted one block in each layer is activated and optimized in one iteration to simulate the discrete neural architecture in the search space. The trained supernet is treated as a performance estimator to the all neural architectures in the search space. With the trained supernet, we can incorporate different search strategies, e.g., evolution algorithm \cite{greedynas}\cite{uniform_sampling}, to search the best neural architecture under different hardware constraints without retraining the supernet. However, executing the search strategy, e.g., evolution algorithm, is still expensive in previous Single-path methods.
\end{comment}
On top of one-shot NAS, we especially investigate \emph{efficiency} and \emph{flexibility}. For efficiency, we mean that, when the supernet is available, the time required to search the best architecture for a specific hardware constraint. For flexibility, we mean that, when $N$ different hardware constraints are to be met, how much total time required to search for $N$ best architectures. As a comparison instance, GreedyNAS \cite{greedynas} takes almost 24 GPU hours to search for the best neural architecture under a specific constraint. Totally $24N$ GPU hours are required for $N$ different constraints.
In this work, we focus on improving efficiency and flexibility of the search strategy of the single-path method. The main idea is \emph{searching the best architecture by generating it}. First, we decouple supernet training from architecture searching and train the supernet as a single-path method. After obtaining the supernet, we propose to build an \emph{architecture generator} to generate the best architecture directly. Given a hardware constraint as input, the architecture generator can generate the architecture parameter within the inference time of one forward pass. This method is extremely efficient and flexible. The total search time for various hardware constraints of the architecture generator is only 5 GPU hours. Moreover, we do not need to re-execute search strategies or re-train the supernet once the architecture generator is trained. When $N$ different constraints are to be met, the search strategy only needs to be conducted once, which is more flexible than $N$ searches required in previous single-path methods \cite{fairnas}\cite{greedynas}\cite{scarlet}.
\begin{comment}
Search space is a very critical component in NAS. A small search space will lead the searched architecture to fall into the suboptimal. However, it is very difficult and time-consuming to train the supernet stably with the huge search space because of the huge number of supernet parameters, and the over complex supernet structure. For example, GreedyNAS \cite{greedynas} has the 13 types of candidate blocks
(e.g., kernel sizes \{3, 5, 7\}, expansion \{3, 6\}, squeeze-and-excitation (SE) module\cite{se} and the skip operation)
for each layer(The size of the search space is $13^{L}$, $L$ denoted the total layer number). Under this size of the search space, there will be 13 candidate blocks in each layer of the supernet. Because of this, the parameters and complexity of the supernet will grow exponentially when we increase the size of the search space.
\end{comment}
The aforementioned idea is on top of a trained supernet. However, we notice that searching on a single-path supernet still requires a lot of GPU memory and time because of the huge number of supernet parameters and complex supernet structure. Previous single-path NAS methods \cite{uniform_sampling}\cite{greedynas}\cite{scarlet} determine a block for each layer, and there may be different candidate blocks with various configurations. For example, GreedyNAS \cite{greedynas} has 13 types of candidate blocks for each layer, and thus size of the search space is $13^{L}$, where $L$ denotes the total number of layers in the supernet. Inspired by the fine-grained supernet in AtomNAS \cite{atomnas}, we propose a novel single-path supernet called \emph{unified supernet} to reduce GPU memory consumption. In the unified supernet, we only construct a block called \emph{unified block} in each layer. There are multiple sub-blocks in the unified block, and each sub-block can be implemented by different operations. By combining sub-blocks, all configurations can be described in a block. In this way, the number of parameters of the unified supernet is much fewer than previous single-path methods
The contributions of this paper are summarized as follows. With the architecture generator and the unified supernet, we propose Searching by Generating NAS (SGNAS), which is a flexible and efficient one-shot NAS framework. We illustrate the process of SGNAS in Fig.~\ref{fig:SGNAS_overview}. Given various hardware constraints as the input, the architecture generator can generate the best architecture for different hardware constraints instantly in one forward pass. After training the best architecture from scratch, the evaluation results show that SGNAS achieves 77.1\% top-1 accuracy on the ImageNet dataset \cite{imagenet} at around 370M FLOPs, which is comparable with the state-of-the-arts of the single-path methods. Meanwhile, SGNAS outperforms SOTA single-path NAS in terms of efficiency and flexibility.
\section{Related Work}
Recently, one-shot NAS \cite{darts}\cite{fbnet}\cite{uniform_sampling} has received much attention because of reduced search cost brought by supernet. To futher reduce the search cost, a number of methods have been proposed, which can be roughly divided into two types: efficient NAS and flexible NAS.
\subsection{Efficient NAS}
For efficiency, many methods were proposed to improve the supernet training strategy or redesign the supernet architecture. Stamoulis et al. \cite{singlelessfour} proposed a single-path supernet to encode architectures with shared convolutional kernel parameters, which reduce search cost of differential NAS. To reduce the huge cost when training on large-scale datasets, training supernet and searching on proxy datasets like CIFAR10 or part of ImageNet was proposed in \cite{fbnet}\cite{fbnetv2}\cite{tfnas}\cite{darts}\cite{cars}\cite{darts}. PC-DARTS \cite{pcdarts:} only sampled a small part of the supernet for training in each iteration to reduce computation cost. DA-NAS \cite{danas} designed a data-adaptive pruning strategy for efficient architecture search.
\subsection{Flexible NAS}
For flexibility, in OFA \cite{onceforall} a single full network is carefully trained. Sub-networks inherit weights from the network and can be directly deployed without training from scratch. An accuracy predictor is trained after training supernet to guide the process for searching a specialized sub-network. FBNetV3 \cite{fbnetv3} trained a predictor on a proxy dataset. The accuracy predictor estimates performance of a candidate sub-network. However, it is still time-consuming to train an accuracy predictor.
In this work, we focus on improving the \emph{search strategy} in terms of both efficiency and flexibility. Note that our search strategy can be incorporated with the methods mentioned above.
\section{Searching by Generating NAS}
\subsection{Background}
\begin{comment}
In Differentiable NAS, we can formulate neural architecture search problem as a bi-level optimization:
\begin{equation}
\operatorname*{min}_{a \in A}\mathcal{L}_{val}(w^*, a)
\label{eq:bi-op_1}
\end{equation}
\begin{equation}
\text{s.t. }w^*(a) = \operatorname*{argmin}_{w}\mathcal{L}_{train}(w, a)
\label{eq:bi-op_2}
\end{equation}
where $a$ denotes the sub-network of supernet $w$ denotes the weight of the supernet.
\end{comment}
Given a supernet $A$ represented by weights $\boldsymbol{w}$, to find an architecture that achieves the best performance while meeting a specific hardware constraint, we need to find the best sub-network $a^*$ from $A$ which achieves the minimum validation loss $\mathcal{L}_{val}(a, \boldsymbol{w})$. Sampling $a$ from $A$ is a non-differentiable process. To optimize $a$ by the gradient descent algorithm, DNAS \cite{fbnet}\cite{proxylessnas}\cite{darts} relaxes the problem as finding a set of continuous architecture parameters $\boldsymbol{\alpha}$, and computes the weighted sum of outputs of candidate blocks by the Gumbel softmax function \cite{gumbel_softmax}:
\begin{equation}
x_{l+1} = \sum_i m^i_l\cdot b^i_l(x_l),
\label{eq:gradient_op1}
\end{equation}
\begin{equation}
m^i_l = \frac{exp(\alpha^i_l+g^i_l/\tau)}{\sum^K_{k=1}exp(\alpha^k_l+g^k_l/\tau)},
\label{eq:gradient_op2}
\end{equation}
where $x_l$ is the input tensor of the $l$th layer, $b_l^i$ is the $i$th block of the $l$th layer, and thus $b^i_l(x_l)$ denotes the output of the $i$th block. The term $\alpha_l^i$ is the weight of the $i$th block in the $l$th layer. The term $g^i_l$ is a random variable sampled from the Gumbel distribution $Gumbel(0, 1)$ and $\tau$ is the temperature parameter. The value $m_l^i$ is the weight for the output $b^i_l(x_l)$.
After relaxation, DNAS can be formulated as a bi-level optimization:
\begin{equation}
\boldsymbol{\alpha}^* = \operatorname*{min}_{\boldsymbol{\alpha}}\mathcal{L}_{val}(\boldsymbol{w}^*, \boldsymbol{\alpha})
\label{eq:bi-op_3}
\end{equation}
\begin{equation}
\text{s.t. }\boldsymbol{w}^* = \operatorname*{argmin}_{\boldsymbol{w}}\mathcal{L}_{train}(\boldsymbol{w}, \boldsymbol{\alpha})
\label{eq:bi-op_4}
\end{equation}
where $\mathcal{L}_{train}(\boldsymbol{w}, \boldsymbol{\alpha})$ is the training loss.
Because of the bi-level optimization of $\boldsymbol{w}$ and $\boldsymbol{\alpha}$, the best architecture $\boldsymbol{\alpha}^*$ sampled from the supernet is only suitable to a specific hardware constraint. With this searching process, for $N$ different hardware constraints, the supernet and architecture parameters should be retrained for $N$ times. This makes DNAS less flexible.
In contrast, single-path methods \cite{uniform_sampling}\cite{fairnas}\cite{scarlet}\cite{mixpath} decouple supernet training from architecture searching. For supernet training, only a single path consisting of one block in each layer is activated and is optimized in one iteration to simulate discrete neural architecture in the search space. We can formulate the process as:
\begin{equation}
\boldsymbol{w}^* = \operatorname*{argmin}_{\boldsymbol{w}}\mathbb{E}_{a\sim \Gamma(A)}(\mathcal{L}_{train}(\boldsymbol{w}(a)))
\label{eq:training_supernet}
\end{equation}
where $\boldsymbol{w}(a)$ denotes the subset of $\boldsymbol{w}$ corresponding to the sampled architecture $a$, and $\Gamma(A)$ is a prior distribution of $a \in A$. The best weights $\boldsymbol{w}^*$ to be determined are the ones yielding the minimum expected training loss. After training, the supernet is treated as a performance estimator to all architectures in the search space. With the pretrained supernet weights $w^*$, we can search the best architecture $a^*$:
\begin{equation}
a^* = \operatorname*{argmin}_{a \in A}\mathcal{L}_{val}(\boldsymbol{w}^*(a)).
\end{equation}
Single-path methods are more flexible than DNAS, because supernet training and architecture search are decoupled. Once the supernet is trained, for $N$ different constraints, only architecture search should be conducted for $N$ times.
In this work, we propose to decouple supernet training from architecture searching and train supernet as in single-path NAS (Eq.~(\ref{eq:training_supernet})). After supernet training, we search the best architecture by the gradient descent algorithm as in DNAS (Eq.~(\ref{eq:bi-op_3})). Instead of training architecture parameters for one specific hardware constraint, we propose a novel search strategy called \emph{architecture generator} to largely increase flexibility and efficiency.
\subsection{Architecture Generator}
\subsubsection{Essential Idea}
Given the target hardware constraint $C$, the architecture generator can generate the best architecture parameters for the hardware constraint $C$. The process of the architecture generator can be described as $\boldsymbol{\alpha} = G(C)$ such that $Cost(\boldsymbol{\alpha}) < C$. With the architecture generator $G$, the objective function of the architecture searching in Eq.~(\ref{eq:bi-op_3}) can be reformulated as :
\begin{equation}
G^* = \operatorname*{min}_{G}\mathcal{L}_{val}(\boldsymbol{w}^*, \boldsymbol{\alpha}).
\label{eq:generator_ce}
\end{equation}
To make $G^*$ generate the best architecture parameters for different hardware constraints accurately, we propose the hardware constraint loss $\mathcal{L}_C$ as:
\begin{equation}
\mathcal{L}_C(\boldsymbol{\alpha}, C) = (Cost(\boldsymbol{\alpha}) - C)^2,
\end{equation}
where the cost yielded by the generated architecture $Cost(\boldsymbol{\alpha})$ is estimated by:
\begin{equation}
Cost(\boldsymbol{\alpha}) = \sum_{l}\sum_{i}m^i_l\cdot Cost(b^i_l).
\label{eq:hc_count}
\end{equation}
The term $Cost(b^i_l)$ is the constant cost of the $i$th block in the $l$ layer and $m^i_l$ is the weight of different blocks described in Eq.~(\ref{eq:gradient_op2}). The cost $Cost(\boldsymbol{\alpha})$ is differentiable with respect to $m^i_l$ and $\boldsymbol{\alpha}$, similarly in \cite{fbnet}\cite{tfnas}. Note that Eq.~(\ref{eq:hc_count}) is also highly correlated with latency, as
mentioned in \cite{fbnet} and \cite{tfnas}. By combining the hardware constraint loss $\mathcal{L}_C$ and the cross entropy loss $\mathcal{L}_{val}$ defined in Eq.~(\ref{eq:generator_ce}), the overall loss of the architecture generator $\mathcal{L}_G$ is:
\begin{equation}
\mathcal{L}_G = \mathcal{L}_{val}(\boldsymbol{w}^*, \boldsymbol{\alpha}) + \lambda\mathcal{L}_C(\boldsymbol{\alpha}, C).
\label{eq:loss_g_1}
\end{equation}
where $\lambda$ is a hyper-parameter to trade-off the validation loss and hardware constraint loss.
\subsubsection{Accurate Generation with Random Prior}
\label{sec:randomprior}
In practice, we found that the architecture generator easily overfits to a specific hardware constraint. The reason is that it is too difficult to generate complex and high-dimensional architecture parameters based on a given simple integer hardware constraint $C$.
To address this issue, a prior is given as input to stabilize the architecture generator. We randomly sample a neural architecture from the search space, and encode the neural architecture into a one-hot vector to be the prior knowledge of architecture parameters. We name it as a random prior $B = {B_1, ..., B_L}$. Formally, $B_l = one\_hot(a_l)$, $l = 1, ..., L$, where $a_l$ is the $l$th layer of the neural architecture $a$ randomly sampled from $A$, and $L$ is the total number of layers in the supernet. With the random prior, the architecture generator is to learn the residual from the random prior to the best architecture parameters, making training architecture generator more stable and accurately (blue line in Fig.~\ref{fig:generator_random prior}(a)), and the process of the architecture generator can be reformulated as $\boldsymbol{\alpha} = G(C, B)$ such that $Cost(\boldsymbol{\alpha}) < C$.
\subsubsection{The Generator Training Algorithm}
We illustrate the algorithm of architecture generator training in Algorithm~\ref{alg:training_generator}. In each iteration, given the target constraint and the random prior, the architecture generator can generate the architecture parameters $\boldsymbol{\alpha}$ (as illustrated in Fig.~\ref{fig:SGNAS_overview}). With $\boldsymbol{\alpha}$, the corresponding cost $C_{\boldsymbol{\alpha}}$ can be calculated by Eqn.~(\ref{eq:hc_count}). We can predict $\hat{y}$ based on the pretrained supernet $N$ with $\boldsymbol{\alpha}$. The total loss is given by Eqn.~(\ref{eq:loss_g_1}). No matter what constraint is given, the architecture generator generates architecture parameters to get the best prediction results. Therefore, training the generator is equivalent to searching the best architectures for various constraints in the proposed SGNAS.
\subsubsection{The Architecture of the Generator}
Fig.~\ref{fig:generator_structure} illustrates the architecture of the generator. We set the channel size of all convolutional layers as 32 and set the stride as 1, making sure the output shape same as the shape of the random prior. Please refer to supplementary materials for detailed configurations of the architecture generator and random prior representations.
\begin{algorithm}[h]
\caption{Training Architecture Generator}
\label{alg:training_generator}
\begin{algorithmic}[1]
\Require
$B$: Random prior;
$N$: Unified supernet;
$G$: Generator;
$[C_L, C_H]$: Pre-define hardware constraint interval;
$D_{val}$: Validation dataset;
$T$: Max iterations;
\For{$t=1,...,T$};
\State Get a data batch $X$ and $y$ from $D_{val}$
\State Randomly sample $C_{target}$ from $[C_L, C_H]$
\State $\boldsymbol{\alpha} = G(B, C_{target})$
\State $C_{\boldsymbol{\alpha}} = Cost(\boldsymbol{\alpha})$
\State $\hat{y} = N(X, \boldsymbol{\alpha})$
\State $\mathcal{L}_{total} = \mathcal{L}_{val}(y, \hat{y}) + \lambda\mathcal{L}_C(C_{target}, C_{\boldsymbol{\alpha}})$
\State Calculate gradients from $\mathcal{L}_{total}$
\State Update $G$ from gradients
\EndFor
\end{algorithmic}
\end{algorithm}
\begin{figure}[t]
\begin{center}
\includegraphics[width=1\linewidth]{image/generator_architecture.png}
\end{center}
\caption{Structure of the architecture generator. Given the input target hardware constraint, the expansion layer expands the input to a tensor with shape same as the random prior.}
\label{fig:generator_structure}
\end{figure}
\subsection{Unified Supernet}
Previous single-path NAS \cite{fairnas}\cite{scarlet}\cite{uniform_sampling}\cite{greedynas} adopts the MobilenetV2 inverted bottleneck \cite{mobilenetv2} as the basic building block. Given the input tensor $X$, the corresponding output $Y_{out}$ is obtained by
\begin{equation}
Y_{out} = P^{c_3, c_2}(D_{K\times K}(P^{c_2, c_1}(X))),
\label{eq:basic_block}
\end{equation}
\begin{comment}
\begin{equation}
c_2 = e \times c_1,
\label{eq:pre_expansion}
\end{equation}
\end{comment}
where $P^{c_j,c_i}$ denotes the pointwise convolution with the input channel size $c_i$ and output channel size $c_j$, $D_{K\times K}$ denotes the depthwise convolution with $K\times K$ kernel size.
Eq.~(\ref{eq:basic_block}) represents that $X$ of $c_1$ channels is first expanded to a tensor of $c_2$ channels, which can be described as $c_2 = e \times c_1$, and then a depthwise convolution is conducted. The term $e$ denotes the expansion rate of the inverted bottleneck. After that, the tensor of $c_2$ channels is embedded into the output tensor of $c_3$ channels. Because one basic building block can only represent one configuration with one kernel size and one expansion rate, previous single-path NAS \cite{proxylessnas}\cite{scarlet}\cite{fairnas} needs to construct blocks of various configurations in each layer, which leads an exponential increase in parameter numbers and complexity of the supernet.
In this work, we propose a novel single-path supernet called \emph{unified supernet} to improve efficiency and flexibility of the architecture generator. The only type of block, i.e., \emph{unified block}, is constructed in each layer. The unified block is built with only the maximum expansion rate $e_{max}$, i.e., $c_2 = e_{max} \times c_1$.
Fig.~\ref{fig:SGNAS_supernet} illustrates the idea of a unified block. To make the unified block represent all possible configurations, we replace the depthwise convolution $D$ by $e_{max}$ sub-blocks, and each sub-block can be implemented by different operations or skip connection. The output tensor of the first pointwise convolution $Y_1$ is equally split into $e_{max}$ parts, $Y_{1,1}, Y_{1,2}, ..., Y_{1,e_{max}}$. With the sub-blocks $d_i$ and split tensors $Y_{1,i}$, we can reformulate $D_{K \times K}$ in Eq.~(\ref{eq:basic_block}) as:
\begin{equation}
d_1(Y_{1,1}) \circ d_2(Y_{1,2}) \circ \cdots \circ d_{e_{max}}(Y_{1,e_{max}}),
\end{equation}
where $\circ$ denotes the channel concatenation function.
With the sub-blocks implemented by different operations, we can simulate blocks with various expansion rates, as shown in Fig.~\ref{fig:expansion_simulation}. The unified supernet thus can significantly reduce the parameters and GPU memory consumption. It is interesting that the MixConv in MixNet \cite{mixconv} is a special case of our search space if different sub-blocks are implemented by different kernel sizes.
\begin{figure}
\begin{center}
\includegraphics[width=1\linewidth]{image/SGNAS_supernet.png}
\end{center}
\caption{Illustration of a unified block. Given the input, the first pointwise convolution expands the input channel size to $e_{max}$ times. The channel split layer splits the tensor into $e_{max}$ parts and feeds to each sub-block, respectively. }
\label{fig:SGNAS_supernet}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=1\linewidth]{image/expansion_simulation.png}
\end{center}
\caption{Different expansion rates can be simulated by sub-blocks with different operations. For example, the expansion rate 6 can be simulated if no skip connection is implemented, and the expansion rate 2 can be simulated if four skip connections are implemented. }
\label{fig:expansion_simulation}
\end{figure}
\begin{comment}
\begin{figure}
\begin{center}
\includegraphics[width=1\linewidth]{image/architecture_redundancy.png}
\end{center}
\caption{Illustration of architecture redundancy and forced sampling. In the four unified blocks, four depthwise convolution with kernel size $3\times 3$ and two skip connections are used in different sub-blocks. However, the three situations on the left are treated as different because of different arrangements. With FS, we enforce arrangement of the operations to be unique.}
\label{fig:architecture_redundancy}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.8\linewidth]{image/variability_statistics.png}
\end{center}
\caption{One-Shot MixConv Superblock.}
\label{fig:variability_statistics}
\end{figure}
\end{comment}
\subsubsection{Large Variability of BNs Statistics}
As in \cite{atomnas}, \cite{universlimm}, and \cite{aows}, we suffer from the problem of unstable running statistics of batch normalization (BN). In the unified supernet, because one unified block would represent different expansion rates, the BN scales change more dramatically during training. To address the problem, BN recalibration \cite{atomnas}\cite{universlimm}\cite{aows} is used to recalculate the running statistics of BNs by forwarding thousands of training data after training. On the other hand, shadow batch normalization (SBN) \cite{mixpath} or switchable batch normalization \cite{slimmable} are used to stabilize BN. In this work, we utilize SBN to address the large variability issue, as illustrated in Fig.~\ref{fig:expansion_simulation}. In our setting, there are five different expansion rates, i.e., 2, 3, 4, 5, and 6. We thus take five BNs after the second pointwise convolution block to capture the BN statistics for different expansion rates. With SBN, we can capture different statistics and make supernet training more stable.
\begin{comment}
In our search space, there are total of 5 different expansion rates(e.g., {2, 3, 4, 5, 6}), so we take 5 BNs after the second pointwise convolution block to capture the BN statistic for different expansion rates.
And we illustrate SBN in the unified block in Fig.\ref{fig:expansion_simulation}. With SBN, we can capture the different statistics and make supernet training more stable.
\end{comment}
\begin{comment}
In Fig.\ref{fig:variability_statistics}, we plot the curve of the BN running variance and the BN running mean of each expansion rate. We can found that the greater the difference in expansion rate, the greater the difference in BNs statistic. With SBN, we can capture the different statistics and make supernet training more stable.
\end{comment}
\subsubsection{Architecture Redundancy}
Denote two sub-blocks as $b_1$ and $b_2$. In the unified supernet, for example, the case of $b_1$ using $3 \times 3$ kernel size and $b_2$ using skip connection is distinct from the case of $b_1$ using skip connnection and $b_2$ using $3 \times 3$ kernel size. However these two cases actually correspond to the same sub-network, and thus the architecture redundancy problem arises. This redundancy makes the unified supernet more complex and hard to train. To address this issue, we force that skip connection can only be used in sub-blocks with higher index. For example, if we want to train a unified block with expansion rate 3, only the last three sub-blocks can be skip connection. We call this strategy forced sampling (FS). Please refer to supplementary materials for details of architecture redundancy and forced sampling.
\begin{comment}
\begin{table}[]
\caption{Macro-architecture of the search space. MBConv $3\times 3$ denotes MobileNetV2 block with kernel size 3. Column-C denotes the output channel number of a block. Column-C denotes the output channel number of a block. Column-N denotes the number of the blocks. Column-S denotes the stride of of first block when stacked for multiple blocks. Column-E denotes the expansion rate of the blocks, and the tuples of three values represent the lowest value, highest, and steps between options (low, high, steps).}
\begin{center}
\begin{tabular}{c|c|c|c|c|c}
\hline
Input shape & Block & C & N & S & E \\ \hline
$224^2 \times 3$ & Conv 3$\times$3 & 32 & 1 & 2 & - \\
$112^2 \times 32$ & MBConv 3$\times$3 & 16 & 1 & 1 & 1 \\ \hline
$112^2 \times 16$ & Unified Block & 32 & 2 & 2 & (2, 6, 1) \\
$56^2 \times 32$ & Unified Block & 40 & 4 & 2 & (2, 6, 1) \\
$28^2 \times 40$ & Unified Block & 80 & 4 & 2 & (2, 6, 1) \\
$14^2 \times 80$ & Unified Block & 96 & 4 & 1 & (2, 6, 1) \\
$14^2 \times 96$ & Unified Block & 192 & 4 & 2 & (2, 6, 1) \\
$7^2 \times 192$ & Unified Block & 320 & 1 & 1 & (2, 6, 1) \\
$7^2 \times 320$ & Unified Block & 1280& 1 & 1 & (2, 6, 1) \\ \hline
$7^2 \times 1280$ & Avg pool & - & 1 & 1 & - \\
$1280$ & FC & 1000& 1 & - & - \\ \hline
\end{tabular}
\end{center}
\end{table}
\end{comment}
\section{Experiments}
\subsection{Experimental Settings}
We adopt the macro structure of supernet (e.g., channel size of each layer and layer number) same as \cite{fairnas} and \cite{proxylessnas}, but utilize the proposed unified blocks to reduce GPU memory consumption and the number of parameters. Each sub-block in the unified blocks can be realized based on convolutional kernel sizes 3, 5, or 7, or the skip connection. We set the minimum and maximum expansion rates as 2 and 6, respectively. The size of our search space is $80^{19}$. Please refer to supplementary materials for more details.
\begin{comment}
We perform all experiments based on the ImageNet dataset \cite{imagenet}. Same as the settings in \cite{proxylessnas}\cite{greedynas}\cite{scarlet}, we randomly sample 50,000 images (50 images for each class) from the training set as our validation set, and the rest is kept as our training set. The original validation set is taken as our test set to measure the final performance of each model.
\end{comment}
For experiments on the ImageNet dataset \cite{imagenet}, we train the unified supernet for 50 epochs using batch size 256 and adopt the stochastic gradient descent optimizer.
\begin{comment}
with a momentum of 0.9 and weight decay of $4 \times 10^{-5}$.
\end{comment}
The learning rate is decayed with the cosine annealing strategy \cite{cosine_ann} from the initial value 0.045. After supernet training, the architecture generator is trained for 50 epochs by the Adam optimizer with the learning rate 0.001.
\begin{comment}
The temperature $\tau$ of Gumbel softmax \cite{gumbel_softmax} in Eq.~(\ref{eq:gradient_op2}) is
initially set as 5.0 and annealed by a factor of 0.95 for each epoch, and the trade-off parameter $\lambda$ in Eq.~(\ref{eq:loss_g_1}) is set as 0.002 in our experiment.
\end{comment}
After searching/generating the best architecture under hardware constraints, we adopt the RMSProp optimizer with 0.9 momentum \cite{mnasnet} to train the searched architecture from scratch. Learning rate is increased from 0 to 0.16 in the first 5 epochs with batch size 256, and then decays 0.03 every 3 epochs.
\subsection{Experiments on ImageNet}
\subsubsection{Comparison with Baselines}
Li and Talwalkar \cite{randomsearch} presented that a random search approach usually achieves satisfactory performance. To make comparison, we randomly select 1,000 candidate architectures with FLOPs under 320 millions (320M) from the unified supernet and pick the architecture yielding the highest top-1 accuracy, as mentioned in \cite{randomsearch}. Besides, we also search the network with FLOPs under 320M by the evolution algorithm \cite{uniform_sampling} as another baseline.
Table~\ref{table:baseline} shows the comparison results. As can be seen, with around 320M FLOPs, the proposed SGNAS achieves the highest top-1 accuracy. Both baselines take around 34 GPU hours to complete the search. For $N$ different hardware constraints, the search strategy should be re-executed for $N$ times, and the search time of each of two baselines is $34N$ GPU hours totally. In contrast, SGNAS only takes 5 GPU hours totally for $N$ different hardware constraints, which is much more efficient and flexible than the baselines.
\begin{table}[]
\caption{Performance comparison with baselines.}
\centering
\vspace{1mm}
\begin{tabular}{c|l|l|c|c|c}
\multicolumn{3}{c|}{Search Strategy} & \begin{tabular}[c]{@{}c@{}}Search Time\\ (GPU hrs)\end{tabular} & \begin{tabular}[c]{@{}c@{}}FLOPs\\ (M)\end{tabular} & \begin{tabular}[c]{@{}c@{}}Top-1\\ (\%)\end{tabular} \\ \hline
\multicolumn{3}{c|}{Random search} & $34N$ & 322 & 74.63 \\
\multicolumn{3}{c|}{Evolution algorithm} & $34N$ & 318 & 74.67 \\ \hline
\multicolumn{3}{c|}{\textbf{SGNAS}} & \textbf{5} & 324 & \textbf{74.87}
\end{tabular}
\label{table:baseline}
\end{table}
\begin{comment}
\begin{table}[]
\caption{Comparison of performance with the different search strategy and the supernet. RS: random search. EA: evolution algorithm. Our's : architecture generator.}
\centering
\vspace{1mm}
\begin{tabular}{c|ccc|c|c}
\multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Unified\\ Supernet\end{tabular}} & \multicolumn{3}{c|}{Search Strategy} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Search Time\\ (GPU hrs)\end{tabular}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Top-1\\ (\%)\end{tabular}} \\ \cline{2-4}
& \multicolumn{1}{c|}{RS} & \multicolumn{1}{c|}{EA} & Our's & & \\ \hline
$\times$ & $\checkmark$ & & & & \\
$\times$ & & $\checkmark$ & & & \\
$\times$ & & & $\checkmark$ & 11 & \\ \hline
$\checkmark$ & $\checkmark$ & & & $34N$ & 74.63 \\
$\checkmark$ & & $\checkmark$ & & $34N$ & 74.67 \\
$\checkmark$ & & & $\checkmark$ & \textbf{5} & \textbf{74.87}
\end{tabular}
\end{table}
\end{comment}
\subsubsection{Comparison with SOTAs}
\begin{table}[]
\centering
\caption{Comparison with the SOTAs for different hardware constraints. $^\dagger$: training with AutoAugment \cite{autoaugment}. $^\ddagger$: searching on a proxy dataset. The unit of search time and train time is GPU hours. }
\vspace{1mm}
\begin{tabular}{l|c|c|c|c}
Method & \begin{tabular}[c]{@{}c@{}}FLOPs\\ (M)\end{tabular} & \begin{tabular}[c]{@{}c@{}}Top-1\\ (\%)\end{tabular} &
\begin{tabular}[c]{@{}c@{}}Train\\ time\end{tabular} &
\begin{tabular}[c]{@{}c@{}}Search \\ time\end{tabular} \\ \hline
MobileNetV2 \cite{mobilenetv2} & 300 & 72.0 & -- & -- \\\hline
EfficientNet B0 \cite{efficientnet} & 390 & 76.3 & -- & -- \\
MixNet-M \cite{mixconv} & 360 & 77.0 & -- & -- \\
MixPath-A \cite{mixpath} & 349 & 76.9 & 240 & -- \\
AtomNAS-C \cite{atomnas} & 363 & 77.6 & 0 & $816N$\\
PC-DARTS \cite{pcdarts:} & 597 & 75.8 & 0 & $91N$ \\
ScarletNAS-A \cite{scarlet} & 365 & 76.9 & 240 & $48N$ \\
GreedyNAS-A \cite{greedynas} & 366 & 77.1 & 168 & $\sim 24N$ \\
\textbf{SGNAS-A (Ours)} & 373 & 77.1 & 280 & \textbf{5} \\ \hline
FBNetV2-L1 \cite{fbnetv2} & 325 & 77.2 & 0 & $600N^\ddagger$\\
Proxyless-R \cite{proxylessnas} & 320 & 74.6 & 0 & $200N$\\
FairNAS-C \cite{fairnas} & 325 & 76.7$^\dagger$ & 240 & $48N$\\
ScarletNAS-B \cite{scarlet} & 329 & 76.3 & 240 & $48N$ \\
SPOS \cite{uniform_sampling} & 326 & 74.5 & 288 & $\sim 24N$ \\
GreedyNAS-B \cite{greedynas} & 324 & 76.8 & 168 & $\sim 24N$ \\
\textbf{SGNAS-B (Ours)} & 326 & 76.8 & 280 & \textbf{5} \\ \hline
MobileNetV3-L \cite{mobilenetv3} & 219 & 75.2 & -- & -- \\
ScarletNAS-C \cite{scarlet} & 280 & 75.6 & 240 & $48N$ \\
GreedyNAS-C \cite{greedynas} & 284 & 76.2 & 168 & $\sim 24N$ \\
\textbf{SGNAS-C (Ours)} & 281 & 76.2 & 280 & \textbf{5}
\end{tabular}
\label{table:sota}
\end{table}
This section is dedicated to compare with various SOTA one-shot NAS methods that utilize the augmented techniques (e.g., Swish activation function \cite{swish} and Squeeze-and-Excitation \cite{se}). We directly modify the searched architecture by replacing all ReLU activation with H-Swish \cite{mobilenetv3} activation and equip it with the squeeze-and-excitation module as in AtomNAS \cite{atomnas}.
For comparison, similar to the settings in ScarletNAS \cite{scarlet} and GreedyNAS \cite{greedynas}, we search architectures under 275M, 320M, and 365M FLOPs, and denote the searched architecture as SGNAS-C, SGNAS-B, and SGNAS-A, respectively. The comparison results are shown in Table~\ref{table:sota}. The column \enquote{Train time} denotes that the time needed to train the supernet, and the column \enquote{Search time} denotes that the time needed to search the best architecture based on the pre-trained supernet. Because DNAS couples architecture searching with supernet optimization, we list the time needed for the entire pipeline in the \enquote{Search time} column. As can be seen, our SGNAS is competitive with SOTAs in terms of top-1 accuracy under different FLOPs. For example, SGNAS-A achieves 77.1\% top-1 accuracy, which outperforms ScarletNAS \cite{scarlet} by 0.2\%, outperforms MixNet-M \cite{mixconv} by 0.1\%, outperforms MixPath-A \cite{mixpath} by 0.2\%, and is comparable with GreedyNAS-A \cite{greedynas}.
\begin{comment}
Our SGNAS-C achieves 76.2\% top-1 accuracy with 280M FLOPs which outperforms ScarletNAS \cite{scarlet}(+0.6)\% and comparable with GreedyNAS \cite{greedynas}. SGNAS-B achieves 76.2\% top-1 accuracy with 320M FLOPs which outperforms ScarletNAS \cite{scarlet}(+0.5\%), FairNAS-C \cite{fairnas}(+0.1\%), and comparable with GreedyNAS \cite{greedynas}. SGNAS-A achieves 77.1\% top-1 accuracy with 360M which outperforms ScarletNAS \cite{scarlet}(+0.2\%), MixNet-M \cite{mixconv}(+0.1\%), MixPath-A \cite{mixpath}(+0.2\%), and comparable with GreedyNAS \cite{greedynas} respectively.
\end{comment}
More importantly, SGNAS achieves much higher search efficiency. With the architecture generator and the unified supernet, even for $N$ different architectures under $N$ different hardware constraints, totally only 5 GPU hours are needed for SGNAS on a Tesla V100 GPU. However, FairNAS \cite{fairnas}, GreedyNAS \cite{greedynas}, and ScarletNAS \cite{scarlet} need $48N$, $24N$, and $48N$ GPU hours, respectively, because of the cost of re-executing search. Supernet retraining is needed for FBNetV2 \cite{fbnetv2} and AtomNAS \cite{atomnas}, which makes search very inefficient.
Note that after finding the best architecture, training from scratch is required in most methods in Table~\ref{table:sota} (including SGNAS, except for AtomNAS \cite{atomnas}). However, training a supernet that can be directly deployed to many constraints (like AtomNAS) needs expensive computation. Even with the time for training from scratch, SGNAS is still more efficient and flexible than AtomNAS.
\subsection{Experiments on NAS-Bench-201}
\begin{table*}[]
\caption{Performance comparison on different datasets in the NAS-Bench-201 benchmark.}
\centering
\footnotesize
\begin{tabular}{c|c|cc|cc|cc}
\multirow{2}{*}{Method} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Search Time\\ (GPU hrs)\end{tabular}} & \multicolumn{2}{c|}{CIFAR-10} & \multicolumn{2}{c|}{CIFAR-100} & \multicolumn{2}{c}{ImageNet-16-120} \\ \cline{3-8}
& & Val & Test & Val & Test & Val & Test \\ \hline
optimal & N/A & 91.61 & 94.37 & 73.49 & 73.51 & 46.77 & 47.31 \\ \hline
RSPS \cite{randomsearch} & 2.6 & 84.16$\pm$1.69 & 87.66$\pm$1.69 & 59.00$\pm$4.60 & 58.33$\pm$4.34 & 31.56$\pm$3.28 & 31.14$\pm$3.88 \\
DARTS \cite{darts} & $2.2N$ & 39.77$\pm$0.00 & 54.30$\pm$0.00 & 15.03$\pm$0.00 & 15.61$\pm$0.00 & 16.43$\pm$0.00 & 16.32$\pm$0.00 \\
SETN \cite{setn} & $7.9N$ & 82.25$\pm$5.17 & 86.19$\pm$4.63 & 56.86$\pm$7.59 & 56.87$\pm$7.77 & 32.54$\pm$3.63 & 31.90$\pm$4.07 \\
GDAS \cite{gdas} & $6.6N$ & 90.00$\pm$0.21 & 93.51$\pm$0.13 & \textbf{71.14$\pm$0.27} & \textbf{70.61$\pm$0.26} & 41.70$\pm$1.26 & 41.84$\pm$0.90 \\
\textbf{SGNAS (Ours)} & \textbf{2.5} & \textbf{90.18$\pm$0.31} & \textbf{93.53$\pm$0.12} & 70.28$\pm$1.2 & 70.31$\pm$1.09 & \textbf{44.65$\pm$2.32} & \textbf{44.98$\pm$2.10}
\end{tabular}
\label{table:nas-bench-201}
\end{table*}
To demonstrate efficiency and robustness of SGNAS more fairly, we evaluate it based on a NAS benchmark dataset called NAS-Bench-201 \cite{NAS-Bench-201:}. NAS-Bench-201 includes 15,625 architectures in total. It provides full information of the 15,625 architectures (e.g., top-1 accuracy and FLOPs) on CIFAR-10, CIFAR-100, and ImageNet-16-120 datasets \cite{iamgenet16}, respectively.
Based on the search space defined by NAS-Bench-201, we follow SETN \cite{setn} to train the supernet by uniform sampling. After that, the architecture generator is applied to search architectures on the supernet. We search based on the CIFAR-10 dataset and look up the ground-truth performance of the searched architectures on CIFAR-10, CIFAR-100, and ImageNet-16-120 datasets, respectively. This process is run for three times, and the average performance is calculated as in Table~\ref{table:nas-bench-201}. We see that the architectures searched by SGNAS outperform previous methods on both CIFAR-10 and ImageNet16-120. It is worth noting that, with the supernet training strategy same as SETN \cite{setn}, our result greatly surpasses SETN \cite{setn} on all three datasets. Moreover, the required search time of SGNAS is only 2.5 GPU hours even for $N$ different hardware constraints.
We show the 15,625 architectures in NAS-Bench-201 on each dataset as gray dots in Fig.~\ref{fig:nas_bench_201}, and draw the architectures searched by the architecture generator under different FLOPs as blue rectangles. After searching once, the architecture generator can generate all blue rectangles directly without re-searching. Moreover, various generated architectures approach the best among all choices.
\begin{figure}[t]
\centering
\subfigure[]{
\begin{minipage}[t]{0.32\linewidth}
\centering
\includegraphics[width=\textwidth]{image/SGNAS_bench_201_cifar10.png}
\end{minipage}%
\label{fig:SGNAS_bench_201_cifar10}
}%
\subfigure[]{
\begin{minipage}[t]{0.32\linewidth}
\centering
\includegraphics[width=\textwidth]{image/SGNAS_bench_201_cifar100.png}
\end{minipage}%
\label{fig:SGNAS_bench_201_cifar100}
}%
\subfigure[]{
\begin{minipage}[t]{0.32\linewidth}
\centering
\includegraphics[width=\textwidth]{image/SGNAS_bench_201_imagenet16.png}
\end{minipage}%
\label{fig:SGNAS_bench_201_imagenet16}
}%
\caption{Search results of SGNAS on the CIFAR10, CIFAR100, and ImageNet16-120 datasets. (a) Result on CIFAR-10; (b) Result on CIFAR100; (c) Result on ImageNet16-120.}
\label{fig:nas_bench_201}
\centering
\end{figure}
\subsection{Performance on Object Detection}
\begin{table}[]
\caption{Performance comparison on the COCO object detection. $^\dagger$: Our implementation result. $^*$ reported in \cite{scarlet}\cite{fairnas}.}
\centering
\footnotesize
\begin{tabular}{c|cc|c}
Model & FLOPs(M) & Top-1 (\%) & mAP (\%) \\ \hline
MobileNetV2$^*$ \cite{mobilenetv2} & 300 & 72.0 & 28.3 \\
MixNet-M$^*$ \cite{mixconv} & 360 & 77.0 & 31.3 \\
FairNAS-A$^*$ \cite{fairnas} & 392 & 77.5 & 32.4 \\
Scarlet-A$^*$ \cite{scarlet} & 365 & 76.9 & 31.4 \\ \hline
MobileNetV2$^\dagger$ & 300 & 72.0 & 29.4 \\
\textbf{SGNAS-A (Ours)} & 373 & 77.1 & 33.9
\end{tabular}
\label{tb:objectdetection}
\end{table}
To verify the transferability of SGNAS on object detection, we adopt the RetinaNet \cite{retinanet} implemented in MMDetection \cite{mmdetection} to do object detection, but replace its backbone by the network searched by SGNAS. The models are trained and evaluated on the MS COCO dataset \cite{coco} (train2017 and val2017, respectively) for 12 epochs with batch size 16 \cite{fairnas}\cite{scarlet}. We use the SGD optimizer with 0.9 momentum and 0.0001 weight decay. The initial learning rate is 0.01, and is multiplied by 0.1 at epochs 8 and 11. Table~\ref{tb:objectdetection} shows that SGNAS has better transferability than the baselines, especially in terms of mAP.
\section{Ablation Studies}
\subsection{Analysis of SGNAS}
In Fig.~\ref{fig:supernet_evaluate}(a), we randomly sample 360 architectures from the search space and illustrate the corresponding top-1 validation accuracies as blue dots. Moreover, we draw the architectures searched by SGNAS under different hardware constraint as red rectangles. As can be seen, the architectures searched by SGNAS are almost always the best.
\begin{figure}[t]
\centering
\subfigure[]{
\begin{minipage}[t]{0.52\linewidth}
\centering
\includegraphics[width=\textwidth]{image/supernet_evaluation_avg.png}
\end{minipage}%
}%
\subfigure[]{
\begin{minipage}[t]{0.5\linewidth}
\centering
\includegraphics[width=\textwidth]{image/search_space.png}
\end{minipage}%
}%
\caption{(a) Top-1 validation accuracy of randomly sampled architectures (blue dots) and the architectures searched by SGNAS (red rectangles). Performance of other variants is also shown. (b) The relationship between the number of supernet parameters and the number of candidate operations in each layer.}
\label{fig:supernet_evaluate}
\centering
\end{figure}
\begin{comment}
\begin{figure}
\begin{center}
\includegraphics[width=0.8\linewidth]{image/search_space.png}
\end{center}
\caption{Comparison of supernet parameters and candidate operation numbers in each layer. We construct the approximate supernets according to previous single-path methods.}
\label{fig:search_space}
\end{figure}
\end{comment}
\begin{comment}
\textcolor{blue}{
\subsection{Analysis of Architecture Generator}
\subsubsection{The Architecture of the Generator}
We experimented with 2 structures of the architecture generator to validate the generating ability of different architectures. The 2 structures are constructed by convolution layer (Conv) and fully connected layer (FC), respectively. Fig~.\ref{fig:search_space} shows the corresponding top-1 validation accuracies generated by different architecture of the generator under different hardware constraint. We also calculate the Kendall tau rank correlation \cite{kendall} and the Pearson correlation coefficient to measure the correlation.
}
\end{comment}
\subsection{Analysis of Unified Supernet}
\begin{comment}
\begin{figure}
\begin{center}
\includegraphics[width=1\linewidth]{image/supernet_training_curve.png}
\end{center}
\caption{The visulization of trainin and validation curve.}
\label{fig:supernet_training_curve}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=1\linewidth]{image/supernet_evaluation_avg.png}
\end{center}
\caption{The visualization of the validation top-1 accuracy of the architectures randomly sampled from the supernet trained with different methods. And we also visualize the top-1 accuracy and FLOPs of the architectures searched once by SGNAS.}
\label{fig:supernet_evaluation_avg}
\end{figure}
\end{comment}
\subsubsection{Efficiency of Unified Supernet}
\begin{table}[]
\caption{Comparison in terms of of GPU memory consumption and search time of the architecture generator.}
\vspace{1mm}
\centering
\footnotesize
\begin{tabular}{c|c|c|c|c}
\begin{tabular}[c]{@{}c@{}}Unified \\ Supernet\end{tabular} & \begin{tabular}[c]{@{}c@{}}Search\\ Space\end{tabular} & \begin{tabular}[c]{@{}c@{}}Batch\\ Size\end{tabular} & \begin{tabular}[c]{@{}c@{}}Search Time\\ (GPU hrs)\end{tabular} & \begin{tabular}[c]{@{}c@{}}Memory Cost\\ (GPU)\end{tabular} \\ \hline
\multirow{2}{*}{$\times$} & \multirow{2}{*}{$6^{19}$} & 32 & \multirow{2}{*}{11} & 10.5GB \\
& & 128 & & 40 GB \\ \hline
\multirow{2}{*}{$\checkmark$} & \multirow{2}{*}{\textbf{$80^{19}$}} & 32 & \multirow{2}{*}{\textbf{5}} & \textbf{9GB} \\
& & 128 & & \textbf{28GB}
\end{tabular}
\label{table:super_efficient}
\end{table}
To show efficiency of the proposed unified supernet, we report the relationship between the total number of parameters in the (unified) supernet and the number of candidate operations per layer in Fig.~\ref{fig:supernet_evaluate}(b). For fair comparison, we calculate the number of parameters of different supernets all based on 19 layers. As can be seen, the number of possible operations of the unified supernet in each layer is 7 times larger than GreedyNAS \cite{greedynas} and ScarletNAS \cite{scarlet}, but the number of parameters needed to represent this unified supernet is only 1/6 times of them. The number of possible operations is 13 times larger than FairNAS \cite{fairnas} and ProxylessNAS \cite{proxylessnas}, but the number of supernet parameters for the unified supernet is only half of them. To compare under the same size of search space, we estimate the number of required supernet parameters in previous single-path methods \cite{fairnas}\cite{greedynas}\cite{scarlet}\cite{proxylessnas} when the number of possible operations in each layer increases to 25, 50, 60, and 80, and show them by green squares. Fig.~\ref{fig:supernet_evaluate}(b) shows that the required parameters are significantly boosted when the number of possible operations increases, while the unified supernet avoids this intractability. Under the same size of search space, the number of needed parameters to represent the unified supernet is only 1/35 times of estimated supernets.
Table~\ref{table:super_efficient} shows the comparison in terms of GPU memory consumption and search time of the architecture generator when it works based on the unified supernet or based on the previous single-path supernet \cite{fairnas}\cite{proxylessnas}. Based on the unified supernet, the GPU memory consumption reduces to 12 GB and the search time is only $0.45$ times of that based on the previous supernet.
\subsubsection{Training Stabilization}
Although the unified supernet largely reduces supernet parameters, the large search space makes the supernet hard to train. To study the effect of force sampling (FS) and shadow batch normalization (SBN) \cite{mixpath} on supernet training, we train the supernet based on different settings, including baseline, with FS, with SBN, and with both FS and SBN. After training the supernet, we randomly sample 360 architectures from the search space and show the corresponding top-1 accuracies in Fig.~\ref{fig:supernet_evaluate}(a). Without FS and SBN, because of large variability and complex architecture, the baseline supernet is hard to train. After utilizing SBN, variability can be well characterized, and the performance becomes more stable. After applying FS, complexity of the supernet is greatly reduced by reducing architecture redundancy. Performance is largely boosted when redundancy is reduced. With both FS and SBN, the unified supernet can more consistently represent architectures with better performance.
\begin{comment}
In Fig.~\ref{fig:supernet_evaluate}(b), we plot the training curves and validation curves of supernets trained based on different strategies. The supernet without FS and SBN encounter overfitting seriously. With both SBN and FS, the trained supernet is more stable and achieves the highest top-1 accuracy.
\end{comment}
\subsection{Study of Random Priors}
\begin{comment}
\begin{figure}
\begin{center}
\includegraphics[width=0.8\linewidth]{image/generator_random prior_evaluation.png}
\end{center}
\caption{One-Shot MixConv Superblock.}
\label{fig:generator_random prior_evaluation}
\end{figure}
\end{comment}
\begin{figure}[t]
\centering
\subfigure[]{
\begin{minipage}[t]{0.49\linewidth}
\centering
\includegraphics[width=\textwidth]{image/generator_backbone_evaluation.png}
\end{minipage}%
\label{fig:SGNAS_generator_random prior_evaluation}
}%
\subfigure[]{
\begin{minipage}[t]{0.49\linewidth}
\centering
\includegraphics[width=\textwidth]{image/backbone_quality.png}
\end{minipage}%
\label{fig:SGNAS_random prior_quality}
}%
\caption{(a) The relationship between target FLOPs and the FLOPs of generated architecture. (b) Performance of the architectures randomly sampled from the unified supernet (blue dots) and those generated by the architecture generator trained based on four random priors. }
\label{fig:generator_random prior}
\centering
\end{figure}
To enable the generator to generate architectures under various hardware constraints accurately, random prior is given as the input of the generator. In Fig.~\ref{fig:generator_random prior}(a), we show the correlation between the target FLOPs and the FLOPs of the generated architectures. With the random prior, the generator can generate architectures much more accurately.
With the random prior, the Kendall tau correlation between the target FLOPs and the generated is 1, while the Pearson correlation is 0.99, which are significantly positive.
We randomly sample four sub-networks A, B, C, and D from the unified supernet as four priors to train the architecture generator. Inherited from the weights of the unified supernet, the top-1 validation accuracy of these four sub-networks are 58.02\%, 63.36\%, 66.48\%, and 68.48\%, respectively. Fig.~\ref{fig:generator_random prior}(b) shows that, no matter starting from good priors or bad priors, the corresponding trained architecture generators are able to generate the architecture yielding the best performance. This shows that random priors are not to improve the top-1 accuracy, but to give reasonable priors to make the architecture generator generate good architectures under the target constraints.
\begin{comment}
\begin{table}[]
\caption{Comparison of accuracy performance of searched architecture and the GPU memory consumption by our architecture generator w.r.t. different batch size on previous single-path supernet and our unified supernet.}
\centering
\begin{tabular}{c|c|c|c}
Name & Formulation & \begin{tabular}[c]{@{}c@{}}Kendall \\ tau\end{tabular} & \begin{tabular}[c]{@{}c@{}}L2 \\ distance\end{tabular} \\ \hline
FBNet & $\lambda_1 log[(Cost(\alpha , \beta)]^{\lambda_2}$ & & \\
ProxylessNAS & $\lambda_1(Cost(\alpha , \beta))$ & & \\
TF-NAS & $\lambda_1 max(\frac{Cost(\alpha , \beta)}{Cost_{target}}-1, 0)$ & & \\ \hline
Our & $(Cost(\alpha) - Cost_{target})^2$ & &
\end{tabular}
\end{table}
\end{comment}
\section{Conclusion}
To improve efficiency and flexibility of finding best sub-networks from the supernet under various hardware constraints, we propose the idea of architecture generator that searches the best architecture by generating it. This approach is very efficient and flexible for that only one forward pass is needed to generate good architectures for various constraints, comparing to previous one-shot NAS methods. To ease GPU memory consumption and boost searching, we propose the idea of unified supernet which consists of a stack of unified blocks. We show that the proposed one-shot framework, called SGNAS (searching by generating NAS), is extremely efficient and flexible by comparing with state-of-the-art methods. We further comprehensively investigate the impact of architecture generator and unified supernet from multiple perspectives. Please refer to supplementary materials for the limitation of SGNAS.
\textbf{Acknowledgement}
This work was funded in part by Qualcomm through a Taiwan University Research Collaboration Project and in part by the Ministry of Science and Technology, Taiwan, under grants 108-2221-E-006-227-MY3, 107-2923-E-006-009-MY3, and 109-2218-E-002-015.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
This work is concerned with the minimization of the von Neumann entropy
$$
S(\varrho)= \Tr (\varrho \log \varrho),
$$
where $\varrho$ is a nonnegative, trace class operator with trace one, on some infinite-dimensional Hilbert space (we will refer to such operator in the sequel as a \textit{state}, which are self-adjoint since nonnegative). Note the sign change in the entropy compared to the standard definition in the physics literature. The problem is motivated by the work of Nachtergaele and Yau addressed in \cite{nachter}, where they derive the Euler equations of fluid mechanics from quantum dynamics. More precisely, they consider the entropy minimization problem in the context of many-body quantum mechanics where the underlying Hilbert space is the \textit{Fermionic Fock space}. The latter is defined as follows: let $\mathfrak{h}=L^2(\mathbb R^d)$ for $d \geq 1$; the Fermionic Fock space $\mathfrak{F}_f$ is the direct sum
\begin{equation*}
\mathfrak{F}_f : = \bigoplus_{n = 0}^{+\infty} \mathfrak{F}^{(n)}_f,
\end{equation*}
where $\mathfrak{F}_f^{(n)} : = \mathfrak{h}^{\wedge_s n}$ is the $n$-fold antisymmetric tensor product of $\mathfrak{h}$, with the convention $\mathfrak{h}^{\wedge_s 0} = \mathbb{C}$. We have
$$
\mathfrak{F}_f^{(n)}=L^2_a\big((\mathbb R^d)^n\big),
$$
where $L^2_a\big((\mathbb R^d)^n\big)$ is the space of antisymmetric square integrable complex functions on $(\mathbb R^d)^n$, that is, for $x_\ell \in \mathbb R^d$ , $\ell=1,\cdots,n$, $1 \leq, i,j \leq n$, $$f(x_1,\cdots,x_i,\cdots,x_j,\cdots,x_n)=-f(x_1,\cdots,x_j,\cdots, x_i,\cdots,x_n)$$
when $f \in L^2_a((\mathbb R^d)^n)$. This setting is usually referred to as the grand canonical picture since the system is not fixed to a particular subspace with $n$ particles.
Nachtergaele and Yau consider the minimization of $S(\varrho)$ over states $\varrho$ with \textit{prescribed particle density, current and energy densities at any given point $x \in \mathbb R^d$}. This results in an infinite dimensional constrained optimization problem, whose main mathematical difficulty is to handle the local nature of the constraints. The solution can be seen as the quantum many-body equivalent of the classical Maxwellian obtained by minimizing the Boltzmann entropy under local constraints of density, current and energy.
Using the formalism of the second quantization, the local constraints can be defined as follows: let $\{e_i\}_{i \in \mathbb N}$ be an orthonormal basis of $\mathfrak{h}$, let $x \in \mathbb R^d$, and consider (formally) the following family of operators parametrized by $x$,
$$
a_x=\sum_{i \in \mathbb N} e_i^*(x) a(e_i),
$$
where $a(\cdot)$ is the annihilation operator and $e_i^*$ is the complex conjugate of $e_i$. The adjoint of $a_x$ in $\mathfrak F_f$ is denoted by $a^*_x$. We do not give the explicit definition of $a(\cdot)$ since it will not be needed in the sequel, and we point the reader to \cite{arai} for instance for more details. For a state $\varrho$ and $\Tr(\cdot)$ the trace in $\mathfrak{F}_f$, we introduce the following functions of $x$,
\mathbf e} \newcommand{\bh}{\mathbf h \label{consint}
\left\{
\begin{array}{ll}
\displaystyle n[\varrho](x):=\Tr (a^*_x a_x\, \varrho), \qquad &\textrm{local density}\\
\displaystyle u[\varrho](x): = \Im \Tr (a^*_x \nabla a_x\, \varrho),\qquad &\textrm{local current}\\
\displaystyle k[\varrho](x):=\Tr (\nabla a^*_x \cdot \nabla a_x\, \varrho)\qquad &\textrm{local kinetic energy},
\end{array}
\right.
\end{equation}
where the gradient $\nabla$ is taken with respect to the variable $x$, and $\Im$ denotes imaginary part. We defined above the local kinetic energy instead of the total energy which includes (two-body) interactions between the particles. The latter will be defined further on. The formulas in \fref{consint} are similar in structure to the usual definitions of the density, current and kinetic energy for the one-particle setting. We will introduce in the sequel equivalent definitions based on one-particle density matrices that are more amenable to mathematical rigor.
Introducing the potential energy $e_P[\varrho]= V n[\varrho]$ for some potential $V$, and the total energy $e[\varrho]=k[\varrho]+e_P[\varrho]+e_I[\varrho]$ for some two-body interaction term $e_I[\varrho]$, Nachtergaele and Yau assume that the minimization problem with constraints on $n[\varrho]$, $u[\varrho]$ and $e[\varrho]$ admits a unique solution, referred to as a \textit{local Gibbs state} (or more accurately they give an informal expression of the statistical equilibrium that is the solution to the constrained minimization problem). Then they prove that a state $\varrho_t$ solution to the quantum Liouville equation
$$
i \partial_t \varrho_t=[H,\varrho_t], \qquad [H,\varrho_t]=H\varrho_t-\varrho_t H, \qquad H=-\Delta+V+W,
$$
with a local Gibbs state with constraints $\{n_0,u_0,e_0\}$ as initial condition, converges, in an appropriate limit that we do not detail here, to a local Gibbs state with constraints $\{n_0(t),u_0(t),e_0(t)\}$. These latter constraints are solutions to the Euler equations with initial condition $\{n_0,u_0,e_0\}$. Above, $W$ is a two-body interaction potential used to define $e_I[\varrho]$.
In the one-body case, a similar constrained entropy minimization problem is central to the work of Degond and Ringhofer in their derivation of quantum fluid models from quantum dynamics, see \cite{DR}.
Under appropriate conditions on $\{n_0,u_0,e_0\}$, our main result in this work is justify rigorously the introduction of these local Gibbs states, and therefore to prove that indeed the constrained minimization problem admits a unique solution, \textit{for both the fermionic and bosonic cases}. We addressed in \cite{DP-CMP} the one-body problem for various quantum entropies, and the key difficulty is the lack of compactness required to handle the local energy constraint. The many-body case treated here introduce new difficulties, in particular the fact that there is now also a lack of compactness to treat the local density constraint, and as a consequence the current constraint. In the one-body setting, there is no such issue with the density since sequences of states with bounded energy have automatically sufficient compactness to pass to the limit in the density constraint, while this is not true in the many-body case. This latter fact is related to the convergence of one-body density matrices that will be defined further.
The main idea to go around the issue is to define two auxiliary optimization problems with \textit{global} constraints, and to prove the monotonicity of the entropy of the corresponding minimizers with respect to these global constraints. Along with classical compactness theorems for trace class operators, this allows us to prove that, while arbitrary sequences of states with bounded energy do not have sufficient compactness, the minimizing sequences of the entropy converge in a sufficiently strong sense that allows us to pass to the limit in the local constraints. The fermionic and bosonic cases are treated in the same fashion with essentially identical proofs.
The article is structured as follows: in Section \ref{prelim}, we introduce some background on second quantization; we next state our main result in Section \ref{mainR}, and prove the main theorem in Section \ref{proofT}. Finally, some standard technical results are given in an appendix
\paragraph{Acknowledgment.} OP's work is supported by NSF CAREER Grant DMS-1452349 and NSF grant DMS-2006416.
\section{Preliminaries} \label{prelim}
We introduce in this section some background that will be used throughout the paper.
\paragraph{Second quantization formalism.} We have already defined the fermionic Fock space in the introduction, and define now the bosonic version, denoted by $\mathfrak{F}_b$. It is given by the direct sum
\begin{equation*}
\mathfrak{F}_b : = \bigoplus_{n = 0}^{+\infty} \mathfrak{F}^{(n)}_b
\end{equation*}
where $\mathfrak{F}_b^{(n)} : = \mathfrak{h}^{\otimes_s n}$ is the $n$-fold symmetric tensor product of $\mathfrak{h}$ and $\mathfrak{h}^{\otimes_s 0} := \mathbb{C}$. We have
$$
\mathfrak{F}_b^{(n)}=L^2_s\big((\mathbb R^d)^n\big),
$$
where $L^2_s\big((\mathbb R^d)^n\big)$ is the space of symmetric square integrable complex functions on $(\mathbb R^d)^n$, that is, for $x_\ell \in \mathbb R^d$ , $\ell=1,\cdots,n$, $1 \leq, i,j \leq n$,
$$f(x_1,\cdots,x_i,\cdots,x_j,\cdots,x_n)=f(x_1,\cdots,x_j,\cdots,x_i,\cdots,x_n)$$
when $f \in L^2_s((\mathbb R^d)^n)$.
We denote by $\mathfrak{F}_{b/f}$ either the bosonic or fermionic Fock space, and represent an element $\psi$ of $\mathfrak{F}_{b/f}$ by the sequence $\psi=\{\psi^{(n)}\}_{n \in \mathbb N}$, where $\psi^{(n)} \in \mathfrak{F}_{b/f}^{(n)}$. The spaces $\mathfrak{F}_{b/f}$ are Hilbert spaces when equipped with the norm
$$
\| \psi\| = \left( \sum_{n \in \mathbb N} \| \psi^{(n)}\|^2_n\right)^{1/2}, \qquad \| \psi^{(n)}\|_n:=\| \psi^{(n)}\|_{L^2_{s/a}((\mathbb R^{d})^n)}.
$$
We use the same notation for the norms in $\mathfrak{F}_{s}$ and $\mathfrak{F}_{a}$ since there will be no possible confusion in the sequel. The inner products associated with $\|\cdot \|$ and $\| \cdot \|_n$ are denoted by $(\cdot,\cdot)$ and $(\cdot,\cdot)_n$.
We denote by $\mathcal J_1:=\mathcal J_1(\mathfrak{F}_{b/f})$ the space of trace class operators on $\mathfrak{F}_{b/f}$. The trace with respect to $\mathcal J_1$ is denoted simply by $\Tr(\cdot)$, while the trace with respect to $\mathcal J_1(E)$ for $E$ a Hilbert space is denoted by $\Tr_E(\cdot)$. The space of bounded operators on $E$ is denoted $\mathcal L(E)$.
We will refer to a ``state'', as a nonnegative, trace class operator on $\mathfrak{F}_{b/f}$ with trace equal to one. The set of states is denoted by $\mathcal S$, i.e.
$$
\mathcal S=\left\{\varrho \in \mathcal J_1: \; \varrho \geq 0,\quad \Tr(\varrho)=1 \right\}.
$$
\begin{definition}(Second quantization)
Let $A$ be an operator acting on $\mathfrak{F}^{(k)}_{b/f}$, $k \geq 1$. Its second quantization, denoted $\mathbb A$, is defined by
$$
\mathbb{A} := \underbrace{0\oplus\cdots\oplus 0}_{k\;\mathrm{times}} \oplus \bigoplus_{n = k}^{+\infty}\; \sum_{1\leq i_1<\cdots< i_k \leq n} (A)_{i_1,\cdots,i_k}
$$
where $(A)_{i_1,\cdots,i_k}$ is the operator $A$ acting on the variables labeled $i_1,\cdots,i_k$ in $\mathfrak{F}_{b/f}^{(k)}$ and leaving the other variables invariant.
\end{definition}
It is customary to denote $d\Gamma(A)=\mathbb A$. The second quantization of the identity on $\mathfrak{h}$ is the number operator
\begin{equation*}
\mathcal{N} : = d\Gamma(\mathrm{Id}_{\mathfrak{h}}).
\end{equation*}
The identity on $\mathfrak{F}_{b/f}$ is simply denoted by $\mathrm{Id}$. The operator $\mathcal N$ is self-adjoint on $\mathfrak{F}_{b/f}$ when equipped with the domain
$$
D(\mathcal N)=\left\{ \psi=\{\psi^{(n)}\}_{n \in \mathbb N} \in \mathfrak{F}_{b/f}: \; \sum_{n \in \mathbb N} n^2 \|\psi^{(n)}\|^2_n < \infty\right\}.
$$
We consider states that are not necessarily diagonal on $\mathfrak{F}_{b/f}$, and therefore that do not commute in general with $\mathcal N$. In particular, $\mathcal N \varrho$ is not necessarily positive when $\varrho \geq 0$, and this leads us to introduce $\mathcal N^{1/2}$, which is self-adjoint on $\mathfrak{F}_{b/f}$ with domain
$$
D(\mathcal N^{1/2})=\left\{ \psi=\{\psi^{(n)}\}_{n \in \mathbb N} \in \mathfrak{F}_{b/f}: \; \sum_{n \in \mathbb N} n \|\psi^{(n)}\|^2_n < \infty\right\}.
$$
We then denote by $\mathcal S_0$ the set of states $\varrho$ with finite average particle number, that is such that
$$
\Tr \left(\overline{\mathcal N^{1/2} \varrho\, \mathcal N^{1/2}}\right)<\infty,
$$
where $\overline{\mathbb A}$ denotes the extension of an operator $\mathbb A$ to $\mathfrak{F}_{b/f}$. We will drop the extension sign in the sequel for simplicity.
\begin{definition}(1-particle density matrix) \label{onebody}
Let $A$ be a bounded operator on $\mathfrak{h}$ and consider its second quantization $\mathbb A$. For a state $\varrho \in \mathcal S_0$, the 1-particle density matrix $\varrho^{(1)}$ is the unique nonnegative operator in $\mathcal J_1(\mathfrak{h})$ such that
\begin{equation} \label{def1}
\mathrm{Tr}_{\mathfrak{h}}\left(A \varrho^{(1)}\right) = \mathrm{Tr} \left (\mathbb A \varrho \right).
\end{equation}
\end{definition}
The fact that $\varrho^{(1)}$ is well-defined is classical and is established in Appendix for the sake of completeness. Note that since $\mathbb A$ is not bounded in $\mathfrak F_{b/f}$, relation \fref{def1} has actually to be understood as
$$
\mathrm{Tr}_{\mathfrak{h}}\left(A \varrho^{(1)}\right) = \mathrm{Tr} \left (\mathbb B \mathcal N^{1/2} \varrho\,\mathcal N^{1/2} \right),
$$
where
\mathbf e} \newcommand{\bh}{\mathbf h \label{defB}
\mathbb B:= 0 \oplus \bigoplus_{n = 1}^{+\infty} n^{-1}\mathbb{A}_{(n)},
\end{equation}
and then belongs to $\mathcal L(\mathfrak{F}_{b/f})$ when $A \in \mathcal L(\mathfrak{h})$. In \fref{defB}, $\mathbb{A}_{(n)}$ is the component of $\mathbb A$ on the sector $\mathfrak F_{b/f}^{(n)}$. Note that by setting $A=\mathrm{Id}_{\mathfrak{h}}$, we have the relation
$$
\mathrm{Tr}_{\mathfrak{h}}\left(\varrho^{(1)}\right) = \mathrm{Tr} \left( \mathcal N^{1/2} \varrho\ \mathcal N^{1/2} \right).
$$
We will need as well the 2-particle density matrix for the definition the interaction potential. It is justified in the same manner as Definition \ref{onebody}.
\begin{definition}(2-particle density matrix) Let $A$ be a bounded operator on $\mathfrak F^{(2)}_{b/f}$ and consider its second quantization $\mathbb A$. For a state $\varrho \in \mathcal S$ such that $\Tr \left(\mathcal N \varrho\, \mathcal N\right)$ is finite, the 2-particle density matrix $\varrho^{(2)}$ is the unique nonnegative operator in $\mathcal J_1(\mathfrak F^{(2)}_{b/f})$ such that
$$
\mathrm{Tr}_{\mathfrak F^{(2)}_{b/f}}\left(A \varrho^{(2)}\right) = \mathrm{Tr} \left (\mathbb A \varrho \right).
$$
\end{definition}
We define now the local constraints, first the density, current, and kinetic energy.
\paragraph{Local density, current, and kinetic energy constraints.}
Consider a state $\varrho \in \mathcal S_0$ and its associated 1-particle density matrix $\varrho^{(1)} \in \mathcal J_1(\mathfrak{h})$. The local (1-particle) density $n[\varrho]$ of $\varrho$ is defined by duality by, for any $\varphi\in C^{\infty}_0(\mathbb{R}^d)$:
\begin{equation*}
\int_{\mathbb{R}^d} n[\varrho](x) \varphi(x) dx = \mathrm{Tr}_{\mathfrak{h}}\left(\varrho^{(1)} \varphi\right),
\end{equation*}
where we identify $\varphi$ and its associated multiplication operator.
Let $h_0=-\Delta$, equipped with domain $H^2(\mathbb R^d)$, and for $ \mathbb H_0=d\Gamma(h_0)$, let $\mathcal E_0$ be the following set:
$$
\mathcal E_0=\left\{ \varrho \in \mathcal S_0: \; \Tr\left( \mathbb H_0^{1/2} \varrho \,\mathbb H_0^{1/2} \right)<\infty \right\}.
$$
$\mathcal E_0$ is the set of states with finite particle number and finite kinetic energy. We will need the following lemma in order to define the current and energy contraints. The straightforward proof is given in the Appendix for convenience of the reader.
\begin{lemma} \label{enrho} Let $\varrho \in \mathcal E_0$. Then $\varrho^{(1)}$ verifies $\mathrm{Tr}_\mathfrak{h} \left( h_0^{1/2} \varrho^{(1)}\ h_0^{1/2} \right)<\infty$ and
\mathbf e} \newcommand{\bh}{\mathbf h \label{eqkin} \mathrm{Tr}_\mathfrak{h} \left( h_0^{1/2} \varrho^{(1)} h_0^{1/2} \right)= \Tr\left( \mathbb H_0^{1/2} \varrho \,\mathbb H_0^{1/2} \right).\end{equation}
\end{lemma}
The current $u[\varrho]$ can be now defined by, for any $\Phi \in (C^{\infty}_0(\mathbb{R}^d))^d$,
\begin{equation*}
\int_{\mathbb{R}^d} u[\varrho](x) \cdot \Phi(x) dx = - i \mathrm{Tr}_{\mathfrak{h}}\left(\Phi \cdot \nabla \varrho^{(1)} +\frac{1}{2} \nabla \cdot \Phi \right),
\end{equation*}
and the kinetic energy $k[\varrho]$ by
\begin{equation*}
\int_{\mathbb{R}^d} k[\varrho](x) \varphi(x) dx = - \mathrm{Tr}_{\mathfrak{h}}\left(\nabla \cdot (\varphi \nabla)\varrho^{(1)} \right).
\end{equation*}
Note that $u[\varrho]$ and $k[\varrho]$ are well-defined since Lemma \ref{enrho} implies that $\nabla \varrho^{(1)}$ and $\nabla \varrho^{(1)} \nabla$ are trace class. Formal calculations also show that $n[\varrho]$, $u[\varrho]$ and $k[\varrho]$ agree with the definitions given in \fref{consint} in the introduction based on the annihilation operator. Moreover, if $\{\mu_p\}_{p \in \mathbb N}$ and $\{\varphi_p\}_{p \in \mathbb N}$ denote the eigenvalues and eigenfunctions of $\varrho^{(1)}$, we have the familiar relations
$$
\left\{
\begin{array}{ll}
\displaystyle n[\varrho]=\sum_{p \in \mathbb N} \mu_p |\varphi_p|^2, \qquad &\textrm{local density}\\
\displaystyle u[\varrho] = \sum_{p\in\mathbb{N}} \mu_p \Im\left( \varphi_p^* \nabla \varphi_p\right),\qquad &\textrm{local current}\\
\displaystyle k[\varrho]=\sum_{p \in \mathbb N} \mu_p |\nabla \varphi_p|^2, \qquad &\textrm{local kinetic energy}.
\end{array}
\right.
$$
The functions $n[\varrho]$, $u[\varrho]$, and $k[\varrho]$ are all in $L^1(\mathbb R^d)$ when $\varrho \in \mathcal E_0$, and the series above converge in $L^1(\mathbb R^d)$. We have moreover the relations
$$
\|n[\varrho]\|_{L^1}=\Tr\left( \mathcal N^{1/2} \varrho \,\mathcal N^{1/2} \right), \qquad \|k[\varrho]\|_{L^1}=\Tr\left( \mathbb H_0^{1/2} \varrho \,\mathbb H_0^{1/2} \right).
$$
\paragraph{Definition of local total energy.} We define now the local potential and local interaction energy constraints. For this, let $v=v_+-v_-$, $v_\pm \geq 0$, and $w$ even, all real-valued, such that
$$
v_+ \in L^1_{\rm{loc}}(\mathbb R^d), \qquad v_-, w \in L^p(\mathbb R^d)+L^\infty(\mathbb R^d),
$$
with $p=1$ when $d=1$, $p>1$ when $d=2$, and $p=d/2$ when $d \geq 3$. We suppose that $w$ is classically stable of the second kind, that is, there exists a constant $C_0>0$ such that
\mathbf e} \newcommand{\bh}{\mathbf h \label{class}
\forall n \geq 2, \qquad \sum_{1 \leq i<j \leq n} w(x_i-x_j) \geq -C_0 n, \qquad \textrm{a.e. on } (\mathbb R^d)^n.
\end{equation}
An example of such $w$ is the standard Coulomb potential $w(x)=|x|^{-1}$ when $d=3$.
For a state $\varrho$ with $\Tr \left(\mathcal N \varrho\, \mathcal N\right)<\infty$, the local interaction energy is formally defined by
$$
\widetilde e_{I}[\varrho](x)=\int_{\mathbb R^d} n[\varrho^{(2)}](x,y) w(x-y) dy,
$$
where $n[\varrho^{(2)}](x,y)$ is the local density associated with the 2-particle density matrix $\varrho^{(2)}$ of $\varrho$. It is defined by duality by
\begin{equation*}
\int_{\mathbb{R}^d \times \mathbb R^d} n[\varrho^{(2)}](x,y) \varphi(x,y) dx dy = \mathrm{Tr}_{\mathfrak{F}_{b/f}^{(2)}}\left(\varrho^{(2)} \varphi\right),
\end{equation*}
for any symmetric test function $\varphi \in C^\infty_0(\mathbb R^d \times \mathbb R^d)$. The condition $\Tr \left(\mathcal N \varrho\, \mathcal N\right)<\infty$ is not natural in the minimization problem since the total energy constraint only involves $\widetilde{e}_I[\varrho]$ and not $\mathcal N \varrho\, \mathcal N$, and it is therefore not clear how to define $n[\varrho^{(2)}]$ rigorously if $\Tr \left(\mathcal N \varrho\, \mathcal N\right)$ is not finite. We then introduce a modified interaction energy $e_I[\varrho]$ as follows. Let $r_0>C_0$ (for the $C_0$ defined in \fref{class}) and let
$$
\mathbb W= 0 \oplus 0 \oplus \bigoplus_{n = 2}^{+\infty} \left(\sum_{1 \leq i<j \leq n} w(x_i-x_j) \right).
$$
The operator $\mathbb W+r_0 \mathcal N$ is strictly positive according to \fref{class}. Consider a state $\varrho$ such that
$$\Tr \left((\mathbb W+r_0 \mathcal N )^{1/2}\varrho\, (\mathbb W+r_0 \mathcal N )^{1/2}\right)<\infty,$$
and let
$$
\mathbb B:=0 \oplus r_0^{1/2} \oplus \bigoplus_{n = 2}^{+\infty}\left(\sum_{1 \leq i<j \leq n} w(x_i-x_j) + n r_0\right)^{1/2} n^{-1}.
$$
The operator $\sigma_\varrho=\mathbb B \varrho\, \mathbb B$ verifies as a consequence $\Tr \left(\mathcal N \sigma_\varrho\, \mathcal N\right)<\infty$, and therefore has a 2-particle density matrix $\sigma^{(2)}_\varrho $. We then define
$$
e_I[\varrho](x):=\int_{\mathbb R^d} n[\sigma^{(2)}_\varrho](x,y)dy \geq 0 \qquad a.e.
$$
Note that $e_I[\varrho]$ is integrable since $n[\sigma^{(2)}] \in L^1(\mathbb R^d \times \mathbb R^d)$, and that
$$
\Tr \left((\mathbb W+r_0 \mathcal N )^{1/2}\varrho\, (\mathbb W+r_0 \mathcal N )^{1/2}\right)=\|e_I[\varrho]\|_{L^1}.
$$
When $\Tr \left(\mathcal N \varrho\, \mathcal N\right)$ is finite, the two definitions of the interaction energy given above are equivalent for the minimization problem since
$$
e_I[\varrho]=\widetilde e_{I}[\varrho]+r_0 n[\varrho],
$$
and prescribing both $e_I[\varrho]$ and $n[\varrho]$ is then equivalent to prescribing both $\widetilde{e}_I[\varrho]$ and $n[\varrho]$.
Regarding the potential energy, the density $n[\varrho]$ must have sufficient decay at the infinity for the entropy of a state $\varrho$ to be bounded below. We then introduce a nonnegative confining potential $v_c \in L^1_{\rm{loc}}(\mathbb R^d)$ with $v_c \to +\infty$ as $|x| \to +\infty$ such that
\mathbf e} \newcommand{\bh}{\mathbf h \label{GT}
\forall t>0, \qquad \int_{\mathbb R^d} e^{- t v_c(x)} dx < \infty,
\end{equation}
and suppose that $v_c n[\varrho] \in L^1(\mathbb R^d)$. With $h_c=h_0+v_c$, defined in the sense of quadratic forms and self-adjoint on an appropriate domain, the condition \fref{GT} ensures by the Golden-Thompson inequality that the operator $e^{-t h_c}$ is trace class for all $t>0$.
Let finally $V=v+v_c$. The local potential energy is defined by
$$
e_P[\varrho]= V n[\varrho].
$$
Again, for the minimization problem, prescribing both $V n [\varrho]$ and $n[\varrho]$ is equivalent to prescribing both $v n [\varrho]$ and $n[\varrho]$, and therefore the introduction of $v_c$ in the constraint does not change the minimizer. The local total energy of a state $\varrho$ is then
$$
e[\varrho]=k[\varrho]+e_P[\varrho]+e_I[\varrho].
$$
\paragraph{The energy space.} For $n\geq 2$, let the symmetric $n$-body operator
$$
H_n=\sum_{j=1}^n \left(-\Delta_{x_j} + V(x_j)+r_0 \right)+\sum_{1 \leq i<j \leq n} w(x_i-x_j),
$$
with for $n=1$,
$$
H_1=-\Delta+V+r_0.
$$
The regularity assumptions on $v_-$ and $w$ imply that the $n$-body potential in $H_n$
$$
\sum_{j=1}^n v_-(x_j)+\sum_{1 \leq i<j \leq n} w(x_i-x_j)
$$
is infinitesimally $(-\Delta)$-form bounded as an adaptation of Kato's theorem, see e.g. \cite{RS-80-2} in the case $d=3$. A first consequence of this is that there exists a constant $\gamma \in (0,1)$ independent of $n$, and that $r_0$ can be chosen sufficiently large, such that
\mathbf e} \newcommand{\bh}{\mathbf h \label{Hbelow}
\gamma\, H_n^c + \gamma n\leq H_n, \qquad H_n^c=\sum_{j=1}^n \left(-\Delta_{x_j} + v_c(x_j) \right),
\end{equation}
in the sense of operators. Second, $H_n$ is associated with a quadratic form closed on the space of functions $\psi \in \mathfrak F_{b/f}^{(n)}$ such that $\psi \in H^1((\mathbb R^d)^n)$ and
$$
\int_{(\mathbb R^d)^n} (v_+(x_1)+v_c(x_1))|\psi(x_1,\cdots,x_n)|^2 dx_1 \cdots dx_n < \infty.
$$
With an abuse of notation, we will also denote by $H_n$ the self-adjoint realization with domain $D(H_n)$ of the quadratic form. In the same way, $H_1$ is the self-adjoint realization of $-\Delta+V+r_0$ defined in the sense of quadratic forms.
Let $\mathbb H$ be the second quantization of $H_n$, that is
$$\mathbb H=d\Gamma(h_0+V+r_0)+\mathbb W.$$
It is self-adjoint with domain
$$
D(\mathbb H)=\mathbb C \oplus D(h_0+V)\oplus \bigoplus_{n = 2}^{+\infty} D(H_n),
$$
see \cite[Theorem 4.2]{arai}. The energy space that we will use in the minimization is finally the following:
$$
\mathcal E=\left\{ \varrho \in \mathcal S: \; \Tr\left( \mathbb H^{1/2} \varrho \,\mathbb H^{1/2} \right)<\infty \right\}.
$$
Note that
\mathbf e} \newcommand{\bh}{\mathbf h \label{eqeneg}
\Tr\left( \mathbb H^{1/2} \varrho \,\mathbb H^{1/2} \right)=\|e[\varrho]\|_{L^1},
\end{equation}
and that condition \fref{Hbelow} yields
\mathbf e} \newcommand{\bh}{\mathbf h \label{below2}
\gamma \Tr\left( d\Gamma(h_c)^{1/2} \varrho \,d\Gamma(h_c)^{1/2} \right) + \gamma \Tr\left( \mathcal N^{1/2} \varrho \, \mathcal N^{1/2} \right)
\leq \Tr\left( \mathbb H^{1/2} \varrho \,\mathbb H^{1/2} \right).
\end{equation}
This implies in particular that $\mathcal E \subset \mathcal E_0$.\\
We are now in position to state our main result.
\section{Main result} \label{mainR}
The entropy of a state $\varrho \in \mathcal S$ is defined by
$$
S(\varrho)= \Tr (\varrho \log \varrho)= \sum_{i \in \mathbb N} \rho_i \log \rho_i,
$$
for $\{\rho_i\}_{i \in \mathbb N}$ the eigenvalues of $\varrho$ (counted with multiplicity and forming a nonincreasing sequence; if $\varrho$ has a finite rank, then $\rho_i=0$ when $i \geq N$ for some $N$). Note that $S$ is always well-defined in $[-\infty,0]$ since $0 \leq \rho_i \leq 1$ as $\Tr(\varrho)=1$.
The set of admissible local constraints (we will sometimes refer to these as ``moments'' in the sequel) is defined by
\begin{align*}
\mathcal M = \Big\{(n,u,e)\in L_+^1(\mathbb{R}^d)\times (L^1(\mathbb{R}^d))^d\times L_+^1(\mathbb{R}^d)\hspace{5cm}
\\ \qquad \qquad\textrm{such that }\quad (n,u,e) = (n[\varrho],u[\varrho],e[\varrho])\quad\textrm{for at least one }\varrho\in\mathcal E \Big\}.
\end{align*}
Above, $L^1_+(\mathbb{R}^d) = \{ \varphi\in L^1(\mathbb{R}^d):\; \varphi\geq 0\; a.e.\}$. In other terms, $\mathcal M$ consists of the set of functions $(n,u,e)$ that are the local density, current and total energy of at least one state with finite energy. It is not difficult to construct admissible constraints, for instance by taking moments of the Gibbs state
$$
\frac{e^{-\mathbb H}}{ \Tr (e^{-\mathbb H})}.
$$
To the best of our knowledge, the characterization of $\mathcal M$ remains to be done.
For $(n_0,u_0,e_0) \in \mathcal M$, the feasible set is then given by
\begin{equation*}
\mathcal A(n_0,u_0,e_0) = \Big\{\varrho\in\mathcal E:\; n[\varrho] = n_0,\; u[\varrho] = u_0\;\textrm{and}\; e[\varrho] = e_0 \Big\}.
\end{equation*}
The set $\mathcal A(n_0,u_0,e_0)$ is not empty by construction since $(n_0,u_0,e_0)$ is admissible. \\
Our main result is the next theorem.
\begin{theorem}\label{thm:Main} Let $(n_0,u_0,e_0) \in \mathcal M$, with $n_0 v_c \in L^1(\mathbb R^d)$. Then, the minimization problem
$$
\inf_{\mathcal A(n_0,u_0,e_0)} S
$$
admits a unique solution.
\end{theorem}
We expect the minimizer $\varrho^\star$ to be a local Gibbs state with Hamiltonian $\mathbb H^\star$, for $\mathbb H^\star$ the second quantization of some two-body interaction Hamiltonian involving the Lagrange multipliers (which are functions here) associated with the constraints. While a formal derivation using standard calculus of variations techniques is quite straightforward (this is actually the formal expression given in the work of Nachtergaele and Yau), a rigorous derivation appears to be quite difficult. It was achieved in the one-particle situation in \cite{MP-JSP, DP-CVPDE, DP-JFA} in various settings.
\paragraph{Outline of the proof.} It will be shown in Section \ref{secpropS} that the entropy of states with fixed total energy is bounded below, and that the entropy is lower semi-continuous on the energy space $\mathcal E$. The proof of the theorem therefore hinges upon showing that minimizing sequences satisfy the constraints in the limit. Standard weak-$*$ compactness theorems in the space of trace class operators show that minimizing sequences converge in some weak sense to an operator $\varrho^\star$ with finite energy, and the convergence is sufficiently strong to obtain that $\varrho^\star$ is a state, i.e. that $\Tr(\varrho^\star)=1$. Lower semi-continuity of the entropy yields moreover
\mathbf e} \newcommand{\bh}{\mathbf h \label{ineqS}
S(\varrho^\star) \leq \inf_{\mathcal A(n_0,u_0,e_0)} S.
\end{equation}
It is not possible to identify at that stage the local moments of $\varrho^\star$, and as a consequence to show that $\varrho^\star$ belongs to the feasible set $\mathcal A(n_0,u_0,e_0)$. The core of the proof consists then in showing that the moments of the minimizing sequences converge in a strong sense, allowing us to obtain that
\mathbf e} \newcommand{\bh}{\mathbf h \label{egalconst}
n[\varrho^\star] = n_0,\qquad u[\varrho^\star]=u_0, \qquad e[\varrho^\star] = e_0,\end{equation}
which, together with \fref{ineqS} and the strict convexity of $S$, proves Theorem \ref{thm:Main}.
Our strategy to recover strong convergence follows the general method we introduced in \cite{DP-CMP}, with some important differences needed to handle the many-body nature of the problem. It is based on defining minimization problems with global constraints. Consider the following sets: for $a>0$, let
\begin{equation} \label{Ag}
\mathcal A_g(a) = \Big\{\varrho\in\mathcal E:\; \|n[\varrho]\|_{L^1}=a \Big\},
\end{equation}
and
\begin{equation} \label{Age}
\mathcal A_{g,e}(a) = \Big\{\varrho\in\mathcal E:\; \|e[\varrho]\|_{L^1}=a \Big\}.
\end{equation}
Starting from \fref{ineqS}, we will prove the following crucial two inequalities: if
$$\|e[\varrho^\star]\|_{L^1}=b^\star \leq b=\|e[\varrho]\|_{L^1}, \qquad \|n[\varrho^\star]\|_{L^1}=a^\star \leq a=\|n[\varrho]\|_{L^1},$$
then
\mathbf e} \newcommand{\bh}{\mathbf h \label{inimp}
\inf_{\mathcal A_{g}(a^\star)} F_\beta \leq \inf_{\mathcal A_{g}(a)} F_\beta, \qquad \inf_{\mathcal A_{g,e}(b^\star)} S \leq \inf_{\mathcal A_{g,e}(b)} S,
\end{equation}
where $F_\beta$ is the free energy at temperature $\beta^{-1}$,
$$
F_\beta(\varrho)= \beta^{-1} S(\varrho) + \Tr (\mathbb H^{1/2} \varrho \mathbb H^{1/2} ), \qquad \varrho \in \mathcal E.
$$
We will prove that the minima of $F_\beta$ in $\mathcal A_{g}(a)$ and of $S$ in $\mathcal A_{g,e}(b)$ are achieved by using (global) Gibbs states of the form
$$
\varrho_{\alpha,\beta}=\frac{e^{- \mathbb H_{\alpha,\beta}}}{\Tr (e^{- \mathbb H_{\alpha,\beta}})},
$$
where for $(\alpha,\beta) \in \mathbb R_+ \times \mathbb R_+^*$, $\mathbb H_{\alpha,\beta}= \beta \mathbb H+\alpha \mathcal N$.
Intuitively, \fref{inimp} is only possible if $b^\star=b$, $a^\star=a$, and if the inequalities are equalities. Call indeed $-S$ the physical entropy. If we accept the heuristics that the equilibrium physical entropy maximizes disorder, the equilibrium state with the largest energy should have the largest physical entropy, which contradicts the second inequality in \fref{inimp} if $b^\star<b$. In the same way, we expect the equilibrium state with the largest average number of particles to lose the largest amount of energy to thermal fluctuations, and therefore to have a lower equilibrium free energy than the equilibrium state with less particles. This contradicts the first inequality in \fref{inimp} if $a^\star<a$. An important part of the proof is to make these arguments rigorous.
Once we know that $\|e[\varrho^\star]\|_{L^1}=\|e_0\|_{L^1}$, and $\|n[\varrho^\star\|_{L^1}=\|n_0\|_{L^1}$, arguments for nonnegative operators of the type ``weak convergence plus convergence of the norms imply strong convergence'' lead to \fref{egalconst}.
Note that it is important to treat the density and energy constraints separately. If one were to set an optimization problem with global constraints both on the density and the energy, one would have to study the minimal possible energy of Gibbs states of the form $\varrho_{\alpha,\beta}$ for a fixed average number of particles. This is not direct as this requires to investigate the ground state of the Hamiltonian $\mathbb H$ for the interaction potential $w$ for both the fermionic and bosonic cases. By separating the two constraints, we circumvent this issue and can then achieve arbitrary low energy states by simply decreasing the temperature and the chemical potential (which is $-\alpha$ here).
In addition, it is necessary to treat the energy constraint first and to obtain that $\|e[\varrho^\star]\|_{L^1}=\|e_0\|_{L^1}$ before handling the density. This allows us to introduce the free energy in \fref{inimp}, which is bounded below and admits a minimizer on $\mathcal A_{g}(a)$. For otherwise we would work with the problem
$$
\inf_{\mathcal A_{g}(a)} S,
$$
which does not have a solution since $S$ is not bounded below on $\mathcal A_{g}(a)$.\\
The rest of the article is dedicated to the proof of Theorem \ref{thm:Main}.
\section{Proof of the theorem} \label{proofT}
In Section \ref{secpropS}, we show that the entropy is bounded below and lower semi-continuous for states in the feasible set $\mathcal A(n_0,u_0,e_0)$. Section \ref{secseq} consists in the core of the proof where we show that minimizing sequences converge in a strong sense. In Sections \ref{propproof1}, \ref{propproof2}, and \ref{propproof3}, we give the proofs of some important results that we were needed in Section \ref{secseq}. Finally, an appendix collects the proofs of some technical results.
\subsection{Properties of the entropy} \label{secpropS}
We will use the relative entropy between two states $\varrho$ and $\sigma$, which is defined by
$$
\mathcal F(\varrho,\sigma)= \Tr \big(\varrho (\log \varrho-\log \sigma) \big) \in [0, \infty].
$$
It is set to the infinity when the kernel of $\sigma$ is not included in the kernel of $\varrho$. See e.g. \cite{wehrl} for more details about the relative entropy. We recall that $e^{-\beta h_c}$ is trace class for any $\beta>0$ by assumption \fref{GT}. According to \cite[Prop. 5.2.27]{brat2} for the bosonic case, and \cite[Prop. 5.2.22]{brat2} for the fermionic case, this implies that $e^{-\beta d\Gamma(h_c)}$ is trace class as well. These trace results are essentially the only parts in the proof where a distinction between fermions and bosons is made. Let then
$$
\varrho_c=\frac{e^{-\beta d\Gamma(h_c)}}{ \Tr (e^{-\beta d\Gamma(h_c)})},
$$
which is a state. We have the following result, which is a straightforward consequence of the nonnegativity of the relative entropy.
\begin{lemma} \label{Sbelow} Let $\varrho \in \mathcal S$ with $\Tr\big( d\Gamma(h_c)^{1/2} \varrho\, d\Gamma(h_c)^{1/2} \big)$ finite. Then
$$
S(\varrho) \geq - \beta \Tr\big( d\Gamma(h_c)^{1/2} \varrho\, d\Gamma(h_c)^{1/2} \big) - \log \Tr (e^{-\beta d\Gamma(h_c)}).
$$
\end{lemma}
\begin{proof}
Let $\varrho$ satisfy the assumptions in the lemma. Then,
$$
\mathcal F(\varrho,\varrho_c)=S(\varrho)+\beta \Tr\big( (d\Gamma(h_c))^{1/2} \varrho (d\Gamma(h_c))^{1/2} \big)+ \log\Tr (e^{-d\Gamma(h_c)}) \geq 0,
$$
which proves the result. Note that there is a formal calculation that has to be justified, i.e. $\Tr(\varrho\, d\Gamma(h_c) )=\Tr\big( (d\Gamma(h_c))^{1/2} \varrho (d\Gamma(h_c))^{1/2} \big)$. When $\varrho\, d\Gamma(h_c)$ is not trace class, this is done by a regularization that we do not detail.
\end{proof}
\medskip
The next result follows also directly from the properties of the relative entropy.
\begin{lemma} \label{contS} Let $\varrho \in \mathcal S$, and consider a sequence of states such that $\varrho_m$ converges to $\varrho$ weak-$*$ in $\mathcal J_1$ as $m \to \infty$. Suppose moreover that there exists $C>0$ independent of $m$ such that
$$
\Tr\big( d\Gamma(h_c)^{1/2} \varrho\, d\Gamma(h_c)^{1/2} \big)+\Tr\big( d\Gamma(h_c)^{1/2} \varrho_m\, d\Gamma(h_c)^{1/2} \big) \leq C.$$
Then
$$
S(\varrho) \leq \liminf_{m \to \infty} S(\varrho_m).
$$
\end{lemma}
\begin{proof} Write
$$
S(\varrho_m)=\mathcal F(\varrho_m,\varrho_c)-\beta \Tr\big( (d\Gamma(h_c))^{1/2} \varrho_m (d\Gamma(h_c))^{1/2} \big)- \log\Tr (e^{-\beta d\Gamma(h_c)}),
$$
so that
$$
S(\varrho_m)\geq \mathcal F(\varrho_m,\varrho_c)-C\beta - \log\Tr (e^{-\beta d\Gamma(h_c)}).
$$
According to \cite[Theorem 2]{LS}, the relative entropy is weakly lower semicontinuous, and therefore
\begin{eqnarray*}
\liminf_{m \to \infty} S(\varrho_m) &\geq& \mathcal F(\varrho,\varrho_c)-C\beta - \log\Tr (e^{-\beta d\Gamma(h_c)})\\
&\geq & S(\varrho)+\beta \Tr\big( (d\Gamma(h_c))^{1/2} \varrho (d\Gamma(h_c))^{1/2} \big)-C\beta.
\end{eqnarray*}
Sending $\beta$ to zero then yields the result.
\end{proof}
\begin{remark} Under the conditions of Lemma \ref{Sbelow}, we have in fact that
$$
S(\varrho) =\lim_{m \to \infty} S(\varrho_m).
$$
Indeed, a direct adaptation of Lemma \ref{lem:comp} further shows that $\{\varrho_m\}_{m \in \mathbb N}$ actually converges strongly to $\varrho$ in $\mathcal J_1$, and as a consequence that the eigenvalues $\{\rho_j^{(m)}\}_{j \in \mathbb N}$ of $\varrho_m$ converge to those of $\varrho$, denoted $\{\rho_j\}_{j \in \mathbb N}$, as $m \to \infty$. Fatou's lemma for sequences then yields
$$
\sum_{j\in \mathbb N} - \rho_j \log \rho_j \leq \liminf_{m \to \infty} \sum_{j\in \mathbb N} - \rho^{(m)}_j \log \rho_j^{(m)},
$$
which corresponds to
$$
S(\varrho) \geq \limsup_{m \to \infty} S(\varrho_m).
$$
\end{remark}
\subsection{Minimizing sequences} \label{secseq} The starting point is that $S$ is bounded below on $\mathcal A(n_0,u_0,e_0)$. Indeed, estimate \fref{below2} shows that states in $\mathcal A(n_0,u_0,e_0)$ satisfy the assumptions of Lemma \ref{Sbelow}, and as a consequence, using again \fref{below2} together with \fref{eqeneg},
$$
S(\varrho) \geq - \gamma^{-1}\Tr\big( \mathbb H^{1/2} \varrho\, \mathbb H^{1/2} \big) - \log \Tr (e^{-d\Gamma(h_c)})=-\gamma^{-1}\|e_0\|_{L^1}- \log \Tr (e^{-d\Gamma(h_c)}),$$
for all $\varrho \in \mathcal A(n_0,u_0,e_0)$. There exists then a minimizing sequence $\{\varrho_m\}_{m\in \mathbb N}$ in $\mathcal A(n_0,u_0,e_0)$ such that
$$
\lim_{m \to \infty}S(\varrho_m)= \inf_{\mathcal A(n_0,u_0,e_0)} S.
$$
We have the following compactness result.
\begin{lemma} \label{lem:comp}
Let $\{\varrho_m\}_{m \in \mathbb N}$ be a sequence in $\mathcal E$ with
\mathbf e} \newcommand{\bh}{\mathbf h
\label{boundlem}
\Tr\left( \mathbb H^{1/2} \varrho_m \,\mathbb H^{1/2} \right) \leq C,
\end{equation} for some $C$ independent of $m$. Then, there exists $\varrho^\star \in \mathcal E$ and a subsequence $\{\varrho_{m_j}\}_{j\in \mathbb N}$ that converges to $\varrho^\star$ strongly in $\mathcal J_1$, and such that
\begin{align*}
\mathcal{N}^{1/2}\varrho_{m_j}\mathcal{N}^{1/2} &\underset{j\to+\infty}\to \mathcal{N}^{1/2}\varrho^\star \mathcal{N}^{1/2},\quad \textrm{weak-$*$}\textrm{ in }\mathcal{J}_1,
\\ \mathbb H^{1/2}\varrho_{m_j}\mathbb H^{1/2}&\underset{j\to+\infty}\to \mathbb H^{1/2}\varrho^\star \mathbb H^{1/2},\quad \textrm{weak-$*$}\textrm{ in }\mathcal{J}_1,
\end{align*}
with
\begin{eqnarray} \label{inflim}
\left\{\begin{array}{ll}
\displaystyle \Tr\left( \mathcal N^{1/2} \varrho^\star \,\mathcal N^{1/2} \right)\leq \liminf_{j \to \infty} \Tr\left( \mathcal N^{1/2} \varrho_{m_j} \,\mathcal N^{1/2} \right),
\\ \displaystyle \Tr\left( \mathbb H^{1/2} \varrho^\star \,\mathbb H^{1/2} \right)\leq \liminf_{j \to \infty} \Tr\left( \mathbb H^{1/2} \varrho_{m_j} \,\mathbb H^{1/2} \right).
\end{array}\right.
\end{eqnarray}
\end{lemma}
The proof of Lemma \ref{lem:comp} is classical and is given below for the reader's convenience.
\begin{proof} First of all, since $\varrho_m$ is a state, we have $\Tr (\varrho_m)=1$, which, together with \fref{boundlem}, and the fact that the space of trace class operators is the dual of the space of compact operators, implies that there exist $\varrho^\star \in \mathcal J_1$ and $\sigma \in \mathcal J_1$, and a subsequence such that $\varrho_{m_j}$ and $\mathbb H^{1/2} \varrho_{m_j} \,\mathbb H^{1/2}$ converge to $\varrho^\star$ and $\sigma$ in $\mathcal J_1$ weak-$*$ as $j \to \infty$, respectively. It is direct to identify $\sigma$: let $K$ compact in $\mathfrak F_{b/f}$, and let $\mathbb B=(\textrm{Id}+\mathbb H)^{-1}$, which is bounded. Then:
\begin{eqnarray*}
\lim_{j \to \infty} \Tr (\mathbb H^{1/2} \varrho_{m_j} \,\mathbb H^{1/2} \mathbb B K \mathbb B)&=&\Tr (\sigma \mathbb B K \mathbb B)\\
&=& \lim _{j \to \infty} \Tr ( \varrho_{m_j} \,\mathbb H^{1/2} \mathbb B K \mathbb B \mathbb H^{1/2})\\
&=&\Tr (\varrho^\star \mathbb H^{1/2}\mathbb B K \mathbb B \mathbb H^{1/2}).
\end{eqnarray*}
In the last line, we used that $\mathbb H^{1/2}\mathbb B K \mathbb B \mathbb H^{1/2}$ is compact. This shows that $\sigma=\mathbb H^{1/2} \varrho^\star \,\mathbb H^{1/2}$. We proceed in the same way for the limit of $\mathcal N^{1/2} \varrho_{m_j} \,\mathcal N^{1/2}$.
The limits in \fref{inflim} follow from the weak-$*$ convergence and the fact that $\Tr(\varrho)=\|\varrho\|_{\mathcal J_1}$ when $\varrho \geq 0$.
Regarding the strong convergence to $\varrho^\star$ in $\mathcal J_1$, we claim first that $(\textrm{Id}+\mathbb H)^{-1}$ is compact. Indeed, we have
$$
(\textrm{Id}+\mathbb H)^{-1}=1 \oplus (\textrm{Id}_\mathfrak{h}+h_0+V)^{-1}\oplus \bigoplus_{ n= 2}^{+\infty} (\textrm{Id}_{\mathfrak F^{(n)}_{b/f}}+H_n)^{-1}.
$$
The operator $h_0+v_c$ has a compact resolvent according to \cite[Theorem XIII.47]{RS-80-4}, and so does $H_n^c$. Then, for $n \geq 2$, each of the $H_n$ have a compact resolvent as perturbations of $H_n^c$ as an application of \cite[Theorem 3.4, Chapter 6 \S 3]{kato}. To obtain that $(\textrm{Id}+\mathbb H)^{-1}$ is compact, it just remains to show that $\|(\textrm{Id}_{\mathfrak F^{(n)}_{b/f}}+H_n)^{-1}\|_{\mathcal L(\mathfrak F^{(n)}_{b/f})} \to 0$ as $n \to \infty$, see \cite[Theorem 4.1]{arai}. This is a consequence of \fref{Hbelow}, that yields
$$
\|(\textrm{Id}_{\mathfrak F^{(n)}_{b/f}}+H_n)^{-1}\|_{\mathcal L(\mathfrak F^{(n)}_{b/f})} \leq (1+\gamma n)^{-1}.
$$
Second of all, it is not difficult to establish that the weak-$*$ convergence of $\varrho_{m_j}$ and $\mathbb H^{1/2} \varrho_{m_j} \,\mathbb H^{1/2}$ imply the weak-$*$ convergence of $(\textrm{Id}+\mathbb H)^{1/2} \varrho_{m_j} \,(\textrm{Id}+\mathbb H)^{1/2}$ to $(\textrm{Id}+\mathbb H)^{1/2} \varrho^\star \,(\textrm{Id}+\mathbb H)^{1/2}$. Then,
\begin{eqnarray*}
\lim_{j \to \infty} \Tr (\varrho_{m_j})&=&\Tr \Big( (\textrm{Id}+\mathbb H)^{1/2} \varrho_{m_j} \,(\textrm{Id}+\mathbb H)^{1/2} (\rm{Id}+\mathbb H)^{-1} \Big)\\
&=& \Tr \Big( (\textrm{Id}+\mathbb H)^{1/2} \varrho^\star \,(\textrm{Id}+\mathbb H)^{1/2} (\rm{Id}+\mathbb H)^{-1}\Big) \\
&=&\Tr (\varrho^\star).
\end{eqnarray*}
Finally, according to \cite[Theorem 2.21, Addendum H]{Simon-trace}, weak convergence in sense of operators together with the convergence of the norm in $\mathcal J_1$ implies strong convergence in $\mathcal J_1$. Since weak-$*$ convergence in $\mathcal J_1$ implies weak convergence in the sense of operators, we obtain that $\varrho_{m_j}$ converges strongly to $\varrho^\star$ in $\mathcal J_1$. In particular, $\varrho^\star$ is a state. This ends the proof.
\end{proof}
\medskip
We now continue the study of minimizing sequences. Since $\varrho_m$ is in $\mathcal A(n_0,u_0,e_0)$ and therefore satisfies the constraints, we have, for all $m \in \mathbb N$,
\begin{align}
&\Tr\left( \mathcal N^{1/2} \varrho_m \,\mathcal N^{1/2} \right)=\|n[\varrho_m]\|_{L^1}=\|n_0\|_{L^1} \label{equa1}\\
&\Tr\left( \mathbb H^{1/2} \varrho_m \,\mathbb H^{1/2} \right)=\|e[\varrho_m]\|_{L^1}=\|e_0\|_{L^1}. \label{equa2}
\end{align}
According to Lemma \ref{lem:comp}, there exists then a subsequence (that we still denote abusively by $\{\varrho_m\}_{m\in \mathbb N}$) that converges in the weak-$*$ topology of $\mathcal{J}_1$ to a state $\varrho^\star\in\mathcal E$. Since the continuity result given in Lemma \ref{contS} shows that
\mathbf e} \newcommand{\bh}{\mathbf h \label{ineqSS}
S(\varrho^\star) \leq \lim_{m \to \infty}S(\varrho_m)= \inf_{\mathcal A(n_0,u_0,e_0)} S,
\end{equation}
we are left to prove that $\varrho^{\star}\in\mathcal{A}(n_0,u_0,e_{0})$, i.e. $\varrho^\star$ verifies the local constraints. For this, we have from Lemma \ref{lem:comp},
\begin{align} \label{infn}
\|n[\varrho^\star]\|_{L^1} = \mathrm{Tr}\left(\mathcal{N}^{1/2}\varrho^\star\mathcal{N}^{1/2}\right) &\leq \underset{m\to+\infty}{\liminf} \; \mathrm{Tr}\left(\mathcal{N}^{1/2}\varrho_m\mathcal{N}^{1/2}\right) = \|n_0\|_{L^1}
\end{align}
and
\begin{align} \label{infe}
\|e[\varrho^\star]\|_{L^1} = \mathrm{Tr}\left(\mathbb{H}^{1/2}\varrho^\star\mathbb{H}^{1/2}\right) &\leq \underset{m\to+\infty}{\liminf} \; \mathrm{Tr}\left(\mathbb{H}^{1/2}\varrho_m\mathbb{H}^{1/2}\right) =\|e_0\|_{L^1}.
\end{align}
As already mentioned in the introduction, a one-body version of Lemma \ref{lem:comp} yields directly that $n[\varrho^\star]=n_0$. It is not true in the many-body case. The issue is the following, and is related to the identification of the 1-particle density matrices. From \fref{equa1} and \fref{equa2}, it is possible to show that the sequence of 1-particle density matrices $\varrho_m^{(1)}$ converges to some $\sigma^{(1)}$ strongly in $\mathcal J_1(\mathfrak{h})$. In particular, the local density of $\sigma^{(1)}$ is $n_0$. The difficulty is to identify $\sigma^{(1)}$ with the 1-particle density matrix of $\varrho^\star$, which is not possible at that stage. Indeed, we have, for all $\varphi \in L^\infty(\mathbb R^d)$,
\begin{eqnarray*}
\int_{\mathbb{R}^d} n[\varrho_m](x) \varphi(x) dx &=& \mathrm{Tr}_{\mathfrak{h}}\left(\varrho_m^{(1)} \varphi\right)\\
&=&\mathrm{Tr}\left(\mathcal N^{1/2}\varrho_m \mathcal N^{1/2} \mathcal N^{-1} d\Gamma(\varphi)\right).
\end{eqnarray*}
The operator $\mathcal N^{-1} d\Gamma(\varphi)$ is bounded in $\mathfrak F_{b/f}$, but not compact, which does not allow us to pass to the limit above since $\mathcal N^{1/2}\varrho_m \mathcal N^{1/2}$ converges only weak-$*$ in $\mathcal J_1$. One could replace $\mathcal N$ by $\mathbb H$ above, but while the projections of $ \mathbb H^{-1} d\Gamma(\varphi)$ on each $\mathfrak F^{(n)}_{b/f}$ are compact, $ \mathbb H^{-1} d\Gamma(\varphi)$ is not compact since these projections do not tend to zero in $\mathcal L(\mathfrak F^{(n)}_{b/f})$ as $n \to \infty$.\\
We have then the following proposition, proved in Section \ref{propproof1}:
\begin{proposition} \label{propegal} Assume that $\|n[\varrho^\star]\|_{L^1}=\|n_0\|_{L^1}$ and that $\|e[\varrho^\star]\|_{L^1}=\|e_0\|_{L^1}$. Then $\varrho^\star \in \mathcal{A}(n_0,u_0,e_{0})$.
\end{proposition}
Based on this last result, the main difficulty is therefore to prove that $\|n[\varrho^\star]\|_{L^1}<\|n_0\|_{L^1}$ and $\|e[\varrho^\star]\|_{L^1}<\|e_0\|_{L^1}$ is not possible. We will use for this the next lemma, where $\mathcal A_g(a)$ and $\mathcal A_{g,e}(a)$ are defined in \fref{Ag} and \fref{Age}.
\begin{lemma} \label{comS}(i) Suppose that the problem
$$
\inf_{\mathcal A_g(a)} F_\beta
$$
admits a solution. Then,
\mathbf e} \newcommand{\bh}{\mathbf h \label{eq1}
\inf_{\mathcal A_g(a)} F_\beta=\inf_{
\scriptsize{\left\{\begin{array}{l}
(n,u,e) \in \mathcal M\\
n \in L^1_+ \textrm{with} \; \|n\|_{L^1}=a \\
u \in (L^1)^d, e \in L^1_+
\end{array} \right.}
} \inf_{\mathcal A(n,u,e)} F_\beta.
\end{equation}
(ii) Suppose that the problem
$$
\inf_{\mathcal A_{g,e}(a)} S
$$
admits a solution. Then,
$$
\inf_{\mathcal A_{g,e}(a)} S=\inf_{
\scriptsize{\left\{\begin{array}{l}
(n,u,e) \in \mathcal M\\
e \in L^1_+ \textrm{with} \; \|e\|_{L^1}=a \\
u \in (L^1)^d, n \in L^1_+
\end{array} \right.}
} \inf_{\mathcal A(n,u,e)} S.
$$
\end{lemma}
\begin{proof} We start with (i). Let $(n,u,e) \in \mathcal M$ and $\varrho \in \mathcal A(n,u,e)$ with $\|n\|_{L^1}=a$. Then $\varrho \in \mathcal A_g(a)$, and as a consequence
$$
\inf_{\mathcal A_g(a)} F_\beta \leq \inf_{\mathcal A(n,u,e)} F_\beta.
$$
Taking the infimum as in \fref{eq1} then yields a first inequality in \fref{eq1}. For the reverse inequality, denote by $\sigma$ a minimizer of $F_\beta$ in $\mathcal A_g(a)$ and let
$$
G(n,u,e)=\inf_{\mathcal A(n,u,e)} F_\beta.
$$
By construction $\|n[\sigma]\|_{L^1}=a$, $u[\sigma] \in (L^1(\mathbb R^d))^d$ and $e[\sigma] \in L^1_+(\mathbb R^d)$, and clearly $(n[\sigma],u[\sigma],e[\sigma])$ is admissible. Hence,
$$
\inf_{
\scriptsize{\left\{\begin{array}{l}
(n,u,e) \in \mathcal M\\
n \in L^1_+ \textrm{with} \; \|n\|_{L^1}=a \\
u \in (L^1)^d, e \in L^1_+
\end{array} \right.}
}G(n,u,e)
\leq G(n[\sigma],u[\sigma],e[\sigma]).
$$
It remains to prove that
$$
G(n[\sigma],u[\sigma],e[\sigma])=\min_{\mathcal A_g(a)} F_\beta,
$$
which is straightforward since
$$
\min_{\mathcal A_g(a)} F_\beta \leq G(n[\sigma],u[\sigma],e[\sigma])=\inf_{\mathcal A(n[\sigma],u[\sigma],e[\sigma])} F_\beta \leq F_\beta(\sigma)=\min_{\mathcal A_g(a)} F_\beta.
$$
This proves (i). Item (ii) follows in the same manner by replacing $n$ by $e$ and $F_\beta$ by $S$. This ends the proof.
\end{proof}
\bigskip
Let now $\|e[\varrho^\star]\|_{L^1}=b^\star$ and $\|e_0\|_{L^1}=b$. Since $\varrho^\star \in \mathcal E$, we have
$$
\inf_{\mathcal A_{g,e}(b^\star)} S \leq S(\varrho^\star).
$$
Together with \fref{ineqSS}, this gives
$$
\inf_{\mathcal A_{g,e}(b^\star)} S \leq \inf_{\mathcal A(n_0,u_0,e_0)} S.
$$
Assuming for the moment that $S$ admits a minimizer on $\mathcal A_{g,e}(b)$, item (ii) of Lemma \ref{comS} yields
$$
\inf_{\mathcal A_{g,e}(b^\star)} S \leq
\inf_{
\scriptsize{\left\{\begin{array}{l}
(n,u,e) \in \mathcal M\\
e \in L^1_+ \textrm{with} \; \|e\|_{L^1}=\|e_0\|_{L^1} \\
u \in (L^1)^d, n \in L^1_+
\end{array} \right.}
} \inf_{\mathcal A(n,u,e)} S=\inf_{\mathcal A_{g,e}(b)} S,
$$
which results in
\mathbf e} \newcommand{\bh}{\mathbf h \label{cont1}
\inf_{\mathcal A_{g,e}(b^\star)} S \leq \inf_{\mathcal A_{g,e}(b)} S.
\end{equation}
The next result shows that there is a contradiction above if $b^\star<b$.
\begin{proposition} \label{gminS}
Let $a>0$. Then, the minimization problem
$$
\inf_{\mathcal A_{g,e}(a)} S
$$
admits a unique solution. Let moreover $f(a)=\inf_{\mathcal A_{g,e}(a)} S$. Then, $f$ is a strictly decreasing continuous function on $\mathbb R_+$.
\end{proposition}
The proof of Proposition \ref{gminS} is given in Section \ref{propproof2}. Suppose that $\|e[\varrho^\star]\|_{L^1}=b^\star<b=\|e_0\|_{L^1}$. Based on the previous proposition, we have $f(b)<f(b^\star)$, which contradicts \fref{cont1}, and we obtain therefore the equality
\mathbf e} \newcommand{\bh}{\mathbf h \label{eqE}\|e[\varrho^\star]\|_{L^1}=\|e_0\|_{L^1}.
\end{equation}
It remains to prove that
\mathbf e} \newcommand{\bh}{\mathbf h \label{eqn}\|n[\varrho^\star]\|_{L^1}=\|n_0\|_{L^1}.
\end{equation}
For this, let $\beta>0$. From \fref{ineqSS} and \fref{eqE}, we find
\mathbf e} \newcommand{\bh}{\mathbf h \label{ineqN2}
F_\beta(\varrho^\star)=\beta^{-1} S(\varrho^\star)+ \|e[\varrho^\star]\|_{L^1} \leq \beta^{-1} \inf_{\mathcal A(n_0,u_0,e_0)} S+ \|e_0\|_{L^1}
=\inf_{\mathcal A(n_0,u_0,e_0)} F_\beta .\end{equation}
The next proposition is similar to Propositon \ref{gminS} and is proved in Section \ref{propproof3}.
\begin{proposition} \label{gminF}
Let $a_0>0$. Then, there exists $\beta_0(a_0)>0$, such that the minimization problem
$$
\inf_{\mathcal A_{g}(a)} F_\beta
$$
admits a unique solution for any $\beta \leq \beta_0(a)$ and any $a \leq a_0$. Let moreover $g(a)=\inf_{\mathcal A_{g}(a)} F_\beta$. Then, $g$ is a strictly decreasing continuous function on $(0,a_0]$.
\end{proposition}
We apply Proposition \ref{gminF} as follows: let $\|n[\varrho^\star]\|_{L^1}=a^\star$ and $\|n_0\|_{L^1}=a$. We have then from \fref{ineqN2},
$$
\inf_{\mathcal A_g(a^\star)} F_\beta \leq F_\beta(\varrho^\star)\leq \inf_{\mathcal A(n_0,u_0,e_0)} F_\beta .$$
In Proposition \ref{gminF}, choose $a_0=\|n_0\|_{L^1}$. Since $\|n[\varrho^\star]\|_{L^1} \leq \|n_0\|_{L^1}$, then both minimization problems on $\mathcal A_{g}(a)$ and $\mathcal A_{g}(a^\star)$ admit a unique solution for $\beta \leq \beta_0(a_0)$.
Then, according to item (i) of Lemma \ref{comS},
$$
\inf_{\mathcal A_{g}(a^\star)} F_\beta \leq
\inf_{
\scriptsize{\left\{\begin{array}{l}
(n,u,e) \in \mathcal M\\
n \in L^1_+ \textrm{with} \; \|n\|_{L^1}=\|n_0\|_{L^1} \\
u \in (L^1)^d, e \in L^1_+
\end{array} \right.}
} \inf_{\mathcal A(n,u,e)} F_\beta=\inf_{\mathcal A_{g}(a)} F_\beta,
$$
which results in
$$
\inf_{\mathcal A_{g}(a^\star)} F_\beta \leq \inf_{\mathcal A_{g}(a)} F_\beta .
$$
But based on the previous proposition, we have $g(a)<g(a^\star)$, which is a contradiction if $a^\star<a$. We therefore obtain \fref{eqn}, which ends the proof of Theorem \ref{thm:Main} as an application of Proposition \ref{propegal}.
\subsection{Proof of Proposition \ref{propegal}} \label{propproof1} When $\|n[\varrho^\star]\|_{L^1}=\|n_0\|_{L^1}$ and $\|e[\varrho^\star]\|_{L^1}=\|e_0\|_{L^1}$, we obtain from \fref{infn} and $\fref{infe}$ that
\begin{align*}
\mathrm{Tr}\left(\mathcal{N}^{1/2}\varrho^\star\mathcal{N}^{1/2}\right) =\underset{m\to+\infty}{\lim} \; \mathrm{Tr}\left(\mathcal{N}^{1/2}\varrho_m\mathcal{N}^{1/2}\right)
\end{align*}
and
\begin{align*}
\mathrm{Tr}\left(\mathbb{H}^{1/2}\varrho^\star\mathbb{H}^{1/2}\right) =\underset{m\to+\infty}{\lim} \; \mathrm{Tr}\left(\mathbb{H}^{1/2}\varrho_m\mathbb{H}^{1/2}\right).
\end{align*}
According to \cite[Theorem 2.21, Addendum H]{Simon-trace}, this implies, together with the corresponding weak-$*$ convergences, that $\mathcal{N}^{1/2}\varrho_m\mathcal{N}^{1/2}$ and $\mathbb{H}^{1/2}\varrho_m\mathbb{H}^{1/2}$ converge strongly in $\mathcal J_1$ to $\mathcal{N}^{1/2}\varrho^\star \mathcal{N}^{1/2}$ and $\mathbb{H}^{1/2}\varrho^\star \mathbb{H}^{1/2}$.
Let now $\varrho_m^{(1)}$ and $\varrho^\star_{(1)}$ be the one-particle density matrices of $\varrho_m$ and $\varrho^\star$. We will show that $\varrho_m^{(1)}$ converges to $\varrho^\star_{(1)}$ strongly in $\mathcal J_1(\mathfrak{h})$. Indeed, with the definition \fref{defB},
\begin{eqnarray*}
\| \varrho_m^{(1)}- \varrho^\star_{(1)}\|_{J_1(\mathfrak{h})}&=&\sup_{\| A \|_{\mathcal L(\mathfrak{h})} \leq 1} \Tr_\mathfrak{h} \big((\varrho_m^{(1)}- \varrho^\star_{(1)}) A \big )\\
&=&\sup_{\| A \|_{\mathcal L(\mathfrak{h})} \leq 1} \Tr \big((\varrho_m- \varrho^\star) \mathbb A \big )\\
&=&\sup_{\| A \|_{\mathcal L(\mathfrak{h})} \leq 1} \Tr \big(\mathcal N^{1/2}(\varrho_m- \varrho^\star)\mathcal N^{1/2} \mathbb B \big )\\
&\leq& \|\mathcal N^{1/2}\varrho_m\mathcal N^{1/2}- \mathcal N^{1/2}\varrho^\star\mathcal N^{1/2}\|_{\mathcal J_1},
\end{eqnarray*}
which yields the result.
We prove similarly that $h_0^{1/2}\varrho_m^{(1)}h_0^{1/2}$ converges to $h_0^{1/2}\varrho^\star_{(1)}h_0^{1/2}$ strongly in $\mathcal J_1(\mathfrak{h})$. An easy consequence of this and of the strong convergence of $\varrho_m^{(1)}$ in $\mathcal J_1(\mathfrak{h})$, is that $(h_0+\textrm{Id}_\mathfrak{h})^{1/2}\varrho_m^{(1)}(h_0+\textrm{Id}_\mathfrak{h})^{1/2}$ converges to $(h_0+\textrm{Id}_\mathfrak{h})^{1/2}\varrho^\star_{(1)}(h_0+\textrm{Id}_\mathfrak{h})^{1/2}$ strongly in $\mathcal J_1(\mathfrak{h})$.
We are now in position to identify $k[\varrho^\star]$ and $u[\varrho^\star]$. For $\varphi \in L^\infty(\mathbb R^d)$, let
$$
A=(h_0+\textrm{Id}_\mathfrak{h})^{-1/2}\nabla \cdot (\varphi \nabla)(h_0+\textrm{Id}_\mathfrak{h})^{-1/2},
$$
which is bounded in $\mathcal L(\mathfrak{h})$. We have
from the definition of $k[\varrho^\star]$,
\begin{align*}
\|k[\varrho_m] -&k[\varrho^\star] \|_{L^1}\\
&=\sup_{\| \varphi \|_{L^\infty} \leq 1} \Tr_\mathfrak{h} \big(\nabla \cdot (\varphi \nabla)(\varrho_m^{(1)}-\varrho^\star_{(1)}) \big )\\
&=\sup_{\| \varphi \|_{L^\infty} \leq 1} \Tr_\mathfrak{h} \big(A (h_0+\textrm{Id}_\mathfrak{h})^{1/2}(\varrho_m^{(1)}-\varrho^\star_{(1)})(h_0+\textrm{Id}_\mathfrak{h})^{1/2} \big )\\
&\leq C \|(h_0+\textrm{Id})^{1/2}\varrho^{(1)}_m(h_0+\textrm{Id}_\mathfrak{h})^{1/2}- (h_0+\textrm{Id})^{1/2}\varrho^\star_{(1)}(h_0+\textrm{Id}_\mathfrak{h})^{1/2}\|_{\mathcal J_1(\mathfrak{h})}.
\end{align*}
This implies that $k[\varrho_m]$ converges to $k[\varrho^\star]$ strongly in $L^1(\mathbb R^d)$, and since $k[\varrho_m]=k_0$ by construction, we have $k[\varrho^\star]=k_0$. The proof that $u[\varrho^\star]=u_0$ is similar.
We have therefore obtained that $\varrho^\star$ satisfies the local constraints, and therefore that it belongs to the feasible set. This ends the proof.
\subsection{Proof of Proposition \ref{gminS}} \label{propproof2}
In preparation of the proof, let, for $(\alpha,\beta) \in \mathbb R_+\times \mathbb R_+^*$, $\mathbb H_{\alpha,\beta}= \beta \mathbb H+\alpha \mathcal N$. The operator $\mathbb H_{\alpha,\beta}$ is self-adjoint on $D(\mathbb H)$. According to \fref{Hbelow}, we have
\mathbf e} \newcommand{\bh}{\mathbf h \label{belowHa}
\gamma \beta d\Gamma(h_c)\leq \gamma \beta (d\Gamma(h_c)+\mathcal N) \leq \mathbb H_{\alpha,\beta}
\end{equation}
in the sense of operators. Hence, $e^{- \mathbb H_{\alpha,\beta}}$ is trace class for all $(\alpha,\beta) \in \mathbb R_+\times \mathbb R_+^*$ since $e^{-\gamma \beta d\Gamma(h_c)}$ is trace class as explained in Section \ref{secpropS}, and the partition function
$$
Z_{\alpha,\beta}=\Tr (e^{- \mathbb H_{\alpha,\beta}})
$$
is well-defined for all $\alpha \geq 0$. Let then the Gibbs state
$$
\varrho_{\alpha,\beta}=\frac{e^{- \mathbb H_{\alpha,\beta}}}{Z_{\alpha,\beta}}.
$$
Note that $\varrho_{\alpha,\beta}$ commutes with $\mathcal N$. The associated average particle number and average energy are defined by, respectively,
$$
N(\alpha,\beta)=\Tr (\mathcal N \varrho_{\alpha,\beta}), \qquad E(\alpha,\beta)=\Tr ( \mathbb H \varrho_{\alpha,\beta}).
$$
They are both well-defined for all $(\alpha,\beta) \in \mathbb R_+\times \mathbb R_+^*$ since in particular, according to \fref{Hbelow},
\begin{eqnarray} \nonumber
Z_{\alpha,\beta} \Tr (\mathcal N \varrho_{\alpha,\beta}) &=& 1+\sum_{n=1}^\infty n \Tr_{\mathfrak F^{(n)}_{b/f}} (e^{- \beta H_n-\alpha n})
\leq 1+\sum_{n=1}^\infty n \Tr_{\mathfrak F^{(n)}_{b/f}} (e^{- \gamma \beta H^c_n-\gamma n})\\
&\leq &1+C\sum_{n=1}^\infty \Tr_{\mathfrak F^{(n)}_{b/f}} (e^{- \gamma \beta H^c_n})=C\Tr (e^{-\gamma \beta d\Gamma(h_c) }). \label{alz}
\end{eqnarray}
We will use the following lemma, helpful as well in the proof of Proposition \ref{gminF}.
\begin{lemma} \label{lemZ}
(i) $Z_{\alpha,\beta} \in C^\infty(\mathbb R_+^* \times \mathbb R^*_+)$ and
\mathbf e} \newcommand{\bh}{\mathbf h \label{expN}
N(\alpha,\beta)=- \partial_\alpha \log Z_{\alpha,\beta}, \qquad E(\alpha,\beta)=- \partial_\beta \log Z_{\alpha,\beta}.
\end{equation}
(ii) For all $\alpha \geq 0$, the function $ \mathbb R_+^* \ni \beta \mapsto E(\alpha,\beta)$ is continuously strictly decreasing with
$$
\lim_{\beta\to + \infty} E(\alpha,\beta)=0, \qquad \lim_{\beta\to 0} E(\alpha,\beta)=+\infty, \qquad \forall \alpha \geq 0.
$$
(iii) For all $\beta > 0$, the function $\mathbb R^*_+ \ni \alpha \mapsto N(\alpha,\beta)$ is continuously strictly decreasing with
$$
\lim_{\alpha\to + \infty} N(\alpha,\beta)=0, \qquad \forall \beta > 0, \qquad \lim_{\beta\to 0} N(0,\beta)=+\infty.
$$
\end{lemma}
\begin{proof} The proof is simplest by using the spectral decomposition of $\mathbb H_{\alpha,\beta}$ instead of each of the $H_n$, $n \geq 2$. We have already observed in the proof of Lemma \ref{lem:comp} that $\mathbb H$ has a compact resolvent, and therefore so does $\mathbb H_{\alpha,\beta}$. Let $\{\psi_\ell\}_{\ell \in \mathbb N}$ be a basis of $\mathfrak F_{b/f}$ of eigenvectors of $\mathbb H$. Then \fref{belowHa} for $\alpha=0$ shows that $\psi_\ell \in D(\mathcal N)$. Then, since $\mathcal N$ and $\mathbb H$ commute, we can pick the basis $\{\psi_\ell\}_{\ell \in \mathbb N}$ such that
$$
\mathcal N \psi_\ell=n_\ell \psi_\ell, \qquad \ell \in \mathbb N,
$$
for some $n_\ell \geq 0$. The eigenvalues of $\mathbb H_{\alpha,\beta}$ are as a consequence of the form $\beta \lambda_\ell+ \alpha n_\ell$, for $\{\lambda_\ell\}_{\ell \in \mathbb N}$ the eigenvalues of $\mathbb H$. The $\lambda_\ell$'s are arranged into a nondecreasing sequence that tends to the infinity. The zero eigenvalue is simple and associated with the vacuum eigenvector $1 \oplus_{n=1}^\infty 0$, and we set $\lambda_0=n_0=0$. We have as well $n_\ell \geq 1$ and $\lambda_\ell \geq \gamma$ for $\ell \geq 1$ according to \fref{Hbelow}. Of course, the $\lambda_\ell$ and $n_\ell$ are different for fermions and bosons.
The partition function $Z_{\alpha,\beta}$ then reads
$$
Z_{\alpha,\beta}=\sum_{\ell \in \mathbb N} e^{-\beta \lambda_\ell -\alpha n_\ell}>1.
$$
It is then direct to show, using dominated convergence for series, that $Z_{\alpha,\beta}$ is continuously infinitely differentiable for $(\alpha,\beta)\in \mathbb R^*_+ \times \mathbb R_+^*$. The expressions in \fref{expN} follow then easily, and as a consequence $N$ and $E$ are also continuously infinitely differentiable for $(\alpha,\beta)\in \mathbb R^*_+ \times \mathbb R_+^*$. In particular,
$$\displaystyle
\partial_\alpha N(\alpha,\beta)=-\frac{\left(\sum_{\ell \in \mathbb N} n^2_\ell e^{-\beta \lambda_\ell -\alpha n_\ell}\right) \left(\sum_{\ell \in \mathbb N} e^{-\beta \lambda_\ell -\alpha n_\ell}\right)-\left(\sum_{\ell \in \mathbb N} n_\ell e^{-\beta \lambda_\ell -\alpha n_\ell}\right)^2}{Z_{\alpha,\beta}^2}.
$$
The Cauchy-Schwarz inequality then shows that $\partial_\alpha N(\alpha,\beta)<0$ since the equality case is not possible as there exist some indices $\ell$ and $\ell'$ for which $n_\ell \neq n_{\ell'}$ (this is easily seen by remarking for instance that if $n_\ell= \bar n <\infty$ for all $\ell \geq 1$, then $\|\mathcal N \psi\|=\bar n \|\psi\|$ for all $\psi \in D(\mathcal N)$, which is absurd since $\mathcal N$ is unbounded). A similar calculation shows that $\partial_\beta E(\alpha,\beta)<0$ for $(\alpha,\beta)\in \mathbb R^*_+ \times \mathbb R_+^*$.
The limits as $\alpha \to +\infty$ for $N$ and as $\beta \to +\infty$ for $E$ follow easily by dominated convergence. For the limit as $\beta \to 0$ of $E$, set $\ell_M$ such that $\lambda_{\ell_M} \geq M$ for some $M>0$ fixed. Since $\sum_{\ell \geq \ell_M} e^{-\beta \lambda_\ell -\alpha n_\ell}$ converges to $+\infty$ as $\beta \to 0$ (for otherwise the operator $e^{-\alpha \mathcal N}$ would be trace class), and since $\sum_{\ell < \ell_M} e^{-\beta \lambda_\ell -\alpha n_\ell}\leq \sum_{\ell < \ell_M}=\ell_M$, there exists $\beta_0(M)$ such that
$$
\sum_{\ell < \ell_M} e^{-\beta \lambda_\ell -\alpha n_\ell} \leq \sum_{\ell \geq \ell_M} e^{-\beta \lambda_\ell -\alpha n_\ell}, \qquad \forall \beta\leq \beta_0(M).
$$
Hence,
$$
\frac{M}{2} \leq \frac{1}{2} \frac{M\sum_{\ell \geq \ell_M} e^{-\beta \lambda_\ell -\alpha n_\ell} }{\sum_{\ell \geq \ell_M} e^{-\beta \lambda_\ell -\alpha n_\ell}} \leq E(\alpha,\beta),
$$
which proves the second limit in (ii).
It remains to treat the second limit in (iii). Since the operator $\mathcal N$ is not bounded on $\mathfrak F_{b/f}$, we can rearrange the sequence $\{n_\ell\}_{\ell \in \mathbb N}$ into a nondecreasing sequence $\{ n_{\ell_j}\}_{j \in \mathbb N}$ such that $n_{\ell_j} \to \infty$ as $j \to \infty$. Proceeding as in the proof of the second limit in (ii), we then find that for all $M>0$, there exist $j_M$ and $\beta_0(M)$ such that
$$
\frac{M}{2} \leq \frac{1}{2} \frac{M \sum_{ j \geq j_M} e^{-\beta\lambda_{\ell_j}}}{\sum_{ j \geq j_M} e^{-\beta \lambda_{\ell_j}}} \leq \frac{\sum_{ j \in \mathbb N} n_{\ell_j}e^{-\beta\lambda_{\ell_j}}}{\sum_{ j \in \mathbb N} e^{-\beta\lambda_{\ell_j}}}=N(0,\beta), \qquad \forall \beta \leq \beta_0(M).
$$
This proves the second limit in (iii) and ends the proof of the lemma.
\end{proof}
\bigskip
We proceed now to the proof of Proposition \ref{gminS}. First of all, according to Lemma \ref{lemZ} (ii), there exists, for all $a>0$, a $\beta_0(a) \in (0,\infty)$ such that
\mathbf e} \newcommand{\bh}{\mathbf h \label{egE}
E(0,\beta_0(a))=a.
\end{equation}
Since $E(0,\beta)$ is continuously differentiable and strictly monotone, the global version of the implicit function theorem implies that $a \mapsto \beta_0(a)$ is continuously differentiable.
Let now $\varrho \in \mathcal E$ and consider the free energy
$$
F_{\beta_0(a)}(\varrho)=\beta^{-1}_0(a) S(\varrho)+ \Tr (\mathbb H^{1/2} \varrho \mathbb H^{1/2} ).
$$
For $\mathcal F$ the relative entropy introduced in Section \ref{secpropS}, we find
$$
\beta_0(a) F_{\beta_0(a)}(\varrho)=\mathcal F(\varrho,\varrho_{0,\beta_0(a)})-\log Z_{0,\beta_0(a)},
$$
and as a consequence, for any $\varrho \in \mathcal A_{g,e}(a)$,
$$
S(\varrho)=\mathcal F(\varrho,\varrho_{0,\beta_0(a)})-\log Z_{0,\beta_0(a)}-a \beta_0(a).
$$
Since $\mathcal F(\varrho,\varrho_{0,\beta_0(a)})=0$ if and only if $\varrho=\varrho_{0,\beta_0(a)}$, we obtain that $\varrho_{0,\beta_0(a)}$ is the unique minimizer of $S$ in $\mathcal A_{g,e}(a)$. This proves the existence part of Proposition \ref{gminS}.
Regarding the monotonicity of the entropy, we have
$$
f(a)=S(\varrho_{0,\beta_0(a)})=-\log Z_{0,\beta_0(a)}-a \beta_0(a),$$
which is continuously differentiable w.r.t. $a$ since $Z_{0,\beta}$ and $\beta_0(a)$ are both $C^1$ w.r.t $\beta$ and $a$, respectively. Then, thanks to Lemma \ref{lemZ} (i) and \fref{egE},
$$
f'(a)=\beta_0'(a) E(0,\beta_0(a))-\beta'_0(a) a -\beta_0(a)=-\beta_0(a)<0.
$$
This ends the proof.
\subsection{Proof of Proposition \ref{gminF}} \label{propproof3}
The proof is very similar to that of Proposition \ref{gminS}.
First of all, according to Lemma \ref{lemZ} (iii), we have
$$
N(\alpha,\beta) \leq N(0,\beta), \qquad \forall (\alpha,\beta) \in \mathbb R_+ \times \mathbb R_+^*.
$$
Note that $N(0,\beta)$ is well-defined for $\beta>0$ as proved in \fref{alz}, and that $N(\alpha,\beta)$ is continuous at $\alpha=0$ for all $\beta>0$ as a direct application of monotone convergence. Pick $a_0>0$. We will set $\beta$ sufficiently small so that all $a>0$ less than $a_0$ are in the range of $N(\alpha,\beta)$ for all $\alpha$ sufficiently large.
Since $N(0,\beta) \to +\infty$ as $\beta \to 0$ according to Lemma \ref{lemZ} (iii), there exists $\beta_0(a_0)$ such that
$$
N(0,\beta) \geq a_0, \qquad \forall \beta \leq \beta_0(a_0).
$$
We fix from now on a $\beta>0$ such that $\beta \leq \beta_0(a_0)$. Then, since $\alpha \mapsto N(\alpha,\beta)$ is strictly decreasing, there exists $\alpha_1>0$ such that
$$
N(\alpha,\beta) \leq a_0, \qquad \forall \alpha \geq \alpha_1,
$$
and for all $a \in (0,a_0]$, there exists $\alpha_0(a) \in [\alpha_1,\infty)$ such that
\mathbf e} \newcommand{\bh}{\mathbf h \label{egN}
N(\alpha_0(a),\beta)=a. \qquad
\end{equation}
Since $N(\alpha,\beta)$ is continuously differentiable and strictly monotone in $\alpha$, the global version of the implicit function theorem implies that $\alpha \mapsto \alpha_0(a)$ is continuously differentiable. As in the proof of Proposition \ref{gminS}, we find that, for any $\varrho \in \mathcal A_{g}(a)$,
\begin{eqnarray*}
\beta F_\beta(\varrho)&=&\mathcal F(\varrho,\varrho_{\alpha_0(a),\beta})-\log Z_{\alpha_0(a),\beta}-\alpha_0(a) \Tr (\mathcal N^{1/2} \varrho \mathcal N^{1/2})\\
&=& \mathcal F(\varrho,\varrho_{\alpha_0(a),\beta})-\log Z_{\alpha_0(a),\beta}-\alpha_0(a) a,
\end{eqnarray*}
which shows that $\varrho_{\alpha_0(a),\beta}$ is the unique minimizer of $F_\beta$ in $\mathcal A_{g}(a)$.
Regarding the monotonicity of the entropy, we have
$$
\beta g(a)=\beta F_\beta(\varrho_{\alpha_0(a),\beta})=-\log Z_{\alpha_0(a),\beta}-a \alpha_0(a),$$
which is continuously differentiable w.r.t. $a$ since $Z_{\alpha,\beta}$ and $\alpha_0(a)$ are both $C^1$ w.r.t $\alpha$ and $a$, respectively. Then, thanks to Lemma \ref{lemZ} (i) and \fref{egN},
$$
\beta g'(a)=\alpha_0'(a) N(\alpha_0(a),\beta)-a \alpha'_0(a)-\alpha_0(a)=-\alpha_0(a)<0.
$$
This ends the proof.
\section{Appendix}
\subsection{Justification of Definition \ref{onebody}}
The fact that $\varrho^{(1)}$ is well-defined is established as follows. Set
$$
\mathbb B= 0 \oplus \bigoplus_{n = 1}^{+\infty} \mathbb{B}_{(n)}:=0 \oplus \bigoplus_{n = 1}^{+\infty} n^{-1}\mathbb{A}_{(n)},
$$
where $\mathbb A$ is the second quantization of $A$ and $\mathbb A_{(n)}$ its component on the $n$-th sector.
Then, we have the estimate
\mathbf e} \newcommand{\bh}{\mathbf h \label{boundB}
\| \mathbb B \|_{\mathcal L(\mathfrak{F}_{b/f})} \leq \|A\|_{\mathcal L(\mathfrak{h})}.
\end{equation}
Indeed, for $\psi =\{\psi^{(n)}\}_{n\in \mathbb N}\in \mathfrak{F}_{b/f}$ and $n \geq 1$, we have
$$
\mathbb B_{(n)} \psi^{(n)}=(\mathbb B \psi)_{(n)}=\frac{1}{n} \left(\sum_{i=1}^n (A)_{i} \psi^{(n)} \right),
$$
for $(A)_{i}$ the operator $A$ acting on the variable $x_i$. This yields directly
$$
\| \mathbb B_{(n)} \psi^{(n)}\|_{n} \leq \|A\|_{\mathcal L(\mathfrak{h})} \| \psi^{(n)}\|_{n}.
$$
Since $\| \mathbb B \|_{\mathcal L(\mathfrak{F}_{b/f})}=\sup_{n \in \mathbb N} \|\mathbb B_{(n)}\|_{n}$, this gives \fref{boundB}.
For $\varrho \in \mathcal S_0$, consider now the linear map
$$
F: A \in \mathcal L(\mathfrak{h}) \mapsto F(A)=\Tr( \mathbb A \varrho).
$$
It is well-defined since, using the above notation for $\mathbb B$,
$$
\left|\Tr( \mathbb A \varrho) \right|=\left|\mathrm{Tr} \left (\mathbb B \mathcal N^{1/2} \varrho\, \mathcal N^{1/2}\right)\right| \leq C \|A\|_{\mathcal L(\mathfrak{h})},
$$
where we used the fact that
$$
\mathrm{Tr} \left (\mathcal N^{1/2} \varrho\, \mathcal N^{1/2}\right)=C<\infty,
$$
since $\varrho \in \mathcal S_0$. When $A$ is compact, $F$ is therefore a linear continuous map on the space of compact operators on $\mathfrak{h}$. By duality, we can then conclude that there exists a unique $\varrho^{(1)} \in \mathcal J_1(\mathfrak{h})$ such that
$$
F(A)=\mathrm{Tr}_{\mathfrak{h}}\left(A \varrho^{(1)}\right),
$$
for all compact operators $A$ on $\mathcal L(\mathfrak{h})$. The case $A$ bounded follows finally by approximation.
The fact that $\varrho^{(1)}$ is nonnegative is established as follows. Let $\varphi \in \mathfrak{h}$ with $\|\varphi\|_{\mathfrak{h}}=1$, and consider the rank one projector $P=\ket{\varphi}\bra{\varphi}$. Then
$$
(\varphi,\varrho^{(1)} \varphi)_\mathfrak{h}=\mathrm{Tr}_{\mathfrak{h}}\left(P \varrho^{(1)}\right) =\mathrm{Tr} \left( d\Gamma(P) \varrho\right).
$$
Denoting by $\{\rho_p\}_{p \in \mathbb N}$ and $\{\psi_p\}_{p \in \mathbb N}$ the (nonnegative) eigenvalues and eigenfunctions of $\varrho \in \mathcal E_0$, the last term is equal to
$$
\sum_{p \in \mathbb N} \sum_{n \in \mathbb N^*} \sum_{j=1}^n \rho_p \left(\psi_p^{(n)}, P_{(j)}\psi_p^{(n)}\right)_{n}= \sum_{p \in \mathbb N} \sum_{n \in \mathbb N^*} \sum_{j=1}^n \rho_p \left(P_{(j)} \psi_p^{(n)}, P_{(j)}\psi_p^{(n)}\right)_{n} \geq 0.
$$
Above $P_{(j)}$ is the operator $P$ acting on $x_j$, and we used the fact that $P^2_{(j)}=P_{(j)}$. This yields the positivity and ends the justification of Definition \ref{onebody}.
\subsection{Proof of Lemma \ref{enrho}}
Denote by $\{\rho_p\}_{p \in \mathbb N}$ and $\{\psi_p\}_{p \in \mathbb N}$ the eigenvalues and eigenfunctions of $\varrho \in \mathcal E_0$. A direct calculation shows first that
$$
\Tr\left( \mathbb H_0^{1/2} \varrho \,\mathbb H_0^{1/2} \right)=\sum_{p \in \mathbb N} \sum_{n \in \mathbb N^*} \rho_p \left\| d\Gamma(h_0)^{1/2}_{(n)} \psi_p^{(n)}\right\|^2_{n},
$$
where $\psi_p^{(n)}$ is the component of $\psi_p$ on the $n$-th sector and
$$
d\Gamma(h_0)^{1/2}_{(n)} = \left(\sum_{j=1}^n -\Delta_{x_j}\right)^{1/2}.
$$
Since $h_0$ is not bounded, we proceed by regularization in order to use \fref{def1} for the definition of the one-body density matrix. For $\varepsilon>0$, set then $h_\varepsilon=h_0(\mathrm{Id}_{\mathfrak{h}}+\varepsilon h_0)^{-1} \in \mathcal L(\mathfrak{h})$. According to \fref{def1}, we have
\mathbf e} \newcommand{\bh}{\mathbf h \label{rel}
\mathrm{Tr}_\mathfrak{h} \left( h_\varepsilon^{1/2} \varrho^{(1)} h_\varepsilon^{1/2} \right)=\mathrm{Tr}_\mathfrak{h} \left( h_\varepsilon \varrho^{(1)} \right)=
\Tr\left( d\Gamma(h_\varepsilon) \varrho \right).
\end{equation}
The last term is equal to
\mathbf e} \newcommand{\bh}{\mathbf h \label{rel1}
\sum_{p \in \mathbb N} \sum_{n \in \mathbb N^*} \rho_p \left\| \mathbb A^\varepsilon_{(n)} \psi_p^{(n)}\right\|^2_{n},
\end{equation}
with
$$
\mathbb A^\varepsilon_{(n)}=\left(\sum_{j=1}^n -\Delta_{x_j} (\mathrm{Id}_{\mathfrak{h}}- \varepsilon \Delta_{x_j})^{-1}\right)^{1/2}.
$$
For $\psi \in \mathfrak{F}_{b/f}^{(n)}$, we find by a Fourier transform
\begin{eqnarray} \label{rel2}
\| \mathbb A^\varepsilon_{(n)} \psi \|^2_{(n)} =&(2 \pi)^{-n} \int_{(\mathbb R^d)^n} \left(\sum_{j=1}^n \frac{|k_j|^2}{1+\varepsilon |k_j|^2}\right)^{1/2} |\widehat \psi(k_1, \cdots,k_d)|^2 d k_1 \cdots d k_d\\\nonumber
\leq& \| d\Gamma(h_0)^{1/2}_{(n)} \psi \|^2_{(n)} .
\end{eqnarray}
Hence,
$$
\mathrm{Tr}_\mathfrak{h} \left( h_\varepsilon^{1/2} \varrho^{(1)} h_\varepsilon^{1/2} \right) \leq \Tr\left( \mathbb H_0^{1/2} \varrho \,\mathbb H_0^{1/2} \right),
$$
and there exists therefore $\alpha \in \mathcal J_1(\mathfrak{h})$ and a subsequence such that $h_{\varepsilon_\ell}^{1/2} \varrho^{(1)} h_{\varepsilon_\ell}^{1/2} \to \alpha$ in $\mathcal J_1(\mathfrak{h})$ weak-$*$ as $\ell \to \infty $ with
$$ \mathrm{Tr}_\mathfrak{h} \left( \alpha \right) \leq \Tr\left( \mathbb H_0^{1/2} \varrho \,\mathbb H_0^{1/2} \right).$$
We now identify $\alpha$ with $h_0^{1/2} \varrho^{(1)} h_0^{1/2}$. Let $K$ be a compact operator on $\mathfrak{h}$. Then,
$$
\mathrm{Tr}_\mathfrak{h} \left( (\mathrm{Id}_{\mathfrak{h}}+h_0)^{-1} K (\mathrm{Id}_{\mathfrak{h}}+h_0)^{-1} h_\varepsilon^{1/2} \varrho^{(1)} h_\varepsilon^{1/2} \right)= \mathrm{Tr}_\mathfrak{h} (K_\varepsilon \varrho^{(1)}),
$$
with
$$
K_\varepsilon=h_\varepsilon^{1/2} (\mathrm{Id}_{\mathfrak{h}}+h_0)^{-1} K (\mathrm{Id}_{\mathfrak{h}}+h_0)^{-1} h_\varepsilon^{1/2}.
$$
As $\varepsilon \to 0$, the operator $K_\varepsilon$ converges strongly to $h_0^{1/2} (\mathrm{Id}_{\mathfrak{h}}+h_0)^{-1} K (\mathrm{Id}_{\mathfrak{h}}+h_0)^{-1} h_0^{1/2}$ in $\mathcal L(\mathfrak{h})$. As a consequence
\begin{align*}
\mathrm{Tr}_\mathfrak{h} \big( (\mathrm{Id}_{\mathfrak{h}}+h_0)^{-1} K& (\mathrm{Id}_{\mathfrak{h}}+h_0)^{-1} \alpha \big)\\
&= \mathrm{Tr}_\mathfrak{h} \left( K h_0^{1/2} (\mathrm{Id}_{\mathfrak{h}}+h_0)^{-1} \varrho^{(1)} (\mathrm{Id}_{\mathfrak{h}}+h_0)^{-1} h_0^{1/2} \right),
\end{align*}
which allows us to identify $\alpha$ with $h_0^{1/2} \varrho^{(1)} h_0^{1/2}$. The relation \fref{eqkin} is obtained by passing to the limit in \fref{rel}: in the l.h.s, we use the fact that $h_0^{1/2} \varrho^{(1)} h_0^{1/2} \in \mathcal J_1(\mathfrak{h})$, and in the r.h.s., we use \fref{rel1}, \fref{rel2}, and monotone convergence. This ends the proof.
\bibliographystyle{plain}
|
\section{Scattering processes and Snell's law}
The local and nonlocal conductance in the main text is determined by the possible scattering processes available in the junction. For brevity, we here consider only the antiferromagnet--superconductor--antiferromagnet (AF--S--AF) junction where the spin is degenerate. The left (right) AF will be referred to as lead $0$ $(1)$.
Consider an incident electron $(e)$ with angle of incidence $\theta$ in lead 0. The possible scattering processes are:
\begin{enumerate}
\item Reflected electron into lead $0$ (NR) with angle of reflection $\theta_e^0$.
\item Reflected hole into lead $0$ (AR) with angle of reflection $\theta_h^0$.
\item Transmitted electron into lead $1$ (CT) with angle of transmission $\theta_e^1$.
\item Transmitted hole into lead $1$ (CAR) with angle of transmission $\theta_h^1$.
\end{enumerate}
Due to the translational invariance in the $y$-direction we can write down Snell's law for the possible scattering processes,
\begin{equation}
\begin{split}
k_e^0 \sin \theta = k_e^0 \sin \theta_e^0 = k_h^0 \sin \theta_h^0 =k_e^1 \sin \theta_e^1 = k_h^1 \sin \theta_h^1.
\end{split}
\label{eq:Snells_Law}
\end{equation}
Here the wavenumbers are
\begin{equation}
\hbar v_{\mathrm{F}}k^{j}_{e(h)} = \sqrt{\left(E_{\mathrm{F}j} \pm E\right)^2 - J_j^2},
\label{eq:electron_dispersion}
\end{equation}
and the index $j=\{0,1\}$ labels the lead. Equations \eqref{eq:Snells_Law} and \eqref{eq:electron_dispersion} are used to determine the angles of reflection $\theta_{e(h)}^0$ and transmission $\theta_{e(h)}^1$, for a given angle of incidence $\theta$. Furthermore, the scattering processes AR, CT, and CAR have the corresponding critical angles
\begin{equation}
\begin{split}
&\theta_{c,h}^0=\arcsin{\frac{k_h^0}{k_e^0} }, \quad \theta_{c,e}^1=\arcsin{\frac{k_e^1}{k_e^0} },\quad\, \mathrm{and}\quad \theta_{c,h}^1=\arcsin{\frac{k_h^1}{k_e^0} },
\end{split}
\end{equation}
respectively. For angles of incidence $\theta$ greater (smaller) than a given critical angle the corresponding scattering processes is an evanescent (propagating) state, giving zero (finite) contribution to the relevant conductance.
\section{Solution Ansatz of the BdG equation}
We can express the wavefunctions in lead $0$, the superconductor, and lead $1$ as
\begin{equation}
\begin{split}
&\Psi_0 = \Psi_{\mathrm{In}} + \sum_{s=\{\mathrm{P},\mathrm{AP}\}}\left( r_{e,s} \Psi_{r_{e,s}} + r_{h,s} \Psi_{r_{h,s}}\right),\\
&\Psi_{\mathrm{S}}=\sum_{i=1}^8 a_i \Phi_i,\\
&\Psi_1 = \sum_{s=\{\mathrm{P},\mathrm{AP}\}}\left( t_{e,s} \Psi_{t_{e,s}} + t_{h,s} \Psi_{t_{h,s}}\right),\\
\end{split}
\end{equation}
respectively. Here $\Psi_{r_{e(h),s}}$ and $\Psi_{t_{e(h),s}}$ are the eigenstates corresponding to the scattering processes NR (AR) and CT (CAR) in the leads. The eigenstates in the superconductor is denoted by $\Phi_i$. The spin-index $s = \mathrm{(A)P}$ refers to the spin polarization being (anti)parallel to the N\'eel vector. The reflection $r_{e(h),s}$ and transmission $t_{e(h),s}$ amplitudes are then determined by requiring continuity of the wavefunction
\begin{equation}
\Psi_0\left(x=-L_{\mathrm{S}}/2\right)=\Psi_S\left(x=-L_{\mathrm{S}}/2\right)\, \mathrm{and}\, \Psi_S\left(x=L_{\mathrm{S}}/2\right)=\Psi_1\left(x=L_{\mathrm{S}}/2\right)
\end{equation}
at the AF--S interfaces.
\section{Robustness of the CAR dominated signal}
In the main text we argued that a perfect CAR dominated signal can be obtained on the voltage interval $\left(0, \Delta_0\right)$ if the gate voltage is tuned such that $E_{\mathrm{F}} = J$, and $E_{\mathrm{F}}/\Delta_0>1/2$. We also discussed the robustness of deviations from these assumptions. If the deviation of $E_{\mathrm{F}}$ from $J$ is greater than the voltage bias then CT dominates the sign of the nonlocal conductance. In this section, we quantitatively demonstrate the competition between CT and CAR.
In Fig. \ref{fig:robust} we have plotted the nonlocal conductance as a function of the parameters $J$ and $E_{\mathrm{F}}$. Fig. \ref{fig:robust} exhibits the general behaviour of Fig. 4 in the main text. The system is insulating when $J>eV + E_{\mathrm{F}}$, and exhibits perfect CAR when $J> \pm \left( eV-E_{\mathrm{F}} \right)$. When these inequalities are not satisfied there is a competition between CAR and CT, that determines the sign of the nonlocal conductance.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\columnwidth]{fig1sm.jpg}
\caption{Nonlocal conductance as a function of $J$ and $E_{\mathrm{F}}$. The figure is an example of the general behaviour discussed in Fig. 4 of the main text. We have used the concrete parameters $L_{\mathrm{S}}/\xi = 1$ and $eV/\Delta_0 = 0.8$.}
\label{fig:robust}
\end{figure}
We can understand the general features of Fig. \ref{fig:robust} analytically by considering the band structures of lead 1 and the superconductor in the subgap regime $eV/\Delta_0<1$ shown in Fig. \ref{fig:banddiagramsupplement.png}. There are in total four cases to consider:
\begin{enumerate}
\item For $\Delta_0 < E_{\mathrm{F}} - J$, CT and CAR competes for all $eV<\Delta_0$.
\item For $ E_{\mathrm{F}} - J<\Delta_0 < E_{\mathrm{F}} + J$, we find perfect CAR if $E_{\mathrm{F}}- J<eV < \Delta_0$, and competition between CT and CAR if $eV < E_{\mathrm{F}} - J$.
\item For $\Delta_0 > E_{\mathrm{F}} + J$, we find perfect CAR if $E_{\mathrm{F}}- J < eV < E_{\mathrm{F}}+ J$, and competition between CT and CAR on the intervals $eV < E_{\mathrm{F}}- J$ and $E_{\mathrm{F}}+ J<eV<\Delta_0$.
\item For $E_{\mathrm{F}} < J$ and $eV < J - E_{\mathrm{F}}$ the system is insulating.
\end{enumerate}
By solving these inequalities in terms of $J$ we obtain Fig. 4 in the main text.
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{fig2sm.png}
\caption{The possible scattering processes allowed in lead 1, depending on the relative values of $J$, $E_{\mathrm{F}}$, and $\Delta_0$.}
\label{fig:banddiagramsupplement.png}
\end{figure}
In Fig. \ref{fig:deviationsweep} we have plotted the contributions from CT, CAR, and the total nonlocal conductance $\mathcal{G}_{\mathrm{NL}}$ as a function of $eV/\Delta_0$. We have parameterized the exchange interaction as
\begin{equation}
J = E_{\mathrm{F}} - l\,\Delta_0.
\end{equation}
Here $l$ is a numerical parameter that determines the deviation between $E_{\mathrm{F}}$ and $J$ in terms of $\Delta_0$. In Fig. \ref{fig:deviationsweep} we have performed a parameter sweep over $l$. Concretely we have let $l$ be on the interval $\left(0,2\right)$. The figure shows that as the deviation parameter increases, CT gradually starts to dominate CAR resulting in a sign change of the nonlocal conductance.
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{fig3sm.jpg}
\caption{A parameter sweep over the deviation parameter $l$. In the figure we have used the values $l = \{0, 0.2, 0.4, 0.6, 0.8, 1, 1.2, 1.4, 1.6, 1.8, 2\}$, and fixed $E_{\mathrm{F}}= 10\Delta_0$. The arrows indicate curves with increasing deviation parameter $l$. As the deviation parameter increases CT starts to gradually dominate CAR. Other choices of $E_{\mathrm{F}}$ result in the same qualitative behaviour.}
\label{fig:deviationsweep}
\end{figure}
\newpage
\section{Oscillations in the Nonlocal conductance}
In Fig. \ref{fig:oscillationls.jpg} we have plotted the nonlocal conductance $\mathcal{G}_{\mathrm{NL}}$ as a function of the length of the superconductor $L_{\mathrm{S}}/\xi$, for $eV/\Delta_0 = 0.5$. We observe rapid oscillations, due to the formation of resonant transmission levels inside the superconductor. The oscillation frequency is determined by the ratio $\Delta_0/E_{\mathrm{FS}}$. The oscillating behaviour is also present in NM--S--NM and FM--S--FM junctions, as confirmed by previous studies referenced in the main text. Importantly, CT is suppressed for all lengths $L_{\mathrm{S}}/\xi$, when $E_{\mathrm{F}} = J$ and $E_{\mathrm{F}}/\Delta_0>1/2$, resulting in a CAR dominant signal. The nonlocal conductance is exponentially suppressed in the limit $L_{\mathrm{S}}/\xi \rightarrow \infty$, where the spatial separation of the electrons in lead 0 and lead 1 is greater than the approximate length of a corresponding Cooper pair.
\begin{figure}[h]
\centering
\includegraphics[width=0.5\columnwidth]{fig4sm.jpg}
\caption{Normalized nonlocal conductance $\mathcal{G}_{\mathrm{NL}}$ as a function of the length of the superconductor $L_{\mathrm{S}}/\xi$. Here $\xi=\hbar v_{\mathrm{F}}/\Delta$ is the approximate length of a Cooper pair. We have set $eV/\Delta_0 = 0.5$, $E_{\mathrm{F}}/\Delta_0 = J/\Delta_0 = 10$, and $E_{\mathrm{FS}}/\Delta_0 = 1000$.}
\label{fig:oscillationls.jpg}
\end{figure}
\newpage
\section{Anisotropy in conductance}
In Fig. \ref{fig:anisotropy} we have plotted the normalized nonlocal conductance $\mathcal{G}_{\mathrm{NL}}$ as a function of the angle of misalignment $\delta \gamma = \arccos{\left(\bm{n}_0\cdot\bm{n}_1\right)}$ between the order parameters $\bm{n}_0$ and $\bm{n}_1$ in lead 0 and lead 1 respectively. We have considered the cases where lead 1 can be: electron (hole) doped (anti)ferromagnets. We have fixed $J = E_{\mathrm{F}} = E_{\mathrm{F}0} = - E_{\mathrm{F}1}$, $L_{\mathrm{S}}/\xi = 1$, and $eV/\Delta_0 = 0.5$. In accordance with the literature mentioned in the main text we find a sign-change effect for the electron doped $\mathrm{FM}_1$. Since, the energy bands are spin-degenerate in $\mathrm{AF}_1$ no sign-change effect is observed. However, all magnetic junctions exhibit an anisotropy as a consequence of the misalignment between the order parameters in lead 0 and lead 1. Notice that for the hole doped $\mathrm{AF}_1$ the nonlocal conductance is negative for any misalignment.
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{fig5sm.jpg}
\caption{Normalized nonlocal conductance $\mathcal{G}_{\mathrm{NL}}$ as a function of the misalignment angle $\delta \gamma$ between the order parameter vectors. In all cases lead $0$ is electron doped. These plots clearly show that the sign of the total nonlocal conductance is very sensitive to the order parameter direction in FM leads, while it remains unchanged in AF leads.}
\label{fig:anisotropy}
\end{figure}
\section{Interface effects}
In the main text, we have assumed that the interface is $i)$ ideal and $ii)$ magnetically compensated. We here want to investigate the effect of lifting these assumptions quantitatively. In the BdG Hamiltonian, interface effects can be modelled by a delta-function-like potential. The interface Hamiltonian, incorporating $i)$ and $ii)$ is
\begin{equation}
H_\mathrm{I} =
\begin{bmatrix}
H_{V} + H_{h} & 0 \\
0 & -H_{V} - H_{h}^T
\end{bmatrix}\delta(x-x_0),
\label{eq:BdG_interface_term}
\end{equation}
where $x_0$ is the position of the interface.
We model a non--ideal interface with a spin-independent potential barrier of strength $V>0$,
\begin{equation}
H_{V} = V s_0 \otimes \sigma_0.
\end{equation}
To incorporate the effect of an uncompensated interface, we introduce the following spin-dependent potential
\begin{equation}
H_{h} = \frac{h}{2} \left(\mathbf{n}(x)\cdot \mathbf{s}\right)\otimes \left(\sigma_0 + \sigma_z\right),
\end{equation}
modelling a net magnetization with strength $h$. The Pauli matrices $\bm{s}$ and $\bm{\sigma}$ denote spin and sublattice degrees of freedom.
Adding the interface potential, Eq. \eqref{eq:BdG_interface_term}, to the BdG Hamiltonian of the main text, and integrating over the interface at $x_0$, we obtain the following continuity condition
\begin{equation}
i \hbar v_{\mathrm{F}}\, \tau_z \otimes s_0 \otimes \sigma_x \left[ \psi(x_0^+) - \psi(x_0^-) \right]=
\left[V \tau_z \otimes s_0 \otimes \sigma_0
+ \frac{h}{2} \tau_4 \otimes \left(\mathbf{n}(x)\cdot \mathbf{s}\right)\otimes \left(\sigma_0 + \sigma_z\right) \right]\psi(x_0)
\end{equation}
where we introduced the Pauli matrices for charge degrees of freedom (electron/hole) $\bm{\tau}$ and the matrix $\tau_4 = \mathrm{diag}\left(1,-T\right)$, where $T$ represents the transpose operator.
To quantitatively investigate the effect of imperfect or uncompensated interfaces, we introduce the dimensionless barrier strength $Z_V = V/\hbar v_{\mathrm{F}}$ and the dimensionless uncompensated magnetic moment $Z_h = h/\hbar v_{\mathrm{F}}$ respectively. In Fig. \ref{fig:fig6sm}, we demonstrate how the interface effects changes the non--local conductance.
An imperfect interface, modelled by spin-independent potential, does not suppress or qualitatively change the non--local conductance, only a shift in the position of the conductance peak is observed. We attribute this behaviour to the ''relativistic'' effects associated with Klein-like tunnelling, masking the potential barrier for hexagonal lattices.
In non-relativistic Hamiltonians, like a square lattice AF system, we expect that the non--local conductance decreases with increasing the barrier strength.
The effect of an uncompensated interface that leads to a net interface magnetization can be more dramatic. For moderate net magnetizations at the interface, the amplitude of the non--local conductance is slightly reduced. For larger net magnetizations at the interface, the conductance gradually decreases and vanishes completely in the limit of very strong net magnetization. This occurs because a strong fixed-direction magnetization closes the spin-flip process channels in the system. Since finite local and non--local ARs need a spin-flip channel, both are suppressed in the strong magnetization regime. This leaves NR as the dominant scattering process and thus, both local and non--local conductances are suppressed.
\begin{figure}[h]
\centering
\includegraphics[width=1\columnwidth]{fig6sm.jpg}
\caption{Normalized nonlocal conductance $\mathcal{G}_{\mathrm{NL}}$ as a function of the potential barriers $Z_V$ and $Z_h$ parameterizing the transparency of the interface and degree of uncompensated magnetic moments respectively.}
\label{fig:fig6sm}
\end{figure}
\end{document} |
\section{Introduction}
\label{intro}
Here, we examine the near-wall scaling behavior of turbulent flows on smooth walls. These flows include two-dimensional zero-pressure gradient boundary layers, and fully developed pipe and channel flows. The focus is on the region $y^+<100$, which includes the peaks in the streamwise and spanwise turbulent stresses. The superscript $^+$ denotes non-dimensionalization using the fluid kinematic viscosity $\nu$ and the friction velocity $u_\tau=\sqrt{\tau_w/\rho}$, where $\tau_w$ is the mean wall shear stress and $\rho$ is the fluid density.
For isothermal, incompressible flow, it is commonly assumed that for the region close to the wall
$$[U_i,\overline{({u_i u_j})^+}] = f(y, u_\tau, \nu, \delta),$$
where $U_i$ and $u_i$ are the mean and fluctuating velocities in the $i$th direction. The overbar denotes ensemble averaging, and the outer length scale $\delta$ is, as appropriate, the boundary layer thickness, the pipe radius, or the channel half-height. That is,
\begin{equation}
[U_i^+,\overline{({u_i u_j})^+}] = f(y^+, Re_\tau).
\label{dim_analysis}
\end{equation}
where the friction Reynolds number $Re_\tau=\delta u_\tau/\nu$. The Reynolds number dependence of $\overline{u^2}$ (streamwise direction), $\overline{v^2}$ (wall-normal direction), $\overline{w^2}$ (spanwise direction), and $-\overline{uv}$ (Reynolds shear stress) is illustrated in figure~\ref{all_stresses}.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{figures/u2_DNS_no180}
\includegraphics[width=0.47\textwidth]{figures/v2_DNS_no180} \\
\includegraphics[width=0.47\textwidth]{figures/w2_DNS_no180}
\includegraphics[width=0.47\textwidth]{figures/uv_DNS_no180}
\caption{ Profiles of turbulent stresses in channel flow, as computed by DNS $Re_\tau =$ 550, 1000, 2000, 5200 \citep{lee2015}. }
\label{all_stresses}
\end{figure}
\cite{hultmark2021smits} considered the scaling of the streamwise component $\overline{u^2}$. Specifically, they showed that a new velocity scale, defined by $b_1 u_\tau$ collapsed the profiles of $\overline{u^2}$ onto a single curve for the whole of the region $y^+ < 20 $ (including the inner peak in $\overline{u^2}$). That is,
\begin{equation}
\overline{{u^2}^+} = \overline{b_1^2} (Re_\tau) f(y^+).
\label{fg}
\end{equation}
The parameter $\overline{b_1^2}$ scales with the mean square value of the fluctuating wall shear stress $\tau_w'$, and it was derived using a Taylor Series expansion of the fluctuating velocity very close to the wall. The analysis used the channel flow direct numerical simulations (DNS) by \cite{lee2015}, with strong support from high Reynolds number experimental data by \cite{Vallikivi2015_HRTF} and \cite{samie2018fully}. Here, we generalize this scaling to include the other components of the Reynolds stress tensor $\overline{v^2}$ (wall-normal direction), $\overline{w^2}$ (spanwise direction), and $-\overline{uv}$ (Reynolds shear stress).
\section{Taylor Series expansions}
We start in the same way as \cite{hultmark2021smits} by writing the Taylor series expansions for $u_i$ in the vicinity of the wall. Instantaneously \citep{bewley2004skin},
\begin{eqnarray}
\underline u^+ & = & a_1 + b_1 y^+ + c_1{y^+}^2 + O({y^+}^3) \label{Taylor1u} \\
\underline v^+ & = & a_2 + b_2 y^+ + c_2{y^+}^2 + O({y^+}^3) \label{Taylor1v} \\
\underline w^+ & = & a_3 + b_3 y^+ + c_3{y^+}^2 + O({y^+}^3) \label{Taylor1w}
\end{eqnarray}
where $\underline u=U+u$, etc. The no-slip condition gives $a_1=a_2=a_3=0$, and by continuity $\partial v /\partial y|_w = b_2 = 0$.
Also
\begin{eqnarray}
b_1 & = & (\partial \underline u^+/ \partial y^+)_w, \label{b1} \\
b_3 & = & (\partial \underline w^+/ \partial y^+)_w, \label{b3} \\
c_2 & = & -{\textstyle {1 \over 2}} (\partial b_1/ \partial x^+ + \partial b_3/ \partial z^+). \label{c2}
\end{eqnarray}
For the time-averaged quantities
\begin{eqnarray}
\overline{{u^2}^+} & = & \overline{{b_1}^2}{y^+}^2 + \dots \label{Taylor1u2} \\
\overline{{v^2}^+} & = & \overline{{c_2}^2}{y^+}^4 + \dots \label{Taylor1v2} \\
\overline{{w^2}^+} & = & \overline{{b_3}^2}{y^+}^2 + \dots \label{Taylor1w2} \\
-\overline{{uv}^+} & = & \overline{b_1 c_2}{y^+}^3 + \dots \label{Taylor1uv}
\end{eqnarray}
These expansions can only hold very close to the wall. Since the mean velocity is linear in this region
\begin{eqnarray}
\tilde u/U & = & \tilde b_{1} \label{Taylor_rms_u} \\
\tilde v/U & = & \tilde c_{2} y^+ \label{Taylor_rms_v} \\
\tilde w/U & = & \tilde b_{3} \label{Taylor_rms_w} \\
\widetilde{-uv}/U & = & \widetilde {b_{1} c_2} \sqrt{y^+} \label{Taylor_rms_uv}
\end{eqnarray}
where $ \tilde u = \sqrt{\overline{{u^2}^+}}$, $ \tilde b_1 = \sqrt{\overline{b_1^2}}$, $ \widetilde {b_{1} c_2} = \sqrt{\overline{b_1c_1}}$, and $ \widetilde {u v} = \sqrt{\overline{u v}}$.
The channel flow DNS data given in figure~\ref{u_U_DNS} shows that at a given Reynolds number $ \tilde b_1$, $ \tilde c_2$ , $ \tilde b_3$ and $ \widetilde {b_{1} c_2}$ all become constant as the wall is approached.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{figures/u_U_DNS_no180} \hspace{3mm}
\includegraphics[width=0.47\textwidth]{figures/v_U_y_DNS_no180} \\
\includegraphics[width=0.45\textwidth]{figures/w_U_DNS_no180}
\includegraphics[width=0.51\textwidth]{figures/uv_U_y_DNS_no180}
\caption{ Top left: Profiles of $\tilde u/U$; intercept at $y^+=0$ is $\tilde b_{1}$. Top right: Profiles of $(\tilde v/U)/y^+$; intercept at $y^+=0$ is $\tilde c_{2}$. Bottom left: Profiles of $\tilde w/U$; intercept at $y^+=0$ is $\tilde b_{3}$. Bottom right: Profiles of $(\widetilde{-uv}/U)/\sqrt{y^+}$; intercept at $y^+=0$ is $\widetilde {b_{1} c_2}$. $Re_\tau =$ 550, 1000, 2000, 5200 \citep{lee2015}. }
\label{u_U_DNS}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{figures/constants_no180_2} \caption{ Reynolds number variation of $\overline{{b_1}^2}$ (black); $1000\overline{{c_2}^2}$ (blue); $\overline{{b_3}^2}$ (red); $-100\overline{b_1c_2}$ (green). See equations~\ref{constants2_log_b1} to \ref{constants2_log_b3}, Channel flow DNS \citep{lee2015}. }
\label{constants}
\end{figure}
The Reynolds number dependencies of $\overline{{b_1}^2}$, $\overline{{c_2}^2}$, $\overline{{b_3}^2}$, and $\overline{b_1 c_2}$ are shown in figure~\ref{constants}, where it could be argued that $\overline{{b_1}^2}$ and $\overline{{b_3}^2}$ for this range of Reynolds numbers the variation with Reynolds number is close to logarithmic, with
\begin{eqnarray}
\overline{{b_1}^2} & = & 0.0796 + 0.0139 \ln{Re_\tau} \label{constants2_log_b1} \\
\overline{{b_3}^2} & = & 0.012 + 0.00091 \ln{Re_\tau} \label{constants2_log_b3}
\end{eqnarray}
Note that $-\overline{b_1 c_2}$ cannot increase without limit since $-\overline{uv}/u_\tau^2$ cannot exceed one in channel flow; $\overline{c_2^2}$ must therefore also be limited. In addition, \cite{chen2021reynolds}, by considering the boundedness of the near-wall dissipation rate, argued that the logarithmic increase in $\overline{b_1^2}$ cannot hold for all Reynolds numbers.
From equations~\ref{b1} and \ref{b3}
\begin{equation}
\overline{b_1^2} = \frac{\overline{{\tau'_{wx}}^2}}{\tau_w^2} \quad \hbox{and} \quad \overline{b_3^2} = \frac{\overline{{\tau'_{wz}}^2}}{\tau_w^2}.
\label{uadef}
\end{equation}
The controlling parameter in the near-wall scaling for $u$ is therefore the mean square of the fluctuating wall stress in the $x$-direction $\tau'_{wx}$, and for $w$ it is the mean square of the fluctuating wall stress in the $z$-direction $\tau'_{wz}$. It is more difficult to give a precise meaning to $c_2$ and $b_1c_2$, but they express correlations between spatial gradients of the instantaneous wall stress fluctuations, as well as the fluctuating wall stress itself. It is therefore to be expected that $\tilde c_2$ and $\widetilde{b_1c_2}$ are more connected to the small-scale motions than either $\tilde b_1$ or $\tilde b_3$.
As pointed out by \cite{hultmark2021smits}, this connection with the fluctuating wall stress helps explain the Reynolds number dependence seen in figure~\ref{constants} since with increasing Reynolds number the large-scale (outer layer) motions contribute more and more to the fluctuating wall stress. Because $\tilde c_2$ and $\widetilde{b_1c_2}$ are more connected to the small-scale motions than either $\tilde b_1$ or $\tilde b_3$, it might be expected that they feel the effects of modulation more than superimposition by the large-scale motions \citep{Marusic_Science}.
\section{Scaling the $\widetilde{{u_1u_j}^+}$ profiles}
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{figures/u2_DNS_scaled_no180_2} \hspace{1mm}
\includegraphics[width=0.51\textwidth]{figures/v2_DNS_scaled_no180_2} \\
\includegraphics[width=0.46\textwidth]{figures/w2_DNS_scaled_no180_2} \hspace{1mm}
\includegraphics[width=0.51\textwidth]{figures/uv_DNS_scaled_no180_2}
\caption{ Profiles of scaled stresses for $Re_\tau =$ 550, 1000, 2000, 5200. Channel flow DNS \citep{lee2015}. }
\label{scaled_stresses}
\end{figure}
\cite{hultmark2021smits} showed that when each $\overline{{u^+}^2} $ profile is scaled with the value of $\overline{b_1^2}$ at its respective Reynolds number, we obtain the results shown in figure~\ref{scaled_stresses}. The collapse of the data for $y^+<20$ is remarkable, including the almost perfect agreement on the scaled peak value $\overline{{u^2_p}^+}$. Furthermore, $\overline{{b_1}^2}$ and $\overline{{u_p^2}^+}$ approach a constant ratio to each other such that for channel flow,
\begin{equation}
\overline{{u_p^2}^+} \approx 46\overline{{b_1}^2}.
\label{u2b12}
\end{equation}
In other words, the magnitude of the peak at $y^+ \approx 15$ tracks precisely with $\overline{{b_1}^2}$, a quantity that is evaluated at $y^+=0$. A similar conclusion was also proposed by \cite{chen2021reynolds}.
The results for scaling the other components of the stress tensor in a similar way are also shown in figure~\ref{scaled_stresses}. The collapse of the data is not quite as impressive as for the streamwise component, but it is nevertheless encouraging, especially when compared to the original unscaled data shown in figure~\ref{all_stresses}.
If we now return to using the normalization by the local mean velocity, originally given in figure~\ref{u_U_DNS}, and apply the scaling suggested by equations~\ref{Taylor_rms_u} to \ref{Taylor_rms_uv} we obtain the results shown in figure~\ref{u_v_U_scaled_DNS}. For $\tilde u$ and $\tilde w$, the collapse holds for $y^+<20$ and $y^+<10$, respectively, which is comparable to that seen for the scaling without $U$, as given in figure~\ref{scaled_stresses}. However, for $\tilde v$, and $\widetilde{-uv}$ the general collapse is much improved, and it extends as far out from the wall as $y^+= 100$.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{figures/u_U_b1_DNS_no180} \hspace{2mm}
\includegraphics[width=0.49\textwidth]{figures/v_U_c2_DNS_no180} \\
\includegraphics[width=0.46\textwidth]{figures/w_U_b3_DNS_no180}
\includegraphics[width=0.51\textwidth]{figures/uv_U_b1c2_DNS_no180}
\caption{ Profiles of scaled local stresses. $Re_\tau =$ 550, 1000, 2000, 5200 \citep{lee2015}. }
\label{u_v_U_scaled_DNS}
\end{figure}
\section{Conclusions}
By expanding the velocity in a Taylor Series with distance from the wall, the Reynolds number dependence of the near-wall distributions of the Reynolds stresses was traced to the magnitude of the fluctuating wall shear stress and its spatial gradients, which are increasingly affected by the superimposition and modulation of the near-wall motions due to large-scale, outer-layer motions as the Reynolds number increases \citep{Marusic_Science}.
The Taylor series expansion also suggests a separate scaling for each component of the Reynolds stress. For the streamwise component, its scaling collapses the data for $y^+ < 20$, a region that includes the near-wall peak, as shown earlier by \cite{hultmark2021smits}. For the spanwise component, the scaling holds only up to about $y^+ < 10$, and fails to capture the growth of its inner peak. For the wall-normal component and the Reynolds shear stress, the proposed scaling uses the local mean velocity rather than the friction velocity (as suggested by the analysis), and gives a very satisfactory collapse up to about $y^+=100$. Hence, for the stresses that depend on the wall-normal component of the velocity fluctuation, the near-wall behavior is governed both by the fluctuating wall shear stress and its spatial gradients, as well as the local mean velocity.
Revisiting the dimensional analysis given in equation~\ref{dim_analysis}, we can now be more precise and write for the Reynolds stresses in a two-dimensional wall-bounded flow
\begin{equation}
{\widetilde{u_i u_j}}/U = g_1(Re_\tau) g_2(y^+),
\label{dim_analysis2}
\end{equation}
where $g_1$ and $g_2$ and their ranges of applicability are different for each component. Specifically, with $U^+=f(y^+)$ absorbed into $g_2$, a consistent framework for all component is given by
\begin{eqnarray*}
\tilde u: & \quad & g_1=\tilde b_1 \hspace{8mm} (y^+<20); \\
\tilde v: & \quad & g_1=\tilde c_2 \hspace{8mm} (y^+<100); \\
\tilde w: & \quad & g_1=\tilde b_3 \hspace{8mm} (y^+<10); \\
-\widetilde{uv}: & \quad & g_1=\widetilde{b_1c_2} \hspace{5mm} (y^+<100).
\end{eqnarray*}
Our conclusions were supported only by channel flow DNS. Although we anticipate that they would extend to boundary layer flows (as Hultmark \& Smits 2021 showed for the streamwise component) and fully developed pipe flows, high Reynolds number DNS for such flows will be needed to confirm these expectations.
It may also be remarked that because the scaling is different for each component of the stress, any isotropic definition of eddy viscosity will obviously fail in the near-wall region. The use of an eddy viscosity in this region is further invalidated by the important role of the mean velocity in scaling the stresses that depend on the wall-normal component of the velocity fluctuation, rather than its gradient. In this respect, \cite{Hultmark2013a} noted that in the overlap region, where both the mean velocity $U$ and the streamwise stress $\overline{{u^2}^+}$ follow a logarithmic distribution, $\overline{{u^2}^+}$ depends on $U$ rather than its gradient.
\subsection*{Acknowledgments}
This work was supported by ONR under Grant N00014-17-1-2309 (Program Manager Peter Chang). We thank Myoungkyu Lee and Robert Moser for sharing their data (available at http://turbulence.ices.utexas.edu).
\bibliographystyle{jfm}
|
\section{Introduction}\label{sec1}
Bismuth is a semimetal with an extremely low density of highly mobile carriers of both signs~\cite{edelman1976,issi1979,fuseya2015}. The long Fermi wavelength of its carriers extends over several tens of lattice parameters. Therefore, only extended defects (such as dislocations) can decay the charge current. In crystals lacking such spatially extended disorder, carriers become ballistic~\cite{hartman}, and their mobility (and as we will see below their magnetoresistance) easily exceeds any other solid hitherto explored~\cite{ali2014}.
Recent research has shown that an extended Dirac Hamiltonian combined to the Fermi surface structure derived by a tight binding model~\cite{liu1995} can explain the complex angle dependent Landau spectrum of electrons and holes~\cite{Zhu_pnas}. The approach successfully accounts for the total evacuation of one or two electron pockets at strong magnetic field aligned along different axes~\cite{zhu2017}. The angle dependent magnetoresistance and its rich structure~\cite{collaudin2015} are also accessible to semiclassical transport theory treating mobility as a tensor~\cite{collaudin2015}.
Open questions remain, however. The origin of the loss of rotational threefold symmetry in presence of strong magnetic fields~\cite{zhu2012,Kuchler2014,collaudin2015} is yet to be understood. Such a `nematicity' was also observed on the surface of bismuth crystals \cite{Feldman2016}, as well as in other solids \cite{JO2014}. In addition to `valley nemeticity' \cite{Abainin,Parames}, other theoretical possibilities for its origin were proposed \cite{Mikitik}.
Another open question is the topology of the electron wave function~\cite{Fu2007,Teo2008}, its consequences for the metallic surface states in bismuth~\cite{Hofmann2006} and the evolution of the latter with thickness and Sb substitutions~\cite{Fuseya2018}. The topology of surface states and their status in the trivial/non-trivial dichotomy has been a subject of ongoing debate~\cite{Ohtsubo_2013,Ohtsubo_2016,Schindler2018,Hsu2019,Aguilera,Hsu13255,Nayak2019}. A recent popular theory identifies bismuth as a higher order topological solid with topologically protected hinge states ~\cite{Schindler2018}. This hypothesis has been invoked to explain the experimental observation of ballistic transport in micrometre-long bismuth nanowires~\cite{Murani2017}.
Here, we present a study of magnetoresistance in prismatic crystals of bismuth with ballistic carriers with unexpected consequences for both these issues. By choosing specific crystallographic planes as faces of the prisms, we uncover a specific contribution to electric conductivity when the magnetic field is aligned parallel to a two-dimensional boundary between the three-dimensional crystal and vacuum. The absence of this effect in antimony crystals of identical shapes points the role played by the band structure topology in tuning the edge-bulk
7
correspondence in macroscopic crystals in the high-field limit ($\omega_c \tau \gg 1$). It confirms that the interruption of cyclotron orbits at the boundary of a macroscopic three-dimensional crystal can generate a highly conducting boundary state in which bulk magnetoresistance is absent~\cite{Azbel1962,Peschanskii}.
It is known that a periodically driven Hamiltonian can provide topological protection
7
in the so-called Floquet systems \cite{Kitagawa,Lindner2011,Rudner2013}. In presence of quantizing magnetic fields, cyclotron orbits interrupted at the edge may be conceived as a periodical perturbation to the local electrons, but we are not aware of any available theory on this.
Our results identify the source of the loss of rotational symmetry and apparent 'nematicity'~\cite{zhu2012,Kuchler2014,collaudin2015} in bismuth crystals. The existence of distinct edge states surrounding bulk states would also explain why identical bismuth tilted crystals across a twin boundary can keep different chemical potentials at high magnetic field~\cite{Zhu_pnas}.
\section{Results}
\subsection{Samples, carrier mobility and orbital magnetoresistance}
The bismuth (Bi) and the antimony (Sb) crystals used in this study are listed in table \ref{Tab1}. The residual resistivities are remarkably low, considering the low carrier density of these two semimetals ($n=p=3 \times 10^{17}$~cm$^{-3}$ in bismuth and $n=p= 5.5 \times 10^{19}$~cm$^{-3}$ in antimony~\cite{liu1995}). For Bi, $\rho_0=0.18~\mu \Omega$\,cm corresponds to an average mobility of $<\mu_e + \mu_h>\ = 1.2 \times 10^8$~cm$^2$\,V$^{-1}$\,s$^{-1}$. For Sb, a residual resistivity of $\rho_0=0.07\ \mu\Omega$\,cm corresponds to an average mobility of $<\mu_e + \mu_h>\ = 1.7 \times 10^7$~cm$^2$\,V$^{-1}$\,s$^{-1}$. These mobilities exceed those of the samples used in previous studies of magnetoresistance on Bi~\cite{zhu2012,collaudin2015,zhu2017} and Sb~\cite{fauque2018}.
Note that the average mobility deduced from residual resistivity ignores the fact that in presence of anisotropic Fermi pockets of electrons and holes, the electrons and holes in different pockets and along different orientations have different mobilities. It is safe to assume that some carriers are ballistic given the size dependence of the residual resistivity~\cite{Jaoui2021}.
Carrier mobility in Bi is probably the highest known in any solid. Bi crystals with a RRR of $\simeq$ 600 were reported in old scientific literature (Supplementary note 1~\cite{SM}). However, the high-field magnetoresistance of such samples was not reported before. Bismuth samples subject to pulsed magnetic fields~\cite{Fauque2009,zhu2017} were small pieces cut from larger crystals and hosted extended scattering centers such as twin boundaries and dislocations. The presence of such disorder led to a shorter electronic mean free path and a significantly lower magnetoresistance compared to the crystals studied here.
The typical magnetoresistance in our crystals is shown in Fig.~\ref{fig:MR}. One can see that the 14 T magnetoresistance of the Bi sample with RRR = 683 is orders of magnitude higher than what was observed in WTe$_2$ at 60~T~\cite{ali2014,zhu2015}. The latter was dubbed `extremely large magnetoresistance' by many authors~\cite{Shekhar2015}. However, such a large non-saturating magnetoresistance was reported by Kapitza back in 1928~\cite{Kapitza} and is a generic feature of low-density semimetals~\cite{Pippard-book}.
\subsection{Triangular prismatic crystals: Bi \textit{vs.}\ Sb}
Bismuth and antimony crystallize in the rhombohedral A7 crystal structure shown in Fig.~\ref{fig:Bi-Sb}a. The three axes of high symmetry are known as trigonal (or C$_3$), binary (C$_2$) and bisectrix (C$_1$)~\cite{issi1979,edelman1976}. As seen in Fig.~\ref{fig:Bi-Sb}b, in the trigonal plane, there are three C$_1$ and three C$_2$ axes, which are equivalent upon 2$\pi$/3 rotation. Our experiments consisted in measuring the magnetoresistance of Bi and Sb crystals with the electric current applied along C$_3$ and the magnetic field rotating in the trigonal plane. The orientation of the magnetic field is given by $\theta$, which is defined as the angle between the field and the C$_1$ axis. As reported previously~\cite{zhu2012,collaudin2015}, despite the constant angle between current and field, orbital magnetoresistance varies with angle reflecting the in-plane anisotropy of the Fermi velocity in the three electron pockets.
Our main observation is illustrated in Fig.~\ref{fig:Bi-Sb}c. It shows the angular dependence of electrical conductivity $\sigma$ at a magnetic field of 12 T and 2 K in a pair of Bi crystals tailored identically as triangular prisms. (Note that since the Hall resistivity is negligible compared to the magnetoresistance, $\sigma\approx1/\rho$). The only difference between the two crystals was that in one case each of the three square faces of the prism was a binary (C$_2$) plane, while in the other case, it was a bisectrix (C$_1$) plane. The angle-dependent magnetoresistance is clearly different in the two crystals. In one there are conductivity peaks each time the field is along a bisectrix axis. In the other, there are minima (instead of maxima) at the same field orientations.
The same experiment was performed in a pair of Sb crystals tailored in the same way as the bismuth crystals. As one can see in the figure, no difference is visible between the two crystals.
Thus, angle-dependent magnetoresistance depends on the choice of specific crystal planes as faces of each prism in bismuth crystals, but not in antimony crystals. We reproduced this observation in two other pairs of Bi crystals and one other pair of Sb crystals.
\subsection{Identifying the source of excess conductivity in bismuth}
The shape dependence of the orbital magnetoresistance sheds light on the origin of the loss of the threefold symmetry in bismuth crystals of various geometry~\cite{zhu2012,collaudin2015}. As reported previously~\cite{zhu2012,collaudin2015}, this effect emerges only at sufficiently low temperature and high magnetic field. Let us now consider the amplitude of the required magnetic field.
Fig.~\ref{fig:T-dep} shows the change of angle-dependent conductivity with temperature in a pair of Bi triangular prisms at two different fields, namely $B=14$~T and $B=0.2$~T. The non-trivial evolution of angle-dependent magnetoconductivity can be quantitatively described by invoking the anisotropy of the effective mass and the evolution of scattering time and carrier density among pockets with temperature and magnetic field. At B = 0.2 T and T~= 40 K, angle-dependent magnetoconductivity displays a star-like shape. With cooling, the anisotropy is lowered due to a partial compensation of the mass anisotropy by the emerging anisotropy in the scattering time~\cite{hartman}. At B~= 14 T, the anisotropy of mobility is compensated by an anisotropy in the distribution of carriers among the three pockets~\cite{collaudin2015,zhu2017,Yamada2020}.
As can be seen in Fig.~\ref{fig:T-dep}, at both fields, the two samples show an identical angle-dependent conductivity at 40~K, but not at low temperature. Upon cooling, additional features emerge in the first triangular prismatic crystal, which are absent in the second one. Now, at B~= 14 T, all electrons are confined to their lowest Landau level, but not at B~= 0.2 T. This implies that the observed shape dependence of magnetoconductivity does not require proximity to the quantum limit $\hbar\omega_c\approx E_F$, but the passage to the high-field limit $\omega_c\tau > 1$.
As for the high-field limit, two relaxation times are to be distinguished. The Dingle scattering time $\tau_{D}$, extracted from quantum oscillations is almost fifty times shorter than the transport scattering rate $\tau_{tr}$ in Bi~\cite{Bhargava}. Such a large difference between $\tau_{tr}$ and $\tau_{D}$ has been observed in several other dilute metals~\cite{Liang2015,Kumar2017,Jaoui2021}. The semiclassical high-field limit ($\omega_c\tau_{tr} \approx 1$) is satisfied when the cyclotron radius becomes shorter than the mean free path. The quantum high-field limit ($\omega_c\tau_{D} \approx 1$) is satisfied when the distance between Landau levels becomes smaller than the broadening caused by temperature and disorder. At T~= 40~K, the first criterion is satisfied, but not the second and there is no shape dependence. As the sample is cooled down, the shape dependence and quantum oscillations emerge concomitantly (supplementary note 2 and supplementary Fig.1~\cite{SM}). Therefore, one can safely conclude that what matters is the $\omega_c\tau_D \approx 1$ criterion.
The angle-dependent Landau spectrum in each sample is revealed by taking the second derivative of magnetoresistance. It remains identical in the two samples in spite of the difference in the sheer amplitude of the magnetoresistance (supplementary note 3 and supplementary figure 2~\cite{SM}). This observation implies that their bulk Fermi surface is identical and can therefore be excluded as the origin of the shape dependence.
The origin of the additional features in the angle-dependent magnetoconductivity was pinned down by studying two other samples with a square cross section. Samples dubbed Bi-Cub-1 and Bi-Cub-2 (See table \ref{Tab1}) were cubic samples with identical dimensions. Both had two trigonal faces, but their four other faces were different. In Bi-Cub-2, the four other faces were two pairs of bisectrix and binary planes. On the other hand, in Bi-Cub-1, the pairs of faces other than trigonal were not aligned along a high-symmetry plane. They were rotated by a finite angle ($\approx \pi/4$) around the trigonal axis with respect to the two crystallographic planes (See insets in Fig.~\ref{fig:Shape}).
Figure~\ref{fig:Shape} compares the angle-dependent magnetoconductivity of four Bi crystals with different shapes. The temperature and the magnetic field are identical in all cases and the current is always applied along the trigonal axis and the field is rotating in the trigonal plane. Two samples are prisms with triangular cross sections and two are cubes as detailed above. The magnitude of conductance is roughly similar in the four samples, which have comparable dimensions and mobilities. The striking difference is the presence of additional peaks in magnetoconductance and their angular locations. In all the samples, magnetoconductivity peaks when the magnetic field is along the binary axis ($B\parallel C_2$). On top of these peaks, in all the samples there is another set of peaks, which appear when the magnetic field is parallel to a face of the sample.
In the two triangular prisms, these peaks appear with a periodicity of $\pi/3$ and there are six of them. The difference between the two is that in one case the surface peaks and the binary peaks are concomitant and in the other there is a $\pi/6$ shift between the two sets of peaks. In prisms with a square cross section, on the other hand, the additional peaks appear with a periodicity of $\pi/2$ and there are four of them. They occur each time the field is parallel to one of the four faces. If this face happens to be parallel to the binary axis (i.e.\ if it is a crystallographic bisectrix plane), than the peak is more pronounced, as shown in Fig.~\ref{fig:Shape}d. One can also see that when distinct, the two types of peaks have different angular width and slightly different amplitudes. The surface peaks (marked by blue arrows) are typically twice wider and twice higher than the bulk binary peaks (marked by red arrows).
This observation clearly establishes that the additional contribution to conductivity emerges when the magnetic field (kept always perpendicular to the charge current flowing along the trigonal axis) lies parallel to a two-dimensional boundary of the three-dimensional sample. Moreover, it does not matter at all for this boundary to be a specific crystal plane. The magnitude of the additional contribution remains the same when the boundary in question is the binary plane, the bisectrix plane or a low-symmetry plane.
\subsection{The boundary contribution and its relevant length scales}
Having demonstrated that the boundary conductivity emerges whenever the magnetic field is parallel to one of the surfaces of a prismatic crystal, let us now consider its evolution with magnetic field.
The relative contribution of the boundary conductance to the total conductivity can be estimated by subtracting the angle-dependent conductivity in two triangular prisms with different crystal planes as faces. This assumes that bulk magnetoconductivity is identical in the two, which is reasonable, but subject to caution given the slight difference in mobility, which implies a difference in the expected magnitude of orbital magnetoresistance.
Figure~\ref{fig:Sub}a shows the relative change in the conductivity between sample Bi-Tri-1a and sample Bi-Tri-2a. What is shown is the evolution of $r= \frac{\sigma_{\#1}-\sigma_{\#2}}{\sigma_{\#1}+\sigma_{\#2}}$ with magnetic field and the angle between field and the crystal axes. The dimensionless $r$ alternates between 0.2 and -0.2. Vertical red stripes show the excess conductivity in sample 1 and vertical blue stripes show the excess conductivity in sample 2. Remarkably, the width of the stripes or their color does not vary with increasing magnetic field. The relative amplitude of the excess boundary conductivity does not change even when the field increases by two orders of magnitude and the amplitude of conductivity decreases by almost four orders of magnitude.
Figure~\ref{fig:Sub}b shows the dependence of $r$ on angle at $B=0.2$~T and $B=10$~T. It is striking to observe how the two curves superpose on each other, in spite of a 50-fold change in magnetic field. Thus, the correction to conductivity brought by the B $\parallel$ surface configuration both in amplitude and in angular dependence does not evolve with magnetic length $\ell_B=\sqrt {\frac{\hbar}{eB}}$, which changes by a factor of 7 between the two fields.
The data presented in Fig.~\ref{fig:Sub}b contains another important feature. Within an angular window of $\pm 4$ degrees, we can fit each peak with a $\cos (q\theta)$ function with $q$ as a free fitting parameter. The fact that a simple cosine fits the data implies that the excess conductivity detected here is not singular, as observed in other contexts~\cite{Suzuki2019}. Moreover, we find that when the surface becoming parallel to the magnetic field was a bisectrix crystallographic plane $q= 20.3 \pm 3$ and when it was a binary crystallographic plane, it was $q= 8.1 \pm 1$. $q$ quantifies the sharpness of the peak, presumably caused by the anisotropy of the relevant length scales. Now, the Fermi momentum and wavelength of electrons is fourteen times longer along the bisectrix~(C$_1$) axis than along the binary axis (C$_2$)~\cite{liu1995}. This is a consequence of the huge (200-fold) anisotropy of the in-plane electron mass. Therefore, the significant difference between the angular width of the excess conductivity brings us to suspect a key role played by the Fermi wavelength of electron pockets in any plausible scenario.
\section{Discussion}
\subsection{Cyclotron orbits and the `static skin effect' }
Decades ago, Azbel and Peschanskii put forward the concept of a static skin effect at high magnetic fields in metals~\cite{Azbel1962,Peschanskii}. This idea provides the departing point of a plausible scenario to explain our observation. In the semiclassical picture, the magnetic field bends the electron trajectory. When the cyclotron radius is shorter than the mean free path of carriers, there is a large magnetoresistance. In semi-metals, this magnetoresistance does not saturate even in the high-field limit. Now, at the surface of the sample, the cyclotron orbits are interrupted and what matters is the scattering of the carriers by the edge. If their collision results in a specular reflection then the conductivity at the edge is much higher. As a result, most of the current will flow near the boundaries of the sample where orbital magnetoresistance is absent. This phenomenon was dubbed 'static skin effect' in analogy with the skin effect in metals. The latter refers to the fact that the density of an alternating current (AC) is largest near the surface of the conductor and decreases exponentially with increasing depth. Note, however, that the conductivity profile in the 'static' version of the skin effect has a completely different origin.
The large magnetoresistance of bismuth can be understood in the semiclassical picture of cyclotron orbits shrinking with increasing magnetic field (Fig.~\ref{fig:cyclotron}a). When the magnetic field is parallel to a surface, within a thickness of the order of cyclotron radius, electrons can conduct much better than in the bulk and generate a sizeable contribution to the total conductivity (Fig.~\ref{fig:cyclotron}b).
However, this semi-classical scenario fails to explain two key observations. The first is the fact that, as we saw above, the amplitude of the effect is unchanged when the magnetic field is changed by a factor of 50 (See Fig.~\ref{fig:Sub}). This is puzzling in the `static skin effect' scenario where the distribution of current depends on the ratio of the cyclotron radius and the effective sample thickness~\cite{Peschanskii}. As illustrated in Fig.~\ref{fig:cyclotron}c, increasing the magnetic field will reduce the width of the cyclotron edge and will enhance the difference in the conductivity of the bulk \textit{vs.}\ edge. One may argue that these two tendencies may approximately cancel out generating an additional conductivity which does not vary much with the amplitude of the field. However, a perfect cancellation would be mysterious given the difference in the evolution of the the magnetoresistance and the cyclotron radius. Additionally, how to explain the indifference of the angular width of the peak to the magnitude of the magnetic field? The observation implies that the current profile has remained the same despite a fifty-fold shrink in the size of the cyclotron radius.
The second observation is the absence of this phenomenon in Sb. One may be tempted to invoke a possible difference in surface rugosity. However, there is no evidence for such a difference. In addition, it is unlikely for a quantitative difference in surface quality to totally erase the effect and make the outcome qualitatively different.
The contrast between Bi and Sb can be traced to a fundamental feature of their electronic properties involving the symmetry of their band structure.
\subsection{Band inversion and topological invariants: Bi \textit{vs.}\ Sb}
The third-neighbor tight binding model conceived by Liu and Allen~\cite{liu1995} gives a successful account of the electronic band structure of bulk bismuth and antimony, as documented by numerous experiments. The model quantifies hopping energies between first, second and third neighbours with unprimed ($V$), primed ($V^\prime$) and double-primed ($V^{\prime\prime}$) parameters, respectively. The crystal lattice has two sublattices, i.e., the unit cell includes two atoms. The third neighbor of the original atom is the closest neighbor on the same sublattice and both atoms reside in the same monolayer. The three first and the second neighbor atoms belong to a different sublattice and lie in other monolayers above and below the original atom (See Fig.~2a). The 14 adjustable parameters of the model were chosen to give the best agreement with experiment. An additional parameter was spin-orbit coupling (SOC), $\lambda$, which was taken to be 0.6 eV for Sb and 1.5 eV for Bi. This model gives a reasonable account of the Fermi surface pockets of electrons and holes and the direct and the indirect gaps of Bi and Sb~\cite{liu1995}.
In 2007, Bi$_{1-x}$Sb$_x$ alloys were identified as the first bulk topological insulators~\cite{Fu2007,Fu2007b,Teo2008}, based on an important difference between Sb and Bi band symmetries. The starting point of this identification was the band inversion at the high-symmetry $L$-point in the bulk Brillouin zone, found in this tight binding model, as well as in previous works~\cite{Golin1968,gonze1990}. The symmetry of the wave function at the $L$-point can be classified into symmetric ($L_s$) and antisymmetric ($L_a)$ with respect to space inversion, where the eigenvalues of parity operator are $+1$ for $L_s$ and $-1$ for $L_a$. As one can see in panels a and b of Fig.~\ref{fig:theory}, in bismuth the conduction band at the $L$-point is symmetric and the valence band is antisymmetric, while the inverse is true in the case of antimony.
There are 8 high-symmetry (one $\Gamma$, one $T$, three $X$ and three $L$) points in the Brillouin zone (Fig.~\ref{fig:theory}c). They remain invariant under inversion and time reversal operators. At $\Gamma$-, $T$- and $X$-points, there is no difference in parity invariants between Sb and Bi. On the other hand, there is one for the $L$-points. Kane and collaborators argued that the difference in parity invariants at the $L$-points leads to a Z$_2$ topological invariant dichotomy between the two systems. As a result, at zero magnetic field, topology of the system is trivial in Bi and non-trivial in Sb~\cite{Teo2008}.
Let us briefly discuss what drives this band inversion. Both Bi and Sb crystallize in the A7 rhombohedral crystal structure, which can be assimilated to an assembly of two distorted FCC sub-lattices. As seen in Table~\ref{Tab2}, there is a significant difference between tight-binding parameters of Bi and Sb. $V_{pp\sigma}$ is the hopping energy of sigma bonding of $p$-orbitals of the first neighbours and $V'_{pp\sigma}$ is the same quantity for second neighbors. One can see that their relative difference is much larger in Sb than in Bi~\cite{liu1995}. As a consequence, the Peierls gap is larger in Sb than in Bi. The larger gap hinders the band crossing and the reversal of $L_s/L_a$ hierarchy.
In a conventional Peierls transition, the energy of the symmetric band is lower than that of the antisymmetric band~\cite{Burdett1984,Hoffmann1972}. (An intuitive picture of this hierarchy is sketched in the next subsection.) This is exactly what happens in Sb. However, the SOC can alter this energy hierarchy. In Fig.~\ref{fig:theory}d and \ref{fig:theory}e, we plot the energies of the conduction and valence bands at the $L$-point for Bi and Sb as a function of the magnitude of SOC using the Liu-Allen model. The conduction and valence bands of Sb are hardly affected and the energy hierarchy is unchanged by the SOC. On the other hand, the hierarchy is inverted by the SOC for Bi. This hierarchy alternation happens only in Bi, because the band gap (i.e., the lattice distortion) is much smaller and the SOC is larger than in Sb. Actually, the bands of Sb would be inverted if the SOC were unrealistically large ($\sim 20$ eV).
Interestingly, the $L_s$ and $L_a$ bands are distinguished by the parity inversion \cite{Teo2008,Falikov1965}.
\subsection{Parity and symmetry of the wave functions}
The gap opening at the $L$-point originates from the Peierls distortion \cite{Peierls_book2,fuseya2015}. The real-space image of the wave functions of conduction and valence bands are depicted in Fig.~\ref{fig:ref} \cite{Burdett1984,Hoffmann1972}. Peierls distortion can be understood as a dimerization, where the lattice is distorted to generate pairs as seen in Fig.~\ref{fig:ref}a. (Each dimer corresponds to a unit cell of Bi or Sb. Two atoms in the dimers are the first nearest neighbor with each other (Fig.~\ref{fig:Bi-Sb}a). There are three Peierls chains crossing at each atom in the rhombohedral structure.) The wave functions of a single dimer are given in terms of the bonding and anti-bonding orbitals. The wave function of the symmetric band ($\psi_s$) is given by the periodic array of bonding orbitals, whereas the wave function of the anti-symmetric band ($\psi_a$) is given by that of anti-bonding orbitals (Fig.~\ref{fig:ref}b). It is clear from Fig.~\ref{fig:ref}b that, in absence of dimerization, the energy of $\psi_s$ is degenerate with that of $\psi_a$. (If one removes the dashed boxes from Fig.~\ref{fig:ref}b, one finds that the two wave functions are equivalent in an infinite system.) Dimerization lowers the energy of $\psi_s$ compared to $\psi_a$, because the energy of bonding orbitals should be lower than that of anti-bonding orbitals. The magnitude of the energy gap between $\psi_s$ and $\psi_a$ is determined by the degree of dimerization, which is roughly given by the difference between the intra- and inter-dimer hopping, i.e., the difference between $V_{pp\sigma}$ and $V_{pp\sigma}'$. It is evident from Fig.~\ref{fig:ref}b that $\psi_s$ is symmetric and $\psi_a$ is anti-symmetric for the space inversion, where the inversion center of the crystal locates at the bond center in the dimer.
Now, let us consider the reflection of cyclotron orbits at the boundary with these wave functions. We only consider specular reflection normal to the surface for the sake of simplicity. (Although the incidence angle is not restricted to be normal in general, normal reflection is expected to play a major role.) By the normal reflection, the wave vector of electrons changes as $\bm{k}$ to $-\bm{k}$, which corresponds to the parity operation $P$. The parity operation results in $P \psi_s = +\psi_s$ and $P \psi_a = -\psi_a$\cite{Golin1968,gonze1990,Teo2008}. The sign of the anti-symmetric wave function is changed by the reflection at the boundary (Fig.~\ref{fig:ref}) only for $\psi_a$. Therefore, naively, one expects a qualitative difference in boundary reflection between $\psi_s$ and $\psi_a$. Further theoretical investigations are required to shed light on this subject.
\subsection{Electron topology at the cyclotron edge}
The static skin effect picture~\cite{Azbel1962,Peschanskii} is a semiclassical approach which does not take into account the phase of the electrons' wave function. The interruption of cyclotron orbits was framed in a specular-diffusive dichotomy. If the reflection is perfectly specular, momentum is conserved and there is an additional contribution to conductivity. The effect will weaken if the reflection becomes partially diffusive. However, reflected electron waves can interfere with incoming waves. The electronic Fabry interferometers employed in two-dimensional electron gases~\cite{Zhang2009} are an eloquent demonstration of this fact.
In a quantum treatment of the interruption of the cyclotron orbits by sample boundaries, it is crucial to consider the fate of the electron wave function and its phase following a mirror transformation. As we saw above, quasi-particles residing in the electron pockets have opposite symmetries in Bi and in Sb. The symmetric conduction band of bismuth, in contrast to the antisymmetric band of Sb, allows a constructive interference upon a mirror reflection at the boundary. However, it remains to be seen how this difference survives in presence of quantizing magnetic field.
Even in the simple isotropic case, large-index Landau wave functions have a non-trivial angular distribution in real space~\cite{Mikhailov2001}, which is to be affected by the zero-field anisotropy of the Fermi momentum. The anisotropic cyclotron orbits of Bi surface states have recently become accessible to experiment, thanks to scanning tunneling microscopy studies~\cite{Feldman2016}.
It is tempting to draw an analogy between the present context and Floquet systems in which topological protection is provided by a periodic perturbation~\cite{Kitagawa,Lindner2011,Rudner2013}. At zero magnetic field, the electronic surface states of a solid are distinct from bulk states by the abrupt interruption of the lattice. There is another distinction, which emerges at high magnetic field. The surface states are periodically disturbed by cyclotron orbits of the bulk. This leads to a spatio-temporal discrete translation symmetry~\cite{keyserlingk2016,Else2016}. An edge atom at a given position and time is not instantaneously equivalent to its neighbor. One of the two may be perturbed by an electron from the bulk in its cyclotron orbit, in contrast to the other. On the other hand, the two atoms remain equivalent if the temporal periodicity is taken in to account. In other words, the discrete symmetries of space and time become intertwined: $\vec (r, t) \rightarrow (\vec r+ \vec a, t+\frac{2\pi}{\omega_c})$. Future studies will tell if this analogy plays any role in explaining our observation.
One of the motivations of the present study, was the theoretical proposal that bismuth is a higher-order non-trivial topological system~\cite{Schindler2018}. This was put forward to explain the origin of ballistic transport in Bi nanowires detected by superconducting proximity studies~\cite{Murani2017}. Note that the boundary ballistic transport detected in our experiment is two-dimensional and does not appear to arise from one-dimensional channels expected in the case of topologically protected hinge states~\cite{Schindler2018}.
In conclusion, we found that in bulk crystals of bismuth, there is a robust contribution to conductance when the magnetic field is aligned parallel to a two-dimensional boundary of the sample. The absence of this effect in antimony implies that the difference in symmetry of the conduction band has a significant outcome.
A satisfactory explanation of our results is missing and remains a challenge for theory. While the `static skin effect' explains the existence of boundary conductance in a macroscopic crystal, it fails to explain its absence in antimony as well as the robust behavior of the conducting channel in presence of strong magnetic fields. We note that the semi-classical `static skin effect' has not yet been formulated in a quantum-mechanical frame incorporating the known contrast between the parity of the Bloch waves in Bi and in Sb at the L-point. We argued that the latter may affect reflection at the crystal boundaries. To the best of our knowledge, this has not yet been addressed by theory.
Our result has implications for several puzzling observations previously reported in bismuth. The loss of threefold symmetry in transport measurements~\cite{zhu2012,collaudin2015} finds a natural explanation. It may also be invoked to explain the loss of symmetry seen by thermodynamic probes~\cite{Kuchler2014}. The anomaly caused by the evacuation of a Landau level is a van Hove singularity with a cut-off due to disorder and finite size. The latter correlates with the shape of the sample. This would imply that the finite size cut-off of the van-Hove singularity may be different for different field orientations. Thermodynamic measurements on samples with different shapes and different sizes will be instructive to check this. Finally, our observation may indicate that boundaries of a bismuth crystal in presence of magnetic field provide a topological barrier. This would provide a possible solution to the puzzle of distinct chemical potentials between twinned crystals of bismuth~\cite{Zhu_pnas}.
\section{Methods} Bi and Sb crystals were commercially obtained through MaTecK GmbH, which oriented and cut them to the desired shape and dimensions. The sample surface was not polished and did not go through any other specific treatment. The experiments were performed in two different locations (Ewha University and ESPCI) and with two different set-ups. A home-made set-up was used in Ewha University and a Quantum Design PPMS was used in \mbox{ESPCI} Paris. In both cases resistivity was measured with a standard 4-wire configuration and electrical contacts were made with silver paste. In order to ensure the homogeneity of charge flow, current electrodes were applied to a thick layer of silver paste (Dupont 4929N) covering most of the end surfaces of the prism except for a small circular area in the center. Then a small island of silver paste was created in the center and voltage electrodes were placed on it. At Ewha University, magnetoresistance was measured with an AC method with a typical current of 100~$\mu$A at frequencies between 13~Hz and 17~Hz. For resistivity measurements as a function of temperature, a DC method was used with a typical current of 1~mA. In Paris, measurements were performed using currents between 1~mA and 5~mA and the standard AC method of the PPMS (50~Hz square wave excitation).
\section{Data availability}
All data generating the plots presented of this study can be obtained from the corresponding author upon request.
|
\section{Introduction}
Let $h : \overline{\mathbb{Q}} \to \mathbb{R}$ denote the (absolute, logarithmic) Weil height.
It is a positive function vanishing precisely at $\mu_\infty$, the set of all roots of unity, and $0$ by a theorem of Kronecker.
It satisfies $h(\alpha^n)= \vert n \vert h(\alpha)$ and $h(\zeta\alpha)=h(\alpha)$ for all $\alpha\in \overline{\mathbb{Q}}, \zeta\in \mu_\infty$ and all $n\in \mathbb{Z}$ as well as the inequality $h(\alpha\beta)\leq h(\alpha)+h(\beta)$ for all $\alpha,\beta\in \overline{\mathbb{Q}}$.
For further information on this height, we refer to \cite{BombieriandGubler}.
Given a field $K\subset \overline{\mathbb{Q}}$, an interesting question is whether there exists a positive constant $c$ such that $h(\alpha) \geq c$ for all $\alpha\in K^* \backslash \mu_\infty$ (here $K^*=K\backslash\{0\}$).
Such a field is said to have the Bogomolov property, or short, property $(B)$.
This notion was introduced by Bombieri and Zannier \cite{BombieriZannier}.
Note that the field $\overline{\mathbb{Q}}$ does not have the property $(B)$ since $h(2^{1/n})=(\log 2)/n \to 0$.
By Northcott's theorem, each number field has the property $(B)$.
Schinzel gave the first example of an infinite extension of $\mathbb{Q}$ with the property $(B)$ \cite{Schinzel}, namely the maximal totally real field extension $\mathbb{Q}^{tr}$ of $\mathbb{Q}$.
The $p$-adic version of this theorem was proved by Bombieri and Zannier \cite{BombieriZannier}.
They proved that the maximal totally $p$-adic extension $\mathbb{Q}^{tp}$ of $\mathbb{Q}$ has the property $(B)$.
In recent years, the study of this property mushroomed, see for example \cite{AmorosoDvornicich, AmorosoZannier10, Habegger, AmorosoDavidZannier, Grizzard, FiliMilner, Galateau, Frey, Plessismino, PazukiPengo}.
The study of this property is not limited to this situation and we can easily define it for abelian varieties.
Let $A$ be an abelian variety defined over a number field $K$, and let $\mathcal{L}$ be a symmetric ample line bundle on $A/K$.
Let $\hat{h}_A : A(\overline{K}) \to \mathbb{R}$ denote the Néron-Tate height attached to $\mathcal{L}$.
It is a positive function vanishing precisely at $A_\mathrm{tors}$, the group of torsion points of $A$.
Again, given a field $L\subset \overline{K}$, the group $A(L)$ is said to have the property $(B)$ (with respect to $\mathcal{L}$) if there exists a positive constant $c$ such that $\hat{h}_A(P) \geq c$ for all $P\in A(L)\backslash A_\mathrm{tors}$.
It is well known that $A(\overline{K})$ does not have the property $(B)$.
Northcott's theorem cited above also states that $A(L)$ has the property $(B)$ if $L$ is a number field.
Zhang showed the abelian analogue of Schinzel's theorem, that is $A(\mathbb{Q}^{tr})$ has the property $(B)$ \cite{Zhang}.
Later, Baker and Petsche proved that $A(\mathbb{Q}^{tp})$ has the property $(B)$ when $p>2$ and $A/\mathbb{Q}$ is an elliptic curve with semistable reduction at $p$ \cite[Theorem 6.6]{BakerPetsche}.
For more examples concerning the property $(B)$ in the case of an abelian variety, see \cite{BakerSilverman, PazukiPengo} (which handle the case of any abelian variety) and \cite{Baker, Silverman, Habegger, Pottmeyer, Plessisiso} (which treat the special case of an elliptic curve).
A very special case of a recent conjecture due to the author predicts the following \cite[Conjecture 1.4]{Plessisiso}.
\begin{conj} \label{conj0}
Let $A$ be an abelian variety defined over a number field $K$, let $\mathcal{L}$ be a symmetric ample line bundle on $A/K$, and let $L/K$ be a finite extension.
Then $L(A_\mathrm{tors})$ and $A(L(A_\mathrm{tors}))$ have the property $(B)$.
\end{conj}
\begin{rqu}
\rm{The abelian part of this conjecture is due to David.}
\end{rqu}
This conjecture was proved to be true when $A$ has complex multiplication (CM).
More precisely, the toric part is due to Amoroso, David and Zannier (see Theorem \ref{thm5} below for a more general statement) and the abelian part was proved by Baker and Silverman \cite[Section 9]{BakerSilverman}, see also \cite[Théorème 1.8]{Carrizosa}.
The case where $A$ has no CM is much harder.
Up to my knowledge, Habegger was the first one to provide a result going in the direction of Conjecture \ref{conj0}.
\begin{thm} [Habegger, \cite{Habegger}] \label{thm-1}
Let $E/\mathbb{Q}$ be an elliptic curve.
Then $\mathbb{Q}(E_\mathrm{tors})$ and $E(\mathbb{Q}(E_\mathrm{tors}))$ have the property $(B)$.
\end{thm}
Recently, Frey pointed out a quite remarkable fact: Conjecture \ref{conj0} can be true for some infinite extensions $L/\mathbb{Q}$.
\begin{thm} [Frey, \cite{Freypubli}] \label{thm0}
Let $E/\mathbb{Q}$ be an elliptic curve, and let $L/\mathbb{Q}$ be a Galois extension such that the exponent of its Galois group is finite.
Then $L(E_\mathrm{tors})$ has the property $(B)$.
\end{thm}
For an elliptic curve $E$ and an integer $N\in\mathbb{N}=\{1,2,\dots\}$, write $E[N]$ for the group of $N$-torsion points of $E$ and define $j_E$ to be the $j$-invariant of $E$.
Let $K$ be a number field, and let $v$ be a finite place of $K$.
We define $K^{tv}$ as the maximal totally $v$-adic field, that is the set of $\alpha\in \overline{K}$ such that $v$ is totally split in $K(\alpha)$.
It is Galois over $K$.
Then for any extension $L/K$ of number fields and any finite place $w$ of $L$ above $v$, we denote by $e_w(L\vert K)$, resp. $f_w(L\vert K)$, the ramification index, resp. inertia degree, of the extension $w\vert v$.
In \cite{Habegger, Freypubli}, the different authors mentioned that the CM case of their results was already known.
They thus reduced their proofs to the case that $E$ has no CM.
Then they worked with a rational prime $p\geq 5$ satisfying:
\begin{align*}
(P_1) : & \; E \; \text{has supersingular reduction at} \; p \; \text{and} \; j_E \not\equiv 0, 1728 \; (p); \\
(P_2) : & \; \text{the natural representation} \; \mathrm{Gal}(\mathbb{Q}(E[p])/\mathbb{Q}) \to \mathrm{Aut} E[p] \; \text{is surjective}.
\end{align*}
By Elkies' theorem \cite{Elkies2}, $E$ has infinitely many rational primes of supersingular reduction and by Serre's open image theorem \cite{Serrejp}, $(P_2)$ is satisfied for all rational primes large enough.
This ensures us the existence of such a rational prime $p$.
However, the combination of $(P_1)$ and $(P_2)$ shows that the ideas developed in \cite{Habegger, Freypubli} cannot be adapted to deal with the case where the pair $E/K$ has only finitely many places of supersingular reduction.
The main purpose of this article is to propose an alternative proof of Theorem \ref{thm-1} that does not involve Serre's open image theorem.
This will allow us to obtain the generalization below, a proof of which will be given in Section \ref{section 4} and Section \ref{section 5}.
\begin{thm} \label{thm2}
Let $E$ be an elliptic curve defined over a number field $K$.
Assume that there exists a finite place $v$ of $K$ such that:
\begin{align*}
(P_0) : & \; f_v(K \vert \mathbb{Q}) \leq 2; \\
(P_1) : & \; p\geq 5, \; \text{where} \; p\mathbb{Z}=v \cap \mathbb{Z}, \; E \; \text{has supersingular reduction at} \; v, \; e_v(K\vert \mathbb{Q})=1 \\
& \text{and} \; j_E \not\equiv 0, 1728 \; (v).
\end{align*}
Then $K^{tv}(E_\mathrm{tors})$ and $E(K^{tv}(E_\mathrm{tors}))$ have the property $(B)$.
\end{thm}
\begin{rqu}
\rm{We immediately infer that $\mathbb{Q}(\sqrt{D}, E_\mathrm{tors})$ and $E(\mathbb{Q}(\sqrt{D}, E_\mathrm{tors}))$ have the property $(B)$ for each pair $E/K= E/\mathbb{Q}(\sqrt{D})$ with $D\in \mathbb{Z}_{\geq 0}$.
This arises from the fact that $(P_0)$ is satisfied for all places $v$ and that $(P_1)$ is satisfied for infinitely many places $v$ by Elkies' thesis \cite{Elkies}.
Nonetheless, Theorem \ref{thm2} does not permit us to treat the case $D<0$ in full generality.
For example, we do not know so far if the elliptic curve
\begin{equation*}
E : iy^2 = x^3+(i-2)x^2+x
\end{equation*}
defined over $\mathbb{Q}(i)$ has at least one place of supersingular reduction (it is however conjectured that there exist infinitely many) \cite[Section 5.2]{Elkies}. }
\end{rqu}
The exponent of the Galois group $\mathrm{Gal}(K^{tv}/K)$ is infinite.
Theorem \ref{thm0} and Theorem \ref{thm2} therefore show that Conjecture \ref{conj0} can be enlarged in at least $2$ different ways.
So it is of interest to suggest a new conjecture including these two theorems.
Throughout this text $K_v$ denotes the completion of a number field $K$ with respect to a place $v$ of $K$.
\begin{defn} \label{defn1}
\rm{Let $K$ be a number field, let $v$ be a finite place of $K$, and let $L/K$ be an algebraic extension.
We say that $L$ has bounded local degree at $v$ if $d_v(L)=\sup_w [L_w : K_v]$ is finite, where $w$ ranges over all extensions of $v$ to $L$. }
\end{defn}
Clearly, $d_v(K^{tv})=1$.
Moreover, by a theorem of Checcoli \cite[Theorem 2.2.2]{Checcoli}, if $L/K$ is Galois, then the exponent of $\mathrm{Gal}(L/K)$ is finite if and only if $\sup_v\{d_v(L)\}$ is finite.
This leads us to propose the following conjecture.
\begin{conj} \label{conj1}
Let $A$ be an abelian variety defined over a number field $K$.
Let $\mathcal{L}$ be a symmetric ample line bundle on $A/K$, and let $L/K$ be an algebraic extension.
If $d_v(L)$ is finite for at least one finite place $v$ of $K$, then $L(A_\mathrm{tors})$ and $A(L(A_\mathrm{tors}))$ have the property $(B)$.
\end{conj}
Although this conjecture has been motivated by Theorem \ref{thm0} and Theorem \ref{thm2}, the best argument in favor of this statement is probably the result below.
\begin{thm} [Amoroso-David-Zannier] \label{thm5}
Let $A$ be a CM abelian variety defined over a number field $K$.
Let $\mathcal{L}$ be a symmetric ample line bundle on $A/K$, and let $L/K$ be a Galois extension.
If $d_v(L)$ is finite for at least one finite place $v$ of $K$, then $L(A_\mathrm{tors})$ has the property $(B)$.
\end{thm}
\begin{proof}
As $A$ is CM, there exits a finite Galois extension $M/K$ such that $M(A_\mathrm{tors})/M$ is abelian.
Choose $\sigma\in \mathrm{Gal}(LM(A_\mathrm{tors})/LM)$ and $\tau\in \mathrm{Gal}(LM(A_\mathrm{tors})/M)$.
If $\alpha\in M(A_\mathrm{tors})$, then $\sigma\tau\alpha = \tau\sigma \alpha$ since $M(A_\mathrm{tors})/M$ is abelian.
If $\alpha\in LM$, then $\sigma(\tau \alpha) = \tau(\sigma\alpha) = \tau\alpha$ since $LM/M$ is Galois (because $L/K$ is by assumption) and $\sigma$ fixes the elements of $LM$.
From all this, we get $\sigma\tau=\tau\sigma$, i.e., $\mathrm{Gal}(LM(A_\mathrm{tors})/LM)$ is contained in the center of $\mathrm{Gal}(LM(A_\mathrm{tors})/M)$.
As $d_v(LM)$ is bounded from above by $d_v(L)d_v(M)< +\infty$, the theorem now arises from \cite[Theorem 1.2]{AmorosoDavidZannier}.
\end{proof}
Actually, Frey gave a more precise conclusion than that of Theorem \ref{thm0}.
\begin{thm} [Frey, \cite{Freypubli}, Theorem 7.1] \label{thm1}
Let $E/\mathbb{Q}$ be a non-CM elliptic curve, and let $L/\mathbb{Q}$ be a Galois extension such that the exponent $\exp(L)$ of its Galois group is finite, i.e., $d=\sup_q \{d_q(L)\}$ is finite, where $q$ runs over all rational primes.
Then there exists a rational prime $p\geq 5$ satisfying:
\begin{itemize}
\item [(a)] $E$ has supersingular reduction at $p$ and $j_E \not\equiv 0, 1728 \; (p)$;
\item [(b)] the natural representation $\mathrm{Gal}(\mathbb{Q}(E[p])/\mathbb{Q}) \to \mathrm{Aut} E[p]$ is surjective;
\item [(c)] $p \geq \max\{2d+2; \exp(L)\}$,
\end{itemize}
and for such a $p$, we have, for all $\alpha\in L(E_\mathrm{tors})^* \backslash \mu_\infty$,
\begin{equation*}
h(\alpha) \geq \frac{(\log p)^4}{10^{15}2^p p^{4(p^4+4)}} \geq \frac{(\log p)^4}{p^{5p^4}}.
\end{equation*}
\end{thm}
Our argument consisting in proving Theorem \ref{thm-1} without using Serre's open image theorem also allows us to remove several technical assumptions in Theorem \ref{thm1}.
We will prove the following in the last section of this paper.
\begin{thm} \label{thm3}
Let $E/\mathbb{Q}$ be a non-CM elliptic curve, and let $L/\mathbb{Q}$ be a Galois extension.
Assume that there exists a rational prime $p\geq 5$ such that:
\begin{itemize}
\item [(a)] $d_p(L)$ is finite and $p>2 d_p(L)$;
\item [(b)] $E$ has supersingular reduction at $p$ and $j_E \not\equiv 0,1728 \; (p)$.
\end{itemize}
Then, for all $\alpha\in L(E_\mathrm{tors})^* \backslash \mu_\infty$, we have
\begin{equation*}
h(\alpha) \geq \frac{(\log p)^4}{10^{15}2^p p^{4(p^3+3)}}\geq \frac{(\log p)^4}{p^{5p^3}}.
\end{equation*}
\end{thm}
We will explain in Remark \ref{rqu1} why we have the improvement $p^3+3$, against $p^4+4$ for Frey, in the denominator of our lower bound.
\begin{rqu}
\rm{Theorem \ref{thm3} also holds for CM elliptic curves defined over $\mathbb{Q}$.
But this situation is a special case of Theorem \ref{thm5} and our lower bound seems weaker than the one that we can compute in \cite[Theorem 1.2]{AmorosoDavidZannier}.
This explains why we decided to rule out this case in the statement of our theorem. }
\end{rqu}
However, in the non-CM case, our lower bound is stronger than that in Theorem \ref{thm1}.
Let us see this through a concrete example.
Consider the elliptic curve
\begin{equation*}
E : y^2+y=x^3-x^2-10x-20
\end{equation*}
defined over $\mathbb{Q}$.
According to \cite[elliptic curve 11.a2]{LMFDB}, $E$ has conductor $N=11$ and $j$-invariant $j_E=-2^{12} \; 11^{-5} \; 31^3$.
By the same reference, $(P_2)$ is satisfied for all $p \geq 7$.
Next $j_E\not\equiv 0 \; (p)$ for all $p\notin \{2;11;31\}$ and $j_E\not\equiv 1728 \; (p)$, that is
\begin{equation*}
2^6 \; 41^2 \; 61^2 = 11^5 \; 1728+2^{12} \; 31^3 \not\equiv 0 \; (p),
\end{equation*}
for all $p \notin \{2; 41; 61\}$.
Finally, $E$ has supersingular reduction at $p=19$ \cite[Chapter 5, Example 4.6]{Silvermanarithmetic}.
From all this, Theorem \ref{thm3} claims that
\begin{equation*}
\forall \alpha\in \mathbb{Q}^{t19}(E_\mathrm{tors})^* \backslash \mu_\infty, \; h(\alpha) \geq \frac{(\log 19)^4}{19^{5.19^3}}.
\end{equation*}
Note that we cannot deduce this lower bound from Theorem \ref{thm1} because $\mathrm{Gal}(\mathbb{Q}^{t19}/\mathbb{Q})$ has infinite exponent.
Consider a number field $M\subset \mathbb{Q}^{t19}$ of degree $d\geq 9$.
Even under this restriction, it is not always possible to get the lower bound above from Theorem \ref{thm1} since $p=19$ is not a suitable choice here, item $(c)$ being not satisfied.
Let $n\in \mathbb{N}$ be an integer and write $\mathcal{V}(n)=\sum_{p \; \text{prime} \; \leq n} \log p$.
We have the inequality $\mathcal{V}(n) < 1.01624n$ \cite{RosserSchoenfeld}.
Applying \cite[Theorem 4.13]{Frey} to $M=2d+2$ proves the existence of a rational prime $p$ between $n=\max\{2d+2, 7654\}$ and
\begin{equation*}
e^{1.3 \; 10^8 \; e^{2\mathcal{V}(n)+\frac{11}{15}e^{\mathcal{V}(n)}}} < e^{e^{e^{\mathcal{V}(n)}}} < e^{e^{e^{1,01624 n}}}
\end{equation*}
such that $E$ has supersingular reduction at $p$.
For such a choice of $p$, items $(a)-(c)$ of Theorem \ref{thm1} are all three satisfied, which leads to the lower bound
\begin{equation*}
\forall \alpha\in M(E_\mathrm{tors})^* \backslash \mu_\infty, \; h(\alpha) \geq \left(e^{e^{e^{e^{1.02n}}}}\right)^{-1}.
\end{equation*}
We can compare the two lower bounds above and check that ours is much better.
\subsection*{Acknowledgement}
I thank P. Habegger for replying to my questions as well as L. Pottmeyer for pointing out that my arguments aiming to prove Theorem \ref{thm2} without using Serre's open image theorem should also work to generalize Theorem \ref{thm1} without assuming that $\mathrm{Gal}(L/\mathbb{Q})$ has finite exponent.
This work was funded by Morningside Center of Mathematics, CAS.
\section{Equidistribution} \label{section 2}
Let $K$ be a number field, and let $v$ be a finite place of $K$.
Write $\vert . \vert_v$ for the normalized $v$-adic absolute value on $K$, that is $\vert p \vert_v = p^{-1}$, where $p$ is the rational prime below $v$.
The completion of $K$ with respect to $\vert. \vert_v$ is denoted with $K_v$.
We use the same symbol $\vert . \vert_v$ for the absolute value on $K_v$, which we then extend to $\mathbb{C}_v$, the completion of a fixed algebraic closure $\overline{K_v}$ of $K_v$.
The text below is largely inspired by \cite[Introduction]{BakerRumelyBookBerkovich}.
As a set, the Berkovich affine line $\mathbb{A}^1_\mathrm{Berk}(\mathbb{C}_v)$ over $\mathbb{C}_v$ consists of all multiplicative seminorms on $\mathbb{C}_v[X]$ which extend $\vert .\vert_v$.
For $a\in\mathbb{C}_v$ and $r\in \mathbb{R}_{\geq 0}$, write $D(a,r)$ for the closed disc centered at $a$ with radius $r$.
By convention, $D(a,0)=\{a\}$.
Using an elementary consequence of Gauss' lemma, we easily prove that the map \[ \begin{array}{ccccc}
\zeta_{a,r} & : & \mathbb{C}_v[X] & \to & \mathbb{R} \\
& & P & \mapsto & \sup_{z\in D(a,r)} \vert P(z)\vert_v \\
\end{array} \] is a seminorm, that is $\zeta_{a,r}\in \mathbb{A}^1_\mathrm{Berk}(\mathbb{C}_v)$.
More generally, it is easy to see that if $(D(a_n,r_n))_n$ is any decreasing nested sequence of closed discs, then the map \[ \begin{array}{ccccc}
& & \mathbb{C}_v[X] & \to & \mathbb{R} \\
& & P & \mapsto & \lim\limits_{n\to +\infty} \zeta_{a_n,r_n}(P) \\
\end{array}\] lies in $\mathbb{A}^1_\mathrm{Berk}(\mathbb{C}_v)$.
As $\mathbb{C}_v$ is not spherically complete, the intersection of all $D(a_n,r_n)$ can be empty.
So the map above produces new elements belonging to $\mathbb{A}^1_\mathrm{Berk}(\mathbb{C}_v)$.
Berkovich's classification theorem \cite{Berkovich} tells us that each point of $\mathbb{A}^1_\mathrm{Berk}(\mathbb{C}_v)$ corresponds to a decreasing nested sequence $(D(a_n,r_n))_n$ of closed discs.
The Berkovich affine line over $\mathbb{C}_v$ is a locally compact, Hausdorff, and uniquely path-connected (i.e., two distinct points can be joined by a unique arc) topological space which contains $\{\zeta_{a,0}, a\in \mathbb{C}_v\}\simeq \mathbb{C}_v$ as a dense subspace.
The topology on $\mathbb{A}^1_\mathrm{Berk}(\mathbb{C}_v)$ is the weakest one for which $\zeta \mapsto \zeta(P)$ is continuous for all $P\in \mathbb{C}_v[X]$.
The Berkovich projective line $\mathbb{P}^1_\mathrm{Berk}(\mathbb{C}_v)$ over $\mathbb{C}_v$ is obtained by adjoining to $\mathbb{A}^1_\mathrm{Berk}(\mathbb{C}_v)$ in a suitable manner a point at infinity; the resulting space is a compact, Hausdorff, uniquely path-connected topological space which contains $\mathbb{P}^1(\mathbb{C}_v)$ (with its natural topology) as a dense subspace.
The topology on $\mathbb{P}^1_\mathrm{Berk}(\mathbb{C}_v)$ is that of the one-point compactification.
Thus the map $\zeta\mapsto \min\{a, \zeta(P)\}$ is continuous on $\mathbb{P}^1_\mathrm{Berk}(\mathbb{C}_v)$ for all $a\in\mathbb{C}_v$ and all $P\in\mathbb{C}[X]$.
For $\phi\in \mathbb{C}_v[X]$, we can define a canonical measure on $\mathbb{P}^1_{\mathrm{Berk}}(\mathbb{C}_v)$ attached to $\phi$.
When $\phi(X)=X^2$, it is the Dirac measure supported at $\zeta_{0,1}$ \cite[Proposition 10.5]{BakerRumelyBookBerkovich}.
Applying \cite[Theorem 2.3]{BakerRumely} to $\phi(X)=X^2$ leads to the following.
\begin{thm} [Baker-Rumely]\label{thm4}
Fix a field embedding $\overline{K} \to \overline{K_v}$.
Let $(z_n)_n$ be a sequence of points in $\overline{K}^* \backslash \mu_\infty$ with $h(z_n) \to 0$.
For all $n$, choose a finite extension $K_n/K(z_n)$.
Then, for every continuous function $f : \mathbb{P}_{Berk}^1(\mathbb{C}_v) \to \mathbb{R}$, we have
\begin{equation} \label{eq1}
\frac{1}{[K_n : K]}\sum_\sigma f(\zeta_{\sigma z_n,0}) \to f(\zeta_{0,1}),
\end{equation}
where $\sigma$ ranges over all field embeddings $K_n\to \overline{K_v}$ which are the identity over $K$.
\end{thm}
\begin{rqu}
\rm{$a)$ In \cite[Theorem 2.3]{BakerRumely}, $(z_n)_n$ is a sequence of pairwise distinct elements of $\overline{K}$ such that $h(z_n)\to 0$.
But a quick look on \cite[§ 4.4]{BakerRumely} (and especially Theorem 4.6) shows that this condition can be relaxed in "$[K(z_n) : K] \to +\infty$ and $h(z_n)\to 0$", which is satisfied when $h(z_n)\to 0$ and $z_n\in \overline{K}^*\backslash \mu_\infty$ for all $n$ (indeed, if a sequence $(x_n)_n$ satisfies $h(x_n)\to 0$ and $[K(x_n) : K]\to d$ for some $d\in \mathbb{N}$, then $x_n\in\mu_\infty$ for all $n$ large enough by Northcott's theorem).
$b)$ The left-hand side in \eqref{eq1} does not depend on the choice of $K_n$.
This comes from the fact that any field embedding $ K(z_n) \to \overline{K_v}$ which is the identity on $K$ extends in $[K_n : K(z_n)]$ different ways to a field embedding $K_n \to \overline{K_v}$. }
\end{rqu}
\section{Reduction step and key result}
Let $v$ and $w$ be two places.
The notation $w\vert v$ is shorthand for "$w$ is above $v$".
The proof of the lemma below is an easy adaptation of that of \cite[Lemma 5.8]{Pottmeyer}.
\begin{lmm} \label{lmm1}
Let $K$ be a number field, and let $v$ be a finite place of $K$.
Take $\alpha\in \overline{K}$.
Then $K^{tv}(\alpha)= \bigcap_{w\mid v} K(\alpha)^{tw}$, where $w\vert v$ ranges over all places of $K(\alpha)$.
\end{lmm}
\begin{proof}
Let $w\vert v$ be a place of $K(\alpha)$.
As $v$ is totally split in $K^{tv}$, it follows that $w$ is totally split in $K^{tv}(\alpha)$.
Consequently, $K^{tv}(\alpha) \subset K(\alpha)^{tw}$ and the left-hand side is therefore contained in the right-hand side.
Let $\beta\in \bigcap_{w\mid v} K(\alpha)^{tw}$, and let $\nu \vert v$ be a place of $K(\alpha, \beta)$.
Denote by $L^\nu$ the maximal intermediate subfield of $K(\alpha, \beta)/K$ such that $(L^\nu)_{\nu'}=K_v$, where $\nu'=\nu\cap L^\nu$, that is $e_{\nu'}(L^\nu\vert K)= f_{\nu'}(L^{\nu}\vert K)=1$.
Put $\nu''=\nu \cap L^\nu(\alpha)$.
Our choice of $\beta$ gives $e_{\nu}(K(\alpha, \beta)\vert K(\alpha))=f_{\nu}(K(\alpha,\beta)\vert K(\alpha))=1$.
Moreover, the definition of $L^\nu$ proves that $\nu$ is the unique place of $K(\alpha, \beta)$ above $\nu'$.
From all this, and from the equality $[K(\alpha,\beta):L^\nu(\alpha)]= \sum_{w\vert \nu''} [K(\alpha,\beta)_w : L^\nu(\alpha)_{\nu''}]$, we conclude $K(\alpha,\beta)=L^\nu(\alpha)$ for all places $\nu\vert v$ of $K(\alpha,\beta)$.
This enables us to choose a number field $L\subset \bigcap_{\nu\vert v} L^\nu$ such that $\beta\in L(\alpha)$.
By definition de $L^\nu$ and construction of $L$, we easily check that $v$ is totally split in $L$; whence $L\subset K^{tv}$, and so $\beta\in K^{tv}(\alpha)$.
\end{proof}
We can now reduce the proof of Theorem \ref{thm2} to the following situation.
\begin{lmm} [Reduction step] \label{lmm1.5}
If Theorem \ref{thm2} is true when we replace $(P_0)$ with the stronger condition $f_v(K\vert\mathbb{Q})=2$, then it is true in full generality.
\end{lmm}
\begin{proof}
Obvious if $f_v(K\vert\mathbb{Q})=2$.
So assume that $f_v(K\vert \mathbb{Q})=1$.
Let $L/K$ be a quadratic extension such that $v$ is inert in $L$.
Then $e_v(L \vert \mathbb{Q})=1$ and $f_v(L \vert \mathbb{Q})=2$.
Our assumption ensures us that $L^{tv}(E_\mathrm{tors})$ and $E(L^{tv}(E_\mathrm{tors}))$ have the property $(B)$.
The lemma follows since $L^{tv}=K^{tv}L$ by Lemma \ref{lmm1} applied to $K(\alpha)=L$.
\end{proof}
Let $L/K$ be a Galois extension of number fields.
Let $v$ be a place of $K$, and let $w\vert v$ be a place of $L$.
Denote by $D(w \vert v)\subset \mathrm{Gal}(L/K)$ the decomposition group of $w \vert v$.
Then if $H$ is a subgroup of $\mathrm{Gal}(L/K)$, we define $L^H$ as the subfield of $L$ fixed by $H$.
Finally, for a subset $X$ of a group $G$, write $\langle X\rangle$ for the subgroup of $G$ generated by $X$.
The next lemma is the key argument to bypass Serre's open image theorem in the proof of our theorems, see the proof of Lemma \ref{lmm18} for more details.
\begin{lmm} [Key result] \label{lmm2}
Let $K$ be a number field, and let $v$ be a finite place of $K$.
Take a finite Galois extension $M/K$, an extension $L/K^{tv}$ and a number field $F \subset L$ containing $K$.
Assume that the restriction map $\mathrm{Gal}(LM/L) \to \mathrm{Gal}(FM/F)$ is an isomorphism.
Then, for any intermediate subfield $N$ of $M/K$, we have $\mathrm{Gal}(FM/FN)= \langle \bigcup_{w\mid v} D(w \vert w\cap FN)\rangle$, where $w\vert v$ runs over all places of $FM$.
\end{lmm}
\begin{proof}
For brevity, put $X=\bigcup_{w\mid v} D(w \vert w\cap FN)$.
Let $v'\vert v$ be a place of $FN$.
Algebraic number theory tells us that for all places $w\vert v'$ of $FM$, we have
\begin{equation*}
e_w((FM)^{D(w\vert v')} \vert FN)=f_w((FM)^{D(w\vert v')}\vert FN)=1.
\end{equation*}
The place $v'$ is therefore totally split in $\bigcap_{w\vert v'} (FM)^{D(w\vert v')}$, which gives us the inclusion $\bigcap_{w\vert v'} (FM)^{D(w\vert v')} \subset (FN)^{tv'}$.
Hence
\begin{equation*}
(FM)^{\langle X\rangle}= \bigcap_{w\vert v}(FM)^{D(w \vert w\cap FN)}= \bigcap_{v'\vert v}\bigcap_{w\vert v'}(FM)^{D(w \vert v')} \subset \bigcap_{v'\vert v} (FN)^{tv'}.
\end{equation*}
Lemma \ref{lmm1} applied to $K(\alpha)=FN$ proves that $\bigcap_{v'\vert v}(FN)^{tv'}=K^{tv}FN$; whence $(FM)^{\langle X\rangle} \subset K^{tv}FN$.
On the one hand, $L$ contains the field $K^{tv}F$, which implies $(FM)^{\langle X\rangle} \subset LN$.
On the other hand, $FN \subset (FM)^{\langle X\rangle}$ by definition of $X$, and so $LN \subset L(FM)^{\langle X\rangle}$.
In conclusion, $L(FM)^{\langle X\rangle}= LN$.
As the restriction map $\mathrm{Gal}(LM/L) \to \mathrm{Gal}(FM/F)$ is an isomorphism, we finally get $(FM)^{\langle X\rangle} = FN$.
\end{proof}
\section{Preliminaries for Theorem \ref{thm2}} \label{section 3}
Fix until the end of Section \ref{section 5} a number field $K$, an elliptic curve $E$ defined over $K$ with neutral element $O$ and a finite place $v$ of $K$ satisfying $f_v(K\vert \mathbb{Q})=2$ and the condition $(P_1)$ of Theorem \ref{thm2}.
Note that this implies $K_v = \mathbb{Q}_{p^2}$, where $\mathbb{Q}_{p^2}$ denotes the unramified extension of degree $2$ of $\mathbb{Q}_p$ inside $\overline{\mathbb{Q}_p}$.
We also fix once and for all a field embedding $\overline{K} \to \overline{K_v}$, which allows us to see from now $\overline{K}$ as a subfield of $\overline{K_v}$.
Thus $K\subset K^{tv}\subset K_v=\mathbb{Q}_{p^2}$ and our elliptic curve is therefore defined over $\mathbb{Q}_{p^2}$.
Finally, thanks to Lemma \ref{lmm1.5}, it suffices to prove Theorem \ref{thm2} in this situation to deduce it in full generality.
As $E$ is fixed, we ease notation by putting $L(N)=L(E[N])$ for any field $L$ and any integer $N$.
The lemmas of this section are mainly based on results of \cite{Habegger}.
\begin{lmm} \label{lmm2.5}
Let $n,M\in \mathbb{N}$ be two integers with $M$ coprime to $p$.
\begin{enumerate} [(i)]
\item The image of the representation $\mathrm{Gal}(\mathbb{Q}_{p^2}(p^n)/\mathbb{Q}_{p^2})\to \mathrm{Aut} E[p^n]$ contains multiplication by $\pm M$ and acts transitively on torsion points of order $p^n$.
\item The restriction map $\mathrm{Gal}(\mathbb{Q}_{p^2}(p^nM)/\mathbb{Q}_{p^2}(M))\to \mathrm{Gal}(\mathbb{Q}_{p^2}(p^n)/\mathbb{Q}_{p^2})$ is an isomorphism.
\end{enumerate}
\end{lmm}
\begin{proof}
For $(i)$, see \cite[Lemma 3.3 $(iii)$]{Habegger}.
For $(ii)$, see \cite[Lemma 3.4 $(iii)$]{Habegger}.
\end{proof}
For $m\in \mathbb{N}$, write $\mu_m$ for the group of $m$-roots of unity.
Fix for this section an integer $N\in \mathbb{N}$.
The proof of the next lemma is largely inspired by that of \cite[Lemma 6.4]{Habegger}.
\begin{lmm} \label{lmm3}
Put $N=p^nM$ with $n\geq 0$ and $M\in\mathbb{N}$ coprime to $p$.
Let $F/K$ be a finite Galois extension with $F\subset \mathbb{Q}_{p^2}$.
Then there exists $\sigma\in \mathrm{Gal}(F(N)/F)$ lying in the center of $\mathrm{Gal}(F(N)/K)$ and acting on $E[p^n]$ as multiplication by $2$.
Furthermore, $\sigma\zeta = \zeta^4$ for all $\zeta\in \mu_{p^n}$.
\end{lmm}
\begin{proof}
Take $\sigma$ to be the identity element if $n=0$.
Now assume that $n\geq 1$.
Since $p$ is odd, Lemma \ref{lmm2.5} $(i)$ implies that there is $\sigma'\in \mathrm{Gal}(\mathbb{Q}_{p^2}(p^n)/\mathbb{Q}_{p^2})$ acting on $E[p^n]$ as multiplication by $2$.
Moreover, the properties of the Weil pairing give $\sigma' \zeta = \zeta^4$ for all $\zeta\in \mu_{p^n}$, the exponent $4=2^2$ arising from the bilinearity of the Weil pairing.
Next $\sigma'$ lifts uniquely to $\sigma\in \mathrm{Gal}(\mathbb{Q}_{p^2}(N)/\mathbb{Q}_{p^2}(M))$ according to Lemma \ref{lmm2.5} $(ii)$.
Its restriction to $F(N)$ belongs to $\mathrm{Gal}(F(N)/\mathbb{Q}_{p^2}(M)\cap F(N))$; call it $\sigma$ again.
Let $\tau\in \mathrm{Gal}(F(N)/K)$.
The element $\sigma$ acts trivially on $E[M]$ and the isomorphism $E[p^n] \times E[M] \simeq E[N]$ is compatible with the action of the Galois group $\mathrm{Gal}(F(N)/F)$.
From all the foregoing, we get $\sigma(\tau T) = \tau(\sigma T)$ for all $T\in E[N]$.
Finally, as $F/K$ is Galois and $\sigma$ fixes the elements of $F$, we obtain $\sigma(\tau \alpha)=\tau(\sigma \alpha)=\tau\alpha$ for all $\alpha\in F$.
In conclusion, $\sigma\tau=\tau\sigma$, which completes the proof of the lemma.
\end{proof}
As in \cite{Habegger}, the proof of Theorem \ref{thm2} will happen by distinguishing two cases: The one where $p\nmid N$ and the one where $p\vert N$.
The rest of statements of this section are split into two subsections, each one of them handling one of these two cases.
\subsection{The case $p\nmid N$} \label{subsection4.1}
Let us start by treating the case $p\nmid N$.
Denote by $\mathbb{Q}_{p^2}^{nr}$ the maximal unramified extension of $\mathbb{Q}_{p^2}$ and by $\phi\in \mathrm{Gal}(\mathbb{Q}_{p^2}^{nr}/\mathbb{Q}_{p^2})$ the Frobenius element.
By the criterion of Néron–Ogg–Shafarevich \cite[Chapter VII, Theorem 7.1]{Silvermanarithmetic}, we have $E[M]\subset \mathbb{Q}_{p^2}^{nr}$ for all $M\in\mathbb{N}$ coprime to $p$.
In particular, $\mathbb{Q}_{p^2}(N)\subset \mathbb{Q}_{p^2}^{nr}$.
\begin{lmm} \label{lmm3.5}
Let $l\neq p$ be a rational prime.
Then $\phi$ acts on $E[l]$ as multiplication by $b:=\pm p\in \mathbb{Z}$.
\end{lmm}
\begin{proof}
See \cite[Lemma 3.2]{Habegger}.
\end{proof}
The proof of the following lemma is an adaptation of that of \cite[Lemma 5.1]{Habegger}.
\begin{lmm} \label{lmm4}
Let $F/K$ be a finite Galois extension with $F\subset \mathbb{Q}_{p^2}$.
Then $\phi$ acts on $E[N]$ as multiplication by $b$.
Moreover, $\phi$ lies in the center of $\mathrm{Gal}(F(N)/K)$.
\end{lmm}
\begin{proof}
Write $N= \prod_l l^{v_l}$ the decomposition of $N$ into a product of rational primes.
Let $l$ be a rational prime dividing $N$.
Then $l\neq p$ since $p\nmid N$ by hypothesis.
Lemma \ref{lmm3.5} implies that $\phi$ acts on $E[l^{v_l}]$ as multiplication by $b$.
The isomorphism $\bigoplus_l E[l^{v_l}] \simeq E[N]$ is compatible with the action of the Galois group.
Hence $\phi$ acts on $E[N]$ as multiplication by $b$ and the first claim follows.
Now let $\tau\in \mathrm{Gal}(F(N)/K)$.
By the foregoing, we have $\tau(\phi T) = \phi(\tau T)$ for all $T\in E[N]$.
Moreover, $\tau(\phi \alpha)=\phi(\tau\alpha)=\tau\alpha$ for all $\alpha\in F$ since $F/K$ is Galois and $\phi$ fixes the elements of $F$.
The equality $\tau\phi=\phi\tau$ follows, which ends the proof.
\end{proof}
A proof of the next result can be found in \cite[Lemma 4.1]{Habegger}.
\begin{lmm} \label{lmm5}
For all $\alpha\in \mathbb{Q}_{p^2}(N)$, we have \[ \vert \phi\alpha-\alpha^{p^2}\vert_v \leq p^{-1}\max\{1, \vert \phi\alpha\vert_v\}\max\{1, \vert \alpha\vert_v\}^{p^2}. \]
\end{lmm}
Fix until the end of Section \ref{section 5} a Weierstrass model of $E$.
For $P\in E(\mathbb{C}_v)\backslash \{O\}$, put $\lambda_v(P)= (1/2)\max\{0; \log \vert x(P)\vert_v\}$, where $x(P)$ is the first coordinate of $P$ with respect to this model.
Let $[m]$ denote the multiplication-by-$m$ map on $E$.
\begin{lmm} \label{lmm6}
For all $P\in E(\mathbb{Q}_{p^2}(N))$ such that $\phi P \neq [b]Q$, we have
\begin{equation*}
\lambda_v(\phi P-[b]P) \geq \frac{\log p}{2}.
\end{equation*}
\end{lmm}
\begin{proof}
See \cite[Lemma 8.3]{Habegger}.
\end{proof}
\subsection{The case $p\mid N$}
In this subsection, we assume that $p\mid N$.
\begin{lmm} \label{lmm7}
For all $\alpha\in \mathbb{Q}_{p^2}(N)$ and all $\psi\in \mathrm{Gal}(\mathbb{Q}_{p^2}(N)/\mathbb{Q}_{p^2})$, we have \[ \vert \psi(\alpha)^{p^2}-\alpha^{p^2}\vert_v \leq p^{-1}\max\{1, \vert \psi(\alpha)\vert_v\}^{p^2}\max\{1, \vert \alpha\vert_v\}^{p^2}. \]
\end{lmm}
\begin{proof}
See \cite[Lemma 4.2]{Habegger}.
\end{proof}
A proof of the next result can be found in \cite[Lemma 8.4]{Habegger}.
\begin{lmm} \label{lmm8}
For all $P\in E(\mathbb{Q}_{p^2}(N))$ and all $\psi\in \mathrm{Gal}(\mathbb{Q}_{p^2}(N)/\mathbb{Q}_{p^2})$ such that $\psi P \neq P$, we have
\begin{equation*}
\lambda_v (\psi P - P) \geq \frac{\log p}{2(p^2-1)}.
\end{equation*}
\end{lmm}
Let $F/K$ be a Galois extension with $F\subset K^{tv}\subset \mathbb{Q}_{p^2}$.
Note that $F_w=\mathbb{Q}_{p^2}$ for all places $w\vert v$ of $F$ since $K_v=\mathbb{Q}_{p^2}$.
Set $X=\bigcup_{w\vert v} D(w\vert w\cap F(N/p))$, where $w\vert v$ ranges over all places of $F(N)$.
It is a subset of $\mathrm{Gal}(F(N)/F(N/p))$.
Let $\alpha\in F(N)^*$.
To show that $h(\alpha)$ is bounded from below by a positive constant independent of $\alpha, F$ and $N$, unless $\alpha\in\mu_\infty$, one of arguments resides in a Kummerian descent.
More precisely, it consists in proving that $\zeta\alpha\in F(N/p)$ for some $\zeta\in\mu_\infty$.
However, at this stage, we can only show that $\zeta\in F(N)^{\langle X\rangle}$.
Thus, to realize this descent, we need to get the following:
\begin{equation} \label{eq2}
\mathrm{Gal}(F(N)/F(N/p))=\langle X\rangle.
\end{equation}
Denote by $v_0\vert v$ the place of $F(N)$ associated to our fixed field embedding $\overline{K}\to\overline{K_v}$.
Let $w\vert v$ be a place of $F(N)$.
There is a homomorphism $\psi_0\in \mathrm{Gal}(F(N)/K)$ satisfying $w=\psi_0^{-1}v_0$, i.e. $D(w\vert v)=\psi_0 D(v_0\vert v)\psi_0^{-1}$.
The group $\mathrm{Gal}(F(N)/F(N/p))$ being normal in $\mathrm{Gal}(F(N)/K)$, we get \[D(w\vert w\cap F(N/p))=\psi_0 D(v_0\vert v_0 \cap F(N/p))\psi_0^{-1}.\]
We can thus rephrase \eqref{eq2} as follows: The subgroup of $\mathrm{Gal}(F(N)/K)$ generated by all conjugates of \[ D(v_0\vert v_0 \cap F(N/p))= \mathrm{Gal}(\mathbb{Q}_{p^2}(N)/\mathbb{Q}_{p^2}(N/p))\] is equal to $\mathrm{Gal}(F(N)/F(N/p))$.
It is under this viewpoint that Habegger, then Frey, proved \eqref{eq2} (see \cite[Lemma 6.2]{Habegger} and \cite[Lemma 6.5]{Freypubli}).
To get it, they assumed, among other things, that the natural representation $\mathrm{Gal}(\mathbb{Q}(p)/\mathbb{Q})\to \mathrm{Aut} E[p]$ is surjective.
With ours, we can get \eqref{eq2} by using another conditions.
\begin{lmm} \label{lmm9}
Let $F/K$ be a finite extension with $F\subset K^{tv}$.
If the restriction map $\mathrm{Gal}(K^{tv}(N)/K^{tv}) \to \mathrm{Gal}(F(N)/F)$ is an isomorphism, then $\mathrm{Gal}(F(N)/F(N/p)) = \langle \bigcup_{w\vert v} D(w \vert w\cap F(N/p)) \rangle$, where $w\vert v$ ranges over all places of $F(N)$.
\end{lmm}
\begin{proof}
Apply Lemma \ref{lmm2} to $L=K^{tv}, M=K(N)$ and $N=K(N/p)$.
\end{proof}
Here is the last statement of this section.
\begin{lmm} \label{lmm10}
Let $\psi \in \mathrm{Gal}(\mathbb{Q}_{p^2}(N)/\mathbb{Q}_{p^2}(N/p))$, and let $\alpha\in \mathbb{Q}_{p^2}(N)^*$, resp. $P\in E(\mathbb{Q}_{p^2}(N))$ such that $(\psi \alpha)/\alpha\in \mu_\infty$, resp. $\psi P - P \in E_\mathrm{tors}$.
Then
\begin{equation*}
\frac{\psi \alpha}{\alpha} \in \mu_{Q(n)}, \; \text{resp.} \; \psi P - P \in E[Q(n)], \; \text{where} \; Q(n)= \begin{cases}
p^2 \; \text{if} \; p^2 \mid N \\
p^2(p^2-1) \; \text{otherwise}
\end{cases}.
\end{equation*}
\end{lmm}
\begin{proof}
See \cite[Lemma 3.6]{Habegger}, resp. \cite[Lemma 8.2]{Habegger}.
\end{proof}
\section{Proof of Theorem \ref{thm2}: toric case} \label{section 4}
The lemma below is a simple combination of Lemma \ref{lmm5} and Lemma \ref{lmm7}.
\begin{lmm} \label{lmm10.5}
Let $F/K$ be a finite Galois extension with $F\subset \mathbb{Q}_{p^2}$.
Let $N\in\mathbb{N}$ be an integer, and let $\Psi\in\mathrm{Gal}(\mathbb{Q}_{p^2}(N)/\mathbb{Q}_{p^2})$.
If $p\nmid N$, we assume that $\Psi$ is the Frobenius element defined in Subsection \ref{subsection4.1}.
Then, for all $\beta\in F(N)^*$ and all $\tau\in\mathrm{Gal}(F(N)/K)$ commuting with $\Psi$ on $F(N)$, we have \[ \left\vert \tau\left(\frac{\Psi\beta^q}{\beta^{p^2}}\right)-1\right\vert_v\leq p^{-1}\vert \tau\beta\vert_v^q \min\{1, \vert \tau\beta\vert_v\}^{-(p^2+q)}, \] where $q=1$ if $p\nmid N$ and $q=p^2$ otherwise.
\end{lmm}
\begin{proof}
Applying Lemma \ref{lmm5} and Lemma \ref{lmm7} to $\alpha=\tau\beta$ gives
\begin{align*}
\vert \Psi\tau\beta^q - \tau\beta^{p^2}\vert_v \leq p^{-1}\max\{1, \vert \Psi\tau\beta\vert_v\}^q \max\{1, \vert \tau\beta\vert_v\}^{p^2} = p^{-1} \max\{1, \vert \tau \beta\vert_v\}^{p^2+q},
\end{align*}
the equality coming from the fact that two conjugate elements over $K_v=\mathbb{Q}_{p^2}$ have the same $v$-adic absolute value \cite[Chapitre II, Corollaire 3]{Corpslocaux}.
Using the fact that $\tau$ commutes with $\Psi$, then dividing the inequality above by $\vert \tau\beta^{p^2}\vert_v$, we get \[\left\vert \tau\left(\frac{\Psi\beta^q}{\beta^{p^2}}\right)-1\right\vert_v \leq p^{-1} \vert \tau\beta\vert_v^{-p^2}\max\{1, \vert \tau \beta\vert_v\}^{p^2+q}.\]
The lemma follows by noticing that $\min\{1,x\}\max\{1,x\}=x$ for all $x\in\mathbb{R}$.
\end{proof}
A proof of the well-known result below can be found in \cite[Lemma 2 (i)]{Dobrowolski}.
\begin{lmm} \label{lmm10.625}
Let $x\in\overline{\mathbb{Q}}^*$, and let $\sigma\in\mathrm{Gal}(\overline{\mathbb{Q}}/\mathbb{Q})$.
If $(\sigma x)^n/x^m\in\mu_\infty$ for some $n,m\in\mathbb{N}$ distinct, then $x\in\mu_\infty$.
\end{lmm}
Let $(\beta_n)_n$ be a sequence of points in $K^{tv}(E_\mathrm{tors})^*\backslash \mu_\infty$ with $h(\beta_n) \to 0$.
For all $n$, let $N_n\in \mathbb{N}$ be any integer such that $\beta_n\in K^{tv}(N_n)$.
As $K^{tv}/K$ is Galois, there exists a finite Galois extension $F_n/K$, with $F_n \subset K^{tv}\subset K_v= \mathbb{Q}_{p^2}$, such that $\beta_n\in F_n(N_n)$ and $[K^{tv}(N_n) : K^{tv}] = [F_n(N_n) : F_n]$.
For brevity, put $K_n=F_n(N_n)$.
So the restriction map induces an isomorphism
\begin{equation} \label{eq6}
\mathrm{Gal}(K^{tv}K_n/K^{tv}) \to \mathrm{Gal}(K_n/F_n).
\end{equation}
Recall that we fixed a field embedding $\overline{K} \to \overline{K_v}$.
Its restriction to $K_n$ is associated to a place $v_n\vert v$ of $K_n$.
Assume in this paragraph that $p\mid N_n$.
Put $X_n= \bigcup_{w\mid v} D(w \vert w\cap F_n(N_n/p))$, where $w\vert v$ ranges over all places of $K_n$.
Choose $\psi_n\in X_n$.
There is a place $w_n\vert v$ of $K_n$ such that $\psi_n \in D(w_n\vert w_n\cap F_n(N_n/p))$.
It is well known that there exists $\rho_n\in \mathrm{Gal}(K_n/K)$ such that $w_n=\rho_n^{-1} v_n$.
We clearly have $\Psi_n v_n=v_n$, where $\Psi_n = \rho_n\psi_n\rho_n^{-1}$.
Note that $\Psi_n\in \mathrm{Gal}(K_n/F_n(N_n/p))$ since this group contains $\psi_n$ and is normal in $\mathrm{Gal}(K_n/K)$.
Moreover, as $(F_n)_{v_n}=\mathbb{Q}_{p^2}$ since $K\subset F_n\subset \mathbb{Q}_{p^2}=K_v$, we get $\Psi_n\in D(v_n\vert v_n\cap F_n(N_n/p)) = \mathrm{Gal}(\mathbb{Q}_{p^2}(N_n)/\mathbb{Q}_{p^2}(N_n/p))$.
If $p\nmid N_n$, we define $\psi_n$ to be the Frobenius element defined in Subsection \ref{subsection4.1} and $\rho_n$ to be the identity on $K_n$.
Then set $\Psi_n = \rho_n\psi_n\rho_n^{-1}=\psi_n\in \mathrm{Gal}(\mathbb{Q}_{p^2}(N_n)/\mathbb{Q}_{p^2})$.
Put $\gamma_n = \rho_n\left((\psi_n\beta_n^{q_n})/\beta_n^{p^2}\right)\in K_n$, where $q_n=1$ if $p\nmid N_n$ and $q_n=p^2$ otherwise.
\begin{lmm} \label{lmm10.7}
We have $h(\rho_n\beta_n)\to 0, h(\gamma_n)\to 0$ and $\rho_n\beta_n\notin \mu_\infty$ for all $n$.
\end{lmm}
\begin{proof}
The first limit is clear since $h(\rho_n\beta_n)=h(\beta_n)$.
The second one arises from basic properties of $h$ recalled in the introduction since
\[0\leq h(\gamma_n)=h((\psi_n\beta_n^{q_n})/\beta_n^{p^2})\leq h(\psi_n \beta_n^{q_n})+h(\beta_n^{p^2})= (q_n+p^2)h(\beta_n)\leq 2p^2 h(\beta_n).\]
Finally, the third assertion is obvious because by hypothesis, $\beta_n\notin\mu_\infty$ for all $n$.
\end{proof}
Finally, write $d_n=p^{-1/(2(q_n+p^2))}$ and define $Y_n$ as the set of $\tau\in\mathrm{Gal}(K_n/K)$ for which $\vert \tau\rho_n\beta_n\vert_v\leq d_n$.
The cardinality of a finite set $Y$ is denoted with $\#Y$.
\begin{lmm} \label{lmm10.75}
We have $\#Y_n/[K_n : K] \to 0$.
\end{lmm}
\begin{proof}
Applying Theorem \ref{thm4} to $z_n=\rho_n\beta_n$ and $f(\zeta)=\min\{1, \zeta(X)\}$, which is continuous on $\mathbb{P}^1_\mathrm{Berk}(\mathbb{C}_v)$ by Section \ref{section 2}, we deduce that \[ \frac{1}{[K_n:K]}\sum_{\tau\in \mathrm{Gal}(K_n/K)} \min\{1, \vert \tau\rho_n\beta_n\vert_v\} \to \min\{1,1\}=1.\]
Splitting the sum above up when $\tau\in Y_n$ and $\tau\notin Y_n$, we get \[\frac{1}{[K_n:K]}\sum_{\tau\in \mathrm{Gal}(K_n/K)} \min\{1, \vert \tau\rho_n\beta_n\vert_v\} \leq \frac{d_n\#Y_n}{[K_n:K]}+\frac{[K_n:K]-\# Y_n}{[K_n:K]}. \]
A small calculation involving the fact that $d_n<1$ gives \[0\leq \frac{\#Y_n}{[K_n:K]}\leq (d_n-1)^{-1} \left(-1+\frac{1}{[K_n:K]}\sum_{\tau\in \mathrm{Gal}(K_n/K)} \min\{1, \vert \tau\rho_n\beta_n\vert_v\}\right)\]
and the lemma arises from the squeeze theorem ($d_n$ can only take two values).
\end{proof}
Denote by $t_n$ the $p$-adic valuation of $N_n$.
Then for $i\in\mathbb{N}$, set \[f(\zeta)=\min\{1, \zeta(X-1)\} \; \text{and} \; g_i(\zeta)=\min\{1, p^{-1/2}\zeta(X^i)\},\] where $\zeta\in \mathbb{P}^1_{\mathrm{Berk}}(\mathbb{C}_v)$.
They are all continuous on $\mathbb{P}^1_{\mathrm{Berk}}(\mathbb{C}_v)$ by Section \ref{section 2}.
\begin{lmm} \label{lmm10.93}
We have $t_n\geq 1$ and $(\psi_n\beta_n)/\beta_n\in\mu_\infty$ for all $n$ large enough.
\end{lmm}
\begin{proof}
Let $C_n$ be the centralizer of $\Psi_n$ in $G_n=\mathrm{Gal}(K_n/K)$, and put $C_n'=C_n\backslash Y_n$.
Let $\tau\in C_n'$.
Lemma \ref{lmm10.5} applied to $F=F_n, N=N_n, \Psi=\Psi_n$ and $\beta=\rho_n\beta_n$ gives \[f(\zeta_{\tau\gamma_n}) = \min\{1, \vert \tau\gamma_n -1\vert_v\} \leq \min\left\{1, p^{-1} \vert\tau\rho_n\beta_n\vert_v^{q_n}\min\{1, \vert \tau\rho_n\beta_n \vert_v\}^{-(p^2+q_n)}\right\}. \]
Then the fact that $\vert \tau\rho_n\beta_n\vert_v>p^{-1/(2(q_n+p^2))}$ by definition of $C'_n$ leads to
\[f(\zeta_{\tau\gamma_n}) \leq \min\{1, p^{-1/2} \vert \tau\rho_n\beta_n\vert_v^{q_n}\}=g_{q_n}(\zeta_{\tau\rho_n\beta_n,0}).\]
Splitting the sum below up when $\tau\notin C'_n$ and $\tau\in C'_n$, we conclude \[\frac{1}{[K_n : K]}\sum_{\tau\in G_n} f(\zeta_{\tau\gamma_n})^9 \leq \frac{[K_n : K]- \#C'_n}{[K_n : K]}+ \frac{1}{[K_n : K]}\sum_{\tau\in C'_n} g_{q_n}(\zeta_{\tau\rho_n\beta_n,0})^9.\]
If $p\nmid N_n$, then $\Psi_n$ is the Frobenius element and Lemma \ref{lmm4} tells us that $C_n=G_n$, i.e. $\#C_n=[K_n:K]$.
If $p\mid N_n$, then $\mathrm{Gal}(K_n/F_n(N_n/p))\ni \Psi_n$ is normal in $G_n$ and its cardinality is bounded from above by $p^4$.
The orbit-stabilizer theorem implies $\# C_n \geq [K_n : K]/p^4$.
As $g_{q_n}$ is non-negative, it follows from all the foregoing that
\begin{multline} \label{eq8}
\frac{1}{[K_n : K]}\sum_{\tau\in G_n} f(\zeta_{\tau\gamma_n})^9 \leq 1-\frac{1}{p^4}+\frac{\#Y_n}{[K_n:K]} + \frac{1}{[K_n : K]} \sum_{\tau\in G_n} g_{q_n}(\zeta_{\tau\rho_n\beta_n,0})^9.
\end{multline}
Assume by contradiction that $\gamma_n\notin\mu_\infty$ for infinitely many $n$.
By passing to a subsequence if needed, we can assume that $\gamma_n\notin\mu_\infty$ for all $n$ and that the sequence $(q_n)_n$ is constant, say to $q$.
On the one hand, Lemma \ref{lmm10.75}, Lemma \ref{lmm10.7} and Theorem \ref{thm4} applied to $z_n=\rho_n\beta_n$ and $f=g_q^9$ show that the right-hand side in \eqref{eq8} goes to $1-1/p^4+\min\{1, p^{-1/2}\}^9$.
On the other hand, Lemma \ref{lmm10.7} and Theorem \ref{thm4} applied to $z_n=\gamma_n$ and $f=f^9$ prove that the left-hand side in \eqref{eq8} goes to $1$.
Consequently, $1\leq 1-1/p^4+1/p^{9/2}<1$, a contradiction.
Thus $\gamma_n\in\mu_\infty$ for all $n$ large enough.
Choose such a $n$.
The definition of $\gamma_n$ gives us $(\psi_n \beta_n^{q_n})/\beta_n^{p^2}\in\mu_\infty$.
As $\beta_n\notin \mu_\infty$, Lemma \ref{lmm10.625} tells us that $q_n=p^2$ implying $t_n\geq 1$ and $(\psi_n \beta_n)/\beta_n\in\mu_\infty$.
\end{proof}
Recall that $\Psi_n=\rho_n\psi_n\rho_n^{-1}$ and note that
\begin{equation} \label{eq7}
\frac{\Psi_n(\rho_n\beta_n)}{\rho_n\beta_n}= \rho_n\left(\frac{\psi_n\beta_n}{\beta_n}\right).
\end{equation}
\begin{lmm} \label{lmm10.975}
We have $t_n\geq 2$ for all $n$ large enough.
\end{lmm}
\begin{proof}
Assume by contradiction that $t_n <2$ for infinitely many $n$.
By passing to a suitable subsequence, we can assume thanks to Lemma \ref{lmm10.93} that $t_n=1$ and $(\psi_n\beta_n)/\beta_n\in\mu_\infty$ for all $n$.
Combining \eqref{eq7} with Lemma \ref{lmm10} applied to $\psi=\Psi_n, N=N_n$ and $P=\rho_n\beta_n$, we obtain $(\psi_n \beta_n)/ \beta_n \in \mu_{p^2(p-1)}$.
Putting $\beta'_n = \beta_n^{p^2(p-1)}$, we get $\psi_n \beta'_n = \beta'_n$.
The choice of $\psi_n\in X_n$ being purely arbitrary, we conclude $\psi \beta'_n = \beta'_n$ for all $\psi\in X_n$.
Thus $\beta'_n\in K_n^{\langle X_n\rangle}=F_n(N_n/p)$,
the equality coming from Lemma \ref{lmm9} applied to $F=F_n$ (the condition involving the isomorphism is satisfied by \eqref{eq6'}).
Clearly $(\beta'_n)_n$ is a sequence of $K^{tv}(E_\mathrm{tors})^*\backslash\mu_\infty$ such that $h(\beta'_n)\to 0$.
Moreover, combining \eqref{eq6} with the fundamental theorem of Galois theory, we infer that the restriction map $\mathrm{Gal}(K^{tv}(N_n/p)/K^{tv})\to \mathrm{Gal}(F_n(N_n/p)/F_n)$ is an isomorphism.
Thus Lemma \ref{lmm10.93} applied to $\beta_n\beta'_n$ and $N_n=N_n/p$ shows that the $p$-adic valuation of $N_n/p$ is at least $1$ for all $n$ large enough, contradicting $t_n=1$.
\end{proof}
\textit{Proof of Theorem \ref{thm2}, toric case}: Let $(\alpha_n)_n$ be a sequence of points in $K^{tv}(E_\mathrm{tors})^*$ such that $h(\alpha_n)\to 0$.
We want to show that $\alpha_n\in\mu_\infty$ for all $n$ large enough.
For this, assume by contradiction that $\alpha_n\notin \mu_\infty$ for infinitely many $n$, that is for all $n$ by considering a suitable subsequence if needed.
For all $n$, let $N'_n\in \mathbb{N}$ be any integer such that $\alpha_n\in K^{tv}(N'_n)$.
As $K^{tv}/K$ is Galois, there exists a finite Galois extension $F_n/K$, with $F_n \subset K^{tv}$, such that $\alpha_n\in F_n(N'_n)$ and $[K^{tv}(N'_n) : K^{tv}] = [F_n(N'_n) : F_n]$.
The second condition is equivalent to say that the restriction map induces an isomorphism
\begin{equation} \label{eq23}
\mathrm{Gal}(K^{tv}(N'_n)/K^{tv}) \to \mathrm{Gal}(F_n(N'_n)/F_n).
\end{equation}
Let $\sigma_n\in \mathrm{Gal}(F_n(N'_n)/F_n)$ be the homomorphism $\sigma$ defined in Lemma \ref{lmm3} and put $\beta_n=(\sigma_n \alpha_n)/\alpha_n^4$.
Note that $\beta_n\notin \mu_\infty$ according to Lemma \ref{lmm10.625}.
Moreover, the basic properties of $h$ recalled in the introduction give $h(\beta_n) \to 0$ since
\begin{equation*}
0 \leq h(\beta_n) \leq h(\sigma_n \alpha_n)+4h(\alpha_n)=5h(\alpha_n).
\end{equation*}
Let $N_n\in\mathbb{N}$ be the smallest divisor of $N'_n$ such that $\beta_n \in K_n=F_n(N_n)\subset F_n(N'_n)$.
The isomorphism \eqref{eq23}, together with the fundamental theorem of Galois theory, proves that the restriction map induces an isomorphism
\begin{equation} \label{eq60}
\mathrm{Gal}(K^{tv}K_n/K^{tv}) \to \mathrm{Gal}(K_n/F_n).
\end{equation}
We now introduce the same notation as those after \eqref{eq6}.
Combining Lemma \ref{lmm10.93} and Lemma \ref{lmm10.975}, we have $t_n\geq 2$ and $(\psi_n\beta_n)/\beta_n\in\mu_\infty$ for all $n$ large enough.
Choose such an integer $n$.
Note that $\psi_n$ and $\sigma_n$ commute by Lemma \ref{lmm3}.
Thus
\begin{equation*}
\frac{\psi_n \beta_n}{\beta_n} = \frac{\psi_n (\sigma_n \alpha_n)}{\psi_n \alpha_n^4} \frac{\alpha_n^4}{\sigma_n\alpha_n} = \frac{\sigma_n\left(\frac{\psi_n \alpha_n}{\alpha_n}\right)}{\left(\frac{\psi_n\alpha_n}{\alpha_n}\right)^4}\in \mu_\infty
\end{equation*}
and Lemma \ref{lmm10.625} implies $(\psi_n \alpha_n)/\alpha_n \in \mu_\infty$.
As $t_n \geq 2$ and $N_n$ divides $N'_n$, we deduce that $p^2\mid N'_n$.
Lemma \ref{lmm10} applied to $\psi=\Psi_n, N=N'_n$ and $\alpha=\rho_n\alpha_n$ gives \[\frac{\Psi_n(\rho_n \alpha_n)}{\rho_n\alpha_n}= \rho_n\left(\frac{\psi_n\alpha_n}{\alpha_n}\right)\in\mu_{p^2} .\]
Thus $(\psi_n\alpha_n)/\alpha_n \in \mu_{p^2}$ and Lemma \ref{lmm3} leads to $\sigma_n((\psi_n \alpha_n)/\alpha_n)=(\psi_n \alpha_n/\alpha_n)^4$.
As $\psi_n$ and $\sigma_n$ commute, we conclude $\psi_n \beta_n = \beta_n$.
Our choice of $\psi_n\in X_n$ being totally arbitrary, we conclude $\psi \beta_n = \beta_n$ for all $\psi\in X_n$.
In other words, $\beta_n\in K_n^{\langle X_n\rangle}= F_n(N_n/p)$, the equality coming from Lemma \ref{lmm9} applied to $F=F_n$ (the condition involving the isomorphism is satisfied by \eqref{eq60}).
This contradicts the minimality of $N_n$, which ends the proof. \qed
\section{Proof of Theorem \ref{thm2}: elliptic case} \label{section 5}
Recall that $\lambda_v(P)= (1/2)\max\{0; \log \vert x(P)\vert_v\}\geq 0$ for all $P\in E(\mathbb{C}_v)\backslash \{O\}$ and that $b=\pm p$ is defined in Lemma \ref{lmm3.5}.
The proof of Theorem \ref{thm2} in the elliptic case relies on the same arguments as the toric case.
\begin{lmm} \label{lmm13.5}
Let $F/K$ be a finite Galois extension with $F\subset \mathbb{Q}_{p^2}$.
Let $N\in\mathbb{N}$ be an integer, and let $\Psi\in\mathrm{Gal}(\mathbb{Q}_{p^2}(N)/\mathbb{Q}_{p^2})$.
If $p\nmid N$, we assume that $\Psi$ is the Frobenius element defined in Subsection \ref{subsection4.1}.
Put $q=b$ if $p\nmid N$ and $q=1$ otherwise.
Then, for all $Q\in E(F(N))$ such that $\Psi Q\neq [q]Q$ and all $\tau\in\mathrm{Gal}(F(N)/K)$ commuting with $\Psi$ on $F(N)$, we have \[\lambda_v(\tau(\Psi Q-[q]Q))\geq \frac{\log p}{2(p^2-1)}. \]
\end{lmm}
\begin{proof}
By assumption, we clearly have $\Psi\tau Q - [q]\tau Q=\tau(\Psi Q-[q]Q)\neq O$.
Applying Lemma \ref{lmm6} and Lemma \ref{lmm8} to $P=\tau Q$ allows us to deduce that
\[\lambda_v(\Psi \tau Q-[q]\tau Q)\geq \frac{\log p}{2(p^2-1)}, \] which finishes the proof of this lemma.
\end{proof}
Recall that $\hat{h}=\hat{h}_E : E(\overline{K}) \to \mathbb{R}$ denotes the Néron-Tate height.
It is non-negative, invariant under the action of $\mathrm{Gal}(\overline{K}/K)$ and vanishes precisely at $E_\mathrm{tors}$.
It is also quadratic, that is
\begin{equation*}
\forall m\in \mathbb{Z}, \forall P\in E(\overline{K}), \; \hat{h}([m]P)=m^2\hat{h}(P).
\end{equation*}
This implies
\begin{equation*}
\forall P\in E(\overline{K}), \forall T\in E_\mathrm{tors}, \; \hat{h}(P+T)=\hat{h}(P).
\end{equation*}
Finally, it also satisfies the parallelogram law, that is
\begin{equation*}
\forall P,Q\in E(\overline{K}), \hat{h}(P+Q)+\hat{h}(P-Q)=2(\hat{h}(P)+\hat{h}(Q)).
\end{equation*}
For more information on $\hat{h}$, we refer to \cite[Chapter VIII, §9]{Silvermanarithmetic}.
\begin{lmm} \label{lmm13.625}
Let $P\in E(\overline{K})$, and let $\sigma\in\mathrm{Gal}(\overline{K}/K)$.
If $[n]\sigma P - [m]P\in E_\mathrm{tors}$ for some $n,m\in\mathbb{N}$ distinct, then $P\in E_\mathrm{tors}$.
\end{lmm}
\begin{proof}
Some properties of $\hat{h}$ recalled above show that \[m^2\hat{h}(P)=\hat{h}([m]P)=\hat{h}([m]P+[n]\sigma P-[m]P)=n^2\hat{h}(\sigma P)=n^2\hat{h}(P) \] and the lemma follows since $n,m\in\mathbb{N}$ are distinct.
\end{proof}
Let $(Q_n)_n$ be a sequence of non-torsion points in $E(K^{tv}(E_\mathrm{tors}))$ with $\hat{h}(Q_n) \to 0$.
For all $n$, let $N_n\in \mathbb{N}$ be any integer such that $Q_n\in E(K^{tv}(N_n))$.
As $K^{tv}/K$ is Galois, there exists a finite Galois extension $F_n/K$, with $F_n \subset K^{tv}\subset K_v= \mathbb{Q}_{p^2}$, such that $Q_n\in E(F_n(N_n))$ and $[K^{tv}(N_n) : K^{tv}] = [F_n(N_n) : F_n]$.
For brevity, put $K_n=F_n(N_n)$.
So the restriction map induces an isomorphism
\begin{equation} \label{eq6'}
\mathrm{Gal}(K^{tv}K_n/K^{tv}) \to \mathrm{Gal}(K_n/F_n).
\end{equation}
Recall that we fixed a field embedding $\overline{K} \to \overline{K_v}$.
Its restriction to $K_n$ is associated to a place $v_n\vert v$ of $K_n$.
Assume in this paragraph that $p\mid N_n$.
Put $X_n= \bigcup_{w\mid v} D(w \vert w\cap F_n(N_n/p))$, where $w\vert v$ ranges over all places of $K_n$.
Choose $\psi_n\in X_n$.
There is a place $w_n\vert v$ of $K_n$ such that $\psi_n \in D(w_n\vert w_n\cap F_n(N_n/p))$.
It is well known that there exists $\rho_n\in \mathrm{Gal}(K_n/K)$ such that $w_n=\rho_n^{-1} v_n$.
We clearly have $\Psi_n v_n=v_n$, where $\Psi_n = \rho_n\psi_n\rho_n^{-1}$.
Note that $\Psi_n\in \mathrm{Gal}(K_n/F_n(N_n/p))$ since this group contains $\psi_n$ and is normal in $\mathrm{Gal}(K_n/K)$.
Moreover, as $(F_n)_{v_n}=\mathbb{Q}_{p^2}$ since $K\subset F_n\subset \mathbb{Q}_{p^2}=K_v$, we get $\Psi_n\in D(v_n\vert v_n\cap F_n(N_n/p)) = \mathrm{Gal}(\mathbb{Q}_{p^2}(N_n)/\mathbb{Q}_{p^2}(N_n/p))$.
If $p\nmid N_n$, we define $\psi_n$ to be the Frobenius element defined in Subsection \ref{subsection4.1} and $\rho_n$ to be the identity on $K_n$.
Then set $\Psi_n = \rho_n\psi_n\rho_n^{-1}=\psi_n\in \mathrm{Gal}(\mathbb{Q}_{p^2}(N_n)/\mathbb{Q}_{p^2})$.
Put $R_n = \rho_n\left(\psi_n Q_n- [q_n]Q_n\right)$, where $q_n=b$ if $p\nmid N_n$ and $q_n=1$ otherwise.
\begin{lmm} \label{lmm13.7}
We have $\hat{h}(\rho_n Q_n)\to 0, \hat{h}(R_n)\to 0$ and $\rho_n Q_n\notin E_\mathrm{tors}$ for all $n$.
\end{lmm}
\begin{proof}
The first limit is clear since $\hat{h}(\rho_n Q_n)=\hat{h}(Q_n)$.
The second one arises from basic properties of $\hat{h}$ recalled above since
\[0\leq \hat{h}(R_n)=\hat{h}(\psi_n Q_n- [q_n]Q_n)\leq 2(\hat{h}(\psi_n Q_n)+ \hat{h}([q_n]Q_n))\leq 2(1+p^2) \hat{h}(Q_n).\]
Finally, the third assertion is obvious because $Q_n\notin E_\mathrm{tors}$ for all $n$.
\end{proof}
Denote by $t_n$ the $p$-adic valuation of $N_n$ and recall that $\lambda_v(.)= (1/2)\max\{0; \log \vert x(.)\vert_v\}$.
\begin{lmm} \label{lmm13.93}
We have $t_n\geq 1$ and $\psi_n Q_n-Q_n\in E_\mathrm{tors}$ for all $n$ large enough.
\end{lmm}
\begin{proof}
For any $P\in E(K_n)\backslash \{O\}$ and any place $w$ of $K$, put \[\hat{h}_w(P) = \frac{1}{[K_n : \mathbb{Q}]}\sum_{w'\mid w} [(K_n)_{w'} : \mathbb{Q}_w] \lambda_{w'}(P),\] where $w'\vert w$ runs over all places of $K_n$, and $\lambda_{w'} : E((K_n)_{w'})\backslash \{O\} \to \mathbb{R}$ denotes the local N\'eron height function on $E$ associated to $w'$ \cite[Chapter VI]{Silvermanadvanced}.
As $E$ has good reduction at $v$, we deduce from \cite[Chapter VI, Theorem 4.1]{Silvermanadvanced} that if $w'\vert v$, then $\lambda_{w'}(.)=(1/2)\max\{0,\log \vert x(.)\vert_{w'}\}$ on $K_n$.
In particular, $\lambda_{v_n}=\lambda_v$ on $K_n$.
From \cite[Theorem 1.1 (c)]{Silvermanadvanced}, it is easy to check that the function $\hat{h}_w$ remains unchanged if we replace $K_n$ with any number field $F$ such that $P\in E(F)$.
This explains why we do not mention the dependence on $K_n$ in this notation.
As $\mathrm{Gal}(K_n/K)$ acts transitively on the places of $K_n$ above $v$, we infer that
\begin{align*}
\hat{h}_v(P) & = \frac{[K_v : \mathbb{Q}_p]}{[K_n : \mathbb{Q}]} \sum_{w' \vert v} [(K_n)_{w'} : K_v] \lambda_{w'}(P) = \frac{[K_v : \mathbb{Q}_p]}{[K_n : \mathbb{Q}]} \sum_{\tau\in \mathrm{Gal}(K_n/K)} \lambda_{\tau^{-1}v_n}(P) \\
& = \frac{[K_v : \mathbb{Q}_p ]}{[K_n : \mathbb{Q}]} \sum_{\tau\in \mathrm{Gal}(K_n/K)} \lambda_{v_n}(\tau P) = \frac{[K_v : \mathbb{Q}_p ]}{[K_n : \mathbb{Q}]} \sum_{\tau\in \mathrm{Gal}(K_n/K)} \lambda_v(\tau P).
\end{align*}
Let $C_n$ be the centralizer of $\Psi_n$ in $G_n=\mathrm{Gal}(K_n/K)$.
If $p\nmid N_n$, then $\Psi_n$ is the Frobenius element and Lemma \ref{lmm4} tells us that $C_n=G_n$.
If $p\mid N_n$, then $\mathrm{Gal}(K_n/F_n(N_n/p))\ni \Psi_n$ is normal in $G_n$ and its cardinality is bounded from above by $p^4$.
The orbit-stabilizer theorem implies $\# C_n \geq [K_n : K]/p^4$.
Assume by contradiction that $R_n\notin E_\mathrm{tors}$ for infinitely many $n$, that is for all $n$ by considering a suitable subsequence.
Note that $\Psi_n \rho_n Q_n - [q_n]\rho_nQ_n = R_n\neq O$.
Let $\tau\in C_n$.
Lemma \ref{lmm13.5} applied to $\Psi=\Psi_n, N=N_n$ and $Q=\rho_nQ_n$ proves that \[ \lambda_v(\tau R_n)= \lambda_v(\tau(\Psi_n \rho_n Q_n - [q_n]\rho_nQ_n)) \geq \frac{\log p}{2(p^2-1)}. \]
As $\lambda_v$ is non-negative and $\#C_n\geq [K_n : K]/p^4$, we get \[\hat{h}_v(R_n)= \frac{[K_v : \mathbb{Q}_p ]}{[K_n : \mathbb{Q}]} \sum_{\tau\in \mathrm{Gal}(K_n/K)} \lambda_v(\tau R_n) \geq \frac{[K_v : \mathbb{Q}_p]\log p}{2[K:\mathbb{Q}]p^4(p^2-1)}. \]
By \cite[Chapter VI, Theorem 2.1]{Silvermanadvanced}, we have $\hat{h}=\sum_w \hat{h}_w$ on $K_n$.
Let $\mathcal{M}$ denote the (finite) set containing all infinite places of $K$ and all finite places $w$ of $K$ for which $E$ does not have potential good reduction at $w$.
We infer from \cite[Lemme 4.4]{Plessisiso} that \[ \hat{h}(R_n)= \hat{h}_v(R_n)+ \sum_{w\neq v} \hat{h}_w(R_n) \geq \hat{h}_v(R_n)+ \sum_{w\in \mathcal{M}} \hat{h}_w(R_n) \geq \frac{[K_v : \mathbb{Q}_p]\log p}{4[K : \mathbb{Q}]p^4(p^2-1)} \] for all $n$ large enough, contradicting $\hat{h}(R_n)\to 0$.
Thus $R_n\in E_\mathrm{tors}$ for all $n$ large enough.
Choose such a $n$.
The definition of $R_n$ implies $\psi_n Q_n - [q_n]Q_n\in E_\mathrm{tors}$.
As $Q_n\notin E_\mathrm{tors}$, Lemma \ref{lmm13.625} leads to $q_n=1$, and so $t_n\geq 1$ and $\psi_n Q_n - Q_n\in E_\mathrm{tors}$.
\end{proof}
Recall that $\Psi_n=\rho_n\psi_n\rho_n^{-1}$ and note that
\begin{equation} \label{eq7'}
\Psi_n(\rho_n Q_n) - \rho_n Q_n= \rho_n\left(\psi_n Q_n - Q_n\right).
\end{equation}
\begin{lmm} \label{lmm13.975}
We have $t_n\geq 2$ for all $n$ large enough.
\end{lmm}
\begin{proof}
Assume by contradiction that $t_n <2$ for infinitely many $n$.
By passing to a suitable subsequence, we can assume thanks to Lemma \ref{lmm13.93} that $t_n=1$ and $\psi_n Q_n - Q_n\in E_\mathrm{tors}$ for all $n$.
Combining \eqref{eq7'} with Lemma \ref{lmm10} applied to $\psi=\Psi_n, N=N_n$ and $\alpha=\rho_nQ_n$, we obtain $\psi_n Q_n - Q_n \in E[p^2(p-1)]$.
Putting $Q'_n = [p^2(p-1)]Q_n$, we get $\psi_n Q'_n = Q'_n$.
The choice of $\psi_n\in X_n$ being purely arbitrary, we conclude $\psi Q'_n = Q'_n$ for all $\psi\in X_n$.
Thus $Q'_n\in E(K_n^{\langle X_n\rangle})=E(F_n(N_n/p))$, the equality coming from Lemma \ref{lmm9} applied to $F=F_n$ (the condition involving the isomorphism is satisfied by \eqref{eq6'}).
Clearly $(Q'_n)_n$ is a sequence of $E(K^{tv}(E_\mathrm{tors}))\backslash E_\mathrm{tors}$ such that $\hat{h}(Q'_n)\to 0$.
Moreover, combining \eqref{eq6'} with the fundamental theorem of Galois theory, we infer that the restriction map $\mathrm{Gal}(K^{tv}(N_n/p)/K^{tv})\to \mathrm{Gal}(F_n(N_n/p)/F_n)$ is an isomorphism.
Thus Lemma \ref{lmm13.93} applied to $Q_n=Q'_n$ and $N_n=N_n/p$ proves that the $p$-adic valuation of $N_n/p$ is at least $1$ for all $n$ large enough, contradicting $t_n=1$.
\end{proof}
\textit{Proof of Theorem \ref{thm2}, elliptic case}: Let $(P_n)_n$ be a sequence of points in $E(K^{tv}(E_\mathrm{tors}))$ such that $\hat{h}(P_n)\to 0$.
We want to show that $P_n\in E_\mathrm{tors}$ for all $n$ large enough.
For this, assume by contradiction that $P_n\notin E_\mathrm{tors}$ for infinitely many $n$, that is for all $n$ by considering a suitable subsequence if needed.
For all $n$, let $N'_n\in \mathbb{N}$ be any integer such that $P_n\in E(K^{tv}(N'_n))$.
As $K^{tv}/K$ is Galois, there exists a finite Galois extension $F_n/K$, with $F_n \subset K^{tv}$, such that $P_n\in E(F_n(N'_n))$ and $[K^{tv}(N'_n) : K^{tv}] = [F_n(N'_n) : F_n]$.
The second condition is equivalent to say that the restriction map induces an isomorphism
\begin{equation} \label{eq24}
\mathrm{Gal}(K^{tv}(N'_n)/K^{tv}) \to \mathrm{Gal}(F_n(N'_n)/F_n).
\end{equation}
Let $\sigma_n\in \mathrm{Gal}(F_n(N'_n)/F_n)$ be the homomorphism $\sigma$ defined in Lemma \ref{lmm3} and put $Q_n=\sigma_n P_n - [2]P_n$.
Note that $Q_n\notin E_\mathrm{tors}$ according to Lemma \ref{lmm13.625}.
Moreover, the basic properties of $\hat{h}$ recalled previously give $\hat{h}(Q_n) \to 0$ since
\begin{equation*}
0 \leq \hat{h}(Q_n) \leq 2(\hat{h}(\sigma_n P_n)+4\hat{h}(P_n))=10 \hat{h}(P_n).
\end{equation*}
Let $N_n\in\mathbb{N}$ be the smallest divisor of $N'_n$ such that $Q_n \in E(K_n)$, where $K_n=F_n(N_n)\subset F_n(N'_n)$.
The isomorphism \eqref{eq24}, together with the fundamental theorem of Galois theory, proves that the restriction map induces an isomorphism
\begin{equation} \label{eq60'}
\mathrm{Gal}(K^{tv}K_n/K^{tv}) \to \mathrm{Gal}(K_n/F_n).
\end{equation}
We now introduce the same notation as those after \eqref{eq6'}.
Combining Lemma \ref{lmm13.93} and Lemma \ref{lmm13.975}, we have $t_n\geq 2$ and $\psi_n Q_n - Q_n\in E_\mathrm{tors}$ for all $n$ large enough.
Choose such an integer $n$.
Note that $\psi_n$ and $\sigma_n$ commute by Lemma \ref{lmm3}.
Thus
\begin{align*}
\psi_n Q_n - Q_n & = \sigma_n\psi_n P_n -[2]\psi_n P_n - \sigma_n P_n+[2]P_n \\
& = \sigma_n(\psi_n P_n - P_n)- [2](\psi_n P_n - P_n)\in E_\mathrm{tors}
\end{align*}
and Lemma \ref{lmm13.625} implies $\psi_n P_n - P_n \in E_\mathrm{tors}$.
As $t_n \geq 2$ and $N_n$ divides $N'_n$, we deduce that $p^2\mid N'_n$.
Lemma \ref{lmm10} applied to $\psi=\Psi_n, N=N'_n$ and $P=\rho_nP_n$ gives \[\Psi_n(\rho_n P_n)-\rho_nP_n= \rho_n(\psi_n P_n - P_n)\in E[p^2] .\]
Thus $\psi_n P_n - P_n \in E[p^2]$ and Lemma \ref{lmm3} leads to $\sigma_n(\psi_n P_n - P_n)=[2](\psi_n P_n - P_n)$.
As $\psi_n$ and $\sigma_n$ commute, we get $\psi_n Q_n = Q_n$.
Our choice of $\psi_n\in X_n$ being totally arbitrary, we conclude $\psi Q_n = Q_n$ for all $\psi\in X_n$.
In other words, $Q_n\in E(K_n^{\langle X_n\rangle})= E(F_n(N_n/p))$, the equality coming from Lemma \ref{lmm9} applied to $F=F_n$ (the condition involving the isomorphism is satisfied by \eqref{eq60'}).
This contradicts the minimality of $N_n$, which proves Theorem \ref{thm2}. \qed
\section{Preliminaries for Theorem \ref{thm3}}
\subsection{A crucial lemma}
In this subsection, $L/K$ is a Galois extension of number fields, and $w$ is a finite place of $L$ lying over a place $v$ of $K$.
By \cite[Chapter 1, Proposition 1.3.1]{BombieriandGubler}, there is a field embedding $\iota_w: L\to L_w$ such that $\vert \iota_w(.)\vert_v=\vert .\vert_w$ on $L$.
The well-known lemma below will be used several times throughout this text.
\begin{lmm} \label{lmm decompo}
For all $\psi\in D(w\vert v)$, there is $\psi_w\in \mathrm{Gal}(L_w/K_v)$ such that $\iota_w \psi = \psi_w \iota_w$.
\end{lmm}
\begin{proof}
By \cite[Chapitre II, §3, Corollaire 4]{Corpslocaux}, $\psi$ can extend to an element $\tau\in \mathrm{Gal}(L_w/K_v)$.
Of course, here "extend" means "after embedding $L$ into $L_w$ via $\iota_w$".
Thus $\tau(\iota_w x)= \iota_w(\psi x)$ for all $x\in L$.
The lemma follows by putting $\psi_w=\tau$.
\end{proof}
\subsection{Some results extracted from \cite{Freypubli}}
Fix from now a non-CM elliptic curve $E/\mathbb{Q}$.
Again, put $K(N)=K(E[N])$ for any field $K$ and any integer $N\in\mathbb{N}$.
In \cite{Freypubli}, $L/\mathbb{Q}$ is a Galois extension whose Galois group has finite exponent $\mathrm{exp}(L)$.
Recall that according to a theorem of Checcoli, this is equivalent to saying that $d=\sup_q\{d_q(L)\}$ is finite (see Definition \ref{defn1}), where $q$ ranges over all rational primes.
Then Frey chose a rational prime $p\geq 5$ satisfying:
\begin{itemize}
\item [(a)] $E$ has supersingular reduction at $p$ and $j_E \not\equiv 0, 1728 \; (p)$;
\item [(b)] the natural representation $\mathrm{Gal}(\mathbb{Q}(E[p])/\mathbb{Q}) \to \mathrm{Aut} E[p]$ is surjective;
\item [(c)] $p \geq \max\{2d+2; \exp(L)\}$.
\end{itemize}
Next she fixed $\alpha\in L(E_\mathrm{tors})^*\backslash \mu_\infty$ and a number field $K\subset L$, Galois over $\mathbb{Q}$, such that $\alpha\in K(E_\mathrm{tors})$.
Finally, she put $F=K_v\mathbb{Q}_{p^2}$, where $v$ denotes the place of $K$ associated to a fixed field embedding $\overline{\mathbb{Q}}\to\overline{\mathbb{Q}_p}$.
Actually, we can prove \cite[Lemma 3.1-Lemma 6.1]{Freypubli} without involving most of conditions above.
Strictly speaking, we should reprove them all using only the minimal conditions.
But they are very technical, making it impossible without considerably burden this text.
As a compromise, for each one of these lemmas, we mention below the hypotheses/references that Frey used to prove it.
Then we detail one by one the requested conditions to use these references.
In \cite[Lemma 3.1]{Freypubli}, she only used $[F:\mathbb{Q}_p]<p$ and \cite[Lemma 3.4]{Habegger}.
In \cite[Lemma 3.2]{Freypubli}, she only used $[F:\mathbb{Q}_p]<p$ and \cite[Lemma 3.3, Lemma 3.4]{Habegger}.
In \cite[Lemma 3.3]{Freypubli}, she only used $[F:\mathbb{Q}_p]<p$ and \cite[Lemma 3.3]{Habegger}.
In \cite[Lemma 3.4]{Freypubli}, she only used \cite[Lemma 2.1, Lemma 3.3, Lemma 3.4]{Habegger} and \cite[Proposition II.7.12]{Neukirch}.
In \cite[Lemma 3.5]{Freypubli}, she only used \cite[Lemma 2.1, Lemma 3.3]{Habegger} and \cite[Lemme IV.5, Proposition IV.12]{Corpslocaux}.
In \cite[Lemma 3.6]{Freypubli}, she only used \cite[Lemma 3.5]{Habegger}, \cite[Proposition II.7.13]{Neukirch}, Goursat's lemma and $[F:\mathbb{Q}_p]<p$.
In \cite[Lemma 3.7]{Freypubli}, she only used \cite[Proposition II.7.12]{Neukirch}.
In \cite[Lemma 3.8]{Freypubli}, she only used results of \cite{Freypubli}.
In \cite[Lemma 4.1]{Freypubli}, she only used \cite[Lemma 2.1, Lemma 3.1, Lemma 5.1]{Habegger} and $K/\mathbb{Q}$ is a finite Galois extension.
In \cite[Lemma 4.2-Lemma 4.5]{Freypubli}, she only used results of \cite{Freypubli}.
In \cite[Lemma 4.6]{Freypubli}, she only used \cite[Theorem 1]{Dobrowolski} and $K/\mathbb{Q}$ is a finite Galois extension.
In \cite[Lemma 5.1-Lemma 5.2]{Freypubli}, she only used results of \cite{Freypubli}.
In \cite[Lemma 5.3]{Freypubli}, she only used \cite[Lemma 5.2]{Habegger}.
In \cite[Lemma 5.4]{Freypubli}, she only used \cite[Theorem 1]{Dobrowolski} and $K/\mathbb{Q}$ is a finite Galois extension.
In \cite[Lemma 6.1]{Freypubli}, she only used the fact that $K/\mathbb{Q}$ is a finite Galois extension.
Thus, to prove \cite[Lemma 3.1-Lemma 6.1]{Freypubli}, we only need $[F:\mathbb{Q}_p]<p, K/\mathbb{Q}$ is a finite Galois extension, and necessary conditions so that \cite[Lemma 2.1, Lemma 3.3-Lemma 3.5, Lemma 5.1-Lemma 5.2]{Habegger}, \cite[Proposition II.7.12-Proposition II.7.13]{Neukirch}, \cite[Lemme IV.5, Proposition IV.12]{Corpslocaux}, \cite[Theorem 1]{Dobrowolski} and Goursat's lemma hold.
Goursat's lemma is a general fact about group theory, \cite[Theorem 1]{Dobrowolski} is the well-known Dobrowolski's Theorem, \cite[Lemme IV.5, Proposition IV.12]{Corpslocaux} are general results about ramification groups and \cite[Proposition II.7.12- Proposition II.7.13]{Neukirch} are general facts about cyclotomic fields.
Then \cite[Lemma 2.1]{Habegger} is a general lemma on local fields.
Finally, the results in \cite[Section 3, Section 5]{Habegger} hold for any rational prime $p\geq 5$ satisfying $(a)$ above.
From all this, we get the following.
\begin{lmm} \label{lmm14.5}
The lemmas \cite[Lemma 3.1- Lemma 6.1]{Freypubli} hold in the following setting:
\begin{itemize}
\item $K/\mathbb{Q}$ is a finite Galois extension;
\item there exists a rational prime $p\geq 5$ such that:
\begin{itemize}
\item $E$ has supersingular reduction at $p$ and $j_E \not\equiv 0, 1728 \; (p)$;
\item $[K_w\mathbb{Q}_{p^2} : \mathbb{Q}_p]<p$ for all places $w\vert p$ of $K$.
\end{itemize}
\end{itemize}
\end{lmm}
Note that viewed as a field, $K_w$ does not depend on $w\vert p$ since $K/\mathbb{Q}$ is Galois.
We can now extract from \cite[Lemma 3.1- Lemma 6.1]{Freypubli} some lemmas that will be useful to prove Theorem \ref{thm3}.
Throughout this text, write $\mu_{p^\infty}=\bigcup_{n\in\mathbb{N}} \mu_{p^n}$.
\begin{lmm} \label{lmm14}
Let $K$ and $p$ be as in Lemma \ref{lmm14.5}.
Let $N\in\mathbb{N}$ such that its $p$-adic valuation $n$ is at least $2$.
Then $K(N)\cap \mu_{p^\infty}=\mu_{p^n}$.
\end{lmm}
\begin{proof}
See \cite[Lemma 3.6]{Freypubli}.
\end{proof}
\begin{lmm} \label{lmm15}
Let $K$ and $p$ be as in Lemma \ref{lmm14.5}.
Let $N\in\mathbb{N}$ such that $p^2\nmid N$, and let $a\in K(N)^*\backslash \mu_\infty$.
Then \[ h(a)\geq \left( \frac{2 \log p}{p^3(1+p^{p^3})(1+5\cdot2^{11})}\right)^4. \]
\end{lmm}
\begin{proof}
Applying \cite[Lemma 4.6]{Freypubli} to $\mathcal{E}=(p^2-1)[K_v\mathbb{Q}_{p^2}:\mathbb{Q}_{p^2}]< p^3/2$.
\end{proof}
\begin{rqu} \label{rqu1}
\rm{In the proof of \cite[Theorem 7.1]{Freypubli}, Frey had to use this lemma with $\mathcal{E}=(p^2-1)[K_v\mathbb{Q}_{p^2}:\mathbb{Q}_{q^2}]\exp(L)<p^4/2$.
This explains why we have $p^3+3$ instead of $p^4+4$ in the lower bound of Theorem \ref{thm3}.}
\end{rqu}
In \cite{Freypubli}, Frey made extensive use of the fact that $K\subset F$.
But for our purpose, it is not convenient to fix a field embedding $\overline{\mathbb{Q}}\to \overline{\mathbb{Q}_p}$.
Consequently, we cannot use the inclusion $K\subset F$.
See below how to remedy this problem.
Let $N\in\mathbb{N}$ be an integer divisible by $p^2-1$, and let $w\vert v$ be the place of $K(N)$ associated to the fixed field embedding $\overline{\mathbb{Q}}\to\overline{\mathbb{Q}_p}$.
Properties of the Weil pairing show that $\zeta_{p^2-1}\in \mathbb{Q}(N)$.
As $\mathbb{Q}_{p^2}=\mathbb{Q}_p(\zeta_{p^2-1})$, we easily get $K(N)_w=K_v(N)=K_v\mathbb{Q}_{p^2}(N)=F(N)$.
Similarly, $K(\zeta_{p^2-1})_w=K_v(\zeta_{p^2-1})=K_v\mathbb{Q}_{p^2}=F$.
Thus, for our purpose, it is more convenient to adopt the viewpoint "$a\in K(N)$ implies $\iota_w a\in K(N)_w$" rather than "$K(N)\subset F(N)$" and use $D(w\vert w\cap K(\zeta_{p^2-1}))$ instead of $\mathrm{Gal}(F(N)/F)$ (these groups are equal).
Finally, note that for any place $w'\vert p$ of $K(N)$, there is a field embedding $\overline{\mathbb{Q}}\to\overline{\mathbb{Q}_p}$ whose restriction to $K(N)$ is associated to $w'$.
From all this, we get the following.
\begin{lmm} \label{lmm16}
Let $K$ and $p$ be as in Lemma \ref{lmm14.5}.
Take $N\in\mathbb{N}$ divisible by $p^2(p^2-1)$ and a place $w\vert p$ of $K(N)$.
If $a\in K(N)^*\backslash \mu_\infty$ with $\iota_w a^{p^2}\notin K(N/p)_w$, then \[ h(a)\geq \frac{(\log p)^4}{4\cdot 10^6p^{32}}. \]
\end{lmm}
\begin{proof}
See \cite[Lemma 5.4]{Freypubli}.
\end{proof}
\begin{lmm} \label{lmm17'}
Let $K$ and $p$ be as in Lemma \ref{lmm14.5}, and let $N\in\mathbb{N}$ be an integer divisible by $p^2-1$ such that its $p$-adic valuation $n$ is at least $2$.
Choose a place $w\vert p$ of $K(N)$.
Then there is $\sigma\in D(w\vert w\cap K(\zeta_{p^2-1}))$ lying in the center of $\mathrm{Gal}(K(N)/K)$ such that $\sigma \zeta=\zeta^{4^{[K_w\mathbb{Q}_{p^2}:\mathbb{Q}_{p^2}]}}$ for all $\zeta\in\mu_{p^n}$.
\end{lmm}
\begin{proof}
See \cite[Lemma 6.1]{Freypubli}.
\end{proof}
The last result that we need to prove Theorem \ref{thm3} is \cite[Lemma 6.5]{Freypubli}.
The proof made by Frey of this lemma requires using (b) and the fact that $\mathrm{Gal}(L/\mathbb{Q})$ has finite exponent.
We give below a variant of this lemma, which permits us to obtain the same conclusion as her, but under another conditions.
\begin{lmm} \label{lmm18}
Let $p$ be a rational prime, and let $L/\mathbb{Q}^{tp}$ be an algebraic extension.
Take an integer $N\in\mathbb{N}$ divisible by $p$ and a finite Galois extension $K/\mathbb{Q}$ such that the restriction map $\mathrm{Gal}(L(N)/L)\to\mathrm{Gal}(K(N)/K)$ is an isomorphism.
Let $\gamma\in K(N)$.
If $\iota_w \gamma\in K(N/p)_w $ for all places $w\vert p$ of $K(N)$, then $\gamma\in K(N/p)$.
\end{lmm}
\begin{proof}
Set $X=\bigcup_{w\vert p} D(w \vert w\cap K(N/p))$, where $w\vert p$ ranges over all places of $K(N)$.
Lemma \ref{lmm2} applied to $K=\mathbb{Q}, v=p, M=\mathbb{Q}(N), F=K$ and $N=\mathbb{Q}(N/p)$ gives $\langle X\rangle = \mathrm{Gal}(K(N)/K(N/p))$.
Let $\psi\in X$.
Then there is a place $w\vert p$ of $K(N)$ such that $\psi\in D(w\vert w\cap K(N/p))$.
By Lemma \ref{lmm decompo}, there exists $\psi_w\in \mathrm{Gal}(K(N)_w/K(N/p)_w)$ such that $\psi_w\iota_w= \iota_w\psi$.
Hence $\iota_w(\psi\gamma) = \psi_w(\iota_w \gamma)=\iota_w\gamma$ by assumption.
The injectivity of $\iota_w$ leads to $\psi\gamma=\gamma$ for all $\psi\in X$, and so for all $\psi\in \langle X\rangle$.
The lemma follows by the foregoing.
\end{proof}
\section{Proof of Theorem \ref{thm3}} \label{section 6}
Let $L/\mathbb{Q}$ be a Galois extension for which there is a rational prime $p\geq 5$ satisfying:
\begin{itemize}
\item $d_p(L)$ is finite and $p>2d_p(L)$ (see Definition \ref{defn1});
\item $E$ has supersingular reduction at $p$ and $j_E\not\equiv 0,1728 \; (p)$.
\end{itemize}
The goal of this section is to give an explicit lower bound, depending only on $p$, for the Weil height of an element belonging to $L(E_\mathrm{tors})^*\backslash \mu_\infty$.
As $d_p(\mathbb{Q}^{tp})=1$, we get $d_p(L)=d_p(L\mathbb{Q}^{tp})$.
As $L\mathbb{Q}^{tp}/\mathbb{Q}$ is Galois, we can assume that $\mathbb{Q}^{tp}\subset L$.
Fix $\alpha\in L(E_\mathrm{tors})^*\backslash \mu_\infty$, then an integer $N\in\mathbb{N}$ such that $\alpha\in L(N)$.
By enlarging $N$ if needed, we can assume that $p^2-1$ divides $N$.
As $L/\mathbb{Q}$ is Galois, there exists a finite Galois subextension $K/\mathbb{Q}$ such that $\alpha\in K(N)$ and $[L(N):L]=[K(N):K]$.
Equivalently, the restriction map $\mathrm{Gal}(L(N)/L)\to \mathrm{Gal}(K(N)/K)$ is an isomorphism.
Finally, the fundamental theorem of Galois theory ensures us that for all divisors $N'$ of $N$ (implying $L(N')\subset L(N)$ and $K(N')\subset K(N)$), the restriction map induces an isomorphism $\mathrm{Gal}(L(N')/L)\to \mathrm{Gal}(K(N')/K)$.
Note that $[K_w\mathbb{Q}_{p^2}:\mathbb{Q}_p]\leq 2[K_w:\mathbb{Q}_p]\leq 2d_p(L)<p$ for all places $w\vert p$ of $K$.
Hence $K$ and $p$ satisfy the conditions of Lemma \ref{lmm14.5}.
Write $N=p^nM$ with $M$ coprime to $p$.
Note that $p^2-1$ divides $M$.
If $n\leq 1$, Theorem \ref{thm3} is then an immediate consequence of Lemma \ref{lmm15} applied to $a=\alpha$.
So assume that $n\geq 2$.
Let $\sigma\in\mathrm{Gal}(K(N)/K(\zeta_{p^2-1}))$ be as in Lemma \ref{lmm17'} and define
\begin{equation} \label{eq30}
\gamma= \frac{\sigma\alpha}{\alpha^D}\in K(N),
\end{equation}
where $D=4^{[K_w\mathbb{Q}_{p^2}:\mathbb{Q}_{p^2}]}=2^{[K_w\mathbb{Q}_{p^2}:\mathbb{Q}_p]}<2^p$ (this number is independent of $w$ since $K/\mathbb{Q}$ is Galois).
As $\alpha\notin\mu_\infty$, it arises from Lemma \ref{lmm10.625} that $\gamma\notin\mu_\infty$.
Moreover,
\begin{equation} \label{eq31}
h(\gamma)\leq h(\sigma \alpha)+h(\alpha^D)= (1+D)h(\alpha)\leq (1+2^p) h(\alpha).
\end{equation}
Let $n'\leq n$ be the least integer such that $\gamma\in K(p^{n'}M)$.
For brevity, put $N'=p^{n'}M$.
\underline{Case $n'\leq 1$}.
As $\gamma$ is neither $0$ nor a root of unity, we can apply Lemma \ref{lmm15} to $N=N'$ and $a=\gamma$.
Hence it follows from \eqref{eq31} that \[h(\alpha) \geq \frac{1}{2^p+1}\left(\frac{2\log p}{p^3(1+p^{p^3})(1+5\cdot 2^{11})}\right)^4.\]
Our theorem follows by noticing that $(1+2^p)(1+p^{p^3})^4(1+5\cdot 2^{11})^4\leq 16\cdot 10^{15}2^p p^{4p^3}$.
\underline{Case $n'\geq 2$}.
Recall that the restriction map $\mathrm{Gal}(L(N')/L)\to \mathrm{Gal}(K(N')/K)$ is an isomorphism.
Thus the minimality of $n'$, together with Lemma \ref{lmm18} applied to $N=N'$, proves that there is a place $w\vert p$ of $K(N')$ such that $\iota_w \gamma\notin K(N'/p)_w$.
We want to apply Lemma \ref{lmm16} to $N=N'$.
As $\gamma\notin\mu_\infty$ and $p^2-1$ divides $M$, it remains to show that $\iota_w \gamma^{p^2} \notin K(N'/p)_w$.
Assume by contradiction that it is the case.
Since $\iota_w\gamma \notin K(N'/p)_w$, there is $\psi_w\in\mathrm{Gal}(K(N')_w/K(N'/p)_w)$ such that $\psi_w\iota_w\gamma \neq \iota_w\gamma$.
Furthermore, $\psi_w\iota_w\gamma^{p^2}= \iota_w\gamma^{p^2}$ by assumption.
By Lemma \ref{lmm decompo} and injectivity of $\iota_w$, we conclude that there is $\psi\in \mathrm{Gal}(K(N')/K(N'/p))$ such that $\psi\gamma\neq\gamma$ and $\psi\gamma^{p^2}=\gamma^{p^2}$.
Thus $\psi\gamma=\zeta\gamma$ for some $\zeta\in\mu_{p^2}\backslash \{1\}$.
Put $\eta=(\psi \alpha)/\alpha\in K(N)^*$.
As $\psi$ and $\sigma$ commute by Lemma \ref{lmm17'}, it follows from \eqref{eq30} that
\begin{align*}
\zeta & = \frac{\psi\gamma}{\gamma} = \frac{\psi((\sigma\alpha)/\alpha^D)}{(\sigma\alpha)/\alpha^D} = \frac{\sigma\psi\alpha}{\psi\alpha^D} \frac{\alpha^D}{\sigma\alpha} = \frac{\sigma\eta}{\eta^D}.
\end{align*}
As $\zeta\in\mu_\infty$, it arises from Lemma \ref{lmm10.625} that $\eta\in\mu_\infty$.
Let $T\in\mathbb{N}$ be an integer coprime to $p$ such that $\eta^T$ has order a power of $p$.
Lemma \ref{lmm14} gives $\eta^T\in\mu_{p^n}$ and by Lemma \ref{lmm17'}, we have $\sigma\eta^T=(\eta^T)^D$.
We finally conclude $\zeta^{p^2}=\zeta^T=1$, and so $\zeta=1$ since $T$ and $p$ are coprime, a contradiction.
In conclusion, $\iota_w \gamma^{p^2} \notin K(N'/p)_w$.
Thus, combining \eqref{eq31} with Lemma \ref{lmm16} applied to $N=N'$ gives \[ \frac{(\log p)^4}{4\cdot 10^6 p^{32}}\leq h(\gamma)\leq (1+4^{p/2})h(\alpha). \]
The theorem in the case $n'\geq 2$ immediately follows. \qed
\bibliographystyle{plain}
|
\section{Introduction}
Let $G=(V,E)$ be a graph and let $[q]\coloneqq\{1,\ldots,q\}$ be a set of spins or colours for an integer $q\geq2$. A function $\sigma:V\to [q]$ is called a $q$-spin configuration or colouring.
The Gibbs measure of the $q$-state Potts model on $G=(V,E)$ is a probability distribution on the set of all $q$-spin configurations $\{\sigma:V\to [q]\}$.
For an \emph{interaction parameter} $w>0$, the Gibbs distribution $\mu_{\mathrm{Potts}}\coloneqq\mu_{\mathrm{Potts},G;q,w}$ is defined by
\begin{equation}\label{eq:prob pots}
\mu_{\mathrm{Potts}}[\sigma]\coloneqq\frac{w^{m(\sigma)}}{\sum_{\tau:V \to [q]} w^{m(\tau)}},
\end{equation}
where, for a given $q$-spin configuration $\tau$, $m(\tau)$ denotes the number of edges $\{u,v\}$ of $G$ for which $\tau(u)=\tau(v)$.
The denominator of the fraction~\eqref{eq:prob pots} is called the \emph{partition function of the Potts model} and is denoted by $Z_{\mathrm{Potts}}(G;q,w)$.
The regime $w\in(0,1)$ is known as the anti-ferromagnetic Potts model, and $w \in (1,\infty)$ as the ferromagnetic Potts model. Furthermore, values of $w$ close to 1 are referred to as high temperature, whereas values close to 0 or infinity are referred to as low temperature. This comes from the physical interpretation in which one writes $w= e^{J\beta}$ with $J>0$ being the interaction energy between same spin sites and $\beta$ the inverse temperature.
We will be concerned with the algorithmic problem of approximately sampling from $\mu_{\mathrm{Potts}}$ as well as approximately computing $Z =Z_{\mathrm{Potts}}(G;q,w)$ for $w$ close to infinity (that is in the low temperature ferromagnetic regime). Given error parameters $\epsilon,\delta \in (0, 1)$, an $\epsilon$-approximate counting
algorithm for $Z$ outputs a number $Z'$ so that $(1- \epsilon) \leq Z/Z' \leq (1 + \epsilon)$, and a $\delta$-approximate sampling
algorithm for $\mu=\mu_\mathrm{Potts}$ outputs a random sample $I$ with distribution $\hat{\mu}$ so that the total variation distance
satisfies $\|\mu-\hat{\mu}\|_{\mathrm{TV}}\leq \delta$.
It was shown in \cite{GSVY16} that, for graphs of a fixed maximum degree $\Delta \geq 3$, there is a critical parameter $w_\Delta>1$, corresponding to a phase transition of the model on the infinite $\Delta$-regular tree, such that approximating the partition function is computationally hard\footnote{Technically they showed that the problem is \#BIS hard, a complexity class introduced in~\cite{DGGJBIS} and known to be as hard as \#BIS, that is the problem of counting the number of independent sets in a bipartite graph. The exact complexity of \#BIS is unknown, but it is believed that no fully polynomial time randomised approximation scheme exists for \#BIS, but also that \#BIS is not \#P-hard}.
This result indicates that it might be hard to compute the partition function of the ferromagnetic Potts model for large values of $w$.
However, recently several results emerged, showing that for certain finite subgraphs of $\mathbb{Z}^d$~\cite{BR19,HPR20,BCHPT20} as well as $\Delta$-regular graphs satisfying certain expansion properties~\cite{JKP20,HJP20,CGGPSV21} it is in fact possible to approximate the partition function of the ferromagnetic Potts model for $w$ large enough. In fact the algorithms in~\cite{BCHPT20,HJP20} even work for all values $w\geq 1$ under the assumption that the number of colours, $q$, is suitably large in terms of the maximum degree.
The running times of all these aforementioned algorithms are polynomial in the number of vertices of the underlying graph, but typically with a large exponent. The exception is \cite{CGGPSV21}, in which the cluster expansion techniques of \cite{JKP20} for expander graphs are extended to a Markov chain setting giving running times of the form $O(n^2 \log n)$ for approximating the partition function, where $n$ is the number of vertices of the input graph.
In this paper we present Markov chain based algorithms for approximating the partition function of the ferromagnetic Potts model at sufficiently low temperatures with similar running times as \cite{CGGPSV21}. While most results in this area focus on graphs of bounded maximum degree, the graph parameters of interest for us are different and so our methods, as well as being able to handle subgraphs of the grid $\mathbb{Z}^d$ (although not for all temperatures), can also handle certain graphs classes of unbounded degree (cf. Lemma~\ref{lem:contraction is good}). The parameters of interest for us are in fact similar to those in \cite{BR19}; here we achieve better running times for our algorithms, while \cite{BR19} achieves better parameter dependencies.
We show how to efficiently generate a sample from the Potts model using a rapidly mixing Markov chain and then use this to approximate the partition function. The Markov chain however is not supported on $q$-spin configurations\footnote{See e.g.~\cite{BGP16} for an analysis of the usual Glauber dynamics for the ferromagnetic Potts model (at high temperatures).} but on flows taking values in $\mathbb Z_q\coloneqq\mathbb Z/q\mathbb Z$.
For planar graphs, this Markov chain on flows may be interpreted as Glauber dynamics of $q$-spin configurations on the dual graph; see Section~\ref{se:slow} for an example of this.
We use this Markov chain on flows together with another trick to show that we can efficiently approximate a certain partition function on flows at high temperatures, which in turn can be used to approximate the Potts partition function at low temperatures. Below we state our main results.
\subsection{Main results}
To state our main results, we need some definitions.
In the present paper we deal with multigraphs and the reader should read multigraph whenever the word graph is used.
A graph is called \emph{even} if all of its vertices have even degree. In what follows we often identify a subgraph of a given graph with its edge set.
Given a graph $G$, fix an arbitrary orientation of its edges. For any even subgraph $C$ of $G$, we can associate to it a signed indicator vector $\chi_C\in \mathbb{Z}^E$ as follows: choose an Eulerian orientation of (each of the components of) $C$. Then for $e\notin C$ we set $\chi_C(e)=0$ and for $e\in C$, we set $\chi_C(e)=1$ if $e$ has the same direction in both $C$ and $G$, and we set $\chi_C(e)=-1$ otherwise. We often abuse notation and identify the indicator vector $\chi_C$ with the set of edges in $C$.
A \emph{$\mathbb{Z}$-flow}, is a map $f:E\to \mathbb{Z}$ satisfying
\[
\sum_{e:\, e \text { directed into } v}f(e)=\sum_{e:\, e \text{ directed out of }v}f(e) \quad \text{ for all $v\in V$}.
\]
We denote the collection of $\mathbb{Z}$-flows by $\mathcal{F}(G)$; note that $\mathcal{F}(G)$ with the obvious notion of addition is known as the first homology group of $G$, and also as the cycle space of $G$.
Clearly, when viewing $\chi_C$ as a function on $E$, we have $\chi_C\in \mathcal{F}(G)$ for any even subgraph $C$.
It is well known that $\mathcal{F}(G)$ has a generating set (as a $\mathbb{Z}$-module) consisting of indicator vectors of even subgraphs; see e.g. \cite[Section 14]{GRbook}.\footnote{In fact there is even a basis consisting of indicator functions of cycles. For later purposes we however need to work with even subgraphs.}
We call such a generating set an \emph{even generating set} for the cycle space.
Let $\mathcal{C}$ be an even generating set of $\mathcal{F}(G)$; we define some parameters associated to $\mathcal{C}$ (see below for some examples of even generating sets and associated parameters).
For $C\in \mathcal{C}$, let $d(C)\coloneqq|\{D\in \mathcal{C}\setminus \{C\}\mid C\cap D\neq \emptyset\}|$, and let
\begin{equation}\label{eq:def d}
d(\mathcal{C})\coloneqq\max \{d(C)\mid C\in \mathcal{C}\}.
\end{equation}
We write
\begin{equation}\label{eq:def iota}
\iota(\mathcal{C})\coloneqq\max \{|C_1\cap C_2| \mid C_1,C_2\in \mathcal{C} \text{ with } C_1\neq C_2\}.
\end{equation}
Define
\begin{equation}\label{eq:def ell}
\ell(\mathcal{C})\coloneqq \max\{|C|\mid C\in \mathcal{C}\}.
\end{equation}
Finally, for an edge $e\in E$, define $s(e)$ to be the number of even subgraphs $C\in \mathcal{C}$ that $e$ is contained in and
\begin{equation}\label{eq:def s}
s\mathcal(\mathcal{C})\coloneqq\max\{s(e)\mid e\in E\}.
\end{equation}
We now present our approximate sampling and counting results. All of our results are based on randomised algorithms that arise from running Markov chains. For us, simulating one step of these Markov chains always includes choosing a random element from a set of $t$ elements with some (often uniform) probability distribution, where $t$ is at most polynomial in the size of the input graph. We take the time cost of such a random choice to be $O(1)$ as in the (unit-cost) RAM model of computation; see e.g.\ \cite{MotRag}.
Our main sampling results read as follows.
\begin{theorem}\label{thm:main sampling}
Fix a number of spins $q\in \mathbb{N}_{\geq 2}$.
\begin{itemize}
\item[(i)] Fix integers $d\geq 2$ and $\iota\geq 1$ and let $\mathcal{G}$ be the set of graphs $G=(V,E)$ for which we have an even generating set $\mathcal{C}$ for $G$ of size $O(|E|)$ such that $d(\mathcal{C}) \leq d$ and $\iota(\mathcal{C})\leq \iota$.
For any $w>\tfrac{(d+1)\iota }{2}q-(q-1)$ and $\delta \in (0,1)$,
there exists a $\delta$-approximate sampling algorithm for $\mu_{\mathrm{Potts},G;q,w}$, on all $m$-edge graphs $G \in \mathcal{G}$
with running time $O(m^2\log(m\delta^{-1}))$.
\item[(ii)]
Fix integers $\ell\geq 3$ and $s\geq 2$
and let $\mathcal{G}$ be the set of graphs $G=(V,E)$ for which we have an even generating set
$\mathcal{C}$ for $G$ of size $O(|E|)$ such that $\ell(\mathcal{C}) \leq \ell$ and $s(\mathcal{C})\leq s$.
For any $w>(q-1)(\ell s-1)$ and $\delta \in (0,1)$
there exists a $\delta$-approximate sampling algorithm for $\mu_{\mathrm{Potts},G;q,w}$ on all $m$-edge graphs $G \in \mathcal{G}$
with running time $O(m\log(m\delta^{-1}))$.
\end{itemize}
\end{theorem}
While parts (i) and (ii) are not directly comparable, we note that when $\iota=1$, part (i) has a better range for $w$.
Our main approximate counting results read as follows.
\begin{theorem}\label{thm:main counting}
Fix a number of spins $q\in \mathbb{N}_{\geq 2}$.
\begin{itemize}
\item[(i)]
Fix integers $d\geq 2$ and $\iota\geq 1$ and let $\mathcal{G}$ be the set of graphs $G=(V,E)$ for which we have an even generating set
$\mathcal{C}$ for $G$ of size $O(|E|)$ such that $d(\mathcal{C}) \leq d$ and $\iota(\mathcal{C})\leq \iota$. For $w>\tfrac{(d+1)\iota }{2}q-(q-1)$ and $\epsilon \in (0,1)$,
there exists a randomised $\epsilon$-approximate counting algorithm for $Z_{\mathrm{Potts}}(G;q,w)$ on all $n$-vertex and $m$-edge graphs $G \in \mathcal{G}$
that succeeds with probability at least $3/4$ and has
running time $O(n^2m^2\epsilon^{-2}\log(nm\epsilon^{-1}))$.
\item[(ii)]
Fix integers $\ell\geq 3$ and $s\geq 2$ and let $\mathcal{G}$ be the set of graphs $G=(V,E)$ for which we have an even generating set
$\mathcal{C}$ for $G$ of size $O(|E|)$ such that $\ell(\mathcal{C}) \leq \ell$ and $s(\mathcal{C})\leq s$.
For any $w>(q-1)(\ell s-1)$ and $\epsilon \in (0,1)$
there exists a randomised $\epsilon$-approximate counting algorithm for $Z_{\mathrm{Potts}}(G;q,w)$ on all $n$-vertex and $m$-edge graphs $G \in \mathcal{G}$
that succeeds with probability at least $3/4$ and has running time $O(n^2m \epsilon^{-2}\log(nm\epsilon^{-1}))$.
\end{itemize}
\end{theorem}
\begin{remark}
We note that the dependence of the Potts model parameter $w$ on the parameters $\ell(\mathcal{C})$ and $s(\mathcal{C})$ is similar as in~\cite{BR19}, except there the dependence on $s$ is order $\sqrt{s}$, which is better than our linear dependence.
This of course raises the question whether our analysis can be improved to get the same dependence.
\end{remark}
We now give a few examples of applications of our results.
\begin{example}
(i) Let $G_1$ and $G_2$ be two graphs that both contain a connected graph $H$ as induced subgraph. Let $G_1\cup_H G_2$ be the graph obtained from $G_1$ and $G_2$ by identifying the vertices of the graph $H$ in both graphs.
If both $G_1$ and $G_2$ have an even generating set consisting of cycles of length at most $\ell$ for some $\ell$, then the same holds for $G_1\cup_H G_2$.
Now use this procedure to build a subgraph $G=(V,E)$ of $\mathbb{Z}^d$, $d\geq 2$ from the union of finitely many copies of elementary cubes ((${\{0,1\}}^d$).
Since an an elementary cube has a generating set consisting of $4$-cycles, as is seen by induction on $d$, and so the resulting graphs has an even generating set $\mathcal{C}$ consisting only of $4$-cycles. The relevant parameters of $\mathcal{C}$ are $d(\mathcal{C})=8(d-1)-4$, $\iota(\mathcal{C})=1$, $\ell(\mathcal{C})=4$, $s(\mathcal{C})=2(d-1)$, and $|\mathcal{C}|\leq (d-1)|E|/2$.
(ii) In a similar manner as in (i) one can also construct graphs with concrete parameters from lattices such as the triangular lattices (here $d=3$, $\iota=1$, $\ell=3$ and $s=2$) or its dual lattice, the honeycomb lattice (here $d=6$, $\iota=1$, $\ell=6$, and $s=2$).
(iii) For any (multi)graph $G=(V,E)$ with even generating set $\mathcal{C}$, the graph $G/e$ obtained by contracting some edge $e \in E$ has an even generating set $\mathcal{C}/e \coloneqq \{C/e: C \in \mathcal{C} \}$. One can check that $\mathcal{C}$ and $\mathcal{C}/e$ have the same parameters $d, \iota, \ell, s$ (see Lemma~\ref{lem:contraction is good}). This allows us to apply our algorithms to many graph classes of unbounded degree e.g.\ any graph that can be obtained from $\mathbb{Z}^d$ by a series of contractions.
\end{example}
As we shall see in the next subsection, the Markov chains on flows that we introduce are a natural means of studying the ferromagnetic Potts model at low temperatures. The examples above show that it is easy to generate many graphs (also of unbounded degree) for which these chains mix rapidly and therefore for which our results above apply. In this paper, we begin the analysis of these Markov chains on flows, but we believe there is a lot of scope for further study of these chains to obtain better sampling and counting algorithms for the ferromagnetic Potts model at low temperature.
\subsection{Approach and discussion}
The key step in our proof of Theorems~\ref{thm:main sampling} and ~\ref{thm:main counting} is to view the partition function of the Potts model as a generating function of flows taking values in an abelian group of order $q$.
Although well known to those acquainted with the Tutte polynomial and its many specializations, this perspective has not been exploited
in the sampling/counting literature (for $q\geq 3$) to the best of our knowledge.
For the special case of the Ising model, that is, $q=2$, this perspective is known as the even `subgraphs world' and has been key in determining an efficient sampling/counting algorithm for the Ising model (with external field) by Jerrum and Sinclair~\cite{JSising}, although the Markov chain used there is defined on the collection of all subsets of the edge set $E$ rather than on just the even sets.
We however define a Markov chain on a state space which, for $q=2$, is supported only on the even sets.
For $q=2$ one could interpret our Markov chain as Glauber dynamics with respect to a fixed basis of the space of even sets (which forms a vector space over $\mathbb{F}_2$), that is, we move from one even subgraph to another by adding/subtracting elements from the basis.
In the general case $(q\geq 3)$ the even subgraphs need to be replaced by flows, but, aside from some technical details, our approach remains the same.
We analyse the Markov chain using the well known method of path coupling~\cite{BD97,Jbook} to obtain our first sampling result Theorem~\ref{thm:main sampling}(i), and
the proof of Theorem~\ref{thm:main counting}(i) then follows by standard arguments after a suitable self-reducibility trick.
Another well known way of representing the partition function of the Potts model is via the random cluster model.
Only recently, it was shown that a natural Markov chain called random cluster dynamics is rapidly mixing for the Ising model~\cite{GJ18}, yielding another way of obtaining approximation algorithms for the partition function of the Ising model.
In the analysis a coupling due to Grimmet and Jansson\cite{GM09} between the random cluster model and the even subgraphs world was used.
We extend this coupling to the level of flows and we analyse the Glauber dynamics on the joint space of flows and clusters to obtain a proof of part (ii) of Theorems~\ref{thm:main sampling} and ~\ref{thm:main counting}.
\paragraph{Organization}
In the next section we introduce the notion of flows and the flow partition function, showing the connection to the Potts model and the random cluster model. We also give some preliminaries on Markov chains. In Section~\ref{se:flowchain}, we introduce and analyse the flow chain and prove Theorem~\ref{thm:main sampling}(i). In Section~\ref{se:jointchain} we introduce and analyse the joint flow-random cluster Markov chain, which allows us to prove Theorem~\ref{thm:main sampling}(ii). In Section~\ref{se:count} we examine the subtleties involved in showing that our sampling algorithms imply corresponding counting algorithms: we deduce Theorem~\ref{thm:main counting} from Theorem~\ref{thm:main sampling} in this section. Finally in Section~\ref{se:slow}, we use the duality between flows and Potts configurations to deduce a slow mixing result for our flow chain (on $\mathbb{Z}^2$) from existing results about slow mixing for the Potts model.
\section{Preliminaries}
\subsection{The flow partition function}
Let $G=(V,E)$ be a graph. Throughout, if it is unambiguous, we will take $n\coloneqq|V|$ and $m\coloneqq|E|$.
In order to define a flow on $G$, we first orient the edges of $G$. (We will assume from now on that the edges of graphs have been given a fixed orientation even if this is not explicitly stated).
For an abelian group $\Gamma$, a $\Gamma$-flow (on $G$) is an assignment $f: E \rightarrow \Gamma$ of a value of $\Gamma$ to every edge of $G$ such that, for every vertex, the sum (in $\Gamma$) at the incoming edges is the same as the sum (in $\Gamma$) at the outgoing edges.
For a positive integer $q$, the \emph{flow partition function} is defined as
\[
Z_\mathrm{flow}(G;q,x) = \sum_{f:E \to \mathbb{Z}_q \text{ flow}} x^{\#\text{non-zero edges in }f}.
\]
Note that $Z_\mathrm{flow}$ only depends on the underlying graph and not on the orientation of $G$.
It is moreover well known that in the definition of the partition function we can replace the group $\mathbb{Z}_q$ by any abelian group $\Gamma$ of order $q$, without changing the partition function.
We will however make no use of this and solely work with the group $\mathbb{Z}_q$.
Recall from the introduction that $\mathcal{F}(G)$ denotes the set of $\mathbb Z$-flows. We write $\mathcal{F}_q(G)$ for the set of $\mathbb Z_q$-flows (namely the set of all flows $f:E\to \mathbb{Z}_q$), and for $F\subseteq E$ we denote by $\mathcal{F}_q(V,F)$ the set of all flows $f:F\to \mathbb{Z}_q$.
The \emph{support} of a flow $f$ is the collection of edges that receive a nonzero flow value and is denoted by $\supp(f)$.
We denote by $\mathrm{nwz}(F;q)$ the number of flows $f:F\to \mathbb{Z}_q$ such that $\supp(f)=F$ (where nwz stands for nowhere zero).
Finally, for positive $x$, there is a natural probability measure $\mu_{\mathrm{flow}}$ on $\mathcal{F}_q(V,E)$, defined by
\begin{equation}
\mu_{\mathrm{flow}}(f)\coloneqq\frac{x^{|\supp(f)|}}{Z_{\mathrm{flow}}(G;q,x)}
\end{equation}
for each $f \in \mathcal{F}_q(V,E)$.
The following fact is well known and goes back to Tutte~\cite{Tutte54}.
\begin{lemma}\label{lem:tutte}
Let $q\in \mathbb{N}_{\geq 1}$ and let $x\in \mathbb{C}\setminus\{1\}$. Let $G=(V,E)$ be a graph. Then
\begin{equation}
q^{|V|} Z_\mathrm{flow}(G;q,x) = (1-x)^{|E|} Z_\mathrm{Potts}\left(G;q,\frac{1+(q-1)x}{1-x}\right).
\end{equation}
\end{lemma}
This lemma follows by combining \eqref{eq:pf Potts=pf RC} and \eqref{eq:lem: pf flow=pf RC} below: it illustrates a useful coupling between random flows and the random cluster model.
We remark that the function $x\mapsto \frac{1+(q-1)x}{1-x}$ (seen as a function from $\mathbb C\cup \{\infty\}\to \mathbb C\cup \{\infty\}$) has the property that it sends $0$ to $1$, $1$ to $\infty$ and the interval $[0,1]$ to $[1,\infty]$ in an orientation preserving way.
So approximating the partition function of the $q$-state ferromagnetic Potts model at low temperatures ($1\ll w$) is equivalent to approximating the flow partition function for values $x\in (0,1)$ close to $1$.
\subsection{The random cluster model and a useful coupling}
We view the \emph{partition function of the random cluster model} for a fixed positive integer $q$ as a polynomial in a variable $y$. It is defined for a graph $G=(V,E)$ as follows:
\begin{equation}
Z_{\mathrm{RC}}(G;q,y)\coloneqq\sum_{F\subseteq E} q^{c(F)}y^{|F|},
\end{equation}
where $c(F)$ denotes the number of components of the graph $(V,F)$.
For $y\geq 0$, we denote the associated probability distribution on the collection of subsets of the edges $\{F\mid F\subseteq E\}$ by $\mu_{\mathrm{RC}}$, i.e., for $F\subseteq E$ we have
\[
\mu_{\mathrm{RC}}(F)=\frac{q^{c(F)}y^{|F|}}{Z_{\mathrm{RC}}(G;q,y)}.
\]
It is well known, see e.g.~\cite{ES88}, that
\begin{equation}
Z_{\mathrm{Potts}}(G;q,w)=Z_{\mathrm{RC}}(G;q,w-1).
\label{eq:pf Potts=pf RC}
\end{equation}
To describe the connection between $Z_\mathrm{RC}$ and $Z_\mathrm{flow}$ and a coupling between the associated probability distributions, it will be useful to consider the following partition function for a graph $G=(V,E)$:
\begin{equation}
Z(G;q,x)\coloneqq(1-x)^{|E|}q^{|V|}\sum_{A\subseteq E}\mathrm{nwz}(A;q,x)\sum_{\substack{F\subseteq E \\ A\subseteq F}}\left(\frac{x}{1-x}\right)^{|F|}. \label{eq:pf general}
\end{equation}
The associated probability distribution $\mu$ is on pairs $(f,F)$ such that $f$ is a $\mathbb{Z}_q$-flow on $G$ with $\supp(f)\subseteq F$.
By~\eqref{eq:pf Potts=pf RC}, the next lemma directly implies Lemma~\ref{lem:tutte}; the lemma and the coupling it implies extend~\cite{GM09}.
\begin{lemma}\label{lem: pf flow=pf RC}
Let $q\in \mathbb{N}_{\geq 1}$ and let $x\in \mathbb{C}\setminus\{1\}$. Let $G=(V,E)$ be a graph.
\begin{equation}
q^{|V|}Z_{\mathrm{flow}}(G;q,x)=Z(G;q,x)=(1-x)^{|E|}Z_{\mathrm{RC}}(G;q,\tfrac{qx}{1-x}). \label{eq:lem: pf flow=pf RC}
\end{equation}
\end{lemma}
\begin{proof}
The first equality follows by the following sequence of identities:
\begin{align*}
&(1-x)^{|E|}q^{|V|}\sum_{A\subseteq E}\mathrm{nwz}(A;q)\sum_{\substack{F\subseteq E \\ A\subseteq F}}\left(\frac{x}{1-x}\right)^{|F|}=q^{|V|}\sum_{A\subseteq E}\mathrm{nwz}(A;q)\sum_{\substack{F\subseteq E \\ A\subseteq F}}x^{|F|}(1-x)^{|E\setminus F|}
\\
&=q^{|V|}\sum_{A\subseteq E}\mathrm{nwz}(A;q)x^{|A|}\sum_{\substack{F\subseteq E \\ A\subseteq F}}x^{|F\setminus A|}(1-x)^{|E\setminus F|}=q^{|V|}\sum_{A\subseteq E}\mathrm{nwz}(A;q)x^{|A|}=q^{|V|}Z_{\mathrm{flow}}(G;q,x).
\end{align*}
For the second equality we use the well known fact that $|\mathcal{F}_q(V,F)|$ (the number of all flows on the graph $(V,F)$ taking values in an abelian group of order $q$), satisfies
\begin{equation}\label{eq:flow count}
|\mathcal{F}_q(V,F)| = Z_\mathrm{flow}((V,F);q,1)=q^{|F|-|V|+c(F)}.
\end{equation}
To see it, note first that we may assume $(V,F)$ is connected since both sides of the identity are multiplicative over components.
Fix a spanning tree $T\subseteq F$ and assign values from $\mathbb{Z}_q$ to $F\setminus T$. It is not hard to see that these values can be uniquely completed to a flow by iteratively `removing' a leaf from $T$.
We then have the following chain of equalities:
\begin{align*}
&(1-x)^{|E|}q^{|V|}\sum_{A\subseteq E}\mathrm{nwz}(A;q)\sum_{\substack{F\subseteq E \\ A\subseteq F}}\left(\frac{x}{1-x}\right)^{|F|}=(1-x)^{|E|}\sum_{F\subseteq E}\left(\frac{x}{1-x}\right)^{|F|}q^{|V|}\sum_{A\subseteq F}\mathrm{nwz}(A;q)
\\
&=(1-x)^{|E|}\sum_{F\subseteq E}\left(\frac{x}{1-x}\right)^{|F|}q^{|V|}
|\mathcal{F}_q(V,F)|
=(1-x)^{|E|}\sum_{F\subseteq E}\left(\frac{x}{1-x}\right)^{|F|}q^{|F|+c(F)}
\\
&=(1-x)^{|E|}Z_{\mathrm{RC}}(G;q,\tfrac{qx}{1-x}).
\end{align*}
\end{proof}
The previous lemma in fact gives a coupling between the probability measures $\mu_{\mathrm{flow}}$ and $\mu_{\mathrm{RC}}$ (with the same parameters as in the lemma).
More concretely, given a random flow $f$ drawn from $\mu_{\mathrm{flow}}$ let $A$ be the support of $f$. Next select each edge $e\in E\setminus A$ independently with probability $x$.
The resulting set $F$ is then a sample drawn from $\mu_{\mathrm{RC}}$. To see this, observe that the probability of selecting the set $F$ is given by
\begin{align*}
\sum_{A\subseteq F}\frac{\mathrm{nwz}(A;q)x^{|A|}}{Z_{\mathrm{flow}}(G;q,x)}x^{|F\setminus A|}(1-x)^{|E\setminus F|}
=\frac{|\mathcal{F}_q(V,F)|}{Z_{\mathrm{flow}}(G;q,x)}x^{|F|}(1-x)^{|E\setminus F|}
=\mu_{\mathrm{RC}}(F),
\end{align*}
where the last equality follows by the lemma above and~\eqref{eq:flow count} and the definition of $\mu_{\mathrm{RC}}$.
Conversely (by a similar calculation), given a sample $F$ drawn from $\mu_{\mathrm{RC}}$ one can obtain a random flow drawn from $\mu_{\mathrm{flow}}$ by choosing a uniform flow on $(V,F)$.
For any $\delta>0$, this procedure transforms a $\delta$-approximate sampler $\hat{\mu_{\mathrm{flow}}}$ for $\mu_{\mathrm{flow}}$ with parameters $q$ and $x\in (0,1)$ into a $\delta$-approximate sampler $\hat{\mu_{\mathrm{RC}}}$ for $\mu_{\mathrm{RC}}$ with parameters $q, \tfrac{qx}{1-x}$ in time bounded by $O(|E|)$.
Indeed, denoting for a flow $f$, $\delta_{\mathrm{flow}}(f)\coloneqq\hat{\mu_{\mathrm{flow}}}(f)-\mu_{\mathrm{flow}}(f)$, we have by the triangle inequality
\begin{align*}
\sum_{F\subseteq E} |\hat{\mu_\mathrm{RC}}(F) - \mu_\mathrm{RC}(F)|
&= \sum_{F\subseteq E} \left| \sum_{f\in \mathcal{F}_q(V,F)} \delta_\mathrm{flow}(f) x^{|F\setminus \supp(f)|} (1-x)^{|E\setminus F|}\right| \\
& \leq \sum_{F\subseteq E} \sum_{f\in \mathcal{F}_q(V,F)} |\delta_\mathrm{flow}(f)| x^{|F\setminus \supp(f)|} (1-x)^{|E\setminus F|}\\
&= \sum_{f\in\mathcal{F}_q(V,F) } |\delta_\mathrm{flow}(f)| \sum_{F\supseteq \supp(f)}x^{|F\setminus \supp(f)|} (1-x)^{|E\setminus F|} \\
&=\sum_{f \in \mathcal{F}_q(V,F) } |\delta_\mathrm{flow}(f)| \leq 2\delta.
\end{align*}
The Edwards-Sokal coupling~\cite{ES88} allows us to generate a sample from the Potts model (with parameters $q, w+1$), given a sample $F$ from the random cluster model (with parameters $q, w$): for each component of $(V,F)$ uniformly and independently choose a colour $i\in [q]$ and colour each of the vertices in this component with this colour.
Again if we have a $\delta$-approximate sampler $\hat{\mu_{\mathrm{RC}}}$ for $\mu_{\mathrm{RC}}$ this will be transformed into a $\delta$-approximate sampler $\hat{\mu_{\mathrm{Potts}}}$ for $\mu_{\mathrm{Potts}}$ in time bounded by $O(|E|)$.
We summarize the discussion above in a proposition.
\begin{prop} \label{prop:flow sampler to RC/Potts}
Let $G=(V,E)$ be a graph and let $q\in \mathbb{N}_{\geq 2}$ and $x>0$.
Let $\delta>0$.
Given an approximate $\delta$-approximate sampler $\hat{\mu_{\mathrm{flow}}}$ for $\mu_{\mathrm{flow}}$ with parameters $q$ and $x\in (0,1)$, we can obtain $\delta$-approximate approximate samplers from
\begin{itemize}
\item $\mu_{\mathrm{RC}}$ with parameters $q$ and $\tfrac{qx}{1-x}$ in time $O(|E|)$,
\item $\mu_{\mathrm{Potts}}$ with parameters $q$ and $\tfrac{1+(q-1)x}{1-x}$ in time $O(|E|)$.
\end{itemize}
\end{prop}
\subsection{Generating sets and bases of flows}
In this subsection we give some useful properties of the set of flows and their even generating sets that will allow us to define Markov chains for sampling from $\mu_{\mathrm{flow}}$ in the next section.
In particular we show that an even generating set for the cycle space also generates the collection of $\mathbb{Z}_q$-flows in an appropriate sense to be made precise below.
Let $G=(V,E)$ be a connected graph and recall (from the Introduction) that $\mathcal{F}(G)$ is the set of $\mathbb{Z}$-flows on $G$ and let $\mathcal{C}$ be an even generating set of $\mathcal{F}(G)$. We already mentioned that $\mathcal{F}(G)$ forms a $\mathbb{Z}$-module; in fact it is a free-module of dimension $|E|-|V|+1$, cf.~\cite[Section 14]{GRbook}.
Similarly, the collection of $\mathbb Z_q$-flows on $G$ is closed under adding two flows and multiplying a flow by an element of $\mathbb Z_q$, making the space of $\mathbb Z_q$-flows into a $\mathbb Z_q$-module; it is also a free module of dimension $|E|-|V|+1$ by the same argument as for $\mathbb{Z}$, cf.~\cite[Section 14]{GRbook}. (Note that this fact also implies~\eqref{eq:flow count}.)
\begin{lemma}\label{lem:generating set for all}
Let $\mathcal{C}$ be an even generating set for $\mathcal{F}(G)$. Then $\mathcal{C}$ is a generating set for $\mathcal{F}_q(G)$ for any positive integer $q$.
\end{lemma}
\begin{proof}
Let $f\in\mathcal{F}_q(G)$ be a flow, we will construct a $\mathbb{Z}$-flow $f'$ which reduces modulo $q$ to $f$. Just as in the proof of Lemma~\ref{lem: pf flow=pf RC}, fix a spanning tree $T\subset E$, and now assign to every edge $e\in E\setminus T$ an integer from the residue class $f(e)$. These assignments can be completed iteratively into the flow $f'$ by choosing the edge towards a leaf, assigning a value to satisfy the flow condition in the leaf, and removing the edge from $T$. These new values are also in the residue class prescribed by $f$, because $f$ itself satisfies the flow condition in every leaf encountered.
Writing $f'$ as a linear combination of $\chi_C$ for $C\in\mathcal{C}$ and reducing modulo $q$, we obtain $f$ as a $\mathbb{Z}_q$-linear combination of $\chi_C$.
\end{proof}
Finally, we will require the following lemma
for our reduction of sampling to counting in Section~\ref{se:count}.
For a graph $G=(V,E)$, a subgraph $H$ of $G$ and an edge $e\in E$, $H/e$ denotes the graph obtained from $H$ by contracting the edge $e$. (If $e$ is not an edge of $H$, then $H/e$ is just $H$.)
\begin{lemma}\label{lem:contraction is good}
Let $G=(V,E)$ be a graph and let $q\in \mathbb{N}$. Let $\mathcal{C}=\{C_1,\ldots,C_r\}$ be an even generating set for the space of $\mathbb{Z}_q$-flows.
Let $e\in E$ be a non-loop edge.
Then $\mathcal{C}'\coloneqq\{C_1/e,\ldots,C_r/e\}$ is an even generating set for the space of $\mathbb{Z}_q$-flows of the graph $G/e$ satisfying $d(\mathcal{C}')\leq d((\mathcal{C})$, $\iota(\mathcal{C}')\leq \iota(\mathcal{C})$, $\ell(\mathcal{C}')\leq \ell(\mathcal{C})$ and $s(\mathcal{C}')\leq s(\mathcal{C})$.
\end{lemma}
\begin{proof}
This follows from the fact that any flow $f'$ on $G/e$ uniquely corresponds to a flow $f$ on $G$. The value on the edge $e$ for $f$ can be read off from the values of the edges incident to the vertex in $G/e$ corresponding to the two endpoints of the edge $e$.
So, writing $f=\sum_{i=1}^r a_i\chi_{C_i}$ for certain $a_i\in \mathbb{Z}_q$, we get $f'=\sum_{i=1}^r a_i\chi_{C_i/e}$, proving the claim.
The claimed inequalities for the parameters are clear.
\end{proof}
\begin{remark}
Suppose $q$ is a prime in which case $\mathbb{Z}_q$ is a field and $\mathcal{F}_q(G)$ is a vector space over $\mathbb Z_q$.
Then given an even generating set $\mathcal{C}$ for $\mathcal{F}_q(G)$ there exists a basis $\mathcal{C}'$ consisting only of cycles for which the parameters $d,\iota,\ell$ and $s$ are all not worse.
To see this note that if $\mathcal{C}$ is a generating set and not a basis, we can always remove elements from it to make it into a basis.
If $\mathcal{C}$ forms a basis and some $C\in \mathcal{C}$ is the edge disjoint union of two nonempty even subgraphs $K_1$ and $K_2$, we have that either $(\mathcal{C}\setminus \{C\})\cup \{K_1\}$ or $(\mathcal{C}\setminus \{C\})\cup \{K_2\}$ forms a basis.
This is generally not true for composite $q$ and therefore we work with even generating sets.
\end{remark}
\subsection{Preliminaries on Markov chains}
To analyse the mixing time of our Markov chains, we will use the path coupling technique. We briefly recall the following results from Section 2 in \cite{DG98}.
Let $\mathcal{M} = (Z_t)_{t=0}^\infty$ be an ergodic, discrete-time Markov chain on a finite state space $\Omega$ with transition matrix $P$. Let $\mu_t$ be the distribution of $Z_t$ and let $\mu$ be the (unique) stationary distribution of $\mathcal{M}$.
Two distributions on $\Omega$ are said to be \emph{$\delta$-close} if the total variation distance between them is at most $\delta$. The \emph{$\delta$-mixing time} of $\mathcal{M}$ is the minimum number of steps after which $\mathcal{M}$ is $\delta$-close to its stationary distribution (i.e.\ the smallest $t$ such that $\|\mu_t - \mu \|_{\mathrm{TV}} \leq \delta$).
A \emph{coupling} for $\mathcal{M}$ is a stochastic process $(X_t,Y_t)$ on $\Omega^2$, such that each of $X_t$ and $Y_t$, considered independently, transition according to $P$. More precisely, the coupling can be defined by its transition matrix $P'$: given $(x, y)$ and $(x',y') \in \Omega^2$, $P'((x,y),(x',y'))$ is the probability that $(X_{t+1}, Y_{t+1}) = (x',y')$ given that $(X_t,Y_t) = (x,y)$. For $P'$ to describe a valid coupling, it must satisfy for each $(x,y) \in \Omega^2$, that
\begin{align}
\sum_{y' \in \Omega}P'((x,y),(x',y')) &= P(x,x') \quad \text{ for all } x' \in \Omega; \nonumber \\
\sum_{x' \in \Omega}P'((x,y),(x',y')) &= P(y,y') \quad \text{ for all } y' \in \Omega .
\label{eq:coupling}
\end{align}
For our use of path coupling, we require an integer-valued distance function $d$ on $\Omega$ such that between any two states $x,y\in\Omega$ there exists a sequence $x=x_0,x_1,\dots,x_s=y$ in which consecutive states are at distance $1$.
If we can define a coupling on the set of pairs $(x,y) \in \Omega^2$ for which $d(x,y) =1$. (that is, we define transition probabilities $P'((x,y),(x',y'))$ for all $(x,y)$ such that $d(x,y)=1$, and $(x',y')\in \Omega^2$ that satisfy equations \eqref{eq:coupling}) then this can be extended to a complete coupling on $\Omega^2$. We can use such a (partial) coupling to bound the mixing time of $\mathcal{M}$ via the following result:
\begin{theorem}[Theorem 2.2 in \cite{DG98}]\label{thm:path coupling}
Let $\mathcal{M}$ be a Markov chain on $\Omega$ and $d$ an integer-valued distance on $\Omega$ as above with maximum distance $D$. Assume there is a coupling $(X_t,Y_t) \mapsto (X_{t+1},Y_{t+1})$ defined for all pairs with $d(X_t,Y_t)=1$ (as described above) such that
\[
\mathbb{E}(d(X_{t+1},Y_{t+1}) \mid (X_t,Y_t)) \leq 1-\alpha
\]
for some $\alpha>0$. Then the Markov chain $\mathcal{M}$ has $\delta$-mixing time at most $\frac{\log(D\delta^{-1})}{\alpha}$.
\end{theorem}
\section{Flow Markov chain}
\label{se:flowchain}
In this section, we introduce and analyse the flow Markov chain and use it to prove Theorem~\ref{thm:main sampling}(i).
\begin{definition}
Let $G=(V,E)$ be a graph and $\mathcal{C}$ an even generating set of $\mathcal{F}_q(V,E)$ of size $r$.
The \emph{flow Markov chain} for $(G,\mathcal{C})$ is a Markov chain on the state space $\mathcal{F}_q(V,E)$. For every flow $f\in\mathcal{F}_q(V,E)$, $t\in\mathbb{Z}_q\setminus\{0\}$ and $C\in\mathcal{C}$, the transition probabilities of the Markov chain are given by:
\begin{align*}
P_\mathrm{flow}(f,f+t\chi_C) &=\frac{1}{r} \frac{\mu_\mathrm{flow}(f+t\chi_C)}{\sum_{u\in\mathbb Z_q} \mu_\mathrm{flow}(f+u\chi_C)},\\
P_\mathrm{flow}(f,f) &= \frac{1}{r}\sum_{C\in\mathcal{C}}\frac{\mu_\mathrm{flow}(f)}{\sum_{u\in\mathbb Z_q} \mu_\mathrm{flow}(f+u\chi_C)},
\end{align*}
and all other transition probabilities are zero.
\end{definition}
We see easily that the measure $\mu_\mathrm{flow}$ satisfies the detailed balance equation \[\mu_\mathrm{flow}(f) P_\mathrm{flow}(f,f+t\chi_C) = \mu_\mathrm{flow}(f+t\chi_C) P_\mathrm{flow}(f+t\chi_C,f),\] so $\mu_\mathrm{flow}$ is the stationary distribution of the flow Markov chain.
We can simulate one step of this Markov efficiently by first selecting $C\in\mathcal{C}$ uniformly at random, and for $t\in\mathbb Z_q$, selecting $f+t\chi_C$ with probability proportional to
\[ \mu_\mathrm{flow}(f+t\chi_C)/\mu_\mathrm{flow}(f) = x^{\#\{e\in C \mid f(e)=0\} - \#\{e\in C \mid f(e)+t\chi_C(e)=0\}}.
\]
For fixed $q$, simulating one step of the Markov chain requires $O(\ell)$ time (where $\ell = \max_{C \in \mathcal{C}}|C|$) in order to compute $f + t \chi_C$ and its support. We bound this by $O(m)$.
\subsection{Rapid mixing of flow Markov chain}
\begin{theorem}\label{thm:mixing time flow MC}
Let $q,d\geq2, \iota\geq1$ be integers and $1>x>1-\frac{2}{(d+1)\iota}$.
Write $\xi=x-\left(1-\frac{2}{(d+1)\iota}\right)$ and let $\delta>0$.
Now let $G=(V,E)$ be a graph and $\mathcal{C}$ an even generating set of $\mathcal{F}_q(G)$ of size $r$ satisfying $d(\mathcal{C})\leq d$ and $\iota(\mathcal{C})\leq \iota$,
then the $\delta$-mixing time of the flow Markov chain for $(G,\mathcal{C})$ with parameter $x$ is at most $\frac{4r}{d\iota}\log(r \delta^{-1})\xi^{-1}$.
\end{theorem}
\begin{remark}
\label{re:runtime}
Because $\xi <\frac{2}{(d+1)\iota}\leq\frac{2}{d\iota}$, the upper bound in this Theorem is always at least $2r\log(r\delta^{-1})$. This shows the upper bound doesn't get better with larger $d$ and $\iota$, even though they are in the denominator.
\end{remark}
For the given range of $x$, the flow Markov chain therefore gives an efficient, randomised algorithm for approximately sampling flows according to $\mu_\mathrm{flow}$.
Combining this with \cref{prop:flow sampler to RC/Potts}, we obtain the following Corollary; it directly implies \cref{thm:main sampling}(i) by \cref{lem:generating set for all}.
\begin{cor}
Fix integers $q,d\geq 2$ and $\iota\geq 1$. For any $w>\frac{(d+1)\iota}{2}q-(q-1)$ and $\delta>0$, there exists an algorithm that on input of an $m$-edge graph $G$ and even generating set $\mathcal{C}$ of $\mathcal{F}_q(G)$ of size $r$ satisfying $d(\mathcal{C})\leq d$ and $\iota(\mathcal{C})\leq \iota$ outputs a $q$-state Potts colouring $\sigma:V\to [ q ] $ within total variation distance $\delta$ of the $q$-state Potts-measure $\mu_\mathrm{Potts}$ with parameter $w$.
This is obtained by running the flow Markov chain for at most $O(r\log(r \delta^{-1}))$ steps where each step takes $O(m)$ time.
\end{cor}
The following technical lemma will be used in the proof of \cref{thm:mixing time flow MC}. Note that the lower bound is actually attained in the limit case $(a_1,\ldots,a_q)=(\iota,0,-\infty,\ldots,-\infty), (b_1,\ldots,b_q)=(0,\iota,-\infty,\ldots,-\infty)$. The proof is postponed to the end of this section.
\begin{lemma}\label{lem:sum of minima}
Let $x\in(0,1)$ be a real number, and $\iota\geq0$ and $a_1,\ldots,a_q,b_1,\ldots,b_q$ integers satisfying the following constraints:
\begin{itemize}
\item $\sum_i a_i = \sum_i b_i$;
\item $\sum_i |a_i-b_i| \leq 2\iota$.
\end{itemize}
Then
\[
S \coloneqq \sum_i \min\left(\frac{x^{-a_i}}{\sum_j x^{-a_j}},\frac{x^{-b_i}}{\sum_j x^{-b_j}}\right) \geq 1- \frac{1-x^\iota}{1+x^\iota}.
\]
\end{lemma}
\begin{proof}[Proof of \cref{thm:mixing time flow MC}]
To prove the theorem we determine an upper bound for the mixing time of the flow Markov chain by using path coupling. For this we define the distance between two flows as the minimal number of steps the flow Markov chain needs to go from one to the other. By \cref{thm:path coupling} it is now enough to define a coupling for states at distance 1. If the expected distance after one step of this coupling is at most $1-\alpha$, the mixing time of the Markov chain is at most $T \coloneqq \frac{\log(r\delta^{-1})}{\alpha}$. (The maximal distance in $\mathcal{F}_q(V,E)$ is at most $r$, because in $r$ steps the coefficients of every even set in $\mathcal{C}$ can be adjusted to the desired value.)
We will construct a coupling on states at distance $1$ for which $\alpha=\frac{(d+1)x^{\iota}-(d-1)}{2r} \geq \frac{d\iota}{4r}\xi$. Therefore the running time of the sampler is bounded by $T\leq \frac{4r}{d\iota}\log(r\delta^{-1})\xi^{-1}$ steps of the flow Markov chain.
Consider a pair of flows $(f,g)$ which differ by a multiple of $\chi_C$. To construct the coupling we first select u.a.r.\@ an even set $D\in\mathcal{C}$. We will separate three cases, and define the transition probabilities in each of these cases. The cases are (a) when $C=D$, (b) when $C$ and $D$ have no common edges, and (c) when $C$ and $D$ do have common edges.
\begin{enumerate}[(a)]
\item We get a valid coupling by making the transition $(f,g) \to (f+t\chi_D,f+t\chi_D)$ with probability $\frac{\mu_\mathrm{flow}(f+t\chi_D)}{\sum_{u\in\mathbb Z_q} \mu_\mathrm{flow}(f+u\chi_D)}$. Then the distance will always drop from 1 to 0.
\item Now the edges of $D$ have the same values in $f$ and $g$, and we see that $\mu_\mathrm{flow}(f+t\chi_D)/\mu_\mathrm{flow}(f) = \mu_\mathrm{flow}(g+t\chi_D)/\mu_\mathrm{flow}(g)$ for all $t$. Therefore we get a valid coupling by making the transition $(f,g) \to (f+t\chi_D , g+t\chi_D)$ with probability $\frac{\mu_\mathrm{flow}(f+t\chi_D)}{\sum_{u\in\mathbb Z_q} \mu_\mathrm{flow}(f+u\chi_D)}=\frac{\mu_\mathrm{flow}(g+t\chi_D)}{\sum_{u\in\mathbb Z_q} \mu_\mathrm{flow}(g+u\chi_D)}$. In this case the distance between the two new states remains 1.
\item The coupling in this case is more complicated, as the values of $f$ and $g$ on $D$ are different. Below we prove the following:
\begin{claim}
There is a coupling where the total probability for all transitions $(f,g) \to (f+t\chi_D,g+t\chi_D)$ is at least $1-\frac{1-x^\iota}{1+x^\iota}$.
\end{claim}
In all these transitions the distance remains 1, and therefore the probability of the distance increasing to 2 is at most $\frac{1-x^\iota}{1+x^\iota}$.
\end{enumerate}
We can now calculate the expected distance after one step of this coupling. Case (a) occurs with probability $1/r$, and case (c) with probability at most $d/r$. Hence the expected distance is at most
\begin{align*}
&1-\frac{1}{r} + \frac{d}{r}\cdot\frac{1-x^\iota}{1+x^\iota}\\
&=1- \frac{1+x^\iota-d(1-x^\iota)}{r(1+x^\iota)}\\
&=1- \frac{(d+1)x^\iota-(d-1)}{r(1+x^\iota)}\\
&\leq 1- \frac{(d+1)x^\iota-(d-1)}{2r} = 1-\alpha.
\end{align*}
We see that $\alpha$ is positive for $x>1-\frac{2}{(d+1)\iota}>\sqrt[\iota]{1-\frac{2}{d+1}}= \sqrt[\iota]{\frac{d-1}{d+1}}$. Further, we see for these $x$ that the derivative of $\alpha$ with respect to $x$ satisfies,
\[
\frac{\mathrm{d}\alpha}{\mathrm{d}x} = \frac{(d+1)\iota x^{\iota-1}}{2r} \geq \frac{(d+1)\iota x^\iota}{2r} \geq \frac{(d-1)\iota}{2r} \geq \frac{d\iota}{4r}.
\]
Hence we find that $\alpha \geq \frac{d\iota}{4r}\xi$.
We finish by proving the Claim in case (c).
Explicitly the transition probabilities in this case are given by (writing $p_t=\frac{\mu_\mathrm{flow}(f+t\chi_D)}{\sum_{u\in\mathbb Z_q} \mu_\mathrm{flow}(f+u\chi_D)}$ and $q_t=\frac{\mu_\mathrm{flow}(g+t\chi_D)}{\sum_{u\in\mathbb Z_q} \mu_\mathrm{flow}(g+u\chi_D)}$)
\[
(f,g) \to (f+t\chi_D,g+t\chi_D) \text{ with probability } \min(p_t,q_t),
\]
and for $s\neq t$
\begin{align*}
(f,g)\to(f+s\chi_D,g+t\chi_D) \text{ with probability } &\frac{(p_s-\min(p_s,q_s))(q_t-\min(p_t,q_t))}{\sum_{u\in\mathbb Z_q} (p_u-\min(p_u,q_u))}
\\
&=\frac{(p_s-\min(p_s,q_s))(q_t-\min(p_t,q_t))}{\sum_{u\in\mathbb Z_q} (q_u-\min(p_u,q_u))}.
\end{align*}
It is easily checked that this yields a valid coupling, i.e.\@ that the first coordinate has transition probabilities $p_t$, and similary $q_t$ for the second coordinate.
Now we wish to bound the sum of the diagonal entries.
To do this we have to take a closer look at the weights occurring in this table. We define $a_i$ to be the number of edges in $D$ with value 0 in the flow $f+i D$. This ensures that $\mu_\mathrm{flow}(f+i \chi_D) \propto x^{-a_i}$ and $p_t=\frac{x^{-a_t}}{\sum_u x^{-a_u}}$. Similarly, we define $b_i$ as the number of edges in $D$ with value 0 in the flow $g+i \chi_D$.
We derive some boundary conditions on the $a_i$'s and $b_i$'s. Ranging $i$ over $\mathbb Z_q$, every edge of $D$ will get value $0$ in exactly one of $f+i \chi_D$. So $\sum_i a_i$ is the length $|D|$. The same holds for the $b_i$'s, so in particular we find that $\sum_i a_i = \sum_i b_i$.
Second we will bound $\sum_i |a_i-b_i|$. If an edge is counted in $a_i$, but not in $b_i$, it must be an edge of $C$. For every such edge it can happen once that it is counted in $a_i$ and not $b_i$, and once vice versa. Hence the total absolute difference $\sum_i |a_i-b_i|$ is bounded by $2|C \cap D|\leq 2\iota$.
Now the sum of all the probabilities on the diagonal is
\[
\sum_i \min \left( \frac{x^{-a_i}}{\sum_j x^{-a_j}} , \frac{x^{-b_i}}{\sum_j x^{-b_j}} \right),
\]
and the numbers $a_i,b_i$ satisfy the conditions of \cref{lem:sum of minima}, so the sum is bounded below by $1-\frac{1-x^\iota}{1+x^\iota}$.\qedhere
\end{proof}
\begin{proof}[Proof of \cref{lem:sum of minima}]
First of all, let us introduce a little terminology: an index $i$ is called $b$-minimal if the minimum of the $i$-term in $S$ is not equal to the $a$-term. Also assume that $\sum_j x^{-a_j} \geq \sum_j x^{-b_j}$. And note that the two conditions imply
\[
2\iota \geq \sum_i |a_i-b_i| \geq |a_j-b_j| + \left| \sum_{i\neq j} a_i-b_i \right| = |a_j-b_j| + |b_j-a_j|=2|a_j-b_j|.
\]
Hence the absolute difference between $a_j$ and $b_j$ is always at most $\iota$.
The proof contains two steps. In the first step, we change the numbers $a_i$ in such a way that the conditions still hold and $S$ does not increase. After the first step there will be at most one $b$-minimal index $i$. This allows us to eliminate the minima from the expression for $S$. In the second step, we give a lower bound for this new obtained expression.
For the first step, assume that two different indices $t,u$ are $b$-minimal, and assume also that $a_t\geq a_u$. Now we increase $a_t$ by 1, and decrease $a_u$ by 1, i.e.\@ define the new sequence
\[
a'_i=\begin{cases}a_t+1 & i=t, \\ a_u-1 & i=u, \\ a_i & \text{otherwise}.\end{cases}
\]
First we note that $\sum_j x^{-a_j'} > \sum_j x^{-a_j}$, simply because
\[
x^{-a'_t}-x^{-a_t}=x^{-(a_t+1)}(1-x) > x^{-a_u}(1-x) = x^{-a_u} - x^{-a'_u}.
\]
Now we will show for every $i$, that the term $\min(x^{-a_i}/\sum_j x^{-a_j},x^{-b_i}/\sum_j x^{-b_j})$ does not increase. For $i\neq t,u$ this is easy, because $x^{-a_i}$ does not change and the sum in the denominator increases. Hence the first term in the minimum decreases and the minimum cannot increase. We also assumed that both $t,u$ were $b$-minimal, and because we don't change the $b_i$'s, the minimum cannot increase.
Further, we have to check that the new sequence still satisfies all the conditions. It is clear that $\sum_i a'_i = \sum_i a_i = \sum_i b_i$ and $\sum_j x^{-a_j'} > \sum_j x^{-a_j} \geq \sum_j x^{-b_j}$. Further we see for $i=t,u$ that
\[
\frac{x^{-a_i}}{ \sum_{j} x^{-a_j}} > \frac{x^{-b_i}}{ \sum_{j} x^{-b_j}} \geq \frac{x^{-b_i}}{ \sum_{j} x^{-a_j}},
\]
hence $a_i > b_i$ for $i=t,u$. Therefore $|a'_t-b_t|= |a_t-b_t +1|= |a_t-b_t|+1$ and $|a'_u-b_u|=|a_u-b_u-1|=|a_u-b_u|-1$ (because $a_u-b_u$ is a positive integer), so the sum of the absolute values remains the same.
After repeating this adjustment with the same indices, eventually one of them will stop being $b$-minimal. Now repeat with two new $b$-minimal indices, as long as they exist. In the end there must be at most one $b$-minimal index.
Now we are ready for step two. If there are no $b$-minimal indices, the sum is equal to 1 and the result holds. Hence we assume wlog that 1 is the only $b$-minimal index and we can write
\[
S= \frac{x^{-b_1}}{\sum_j x^{-b_j}} + \sum_{i\neq 1} \frac{x^{-a_i}}{\sum_j x^{-a_j}} =\frac{x^{-b_1}}{\sum_j x^{-b_j}} + 1-\frac{x^{-a_1}}{\sum_j x^{-a_j}}.
\]
Note that for positive $p,q$, the function $ \frac{-p}{p+q} $ is increasing in $q$ and decreasing in $p$. Because $x^{-a_1} \leq x^{-(b_1+\iota)}$ and $\sum_{j\geq2} x^{-a_j} \geq \sum_{j\geq2} x^{-(b_j-\iota)}$, we can thus estimate that
\[
S \geq \frac{x^{-b_1}}{\sum_j x^{-b_j}} + 1 - \frac{x^{-\iota} x^{-b_1}}{x^{-\iota} x^{-b_1}+x^\iota \sum_{j\geq2} x^{-b_j}}.
\]
Now write $X=x^{-\iota}$, $B_1=x^{-b_1}$ and $B_2=\sum_{j\geq2} x^{-b_j}$, so that the lower bound for $S$ can be written as $ \frac{B_1}{B_1+B_2} +1- \frac{X^2B_1}{X^2B_1+B_2}$. By AM-GM we can estimate that
\[
(B_1+B_2)(X^2B_1+B_2)=X^2B_1^2+B_2^2+(X^2+1)B_1B_2 \geq 2XB_1B_2+(X^2+1)B_1B_2 = (X+1)^2B_1B_2,
\]
so that we find:
\begin{align*}
S \geq 1+ \frac{B_1}{B_1+B_2} - \frac{X^2B_1}{X^2B_1+B_2} &= 1+ \frac{B_1(X^2B_1+B_2)-X^2B_1(B_1+B_2)}{(B_1+B_2)(X^2B_1+B_2)} \\
&= 1- \frac{(X-1)(X+1)B_1B_2}{(B_1+B_2)(X^2B_1+B_2)} \\
&\geq 1- \frac{(X-1)(X+1)B_1B_2}{(X+1)^2B_1B_2} = 1- \frac{X-1}{X+1}. \qedhere
\end{align*}
\end{proof}
\section{Joint flow-random cluster Markov chain}
\label{se:jointchain}
In this section we will consider a different chain that allows us to sample flows. We will again prove rapid mixing by using path coupling, and this holds for roughly the same range of parameters $x$.
To describe the chain let $q\geq 2$ be an integer and let $G=(V,E)$ be a graph $m$ edges. Let $\mathcal{C}$ be an even generating set for the flow space $\mathcal{F}_q(G)$ of size $r$ and let $\ell=\ell(\mathcal{C}).$
\begin{definition}
Let $\Omega_{\mathrm{flow-RC}}$ be the set of pairs $(f,F)$ with $F\subset E$ a set of edges and $f$ a flow on $(V,F)$.
The \emph{joint flow-RC Markov chain} is a Markov chain on the state space $\Omega_{\mathrm{flow-RC}}$ depending on two parameters $x,p\in(0,1)$. The transition probabilities are as follows:\\
For $e\in E \setminus F$:
\[
P_{\mathrm{flow-RC}}[(f,F) , (f,F\cup\{e\})] = \frac{(1-p)x}{m}.
\]
For $e\in F$ such that $f(e)=0$:
\[
P_{\mathrm{flow-RC}}[(f,F) , (f,F\setminus\{e\})] = \frac{(1-p)(1-x)}{m}.
\]
And for $t\in\{1,\ldots,q-1\}$, $C\in\mathcal{C}$ an even set such that $C\subseteq F$:
\[
P_{\mathrm{flow-RC}}[(f,F) , (f+t\chi_C,F)] = \frac{p}{qr}.
\]
All other transition probabilities are zero, except for the stationary probabilities $P_{\mathrm{flow-RC}}[(f,F),(f,F)]$.
\end{definition}
Simulating one step of this Markov chain starting in the state $(f,F)$ can be done as follows. We first select either `flow' or `edges' with probabilities resp.\@ $p$ and $1-p$.
\begin{itemize}
\item If we select `flow', we will update the flow $f$. We choose $C \in \mathcal{C}$ and $t\in\mathbb Z_q$ uniformly at random. If the flow $f + t\chi_C$ is supported on $F$ (for $t\neq 0$ this is equivalent to $C\subseteq F$), we make the transition $(f,F) \to (f+t\chi_C,F)$. Otherwise the chain stays in $(f,F)$.
\item If we select `edges', we will update the set of edges $F$. We choose an edge $e\in E$ uniformly at random. If $e$ is not contained in $F$, we make with probability $x$ the transition $(f,F)\to(f,F\cup\{e\})$. If $e$ is contained in $F$ and $f(e)=0$, we make with probability $1-x$ the transition $(f,F) \to (f, F\setminus\{e\})$. Otherwise the chain stays in $(f,F)$.
\end{itemize}
The total cost of simulating one step of this Markov chain is $O(\ell)$ for checking whether $C \subseteq F$ in the first case.
Further this Markov chain has stationary distribution $\mu_{\mathrm{flow-RC}} \colon (f,F) \mapsto \frac{1}{Z_\mathrm{flow}} x^{|F|}(1-x)^{|E\setminus F|}$. (From \cref{lem: pf flow=pf RC} it follows that the sum over all states is 1.) This follows easily from checking the detailed balance equation.
\subsection{Rapid mixing of joint flow-RC Markov chain}
\begin{theorem}\label{thm:radid mixing joint}
Let $\ell\geq3, q,s\geq2$ be integers and $1>x>1-\frac{q}{(q-1)\ell s}$.
Write $\xi=x-\left(1-\frac{q}{(q-1)\ell s}\right)$ and let $\delta>0$.
Let $G=(V,E)$ be a graph and $\mathcal{C}$ an even generating set of $\mathcal{F}_q(V,E)$ of size $r$ satisfying $\ell(\mathcal{C}) \leq \ell$ and $s(\mathcal{C}) \leq s$, then there is a value of $p$ for which the joint flow-RC Markov chain for $(G,\mathcal{C})$ comes $\delta$-close to $\mu_{\mathrm{flow-RC}}$ with parameter $x$ in at most $\frac{2(m+r)}{\ell}\log((2m+r)\delta^{-1})\xi^{-1}$ steps.
\end{theorem}
\begin{remark}
An exact value for $p$ in the theorem above can be obtained from equation \eqref{eq:determine p} below.
\end{remark}
\begin{remark}
Note again that $\xi > \frac{q}{(q-1)\ell s} > \frac{1}{\ell s}$, and hence the required number of calls in the above theorem is at least $2s(m+r) \log((2m+r)\delta^{-1})$. Again this means the bound does not get better with larger $\ell$, even though it appears in the denominator, and even gets worse with larger $s$.
\end{remark}
It would be interesting to see if the theorem could be used to say anything about possible rapid mixing of the Glauber dynamics for the random cluster model at low temperatures cf.~\cite{GJ18}.
The following corollary is immediate by Proposition~\ref{prop:flow sampler to RC/Potts} and directly implies Theorem~\ref{thm:main sampling}(ii) by \cref{lem:generating set for all}.
\begin{cor}
Fix integers $\ell\geq 3$ and $q,s\geq 2$.
Let $w>(q-1)(\ell s-1)$ and $\delta>0$, then there exists an algorithm that on input an $m$-edge graph $G$ and an even generating set $\mathcal{C}$ for $\mathcal{F}_q(G)$ of size $r$ satisfying $\ell(\mathcal{C}) \leq \ell$ and $s(\mathcal{C})\leq s$, outputs a $q$-state Potts colouring $\sigma:V\to [ q ] $ within total variation distance $\delta$ of the $q$-state Potts-measure $\mu_\mathrm{Potts}$ with parameter $w$. This is obtained by running the joint flow-RC Markov chain for $O((m+r)\log((m+r)\delta^{-1}))$ steps, where each step takes $O(1)$ time (since $\ell$ is fixed).
\end{cor}
\begin{proof}[Proof of \cref{thm:radid mixing joint}]
We will again use path coupling to deduce rapid mixing of the above defined Markov chain. The distance we use on the state space is defined as the least number of steps required in the Markov chain to go from one state to the another.
A crude upper bound on the diameter is given by $2m+r$.
There are two kinds of pairs of states at distance one, which we will treat separately. Just as in the proof of \cref{thm:mixing time flow MC}, we will prove that the expected distance after one step of the coupling is at most $1-\alpha$ for some $\alpha$, and therefore the mixing time is at most $\log((2m+r)\delta^{-1})\alpha^{-1}$.
Consider the states $(f,F)$ and $(f,F \cup \{e\})$. We will make a coupling on them. The transition probabilities of this coupling are as follows:
\[
\begin{pmatrix*}[l] f & F\\ f & F\cup\{e\} \end{pmatrix*} \to
\begin{cases}
\begin{pmatrix*}[l] f & F\cup\{e\} \\ f & F\cup\{e\}\end{pmatrix*} & \frac{(1-p)x}{m},\\
\begin{pmatrix*} f & F \\ f & F \end{pmatrix*} & \frac{(1-p)(1-x)}{m},\\
\begin{pmatrix*}[l] f & F\cup\{e'\} \\ f & F\cup\{e,e'\} \end{pmatrix*} & \frac{(1-p)x}{m} \text{ if $e'\not\in F\cup\{e\}$}, \\
\begin{pmatrix*}[l] f & F\setminus\{e'\} \\ f & F\setminus\{e'\} \cup \{e\} \end{pmatrix*} & \frac{(1-p)(1-x)}{m} \text{ if $e'\in F$ and $f(e')=0$},\\
\begin{pmatrix*}[l] f+t\chi_C & F \\ f+t\chi_C & F\cup\{e\} \end{pmatrix*} & \frac{p}{qr} \text{ if $t\neq0$ and $C\subseteq F$},\\
\begin{pmatrix*}[l] f & F \\ f+t\chi_C & F\cup\{e\} \end{pmatrix*} & \frac{p}{qr} \text{ if $t\neq0$, $e\in C$ and $C\subseteq F\cup\{e\}$}.
\end{cases}
\]
The first two cases each occur exactly once and decrease the distance by one. The last case occurs at most $s(q-1)$ times and increases the distance by one. Therefore the expected distance after one step of the coupling is at most
\[
1-\frac{1-p}{m}+\frac{(q-1)sp}{qr}
\]
in this case.
Next is the coupling on the neighbouring states $(f,F)$ and $(f+t\chi_C,F)$ (with $t \not= 0$). The transition probabilities are as follows:
\[
\begin{pmatrix*}[l]f & F \\ f+t\chi_C & F \end{pmatrix*} \to
\begin{cases}
\begin{pmatrix*}[l]f & F\cup\{e\} \\ f+t\chi_C & F\cup\{e\} \end{pmatrix*} & \frac{(1-p)x}{m} \text{ if $e\not\in F$},\\
\begin{pmatrix*}[l]f & F\setminus\{e\} \\ f+t\chi_C & F\setminus\{e\} \end{pmatrix*} & \frac{(1-p)(1-x)}{m} \text{ if $e\not\in F$ and $e\not\in C$},\\
\begin{pmatrix*}[l]f & F \setminus \{e\}\\ f+t\chi_C & F \end{pmatrix*} & \frac{(1-p)(1-x)}{m} \text{ if $e\in C$ and $f(e)=0$},\\
\begin{pmatrix*}[l]f & F\\ f+t\chi_C & F\setminus\{e\} \end{pmatrix*} & \frac{(1-p)(1-x)}{m} \text{ if $e\in C$ and $f(e)+t\chi_C(e)=0$},\\
\begin{pmatrix*}[l]f+t'\chi_C & F \\f+t'\chi_C & F \end{pmatrix*} & \frac{p}{qr},\\
\begin{pmatrix*}[l]f+t'\chi_{C'} & F \\ f+t\chi_C+t'\chi_{C'} & F\end{pmatrix*} & \frac{p}{qr} \text{ if $t'\neq 0$ and $C'\neq C$}.
\end{cases}
\]
The third and fourth case occur together at most $\ell$ times and increase the distance with one. The fifth case occurs exactly $q$ times and decreases the distance with one. Therefore the expected distance after one step of the coupling is at most
\[
1-\frac{p}{r}+\frac{\ell(1-x)(1-p)}{m}.
\]
To find a useful coupling, both expected distances will have to be smaller than one and we have to solve the following equations (for $p$ and $\alpha$):
\begin{align}
1-\frac{1-p}{m}+\frac{(q-1)sp}{qr} &= 1-\frac{p}{r}+\frac{\ell(1-x)(1-p)}{m} = 1-\alpha, \nonumber
\\
\text{i.e. }\:\frac{1-p}{m}-\frac{(q-1)sp}{qr} &= \frac{p}{r}-\frac{\ell(1-x)(1-p)}{m} = \alpha. \label{eq:determine p}
\end{align}
For $p=0$, the first term is positive while the second is negative, and vice versa for $p=1$. Therefore the solution for $p$ lies indeed in $(0,1)$ and we will not calculate it explicitly. Instead we eliminate $p$ to only calculate the value of $\alpha$:
\begin{align*}
&\frac{1}{qrm}\left( qr + (q-1)sm + qm + qr\ell(1-x) \right) \alpha \\
&= \left( \frac{1}{m} + \frac{(q-1)s}{qr} \right) \left(\frac{1}{r}p+\frac{\ell(1-x)}{m}p - \frac{\ell(1-x)}{m}\right) + \left( \frac{1}{r} + \frac{\ell(1-x)}{m} \right) \left(\frac{1}{m}-\frac{1}{m}p-\frac{(q-1)s}{qr}p \right)\\
&= -\frac{(q-1)\ell s(1-x)}{qrm} + \frac{1}{rm},
\end{align*}
reducing to
\begin{equation*}
\alpha = \frac{q-(q-1)\ell s(1-x)}{qr+(q-1)sm + qm + qr\ell(1-x)}.
\end{equation*}
Since $x>1-\frac{q}{(q-1)\ell s}$, this value of $\alpha$ is positive. Plugging in $1-x=\frac{q}{(q-1)\ell s}-\xi$, we continue to find a bound on $\alpha^{-1}$:
\begin{align*}
\alpha^{-1} &= \frac{qr+(q-1)sm + qm + qr\ell(1-x)}{q-(q-1)\ell s(1-x)} = \frac{qr+(q-1)sm + qm + qr\ell(1-x)}{(q-1)\ell s\xi}\\
&< \frac{qr+(q-1)sm + qm + \frac{q^2r}{(q-1)s}}{(q-1)\ell s\xi} \leq \frac{2(m+r)}{\ell}\xi^{-1}.
\end{align*}
This finishes the proof.
\end{proof}
\section{Computing the partition function using the Markov chain sampler}
\label{se:count}
In this section we prove Theorem~\ref{thm:main counting}.
We will do this with a self-reducibility argument, making use of a connection between removing and contracting edges.
We have the following result.
\begin{prop}\label{prop:sampling implies counting}
Let $x\in [1/3,1]$ and let $q\in \mathbb{N}_{\geq 2}$.
Let $\mathcal{G}$ be a family of graphs which is closed under contracting edges.
Assume we are given an algorithm that for $n$-vertex and $m$-edge graph $G\in \mathcal{G}$ and any $\delta>0$ computes a random $\mathbb{Z}_q$-flow with distribution $\delta$-close to $\mu_{\mathrm{flow}}$ in time bounded by $T(\delta,n,m)$.
Then there is an algorithm that given an $n$-vertex and $m$-edge graph $G\in \mathcal{G}$ and any $\epsilon>0$ computes a number $\zeta$ such that with probability at least $3/4$
\[
e^{-\epsilon}\leq \frac{\zeta}{Z_{\mathrm{flow}}(G;q,x)}\leq e^{\epsilon}
\]
in time $O(n^2\epsilon^{-2}T( \epsilon/n,n,m))$.
\end{prop}
Before proving the proposition, let us show how it implies Theorem~\ref{thm:main counting}.
\begin{proof}[Proof of Theorem~\ref{thm:main counting}]
We prove part (i): part (ii) follows in exactly the same way.
Fix positive integers $\iota$ and $d$ with $d$ at least $2$.
Consider the class of graphs $\mathcal{G}$ that have a basis for the cycle space $\mathcal{C}$ consisting of even sets satisfying $\iota(\mathcal{C})\leq \iota$ and $d(\mathcal{C})\leq d.$
By Lemma~\ref{lem:contraction is good} this class is closed under contracting edges.
By Theorem~\ref{thm:mixing time flow MC} we have an algorithm that for each $m$-edge graph $G\in \mathcal{G}$ and any $\delta>0$ computes a random $\mathbb{Z}_q$-flow with distribution within total variation distance $\delta$ from $\mu_{\mathrm{flow}}$ in time bounded by $T(\delta,n,m) =O(m^2 \log(m \delta^{-1}))$ provided $x>1-\frac{2}{(d+1)\iota}\geq 1/3$; see Remark~\ref{re:runtime}).
The theorem now follows from the previous proposition combined with the fact that $Z_{\mathrm{flow}}(G;q,x)=(1-x)^{|E|}q^{-|V|}Z_{\mathrm{Potts}}(G;q,\tfrac{1+(q-1)x}{1-x})$ by Lemma~\ref{lem:tutte}.
The running time is given by $O(n^2m^2 \epsilon^{-2} \log(nm \epsilon^{-1}))$.
\end{proof}
We now turn to the proof of Proposition~\ref{prop:sampling implies counting}.
\begin{proof}[Proof of Proposition~\ref{prop:sampling implies counting}]
As already mentioned above the proof relies on a self-reducibility argument.
The flow partition function satisfies the following well known deletion-contraction relation: for a graph $G=(V,E)$ and $e\in E$ not a loop, we have
\begin{equation}\label{eq:deletion contraction}
Z_{\mathrm{flow}}(G;q,x)=(1-x)Z_{\mathrm{flow}}(G\setminus e;q,x) + xZ_{\mathrm{flow}}(G/e;q,x).
\end{equation}
This holds because the collection of all flows on $G$ and on $G/e$ are in bijection with each other, while the flows on $G\setminus e$ correspond to the flows on $G$ that take value $0$ on $e$.
We rewrite~\eqref{eq:deletion contraction} as
\begin{equation}\label{eq:ratio flow}
\frac{Z_{\mathrm{flow}}(G/e;q,x)}{Z_{\mathrm{flow}}(G;q,x)}=\frac{1}{x}-\frac{1-x}{x}\cdot\frac{{Z_{\mathrm{flow}}(G\setminus e;q,x)}}{Z_{\mathrm{flow}}(G;q,x)},
\end{equation}
and we interpret the fraction
\[
\frac{{Z_{\mathrm{flow}}(G\setminus e;q,x)}}{Z_{\mathrm{flow}}(G;q,x)}
\]
as the probability that $e$ is assigned the value $0\in \mathbb{Z}_q$ when a flow is sampled from $\mu_{\mathrm{flow}}$.
This probability can be estimated using the assumed sampler. Hence we can use the sampler to estimate \eqref{eq:ratio flow}.
From $G=(V,E)$. we now construct a series of graphs $G=G_0,G_1,\ldots G_t$ where in each step we contract one edge (which is not a loop). We can do this, until every component has been contracted to a single vertex, possibly with some loops attached to it.
This takes $t=|V|-c(G)\leq |V|$ steps, where $c(G)$ denotes the number of components of $G$.
In the end we have $|E|-|V|+c(G)\leq |E|$ edges (loops) left and
the resulting graph $G_t$ thus has flow partition function $Z_{\mathrm{flow}}(G_t;q,x)=(1+(q-1)w)^{|E|-|V|+c(G)}$.
Therefore
\begin{equation}\label{eq:telescope}
\frac{(1+(q-1)x)^{|V|-c(G)}}{Z_{\mathrm{flow}}(G;q,x)} =\frac{Z_{\mathrm{flow}}(G_{r};q,x)}{Z_{\mathrm{flow}}(G_0;q,x)}= \frac{Z_{\mathrm{flow}}(G_1;q,x)}{Z_{\mathrm{flow}}(G_0;q,x)}\cdots \frac{Z_{\mathrm{flow}}(G_{t};q,x)}{Z_{\mathrm{flow}}(G_{t-1};q,x)}.
\end{equation}
Note that for each $i$ and any non-loop edge $e\in E(G_i)$ we have by~\eqref{eq:ratio flow},
\begin{equation}\label{eq:bound prob e notin}
1\leq \frac{Z_{\mathrm{flow}}(G_i/ e;q,x)}{Z_{\mathrm{flow}}(G_i;q,x)}\leq 1/x\leq 3,
\end{equation}
since $x\geq1/3$.
We can now estimate each individual probability on the right-hand side of \eqref{eq:telescope} to get an estimate for $Z_{\mathrm{flow}}(G;q,w)$.
This is rather standard and can be done following the approach in~\cite{Jbook} for matchings. We therefore only give a sketch of the argument, leaving out technical details
For each $i$, let
\[
p_i\coloneqq\frac{Z_{\mathrm{flow}}(G_i\setminus e;q,x)}{Z_{\mathrm{flow}}(G_i;q,x)}.
\]
To estimate $p_i$ we run our sampler $M=O(\epsilon^{-2}t)$ times with $\delta=O(\epsilon/t)$ to generate independent random flows $f_j$ $(j=1,\ldots,M)$. Denote by $X_j$ the random variable that is equal to $1$ if $e$ is not contained in $\supp(f_j)$ and $0$ otherwise.
We are in fact not interested in $p_i$, but rather in
\[
\hat{p}_i\coloneqq\frac{Z_{\mathrm{flow}}(G_i/ e;q,x)}{Z_{\mathrm{flow}}(G_i;q,x)}=\frac{1}{x}-\frac{1-x}{x}p_i.
\]
We therefore define the random variable $Y_j\coloneqq\frac{1}{x}-\frac{1-x}{x}X_j$ and $Y^i\coloneqq1/M\sum_{j=1}^M Y_j$.
Note that $\mathbb{E}[Y^i]=\mathbb{E}[Y_j] = \hat{p}_i$ and it is easy to check that $\text{Var}[Y^i]=1/M\text{Var}[Y_j]=1/M(\mathbb{E}(Y_j)-1)(1/x-\mathbb{E}(Y_j))$ for any $j=1,\ldots,M$.
We note that, by definition of the total variation distance, the fact that $x\geq 1/3$, and \eqref{eq:bound prob e notin}, we have
\begin{equation}\label{eq:delta bound}
\hat{p}_i(1-2\delta)\leq \hat{p}_i-\frac{1-x}{x}\delta\leq \mathbb{E}[Y^i] = \mathbb{E}[Y_j] \leq \hat{p}_i+\frac{1-x}{x}\delta\leq (1+2\delta)\hat{p}_i.
\end{equation}
This implies that
\[
\frac{\text{Var}[Y^i]}{\mathbb{E}[Y^i]^2}=\frac{1}{M}
\frac{(\mathbb{E}(Y_j)-1)(1/x-\mathbb{E}(Y_j))}{\mathbb{E}[Y^i]^2}\leq O(\epsilon^2/t).
\]
Consider next the random variable $Y\coloneqq\prod_{i=1}^t Y^i$. This will, up to a multiplicative factor (cf.\eqref{eq:telescope}), give us the desired estimate.
Since the $Y^i$ are independent we have
\begin{align*}
\frac{\text{Var}[Y]}{\prod_{i=1}^t \mathbb{E}[Y^i]^2}=\prod_{i=1}^t \frac{\mathbb{E}[(Y^i)^2]}{\mathbb{E}[Y^i]^2}-1
=\prod_{i=1}^t \left(1+\frac{\text{Var}[Y^i]}{\mathbb{E}[Y^i]^2}\right)-1\leq O(\epsilon^2).
\end{align*}
Then by Chebychev's inequality $Y$ does not deviate much from $\prod_{i=1}^t \mathbb{E}[Y^i]$ with high probability, which by \eqref{eq:delta bound} and our choice of $\delta$ does not deviates much from $\prod_{i=1}^t \hat{p}_i$.
More precisely, $Y$ will not deviate more than an $\exp(O(\epsilon))$ multiplicative factor from $\prod_{i=1}^t \hat{p}_i$ with high probability, as desired.
We need to access the sampler $O(t/\epsilon^2)$ many times with $\delta=O(\epsilon/t)$ to compute each $Y^i$. So this gives a total running time of $O(n^2 \epsilon^{-2} T(\epsilon/n, n,m))$.
This concludes the proof sketch.
\end{proof}
\section{Slow mixing of the flow chain}
\label{se:slow}
In this section we show that the flow Markov chain cannot mix rapidly for all $x \in (0,1)$. We do this by
using the duality of our Markov chain on flows and Glauber dynamics of the Potts model on the planar grid (although the duality holds more generally on planar graphs). A result of Borgs, Chayes, and Tetali \cite{BCT12} for slow mixing of the Glauber dynamics of the Potts model on the grid (below a critical temperature) then immediately implies slow mixing of our flows Markov chain at the same temperature.
Given a graph $G=(V,E)$, let $\mathcal{F}_q(G)$ be the set of $\mathbb Z_q$-flows on $G$ and let $\Omega_q(G)$ be the set of $\tau: V \rightarrow [q]$ of $q$-spin configurations on $G$. Clearly $|\Omega_q(G)| = q^{|V|}$ and, as noted earlier, $|\mathcal{F}_q(G)| = q^{|E| - |V|-1}$.
Recall that the Glauber dynamics for the $q$-state Potts model for a graph $G$ and parameter $x$ is the following Markov chain with state space $\Omega_q(G)$. Given that we are currently at state $\sigma \in \Omega_q(G)$, we pick a vertex $v \in V$ uniformly at random and update its state as follows: we choose the new state to be $i$ with probability $x^{m(i)}/Z_v$, where $m(i)$ is the number of neighbours of $v$ that have state $i$ in $\sigma$, and $Z_v = \sum_i x^{m(i)}$.
Let $G=(V,E)$ be the $((L+1) \times (L+1))$-grid and $H=(V',E')$ the $(L \times L)$-grid. One can easily check that $|V'| = |E| - |V| + 1$ and so $|\Omega_q(H)| = |\mathcal{F}_q(G)|$. There is a natural bijection $\phi: \Omega_q(H) \rightarrow \mathcal{F}_q(G)$ defined as follows. First note that $H$ is the planar dual of $G$ (ignoring the outer face of $G$). Using this, write $v_1, \ldots, v_{L^2}$ for the vertices of $H$ and $C_1, \ldots, C_{L^2}$ for the corresponding faces (i.e.\ $4$-cycles) of $G$. Given $\sigma \in \Omega_q(H)$, let $\phi(\sigma) = \sum_{i = 1}^{L^2}\sigma(v_i)C_i$. We see that $\phi$ is injective since the $C_i$ form a basis of the cycle space of $G$, and hence $\phi$ must be bijective.
Now it is easy to check that the $q$-state Potts Glauber dynamics on $H$ is equivalent to the $\mathbb{Z}_q$-flow Markov chain on $G$ (where both chains have the same interaction parameter, say $x$) via the correspondence $\phi$ between their state spaces. In other words if $P$ and $Q$ are their respective transition matrices then $P_{\sigma_1 \sigma_2} = Q_{\phi(\sigma_1) \phi(\sigma_2)}$ for all $\sigma_1, \sigma_2 \in \Omega_q(H)$.
Borgs, Chayes, and Tetali \cite{BCT12} showed that the mixing time of the Glauber dynamics of the $q$-state Potts model on the $L \times L$ grid with interaction parameter $x = e^{-\beta}$ is bounded below by $x^{CL}$ for some constant $C$ when $\beta$ is above the critical threshold for the grid, i.e.\ $\beta \geq \beta_0(\mathbb Z^2) = \frac{1}{2} \log q + O(q^{-1/2})$. In particular this shows the same exponential lower bound on the mixing time for the $\mathbb{Z}_q$-flow Markov chain (for the same interaction parameter $x$) on the $(L+1) \times (L+1)$-grid.
\subsection*{Acknowledgement}
We are grateful to an anonymous referee for several useful suggestions.
\bibliographystyle{plain}
|
\section{Introduction}
Let $I=[0,T]$ ($T>0$). We consider in this paper, in a separable
Hilbert space $\mathcal{H}$, the following perturbed evolution
differential inclusion
\begin{equation}\label{1.1} -Du(t) \in A(t) u(t) +f(t,
u(t))\;\;a.e.,\;\;\;u(0)=u_0,\end{equation} where for each $t\in I$,
$A(t)$ is a maximal monotone operator of $\mathcal{H}$, and where
the set-valued map $t \mapsto A(t) $ is right continuous with
bounded variation (BVRC),
in the sense that
there exists a function $\rho: I \rightarrow [0, \infty[$,
which is right continuous on $[0, T[$ and nondecreasing with
$\rho(0) = 0$ and $\rho(T) < \infty$ such that
$$ dis(A(t), A(s)) \leq d\rho(]s, t]) = \rho(t)-\rho(s) \;\;\; 0\leq s \leq t \leq T,$$
here $dis(\cdot, \cdot)$ is the pseudo-distance between maximal
monotone operators introduced by Vladimirov \cite{Vla}; see relation
\eqref{2.1}, and $f:I\times \mathcal{H}\longrightarrow \mathcal{H}$
is measurable in $t$ and satisfies a Lipschitz condition w.r.t the
second variable. For the study of problem \eqref{1.1} without
perturbation, we refer to Theorem 1 in \cite{KM}.
Our result extends to the BVRC time-depend maximal monotone
operators, Theorem 3.1 in \cite{ACM}, dealing with BVC (continuous
with bounded variation) time-dependent maximal monotone operators,
and Theorem 4.1 in \cite{AHT} dealing with BVRC sweeping processes.
It is worth to mention that in the recent work by Azzam et al
\cite{ACM1}, the authors gave a theorem on the existence and
uniqueness of a BVRC solution to problem \eqref{1.1}, by assuming
that the domain of the operator $A(t)$ is ball-compact. In the
present paper, we will establish the same result without any
compactness assumption. Actually, we will prove the Cauchy criteria
to conclude the convergence of our approximate sequence to the
desired solution, by following ideas in the proofs of Theorem 3.1
in \cite{ACM} and Theorem 4.1 in \cite{AHT}.
Let $\lambda$ be the Lebesgue measure on $I$ and $d\rho$ the
Stieljes measure associated with $\rho$. We set $\nu := \lambda
+d\rho$ and $\frac{ d\lambda}{d\nu}$ the density of $ \lambda$ w.r.t
$\nu$. By a solution of \eqref{1.1}, we mean that
there exists a BVRC mapping $u: I\longrightarrow \mathcal{H}$
satisfying
$$
(P_f)
\begin{cases}
u(0)=u_0\in D(A(0));\\
u(t)\in D(A(t))\;\;\;\forall t\in I;\\
-\displaystyle\frac{du}{d\nu }(t)\in A(t) u(t)+ f(t, u(t)) \frac{d\lambda} {d\nu}(t) \;\;\;d\nu-a.e.\,t\in
I,
\end{cases}
$$
where $\frac{du}{d\nu }$ denotes the density of $u$ w.r.t $\nu$.
There is an intensive study on differential inclusions governed by
time-dependent and time and state dependent maximal monotone
operators, where the variation of the operator is absolutely
continuous or continuous with bounded variation; see for instance
\cite{ABCM1, ABCM2, AB, ACM, BAC, Ken, KM, Le, P, SAM, TOL1, Vla}.
However, there are few works concerning $(P_f)$ when the variation
of the operator $A(t)$ is right continuous with bounded variation
(BVRC). Actually, to the best of our knowledge, there is an
existence result to $(P_0)$ $(f=0)$ in \cite{KM} and for the
perturbed problem we refer to the recent reference \cite{ACM1}.
Differential inclusions governed by time-dependent maximal monotone
operators constitute generalization of the so called "sweeping
process", that is differential inclusions governed by the normal
cone to closed and convex moving sets, since this normal cone is a
maximal monotone operator. This process was introduced and widely
studied by J.J. Moreau; we cite for instance \cite{M, M1}. We can
also refer to the book of Monteiro-Marques \cite{Mon}. Following the
results of Moreau, several extensions in diverse directions, in
particular perturbed sweeping processes, have been studied in the
literature, where the moving sets in the normal cone have absolutely
continuous variation or continuous and bounded in variation (BVC).
The BVRC case was also deeply discussed in many interesting papers;
we cite for instance \cite{AHT, ANT, BCSS, CM, ET1, M2, NNT, Thi3}.
The paper is organized as follows. In section 2, we give notations
and recall the preliminary results that we need in the sequel. In
section 3, we prove the existence and uniqueness of a right
continuous with bounded variation solution to the evolution problem
$(P_f)$ when $t \mapsto A(t) $ is BVRC and the perturbation $f:
I\times \mathcal{H}\longrightarrow \mathcal{H}$ is measurable on $I$
and satisfies a Lipschitz condition w.r.t the second variable.
\section{Notations and Preliminaries} From now on, $I:=[0,T]$ $(T>0)$ is an interval of
$\mathbb{R}$ and $(\mathcal{H}, \langle \cdot ,\cdot \rangle)$ is a
separable Hilbert space. Its norm will be denoted by $\|\cdot\|$,
and its unit closed ball (resp. closed ball of center $0$ and radius
$r>0$) will be denoted by $\overline{B}_{\mathcal{H}}$ (resp.
$r\overline{B}_{\mathcal{H}}$). We will denote by $\mathcal{B}(I)$
(resp. $\mathcal{B}(\mathcal{H})$) the Borel tribe on $I$ (resp. on
$\mathcal{H})$. The identity mapping of $\mathcal{H}$ will be
denoted by $Id_{\mathcal{H}}$. For a subset $K$ of $\mathcal{H}$,
$\overline{co}(K)$ will be the closed convex hull of $K$, which is
characterized by:
\begin{equation}\label{co}
\overline{co}(K)=\big\{y\in \mathcal{H}:\;\langle y, x\rangle\leq
\sup_{z\in K}\langle z, x\rangle\;\forall x\in K\big\}.
\end{equation}
If $\mu$ is a positive measure on $I$, we will denote by $L^p(I,
\mathcal{H}; \mu)$ $p \in [1, +\infty[$ (resp. $p=+\infty)$, the
Banach space of classes of $p$-$\mu$-integrable (resp.
$\mu$-essentially bounded) functions equipped with its classical
norm $\|\cdot\|_p$ (resp. $\|\cdot\|_{\infty}$).
We introduce in the following, definitions and some properties of
functions with bounded variation and general vector measures. We
refer to \cite{Mon, M2, M3, MV} for more details.
Let $u:I\longrightarrow \mathcal{H}$. The variation of $u$
in $I$ is the nonnegative extended real number
\begin{equation}\label{1}
var(u ; I):=\sup\sum_{k=1}^{n}\|u(t_k)-u(t_{k-1})\|,
\end{equation}
where the supremum is taken in $[0,+\infty]$ w.r.t all the finite
sequences $t_0<t_1<\cdots<t_n$ of points of $I$ ($n$ is arbitrary).
The function $u$ is said with bounded variation (BV) if and only if
$var(u ; I)<+\infty$, and we write $u\in BV(I, \mathcal{H})$. In this
case we have
\begin{equation*}
\lim_{s\uparrow t}\|u(t)-u(s)\|=\lim_{s\uparrow t} var(u; [s,t]).
\end{equation*}
Now, let $\mu$ be a positive Radon measure on $I$ and $\hat{\mu}$
be an $\mathcal{H}$-valued measure on
$I$ admitting a density $\frac{d \hat{\mu}}{d\mu}\in
L^1_{loc}(I;\mathcal{H};\mu)$. Then for $d \mu$-almost every $t\in
I$, we have
\begin{equation}\label{2}
\frac{d \hat{\mu}}{d\mu}(t)=\lim_{\varepsilon\downarrow 0}\frac{d
\hat{\mu}([t, t+\varepsilon])}{d\mu([t,
t+\varepsilon])}=\lim_{\varepsilon\downarrow 0}\frac{d
\hat{\mu}([t-\varepsilon, t])}{d\mu([t-\varepsilon, t])}.
\end{equation}
The measure $\hat{\mu}$ is absolutely continuous w.r.t $\mu$ if and
only if $\hat{\mu}=\frac{d \hat{\mu}}{d\mu}\mu$, i.e., $\frac{d
\hat{\mu}}{d\mu}$ is a density of $\hat{\mu}$ w.r.t $\mu$. In this
case a mapping $u:I\longrightarrow \mathcal{H}$ is
$\hat{\mu}$-integrable if and only if the mapping $t\mapsto u(t)\frac{d
\hat{\mu}}{d\mu}(t)$ is $\mu$-integrable, and we have
\begin{equation}\label{7}
\int_I u(t) d\hat{\mu}(t)=\int_I u(t)\frac{d
\hat{\mu}}{d\mu}(t) d\mu(t).
\end{equation}
If $u:I\longrightarrow \mathcal{H}$ is BVRC and $du$ is its
differential measure, then we have
\begin{equation}\label{3}
u(t)=u(s)+\int_{]s,t]} du\;\;\;\;\;\forall s, t\in I\;(s\leq t).
\end{equation}
Conversely, if there exists $v\in L^1(I,\mathcal{H};\mu)$ such that
$u(t)=u(0)+\int_{]0, t]} v \,d \mu$ for all $t\in I$, then $u$ is
BVRC and $du=v \, d\mu$, that is $v$ is a density of the vector
measure $du$ w.r.t the measure $\mu$. So that by \eqref{2}, we get
\begin{equation}\label{4}
v(t)=\frac{d u}{d\mu}(t)=\lim_{\varepsilon\downarrow 0}\frac{ du
([t, t+\varepsilon])}{d\mu([t,
t+\varepsilon])}=\lim_{\varepsilon\downarrow 0}\frac{ du
([t-\varepsilon, t])}{d\mu([t-\varepsilon, t])}.
\end{equation}
If $\mu(\{t\})>0$, this last relation shows that
\begin{equation}\label{5}
v(t)=\frac{d u}{d\mu}(t)=\frac{d
u(\{t\})}{d\mu(\{t\})}\;\;\;\textmd{and}\;\;\;\frac{d
\lambda}{d\mu}(t)=0.
\end{equation}
The following lemmas will be needed in our proof.
\begin{lem}\label{Lem2.1}(\cite{M3})
Let $\mu$ be a positive Radon measure on $I$ and $u:I\longrightarrow
\mathcal{H}$ a BVRC mapping such that its differential measure $du$ has a density
$\frac{du}{d\mu}$ w.r.t $\mu$. Then the real function
$\varphi:t\mapsto \varphi(t)=\|u(t)\|^2$ is BVRC and its
differential measure $d\varphi$ satisfies
\begin{equation}\label{6}
d\varphi\leq 2\big\langle u, \frac{du}{d\mu}\big\rangle d\mu.
\end{equation}
\end{lem}
\begin{lem}\label{lem2.6}(\cite{BCG}, Lemma 2.1) Let $\mu$ be a positive Radon measure on $I$. Let $g \in L^1(I, \mathbb{R}; \mu)$ be a nonnegative function
and $\beta \geq 0$ be such that,
$\forall t \in I$, $0\leq \mu ( \{t\} ) g(t) \leq \beta < 1$. Let
$\varphi \in L^\infty(I, \mathbb{R}; \mu)$ be a nonnegative
function satisfying
\begin{equation*}\varphi(t) \leq \alpha +\int_{]0, t]} g(s) \varphi(s)
\mu(ds)\;\;\; \forall t \in I, \end{equation*} where $\alpha$ is a
nonnegative constant. Then
\begin{equation*}\varphi(t) \leq \alpha \exp \Big( \frac{1} {1-\beta} \int_{]0, t]} g(s) \mu(ds)\Big)\;\;\; \forall t \in I. \end{equation*}
\end{lem}
\begin{lem}\label{lem2.5}
Let $(\alpha_i)$, $(\beta_i)$, $(\gamma_i)$ and $(a_i)$ be sequences
of nonnegative real numbers such that $a_{i+1}\leq
\alpha_i+\beta_i\big(a_0+a_1+\cdots+a_{i-1}\big)+(1+\gamma_i)a_i$
for $i\in \mathbb{N}$. Then
$$ a_j\leq \bigg(a_0+\sum_{k=0}^{j-1}
\alpha_k\bigg)\exp\bigg(\sum_{k=0}^{j-1}
\big(k\beta_k+\gamma_k\big)\bigg)\;\;\textmd{for}\;j\in
\mathbb{N}^*.$$
\end{lem}
We finish this section by the definition and some properties of
maximal monotone operators. We refer the reader to \cite{Ba, brezis,
Vra} for these concepts.
Let $A: \mathcal{H}
\rightrightarrows \mathcal{H}$ be a set-valued map. We denote by
$D(A)$, $R(A)$ and $Gr(A)$ its domain, range and graph. We say that
the operator $A$ is monotone, if $\langle y_1 -y_2, x_1 -
x_2\rangle\ge 0$ for all $(x_i, y_i)\in Gr(A)$ $(i=1, 2)$, and
we say that $A$ is a maximal monotone operator of $\mathcal{H}$,
if it is monotone and
its graph could not be contained strictly in the graph of any other
monotone operator.
If $A$ is a maximal monotone operator of $\mathcal{H}$, then for
every $x \in D(A)$, $A(x)$ is nonempty closed and convex. We denote
the projection of the origin on the set $A(x)$ by $A^0(x)$.
For $\eta > 0$, we denote by $J_{\eta}^A= (Id_{\mathcal{H}} +
\eta A)^{-1}$ the resolvent of $A$.
It is
well-known that
this operator
is single-valued and defined on all of $\mathcal{H}$, furthermore
$J^A_{\eta}(x)\in D(A)$, for all $x\in \mathcal{H}$.
Let $A: D(A)\subset \mathcal{H}\rightrightarrows \mathcal{H}$ and
$B: D(B)\subset \mathcal{H}\rightrightarrows \mathcal{H}$ be two
maximal monotone operators. Then, $dis(A, B)$ is the Vladimirov's
pseudo-distance between $A$ and $B$, defined as follows
\begin{equation}\label{2.1} dis(A,B)=\sup\bigg\{\frac{\langle y-y', x'-x\rangle}{1+\|
y\|+\| y'\|}:\;(x,y)\in Gr(A),\;(x',y')\in
Gr(B)\bigg\}.\end{equation}
The following lemmas are also needed. We refer to \cite{KM} for
their proofs.
\begin{lem}\label{lem2.1}
Let $A$ be a maximal monotone operator of $\mathcal{H}$. If $x\in
\overline{D(A)}$ and $y\in \mathcal{H}$ are such that
$$ \langle A^0(z)-y, z-x\rangle\geq 0\;\;\forall z\in D(A),$$
then $x\in D(A)$ and $y\in A(x)$.
\end{lem}
\begin{lem}\label{lem2.2}
Let $A_n$ $(n\in \mathbb{N})$ and $A$ be maximal monotone operators
of $\mathcal{H}$ such that $dis(A_n, A)\to 0$. Suppose also that
$x_n\in D(A_n)$ with $x_n\to x$ and $y_n\in A_n(x_n)$ with $y_n\to
y$ weakly for some $x, y\in E$. Then $x\in D(A)$ and $y\in A(x)$.
\end{lem}
\begin{lem}\label{lem2.3}
Let $A$ and $B$ be maximal monotone operators of $\mathcal{H}$. Then\\
1) for $\eta>0$ and $x\in D(A)$
$$ \|x-J_{\eta}^{B}(x)\| \leq \eta\|A^0(x)\|+dis(A,B)+\sqrt{\eta\big(1+\|A^0(x)\|\big)dis(A, B)}.$$
2) For $\eta>0$ and $x, x'\in E$
$$ \|J_{\eta}^{A}(x)-J_{\eta}^{A}(x')\|\leq
\|x-x'\|.$$
\end{lem}
\begin{lem}\label{lem2.4}
Let $A_n$ $(n\in \mathbb{N})$ and $A$ be maximal monotone operators
of $\mathcal{H}$ such that $dis(A_n, A)\to 0$ and $\|A^0_n(x)\|\leq
c(1+\|x\|)$ for some $c>0$, all $n\in \mathbb{N}$ and $x\in D(A_n)$.
Then for every $z\in D(A)$ there exists a sequence $(\zeta_n)$ such
that
\begin{equation*}\label{2.2}\zeta_n\in
D(A_n),\;\;\;\zeta_n\to z\;\;\textmd{and}\;\;A_n^0(\zeta_n)\to
A^0(z).\end{equation*}
\end{lem}
\vskip2mm
We will establish our main result under the following hypotheses.\\
$(H_1)$ There exists a function $\rho:I\longrightarrow [0, +\infty[$
which is right continuous on $[0, T[$ and nondecreasing with
$\rho(0) = 0$ and $\rho(T)<+\infty$ such that
\begin{equation*}dis(A(t), A(s))\leq d\rho(]s,
t])=\rho(t)-\rho(s)\;\;\textmd{for}\;\;0\leq s\leq t\leq
T.\end{equation*} $(H_2)$ There exists a nonnegative real constant
$c$ such that
\begin{equation*}\|A^0(t,x)\|\leq c(1+\| x\|)\;\;\textmd{for}\;\;t\in I,\;x\in
D(A(t)).\end{equation*} $(H_3)$ There exists a nonnegative real
constant $m$ such that
$$ \|f(t,x)\|\leq m(1+\|x\|)\;\;\;\forall (t,x)\in I\times
\mathcal{H}.$$ $(H_4)$ There exists $\alpha\in L^1(I, \mathbb{R};
\lambda)$ such that
$$ \|f(t,x)-f(t,y)\|\leq \alpha(t)\|x-y\|\;\;\;\forall(t,x,y)\in I\times
\mathcal{H}\times \mathcal{H}.$$
\section{Main result}
Now we are able to state our main theorem.
\begin{thm}\label{Theorem 3.2} Let for
every $t\in I$, $A(t):D(A(t))\subset \mathcal{H}\rightrightarrows
\mathcal{H}$ be a maximal monotone operator satisfying $(H_1)$ and
$(H_2)$. Let $f:I\times \mathcal{H}\longrightarrow \mathcal{H}$ be
such that for any fixed $x\in \mathcal{H}$, $f(\cdot, x)$ is
$\big(\mathcal{B}(I),\mathcal{B}(\mathcal{H})\big)$-measurable and
$(H_3)$, $(H_4)$ are satisfied.
\\Then for any $u_0\in D(A(0))$, there exists a unique
BVRC solution $ u :I \longrightarrow \mathcal{H}$ to the problem
$$ (P_f)
\begin{cases}
-\displaystyle\frac{du}{d\nu }(t)\in A(t) u(t)+ f(t,u(t))
\frac{d\lambda} {d\nu}(t) \;\;\;\nu-a.e.\,t\in
I;\\
u(t)\in D(A(t))\;\;\;\forall t\in I;\\
u(0)=u_0.
\end{cases}
$$
\end{thm}
\proof Following Castaing et al \cite{CM}, we can choose a sequence
$(\varepsilon_n)\subset ]0, 1]$ such that $\varepsilon_n \downarrow
0$ and a partition $0=t_0^n<t_1^n<\cdots<t_{q_n}^n=T$ of $I$, for
which we have
\begin{equation}\label{3.1}
\nu(]t^n_i, t^n _{i+1}])=|t_{i+1}^n-t_i^n|+d\rho(]t_i^n,
t_{i+1}^n])\leq
\varepsilon_n\;\;\;\textmd{for}\;i=0,...,q_n-1.\end{equation} For
each $i\in \{0,...,q_n-1\}$, put $I_i^n=]t_i^n, t_{i+1}^n]$,
\begin{equation}\label{3.2} \delta_{i+1}^n=d\rho(]t_i^n,
t_{i+1}^n]), \;\;\;\; \eta^n_{i+1} = t ^n_{i+1}-t ^n_i, \;\;\;\;
\beta ^n_{i+1} = \nu(]t^n_i, t^n _{i+1}]).\end{equation} For every
$n\in \mathbb{N}$, put $u_0^n=u_0\in D(A(0))$, and let us define by
induction, the sequence $(u_i^n)_{0\leq i\leq q_n-1}$ such that
\begin{equation}\label{3.3} u_{i+1}^n= J^n_{i+1}
\Big(u_i^n-\int_{t_i^n}^{t_{i+1}^n} f(s,
u_i^n)d\lambda(s)\Big)\end{equation} where
$J^n_{i+1}:=J^{A(t_{i+1}^n)}_{\beta^n_{i+1}} =\big(Id_{\mathcal{H}}+
\beta^n_{i+1} A(t_{i+1}^n)\big)^{-1}$.
\\ Clearly, by the
properties of the resolvent, we have $u^n_{i+1} \in
D(A(t_{i+1}^n))$ and
\begin{equation}\label{3.5} -\frac{1} {\beta^n_{i+1}}\Big(u^n_{i+1}-
u_i^n+\int_{t_i^n}^{t_{i+1}^n} f(s, u_i^n)d\lambda(s)\Big) \in A(
t_{i+1}^n) u^n _{i+1}. \end{equation} For $t\in [t^n_i, t
^n_{i+1}[$, $i = 0,...,q_n-1$, set
\begin{equation}\label{3.7}
u_n(t)= u_i^n + \frac{ \nu(]t^n_i, t]) }{ \nu(]t^n_i, t
^n_{i+1}])} \Big(u_{i+1}^n-u_i^n+\int_{t_i^n}^{t_{i+1}^n} f(s,u_i^n)
d\lambda(s)\Big) - \int_{t_i^n}^{t}
f(s,u_i^n)d\lambda(s),\end{equation} and $u_n(T)=u_{q_n}^n$. So that
$u_n(t_i^n)=u_i^n$ and $u_n$ is a BVRC mapping on $I$.
\vskip2mm{\bf Step 1.}
We prove in this step that the sequence $(u_n)$ is bounded in norm and
in variation.\\Using Lemma \ref{lem2.3}, we have from
\eqref{3.3}, $(H_1)$, $(H_2)$ and $(H_3)$, for $i=0,1,\cdots,
q_n-1$,
\begin{eqnarray*}
\| u_{i+1}^n-u_i^n\| &\leq& \Big\|J_{i+1}^n\Big(u_i^n-
\int_{t_i^n}^{t_{i+1}^n}
f(s,u_i^n)d\lambda(s)\Big)-J_{i+1}^n(u_i^n)\Big\|+\big\|J_{i+1}^n(u_i^n)-u_i^n\big\|\\
&\leq& \int_{t_i^n}^{t_{i+1}^n} \|f(s,u_i^n)\|
d\lambda(s)+\beta_{i+1}^n\|A^0(t_i^n,
u_i^n)\|+dis\big(A(t_{i+1}^n),A(t_i^n)\big)\\&+&\sqrt{\beta_{i+1}^n\big(1+\|A^0(t_i^n,
u_i^n)\|\big)dis\big(A(t_{i+1}^n),A(t_i^n)\big)}\\
&\leq& m(1+\|u_i^n\|)\beta_{i+1}^n
+\big(1+c(1+\|u_i^n\|)\big)\beta_{i+1}^n+
\sqrt{\big(1+c(1+\|u_i^n\|)\big)(\beta_{i+1}^n)^2}\\
&\leq& m(1+\|u_i^n\|)\beta_{i+1}^n
+\big(1+c(1+\|u_i^n\|)\big)\beta_{i+1}^n
+\big(1+c(1+\|u_i^n\|)\big)\beta_{i+1}^n,
\end{eqnarray*}
that is, \begin{equation}\label{3.8}\| u_{i+1}^n-u_i^n\|\leq \big(
(2c+m)\|u_i^n\|+2(1+c)+m\big)\beta_{i+1}^n,\end{equation} which
entails
$$ \|u_{i+1}^n\|\leq
\big(1+(2c+m)\beta_{i+1}^n\big)\|u_i^n\|+\big(2(1+c)+m\big)\
\beta_{i+1}^n.$$ Applying Lemma \ref{lem2.5}, we get
\begin{eqnarray*}\|u_i^n\|&\leq&
\Big(\|u_0\|+\big(2(1+c)+m\big)\sum_{j=0}^{i-1}
\beta_{j+1}^n\Big)\exp\Big((2c+m)\sum_{j=0}^{i-1}
\beta_{j+1}^n\Big)\\
&\leq&\Big(\|u_0\|+\big(2(1+c)+m\big)\nu(]0,T])\Big)\exp\Big((2c+m)
\nu(]0,T])\Big)=:m_1, \end{eqnarray*} using this last estimate in
\eqref{3.8}, we obtain
$$\|u_{i+1}^n-u_i^n\|\leq\Big((2c+m) m_1+2(1+c)+m\Big)\beta_{i+1}^n=:m_2\beta_{i+1}^n.$$
Consequently,
\begin{equation}\label{3.9}\|u_i^n\|\leq M\;\;\forall\, 0\leq i \leq
q_n\;\;\textmd{and}
\;\;\|u_{i+1}^n-u_i^n\|\leq M\nu(]t_i^n,
t_{i+1}^n])\;\;\forall\, 0\leq i<q_n,
\end{equation}
where $M=\max(m_1, m_2)$. Now, for $t\in [t_i^n, t_{i+1}^n[$, we
have from \eqref{3.1}, \eqref{3.7} and \eqref{3.9},
\begin{eqnarray}\label{3.10}
\|u_n(t)-u_i^n\| &=& \Big\| \frac{ \nu(]t^n_i, t])}{ \nu(]t^n_i, t
^n_{i+1}])} \Big(u_{i+1}^n-u_i^n+\int_{t_i^n}^{t_{i+1}^n} f(s,
u_i^n) d\lambda(s)\Big) - \int_{t_i^n}^{t}
f(s, u_i^n)d\lambda(s)\Big\|\nonumber\\
&\leq& \|u_{i+1}^n-u_i^n\|+2m(1+\|u_i^n\|)\delta_{i+1}^n \leq
M\varepsilon_n+2m(1+M) \varepsilon_n
=:M_1\varepsilon_n.\end{eqnarray} This last relation gives us, for
all $n\in \mathbb{N}$,
\begin{equation*}\|u_n(t)\|\leq M+M_1\varepsilon_n
\leq M+M_1=:M_2\;\;\;\forall t\in I,\end{equation*} that is
\begin{equation}\label{3.11}
\sup_n\|u_n\|=\sup_n\big(\sup_{t\in I}\|u_n(t)\|\big)\leq M_2.
\end{equation}
{\bf Step 2.} Convergence of the sequences $(u_n)$ and $(\frac{du_n}{d\nu})$.\\
Define the functions $\phi_n,\;\theta_n: I\longrightarrow I$ by
$$\phi_n(t)=t_{i}^n,\;\;\;\theta_n(t)=t_{i+1}^n\;\;\textmd{for}\;\; t\in ]t_i^n,
t_{i+1}^n],\;i=0,1,...,q_n-1,\;\; \textmd{and}\;\;
\phi_n(0)=\theta_n(0)=0,
$$ and set for all $t\in I$,
\begin{equation*}
B_n(t)=\sum_{i=0}^{q_n-1}\frac{1}{\nu(]t_i^n, t_{i+1}^n])}\Big(
u_{i+1}^n-u_i^n+\int_{t_i^n}^{t_{i+1}^n}
f(s,u_i^n)d\lambda(s)\Big)\chi_{]t_i^n, t_{i+1}^n]}(t),
\end{equation*}
where $\chi_{J}(.)$ is the characteristic function of the set
$J\subset I$. Whence, by \eqref{3.7}, it is clear that
\begin{equation*}
u_n(t)=u_0+\int_{]0,t]} B_n(s)d\nu(s)-\int_{]0,t]} f(s,
u_n(\phi_n(s)))d\lambda(s).
\end{equation*}
Since $\frac{d\lambda}{d\nu}$ is a density of $\lambda$ w.r.t $\nu$,
then by \eqref{7}, we get for every $t\in I$,
\begin{equation*}
u_n(t)=u_0+\int_{]0,t]} \Big(B_n(s)-f(s,
u_n(\phi_n(s)))\frac{d\lambda}{d\nu}(s)\Big)d\nu(s),
\end{equation*}
so that, $\frac{du_n}{d\nu}$ is a density of the vector measure
$du_n$ w.r.t $\nu$ and for $\nu$-almost every $t\in I$,
\begin{equation*}\label{8}\frac{du_n}{d\nu}(t)=B_n(t)-f(t,
u_n(\phi_n(t)))\frac{d\lambda}{d\nu}(t),\end{equation*} that is, by
the definition of $B_n$, for $\nu$-almost every $t\in I$,
\begin{equation}\label{8}
\frac{du_n}{d\nu}(t)+f(t,
u_n(\phi_n(t)))\frac{d\lambda}{d\nu}(t)=\sum_{i=0}^{q_n-1}\frac{1}{\nu(]t_i^n,
t_{i+1}^n])}\Big( u_{i+1}^n-u_i^n+\int_{t_i^n}^{t_{i+1}^n}
f(s,u_i^n)d\lambda(s)\Big)\chi_{]t_i^n, t_{i+1}^n]}(t).
\end{equation}
From $(H_3)$, \eqref{3.9} and \eqref{8}, it results that
\begin{equation}\label{9}
\Big\|\frac{du_n}{d\nu}(t)+f(t,
u_n(\phi_n(t)))\frac{d\lambda}{d\nu}(t)\Big\|\leq
M+m(1+M)=:M_3\;\;\;\nu-a.e.\,t\in I.
\end{equation}
On the other hand, we know that $\nu=d\rho+\lambda$, whence
$\frac{d\lambda}{d\nu}(t)\leq 1$. So that, by $(H_3)$ and
\eqref{3.9}, it follows that
\begin{equation}\label{10}
\Big\|f(t,
u_n(\phi_n(t)))\frac{d\lambda}{d\nu}(t)\Big\|\leq
m(1+M)=:M_4\;\;\;\nu-a.e.\,t\in I,
\end{equation}
consequently, there is a Borel subset $J\subset I$ with $\nu(J)=0$,
such that
\begin{equation}\label{11}
\Big\|\frac{du_n}{d\nu}(t)\Big\|\leq M_3+M_4=:M_5\;\;\;\forall t\in I\setminus J.
\end{equation}
Now, observe that by \eqref{3.5} and \eqref{8}, for each $n\in \mathbb{N}$, there is a
Borel subset $J_n\subset I$ with $\nu(J_n)=0$ such that
\begin{equation}\label{12}
\frac{du_n}{d\nu}(t)+f(t,
u_n(\phi_n(t)))\frac{d\lambda}{d\nu}(t)\in
-A(\theta_n(t))u_n(\theta_n(t))\;\;\;\forall t\in I\setminus J_n,
\end{equation}
moreover,
\begin{equation}\label{13}
u_n(\theta_n(t))\in D\big(A(\theta_n(t))\big)\;\;\;\forall t\in I.
\end{equation}
We will show in the following, that $(u_n)$ is a Cauchy sequence.\\
Let $m, n\in \mathbb{N}$. Set for all $t\in I$,
$f_n(t):=f(t,u_n(\phi_n(t)))\frac{d\lambda}{d\nu}(t)$. From $(H_1)$
and relations \eqref{12}, \eqref{9}, we have for $\nu$-a.e. $t\in
I$,
\begin{eqnarray*}
&&\Big\langle u_n(\theta_n(t))-u_m(\theta_m(t)),
\frac{du_n}{d\nu}(t)+f_n(t)-\frac{du_m}{d\nu}(t)-f_m(t)\Big\rangle\\
&\leq&
\Big(1+\big\|\frac{du_n}{d\nu}(t)+f_n(t)\big\|+\big\|\frac{du_m}{d\nu}(t)
+f_m(t)\big\|\Big)
dis (A(\theta_n(t)), A(\theta_m(t)))\\
&\leq& (1+2M_3)|\rho(\theta_n(t))-\rho(\theta_m(t))|\leq
(1+2M_3)\big(|\rho(\theta_n(t))-t|+|\rho(\theta_m(t))-t|\big)\\
&\leq& (1+2M_3)\big(d\rho(]t, \theta_n(t)])+d\rho(]t,
\theta_m(t)])\big).
\end{eqnarray*}
On the other hand, from \eqref{9} and \eqref{11},
\begin{eqnarray*}
&&\Big\langle u_n(t)-u_n(\theta_n(t))-u_m(t)+u_m(\theta_m(t)),
\frac{du_n}{d\nu}(t)+f_n(t)-\frac{du_m}{d\nu}(t)-f_m(t)\Big\rangle\\
&\leq&
2M_3\big(\|u_n(t)-u_n(\theta_n(t))\|+\|u_m(t)-u_m(\theta_m(t))\|\big)\\
&\leq& 2M_3M_5\big(\nu(]t, \theta_n(t)])+\nu(]t, \theta_m(t)])\big).
\end{eqnarray*}
Combining these last relations, we get for $\nu$-a.e. $t\in I$,
\begin{eqnarray}\label{14}
&&\Big\langle u_n(t)-u_m(t),
\frac{du_n}{d\nu}(t)-\frac{du_m}{d\nu}(t)\Big\rangle\nonumber\\
&=& \Big\langle u_n(t)-u_n(\theta_n(t))-u_m(t)+u_m(\theta_m(t)),
\frac{du_n}{d\nu}(t)+f_n(t)-\frac{du_m}{d\nu}(t)-f_m(t)\Big\rangle\nonumber\\
&+& \Big\langle u_n(\theta_n(t))-u_m(\theta_m(t)),
\frac{du_n}{d\nu}(t)+f_n(t)-\frac{du_m}{d\nu}(t)-f_m(t)\Big\rangle\nonumber\\
&+&\Big\langle u_n(t)-u_m(t),-f_n(t)+f_m(t)\Big\rangle\nonumber\\
&\leq& 2M_3M_5\big(\nu(]t, \theta_n(t)])+\nu(]t, \theta_m(t)])\big)
+(1+2M_3)\big(d\rho(]t, \theta_n(t)])+d\rho(]t, \theta_m(t)])\big)\nonumber\\
&+&\Big\langle u_n(t)-u_m(t),-f_n(t)+f_m(t)\Big\rangle\nonumber\\
&\leq& c_1\big(\sigma_n(t)+\sigma_m(t)\big)+\Big\langle
u_n(t)-u_m(t),-f_n(t)+f_m(t)\Big\rangle,
\end{eqnarray}
where, $c_1:=2M_3M_5+(1+2M_3)$ and $\sigma_n(t)=\nu(]t,
\theta_n(t)])$. Let us examine the last term in the preceding
estimate. For all $t\in I$, we have by $(H_4)$
\begin{eqnarray*}
&&\Big\langle u_n(t)-u_m(t),-f_n(t)+f_m(t)\Big\rangle\\&=&
\Big\langle
u_n(t)-u_m(t),-f(t,u_n(\phi_n(t)))\frac{d\lambda}{d\nu}(t)+f(t,u_m(\phi_m(t)))\frac{d\lambda}{d\nu}(t)\Big\rangle\\
&\leq&\|u_n(t)-u_m(t)\|\big(\alpha(t)\frac{d\lambda}{d\nu}(t)\|u_n(\phi_n(t))-u_m(\phi_m(t))\|\big)\\
&\leq&\alpha(t)\frac{d\lambda}{d\nu}(t)\|u_n(t)-u_m(t)\|\big(\|u_n(\phi_n(t))-u_n(t)\|+\|u_n(t)-u_m(t)\|+\|u_m(t)-u_m(\phi_m(t))\|\big)\\
&\leq&
\alpha(t)\frac{d\lambda}{d\nu}(t)\|u_n(t)-u_m(t)\|^2+\gamma_{n,m}(t),
\end{eqnarray*}
where,
\begin{eqnarray*}&&\alpha(t)\frac{d\lambda}{d\nu}(t)\|u_n(t)-u_m(t)\|\big(\|u_n(\phi_n(t))-u_n(t)\|+\|u_m(t)-u_m(\phi_m(t))\|\big)\\
&\leq&M_5\alpha(t)\|u_n(t)-u_m(t)\|\big(\nu(]\phi_n(t),
t])+\nu(]\phi_m(t),
t])\big)\frac{d\lambda}{d\nu}(t)=:\gamma_{n,m}(t).
\end{eqnarray*}
Replacing in \eqref{14}, its results that for $\nu$-a.e. $t\in I$,
\begin{eqnarray}\label{15}
\Big\langle u_n(t)-u_m(t),
\frac{du_n}{d\nu}(t)-\frac{du_m}{d\nu}(t)\Big\rangle
&\leq&\alpha(t)\frac{d\lambda}{d\nu}(t)\|u_n(t)-u_m(t)\|^2+\Delta_{n,m}(t),
\end{eqnarray}
where $
\Delta_{n,m}(t):=c_1\big(\sigma_n(t)+\sigma_m(t)\big)+\gamma_{n,m}(t)$.
Since $\rho$ is right continuous, it is clear that for all $t\in I$,
$\sigma_n(t)\to 0$ as $n\to\infty$. On the other hand, we have for
all $t\in I$, $\nu(]\phi_n(t), t])\frac{d\lambda}{d\nu}(t)\to\nu(\{
t\})\frac{d\lambda}{d\nu}(t)=0$ thanks to \eqref{5}. This shows
that $\gamma_{n,m}(t)\to 0$ as $n,m\to\infty$ since from
\eqref{3.11}, $\|u_n(t)\|\leq M_2$. We conclude that
\begin{equation}\label{16}\Delta_{n,m}(t)\longrightarrow 0\;\;\; \textmd{as}\;\; n,m\to\infty,\end{equation} and by the Lebesgue
dominated convergence theorem,
$$ \int_{]0, T]} \Delta_{n,m}(t)\,d\nu(t)\longrightarrow 0\;\;\;
\textmd{as}\;\;n,m\to\infty.$$ Applying Lemma \ref{Lem2.1}, we get
from \eqref{15}, taking account that $\|u_n(0)-u_m(0)\|=0$,
\begin{eqnarray*}
\|u_n(t)-u_m(t)\|^2\leq 2\int_{]0,t]}
\big(\alpha(s)\frac{d\lambda}{d\nu}(s)\|u_n(s)-u_m(s)\|^2\big)d\nu(s)+2\int_{]0,T]}\Delta_{n,m}(s)\,d\nu(s).
\end{eqnarray*}
Since (see \eqref{5})
$\alpha(s)\frac{d\lambda}{d\nu}(s)\nu(\{s\})=0<1$, we get by Lemma
\ref{lem2.6},
\begin{eqnarray*}
\|u_n(t)-u_m(t)\|^2&\leq&\Big(2\int_{]0,T]}\Delta_{n,m}(s)\,d\nu(s)\Big)\exp\Big(2\int_{]0,t]}
\big(\alpha(s)\frac{d\lambda}{d\nu}(s)\big)d\nu(s)\Big)\\&=&
\Big(2\int_{]0,T]}\Delta_{n,m}(s)\,d\nu(s)\Big)\exp\Big(2\int_{]0,t]}
\alpha(s)d\lambda(s)\Big),
\end{eqnarray*}
so that
$$\sup_{t\in I} \|u_n(t)-u_m(t)\|^2\leq\Big(2\int_{]0,T]}\Delta_{n,m}(s)\,d\nu(s)\Big)\exp\Big(2\int_{]0,T]}
\alpha(s)d\lambda(s)\Big).$$ We conclude, by \eqref{16}, that
$(u_n)$ is a Cauchy sequence w.r.t the norm of uniform convergence
on the space of $\mathcal{H}$-valued bounded mappings defined on
$I$. This implies that $(u_n)$ converges uniformly on $I$ to some
mapping $u$.
Now, observe that the sequence $(\frac{du_n}{d\nu})$ is bounded in
$L^2(I, \mathcal{H}; \nu)$ due to relation \eqref{11}, so that it
converges weakly in $L^2(I, \mathcal{H}; \nu)$ to some mapping $w\in
L^2(I, \mathcal{H}; \nu)$. In particular, for all $t\in I$
\begin{equation*}
\int_{]0, t]} \frac{du_n}{d\nu}(s)\, d\nu(s)\longrightarrow\int_{]0,
t]} w(s)\, d\nu(s)\;\;\;\textmd{weakly in }\;\mathcal{H}.
\end{equation*}
Since $\frac{du_n}{d\nu}$ is a density of the vector measure $du_n$
w.r.t the measure $\nu$, we have for all $t\in I$,
\begin{eqnarray*}
u_n(t)=u_0+\int_{]0, t]} \frac{du_n}{d\nu}(s)\,d\nu(s)
\end{eqnarray*}
and since $(u_n(t))$ converges strongly in $\mathcal{H}$ and then
weakly to $u(t)$, we deduce by what precedes that
\begin{equation*}
u(t)=u_0+\int_{]0, t]} w(s)\,d\nu(s)\;\;\;\forall t\in I.
\end{equation*}
This shows that $u$ is right continuous with bounded variation and
for $\nu$-a.e. $t\in I$, $\frac{du}{d\nu}=w$. Consequently,
$(\frac{du_n}{d\nu})$ converges weakly in $L^2(I, \mathcal{H}; \nu)$
to $\frac{du}{d\nu}$.
\vskip2mm
{\bf Step 3.} Existence of solution.\\
First, observe that from \eqref{13}, $ u_n(\theta_n(t))\in
D\big(A(\theta_n(t))\big)$ for all $t\in I $. Also, we have that
for all $t\in I$,
$$dis\big(A(\theta_n(t)),A(t)\big)\leq |\rho(\theta_n(t))-\rho(t)|
\longrightarrow 0\;\;\;\textmd{as}\;\;n\to\infty,$$ and from
\eqref{11},
$$\|u_n(\theta_n(t))-u_n(t)\|\leq M_5\nu(]t, \theta_n(t)])\longrightarrow 0\;\;\;\textmd{as}\;\;n\to\infty,$$
so that, \begin{equation}\label{conv}\|u_n(\theta_n(t))-u(t)\|\leq
\|u_n(\theta_n(t))-u_n(t)\|+\|u_n(t)-u(t)\|\longrightarrow
0\;\;\;\textmd{as}\;\;n\to\infty.\end{equation} On the other hand,
by \eqref{3.11} and $(H_2)$, we have that the sequence
$\big(A^0\big(\theta_n(t), u_n(\theta_n(t))\big)\big)$ is bounded in
$\mathcal{H}$, and hence it is weakly relatively compact. Using all
these facts, we conclude by Lemma \ref{lem2.2}, that $u(t)\in
D(A(t))$ for all $t\in I$.
Next, since the sequence $(\frac{du_n}{d\nu})$ converges weakly in
$L^2(I, \mathcal{H}; \nu)$ to $\frac{du}{d\nu}$, by Mazur's theorem,
there is a sequence $(\xi_n)$ such that for each $n\in\mathbb{N}$,
$\xi_n\in co\{\frac{du_k}{d\nu};\;k\geq n\}$, and $(\xi_n)$
converges strongly in $L^2(I, \mathcal{H}; \nu)$ to
$\frac{du}{d\nu}$. So that, there is a subsequence $(\xi_{n_j})$,
which converges $\nu$-almost every where to $\frac{du}{d\nu}$. This
means the existence of a Borel subset $J'\subset I$, with
$\nu(J')=0$ and for $t\in I\setminus J'$,
$$\xi_{n_j}(t)\longrightarrow \frac{du}{d\nu}(t)\in\bigcap_{n}
\overline{co} \big\{\frac{du_{k}}{d\nu};\;k\geq n\big\}.$$ This
implies, by \eqref{co}, that for any fixed $\eta\in \mathcal{H}$,
\begin{equation}\label{17}
\big\langle \frac{du}{d\nu}(t), \eta
\big\rangle\leq\limsup_{n\to\infty}\big\langle \frac{du_n}{d\nu}(t),
\eta \big\rangle.
\end{equation}
To prove that $u$ is a solution to our considered problem, we will
use Lemma \ref{lem2.1}. Since $u(t)\in D(A(t))$ for all $t\in I$, we
have to show that for $\nu$-almost every fixed $t\in I$ and for any
$z\in D(A(t))$
\begin{equation*}
\big\langle
A^0(t,z)+\frac{du}{d\nu}(t)+f(t,u(t))\frac{d\lambda}{d\nu}(t),z-u(t)\big\rangle\geq
0.
\end{equation*}
Indeed, let $t\in I$. By $(H_2)$, using Lemma \ref{lem2.4}, we can
ensure the existence of a sequence $(\zeta_n)_n$, such that
\begin{equation}\label{18}
\zeta_n\in D(A(\theta_n(t))),\;\;\;\zeta_n\longrightarrow
z\;\;\textmd{and}\;\; A^0\big(\theta_n(t),
\zeta_n\big)\longrightarrow A^0(t,z).
\end{equation}
Since $A(\theta_n(t))$ is monotone, using \eqref{12}, we have for
each $n\in \mathbb{N}$ and $t\in I\setminus J_n$,
\begin{equation}\label{19}
\big\langle u_n(\theta_n(t))-\zeta_n,
\frac{du_n}{d\nu}(t)+f(t,u_n(\phi_n(t)))\frac{d\lambda}{d\nu}(t)+A^0(\theta_n(t),
\zeta_n)\big\rangle\leq 0.
\end{equation}
On the other hand, we have from $(H_4)$ and \eqref{11}
\begin{eqnarray*}
&&\|f(t,u_n(\phi_n(t))\frac{d\lambda}{d\nu}(t)-f(t,u(t))\frac{d\lambda}{d\nu}(t)\|\leq\alpha(t)\|u_n(\phi_n(t))-u(t)\|\frac{d\lambda}{d\nu}(t)\nonumber\\
&\leq&\alpha(t)(\|u_n(\phi_n(t))-u_n(t)\|+\|u_n(t)-u(t)\|)\frac{d\lambda}{d\nu}(t)\nonumber\\
&\leq&\alpha(t)(M_5\nu(]\phi_n(t),
t])+\|u_n(t)-u(t)\|)\frac{d\lambda}{d\nu}(t),
\end{eqnarray*}
whence by the uniform convergence of $(u_n)$ to $u$ and relation
\eqref{5}, it results that for all $t\in I$,
\begin{equation}\label{lip}
f(t,u_n(\phi_n(t))\frac{d\lambda}{d\nu}(t)\longrightarrow
f(t,u(t))\frac{d\lambda}{d\nu}(t)\;\;\;\textmd{as}\;\;n\to\infty.
\end{equation}
Now, let $t\in I\setminus \big(J\cup J'\cup(\underset{n}{\bigcup} J_n)\big)$. From \eqref{17}, we have
\begin{equation*}
\big\langle \frac{du}{d\nu}(t), u(t)-z\big\rangle\leq\limsup_{n\to\infty}\big\langle \frac{du_n}{d\nu}(t),
u(t)-z\big\rangle,
\end{equation*}
and since from \eqref{11} and \eqref{19}, we have for each $n\in
\mathbb{N}$,
\begin{eqnarray*}
&&\big\langle \frac{du_n}{d\nu}(t),
u(t)-z\big\rangle\leq\big\langle
\frac{du_n}{d\nu}(t),u(t)-u_n(\theta_n(t))\big\rangle+ \big\langle\frac{du_n}{d\nu}(t), u_n(\theta_n(t))-\zeta_n\big\rangle+\big\langle
\frac{du_n}{d\nu}(t),\zeta_n-z\big\rangle\\&\leq&
M_5\big(\|u(t)-u_n(\theta_n(t))\|+\|\zeta_n-z\|\big)+\big\langle
\frac{du_n}{d\nu}(t),u_n(\theta_n(t))-\zeta_n\big\rangle
\\&\leq&
M_5\big(\|u(t)-u_n(\theta_n(t))\|+\|\zeta_n-z\|\big)+\big\langle f(t,u_n(\phi_n(t)))\frac{d\lambda}{d\nu}(t)+A^0(\theta_n(t),
\zeta_n),\zeta_n-u_n(\theta_n(t))\big\rangle
\end{eqnarray*}
we conclude, using \eqref{conv}, \eqref{18} and \eqref{lip},
that\begin{equation} \big\langle \frac{du}{d\nu}(t),
u(t)-z\big\rangle\leq\big\langle
f(t,u(t))\frac{d\lambda}{d\nu}(t)+A^0(t, z),z-u(t)\big\rangle,
\end{equation}
which, with the fact that $u(t)\in D(A(t))$ and $u(0)=u_0$, means
that $u$ is a BVRC solution to our problem.
{\bf Step 4.} Uniqueness of the solution.\\
Let $u$, $v$ be two BVRC solutions of $(P_f)$. By the monotonicity
of the operator $A(t)$, we get for $\nu$-a.e. $t\in I$,
\begin{eqnarray*}
\big\langle
-\frac{du}{d\nu}(t)-f(t,u(t))\frac{d\lambda}{d\nu}(t)+\frac{dv}{d\nu}(t)+f(t,v(t))\frac{d\lambda}{d\nu}(t),
u(t)-v(t)\big\rangle\geq 0,
\end{eqnarray*}
that is
\begin{eqnarray*}
\big\langle \frac{du}{d\nu}(t)-\frac{dv}{d\nu}(t),
u(t)-v(t)\big\rangle\leq \big\langle
f(t,v(t))\frac{d\lambda}{d\nu}(t)-f(t,u(t))\frac{d\lambda}{d\nu}(t),
u(t)-v(t)\big\rangle,
\end{eqnarray*}
and by $(H_4)$, we obtain
\begin{eqnarray*}
\big\langle \frac{du}{d\nu}(t)-\frac{dv}{d\nu}(t),
u(t)-v(t)\big\rangle\leq\alpha(t)\frac{d\lambda}{d\nu}(t)\|u(t)-v(t)\|^2.
\end{eqnarray*}
Then, an application of Lemma \ref{Lem2.1} taking account that
$u(0)=v(0)=u_0$, gives us for all $t\in ]0, T]$,
\begin{eqnarray*}
\|u(t)-v(t)\|^2\leq 2\int_{]0, t]}
\alpha(s)\frac{d\lambda}{d\nu}(s)\|u(s)-v(s)\|^2 d\nu(s),
\end{eqnarray*}
from Lemma \ref{lem2.6}, it results that $\|u(t)-v(t)\|^2=0$ for all
$t\in I$, which means the uniqueness of the solution.
$\hfill\square$
|
\section{Introduction}
Based on \cite{Pitt}, complete criteria have been established in \cite{Wang} for the order preservation and positive correlation for diffusion processes corresponding to linear Fokker-Planck equations, where the order preservation links to comparison theorem in the literature of SDEs, and the positive correlation arises from statistics is known as FKG inequality due to \cite{FKG}. In the present paper we aim to extend these criteria to nonlinear Fokker-Planck equations associated with McKean-Vlasov SDEs. In the following we first recall the criteria of \cite{Wang}.
Consider the following second order differential operator on $\R^d$:
$$L=\sum_{i,j=1}^d a_{ij}\partial_i\partial_j+\sum_{i=1}^d b_i\partial_i, $$
where
$$a=(a_{ij})_{1\le i,j\le d}: \R^d \rightarrow}\def\l{\ell}\def\iint{\int}\def\gg{\gamma \R^d\otimes \R^d,\ \ b=(b_i)_{1\le i\le d}: \R^d\rightarrow}\def\l{\ell}\def\iint{\int}\def\gg{\gamma\R^d$$ are continuous, $a$ is positive definite, such that the martingale problem of $L$ is well-posed, or equivalently, there exists a unique $L$-diffusion process for any initial distribution. For any $t\ge 0$ and $\mu\in \scr P$, the space of all probability measures on $\R^d$ equipped with the weak topology, let $P^*_t\mu$ be the distribution of the $L$-diffusion process at time $t$ with initial distribution $\mu$.
We denote $x\le y$ for $x:=(x_i)_{1\le i\le d}, ~y:=(y_i)_{1\le i\le d}\in \R^d$ if $x_i\le y_i$ for any $1\le i\le d.$
Let $\B_b(\R^d)$ be the set of all bounded measurable functions on $\R^d$. Consider the class of bounded measurable increasing functions:
$$\scr U_b:=\big\{f\in\B_b(\R^d)\big| f(x)\le f(y) \ \text{for}\ x,y\in\R^d\ \text{with}\ x\le y\big\},$$
and the family of probability measures of positive correlations:
$$\scr P_{+}:=\big\{\mu\in\scr P|\, \mu(fg)\geq \mu(f)\mu(g)~~f,g\in \scr U_b\big\},$$ where we call $\mu$ satisfies the FKG inequality if $\mu\in \scr P_+$.
Moreover, we write $\mu \preceq \nu$ for any two probability measures $\mu,\nu\in \scr P,$ if $\mu(f)\leq \nu(f)$ holds for any $f\in \scr U_b(\R^d).$
Note that definitions of $\scr P_+$ and $\mu \preceq \nu$ does not change if we replace $\B_b(\R^d)$ by $C_b^k(\R^d)$ for $k\in \mathbb Z_+\cup\{\infty\}$, where $C_b^0(\R^d)=C_b(\R^d)$ denotes the set of all bounded continuous functions on $\R^d$, while when $k\ge 1$ the class $C_b^k(\R^d)$ consists of bounded functions on $\R^d$ having bounded derivatives up to order $k$.
Now, for the $L$-diffusion process, we denote $P_t^*\in \scr \scr P_t$ if $ P_t^* \scr P_+\subset \scr P_+$. In this case we call $P_t^*$ preserves positive correlations.
The following result is implied by \cite[Theorem 1.3 and Theorem 1.4]{Wang}.
\begin} \def\beq{\begin{equation}} \def\F{\scr F{thm}[\cite{Wang}]\label{CW1} $P_t^*\in \scr P_+$ for any $t\ge 0$ if and only if the following conditions hold:
\begin} \def\beq{\begin{equation}} \def\F{\scr F{enumerate}
\item[$(a)$] For any $1\le i,j\le d,$ $a_{ij}\ge 0,$ and $a_{ij}(x)$ depends only on $x_i$ and $x_j.$
\item[$(b)$] For any $1\le i\le d$, $b_i(x)$ is increasing in $x_j$ for $j\neq i.$
\end{enumerate}
\end{thm}
Next, let $\bar P_t^*$ be associated with the diffusion process generated by another operator $\bar L$ satisfying the same assumption on $L$:
$$\bar L=\sum_{i,j=1}^d \bar a_{ij}\partial_i\partial_j+\sum_{i=1}^d \bar b_i\partial_i.$$
We denote $\bar P_t^* \preceq P_t^*$ if
$$\bar P_t^*\mu \preceq P_t^* \nu, \ \ \mu,\nu\in \scr P, \mu \preceq \nu.$$
In this case we call these two diffusion processes (or $\bar P_t^*$ and $P_t^*$) order-preserving, and when $L=\bar L$, we call the $L$-diffusion process (or $P_t^*$) monotone.
The next result follows from \cite[Theorem 1.3]{Wang}.
\begin} \def\beq{\begin{equation}} \def\F{\scr F{thm}[\cite{Wang}]\label{CW2} $\bar P_t^* \preceq P_t^* $ for any $t\ge 0$, if and only if the following conditions hold:
\begin} \def\beq{\begin{equation}} \def\F{\scr F{enumerate}
\item[$(a)$] For any $1\le i,j\le d,$ $\bar a_{ij}=a_{ij},$ and $a_{ij}(x)$ depends only on $x_i$ and $x_j.$
\item[$(b)$] For any $1\le i\le d$, $\bar b_i(x)\le b_i(y) $ for $x\le y$ with $x_i=y_i$.
\end{enumerate}
\end{thm}
According to Theorem \ref{CW1} and Theorem \ref{CW2} with $L=\bar L$, we see that $P_t^*\in \scr P_+$ implies the monotonicity.
To conclude this subsection, we recall the link of diffusion processes and linear Fokker-Planck equations.
By the definition of martingale problem, we see that $\mu_t:=P_t^*\mu$ solves the following linear Fokker-Planck equation on $\scr P$:
\beq\label{LFK} \pp_t\mu_t= L^*\mu_t,\end{equation} in the sense that $\mu\in C([0,\infty);\scr P)$ satisfies
$$\mu_t(f):=\int_{\R^d} f\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D\mu_t= \mu_0(f) +\int_0^t\mu_s(Lf)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D s,\ \ t\ge 0, f\in C_0^{\infty}(\R^d).$$
On the other hand, according to the superposition principle, see \cite{TR, SV}, a solution $\mu_t$ of of \eqref{LFK} with
$$\int_0^t \mu_s\big(\|a\|+|b|\big)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D s <\infty,\ \ t\ge 0$$
is given by $\mu_t:= P_t^*\mu_0, t\ge 0.$ Since the order preservation and positive correlation are distribution properties of diffusion processes,
they are indeed properties of solutions to linear Fokker-Planck equations.
As mentioned above, we aim to extend Theorems \ref{CW1} and \ref{CW2} to nonlinear Fokker-Planck equations. Consider the Wasserstein space
$$\scr P_2=\{\mu\in \scr P: \mu(|\cdot|^2)< \infty \},$$
which is a Polish space under the Wasserstein distance
$$\W_2(\mu,\nu):=\inf_{\pi\in \scr C} \def\aaa{\mathbf{r}} \def\r{r(\mu,\nu)} \bigg(\int_{\R^d\times \R^d}|x-y|^2\pi(\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D x,\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D y)\bigg)^{\ff 1 2},$$
where $\scr C} \def\aaa{\mathbf{r}} \def\r{r(\mu,\nu)$ is the space of all couplings of $\mu$ and $\nu$.
Consider the following time-distribution dependent second order differential operators
\beq\label{N2} \begin} \def\beq{\begin{equation}} \def\F{\scr F{split} &L_{t,\mu}
:=\sum_{i,j=1}^d a_{ij}(t,\cdot,\mu)\partial_i\partial_j+\sum_{i=1}^d b_i(t,\cdot, \mu)\partial_i,\\
& \bar L_{t,\mu}
:=\sum_{i,j=1}^d \bar a_{ij}(t,\cdot,\mu)\partial_i\partial_j+\sum_{i=1}^d \bar b_i(t,\cdot, \mu)\partial_i,\end{split} \end{equation}
where
\begin} \def\beq{\begin{equation}} \def\F{\scr F{align*} &a=(a_{ij})_{1\le i,j\le d},\ \bar a=(\bar a_{ij})_{1\le i,j\le d}: [0,\infty)\times\R^d\times \scr P_2\rightarrow}\def\l{\ell}\def\iint{\int}\def\gg{\gamma \R^d\otimes \R^d,\\
&b=(b_{i})_{1\le i\le d},\ \bar b=(\bar b_{i})_{1\le i\le d}: [0,\infty)\times\R^d\times \scr P_2\rightarrow}\def\l{\ell}\def\iint{\int}\def\gg{\gamma \R^d \end{align*} are continuous.
The nonlinear Fokker-Planck equations for $L$ and $\bar L$ are formulated as
\beq\label{NFP} \pp_t \mu_t= L_{t,\mu_t}\mu_t,\ \ \pp_t \bar \mu_t= \bar L_{t,\bar\mu_t}\bar \mu_t,\ \ \mu_s=\bar\mu_s=\mu, t\ge s.\end{equation}
We call $(\mu_t,\bar\mu_t)_{t\ge s} \in C([s,\infty);\scr P_2)\times C([s,\infty);\scr P_2)$ a solution to \eqref{NFP}, if
\begin} \def\beq{\begin{equation}} \def\F{\scr F{align*} & \mu_t(f)= \mu(f)+\int_s^t \mu(L_{r,\mu_r}f)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r,\\
&\bar \mu_t(f)= \mu(f)+\int_s^t\bar \mu(\bar L_{r,\bar \mu_r}f)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r,\ \ t\ge s, f\in C_0^\infty(\R^d).\end{align*}
The nonlinear Fokker-Planck equations can be characterized by distribution dependent SDEs, also called McKean-Vlasov or mean field SDEs.
Let $W_t$ be a $d$-dimensional Brownian motion on a complete filtration probability space $(\OO, \{\F_t\}_{t\ge }, \P)$. Consider the distribution dependent SDEs
\beq\label{DDSDE} \begin} \def\beq{\begin{equation}} \def\F{\scr F{split} &\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D X_t= b(t,X_t,\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t +\ss{2 a(t,X_t,\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t})}\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_t, \\
& \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \bar X_t= \bar b(t,\bar X_t,\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\bar X_t})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t +\ss{2 \bar a(t,\bar X_t,\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\bar X_t})}\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_t,\end{split}\end{equation}
where $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\xi}$ denotes the distribution of a random variable $\xi$.
We call these SDEs well-posed if for any $s\ge 0$ and any $\F_s$-measurable random variables $X_s$ and $\bar X_s$ with $\E [|X_s|^2+|\bar X_s|^2]<\infty,$
they have unique solutions with $(\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t})_{t\ge s}, (\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\bar X_t})_{t\ge s}\in C([s,\infty);\scr P_2).$ In this case, we denote
$$P_{s,t}^*\mu:=\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t},\ \bar P_{s,t}^*\mu:=\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\bar X_t}, ~\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_s}=\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\bar X_s}=\mu\in \scr P_2,\ \ t\ge s.$$
By It\^o's formula, $(\mu_t:= P_{s,t}^*\mu, \bar\mu_t:= \bar P_{s,t}^*\mu)$ solves \eqref{NFP}. On the other hand, by the superposition principle, see \cite{BR}, if $(\mu_t,\bar\mu_t)$ solves \eqref{NFP} with
$$\int_s^t \mu_r\big(|b(r,\cdot,\mu_r)|+\|a(r,\cdot,\mu_r)\|\big)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r+ \int_s^t \bar\mu_r\big(|\bar b(r,\cdot,\mu_r)|+\|\bar a(r,\cdot,\bar \mu_r)\|\big)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r<\infty,\ \ t\ge s,$$
then $\mu_t= P_{s,t}^*\mu$ and $\bar\mu_t=\bar P_{s,t}^*\mu, t\ge 0$. Thus, to investigate the nonlinear Fokker-Planck equations \eqref{NFP} is equivalent to study the solutions of
the distribution dependent SDEs \eqref{DDSDE}.
Unlike in the setting of standard Markov processes or linear Fokker-Planck equations, the distribution $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{(X_t)_{t\ge s}}$
on the path space is no longer determined by its time-marginals $P_{s,t}^*\mu$, so we will also consider the order preservation and positive correlations for
$$ \LL_s\mu:= \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{(X_t)_{t\ge s}},\ \ \ \bar \LL_s\mu=\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{(\bar X_t)_{t\ge s}},\ \ \ s\ge 0,\mu\in \scr P_2.$$
In Section 2, we state our main results on the order preservation for $(P_{s,t}^*, \bar P_{s,t}^*)$ and $(\LL_s, \bar\LL_s)$, as well as on the positive correlations for $\LL_s$.
To prove these results, In Section 3 we extend Theorems \ref{CW1} and \ref{CW2} to the time inhomogeneous setting which are also new in the literature.
Finally, the main results are proved in Section 4.
\section{Main results}
To ensure the well-posedness and to apply the superposition principle, we make the following assumption.
\begin} \def\beq{\begin{equation}} \def\F{\scr F{enumerate}
\item[({\bf A})] $b,\bar b, a, \bar a$ are continuous on $[0,\infty)\times\R^d\times\scr P_2$, $a$ and $\bar a$ are positive definite,
$$ \mu\big(|b(t,\cdot,\mu)|+\|a(t,\cdot,\mu)\|\big):=\int_{\R^d} \big(|b(t,\cdot,\mu)|+\|a(t,\cdot,\mu)\|\big)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D\mu$$ is locally bounded in $(t,\mu)\in [0,\infty)\times \scr P_2$, and
there exits an increasing function $K:[0,\infty)\rightarrow}\def\l{\ell}\def\iint{\int}\def\gg{\gamma [0, \infty)$ such that $b, \bar b, \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}:=\ss{2a}$ and $\bar\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}:=\ss{2\bar a}$ satisfy
\beq\label{A1}
\begin{split}
\max\Big\{&2\<b(t,x,\mu)-b(t, y,\nu), x-y\>+\|\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}(t,x,\mu)-\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}(t,x,\nu)\|_{HS}^2,\\
&2\<\bar b(t,x,\mu)-\bar b(t, y,\nu), x-y\>+\|\bar \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}(t,x,\mu)-\bar \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}(t,x,\nu)\|_{HS}^2\Big\}\\
&\leq K(t)(|x-y|^2+\W_2(\mu,\nu)^2), ~~~t\ge 0,~~x,y\in \R^d,~~\mu,\nu \in \scr P_2.
\end{split}
\end{equation}
\end{enumerate}
According to \cite{Wang1} and the superposition principle in \cite{BR}, {\bf (A)} implies the well-posedness of \eqref{DDSDE} and \eqref{NFP}, and
for any $s\ge0, \mu\in \scr P_2$,
$$\mu_t:=P_{s,t}^*\mu,\ \ \ \bar \mu_t:= \bar P_{s,t}^*\mu,\ \ t\ge s$$ give the unique solution of \eqref{NFP}.
To state the main results, we first define the order preservation and positive correlations in the present setting.
For any
$$\xi,\eta\in C_s:=C([s,\infty); \R^d), $$ we denote $\xi\le \eta$ if $\xi_t\le \eta_t$ for all $t\ge s.$ For any two probability measures $\Phi_1,\Phi_2$ on the path space $C_s$, we denote
$\Phi_1 \preceq \Phi_2$ if $\Phi_1(F)\le \Phi_2(F)$ holds for any bounded increasing function $F$ on $C_s$.
Similarly, let $\scr P_+^s$ denote the set of probability measures on $C_s$ satisfying the FKG inequality for bounded increasing functions on $C_s$.
\begin} \def\beq{\begin{equation}} \def\F{\scr F{defn} Let $t\ge s\ge 0.$
\begin} \def\beq{\begin{equation}} \def\F{\scr F{enumerate} \item[(1)] We write $\bar P_{s,t}\preceq P_{s,t}^*$, if $\bar P_{s,t}^*\mu \preceq P_{s,t}^*\nu$ holds for any $\mu,\nu\in \scr P_2$ with $\mu\preceq \nu.$
\item[(2)] We write $\bar \LL_s\preceq \LL_s$, if $\bar \LL_s\mu\preceq \LL_s\nu$ holds for any $\mu,\nu\in \scr P_2$ with $\mu\preceq \nu.$
\item[(3)] We write $P_{s,t}^*\in \scr P_+$ if $P_{s,t}^*\scr P_+\subset \scr P_+$; and $\LL_s \in \scr P_+^s$ if $\LL_s\mu\in \scr P_+^s$ holds for all $\mu\in \scr P_+.$
\end{enumerate}
\end{defn}
Obviously, $\bar \LL_s \preceq \LL_s$ for all $s\ge 0$ implies $\bar P_{s,t}\preceq P_{s,t}^*$ for all $t\ge s\ge 0$, but the inverse may not true in the nonlinear setting. Similarly,
$\LL_s\in \scr P_+^s$ implies $P_{s,t}^*\in \scr P_+$ for any $t\ge s$ but the inverse may not be true.
\subsection{Order preservation}
The following result provides sufficient conditions for the order preservation.
\begin} \def\beq{\begin{equation}} \def\F{\scr F{thm}\label{T1} Assume {\bf (A)} and the following two conditions:
\begin} \def\beq{\begin{equation}} \def\F{\scr F{enumerate} \item[$(1)$] For any $1\le i\le d$ and $s\ge 0$, $\bar b_i(s, x, \nu) \leq b_i(s,y, \mu)$ holds for $x\preceq y $ with $ x_i=y_i$ and $ \nu\preceq \mu;$
\item[$(2)$] $a=\bar a$, and for any $1\le i,j\le d, s\ge 0$ and $\mu\in\scr P_2$, $ a_{ij}(s,x,\mu)$ depends only on $x_i$ and $x_j.$
\end{enumerate}
Then $\bar \Lambda_s \preceq \Lambda_s$ for all $s\ge 0$. Consequently, $\bar P^*_{s,t}\preceq P^*_{s,t}$ for $t\geq s.$
\end{thm}
The next two results include necessary conditions for the order preservation, which are weaker than the sufficient ones given in Theorem \ref{T1}.
However, they coincide with the sufficient conditions and hence become sufficient and necessary conditions when $b(t,x,\mu)$ and $a(t,x,\mu)$ do not depend on $\mu$, and hence our first three results recover Theorems \ref{CW1} and \ref{CW2}.
For any $\mu\in \scr P$ and $I\subset \{1,\cdots,d\},$ let
$$\mu_I(A):= \mu(\{x\in \R^d: x_I\in A\}),\ \ A\in \B(\R^{\# I})$$ be the marginal distribution of $\mu$ with respect to components indexed by $I$, where $\# I$ denotes the number of elements in $I$. In particular, we simply denote $\mu_i=\mu_{\{i\}}$
\begin} \def\beq{\begin{equation}} \def\F{\scr F{thm}\label{T2}
If $\bar \Lambda_s\preceq \Lambda_s$ for all $s\geq 0$, then the following conditions hold:
\begin} \def\beq{\begin{equation}} \def\F{\scr F{enumerate}
\item[$(i)$] for any $\nu \preceq \mu$ with $\nu_i=\mu_i,$ $1\leq i \leq d$, there exists a coupling $\pi \in \scr C} \def\aaa{\mathbf{r}} \def\r{r(\nu,\mu)$ with $\pi \big(\{x\leq y\}\big)=1$ such that $$\bar b_i(s, x, \nu) \leq b_i(s,y, \mu),~~~s\ge 0,~(x,y)\in \supp\pi.$$
Consequently, $\bar b_i(s,x,\mu) \leq b_i(s,x,\mu) $ for $s\ge0,x\in\R^d,\mu\in\scr P_2.$
\item[$(ii)$] for any $ \nu\preceq \mu$ with $\nu_{ij}=\mu_{ij},1\leq i,j \leq d,$ there exists a coupling $\pi \in \scr C} \def\aaa{\mathbf{r}} \def\r{r(\nu,\mu)$ with $\pi \big(\{x\leq y\}\big)=1$ such that $$\bar a_{ij}(s, x, \nu)=a_{ij}(s,y,\mu),~s\geq0, ~(x,y)\in \supp\pi.$$
Consequently, $a(s,x,\mu)=\bar a(s,x,\mu)$ for any $s\ge0,x\in\R^d,\mu\in\scr P_2.$
\end{enumerate}
\end{thm}
Since $\bar \Lambda_s\preceq \Lambda_s$ implies $\bar P_{s,t}^* \preceq P_{s,t}^*$ for $t\ge s$, conditions in the following result are also necessary for $\bar \Lambda_s\preceq \Lambda_s$.
\begin{thm}\label{T3}
If $\bar P_{s,t}^* \preceq P_{s,t}^*~\mbox{for}~t\geq s\ge 0,$ then the following conditions hold:
\begin} \def\beq{\begin{equation}} \def\F{\scr F{enumerate}
\item[$(i)$] For any $s\ge 0$ and $1\le i\le d$, $\nu(\bar b_i(s,\cdot, \nu)) \leq \mu(b_i(s,\cdot, \mu))$ holds for $\nu\preceq \mu$ with $\nu_i=\mu_i.$
\item[$(ii)$] For any $s\ge 0$ and $1\le i,j\le d$, $\bar a_{ij}(s, x, \delta_x)=a_{ij}(s, x, \delta_x)$ holds and $a_{ij}(x, \delta_x)$ depends only on $x_i$ and $x_j.$
\end{enumerate}
\end{thm}
\subsection{Positive correlations}
We first present sufficient conditions for the positive correlations.
\begin{thm}\label{T4}
Assume {\bf (A)} and suppose further
\begin} \def\beq{\begin{equation}} \def\F{\scr F{enumerate}
\item[$(1)$] For any $s\ge 0$ and $1\le i\le d$, $b_i(s,x, \nu)\leq b_i(s,y, \mu),~ \nu\preceq \mu, x\leq y~\mbox{with}~x_i=y_i.$
\item[$(2)$] For any $1\le i,j\le d$, $a_{ij}\geq 0$, and for any $\mu \in \scr P_+,$ $a_{ij}(s,x,\mu)$ depends only on $x_i$ and $x_j.$
\end{enumerate}
Then $\Lambda_s \in \scr P_+^s.$ Consequently, $P^*_{s,t}\in \scr P_+$ for any $t\geq s\geq 0. $
\end{thm}
Similarly to the above results on the order preservation, necessary conditions for positive correlations presented in the next result are weaker than the above sufficient ones, but they coincide
and hence become necessary and sufficient conditions when $b$ and $a$ do not depend on the distribution.
\begin} \def\beq{\begin{equation}} \def\F{\scr F{thm}\label{T5} If $ \Lambda_s^{\mu} \in \scr P_+^s$ for $s\ge0$ and $\mu=\mu_{\{ij\}}\times\mu_{\{ij\}^c}\in\scr P_+$, then the following assertions hold:
\begin} \def\beq{\begin{equation}} \def\F{\scr F{enumerate}
\item[$(1)$] For any $s\ge 0$, $1\le i,j\le d$, $a_{ij}(s,x,\mu)$ depends only on $x_i$ and $x_j.$
\item[$(2)$] For any $s\ge 0$, $1\le i\le d$ and $f \in \scr U_b$ independent on $x_i,$
$$\mu(b_i(s,\cdot, \mu) f) \geq \mu(f)\mu(b_i(s,\cdot,\mu)).$$
\end{enumerate}
\end{thm}
\section{Time-inhomogeneous diffusion processes}
Consider the time-dependent second order diffusion operators: for $t\geq 0$ and $ x\in\R^d$,
\begin{equation}\label{OP}
\begin{split}
L_t:&=\frac{1}{2}\sum_{i,j=1}^d a_{ij}(t,x)\pp_i\pp_j+\sum_{i=1}^d b_i(t,x)\pp_i, \\
\bar L_t:&=\frac{1}{2}\sum_{i,j=1}^d \bar a_{ij}(t,x)\pp_i\pp_j+\sum_{i=1}^d\bar b_i(t,x)\pp_i,
\end{split}
\end{equation}
where $a_{ij},\bar a_{ij},b_i,\bar b_i\in C([0,\infty)\times\R^d)$. Assume that the martingale problems associated with $(L_t)_{t\ge0}$ and $(\bar L_t)_{t\ge0}$ are well-posed
so that there exist unique time-inhomogeneous diffusion processes $(X_{s,t})_{t\ge s\ge0}$ and $(\bar X_{s,t})_{t\ge s\ge0}$ corresponding to $(L_t)_{t\ge0}$ and $(\bar L_t)_{t\ge0}$, respectively.
Let $(P_{s,t})_{t\geq s\ge0}$ and $(\bar P_{s,t})_{t\geq s\ge0}$ be the Markov semigroups generated by $(X_{s,t}^x)_{\{ij\}}\times\mu_{\{ij\}^c}$ and $(\bar X_{s,t}^x)_{\{ij\}}\times\mu_{\{ij\}^c}$ with the initial value $X_{s,s}=\bar X_{s,s}=x$, respectively, i.e.,
\begin{equation}\label{semi}
P_{s,t}f(x)=\E f(X_{s,t}^x),\, \bar P_{s,t}f(x)=\E f(\bar X_{s,t}^x), \, \quad f\in \scr B_b(\R^d).
\end{equation}
It is well known that for any $f \in C_0^\infty(\R^d)$
\begin{equation}\label{PP}
\frac{\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D}{\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D s} P_{s,t}f(x)= -P_{s,t} L_sf,~~~~~\frac{\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D}{\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t} P_{s,t}f(x)= L_tP_{s,t}f,\quad t\ge s\ge0.
\end{equation}
For any $x,y\in\R^d$ with $x\leq y $, $f \in \scr U_b$ and $t\ge s\ge0$, if $\bar P_{s,t}f(x)\leq P_{s,t}f(y),$ we call $P_{s,t}$ preserving order, written as $\bar P_{s,t}^*\preceq P_{s,t}^*,$
where for any $\mu\in \scr P$, $P_{s,t}^*\mu, \bar P_{s,t}^*\mu\in \scr P$ is given by
$$(P_{s,t}^*\mu)(f):= \mu(P_{s,t}f),\ \ (\bar P_{s,t}^*\mu)(f):= \mu(\bar P_{s,t}f),\ \ f\in \B_b(\R^d).$$
Moreover, we denote $P_{s,t}^*\in\scr P_+$ if $P_{s,t}^*\scr P_+\subset \scr P_+$.
For any $\mu\in \scr P$, let $\LL_s\mu$ and $\bar \LL_s\mu$ be the distributions of the processes starting at $\mu$ from time $s$ generated by $L$ and $\bar L$ respectively.
By the standard Markov property we see that $\bar P_{s,t}^*\preceq P_{s,t}^*$ for $t\ge s\ge 0$ if and only if $\bar\LL_s\preceq \LL_s$ for $s\ge 0$, while $P_{s,t}^*\in \scr P_+$ for $t\ge s$ is equivalent to $\LL_s\in \scr P_+^s.$
\subsection{Main results}
The following two results extend Theorems \ref{CW1} and \ref{CW2} to the present time-inhomogeneous setting.
\begin} \def\beq{\begin{equation}} \def\F{\scr F{thm}\label{T1.3} If $\bar P_{s,t}^* \preceq P_{s,t}^* $ for $ t\geq s\ge0$, equivalently $\bar \LL_s\preceq \LL_s$ for $s\ge 0$, if and only if the following conditions hold:
\begin} \def\beq{\begin{equation}} \def\F{\scr F{enumerate}
\item[$({\bf 1})$] For any $s\ge 0$ and $1\le i\le d$,
$\bar b_i(s,x)\leq b_i(s,y) $ with $x\leq y$ and $ x_i=y_i.$
\item[$({\bf 2}$)] For any $s\ge 0$ and $1\le i,j\le d$, $ \bar a_{ij}=a_{ij}$ and $a_{ij}(s,x)$ only depends on $x_i$ and $x_j$.
\end{enumerate}
\end{thm}
\begin} \def\beq{\begin{equation}} \def\F{\scr F{thm}\label{T1.6} $P_{s,t}^*\in \scr P_+$ for $t\ge s$, equivalently $\LL_s\in \scr P_+^s$ for $s\ge 0$, if and only if the following conditions hold:
\begin} \def\beq{\begin{equation}} \def\F{\scr F{enumerate}
\item[$(1)$] For any $s\ge 0$ and $1\le i\le d$, $b_i(s, x)\leq b_i(s, y) $ with $x\leq y$ and $x_i=y_i; $
\item[$(2)$] For any $s\ge 0$ and $1\le i\le d$, $a_{ij}\ge 0$ and $ a_{ij}(s,x)\ge 0$ depends only on $x_i$ and $x_j.$
\end{enumerate}
\end{thm}
\subsection{Proofs}
\begin} \def\beq{\begin{equation}} \def\F{\scr F{proof}[Proof of Theorem \ref{T1.3}] (a) We first prove the necessity.
For any $t\ge s\ge0$ and $x\in\R^d$, let $\Lambda _{s}^x$ (resp. $\bar \Lambda _{s}^x$) be the distribution of the $L_t$-diffusion (resp. $\bar L_t$-diffusion) process on the path space $C_s:=C([s,\infty);\R^d)$ starting from $x$ at time $s.$
For $x\in\R^d$ and $0\le s_0\leq s_1< s_2<\cdots <s_n$, let $\Lambda^x_{s_0, s_1,\cdots, s_n}$ be the marginal distribution of $\Lambda_{s_0}^x$ at the time sequence $(s_1,\cdots, s_n)$, which can be expressed via the Markov property as below
$$\Lambda^x_{s_0, s_1,\cdots, s_n}(\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D y_1,\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D y_2,\cdots,\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D y_n)=P_{s_0,s_1}(x,\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D y_1)P_{s_1,s_2}(y_1,\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D y_2)\cdots P_{s_{n-1},s_n}(y_{n-1},\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D y_n).$$
Then, by an inductive argument, together with the Markov property of the associated Markov process, $\bar P_{s,t}^*\preceq P_{s,t}^*$ implies $\bar \Lambda _{s}^x\leq \Lambda _{s}^y$ (i.e., $\bar \Lambda_{s }^x(f)\leq \Lambda_{s }^y(f)$ for any $f \in \scr U_b\cap C_s$. Therefore,
there exists a coupling $\P_{s }^{x,y} \in \scr C} \def\aaa{\mathbf{r}} \def\r{r (\bar \Lambda^x_{s }, \Lambda^y_{s })$ such that
\begin{equation}\label{D}
\P_{s }^{x,y} \big( (\xi,\eta)\in C_{s }\times C_{s }: \eta \preceq \xi \big)=1.
\end{equation}
Let $(\Omega,\F,\P)=(C_{s }\times C_{s },\B(C_{s }\times C_{s} ),\P_{s }^{x,y})$ with the natural filtration $(\F_t)_{t\geq s }$ induced by the coordinate process $(\xi_t,\eta_t)_{t\geq s } $ solving
\begin{equation}\label{coor}
\begin{cases}
\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \xi_t=b(t,\xi_t)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}(t, \xi_t) \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_t^1,\quad & \xi_{s}=y\\
d \eta_t=\bar b(t,\eta_t)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+\bar \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}(t, \eta_t) \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_t^2,\quad & \eta_{s}=x
\end{cases}
\end{equation}
for some $d$-dimensional Brownian motions $(B_t^1)_{t\geq s}$ and $(B_t^2)_{t\geq s},$ and some measurable mappings $\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}, \bar \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}: [s,\infty)\times \R^d \rightarrow \R^d \otimes \R^d$ with
$a=\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s} \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}^*,~\bar a=\bar \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}\, \bar \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}^*$. Then, from \eqref{D}, we have
$\xi_t\geq \eta_t$, $\P_{s }^{x,y}$-a.s., for all $t\ge s$.
Let $x\leq y$ with $x_i=y_i$. Since $\xi_t\geq \eta_t$, $\P_{s}^{x,y}$-a.s., and $(\xi_s)_i=(\eta_s)_i$ due to $x_i=y_i$, we derive from \eqref{coor} that
$$\int^t_{s} \big(b_i(r,\xi_r)-\bar b_i(r,\eta_r)\big)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r\geq \int^t_{s} \big\<\bar \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{i\cdot}(r,\xi_r), \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_r^2\big\>-\int^t_{s} \big\<\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{i\cdot}(r,\xi_r), \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_r^1\big\>,$$
where $\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{i\cdot}$ means the $i$-th row of $\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}.$
Taking conditional expectation $\P_{s_0}^{x,y}(\cdot| \F_{s_0})$ on both sides yields
$$\int^t_{s } \E \big(\big(b_i(r,\xi_r)-\bar b_i(r,\eta_r)\big)|\F_{s }\big)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r\geq 0,~~~t\geq s.$$
This implies the assertion ({\bf 1}) by taking the continuity of
$b_i,~\bar b_i$ and $(\xi_\cdot,~\eta_\cdot)$ into account.
Let $x\leq y$ with $(x_i,x_j)=(y_i, y_j)$. Then, by using $\xi_t\geq \eta_t, \P_{s_0}^{x,y}$-a.s., again, we have
\begin{equation}\label{coor1}
\int_{s_0}^t b_k(s,\xi_s)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D s +\int_{s_0}^t \big\<\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{k\cdot}(s,\xi_s), \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_s^1\big\>\geq \int_{s_0}^t \bar b_k(s,\eta_s)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D s +\int_{s_0}^t \big\<\bar \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{k\cdot}(s,\eta_s), \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_s^2\big\>,\, k=i,j.
\end{equation}
Note that as $t\downarrow s_0,$
$$\frac{1}{\sqrt{t-s_0}}\Big( \int_{s_0}^t\big\<\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{i\cdot}(s,\xi_s), \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_s^1\big\>, \int_{s_0}^t\big\<\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{j\cdot}(s,\xi_s), \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_s^1\big\> \Big)\overrightarrow{weakly} ~~N\left(
0,
\left(\begin{array}{c}
a_{ii}(s, y)~~ a_{ij}(s,y) \\
a_{ji}(s, y)~~ a_{jj}(s,y)
\end{array}\right)
\right) =:\mu,$$
and that
$$\frac{1}{\sqrt{t-s_0}}\Big( \int_{s_0}^t\big\<\bar \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{i\cdot}(s,\xi_s), \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_s^2\big\>, \int_{s_0}^t\big\<\bar \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{j\cdot}(s,\xi_s), \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_s^1\big\> \Big)\overrightarrow{weakly} ~~N\left(
0,
\left( \begin{array}{c}
\bar a_{ii}(s, y)~~ \bar a_{ij}(s,y) \\
\bar a_{ji}(s, y)~~ \bar a_{jj}(s,y)
\end{array}\right)
\right) =:\bar \mu.$$
Then \eqref{coor1} implies $\bar \mu\preceq \mu$. On the other hand, by the symmetry of $\mu$ and $\bar \mu$ due to the symmetry of $a$ and $\bar a$, then $\bar \mu\preceq \mu $ implies $\bar \mu \preceq \mu.$ Therefore, we have $\mu=\bar \mu $ so that $a=\bar a$. For the assertion that $a_{ij}$ depends only on $x_i$ and $x_j$ of
({\bf 2}),
it can be available by following exactly the arguments of \cite[Lemmas 2.1 \& Lemma 2.3]{Wang}.
(b) Following exactly the arguments of \cite[Lemmas 2.4, 2.5 \& Theorem 1.3]{Wang} by replacing time homogeneous semi-group $P_t$ by time inhomogeneous semi-group $P_{s,t}$, we prove the sufficiency by the following Theorem \ref{T1.4} on the monotonicity. \end{proof}
\begin} \def\beq{\begin{equation}} \def\F{\scr F{thm}\label{T1.4} $ P_{s,t}^*$ is monotone, i.e., $P_{s,t}^*\preceq P_{s,t}^*$ for $t\ge s\ge 0$, provided the following two conditions hold:
\begin} \def\beq{\begin{equation}} \def\F{\scr F{enumerate}
\item[$({\bf1'})$] $b_i(s, x)\leq b_i(s, y) $ with $x\preceq y$ and $x_i=y_i; $
\item[$({\bf2'})$] $ a_{ij}(s,x)$ depends only on $x_i$ and $x_j.$
\end{enumerate}
\end{thm}
\begin{proof}
To get $P_{s,t}f \in \scr U_b$ for $t\geq s$ and $f\in \scr U_b$,
it suffices to show $$\nabla P_{s,t} f(x)\geq 0,\quad t\geq s, \, f\in \scr U_b \cap C_b^{\infty}(\R^d)$$
since $\scr U_b \cap C_b^{\infty}(\R^d)$ is dense in $\scr U_b$. Below, we assume $f\in \scr U_b \cap C_b^{\infty}(\R^d)$. Let $u_{s,t}=P_{s,t} f,~t\geq s.$ Then by \eqref{PP}, we have
$$\pp_t u_{s,t}=L_t u_{s,t},~~~t\geq s,~~~~~u_{s,s}=f.$$
Taking the partial derivative w.r.t. the $k$-th component (i.e., $\pp_k$) on both sides yields
\begin} \def\beq{\begin{equation}} \def\F{\scr F{equation}\label{Ust}
\pp_t(\pp_k u_{s,t})=\pp_k\pp_t u_{s,t}=L_t^k(\pp u_{s,t})+\sum_{j=1}^d \alpha_{kj}(t,\cdot)\pp_j u_{s,t},
\end{equation}
where $$L_t^k:= A_t+\sum_{j=1}^d\big[(1-\frac{1}{2})\pp_k a_{jk}(t,\cdot)\big]\pp_j+\pp_k b_k(t,\cdot), \quad \alpha_{kj}(t,\cdot):=(\pp_kb_j(t,\cdot))I_{\{k\neq j\}}.$$
Since $L_t^k$ is a time-inhomogeneous Schr\"odinger operator, it generates a positivity-preserving semigroup $\big(T_{s,t}^k\big)_{t\geq s}.$ So, the operator $L_t:= \big(L_t^k\big)_{1\leq k \leq d}$ defined on $C^2\big(\R^d;\R^d\big)$ by $L_tV:=\big(L_t^k V_k\big)_{1\leq k \leq d}$ generates a positivity preserving semigroup$$T_{s,t}:=\big(T_{s,t}^k\big)_{1\leq k \leq d}~,~~~~~t\geq s. $$
Let $D_r=\big(\alpha_{kj}(r,\cdot)_{1\leq k,j\leq d}\big)$ and $V_{s,t}=\nabla P_{s,t}f=\nabla u_{s,t}$. Then $\eqref{Ust}$ implies
\begin{equation*}\label{V}
\pp_t V_{s,t}=L_t V_{s,t}+D_t V_{s,t},~~~ t\geq s, ~~~V_{s,s}=\nabla f.
\end{equation*}
This, together with Duhamel's formula, gives
$$V_{s,t}=T_{s,t}V_{s,s}+\int_s^t T_{r,t} D_r V_{s,r} \mbox{d}r,~~~t\geq s.$$
Thus, we conclude that $V_{s,t}=\nabla P_{s,t}f \geq 0$ since $V_{s,s}=\nabla f\geq 0$ due to $f\in \scr U_b$ and $T_{s,t},~D_r$ are positivity preserving.
\end{proof}
\begin} \def\beq{\begin{equation}} \def\F{\scr F{rem}
Different from the proof in \cite{Pitt} for the time-homogeneous setting, we adopt the Duhamel's formula instead of the Trotter product formula which is less explicit in the present setting.
\end{rem}
\begin} \def\beq{\begin{equation}} \def\F{\scr F{proof}[Proof of Theorem \ref{T1.6}]
Theorem \ref{T1.6} can be proved using the same arguments in \cite[Proof of Proposition 4.1]{Pitt} by combining Theorem \ref{T1.3} and \ref{T1.4}. So, we omit the details to save space. \end{proof}
\section{Proofs of Theorems \ref{T1}, \ref{T2} and \ref{T3}}
\begin{proof}[Proof of Theorem \ref{T1}]
Note that $\bar P_{s,t}^* \nu$ and $ P_{s,t}^* \mu$ are marginal distribution at time $t$ of $\bar \Lambda_s\nu$ and $\Lambda_s\mu$, respectively, we infer that $\bar P^*_{s,t}\preceq P^*_{s,t}$ for $t\geq s $ once $\bar \Lambda_s \preceq \Lambda_s $ is available. Therefore,
to obtain the desired assertion, it is sufficient to show $\bar \Lambda_s \nu \preceq \Lambda_s \mu$. Below, we set $\mu, \nu \in \scr P_2$ with $\nu \preceq \mu$. For any $T>s$,
set
$$\scr P^{\nu, \mu}_{s,T}:=\big\{ (\mu^{(1)},\mu^{(2)}) \in C([s,T]; \scr P_2\times \scr P_2):\mu_t^{(1)}\preceq \mu_t^{(2)},~t\in [s,T],~\mu_s^{(1)}=\nu, \mu_s^{(2)}=\mu\big\},$$
which is a complete metric space under the metric for $\lambda>0$,
$$\rr_{\lambda}\big( (\mu^{(1)},\mu^{(2)}), (\tilde \mu^{(1)}, \tilde \mu^{(2)}) \big):=
\sup_{t\in [s,T]} \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-\lambda t} \big\{\W_2(\mu^{(1)}_t, \tilde \mu^{(1)}_t)+\W_2(\mu^{(2)}_t, \tilde \mu^{(2)}_t)\big\}.$$
For any $(\mu^{(1)},\mu^{(2)})\in \scr P_{s,T}^{\nu, \mu},$ consider the following time-dependent SDEs:
\begin{equation}\label{ddsde1}
\begin{cases}
\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D X_t^{(1),\mu^{(1)}}=\bar b(t, X_t^{(1),\mu^{(1)}}, \mu^{(1)}_t)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+ \bar \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}(t, X_t^{(1),\mu^{(1)}}, \mu^{(1)}_t) \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_t & t\geq s,\quad X_s^{(1),\mu^{(1)}}=\xi\sim\nu,\\
\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D X_t^{(2),\mu^{(2)}}=b(t, X_t^{(2),\mu^{(2)}}, \mu^{(2)}_t)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+ \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}(t, X_t^{(2),\mu^{(2)}}, \mu^{(2)}_t) \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_t & t\geq s,\quad X_s^{(2),\mu^{(2)}}=\eta\sim\mu,
\end{cases}
\end{equation} where $\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}=\ss{2a}$ and $\bar\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}=\ss{2\bar a}$, and $\xi \sim\nu$ means $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\xi}=\nu$.
Define the mapping on $\scr P^{\nu, \mu}_{s,T}$ by
\begin{equation}\label{N1}
H\big( (\mu^{(1)},\mu^{(2)})\big)(t)=\Big(\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t^{(1),\mu^{(1)}}}, \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t^{(2),\mu^{(2)}}} \Big),\quad t\ge s.
\end{equation}
Since $\mu^{(1)}_t\preceq \mu^{(2)}_t$, the standard Banach fixed point theorem yields
\begin{equation}\label{N3}
\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_{[s,T]}^{(1),\mu^{(1)}}} \preceq \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_{[s,T]}^{(2),\mu^{(2)}}},
\end{equation}
In the sequel, we aim to prove that $H$ is contractive under the metric $\rr_{\lambda}$ for large enough $\lambda>0.$
Let $(\mu^{(1)},\mu^{(2)}),(\tilde \mu^{(1)},\tilde \mu^{(2)}) \in \scr P^{\nu, \mu}_{s,T}.$
By It\"o's formula and the assumption $({\bf A})$ , we get
\begin{equation}
\begin{split}
&\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \big( \big|X_t^{(1),\mu^{(1)}}-X_t^{(1),\tilde \mu^{(1)}}\big|^2+\big|X_t^{(2),\mu^{(2)}}-X_t^{(2),\tilde \mu^{(2)}}\big|^2\big)\\
&\quad\leq K(t)\Big(|X_t^{(1),\mu^{(1)}}-X_t^{(1),\tilde \mu^{(1)}}\big|^2\\
&\quad+|X_t^{(2),\mu^{(2)}}-X_t^{(2),\tilde \mu^{(2)}}\big|^2+\W_2(\mu^{(1)}_t, \tilde \mu^{(1)}_t)^2+ \W_2(\mu^{(2)}_t, \tilde \mu^{(2)}_t)^2 \Big)+\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D M_t
\end{split}
\end{equation}
for some martingale $M_t.$
Then, taking expectation on both sides, using Grownwall's inequality and taking $X_s^{(1),\mu^{(1)}}=X_s^{(1),\tilde \mu^{(1)}}=\xi$ and $X_s^{(2),\mu^{(2)}}=X_s^{(2),\tilde \mu^{(2)}}=\eta$ into consideration, we have
\begin{equation*}\label{esti}
\begin{split}
&\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-2\lambda t}\big( \E |X_t^{(1),\mu^{(1)}}-X_t^{(1),\tilde \mu^{(1)}}\big|^2+\E|X_t^{(2),\mu^{(2)}}-X_t^{(2),\tilde \mu^{(2)}}\big|^2\big)\\
&\quad\leq \frac{1}{2\lambda} K(T)\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{K(T)(T-s)} \sup_{t\in[s,T]} \big[ \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-2\lambda t}\big( \W_2(\mu^{(1)}_t, \tilde \mu^{(1)}_t)^2+ \W_2(\mu^{(2)}_t, \tilde \mu^{(2)}_t)^2\big) \big].
\end{split}
\end{equation*}
This yields
\begin{equation*}
\rr_{\lambda}\big(H(\mu^{(1)},~\mu^{(2)}), H(\tilde \mu^{(1)},~\tilde\mu^{(2)})\big)\leq \frac{1}{2\lambda} K(T)\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{K(T)(T-s)} \rr_{\lambda} \big((\mu^{(1)},~\mu^{(2)}),~(\tilde \mu^{(1)},~\tilde\mu^{(2)}) \big).
\end{equation*}
Hence, for $\lambda>0$ large enough, $H$ is contractive under the metric $\rr_{\lambda}$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{T2}]
Let $s\geq 0$ and $\nu\preceq \mu$ with $\nu_i=\mu_i$. By $ \bar \Lambda_s \preceq \Lambda_s,$
we have $\bar \Lambda_s \nu \preceq \Lambda_s \mu.$ According to \cite[Theorem 5]{ KKO}, there exists $\P_s \in \scr C} \def\aaa{\mathbf{r}} \def\r{r(\bar \Lambda_s\nu, \Lambda_s\mu)$ such that
\begin{equation}\label{coup}
\P_s\big(\{(\xi,\eta)\in C_s \times C_s: \xi_t\geq \eta_t, t\geq s\}\big)=1.
\end{equation}
Since $\bar \Lambda_s\nu$ and $ \Lambda_s\mu$ are solutions to the martingale problems associated with the operators $\bar L$ and $L$ in \eqref{N2}, respectively, according to the superposition principle (see \cite{SV}), we have $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{(\xi, \eta)}=\P_s$, where $(\xi_t,\eta_t)$ solves
\begin{equation}\label{PSDE1}
\begin{cases}
&\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \xi_t=\bar b(t, \eta_t, \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\eta_t}) \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+\bar \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}(t, \eta_t, \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\eta_t}) \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_t^1,~~t\geq s\\
& \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \eta_{t}=b(t,\xi_{t}, \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\xi_{t}}) \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}(t,\xi_{t}, \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\xi_{t}}) \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_t^2,~~ t\geq s
\end{cases}
\end{equation}
for
some $2d$-dimensional Brownian motions $(B_t^1, B_t^2)_{t\geq s}$ on the probability space $(C_s\times C_s, \B(C_s\times C_s),\{\F_t\}_{t\geq s}, \P_s),$
where $ \{\F_t\}_{t\geq s}$ is the natural filtration induced by the coordinate processes $(\xi_t, \eta_t)_{t\geq s}$.
Since $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{(\xi, \eta)}=\P_s$ satisfying \eqref{coup}, then we have $\xi_t\geq \eta_t$ for all $t\geq s.$ Moreover, note that $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{(\xi_s, \eta_s)}\in \scr C} \def\aaa{\mathbf{r}} \def\r{r(\nu,\mu)$ and $\nu_i=\mu_i$ imply $\xi_s^i=\eta_s^i.$ Thus, we find $\P_s$-a.s.
\begin{equation}\label{NN3}
\int_s^t \bar b_i(r, \xi_r, \mu^{(1)}_{r})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r+\int_s^t \bar \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{i\cdot}(r, \xi_r, \mu^{(1)}_{r})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_t^1 \leq \int_s^t b_i(r, \eta_r, \mu^{(2)}_{r})+ \int_s^t \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{i\cdot}(r, \eta_r, \mu^{(2)}_{r})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_t^1,\, t\ge s.
\end{equation}
Taking conditional expectation with respect to $\F_s$, we drive
\begin{equation*}
\int_s^t \E\big(\bar b_i(r, \xi_r, \mu_r^{(1)})| \F_s\big)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r \leq \int_s^t \E\big(b_i(r, \eta_r, \mu_r^{(2)})| \F_s\big)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r,~~~~t\geq s.
\end{equation*}
By the continuity of $\bar b$ and $b$ and $\mu_r^{(1)}\rightarrow}\def\l{\ell}\def\iint{\int}\def\gg{\gamma \nu, \mu_r^{(2)}\rightarrow}\def\l{\ell}\def\iint{\int}\def\gg{\gamma \mu$ weakly as $r\downarrow s$, we obtain
\begin{equation*}
\bar b_i(s, \xi_s, \nu)\le \bar b_i(s, \eta_s, \nu),~~~t\geq s,~\P_s-a.s.
\end{equation*}
Consequently, for $\pi:=\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{(\xi_s, \eta_s)}\in \scr C} \def\aaa{\mathbf{r}} \def\r{r(\nu, \mu)$ with $\pi \big(\{x\leq y\}\big)=1$,
$$\bar b_i(s, x, \nu) \leq b_i(s,y, \mu), ~(x,y)\in \supp \pi.$$
Thus, the first assertion of (i) holds true. Hence, for $\nu=\mu,$ $\pi (\{x\leq y\})=1$ implies $x=y, \pi-a.s.$ Whence, we have
\begin{equation*}\label{NN4}
\bar b_i(s, x, \mu) \leq b_i(s,x, \mu), ~x\in \supp \mu.
\end{equation*}
In general, for any $x\in \R^d,$ let $\mu_{\varepsilon}\def\ddd{D^*}=(1-\varepsilon}\def\ddd{D^*)\mu+\varepsilon}\def\ddd{D^*\delta_x.$ It is easy to see that $x\in \supp\mu_{\varepsilon}\def\ddd{D^*}.$ Thus applying \eqref{NN4} with $\mu_{\varepsilon}\def\ddd{D^*}$ replaced by $\mu$ yields
$$\bar b_i(s, x, \mu_\vv) \leq b_i(s, x, \mu_\vv), ~s\geq 0,~\varepsilon}\def\ddd{D^*>0.$$
Consequently, the second assertion in $(i)$ follows by taking $\varepsilon}\def\ddd{D^* \downarrow 0$.
Below we assume $\nu\preceq \mu$ with $\nu_{ij}=\mu_{ij}$ so that
$\nu_i=\mu_i,\nu_j=\mu_j.$ Thus, we deduce from \eqref{NN3} that for any $\varepsilon}\def\ddd{D^* \in [0,1],$
\begin{equation*}
\begin{split}
&\int_s^t \big[\varepsilon}\def\ddd{D^* \bar b_i(r, \xi_r, \mu_r^{(1)})+(1-\varepsilon}\def\ddd{D^*)\bar b_j(r, \xi_r, \mu_r^{(1)})\big]\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r+\int_s^t \big[\varepsilon}\def\ddd{D^* \bar \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{i\cdot}(r, \xi_r, \mu_r^{(1)})+(1-\varepsilon}\def\ddd{D^*)\bar \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{j\cdot}(r, \xi_r, \mu_r^{(1)})\big]\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_r^1\\
&\leq \int_s^t \big[\varepsilon}\def\ddd{D^* \bar b_i(r, \eta_r, \mu_r^{(2)})+(1-\varepsilon}\def\ddd{D^*)\bar b_j(r, \eta_r, \mu_r^{(2)})\big]\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r+\int_s^t \big[\varepsilon}\def\ddd{D^* \bar \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{i\cdot}(r, \eta_r, \mu_r^{(2)})+(1-\varepsilon}\def\ddd{D^*)\bar \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{j\cdot}(r, \eta_r, \mu_r^{(2)})\big]\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_r^2
\end{split}
\end{equation*}
Dividing both side by $\frac{1}{\sqrt{t-s}}$ and letting $t\downarrow s,$ we find
\begin{equation*}
\begin{split}
&N\left(
0, \varepsilon}\def\ddd{D^*^2 \bar a_{ii}(s,\xi_s, \nu)+2\varepsilon}\def\ddd{D^*(1-\varepsilon}\def\ddd{D^*)\bar a_{ij}(s,\xi_s, \nu)+(1-\varepsilon}\def\ddd{D^*)^2 \bar a_{jj}(s,\xi_s, \nu)\right)\\
& \le N\left(
0, \varepsilon}\def\ddd{D^*^2 a_{ii}(s,\eta_s, \mu)+2\varepsilon}\def\ddd{D^*(1-\varepsilon}\def\ddd{D^*) a_{ij}(s,\eta_s, \mu)+(1-\varepsilon}\def\ddd{D^*)^2 a_{jj}(s,\eta_s, \mu)\right) .
\end{split}\end{equation*}
By the symmetry of centred normal distribution, this further implies
\begin{equation*}
\begin{split}&\varepsilon}\def\ddd{D^*^2 \bar a_{ii}(s,\xi_s, \nu)+2\varepsilon}\def\ddd{D^*(1-\varepsilon}\def\ddd{D^*)\bar a_{ij}(s,\xi_s, \nu)+(1-\varepsilon}\def\ddd{D^*)^2 \bar a_{jj}(s,\xi_s, \nu)\\&=\varepsilon}\def\ddd{D^*^2 a_{ii}(s,\eta_s, \mu)+2\varepsilon}\def\ddd{D^*(1-\varepsilon}\def\ddd{D^*) a_{ij}(s,\eta_s, \mu)+(1-\varepsilon}\def\ddd{D^*)^2 a_{jj}(s,\eta_s, \mu), \varepsilon}\def\ddd{D^* \in [0,1]. \end{split}
\end{equation*}
Consequently, dividing by $\varepsilon}\def\ddd{D^*^2$ on both sides yields
\begin{equation}\label{NN5}
\bar a_{ij}(s, \xi_s, \nu) = a_{ij}(s,\eta_s, \mu),~\P_s-a.s.,
\end{equation}
which gives for $\pi=\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{(\xi_s, \eta_s)} \in \scr C} \def\aaa{\mathbf{r}} \def\r{r(\nu, \mu),$
$$\bar a_{ij}(s, x, \nu)=a_{ij}(s,y,\mu),~~(x,y)\in \supp \pi, ~s\geq 0.$$
Thus, by the approximation trick above, we can obtain the second assertion in (ii).
\end{proof}
\begin{proof}[Proof of Theorem \ref{T3}]
Due to $\bar P_{s,t}^* \preceq P_{s,t}^*$, we have $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\bar X_{s,t}}=\bar P_{s,t}^*\nu \preceq P_{s,t}^*\mu=\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_{s,t}}$ for $\nu\preceq \mu.$ Therefore, in particular for $f(x)=x_i \in \scr U$, we obtain
$$\E(\bar X_{s,t})_i\leq \E(X_{s,t})_i.$$
Since $\nu_i(f)=\mu_i(f)$, we then deduce from \eqref{PSDE1} with $\xi_t$ and $\eta_t$ replaced by $X_{s,t}$ and $\bar X_{s,t}$ in \eqref{ddsde1} that
$$\int_s^t \E(\bar b_i(s,\bar X_{s,r}, \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\bar X_{s,r}} ))\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r\leq \int_s^t \E( b_i(s, X_{s,r}, \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_{s,r}}))\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r.$$
Dividing by $t-s$ on both side followed by $t\downarrow s$, we get $(i).$
Since $\bar P_{s,t}^*\mu \leq P_{s,t}^*\mu,$ for
$f\in \scr U_b\cap C^{\infty}_b(\R^d)$, we have $\mu(\bar L_{s,\mu}f)\leq \mu( L_{s,\mu}f).$
In particular, taking $\mu=\delta} \def\DD{\Delta} \def\vv{\varepsilon} \def\rr{\rho_x$ yields $ \bar L_{s,\delta} \def\DD{\Delta} \def\vv{\varepsilon} \def\rr{\rho_x}f(x)\leq L_{s,\delta} \def\DD{\Delta} \def\vv{\varepsilon} \def\rr{\rho_x}f(x).$ With this at hand, we can get the assertion (ii) by following exactly the argument of \cite[Lemma 3.4]{Wang}.
\end{proof}
\subsection{Proofs of Theorems \ref{T4} and \ref{T5} }
We first present some lemmas.
\begin{lem}\label{Lem}
Let $\mu=\frac{1}{2} (\mu^{(1)}+\mu^{(2)})$, where $\mu^{(1)}, \mu^{(2)}\in \scr P_+$ such that $\mu^{(1)}\preceq \mu^{(2)}$. Then, $\mu\in\scr P_+.$
\end{lem}
\begin{proof}
Due to $\mu^{(1)}\preceq \mu^{(2)}$, we have for any $f,g\in\scr U_b$
$$\big(\mu^{(1)}(f)-\mu^{(2)}(f)\big)\big(\mu^{(1)}(g)-\mu^{(2)}(g)\big)\ge0.$$
The above inequality is equivalent to the following inequality
\begin{equation}\label{E1}
2\big(\mu^{(1)}(f)\mu^{(1)}(g)+\mu^{(2)}(f)\mu^{(2)}(g)\big)\ge \big(\mu^{(1)}(f)+\mu^{(2)}(f)\big) \big(\mu^{(1)}(g)+\mu^{(2)}(g)\big).
\end{equation}
Furthermore, in terms of $\mu^{(1)}, \mu^{(2)}\in \scr P_+$, we deduce for any $f,g\in\scr U_b$,
$$\mu^{(1)}(fg)\ge \mu^{(1)}(f)\mu^{(1)}(g),\quad\mu^{(2)}(fg)\ge \mu^{(2)}(f)\mu^{(2)}(g).$$ Substituting this \eqref{E1} yields
\begin{equation*}
\ff{1}{2}\big(\mu^{(1)}(fg)+\mu^{(2)}(fg) \big)\ge \ff{1}{4}\big(\mu^{(1)}(f)+\mu^{(2)}(f)\big) \big(\mu^{(1)}(g)+\mu^{(2)}(g)\big)
\end{equation*}
so that $\mu\in\scr P_+.$
\end{proof}
\begin{lem}\label{lem1}
Suppose that $(P_{s,t}^*)_{t\ge s}$ preserves positive correlations and let $\mu$ be the same as that in Lemma \ref{Lem}. Then, for
$f,g\in \scr U_b\cap C_b^\infty(\R^d)$ such that
\begin{equation}\label{EE}
\mu(fg)=\mu(f)\mu(g),
\end{equation}
we have
\begin{equation}\label{in1}
\begin{split}
2 \big(\mu^{(1)}(L_{s, \mu}(fg))+\mu^{(2)}(L_{s, \mu}(fg))\big) &\geq \big( \mu^{(1)}(L_{s, \mu}(f))+\mu^{(2)}(L_{s, \mu}(f))\big)\big(\mu^{(1)}(g)+\mu^{(2)}(g)\big)\\
&\quad+\big( \mu^{(1)}(L_{s, \mu}(g))+\mu^{(2)}(L_{s, \mu}(g))\big)\big(\mu^{(1)}(f)+\mu^{(2)}(f)\big).
\end{split}
\end{equation}
Equivalently,
\begin{equation}\label{in2}
\begin{split}
2 \big(\mu^{(1)}(\Gamma_1(f,g))+\mu^{(2)}(\Gamma_1(f,g))\big) &\geq \big( \mu^{(2)}(L_{s, \mu}(g))-\mu^{(1)}(L_{s, \mu}(g))\big)\big(\mu^{(1)}(f)-\mu^{(2)}(f)\big)\\
&\quad+\big(\mu^{(2)}(L_{s, \mu}(f))+\mu^{(1)}(L_{s, \mu}(f))\big)\big(\mu^{(1)}(g)-\mu^{(2)}(g)\big),
\end{split}
\end{equation}
where
$$\Gamma_1(f,g)(x):=L_{s, \mu}(fg)(x)-f(x)L_{s, \mu}g(x)-g(x)L_{s, \mu}f(x)=\sum_{i,j}a_{i,j}(t,x,\mu)\frac{\pp}{\pp x_i}f(x)\frac{\pp}{\pp x_j}g(x).$$
\end{lem}
\begin{proof}
By a direct calculation, we obtain that \eqref{in1} is equivalent to \eqref{in2}. Therefore, it is sufficient to show that \eqref{in1} holds true.
From Lemma \ref{Lem}, we have $\mu\in\scr P_+.$ Since $(P_{s,t}^*)_{t\ge s}$ preserves positive correlations, we deduce for any $f,g\in \scr U_b\cap C_b^\infty(\R^d)$,
$$(P_{s,t}^*\mu)(fg)\ge (P_{s,t}^*\mu)(f)(P_{s,t}^*\mu)(g),\quad t\ge s. $$
This, together with $\mu=\frac{1}{2} (\mu^{(1)}+\mu^{(2)})$, yields
\begin{equation*}
2\big((P_{s,t}^*\mu^{(1)})(fg)+(P_{s,t}^*\mu^{(2)})(fg)\big)\ge \big((P_{s,t}^*\mu^{(1)})(f)+(P_{s,t}^*\mu^{(2)})(f)\big)\big((P_{s,t}^*\mu^{(1)})(g)+(P_{s,t}^*\mu^{(2)})(g)\big).
\end{equation*}
Thus, by taking \eqref{EE} into consideration, we derive that
\begin{equation*}
\begin{split}
&\ff{2}{t-s}\big((P_{s,t}^*\mu^{(1)})(fg)+(P_{s,t}^*\mu^{(2)})(fg)-(\mu^{(1)})(fg)+(\mu^{(2)})(fg)\big)\\
&\ge \ff{1}{t-s}\big\{\big((P_{s,t}^*\mu^{(1)})(f)+(P_{s,t}^*\mu^{(2)})(f)\big)\big((P_{s,t}^*\mu^{(1)})(g)+(P_{s,t}^*\mu^{(2)})(g)\big)\\
&\quad-
\big( \mu^{(1)} (f)+ \mu^{(2)} (f)\big)\big( \mu^{(1)} (g)+ \mu^{(2)} (g)\big)\big\}.
\end{split}
\end{equation*}
Consequently, the assertion \eqref{in1} follows by taking $t\downarrow s.$
\end{proof}
\begin{lem}\label{lem4}
If $(P_{s,t}^*)_{t\ge s}$ preserves positive correlation and $\mu\in\scr P_+$ same as in Lemma \ref{Lem}, then $a_{ij}(s,x,\mu)\geq 0$ for any $s\ge0, x\in\R^d$.
\end{lem}
\begin{proof}
For $\vv>0$ and $x\in\R^d$, let $B_{\varepsilon}\def\ddd{D^*}(x)=\{y\in \R^d |\, |x-y|\leq \varepsilon}\def\ddd{D^*\}$, the closed ball centred at the point $x$ with the radius $\varepsilon}\def\ddd{D^*.$
For $f,g\in\scr U_b\cap C_b^\infty(\R^d)$, let $f_{\varepsilon}\def\ddd{D^*}, g_{\varepsilon}\def\ddd{D^*}\in \scr U_b\cap C_b^\infty(\R^d)$ with $f_{\varepsilon}\def\ddd{D^*} \in[f(x)-\vv,f(x)+\vv]$ and $g_{\varepsilon}\def\ddd{D^*} \in[g(x)-\vv,g(x)+\vv]$ on $B_{\varepsilon}\def\ddd{D^*}(x)^c$
be the point-wise approximation of $f,g\in\scr U_b\cap C_b^\infty(\R^d)$ (i.e., $f_\vv\rightarrow}\def\l{\ell}\def\iint{\int}\def\gg{\gamma f$ and $g_\vv\rightarrow}\def\l{\ell}\def\iint{\int}\def\gg{\gamma g$ as $\vv\to0$).
Below we assume that $\mu\in\scr P_+$ is given in Lemma \ref{Lem}.
By Lemma \ref{lem1}, we obtain for $f_{\varepsilon}\def\ddd{D^*}, g_{\varepsilon}\def\ddd{D^*}\in \scr U_b\cap C_b^\infty(\R^d)$,
$$\mu(L_{s,\mu}(f_{\varepsilon}\def\ddd{D^*}g_{\varepsilon}\def\ddd{D^*}))(x)\geq \mu(f_{\varepsilon}\def\ddd{D^*})\mu(L_{s,\mu}g_{\varepsilon}\def\ddd{D^*})(x)+\mu(g_{\varepsilon}\def\ddd{D^*})\mu(L_{s,\mu}f_{\varepsilon}\def\ddd{D^*})(x), ~~ s\geq 0.$$
Then, combining the fact that $f_{\varepsilon}\def\ddd{D^*}, g_{\varepsilon}\def\ddd{D^*}\in \scr U_b\cap C_b^\infty(\R^d)$ are constants on $B_{\varepsilon}\def\ddd{D^*}(x)^c$,
we have
\begin{equation}\label{EE2}
\begin{split}
\int_{\R^d} (L_{s,\mu}(f_{\varepsilon}\def\ddd{D^*}g_{\varepsilon}\def\ddd{D^*}))(y)\frac{I_{B_{\varepsilon}\def\ddd{D^*}(x)}(y)(dy)}{\mu(B_{\varepsilon}\def\ddd{D^*}(x))}&\geq
\mu(f_{\varepsilon}\def\ddd{D^*})\int_{\R^d}L_{s,\mu}g_{\varepsilon}\def\ddd{D^*}(y))\frac{I_{B_{\varepsilon}\def\ddd{D^*}(x)}(y)(dy)}{\mu(B_{\varepsilon}\def\ddd{D^*}(x))}\\
&\quad+
\mu(g_{\varepsilon}\def\ddd{D^*})\int_{\R^d}L_{s,\mu}f_{\varepsilon}\def\ddd{D^*}(y))\frac{I_{B_{\varepsilon}\def\ddd{D^*}(x)}(y)(dy)}{\mu(B_{\varepsilon}\def\ddd{D^*}(x))}.
\end{split}
\end{equation}
Observe that
\begin{equation*}
\mu(f_{\varepsilon}\def\ddd{D^*})=\int_{B_\vv(x)}(f_\vv(z)-f(x))\mu(\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D z)+\int_{B_\vv(x)^c}(f_\vv(z)-f(x))\mu(\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D z)+f(x)\rightarrow f(x)
\end{equation*}
as $\vv\to0$, where the first integral goes to zero since $f_\vv$ is uniformly continuous on $B_\vv(x)$, and the second integral tends to zero due to $f_\vv\in[f(x)-\vv,f(x)+\vv]$ on $B_\vv(x)^c. $ Similarly, we obtain $\mu(g_\vv)\rightarrow}\def\l{\ell}\def\iint{\int}\def\gg{\gamma g(x)$ as $\vv\rightarrow}\def\l{\ell}\def\iint{\int}\def\gg{\gamma 0.$ Furthermore,
note that $\frac{\mu(B_{\varepsilon}\def\ddd{D^*}(x) \cap \cdot)}{\mu(B_{\varepsilon}\def\ddd{D^*}(x))}$ converges weakly to $\delta_x$ as $\varepsilon}\def\ddd{D^* \downarrow0.$ Thus, taking $\vv\downarrow0$ on both sides of \eqref{EE2} yields
\begin{align*}
L_{s,\mu}(fg)(x)\geq f(x)L_{s,\mu}g(x)+g(x)L_{s,\mu}f(x).
\end{align*}
Thus, by choosing $f,g\in \scr U_b\cap C_b^\infty(\R^d)$ such that in a neighbourhood of $x$
$$f(z)=z_i,~~~g(z)=z_j,$$
we deduce that $$x_jb_i(s,x,\mu)+x_ib_j(s,x,\mu)+a_{ij}(s,x,\mu)\geq x_jb_i(s,x,\mu)+x_ib_j(s,x,\mu)$$
so that $a_{ij}(t,x,\mu)\geq 0.$
\end{proof}
\begin{lem}\label{lem2}
Assume $\mu^{(1)}:=\mu^{(1)}_i\times \mu^{(1)}_{\{i\}^c}$ and $\mu^{(2)}:=\mu^{(2)}_i\times \mu^{(2)}_{\{i\}^c}$ with $\mu^{(1)}_i= \mu^{(2)}_i$, where $\mu^{(1)}_i$, $\mu^{(2)}_i$, $\mu^{(1)}_{\{i\}^c}$, $ \mu^{(2)}_{\{i\}^c}\in \scr P_+, $ and suppose further that
$(P_{s,t}^*)_{t\ge s}$ preserves positive correlation. Then,
$$\mu^{(1)}( b_i(t, \cdot, \mu^{(1)}))\leq \mu^{(2)}( b_i(t, \cdot, \mu^{(2)})).$$
\end{lem}
\begin{proof}
Since $\mu^{(1)}_i$, $\mu^{(2)}_i$, $\mu^{(1)}_{\{i\}^c}$, $ \mu^{(2)}_{\{i\}^c}\in \scr P_+, $ we deduce $\mu^{(1)},\mu^{(2)}\in\scr P_+.$
For given $i$ and $k\neq i$, take $f,g \in \scr U_b\cap C_b^\infty(\R^d)$ such that in a neighbourhood of $x$,$$f(z)=z_i -\int_\R r\mu^{(1)}_i(\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r),~~~~g(z)=\frac{\frac{h(z_k)}{1+h(z_k)}-\mu^{(1)}(\frac{ h}{1+ h})}{\mu^{(2)}(\frac{ h}{1+ h})-\mu^{(1)}(\frac{ h}{1+ h})},$$
where $h\in C^{\infty}(\R;\R_+)$ is an increasing function. Note that
$$
\mu(fg)=\mu(f)=0,\quad \mu=(\mu^{(1)}+\mu^{(2)})/2.
$$
So applying Lemma \ref{lem1} gives
$$\int b_i(t, x, \mu^{(1)})\mu^{(1)}(\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D x)\leq \int b_i(t, x,\mu^{(2)})\mu^{(2)}(\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D x).$$
\end{proof}
\begin{proof}[Proof of Theorem \ref{T4}] Since $\bar P_{s,t}^*\mu$ is the marginal distributions of $\LL_s\mu$ at time $t$,
$ \LL_s\in \scr P_+$ implies $ P_{s,t}^*\in \scr P_+$ for $t\ge s.$ So, it suffices to prove $ \LL_s\in \scr P_+^s$.
To this end, we only need to prove that for any $\mu_0\in \scr P_+$ and $T>s\ge 0,$ the marginal distribution $\LL_{s,T}\mu_0$ of $\LL_s\mu_0$ on $C_{s,T}:= C([s,T];\R^d)$ satisfies
\beq\label{*G} (\LL_{s,T}\mu_0) (FG)\ge (\LL_{s,T}\mu_0) (F)\LL_{s,T}(G)\end{equation}
for any bounded increasing functions $F,G$ on $C_{s,T}.$
To achieve this, let
$$\scr D_+=\big\{\nu\in C([s,T];\scr P_2(\R^d)):\nu_s=\mu_0,~\nu_t\in \scr P_+, ~t\in [s,T]\big\},$$ which is a Polish space under the metric for $\lambda >0:$
\beq\label{C2}
\W_{2,\lambda}(\mu,\nu):=\sup_{t \in [s,T]} (\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-\lambda t} \W_2(\mu_t,\nu_t)).
\end{equation}
For $\nu \in C([s,T];\scr P_2(\R^d)), x\in\R^d$ and $t\in [s,T]$, let
$$b_t^{\nu}(x)=b_t(x, \nu_t),~~~\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_t^{\nu}(x)=\ss{2 a_t(x, \nu_t)}.$$
For $\nu\in \scr D_+ $, consider the following time-dependent SDE
\beq\label{EQ3}
\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D X_t^{\nu}=b_t^{\nu}(X_t^{\nu})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_t^{\nu}(X_t^{\nu})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_t,\quad t\in[s,T],~~X_s^{\nu}=X_s\sim \mu_0.
\end{equation}
For $\nu\in \scr D_+,$ we define the mapping $\nu\mapsto \Phi(\nu)$ as below, $$(\Phi(\nu))_t:=\scr L_{X_t^{\nu}},\quad t\in[s,T].$$
Under $(1)$ and $(2)$, by $\mu_0\in \scr P_+$ and Theorem \ref{T1.6}, we have $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t^{\mu}} \in \scr P_+ $ so that $\Phi: \D \rightarrow}\def\l{\ell}\def\iint{\int}\def\gg{\gamma \D. $
Below, we assume that $\nu^1, \nu^2\in \scr D.$ By It\"o's formula and \eqref{A1}, it follows that
\begin{equation*}
\begin{split}
\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D|X_t^{\nu^1}-X_t^{\nu^2}|^2&=\big\{2\<X_t^{\nu^1}-X_t^{\nu^2}, ~b_t^{\nu^1}(X_t^{\nu^1})-b_t^{\nu^2}(X_t^{\nu^2})\>+\|\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_t^{\nu^1}(X_t^{\nu^1})-\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_t^{\nu^2}(X_t^{\nu^2})\|_{HS}^2 \big\}\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t \\
&\quad+2\<X_t^{\nu^1}-X_t^{\nu^2}, (\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_t^{\nu^1}(X_t^{\nu^1})-\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_t^{\nu^2}(X_t^{\nu^2}))\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_t\>\\
&\leq K (|X_t^{\nu^1}-X_t^{\nu^2}|^2+\W_2(\nu_t^1, \nu_t^2)^2)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+2\<X_t^{\nu^1}-X_t^{\nu^2}, (\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_t^{\nu^1}(X_t^{\nu^1})-\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_t^{\nu^2}(X_t^{\nu^2}))\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_t\>.
\end{split}
\end{equation*}
Thus, taking expectation on both side yields,
$$\E|X_t^{\nu^1}-X_t^{\nu^2}|^2\leq \E|X_s^{\nu^1}-X_s^{\nu^2}|+K(T)\int_s^t\big(\E|X_r^{\nu^1}-X_r^{\nu^2}|^2+\W_2(\nu^1_r,\nu^2_r)\big)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r.$$
Noting that, $X_s^{\nu^1}=X_s^{\nu^2},$ so we have
$$\E|X_t^{\nu^1}-X_t^{\nu^2}|^2\leq K(T)\int_s^t\W_2(\nu^1_r,\nu^2_r)^2\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r+K(T)\int_s^t \E|X_r^{\nu^1}-X_r^{\nu^2}|^2\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r.$$
Then Growall's inequality gives
\begin{equation*}
\E|X_t^{\nu^1}-X_t^{\nu^2}|^2\leq K(T)\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{K(T)T}\int_s^t\W_2(\nu^1_r,\nu^2_r)^2\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r.
\end{equation*}
This implies for any $\lambda>0,$
\begin{equation}\label{EQ5}
\begin{split}
\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-2\lambda t}\E|X_t^{\nu^1}-X_t^{\nu^2}|^2&\leq K(T)\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{K(T)T}\int_s^t \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-2\lambda(t-r)} \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-2\lambda r}\W_2(\nu^1_r, \nu^2_r)^2\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r\\
&\leq \frac{K(T)\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{K(T)T}}{2\lambda}\W_{2,\lambda}(\nu^1, \nu^2)^2.
\end{split}
\end{equation}
Observe that
\begin{equation*}
\begin{split}\W_{2,\lambda}(\Phi(\nu^1),\Phi(\nu^2))&= \sup_{t \in [s,T]} \big(\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-\lambda t} \W_2\big(\Phi(\nu^1))_t,(\Phi(\nu^2))_t\big)\big) \leq \sup_{t\in[s,T]} \big(\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-2\lambda t}\E|X_t^{\nu^1}-X_t^{\nu^2}|^2\big)^{\frac{1}{2}}\\
&\leq \Big(\frac{K(T)\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{K(T)T}}{2\lambda}\Big)^{\frac{1}{2}}\W_{2,\lambda}(\nu^1, \nu^2),
\end{split}
\end{equation*}
where the last inequality is due to \eqref{EQ5}. Then, by taking $\lambda=4K(T)\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{K(T)T},$ we conclude
$$W_{2,\lambda}(\Phi(\nu^1),\Phi(\nu^2))\leq \frac{1}{2}\W_{2,\lambda}(\nu^1,\nu^2).$$
This implies that $\scr D_+\ni \nu \mapsto \Phi(\nu)$ is contractive under the metric $\W_{2,\lambda}.$ Therefore, the Banach fixed point theorem implies that the mapping $\nu \mapsto \Phi(\nu)$ has a unique fixed point, denoted by $\nu.$ Consequently, we have
$$(\Phi(\nu))_t=\nu_t=\scr L_{X_t^{\nu}}\in\scr P_+, ~~t\in[s,T],$$ so that $\LL_{s,T}:=\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{(X_t^\nu)_{t\in [s,T]}}$.
Therefore, by applying Theorem \ref{T1.6} to the diffusion process generated by $L_t$ with coefficients $(b^\nu,a^\nu)$, we conclude that the present conditions (1) and (2) imply \eqref{*G}
as desired.
\end{proof}
\begin{proof}[Proof of Theorem \ref{T5}]
Consider the decoupled SDE
\begin{equation}
\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D X_{s,t}^{x,\mu}=b(t, X_{s,t}^{x,\mu}, P_{s,t}^*\mu)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}(t, X_{s,t}^{x,\mu}, P_{s,t}^*\mu)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_t,~\quad t\ge s, \, ~X_{s,s}^{x,\mu}=x,
\end{equation}
where $P_{s,t}^*\mu$ is the marginal distribution of $\Lambda_s$ at the time $t.$
For $x\in \R^d,$ let $\Lambda_s^{x,\mu}=\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_{[s,\infty)}^{x,\mu}}$. Then, for any $\nu \in \scr P,$
$$\Lambda_s^{\nu,\mu}:=\int_{\R^d} \Lambda_s^{x,\mu}\nu(dx)$$is the law of $X_{[s,\infty)}^{\nu,\mu}$ with initial distribution $\nu.$ Note that
$$\Lambda_s^{\mu}=\Lambda_s^{\mu,\mu}=\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_{[s,\infty)}^{\mu}}.$$
Since $\Lambda_s^{\mu}\in\scr P_+^s$, we have
\begin{equation}\label{FKG3}
\Lambda_s^{\mu}(FG)\geq \Lambda_s^{\mu}(F)\Lambda_s^{\mu}(G),~~F,G\in \scr U(C_s), \mu\in\scr P_+.
\end{equation}
For $\gamma\in C_s$, let $F(\gamma)=f(\gamma_s)$ with $0\le f\in \scr U(\R^d).$ Then \eqref{FKG3} becomes
$$\Lambda_s^{\nu, \mu}(G)\geq \Lambda_s^{\mu, \mu}(G),~G\in \scr U(C_s),$$
where $\nu(dx):=\frac{f(x)\mu(dx)}{\mu(f)}.$ That is, $\Lambda_s^{\nu, \mu} \geq \Lambda_s^{\mu, \mu}$. Then there exit $\pi_s \in \scr C} \def\aaa{\mathbf{r}} \def\r{r(\Lambda_s^{\nu, \mu},\Lambda_s^{\mu, \mu})$ and Brownian motions $B_t^1~\mbox{and}~B_t^2$ on $(\Omega, \F_t, \P):=(C_s, \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}(\gamma_r: r\in [s,t]), \pi_s)$ such that
\begin{equation}\label{PSDE3}
\begin{cases}
&\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \xi_t= b(t, \xi_t, P_{s,t}^*\mu) \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+ \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}(t, \xi_t, P_{s,t}^*\mu) \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_t^1,~~\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\xi_s}=\nu,~~t\geq s\\
& \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \eta_{t}=b(t,\eta_{t}, P_{s,t}^*\mu) \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}(t,\eta_{t}, P_{s,t}^*\mu) \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_t^2,~~\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\eta_s}=\mu,~~ t\geq s.
\end{cases}
\end{equation}
satisfy $\pi_s(\xi\geq\eta)=1.$
For any increasing $0\le f\in \scr U_b$, which does not depend on $x_i, x_j,$ we have $\nu:=\frac{f d\mu}{\mu(f)}$ with $\mu=\mu_{\{ij\}}\times\mu_{\{ij\}^c}$ such that $\nu_{ij}=\mu_{ij}.$ Thus, $\xi_s^i=\eta_s^i,~\xi_s^j=\eta_s^j.$ So,
\begin{equation}\label{PSDE4}
\begin{split}
&\xi_t^k-\eta_t^k= \int_s^t \big( b_k(r, \xi_r, P_{s,r}^*\mu)- b_k(r, \eta_r, P_{s,r}^*\mu)\big) \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r\\
&\quad+ \int_s^t \big\< \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{k\cdot}(r,\eta_{r}, P_{s,r}^*\mu), \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_t^1\big\>-\int_s^t \big\< \sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{k\cdot}(r,\eta_{r}, P_{s,r}^*\mu), \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_t^2\big\>\geq 0,\quad k=i,j.
\end{split}
\end{equation}
Thus, by following the argument to derive \eqref{NN5}, we have
$$a_{ij}(s,\xi_s,\mu)=\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{i\cdot}(s,\xi_s,\mu)\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{i\cdot}(s,\xi_s,\mu)=\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{i\cdot}(s,\eta_s,\mu)\sigma} \def\ess{\text{\rm{ess}}}\def\s{{\bf s}_{j\cdot}(s,\eta_s,\mu)=a_{ij}(s,\eta_s,\mu).$$
This, together with $\nu=\frac{f d\mu}{\mu(f)}$, $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\xi_s}=\nu$ and $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\eta_s}=\mu$, leads to
\begin{equation}\label{PSDE5}
\begin{split}
\int f(x) a_{ij}(s,x,\mu)\mu(dx)&= \mu(f)\E a_{ij}(s,\xi_s,\mu)\\
&= \mu(f)\E a_{ij}(s,\eta_s,\mu)=\mu(f)\int a_{ij}(s,x,\mu)\mu(dx).
\end{split}
\end{equation}
Let $g$ be a function such that $$\mu(fg)=\mu(f)\mu(g).$$ Then, for $f(x)=I_{A}(x_k: k\neq i,j)$ with $A\in \scr B(\R^{(d-2})),$
we obtain
$$\E^{\mu}(I_{A}g)=\int _{\R^d} I_{A}(x)g(x)\mu(dx)=\mu(A)\mu(g).$$
Now, by the definition of conditional expectation we get
$$\E^{\mu}\big( g|x_k:k\neq i,j\big)=\mu(g),$$ which obviously implies that $g$ depends only on $x_i,x_j$. Thus, \eqref{PSDE5} yields the first assertion.
Dividing by $t-s$ on both side of \eqref{PSDE4} and taking $t\rightarrow}\def\l{\ell}\def\iint{\int}\def\gg{\gamma s,$ we get
$$\E b_i(s, \xi_s, \mu)\geq \E b_i(s, \xi_s, \mu).$$
This, together with $\nu=\frac{f d\mu}{\mu(f)}$, $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\xi_s}=\nu$ and $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\eta_s}=\mu$, leads to the second assertion. \end{proof}
\paragraph{Acknowledgement.} The author would like to thank Professor Feng-Yu Wang for his helpful comments.
|
\section{Introduction}
Lie algebras are vector spaces
equipped with an antisymmetric bracket satisfying the Jacobi identity. Many interesting physical theories can be cast in this language, but others require a suitable generalization of this program. This is the case of Double Field Theory (DFT) \footnote{For reviews check \cite{ReviewDFT} and references therein.} \cite{Siegel} \cite{DFT}, a proposal to incorporate T-duality as a symmetry of a field theory, since it contains a non-trivial Jacobiator and therefore satisfies the Jacobi identity up to homotopy. For this reason its algebraic structure requires a set of brackets defined on a graded vector space satisfying a generalized notion of the Jacobi identities. Such structures are known as $L_{\infty}$ algebras and were initially described in the context of closed string field theory \cite{Zwiebach:1992ie} and, in the mathematics literature, in topology \cite{Topology}.
One way of organizing the algebraic structure of DFT in an $L_\infty$ structure turns from noticing that the Courant algebroids can be cast in this language \cite{Roytenberg:1998vn}, as well as their duality covariant counterparts \cite{Hull:2009zb} \cite{Deser:2016qkw}. Moreover, when dynamics is taken into account the full DFT, written in the generalized metric approach, also fit in an $L_{\infty}$ structure, as described in \cite{Hohm:2017pnh} \cite{Otros}.
In this work we are interested in Gauged Double Field Theory (GDFT) \cite{GDFT} with $O(D,D+n)$ as global duality group, where $n$ is the dimension of a gauge group. This formalism is a generalization of DFT and requires a frame \cite{frame} or flux formalism \cite{flux} in order to introduce the generalized version of the structure constants $f_{M N P}$. Additionally, the generalized Lie derivative acting on a generic vector $V_{M}$ with $M=0,\dots,2D-1+n$ is consistently deformed,
\begin{eqnarray}
\widehat{\cal L}_\xi V_M = {\cal L}_{\xi} V_{M} + f_{M N P} \xi^{N} V^{P} \, ,
\end{eqnarray}
and the closure is given by a deformed bracket
\begin{eqnarray}
[\xi_1, \xi_{2} ]^{ M}_{(C_{f})}=2\xi^{ P}_{[1}\partial_{ P}\xi_{2]}^{ M}-\xi_{[1}^{ N}\partial^{ M}\xi_{2] N}+f_{{ PQ}}{}^{ M} \xi_{1}^{ P} \xi_2^{ Q}\, ,
\end{eqnarray}
which reduces to the C-bracket when the structure constants vanish. As expected, the Jacobiator is also deformed
\begin{equation}
J(\xi_{1},\xi_{2},\xi_3)^{M} = \frac{3}{2}\partial^{M}(\xi_{[1}^{N} \xi_{2}^{P} \partial_{N} \xi_{3]P} + \frac13 f_{N P Q} \xi_{1}^{N} \xi_{2}^{P} \xi_{3}^{Q}) \, .
\end{equation}
The inclusion of the generalized frame/fluxes introduces a double Lorentz symmetry given by $O(D-1,1)_{L} \times O(1,D-1+n)_{R}$. From a $L_{\infty}$ point of view all these new ingredients enrich the algebraic structure of DFT or, in other words, define the algebraic structure of GDFT.
The products related to the dynamics of the theory can be cast in a closed form if we restrict our study to a family of theories given by the generalized Kerr-Schild ansatz. This ansatz was introduced in the context of DFT in \cite{KL}, extended to heterotic DFT in \cite{KL2} and \cite{alpha}, and further explore in the context of duality covariant theories in \cite{GKSA}. In this ansatz the perturbation of the generalized exact frame is given by
\begin{eqnarray}
{\cal E}_{M}{}^{\overline A} = E_{M}{}^{\overline A} + \frac12 \kappa E_{M}{}^{\underline B} K_{\underline B} {\bar K}^{\overline A} \, , \nonumber \\ {\cal E}_{M}{}^{\underline A} = E_{M}{}^{\underline A} - \frac12 \kappa E_{M}{}^{\overline B} {\bar K}_{\overline B} K^{\underline A}\, ,
\label{introan}
\end{eqnarray}
where $K_{\underline A}$ and $\bar K_{\overline A}$ are a pair of generalized null vectors,
\begin{eqnarray}
K_{\underline A} K^{\underline A} & = & \bar K_{\overline A} \bar K^{\overline A} = 0 \, ,
\end{eqnarray}
and $\kappa$ in (\ref{introan}) is an order parameter. We use $\underline A$, $\overline A$ as the flat left and right projections of the $M,N$ indices. The vectors $K_{\underline A}$ and $\bar K_{\overline A}$ satisfy the equivalent of a geodesic condition in the context of DFT,
\begin{eqnarray}
K^{\underline A} D_{\underline A} {\bar K}^{\overline B} =
{\bar K}^{\overline A} D_{\overline A} {K}^{\underline B} = 0 \, ,
\end{eqnarray}
where $D_{A}$ is a generalized covariant derivative. The ansatz (\ref{introan}) plus a linear expansion for the generalized dilaton,
\begin{eqnarray}
d = d_{o} + \kappa f
\end{eqnarray}
with $K^{\underline A} E_{\underline A} f = K^{\overline A} E_{\overline A} f = 0$ provide a family of exact solutions in a perturbative framework. In this sense, all the non-trivial products of the $L_{\infty}$ structure of GDFT can be explicitly/exactly computed, as we show. Considering a $L^{\rm{gauge+fields}}_{\infty}$ structure the theory can be cast in an $L_{3}$ algebra, where new brackets related to the generalized structure constants and the double Lorentz transformations are computed. When one also considers the equation of motion of the fields the algebraic structure is exactly promoted to an $L_{4}$ algebra.
This work is organized as follows: in Section \ref{GKSAA} we introduce GDFT in the generalized metric/flux formulation. Here we present the generalized Kerr-Schild ansatz (GKSA) for flat backgrounds. In Section \ref{Linf} we start by reviewing the way to obtain the products for a generic $L_{\infty}$ algebra. Then we cast the algebraic structure of both DFT and GDFT when the GKSA is considered. The present computations show the algebraic structure of the fundamental charged heterotic string and (Bosonic) Enhanced Double Field Theory, as we show in Section \ref{App}. Finally in Section \ref{Con} we summarize our work.
\section{The generalized Kerr-Schild ansatz}
\label{GKSAA}
\subsection{The DFT approach in metric formalism}
The GKSA is given by an exact and linear perturbation of the generalized background metric $H_{M N}$ ($M,N=0, \dots, 2D-1$) and an exact perturbation of the generalized background dilaton $d_{o}$. In this work we will consider linear perturbations in both fields.
The perturbation of the generalized background metric $H_{M N}$ is given by a pair of generalized vectors, $K_{M}$ and $\bar{K}_{M}$, and an order parameter $\kappa$, such that
\begin{eqnarray}
{\cal H}_{MN} = H_{MN} + \kappa (\bar{K}_{M} K_{N} + {K}_{M} \bar{K}_{N} ) \, ,
\label{DFTKS}
\end{eqnarray}
while the vectors satisfy
\begin{eqnarray}
\bar{K}_{M} & = & \frac12({\eta}_{M N} + {H}_{M N}) \bar{K}^{N} = \bar{P}_{M N} \bar{K}^{N} \, , \nonumber \\
K_{M} & = & \frac12({\eta}_{M N} - {H}_{M N}) K^{N} = {P}_{M N} {K}_{N} \, ,
\end{eqnarray}
and the generlized null conditions,
\begin{eqnarray}
\eta^{MN} \bar{K}_{M} \bar{K}_{N} & = & \eta^{MN} K_{M} K_{N} = \eta^{MN} \bar{K}_{M} K_{N} = 0 \, .
\label{nulldft}
\end{eqnarray}
or, equivalently,
\begin{eqnarray}
H^{MN} \bar{K}_{M} \bar{K}_{N} & = & H^{MN} K_{M} K_{N} = H^{MN} \bar{K}_{M} K_{N} = 0 \, .
\label{nullHdft}
\end{eqnarray}
The generalized background dilaton $d_{o}$ is perturbed in a similar way\footnote{We work with a linear perturbation for simplicity, \textit{i.e.,} $f=\textrm{const.}$. In the general case $f = \sum_{n=0}^{\infty}\kappa^{n}f_{n}$},
\begin{equation}
d = d_{o} + \kappa f\, .
\label{dilaton}
\end{equation}
The perturbations of the GKSA satisfy the following extra conditions
\begin{eqnarray}
{\bar K}^{P} \partial_{P} K^{M} + K_{P} \partial^{M}{\bar K}^{P} - K^{P} \partial_{P}{\bar K}^{M} & = & 0 \nonumber \, , \\
K^{M} \partial_{M}f = {\bar K}^{M} \partial_{M}f & = & 0 \, ,
\label{geodesic2}
\end{eqnarray}
which play the role of generalized geodesic equations.
In addition to the global $O(D,D)$ symmetry, the principle action of DFT is invariant under generalized diffeomorphisms generated infinitesimally by $\xi^{M}$ through the generalized Lie derivative. Acting on an arbitrary vector it reads,
\begin{eqnarray}
{\cal L}_\xi V_M = \xi^{N} \partial_N V_M + (\partial_M \xi^N - \partial^N \xi_{M}) V_N + \omega (\partial_{N} \xi^{N})V_{M} \, ,
\label{glie}
\end{eqnarray}
where $\omega$ is a weight constant. The generalized metric ${\cal H}_{M N}$ and the generalized background metric $H_{M N}$ are tensors with $\omega=0$ with respect to generalized diffeomorphisms, and $\omega(e^{-2d})= \omega(e^{-2d_{o}})=1$. It is straightforward to check that conditions (\ref{geodesic2}) are covariant under generalized diffeomorphism transformations.
The Lagrangian of DFT is defined as,
\begin{eqnarray}
{\cal L}_{DFT} & = & e^{-2d}(\frac18 {\cal H}^{MN} \partial_{M}{\cal H}^{KL}\partial_{N}{\cal H}_{KL} - \frac12 {\cal H}^{MN}\partial_{N}{\cal H}^{KL}\partial_{L}{\cal H}_{MK} \nonumber \\ && + 4 {\cal H}^{MN} \partial_{M}d\partial_{N}d - 2 \partial_{M}{\cal H}^{MN} \partial_{N}d) \, ,
\label{scalarDFT}
\end{eqnarray}
while the equations of motion can be written in terms of generalized curvatures,
\begin{eqnarray}
{\cal R}_{\underline P \overline Q} = {\cal P}_{P}{}^{M} \bar {\cal P}_{Q}{}^{N} \Big(\frac 18 \partial_M {\cal H}^{KL} \partial_N {\cal H}_{KL} -
\frac 1 4 (\partial_L - 2 (\partial_L d)) ({\cal H}^{LK} \partial_K {\cal
H}_{MN}) + 2 \partial_M \partial_N d \nonumber \\
- \frac 1 2 \partial_{(M|} {\cal H}^{KL} \partial_L {\cal H}_{|N)K} + \frac
1 2 (\partial_L - 2 (\partial_L d)) ({\cal H}^{KL} \partial_{(M} {\cal H}_{N)K}
+ {\cal H}^K{}_{(M|}\partial_K {\cal H}^L{}_{|N)})\Big) = 0 \, ,
\label{Ricci}
\end{eqnarray}
and
\begin{eqnarray}
{\cal R} = && \frac18 {\cal H}^{MN} \partial_{M}{\cal H}^{KL}\partial_{N}{\cal H}_{KL} - \frac12 {\cal H}^{MN}\partial_{N}{\cal H}^{KL}\partial_{L}{\cal H}_{MK} + 4 {\cal H}^{MN} \partial_{M}\partial_{N}d \nonumber \\ && + 4 \partial_{M}{\cal H}^{MN} \partial_{N}d - 4 {\cal H}^{MN} \partial_{M}d \partial_{N}d - \partial_{M} \partial_{N} {\cal H}^{MN} = 0 \, .
\label{scalar}
\end{eqnarray}
\subsection{Extension to GDFT in flux formalism}
The ansatz (\ref{DFTKS}) and (\ref{dilaton}) are powerful tools to work pertubatively since the generalized null and geodesic conditions provide finite contributions to the action principle and the equations of motion. Interestingly enough (\ref{DFTKS}) admits an extension to the flux formulation of DFT, which is a mandatory step to consider a GDFT. In this case we consider perturbations of the form,
\begin{eqnarray}
{\cal E}_{M}{}^{\overline A} = E_{M}{}^{\overline A} + \frac12 \kappa E_{M}{}^{\underline B} K_{\underline B} {\bar K}^{\overline A} \, , \nonumber \\ {\cal E}_{M}{}^{\underline A} = E_{M}{}^{\underline A} - \frac12 \kappa E_{M}{}^{\overline B} {\bar K}_{\overline B} K^{\underline A}\, ,
\label{GKSA}
\end{eqnarray}
where $K_{\underline A} = {\cal E}^{M}{}_{\underline A} K_{M}={E}^{M}{}_{\underline A} K_{M}$ and $\bar{K}_{\overline A} = {\cal E}^{M}{}_{\overline A} \bar{K}_{M}=E^{M}{}_{\overline A} \bar{K}_{M}$ and ${\cal E}_{MA}$ is an $O(D,D+n)/O(D-1,1)_{L} \times O(1,D-1+n)_{R}$ frame. Here $\underline A= 0, \dots, D-1$ and $\overline A=0, \dots, D-1+n$ are $O(D-1,1)_{L}$ and $O(1,D-1+n)_{R}$ indices, respectively. In agreement with the previous section, we are going to consider a constant generalized frame background, \textit{i.e.}, $\partial_{M}E_{N A}=0$ and a constant generalized dilaton background $\partial_{M}{d_{o}}=0$ .
Defining $\eta_{A B}$ and $H_{A B}$ as the invariant metrics of $O(D-1,1)_{L} \times O(1,D-1+n)_{R}$, we have,
\begin{eqnarray}
\eta_{AB} = {\cal E}_{M A}\eta^{MN} {\cal E}_{N B} = E_{M A}\eta^{MN} E_{N B} \, , \\
{H}_{AB} = {\cal E}_{MA} {\cal H}^{MN} {\cal E}_{N B} = E_{MA} {H}^{MN}E_{N B} \, .
\end{eqnarray}
The generalized fluxes take the form
\begin{eqnarray}
{\cal F}_{ABC} & = & 3 {\cal E}_{[A}({\cal E}^{M}{}_{B}){\cal E}_{M C]} + \sqrt{2} f_{M N P} {\cal E}^{M}{}_{A} {\cal E}^{N}{}_{B} {\cal E}^{P}{}_{C} \, , \nonumber \\ {\cal F}_{A} & = & \sqrt{2}e^{2d}\partial_{M}\left({\cal E}^{M}{}_{A}e^{-2d}\right) \, ,
\end{eqnarray}
where $f_{M N P}$ plays the role of generalized structure constants and therefore satisfy
\begin{eqnarray}
f_{MNP}=f_{[MNP]}\, , \qquad f_{[MN}{}^{ R}f_{{P}]R}{}^{Q}=0\, , \label{consf}
\end{eqnarray}
and
\begin{eqnarray}
f_{ {MN}}{}^{ P}\partial_{P}\cdots =0 \, .
\label{fcond}
\end{eqnarray}
The generalized Lie derivative is deformed as,
\begin{eqnarray}
\widehat{\cal L}_\xi V_M = {\cal L}_{\xi} V_{M} + f_{M N P} \xi^{N} V^{P} \, ,
\label{glie2}
\end{eqnarray}
and, in addition, the theory is invariant under $O(D-1,1)_{L} \times O(1,D-1+n)_{R}$ or double Lorentz transformations,
\begin{eqnarray}
\delta_{\Gamma} V^{A} = V^{B} \Gamma_{B}{}^{A} \, ,
\end{eqnarray}
where $V^{A}$ is a generic vector and $\Gamma_{A B} = - \Gamma_{B A}$ an arbitrary parameter. The previous transformations close with the following parameters
\begin{eqnarray}\label{par0}
\xi^{ M}_{12} & = & [\xi_1, \xi_{2} ]^{ M}_{(C_f)} \, ,\\
\Gamma_{12 { A} { B}} & = & 2 \xi_{[1}^{ P} \partial_{ P} \Gamma_{2] { A} { B}} - 2 \Gamma_{[1 A}{}^{ C} \Gamma_{2] {C B}}
\, , \label{Lorentzbrack}
\end{eqnarray}
where the $C_f$-bracket is a deformation of the $C$-bracket given by
\begin{eqnarray}
[\xi_1, \xi_{2} ]^{ M}_{(C_{f})}=2\xi^{ P}_{[1}\partial_{ P}\xi_{2]}^{ M}-\xi_{[1}^{ N}\partial^{ M}\xi_{2] N}+f_{{ PQ}}{}^{ M} \xi_{1}^{ P} \xi_2^{ Q}\, , \label{Cbrack}
\end{eqnarray}
where (\ref{fcond}) is required for consistency.
A flat covariant derivative acting on a generic vector is given by
\begin{eqnarray}
{\cal D}_{A} V_{B} = {\cal E}_{A} V_{B} + {\cal W}_{AB}{}^{C} V_{C} \, ,
\label{covder}
\end{eqnarray}
where ${\cal E}_{A} = \sqrt{2} {\cal E}^{M}{}_{A} \partial_{M}$. The covariant derivative as well as the flat derivative can also be defined for background fields in a similar fashion. In (\ref{covder}) ${\cal W}_{AB}{}^{C}$ is the generalized spin connection, which is partially identified with the generalized fluxes according to
\begin{eqnarray}
{\cal W}_{[ABC]} & = & -\frac13{\cal F}_{ABC}\, , \\
{\cal W}_{BA}{}^{B} & = & - {\cal F}_{A}\, ,
\end{eqnarray}
in order to have, on the one hand, frame compatibility under covariant derivation and, on the other, partial integration with respect to the dilaton density, \textit{i.e.},
\begin{equation}
\int e^{-2d}V {\cal D}_{ A}V^{ A}=-\int e^{-2d}V^{ A} {\cal D}_{ A}V\, .
\end{equation}
Considering the flat projectors as $P_{A B}=\frac12 \eta_{A B} - \frac12 H_{A B}$ and $\bar P_{A B}=\frac12 \eta_{A B} + \frac12 H_{A B}$, and the notation
\begin{eqnarray}
V_{A}=V_{\underline A} + V_{\overline A}= P_{\underline A}{}^{\underline B} V_{\underline B} + \bar P_{\overline A}{}^{\overline B} V_{\overline B} \, ,
\end{eqnarray}
the generalized curvatures (\ref{Ricci}) and (\ref{scalar}) are rewritten as
\begin{eqnarray}
\label{GRicci_scalar}
{\cal R} & = & 2{\cal E}_{\underline{A}}{\cal F}^{\underline{A}} + {\cal F}_{\underline{A}}{\cal F}^{\underline{A}} - \frac16 {\cal F}_{\underline{ABC}} {\cal F}^{\underline{ABC}} - \frac12{\cal F}_{\overline{A}\underline{BC}}{\cal F}^{\overline{A}\underline{BC}} \, , \\
{\cal R}_{\overline{A}\underline{B}} & = & {\cal E}_{\overline{A}}{\cal F}_{\underline{B}} - {\cal E}_{\underline{C}}{\cal F}_{\overline{A}\underline{B}}{}^{\underline{C}} + {\cal F}_{\underline{C}\overline{DA}}{\cal F}^{\overline{D}}{}_{\underline{B}}{}^{\underline{C}} - {\cal F}_{\underline{C}}{\cal F}_{\overline{A}\underline{B}}{}^{\underline{C}} \, .
\label{GRicci_tensor}
\end{eqnarray}
while the relevant projections of the fluxes are written in terms of $K_{A}$ and $\bar K_{\bar A}$ in the following way,
{\footnotesize
\begin{eqnarray}
{\cal F}_{\underline{ABC}} & = & \sqrt{2} f_{M N P} \Bigg( { E}^{M}{}_{{\underline A}} { E}^{N}{}_{{\underline B}} {E}^{P}{}_{\underline C} -\frac{1}{2} \kappa K_{{\underline A}} {\overline K}_{{\overline B}} E^{M{\overline B}} E^{N}{}_{{\underline B}} E^{P}{}_{{\underline C}} -\kappa E^{M}{}_{{\underline A}}K_{[{\underline B}|} {\overline K}_{{\overline C}} E^{N{\overline C}} E^{P}{}_{|{\underline C}]} \Bigg), \nonumber \\
{\cal F}_{\underline{A}\overline{BC}} & = & \kappa\left(\bar K{}_{[\overline{C}}D{}_{\overline{B}]}K_{\underline{A}} + K_{\underline{A}}E_{[\overline{B}}\bar K_{\overline{C}]} \right) + \sqrt{2} f_{M N P} \Bigg( { E}^{M}{}_{{\underline A}} { E}^{N}{}_{{\overline B}} {E}^{P}{}_{\overline C} - \frac12 \kappa K_{\underline A} \bar K_{\overline D} E^{M \overline D} E^{N}{}_{\overline B} E^{P}{}_{\overline C} \Bigg)\, , \nonumber \\
{\cal F}_{\overline{A}\underline{BC}} & = & - \kappa\left(K_{[\underline{C}}D_{\underline{B}]}\bar K_{\overline{A}} + \bar K_{\overline{A}}E_{[\underline{B}}K_{\underline{C}]} \right) + \sqrt{2} f_{M N P} \Bigg( { E}^{M}{}_{{\overline A}} { E}^{N}{}_{{\underline B}} {E}^{P}{}_{\underline C} + \frac12 \kappa \bar K_{\overline A} K_{\underline D} E^{M \underline D} E^{N}{}_{\underline B} E^{P}{}_{\underline C} \Bigg) \, , \nonumber \\
{\cal F}^{\underline{A}} & = & - \frac{1}{2}\kappa\left( (E_{\bar B}\bar K^{\overline B})K^{\underline A} + (E_{\overline B} K^{\underline A}) \bar K^{\overline B} + 4E^{\underline{A}}f\right)\, .
\label{constrained_fluxes}
\end{eqnarray}}
The flat version of the null conditions reads
\begin{eqnarray}
K_{\underline A} K^{\underline A} & = & \bar K_{\overline A} \bar K^{\overline A} = 0 \, , \label{flatnull}
\end{eqnarray}
and the flat geodesic conditions now contain a contribution related to the generalized structure constants,
\begin{eqnarray}
K^{\underline A} E_{\underline A} {\bar K}^{\overline C} + \sqrt{2} {K}^{\underline A} \bar K^{\overline B} f_{M P Q} E^{M}{}_{\underline A} E^{P}{}_{\overline B} E^{Q \overline C} & = & 0 \, , \\
{\bar K}^{\overline A} E_{\overline A} {K}^{\underline C} + \sqrt{2} {\bar K}^{\overline A} K^{\underline B} f_{M P Q} E^{M}{}_{\overline A} E^{P}{}_{\underline B} E^{Q \underline C} & = & 0 \, , \\
K^{\underline A} E_{\underline A} f = {\bar K}^{\overline A} E_{\overline A} f & = & 0 \, .
\label{flatgeo}
\end{eqnarray}
\section{$L_{\infty}$ algebras}
\label{Linf}
In this section we start by reviewing how to fit DFT in an $L_{\infty}$ algebra and then we show the extension to GDFT. We always consider the GKSA in order to obtain closed expressions when dynamics is taken into account and we dedicate next section to discuss about the family of theories that can be described within this approach.
\subsection{Basics}
\label{Basics}
Let us consider a vector graded space $X$ which is the direct sum of vector spaces $X_n$, each of which has degree $n$
\begin{equation}
X = \bigoplus_{n} X_n \,, \quad n \in \mathbb{Z} \ .
\end{equation}
We will denote by $x$ an element of $X$ with definite degree, $i.e$, $x\in X_p$ for some fixed $p$.
We consider multilinear products $\ell_k$
\begin{equation}
\ell_k: X^{\otimes k} \rightarrow X \ ,
\end{equation}
with degree given by
\begin{equation}
\hbox{deg}(\ell_k(x_1,x_2,...,x_k))= k-2 + \sum_{i=1}^k \hbox{deg} (x_i) \ .
\end{equation}
For a permutation $\sigma$ of $k$ labels we have
\begin{equation}
\ell_k ( x_{\sigma(1)} , \ldots , x_{\sigma(k)} ) \ = \ (-1)^\sigma \epsilon(\sigma;x ) \,
\ell_k (x_1 , \ldots \,, x_k) \ .
\end{equation}
The $(-1)^\sigma$ factor gives a plus or minus sign if the permutation is even or odd, respectively. The $\epsilon(\sigma;x )$ factor is the Koszul sign. For a graded commutative algebra $\Lambda (x_1, x_2, \cdots )$ with
\begin{equation}
x_i \wedge x_j \ = \ (-1)^{{\rm deg}(x_i) {\rm deg}(x_j)} \, x_j \wedge x_i \,, \quad \forall i, j\ ,
\end{equation}
the Koszul sign for a general permutation is given by
\begin{equation}
x_1\wedge \ldots \wedge x_k = \epsilon (\sigma; x) \ x_{\sigma(1)} \wedge \ldots \wedge \, x_{\sigma(k)} \ .
\end{equation}
It is convenient to abuse with the notation in the following way
\begin{equation}
\ (-1)^{{\rm deg}(x_i) {\rm deg}(x_j)}\equiv(-1)^{x_i x_j}\ .
\end{equation}
The $L_\infty$ relations are labeled by a positive integer $n$ given by the number of inputs. Explicitly they are
\begin{equation}
\label{main-Linty-identity}
\sum_{i+j= n+1} (-1)^{i(j-1)} \sum_\sigma (-1)^\sigma \epsilon (\sigma; x) \, \ell_j \, \bigl( \, \ell_i ( x_{\sigma(1)} \,, \, \ldots\,, x_{\sigma(i)} ) \,, \, x_{\sigma(i+1)}, \, \ldots \, x_{\sigma (n)} \bigr) \ = \ 0\ .
\end{equation}
The sum over $\sigma$ is a sum over ``unshuffles'', it includes only the terms which satisfy
\begin{equation}
\sigma(1) < \, \cdots \, < \, \sigma(i) \,, \qquad
\sigma(i+1) < \, \cdots \, < \, \sigma(n) \ .
\end{equation}
It is common to write these relations as
\begin{equation}
\label{main-Linty-identity-schem}
\sum_{i+j= n+1} (-1)^{i(j-1)} \ell_j \, \ell_i \ = \ 0\ ,
\end{equation}
such that
\begin{eqnarray}
n = 1 \ \ \ \ \ \ \ \ 0 &=& \ell_1 \ell_1 \\
n = 2 \ \ \ \ \ \ \ \ 0 &=& \ell_1 \ell_2 - \ell_2 \ell_1 \\
n= 3 \ \ \ \ \ \ \ \ 0 &=& \ell_1 \ell_3 + \ell_2 \ell_2 + \ell_3 \ell_1 \\
n = 4 \ \ \ \ \ \ \ \ 0 &=& \ell_1 \ell_4 - \ell_2 \ell_3 + \ell_3 \ell_2 - \ell_4 \ell_1 \ , \ \dots
\end{eqnarray}
For instance, the $n=3$ case is given by
\begin{eqnarray}
0 & = & \ell_2(\ell_2(x_1,x_2),x_3) + (-1)^{(x_1+ x_2) x_3}\ell_2(\ell_2(x_3,x_1),x_2)
+(-1)^{(x_2+ x_3) x_1 }\ell_2(\ell_2(x_2,x_3),x_1) \nonumber \label{L3L1}\\
&&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! + \ell_1(\ell_3 (x_1,x_2, x_3)) + \ell_3(\ell_1 (x_1) ,x_2, x_3)
+ (-1)^{x_1} \ell_3( x_1 ,\ell_1(x_2), x_3)
+ (-1)^{x_1+ x_2} \ell_3( x_1 ,x_2, \ell_1(x_3)) \ . \nonumber
\label{n=3}
\end{eqnarray}
One must assign a given degree $p$ to gauge parameters, fields, EOM's, etc., and so specify to what vector subspace $X_p$ they belong. In this work we consider that the space of degree two contains the constants ($c$), the space of degree one contains
functions ($\chi$), the space of degree zero contains the gauge parameters ($\zeta$), the space of degree minus one contains the fields ($\Psi$) and, finally, the space of degree minus two the dynamics (${\cal F}$). In general the products can be read from the symmetries and dynamics of a given field theory. The symmetry transformations define the brackets $\ell_{n+1}(\zeta, \Psi^n)$ as follows
\begin{equation}
\delta_{\xi}\Psi =\sum_{n\ge 0} \frac{1}{n!}
(-1)^{{n(n-1)}/{2}}\,
\ell_{n+1}(\xi, \Psi^n)
\ , \label{gaugel}
\end{equation}
where $
\Psi^k = \underbrace{\Psi,...,\Psi}_{k\;\text{times}}$. The equations of motion define the $l_n (\Psi^n)$ brackets as follows
\begin{equation}
{\cal F}(\Psi) = \sum_{n=1}^\infty
\frac{(-1)^{n(n-1)/ 2}}{n!} \ell_n(\Psi^n) \, .
\label{eoml}
\end{equation}
Both (\ref{gaugel}) and (\ref{eoml}) are fundamental relations that can be used to read non-trivial products, and then extra products can appeared upon checking the $L_{\infty}$ relations (\ref{main-Linty-identity}).
\subsection{GKSA-DFT as an $L_{3}$ algebra}
\label{GKSA-DFT}
Here we follow the construction presented in \cite{Hohm:2017pnh}. In that work the authors show that when the arguments of $l_2$ are the DFT gauge parameters, this product is related to the C-bracket. Moreover, the first line in (\ref{n=3}) coincides with the Jacobiator and the last line characterizes the non-trivial Jacobiator of DFT given by
\begin{equation}
J(\xi_{1},\xi_{2},\xi_3)^{M} = \frac{3}{2}\partial^{M}(\xi_{[1}^{N} \xi_{2}^{P} \partial_{N} \xi_{3]P}) = N^M \, .
\end{equation}
Considering the following relation derived from (\ref{gaugel}),(\ref{eoml}) and (\ref{main-Linty-identity}),
\begin{eqnarray}
\label{commurel}
[\delta_{\zeta_1},\delta_{\zeta_2}] \Psi=\delta_{-\mathbf C(\zeta_1,\zeta_2)} \Psi \, ,
\end{eqnarray}
with $\mathbf C(\zeta_1,\zeta_2) \equiv \;\ell_{2}(\zeta_1,\zeta_2)\ ,$
the non-trivial products are
\begin{eqnarray}
\ell_{1}(\chi) = \partial \chi \in X_{0}, \\
\ell_{1}(c) = \iota c \in X_{1},\\
\ell_{2}(\xi_{1},\xi_{2}) = \big[ \xi_{1},\xi_{2} \big]_{C} \in X_{0},\\
\ell_{2}(\xi , \chi) = \frac{1}{2} \xi^{K}\partial_{K}\chi \in X_{1},\\
\ell_{3}(\xi_{1},\xi_{2},\xi_{3}) = - N(\xi_{1},\xi_{2},\xi_{3}) \in X_{1}.
\end{eqnarray}
On the other hand, considering
\begin{eqnarray}
\delta_{\xi} {\cal H}_{M N} & = & \xi^{P} \partial_{P} {\cal H}_{M N} + 2 (\partial_{(M} \xi^{P} - \partial^{P} \xi_{(M}) {\cal H}_{N) P} \nonumber \\
\delta_{\xi} d & = & \xi^{P} \partial_{P}d - \frac12 \partial_{P}\xi^{P} \, ,
\end{eqnarray}
and (\ref{gaugel}), and invoking the GKSA it is straightforward to find
\begin{eqnarray}
\ell_{1}(\xi)_{t} & = & 2(\partial_{\underline M} \xi_{\overline N} - \partial_{\overline N} \xi_{\underline M}) \, , \\
\ell_{1}(\xi)_{s} & = & - \frac12 \partial_{P} \xi^{P} \, , \\
\ell_{2}(\xi,K)_{v} & = & \delta_{\xi} K_{M} \, , \\
\ell_{2}(\xi,\bar K)_{v} & = & \delta_{\xi} \bar K_{M} \, , \\
\ell_{2}(\xi,f)_{s} & = & \xi^{P}\partial_{P}f \, ,
\end{eqnarray}
where the letters $s$,$v$,$t$ means that we are considering the scalar, vectorial or tensorial part of the product, respectively.
\subsubsection{Pertubative DFT as an exact $L_{3}$ algebra}
The closed expressions for the dynamics can be easily obtained from (\ref{eoml}). Considering the equation of motion for the generalized dilaton we identify,
\begin{eqnarray}
\ell_{1}(f)_s & = & 4\kappa{H}^{K L} \partial_{K}\partial_{L}{f} \\
\ell_{2}(f,f)_s & = & 8\kappa^{2} {H}^{K L} \partial_{K}{f} \partial_{L}{f} \\
\ell_{2}(\bar K,K)_s & = & 4 \kappa \partial_{K} \partial_{L} ( K^{K} \bar{K}^{L}) \, ,
\end{eqnarray}
and, analogously, from the generalized metric equation we obtain,
\begin{eqnarray}
\ell_{1}(f)_t & = & 4 \kappa P_{K}{}^{M} \bar{P}_{L}{}^{N}\partial_{M N}{f} \\
\ell_{2}(\bar K,K)_t & = & \kappa \Big[ {H}^{M N} \partial_{M N}\big(K_{K} \bar{K}_{L}\big)-2\partial_{M N} \big( K^{N} \bar{K}_{L} P_{K}{}^{M} - K_{K} \bar{K}^{N} \bar{P}_{L}{}^{M}\big) \Big] \\
\ell_{3}(f,\bar K, K)_{t} & = & -6 \kappa^{2} \Big[\ {H}^{MN}\partial_{M}f \partial_{N}\big( K_{K} \bar{K}_{L}\big) -2 P_{K}{}^{M} \partial_{M}\big( K^{N}\bar{K}_{L}\partial_{N}f\big) \nonumber \\ && + 2\bar{P}_{L}{}^{M}\partial_{M}\big( K_{K}\bar{K}^{N}\partial_{N}f\big) \ \Big] \, ,
\end{eqnarray}
where $\partial_{M N}= \partial_{M} \partial_{N}$.
In order to verify the $L_{\infty}$ relations given by (\ref{main-Linty-identity})
it is necessary to only include extra products related to the gauge transformation of the equations of motion, $\ell_{2}(\xi,{\cal R})_s=\delta_{\xi}{\cal R}$ and $\ell_{2}(\xi,{\cal R}_{\underline M \overline N})_t=\delta_{\xi}{\cal R}_{\underline M \overline N}$, while the remaining products are null.
\subsection{GKSA-GDFT as an $L_{4}$ algebra}
\label{GKSA-GDFT}
The extension to GDFT will be performed in several steps. We start by considering only the subspaces related to the brackets algebra ($X_2$,$X_1$,$X_0$), then we include the subspace of fields $X_{-1}$ and finally we include both fields and their dynamics $X_{-2}$.
\subsubsection{GDFT bracket algebra as an $L_{3}$ algebra}
We start by discussing the
subalgebra corresponding to the pure gauge structure, given by the $C_f$-bracket algebra (\ref{Cbrack}) and the Double Lorentz bracket (\ref{Lorentzbrack}). The graded vector space is taken to contain three spaces of fixed degree,
\begin{eqnarray}
0 & \rightarrow X_2 & \rightarrow X_1 \rightarrow X_0 \nonumber \\ & \quad c & \quad \, \, \, \chi \, \, \, \quad \quad \, \zeta
\end{eqnarray}
where $\zeta=(\xi,\Lambda)$ is a generic parameter and the above arrows define the $\ell_1$ action. From $X_2$ to $X_1$ the action is given by the inclusion map, while from $X_1$ to $X_0$ the action is given by the partial derivative. Acting on $X_0$ the map $\ell_1$ is null since we are not considering the fields yet. At this level the non-trivial products are
\begin{eqnarray} \label{prod1}
\ell_{1}(\chi) = \partial \chi \in X_{0}, \\
\ell_{1}(c) = \iota c \in X_{1},\\
\ell_{2}(\xi_{1},\xi_{2}) = \big[ \xi_{1},\xi_{2} \big]_{C_{f}} \in X_{0},\\
\ell_{2}(\xi , \chi) = \frac{1}{2} \xi^{K}\partial_{K}\chi \in X_{1},\\
\ell_{3}(\xi_{1},\xi_{2},\xi_{3}) = - N_{o}(\xi_{1},\xi_{2},\xi_{3}) - N_{f}(\xi_{1},\xi_{2},\xi_{3}) \in X_{1} \, , \\
\ell_{2}(\xi,\Gamma) = \xi^{ P} \partial_{ P} \Gamma_{{ A} { B}} \in X_{0}, \label{extra}\\
\ell_{2}(\Gamma_{1},\Gamma_{2}) = - \Gamma_{1 A}{}^{ C} \Gamma_{2 {C B}} \in X_{0} \, , \label{prod7}
\end{eqnarray}
where $N_o$ and $N_f$ can be computed from the Jacobiator of GDFT,
\begin{equation}
J(\xi_{1},\xi_{2},\xi_3)^{M} = \frac{3}{2}\partial^{M}(\xi_{[1}^{N} \xi_{2}^{P} \partial_{N} \xi_{3]P} + \frac13 f_{N P Q} \xi_{1}^{N} \xi_{2}^{P} \xi_{3}^{Q}) = N_{o}^M + N_{f}^M \, .
\end{equation}
The bracket $\Gamma_{[1A}{}^{C} \Gamma_{2]C B}$ encodes the algebra of matrix multiplication and therefore the analogous of the Jacobiator for the Double Lorentz symmetry is trivially null. Moreover from (\ref{extra}) it is straightforward to show the following relation,
\begin{eqnarray}
\ell_{2}(\Gamma,\partial \chi)=\partial \ell_{2}(\Gamma,\chi) \, .
\end{eqnarray}
Using the previous relation and the products (\ref{prod1})-(\ref{prod7}) it is straightforward to show that the relations $n\geq4$ are trivial.
\subsubsection{Off-shell GDFT as extended $L_{3}$ algebra}
Now we extend the $L_{3}$ algebra describing the $C_{f}$ and the double Lorentz brackets to include the fields and the symmetry
transformations. We recall at this point that the generalized metric formalism has to be abandoned in order to describe the GDFT structure. The graded vector space now contains four spaces,
\begin{eqnarray}
0 \rightarrow & X_2 & \rightarrow X_1 \rightarrow X_0 \rightarrow X_{-1} \nonumber \\ & c & \quad \, \, \, \chi \, \, \, \quad \quad \zeta \, \, \, \quad \, \, \, \Psi
\end{eqnarray}
where $\Psi=(K,\bar K , f)$ and $\ell_{n}\Psi^{n}=0$ with $n\geq1$ since there is no dynamics at this point. From the symmetry transformations we read the following products,
\begin{eqnarray}
\ell_{1}(\xi)_{t} & = & 2(\partial_{\underline M} \xi_{\overline N} - \partial_{\overline N} \xi_{\underline M}) \, , \\
\ell_{1}(\xi)_{s} & = & - \frac12 \partial_{P} \xi^{P} \, , \\
\ell_{2}(\xi,\bar K)_{\overline v} & = & \widehat{\mathcal L}_{\xi} \bar K_{\overline A} \, , \\
\ell_{2}(\Gamma,\bar K)_{\overline v} & = & \delta_{\Gamma} \bar K_{\overline A} \, , \\
\ell_{2}(\xi, K)_{\underline v} & = & \widehat{\mathcal L}_{\xi} K_{\underline A} \, , \\
\ell_{2}(\Gamma, K)_{\underline v} & = & \delta_{\Gamma} K_{\underline A} \, , \\
\ell_{2}(\xi,f)_{s} & = & \xi^{P}\partial_{P}f \, .
\end{eqnarray}
The $L_{\infty}$ relations can be probed considering the previous list and the one from the previous section. The relations $n=1$ and $n=2$ are trivial. The relation $n=3$ is not trivial for the case $x_1=\Psi$, $x_2=\zeta_2$, $x_3=\zeta_3$,
\begin{eqnarray}
0 & = & \ell_2(\ell_2(\Psi,\zeta_2),\zeta_3) + \ell_2(\ell_2(\zeta_3,\Psi,),\zeta_2)
+\ell_2(\ell_2(\zeta_2,\zeta_3),\Psi)
\ . \nonumber
\end{eqnarray}
The previous expression can be rewritten in the following form,
\begin{eqnarray}
[\delta_{\zeta_2},\delta_{\zeta_3}]\Psi = \delta_{\zeta_{23}} \Psi
\end{eqnarray}
and therefore it is satisfied using the closure condition for the deformed generalized diffeomorphisms and double Lorentz transformations. Relations with $n \geq 4$ are trivial.
\subsubsection{Pertubative GDFT as an exact $L_{4}$ algebra}
Finally we extend the $L_{3}$ algebra describing the $C_{f}$ and the double Lorentz brackets algebra to include the dynamics. The graded vector space now contains five spaces,
\begin{eqnarray}
0 \rightarrow & X_2 & \rightarrow X_1 \rightarrow X_0 \rightarrow X_{-1} \rightarrow X_{-2} \nonumber \\ & c & \quad \, \, \, \chi \, \, \, \quad \quad \zeta \, \, \, \quad \, \, \, \Psi \quad \quad \, \, {\cal F}
\end{eqnarray}
where the perturbative equations of motion are related to the equations of the generalized dilaton and the generalized metric ${\cal F}=(\cal R, \cal R_{\overline A \underline B})$, but considering the GKSA and the linear perturbation for the generalized dilaton. From (\ref{eoml}) we have
\begin{eqnarray}
\ell_{1}(f)_s & = & - 4 \kappa E^{\underline A}\left(E_{\underline{A}}f\right) \\ \ell_{2}(f,f)_s & = & - 8 \kappa^2 E_{\underline{A}}f E^{\underline{A}}f \\ \ell_{2}(K,\bar K)_s & = & 2 \kappa E^{\underline A}\left(K_{\underline A} E_{\overline B}\bar K^{\overline B} + \bar K^{\overline B} E_{\overline B} K_{\underline A}\right) - 4 \kappa f_{\overline A \underline B \underline C} f^{\underline D \underline B \underline C} \bar K^{\overline A} K_{\underline D} \\ \ell_{4}(K,K,\bar K,\bar K)_s & = &
- 6 \kappa^2 \bar K_{\overline B} \bar K_{\overline C} \left[ (E^{\overline C} K^{\underline A}) E^{\overline B} K_{\underline A} - 2 f^{\overline B}{}_{\underline B \underline C} K_{{\underline A}} f^{\overline C \underline A \underline C} K^{{\underline B}} \right]
\end{eqnarray}
where we use the compact notation $f_{A B C} = E^{M}{}_{A} E^{N}{}_{B} E^{P}{}_{C} f_{M N P}$. The previous contributions come from the GDFT Lagrangian up to a cosmological term that requires a field redefinition. Analogously, from the generalized flat Ricci scalar we read
\begin{eqnarray}
\ell_{1}(f)_t & = & 2\kappa \left[ \sqrt{2} f_{\overline A \underline B \underline C} E^{\underline{C}}f - E_{\overline A}[E_{\underline{B}}f] \right] \\ \ell_{2}(K,\bar K)_t & = & \kappa\left[ 2 f^{\overline D}{}_{\underline B \underline C} K^{\underline C} \bar K_{\overline E} f^{\overline E}{}_{\overline D \overline A} -2 f^{\underline C}{}_{\overline D \overline A} \bar K^{\overline D} K_{\underline D} f^{\underline D}{}_{\underline B \underline C} \right. \nonumber \\ &&
- 2 \sqrt{2} f^{\underline C}{}_{\overline D \overline A} \left(- \frac12 K_{\underline{C}}D_{\underline{B}}\bar K^{\overline{D}} - \frac12 \bar K^{\overline{D}}E_{\underline{B}}K_{\underline{C}} + \frac12 K_{\underline{B}}D_{\underline{C}}\bar K^{\overline{D}} \right. \nonumber \\&& \left. + \frac12 \bar K^{\overline{D}}E_{\underline{C}}K_{\underline{B}} \right) - \sqrt{2} \left((E_{\bar B}\bar K^{\overline B})K^{\underline C} + (E_{\overline B} K^{\underline C}) \bar K^{\overline B} \right) f_{\overline A \underline B \underline C} \nonumber \\ && +2 E^{\underline C}\left[ - K_{[\underline{C}}D_{\underline{B}]}\bar K_{\overline{A}} - \bar K_{\overline{A}}E_{[\underline{B}}K_{\underline{C}]} + \frac{1}{\sqrt{2}} \bar K_{\overline A} K_{\underline D} f^{\underline D}{}_{\underline B \underline C} \right] \nonumber \\
&&
+ E_{\overline A}\left[(E_{\bar C}\bar K^{\overline C})K_{\underline B} + (E_{\overline C} K_{\underline B}) \bar K^{\overline C} \right]
\nonumber \\&& \left. - 2 \sqrt{2} f^{\overline D}{}_{\underline B \underline C}\left(\bar K{}_{[\overline{A}}D{}_{\overline{D}]}K^{\underline{C}} + K^{\underline{C}}E_{[\overline{D}}\bar K_{\overline{A}]} \right) \right] \\ \ell_{3}(f,K,\bar K)_t & = & \kappa^2 \left[-\frac{3}{2} \left(4E^{\underline{C}}f\right) \left( K_{\underline{B}}D_{\underline{C}}\overline{K}_{\overline{A}} \right) + 6 K_{\underline C}\bar{K}_{\overline A} E^{\underline C}\left[E_{\underline{B}}f\right] \right. \nonumber \\ &&
- 6 \left( E^{\underline{C}}f\right) K_{\underline{B}}D_{\underline{C}}\overline{K}_{\overline{A}} + 6 \left( E^{\underline{C}}f\right) \overline{K}_{\overline{A}}E_{\underline{B}}K_{\underline{C}} \nonumber \\ && \left.
- 6 \left( E^{\underline{C}}f\right) \overline{K}_{\overline{A}}E_{\underline{C}}K_{\underline{B}} - 6 \sqrt{2} \left( E^{\underline{C}}f\right) \bar K_{\overline A} K_{\underline D} f^{\underline D}{}_{\underline B \underline C} \right] \\ \ell_{4}(K,K, \bar K,\bar K)_t & = & 6 \kappa^2 \left[ \left( (E_{\overline B} K^{\underline C}) \bar K^{\overline B} \right) \left( K_{\underline{B}}D_{\underline{C}}\overline{K}_{\overline{A}} \right) - K_{\underline C}\bar{K}_{\overline A} E^{\underline C}\left[(E_{\bar C}\bar K^{\overline C})K_{\underline B} \right. \right. \nonumber \\ && \left. + (E_{\overline C} K_{\underline B}) \bar K^{\overline C} \right] - K^{\underline C} \bar K_{\overline D} \bar K_{\overline{A}}E^{\overline D}\left[E_{\underline{B}}K_{\underline{C}}\right]
+ K^{\underline C} \bar K_{\overline D} K_{\underline{B}}E^{\overline D}\left[D_{\underline{C}}\bar K_{\overline{A}}\right] \nonumber \\ &&
+ K^{\underline C} \bar K_{\overline D} E^{\overline D} \left[\bar K_{\overline{A}}E_{\underline{C}}K_{\underline{B}} \right]
+ \sqrt{2} K^{\underline C} \bar K_{\overline D} K_{\overline A} E^{\overline D} \bar K_{\underline D} f^{\underline D}{}_{\underline B \underline C} \nonumber \\ &&
+ K_{\underline{B}}D_{\underline{C}}\bar K^{\overline{D}} \bar K{}_{\overline{A}}D{}_{\overline{D}}K^{\underline{C}} + \bar K^{\overline{D}}E_{\underline{C}}K_{\underline{B}} K^{\underline{C}}E_{\overline{D}}\bar K_{\overline{A}} \nonumber \\
&&
+ \left( (E_{\overline B} K^{\underline C}) \bar K^{\overline B} \right) K_{\underline{B}}D_{\underline{C}}\overline{K}_{\overline{A}} - \left((E_{\overline B} K^{\underline C}) \bar K^{\overline B} \right) \overline{K}_{\overline{A}}E_{\underline{B}}K_{\underline{C}} \nonumber \\ &&
+ \left((E_{\bar B}\bar K^{\overline B})K^{\underline C} + (E_{\overline B} K^{\underline C}) \bar K^{\overline B} \right) \overline{K}_{\overline{A}}E_{\underline{C}}K_{\underline{B}} \nonumber \\ && \left. + \sqrt{2} \left(\bar K^{\overline B} E_{\overline B} K^{\underline C} \right) \bar K_{\overline A} K_{\underline D} f^{\underline D}{}_{\underline B \underline C} \, \right].
\end{eqnarray}
At this point we include products related to the gauge transformation of the equations of motion, $\ell_{2}(\xi,{\cal R})_s=\delta_{\xi}{\cal R}$, and $\ell_{2}(\xi,{\cal R}_{\overline A \underline B})_t=\delta_{\xi}{\cal R}_{\overline A \underline B}$, $\ell_{2}(\Lambda,{\cal R}_{\overline A \underline B})_t=\delta_{\Lambda}{\cal R}_{\overline A \underline B}$, where the last contribution can be easily computed considering that each index of ${\cal R}_{\overline A \underline B}$ transforms as a projected double Lorentz vector. The products related to the transformation of the equations of motion are required to check the $n=2$ relation. In this context, the absence of a $\ell_{3}(\zeta_{1},\zeta_2,{\cal F})$ implies that the closure of the gauge algebra holds off-shell. The remaining products are also null as can be easily verified.
\section{Applications}
\label{App}
\subsection{Fundamental charged heterotic string}
The most simple theory that lies inside the family of low energy effective field theories that can be reproduced with the GKSA is the fundamental charged heterotic string solution in $D=10$ \cite{Sen}. The duality approach can be easily constructed considering the following parametrization,
\begin{eqnarray}
{H}_{M N} = \left(\begin{matrix} g_{o}^{\mu \nu} & - g_{o}^{\mu \rho} C_{o\rho \nu} & - g_{o}^{\mu \rho} A_{o \rho i} \\
- g_{o}^{\nu \rho} C_{o\rho \mu} & g_{o\mu \nu} + C_{o\rho \mu} C_{o\sigma \nu} g_{o}^{\rho \sigma} + A_{o\mu}{}^i \kappa_{ij} A_{o\nu}{}^j &
C_{o\rho \mu} g_{o}^{\rho \sigma} A_{o\sigma i} + A_{o\mu}{}^j \kappa_{ji} \\
- g_{o}^{\nu \rho} A_{o\rho i} & C_{o\rho \nu} g_{o}^{\rho \sigma} A_{o\sigma i} + A_{o\nu}{}^j \kappa_{ij} & \kappa_{ij} + A_{o\rho i} g_{o}^{\rho \sigma} A_{o\sigma j}\end{matrix}\right) \ ,
\label{Gmetric}
\end{eqnarray}
with $\kappa_{ij}$ a Cartan-Killing metric. Since the generalized structure constants force us to describe the theory with the generalized frame/flux formalism, compatibility with the ansatz forces that the gauge field remains unperturbed as in \cite{KL}. Similarly, the generalized null vectors $K_{M}$ and ${\bar K}_{M}$ can be parametrized in terms of a pair of null vectors $l$ and $\bar l$ in the following way,
\begin{eqnarray}
K_{M} = \, \frac{1}{\sqrt{2}} \left( \begin{matrix} l^{\mu} \\ - l_{\mu} - C_{o \rho \mu} l^{\rho } \\ - A_{oi \rho} {l}^{\rho} \end{matrix} \right) \, , \quad
\bar{K}_{M} = \, \frac{1}{\sqrt{2}} \left( \begin{matrix} {\bar l}^{\mu} \\ {\bar l}_{\mu} - C_{o \rho \mu} {\bar l}^{\rho} \\- A_{oi \rho} {\bar l}^{\rho} \end{matrix} \right) \, ,
\end{eqnarray}
where $C_{o\mu \nu}=b_{o\mu \nu} + \frac12 A_{o\mu}{}^{i} A_{o\nu i}$ and the null vectors are constraint by (\ref{flatgeo}). The parametrization of the dilaton is $
e^{-2d_{o}} = \sqrt{g_{o}} e^{-2 \phi}$. In this case the perturbed solution is given by
\begin{eqnarray}
ds^{2} = \frac{1}{1+NH(r)}(-dt^{2}+ (dx^{9})^2) + \frac{q^2 H(r)}{4N(1+NH(r))^2}(dt+dx^{9})^2 + \sum_{i=1}^{8} dx^{i} dx^{i} \, ,
\end{eqnarray}
with $H(r)$ a Green function and $N$ a constant. The non-vanishing components of the two form and gauge field are
\begin{eqnarray}
b_{9t} & = &\frac{NH(r)}{1+NH(r)} \,, \\
A^1_{0} & = & A^{1}_{9} = \frac{qH(r)}{1+NH(r)} \, ,
\end{eqnarray}
with $q$ a charge and $\phi=-\frac12 \textrm{ln}(1+N H(r))$.
At the level of the symmetry transformations the algebraic structure of the duality covariant approach of this theory is an $L_{3}$ algebra, given by the transformations of $K_{M}$, $\bar K_{M}$ and $f$ under generalized diffeomorphisms and Double Lorentz transformations. While the former encodes ordinary diffeomorphisms and abelian/non abelian gauge transformations for $b_{0 \mu \nu}/A_{o\mu i}$, the latter transforms the flat version of the null vectors with a 10-dimensional Lorentz parameter $\Lambda_{ab}$ such that,
\begin{equation}
\delta_{\Lambda} l_{a} = \Lambda_{a}{}^{b} l_{b} \, , \quad \delta_{\Lambda} \bar l_{a} = \Lambda_{a}{}^{b} \bar l_{b} \, ,
\end{equation}
where $l_{a} = e_{o}^{\mu}{}_{a} l_{\mu}$ and $\bar l_{a} = e_{o}^{\mu}{}_{a} \bar l_{\mu}$. The full perturbative GDFT for this solution can be cast in an $L_{4}$ algebra with $\kappa^2$ contributions in the equations of motion as we showed in the previous section.
\subsection{(Bosonic) Enhanced DFT}
In compactifications of the bosonic string on k-dimensional tori,
the $U(1)_{L} \times U(1)_R$ symmetry of the Kaluza-Klein reduction gets enhanced at special points in moduli space and there are new massless scalars transforming in the adjoint representation of the enhanced symmetry groups. The new gauge group corresponds to $G_{L} \times G_{R}$ where $G_{L}=G_{R}$, and reduces to the standard $U(1)_{L} \times U(1)_R$ outside the points that provide the enhancement. As we mentioned, DFT incorpores T-duality
as a global symmetry group and therefore it is expected that there exists a formulation that captures these new states in a duality covariant way \cite{Enh}. When $G_{L}\times G_{R}$ has non-simple roots, it was shown in \cite{Mariana} that the C-bracket can be deformed in a way that preserves the duality covariance. The deformation accounts for
the cocycle factors that are necessary in the vertex representation of the current algebra. In terms of the generalized Lie derivative this deformation is
\begin{eqnarray}
({\cal \widehat L}_{E_A}E_B)^M= ({\cal L}_{E_A}E_B)^M + \widehat \Omega_{AB}{}^C E^{M}{}_{C} \, ,
\end{eqnarray}
where $ \widehat\Omega_{ABC}$ vanishes if one or more indices correspond to Cartan generators and if $A,B,C$\footnote{We use the same notation, but these indices $A,B,C \dots$ must be thought as double internal index of a generalized parallelizable manifold.} are associated with roots of the enhancement algebra, say $\alpha,\beta,\gamma$, respectively,
\begin{eqnarray}
\widehat\Omega_{ABC}=\left\{\begin{matrix} (-1)^{\alpha * \beta}\;\delta_{\alpha+\beta+\gamma} &\ \ {\rm if \ two \ roots \ are \ positive,}\\
-(-1)^{\alpha * \beta}\;\delta_{\alpha+\beta+\gamma} &\ \ {\rm if \ two \ roots \ are \ negative.} \\
\end{matrix}\right.
\label{deform}
\end{eqnarray}
The tensor $\widehat\Omega_{ABC}$ satisfies
\begin{eqnarray}
\widehat\Omega_{ABC}=\widehat\Omega_{[ABC]}\, , \qquad \widehat\Omega_{[AB}{}^D\widehat\Omega_{C]D}{}^E=0\, , \qquad \widehat\Omega_{ABC}\partial^C\cdots =0\, ,
\end{eqnarray}
and therefore (\ref{deform}) can be easily identified with the generalized structure constants $f_{ABC}$ upon trivially extended $O(D,D+n)\rightarrow O(D+n,D+n)$ in (\ref{glie2}). In this sense, the algebraic structure of enhanced DFT can be cast in the $L_{3}$ framework at the level of $C_{\Omega}$-bracket algebra according to the results of this work.
\section{Summary}
\label{Con}
In this work we show that GDFT can be cast in an $L_{\infty}$ structure. The presence of a deformed generalized Lie derivative and the double Lorentz transformation enrich the algebraic structure including non-trivial products to the well known $L_{\infty}$ structure of DFT. The frame formalism is needed to compute the generalized fluxes, which are deformed with a generalized version of the structure constants. At the level of the symmetry transformations the algebraic structure of GDFT is given by an $L_{3}$ algebra. We also show that the study of the dynamics can be performed in a closed form considering a GKSA for the generalized frame and a linear perturbation for the generalized dilaton. When dynamics is taken into account, the structure is promoted to an $L_{4}$ algebra with $\kappa^2$ corrections. The present computation has direct implications in the low energy effective action principle of the fundamental heterotic charged string, and bosonic string compactified on an specific internal k-dimensional torus with enhanced gauge symmetry. The latter is described by a enhanced DFT, which can be understood as a particular case of GDFT and therefore, at the level of the deformed bracket, the algebraic structure is an $L_{3}$ algebra.
\subsection*{Acknowledgements}
We thank D. Marqués for useful discussions. M.M. thanks the Max-Planck institute for kind hospitality during the final stages of this project. This work is partially supported by CONICET grant PIP-11220110100005 and PICT-2016-1358.
|
\section{Introduction}
Compartmental models for disease propagation have a long and illustrious history
\cite{AM92,KM27}, and they
remain a fundamental
predictive tool
\cite{EE2020,GBBCFPC20}.
For a stochastic, individual-level model, it has been suggested recently that
hyperedge information
should be incorporated \cite{APM20,simplicialSocialContagion,heterogeneityHypergraph}.
Hyperedges allow us to
account directly for
group interactions of any size,
rather than, as in, for example,
\cite{epidemicsSpread,virusSpreadInNetworks,spreadingEigenvalue},
treating them as
a collection of essentially independent
pairwise encounters.
In this work we
contribute to the
modeling and analysis
of disease spreading on a hypergraph.
We include the case where the number of infected nodes in a hyperedge contributes nonlinearly to the overall infection rate; this covers the
so-called \emph{collective contagion model}
setting and a new alternative
that we call a
\emph{collective suppression model}.
The main contributions of our work are
\begin{itemize}
\item a mean field approximation (\ref{dynamical system})--(\ref{eq:gidef}) with a spectral condition for local asymptotic stability of the zero-infection state
(Theorem~\ref{thm: local asymptotic stability})
and
an extension to global asymptotic stability
(Theorem~\ref{thm: global asymptotic condition})
when the nonlinear infection function is concave,
\item for the exact, individual-level model, a spectral condition for exponential decay
of the non-extinction probability
in the concave case (Theorem~\ref{thm: exponential decay}) and a spectral bound on the expected time to extinction (Corollary~\ref{cor:et}),
\item extensions of these results to more general
partitioned hypergraph models, where distinct infection rates apply to different categories of hyperedge (\ref{eq:partition}),
\item results for the non-concave collective
contagion model (Theorem~\ref{thm: collective contagion model} and
Theorem~\ref{thm: max collective contagion model}),
\item a complementary condition that rules out extinction of the disease (Theorem~\ref{thm: non-zero steady-state}),
\item interpretations of these mathematical results: the spectral
thresholds for disease extinction naturally distinguish between
the
inherent biological infectiousness of the disease and behavioural choices of the individuals in the population, allowing us to account for intervention strategies
(Section~\ref{sec:disc}).
\end{itemize}
The manuscript is organized as follows.
In section~\ref{sec:SIS}
we introduce the traditional
graph-based susceptible-infected-susceptible
(SIS) model and quote a spectral condition
that
characterizes
control of the disease.
We then
discuss the generalization to hyperedges, and
motivate the
use of infection rates that
do not scale linearly with
respect to the number of infected neighbors.
Section~\ref{sec:hyp_model}
formalizes the hypergraph
model and
shows how it may be simulated.
In section~\ref{sec:mean_field}
we derive a mean field approximation
to characterize the behaviour of the model, and
in section~\ref{sec:comp} we
give computational results
to illustrate its relevance.
Section~\ref{sec:stab}
analyses the deterministic mean field setting and
gives a spectral condition for
long-term decay of the disease.
The result is local for general infection rates and global (independent of the initial condition)
for the concave case.
This spectral condition generalizes a well-known result concerning disease
propagation on a graph.
Section~\ref{sec:more_sim} provides further computational simulations to illustrate the spectral threshold.
The full stochastic model is then studied in
section~\ref{sec:exact_model},
where we extend
the analysis in \cite{epidemicsSpread} to our hypergraph setting.
Here we study extinction of the disease
in the case where
the nonlinearity in the infection rate is concave.
We also derive conditions for non-extinction.
In section~\ref{sec:cim}
we extend
our analysis to the
so-called
collective contagion model
proposed in
\cite{APM20,simplicialSocialContagion,heterogeneityHypergraph}.
In section~\ref{sec:disc}
we summarize and interpret our results, and
discuss follow-on work.
For a review of recent studies of spreading processes on hypergraphs, including
the dissemination of rumours, opinions and knowledge,
we recommend \cite[subsection~7.1.2]{BCILLPYP21}.
The model that we study fits into the framework of
\cite{SISonHypergraphs}. This work introduced the idea
of a nonlinear ``infection pressure'' from each hyperedge, and derived a mean field approximation that was compared with microscale-level simulation results.
In \cite{simplicialSocialContagion}, the authors studied
this type of model on simplicial complexes of degree up to two (a subclass of the more general hypergraph setting) and also studied a mean field approximation.
These authors examined the mean field system from a dynamical systems perspective and analysed issues such
as bistability, hysteresis and discontinuous transitions.
Similarly, in \cite{APM20,heterogeneityHypergraph}, a hypergraph version was considered. Our work differs from these studies in
(a) focusing on the derivation of spectral thresholds
for extinction of a disease in both the exact and mean field settings and (b)
seeking to interpret the results from a
mathematical modelling perspective.
We mention that it would also be of interest to
develop corresponding thresholds for the
mean field models in
\cite{SISonHypergraphs,APM20,simplicialSocialContagion}.
\section{Stochastic SIS models}
\label{sec:SIS}
\subsection{Stochastic SIS model on a graph}
\label{subsec:graph}
Classical ODE compartmental models are based on the assumption that
any pair of individuals is equally likely to interact---this
is the homogeneous mixing case \cite{AM92}.
If, instead, we have knowledge of all
possible pairwise interactions between individuals,
then this information may be
incorporated via a contact graph and used in a stochastic model.
Here, each node represents an individual, and
an edge between nodes $i$ and $j$ indicates that
individuals $i$ and $j$ interact.
For a population with $n$ individuals,
we may let
$A \in \mathbb{R}^{n \times n}$ denote the
corresponding symmetric adjacency matrix, so nodes $i$ and $j$ interact if and only
$A_{ij} = 1$.
In this setting, a stochastic SIS model
uses the two-state random variable
$X_i(t)$
to represent the §
status of node $i$ at time $t$, with $X_i(t) = 0$
for a susceptible node and
$X_i(t) = 1$
for an infected node. Each
$X_i(t)$
then
follows
a continuous time Markov process where
the infection rate is given by
\begin{equation}
\beta \sum_{j=1}^n A_{ij} X_j(t)
\label{eq:graph_inf}
\end{equation}
and the recovery rate is $\delta$. Here,
$\beta > 0$ and $\delta > 0$ are parameters governing the strength of the two effects.
In this model, we see from (\ref{eq:graph_inf}) that the
current
chance of infection increases linearly in proportion to the
current number of infected neighbors.
This model was studied in
\cite[Theorem~1]{spreadingEigenvalue}, where it was argued that the
condition
\begin{equation}
\lambda(A) \frac{\beta}{\delta} <1
\label{eq:graph_epi}
\end{equation}
guarantees the disease will die out.
Here, $\lambda(A)$ denotes the largest eigenvalue of the symmetric matrix $A$.
Further justification for this result may be found, for example,
in
\cite{epidemicsSpread,virusSpreadInNetworks}.
We note that (\ref{eq:graph_epi}) gives an elegant generalization of the homogeneous mixing case (where $A$ corresponds to the complete graph).
\subsection{Why use a hypergraph?}
\label{subsec:hypergraph}
It has been argued
\cite{ABAMPL21,BCILLPYP21,BASJK18,benson2016higher,ER06}
that
in many network science applications
we lose information by
recording only pairwise interactions.
For example, emails can be sent to groups of recipients,
scholarly articles may have multiple coauthors,
and
many proteins may interact to form a complex.
In such cases, recording the relevant lists of interacting
nodes gives a more informative picture than reducing
these down to a collection of edges.
In the setting of an SIS model, we may argue that
individuals typically come together in
well-defined groups, for example, in a household,
a workplace or a social setting.
Such groups may be handled by the use of hyperedges,
leading to a hypergraph; these
concepts are formalized in the next section.
With a classic graph model, as described in section~\ref{subsec:graph},
the rate of infection of a node is linearly proportional to the number of infectious neighbors.
With a hypergraph we may consider more intricate contagion mechanisms.
For example,
using the terminology of
\cite{heterogeneityHypergraph},
the \textit{collective contagion} model
is used in \cite{APM20,simplicialSocialContagion,heterogeneityHypergraph}). Here, infection only starts spreading within a hyperedge after a certain threshold number of infectious neighbors has been reached.
This type of behaviour is relevant, for example, in an office environment. A small number of workers may be able to
socially distance in way that effectively eliminates the
risk of infection.
However,
if the number of
individuals (size of the hyperedge) is too large, then
the disease may spread.
We mention here that an alternative type of mechanism may also
operate, which we call \textit{collective suppression}.
Imagine that a disease may be contracted through
contact with a surface that was previously touched by an infected individual.
Now suppose that
a group of individuals is likely to use the
same physical object, such as a door handle, hand rail,
cash machine, or water cooler.
If an infected individual contaminates the object, then further
contamination by other individuals is less relevant.
In this case,
doubling the number of common users will increase the
risk of infection by a factor less than two; generally
risk grows sublinearly as a function of the size of the hyperedge.
These arguments motivate us to study the case where
the rate of infection of a node within a hyperdege is dependent on a generic function $f$ of the number of infectious neighbors in a hyperedge; this approach was also taken in \cite{SISonHypergraphs}.
We will be particularly concerned with the
case where $f$ is concave, since this
is tractable for analysis and allows us to
draw conclusions about the
collective contagion model.
We note that if $f$ is the identity, then we recover linear dependence on the number of infectious neighbors and the hypergraph model is equivalent to a virus spreading on the clique graph of the hypergraph.
\section{SIS on a Hypergraph}
\label{sec:hyp_model}
\subsection{Background}
We continue with some standard definitions
\cite{bretto2013hypergraph}.
\begin{defn}
A \emph{hypergraph} is a tuple $\mathcal H:=(V,E)$ of \emph{nodes} $V$ and \emph{hyperedges} $E$ such that $E\subset {\mathcal P}(V)$.
Here, ${\mathcal P}(V)$ denotes the power set of $V$.
\end{defn}
We will let $n$ and $m$ denote the number of nodes
and hyperedges, respectively; that is,
$|V| =n$ and $|E| = m$. Loosely, a hypergraph generalizes
the concept of a graph by allowing an ``edge'' to be a list of more than two nodes.
\begin{defn}
Consider a hypergraph $\mathcal H:=(V,E)$.
The \emph{incidence matrix},
${\mathcal I}$,
is the $n\times m$ matrix such that
$\mathcal I_{i h}=1$ if node $i$ belongs to hyperedge $h$ and $\mathcal I_{i h}=0$ otherwise.
\end{defn}
It is also useful to introduce
$W:={\mathcal I} {\mathcal I}^T$.
This $n\times n$ matrix has the property that
$W_{i j}$ records the number of hyperedges containing both nodes $i$ and $j$.
In particular, if $\mathcal H$ is a graph then $W$ is the affinity matrix of the graph.
\subsection{General infection model}
\label{subsec:model}
In our context, the nodes represent individuals and
a
hyperedge records a collection of individuals who
are known to interact as a group.
As in the graph case introduced in subsection~\ref{subsec:graph},
we use a state vector $X(t)$
which follows a continuous time Markov process,
where, for each $1 \le i \le n$,
$X_i(t) = 1$ if node $i$ is infected at time
$t$ and
$X_i(t) = 0$ otherwise.
We continue to assume that
an infectious node becomes susceptible with
constant recovery rate $\delta>0$.
However,
generalizing
(\ref{eq:graph_inf}), we now assume
that a susceptible node $i$ becomes infectious with
rate
\begin{equation}\label{infection rate random}
\beta\sum_{h\in E}\mathcal I_{i h}f\big(\sum_{j = 1}^{n} X_j(t)\mathcal I_{j h}\big),
\end{equation}
where $\beta>0$ is a constant.
In (\ref{infection rate random}),
$f:\mathbb{R}_+\to \mathbb{R}_+$
specifies the
manner in which
the contribution to the overall level of
infectiousness
from each hyperedge involving node $i$
is assumed to increase
in proportion to the number of
infected nodes in that hyperedge.
Throughout our analysis we will always assume that
$f(0)=0$ and
$f$ is $C^1$ in a neighborhood of $0$.
If $f$ is the identity, the rate of infection reduces to $\beta \sum_{j=1}^{n} W_{i j}X_j(t)$. This gives a weighted version of the infection rate of an SIS model on a graph.
As discussed in subsection~\ref{subsec:hypergraph},
it may be appropriate to choose
nonlinear $f$ in certain circumstances.
We note that in \cite{SISonHypergraphs} the authors have in mind functions which behave like the identity near the origin and have a horizontal asymptote. Instances of such functions are
$x\mapsto \mathrm{arctan}(x)$ and $x\mapsto \min\{x,c\}$ for some $c>0$. Relaxing these conditions, we may ask more generally in such a
setting that the function be concave. On the other hand, the authors in \cite{APM20,simplicialSocialContagion} consider a \emph{collective contagion model}, where infection spreads within a hyperdge only if a certain threshold of infectious vertices is reached in that hyperedge.
A collective contagion model may be represented via the function $x\mapsto c_2\mathbbm{1}(x\geq c_1)$ for some $c_1,c_2>0$, or $x\mapsto \max\{0,x-c\}$ for some $c>0$.
\subsection{Partitioned hypergraph model}
\label{subsec:partitoned}
We also introduce a more general case where we partition the hyperedges into $K$ disjoint categories with each category $1 \le k \le K$ having its own distinct rate of infection in response to the number infected nodes in a hyperedge, represented by a function $f_k$.
For example, the categories may correspond to different types of housing, workplaces, hospitality venues or sports facilities.
We may then represent the infection rate of node $i$ as
\begin{equation}\label{eq:partition}
\beta \sum_{k=1}^{K}\sum_{h\in E}{\mathcal I}^{(k)}_{i h}f_k\left(\sum_{j=1}^{n}X_j(t){\mathcal I}^{(k)}_{j h}\right),
\end{equation}
where we let ${\mathcal I}^{(k)}_{i h}=1$ if $i$ belongs to hyperedge $h$ in the category $k$ and ${\mathcal I}^{(k)}_{i h}=0$ otherwise; so
${\mathcal I}^{(k)}$ is the incidence matrix of the subhypergraph consisting of only the hyperedges from category $k$.
We will refer to this as a
\emph{partitioned hypergraph model}.
In this generalized case, a collective contagion model could be defined by first organizing the hyperedges into categories depending on their size, so that category $k$ is the set of hyperedges of size $k+1$. A collective contagion model may then represented, for example, via the functions $f_1:x\mapsto x$, and $f_k:x\mapsto c_{2,k}\mathbbm{1}(x\geq c_{1,k})$, $k\in \{2,\dots,K\}$.
\section{Mean Field Approximation}
\label{sec:mean_field}
A classic approach to studying processes such as
(\ref{infection rate random}), where
infection rates are random,
is to develop
a mean field approximation for the expected process $$\big(\mathbb{E}[X_i(t)]\big)_{t\geq 0}=\big(\mathbb P(X_i(t)=1)\big)_{t\geq 0}=:(p_i(t))_{t\geq 0},$$
with
deterministic rates.
In our case,
the rate of recovery $\delta$ is constant, so can remain unchanged. Let us express the rate of infection (\ref{infection rate random}) of a node solely in terms of the expected processes
$\{p_i(t)\}_{i=1}^{n}$. To do this we can substitute the $X_j(t)$ appearing in (\ref{infection rate random}) by their expected values $p_j(t)$. The approximate rate of infection for node $i$ then becomes
\begin{equation}\label{infection rate approximate}
\beta \sum_{h\in E}\mathcal I_{i h}f\big(\sum_{j=1}^{n} p_j(t)\mathcal I_{j h}\big).
\end{equation}
We arrive at the deterministic mean field ODE
\begin{equation}\label{dynamical system}
\frac{d P(t)}{dt}=g(P(t)),
\end{equation}
where $g_i:\mathbb{R}^n\to \mathbb{R}$ is defined by
\begin{equation}
g_i(P(t)):=\beta\sum_{h\in E}\mathcal I_{i h}f\left(\sum_{j=1}^{n}p_j(t)\mathcal I_{j h}\right)(1-p_i(t))-\delta p_i(t).
\label{eq:gidef}
\end{equation}
\section{Simulations and Comparison between Exact and Mean Field Models}
\label{sec:comp}
Let us emphasize that the approximate infection rates in (\ref{infection rate approximate}) differ
in general from the expectation of the random rates in (\ref{infection rate random}). When the function $f$ is concave, however, Jensen's reverse inequality indicates that the rates in (\ref{infection rate approximate}) are greater than the expectation of the rates in (\ref{infection rate random}). Hence, in this case the expected quantities $p_i(t)$ are overestimated by
(\ref{dynamical system})--(\ref{eq:gidef}).
This is fine since we are looking for conditions for the disease to vanish. If $f$ is not concave (e.g., for a collective contagion model), these expected quantities are underestimated and it is not clear a priori whether the exact model is well approximated by the mean field ODE.
In this section we therefore present
results of computational simulations in order to gain insight into the accuracy of our mean field
approximation.
\subsection{Simulation algorithm}
Before presenting numerical results, we
summarize our approach for simulating the
individual-level stochastic model, which is based
on a standard
time discretization;
see, for example, \cite{SISonHypergraphs}.
Using a small fixed time step $\Delta t$, we advance from time $t$ to $t+\Delta t$ as follows.
First, let $r\in [0,1]^n$ be a random vector
of i.i.d.\ values uniformly sampled from $[0,1]$. For every node $1 \le i \le n$,
\begin{itemize}
\item when $X_i(t)=0$, set $X_i(t+\Delta t)=1$ if
$$
r_i<1-\exp\left(-\beta \sum_h\mathcal I_{i h}f(\sum_j X_j(t)\mathcal I_{j h})\Delta t\right),
$$
and set $X_i(t+\Delta t)=0$ otherwise;
\item when $X_i(t)=1$, set $X_i(t+\Delta t)=0$ if
$$
r_i<1-\exp\left(-\delta\Delta t\right),
$$
and set $X_i(t+\Delta t)=1$ otherwise.
\end{itemize}
\subsection{Computational results}
\label{subsec:comp}
In the simulations we chose $n=400$ nodes with fixed
recovery rate $\delta =1$.
We look at results for different choices of
infection strength $\beta$ and $i_0$, the latter denoting the (independent) initial probability for each node to be infectious.
We simulated the mean field ODE using Euler's method
with time step $\Delta t=0.05$. The largest size of a hyperedge was $5$ and we distributed the number of hyperedges for the hypergraph randomly as follows: $300$ edges, $200$ hyperedges of size $3$, $100$ hyperedges of size $4$ and $50$ hyperedges of size $5$. To give a feel for the level of fluctuation, the individual-level paths are averaged over $10$ runs, each
with the same hypergraph connectivity and initial state.
Figures~\ref{fig: min},
\ref{fig: log} and
\ref{fig: atan}
show results for three concave choices of $f$; respectively,
\begin{itemize}
\item $f(x)=\min(x,3)$,
\item $f(x) = \log(1+x)$,
\item $f(x) = \mathrm{arctan}(x)$.
\end{itemize}
For
Figure~\ref{fig: max}
we used a collective contagion model on a partitioned
hypergraph. Assigning each hyperedge to a category in $\{1,2,3,4\}$, where category $k$ contains the hyperedges of size $k+1$, we chose the following associated functions to determine the infection rates: $f_1(x):=x$, and for $k\in \{2,3,4\}$, $f_k(x):=(k-1)\mathbbm{1}(x\geq k-1)$.
The four figures show the proportion of infectious individuals as a function of time.
In these simulations, and others not reported here,
we observe that the initial value $i_0$ does not affect the asymptotic behavior of the process: the process vanishes or converges to a non-zero equilibrium depending on the value of $\beta$ but regardless of the value of $i_0$.
In
Figures~\ref{fig: min},
\ref{fig: log} and
\ref{fig: atan}, where $f$ is concave, we know that
the mean field model gives an upper bound
on the expected proportion of infected individuals
in the microscale model. We also see
that the mean field model provides a reasonably sharp approximation.
Moreover, we see a similar level of sharpness in
Figure~\ref{fig: max} for the
collective contagion model, where
$f$ is not concave.
\begin{figure}[htp]
\centering
\includegraphics[width=\textwidth]{alt_f_min_n_400_K_5_T_40_dt_0.05.png}
\caption{Here, $f(x) = \min(x,3)$.
Red dashed line: mean field approximation from
(\ref{dynamical system})--(\ref{eq:gidef}).
Blue solid line:
proportion of infected individuals,
$\sum_{i} X_i(t)/n$, from the individual-level stochastic model
(\ref{infection rate random}), averaged over $10$ runs.
}
\label{fig: min}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=\textwidth]{alt_f_log_n_400_K_5_T_40_dt_0.05.png}
\caption{Here, $f(x) = \log(1+x)$.
Red dashed line: mean field approximation from
(\ref{dynamical system})--(\ref{eq:gidef}).
Blue solid line:
proportion of infected individuals,
$\sum_{i} X_i(t)/n$, from the individual-level stochastic model
(\ref{infection rate random}), averaged over $10$ runs.
}
\label{fig: log}
\end{figure}
\begin{figure}[htp]\label{fig: atan}
\centering
\includegraphics[width=\textwidth]{alt_f_atan_n_400_K_5_T_40_dt_0.05.png}
\caption{Here, $f(x) = \mathrm{arctan}(x)$.
Red dashed line: mean field approximation from
(\ref{dynamical system})--(\ref{eq:gidef}).
Blue solid line:
proportion of infected individuals,
$\sum_{i} X_i(t)/n$, from the individual-level stochastic model
(\ref{infection rate random}), averaged over $10$ runs.
}
\label{fig: atan}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=\textwidth]{alt_f_collective_n_400_K_5_T_40_dt_0.05.png}
\caption{Collective contagion model on a partitioned hypergraph.
Red dashed line: mean field approximation from
(\ref{dynamical system}) with (\ref{eq:mfpart}).
Blue solid line:
proportion of infected individuals,
$\sum_{i} X_i(t)/n$, from the individual-level stochastic model
(\ref{infection rate random}), averaged over $10$ runs.
}
\label{fig: max}
\end{figure}
A key advantage of the mean field approximation is that it gives rise to a deterministic autonomous dynamical system for which there exists a rich theory to study the asymptotic stability of equilibrium points.
This motivates the analysis in the next section.
\section{Stability Analysis}
\label{sec:stab}
We provide below spectral conditions which imply that
the infection-free solution $0 \in \mathbb{R}^n$ is a locally or globally asymptotically stable equilibrium of
(\ref{dynamical system})--(\ref{eq:gidef}).
We will find that local asymptotic stability can be shown with no structural assumptions on $f$.
We will also find that
global asymptotic stability
follows \emph{under the same conditions} when $f$ is concave.
Our conclusions fit into a framework that generalizes
the graph case (\ref{eq:graph_epi}):
the spectral threshold takes the form
\[
\lambda(W)\frac{c_f \beta}{\delta} <1
\]
for some constant $c_f>0$ depending only on the choice of $f$.
Throughout this work, to be concrete we
let $\| \cdot \|$ denote the Euclidean norm.
\subsection{Local asymptotic stability}
\begin{thm}\label{thm: local asymptotic stability}
If
\begin{equation}
\lambda(W)
\frac{ f'(0) \beta}{\delta} < 1
\label{eq:W1}
\end{equation}
then
$0 \in \mathbb{R}^n$ is a locally asymptotically stable equilibrium of (\ref{dynamical system})--(\ref{eq:gidef}); that is,
there exists a positive $\gamma$ such that
$||P(0)||<\gamma \Rightarrow\lim_{t\to\infty}||P(t)||=0$.
\end{thm}
\begin{proof}
We see that $g(0)=0$, so $0 \in \mathbb{R}^n$ is an equilibrium for (\ref{dynamical system}). It remains to show that this solution is locally asymptotically stable.
Appealing to a standard linearization
result~\cite{Ve90}, it suffices to show that every eigenvalue of the Jacobian matrix $\nabla g(0)$ has a negative real part.
We compute
\begin{align*}
\frac{\partial g_i}{\partial p_{j_0}}&=\begin{cases}
\beta\sum_h\mathcal I_{i h}\mathcal I_{j_0 h}f'(\sum_j p_j\mathcal I_{jh})(1-p_i),&\text{ }j_0\neq i,\\
\beta\sum_h\mathcal I_{i h}\mathcal I_{j_0 h}f'(\sum_j p_j\mathcal I_{jh})(1-p_i)-\beta\sum_{h}\mathcal I_{ih}f(\sum_j p_j\mathcal I_{j h})-\delta,&\text{ }j_0=i.
\end{cases}
\end{align*}
We see that $ \nabla g(0)= \beta f'(0)W-\delta I$.
This matrix is symmetric and therefore has real eigenvalues.
Hence,
it suffices that the
largest eigenvalue of
$\beta f'(0)W$ does not exceed
$\delta$, and the result follows.
\end{proof}
Theorem~\ref{thm: local asymptotic stability} extends to the partitioned
model in
(\ref{eq:partition}).
In this case $g_i(P(t))$ in the mean field ODE
(\ref{dynamical system}) is defined as
\begin{equation}
g_i(P(t)):=\beta\sum_{k=1}^{K}\sum_{h\in E}\mathcal I_{i h}^{(k)}f_k\big(\sum_{j=1}^{n}p_j(t)\mathcal I_{j h}^{(k)}\big)(1-p_i(t))-\delta p_i(t),
\label{eq:mfpart}
\end{equation}
and we let
$W^{(k)}:= {\mathcal I}^{(k)} {{\mathcal I}^{(k)}}^T$.
\begin{thm}\label{thm: generalized local asymptotic stability}
If
\begin{equation}
\lambda\left(\sum_{k=1}^{K} f_k'(0)W^{(k)}\right)\frac{\beta}{\delta} < 1
\label{eq:Wk}
\end{equation}
then
$0 \in \mathbb{R}^n$ is a locally asymptotically stable equilibrium
of (\ref{dynamical system}),
with $g_i$ defined in (\ref{eq:mfpart}).
\end{thm}
\begin{proof}
The proof of
Theorem~\ref{thm: local asymptotic stability}
extends straightforwardly.
We compute
\begin{align*}
\frac{\partial g_i}{\partial p_{j_0}}&=\begin{cases}
\beta\sum_{k}\sum_h\mathcal I^{(k)}_{i h}\mathcal I^{(k)}_{j_0 h}f_k'(\sum_j p_j\mathcal I^{(k)}_{jh})(1-p_i),&\text{ }j_0\neq i,\\
\beta\sum_{k}\sum_h\mathcal I^{(k)}_{i h}\mathcal I^{(k)}_{j_0 h}f_k'(\sum_j p_j\mathcal I^{(k)}_{jh})(1-p_i)-\beta\sum_{k}\sum_{h}\mathcal I^{(k)}_{ih}f_k(\sum_j p_j\mathcal I^{(k)}_{j h})-\delta,&\text{ }j_0=i,
\end{cases}\\
\Rightarrow\ &\frac{\partial g_i}{\partial p_{j_0}}|_{P=0}= \begin{cases}
\beta \sum_{k}W^{(k)}_{i j}f_k'(0),&\text{ }j_0\neq i,\\
\beta \sum_{k}W^{(k)}_{i j}f_k'(0)-\delta,&\text{ }j_0=i,
\end{cases}
\end{align*}
and note that
$$
\lambda(\beta \sum_{k}f'_k(0)W^{(k)}-\delta I)<0
\quad
\Leftrightarrow
\quad
\lambda(\sum_{k}f'_k(0)W^{(k)})<\frac{\delta}{\beta}.
$$
\end{proof}
\subsection{Global asymptotic stability for the concave infection model}
We now show that
when $f$ is concave
the
condition in
Theorem~\ref{thm: local asymptotic stability}
ensures global stability of the zero equilibrium, and hence guarantees that the disease dies out
according to the mean field approximation.
\begin{defn}
Given a matrix $A$, define its symmetric version to be
$$
A^{(S)}:=(A+A^T)/2.
$$
\end{defn}
\begin{lemma}\label{lemma: spectrum inequality}
Suppose that $A$ and $B$ are $n\times n$ real matrices, and suppose that there exists a diagonal matrix $\Lambda$ such that for all $i\in \{1,2,\dots,n\},\ \Lambda_{i i}\geq 0$, and
$$
A=B-\Lambda.
$$
Then the largest eigenvalues of $A$ and $B$ satisfy $\lambda(A)\leq \lambda (B)$, and the largest eigenvalues of $A^{(S)}$ and $B^{(S)}$ also satisfy $\lambda(A^{(S)})\leq \lambda (B^{(S)})$.
\end{lemma}
\begin{proof}
Let $x$ be a unit eigenvector associated with $\lambda(A^{(S)})$. We have
\begin{align*}
2\lambda(A^{(S)})&=x^TAx+x^TA^Tx\\
&=x^TBx+x^TB^Tx - 2x^T\Lambda x\\
&=\sum_{i,j=1}^n(b_{ij}+b_{ji})x_ix_j-2\sum_{i=1}^n\Lambda_{i i} x_i^2\\
&\leq x^T(B+B^T) x\\
&\leq \max\{x^T(B+B^T)x\ |\ x^Tx = 1\} = 2\lambda(B^{(S)}).
\end{align*}
The inequality $\lambda(A)\leq \lambda(B)$ may be shown similarly.
\end{proof}
\begin{thm}\label{thm: global asymptotic condition}
Suppose $f$ is concave.
If
\[
\lambda(W) \frac{f'(0) \beta}{\delta} < 1,
\]
then $0 \in \mathbb{R}^n$ is a globally asymptotically stable equilibrium of
(\ref{dynamical system})--(\ref{eq:gidef}); so
$\lim_{t\to\infty}||P(t)||=0$
for any valid initial condition (that is, with
$0 \le p(0)_i \le 1$).
\end{thm}
\begin{proof
From the global asymptotic stability result in \cite[Lemma $1'$ ]{GSAcriterion}
it is sufficient to show that
all eigenvalues of the symmetric matrix $(\nabla g(P))^{(S)}$
are strictly less than $0$, for all $P\neq 0$.
We have
\begin{align*}
\frac{\partial g_i}{\partial p_{j_0}}&=\begin{cases}
\beta\sum_h\mathcal I_{i h}\mathcal I_{j_0 h}f'(\sum_j p_j\mathcal I_{jh})(1-p_i),&\ j_0\neq i,\\
\beta\sum_h\mathcal I_{i h}\mathcal I_{j_0 h}f'(\sum_j p_j\mathcal I_{jh})(1-p_i)-\beta\sum_{h}\mathcal I_{ih}f(\sum_j p_j\mathcal I_{j h})-\delta,&\ j_0=i.\\
\end{cases}\\
\end{align*}
Letting $B$ denote the $n\times n$ matrix given by
\begin{align*}
B_{i j_0}=\begin{cases}
\beta\sum_h\mathcal I_{i h}\mathcal I_{j_0 h}f'(\sum_j p_j\mathcal I_{jh})(1-p_i),&\ j_0\neq i,\\
\beta\sum_h\mathcal I_{i h}\mathcal I_{j_0 h}f'(\sum_j p_j\mathcal I_{jh})(1-p_i)-\delta,&\ j_0=i,\\
\end{cases}\\
\end{align*}
we have $\nabla g(P)= B-\Lambda$, where $\Lambda$ is the $n\times n$ diagonal matrix where for all $i\in \{1,2,\dots\}$, $\Lambda_{i i}:=\beta\sum_h \mathcal I_{i h}f(\sum_j p_j \mathcal I_{j h})\geq 0$. On the one hand Lemma~\ref{lemma: spectrum inequality} now yields $\lambda((\nabla g(P))^{(S)})\leq \lambda(B^{(S)})$; on the other hand, note that $B+\delta I\leq \nabla g(0)+\delta I$, where we interpret the inequality in a componentwise sense, and where we use $f'(\sum_j p_j \mathcal I_{j h})\leq f'(0)$,
since $f$ is concave. Hence $B^{(S)}+\delta I\leq \nabla g(0)+\delta I$, and since $B^{(S)}+\delta I$ has only non-negative entries, appealing to the Perron-Frobenius theorem, we have $\lambda(B^{(S)})\leq \lambda(\nabla g(0))$.
Combining these inequalities and using the spectral condition in the statement of the theorem, we deduce that
\[
\lambda\left( (\nabla g(P))^{(S)} \right)
\leq
\lambda(\nabla g(0)) = \lambda \left( \beta f'(0)W-\delta I \right) < 0,
\]
as required.
\end{proof}
A straightforward
adaptation of the proof of Theorem~\ref{thm: global asymptotic condition} yields the following global asymptotic stability result for
the more general partitioned model.
\begin{thm}\label{thm: generalized global asymptotic stability}
Suppose all $f_k$ are concave.
If
\[
\lambda(\sum_{k = 1}^{K}f_k'(0)W^{(k)})
\frac{\beta}{\delta}
< 1,
\] then
$0 \in \mathbb{R}^n$ is a globally asymptotically stable
equilibrium of (\ref{dynamical system}), with
$g_i$ defined in
(\ref{eq:mfpart}).
\end{thm}
\section{Simulations to Test the Spectral Condition}
\label{sec:more_sim}
We now show the results of experiments that
test the sharpness of our spectral vanishing condition.
Here, we used
the concave functions $f(x)=2\log(1+x)$ (on the left of Figure~\ref{fig: both_i_infty vs beta} and in
Figure~\ref{fig: 2log i_t vs t}) and $f(x) = \mathrm{arctan}(x)$ (on the left in
Figure~\ref{fig: both_i_infty vs beta} and in
Figure~\ref{fig: atan i_t vs t}) to construct partitioned models
with $f_1(x)=x$ and $f_k(x)=f(x)$ for all $k\geq 2$. We fixed a hypergraph with $n = 400$ nodes, $400$ edges, $200$ hyperedges of size $3$, $100$ hyperedges of size $4$ and $50$ hyperedges of size $5$.
At time zero, each node
was infected with independent probability
$i_0=0.5$
and we used a recovery rate of
$\delta =1$.
In addition to the mean field ODE, we also simulated the microscale model, averaged over $5$ runs, using
the discretization scheme
described in Section~\ref{subsec:comp},
with $\Delta t=0.1$.
In Figure~\ref{fig: both_i_infty vs beta}
the
asterisks (red) show the corresponding
proportion of infected individuals
according to the
mean field model,
$\sum_i p_i(t)/n$,
at time $t = 150$,
for a range of different
$\beta$ between $0$ and $0.2$: $\beta\in \{(0.01)k\ |\ k\in \{0,\dots,20\}\}$.
The crosses (blue) show the
corresponding proportion of infected individuals
from the microscale model,
$\sum_i X_i(t)/n$,
at time $t=150$.
The vertical green line represents the critical value $\beta_c:=\delta/\lambda(\sum_{k=1}^{K}f_k'(0)W^{(k)})$ (we have $\beta_c\approx 0.0265$ on the left and $\beta_c\approx 0.0490$ on the right).
For the mean field model,
we know from
Theorem~\ref{thm: generalized global asymptotic stability} that
$\beta < \beta_c$
guarantees global stability of the
zero-infection state. We see that $\beta_c$ also lies close to the threshold beyond which
extinction
of the disease is lost in the mean field model.
For the individual-level stochastic model,
Theorem~\ref{thm: generalized exponential decay}
below
shows that
$\beta < \beta_c$ is also
sufficient for eventual extinction of the disease.
This is consistent with the results in Figure~\ref{fig: both_i_infty vs beta}.
\begin{figure}[htp]
\centering
\includegraphics[width=0.45\textwidth]{2log_i_infty_vs_beta_T_150.png}
\includegraphics[width=0.45\textwidth]{atan_i_infty_vs_beta_T_150_new.png}
\caption{Left: infection function based on $2\log(1+x)$.
Right: infection function based on $\mathrm{arctan}(x)$.
For different choices of infection strength
$\beta$ (horizontal axis), we show
the proportion of infection individuals at time
$ t = 150$ (vertical axis) for the mean
field approximation
(\ref{dynamical system}) with (\ref{eq:mfpart})
in red asterisks and
for the
individual-level stochastic model
(\ref{infection rate random})
in blue crosses.
The spectral bound arising from our analysis is show as a green vertical line.
}
\label{fig: both_i_infty vs beta}
\end{figure}
In the left of
Figures~\ref{fig: 2log i_t vs t}
and \ref{fig: atan i_t vs t}
we show individual
trajectories
of the
proportion of infected individuals,
$\sum_i p_i(t)/n$,
according to the
mean field model,
for a range of $\beta$ values.
For the same range of $\beta$ values, the plots on the right of these figures show the
corresponding proportion of infected individuals
from the microscale model,
$\sum_i X_i(t)/n$.
The curves are colored in red if the spectral vanishing condition $\beta < \beta_c$ is satisfied.
We see qualitative agreement between the
mean field and individual-level models, and
extinction for the $\beta$ values below the spectral threshold.
\begin{figure}[htp]
\centering
\includegraphics[width=15cm, height=7cm]{2log_beta_function_of_t_T_150_new.png}
\caption{Results with the arctan infection function.
Left: proportion of infected individuals using the
mean field model.
Right: proportion of infected individuals using the
individual-level model.
From bottom to top, the $\beta$ values used were
From bottom to top, the $\beta$ values used were
$\beta\in \{(0.02)k\ |\ k\in \{0,\dots,10\}\}$.
Cases where $\beta$ is below the spectral
bound are colored in red.
}
\label{fig: 2log i_t vs t}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=15cm, height=7cm]{atan_beta_function_of_t_T_150_new.png}
\caption{Results with the arctan infection function.
Left: proportion of infected individuals using the
mean field model.
Right: proportion of infected individuals using the
individual-level model.
From bottom to top, the $\beta$ values used were
$\beta\in \{(0.02)k\ |\ k\in \{0,\dots,10\}\}$.
Cases where $\beta$ is below the spectral
bound are colored in red.
}
\label{fig: atan i_t vs t}
\end{figure}
Having derived and tested spectral conditions
that concern extinction of the disease
at the mean field approximation level,
in the next section
we study the microscale model directly.
\section{Exact Model}
\label{sec:exact_model}
To proceed, we recall our assumption that at time zero each node has
the same, independent, probability, $i_0$, of being infectious;
so $\mathbb{P} ( X_j(0) = 1 ) = i_0$ for
all $1 \le j \le n$.
This implies that
$n\, i_0$ is the expected number of infectious individuals at time zero.
We are interested in the stochastic process
$\sum_i X_i(t)$, which records the number of
infected individuals.
Our analysis generalizes arguments in
\cite{epidemicsSpread}, which considered a stochastic
SIS
model on a graph
with $f$ as
the identity map.
\subsection{Extinction}
\label{subsec:ext}
Our first result shows that the spectral condition
arising from the mean field analysis in Theorems~\ref{thm: local asymptotic stability}
and \ref{thm: global asymptotic condition} is
also
relevant
to the probability of extinction in the
individual-level model.
\begin{thm}\label{thm: exponential decay}
Suppose $f$ is concave in the hypergraph infection model (\ref{infection rate random}). Then
\[
\mathbb P\left(\sum_i X_i(t)> 0\right)\leq n\, i_0\, \exp\left((\beta f'(0)\lambda(W)-\delta)t\right).
\]
Hence, if
$
\lambda(W)
f'(0)
\beta/\delta <1
$ then the disease vanishes at an exponential rate.
\end{thm}
\begin{proof}
Consider the continuous time Markov process $\{(Y_i(t))_{t\geq 0}\}_{i=1}^{n}$
taking values in $\mathbb{N}^n$, with transition of states defined for every $1 \le i\le n$ and $\ t\geq 0$ by
$$
\begin{cases}
k&\rightarrow\ k+1\text{, with rate }\beta f'(0)\sum_j W_{i j}Y_j(t),\\
k&\rightarrow\ k-1\text{, with rate }\delta.
\end{cases}
$$
This new process is introduced here purely for the purpose of analysis. However, it may be interpreted
as a disease model where the state of each individual is represented by a non-negative integer that indicates severity of infection. Here, exposure to
highly infected individuals raises the chance
of an increase in infection severity.
Suppose also that $X_i(0)=Y_i(0)$
for all $1 \le i \le n$. Since $f$ is concave,
\[
\beta \sum_h\mathcal I_{i h}f(\sum_j X_j(t)\mathcal I_{j h})
\leq
\beta \sum_h\mathcal I_{i h}f'(0) \sum_j X_j(t)\mathcal I_{j h} =
\beta f'(0)\sum_{i j}W_{i j}X_j(t),
\]
from which we see that $Y_i$ stochastically dominates $X_i$. Hence
$$
\mathbb P\left(\sum_i X_i(t)>0\right)\leq \mathbb P\left(\sum_i Y_i(t)>0\right)\leq \sum_i q_i(t),
$$
where $q_i(t):=\mathbb{E}[Y_i(t)]$.
In terms of the
Chapman--Kolmogorov Equation,
or Chemical Master Equation,
\cite{Gil91},
we have
\[
\frac{d q_i(t)}{dt}=\beta f'(0)\sum_{j}W_{i j}q_j(t)-\delta q_i(t).
\]
Letting $Q(t) = [q_1(t), q_2(t), \ldots, q_n(t)]^T$,
this linear ODE system solves to give
\[
Q(t)=\exp\left(t(\beta f'(0)W-\delta I)\right)Q(0).
\]
The matrix $\exp\left(t(\beta f'(0)W-\delta I)\right)$ is symmetric and has spectral radius
$\exp\left((\beta f'(0)\lambda(W)-\delta I)t\right)$.
Hence, in Euclidean norm,
$$
||Q(t)|| \leq \exp((\beta f'(0)\lambda(W)-I)t)||Q(0)||.
$$
Since $||Q(0)|| =\sqrt{n} \, i_0 $ and,
by Cauchy--Schwarz,
\[
\sum_i q_i(t)\leq \sqrt{n} ||Q(t)||,
\]
the proof is complete.
\end{proof}
We deduce, analogously to \cite{epidemicsSpread}, the following corollary.
\begin{corollary}
\label{cor:et}
Suppose $f$ is concave in the hypergraph infection model (\ref{infection rate random}).
Let $\tau$ denote the time of extinction of the disease and suppose
$\lambda(W) f'(0) \beta <\delta$, then
\[
\mathbb{E}[\tau]\leq\frac{\log n+1}{\delta-f'(0)\beta \lambda (W)}.
\]
\end{corollary}
\begin{proof}
Using Theorem~\ref{thm: exponential decay},
\begin{align*}
\mathbb{E}[\tau]&=\int_0^\infty\mathbb P(\tau >t)dt\\
&=\int_0^\infty\mathbb P(\sum_i X_i(t)>0)dt\\
&\leq \frac{\log n}{\delta-f'(0)\beta \lambda (W)} + \int_{(\log n)/(\delta - f'(0)\beta \lambda(W))}^\infty n\exp((\beta f'(0)\lambda(W)-\delta)t)dt\\
&\leq \frac{\log n+1}{\delta - f'(0)\beta \lambda(W)}.
\end{align*}
\end{proof}
Likewise the partitioned case yields the following result.
\begin{thm}\label{thm: generalized exponential decay}
Suppose every $f_k$ is concave in the partitioned
hypergraph model with infection rate
(\ref{eq:partition}).
Then
$$
\mathbb P\left(\sum_i X_i(t)> 0\right)\leq n \, i_0 \, \exp\left(\beta \lambda(\sum_{k=1}^{K}f'_k(0)W^{(k)})-\delta\right).
$$
Hence, if
$
\lambda(\sum_{k=1}^{K}f'_k(0)W^{(k)})
\beta /\delta < 1
$ then the disease vanishes at an exponential rate.
\end{thm}
We also have the following analogue of Corollary~7.2 on the expected time to extinction for the partitioned case.
\begin{corollary}
Suppose every $f_k$ is concave in the partitioned hypergraph model with infection rate (\ref{eq:partition}).
Let $\tau$ denote the time of extinction of the disease and suppose
$\lambda(\sum_{k=1}^{K}f'_k(0)W^{(k)})
\beta /\delta < 1$, then
\[
\mathbb{E}[\tau]\leq\frac{\log n+1}{\delta-\beta \lambda (\sum_{k=1}^{K}f'_k(0)W^{(k)})}.
\]
\end{corollary}
\subsection{Conditions that preclude extinction}
\label{subsec:hyp}
So far, we have focused on deriving thresholds
that imply extinction.
In this subsection, following
ideas from \cite{epidemicsSpread}, we
derive a condition under which
the disease will
persist.
Note that our analysis does not require
the graph
associated with $W$
to be connected.
The disconnected setting
is relevant, for example, when interventions have been imposed in order to limit interactions.
We let
$\Delta:=D-W$ denote the Laplacian,
and let $\lambda_c(\Delta)>0$ denote the smallest non-zero eigenvalue of $\Delta$.
We also let $e_{\max}$ denote the size of the largest hyperedge.
\begin{defn}
Given a hypergraph $\mathcal{H}$,
a function $f : \mathbb{R}_{+} \to \mathbb{R}_{+}$ and
a subset of the nodes $S \subset V$, let
$$
E(S,f):=\sum_{i\in S}\sum_{h\in E}\mathcal I_{i h} f(\sum_{j\in S^c}\mathcal I_{j h}),
$$
where $S^c:= V\backslash S$ is the complement of $S$.
Also define for integer $1\leq m\leq \floor{n/2}$
$$
\eta(\mathcal H,m,f):=\inf\left\{\frac{E(S,f)}{|S|}\ |\ \ 1\leq |S|\leq m\right\},
$$
and let $\eta(\mathcal H,m):=\eta(\mathcal H,m,Id)$.
\end{defn}
Notice that when $S$ consists of those nodes for which
$X_i(t) = 0$,
we can write the infection transition rate of $\sum_{i=1}^{n}X_i(t)$ as $\beta E(S,f)$.
More generally,
$\beta E(S,f)$ may be regarded as the rate at which
nodes in the set $S$ may be infected by nodes in
the remainder of the network. When $f=Id$ and $m=\floor{n/2}$, $\eta(\mathcal H,m)$ is the Cheeger constant, or isoperimetric number, associated with the weighted graph induced by $W=\mathcal I\mathcal I^T$.
We may also regard $\eta(\mathcal H,m,f)$
as the smallest
average infection rate over all subsets consisting of no more than half of the network.
The next theorem gives a probabilistic lower bound on the time to extinction.
\begin{thm}\label{thm: non-zero steady-state}
Recall that $\tau$ denotes the hitting time of the state $0$ for the process $(\sum_j X_j(t))_{t\geq 0}$
in the hypergraph model (\ref{infection rate random}).
If $f$ is concave and $\lambda_c(\Delta)>\left(2\frac{e_{\max}-1}{f(e_{\max}-1)}\right)\frac{\delta}{\beta}$, then
$$
\mathbb P\left(\tau >\frac{\floor{r^{-m+1}}}{2m}\right)\geq \frac{1-r}{e}(1+O(r^m)),
$$
where $r:=\frac{(e_{\max}-1)\delta}{ f(e_{\max}-1)\, \beta\, \eta(\mathcal H,m)}<1$ and $m:=\floor{\frac{n}{2}}$.
\end{thm}
From standard Cheeger inequalities \cite{CheegerInequalitiesSharp}, we know that the Cheeger constant of the graph induced by $W$ satisfies $2 \eta(\mathcal H, m)\geq \lambda_c(\Delta)$, hence using the assumptions on $\lambda_c(\Delta)$ in the theorem, we see that $r<1$ indeed.
In order to prove this result, we introduce the following lemma.
\begin{lemma}
If $f$ is concave and non-decreasing, then
$$\frac{f(e_{\max}-1)}{e_{\max}-1}\eta(\mathcal H,m)\leq \eta(\mathcal H,m,f).$$
\end{lemma}
\begin{proof}
The proof is immediate once we see that by concavity of $f$, for all $x\in \{0,\dots,e_{\max}-1\}$
$$
\frac{f(e_{\max}-1)}{e_{\max}-1}x\leq f(x).
$$
\end{proof}
Now, to prove Theorem~\ref{thm: non-zero steady-state} consider the Markov process $(Z(t))_{t\geq 0}$ valued in $\{0,\dots,m\}$, with transition of states given by
$$
\begin{cases}
k\ \rightarrow\ k+1\text{ with transition rate }k\beta\frac{f(e_{\max}-1)}{e_{\max}-1}\eta(\mathcal H, m),\\
k\ \rightarrow\ k-1\text{ with transition rate }k\delta.
\end{cases}
$$
This Markov process is stochastically dominated by $((\sum_{i=1}^{n} X_i(t))_{t\geq 0})$, which has
the same downward transition rate, and an upward transition rate that is at least as large:
\begin{align*}
k\beta \frac{f(e_{\max}-1)}{e_{\max}-1}\eta(\mathcal H,m)&\leq k\beta\eta(\mathcal H,m,f)\\
&\leq \beta E(S,f),
\end{align*}
where $S:=\{i\in \{1,2,\ldots,n\}\ |\ X_i(t)=0\}$. Thus, to show Theorem~\ref{thm: non-zero steady-state} it suffices to find a suitable lower bound for
$\mathbb{P}\left(\widehat{\tau} >\frac{\floor{r^{-m+1}}}{2m}\right)$, where
$\widehat{\tau} $ is the hitting time of $0$ for the process $(Z(t))_{t\geq 0}$. This follows by applying Theorem $4.1$ of \cite{epidemicsSpread} to the process $(Z(t))_{t\geq 0}$.
\section{Collective Contagion Models}
\label{sec:cim}
We now consider the collective contagion models from
\cite{APM20,simplicialSocialContagion,heterogeneityHypergraph}, where infection only starts spreading within a hyperedge once a threshold number of infectious nodes in that hyperedge has been reached.
As discussed in subsection~\ref{subsec:model}, collective contagion models can be represented by nonlinear functions of the form $f(x):=\max\{0,x-c\}$ for some $c>0$, or $f(x):=c_2\mathbbm{1}(x\geq c_1)$ for some $c_1,c_2>0$.
In these cases it is obvious that the zero-infection state for the mean field approximation is locally asymptotically stable (and, indeed,
Theorem~\ref{thm: local asymptotic stability} applies).
However,
because the functions
are not concave, the theory found in
Section~\ref{sec:exact_model}
for the exact model does not directly apply.
Nonetheless, we can still derive similar spectral conditions for the vanishing of the disease by finding concave functions which serve as upper bounds for $f$. For instance using $c_2\mathbbm{1}(x\geq c_1)\leq \frac{c_2}{c_1}x\mathbbm{1}(x\leq c_1)+c_2\mathbbm{1}(x\geq c_1)$,
the bounds in
Theorem~\ref{thm: exponential decay}
and Corollary~\ref{cor:et}
lead to the
the following result.
\begin{thm}\label{thm: collective contagion model}
Suppose that $f(x):=c_2\mathbbm{1}(x\geq c_1)$ for some $c_1,c_2>0$.
Then
$$
\mathbb P\left(\sum_i X_i(t)> 0\right)\leq n\, i_0\, \exp\left(\beta \frac{c_2}{c_1}\lambda(W)-\delta\right).
$$
In particular if $\lambda(W)<\frac{c_1}{c_2}\frac{\delta}{\beta}$, then the disease asymptotically vanishes with exponential decay and the extinction time $\tau$ satisfies
$$
\mathbb{E}[\tau]\leq\frac{\log n +1}{\delta -
\beta \lambda(W) c_2 / c_1}.
$$
\end{thm}
Likewise note that
$
\max\{0,x-c\}\leq \frac{e_{\max}-1-c}{e_{\max}-1}x,
$
where we recall that $e_{\max}$ is the largest size of a hyperedge of $\mathcal H$.
Hence we deduce the following result.
\begin{thm}\label{thm: max collective contagion model}
Suppose that $f(x):=\max\{0,x-c\}$ for some $c>0$. Then
$$
\mathbb P\left(\sum_i X_i(t)> 0\right)\leq n\, i_0\, \exp\left(\beta \frac{e_{\max}-1-c}{e_{\max}-1}\lambda(W)-\delta\right).
$$
In particular if
\[
\lambda(W)< \left( \frac{e_{\max}-1}{e_{\max}-1-c} \right) \frac{\delta}{\beta},
\]then the disease asymptotically vanishes with exponential decay and the extinction time $\tau$ satisfies
\[
\mathbb{E}[\tau]\leq\frac{\log n+1}{\delta-\beta \frac{e_{\max}-1-c}{e_{\max}-1} \lambda (W)}.
\]
\end{thm}
\section{Discussion}
\label{sec:disc}
In this work we derived several
spectral conditions that control the
spread of disease in an SIS model on a hypergraph.
The conditions have the general form
\begin{equation}\label{new vanishing condition}
\beta \, \lambda(W) \, c_f/ \delta < 1,
\end{equation}
where $c_f>0$ is a constant depending on the
function $f$
that determines the nonlinear infection rate within a hyperedge.
We note that in the special case where
(i) the hypergraph
is an undirected graph and hence $W$ becomes the binary adjacency matrix,
and
(ii) we have linear dependence on the number of infectious neighbors for the infection rate of a node, so $f$ is the identity function,
the condition
(\ref{new vanishing condition})
reduces to the
well-known vanishing spectral condition
studied in, for example,
(\cite{epidemicsSpread,virusSpreadInNetworks,spreadingEigenvalue}).
There are two important points to be made about the
general form of
(\ref{new vanishing condition}).
First, the hypergraph structure appears only via the presence of the symmetric matrix
$W \in \mathbb{R}^{n \times n}$.
Recall that
$W_{ij}$ records
the number of times that
$i$ and $j$ both appear in the same
hyperedge.
Such weighted but \emph{pairwise}
information is all that feeds into this
spectral
threshold.
On a positive note, this implies that useful
predictions can be made about
disease spread on a hypergraph without
full knowledge of the types of hyperedge
present and the distribution of nodes within them.
(For example, when collecting human interaction data it is more reasonable to ask an individual to list each
neighbour and state how many different ways they
interact with that neighbour than to ask
an individual to list all hyperedges they take part in.)
However,
this observation also raises the possibility that
more refined analysis might lead to
sharper
bounds, perhaps at the expense of simplicity and interpretability.
Our second point is that
the new vanishing condition
(\ref{new vanishing condition})
neatly separates three aspects:
\begin{description}
\item[(a)]
The biologically-motivated infection parameter, $\beta$.
\item[(b)]
The interaction
structure, captured in $\lambda(W)$.
\item[(c)]
The
coefficient $c_f$ that arises from
modelling the nonlinear infection process.
For instance, Theorem~\ref{thm: local asymptotic stability} and Theorem~\ref{thm: global asymptotic condition} have $c_f=f'(0)$. In the collective contagion model case $f(x)=c_2\mathbbm{1}(x\geq c_1)$, Theorem~\ref{thm: collective contagion model} indicates that we can take $c_f=c_2/c_1$.
\end{description}
We may view $\beta$ as an invariant biological constant
that reflects the underlying virulence of the disease
and is not affected by human behaviour.
The factor $\lambda(W)$, which arises from the interaction structure, will be determined by
regional and cultural issues, including
population density,
age demographics,
typical household sizes, and the
nature of prevalent commercial and manufacturing activities.
Interventions, including full or partial lockdowns,
could be modeled through a change in $\lambda(W)$.
The third factor, $c_f$,
is strongly dependent upon
human behaviour and may be adjusted to reflect
individual-based containment strategies such as social distancing, mask wearing or more frequent hand washing.
This work has focused on modelling, analysis and interpretation at the abstract level, concentrating on the
fundamental question of disease extinction.
Having developed this theory,
it would, of course, now be of great interest to
perform practical experiments using realistic interaction and infection data, with the aim of
\begin{itemize}
\item calibrating model parameters,
\item testing hypotheses about the appropriate functional form of the infection rate,
\item testing the predictive power of the modeling framework, especially in comparison with
simpler homogeneous mixing and pairwise interaction versions,
\item quantifying the effect of different interventions.
\end{itemize}
\bibliographystyle{siam}
|
\section{Introduction\label{Introduction}}
The intriguing behavior of glass-forming liquids is attracting continued interest from many researchers \cite{bapst2020unveiling,jensen2018slow,hecksher2017toward,klieber2013mechanical,blazhnov2004temperature,niss2018perspective,klieber2015nonlinear,gundermann2011predicting}.
By virtue of its ability to simultaneously probe multiple relaxation processes (thermal expansion, acoustic and even orientational response \cite{glorieux2002thermal,silence1992structural}), the use of impulsive stimulated scattering (ISS) in a periodical grating geometry has been successful in obtaining new insights from the thermoelastic response (measured via the accompaning coherent diffraction of a probe laser beam) to impulsive photothermal excitation of different glassformers \cite{yang1995t,yang1995impulsive2,paolucci2000impulsive,halalay1992liquid,halalay1992time,silence1992structural,silence1990impulsive}.
Standard thermo-mechanical modelling, based on the assumption of frequency independent or non-relaxing heat capacity and thermal expansion coefficient, have been shown not to be adequate to characterize the dynamics triggered in an ISS experiment, especially for viscous systems.
Along with the first experimental ISS results, a semi-empirical model (SEM) \cite{yang1995impulsive1}, relying on a stretched-exponential function to describe the nontrivial initial thermal expansion rise of the ISS signal, has proved to be effective to describe the ISS response of glycerol, salol, and oil DC705 \cite{yang1995t,yang1995impulsive1,yang1995impulsive2,paolucci2000impulsive}.\\
Inspired by successful descriptions in literature of experimental results for the temperature response to heating \cite{birge1985specific,bentefour2003broadband,bentefour2004thermal,niss2012dynamic} by invoking a frequency dependent heat capacity $C$, and indications for a frequency dependent thermal expansion coefficient $\gamma$ \cite{blazhnov2004temperature}, here, we derive analytically a generalized ISS model to take into account the relaxation of $C$ and $\gamma$, which are not implicitly considered in the SEM.
We start from frequency domain versions of the thermal diffusion equation and the thermoelastic equation and we impose a frequency dependent heat capacity and thermal expansion coefficient according to Debye and Havriliak-Negami (HN) relaxation models.
We then investigate to what extent the introduced physical model is consistent with the empirical model by conducting a case study on ISS results of glycerol reported in Refs. \cite{paolucci2000impulsive,Liu2021}.
We also present an interpretation of the Debye assumption for the frequency dependent heat capacity and thermal expansion coefficient in the framework of a two-temperature model (TTM). Furthermore, a set of experimentally recorded ISS signals of a supercooled glycerol is analysed with the developed models to extract $C(\omega)$ and $\gamma(\omega)$ up to sub-100 MHz. This largely extends the upper limit of the previously accessible bandwidth, 100 kHz \cite{bentefour2003broadband} and 1 Hz \cite{niss2012dynamic} respectively for the spectroscopy of \textit{C} and $\gamma$, enabling the comparison of fragility by thermal, mechanical, and dielectric susceptibilities in a broader frequency/temperature range. \\
The manuscript is organized as follows:
In Section \ref{Sec:Temperature} we present analytical expressions for the temperature response to impulse photothermal excitation in a grating geometry in two scenarios: frequency independent and frequency dependent (according to Debye and HN functions) heat capacity.
In Section \ref{Sec:displacement} a continuum mechanics thermoelastic model is used to calculate the response of the material strain to photothermal excitation, by considering the temperature change as source term in the equation of motion, in which Debye and HN relaxation behavior of thermal expansion is coupled.
A comparison between results obtained by the newly proposed approach and simulations by the empirical model for literature values on glycerol \cite{paolucci2000impulsive} is presented in Section \ref{Sec:comparison_with_liter}. In Section \ref{Sec:two_temp_model} the compatibility between Debye frequency dependence of the heat capacity and of thermal expansion coefficient and the TTM is verified. Finally, in Section \ref{experiment}, we apply the developed models to the concrete case of the experimental ISS signals recorded on a supercooled glycerol.\\
The present work accompanies the results presented in Ref. \cite{Liu2021}, in which the thermal relaxation dynamics of glycerol is investigated by a combination of ISS and thermal lens spectroscopy \cite{ThermalLens}.\\
\section{Temperature response to impulsive photothermal excitation in a periodical grating geometry}
\label{Sec:Temperature}
\subsection{Scenario with frequency-independent heat capacity \label{Subsec:Frequency_independent_heat_capacity}}
In this section we calculate the temperature evolution of a system that is subject to impulsive photothermal excitation generating a transient thermal grating (TTG). For now, we assume that the heat capacity is frequency independent.\\
The starting point is the thermal diffusion equation in the temperature $T$ for a 1D infinite geometry \cite{gandolfi2019accessing}:
\begin{equation}
\frac{\partial^2 T}{\partial x^2 }-\frac{\rho C}{\kappa_T} \frac{\partial T}{\partial t}=-\frac{Q(x,t)}{\kappa_T},
\label{diffusion_equation2}
\end{equation}
where $\rho$ (kg m$^{-3}$), $\kappa_T$ (W m$^{-1}$ K$^{-1}$) and $C$ (J kg$^{-1}$ K$^{-1}$)) are the mass density, the thermal conductivity and the frequency-independent heat capacity per unit mass, while $Q(x,t)$ (W m$^{-3}$) is the heat source. In ISS experiments, the heat input is impulsive in time and periodical in space:
\begin{equation}
Q(x,t)=Q_0\cos(qx)\delta(t),
\label{Q_x_t}
\end{equation}
where $Q_0$ (J m$^{-3}$) is the supplied heat density and $q$ (m$^{-1}$) is the wavenumber, defined as $2\pi$ times the inverse of spatial period of periodical light intensity pattern. \\
Prior to excitation, the system is at equilibrium at constant temperature $T_0$.\\
After taking a Fourier transform, the following frequency domain expression is obtained:
\begin{equation}
\frac{\partial^2 \tilde{T}}{\partial x^2 }-i\omega\frac{\rho C}{\kappa_T} \tilde{T}=-\frac{1}{\kappa_T}\tilde{Q}(x,\omega),
\label{diffusion_equation_in_freq_domain_AA}
\end{equation}
and the solution for temperature field reads:
\begin{equation}
\tilde{T}(x,\omega)=T_0\delta(\omega)+\frac{Q_0}{2\pi i \rho C( \omega-i\alpha q^2)}\cos(qx),
\label{temperature_in_freq}
\end{equation}
where $\alpha=\kappa_T/(\rho C)$ (m$^2$ s) is the thermal diffusivity.
By taking an inverse Fourier transform, the following expression is obtained for the temperature evolution:
\begin{equation}
T(x,t)=T_0+\frac{Q_0}{\rho C}\cos(qx)\exp\left(-\alpha q^2 t\right)\theta(t),
\label{temperature_in_time}
\end{equation}
where $\theta(t)$ is the Heaviside step function.\\
\subsection{Scenario with frequency dependent heat capacity}
\label{Subse:Frequ-dependent heat capacity}
\subsubsection{Debye model \label{Subsum:Debye}}
The following Debye expression for the frequency dependent heat capacity \cite{fivez2011dynamics}:
\begin{equation}
C(\omega)=C_\infty+\frac{\Delta C}{1+i\omega \tau_C}=C_\infty+\frac{\Delta C}{1+i\frac{\omega}{\omega_C}},
\label{C_omega}
\end{equation}
with $C_\infty$ the part of the heat capacity related to the higher frequency response, or, in time domain, the instantaneous response of the temperature to impulsive heating. $\Delta C$ is the additional part of the heat capacity that determines the reduction of the temperature response at low frequencies (lower than the relaxation frequency $\omega_C=\tau_C^{-1}$), or, in time domain, at times longer than the relaxation time $\tau_C$.
Upon substitution of the expression for $C(\omega)$ into Eq. \ref{diffusion_equation_in_freq_domain_AA} we get the following differential equation:
\begin{equation}
\frac{\partial^2 \tilde{T}}{\partial x^2 }-i\omega\frac{\rho}{\kappa_T} \left(C_\infty+\frac{\Delta C}{1+i{\frac{\omega}{\omega_C} }}\right)\tilde{T}=-\frac{1}{\kappa_T}\tilde{Q}(x,\omega).
\label{diffusion_equation_in_freq_C_omega}
\end{equation}
Inserting the expression for the heat source $\tilde{Q}(x,\omega)$ (obtained transforming in frequency domain Eq. \ref{Q_x_t}), we obtain the following solution:
\begin{equation}
\tilde{T}(x,\omega)=T_0\delta(\omega)-\frac{iQ_0\left(\omega-i\omega_C\right)}{2\pi \rho C_\infty\left(\omega-\omega_1\right)\left(\omega-\omega_2\right)}\cos(qx),
\label{temperature_in_freq_C_omega}
\end{equation}
with $\alpha_\infty=\kappa_T/(\rho C_\infty)$ the high frequency limit of the thermal diffusivity and the frequences
\begin{widetext}
\begin{eqnarray}
\omega_1=\frac{i}{2}\left\{\left[\alpha_\infty q^2+\omega_C\left(1+\frac{\Delta C}{C_\infty}\right)\right]-\sqrt{\left[\alpha_\infty q^2+\omega_C\left(1+\frac{\Delta C}{C_\infty}\right)\right]^2-4\alpha_\infty q^2\omega_C}\right\}, \label{omega_1} \label{omega1_def}\\
\nonumber\\
\omega_2=\frac{i}{2}\left\{\left[\alpha_\infty q^2+\omega_C\left(1+\frac{\Delta C}{C_\infty}\right)\right]+\sqrt{\left[\alpha_\infty q^2+\omega_C\left(1+\frac{\Delta C}{C_\infty}\right)\right]^2-4\alpha_\infty q^2\omega_C}\right\}, \label{omega2_def}\\
\nonumber
\end{eqnarray}
The expression for the temperature in time domain can be obtained by applying the inverse Fourier transform to Eq. \ref{temperature_in_freq_C_omega} (see Appendix \ref{appen:res_th} for more details).
Thus, we obtain:
$$T(x,t)=T_0+\frac{Q_0}{\rho C_\infty}\cos(qx)\times\left[\frac{\left(\omega_1-i\omega_C\right)}{\left(\omega_1-\omega_2\right)}\exp\left(i\omega_1 t\right)+\frac{\left(\omega_2-i\omega_C\right)}{\left(\omega_2-\omega_1\right)}\exp\left(i\omega_2 t\right)\right]\theta(t)=$$
$$=T_0+\frac{Q_0}{\rho C_\infty}\cos(qx)\exp\left\{-\frac{t}{2}\left[\alpha_\infty q^2+\omega_C\left(1+\frac{\Delta C}{C_\infty}\right)\right]\right\}\times$$
$$\times\left\{\cosh\left({t\over2}\sqrt{\left[\alpha_\infty q^2+\omega_C\left(1+\frac{\Delta C}{C_\infty}\right)\right]^2-4\alpha_\infty q^2\omega_C}\right)\right.+$$
\begin{equation}
-\frac{\left[\alpha_\infty q^2+\omega_C\left(\frac{\Delta C}{C_\infty}-1\right)\right]}{\sqrt{\left[\alpha_\infty q^2+\omega_C\left(1+\frac{\Delta C}{C_\infty}\right)\right]^2-4\alpha_\infty q^2\omega_C}}\left.\sinh\left({t\over2}\sqrt{\left[\alpha_\infty q^2+\omega_C\left(1+\frac{\Delta C}{C_\infty}\right)\right]^2-4\alpha_\infty q^2\omega_C}\right)\right\}\theta(t).
\label{T_t_expression_final}
\end{equation}
\subsubsection{Havriliak-Negami model \label{Subsum:Havriliak}}
Simple Debye relaxation behavior has turned out not to be a fully adequate description for the dynamic behavior of many glass-forming materials.
By virtue of two additional model parameters $a_C$ and $b_C$, the generalized \textit{Havriliak-Negami} (HN) model \cite{havriliak1966complex}:
\begin{equation}
C(\omega)=C_\infty+\frac{\Delta C}{\left[1+\left(i\omega \tau_C\right)^{a_C}\right]^{b_C}}=C_\infty+\frac{\Delta C}{\left[1+\left(i\frac{\omega}{\omega_C}\right)^{a_C}\right]^{b_C}},
\label{C_omega_HN}
\end{equation}
has been shown to be more effective (note that the HN model tends to the Debye model when $a_C=b_C=1$).\\
In this case, the temperature response in the frequency domain is given by:
\begin{equation}
\tilde{T}(x,\omega)=T_0\delta(\omega)-\frac{iQ_0\left[\omega_C^{a_C}+(i\omega)^{a_C}\right]^{b_C}\cos(qx)}{2\pi \rho C_\infty\left\{\omega\left[\omega_C^{a_C}+(i\omega)^{a_C}\right]^{b_C}+\frac{\Delta C}{C_\infty}\omega\omega_C^{a_Cb_C}-i\alpha_\infty q^2\left[\omega_C^{a_C}+(i\omega)^{a_C}\right]^{b_C}\right\}}.
\label{temperature_in_freq_C_omega_HN}
\end{equation}
As the exponents $a_C$ and $b_C$ are typically non-integer, analytical derivation of the inverse Fourier transform of the latter expression is cumbersome. Therefore, in this work, this inverse Fourier was performed numerically.\clearpage
\end{widetext}
\section{ISS Signal}
\label{Sec:displacement}
\subsection{Constitutive equation}
\label{SubSec:Constitutive_Eq}
Impulsive stimulated scattering occurs due to coherent diffraction of a probe beam that trespasses a medium of interest in which, via its relation with the refractive index, a spatially periodic strain pattern is present. The ISS signal is therefore proportional to the magnitude of the strain grating in the medium.
In this subsection we derive, for different scenarios for the relaxation behavior an expression for the displacement and strain.
We assume that the viscoelastic behavior of the material can be described by the Kelvin-Voigt model, corresponding with a lumped model containing a spring and a dashpot in parallel (as described on page 87 of Ref. \cite{auld1973acoustic}).\\
Under this assumption, the constitutive equations are:
\begin{equation}
\left\{\begin{array}{l}
\displaystyle{\rho \frac{\partial^2 \mathbf{u}}{\partial t^2}=\nabla\cdot \sigma}\\
\\
\displaystyle{\sigma=\mathbf{C}\varepsilon+\eta\frac{\partial \varepsilon}{\partial t},}
\end{array}\right.
\label{constitutive_relations}
\end{equation}
where $\mathbf{u}$ (m) is the displacement, $\sigma$ (Pa) is the stress, $\mathbf{C}$ (Pa) is the stiffness matrix and $\eta$ (Pa s) is the viscosity tensor. The strain $\varepsilon$ can be written as:
\begin{equation}
\varepsilon=\nabla_S\mathbf{u}-\gamma_M \Delta T,
\label{definition_of_strain_C}
\end{equation}
where $\nabla_S\mathbf{u}=\frac{\nabla \mathbf{u}+\nabla^T\mathbf{u}}{2}$, $\gamma_M$ [$K^{-1}$] is the matrix of linear expansion and $\Delta T$ is temperature variation that drives the mechanics \cite{gandolfi2020optical}. This approach is in agreement with Green-Lindsay theory for thermoviscoelastic media \cite{mukhopadhyay1999relaxation,othman2012fundamental}.\\
We write the viscoelastic tensor as $\eta=\tau_\eta \mathbf{C}$
\footnote{As indicated on page 88 of Ref. \cite{auld1973acoustic}, the viscosity tensor has same form of the stiffness matrix.
Hence, we can write viscosity tensor as $\eta=\tau_\eta C$, where $C$ is the stiffness matrix \cite{mukhopadhyay1999relaxation,othman2012fundamental}.}, where $\tau_\eta$ representes the damping time and we assume that the medium is homogeneous and isotropic. Upon these assumptions, the equation ruling the displacement reads:
\vspace{3cm}
\begin{widetext}
\begin{equation}
\frac{\partial^2 u_x}{\partial t^2}=
c_L^2\left(1+\tau_\eta\frac{\partial}{\partial t}\right)\frac{\partial^2 u_x}{\partial x^2}-(3c_L^2-4c_T^2)\gamma\left(1+\tau_\eta\frac{\partial}{\partial t}\right)\frac{\partial T}{\partial x},
\label{Expansion_isotropic}
\end{equation}
where $c_L=\sqrt{(\lambda+2\mu)/\rho}$ and $c_T=\sqrt{\mu/\rho}$ are the longitudinal and transverse velocities (m/s), with $\lambda$ (Pa) and $\mu$ (Pa) the two Lam\'{e} coefficients. $\gamma$ is the linear expansion coefficient.
\end{widetext}
\subsection{ISS response in case of frequency independent heat capacity and thermal expansion}
\label{SubSec:Freq_ind_u}
In this subsection we derive, for a non-relaxing medium, a general expression for the displacement occurring when the system is excited by a transient optical grating. \\
To do this, we apply the temporal Fourier transform to Eq. \ref{Expansion_isotropic} to get:
$$-\omega^2 \tilde{u}_x=
c_L^2(1+i\omega\tau_\eta)\frac{\partial^2 \tilde{u}_x}{\partial x^2}+$$
\begin{equation}
-\left(3c_L^2-4c_T^2\right)\left(1+i\omega \tau_\eta\right)\gamma\frac{\partial \tilde{T}}{ \partial x}.
\label{Expansion_isotropic_FT}
\end{equation}
By defining:
\begin{equation}
c^2(\omega)=c_L^2(1+i\omega\tau_\eta)
\label{def_c}
\end{equation}
and
\begin{equation}
\xi(\omega)=3-4\frac{c_T^2}{ c_L^2},
\label{def_xi}
\end{equation}
we can write Eq. \ref{Expansion_isotropic_FT} in the more compact form:
\begin{equation}
\frac{\partial^2 \tilde{u}_x}{\partial x^2}+\frac{\omega^2}{c^2(\omega)} \tilde{u}_x=\xi(\omega)\gamma\frac{\partial \tilde{T}}{ \partial x}.
\label{Expansion_isotropic_FT_compact}
\end{equation}
In order calculate the displacement occurring due to the TTG excitation, we use the solution for the temperature in frequency domain, as derived in Section \ref{temperature_in_freq}:
\begin{equation}
\frac{\partial^2 \tilde{u}_x}{\partial x^2}+\frac{\omega^2}{c^2(\omega)} \tilde{u}_x=Z(\omega) \sin(qx),
\label{Expansion_isotropic_FT_compact2}
\end{equation}
where
\begin{equation}
Z(\omega)=-\frac{qQ_0\xi(\omega)\gamma}{2\pi i \rho C( \omega-i\alpha q^2)}.
\label{def_Z}
\end{equation}
The general solution of Eq. \ref{Expansion_isotropic_FT_compact2} is $\tilde{u}(x,\omega)=z(x,\omega)+z_p(x,\omega)$, where $z_p(x,\omega)$ is a particular solution of Eq. \ref{Expansion_isotropic_FT_compact2}, while $z(x,\omega)$ is the solution of the associated homogeneous differential equation.
It can be shown that
\begin{equation}
z_p(x,\omega)=\frac{Z(\omega)c^2(\omega)}{\omega^2- q^2c^2(\omega)}\sin(qx)
\end{equation}
is a particular solution of Eq. \ref{Expansion_isotropic_FT_compact2}.\\
In order to have the system at rest before the excitation (i.e. $u(x,t)=0$ and $\frac{du}{dt}(x,t)=0$ for negative times) and avoid divergence of the displacement at infinity, we must have $z(x,\omega)=0$ $\forall \omega$.\\
Hence the final solution is:
\begin{equation}
\tilde{u}(x,\omega)=z_p(x,\omega)=\frac{Z(\omega)c^2(\omega)}{\omega^2- q^2c^2(\omega)}\sin(qx)
\end{equation}
Substituting back the expressions for $Z(\omega)$, $\xi(\omega)$ and for $c(\omega)$ in the latter equation we obtain:
\begin{equation}
\tilde{u}(x,\omega)=-\frac{qQ_0\gamma\left(3c_L^2-4c_T^2\right)\left(1+i\omega \tau_\eta\right)}{2\pi i \rho C(\omega-i\alpha q^2)\left(\omega-\omega_3\right)\left(\omega-\omega_4\right)}\sin(qx),
\label{u_omega_substituted2}
\end{equation}
where
\begin{equation}
\omega_3=i\frac{q^2}{2}\left[c_L^2\tau_\eta-\sqrt{c_L^4\tau_\eta^2-\frac{4\rho}{q^2}c_L^2} \right].
\label{def_of_omega3}
\end{equation}
and
\begin{equation}
\omega_4=i\frac{q^2}{2}\left[c_L^2\tau_\eta+\sqrt{c_L^4\tau_\eta^2-\frac{4\rho}{q^2}c_L^2} \right].
\label{def_of_omega4}
\end{equation}
The time domain expression for the displacement can be obtained by applying an inverse Fourier transform to Eq. \ref{u_omega_substituted2} (see Appendix \ref{appen:res_th}), resulting in:
\begin{widetext}
$$u(x,t)=-\frac{qQ_0\gamma}{\rho C}\sin(qx)\left(3c_L^2-4c_T^2\right)\left\{\frac{1-\alpha q^2 \tau_\eta}{\left(i\alpha q^2-\omega_3\right)\left(i\alpha q^2-\omega_4\right)}\exp\left(-\alpha q^2t\right)\right.+$$
\begin{equation}
\left.+\frac{1+i\omega_3 \tau_\eta}{(\omega_3-i\alpha q^2)\left(\omega_3-\omega_4\right)}\exp\left(i \omega_3 t\right)+\frac{1+i\omega_4 \tau_\eta}{(\omega_4-i\alpha q^2)\left(\omega_4-\omega_3\right)}\exp\left(i \omega_4 t\right)\right\}\theta(t)
\label{u_t_non_dep_on_freq}
\end{equation}
\subsection{ISS response in case of frequency dependent heat capacity and thermal expansion described by the Debye model}
\label{SubSec:Freq_dep_u}
In this section we assume that the heat capacity depends on frequency according to Debye model, in analogy with Subsection \ref{Subsum:Debye}.
Hence, we substitute Eq. \ref{C_omega} into Eq. \ref{u_omega_substituted2} to get:
\begin{equation}
\tilde{u}(x,\omega)=-\frac{qQ_0\gamma\left(3c_L^2-4c_T^2\right)\left(1+i\omega \tau_\eta\right)(\omega-i\omega_C)}{2\pi i \rho C_\infty(\omega-\omega_1)(\omega-\omega_2)\left(\omega-\omega_3\right)\left(\omega-\omega_4\right)}\sin(qx),
\label{u_omega_substituted_C_omega}
\end{equation}
\end{widetext}
where $\omega_1$ and $\omega_2$ are defined in Eq.s \ref{omega1_def} and \ref{omega2_def}.\\
Furthermore, we also assume the linear expansion coefficient to be frequency dependent following the Debye expression:
\begin{equation}
\gamma(\omega)=\gamma_\infty+\frac{\Delta \gamma}{1+i\omega \tau_\gamma}=\gamma_\infty+\frac{\Delta \gamma}{1+i\frac{\omega}{\omega_\gamma}},
\label{gamma_omega}
\end{equation}
where $\gamma_\infty$ and $\Delta \gamma$ represent the instantaneous and additional relaxing contribution of the response of the volume to a temperature change, respectively.
$\omega_\gamma=\tau_\gamma^{-1}$ is the associated relaxation frequency.\\
\\
For the sake of simple analytical treatment, and given that the focus of this work is on the thermal expansion part of the signal and not on the superposed acoustic part, in the following, we neglect the frequency and temperature dependence of the elastic moduli and of the density \cite{jensen2018slow,hecksher2017toward,klieber2013mechanical,blazhnov2004temperature}. \\
With this choice, the equation to be solved reduces to:
\begin{equation}
\frac{\partial^2 \tilde{u}_x}{\partial x^2}+\frac{\omega^2}{c^2(\omega)} \tilde{u}_x=\xi(\omega)\left(\gamma_\infty+\frac{\Delta \gamma}{1+i\frac{\omega}{\omega_\gamma}}\right)\frac{\partial \tilde{T}}{ \partial x}.
\label{Expansion_isotropic_FT_compact_gamma_omega}
\end{equation}
The expression for $\gamma(\omega)$ can be rewritten as:
\begin{equation}
\gamma(\omega)=\gamma_\infty\frac{\omega-i\omega_\gamma\left(1+\frac{\Delta \gamma}{\gamma_\infty}\right)}{\omega-i\omega_\gamma}=\gamma_\infty\frac{\omega-\omega_6}{\omega-\omega_5},
\label{gamma_omega_2}
\end{equation}
where
\begin{equation}
\omega_5=i\omega_\gamma
\label{omega5_def}
\end{equation}
and
\begin{equation}
\omega_6=i\omega_\gamma\left(1+\frac{\Delta \gamma}{\gamma_\infty}\right).
\label{omega6_def}
\end{equation} \vspace{0.5 cm}
Substituting Eq. \ref{gamma_omega_2} into Eq. \ref{u_omega_substituted_C_omega} we get:
\begin{widetext}
\begin{equation}
\tilde{u}_x(x,\omega)=-\frac{qQ_0\gamma_\infty\left(3c_L^2-4c_T^2\right)\left(1+i\omega \tau_\eta\right)(\omega-i\omega_C)(\omega-\omega_6)}{2\pi i \rho C_\infty\prod_{j=1}^{5}(\omega-\omega_j)}\sin(qx).
\label{u_omega_substituted_C_omega2}
\end{equation}
By inverse Fourier inverse transforming the latter expression we obtain the solution for the displacement in time domain:
\begin{equation}
u_x(x,t)=-\left[\frac{qQ_0\gamma_\infty \sin(qx)}{\rho C_\infty}\right]\left(3c_L^2-4c_T^2\right)
\sum_{l=1}^{5}\left\{\left(1+i\omega_l \tau_\eta\right)(\omega_l-i\omega_C)(\omega_l-\omega_6)\left(\prod_{\substack{j=1 \\ j\neq l}}^{5}{1\over\omega_l-\omega_j}\right)\exp(i\omega_l t)\right\}\theta(t).
\label{u_t_substituted_C_omega}
\end{equation}
The ISS signal $U_{ISS}(t)$ is proportional to the amplitude of the strain grating \cite{fivez2011dynamics,yan1987impulsive} $\Delta \rho/\rho$.
Hence, the strain of the 1D displacement pattern equally its spatial derivative, the ISS signal can be derived from Eq. \ref{u_t_substituted_C_omega} as:
\begin{equation}
U_{ISS}(t)\propto \max_{x}\left[\frac{\partial u_{x}(x,t)}{\partial x}\right].
\end{equation}
\subsection{ISS response in case of frequency dependent heat capacity and thermal expansion described by the Havriliak-Negami model}
\label{SubSec:Freq_dep_u_HN}
In the HN scenario for the thermal expansion response,
\begin{equation}
\gamma(\omega)=\gamma_\infty+\frac{\Delta \gamma}{\left[1+\left(i\omega \tau_\gamma\right)^{a_\gamma}\right]^{b_\gamma}}=\gamma_\infty+\frac{\Delta \gamma}{\left[1+\left(i\frac{\omega}{\omega_\gamma}\right)^{a_\gamma}\right]^{b_\gamma}},
\label{gamma_omega_HN}
\end{equation}
with $a_\gamma$ and $b_\gamma$ are additional model parameters.\\
Taking also the heat capacity behavior according to the HN model, as described in Subsection \ref{Subsum:Havriliak} we get:
\begin{equation}
\tilde{u}(x,\omega)=-\frac{qQ_0\gamma_\infty\left(3c_L^2-4c_T^2\right)\left(1+i\omega B_C\tau_\eta\right)\left[\omega_C^{a_C}+(i\omega)^{a_C}\right]^{b_C}\left\{\left[\omega_\gamma^{a_\gamma}+(i\omega)^{a_\gamma}\right]^{b_\gamma}+\frac{\Delta \gamma}{\gamma_\infty}\omega_\gamma^{a_\gamma b_\gamma}\right\}\sin(qx)}{2\pi i \rho C_\infty\left\{\omega\left[\omega_C^{a_C}+(i\omega)^{a_C}\right]^{b_C}+\frac{\Delta C}{C_\infty}\omega\omega_C^{a_Cb_C}-i\alpha_\infty q^2\left[\omega_C^{a_C}+(i\omega)^{a_C}\right]^{b_C}\right\}\left(\omega-\omega_3\right)\left(\omega-\omega_4\right)\left[\omega_\gamma^{a_\gamma}+(i\omega)^{a_\gamma}\right]^{b_\gamma}}.
\label{u_omega_substituted_HN}
\end{equation}
The ISS signal can then be obtained in analogy with \ref{SubSec:Freq_dep_u}.
\end{widetext}
\section{Comparison with stretched exponential model \label{Sec:comparison_with_liter}}
\begin{figure}
\begin{center}
\includegraphics[width=0.38\textwidth]{Figure1_PRB.pdf}
\caption{Plot of the time dependence of the ISS signal obtained with the SEM (black curves), most squares fit of this signal with the Debye model (blue curves) and the HN model (red curves), for different temperature-wavenumber combinations, based on material parameters of glycerol reported in \cite{paolucci2000impulsive} and listed in Table \ref{table_Param}. The small fitting residues (fitting curve minus SEM curve) indicate that the Debye based model is adequate. Thanks to the two additional model parameters, the HN model is fitting even better.
In each panel, all the curves are normalized to the maximum of the SEM ISS signal.}
\label{ISS_strVsHN}
\end{center}
\end{figure}
A semi-empirical model (SEM) describing ISS signals in glassformers has been introduced along with the first experimental reports on ISS signals in salol \cite{yang1995t,yang1995impulsive1,yang1995impulsive2} and has been successfully used to fit ISS data in glycerol \cite{paolucci2000impulsive}.
The SEM expression describing the ISS signal is \footnote{
Since in this work we study the ISS signal detected with a heterodyne experimental setup, the right-hand-side of Eq. \ref{Str_exp_expr} is not squared. This choice is at variance with respect to Ref. \cite{paolucci2000impulsive}, which is based on homodyne detection.
}:
$$I(t)=(A+B)\exp\left(-\Gamma_Ht\right)+$$
\begin{equation}
-A\exp\left(-\Gamma_At\right)\cos\left(\omega_At\right)-B\exp\left[-\left(\Gamma_Rt\right)^\beta\right],
\label{Str_exp_expr}
\end{equation}
where the first term is associated to the thermal diffusion ($\Gamma_H$ being the thermal decay rate), while the second one corresponding to the acoustics ($\omega_A$ being the acoustic oscillation frequency and $\Gamma_A$ the acoustic damping rate).
The use of a stretched exponential - also known as Kohlrausch-Williams-Watts (KWW) - term was inspired by other response functions in the physics of supercooled liquids, and was aimed at coping with the empirical observation that the initial thermal expansion cannot be fitted by a simple exponential.
$\Gamma_R$ is the structural relaxation rate and $0<\beta\leq 1$ the stretching exponent.
The coefficients $A$ and $B$ account for the weights of each term contributing to the total ISS signal.\\
The SEM has proved useful to describe and fit the ISS signal measured on supercooled glycerol, as described by Paolucci et al. \cite{paolucci2000impulsive}.
The fitting parameters of interesting temperature-wavenumber combinations treated by Paolucci et al. are recalled in Table \ref{table_Param} and were used to simulate respective theoretical ISS signals, shown in Fig. \ref{ISS_strVsHN} (black curves).
\begin{table}[h]
\begin{center}
\begin{tabular}{|c|c|c|c|c|}\hline
Case &\#1 & \#2&\#3&\#4 \\ \hline
$T_0$ (K) & 250 & 250 & 230 & 230 \\ \hline
$q$ (m$^{-1}$) & 3.05$\times 10^5$ & 1.036$\times 10^6$ & 3.05$\times 10^5$ & 1.036$\times 10^6$ \\ \hline
$\Gamma_H$ (s$^{-1}$)\footnote{
The thermal decay rate was estimated as $\Gamma_H=\kappa_T q^2/(\rho\ C)$, where $\kappa_T=0.28\ \um{W/m\ K}$, $\rho=1260\ \um{kg/m^3}$ \cite{gupta2012scope} and $C=1800$ J/kg K (for T$_0=250$ K) or $C=1500$ J/kg K (for T$_0$=230 K), in agreement with Refs. \cite{bentefour2003broadband,bentefour2004thermal}.
} & 1.14$\times 10^4$ & 1.33$\times 10^5$ & 1.38$\times 10^4$ & 1.59$\times 10^5$ \\ \hline
$\Gamma_A$ [s$^{-1}$]\footnote{
The acoustic damping rate was taken from Fig. 3 of Ref. \cite{paolucci2000impulsive}.
} & 3.5$\times 10^7$ & 8.5$\times 10^7$ & 2.0$\times 10^6$ & 2.5$\times 10^6$ \\ \hline
$\omega_A$ (Grad/s)\footnote{
The acoustic oscillation frequency was obtained as the product between $q$ and the speed of sound reported in Fig. 3 of Ref. \cite{paolucci2000impulsive}.
} & 0.98 & 3.32 & 1.03 & 3.49 \\ \hline
$\beta$\footnote{
The stretch exponent was taken from Fig. 5 of Ref. \cite{paolucci2000impulsive}.
} & 0.6 & 0.6 & 0.6 & 0.6 \\ \hline
$\Gamma_R$ (s$^{-1}$)\footnote{
The structural relaxation rate was obtained as $\Gamma_R=\Gamma({1/ \beta})/\left(\langle\tau\rangle\beta \right)$, where $\Gamma$ is the Gamma function, and $\langle\tau\rangle$ was taken from Fig. 4 of Ref. \cite{paolucci2000impulsive}.
} & 5.5$\times 10^6$ & 5.5$\times 10^6$ &1.1$\times 10^5$ & 1.1$\times 10^5$ \\ \hline
$B/A$\footnote{
The ratio between the coefficients $B$ and $A$ was calculated as $B/A=f/(1-f)$, where $f=0.67$ is the Debye-Waller factor taken from Fig. 7 of Ref. \cite{paolucci2000impulsive}.
} & 2.03 & 2.03 & 2.03 & 2.03 \\ \hline
\end{tabular}
\end{center}
\caption{Material parameters reported by Paolucci et al. based on SEM fits of ISS signals in supercooled glycerol.
The parameters reported in case \#4 imply a SEM model ISS signal with the unphysical negative tail, as reported in Fig. \ref{Case4_ISS_signal}. This unphysical behavior is not present in the other cases.}
\label{table_Param}
\end{table}
Three cases are considered in Fig. \ref{ISS_strVsHN}: in panel a and b the ISS of glycerol is reported for the two grating wave numbers $q=3.05\times10^5\ \um{m^{-1}}$ and $q=1.036\times10^6\ \um{m^{-1}}$ respectively, at the same temperature $T_0=250$ K (i.e. cases \#1 and \#2 in Table \ref{table_Param}, respectively).
Panel c reports the ISS signal for a lower temperature $T_0=230$ K and for the shortest grating wave vector available, i.e. $q=3.05\times10^5\ \um{m^{-1}}$ (case \#3 in Table \ref{table_Param}). For each panel, the value of the coefficient $A$ was chosen in order to have the maximum of the ISS signal normalized to 1.\\
In order to verify if the Debye and HN models developed in the previous section are able to reproduce the SEM based ISS signal, we have fitted the black curves in Fig. \ref{ISS_strVsHN} with the respective expressions (blue curves: Debye, red curves: HN).
The fitting was carried out by implementing a most-squares fitting (MSF) protocol \cite{jackson1976most,salenbien2011laser} to search for the minimum of the cost function, defined as the sum of the squared residuals (SSR).
MSF is advantageous over the commonly used least-squares fitting (LSF) as it is able to take into account the possible co-variance of the multiple fitting variables, namely, different combinations of fitting parameters yielding a statistically indistinguishable cost function value SSR (local minima).
Both models fit very well, with the residues of the HN model being the smallest, thanks to the additional two fitting parameters.
\begin{table*}
\begin{center}
\begin{tabular}{|c|c|c|c|}\hline
Case & \#1 & \#2 & \#3 \\ \hline
$C_{\infty}$ Debye (J/kg K) & 452 & 460 & 360 \\ \hline
$C_{\infty}$ HN (J/kg K) & $503\in[0,6\times 10^5]$ & $503\in[0,6\times 10^5]$ & $390\in [388,400]$\\ \hline
$C_0$ Debye (J/kg K) & 1837 & 1882 & 1645 \\ \hline
$C_0$ HN (J/kg K) & $1777\in[1775,1798]$ & $1777\in[1775,1798]$ & $1345\in[1340,1351]$ \\ \hline
$\Delta \gamma/\gamma_\infty$ Debye & 10 & 10 & 10 \\ \hline
$\Delta \gamma/\gamma_\infty$ HN & $9.4\in[9.3,9\times 10^3]$ & $9.4\in[9.3,9\times 10^3]$ & $9.8\in[9.7,10]$ \\ \hline
$\omega_C$ (Mrad/s) Debye & 5.1 & 6.6 & 0.2 \\ \hline
$\omega_C$ (Mrad/s) HN &$3.5\in[0,4\times10^3]$&$3.5\in[0,4\times10^3]$&$1.5\in[1.1,500]$\\ \hline
$\omega_\gamma$ (krad/s) Debye & 3190 & 4080 & 150 \\ \hline
$\omega_\gamma$ (krad/s) HN & $2260\in[2230,2320] $ & $2260\in[2230,2320] $& $38\in[0,39]$ \\ \hline
$a_c$ HN & $0.89\in[0.06,0.9]$ & $0.89\in[0.06,0.9]$ & $0.544\in[0.543,0.554]$ \\ \hline
$b_c$ HN & $0.52\in[0.02, 0.55]$ & $0.52\in[0.02, 0.55]$ & $0.78\in[0.75,1]$ \\ \hline
$a_\gamma$ HN & $0.9\in[0,1]$&$0.9\in[0,1]$ &$0.7\in[0,1]$ \\ \hline
$b_\gamma$ HN & $0.68\in[0.67,1]$ & $0.68\in[0.67,1]$ & $0.5\in[0.1,1]$ \\ \hline
\end{tabular}
\end{center}
\caption{Fitting parameter values obtained by fitting the SEM curves reported in Fig. \ref{ISS_strVsHN} with Debye and HN model based expressions for the ISS signal. For the HN model, we report also confidence interval next to the best fit parameter.}
\label{table_fitting_coeff}
\end{table*}
The obtained fitting parameters are summarized in Table \ref{table_fitting_coeff}. For the HN model, for each parameter, the fitting error was determined by most squares analysis of the cost function in the multidimensional space of fitting parameters and thus includes the effect of covariance with other fitting parameters \cite{jackson1976most,salenbien2011laser}.
Instead of displaying the parameter $\Delta C$ obtained from the fit, in Table \ref{table_fitting_coeff} we have reported the zero-frequency heat capacity $C_0=C_\infty+\Delta C$. The latter definition was retrieved from Eq. \ref{C_omega} in the limit $\omega\rightarrow0$.
The fits with the HN model expression have been performed simultaneously on cases $\#1$ and $\#2$, where the glycerol equilibrium temperature is the same, but the grating wave numbers are different. In ISS signals, besides a change of acoustic frequency and damping, a difference in wave number results (when viewing the signal on a logarithmic time scale) in a different "onset" time ($1/q^2\alpha$) of the thermal diffusion driven washing out of the thermal (expansion) grating. Once ongoing, this exponential decay dominates the signal behavior and masks the influence of the parameters that determine the onset of the relaxation of the temperature to heat and thermal expansion to temperature response. Simultaneously fitting signals at two wavenumbers, and thus considering signals containing two mixing ratios of the respective influences, limits the possibilities for covariant influences of fitting parameter values on the signals, and thus leads to smaller uncertainties.
Interestingly, the uncertainty on $C_0$ is always small. This can be explained by the strong influence of the low frequency/late time limit of the heat capacity on the signal, via the always significant thermal diffusion related exponential decay tail of the signal. For the chosen signals, this tail occurs later than the relaxation times, and it is thus not affected by possible degeneracy of the effect of $C_0$ on the signal with relaxation influenced thermal expansion parameters. The influence of $C_\infty$ on the initial part of the signal goes along with influence of the thermal expansion and the acoustic wave related parameters. This leads to a larger fitting covariance.
\begin{figure}
\begin{center}
\includegraphics[width=0.4\textwidth]{Figure2_PRB.pdf}
\caption{Time evolution of the temperature obtained for the case of Debye and HN models (blue and red curves, respectively), simulated for different temperature-wavenumber combinations.
In each panel, all the curves are normalized to the value of the Debye model temperature at the shortest displayed time\protect\footnote{
Eq.s \ref{T_t_expression_final} and \ref{temperature_in_freq_C_omega_HN} depend on the spatial coordinate.
The curves in Fig. \ref{temperature_strVsHN} were evaluated at the same spatial coordinate.
The particular choice of the latter is irrelevant thanks to the proposed normalization.}.}
\label{temperature_strVsHN}
\end{center}
\end{figure}
The limited effect of $C_\infty$ on the ISS signal can be further understood by looking at the calculated temperature evolution after impulsive illumination, as depicted in Fig. \ref{temperature_strVsHN} for the Debye and HN models (blue and red curves, respectively).
The temperature evolution for the Debye model was obtained by evaluating Eq. \ref{T_t_expression_final} upon insertion of the Debye model parameters listed in Table \ref{temperature_strVsHN}.
For the calculation based on the HN model, the thermal parameters were first inserted into Eq. \ref{temperature_in_freq_C_omega_HN} to calculate the HN temperature in frequency domain.
The latter was then Fourier transformed numerically to time domain, obtaining the red curves in Fig. \ref{temperature_strVsHN}.
For long times the Debye and HN model based temperature evolutions match. However, despite the fact that the Debye and HN model yield a very similar ISS signal (as shown in Fig. \ref{ISS_strVsHN}), the corresponding fitting parameters reported in Table \ref{temperature_strVsHN} imply a very different temperature profile at short times.
This indicates a very strong degeneracy between the early dynamics of the temperature, the thermal expansion and the acoustic wave generation.\\
From another point of view, one may wonder why two different temperature profiles can give rise to the same ISS signal.
This can also be seen by looking into the math: substituting the expression for the temperature into the source term of the Eq. \ref{Expansion_isotropic_FT_compact} for the displacement, we see that the source term is proportional to $\gamma/\left[\rho C\left(\omega-i\alpha q^2\right)\right]$ (as reported in Eq. \ref{def_Z}).
Hence, at high frequencies the source term is proportional to $\gamma/C$.
Fig. \ref{C(omega)_Gamma(omega)} shows that the real part of $C(\omega)$ and $\gamma(\omega)$ for cases $\#3$ follow a very similar trend, both for the Debye and HN model.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.41\textwidth]{Figure3_PRB.pdf}
\caption{Real part of $\left(C(\omega)-C_\infty\right)/C_\infty$ (brown lines, left axis) and of $\left(\gamma(\omega)-\gamma_\infty\right)/\gamma_\infty$ (purple lines, right axis), as a function of frequency (horizontal axis, log scale).
The full and dashed lines refer to the Debye and HN model, respectively.
These curves were calculated for the case $T_0=230$ K and $q=3.05\times 10^5\ \um{m^{-1}}$ (case $\#3$).
}
\label{C(omega)_Gamma(omega)}
\end{center}
\end{figure}
We verified that this is also the case for the imaginary part of these quantities.
Hence, even if the time dependences of $C$ and $\gamma$ are quite different between the Debye and HN scenario, their ratio, and thus the corresponding ISS signal is similar.\\
A similar observation has been made for cases $\#1$ and $\#2$.\\
In conclusion, at short times the parameters $C$ and $\gamma$ are degenerate and their individual values cannot be reliably extracted from fitting.
Conversely, at low frequency the source term is no longer simply proportional to the ratio $\gamma/C$, hence the degeneracy is lifted and at long times the heat capacity and thermal expansion coefficient can be disentangled precisely by the the fitting procedure. A prospective scenario for lifting the large degeneracy between $C_\infty$ and $\gamma_\infty$ could be to further increase the wavenumber so that the thermal diffusion tail occures before the heat capacity relaxation time. In that scenario, the shape of the thermal diffusion tail is dominated by the decay time, which gives direct information on $C_\infty$, with little influence of the other parameters. \\
It is worth to note that the ISS signal obtained for case $\#4$, which is representative for a rather low temperature (long relaxation times) and a rather long grating spacing (long thermal diffusion time) in Fig. \ref{Case4_ISS_signal} reveals a temporal span in which the ISS signal, calculated according to the SEM model, is negative.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.36\textwidth]{Figure4_PRB.pdf}
\caption{Plot of the ISS signal obtained with SEM (black curves), for the case $T_0=230$ K and $q=1.036\times 10^6\ \um{m^{-1}}$ (case $\#4$).
The curve is normalized to 1 at the maximum. The graph shows that the ISS signal goes below zero (dashed red line).}
\label{Case4_ISS_signal}
\end{center}
\end{figure}
This unphysical result, which is a consequence of a particular mix between positive and negative terms in Eq. \ref{Str_exp_expr}, prevents an adequate comparison with the here presented models, both in the Debye and HN scenario.
For the sake of curiosity, we have evaluated the conditions for which the SEM based ISS signal goes negative.
For this evaluation, we have simplified Eq. \ref{Str_exp_expr} by neglecting the acoustic term with respect to the thermal diffusion, because (i) the amplitude of the former ($A$) is smaller than the one of the latter ($A+B$) and (ii) the former decays faster than the latter ($\Gamma_A$ being much larger than $\Gamma_H$).
Upon this simplification, the ISS signal in the SEM model becomes:
\begin{equation}
I(t)=(A+B)\exp\left(-\Gamma_Ht\right)-B\exp\left[-\left(\Gamma_Rt\right)^\beta\right].
\label{Str_exp_expr_simplified}
\end{equation}
By performing some algebric calculations, one obtains that $I(t)$ is positive for time instants satisfying the following inequality:
\begin{equation}
t^\beta\geq\left(\Gamma_H\over \Gamma_R^\beta\right)t-{1\over \Gamma_R^\beta}\ln\left(1+\frac{A}{B}\right).
\label{inequality_I_bigger_0}
\end{equation}
It is evident that for the limit $t\rightarrow0$, Eq. \ref{inequality_I_bigger_0} is satisfied, while in the limit $t\rightarrow +\infty$, Eq. \ref{inequality_I_bigger_0} is violated.
Hence, there must be at least one time where $I(t)$ changes sign.
We call $t^*$ the earliest positive time satisfying $I(t)=0$.\\
Furthermore, since $\beta<1$ for the cases reported in Table \ref{table_Param}, the slope of the left-hand side of Eq. \ref{inequality_I_bigger_0} decreases for increasing time, while the slope of the right-hand side is constant.
Therefore, for times later than $t^*$, the right-hand-side grows faster than the left-hand-side and hence, there are not other time instants satisfying $I(t)=0$.
Summarizing, for $0\leq t\leq t^*$, the ISS signal is positive, while for $t>t^*$ the ISS signal is negative.\\
The derivation of an analytic expression for $t^*$ is challenging.
However, we can have some insight by speculating on the time instant:
\begin{equation}
t^*_{low}=\left(\Gamma_H\over \Gamma_R^\beta\right)^{1\over(\beta-1)},
\label{t_star_low}
\end{equation}
which is a lower bound for $t^*$ \footnote{
$t^*_{low}$ is the time instant solving Eq. \ref{inequality_I_bigger_0} without the last term, the latter reading:
$$t^\beta=\left(\Gamma_H\over \Gamma_R^\beta\right)t.$$
Hence, we have
$$\left(t^*_{low}\right)^\beta=\left(\Gamma_H\over \Gamma_R^\beta\right)t^*_{low}>\left(\Gamma_H\over \Gamma_R^\beta\right)t^*_{low}-{1\over \Gamma_R^\beta}\ln\left(1+\frac{A}{B}\right).$$
The latter relation states that $t^*_{low}>t^*$ or, in other words, $t^*_{low}$ is a lower bound for $t^*$}.\\
In Table \ref{table_t_star_low} we have evaluated $t^*_{low}$ for the four cases considered in the current section.
\begin{table}[h]
\begin{center}
\begin{tabular}{|c|c|c|c|c|}\hline
Case &\#1 & \#2&\#3&\#4 \\ \hline
$t^*_{low}$ (s) & 0.93 & $2.0\times 10^{-3}$ & $1.6\times 10^{-3}$ & $3.6\times 10^{-6}$ \\ \hline
\end{tabular}
\end{center}
\caption{Coefficients reported by Paolucci et al. concerning the fitting of the supercooled glycerol ISS signals with SEM model.}
\label{table_t_star_low}
\end{table}
For cases $\#1$ to $\#3$, the lower bound for $t^*$ (and, hence, $t^*$ itself) occurs on only after several milliseconds.
Therefore, the negative ISS signal is not visible in Fig. \ref{ISS_strVsHN}: it occurs on late times that are not shown in the figure.
For case $\#4$ $t^*_{low}\sim 3.6\ \um{\mu s}$, paving the way to the detection of a negative ISS signal for times $t>t^*\sim 10\ \um{\mu s}$. This is confirmed in Fig. \ref{Case4_ISS_signal}.\\
\section{Debye model vs two-temperature model\label{Sec:two_temp_model}}
\subsection{Frequency dependent heat capacity \label{Subsec:two_temp_model_HC}}
As mentioned earlier, part of the energy that is optically supplied to a relaxing material is channeled, around the relaxation time, to a change of configurational energy that goes along with a structural rearrangement of the amorphous network.
In this section, we describe the network's energy distribution in terms of a temperature $T_N$, and we assume that the configurational energy reservoir is in thermal contact with a kinetic (mainy vibrational) energy reservoir (KER), with physically measurable temperature T.
The energy flux between the two reservoirs is quantified as $G(T-T_N)$, where $G$ (W/(m$^3$ K)) is a (positive) coupling constant; this term indicates that when $T>T_N$, then energy flows from the KER to the network.
The capability of storing and transferring energy within the network are formally described by the network's heat capacity $C_N$, thermal conductivity $k_{T,N}$ and density $\rho_N$.\\
Hence, in this approach, the energy exchange between the KER (vibrational energy) and the network (configurational energy) can be described by a two-temperature model (TTM) \cite{caddeo2017thermal}, yielding the following equations for the KER's temperature $T$ and the network's temperature $T_N$:\\
\begin{equation}
\left\{\begin{array}{l}
\displaystyle{\rho C \frac{\partial T}{\partial t}=\kappa_T\frac{\partial^2 T}{\partial x^2}+Q(x,t)-G(T-T_N)},\\
\\
\displaystyle{\rho_N C_N \frac{\partial T_N}{\partial t}=k_{T,N}\frac{\partial^2 T_N}{\partial x^2}+G(T-T_N)},\\
\end{array}
\right.
\label{two_temperatures_model}
\end{equation}
where $\rho$, $C$ and $\kappa_T$ are the KER's density, heat capacity and thermal conductivity, respectively.
The source term $Q$ (W/m$^3$) enters only in the equation for the KER temperature, indicating that the optical excitation delivers energy directly to the KER. The network modifications follow the dynamics occurring in the KER, hence they are driven by variations of $T$.\\
For the sake of simplicity, we assume that local network reconfigurations only depend on the local energy exchange with the KER, and we neglect possible configurational energy flow within the network.
This yields a very low value for the network's thermal conductivity, allowing to drop the term $k_{T,N}{\partial^2 T_N}/\partial x^2$ in System \ref{two_temperatures_model}.\\
The first equation of System \ref{two_temperatures_model} can be reformulated as:
\begin{equation}
T_N=T+\frac{\rho C}{G} \frac{\partial T}{\partial t}-\frac{\kappa_T}{G}\frac{\partial^2 T}{\partial x^2}-{1\over G}Q(x,t).
\label{first_TTM_rewritten}
\end{equation}
Substituting the latter expression into the second equation of System \ref{two_temperatures_model}, we get to the following differential equation:
$$\frac{\rho_N C_N}{G}\left(\rho C \frac{\partial^2 T}{\partial t^2}-\frac{\partial Q}{\partial t}+G\frac{\partial T}{\partial t}-\kappa_T\frac{\partial^3 T}{\partial t \partial x^2} \right)=$$
\begin{equation}
=-\rho C\frac{\partial T}{\partial t}+Q(x,t)+\kappa_T\frac{\partial^2 T}{\partial x^2}.
\label{decoupled_for_T}
\end{equation}
Applying a Fourier transform to Eq. \ref{decoupled_for_T} we get to:
$$\frac{\rho_N C_N}{G}\left[-\omega^2\rho C \tilde{T}-i\omega \tilde{Q}(x,\omega)+i\omega G\tilde{T}-i\omega \kappa_T\frac{\partial^2 \tilde{T}}{ \partial x^2} \right]=$$
\begin{equation}
=-i\omega\rho C\tilde{T}+\tilde{Q}(x,\omega)+\kappa_T\frac{\partial^2 \tilde{T}}{\partial x^2},
\label{decoupled_for_T_freq}
\end{equation}
which can be rearranged as:
$$\kappa_T\left(1+i\frac{\rho_N C_N}{G}\omega\right)\frac{\partial^2 \tilde{T}}{\partial x^2}+$$
$$+\left[\frac{\rho_N C_N \rho C}{G}\omega^2-i\omega\left(\rho_N C_N +\rho C\right)\right]\tilde{T}+$$
\begin{equation}
+\left(1+i\omega\frac{\rho_N C_N}{G} \right) \tilde{Q}(x,\omega)=0.
\label{decoupled_for_T_freq_2}
\end{equation}
Dividing by $\kappa_T(1+i\omega \rho_N C_N/G)$ and rewriting the second term we have:
\begin{equation}
\frac{\partial^2 \tilde{T}}{\partial x^2}-i\omega\frac{\rho}{\kappa_T}\left[C+\frac{{\rho_N \over \rho} C_N}{1+i\frac{\rho_N C_N}{G}\omega}\right]\tilde{T}=-{1\over \kappa_T}\tilde{Q}(x,\omega).
\label{decoupled_for_T_freq_4}
\end{equation}
The latter equation can be remapped into Eq. \ref{diffusion_equation_in_freq_C_omega} upon substitutions $C\rightarrow C_\infty$, $\rho_N C_N/\rho\rightarrow \Delta C$ and $G/(\rho_N C_N)\rightarrow \omega_C$.
Hence, it can be stated that the frequency dependent heat capacity in terms of the Debye model (as reported in Eq. \ref{C_omega}) is equivalent to the TTM.\\
\subsection{Frequency dependent thermal expansion in the frame of the two-temperature model
\label{Sec:two_temp_model_therm_exp}}
Along with the process of taking up potential energy, the network is undergoing structural changes, and it can thus change its volume.
Hence, it can contributute to the system's thermal expansion and add up to the thermal strain that is related to the increase of vibrational energy (which is connected to the anharmonicity of the intermolecular potential minima).
To the best of our knowledge, no models describing the latter point are reported in literature.
Nevertheless, it seems reasonable to assume that the thermal strain produced both by the KER and the network depends on the history of the network's temperature.
Assuming an isotropic material, the total thermal strain can hence be written as $\left[\gamma \Delta T\circledast \varphi(t)+\gamma_{N} \Delta T_N\circledast \varphi_N(t)\right]I_d$, where $I_d$ is the identity matrix.
Indeed, the two temperatures have been convolved with memory functions $\varphi(t)$ and $\varphi_N$ describing the thermal history of the KER and network respectively.\\
After substituting $\gamma_M \Delta T(x,t)$ with $\left[\gamma \Delta T\circledast \varphi(t)+\gamma_{N} \Delta T_N\circledast \varphi_N(t)\right]I_d$ into Eq \ref{definition_of_strain_C}, we can repeat analogously the derivation presented in Section \ref{Sec:displacement}.\\
In this way, we reach to the following equation for the displacement ($\xi(\omega)$ is defined in Eq. \ref{def_xi}):
$$\frac{\partial^2 \tilde{u}_x}{\partial x^2}+\frac{\omega^2}{c^2(\omega)} \tilde{u}_x=$$
$$=\xi(\omega)\frac{\partial }{ \partial x}\left(\gamma\tilde{T}(x,\omega)\tilde{\varphi}(\omega)+\gamma_N\tilde{T}_N(x,\omega)\tilde{\varphi}_N(\omega)\right)=$$
\begin{equation}
=\xi(\omega)\left(\gamma\tilde{\varphi}(\omega)+\frac{\gamma_N\tilde{\varphi}_N(\omega)}{1+i\frac{\rho_N C_N}{G}\omega}\right)\frac{\partial \tilde{T}}{ \partial x}.
\label{Expansion_isotropic_FT_compact_TN}
\end{equation}
The last step involved the substitution $\tilde{T}_N=\tilde{T}(x,\omega)/\left(1+i\frac{\rho_N C_N}{G}\omega\right)$ (see Appendix \ref{app:Derivation of the} for the proof of the latter expression).\\
We suppose that the network contribution to the thermal strain contains an instantaneous term and a second term accounting for the network's thermal history.
These considerations are well reproduced by a memory function of the type:
\begin{equation}
\varphi_N(t)=2\pi\delta(t)+2\pi\omega_C\left(1-\chi_\gamma\right)\exp(-\omega_\gamma t)\theta(t)
\label{def_of_varphi}
\end{equation}
which corresponds to the following expression for the network's contribution to the thermal strain:
$$\gamma_{N} \Delta T_N\circledast \varphi_N(t)=2\pi \gamma_{N}\Delta T_N(x,t)+$$
\begin{equation}
+2\pi\omega_C\left(1-\chi_\gamma\right)\int_{-\infty}^{t}\exp\left[-\omega_\gamma(t-\tau)\right]\Delta T_N(\tau)d\tau.
\label{equivalence_therm_strain_varphi}
\end{equation}
With the latter definitions, the past thermal events in the network are exponentially less and less important with increasing time in the past.
The temporal cutoff for the exponential is the inverse of a frequency $\omega_\gamma$, which can be written in terms of the frequency $\omega_C=G/(\rho_NC_N)$ (already introduced in Subsection \ref{Subsec:two_temp_model_HC}) as:
\begin{equation}
\omega_\gamma=\chi_\gamma \omega_C=\chi_\gamma\frac{G}{\rho_NC_N}.
\label{chi_gamma}
\end{equation}
$\chi_\gamma$ quantifies to what extent the thermal expansion relaxation is slower than the heat capacity relaxation.\\
The Fourier transform of the memory function reads \footnote{The Fourier transform of $2\pi\delta(t)$ is $1$. Furthermore, the Fourier transform of $e^{-A t}\theta(t)$, with $A$ real and strictly positive, is $(2\pi)^{-1}(i\omega+A)^{-1}$.}:
\begin{equation}
\tilde{\varphi}_N(\omega)=1+\frac{\omega_C\left(1-\chi_\gamma\right)}{i\omega+\omega_\gamma}={1\over\chi_\gamma}\frac{1+i{\omega\over \omega_C}}{1+i{\omega\over \omega_\gamma}}.
\label{phi_omega}
\end{equation}
By introducing Eq. \ref{phi_omega} into Eq. \ref{Expansion_isotropic_FT_compact_TN} we get to:
\begin{equation}
\frac{\partial^2 \tilde{u}_x}{\partial x^2}+\frac{\omega^2}{c^2(\omega)} \tilde{u}_x=\xi(\omega)\left(\gamma\tilde{\varphi}(\omega)+\frac{\gamma_N/\chi_\gamma}{1+i{\omega\over\omega_\gamma}}\right)\frac{\partial \tilde{T}}{ \partial x}.
\label{Expansion_isotropic_FT_compact_semi_final}
\end{equation}
We should also provide an expression for the KER's memory function.
If we assume that only the instantaneous value of $T$ is important for the evaluation of the thermal strain, i.e. $\varphi(t)=2\pi\delta(t)$, then $\tilde{\varphi}(\omega)$ becomes identically 1, yielding:
\begin{equation}
\frac{\partial^2 \tilde{u}_x}{\partial x^2}+\frac{\omega^2}{c^2(\omega)} \tilde{u}_x=\xi(\omega)\left(\gamma+\frac{\gamma_N/\chi_\gamma}{1+i{\omega\over\omega_\gamma}}\right)\frac{\partial \tilde{T}}{ \partial x}.
\label{Expansion_isotropic_FT_compact_final}
\end{equation}
By performing the substitution $\gamma\rightarrow \gamma_\infty$ and $\gamma_N/\chi_\gamma\rightarrow \Delta \gamma$, the latter equation can be mapped on Eq. \ref{Expansion_isotropic_FT_compact_gamma_omega}.
Again, the choice of the Debye model for $\gamma$ (relation \ref{gamma_omega}) can be justified in terms of the TTM.\\
Equivalently, the choice of the thermal expansion in the frame of Debye model implies that the thermal strain is related instantaneously to the KER's temperature. On the other hand, the thermal history of the network has to be accounted for to estimate the thermal strain. The Debye model implies that the memory function for the network is described by Eq. \ref{def_of_varphi}.\\
Analogously, the thermal expansion ruled by the HN model can be justified considering the thermal history of the KER and of the network.
However, the complexity of the HN model, yielding also non-integer exponentials, prevents the possibility of having a simple and general expression for the memory functions.\\
\section{Experimental results and discussion}
\label{experiment}
We have heretofore developed the generalized physical model addressing the ISS response of glass-forming liquids subject to ultrafast photothermal excitation.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.5\textwidth]{Figure5_PRB.png}
\caption{Scheme of the experimental setup based on the heterodyne-detected transient grating technique. A spatially periodical laser pattern from a pulsed pump laser (red) is formed on the sample to create thermoelastic transients, the latter detected by a coaxially aligned probe laser (green and arrows).
In the scheme we sketch the phase mask (PM), the lenses (L1 and L2), the optical cryostat (OC) and the photodetector (PD).}
\label{exp_setup}
\end{center}
\end{figure}
Now, an experimental study of the ISS response of glycerol ($>99.0 \%$ purity) under supercooling is presented. An ultrafast heterodyne-detected transient grating (HD-TG) setup is used for the experiment.
Fig. \ref{exp_setup} shows the scheme of the setup, in which a ps pump laser pulse at 1064 nm (shown in red) is diffracted by a transmission phase mask (PM) into two 1st diffraction orders, namely $\pm1$ orders.
The two are then recombined via a two-lens (4\textit{f}) imaging system into the bulk of the sample. The sample is accommodated in a liquid nitrogen optical cryostat (OC) to allow temperature control over it. The light interference forms a spatially periodical light pattern and creates a transient local density grating (thermoelastic transients), at a wavelength identical to the spacing of the light grating, $d$.
For a given light wavelength light $\lambda$, one can tune the spacing of the excitation pattern by varying the intersecting angle of the two beams $\theta$, namely via $d=\lambda/(2 n) \sin{(\theta/2)}$ with $n$ the optical refractive index of the sample medium.
In this setup, the $\theta$-tuning is implemented by translating a phase mask (PM) array containing multiple PMs of varying period \cite{verstraeten2015determination}.
Alternatively, one can also rotate the PM to realize the $\theta$-tuning \cite{vega2015laser}.
The detection of ISS takes advantage of the optical heterodyne scheme \cite{maznev1998optical}, in which the probe beam from a continuous wave (CW) laser at wavelength 532 nm (shown in green with black arrows in Fig. \ref{exp_setup}), is aligned to be coaxial with the pump beam.
Both beams are sent to the PM and diffracted into excitation and probe/reference beam pairs. This heterodyne scheme has been widely used in the field for studying optical transparent or weakly absorbing liquids \cite{brodard2005application,taschin2008time,glorieux2002thermal} owing to its high sensitivity. More detailed description of the setup can be found in Ref. \cite{salenbien2012photoacoustic}.
In our experiments, the temperature scanning measurements are performed from 320 K to 200 K with a step of 1 K, under the excitation of three different gratings with $d$ of 10, 14, and 20 \textmu m.\\
\begin{figure*}
\begin{center}
\includegraphics[width=1\textwidth]{Figure6_PRB.png}
\caption{a) Experimental ISS signal (colour scale, arbitrary units) of supercooled glycerol over a broad temperature (vertical axis) and time window (horizontal axis).
The grating size is 10 \textmu m (top), 14 \textmu m (central), and 20 \textmu m (bottom).
b) to d) best fit based on Debye, HN models, and the SEM, respectively. A full presentation of the best fit of all the waveforms is shown in the online Movies.}
\label{results_ISS}
\end{center}
\end{figure*}
Fig. \ref{results_ISS}(a) presents the recorded ISS waveform datasets. As DC-temperature decreases, the acoustic ripples at short times shift the oscillation frequencies from low to high, 60-350 MHz covered by the three gratings, with the attenuation reaching a maximum around 280 K.
This observation reflects the undercooling of the sample, the latter undergoing a transition from liquid-like to glassy-like, and solid-like due to reduced molecular mobility \cite{liu1998jamming}.
The overshoot-like response is noteworthy, spanning from the start of the signal (bluish region), where it overlaps with the acoustic oscillations and fast part of thermal expansion, till the late times (reddish region), when it is quenched by the thermal diffusion dominated part (bluish tail).
This process is the manifestation of the relaxation of heat capacity and thermal expansion coefficient, which are strongly (quasi exponentially) temperature dependent.\\
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.45\textwidth]{Figure7_PRB.png}
\caption{Temperature dependent complex longitudinal velocity ($c$) determined with the three gratings.
The top (bottom) panel corresponds to the real (imaginary) part of $c$.}
\label{V_complex}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{Figure8_PRB.png}
\caption{a) low-frequency limit of $C$ (left axis, red) and $\gamma$ (right axis, blue) vs temperatures. b) The relative ratio of each relaxation quantity, $\Delta C$/$C_0$ (left axis, red) and $\Delta\gamma $/$\gamma_0$ (right axis, blue). The data are obtained from the fit in the frame of Debye model and for $d=14$ \textmu m.}
\label{C_gamma_debye_waller}
\end{center}
\end{figure}
A comparative fitting analysis of the acquired ISS datasets is carried out through the two analytical physical models, developed in this work, coupled with Debye and HN relaxation function, and also through the SEM, the latter relying on a single stretched exponential.
A full presentation of the best fits for all signals, obtained with the three models, is summarized in Fig. \ref{results_ISS} (b-d) and also available in the online Movies (1-3) in the Supplemental Material.
Satisfactory fit quality is overall achieved at all temperatures and grating periods by the three models, confirming again the reliability of physical models developed in this work.
ISS signals are information-rich, providing access to the mechanical and thermal relaxation dynamics in a single waveform, which will be discussed in the following.\\
By fitting the experimental traces with our models, for every temperature and light grating we retrieve $c_L$ and $\tau_\eta$.
By inserting these parameters into Eq. \ref{def_c}, we can calculate the complex velocity of the medium at the acoustic frequency imposed by the grating $\omega_a=2\pi c_L/d$.
Fig. \ref{V_complex} shows the obtained complex sound velocity of the supercooled glycerol at different temperatures determined with the three gratings. The real part of $c$ (top panel) increases upon cooling because of the stiffening of the liquid. The imaginary part of $c$ (bottom panel) reaches a maximum around 280 K, where the structural relaxation timescale overlaps with the acoustic frequency. The results are in good agreement with those reported in Ref. \cite{paolucci2000impulsive}. It is interesting to notice that both the real and imaginary part undergoes a transition around 280 K, which is a reflection of the strong coupling between the acoustic motion and structural changes of the network when $1/f_A$ being of the order of the structural relaxation time.
This feature provides a way to study the mechanical relaxation by performing measurements at numerous grating spacings in a broad range \cite{hecksher2017toward}, namely a mechanical spectroscopic analysis like the traditional rheological spectroscopy \cite{jensen2018slow} or ultrasonic spectroscopy \cite{jeong1986ultrasonic,schroyen2020bulk}.\\
In addition to the mechanical relaxation dynamics, the models developed in this work enable the individual and simultaneous determination of the heat capacity $C$ and the thermal expansion coefficient $\gamma$ relaxation.
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{Figure9_PRB.png}
\caption{The relaxation of $C$ is also manifested in the thermal diffusion tail of the signal, via its influence on the effective thermal diffusivity, $\alpha_{eff}$=$\kappa/(\rho C_{eff})$. At the low and high temperature limits, the value of $C_{eff}$ extracted from the thermal diffusion tail corresponds well to the respective asymptotic values $C_0$ and $C_{\infty}$, indicated by the dashed lines. }
\label{C_effective}
\end{center}
\end{figure}
In the following, we focus on the case of $d=14$ \textmu m, analyzed in the frame of Debye model, the other cases yielding the same conclusions.
Fig. \ref{C_gamma_debye_waller} (a) shows the obtained low-frequency limit response of $C$ (left axis, red) and $\gamma$ (right axis, blue), in the frame of Debye model.
Panel (b) displays the fitted ratio of $\Delta C/C_{\infty}$ (left axis, red) and $\Delta \gamma/\gamma_{\infty}$ (right axis, blue) at different temperatures.
Within the margin of uncertainty (error bars in Fig. \ref{C_gamma_debye_waller}), determined by the most square error analysis \cite{salenbien2011laser, ThermalLens}, no temperature dependence is observed for all the parameters.
Large fitting uncertainty was found when $T<230$ K and $T>260$ K.
This is because the slow parts of the responses, which are determined by the relaxation strengths and the relaxation frequencies, occur later than 100 \textmu s (when $T<230$ K), and thus after the thermal diffusion driven decay of the signal, or before 1 ns (when $T> 260$ K), the experimentally accessible time window, respectively. We thus used the values between 230-260 K to calculate the average as a representation. Access to lower temperatures can be enabled by using larger grating spacing.
In an accompanying work, we have demonstrated the use of thermal lens technique \cite{ThermalLens}, with a focused Gaussian beam of about 30 \textmu m, to study the relaxation down to 200 K.\\
The average $C_0$ and $\gamma_0$ from ISS technique are $1980\pm160$ J Kg$^{-1}$ K$^{-1}$ and $(5.5\pm0.7)\times 10^{-4}$ K$^{-1}$, respectively.
The average ratios $\Delta C/C_\infty$ and $\Delta \gamma/\gamma_\infty$ are $1.2\pm0.2$ and $4.9\pm0.7$, for $C$ and $\gamma$, respectively. Using the latter four fitting parameters, one can further calculate the high-frequency limit response, $910\pm150$ J Kg$^{-1}$ K$^{-1}$ for $C_{\infty}$ and $(1.0\pm0.2)\times 10^{-4}$ K$^{-1}$ for $\gamma_{\infty}$, and the relaxation strength ($R_S$), defined as $\Delta C/C_0$ and $\Delta \gamma/\gamma_0$, $0.5\pm0.1$ and $0.81\pm0.04$, respectively.
The obtained results comply well with the data reported in literature, as summarized in Table \ref{table_rel_quan_exp}.\\
By fitting with SEM model, the Debye-Waller factor \cite{paolucci2000impulsive}, $B/(A+B)$ in Eq. \ref{Str_exp_expr}, is used to describe the relaxation strength and we found a value of about $0.65\pm0.05$, which is in good agreement with the one reported in Ref. \cite{paolucci2000impulsive}, the latter reading being $0.66$.
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|c|c|}\hline
& Fit & 3-omega & PPE & DSC\\ \hline
$C_0$ (J Kg$^{-1}$ K$^{-1}$) & $1980\pm160$ & 2071 & 2100 & 2000\\ \hline
$C_\infty$ (J Kg$^{-1}$ K$^{-1}$) & $910\pm150$ & 1070 & 1180 & 1000\\ \hline
$R_{S}$ & $0.5\pm0.1$ & 0.48 & 0.44 & 0.5\\ \hline
Ref. & Current work & \cite{birge1985specific,birge1986specific} & \cite{bentefour2003broadband,bentefour2004thermal} & \cite{wang2002direct}\\ \hline
\multicolumn{5}{c}{}\\ \cline{1-3}
& Fit & Dilatometer & \multicolumn{2}{c}{}\\ \cline{1-3}
$\gamma_0$ ($10^{-4}K^{-1}$) & $5.5\pm0.7$ & 1 & \multicolumn{2}{c}{}\\ \cline{1-3}
$\gamma_\infty$ ($10^{-4}$K$^{-1}$) & $1.0\pm0.2$ & 5 & \multicolumn{2}{c}{}\\ \cline{1-3}
$R_{S}$ & $0.81\pm0.04$ & 0.8 & \multicolumn{2}{c}{}\\ \cline{1-3}
Ref. & Current work & \cite{blazhnov2004temperature} & \multicolumn{2}{c}{}\\ \cline{1-3}
\end{tabular}
\end{center}
\caption{Low-frequency and high-frequency limit of the average relaxing quantity $C$ and $\gamma$ and comparison with results in literature obtained with 3-omega, differential scanning calorimetry (DSC), and photopyroelectric spectroscopy (PPE).
In 3-omega and PPE, one measures thermal effusivity ($e$), from which $C(\omega)$ may be indirectly obtained via $e^2=C\kappa_T$, with $\kappa_T$ the thermal conductivity. To perform the conversion, we used $\kappa_T=$0.29 W m$^{-1}$ K$^{-1}$.
}
\label{table_rel_quan_exp}
\end{table}
The value lies in between the relaxation strength of $C$ and $\gamma$, which is expected in the sense that the two relaxing quantities are implicitly incorporated together into a single stretched exponential function.\\
Interestingly, the asymptotic values of the heat capacity can also be extracted, independently of the used models, from the temperature dependence of the thermal diffusion tail of the signal, as depicted in Fig. \ref{ISS_strVsHN}. Provided the relaxation time of the heat capacity and thermal expansion occur before or after the time window of the thermal diffusion tail, the signal tail evolves simply proportional with $\exp (-q^2\alpha_{eff} t)$ with $\alpha_{eff}$ an effective thermal diffusivity value, connected to the specific heat via $\alpha_{eff}$=$\kappa_T/(\rho C_{eff})$. $\kappa_T$ and $\rho$ denote thermal conductivity and mass density respectively.
In light of their weak temperature dependence \cite{blazhnov2004temperature,minakov2001simultaneous}, in this work the latter two parameters have been assumed as constant, as 0.29 W m$^{-1}$ K$^{-1}$ and 1260 Kg m$^{-3}$, respectively.\\
In Fig. \ref{C_effective} we report $C_{eff}$ as a function of temperature for the three gratings.
The asymptotic values of $C$ for low and high temperatures were found to be $960\pm20$ and $2190\pm30$ J Kg$^{-1}$ K$^{-1}$, as indicated by the two dashed lines, corresponding to a relaxation strength of 0.56, consistent with the value obtained by model fitting, 0.53.
\begin{figure*}
\begin{center}
\includegraphics[width=0.8\textwidth]{Figure10_PRB.png}
\caption{Comparison of the temperature dependent relaxation frequency $f_R$ extracted through Debye model (yellow squares), HN model (blue circles) and SEM (purple circles) and their fit with VFT (solid lines).}
\label{f_relax}
\end{center}
\end{figure*}
The empirical model assumes that the relaxation for $C$ and $\gamma$ occurring on the same time scale and connects their contribution to the ISS signal into a single stretched exponential function.\\
In order to experimentally verify whether the two response functions are indeed characterized by the same time scale and to what extent they can be disentangled, in Fig. \ref{f_relax} we compare the characteristic relaxation frequency of the heat capacity and of the thermal expansion coefficient (defined as $2\pi/\omega_C$ and $2\pi/\omega_\gamma$, respectively), both in the frame of Debye and Hn models, with the relaxation frequency $\Gamma_R$ of the SEM.
In the case of Debye model (Fig. \ref{f_relax} a), the heat capacity relaxation frequency (yellow squares) is systematically higher (about a factor of $1.5\pm 0.1$) than the one of the thermal expansion coefficient (blue diamonds).
This implies that after photothermally supplying energy, first heat is transferred from vibrational energy levels to configurational energy changes and, somewhat later,
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|}\hline
\multicolumn{3}{|c|}{Heat capacity}\\ \hline
&Debye&HN\\ \hline
$\log_{10}$($f_0/1$ Hz)&14.5 &14.5\\ \hline
$B$ (K)&2140 &2100\\ \hline
$T_0$ (K)&127 &124\\ \hline
$m$ & 50.9 &50.9\\ \hline
\multicolumn{3}{c}{}\\ \hline
\multicolumn{3}{|c|}{Thermal expansion coefficient}\\ \hline
&Debye&HN\\ \hline
$\log_{10}$($f_0/1$ Hz)&13.9 &13.9\\ \hline
$B$ (K)&2011 &2195\\ \hline
$T_0$ (K)&130 &125\\ \hline
$m$ & 54.1 &49.9\\ \hline
\multicolumn{3}{c}{}\\ \cline{1-2}
\multicolumn{2}{|c|}{SEM}&\multicolumn{1}{l}{}\\ \cline{1-2}
$\log_{10}$($f_0/1$ Hz)&14.8&\multicolumn{1}{l}{}\\ \cline{1-2}
$B$ (K)&2138&\multicolumn{1}{l}{}\\ \cline{1-2}
$T_0$ (K)&135&\multicolumn{1}{l}{}\\ \cline{1-2}
\end{tabular}
\end{center}
\caption{Summary of the parameters extracted by fitting with the VFT expression.
}
\label{table_VFT_parameters}
\end{table}
the configurational energy changes result in an increase of volume, in agreement with the two-temperature model developed in Section \ref{Sec:two_temp_model}.\\
Similar conclusions can be drawn from the results obtained by the HN model as shown in Fig. \ref{f_relax} (b). However, the results from HN model fitting are more dispersed due to the co-variance with the additional two fitting variables, namely a and b in Eq. \ref{C_omega_HN}.\\
The structural relaxation frequency $\Gamma_R$ (purple circles), obtained by fitting the experimental data with the SEM model, characterizes the (combined thermal and thermal expansion) structural relaxation and turn out to lie somewhat in between the other two relaxation frequencies.\\
The obtained temperature dependence of the relaxation frequencies were fitted to the Vogel-Fulcher-Tamman (VFT) equation (solid lines in Fig. \ref{f_relax}), defined by $f_{relax}=f_0\exp\left[-B/(T-T_0)\right]$, with $f_0$ the relaxation frequency in the high temperature limit, $T_0$ the Vogel-Fulcher temperature, around 130 K for glycerol.
In Table \ref{table_VFT_parameters} we report the fitted VFT parameters based on Debye, HN and SEM. The results for the latter model are in line with values reported in Ref. \cite{paolucci2000impulsive}.\\
From the ratio $D=B/T_0$ we have determined the so-called fragility $m$, via $m= 16+590/D$, which can be considered as a measure for the deviation from Arrhenius behavior (and thus as a measure for the degree of temperature dependence of the potential energy landscape morphology).
The fragility values obtained with our model are summarized in Table \ref{table_VFT_parameters} and are close to 53, the latter being the fragility for glycerol reported in Ref. \cite{bohmer1993nonexponential}.\\
\section{conclusion}
In this manuscript, a model to describe ISS signals generated in relaxing materials has been introduced, which is based on the solution of the thermal diffusion equation and the continuum mechanics equation, in combination with a frequency dependent heat capacity and thermal expansion coefficient. As functional forms for the frequency dependencies, Debye and Havriliak-Negami expressions were assumed. The assumption of a Debye frequency dependence of the heat capacity was shown to be compatible with a two-temperature model, in which the experimentally measured temperature refers to the energy distribution of the kinetic degrees of freedom, and a network temperature describes the state of the amorphous network, which is assumed to be in thermal contact with the kinetic energy reservoir.\\
The obtained physical models for describing ISS response, were shown to fit well ISS signals that had been simulated, for different temperature-wavenumber combinations in glycerol, by a semi-empirical model \cite{yang1995impulsive1} that has been historically used to describe ISS signals in relaxing materials.\\ Furthermore, we have carried out an experimental ISS investigation of glycerol under supercooling and also a comparative model fitting analysis based the physical models developed in this work and the existing empirical model. Satisfied fitting quality has been achieved for all ISS waveforms, confirming the models developed in this work and allowing us to study the relaxation of $C$ and $\gamma$, up to several tens of MHz, largely extending the upper limit of spectroscopy of thermal susceptibility, by nearly 3 and 7 decades for $C(\omega)$ and $\gamma(\omega)$, respectively. The best fit results also suggest that the relaxation of heat capacity and thermal expansivity occur on a slight different time scale, relaxation of $C$ is about 1.5 times faster than that of $\gamma$, which is line with the observation by a thermal lens spectroscopy investigation, reported in an accompanying article \cite{ThermalLens}.
|
\section{Introduction}
A $d$-dimensional space-filling curve is a continuous map
$\gamma\colon[0,1]\to\mathbb R^d$ having image with non-empty interior. The
first examples of such curves for $d=2,3$ were constructed by G.~Peano in 1890.
In the computer era space-filling curves have found many practical applications,
see~\cite{B13}. The point is that these curves allow one to efficiently order
multidimensional lattices, which is important in some computations.
The constructions of space-filling curves are based on the idea of
self-similarity. Let us consider curves with cubic image $[0,1]^d$. Suppose that
the cube is divided on a number of sub-cubes and the curve on each
sub-cube (``fraction'') is similar to the whole curve. Such curve is a \textit{monofractal} and
the number of fractions is curve's \textit{genus}.
In this paper we consider also multifractal curves. We say that a tuple of
curves is a \textit{multifractal}, if each fraction of each curve is similar to
one of the curves (the precise definition is given below). This class is rich
enough to contain many interesting examples, and simple enough to make
computations with it.
An important property of Peano curves is their H\"older continuity.
It is easy to see that space-filling curves $\gamma\colon[0,1]\to\mathbb{R}^d$ are not
$\alpha$-H\"older for $\alpha>1/d$, but fractal curves are indeed
$(1/d)$-H\"older. See~\cite{S10} for more details. So we have a natural measure
of curve smoothness, or ``locality'',
$$
\sup_{t_1<t_2}\frac{\|\gamma(t_2)-\gamma(t_1)\|^d}{t_2-t_1},
$$
that we call the ``dilation''.
In this paper we treat the case of the Euclidean norm $\|x\|=\|x\|_2$.
We introduce an algorithm for a search of extremal fractal curves in
large curve classes. It relies on SAT-solvers~--- heuristic algorithms that
find models for CNF boolean formulas. We used our algorithm to find some new
curves with small dilation.
The monofractal Peano curve of genus $2\times 2$ is unique (up to isometry).
This is the most popular plane--filling curve~--- the Hilbert curve. It has
dilation $6$, as proved by K. Bauman~\cite{B06}.
Peano monofractals of genus $3\times 3$ were studied in detail in the papers of
E.V.~Shchepin and K.~Bauman~\cite{SB08,B11}. It was proven that the minimal
dilation equals $5\frac23$ and the unique minimal diagonal curve was
found. The same curve was found independently by H.~Haverkort and
F.~Walderveen~\cite{HW10} and named ``Meurthe''.
On the other hand, E.V.~Shchepin~\cite{S04} and K.~Bauman~\cite{B14} proved that any
Peano monofractal has the dilation at least $5$.
The best known plane Peano multifractal is the $\beta\Omega$ curve, which has
dilation $5$.
One of our results is a new curve that we call ``YE'', a
monofractal plane curve of genus $5\times 5$ with dilation
$5\frac{43}{73}=5.5890\ldots$, which is best-known among plane
monofractals. Moreover, the YE-curve is the unique minimal
curve among monofractals of genus $\le 6\times 6$. We give a proof of
minimality, which relies both on computations and theoretical
results. This result (together with the algorithm) was announced in~\cite{MS20}.
In the 3D case much of the research was done by H.~Haverkort. We refer to his
papers~\cite{H11} and~\cite{H17}.
We consider classes of facet-gated multifractals, i.e. multifractals with
entrances and exits on facets. These classes are rigid enough to
allow an efficient search, and contain many curves with small dilation. In
dimension $3$ we have found facet-gated bifractals of
genus $2\times2\times 2$ with dilation less than $17$. We give a name ``Spring''
for this family of curves. In dimension $4$ we obtained that
there is a curve with dilation less than $62$. As we are aware, these
new curves are the best known (in terms of the dilation) in both dimensions $3$
and $4$.
Some lower bounds on the dilation for wider classes of cubically
decomposable curves are proved.
Let us outline the structure of the paper. In the next section we give
neccessary definitions. In Section~\ref{sec_results} we state the results of the
paper and compare them with known results, see Table~\ref{tbl_dil}. In
Section~\ref{sec_algo} we describe our algorithm, give a computer-assisted proof
of the minimality of the YE-curve and provide statistics of runs of our
algorithm. Section~\ref{sec_low} is devoted to the proofs of lower
bounds for the dilation. In the end of our paper we discuss possible directions of
further work.
\section{Definitions}
\paragraph{Dilation.}
Recall that a map $f$ between metric spaces $(T,\rho)$ and $(X,\sigma)$ is called
H\"older continuous with exponent $\alpha$ (or $\alpha$-H\"older, for
short) if for some $M$ the inequality holds
$$
\sigma(f(t_1),f(t_2))\le M\rho(t_1,t_2)^\alpha,\quad\forall t_1,t_2\in T.
$$
The smallest such $M$ is called H\"older coeffient, or H\"older seminorm (in case of
normed $X$), it is denoted as $|f|_{H^\alpha(T,X)}$.
Let us consider the $\ell_p^d$-norms in $\mathbb{R}^d$:
$\|x\|_p=(|x_1|^p+\ldots+|x_d|^p)^{1/p}$, $1\le p<\infty$,
$\|x\|_\infty=\max|x_i|$. It is convenient to raise the corresponding
$(1/d)$-H\"older seminorm of $d$-dimensional space-filling curves to the $d$-th power.
\begin{definition}
The $\ell_p$-dilation of a curve $\gamma\colon[a,b]\to\mathbb{R}^d$
is the quantity
\begin{equation}
\label{WD}
\WD_p(\gamma) := \sup_{a\le t_1<t_2\le b} \frac{\|\gamma(t_1)-\gamma(t_2)\|_p^d}{t_2-t_1}.
\end{equation}
\end{definition}
It is obvious that
$$
\WD_p(\gamma) = |\gamma|^d_{H^{1/d}([a,b];\ell_p^d)}.
$$
Let us look at a slightly more geometrical approach to measure curve locality.
We say that a curve $\gamma\colon[0,1]\to\mathbb{R}^d$ is
volume-parametrized, if for any $t_1<t_2$ we have
$\vol\gamma([t_1,t_2])=t_2-t_1$.
Define the $\ell_p$-dilation of a set $S\subset\mathbb{R}^d$ as
\begin{equation}
\label{dilation}
\dil_p(S):=\frac{\diam_p(S)^d}{\vol(S)}.
\end{equation}
It is easy to see that for a curve with volume parametrization we
have
$$
\WD_p(\gamma) = \sup_{t_1<t_2} \dil_p(\gamma([t_1,t_2])).
$$
The last quantity depends only on the scanning order, not on the parametrization
of the curve.
The quantity $\WD_p$ was proposed as a measure of curve quality in the
paper~\cite{GL96}. The notation $\WD_p$ is not well-established. Gotsman and
Lindenbaum in~\cite{GL96} intoduced notation $L_p(\gamma)$; Bader~\cite{B13}
writes $C_p$ for H\"older coefficient and $\widetilde{C}_p$ for its $d$-th power
(so, $\widetilde{C}_p=\WD_p$). Haverkort and Walderveen~\cite{HW10} suggested
the notation $\mathrm{WL}_p$ (Worst-case Locality) for~(\ref{WD});
Haverkort~\cite{H11} added the notation $\WD_p$ (Worst-case Diameter ratio) for
the equivalent quantity with diameter instead of distance. Later,
in~\cite{H17}, he called $\mathrm{WL}_p$ the $L_p$-dilation. We decided to
stick to the ``dilation'' term and the $\WD_p$ notation (Worst-case Dilation).
In this paper we consider only the Euclidean case, i.e. the Euclidean dilation
$\WD_2$, but our algorithms apply to any $p\in[1,\infty]$.
\paragraph{Peano multifractals.}
We say that curves $\gamma_1\colon I_1\to Q_1$ and $\gamma_2\colon I_2\to Q_2$
are \textit{isometric}, if there exist isometries $\alpha\colon I_1\to I_2$ and
$\beta\colon Q_1\to Q_2$, such that $\beta\circ\gamma_1 = \gamma_2\circ\alpha$.
Curves are \textit{similar} with a coefficient $s$, if $s\gamma_1(t/s^d)$ is
isometric to $\gamma_2(t)$.
Fix two parameters: a dimension $d\in\mathbb N$ and a scale $s\in\mathbb N$,
$s\ge 2$. The
unit cube $[0,1]^d$ is divided into the regular grid of $g:=s^d$ so-called
$s$-cubes
$$
\left[\frac{j_1}{s},\frac{j_1+1}{s}\right]\times\cdots
\times\left[\frac{j_d}{s},\frac{j_d+1}{s}\right].
$$
Moreover, we divide $[0,1]$ into $g$ equal segments $I_k^g:=[k/g,(k+1)/g]$. Given
a curve $\gamma$, each of the $g$ subcurves $\left.\gamma\right|_{I_k^g}$ is called
a \textit{fraction} of a curve.
\begin{definition}
A tuple of surjective curves $\gamma_1,\ldots,\gamma_m\colon [0,1]\to[0,1]^d$
forms a \textit{Peano multifractal} of genus $g=s^d$ if for all
$r=1,2,\ldots,m$ each of the $g$ fractions $\left.\gamma_r\right|_{I_k^g}$
maps $I_k^g$ to some $s$-cube $Q$ and is similar to one of the $\gamma_l$
with the coefficient $s$.
\end{definition}
The parameter $m$ is called \textit{the multiplicity} of a multifractal.
For $m=1$ we get a Peano monofractal, for $m=2$ a bifractal, etc. A
related but broader class of polyfractal curves was introduced in~\cite{S15} for
dimension $3$.
Components of Peano multifractals are volume-parametrized curves.
They are $(1/d)$-H\"older continuous and the maximum
in~(\ref{WD}) is attained. If a multifractal $(\gamma_1,\ldots,\gamma_m)$ is
irreducible, i.e. does not contain a non-trivial subset $(\gamma_i)_{i\in I}$ that is a
multifractal itself, then $\WD_p(\gamma_1)=\ldots=\WD_p(\gamma_m)$. For the
generic case, one may formally define the dilation of a multifractal as
$\max_j\WD_p(\gamma_j)$.
Let us give some more definitions. A Peano monofractal of genus $s^d$ defines
the order (of traversal) of $s$-cubes. The sequence of $s$-cubes in that order
is called a \textit{prototype}. Peano multifractal
$(\gamma_1,\ldots,\gamma_m)$ has $m$ prototypes, accordingly. We remark that a
prototype is a special case of a polycubic chain, see Section~\ref{subsect_poly}.
The \textit{entrance} of a curve $\gamma\colon[0,1]\to\mathbb{R}^d$ is the point $\gamma(0)$,
and the \textit{exit} is $\gamma(1)$. They are also called entrance and
exit \textit{gates}.
A \textit{pointed prototype} is a prototype with specified entrance/exit points for
each cube.
Let us give the definition, following~\cite{H17}.
\begin{definition}
A Peano multifractal $(\gamma_1,\ldots,\gamma_m)$ is called \textit{facet-gated} if all
entrances $\gamma_1(0),\ldots,\gamma_f(0)$ and exits $\gamma_1(1),\ldots,\gamma_m(1)$ lie
on the cube facets (i.e., on $(d-1)$-dimensional faces) and not on faces of lower dimension.
\end{definition}
Each fraction $\left.\gamma_r\right|_{I_k^g}$ of a multifractal
$(\gamma_1,\ldots,\gamma_m)$ of genus $g$ is similar to some of the $\gamma_l$,
$l=l_{r,k}$.
That similarity is defined by a pair of isometry of the cube
$[0,1]^d$ (space orientation) and isometry of $[0,1]$ (time orientation). We
will refer to this pair as a \textit{base orientation}. So, a
multifractal is completely defined by
$m$ prototypes, $mg$ base orientations and $mg$ numbers $l_{r,k}$.
Note that base orientations form a group isomorphic to
$\mathbb Z_2^d\times S_d\times \mathbb Z_2$. We remark that base orientations
are not completely defined by a multifractal, if it has symmetries.
Each curve in the multifractal of genus $g$ is divided into $g$ fractions; they
are called fractions of order $1$. If we divide each of the fractions into $g$
sub-fractions, we will obtain $g^2$ fractions of order $2$, etc.
Often we identify fractions (curves) with their images (cubes).
A \textit{junction} $F\to G$ of a curve is a pair of adjacent fractions of some order
$k$. We call $F\to G$ a \textit{facet junction} if $F$ and $G$ intersect by a
common facet . The derived junction $F'\to G'$ is
defined as the junction of $(k + 1)$-th order formed by the last
fraction of the subdivision of $F$ and the first fraction of the subdivision of
$G$. The \textit{depth} of a curve
(see~\cite{S04}) is the maximal order $k$ that generates new junctions.
An important observation made by E.~Shchepin~\cite{S15} is that the maximum in~(\ref{WD})
is attained at a pair of points from some non-adjacent sub-fractions $F_1\subset
F$, $G_1\subset G$ of some junction $F\to G$ (i.e., a pair from $F\times
G\setminus F'\times G'$, where $F'\to G'$ is the derived junction), or from some non-adjacent
first-order fractions $F_1$, $G_1$.
\paragraph{Cubically decomposable curves.}
We will require one more definition, from~\cite{KS18}.
\begin{definition}
Consider a surjective curve $\gamma\colon[0,1]\to[0,1]^d$ and a number
$g=s^d$. We say that $\gamma$ is decomposed into $g$ fractions if
each of the $g$ fractions $\left.\gamma_m\right|_{I_k^g}$
maps $I_k^g$ onto some $s$-cube $Q$.
The curve is called \textit{cubically decomposable} if it can be decomposed
into an arbitrarily large number of fractions.
\end{definition}
Cubically decomposable curves are volume-parametrized. Obviously, Peano
multifractal curves are cubically decomposable.
It is proven in~\cite{KS18} that cubically decomposable curves
have a recursive structure: there is some minimal number $s_1$ such that curve is
decomposable into $s_1^d$ first-order fractions; then there is some $s_2$ such
that each of the first-order fractions is decomposed into $s_2^d$ second-order
fractions, etc. The definition of a junction is the same as for multifractals: it
is a pair of order-$k$ adjacent fractions, for some $k$.
\section{Results}
\label{sec_results}
\subsection{Results on the $\ell_2$-dilation of monofractals.}
One of the main results of this paper is a Peano monofractal with
$\WD_2<5\frac23$.
\begin{theorem}
\label{th_YE}
There exists a unique (up to isometry) plane Peano monofractal curve
$\gamma\colon[0,1]\to[0,1]^2$ of genus $5\times5$ such that
$\WD_2(\gamma)=5\frac{43}{73}$. Moreover, it minimizes $\WD_2$ over
monofractals of genus less or equal $6\times 6$.
\end{theorem}
\begin{figure}[!h]
\begin{center}
\includegraphics[width=12cm]{YE.png}
\end{center}
\caption{YE minimal curve}
\label{fig_YE}
\end{figure}
The fractal curve that we have found is shown in the Figure~\ref{fig_YE}.
More precisely, the figure shows the central broken lines of its prototype
and prototypes of its fractions; that uniquely defines it. The prototype of
this curve resembles stucked together letters Y and E, so we suggest the name
YE-shaped for curves of this prototype. It is proved that the dilation of
the curve shown in the figure is minimal among monofractals of genus $\le36$.
So we will call it minimal YE-shaped curve, or just YE-curve.
\subsection{Computations for facet-gated multifractals.}
Recall that the best known plane Peano multifractal, $\beta\Omega$ curve,
has entrances and exits on the sides of the square $[0,1]^2$ and not in
vertices. So, facet-gated curves are candidates for small $\ell_2$-dilation.
Facet-gated curves are facet-continuous, i.e., adjacent fractions
intersect by facets. But the gate condition is, of course, more
\textit{rigid} (restrictive) and allows us to make an effective search among
such curves.
We suppose that facet-gated curves minimize $\ell_2$-dilation in many
cases. This is true for the class of 2D bifractals of genus $2\times 2$ and for 3D
monofractals of genus $2\times 2\times 2$. Let us consider 3D bifractals.
\begin{definition}
We call any $3D$ facet-gated Peano bifractal of genus $2\times2\times2$ having
$\WD_2<17$ a \textit{Spring} curve.
\end{definition}
The following computation shows that Spring curves exist.
\begin{computation}
\label{comp_spring}
The minimal $\ell_2$-dilation of $3D$ facet-gated Peano bifractals of genus
$2\times 2\times 2$ satisfies $16.9912 < \WD_2 < 16.9913$.
\end{computation}
Our hypothesis is that Spring curves are minimal among all
bifractals of genus $2^3$. Let us describe some properties of that curves.
Our calculations show that all Spring curves have dilation less than
$16.9913$; we suppose that their dilations are equal to
$1533\sqrt{6}/221$.
Moreover, all non-Spring facet-gated bifractals have $\WD_2 > 17.04$.
There are $9$ pointed prototypes containing Spring curves.
The entance/exit points and prototypes of Spring
curves are uniquely determined (up to isometry, of course). Spring bifractal
is a pair $(\gamma_1,\gamma_2)$; for short, we denote the first curve as (a) and
the second one as (b).
Entrance of (a) is $(1/3,1/3,0)$ and exit is $(2/3,0,1/3)$;
entrance of (b) is $(1/3,1/3,0)$ and exit: $(1,1/3,1/3)$.
To describe prototypes for (a) and (b), we use \textit{chain codes}
(see~\cite{KS18}). Fix a standard basis $\{\texttt{i,j,k}\}$ in $\mathbb{R}^3$ and let
$\texttt{I}=-\texttt{i}$, $\texttt{J}=-\texttt{j}$, $\texttt{K}=-\texttt{k}$.
\begin{figure}[h]
\begin{center}
\includegraphics[width=2cm]{ijk.png}
\end{center}
\caption{Standard basis $\texttt{i,j,k}$.}
\end{figure}
Then, we can define a sequence of cubes $Q_1,\ldots,Q_{g}$ by sequence of vectors
$\delta_l$ such that $Q_{l+1}=Q_l+\delta_l$. For facet-continuous curves that
sequence may be represented as a word in the alphabet
$\{\texttt{i,j,k,I,J,K}\}$, called chain code.
In our case, the chain code for prototype (a) is \texttt{jikIJiK}
and for prototype (b) is \texttt{jkJijKJ}.
To give an example of a Spring curve, we have to specify base orientations and (a)/(b) letters
for each of the fractions.
A space orientation (an isometry of $[0,1]^3$) is completely defined by
its linear part, which is determined by images of basis vectors. Say,
$\texttt{KiJ}$ is the isometry with linear part that maps $\texttt{i}\to
\texttt{K}$, $\texttt{j}\to\texttt{i}$, $\texttt{k}\to\texttt{J}$. A time
orientation is either identity or time-reversal; we denote by $(\overline{a})$
time-reversed curve $(a)$ and similarly for $(b)$.
In our example we have the following fractions (see~Fig.~\ref{fig_spring}):
(a): \texttt{KIJ}($\overline{b}$),
\texttt{KIj}(a), \texttt{kji}(b), \texttt{Jki}($\overline{a}$),
\texttt{JkI}(b), \texttt{kIJ}(a), \texttt{KJi}(b), \texttt{JiK}(b).
(b): \texttt{KIJ}($\overline
b$), \texttt{ijK}($\overline{a}$), \texttt{Ikj}($\overline{a}$),
\texttt{KJI}($\overline{a}$), \texttt{kiJ}($\overline{a}$),
\texttt{Ijk}($\overline{a}$), \texttt{IjK}(a), \texttt{iKJ}(b).
\begin{figure}[h]
\begin{center}
\includegraphics[width=12cm]{bi17.png}
\end{center}
\caption{Example of a Spring curve.}\label{fig_spring}
\end{figure}
\vspace{5pt}
We also have a result in the $4D$ case.
\begin{computation}
The minimal $\ell_2$-dilation of $4D$ facet-gated Peano monofractals of genus
$2^4$ safisties $61.935 < \WD_2 < 61.936$.
\end{computation}
Let us give an example of such curve. We use basis $\{\texttt{i,j,k,l}\}$.
Prototype: \texttt{kljKLkiKlkJLKlI}.
Base orientations:
\texttt{iKLJ}(a),
\texttt{kLij}(a),
\texttt{lJki}(a),
\texttt{jklI}(a),
\texttt{lKjI}($\overline{a}$),
\texttt{LKIj}(a),
\texttt{IkLj}($\overline{a}$),
\texttt{ikLj}(a),
\texttt{LKij}($\overline{a}$),
\texttt{lKji}(a),
\texttt{kjlI}(a),
\texttt{lJkI}($\overline{a}$),
\texttt{LkIJ}(a),
\texttt{LKji}($\overline{a}$),
\texttt{iljk}($\overline{a}$),
\texttt{jIlK}($\overline{a}$).
Entrance is $(0,1/3,1/3,1/3)$ and exit is $(1/3,0,1/3,2/3)$.
The dilation is at least $61\frac{29}{31}$ and we suppose that this is sharp.
\subsection{Lower bounds on the dilation.}
The proof of the YE-curve optimality relies on the following result.
\begin{theorem}
\label{th_diag}
If a plane Peano monofractal curve $\gamma$ has a diagonal junction, then
$\WD_2(\gamma)\ge5\frac23$.
\end{theorem}
We also prove some lower bounds on the dilation for cubically decomposable
curves.
\begin{theorem}
\label{th_poly2}
There exists some $\varepsilon>0$, such that for any plane cubically decomposable
curve $\gamma\colon[0,1]\to[0,1]^2$ one has $\WD_2(\gamma)\ge 4+\varepsilon$.
\end{theorem}
Note that the bound $\WD_2\ge4$ (and even stronger bound $\WD_\infty\ge 4$) was
proven in~\cite[Th.4]{HW10}.
\begin{theorem}
\label{th_poly3}
Let $\gamma\colon[0,1]\to[0,1]^3$ be a cubically decomposable curve. Then
$\WD_2(\gamma)\ge 7\sqrt{14}/2>13.09$. If $\gamma$ has a non-facet junction, then
$\WD_2(\gamma)\ge 17.17$.
\end{theorem}
From this Theorem and Computation~\ref{comp_spring} we obtain the following.
\begin{corollary}
$3D$ Peano bifractals of genus $2\times 2\times 2$ with minimal
$\ell_2$-dilation have only facet junctions.
\end{corollary}
We conjecture that minimal bifractals are facet-gated, but we do not have a
proof.
\begin{theorem}
\label{th_poly4}
Let $\gamma\colon[0,1]\to[0,1]^4$ be a cubically decomposable curve. Then
$\WD_2(\gamma)\ge 42.25$. If $\gamma$ has a non-facet junction, then
$\WD_2(\gamma)\ge 62.25$.
\end{theorem}
\begin{corollary}
$4D$ Peano monofractals of genus $2^4$ with minimal $\ell_2$-dilation
have only facet junctions.
\end{corollary}
Finally, we prove a bound on the dilation for arbitrary surjective curves in the
4D case.
\begin{theorem}
\label{th_4dsurj}
For any surjective curve $\gamma\colon[0,1]\to[0,1]^4$ we have
$\WD_2(\gamma)\ge 30$.
\end{theorem}
Table~\ref{tbl_dil} summarizes known results on the minimal $\ell_2$-dilation of
surjective curves~$\gamma\colon[0,1]\to[0,1]^d$. In the last column we normalize the
curve dilation by the dilation of the unit cube; this allows to compare
dilations across different dimensions and metrics.
\begin{table}
\begin{tabular}{|l|l|l|l|l|l|}
\hline
d & \textbf{genus} & \textbf{curve type} & \begin{tabular}{@{}l@{}}\textbf{minimal
curve}\\\textbf{and references}\end{tabular} & $\WD_2$ &
$\frac{\WD_2^{1/d}}{d^{1/2}}$ \\
\hline
\hline
2 & $2\times 2$ & monofractal & Hilbert~\cite{B06} & 6 & 173\% \\
& $3\times 3$ & & Meurthe~\cite{SB08} & $5\frac23$ & 168\% \\
& $\le 6\times 6$ & & YE ($5\times 5$) & $5\frac{43}{73}$ & 167\% \\
\hhline{|~|-----|}
& any & monofractal &
\begin{tabular}{@{}l@{}}\cite{S04,B14}\\best known: YE\end{tabular} &
$\ge 5$ & $> 158\%$ \\
\hhline{|~|-----|}
& $2\times 2$ & bifractal &
\begin{tabular}{@{}l@{}}Wierum's $\beta\Omega$\\\cite{W02,HW10}\end{tabular}
& 5.000 & 158\% \\
\hhline{|~|-----|}
& \multicolumn{2}{l|}{any cubically decomposable} &
\begin{tabular}{@{}l@{}}Theorem~\ref{th_poly2}\\best known: $\beta\Omega$\end{tabular} &
$>4$ & $>141\%$ \\
\hhline{|~|-----|}
& \multicolumn{2}{l|}{special triangle fractal} & Serpinski~\cite{NRS02} & 4 & 141\% \\
\hhline{|~|-----|}
& \multicolumn{2}{l|}{any surjective curve} &
\begin{tabular}{@{}l@{}}\cite[Theorem 2]{MS21}\\best known: Serpinski\end{tabular} &
$>3\frac58$ & $> 134\%$ \\
\hline
\hline
3 & $2\times 2\times 2$ & monofractal & Haverkort's Beta~\cite{H17} & 18.566 & 153\% \\
\hhline{|~|-----|}
& $2\times 2\times 2$ &
\begin{tabular}{@{}l@{}}facet-gated\\bifractal\end{tabular} &
Spring (Computation 1) & 16.991 & 148\% \\
\hhline{|~|-----|}
& \multicolumn{2}{l|}{any cubically decomposable} &
\begin{tabular}{@{}l@{}}Theorem~\ref{th_poly3}\\best known: Spring\end{tabular} &
$> 13.095$ & $> 136\%$ \\
\hhline{|~|-----|}
& \multicolumn{2}{l|}{any surjective} &
\begin{tabular}{@{}l@{}}\cite{NRS02}\\best known: Spring\end{tabular} &
$\ge 11.1$ & $> 128\%$ \\
\hline
\hline
4 & $2^4$ &
\begin{tabular}{@{}l@{}}facet-gated\\monofractal\end{tabular} &
Computation 2 & 61.93 & 140\% \\
\hhline{|~|-----|}
& \multicolumn{2}{l|}{any cubically decomposable} &
\begin{tabular}{@{}l@{}}Theorem~\ref{th_poly4}\\best known: Comp. 2\end{tabular} &
$\ge 42.25$ & $> 127\%$ \\
\hhline{|~|-----|}
& \multicolumn{2}{l|}{any surjective} &
\begin{tabular}{@{}l@{}}Theorem~\ref{th_4dsurj}\\best known: Comp.
2\end{tabular} & $\ge 30$ & $> 117\%$ \\
\hline
\end{tabular}
\caption{Known results on $\ell_2$-dilation}\label{tbl_dil}
\end{table}
\section{Search algorithm and implementation}
\label{sec_algo}
\subsection{Algorithm}
The algorithms in this section work for arbitrary Peano multifractals, but for
simplicity we describe them for monofractals.
Recall that the prototype of a Peano monofractal of genus $s^d$ is the sequence
of $s$-cubes in the curve scanning order and the pointed prototype is th prototype
with specified entry/exit points for each cube.
To define a fractal curve with given pointed prototype, you must specify valid base
orientations for all fractions, i.e., such orientations that transform
entrance/exit points as required (note that time reversal swaps entrance and
exit). An important property of that structure is
that the choices for different fractions are independent.
Our computations are based on the algorithm for the estimation of the minimal dilation
within a class of curves with given pointed prototype.
Let us give an example to show the effeciency of such approach.
Consider curves of genus $5\times 5$ with entrance $(0,0)$ and exit $(1,1)$.
Fix some pointed prototype. It is clear that there are
four possible base orientations (two with time reversal and two without) in each
fraction. This gives $2^{50}$ possible curves. Our algorithm allows us to
consider all that curves simultaneously.
\paragraph{Bisection of the curve dilation.}
It is advisable to consider the case of completely defined curve first.
We describe an algorithm that, given a monofractal curve $\gamma$ and thresholds
$\omega_-<\omega_+$,
either outputs pair of points on a curve providing $\WD_p(\gamma)\ge\omega_-$, or guarantees
that $\WD_p(\gamma)\le\omega_+$.
The dilation of the curve is equal to the maximum of the dilation for all
possible pairs of non-adjacent sub-fractions $F\subset F^\circ,G\subset G^\circ$
of curve junctions $F^\circ\to G^\circ$, or non-adjacent first-order fractions
$F,G$:
$$
\WD_p(F,G) := \max\left\{\frac{\|\gamma(t_1)-\gamma(t_2)\|_p^d}{|t_2-t_1|} \colon \gamma(t_1)\in
F,\;\gamma(t_2)\in G\right\}.
$$
There is a finite number of such pairs (becase the number of distinct junctions
is finite), and for each pair we can get upper $U$ and lower $L$ bounds for the
quantity $\WD_p(F,G)$:
$$
U = \frac{\max\{\|\gamma(t_1)-\gamma(t_2)\|_p\colon \gamma(t_1)\in F,\;\gamma(t_2)\in G\}}
{\min\{|t_2-t_1|\colon \gamma(t_1)\in F,\;\gamma(t_2)\in G\}},
$$
$$
L = \frac{\max\{\|\gamma(t_1)-\gamma(t_2)\|_p\colon \gamma(t_1)\in F,\;\gamma(t_2)\in G\}}
{\max\{|t_2-t_1|\colon \gamma(t_1)\in F,\;\gamma(t_2)\in G\}}.
$$
Three cases are possible:
\begin{itemize}
\item[(A)] If $L\ge\omega_-$, then $\WD_p(\gamma)\ge\omega_-$ and the work is done.
\item[(B)] If $U\le\omega_+$, then we can throw away (do not consider further) that pair.
\item[(C)] If $L < \omega_- < \omega_+ < U$, then we subdivide both fractions and
continue to evaluate pairs of sub-fractions.
\end{itemize}
Eventually, $U-L$ will be small enough to avoid the case (C). So, we either will
get (A), or we will throw away all pairs~--- it will mean that
$\WD_p(\gamma)\le\omega_+$.
\paragraph{Bisection of the minimal dilation for a pointed prototype.}
Now we describe an algorithm that, given a pointed prototype and thresholds
$\omega_-<\omega_+$, either guarantees that for all curves with that pointed prototype
$\WD_p(\gamma)\ge \omega_-$, or finds a curve with $\WD_p(\gamma)\le \omega_+$.
The difference from the previous algorithm is the following: in the case (C), when
subdividing fractions, we need to know the base orientations
on the corresponding fractions. Therefore, this step is
branching in accordance with all
possible base orientations. Next, at the case (A), when we arrive at a configuration with
$L\ge\omega_-$, we cannot state that $\WD_p(\gamma)\ge\omega_-$ for all curves in the class.
Instead, this only gives us a ``ban'' on all curves compatible with
orientations we selected in (C) that led to this configuration.
How to check that the forbidden configurations obtained at (A) cases exhaust all
curves of the class? We can describe the curves using boolean variables. Namely,
for each index of the fraction $i$ and for each valid base orientation $b$ we
introduce a boolean variable $x_{i,b}$, which is true if and only if the base
orientation at fraction $i$ is $b$. Our restrictions can be written as a CNF
formula over these boolean variables. If this formula is unsatisfiable, it
means that all curves have the dilation at least $\omega_-$. Otherwise, if all
pairs are processed and the resulting formula has a model, then it will give us
a curve with $\WD_p\le\omega_+$.
The satisfiability problem of Boolean formulas (SAT-problem) is well studied
(see~\cite{BHM09}). There are many software libraries available to solve SAT tasks
(so-called SAT-solvers). Because the SAT belongs to the NP class, there is
no proven polynomial estimates for SAT-solvers.
The empirical fact we observed is that \textit{modern SAT-solvers work well on CNF
formulas, arising in the problem of finding optimal curves}.
\paragraph{Technical details.}
Let us give some more details of our algorithm that may have significant impact
on it's performance. When we subdivide pairs of fractions, we pick the
pair with the highest upper bound on the dilation. And when we do the subdivision,
we divide only one fraction in a pair (and keep pairs ``balanced'', i.e. the
depth of subdivision differs at most by $1$). If a pair falls into both
cases (A) and (B), i.e. $\omega_- \le L < U \le \omega_+$, then we choose (B),
i.e. the pair is discarded (not to enlarge CNF formula).
The SAT-solver is used in the simplest way: every time we call the solve method,
we rebuild the CNF formula in it from scratch. To reduce computational costs of
this, we do this every $\lfloor1.3^k\rfloor$-th subdivision step.
To estimate the minimal dilation of a pointed prototype we use the bisection
algorithm many times. Suppose that we know some lower and upper bounds $(L,U)$ on the
minimal dilation. We put $\omega_- := \frac23L+\frac13U$ and $\omega_+ :=
\frac13L+\frac23U$. If there is a curve with $\WD\le\omega_+$, then the bounds
are updated as $L':=L$, $U':=\frac13L+\frac23U$. If every curve has
$\WD\ge\omega_-$, then $L'=\frac23L+\frac13U$, $U'=U$.
We estimate the minimal dilation for a sequence of pointed prototypes in
several ``epochs''. In the first epoch we estimate the minimal dilation of each
pointed prototype with some relative error $\varepsilon$ and keep only possibly
minimal prototypes. In the second epoch the relative error is taken smaller,
e.g. $\varepsilon'=\varepsilon/4$. And so on.
\subsection{Computer-assisted proof of Theorem~\ref{th_YE}}
In this section we will give a computer-assisted proof of Theorem 1 that
consists of two steps: finding the exact value of the dilation and proving that
it is minimal.
Our main tool that allows to find the exact value of the dilation is the following result.
\begin{theoremA}[\cite{SB08}, Theorem 7]
If a plane Peano monofractal curve $\gamma$ of genus $g$ satisfies
$\omega_\infty<\omega_2$, where $\omega_p:=\WD_p(\gamma)$,
then the $\ell_2$-dilation of $\gamma$ is attained on a
pair of vertices of fractions of order $k$, where
\begin{equation}
\label{depth}
k < \mathrm{depth}(\gamma)+3 +\log_g\frac{\omega_2^2}{4(\omega_2-\omega_\infty)}.
\end{equation}
\end{theoremA}
Recall that the depth of a curve is the largest number of the curve subdivision
that contains a junction that is not similar to junctions of less order. The minimal
YE-curve $\gamma_{\mathrm{YE}}$, see Fig.1, has depth $1$ because derivatives of
it’s junctions of first subdivision are similar to that junctions. For our
curve we have $g=25$, $\omega_2=5\frac{43}{73}$ and $\omega_\infty=5\frac13$.
Hence from~(\ref{depth}) it follows that $k\le 5$ and we can calculate exact
value of $\WD_2$ using 5-th subdivision only.
The rigorous proof of the equality
$\omega_2=\WD_2(\gamma_{\mathrm{YE}})=5\frac{43}{73}=5.5890...$ goes in the
following way: first, we compute approximations for $\omega_2$ and
$\omega_\infty$ and prove that they differ from presumed
values $5\frac{43}{73}$ and $5\frac13$, say, less than $0.01$. These approximate
computations allow us to estimate the logarithm in~(\ref{depth}) and to find the
value of dilation using 5-th subdivision.
Let us describe the computations that show that YE curve has minimal
$\ell_2$-dilation among all Peano monofractal curves of genus $\le 6\times 6$.
For brevity we will write $\WD=\WD_2$.
As we already mentioned, there is one $2\times2$ curve (Hilbert curve), it has
$\ell_2$-dilation $6$; the minimal $3\times 3$ curve has $\WD=5\frac23$.
Our computations show that minimum $\ell_2$-dilation of plane monofractal curves
of genus $4\times 4$ is greater than $5.9$.
It was proven in~\cite{B14} that plane Peano monofractal curves fall into three categories:
\begin{itemize}
\item side: curves with entrance and exit in the vertices of the same side;
\item diagonal: curves with entrance and exit in the opposite vertices;
\item median: curves with entrance in some vertex and exit in the middle of
the opposite side (or vice versa).
\end{itemize}
In the cases of genus $5\times5$ or $6\times 6$ we have processed only curves without
diagonal steps in prototype; other curves have diagonal junctions and by
Theorem~\ref{th_diag} they satisfy $\WD\ge5\frac23$. Our computations provide the following estimates:
median $6\times6$ curves have $\WD>6.5$; side $6\times6$ curves have $\WD>5.6$;
diagonal $6\times6$ curves do not exist; median $5\times 5$ curves have
$\WD>6.5$; diagonal $5\times 5$ curves have $\WD>5.9$.
The remaining case is side $5\times5$ curves.
Such curves with non-YE prototype have dilation at least $5.7$.
There are $2^{25}$ YE-shaped curves. To prove that our curve is minimal,
for each of the $25$ fractions of the YE-prototype we have fixed
the base orientation that does not match our curve. It turned out that for
curves with such constraints $\WD>5.6$.
\subsection{Software and running stats}
Our code is a Python package \texttt{peano} which can be found on the GitHub,~\cite{gitYM}.
We have used the library Glucose3~\cite{ALS13} and the \texttt{python-sat} (PySAT) wrapper~\cite{IMM18}
for SAT-solving. All calculations are based on rational arithmetic, with
\texttt{quicktions} implementation.
The \texttt{peano} package implements Peano multifractals of
arbitrary dimension, multiplicity and genus. Let us list main capabilities of our library:
\begin{itemize}
\item define Peano multifractals in terms of prototypes and base
orientations;
\item find entrance and exit points;
\item find vertex moments: $\{t\colon \gamma(t)\in\{0,1\}^n\}$ and,
more generally, first/last face moments for cube faces of
arbitrary dimension;
\item generate entrance/exit configurations given dimension, multiplicity
and genus;
\item effectively generate all entrance/exit configurations on facets;
\item generate pointed prototypes given entrance/exit configuration;
\item estimate dilation of a Peano multifractal for any coordinate-monotone
norm (possibly using face moments);
\item estimate minimal dilation for a sequence of pointed prototypes.
\end{itemize}
The code was thoroughly tested on a set of known Peano curves. The main source
of curves in the 3D case was the paper~\cite{H11} with many detailed
descriptions. We have used also~\cite{HW10} for a bunch of 2D curves
and~\cite{KS18} for the Tokarev curve.
In Table~\ref{tbl_run} we give some basic statistics for three search runs:
Computation 1, the search of a minimal curve among facet-gated 3D bifractals of
genus $2\times2\times2$; Computation 2, the search among facet-gated 4D monofractals of genus
$2\times2\times2\times2$; and the proof of minimality of YE curve, Theorem~\ref{th_YE}.
The commands to perform these runs are given in the README file of the package.
Let us give some comments on Table~\ref{tbl_run}.
The row ``Peano curves'' contains the numbers of all Peano curves with pointed
prototypes that were processed during the search. This number is rather small
for the Computation 1; it seems that the Spring curve can be found without SAT
solvers at all. The number of curves in the YE proof is huge; main contribution
comes from $11$ diagonal $5\times 5$ pointed prototypes (without diagonal steps), each
of them contains $4^{25}$ curves.
The size of a SAT problem for a CNF formula is determined by three parameters: number of variables,
number of clauses in the formula and total number of literals in clauses. We
show the SAT problem having the most literals.
\begin{table}
\begin{tabular}{|l|l|l|l|}
\hline
& Comp. 1 & Comp. 2 & YE proof \\
\hline
Gate configurations & 35 & 3 & 3 \\
Pointed prototypes & 909 & 1761 & 2763 \\
Peano curves & 122955 & $115\times10^6$ & $1.24\times10^{16}$ \\
\hline
Bisect steps & 3301 & 8345 & 8472 \\
Pairs of fractions & $10.4\times10^6$ & $143\times10^6$ & $204\times10^6$ \\
Banned pairs & 40840 & 815466 & $1.41\times10^6$ \\
\hline
SAT solve calls & 35764 & 98164 & 47984 \\
SAT largest problem: & & & \\
\hspace{5pt} literals & 19515 & 7825 & 480792 \\
\hspace{5pt} clauses & 7019 & 2405 & 82918 \\
\hspace{5pt} variables & 819 & 275 & 1713 \\
\hline
Running time & 4 min & 45 min & 1 h 20 min \\
\hline
\end{tabular}
\caption{Running stats}\label{tbl_run}
\end{table}
\section{Lower bounds}
\label{sec_low}
In this section we denote $|x|:=\|x\|_2$ for brevity.
Given two points $\gamma(t_1)$, $\gamma(t_2)$ on a curve, we define
the \textit{cube-to-linear ratio} as
$|\gamma(t_1)-\gamma(t_2)|^d/|t_2-t_1|$. The curve dilation equals
the supremum of cube-to-linear ratios over all pairs.
\subsection{Proof of Theorem~\ref{th_diag}}
\def\ref{th_4dsurj}{\ref{th_diag}}
\begin{theorem}
If a plane Peano monofractal curve $\gamma$ has a diagonal junction, then
$\WD_2(\gamma)\ge5\frac23$.
\end{theorem}
Suppose that consecutive fractions $F$ and $G$ of the curve $\gamma$ form a diagonal
junction. Let us put the transition point from $F$ to $G$ to the origin
$O=(0,0)$ and suppose that $G$ lies in the first, positive quadrant (right and above $O$)
and $F$ lies in the third, negative quadrant (left and below $O$). We also consider
fractions $F_1$ and $G_1$ of the next subdivision, such that $F\supset F_1\ni O$
and $G\supset G_1\ni O$. Moreover, let $G_2,G_3,G_4$ be consecutive fractions
that go after $G_1$ and let $F_4,F_3,F_2$ be consecutive fractions before $F_1$
(so, $F_2$ meets $F_1$). For convenience, we assume that $F_i$ and $G_i$ have
unit side length (and passed in the unit time). See Figure~\ref{fig_lem1} for an example.
We will treat several cases, illustrated in Figure~\ref{fig_cases}. Some of them
will require the following Lemma.
\begin{lemma}
\label{lem_AB}
Let $A$ be the left bottom vertex of $F_i$, and $B$ be the right upper
vertex of $G_j$ and the conditions hold true:
\begin{itemize}
\item $i+j=6$,
\item the distance from $A$ to the entrance of $F_i$ is at least
$1$;
\item the distance from $B$ to the exit of $G_j$ is at least
$1$;
\item $|A-B|\ge\sqrt{32}$.
\end{itemize}
Then $\WD_2(\gamma)\ge 5\frac23$.
\end{lemma}
\begin{figure}[!h]
\begin{center}
\includegraphics{th2_lem1.png}
\end{center}
\caption{Lemma~\ref{lem_AB}, $i=j=3$ (example).}
\label{fig_lem1}
\end{figure}
\begin{proof}
Suppose that $\WD_2(\gamma)\le 5\frac23$. Then the curve spends
at least $\frac{3}{17}$ units of time to pass one unit of length.
So the time interval between $A$ and $B$ is
at most $6-\frac{6}{17} = \frac{6\cdot 16}{17}$ and the cube-to-linear ratio for $(A,B)$ is at least
$$
\frac{|A-B|^2}{t_B-t_A} \ge \frac{32\cdot 17}{6\cdot 16}=\frac{17}3=5\frac23.
$$
\end{proof}
Let us return to the proof of the theorem.
\paragraph{Diagonal curves.}
If $\gamma$ is diagonal, then the farthest (from origin) point $A$ of $F_3\cup F_4$ has
$\|A-O\|_1 \ge 5$. The same holds for the farthest point $B$ of $G_3\cup G_4$ and hence
$\|A-B\|_1\ge 10$, so $|A-B|^2 \ge 5^2+5^2=50$. The time interval from $A$
to $B$ is at most $8$, so the cube-to-linear ratio for this pair is at least
$6\frac14$.
\paragraph{Median curves.}
If four fractions $G_1\ldots G_4$ do not lie in a $2\times 2$ square, then the
$\ell_1$-distance from the farthest point $B$ of $G_1\cup\ldots\cup G_4$ to $O$ is
at least $5$. And in any case the $\ell_1$-distance from the farthest point $A$
of $F_1\cup F_2\cup F_3$ to $O$ is at least $4$. Hence $\|A-B\|_1\ge 9$ and
$|A-B|^2\ge 4^2+5^2=41$. The time interval from $A$ to $B$ is at most $7$, so
the cube-to-linear ratio is $\ge\frac{41}{7}>5\frac23$. If fractions $F_1\ldots F_4$
do not lie in a $2\times 2$ square, the same arguments apply.
Finally, if both quads $F_1\ldots F_4$ and $G_1\ldots G_4$ lie in $2\times 2$ squares,
then the dilation bound follows from the Lemma~\ref{lem_AB} with $i=j=3$.
\begin{figure}
\includegraphics[width=15cm]{th2_cases.png}
\caption{Cases in the proof of Theorem~\ref{th_diag}.}\label{fig_cases}
\end{figure}
\paragraph{Side curves.}
If the directions (from entrance to exit) of consecutive fractions coincide,
then we say that the corresponding junction \textit{keeps the direction}.
We will start with the case when $\gamma$ has no diagonal junctions that keep the
direction.
W.l.o.g. let $G_1$ has the horizontal direction, so $F_1$ is vertical. It follows
that $F_2$ is below $F_1$ and $G_2$ is to the right from $G_1$.
\textit{Case 1}. $G_2$ is horisontal. Then $G_3$ is to the right from $G_2$. Let
us consider fraction $F_3$: it is below $F_2$ or to the left from $F_2$; it
cannot be to the left from $F_1$ because in that case $F_3\to F_2$ will be a
diagonal junction that keeps the direction. In both cases we can apply
Lemma~\ref{lem_AB} for $F_3$ and $G_3$.
If $F_2$ is vertical, the reasoning is the same as in the Case 1.
\textit{Case 2}. $G_2$ is vertical, and $F_2$ is horizontal. Then $G_3$ must be
above $G_2$ and $F_3$~--- to the left from $F_2$. We apply Lemma~\ref{lem_AB}
for $F_3$ and $G_3$.
So, it remains to consider the case of a side curve with a diagonal junction keeping
direction. We may assume that the junction $F\to G$ keeps the direction.
Denote by $\tau_1(\gamma),\tau_2(\gamma)$ and $\tau_3(\gamma)$ the time
intervals from the curve entrance to the next visited vertex, then to the third
vertex and, finally, from it to the exit. We say that a curve is
\textit{accelerating} if $\tau_1(\gamma)>\tau_3(\gamma)$ and is
\textit{decelerating} if $\tau_1(\gamma)<\tau_3(\gamma)$. We will apply these
definitions to the fractions of $\gamma$.
\begin{lemma}
\label{lem_accelerate}
If a diagonal junction $F\to G$ keeps the direction and
$\WD_2(\gamma)\le5\frac23$, then the first fraction $F$ is decelerating and
the second, $G$, is accelerating. Moreover, we have
$$
\tau_1(F)=\tau_3(G)\le\frac5{17}<\frac{6}{17}\le\tau_3(F)=\tau_1(G).
$$
\end{lemma}
\begin{proof}
Let us consider first the case when vertices are visited along the edges, as
in Figure~\ref{fig_cases}.
Denote $\tau_1:=\tau_1(F)$, $\tau_3:=\tau_3(F)$, $\tau_1':=\tau_1(G)$,
$\tau_3':=\tau_3(G)$.
Condition $\WD_2(\gamma)\le 5\frac23$ gives us inequalities
$\tau_3 + \tau_1' \ge \frac{4}{5\frac23} = \frac{12}{17}$
and
$2-\tau_1-\tau_3' \ge \frac{8}{5\frac23} = \frac{24}{17}$,
so
$$
\tau_1+\tau_3' \le \frac{10}{17} < \frac{12}{17} \le \tau_3 + \tau_1'.
$$
This excludes case $\tau_1=\tau_1'$, $\tau_3=\tau_3'$. Since $\gamma$ is a
monofractal, the only possible case is
$$
\tau_1=\tau_3' \le \frac{5}{17} < \frac{6}{17} \le \tau_3=\tau_1'.
$$
Now, suppose that the first and the second vertex visited in $F$
lie diagonally opposite of each other (and therefore, so do the third and
the fourth vertex, that is, $O$). Then we have
$\tau_3+\tau_1'\ge\frac{24}{17}$ and $\tau_1\ge\frac{6}{17}$. If
$\tau_1=\tau_1',\tau_3=\tau_3'$, then
$\tau_3+\tau_1'+\tau_3'+\tau_1\ge\frac{48}{17}>2$, which is impossible. If
$\tau_1=\tau_3',\tau_3=\tau_1'$, then $\tau_3\ge\frac{12}{17}$ and
$\tau_1+\tau_3\ge\frac{18}{17}$, which is also impossible. So, this case
cannot take place.
\end{proof}
From Lemma we derive that the junction $F\to G$ is centrally-symmetric. Since
our junction keeps the direction, we may assume that $F_1$ and $G_1$ are both
horizontal.
If $G_2$ is also horizontal, then $F_2$ has the same direction and the squared
diameter of $F_3\cup\ldots\cup G_3$ is at least $40$, which gives
$\WD_2(\gamma)\ge\frac{40}6$.
If $G_2$ is vertical, then $\gamma$ runs the distance at least $2$ in time $1+\tau$,
where $\tau:=\tau_1(G_2)$. Since the whole $F\to G$ junction is
centrally-symmetric, $\gamma$ runs the distance $\ge 4$ in time $2+2\tau$. Hence
$\WD_2(\gamma)\ge \frac{16}{2+2\tau}$. If $\WD_2(\gamma)\le 5\frac23$, then it gives
$\tau\ge\frac{7}{17}$. Lemma~\ref{lem_accelerate} implies that $G_2$ is
accelerating. It also follows that junction $G_2\to G_3$ cannot be diagonal, so
$G_3$ is above $G_2$. Finally, application of Lemma~\ref{lem_AB} for $F_3$ and
$G_3$ finishes the proof.
\subsection{Polycubic chains}
\label{subsect_poly}
In this section we will prove theorems~\ref{th_poly2},~\ref{th_poly3} and~\ref{th_poly4}.
\begin{definition}
\label{def_poly}
A sequence $Q_1,\ldots,Q_n$ of distinct $d$-dimensional cubes in
$d$-dimensional Euclid space if called a \textit{polycubic chain} if in some
coordinates we have $Q_1=[0,1]^d$ and $Q_{i+1}=Q_i+\delta_i$, $1\le i\le
n-1$, for some translation vectors $\delta_i\in\{-1,0,1\}^d$.
\end{definition}
The \textit{length} of a chain is the number of elements (cubes) in it. Elements of a
chain are called chain's \textit{fractions}. The union $Q_1\cup
Q_2\cup\ldots\cup Q_n$ of all fractions is
called the \textit{body} of a chain. A subsequence of cubes $Q_i,\ldots,Q_{i+k}$
is called a \textit{segment} of a chain (of course, it is a polycubic chain
itself). A \textit{junction} if a pair of adjacent fractions (i.e., a segment of
length $2$).
A chain is \textit{facet-continuous} if adjacent fractions intersect by facets.
Note that by definition polycubic chains are continuous, i.e., adjacent
fractions have non-empty intersection.
The \textit{dilation} of a polycubic chain is the maximal dilation of bodies of
its segments. (Recall the definition~(\ref{dilation}) of the set dilation.)
It is obvious that the sequence of fractions (of some order) of a cubically decomposable curve
is a polycubic chain and the dilation of the curve is at least the
dilation of this chain.
We say that fractions of a chain are collinear (coplanar), it the centers of them are
collinear (coplanar).
\paragraph{Dimension 2.}
We will use the following result.
\begin{theoremB}[\cite{MS21}, Theorem 1]
If a plane surjective curve $\gamma\colon[0,1]\to[0,1]^2$ has entrance
$\gamma(0)$ and exit $\gamma(1)$ on the opposite sides of $[0,1]^2$, then
$\WD_2(\gamma)>4$.
\end{theoremB}
\begin{lemma}
\label{lem_square}
A 2D facet-continuous polycubic chain $S_1,\ldots,S_n$ with more than $4$
fractions and square body has a collinear segment of length $3$.
\end{lemma}
\begin{proof}
Consider the square $P=S_1\cup\ldots\cup S_n$.
It has size $s\times s$ with $s^2=n\ge 9$, so $P$ has
has a vertex that is not contained in $(S_1\cup S_2)\cup (S_{n-1}\cup S_n)$.
This vertex is contained in some fraction $S_i$ with $3\le i\le n-2$. Denote by
$S_j$ the only fraction that intersects $S_i$ by a vertex. If $j>i$ then one can
take triple $S_{i-2},S_{i-1},S_i$ and if $j<i$ we take
$S_i,S_{i+1},S_{i+2}$.
\end{proof}
\def\ref{th_4dsurj}{\ref{th_poly2}}
\begin{theorem}
There exists some $\varepsilon>0$, such that for any plane cubically decomposable
curve $\gamma\colon[0,1]\to[0,1]^2$ one has $\WD_2(\gamma)\ge 4+\varepsilon$.
\end{theorem}
\begin{proof}[Proof of Theorem~\ref{th_poly2}]
If all junctions of a cubically decomposable curve $\gamma$ are facet, then one can take a decomposition into $n\ge 9$
squares and apply Lemma~\ref{lem_square} to find a collinear segment
$S_{i-1},S_i,S_{i+1}$. Then the middle fraction $S_i$ has its entrance and
exit on the opposite sides on the square. We apply Theorem B to obtain that
$\WD(\gamma)>4$.
Suppose that there is a diagonal junction $F\to G$. Consider two last sub-fractions
$F_2,F_1$ of $F$ and two first sub-fractions $G_1,G_2$ of $G$ (so, $F_1$
meets $G_1$). Then
$$
\diam(F_2\cup F_1\cup G_1\cup G_2)\ge\min\{|(3,3)|,|(4,2)|\}=\sqrt{18},
$$
which gives $\WD\ge 18/4=4.5$.
We have proved that $\WD(\gamma)>4$. The bound $\WD(\gamma)\ge4+\varepsilon$ with
some absolute $\varepsilon>0$ follows from the compactness argument.
\end{proof}
\paragraph{Dimension 3.}
\begin{lemma}
\label{lem_4empty}
A 3D facet-continuous $4$-cubic chain with empty intersection of its fractions has dilation at
least $7\sqrt{14}/2=13.0958\ldots$.
\end{lemma}
\begin{proof}
Note that if the intersection of fractions of our chain $Q_1,\ldots,Q_4$
is empty, then it is coplanar. In appropriate coordinates, the vector
between two farthest points of $Q_1$ and $Q_4$ (the ``diameter vector'')
equals $(3,2,1)$ or $(4,1,1)$ and
the diameter of the body $Q_1\cup\ldots Q_4$ at least $|(3,2,1)|=\sqrt{14}$, which
gives the dilation bound $\ge 14\sqrt{14}/4$.
\end{proof}
\begin{lemma}
\label{lem_5empty}
If a polycubic chain with cubic body has more than $8$ fractions then it
contains a $5$-cubic segment with empty intersection.
\end{lemma}
\begin{proof}
Let $Q_1,\ldots,Q_n$ be a chain with a cubic body $P$. If $n>8$, then
$P$ is a $s\times s\times s$ cube with $s\ge 3$. Any vertex of $P$ belongs
to some fraction; let $Q_i$ be the fraction that contains the $4$-th vertex
of P (in the chain traverse order). Then $7\le i\le n-8$.
Let $k$ be the maximal number such that $Q_{i-1},Q_{i-2},\ldots,Q_{i-k}$
intersect $Q_i$, and $l$ be the maximal number such that
$Q_{i+1},\ldots,Q_{i+l}$ intersect $Q_i$. There are $7$ fractions
intersecting $Q_i$, so $\min(k,l)\le 3$. Let, e.g., $l\le 3$. Then the chain
$Q_i,\ldots,Q_{i+l+1}$ has at most $5$ fractions and empty intersection.
\end{proof}
\begin{statement}
\label{stm_hyper3d}
Any facet-continuous $3$-dimensional polycubic chain with cubic body and more
than $8$ fractions has the dilation at least $7\sqrt{14}/2$.
\end{statement}
\begin{proof}
If our chain contains a $4$-cubic segment with empty intersection, then
Lemma~\ref{lem_4empty} provides the required bound.
Otherwise, consider a $5$-cubic chain $Q_1,\ldots,Q_5$ with empty intersection, given by
Lemma~\ref{lem_5empty}. Fractions $Q_2,Q_3,Q_4$ must intersect by an edge.
$Q_1$ and $Q_5$ must intersect that edge (otherwise a $4$-cubic chain with
empty intersection appears), but they cannot contain that edge. Instead of that, they contain
opposite endpoints of it. It follows that the diameter of the body
$Q_1\cup\ldots\cup Q_5$ is at least $\sqrt{3^2+2^2+2^2}$ and the dilation is
at least $17^{3/2}/5>14$.
\end{proof}
Now we are ready to prove lower bounds for the dilation of 3D cubically decomposable curves.
\def\ref{th_4dsurj}{\ref{th_poly3}}
\begin{theorem}
Let $\gamma\colon[0,1]\to[0,1]^3$ be a cubically decomposable curve. Then
$\WD_2(\gamma)\ge 7\sqrt{14}/2>13.09$. If $\gamma$ has a non-facet junction, then
$\WD_2(\gamma)\ge 17.17$.
\end{theorem}
\begin{proof}[Proof of Theorem~\ref{th_poly3}]
If all junctions of a cubically-decomposable curve $\gamma$ are facet, then we apply
Statement~\ref{stm_hyper3d}.
Suppose that $\gamma$ has a non-facet junction $F\to G$. If $F$ and $G$
intersect by a vertex, then we have the dilation bound:
$\WD(\gamma)\ge\diam(F\cup G)^3/2\ge 12^{3/2}/2>20$.
So we may suppose that $\gamma$ has no such junctions.
Let $F$ and $G$ intersect by an edge. Consider sub-fractions
$F_3,F_2,F_1\subset F$ and $G_1,G_2,G_3\subset G$, such that $F_1$ meets
$G_1$.
Let us choose coordinates $x,y,z$, such that the edge $F\cap G$ is
parallel to the third (vertical) axis, the projections of $F$ and $G$
onto horizontal plane $\{z=0\}$
lie in the quadrants $\{x,y\le 0\}$ and $\{x,y\ge 0\}$, correspondingly. For definiteness,
let $G_1=[0,1]^3$ and $F_1=G_1 + (-1,-1,0)$.
Consider the fraction $G_2$.
If $G_1$ and $G_2$ are on the same height, then
$$
\diam(F_2\cup F_1\cup G_1\cup G_2)\ge
\min\{|(3,2,2)|,|(3,3,1)|,|(4,2,1)|\}=\sqrt{17},
$$
hence $\WD(\gamma)\ge 17\sqrt{17}/4>17.5$.
If $G_1$ and $G_2$ intersect by an edge, then the diameter of $F_1\cup G_2$
is already too large. So, we may assume that $G_2$ is above:
$G_2=G_1+(0,0,1)$.
Consider the fraction $G_3$. If the junction $G_2\to G_3$ is facet, then one
can show that $\diam(F_1\cup G_3)\ge|(3,2,2)|$, which is sufficient for us.
It remains one (up to isometry) ``difficult'' case, when $G_3$ and $G_2$
intersect by an edge: $G_3 = G_1 + (1,0,0)$.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=10cm]{th_poly3.png}
\end{center}
\caption{Proof of Theorem~\ref{th_poly3}, ``difficult'' case.}
\label{fig_difficult}
\end{figure}
Consider the ``difficult'' case, illustrated in Figure~\ref{fig_difficult}. Take points $A_0=(-1,-1,1)\in F_1$,
$A_1=(1,1,0)\in G_1$, $A_2=(0,0,2)\in G_2$, $A_3=(2,1,0)\in G_3$. There
exists some moments $t_0<t_1<t_2<t_3$, such that $\gamma(t_i)=A_i$. Then by
definition of $\WD$, we have
$$
\sum_{i=0}^2 |A_{i+1}-A_i|^3 \le \WD(\gamma)\sum_{i=0}^2 |t_{i+1}-t_i| \le
4\WD(\gamma),
$$
$$
\WD(\gamma) \ge \frac14\sum_{i=0}^2 |A_{i+1}-A_i|^3 =
\frac14(27+6\sqrt{6}+27)>17.17.
$$
\end{proof}
\paragraph{Dimension 4.}
\begin{lemma}
\label{lem_7chain}
If a $4D$ facet-continuous $7$-cubic chain $Q_1,\ldots,Q_7$ has a collinear subchain $Q_3,Q_4,Q_5$,
then the dilation of $Q_1,\ldots,Q_7$ is more than $64$.
\end{lemma}
\begin{proof}
Let us reduce the problem to a planar situation. Suppose that 5-chain
$Q_2,\ldots,Q_6$ is not coplanar. Then the directions $Q_2Q_3$ and $Q_5Q_6$
are orthogonal to each other and to the direction of $Q_3,Q_4,Q_5$. Then, in
appropriate coordinates, the vector between two farthest points of $Q_2$ and
$Q_6$ (the ``diameter vector'') equals $(3,2,2,1)$ and
the dilation is $18^2/5=64.8$.
So, $Q_2,\ldots,Q_6$ is coplanar to some plane $\pi$. If one of the
directions $Q_2Q_3$ or $Q_5Q_6$ is parallel to $Q_3,Q_4,Q_5$, then be have a
collinear 4-chain. It has dilation $19^2/4>90$.
If $Q_1,\ldots,Q_5$ or $Q_3,\ldots,Q_7$ are not coplanar to $\pi$, then the
proof goes the same way as for $Q_2,\ldots,Q_7$. So, we may assume that the
whole 7-chain is coplanar to $\pi$. The projection onto $\pi$ of one of the
$5$-chains $Q_1\ldots Q_5$, $Q_2\ldots Q_6$, $Q_3\ldots Q_7$ does not fit
into a $2\times 3$ rectangle and it gives a diameter vector $(3,3,1,1)$ or
$(4,2,1,1)$ and the dilation at least $20^2/5=80$.
\end{proof}
\begin{statement}
\label{stm_hyper4d}
The dilation of 4D facet-continuous polycubic chain with cubic body and
more than one fraction is at least $42.25$.
\end{statement}
\begin{proof}
Denote our chain as $Q_1,\ldots,Q_n$. If there is a collinear 3-cubic chain
$Q_i,Q_{i+1},Q_{i+2}$, $3\le i\le n-4$, then by Lemma~\ref{lem_7chain}
the dilation at least $64$. Assume that such 3-cubic chains are not
collinear. Let us prove that there is a 4-cubic chain $Q_i,Q_{i+1},Q_{i+2},Q_{i+3}$ that is
not coplanar. It this is not the case, then the chain
$Q_2,\ldots,Q_{n-1}$ is coplanar, because all adjacent 4-chains in it are
coplanar to the same plane (defined by their three common fractions).
Obviously, our chain cannot have cubic body in that case.
If $Q_i,Q_{i+1},Q_{i+2},Q_{i+3}$ is not coplanar, then $Q_{i+3}$ is a
translation of $Q_i$ by the vector $(1,1,1,0)$ (in appropriate coordinates).
So, $\diam(Q_i\cup\ldots\cup Q_{i+3})\ge \sqrt{13}$ and we have the
dilation bound $13^2/4=42.25$.
\end{proof}
\def\ref{th_4dsurj}{\ref{th_poly4}}
\begin{theorem}
Let $\gamma\colon[0,1]\to[0,1]^4$ be a cubically decomposable curve. Then
$\WD_2(\gamma)\ge 42.25$. If $\gamma$ has a non-facet junction, then
$\WD_2(\gamma)\ge 62.25$.
\end{theorem}
\begin{proof}[Proof of Theorem~\ref{th_poly4}]
If a cubically-decomposable curve $\gamma$ has only facet junctions, then we apply
Statement~\ref{stm_hyper4d}.
Suppose that $\gamma$ has a non-facet junction $F\to G$. If $F$ and $G$
intersect by an edge, then we have the dilation bound:
$$
\WD(\gamma)\ge\frac12\diam(F\cup G)^4 = |(2,2,2,1)|^4/2 = 84.5
$$
and the proof is done. If $F$ and $G$ intersect by a vertex, there bound is even stronger.
Let $F$ and $G$ intersect by a 2-dimensional face. Consider sub-fractions
$F_3,F_2,F_1\subset F$ and $G_1,G_2,G_3\subset G$, such that $F_1$ meets
$G_1$.
Let us choose coordinates $x,y,z,w$, such that $F\cap G$ is
parallel to the third (vertical) axis $Oz$ and the fourth axis $Ow$, the projections of $F$ and $G$
onto horizontal plane $\{z=w=0\}$
lie in the quadrants $\{x,y\le 0\}$ and $\{x,y\ge 0\}$. For definiteness,
let $G_1=[0,1]^4$ and $F_1=G_1 + (-1,-1,0,0)$.
If the projections of $G_1$ and $G_2$ onto $Oxy$ do not coincide, then
$\diam(F_1\cup G_2)\ge|(3,2,1,1)|=\sqrt{15}$, which gives the dilation bound
$15^2/3=75$. So, $G_2$ may differ from $G_1$ only in the coordinates $z$ and $w$.
If it differs in both coordinates, e.g., $G_2=G_1+(0,0,1,1)$, then we
have the bound
$\WD\ge \diam(F_1\cup G_2)^4/3\ge|(2,2,2,2)|^4/3=85\frac13$. So, we may
assume that $G_2$ differs in $3$-rd coordinate: $G_2=G_1+(0,0,1,0)$.
Analogously, $F_2$ differs from $F_1$ in either $3$-rd or $4$-th coordinate.
Suppose, $F_2$ differs from $F_1$ in the $4$-th coordinate, say,
$F_2=F_1+(0,0,0,1)$~--- then $\WD\ge\diam(F_2\cup
G_2)^4/4\ge|(2,2,2,2)|^4/4=64$. So, $F_2=F_1+(0,0,\pm 1,0)$. The case ``-1''
is impossible, since then $\diam(F_2\cup G_2)\ge|(3,2,2,1)|=\sqrt{18}$.
We obtain that $F_2=F_1+(0,0,1,0)$.
Let us summarize the configuration that we have:
$G_1=[0,1]^4$, $G_2=G_1+(0,0,1,0)$,
$F_1=G_1+(-1,-1,0,0)$, $F_2=G_1+(-1,-1,1,0)$.
Now we have to consider $G_3$ and the translation vector $\delta\colon
G_3=G_2+\delta$. Note that $\delta_1,\delta_2\ge0$.
The first case: $\delta_3=1$: we have
$\diam(F_1\cup G_3)\ge|(3,2,2,1)|=\sqrt{18}$, which gives $\WD\ge
18^2/4=81$.
The second case: $\delta_3=0$ (i.e., $G_3$ and $G_2$ are on the same
``height''). As $\delta\ne\mathbf{0}$, we have $\delta_1=1$ or $\delta_2=1$
or $\delta_4=\pm 1$. In any case, we have
$$
\diam(F_1\cup G_3)\ge \min\{|(3,2,2,1)|,|(2,2,2,2)|\}=4,\quad \WD\ge
4^4/4=64.
$$
The last case: $\delta_3=-1$. If $\delta_1=1$ or $\delta_2=1$, then
$\diam(F_2\cup G_3)\ge|(3,2,2,1)|$, $\WD\ge18^2/5=64.8$ and the proof is
done. So we may assume that $\delta_1=\delta_2=0$. As $G_3\ne G_1$, we have $\delta_4\ne0$; so, w.l.o.g. we may
assume that $\delta=(0,0,-1,1)$ and $G_3=G_1+(0,0,0,1)$.
This ``difficult'' case
needs special treatment, as in the proof of the previous theorem.
Take the following points:
$A_0=(-1,-1,1,0)\in F_1$, $A_1=(1,1,0,1)\in G_1$, $A_2=(0,0,2,0)\in G_2$,
$A_3=(1,1,0,2)\in G_3$. There exist some moments $t_0<t_1<t_2<t_3$ such that
$\gamma(t_i)=A_i$. By the definition of $\WD$, we have
$$
\sum_{i=0}^2 |A_{i+1}-A_i|^4 \le \WD(\gamma)\sum_{i=0}^2|t_{i+1}-t_i| \le
4\WD(\gamma),
$$
so $\WD(\gamma)\ge\frac14(10^2+7^2+10^2)=62.25$.
\end{proof}
\subsection{Proof of Theorem~\ref{th_4dsurj}}
\def\ref{th_4dsurj}{\ref{th_4dsurj}}
\begin{theorem}
For any surjective curve $\gamma\colon[0,1]\to[0,1]^4$ we have
$\WD_2(\gamma)\ge 30$.
\end{theorem}
Vertices of the cube $[0,1]^4$ form the boolean cube $\{0,1\}^4$. In fact we will
obtain the dilation bound for any curve whose image contains $\{0,1\}^4$. We
say that vertices $v,v'\in\{0,1\}^4$ are \textit{3-antipodes} if they differ in
exactly $3$ coordinates, and \textit{4-antipodes}, if the differ in all $4$
coordinates.
\begin{lemma}
Any set $S$ of $6$ elements of the boolean cube $\{0,1\}^4$ contains
a pair of $4$-antipodes or two pairs of $3$-antipodes.
\end{lemma}
\begin{proof}
We may assume that there are no $4$-antipodes in $S$.
Let us prove that there is at least one pair of $3$-antipodes.
We may assume that $v=(0,0,0,0)\in S$. If $S$ has no $3$-antipodes, then all
points in $S$ have at most two non-zero coordinates. So, $S$ contains a
point, say, $w=(1,1,0,0)$ with two non-zero coordinates (because there
are only $4$ vertices with one coordinate). But $w$ has $3$-antipodes
$(0,0,1,0)$ and $(0,0,0,1)$, so that points can't lie in $S$. Then $S$
contains one more point $u$ with two non-zero coordinates; as
$u\ne(0,0,1,1)$, we may assume $u=(0,1,1,0)$. Then vertices $(1,0,0,0)$ and
$(0,0,0,1)$ are forbidden, so $S$ must contain at least $4$ points with two
non-zero coordinates. These points contain a pair of $4$-antipodes, that
contradicts our assumption.
So, we have a pair of $3$-antipodes, say, $v=(0,0,0,0)\in S$ and
$v'=(1,1,1,0)\in S$. If there are no other $3$-antipodes, then all vertices
in $S$ have at most two non-zero coordinates. But if the last coordinate of
such point is non-zero, then this point is $3$-antipode to $v'$. So, $S$
lies in the $3$-dimensional sub-cube $\{(x_1,x_2,x_3,x_4)\colon x_4=0\}$.
And in the $3$-dimensional case the statement of Lemma is obvious.
\end{proof}
Now we can finish the proof of Theorem~\ref{th_4dsurj}. Let
$A_0 := \gamma(t_0),\ldots,A_{15} := \gamma(t_{15})$ be the sequence of points of $\{0,1\}^4$ in the
traversal order. Note that $(t_5-t_0)+(t_{10}-t_5)+(t_{15}-t_{10})\le 1$, so one
of the summands is at most $1/3$. Hence we have a sub-sequence
$A_s,\ldots,A_{s+5}$ of six points with $t_{s+5}-t_s\le 1/3$.
We apply Lemma to points $A_s,\ldots,A_{s+5}$ and arrive to
three possible cases: 1) there is a pair of $4$-antipodes 2) there are three
points that give two pairs of $3$-antipodes, 3) there
are four points that give two pairs of $3$-antipodes.
Denote $\omega := \WD_2(\gamma)$.
In the case 1) we have a pair $A_i$, $A_j$ of $4$-antipodes,
$$
16 = |A_i-A_j|^4 \le \omega(t_j-t_i) \le \omega/3,
$$
In the case 2) we have $A_i,A_j,A_k$, $i<j<k$ and two of three
pairwise distances equal $\sqrt{3}$, so
$$
10 \le |A_i-A_j|^4 + |A_j-A_k|^4 \le \omega(t_k-t_i) \le \omega/3.
$$
Consider the last case of four points $A_i,A_j,A_k,A_l$. If the
antipode $B$ of $A_i$ is one of $A_j$, $A_k$, then
$$
10 \le |A_i-B|^4 + |B-t_l|^4 \le \omega(t_l-t_i) \le \omega/3.
$$
If $A_i$ and $A_l$ are antipodes, then $A_j$ and $A_k$ are antipodes too,
$$
11 \le |A_i-A_j|^4 + |A_j-A_k|^4 + |A_k-A_l|^4 \le \omega/3.
$$
In any case $\omega\ge30$ and the Theorem is proven.
\section{Further work}
There remain a lot of problems on the minimal $\ell_p$-dilation of Peano
multifractals, both theoretical and computational. Let us discuss some of them.
\paragraph{Promising directions.}
Our library allows to estimate minimal dilation for Peano
multifractals of arbitrary dimension $d$, multiplicity $m$ and genus $g$. The
number of possible variants grows rapidly as we increase any of the parameters
$d$, $m$, $g$; this raises two difficulties. When we make a brute force search over
all pointed prototypes in some class, the number of pointed prototypes
corresponds to the ``width'' of the search, it is CPU-bounded. Given a pointed
prototype, we use a SAT-solver to approximate minimum dilation and the number of
possible curves corresponds to the ``depth'' of the search, it is memory (RAM)
bounded. For example, for edge 3D trifractals of genus $2\times2\times2$ there
are not much pointed prototyes, but they are very
``deep''. Conversely, there are too many pointed prototypes for edge 2D
monofractals of genus $6\times 6$; we have succeeded in the search among them
only because we could throw away prototypes with diagonal steps (and these
prototypes are not so deep).
Still it is possible to use the library in many interesting
cases. Here is a list of configurations to start with:
\begin{itemize}
\item plane $7\times 7$ monofractals ($\ell_2$-dilation);
\item 3D-bifractals of genus $2^3$ ($\ell_1$-dilation);
\item 3D facet-gated trifractals of genus $2^3$ ($\ell_2$-dilation)
\item 3D-monofractals of genus $3^3$;
\item 4D facet-gated monofractals of genus $2^4$ ($\ell_1$, $\ell_\infty$-dilation);
\end{itemize}
Some of them will require optimizations in the algorithm or additional
constraints (on prototypes, base orientations, junctions, gates, etc) to reduce
the number of curves. A good example is the narrow class of facet-gated curves.
Another interesting constraint is the \textit{symmetry} of a curve.
SAT-solvers may be applied in broader optimization, e.g., in the classes of
curves with fixed prototype~--- just add the clauses for adjacent fractions that
ensure the continuity of a curve.
\paragraph{Other metrics.}
Our algorithm for minimal $\ell_p$-dilation estimation in fact works for any
norms and other curve functionals. This allows, e.g., to investigate curves in
boxes instead of cubes: $\gamma\colon[0,1]\to
[a_1,b_1]\times\cdots\times[a_d,b_d]$; it is equivalent to the consideration of
usual Peano curves in metric $\|x\| := \|(b-a)x\|_p$.
Another interesting family of metrics is based on the bounding boxes of the
sets $\gamma([t_1,t_2])$, see~\cite{HW10}. The structure of our algorithm
allows to minimize them analogously.
\paragraph{Acknowledgement.}
The authors express their gratitude to the anonymous referee for his careful
work and valuable suggestions.
Data availability statement. The datasets generated during and/or analysed
during the current study are available in the github repository~\cite{gitYM}.
|
\subsection{Single mode approximation}
We start by solving Eq.~\eqref{EQM} within a single-mode approximation, by considering one relevant mode $\omega_i = 2\pi f_i$, corresponding to one relevant site,
\begin{equation}
\ddot{u}_{\alpha}(t) + \eta\,\dot{u}_{\alpha}(t) + \omega^2_i u_{\alpha}(t) = \frac{q}{m}\, E^*_\alpha(t).
\label{EQMeffective}
\end{equation}
We choose circularly polarized light, i.e., $\vec{E}^*(t) = E^*_0 \left(\sin(\omega t),\cos(\omega t),0\right)$. In a coarse approximation, from \eqref{EQMeffective}, we notice that the displacement scales linearly with the applied field, $\vec{u} \approx \frac{q \vec{E}^*}{m\omega^2}$. For a harmonic displacement, we can estimate the corresponding time derivative as $\dot{\vec{u}}\approx \omega \vec{u}$. Using equation \eqref{DMF} and replacing the gyromagnetic ratio by $\gamma=\frac{q}{2m}$, we can estimate the asymptotic behavior for the dynamically induced magnetic moment by
\begin{equation}
M_z \sim \frac{q^3 {\vec{E}^*}^2}{m^2 \omega^3}.
\label{asymptotics}
\end{equation}
Hence, the effect increases quadratically in the field strength, but decreases with $\omega^{-3}$ in the driving frequency. The corresponding values for the charge $q$ and the mass $m$ for KTO are given in Tab. \ref{parameters}. The charges calculated using DFT are close to the chemistry picture of an ionic crystal, with integer oxydation states O$^{-2}$, K$^{+1}$, and Ta$^{+5}$.
Equation \eqref{EQMeffective} can be solved exactly. As we are solely interested in the contribution to the atomic displacement emerging due to exposure to an external laser field, we only keep the inhomogeneous part of the solution of Eq.~\eqref{EQM} that can be written as
\begin{equation}
\vec{u}(t)
= \frac{1}{\Delta_{\omega}^4+4\eta^2\omega^2}
\left(
\begin{array}{cc}
\Delta_{\omega}^2 & - 2 \eta \omega \\
2 \eta \omega & \Delta_{\omega}^2
\end{array}
\right)
\frac{q}{m} \vec{E}^*,
\label{phonon1}
\end{equation}
with $\Delta_{\omega}^2 = \omega^2_i-\omega^2$. Evaluating the polarization as $\vec{P}=\frac{q}{V} \vec{u}$, the $\omega$-dependent part of Eq.~\eqref{phonon1} can be interpreted as the susceptibility $\chi$, by transforming it into the well-known expression $\vec{P} = \chi \epsilon_0 \vec{E}$. Hence, we obtain for the magnetization
\begin{equation}
M_z = \frac{q^3 \omega {\vec{E}^*}^2}{2 m^2 \left(\eta^2\omega^2+\Delta_\omega^4\right)}.
\label{Mzanalytical}
\end{equation}
In the limit $\omega \gg \omega_i$, we obtain $\Delta_\omega^4\approx \omega^4$. Neglecting the damping term $\eta^2\omega^2 \ll \omega^4$ gives a similar expression to Eq.~\eqref{asymptotics}.
{\it System driven with a terahertz pulse.} Next, we consider a more realistic terahertz pulse and solve Eq.~\eqref{EQM} numerically. Such terahertz pulses are nowadays available \cite{SALEN20191} and allow for large peak electric field to drive phonons, but with an average deposited energy which is not enough to melt the sample. We set $\vec{u}_\alpha(0) = \dot{\vec{u}}_\alpha(0) = 0$. The pulse is modeled by a Gaussian embedding as follows,
\begin{equation}
\vec{E}(t) = E_0\,e^{-\frac{(t-t_0)^2}{2 \sigma}} \left(
\begin{array}{c}
\sin(\omega t) \\ \cos(\omega t) \\ 0
\end{array}
\right).
\end{equation}
The considered driving frequencies are 3.17 THz and 6.19 THz, being resonant with the phonon modes. We choose a total width of $2$ ps with a peak at $2$ ps and obtain the solution for a window up to 16 ps. The dynamically induced magnetic moments are shown in Fig. \ref{pulsedresult} for various values of the damping parameter ($\eta$ = 0.05 THz, 0.10 THz, 0.15 THz). Depending on the damping factor we observe a slow decay of the dynamically induced magnetic moment. The maximal total dynamically induced magnetic moment is $\approx 0.7 \mu_N$ for small damping of $\eta < 0.1$ THz. The dynamically induced magnetic moment decreases by about one order of magnitude for a driving frequency in resonance with the $T_{1u}$ mode at 6.19 THz.
\begin{figure}[t!]
\centering
\includegraphics[width=0.495\textwidth]{pulsed_muN.pdf}
\caption{Dynamically induced total moment per unit cell for a laser pulse with driving frequencies 3.17 THz (left panel) and 6.19 THz (right panel).}
\label{pulsedresult}
\end{figure}
Due to the opposite local charges of the ions, the induced moments have opposite strength for O, compared to Ta and K. The site resolved dynamically induced moments due to local displacements are shown in Fig. \ref{siteresolved}. We observe that the main contributions to the total induced magnetization per unit cell come from Ta and O, being of the order of $0.2\mu_N$ and $-0.1\mu_N$ for a small value of the damping parameter, $\eta = 0.05$ THz.
\begin{figure}[t!]
\centering
\includegraphics[width=0.4\textwidth]{atomic_pulsed.pdf}
\caption{Site resolved dynamically induced moments within the unit cell. We used the same pulse as in Fig. \ref{pulsedresult}, a driving frequency of 3.17 THz and a damping of $\eta = 0.05$ THz.}
\label{siteresolved}
\end{figure}
{\it Conclusion and Outlook.} We showed that KTO is a prominent candidate for the observation of the dynamical multiferroicity. By performing an \textit{ab initio} analysis, we first find that the T$_{1u}$ soft phonon modes may be relevant for the observation of the effect. We suggest an experimental setup where the KTO sample is exposed to a circularly polarized laser field in the terahertz range to excite phonons resonantly. The dynamically induced magnetization due to locally oscillating dipoles could be measured by the time-resolved Faraday effect using a femtosecond laser pulse in the visible range. The estimated scale of the effect for an experimentally feasible setup is in the order of $10^{-2}~\mu_N$ per unit cell, with $\mu_N$ being the nuclear magneton. In Eq.~\eqref{asymptotics} we show that in an asymptotic limit, the induced moment scales quadratically with the electric field strength and to the the third power in the charge. It also scales inversely with the third power in driving frequency and the mass squared. In particular the latter feature could be of interest.
Here we discussed the ionic movement as a driver for the induced magnetism. We now point out an interesting possibility of induced electron motion that also would produce the magnetic moment. We expect the angular momentum transfer from the moving ions to the electronic charge cloud in the solid. While the exact microscopic details need to be worked out the qualitative argument goes as follows. To estimate the gyromagnetic ratio for the coupling we follow Refs. \cite{rebane1983faraday,khaetskii2020thermal} in a modified form. The position of a charged ion is denoted by $\vec{u}_+$, the average displacement of the electron cloud is $\vec{u}_-$. The respective masses are $m_+$ and $m_-$. We introduce average and relative coordinates $\vec{U}=(m_+\vec{u}_-+m_-\vec{u}_+)/(m_++m_-)$ and $\vec{u}=\vec{u}_+-\vec{u}_-$. We focus on the relative coordinate, having the momentum $p=\mu \dot{\vec{u}}$ with $\mu=m_+m_-/(m_++m_-)$. It follows for the angular momentum of the relative coordinate
\begin{equation}
\vec{L} = \vec{u}\times\vec{p} = \frac{m_+m_-}{m_++m_-} \vec{u}\times\dot{\vec{u}}.
\end{equation}
Setting $m_-\vec{u}_+ = m_+\vec{u}_-$ we obtain for the dynamically induced moment according to Eq.~\eqref{DMF}
\begin{equation}
\vec{M} = \vec{m}_+ + \vec{m}_- = \frac{q}{2}\frac{m_+-m_-}{m_++m_-} \vec{u}\times\dot{\vec{u}}.
\end{equation}
Taking $\vec{M} = \gamma \vec{L}$, we obtain for the gyromagnetic ratio
\begin{equation}
\gamma = \frac{q}{2} \left(\frac{1}{m_-} - \frac{1}{m_+}\right).
\label{gmratio_electron}
\end{equation}
For nonequal charges, this equation generalizes to
\begin{equation}
\gamma = \frac{m_+}{m_-} \frac{q_+}{m_++m_-} - \frac{m_-}{m_+} \frac{q_-}{m_++m_-}.
\label{gmratio_electron2}
\end{equation}
Hence, from Eqs.~\eqref{gmratio_electron} and \eqref{gmratio_electron2} it becomes apparent that the total gyromagnetic ratio of ion and electron is dominated by the electron mass ($m_i/m_e \sim 10^3\dots 10^5$). Here we need to distinguish between a direct coupling of the electron to the external field $\sim \epsilon_0\left(\epsilon_{\alpha\beta} -\delta_{\alpha\beta}\right) E_\beta$ as well as an induced motion of the electrons due to the ionic movement. While the former contribution to the total magnetization should vanish with vanishing electric field, the latter should be present as long as the ionic movement persists. More precise analysis will be a topic of a separate publication.
We propose KTO as a prominent candidate for the observation of the dynamical multiferroicity. Our findings open up a route for the experimental detection of the entangled dynamical orders. They should also motivate further studies of the candidate materials for the realization of the effect.
{\it Acknowledgment.} We are grateful to G. Aeppli, U. Aschauer, M. Basini, M. Pancaldi, O. Tjernberg, I. Sochnikov, N. Spaldin and J. Weissenrieder for useful discussions. We acknowledge support from VILLUM FONDEN via the Centre of Excellence for Dirac Materials (Grant No. 11744), the European Research Council under the European Union Seventh Framework ERS-2018-SYG
810451 HERO, the Knut and Alice Wallenberg Foundation KAW 2018.0104.
V.J. acknowledges the support of the Swedish Research Council (VR 2019-04735) and J.-X.Z. was supported by the Los Alamos National Laboratory LDRD Program. SB acknowledges support from the Swedish Research Council (VR 2018-04611). The computational resources were provided by the Swedish National Infrastructure for Computing (SNIC) via the High Performance Computing Centre North (HPC2N) and the Uppsala Multidisciplinary Centre for Advanced Computational Science (UPPMAX).
|
\section{Introduction}
\vspace{0.3cm}
Finite element (FE) methods and finite difference (FD) methods are among the most studied and employed numerical methods for partial differential equations (PDEs), which serve numerous industrial applications and other research areas, e.g., economics, engineering, chemistry, physics. FD methods are often seen as \cpurple{simple and efficient schemes} for high-order approximations, see e.g., \citet{Mattsson2004,Shu2003}, despite being restricted to structured grids. FE methods are favored in various computational \cpurple{areas} due to their theoretical reliability, generality, and the ability to handle complex \cpurple{domain} shapes. However, the obtained matrices can be storage demanding; and the method itself is often considered to be more computationally expensive. For instance, while solving time-dependent nonlinear diffusion dominated problems, one has to reconstruct FE matrices in each time step, which is time-consuming. The greatest strengths of both methods can be combined by developing a hybrid scheme that allows different methods and refinement levels to be involved simultaneously. Hence, computational efficiency can be enhanced by utilizing the most suitable treatment for each portion of the domain. The major complication is known to consist in designing a method-to-method interface that results in an accurate and stable approximation. \cpurple{A favorable method should also accommodate the demands for simplicity, freedom to choose domain discretization, and minimal modification in the existing schemes to preserve their advantageous native properties.}
\cpurple{A well-designed numerical framework for energy stable approximations of time-dependent problems is the combination of the summation-by-parts (SBP) operators, see \citet{Kreiss1974}, and the simultaneous-approximation-term (SAT) technique, see \citet{Carpenter1994}. The SBP operators are used to approximate the governing equations. By design, they mimic the continuous integrations-by-parts properties. The SAT technique then imposes the boundary conditions in a provably stable manner.} We refer the reader for a detailed discussion on the theory and applications of SBP-SAT in the following papers and references therein, \cpurple{\cite{Carpenter2010,Fernandez2014,Kozdon2016,Lundgren2020,Lundquist2018,Mattsson2004,Nordstrom2001,Zemui2005}}. A crucial benefit of numerical schemes in this form is that a proper formulation strictly prevents nonphysical energy growth in the discrete approximation, a property often referred to as ``\textit{strict stability}''. The vast majority of previous papers on the SBP-SAT framework have been committed to FD schemes due to high-order operators' apparent advantages. Nontrivial geometries are then usually handled with \cpurple{curvilinear grids, see \citet{Nordstrom2001}, \cblue{\citet{Shadpey2020}}, \cblue{curved elements, see \citet{Crean2018}}, and multiblock coupling, see \citet{Carpenter2010}}. To carry out further practical geometries, successful attempts have been made to couple FD methods with unstructured finite volume (FV) methods, see \citet{Nordstrom2006,Nordstrom2009} by modifying the FV scheme at interfaces. The coupling of mixed order schemes on nonconforming grids was, for the first time, \cpurple{proposed} in an energy stable manner by \citet{Mattsson2010} introducing \textit{SBP-preserving interpolation operators}. Fundamental properties of SBP schemes are extended to block-to-block coupling: strict stability, accuracy, and conservation. This technique paves the way for many possibilities to couple different methods which can be written in SBP form. \cpurple{So far, the coupling of FD--FD by \citet{Mattsson2010}, FD--discontinuous Galerkin (dG) method by \citet{Kozdon2016}, and FD--FV by \citet{Lundquist2018} are available in the literature.}
In practical computation, properties of the concerning problem, e.g., nature of the solution, desired precision, geometries, often vary over the domain. An impediment with discretizations by a single conventional method, especially when using uniform precision, is that sometimes the preferable scheme or the required resolution for a minor region may bring up extreme computational necessity over the whole domain, e.g., when complex geometries present. Despite being of great interest, attempts to combine FD methods and FE methods have rarely been successful. Independent techniques have been proposed \cpurple{for} several applications, e.g., seismic wave propagation, see \cite{Galis2008,Ma2004,Gao2019}, electrical modeling, see \cite{Vachiratienchai2010}. For example, \cite{Galis2008,Ma2004} require a transition region between the two methods; the interface solution is updated similarly to imposing Dirichlet boundary conditions with a suitably averaging solution. In \cite{Vachiratienchai2010}, at regions of interest, each rectangular FD block is \cpurple{split} into two FE triangular elements resulting in locally modified schemes; accuracy is observed in cases of sufficiently gradual topographic gradients. Most relevantly, \citet{Gao2019} derive an energy stable coupling of SBP FD schemes and FE schemes with quadrilateral elements under shared interface vertices/nodal points. Inline with other hybrid SBP methods, although \cpurple{the continuous Galerkin FE also has SBP properties, see \citet{Zemui2005}}, \cpurple{one difficulty} in deriving a hybrid method lies within its nondiagonal mass matrix (can be understood as an SBP norm matrix). Note that the most related references, such as \citet{Lundquist2018,Gao2019}, necessitate diagonal norm matrices.
In this paper, \cred{our main focus is to develop} a hybrid method to couple FD and FE methods in a nonconforming multiblock fashion. The proposed technique results in a provably stable, accurate, and energy-conserved unified SBP method. Interface continuity is weakly imposed by the SAT technique using \textit{nondiagonal norm SBP-preserving interpolation operators}. The SAT treatment on the FD side is the same as in our FD--FD coupling \cite{Mattsson2010}. \cpurple{The SAT treatment for FE that we propose in this current paper} is formulated in a general expression in the sense that it does not depend on \cpurple{the coupled scheme or the coupled mesh/grid}. \cred{Furthermore}, we show that even when considering nonlinear conservation laws, nondiagonal norm matrices, namely the continuous FE mass matrices, could be applicable without further \cred{techniques involved}, e.g., mass lumping \cite{Zemui2005}, to the classical Galerkin formulation. \cred{We include two techniques to construct the required interpolation operators}. In the latter one, we introduce a \cred{way} of acquiring the target interpolation operators through existing diagonal norm techniques, for instance, \cite{Almquist2018,Kozdon2016,Lundquist2018,Lundquist2020,Mattsson2010}.
This paper is organized as follows. In Section \ref{sec:sbp_operators}, the numerical schemes of both FE and FD are \cpurple{presented} in the context of SBP framework. \cpurple{Section \ref{sec:coupling} describes the proposed coupling technique}. We show numerical verification and computational results in Section \ref{sec:numerical_results} on an advection-diffusion equation and a nonlinear Burgers' equation. Section \ref{sec:conclusion} is the conclusion. Two approaches to constructing the required interpolation operators are shown in Appendix \ref{section:construct_interpolation_operators}.
\section{Summation-by-parts operators}
\label{sec:sbp_operators}
This section formulates summation-by-parts schemes by FD and FE approximations. Boundary treatment is briefly \cpurple{mentioned}.
\subsection{The continuous problem to a viscous scalar conservation law}
\label{sec:advection_diffusion_continuous}
Let $\Omega$ be a bounded domain in $\mathbb{R}^2$. \cpurple{Consider the following scalar conservation laws}
\begin{equation}
\label{eq:advection_diffusion_fem}
\begin{aligned}
\frac{\partial u}{\partial t} + \nabla \cdot \bm{f}(u) & = \nabla \cdot (\epsilon \nabla u), && (x,y) \in \Omega, t\in (0,+\infty),\\
u & = u_0(x,y), && (x,y) \in \Omega, t = 0,
\end{aligned}
\end{equation}
where the flux vector $\bm{f}(u) = \left(f_1(u),f_2(u)\right)^T$ consists of two linear or nonlinear mappings $f_1,f_2 \in C^1(\Omega; \mathbb R)$; $\epsilon = \epsilon(x,y)$ is a positive-valued function on $\Omega$; and $\mathbf{n}$ is the normal vector pointing outward at the boundary. The notation ``$\nabla$'' refers to the gradient operator $\nabla = \left(\frac{\partial}{\partial x},\frac{\partial}{\partial y}\right)^T$. In analysis of FD schemes, it is more convenient to write \eqref{eq:advection_diffusion_fem} as
\begin{equation}
\label{eq:advection_diffusion_fdm}
u_t + (f_1)_{x}u_x + (f_2)_{y}u_y = (\epsilon u_{x})_x+(\epsilon u_{y})_y, \quad (x,y) \in \Omega,
\end{equation}
where the subscript notations $x,y,t$ indicate the partial derivatives with respect to the corresponding variables. \cblue{Let $\bm f' = (f_1', f_2')^T$ be the first derivative of $\bm f$ with respect to the conserved variable $u$}. We assume that there exists a skew-symmetric splitting with the weight functions $\bm\alpha,\bm{\tilde\alpha}$,
\begin{equation}
\label{eq:skew_symmetric_assumption}
\nabla \cdot \bm{f} = (\bm\alpha\nabla) \cdot \bm{f} + \bm f'(u) \cdot \big(\bm{\tilde\alpha}\nabla\big) u, \quad\bm\alpha = (\alpha_1(x,y),\alpha_2(x,y))^T,\bm{\tilde\alpha}=(1-\alpha_1(x,y),1-\alpha_2(x,y))^T,
\end{equation}
such that by using the alternative flux form, the total energy over time of a solution to \eqref{eq:advection_diffusion_fem} only conserves or decays, on the boundary or due to viscous dissipation, yet to be quantified. It is \cpurple{worth noting} that the above assumption is not too severe, in which the existence of such a splitting is shown for a broad class of fluxes \cpurple{by} \citet{Tadmor1984} by symmetrizing variables. \cpurple{This splitting technique is necessary for obtaining nonlinear energy estimates to \eqref{eq:advection_diffusion_fem}}. In \eqref{eq:skew_symmetric_assumption}, the following notations are used $\bm\alpha\nabla=\left(\alpha_1\partial/\partial x,\alpha_2\partial/\partial x\right)^T,\bm{\tilde\alpha}\nabla=\left(\tilde\alpha_1\partial/\partial x,\tilde\alpha_2\partial/\partial x\right)^T$. Similar notations used below are $(\bm\alpha\bm n)=(\alpha_1 n_1,\alpha_2 n_2)^T,(\bm{\tilde\alpha}\bm n)=(\tilde\alpha_1n_1,\tilde\alpha_2n_2)^T,\bm n=(n_1,n_2)^T$. One option of specifying the boundary condition is
\begin{equation}
\label{eq:advection_diffusion_fem_bc}
\begin{aligned}
\frac{1}{2}\big((\bm\alpha\mathbf{n})\cdot \bm{f}-\text{sign}(u)\abs{(\bm\alpha\mathbf{n})\cdot \bm{f}}\big)+\frac{1}{2}\big((\bm{\tilde\alpha}\mathbf{n})\cdot (u\bm{f}')-\text{sign}(u)\abs{(\bm{\tilde\alpha}\mathbf{n})\cdot (u\bm{f}')}\big)-\mathbf{n}\cdot\epsilon\nabla u & = g_b(x,y,t),\\ & (x,y) \in \partial \Omega, t > 0,
\end{aligned}
\end{equation}
where $\partial \Omega$ denotes the domain boundary. \cgreen{Stability} of the problem \eqref{eq:advection_diffusion_fem} subject to the boundary condition \eqref{eq:advection_diffusion_fem_bc} is analyzed \cpurple{below}.
\begin{definition}
\label{def:continuous_inner_product_norm}
Given two real-valued vector functions $\bm w_1$, $\bm w_2$, and a real-valued function $b(x,y)>0$, the following definitions of a continuous \textit{inner product} and its corresponding \textit{norm} are used
\[
(\bm w_1,\bm w_2)_b = \int_\Omega \bm w_1\cdot \bm w_2 b(x,y) \,dxdy,\quad \norm{\bm w_1}^2_b = (\bm w_1,\bm w_1)_b,\quad(\bm w_1,\bm w_2)_{\partial\Omega,b} = \int_{\partial\Omega} \bm w_1\cdot \bm w_2 b(x,y) \,ds.
\]
\end{definition}
\noindent When $b(x,y)\equiv 1$, the subscript notation $b$ is skipped. Inner products and a norm of two scalar functions are defined analogously by considering $\bm w_1,\bm w_2$ as scalar functions. One can investigate necessary conditions for well-posedness of \eqref{eq:advection_diffusion_fem} with the boundary condition \eqref{eq:advection_diffusion_fem_bc} using the \textit{energy method}, see e.g., \citet{Kreiss1974,Fernandez2014}. For continuous problems that can be written in the form $u_t=F(t,u)$, a basic idea of the energy method is to examine energy growth utilizing the relation $\frac{d}{dt}\norm{u}^2=(u,u_t)+(u_t,u)$. A well-posed problem implies that $\frac{d}{dt}\norm{u}^2$ is nonpositive given zero boundary data. For two smooth scalar functions $w_1,w_2\in C^1(\Omega)$ and a smooth vector function $\bm{w}_3$, continuous integration rules read
\begin{equation}
\label{eq:continuous_intergration_rules}
\begin{aligned}
\left(w_1,\nabla \cdot \bm{w}_3\right) & = \left(w_1,\mathbf{n}\cdot \bm{w}_3\right)_{\partial\Omega} - \left(\nabla w_1,\bm{w}_3\right),\\
\left(w_1,\nabla\cdot\nabla w_2\right) & = \left(w_1,\mathbf{n}\cdot\nabla w_2\right)_{\partial\Omega} - \left(\nabla w_1, \nabla w_2\right).
\end{aligned}
\end{equation}
Applying \eqref{eq:continuous_intergration_rules} to \eqref{eq:advection_diffusion_fem} gives
\begin{align*}
(u,u_t) = &\; \left(\bm\alpha\nabla u,\bm{f}\right)-\left(u,\bm\alpha\mathbf{n}\cdot \bm{f}\right)_{\partial\Omega}+(\tilde{\bm\alpha}\nabla\cdot(u \bm f'),u)-(\tilde{\bm\alpha}\mathbf{n}\cdot(u\bm{f}'),u)_{\partial\Omega}\\
&+(u,\mathbf{n}\cdot\epsilon\nabla u)_{\partial\Omega}-(\epsilon\nabla u,\nabla u),\\
(u_t,u) = &\;-\left(\bm\alpha\nabla \cdot \bm f(u),u\right)-\left(\tilde{\bm\alpha}\bm f'(u) \cdot \nabla u, u \right)\\
&+(\mathbf{n}\cdot \epsilon\nabla u,u)_{\partial\Omega}-(\epsilon\nabla u,\nabla u).
\end{align*}
Adding up the two equalities above yields
\begin{align}\label{eq:continous_energy_estimate}
\frac{d}{dt} \norm{u}^2 + 2\norm{\nabla u}_\epsilon^2 = -(u,\bm\alpha\mathbf{n}\cdot\bm{f})_{\partial\Omega}-(\tilde{\bm\alpha}\bm n \cdot (u\bm f'),u)_{\partial \Omega}+2(u,\mathbf{n}\cdot\nabla u)_{\partial \Omega,\epsilon},
\end{align}
where all the interior terms are cancelled by the following skew-symmetry assumption
\begin{equation}
\label{eq:skew_symmetric_precise}
-\alpha_1[(f_1)_xu-f_1u_x]-\alpha_2[(f_2)_yu-f_2u_y]+(1-\alpha_1)u^2(f_1')_x+(1-\alpha_2)u^2(f_2')_y = 0.
\end{equation}
\cred{If the splitting is defined pointwise}, i.e., $\alpha_1, \alpha_2$ are constant weights, a sufficient spatial-derivative-free argument to \eqref{eq:skew_symmetric_precise} reads
\begin{equation}\label{eq:skew_symmetric_sufficient}
\begin{cases}
\alpha_1f_1(u)-(1-\alpha_1)uf_1'(u) = 0,\\
\alpha_2f_2(u)-(1-\alpha_2)uf_2'(u) = 0.
\end{cases}
\end{equation}
Given a specific flux, one can either solve \eqref{eq:skew_symmetric_precise} or \eqref{eq:skew_symmetric_sufficient} to determine the value of $\bm\alpha$. It becomes apparent that the boundary condition \eqref{eq:advection_diffusion_fem_bc} imposes nonlinear stability to \eqref{eq:advection_diffusion_fem} since bounded energy is followed. Indeed, inserting \eqref{eq:advection_diffusion_fem_bc} \cred{into \eqref{eq:continous_energy_estimate}} with zero boundary data $g_b \equiv 0$ gives
\begin{align}
\label{eq:advection-diffusion-continuous-energy}
\frac{d}{dt} \norm{u}^2 + 2\norm{\nabla u}_\epsilon^2 = -(u,\text{sign}(u)\abs{\bm\alpha
\mathbf{n}\cdot \bm{f}})_{\partial \Omega}-(\text{sign}(u)\abs{\tilde{\bm\alpha}\bm n \cdot (u\bm f')},u)_{\partial \Omega} \leq 0.
\end{align}
\subsection{Properties of the discrete operators}
\cgreen{Considering a general numerical method}, we denote \cblue{$\mathrm N$} the total number of degrees of freedom of the discretization, and \cblue{$\mathrm N_{\partial\Omega}$} the number of degrees of freedom distributed on the boundary. Let $\mathcal{H}$ be an integration operator over the whole domain $\mathcal{H}\bm u \approx \int_{\Omega} u dxdy, \bm u \in \mathbb{R}^{\cblue{\mathrm N}}, \bm u\approx u(x,y)$; $\mathcal{B}$ be an integration operator over the boundary $\mathcal{B}\bm u|_{\partial\Omega} \approx \int_{\partial\Omega} u dxdy$; and $\mathcal{L}$ be a boundary selection operator $\bm u|_{\partial\Omega}=\mathcal{L}\bm u$. Precise definitions of the three matrix operators $\mathcal{H},\mathcal{B},\mathcal{L}$ are often required by conventional summation-by-parts schemes, see e.g., \citet{Kreiss1974,Lundquist2018}. In the current study, $\mathcal{H},\mathcal{B},\mathcal{L}$ are assumed to have the following properties: $(i)$ $\mathcal{H},\mathcal{B}$ are symmetric, positive definite; $\mathcal{H}$ is of size \cblue{$\mathrm N\times\mathrm N$}; and $\mathcal{B}$ is of size \cblue{$\mathrm N_{\partial\Omega}\times\mathrm N_{\partial\Omega}$}; $(ii)$ $\mathcal{L}$ comprises a nonsquare binary matrix of which the only ``1'' element in each row one-to-one corresponds to a boundary index.
\begin{definition}
\label{def:discrete_inner_product_norm}
Given two vectors $\bm{u},\bm{v} \in \mathbb{R}^{\cblue{\mathrm N}}$, the following definitions of a \textit{discrete $\mathcal H$-inner product} and its corresponding \textit{discrete $\mathcal H$-norm} are used,
\[
(\bm{u},\bm{v})_{\mathcal{H}} = \bm{u}^T\mathcal{H}\bm{v};\quad (\mathcal{L}\bm{u},\mathcal{L}\bm{v})_{\mathcal{B}} = (\mathcal{L}\bm{u})^T\mathcal{B}(\mathcal{L}\bm{v});\quad \norm{\bm{u}}^2_{\mathcal{H}} = (\bm u,\bm u)_{\mathcal{H}} = \bm u^T\mathcal{H}\bm u.
\]
\end{definition}
\cgreen{The matrix $\mathcal{H}$ is often referred to as the norm matrix}. Advantages of an SBP scheme come from the fact that its discrete operators, by design, mimic the continuous integration-by-parts. The following definitions introduce discrete operators of which properties are analogous to the continuous integration rules \eqref{eq:continuous_intergration_rules}, see \citet{Carpenter1994,Nordstrom2001}.
\begin{definition}
\label{def:sbp_properties_first}
A matrix operator $\mathcal{H}^{-1}\mathcal{Q}$ approximating $\bm{a}\cdot\nabla,\bm{a}\in\mathbb{R}^2$ is said to be an \textit{SBP advection operator} if
\begin{equation}
\label{eq:sbp_properties_first}
\mathcal{Q}+\mathcal{Q}^T = \mathcal{L}^T \mathcal{B} \;\mbox{Diag}(\bm a\cdot\mathbf{n})\; \mathcal{L},
\end{equation}
where $\mbox{Diag}(\bm{\omega})$ denotes the square diagonal matrix of which diagonal is the discrete vector $\bm{\omega}$.
\end{definition}
\begin{definition}
\label{def:sbp_properties_second}
A matrix operator $\mathcal{H}^{-1}\mathcal{R}^{(\epsilon)}$ approximating $\nabla\cdot\epsilon\nabla$ is said to be an \textit{SBP diffusion operator} if
\begin{equation}
\label{eq:sbp_properties_second}
\mathcal{R}^{(\epsilon)} = \mathcal{L}^T\mathcal{B} \mathcal S-\mathcal{A},
\end{equation}
where $\mathcal{A}=\mathcal{A}^T \geq 0$ and $\mathcal S$ is a consistent approximation of $\mathbf{n}\cdot\epsilon\nabla$ at the boundary.
\end{definition}
The discrete energy method can be analogously applied to SBP schemes utilizing the properties \eqref{eq:sbp_properties_first}, \eqref{eq:sbp_properties_second}. A necessary condition for strict stability is that $\frac{d}{dt}\norm{\bm u}_{\mathcal{H}}^2$ being nonpositive given zero boundary data.
We introduce FD notations in the following subsection.
\subsection{The FD scheme}
\label{sec:sbp_form_fdm}
Consider a one-dimensional domain $[x_l,x_r]$ discretized by a grid of $n$ equally spaced points $\{x_l\equiv x_1 < x_2<\dots<x_n \equiv x_r\}$. We first look at one-dimensional SBP operators for FD. The following two definitions (first presented in \cite{Mattsson2012, Mattsson2004}) are central.
\begin{definition}
A FD operator $\bm D_1 = \bm H^{-1} \bm Q$ approximating $\frac{\partial}{\partial x}$ using $q$th-order interior stencils is called a \textit{$q$th-order accurate first derivative SBP operator} if $\bm H=\bm H^T>0$ and $\bm Q+\bm Q^T=\mbox{Diag}(-1,0,\dots,0,1)$.
\end{definition}
\cgreen{$\bm H$ is the one-dimensional FD norm matrix}. Let $\bm e_1 = (1,0,\dots,0)^T,\bm e_n = (0,\dots,0,1)^T$ be column vectors of length $n$, and $\bm d_1,\bm d_n$ be row vectors approximating first derivatives at $x_1,x_n$.
\begin{definition}
A FD operator $\bm D_2^{(\epsilon)} = \bm H^{-1}(-\bm A-\epsilon_1\bm e_1 \bm d_1+\epsilon_n\bm e_n \bm d_n)$ approximating $\epsilon(x)\frac{\partial^2}{\partial x^2}$ using $q$th-order interior stencils is called a \textit{$q$th-order accurate second derivative SBP operator} if $\bm H= \bm H^T>0, \bm A= \bm A^T\geq0$.
\end{definition}
\begin{figure}
\centering
\includegraphics[scale=1]{FD-domain.eps}
\caption{A rectangle domain discretized by an equidistant grid.}
\label{fig:fd_domain}
\end{figure}
SBP FD schemes for multiple spatial dimensions are naturally extended from the one-dimensional operators. Consider a bounded rectangle domain $\Omega = [x_l,x_r]\times[y_l,y_r] \subset \mathbb{R}^2$ illustrated in Figure \ref{fig:fd_domain}. A FD solution is approximated on $n\times m$ equidistant grid points $\{(x_i,y_j), i =1,2,\dots,n, j = 1,2,\dots,m\}$, where
\begin{align*}
x_i&=x_l+(i-1)h_x, \quad h_x = \frac{x_r-x_l}{n-1},\\
y_j&=y_l+(j-1)h_y, \quad h_y = \frac{y_r-y_l}{m-1}.
\end{align*}
A discrete solution vector $\bm{u}$ approximating a continuous function $u:\Omega\longrightarrow\mathbb{R}$ is then presented by
\[
\bm{u} = \begin{bmatrix} u_{1,1},u_{1,2},\dots,u_{1,m}, & u_{2,1},u_{2,2},\dots,u_{2,m}, & \dots & u_{n,1},u_{n,2},\dots,u_{n,m} \end{bmatrix}^T
\]
where $u_{i,j}$ approximates $u(x_i,y_j)$. It's often convenient to interprete $\bm{u}$ as a ``vector of vectors''. The total number of degrees of freedom is $\cblue{\mathrm N}=n \times m$.\\
Discrete operators for the two-dimensional approximation are listed below. Definition of the Kronecker product ``$\otimes$'' can be found in e.g., \citet{Mattsson2010}. Denote the identity matrix by $\bm I$, and the one of size $k\times k$ by $\bm I_k$. The following FD operators will be frequently used in the \cblue{analysis},
\begin{equation}
\label{eq:fdm_2d_operators}
\begin{aligned}
\H_x & = \bm H_x \otimes\bm I_m, & \H_y & =\bm I_n \otimes\bm H_y,\\
\mathbf{D}_x & = \bm D_1 \otimes\bm I_m, & \mathbf{D}_y & = \bm I_n \otimes \bm D_1,\\
\mathbf{D}_{2x} & = \bm D_2^{(b)} \otimes \bm I_m, & \mathbf{D}_{2y} & =\bm I_n \otimes\bm D_2^{(b)},
\end{aligned}
\end{equation}
where $\bm H_x,\bm H_y$ are the one-dimensional norm matrices in $x-$ and $y-$ directions. For ease of reading, it is worth noting that the first components in all of the above Kronecker products are of size $n\times n$, and the second components are of size $m\times m$. Note that the matrix $\bm D_2^{(b)}$ in $\mathbf{D}_{2x},\mathbf{D}_{2y}$ varies for different rows and columns, namely,
\cblue{$\mathbf{D}_{2y}^{(b)} = \mbox{Diag}(\bm D_2^{(\bm b^1)}, \dots, \bm D_2^{(\bm b^n)})$}, where $\bm b^i,i=1,2,\dots,n$, being a vector of length $m$, corresponds to the value of $b(x,y)$ at $(x_i,y_j),j=1,2,\dots,m$. The corresponding two-dimensional norm matrix is $\H:=\H_x\H_y = \bm H_x\otimes\bm H_y$.
A consistent FD semi-discretization of \eqref{eq:advection_diffusion_fdm} using diagonal norm SBP operators is given by
\begin{equation}
\label{eq:advection_diffusion_fdm_approx}
\begin{aligned}
\bm{u}_t = & \; -\alpha_1 \mathbf{D}_x \bm f_1(\bm u) -(1-\alpha_1)\bm F_1' \mathbf{D}_x \bm{u} - \alpha_2 \mathbf{D}_y \bm f_2 (\bm{u}) -(1-\alpha_2)\bm F_2' \mathbf{D}_y \bm u\\
&+(\mathbf{D}_{2x}^{(\epsilon)}+ \mathbf{D}_{2y}^{(\epsilon)})\bm u + \mbox{SAT},
\end{aligned}
\end{equation}
where \cgreen{$\bm f_1, \bm f_1',\bm f_2, \bm f_2'$ are, respectively, the vectors containing values of $f_1,f_1',f_2,f_2'$ evaluated at nodal points with respect to the discrete variable $\bm u$}. The diagonal matrices $\bm F_1'$ and $\bm F_2'$ are defined as $\bm F_1' = \mbox{Diag} (\bm f_1'(\bm u))$, $\bm F_2' = \mbox{Diag} (\bm f_2'(\bm u))$.
The following proposition can be straightforwardly derived from how the above operators are defined, but useful in connection with FE schemes in the next section.
\begin{proposition}
\cred{Let $\nabla \cdot \bm f$ be a linear advection flux, i.e. $\bm f'= \bm a = (a_1,a_2)^T\in\mathbb{R}^2$. }The advection operator $a_1\mathbf{D}_x+a_2\mathbf{D}_y, \bm a = (a_1,a_2) \in \mathbb{R}^2$, and the diffusion operator $\mathbf{D}_{2x}^{(\epsilon)} + \mathbf{D}_{2y}^{(\epsilon)}$ in \eqref{eq:advection_diffusion_fdm_approx} satisfy the summation-by-parts rules \eqref{eq:sbp_properties_first} and \eqref{eq:sbp_properties_second}, with the norm matrix $\H$.
\end{proposition}
Since this property is intended by definition, no proof is given.\\
The normal vectors at the North, East, South, West boundaries are $\mathbf{n}_N=(0,1)^T, \mathbf{n}_E=(1,0)^T,\mathbf{n}_S=(0,-1)^T$ and $\mathbf{n}_W=(-1,0)^T$, respectively. The boundary condition \eqref{eq:advection_diffusion_fem_bc} can be written more precisely at each boundary,
\[
\begin{cases}
\displaystyle\mbox{North}: \frac{1}{2} \left(\alpha_2f_2(u)-\text{sign}(u)\abs{\alpha_2f_2(u)}\right)+\frac{1}{2} \left(\tilde\alpha_2uf_2'(u)-\text{sign}(u)\abs{\tilde\alpha_2 uf_2' (u)}\right) -\epsilon u_y = g_{bN}(t),\\
\displaystyle\mbox{East}: \frac{1}{2} \left(\alpha_1f_1(u)-\text{sign}(u)\abs{\alpha_1f_1(u)}\right)+\frac{1}{2} \left(\tilde\alpha_1uf_1'(u)-\text{sign}(u)\abs{\tilde\alpha_1 uf_1'(u)}\right) -\epsilon u_x = g_{bE}(t),\\
\displaystyle\mbox{South}: \frac{1}{2} \left(-\alpha_2f_2(u)-\text{sign}(u)\abs{\alpha_2f_2(u)}\right)+\frac{1}{2} \left(-\tilde\alpha_2uf_2'(u)-\text{sign}(u)\abs{\tilde\alpha_2uf_2'(u)}\right) + \epsilon u_y = g_{bS}(t),\\
\displaystyle\mbox{West}: \frac{1}{2} \left(-\alpha_1f_1(u)-\text{sign}(u)\abs{\alpha_1f_1(u)}\right)+\frac{1}{2} \left(-\tilde\alpha_1uf_1'(u)-\text{sign}(u)\abs{\tilde\alpha_1uf_1'(u)}\right)+\epsilon u_x = g_{bW}(t),
\end{cases}
\]
where $g_{bN}, g_{bE}, g_{bS}, g_{bW}$ are parts of $g_b(x,y,t)$ on the North, East, South, West boundaries, respectively. We show that the following penalty terms weakly forcing the boundary condition \eqref{eq:advection_diffusion_fem_bc} yield stability to \eqref{eq:advection_diffusion_fdm_approx},
\begin{equation}
\begin{aligned}
\label{eq:advection_diffusion_fdm_sat}
\mbox{SAT}\;\;\, &= \mbox{SAT}_N+\mbox{SAT}_E+\mbox{SAT}_S+\mbox{SAT}_W,\\
\mbox{SAT}_N &= \tau_N\Big[\frac{1}{2}\left(\alpha_2\bm f_2(\bm u_N)-\mbox{Diag}(\text{sign}(\bm u_N))\abs{\alpha_2\bm f_2(\bm u_N)}\right)\\
&\quad\quad+\frac{1}{2}\left(\tilde\alpha_2{\bm U}_N\bm f_2'(\bm u_N)-\mbox{Diag}(\text{sign}(\bm u_N))\abs{\tilde\alpha_2{\bm U}_N\bm f_2'(\bm u_N)}\right)-\bm\mathcal E_N(\mathbf{D}_y\bm u)_N-\bm g_{bN}\Big]\otimes (\bm H_y^{-1}\bm e_m),\\
\mbox{SAT}_E\, &= \tau_E (\bm H_x^{-1} \bm e_n) \otimes \Big[\frac{1}{2}\left(\alpha_1\bm f_1(\bm u_E)-\mbox{Diag}(\text{sign}(\bm u_E))\abs{\alpha_1\bm f_1(\bm u_E)}\right)\\
&\quad\quad+\frac{1}{2}\left(\tilde\alpha_1{\bm U}_E\bm f_1'(\bm u_E)-\mbox{Diag}(\text{sign}(\bm u_E))\abs{\tilde\alpha_1{\bm U}_E\bm f_1'(\bm u_E)}\right)-\bm\mathcal E_E(\mathbf{D}_x\bm u)_E - \bm g_{bE}\Big],\\
\mbox{SAT}_S\; &= \tau_S \Big[\frac{1}{2}\left(-\alpha_2\bm f_2(\bm u_S)-\mbox{Diag}(\text{sign}(\bm u_S))\abs{\alpha_2\bm f_2(\bm u_S)}\right)\\
&\quad\quad+\frac{1}{2}\left(-\tilde\alpha_2{\bm U}_S\bm f_2'(\bm u_S)-\mbox{Diag}(\text{sign}(\bm u_S))\abs{\tilde\alpha_2{\bm U}_S\bm f_2'(\bm u_S)}\right)+\bm\mathcal E_S(\mathbf{D}_y\bm u)_S -\bm g_{bS}\Big] \otimes (\bm H_y^{-1} \bm e_1),\\
\mbox{SAT}_W &= \tau_W (\bm H_x^{-1} \bm e_1) \otimes \Big[\frac{1}{2}\left(-\alpha_1\bm f_1(\bm u_W)-\mbox{Diag}(\text{sign}(\bm u_W))\abs{\alpha_1\bm f_1(\bm u_W)}\right)\\
&\quad\quad+\frac{1}{2}\left(-\tilde\alpha_1{\bm U}_W\bm f_1'(\bm u_W)-\mbox{Diag}(\text{sign}(\bm u_W))\abs{\tilde\alpha_1{\bm U}_W\bm f_1'(\bm u_W)}\right)+\bm\mathcal E_W(\mathbf{D}_x\bm u)_W - \bm g_{bW}\Big],
\end{aligned}
\end{equation}
where $\bm g_{bN},\bm g_{bE},\bm g_{bS},\bm g_{bW}$ are vectors containing values of $g_b$ respectively evaluated at North, East, South, West boundary grid points; $\bm\mathcal E=\mbox{Diag}(\bm\epsilon)$, where $\bm\epsilon$ is the discrete analogue of $\epsilon$ evaluated at the grid points, $\bm\mathcal E_N$, $\bm\mathcal E_E$, $\bm\mathcal E_S$, $\bm\mathcal E_W$ are parts of $\bm\mathcal E$ at the North, East, South, West boundaries; and
\begin{equation}
\label{eq:fdm_2d_operators2}
\hspace*{-0.5cm}
\begin{aligned}
\bm u_N &= (\bm I_n\otimes \bm e_m^T)\bm u, &\bm u_E&=(\bm e_n^T\otimes\bm I_m)\bm u, &\bm u_S&=(\bm I_n\otimes \bm e_1^T)\bm u, &\bm u_W&=(\bm e_1^T\otimes\bm I_m)\bm u,\\
{\bm U}_N &= \mbox{Diag}(\bm u_N), & {\bm U}_E &= \mbox{Diag}(\bm u_E), & {\bm U}_S &= \mbox{Diag}(\bm u_S), & {\bm U}_W &= \mbox{Diag}(\bm u_W),\\
(\mathbf{D}_y \bm u)_N &= (\bm I_n \otimes \bm d_m) \bm u, &(\mathbf{D}_x \bm u)_E&= (\bm d_n \otimes\bm I_m) \bm u, &(\mathbf{D}_y\bm u)_S&=(\bm I_n \otimes \bm d_1) \bm u, &(\mathbf{D}_x \bm u)_W&= (\bm d_1 \otimes\bm I_m) \bm u.
\end{aligned}
\end{equation}
The following theorem completes the formulation of the FD approximation to \eqref{eq:advection_diffusion_fem}, \eqref{eq:advection_diffusion_fem_bc}.
\begin{theorem}
\label{theorem:advection_diffusion_bt_penalties}
By choosing $\tau_N = \tau_E = \tau_S = \tau_W = 1$, the approximation \eqref{eq:advection_diffusion_fdm_approx} with the outer boundary terms \eqref{eq:advection_diffusion_fdm_sat} is stable.
\end{theorem}
\begin{proof}
Let ${\bm U} = \mbox{Diag}(\bm u)$. Deriving a discrete energy estimate from \eqref{eq:advection_diffusion_fdm_approx} and \eqref{eq:advection_diffusion_fdm_sat}, one \cblue{has}
\begin{align*}
&\;\frac{d}{dt} \norm{\bm u}_{\H}^2 + 2\bm u^T(\bm A\otimes \bm H_y)\bm u+ 2\bm u^T(\bm H_x\otimes \bm A)\bm u\\
=&\; -(\mbox{Diag}(\text{sign}(\bm u)) \bm u)^T(\bm H_x\otimes \bm e_1\bm e_1^T+\bm e_m\bm e_m^T)\left(\abs{\alpha_2\bm{f}_2(\bm u)}+\abs{\hat\alpha_2{\bm U}\bm{f}_2'(\bm u)}\right)\\
&-(\mbox{Diag}(\text{sign}(\bm u)) \bm u)^T(\bm e_1\bm e_1^T+\bm e_n\bm e_n^T\otimes \bm H_y)\left(\abs{\alpha_1\bm{f}_1(\bm u)}+\abs{\hat\alpha_1{\bm U}\bm{f}_1'(\bm u)}\right).
\end{align*}
\cpurple{It can be seen that the portion corresponding to each boundary is nonpositive. For example, the one corresponding to the North boundary is $-(\mbox{Diag}(\text{sign}(\bm u)) \bm u)^T(\bm H_x\otimes \bm e_m\bm e_m^T)\left(\abs{\alpha_2\bm{f}_2(\bm u)}+\abs{\hat\alpha_2{\bm U}\bm{f}_2'(\bm u)}\right)$. Therein, the matrix $(\bm H_x\otimes \bm e_m\bm e_m^T)$ is positive definite. The vector $\abs{\alpha_2\bm{f}_2(\bm u)}+\abs{\hat\alpha_2{\bm U}\bm{f}_2'(\bm u)}$ is nonnegative. The vector $\mbox{Diag}(\text{sign}(\bm u)) \bm u = \abs{\bm u}$ is also nonnegative. Therefore, the right-hand side of the above estimate is less than or equal to zero. Strict stability is thus guaranteed. The above estimate is analogous to the continuous energy estimate \eqref{eq:advection-diffusion-continuous-energy}.}
\end{proof}
\subsection{The FE scheme}
\label{sec:sbp_form_fem}
We formulate and prove the SBP properties of the FE methods in this section.
\subsubsection*{The Galerkin FE approximation}
Let $\{\mathcal T_h\}_{h>0}$ be a triangulation mesh of $\Omega$, and $\mathbb{P}^1(K)$ is the space of all linear functions on each element $K$ of $\{\mathcal T_h\}$. Denote $C^0(\Omega)$ the space of continuous functions on $\Omega$. We seek the solution of \eqref{eq:advection_diffusion_fem} in the space of weakly differentiable functions $H^1(\Omega)$. To this purpose, we define a scalar-valued Lagrange FE space
\[
V_h:= \{ v(\cdot,t) \in C^0(\Omega)\mid v|_K \in \mathbb{P}^1(K), \forall K \in \mathcal T_h \}.
\]
\cblue{Note that all the degrees of freedom of continuous Galerkin FE methods are located on the nodal points. More specifically, using $\mathbb{P}^1$ elements, they are the vertices of the triangulation $\mathcal T_h$}. A FE approximation of \eqref{eq:advection_diffusion_fem} subject to the boundary condition \eqref{eq:advection_diffusion_fem_bc} can then be formulated as: find $u_h \in V_h$ such that
\begin{equation}
\label{eq:gfem_1}
\begin{aligned}
(\partial_tu_h,v)+\cgreen{( \bm\alpha\nabla \cdot \bm{f}(u_h) + \bm f'(u_h) \cdot \bm{\tilde\alpha}\nabla u_h, v)} &\\
+(\epsilon\nabla u_h, \nabla v) =
&\;\frac{1}{2}\int_{\partial\Omega}\big((\bm\alpha\mathbf{n})\cdot \bm{f}(u_h)-\text{sign}(u_h)\abs{(\bm\alpha\mathbf{n})\cdot \bm{f}(u_h)}\big)vds\\
&+\frac{1}{2}\int_{\partial\Omega}\big((\bm{\tilde\alpha}\mathbf{n})\cdot (u_h\bm{f}'(u_h))-\text{sign}(u_h)\abs{(\bm{\tilde\alpha}\mathbf{n})\cdot (u_h\bm{f}'(u_h))}\big)v ds\\
&-\int_{\partial\Omega} g_bv ds, \;\forall v \in V_h.
\end{aligned}
\end{equation}
We can write $V_h = \mbox{span}\{\varphi_i\}_{i=1}^{\cblue{\mathrm N}}$, where $\varphi_i,i=1,\dots,\cblue{\mathrm N}$ are piecewise linear Lagrange basis functions. Inserting the ansatz $u_h = \displaystyle\sum_{j=1}^{\cblue{\mathrm N}}\xi_j(t)\varphi_j(x,y), \xi_i \in \mathbb{R},j=1,2,\dots,\cblue{\mathrm N}$ and $v = \varphi_i,i=1,2,\dots,\cblue{\mathrm N}$ into \eqref{eq:gfem_1} gives
\begin{equation}
\label{eq:gfem_matrix}
\begin{aligned}
\mathbf{M} \dot{\bm{\xi}}(t) = -\mathbf{C}(u_h)\bm{\xi}(t) - \mathbf{A}\bm{\xi}(t)+\frac{1}{2}\mathbf{R}_M\left((\bm\alpha\mathbf{n})\cdot{\bm f}(\bm{\xi}(t))\right)-\frac{1}{2}\mathscr{R}_M(\text{sign}(u_h))\abs{(\bm\alpha\mathbf{n})\cdot{\bm f}(\bm{\xi}(t))}\\
+\frac{1}{2}\mathscr{R}_M\left((\bm{\tilde\alpha}\mathbf{n})\cdot \bm f'(u_h)\right)\bm{\xi}(t)-\frac{1}{2}\mathscr{R}_M\left(\abs{(\bm{\tilde\alpha}\mathbf{n})\cdot \bm f'(u_h)}\right)\bm{\xi}(t)-\bm{r},
\end{aligned}
\end{equation}
where the over dot denotes the time derivative; $\mathbf{M},\mathbf{C}(\phi),\mathbf{A},\mathbf{R}_M,\mathscr{R}_M(\phi),\bm{r}$ are respectively called the mass matrix, the convective flux matrix, the stiffness matrix, the Robin boundary mass matrix, the nonlinear Robin boundary mass matrix, and the Robin boundary vector. A different notation style is used for $\mathscr{R}_M(\phi)$ to indicate that each entry corresponds to a $\phi$-weighted line integral. The matrices are of full size \cblue{$\mathrm N\times \mathrm N$}. To be more precise, for $i,j=1,\dots,\cblue{\mathrm N}$,
\begin{align*}
\mathbf{M}_{ij} & = (\varphi_j,\varphi_i), &
\mathbf{C}_{ij}(\phi) & = (\bm f'(\phi)\cdot\bm\alpha\nabla\varphi_j,\varphi_i)+(\tilde{\bm\alpha}\nabla\cdot (\bm f(\phi)\varphi_j),\varphi_i),&
\mathbf{A}_{ij} & = (\epsilon\nabla\varphi_j,\nabla\varphi_i),\\
\left(\mathbf{R}_M\right)_{ij} & = \int_{\partial\Omega}\varphi_j\varphi_ids, &
\left(\mathscr{R}_M\right)_{ij}(\phi) & = \int_{\partial\Omega}\phi\varphi_j\varphi_ids, &
\bm{r}_i & = \int_{\partial\Omega} g_{\cgreen b}\varphi_i ds. &
\end{align*}
\cblue{As in the standard continuous Galerkin method, given that $\phi, g_b, \epsilon$ present nodal weights, the above integrals can be evaluated exactly, see Remark \ref{remark:fe_quadrature}.} \cgreen{We write $\mathbf C$ instead of $\mathbf C(\phi)$ when $\phi \equiv 1$.} Pseudo-code for the assembly of the above matrices can be found in \citet{Larson2009}, and \citet{Dao2019}.
\cblue{
\begin{remark}\label{remark:fe_quadrature}
To clarify, we use the following quadrature rule to calculate the $\phi$-, $g_b$-, $\epsilon$-dependent matrices,
\[
\int_{K} \phi\varphi_j\varphi_idx\approx \frac{\phi(x_i)+\phi(x_j)}{2} \int_{K}\varphi_j\varphi_idx,
\]
for each element $K$ on $\mathcal T_h$, and
\begin{equation}\label{eq:quadrature_rule}
\int_{E} \phi\varphi_j\varphi_ids\approx \frac{\phi(x_i)+\phi(x_j)}{2} \int_{E}\varphi_j\varphi_ids,
\end{equation}
for each edge $E$ on the boundary $\Gamma$. The above integrals are exact if $\phi, g_b, \epsilon$ are considered as nodal weights. Later we will show that the stability proofs do not rely on these quadrature rules. Thus, any other consistent rules could have been used. The use of \eqref{eq:quadrature_rule} merely simplifies the SATs for coupling FE.
\end{remark}}
\begin{proposition}
Let $\nabla \cdot \bm f$ be a linear advection flux, i.e., $\bm f'= \bm a = (a_1,a_2)^T\in\mathbb{R}^2$. Then, the advection operator $\mathbf{M}^{-1}\mathbf{C}$ and the diffusion operator $-\mathbf{M}^{-1}\mathbf{A} + \mathbf{M}^{-1}\mathbf{R}_C^{\partial \Omega},\mathbf{R}^{\partial \Omega}_{C\,ij} = \int_{\partial \Omega} \mathbf{n}\cdot \epsilon\nabla\varphi_j \varphi_i ds$, $i,j=1,2,\dots,\cblue{\mathrm N}$ used in \eqref{eq:gfem_matrix} satisfy the summation-by-parts rules \eqref{eq:sbp_properties_first} and \eqref{eq:sbp_properties_second}, with the norm matrix $\mathbf{M}$.
\end{proposition}
\begin{proof}
We prove that the FE advection operator under the general notation $\mathcal{Q}=\mathbf{C}$ satisfies \eqref{eq:sbp_properties_first}. By integration-by-parts, we have $\mathbf{C}+\mathbf{C}^T=\mathscr{R}_M(\bm a\cdot \bm n)$. We need to show that $\mathscr{R}_M(\bm a\cdot \bm n)=\mathcal{L}^T \mathcal{B} \;\mbox{Diag}(\bm a\cdot\mathbf{n})\; \mathcal{L}$ with the FE formulation of $\mathcal{B}$ and $\mathcal{L}$. For a general boundary selection operator $\mathcal{L}$, each row of $\mathcal{L}$ must be a one-to-one mapping from its ``one'' element with a node on the boundary. The combination of a left multiplier $\mathcal{L}^T$ and a right multiplier $\mathcal{L}$ thus expands the local boundary operators into the full domain size (with zeros on the extended rows and columns). Relation \eqref{eq:sbp_properties_first} becomes obvious with the local boundary integration operator inserted $\mathcal{B}_{ij}\equiv\left(\mathbf{R}_M\right)_{ij}=\int_{\partial\Omega}\varphi_j\varphi_ids$, for $i,j=1,2,\dots,\cblue{\mathrm N_{ \partial \Omega}}$.
For the diffusion operator $\mathcal{R}=-\mathbf{A}+\mathbf{R}_C^W$, we need to show that $-\mathbf{A}+\mathbf{R}_C^{\partial \Omega}=\mathcal{L}^T\mathcal{B} \mathcal S-\mathcal{A}$ with the FE formulation of $\mathcal{B},\mathcal{L},\mathcal S$, and $\mathcal{H}$. Firstly, we identify $\widetilde{\mathcal S}(u_h) \in V_h, u_h \in V_h$ approximating $\mathbf{n}\cdot\epsilon\nabla u_h$. The operator $\widetilde{\mathcal S}$ satisfies
\[
(\widetilde{\mathcal S}(u_h)-\mathbf{n}\cdot\epsilon\nabla u_h,v)_{\partial\Omega} = 0,\, \forall v \in V_h.
\]
Since $\widetilde{\mathcal S}(u_h) \in V_h$, there exists $\bm{\eta}=(\eta_1,\eta_2,\dots,\eta_{\cblue{\mathrm N}})$ such that $\widetilde{\mathcal S}(u_h) = \sum_{j=1}^{\cblue{\mathrm N}} \eta_j \varphi_j$. Inserting $\widetilde{\mathcal S}(u_h) = \sum_{j=1}^{\cblue{\mathrm N}} \eta_j \varphi_j$ and $u_h=\sum_{j=1}^{\cblue{\mathrm N}}\xi_j\varphi_j$ into the above equation gives
\begin{equation}
\label{eq:fe_sbp_properties_3}
\tilde{\mathcal{B}}\bm\eta = \mathbf{R}_C^{\partial \Omega} \bm{\xi},
\end{equation}
where $\tilde{\mathcal{B}}_{ij}=\int_{\partial\Omega}\varphi_j\varphi_ids,i,j=1,2,\dots,\cblue{\mathrm N}$ and $\mathbf{R}_C^{\partial \Omega}$ is defined above. Since $\int_{\partial\Omega}\varphi_j\varphi_ids = 0$ if either $i$ or $j$ corresponds to a nonboundary node,
\begin{equation}
\label{eq:fe_sbp_properties_4}
\mathcal{L}^T\mathcal{B}\mathcal{L} = \tilde{\mathcal{B}}.
\end{equation}
The multiplication of $\mathcal{L}^T$ and $\mathcal{L}$ is because $\mathcal{B}$ is defined in the SBP framework of size \cblue{$\mathrm N_{\partial\Omega}\times\mathrm N_{\partial\Omega}$}. The matrix $\mathcal S$ is obtained by projecting $\widetilde{\mathcal S}$ onto the boundary, $\mathcal S \bm{\xi}= \mathcal{L}\bm\eta$. Note that $\mathcal S$ is not square because of the matrix dimensions in \eqref{eq:sbp_properties_second}. From \eqref{eq:fe_sbp_properties_3} and \eqref{eq:fe_sbp_properties_4}, we obtain
\begin{equation}
\label{eq:fe_sbp_properties_1}
\mathbf{R}_C^{\partial \Omega} = \mathcal{L}^T\mathcal{B} \mathcal S.
\end{equation}
What is left is to prove the positive semi-definiteness of $\mathbf{A}$. Let $\Psi(u,v):=(\epsilon\nabla u,\nabla v)_\Omega, u,v \in V$. Given $\bm{\nu} = (\nu_1,\nu_2,\dots,\nu_\cblue{\mathrm N})^T\in\mathbb{R}^{\cblue{\mathrm N}}$ and $v_h = \sum_{j=1}^{\cblue{\mathrm N}}\nu_j\varphi_j$, we have
\begin{equation}
\label{eq:fe_sbp_properties_2}
\hspace{-.5cm}
\bm\nu^T \mathbf{A} \bm\nu = \sum_{i=1}^{\cblue{\mathrm N}} \sum_{j=1}^{\cblue{\mathrm N}} \nu_i \mathbf{A}_{ij} \nu_j = \sum_{i=1}^{\cblue{\mathrm N}} \sum_{j=1}^{\cblue{\mathrm N}} \Psi(\nu_i \varphi_i, \nu_j\varphi_j) = \Psi \left(\sum_{i=1}^{\cblue{\mathrm N}} \nu_i \varphi_i, \sum_{j=1}^{\cblue{\mathrm N}} \nu_j\varphi_j\right) = \Psi (v_h,v_h) = \norm{\nabla v}^2_\epsilon \geq 0
\end{equation}
for all $v_h \in V_h$. Therefore, the desired property of $\mathcal{R}$ is achieved.\end{proof}
\begin{remark}
By the energy method, conservation property of the discrete fluxes \eqref{eq:advection_diffusion_fdm_approx} and \eqref{eq:gfem_matrix} approximating the continuous skew-symmetric flux \eqref{eq:skew_symmetric_assumption} is fulfilled under the following condition
\begin{equation}\label{eq:skew_symmetric_sufficient_discrete}
\begin{cases}
\alpha_1\bm f_1 - \tilde\alpha_1 \bm F_1' \bm u = 0,\\
\alpha_2\bm f_2 - \tilde\alpha_2 \bm F_2' \bm u = 0,
\end{cases}
\end{equation}
which naturally holds if the condition \eqref{eq:skew_symmetric_sufficient} of the flux $\bm f$ holds for every point in space.
\end{remark}
\section{Simultaneous-approximation-term technique for interface coupling}
\label{sec:coupling}
This section describes our proposed SAT technique for interface treatment. To make this paper self-contained, we derive the coupling of both FD--FD and FD--FE. The former one is not the main focus, and partially covered in the literature, see \citet{Lundquist2018,Mattsson2010}. \cgreen{Unlike the mentioned references which investigate linear problems, the technique presented in this paper is extended to nonlinear problems, as $\bm f$ can be nonlinear in \eqref{eq:advection_diffusion_fem}}. Description of compatible \textit{SBP-preserving interpolation operators} for nondiagonal norm schemes is included.
\subsection{Continuous analysis}
\label{sec:coupling_continuous}
For simplicity, the analysis is done on the coupling of two domains $\Omega_L$, $\Omega_R$ illustrated in Figure \ref{fig:coupling-interface}, where $u$ and $v$ present the parts of solution on the left and right domains, respectively,
\begin{equation}
\label{eq:advection_diffusion_coupling_continuous}
\begin{aligned}
\frac{\partial u}{\partial t} + \nabla \cdot \bm{f}(u) & = \nabla \cdot (\epsilon \nabla u), & (x,y) \in \Omega_L,\\
\frac{\partial v}{\partial t} + \nabla \cdot \bm{f}(v) & = \nabla \cdot (\epsilon \nabla v), & (x,y) \in \Omega_R.
\end{aligned}
\end{equation}
\begin{figure}[pos=h]
\centering
\includegraphics[scale=1]{coupling-interface.eps}
\caption{Coupling of two adjoining blocks.}
\label{fig:coupling-interface}
\end{figure}
The North, South, West boundaries of $\Omega_L$ and the North, East, South boundaries of $\Omega_R$ will be referred to as the ``outer boundary'' and the shared edge of the two adjoining blocks \cpurple{is referred to} the ``interface'', denoted $\Omega_I$. \cblue{With the setup in Figure \ref{fig:coupling-interface}, $\Omega_I$ can be seen both as the East boundary of $\Omega_L$ and the West boundary of $\Omega_R$}. Applying the energy method on \eqref{eq:advection_diffusion_coupling_continuous}, a total energy estimate of the whole solution on $\Omega_L \cup \Omega_R$ reads
\[
\frac{d}{dt}\left(\norm{u}_{\Omega_L}^2+\norm{v}_{\Omega_R}^2\right) +2\norm{\nabla u}^2_{\Omega_L,\epsilon} +2\epsilon\norm{\nabla v}^2_{\Omega_R,\epsilon} = BT + IT,
\]
where $BT$ contains all the outer boundary terms being bounded with proper boundary treatment as discussed in Section \ref{sec:sbp_operators}, and $IT$ contains the interface terms given by
\begin{align*}
\hspace{-0.5cm}
IT & = \; -\int_{\Omega_I} \alpha_1 u f_1(u) ds+\int_{\Omega_I} \alpha_1 v f_1(v)ds -\int_{\Omega_I} \tilde\alpha_1 u^2 f_1'(u) ds+\int_{\Omega_I} \tilde\alpha_1 v^2 f_1'(v)ds + 2(u,\epsilon u_x)_{\Omega_I}-2(v,\epsilon v_x)_{\Omega_I}.
\end{align*}
Therefore, a sufficient condition for an accurate and energy-conserved coupling is
\begin{equation}
\label{eq:advection_diffusion_continuity}
\begin{cases}
u = v,\\
u_x = v_x
\end{cases}
\mbox{ at } x \in \Omega_I.
\end{equation}
\subsection{Necessary properties of interface interpolation for stability and conservation}
\label{sec:coupling_conditions}
Let $\bm u, \bm v$ be some discrete approximate solutions of $u, v$. An SBP-SAT semi-discretization of \eqref{eq:advection_diffusion_coupling_continuous} is given by
\begin{equation}
\label{eq:advection_diffusion_sbp_sat}
\begin{aligned}
\bm u_t + \mathcal{H}_L^{-1}\mathcal{Q}_L\bm\alpha\bm{f}(\bm u) + \tilde{\bm\alpha}\bm {F}'(\bm u)\mathcal{H}_L^{-1}\mathcal{Q}_L\bm u = \mathcal{H}_L^{-1}\mathcal{R}_L^{(\epsilon)}\bm u + \mbox{SAT}_L+\mbox{SAT}_{IL},\\
\bm v_t + \mathcal{H}_R^{-1}\mathcal{Q}_R\bm\alpha\bm{f}(\bm v) + \tilde{\bm\alpha}\bm {F}'(\bm v)\mathcal{H}_R^{-1}\mathcal{Q}_R\bm v = \mathcal{H}_R^{-1}\mathcal{R}_R^{(\epsilon)}\bm v + \mbox{SAT}_R+\mbox{SAT}_{IR},
\end{aligned}
\end{equation}
where $\mbox{SAT}_{L},\mbox{SAT}_{R}$ present the weak boundary treatment, and $\mbox{SAT}_{IR},\mbox{SAT}_{IR}$ present the weak interface treatment. Another advantage of this weakly forcing technique is that $\mbox{SAT}_{R},\mbox{SAT}_{L},\mbox{SAT}_{IR},\mbox{SAT}_{IL}$ each can be treated separately by the energy method. Therefore, in the following subsections, we assume stable outer boundary treatment and only analyze the estimate terms regarding the additional involvement of $\mbox{SAT}_{IL},\mbox{SAT}_{IR}$.
\begin{figure}[pos=h]
\centering
\includegraphics[scale=1.2]{nonmatching-interface.eps}
\caption{Example of translating nodal values on a nonmatching interface using interpolation operators.}
\label{fig:nonmatching-interface}
\end{figure}
To prove stability and energy conservation, making use of the following class of interpolation operators to translate the shared-interface \cgreen{nodal values}, first introduced in \citet{Mattsson2010}, is necessary.
\cpurple{
\begin{definition}
\label{def:conservation_property_fd_fd}
Let $\mathcal{H}^L_y,\mathcal{H}^R_y$ be the integration operators over the interface of the left and the right domains, respectively. Let $\bm{I}_{L2R}$ and $\bm{I}_{R2L}$ be two $p$-th order accurate interpolation operators, where $\bm{I}_{L2R}$ maps the interface values from the left domain to the right domain, and $\bm{I}_{R2L}$ does the opposite direction,
\begin{equation}
\label{eq:accuracy_requirement}
\bm u_I = \bm{I}_{R2L} \bm v_I + \mathcal{O}(h^p), \bm v_I = \bm{I}_{L2R} \bm u_I + \mathcal{O}(h^p).
\end{equation}
We call $\left\{\bm{I}_{L2R}, \bm{I}_{R2L}\right\}$ a pair of \textit{SBP-preserving interpolation operators} if
\begin{equation}
\label{eq:conservation_property_fd_fd}
\mathcal{H}_y^R\bm{I}_{L2R} = \bm{I}_{R2L}^T\mathcal{H}_y^L,
\end{equation}
where $p = \min(p_L,p_R)$, and $p_L,p_R$ are the orders of accuracy in boundary closure of the two involving methods.
\end{definition}}
The matrix mappings $\bm{I}_{L2R}$ and $\bm{I}_{R2L}$ are illustrated in Figure \ref{fig:nonmatching-interface}. For FD, $\mathcal{H}_y$ is the $\bm H_y$-norm we defined earlier. For FV, this operator is a diagonal matrix containing the Euclid distances between two neighboring interface nodes, \citet{Lundquist2018}. For dG, this operator is a diagonal matrix presenting integration weights for interface edges, \citet{Kozdon2016}. However, it is well-known that in continuous Galerkin FE schemes, the one-dimensional norm matrix, i.e., the mass matrix, is not diagonal. In Appendix \ref{section:construct_interpolation_operators}, we present two techniques of constructing the pair $\left\{\bm{I}_{L2R},\bm{I}_{R2L}\right\}$ which satisfy and \eqref{eq:accuracy_requirement} and \eqref{eq:conservation_property_fd_fd}. The first technique Appendix \ref{section:construction_by_structure} is similar to
\citet{Mattsson2010}. \cgreen{Appendix \ref{section:shortcut_diagonal_norm} describes a way of acquiring the target interpolation operators through existing diagonal norm techniques. This construction is beneficial due to the fact that various techniques to construct the necessary diagonal norm interpolation matrices can be found in the literature, e.g., \citet{Almquist2018,Kozdon2016,Lundquist2018,Lundquist2020,Mattsson2010}.}
\subsection{The FD--FD coupling}
We consider the case where FD schemes are used on both $\Omega_L$ and $\Omega_R$. \cgreen{Let $\bm\omega \in \mathbb R^{\mathrm N}$ be an arbitrary FD solution. We denote by $\bm\omega_I\in\mathbb R^m$ the part of $\bm\omega$ on $\Omega_I$, where $m$ grid points are distributed on $\Omega_I$. For example, in the single-domain notations \eqref{eq:fdm_2d_operators2}, the orientation of $\Omega_L, \Omega_R$ gives that $\bm u_I \equiv \bm u_E, \bm v_I \equiv \bm v_W, (\mathbf{D}_x\bm u)_I \equiv (\mathbf{D}_x\bm u)_E, (\mathbf{D}_x\bm v)_I \equiv (\mathbf{D}_x\bm v)_W$}. The coupling SAT treatment is given by
\cblue{
\begin{equation}
\label{eq:advection_diffusion_sat_interface}
\hspace{-0.5cm}
\begin{aligned}
\mbox{SAT}_{IL} & = \mbox{SAT}_{IL}^{f} + \mbox{SAT}_{IL}^{f'} + \mbox{SAT}_{IL}^{\epsilon},\\
\mbox{SAT}_{IR} & = \mbox{SAT}_{IR}^{f} + \mbox{SAT}_{IR}^{f'} + \mbox{SAT}_{IR}^{\epsilon},
\end{aligned}
\end{equation}
where $\mbox{SAT}_{IL}^{f}, \mbox{SAT}_{IR}^{f}$ weakly couple the flux $f(u) = f (v)$ at $\Omega_I$
\begin{align*}
\mbox{SAT}_{IL}^{f} & = \alpha_L({\bm H_x^L}^{-1} \bm e_n)\otimes \left[(\alpha_1\bm f_1(\bm u))_I-\bm{I}_{R2L}(\alpha_1\bm f_1(\bm v))_I\right],\\
\mbox{SAT}_{IR}^{f} & = \alpha_R({\bm H_x^R}^{-1} \bm e_1) \otimes \left[(\alpha_1\bm f_1(\bm v))_I-\bm{I}_{L2R}(\alpha_1\bm f_1(\bm u))_I\right].
\end{align*}
The terms $\mbox{SAT}_{IL}^{f'}, \mbox{SAT}_{IR}^{f'}$ weakly couple $u=v$ at $\Omega_I$ with the weights of $\alpha_1 f'(u),\alpha_1 f'(v)$,
\begin{align*}
\mbox{SAT}_{IL}^{f'} & =\beta_L({\bm H_x^L}^{-1} \bm e_n)\otimes \left[(\tilde\alpha_1\bm F_1'(\bm u))_I\left(\bm u_I-\bm{I}_{R2L}\bm v_I\right)\right],\\
\mbox{SAT}_{IR}^{f'} & = \beta_R({\bm H_x^R}^{-1} \bm e_1) \otimes \left[(\tilde \alpha_1 \bm F_1'(\bm v))_I\left(\bm v_I-\bm{I}_{L2R}\bm u_I\right)\right],
\end{align*}
where $(\tilde\alpha_1\bm F_1'(\bm u))_I = \mbox{Diag}(\tilde\alpha_1\bm f_1'(\bm u))_I, (\tilde \alpha_1 \bm F_1'(\bm v))_I = \mbox{Diag}(\tilde\alpha_1\bm f_1'(\bm v))_I$. Finally, the terms $\mbox{SAT}_{IL}^{\epsilon}, \mbox{SAT}_{IR}^{\epsilon}$ impose the condition $u_x = u_y$ at $\Omega_I$
\begin{align*}
\mbox{SAT}_{IL}^{\epsilon} & = \delta_L ({\bm H_x^L}^{-1} \bm e_n) \otimes \left[\bm\mathcal E_{IL}(\mathbf{D}_x\bm u)_I - \bm{I}_{R2L}\bm\mathcal E_{IR}(\mathbf{D}_x\bm v)_I\right]\\
&+ \sigma_L ({\bm H_x^L}^{-1}\bm d_n^T)\otimes\left[\bm\mathcal E_{IL}(\bm u_I-\bm{I}_{R2L}\bm v_I)\right],\\
\mbox{SAT}_{IR}^{\epsilon} & = \delta_R ({\bm H_x^R}^{-1} \bm e_1) \otimes\left[\bm\mathcal E_{IR}(\mathbf{D}_xv)_I-\bm{I}_{L2R}\bm\mathcal E_{IR}(\mathbf{D}_x\bm u)_I\right]\\
&+ \sigma_R ({\bm H_x^R}^{-1} \bm d_1^T) \otimes\left[\bm\mathcal E_{IR}(\bm v_I-\bm{I}_{L2R}\bm u_I)\right],
\end{align*}
where ${\bm\mathcal E}_{IL} = \mbox{Diag}(\bm\epsilon_{IL}),{\bm\mathcal E}_{IR}= \mbox{Diag}(\bm\epsilon_{IR})$; $\bm\epsilon_{IL}, \bm\epsilon_{IR}$ respectively contain the viscosity coefficients of the left and the right domain evaluated at the grid points on $\Omega_I$. The terms $\mbox{SAT}_{IL}^{\epsilon}, \mbox{SAT}_{IR}^{\epsilon}$ naturally allow interface jumps of the viscosity $\epsilon$, i.e., when $\bm\epsilon_{IL}\neq \bm\epsilon_{IR}$.} \cgreen{This property can be useful in applications involving discontinuous media, or when a significant difference in the amount of artificial viscosity is needed to stabilize the coupled schemes.}\\
\begin{remark}
Though $\bm e_1,\bm e_n,\bm e_m,\bm d_1,\bm d_n,\bm d_m$ all differ in the $x-$ and $y-$ dimensions, or in the left and right domains, the same notations are used to increase readability.
\end{remark}
\begin{theorem}
\label{theorem:choice_linear_fd_fd}
By choosing $\cblue{\alpha_L=\beta_L=1/2, \alpha_R=\beta_R=-1/2}, \delta_L=-\sigma_R, \sigma_L =-\delta_R$ and $\delta_L+\sigma_L = -1$, assuming proper outer boundary treatment, the approximation \eqref{eq:advection_diffusion_sbp_sat} with the interface SATs \eqref{eq:advection_diffusion_sat_interface} is stable.
\end{theorem}
\begin{proof}
Applying the energy method on each equation of \eqref{eq:advection_diffusion_sbp_sat} and adding them \cblue{together}, a total energy estimate can be written as
\begin{align*}
\frac{d}{dt} \norm{\bm u}_{\H}^2 + 2\bm u^T(\bm{A}\otimes \bm H_y^L)\bm u + 2\bm u^T(\bm H_x^L\otimes \bm{A})\bm u&\\+ \frac{d}{dt} \norm{\bm v}_{\H}^2 + 2\bm v^T(\bm{A}\otimes \bm H_y^R)\bm v+ 2\bm v^T(\bm H_x^R\otimes \bm A)\bm v & = BT + IT,
\end{align*}
where the outer boundary terms all contained in $BT$ are assumed to be finely tuned as given in \eqref{eq:advection_diffusion_fdm_sat} and Theorem \ref{theorem:advection_diffusion_bt_penalties}; and the interface terms consist of
\cblue{
\[
IT = IT^f + IT^{f'} + IT^{\epsilon} + IT^f_{SAT} + IT^{f'}_{SAT} + IT^{\epsilon}_{SAT},
\]
where $IT^f, IT^{f'}$ estimate the energy growth contributed by the discrete operators approximating the flux $\nabla\cdot \bm f(u)$; $IT^{\epsilon}$ concerns the energy growth by the operators approximating $\nabla \cdot (\epsilon \nabla u)$; the terms $IT^f_{SAT}, IT^{f'}_{SAT}, IT^{\epsilon}_{SAT}$ are the contributions of the added SATs \eqref{eq:advection_diffusion_sat_interface}. Concretely, the terms $\mbox{SAT}_{IL}^{f}, \mbox{SAT}_{IR}^{f}$, resulting in the estimate $IT^f_{SAT}$, are used to control $IT^f$,
\begin{align*}
IT^f = & -\bm u^T(\bm e_n\bm e_n^T\otimes \bm H_y^L)(\alpha_1\bm f_1(\bm u))+\bm v^T(\bm e_1\bm e_1^T\otimes \bm H_y^R)(\alpha_1\bm f_1(\bm v)),\\
IT^{f}_{SAT} =&\;2\alpha_L\bm u^T(\bm e_n\bm e_n^T\otimes \bm H_y^L)\alpha_1\bm f_1(\bm u)-2\alpha_L\bm u^T(\bm e_n\bm e_1^T\otimes \bm H_y^L\bm{I}_{R2L})\alpha_1\bm f_1(\bm v)\\
&+2\alpha_R\bm v^T(\bm e_1\bm e_1^T\otimes \bm H_y^R)\alpha_1\bm f_1(\bm v)-2\alpha_R\bm v^T(\bm e_1\bm e_n^T\otimes \bm H_y^R\bm{I}_{L2R})\alpha_1\bm f_1(\bm u).
\end{align*}
The estimate $IT^f$ is derived from the SBP property in Definition \ref{sec:sbp_form_fdm}. By substituting $\alpha_L=1/2, \alpha_R=-1/2$, one can see that $IT^f$ is canceled by $IT^f_{SAT}$. Accordingly, the terms $\mbox{SAT}_{IL}^{f'}, \mbox{SAT}_{IR}^{f'}$ are used to control $IT^{f'}$
\begin{align*}
IT^{f'} = & -\bm u^T(\bm e_n\bm e_n^T\otimes \bm H_y^L)(\bm U\tilde\alpha_1\bm f_1'(\bm u))+\bm v^T(\bm e_1\bm e_1^T\otimes \bm H_y^R)(\bm V\tilde\alpha_1\bm f_1'(\bm v)),\\
IT^{f'}_{SAT} =&\;2\beta_L\bm u^T(\bm e_n\bm e_n^T\otimes \bm H_y^L)(\bm U\tilde\alpha_1\bm f_1'(\bm u))-2\beta_L\bm v^T(\bm e_1\bm e_n^T\otimes \bm{I}_{R2L}^T\bm H_y^L)(\bm U\tilde\alpha_1\bm f_1'(\bm u))\\
&+2\beta_R\bm v^T(\bm e_1\bm e_1^T\otimes \bm H_y^R)(\bm V\tilde\alpha_1\bm f_1'(\bm v))-2\beta_R\bm u^T(\bm e_n\bm e_1^T\otimes \bm{I}_{L2R}^T\bm H_y^R)(\bm V\tilde\alpha_1\bm f_1'(\bm v)),
\end{align*}
in which $IT^{f'}$ is also canceled by $IT^{f'}_{SAT}$ with $\beta_L=1/2, \beta_R=-1/2$. The remaining mixed terms containing $\bm{I}_{L2R}, \bm{I}_{R2L}$ are
\begin{equation}\label{eq:thm_choice_fd_fd_mixed_terms}
\hspace{-0.5cm}
\begin{aligned}
IT^f + IT^{f'} + IT^f_{SAT} + IT^{f'}_{SAT} = & -2\alpha_L\bm u^T(\bm e_n\bm e_1^T\otimes \bm H_y^L\bm{I}_{R2L})\alpha_1\bm f_1(\bm v) -2\beta_R\bm u^T(\bm e_n\bm e_1^T\otimes \bm{I}_{L2R}^T\bm H_y^R)(\bm V\tilde\alpha_1\bm f_1'(\bm v))\\
&-2\beta_L\bm v^T(\bm e_1\bm e_n^T\otimes \bm{I}_{R2L}^T\bm H_y^L)(\bm U\tilde\alpha_1\bm f_1'(\bm u))-2\alpha_R\bm v^T(\bm e_1\bm e_n^T\otimes \bm H_y^R\bm{I}_{L2R})\alpha_1\bm f_1(\bm u).
\end{aligned}
\end{equation}
The SBP property \eqref{eq:conservation_property_fd_fd} implies that $\bm{I}_{L2R}^T\bm H_y^R = \bm H_y^L\bm{I}_{R2L}$. Moreover, we have $\alpha_1 \bm f_1(\bm u) = \bm U\tilde\alpha_1\bm f_1'(\bm u)$ and $\alpha_1 \bm f_1(\bm v) = \bm V\tilde\alpha_1\bm f_1'(\bm v)$ by the discrete skew-symmetry property \eqref{eq:skew_symmetric_sufficient_discrete}. The above choice of the penalty parameters makes $\alpha_L = - \beta_R$ and $\beta_L = - \alpha_R$. Therefore, the sum of two terms in the first line of \eqref{eq:thm_choice_fd_fd_mixed_terms} is zero. The sum of the two remaining terms in the second line is also zero. $SAT^f_{IL}$ and $SAT^{f'}_{IR}$ together imply conservation to the interface interpolation due to the discrete skew-symmetry \eqref{eq:skew_symmetric_sufficient_discrete}. The same can be said for the other pair $SAT^f_{IR}$ and $SAT^{f'}_{IL}$. Similarly, the estimate $IT^{\epsilon} _{SAT}$ of $\mbox{SAT}_{IL}^{\epsilon} , \mbox{SAT}_{IR}^{\epsilon} $ are used to control $IT^{\epsilon}$
\begin{align*}
IT^{\epsilon} = & \; \bm u^T(\bm\mathcal E_L\bm e_n\bm d_n+(\bm\mathcal E_L\bm e_n\bm d_n)^T\otimes \bm H_y^L)\bm u+ \bm v^T(-\bm\mathcal E_R\bm e_1\bm d_1-(\bm\mathcal E_R\bm e_1\bm d_1)^T\otimes \bm H_y^R)\bm v\\
IT^{\epsilon}_{SAT} = & \;\delta_L\left[\bm u^T(\bm\mathcal E_L\bm e_n\bm d_n+(\bm\mathcal E_L\bm e_n\bm d_n)^T\otimes \bm H_y^L)\bm u-\bm u^T(\bm\mathcal E_R\bm e_n\bm d_1+(\bm\mathcal E_R\bm e_n\bm d_1)^T\otimes \bm H_y^L\bm{I}_{R2L})\bm v\right]\\
& +\sigma_L\left[\bm u^T(\bm\mathcal E_L\bm e_n\bm d_n+(\bm\mathcal E_L\bm e_n\bm d_n)^T\otimes \bm H_y^L)\bm u-\bm u^T(\bm\mathcal E_L\bm e_1\bm d_n+(\bm\mathcal E_L\bm e_1\bm d_n)^T\otimes \bm H_y^L\bm{I}_{R2L})\bm v\right]\\
& +\delta_R\left[\bm v^T(\bm\mathcal E_R\bm e_1\bm d_1+(\bm\mathcal E_R\bm e_1\bm d_1)^T\otimes \bm H_y^R)\bm v-\bm v^T(\bm\mathcal E_L\bm e_1\bm d_n+(\bm\mathcal E_L\bm e_1\bm d_n)^T\otimes \bm H_y^R\bm{I}_{L2R})\bm u\right]\\
& +\sigma_R\left[\bm v^T(\bm\mathcal E_R\bm e_1\bm d_1+(\bm\mathcal E_R\bm e_1\bm d_1)^T\otimes \bm H_y^R)\bm v-\bm v^T(\bm\mathcal E_R\bm e_n\bm d_1+(\bm\mathcal E_R\bm e_n\bm d_1)^T\otimes \bm H_y^R\bm{I}_{L2R})\bm u\right].
\end{align*}
One can see that the mixed terms, e.g., $\delta_L\bm u^T(\bm\mathcal E_R\bm e_n\bm d_1+(\bm\mathcal E_R\bm e_n\bm d_1)^T\otimes \bm H_y^L\bm{I}_{R2L})$ and $\sigma_R \bm v^T(\bm\mathcal E_R\bm e_n\bm d_1+(\bm\mathcal E_R\bm e_n\bm d_1)^T\otimes \bm H_y^R\bm{I}_{L2R})\bm u$, are canceled as long as we choose $\delta_L=-\sigma_R, \sigma_L =-\delta_R$. To cancel $IT^{\epsilon}$, we also need $\delta_L+\sigma_L = -1$. Therefore, it is sufficient to conclude that with the above choice of parameters, all the interface terms vanish $IT \equiv 0$. The interface treatment is thus stable and conservative.}
\end{proof}
\subsection{The FD--FE coupling}
\label{sec:coupling_fd_fe}
This section contains the main contribution of this paper. \cpurple{The following definitions align FE schemes with the coupling to FD schemes.}
\begin{definition}
\label{def:fem_boundary_selection_operators}
\cpurple{Let $m$ be the number of degrees of freedom on the concerned part of the boundary, and $\bm L$ be a binary matrix of size $m\times \cblue{\mathrm N}$. The matrix $\bm L$ is called a \textit{FE boundary selection operator} if the following properties are satisfied:}
\begin{itemize}
\item[(i)] Each row of \cblue{$\bm L$} contains exactly one ``1'' element uniquely corresponding to a node on the concerned part of the boundary,
\item[(ii)] \cblue{$\bm L$} rearranges the boundary nodes to the same spatial order of the coupled solution.
\end{itemize}
\end{definition}
The following properties of the above boundary selection operators are useful for the stability proofs in this section.
\cblue{
\begin{property}
\label{property:L_W}
Let $\bm L$ be a FE boundary selection operator. The following statements hold
\begin{itemize}
\item[(i)] $\bm E\bm L = \bm F \implies \bm E =\bm F \bm L^T$,
\item[(ii)] $\bm F \bm L^T = \bm E \implies \bm F = \bm E\bm L$\; if\; $(\bm I_{\cblue{\mathrm N}}-\bm L^T\bm L) \bm F^T = 0$,
\end{itemize}
\end{property}
\noindent where $\bm E, \bm F$ are matrices of appropriate sizes.
\begin{proof}
We prove (i). From $\bm E\bm L=\bm F$, right multiplying both sides by $\bm L^T$ gives $\bm E\bm L\bm L^T=\bm F\bm L^T$. Since each row of $\bm L$ contains exactly one ``1'' element corresponding to the same position on each column of $\bm L^T$, it is clear that $\bm L\bm L^T=\bm I_m$. We have the first statement proven.\\
Similarly, we prove (ii) by right multiplying both sides by $L$ to obtain $\bm F\bm L^T\bm L=\bm E\bm L$. However, $\bm L^T\bm L\neq \bm I_{\cblue{\mathrm N}}$. The additional constraint $(\bm I_{\cblue{\mathrm N}}-\bm L^T\bm L) F^T = 0$ is necessary for getting $\bm F \bm L^T = \bm E$. Assuming that $\bm F$ is a square matrix \cblue{of size $\mathrm N \times \mathrm N$}, this additional constraint can be understood as, all the columns of $\bm F$ of which indices corresponding the nonboundary nodes are zero.
\end{proof}}
\cblue{We consider the setup in Figure \ref{fig:coupling-interface} where FD is used on $\Omega_L$ and FE is used on $\Omega_R$. The interface $\Omega_I$ is thus both the West boundary of $\Omega_R$ and the East boundary of $\Omega_L$. Therefore, the FE operators activating on the West boundary, e.g., $\bm L_W$, and the FD operators activating on the East boundary, e.g., $(\bm e_n^T\otimes\bm I_m)$, are frequently used in this section. The notations $\bm H_x,\bm H_y$ are used instead of $\bm H_x^L,\bm H_y^L$ since there is only one domain involving FD}. \cred{Definition \ref{def:conservation_property_fd_fd} is restated below in the FD--FE coupling context.}
\begin{definition}
\label{def:conservation_property_fd_fe}
$\{\bm{I}_{L2R}, \bm{I}_{R2L}\}$ is called a pair of \textit{SBP-preserving FD--FE interpolation operators} if
\begin{equation}
\label{eq:conservation_property_fd_fe}
\mathbf{M}_I\bm{I}_{L2R} = \bm{I}_{R2L}^T\bm H_y,
\end{equation}
where $\displaystyle \mathbf{M}_I=\bm L_W\mathbf{R}_M^W\bm L_W^T$. Recall that $\cblue{\mathrm N}\times \cblue{\mathrm N}$ matrix $(\mathbf{R}_{M}^W)_{ij}=\int_{\Gamma_W}\varphi_j\varphi_ids,i,j=1,2,\dots,\cblue{\mathrm N}$.
\end{definition}
\cred{Notice that $\mathbf{M}_I$ is analogous to the norm $\bm H_y^R$ in the FD--FD coupling}. Despite it not being diagonal, $\mathbf{M}_I$ has nice properties including being symmetric and positive definite. For example, if the node distribution on $\Gamma_W$ is equidistant with step length $\bm H_y$, and $\mathbb{P}^1$ is used, then $\mathbf{M}_I$ is a banded matrix,
\[
\mathbf{M}_I = h_y\begin{bmatrix}
1/3 & 1/6 & 0 & 0 & \hdots & 0 \\
1/6 & 2/3 & 1/6 & 0 & \hdots & 0 \\
0 & 1/6 & 2/3 & 1/6 & \hdots & 0 \\
\vdots & 0 & \ddots & \ddots & \ddots & \vdots\\
0 & \hdots & 0 & 1/6 & 2/3 & 1/6\\
0 & \hdots & 0 & 0 & 1/6 & 1/3
\end{bmatrix}.
\]
\cgreen{Let $\bm\omega$ be a discrete vector containing nodal values of a FE solution in $V_h$. We denote $\bm \omega_I = \bm L_W \bm \omega$ the vector contains values of $\bm\omega$ at the nodes on $\Omega_I$}. We prove that the following FD--FE coupling SATs impose stability to \eqref{eq:advection_diffusion_sbp_sat},
\cblue{
\begin{equation}
\label{eq:advection_diffusion_fe_interface}
\begin{aligned}
\mbox{SAT}_{IL} & = \mbox{SAT}^{f}_{IL} + \mbox{SAT}^{f'}_{IL} + \mbox{SAT}^{\epsilon}_{IL},\\
\mbox{SAT}_{IR} & = \mbox{SAT}^{f}_{IR} + \mbox{SAT}^{f'}_{IR} + \mbox{SAT}^{\epsilon}_{IR}.
\end{aligned}
\end{equation}
The terms $\mbox{SAT}^{f}_{IL}, \mbox{SAT}^{f'}_{IL}, \mbox{SAT}^{f}_{IR}, \mbox{SAT}^{f'}_{IR}$ couple the flux
\[
\begin{aligned}
\mbox{SAT}^{f}_{IL} & = \alpha_L ({\bm H_x^L}^{-1} \bm e_n)\otimes \left[(\alpha_1\bm f_1(\bm u))_I-\bm{I}_{R2L}(\alpha_1\bm f_1(\bm v))_I\right],\\
\mbox{SAT}^{f'}_{IL} & + \beta_L ({\bm H_x^L}^{-1} \bm e_n)\otimes \left[(\tilde\alpha_1\bm F_1'(\bm u))_I(\bm u_I-\bm{I}_{R2L}\bm v_I)\right],
\end{aligned}
\]
\vspace{-0.3cm}
\begin{equation}\label{eq:advection_diffusion_fe_interface_general}
\begin{aligned}
\mbox{SAT}^{f}_{IR} & =\alpha_R \mathbf{M}^{-1}\mathbf{R}_M^W\bm L_W^T\left[(\alpha_1\bm f_1(\bm v))_I-\bm{I}_{L2R}(\alpha_1\bm f_1(\bm u))_I\right],\\
\mbox{SAT}^{f'}_{IR}&+ \beta_R\mathbf{M}^{-1}\Big[\mathscr{R}_M^W\left( \tilde\alpha_1\bm f_1'(\bm v)\right)\bm v-\bm V\mathbf{R}_M^W\bm L_W^T(\tilde\alpha_1\bm F_1'(\bm v))_I\bm{I}_{L2R}\bm u_I\Big],
\end{aligned}
\end{equation}
and the terms $\mbox{SAT}^{\epsilon}_{IL}, \mbox{SAT}^{\epsilon}_{IR}$ couple the diffusion term}
\[\cblue{
\begin{aligned}
\mbox{SAT}^{\epsilon}_{IL} & = \delta_L(H_x^{-1} \bm e_n) \otimes (\bm\mathcal E_L(\mathbf{D}_x\bm u)_I - \bm{I}_{R2L}\mathbf{D}_x^{W}(\epsilon)\bm v)+\sigma_L(H_x^{-1}\bm d_n^T)\otimes\bm\mathcal E_L(\bm u_I-\bm{I}_{R2L}\bm L_W\bm v),
\end{aligned}}
\]
\[\cblue{
\begin{aligned}
\mbox{SAT}^{\epsilon}_{IR} & = \delta_R\mathbf{M}^{-1}\mathbf{R}_M^W\bm L_W^T(\mathbf{D}_x^W(\epsilon)\bm v-\bm{I}_{L2R}\bm\mathcal E_L(\mathbf{D}_x\bm u)_I) +\sigma_R\mathbf{M}^{-1}(-\mathbf{R}^W_C(\epsilon))^T\bm L_W^T(\bm v_I - \bm{I}_{L2R}\bm u_I),
\end{aligned}}
\]
where $\mathbf{R}^W_{C\,ij}(\epsilon) = \int_{\Gamma_W} \mathbf{n}\cdot\epsilon\nabla\varphi_j \varphi_i ds, \mathbf{D}_x^W(\epsilon) = -\mathbf{M}_I^{-1}\bm L_W\mathbf{R}^W_C(\epsilon)$. \cgreen{Other notations are defined in Section \ref{sec:sbp_form_fem}}. The assembly of the above matrices is described in \citet{Larson2009,Dao2019}. Note that the minus signs coming with $\mathbf{R}_C^W(\epsilon)$ in $\mathbf{D}_x^W(\epsilon)$ and the third component of $\mbox{SAT}_{IR}$ are in order to make $\mathbf{R}_C^W(\epsilon)$ analogous to the FD operator $(\bm d_1\otimes\bm I_m)$ since $\mathbf{n}_W$ is a negative vector.
\begin{lemma}
\label{lemma:R_M_W_nonlinear}
The following equality is fulfilled if the quadrature rule \eqref{eq:quadrature_rule} is used to calculate $\mathscr{R}_M(\phi)$,
\[
\mathscr{R}_M(\phi) = \frac{1}{2}\left(\mathbf{R}_M \mathbf \Phi +\mathbf \Phi\mathbf{R}_M\right),
\]
where $\mathbf \Phi$ is the diagonal matrix of which diagonal is $\left(\phi(x_2), \phi(x_2),\dots, \phi(x_{\mathrm N})\right)^T$.
\end{lemma}
The matrices $\mathscr{R}_M(\phi), \mathbf{R}_M$ are defined in Section \ref{sec:sbp_form_fem}. The result of Lemma \ref{lemma:R_M_W_nonlinear} is direct by writing \eqref{eq:quadrature_rule} in the matrix form. Restricting the integrals on $\Gamma_W$, we have
\[
\mathscr{R}_M^W(\phi) = \frac{1}{2}\left(\mathbf{R}_M^W \mathbf \Phi +\mathbf \Phi\mathbf{R}_M^W\right).
\]
Using the quadrature rule \eqref{eq:quadrature_rule}, the terms $\mbox{SAT}^{f}_{IR}, \mbox{SAT}^{f'}_{IR}$ \eqref{eq:advection_diffusion_fe_interface_general} can then be simplified as
\begin{equation}\label{eq:advection_diffusion_fe_interface_simplified}
\begin{aligned}
\mbox{SAT}^{f}_{IR} & = \alpha_R\mathbf{M}^{-1}\mathbf{R}_M^W\bm L_W^T\left[(\alpha_1\bm f_1(\bm v))_I-\bm{I}_{L2R}(\alpha_1\bm f_1(\bm u))_I\right],\\
\mbox{SAT}^{f'}_{IR} &= \beta_R\mathbf{M}^{-1}(\tilde\alpha_1\bm F_1'(\bm v))\mathbf{R}_M^W\bm L_W^T\left[(\bm v_I-\bm{I}_{L2R}\bm u_I)\right],
\end{aligned}
\end{equation}
which is in the same form as the diagonal norm coupling \eqref{eq:advection_diffusion_sat_interface}. We show the relevance of the quadrature rule \eqref{eq:quadrature_rule} with the SAT \eqref{eq:advection_diffusion_fe_interface_simplified} in the Theorem \ref{theorem:advection_diffusion_fe_interface}.
Strict stability for the FD--FE coupling is shown in the theorem below.
\begin{theorem}
\label{theorem:advection_diffusion_fe_interface}
By choosing $\cblue{\alpha_L=\beta_L=1/2, \alpha_R=\beta_R=-1/2}, \delta_L=-\sigma_R, \sigma_L =-\delta_R$ and $\delta_L+\sigma_L = -1$, assuming proper outer boundary treatment, the approximation \eqref{eq:advection_diffusion_sbp_sat} with the interface SATs \eqref{eq:advection_diffusion_fe_interface} is stable.
\end{theorem}
\begin{proof}
A total energy estimate can be obtained by applying the energy method to both schemes
\begin{align*}
\frac{d}{dt} \norm{\bm u}_{\H}^2 + 2\epsilon \bm u^T(A\otimes H_y)\bm u + 2\epsilon \bm u^T(H_x\otimes A)\bm u+\frac{d}{dt}\norm{\bm v}^2_{\mathbf{M}}+2\epsilon \bm v^T\mathbf{A}\bm v & = BT + IT.
\end{align*}
\cblue{
The structure of this proof is similar to the proof of Theorem \ref{theorem:choice_linear_fd_fd}. We only focus on the key differences coming from FE. The interface energy estimate reads
\[
IT = IT^{f} + IT^{f'} + IT^{\epsilon} + IT^f_{SAT} + IT^{f'}_{SAT} + IT^{\epsilon}_{SAT},
\]
where $IT^{f}, IT^{f'}$ come from the operators approximating the flux; $IT^{\epsilon}$ comes from the operators approximating the diffusion term; $IT^f_{SAT}, IT^{f'}_{SAT}, IT^{\epsilon}_{SAT}$ come from the corresponding SATs. The following estimate is derived from the SBP property of the discrete operators
\begin{align*}
IT^{f} + IT^{f'} = & \; -\bm u^T(\bm e_n\bm e_n^T\otimes \bm H_y^L)(\alpha_1\bm f_1(\bm u)+\bm U\tilde\alpha_1\bm f_1'(\bm u))+\bm v^T\mathbf{R}_M^W\alpha_1 \bm f_1(\bm v)+\bm v^T\mathscr{R}_M^W(\tilde\alpha_1\bm f_1'(\bm v))\bm v.
\end{align*}
The involvement of $\mbox{SAT}^{f}_{IL}, \mbox{SAT}^{f'}_{IL}, \mbox{SAT}^{f}_{IR}, \mbox{SAT}^{f'}_{IR}$ \eqref{eq:advection_diffusion_fe_interface_simplified} mutually stabilizes the discrete flux
\begin{align*}
IT^{f}_{SAT} = &\;2\alpha_L\bm u^T(\bm e_n\bm e_n^T\otimes \bm H_y^L)\alpha_1\bm f_1(\bm u)-2\alpha_L\bm u^T(e_n\otimes H_y\bm{I}_{R2L}\bm L_W)\alpha_1\bm f_1(\bm v)\\
&+2\alpha_R\frac{(\alpha_1 \bm f_1(\bm v))^T\mathbf{R}_M^W\bm v+\bm v^T\mathbf{R}_M^W\alpha_1 \bm f_1(\bm v)}{2}+2\alpha_R\bm v^T(\bm e_n^T\otimes \mathbf{R}_M^W\bm L_W^T\bm{I}_{L2R})\alpha_1 \bm f_1(\bm u),\\
IT^{f'}_{SAT} = &\;2\beta_L\bm u^T(\bm e_n\bm e_n^T\otimes H_y)\tilde\alpha_1\bm f_1'(\bm u)-2\beta_L\bm v^T(e_n^T\otimes \bm L_W^T\bm{I}_{R2L}^TH_y){\bm U}\tilde\alpha_1\bm f_1'(\bm u)\\
&+2\beta_R\frac{(\bm V\tilde\alpha_1\bm f_1'(\bm v))^T\mathbf{R}_M^W\bm v+\bm v^T\mathbf{R}_M^W\bm V\tilde\alpha_1\bm f_1'(\bm v)}{2}-2\beta_R\bm u^T(\bm e_n\otimes \bm{I}_{L2R}^T\bm L_W\mathbf{R}_M^W)\bar v\tilde\alpha_1\bm f_1'(\bm v).
\end{align*}
Property \ref{property:L_W} is used to handle the transpose terms. Under Lemma \ref{lemma:R_M_W_nonlinear}, we have
\[
\bm v^T\mathscr{R}_M^W(\tilde\alpha_1\bm f_1'(\bm v)) \bm v=\bm v^T\frac{1}{2}\left(\mathbf{R}_M^W\tilde\alpha_1\bm F_1'(\bm v) + \tilde\alpha_1\bm F_1'(\bm v)\mathbf{R}_M^W \right)\bm v = \frac{\bm v^T\mathbf{R}_M^W\bm V\tilde\alpha_1\bm f_1'(\bm v) + (\bm V\tilde\alpha_1\bm f_1'(\bm v))^T\mathbf{R}_M^W\bm v}{2}.
\]
Thus, the choice of $\alpha_L=\beta_L=1/2, \alpha_R=\beta_R=-1/2$ makes $-(IT^{f}+IT^{f'})$ appear in the sum $(IT^{f}_{SAT} + IT^{f'}_{SAT})$. Notice that Lemma \ref{lemma:R_M_W_nonlinear} is only necessary to show that the energy term $\bm v^T\mathscr{R}_M^W(\tilde\alpha_1\bm f_1'(\bm v)) \bm v$ coming from the discrete operator is canceled. Because $\mathscr{R}_M^W$ is symmetric, the energy terms are canceled with the same choice of $\beta_L, \beta_R$ as long as the quadrature rule to compute $\mathscr{R}_M^W$ in the scheme and in the coupling SATs are the same. Similar to the proof of Theorem \ref{theorem:choice_linear_fd_fd}, the mixed terms containing the interpolation operators $\bm{I}_{L2R}, \bm{I}_{R2L}$ are also canceled because of the interpolation property \eqref{eq:conservation_property_fd_fe} and the discrete skew-symmetry \eqref{eq:skew_symmetric_sufficient_discrete}. The coupling of the diffusion terms can be proved to be stable in a similar manner.
\begin{align*}
IT^{\epsilon} = &\;\bm u^T(\bm\mathcal E_L\bm e_n\bm d_n+(\bm\mathcal E_L\bm e_n\bm d_n)^T\otimes H_y)\bm u + \bm v^T (\mathbf{R}^W_C + (\mathbf{R}^W_C)^T)\bm v,\\
IT^{\epsilon}_{SAT} = &\; \delta_L \left[ \bm u^T(\bm\mathcal E_L\bm e_n\bm d_n+(\bm\mathcal E_L\bm e_n \bm d_n)^T \otimes H_y)\bm u -\bm u^T(\bm e_n \otimes H_y\bm{I}_{R2L}\mathbf{D}_x^W)\bm v-\bm v^T(\bm e_n^T\otimes (H_y\bm{I}_{R2L}\mathbf{D}_x^W)^T)u\right]\\
& + \sigma_L\left[\bm u^T(\bm\mathcal E_L\bm e_n\bm d_n+(\bm\mathcal E_L\bm e_n\bm d_n)^T\otimes H_y)\bm u-\bm u^T(\bm d_n^T\otimes H_y\bm{I}_{R2L}\bm\mathcal E_L\bm L_W)\bm v-\bm v^T(\bm d_n\otimes \bm L_W^T\bm\mathcal E_L\bm{I}_{R2L}^TH_y)\bm u\right]\\
& + \delta_R\left[-\bm v^T(\mathbf{R}_C^W+(\mathbf{R}_C^W)^T)\bm v-\bm v^T\mathbf{R}_M^W\bm L_W^T\bm{I}_{L2R}\bm\mathcal E_L(d_n\otimes\bm I_m)\bm u-\bm u^T(\bm d_n^T\otimes\bm I_m)\bm\mathcal E_L\bm{I}_{L2R}^T\bm L_W\mathbf{R}_M^W\bm v)\right]\\
& - \sigma_R\bm v^T((\mathbf{R}^W_C)^T\bm L_W^T\bm L_W + \bm L_W^T\bm L_W\mathbf{R}^W_C)\bm v+\sigma_R \bm v^T(\mathbf{R}_C^W)^T\bm L_W^T\bm{I}_{L2R}(\bm e_n^T\otimes\bm I_m)\bm u\\
& + \sigma_R \bm u^T(\bm e_n \otimes\bm I_m) \bm{I}_{L2R}^T\bm L_W\mathbf{R}_C^W\bm v.
\end{align*}
Due to the fact that $\mathbf{R}_C^W$ contains only zeros on the rows which do not correspond to the West boundary, we have $\bm L_W^T\bm L_W\mathbf{R}^W_C=\mathbf{R}^W_C$. As long as $\delta_L=-\sigma_R, \sigma_L =-\delta_R$ and $\delta_L+\sigma_L = -1$ in $IT^{\epsilon}_{SAT}$, $IT^{\epsilon}$ is canceled by $IT^{\epsilon}_{SAT}$. To show that the remaining terms in $IT^{\epsilon}_{SAT}$ containing $\bm{I}_{L2R}, \bm{I}_{R2L}$ also vanish, we need to show that
\begin{equation}\label{eq:advection_diffusion_fe_interface_proof_2}
H_y\bm{I}_{R2L}\mathbf{D}_x^W=-\bm{I}_{L2R}^T\bm L_W\mathbf{R}_C^W.
\end{equation}
Indeed, from the property \eqref{eq:conservation_property_fd_fe} which is $H_y\bm{I}_{R2L} = \bm{I}_{L2R}^T \mathbf{M}_I$, we have $H_y\bm{I}_{R2L} (-\mathbf{M}_I^{-1}\bm L_W \mathbf{R}_C^W) = -\bm{I}_{L2R}^T\bm L_W \mathbf{R}_C^W$, where $-\mathbf{M}_I^{-1}\bm L_W \mathbf{R}_C^W = \mathbf{D}_x^W$.
}
It it sufficient to conclude that the interface terms vanish $IT\equiv 0$ with the above choice of parameters. \cblue{This result comes from the construction of the SATs}. The interface treatment is thus stable and conservative.
\end{proof}
\begin{remark}
Similar to the FD--FD coupling, the use of SBP-preserving interpolation operators used for FD--FE can also be shown to reserve the full order of accuracy. It is known that when narrow stencil FD operators of $q$th-order are used, the accuracy at boundary closures is of $q/2$th-order resulting in $(q/2+1)$th-order of overall accuracy. However, since the solution space $\mathbb{P}^1$ is currently used for FE, we expect that with roughly the same number of degrees of freedom of each method, the overall accuracy is bounded by second-order \cblue{even} if higher-order FD operators are used.
\end{remark}
\section{Numerical examples}
This section includes numerical experiments of the proposed coupling technique on the linear advection-diffusion equation, and a nonlinear viscous Burgers' equation.
\label{sec:numerical_results}
\subsection{Linear advection-diffusion equation}
The following linear equation is obtained by letting $\bm f = \bm a = (a_1,a_2)^T \in \mathbb{R}^2$ being a constant vector,
\begin{equation}\label{eq:linear_fem}
\begin{aligned}
u_t + \bm a \cdot \nabla u & = \epsilon\nabla\cdot(\nabla u), && (x,y) \in \Omega,\\
u & = u_0(x,y), && (x,y) \in \Omega, t = 0.
\end{aligned}
\end{equation}
Inserting $\bm f = \bm a$ into \eqref{eq:skew_symmetric_precise} gives $\alpha_1=\alpha_2=\frac{1}{2}$. From \eqref{eq:advection_diffusion_fem_bc}, a set of well-posed boundary condition reads
\begin{equation}
\label{eq:linear_fem_bc}
\frac{1}{2} (\bm a\cdot\mathbf{n}-\abs{\bm a\cdot\mathbf{n}})u-\epsilon \mathbf{n}\cdot \nabla u = g(x,y,t), \quad(x,y) \in \partial \Omega, t > 0.
\end{equation}
We consider the diffusion coefficient $\epsilon$ as a constant.
\subsubsection*{The FD--FD coupling}
A FD--FD semi-discretization of the coupling problem \eqref{eq:advection_diffusion_coupling_continuous} under the governing equation \eqref{eq:linear_fem} is given by
\begin{equation}
\label{eq:linear_sbp_sat}
\begin{aligned}
\bm u_t &= -a_1\mathbf{D}_x\bm u - a_2\mathbf{D}_y\bm u +\epsilon \mathbf{D}_{2x}\bm u + \epsilon \mathbf{D}_{2y}\bm u + \mbox{SAT}_L+\mbox{SAT}_{IL},\\
\bm v_t &= -a_1\mathbf{D}_x\bm v - a_2\mathbf{D}_y\bm v + \epsilon \mathbf{D}_{2x}\bm v + \epsilon \mathbf{D}_{2y}\bm v + \mbox{SAT}_R+\mbox{SAT}_{IR},
\end{aligned}
\end{equation}
where interface SATs are
\begin{equation}
\label{eq:linear_sat_interface}
\begin{aligned}
\mbox{SAT}_{IL} & = \frac{1}{2}a_1 ({\bm H_x^L}^{-1} \bm e_n)\otimes (\bm u_I-\bm I_{R2L}\bm v_I)\\
&+ \delta_L\epsilon ({\bm H_x^L}^{-1} \bm e_n) \otimes ((\mathbf{D}_x\bm u)_I - \bm I_{R2L}(\mathbf{D}_x\bm v)_I) + \sigma_L\epsilon ({\bm H_x^L}^{-1}\bm d_n^T)\otimes(\bm u_I-\bm I_{R2L}\bm v_I),\\
\mbox{SAT}_{IR} & = -\frac{1}{2}a_1({\bm H_x^R}^{-1}\bm e_1) \otimes (\bm v_I-\bm I_{L2R}\bm u_I)\\
&+\delta_R \epsilon({\bm H_x^R}^{-1}\bm e_1) \otimes ((\mathbf{D}_x\bm v)_I-\bm I_{L2R}(\mathbf{D}_x\bm u)_I)+ \sigma_R \epsilon({\bm H_x^R}^{-1} \bm d_1^T) \otimes (\bm v_I-\bm I_{L2R}\bm u_I),
\end{aligned}
\end{equation}
where $\delta_L=-\sigma_R, \sigma_L =-\delta_R, \delta_L+\sigma_L = -1$.
\subsubsection*{The FD--FE coupling}
A FD--FE semi-discretization of the coupling problem \eqref{eq:advection_diffusion_coupling_continuous} under the governing equation \eqref{eq:linear_fem} is given by
\begin{equation}
\label{eq:linear_fe_sbp_sat}
\begin{aligned}
\bm u_t &= -a_1\mathbf{D}_x\bm u - a_2\mathbf{D}_y\bm u +\epsilon \mathbf{D}_{2x}\bm u + \epsilon \mathbf{D}_{2y}\bm u + \mbox{SAT}_L+\mbox{SAT}_{IL},\\
\bm v_t &= -\mathbf{M}^{-1}\mathbf{C}\bm v-\epsilon \mathbf{M}^{-1}\mathbf{A}\bm v+ \frac{1}{2} (a_2-\abs{a_2})\mathbf R_M^N\bm v+ \frac{1}{2} (a_1-\abs{a_1})\mathbf R_M^E\bm v+\frac{1}{2} (-a_2-\abs{a_2})\mathbf R_M^S\bm v - \bm r +\mbox{SAT}_{IR},
\end{aligned}
\end{equation}
where $\mathbf{R}_M^N,\mathbf{R}_M^E,\mathbf{R}_M^S$ are obtained by limiting the integration of $\mathbf{R}_M$ to $\Gamma_N,\Gamma_E,\Gamma_S$, respectively. The coupling penalty terms are given by
\begin{equation}
\label{eq:advection_diffusion_sat_interface2}
\begin{aligned}
\mbox{SAT}_{IL} & = \frac{1}{2}a_1 (\bm H_x^{-1} \bm e_n)\otimes (\bm u_I-\bm{I}_{R2L}\bm L_W\bm v)\\
&+ \delta_L\epsilon(\bm H_x^{-1} \bm e_n) \otimes ((\mathbf{D}_x\bm u)_I - \bm{I}_{R2L}\mathbf{D}_x^W\bm v)+\sigma_L\epsilon(\bm H_x^{-1}\bm d_n^T)\otimes(\bm u_I-\bm{I}_{R2L}\bm L_W\bm v),\\
\mbox{SAT}_{IR} & = -\frac{1}{2}a_1 \mathbf{M}^{-1}\mathbf{R}_M^W\bm L_W^T(\bm L_W\bm v - \bm{I}_{L2R}\bm u_I)\\
&+\delta_R\epsilon\mathbf{M}^{-1}\mathbf{R}_M^W\bm L_W^T(\mathbf{D}_x^W\bm v-\bm{I}_{L2R}(\mathbf{D}_x\bm u)_I) +\sigma_R\epsilon\mathbf{M}^{-1}(-\mathbf{R}^W_C)^T\bm L_W^T(\bm L_W\bm v - \bm{I}_{L2R}\bm u_I),
\end{aligned}
\end{equation}
where $\delta_L=-\sigma_R, \sigma_L =-\delta_R, \delta_L+\sigma_L = -1$.
\subsubsection{Eigenvalue analysis for validation of the interface treatment}
In this subsection, we show a numerical validation of Theorem \ref{theorem:advection_diffusion_fe_interface}. Consider the model problem \eqref{eq:advection_diffusion_coupling_continuous} subject to the boundary condition \eqref{eq:linear_fem_bc}. Second-order FD operators are used on $\Omega_L = [-1,1]\times[-2,0]$, and a FE scheme using piecewise linear basis functions ($\mathbb{P}^1$) is used on $\Omega_R = [-1,1]\times[0,2]$. The FD domain is discretized with $41^2$ grid points ($m=n=41$). The discretization of FE scheme uses 2796 vertices for an unstructured mesh using Delaunay triangulation with maximum cell diameter $h_{\max}=0.025$ and interface nodes matched to the FD grid, intended for an equally distributed image of eigenvalues. The solution is transported from the left to the right by setting $\bm a = (1,0)^T$.
Figure \ref{fig:fd-fe-2d} shows another example using $41^2$ grid points on the FD domain, and an unstructured mesh using Delaunay triangulation with maximum element diameter $h_{max}=0.2$ on the FE domain (nonmatching coupling with 41 FD grid points, and 51 FE nodal points on the interface). \cblue{The reference solution,
\begin{equation}\label{eq:analytic_inviscid_solution}
u(x,y) = r_1\exp\left(\frac{(x-x_0-a_1t)^2+(y-y_0-a_2t)^2}{r_2^2}\right),
\end{equation}
is initialized at $(x_0,y_0)^T = (-1,0)^T$ and transported from left to right by setting $\bm a = (1,0)^T$}. The diffusion parameter is set to zero $\epsilon = 0$. \cblue{The other parameters are $r_1 = 1, r_2 = 0.2$.} Sixth-order FD operators and $\mathbb{P}^1$ finite elements are used. The solution is examined at $t = 0$ and $t=1$.
\begin{figure}[pos=h]
\begin{subfigure}{\textwidth}
\centering
\includegraphics[scale=0.2]{linear_t0.png}
\caption{Initial data $t = 0$.}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[scale=0.2]{linear_t1.png}
\caption{Computed solution at $t = 1$.}
\end{subfigure}
\caption{An example of FD--FE nonconforming multiblock coupling for the inviscid problem $\epsilon=0$ using a sixth-order FD scheme on the left domain (discretized with $41^2$ grid points) and a FE scheme on the right domain (using 8552 $\mathbb{P}^1$ elements). Notice the unmatching placement of nodal points on the interface.}
\label{fig:fd-fe-2d}
\end{figure}
A semi-discrete problem to \eqref{eq:advection_diffusion_fem} can be written in the form
\cblue{
\begin{equation}
\label{eq:linear-semi-discrete-form}
\begin{bmatrix}
\bm u\\\bm v
\end{bmatrix}_t
= \mathbf{F} \begin{bmatrix}
\bm u\\\bm v
\end{bmatrix},
\end{equation}}
where $\mathbf{F}$ is a square real-valued matrix. In Figure \ref{fig:spectrum_linear_FD-FE}, the eigenvalues of $\mathbf F$ are plotted for several choices of the penalty parameters imposing the interface conditions in Theorem \ref{theorem:advection_diffusion_fe_interface}. \cblue{The horizontal axis indicates the real part of the eigenvalues. The imaginary axis is vertical.} Figure \ref{fig:spectrum_linear_FD-FE}(a) and \ref{fig:spectrum_linear_FD-FE}(c) are with two different sets of penalty parameters taking into account a diffusion term $\epsilon=0.01$ both satisfying Theorem \ref{theorem:advection_diffusion_fe_interface} while the parameters used in Figure \ref{fig:spectrum_linear_FD-FE}(b) and \ref{fig:spectrum_linear_FD-FE}(d) violate the theorem. It can be clearly seen that in the former cases, all the yielding eigenvalues have nonpositive real part. Therefore, with some explicit method to integrate in time, for instance Runge-Kutta 4, and a sufficiently small time step, the fully discrete scheme is stable.
\begin{figure}[pos=h]
\centering
\begin{subfigure}{0.49\textwidth}
\includegraphics[scale=0.3]{spectrum_linear_FD-FE-right.eps}
\caption{With a set of correct penalty parameters $\alpha_L = 0.5,\;\;\;\;\;$ $\alpha_R=-0.5, \delta_L=-\epsilon, \sigma_R=-\epsilon, \sigma_L=\delta_R=0$.}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics[scale=0.3]{spectrum_linear_FD-FE-zero.eps}
\caption{Without coupling terms $\alpha_{L,R}=\delta_{L,R}=\sigma_{L,R}=0$. There exist positive eigenvalues.}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics[scale=0.3]{spectrum_linear_FD-FE-right2.eps}
\caption{With a set of correct penalty parameters $\alpha_L = 0.5,\;\;\;\;\;$, $\alpha_R=-0.5, \delta_L=\sigma_L = -\epsilon/2, \delta_R=\sigma_R=\epsilon/2$.}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics[scale=0.3]{spectrum_linear_FD-FE-wrong.eps}
\caption{With violating parameters $\alpha_L = -0.05, \delta_L=0.001, \alpha_R=-0.05, \sigma_R=0.001, \sigma_L=\delta_R = 0$.}
\end{subfigure}
\caption{Eigenvalues \cblue{in the complex plane} of the system \eqref{eq:advection_diffusion_sbp_sat} coupling second-order FD -- $\mathbb{P}^1$ FE with parameters satisfying Theorem \ref{theorem:advection_diffusion_fe_interface} (on the left column) and violating the theorem (on the right column). All are with proper outer boundary treatment.}
\label{fig:spectrum_linear_FD-FE}
\end{figure}
\subsubsection{Convergence study}
\label{sec:linear_convergence_study}
We use the following two dimensional Gaussian as reference solution
\[
u(x,y) = \frac{r}{4\pi(t-t_0)\epsilon} \exp\left[\frac{-1}{t-t_0}\left(\frac{(x-x_0-a_1t)^2+(y-y_0-a_2t)^2}{4\epsilon}\right)\right].
\]
The solution satisfies the boundary condition \eqref{eq:advection_diffusion_fem_bc} with zero boundary data $g=0$ if the initial solution is small enough at the inflow boundary. For convergence results in this section, we choose $r=0.005$. For the inviscid problem, we set $\epsilon = 0$. The 2D Gaussian \eqref{eq:analytic_inviscid_solution} is used as the reference solution. We set $r_1=1,r_2=0.2$. The $l_2$-error in each FD solution part $\bm u_{FD}$ and in each FE solution part $\bm u_{FD}$ are respectively calculated by
\begin{equation}
\label{eq:discrete_error_measurement}
l_2e(\bm u_{FD}) = \sqrt{(\bm u-\bm u_{e})^T\H(\bm u-\bm u_{e})},\quad l_2e(\bm u_{FE})=\sqrt{(\bm u-\bm u_{e})^T\mathbf{M}(\bm u-\bm u_{e})},
\end{equation}
where $\bm u_e$ is the exact solutions at the time of evaluation. The convergence rate in each method block is calculated as
\begin{equation}
\label{eq:convergence_rate}
Q = \log\left(\frac{l_2e(\bm u)^{(h_1)}}{l_2e(\bm u)^{(h_2)}}\right)/\log\left(\frac{h_1}{h_2}\right),
\end{equation}
for $h_1\neq h_2$ being two different mesh/grid sizes. Since multiple blocks with different methods are involved, the global convergence rate is calculated based on the sum of all $l_2$-errors, and the scale of mesh/grid sizes, e.g., $h_2/h_1=2$ when refining all the meshes/grids of size $h_1$ twice. We use the FD diagonal norm SBP operators formulated in \citet{Mattsson2004}. To demonstrate a proposed technique of constructing nondiagonal norm interpolation operators, the blocks are coupled by the `glue grid' FD-FD interpolation operators \citet{Kozdon2016}, and our proposed procedure Appendix \ref{section:shortcut_diagonal_norm}.
We choose the multiblock setup in Figure \ref{fig:2by2_domain}(a) to investigate convergence with an intention of examining possible degradation at corner points where the normal vector $\mathbf{n}$ is not defined. Tables \ref{table:2by2_multiblock_test1}, \ref{table:2by2_multiblock_test2}, \ref{table:2by2_multiblock_test3}, \ref{table:2by2_multiblock_test4} show the $l_2$-error and convergence rates for the problem \eqref{eq:linear_fem} subject to the boundary condition \eqref{eq:advection_diffusion_fem_bc} with different numbers of grid points. \eqref{eq:linear-semi-discrete-form} is integrated in time using Runge-Kutta 4 with time step $\Delta t = \min_{\text{blocks}}\{\beta h_{\max}^2\}$, where $\beta = 0.1$, and $h_{\max} = 1/(m-1)$ or $h_{\max}=r_{\max}$ the maximum element diameter depending on the method of the block. The solution is examined at $t=2$. The convergence rate is denoted by $Q$ and calculated by \eqref{eq:convergence_rate}. We respectively denote the number of FD grid points in each dimension $m$ (FD) and the maximum element diameter $r_{\max}$ (FE). For example, in the experiments where $m$ (FD) $=91$ and $r_{\max}$ (FE) $=D/91$, the discretization is carried out by $24843$ FD grid points and $27764$ Delaunay triangulated finite elements ($14063$ nodal points).
In Tables \ref{table:2by2_multiblock_test1} and \ref{table:2by2_multiblock_test2}, $m$ is equal to both grid size in the FD domain ($m=n_L=m_L$), and the number of discretization points along each axis of the FE mesh (matching interface). \cblue{In Table \ref{table:2by2_multiblock_test3} and \ref{table:2by2_multiblock_test4}, we perform the same study but when there are twice as many FE mesh points as FD grid points being located on the interfaces}. The obtained rates are expected between second-order FE approximation and fourth-order FD approximation. Order degradation by the edges/corner point is not observed. Simpler experiments involving equal distribution of the two methods can be found in \citet{Dao2019}. There one can see that the global convergence is bounded by the lowest second-order rate in $\mathbb P^1$ FE regions.
\begin{figure}[pos=h]
\centering
\begin{subfigure}{0.33\textwidth}
\centering
\includegraphics[scale=0.5]{2by2_domain.eps}
\caption{}
\end{subfigure}
\begin{subfigure}{0.32\textwidth}
\centering
\vspace{0.1cm}
\includegraphics[scale=0.505]{2by2_multiblock_test1.eps}
\caption{}
\end{subfigure}
\begin{subfigure}{0.32\textwidth}
\centering
\vspace{0.1cm}
\includegraphics[scale=0.505]{2by2_multiblock_test2.eps}
\caption{}
\end{subfigure}
\caption{The computational domain (a) and the two test cases (b) and (c).}
\label{fig:2by2_domain}
\end{figure}
\begin{table}[pos=h]
\centering
\caption{Test case (b), Delaunay triangulated mesh, Fourth-order FD, $\mathbb{P}^1$ FE, $\bm a=(1,0)^T,x_0=-1,y_0=0,t=2$.}
\label{table:2by2_multiblock_test1}
\begin{tabular}{c|l|*4c|*4c}
\hline
\multirow{2}{*}{\scalebox{.8}{$m$ (FD)}} & \multirow{2}{*}{\scalebox{.8}{$r_{\max}$ (FE)}} & \multicolumn{4}{c|}{inviscid $\epsilon=0$} & \multicolumn{4}{c}{viscous $\epsilon=0.01$} \\ \cline{3-10}
& & \scalebox{.8}{$l_2e(\bm u_{FD})$} & \scalebox{.8}{$l_2e(\bm u_{FE})$} & \scalebox{.8}{$l_2e(\bm u)$} & $Q$ & \scalebox{.8}{$l_2e(\bm u_{FD})$} & \scalebox{.8}{$l_2e(\bm u_{FE})$} & \scalebox{.8}{$l_2e(\bm u)$} & $Q$\\ \hline
41 & D/41 & 2.46E-2 & 5.25E-3 & 3.00E-2 & -- &6.45E-3 & 4.07E-3 & 1.06E-2 & -- \\
51 & D/51 & 1.27E-2 & 2.69E-3 & 1.54E-2 & 2.97 & 2.45E-3 & 1.25E-3 & 3.73E-3 & 4.66 \\
61 & D/61 & 7.92E-3 & 1.76E-3 & 9.68E-3 & 2.57 & 9.55E-4 & 3.16E-4 & 1.28E-3 & 5.86 \\
71 & D/71 & 5.49E-3 & 1.31E-3 & 6.79E-3 & 2.30 & 4.16E-4 & 4.36E-4 & 8.50E-4 & 2.66 \\
81 & D/81 & 3.98E-3 & 1.04E-3 & 5.00E-3 & 2.28 & 2.45E-4 & 5.00E-4 & 7.39E-4 & 1.05 \\
91 & D/91 & 2.88E-3 & 7.94E-4 & 3.71E-3 & 2.54 & 1.54E-4 & 4.07E-4 & 5.59E-4 & 2.37 \\ \hline
\end{tabular}
\end{table}
\begin{table}[pos=h]
\centering
\caption{Test case (c), Delaunay triangulated mesh, Fourth-order FD, $\mathbb{P}^1$ FE, $\bm a=(1,1)^T,x_0=-1,y_0=-1,t=2$.}
\label{table:2by2_multiblock_test2}
\begin{tabular}{c|l|*4c|*4c}
\hline
\multirow{2}{*}{\scalebox{.8}{$m$ (FD)}} & \multirow{2}{*}{\scalebox{.8}{$r_{\max}$ (FE)}} & \multicolumn{4}{c|}{inviscid $\epsilon=0$} & \multicolumn{4}{c}{viscous $\epsilon=0.01$} \\ \cline{3-10}
& & \scalebox{.8}{$l_2e(\bm u_{FD})$} & \scalebox{.8}{$l_2e(\bm u_{FE})$} & \scalebox{.8}{$l_2e(\bm u)$} & $Q$ & \scalebox{.8}{$l_2e(\bm u_{FD})$} & \scalebox{.8}{$l_2e(\bm u_{FE})$} & \scalebox{.8}{$l_2e(\bm u)$} & $Q$\\ \hline
41 & D/41 & 4.26E-2 & 4.89E-3 & 4.77E-2 & -- & 1.73E-3 & 3.38E-7 & 1.74E-3 & -- \\
51 & D/51 & 2.29E-2 & 2.81E-3 & 2.59E-2 & 2.73 & 7.94E-4 & 1.25E-7 & 7.92E-4 & 3.54 \\
61 & D/61 & 1.38E-2 & 1.23E-3 & 1.50E-2 & 3.00 & 4.26E-4 & 1.20E-7 & 4.24E-4 & 3.43 \\
71 & D/71 & 8.91E-3 & 9.65E-4 & 9.89E-3 & 2.71 & 2.69E-4 & 1.14E-7 & 2.70E-4 & 2.92 \\
81 & D/81 & 6.30E-3 & 5.37E-4 & 6.78E-3 & 2.83 & 1.82E-4 & 1.00E-7 & 1.83E-4 & 2.91 \\
91 & D/91 & 4.78E-3 & 3.80E-4 & 5.11E-3 & 2.40 & 1.34E-4 & 6.02E-8 & 1.34E-4 & 2.68 \\ \hline
\end{tabular}
\end{table}
\begin{table}[pos=h]
\centering
\caption{Nonconformal Test case (b), Delaunay triangulated mesh, Fourth-order FD, $\mathbb{P}^1$ FE, $\bm a=(1,0)^T,x_0=-1,y_0=0,t=2$.}
\label{table:2by2_multiblock_test3}
\begin{tabular}{c|l|*4c|*4c}
\hline
\multirow{2}{*}{\scalebox{.8}{$m$ (FD)}} & \multirow{2}{*}{\scalebox{.8}{$r_{\max}$ (FE)}} & \multicolumn{4}{c|}{inviscid $\epsilon=0$} & \multicolumn{4}{c}{viscous $\epsilon=0.01$} \\ \cline{3-10}
& & \scalebox{.8}{$l_2e(\bm u_{FD})$} & \scalebox{.8}{$l_2e(\bm u_{FE})$} & \scalebox{.8}{$l_2e(\bm u)$} & $Q$ & \scalebox{.8}{$l_2e(\bm u_{FD})$} & \scalebox{.8}{$l_2e(\bm u_{FE})$} & \scalebox{.8}{$l_2e(\bm u)$} & $Q$\\ \hline
21 & D/41 & 1.54E-1 & 3.39E-2 & 1.86E-1 & -- & 3.31E-2 & 1.55E-2 & 4.90E-2 & -- \\
26 & D/51 & 9.55E-2 & 2.00E-2 & 1.15E-1 & 2.20 & 2.45E-2 & 1.74E-2 & 4.17E-2 & 0.65 \\
31 & D/61 & 5.88E-2 & 1.17E-2 & 7.08E-2 & 2.67 & 1.48E-2 & 1.17E-2 & 2.69E-2 & 2.47 \\
36 & D/71 & 3.71E-2 & 7.41E-3 & 4.47E-2 & 3.00 & 9.33E-3 & 7.76E-3 & 1.70E-2 & 2.93 \\
41 & D/81 & 2.45E-2 & 5.01E-3 & 2.95E-2 & 3.09 & 5.62E-3 & 4.90E-3 & 1.05E-2 & 3.57 \\
46 & D/91 & 1.73E-2 & 3.39E-3 & 2.04E-2 & 3.05 & 3.31E-3 & 3.02E-3 & 6.31E-3 & 4.30 \\ \hline
\end{tabular}
\end{table}
\begin{table}[pos=h]
\centering
\caption{Nonconformal Test case (c), Delaunay triangulated mesh, Fourth-order FD, $\mathbb{P}^1$ FE, $\bm a=(1,1)^T,x_0=-1,y_0=-1,t=2$.}
\label{table:2by2_multiblock_test4}
\begin{tabular}{c|l|*4c|*4c}
\hline
\multirow{2}{*}{\scalebox{.8}{$m$ (FD)}} & \multirow{2}{*}{\scalebox{.8}{$r_{\max}$ (FE)}} & \multicolumn{4}{c|}{inviscid $\epsilon=0$} & \multicolumn{4}{c}{viscous $\epsilon=0.01$} \\ \cline{3-10}
& & \scalebox{.8}{$l_2e(\bm u_{FD})$} & \scalebox{.8}{$l_2e(\bm u_{FE})$} & \scalebox{.8}{$l_2e(\bm u)$} & $Q$ & \scalebox{.8}{$l_2e(\bm u_{FD})$} & \scalebox{.8}{$l_2e(\bm u_{FE})$} & \scalebox{.8}{$l_2e(\bm u)$} & $Q$\\ \hline
21 & D/41 & 1.91E-1 & 5.25E-2 & 2.40E-1 & -- & 2.40E-2 & 1.58E-3 & 2.57E-2 & -- \\
26 & D/51 & 1.29E-1 & 3.09E-2 & 1.58E-1 & 1.88 & 9.33E-3 & 4.90E-5 & 9.33E-3 & 4.50 \\
31 & D/61 & 8.71E-2 & 1.62E-2 & 1.05E-1 & 2.37 & 4.68E-3 & 1.95E-6 & 4.68E-3 & 3.82 \\
36 & D/71 & 5.89E-2 & 1.00E-2 & 6.92E-2 & 2.69 & 2.69E-3 & 1.07E-6 & 2.69E-3 & 3.60 \\
41 & D/81 & 3.89E-2 & 6.76E-3 & 4.57E-2 & 3.03 & 1.66E-3 & 2.24E-7 & 1.66E-3 & 3.70 \\
46 & D/91 & 2.88E-2 & 3.98E-3 & 3.24E-2 & 2.83 & 1.05E-3 & 1.05E-7 & 1.05E-3 & 3.79 \\ \hline
\end{tabular}
\end{table}
\newpage
\newpage
\subsubsection{Superconvergence preservation}
For the convergence results, we use \textit{regular meshes} for the FE discretization, illustrated in Figure \ref{fig:computational_domain}, $\Omega = \Omega_L\cup\Omega_R= ([-2,0] \times [-1,1])\cup ([-2,0] \times [-1,1])$, where FD discretization is used on $\Omega_L$, and FE discretization is used on $\Omega_R$.
\begin{figure}[pos=h]
\centering
\includegraphics[scale=0.7]{regular_mesh.eps}
\caption{The computational mesh for superconvergence study.}
\label{fig:computational_domain}
\end{figure}
\begin{table}[pos=h]
\caption{Superconvergence study on inviscid linear problem $\epsilon=0$, mesh type Figure \ref{fig:computational_domain} with $m$ grid points each direction each domain.}
\label{table:linear_superconvergence}
\centering
\begin{tabular}{c|*4c|*4c}
\hline
\multirow{2}{*}{$m$} & \multicolumn{4}{c|}{$\mathbb{P}^1$ FE to fourth-order FD} & \multicolumn{4}{c}{Fourth-order FD to $\mathbb{P}^1$ FE} \\ \cline{2-9}
& \scalebox{.8}{$l_2e(\bm u_{FD})$} & \scalebox{.8}{$l_2e(\bm u_{FE})$} & \scalebox{.8}{$l_2e(\bm u)$} & $Q$ & \scalebox{.8}{$l_2e(\bm u_{FD})$} & \scalebox{.8}{$l_2e(\bm u_{FE})$} & \scalebox{.8}{$l_2e(\bm u)$} & $Q$\\ \hline
31 & 1.86E-2 & 1.91E-2 & 3.72E-2 & -- & 1.95E-2 & 1.86E-2 & 3.80E-2 & --\\
41 & 7.59E-3 & 6.76E-3 & 1.45E-2 & 3.32 & 7.08E-3 & 7.59E-3 & 1.48E-2 & 3.28 \\
51 & 3.72E-3 & 2.69E-3 & 6.31E-3 & 3.67 & 2.88E-3 & 3.72E-3 & 6.61E-3 & 3.58 \\
61 & 2.14E-3 & 1.29E-3 & 3.39E-3 & 3.40 & 1.48E-3 & 2.14E-3 & 3.63E-3 & 3.26 \\ \hline
\end{tabular}
\end{table}
We test propagating the solution by both direction, from the FE domain to the FD domain $(x_0,y_0)^T=(1,0)^T,\bm a = (-1,0)^T$, and in the opposite direction $(x_0,y_0)^T=(-1,0)^T,\bm a = (1,0)^T$. \cgreen{The solution is evaluated at $t = 2$}. The convergence rates in fourth-order FD cases reported in Table \ref{table:linear_superconvergence} are noticeably close to a fourth-order--fourth-order coupling, cf., FD--FD fourth-order case in \citet{Mattsson2010}. We observe the so-called \textit{superconvergence} phenomenon in FE approximations, see e.g., \citet{Andreev1988}. At some special points, the order of convergence is higher than the convergence rate of the total error if calculated analytically. Those points are called \textit{superconvergence points}, and they happen to be the nodal points of the regular mesh Figure \ref{fig:computational_domain} with the advection operator $\bm a\cdot\nabla v$. In the viscous case, there also exist \cblue{such} points which give rise to this behavior. It turns out that \eqref{eq:discrete_error_measurement} is not the most correct way to assess the discrete error in the sense of mimicking the continuous measurement $\norm{e}_{L^2(\Omega)} = \sqrt{\int_\Omega e^2 dxdy}$. A straightforward fix to avoid this problem is calculating the error with many more sample points added to the computation mesh. From a different angle, it cannot be denied that \eqref{eq:discrete_error_measurement} is desirable once our FE approximation fits in the SBP framework. (Simply, the defined discrete norm was applied.) To that extent, interpreting the advection operator as fourth-order and the diffusion operator as second-order may explain all the rates obtained. The authors have done separate numerical experiments on the FE approximation to verify that the current advection operator is fourth-order on the regular mesh regarding the error assessment \eqref{eq:discrete_error_measurement}. This is strong evidence that the proposed coupling technique preserves the native properties of the existing schemes.
\subsection{Nonlinear Burgers' Equation in 2D}
\label{sec:burger}
We consider the following nonlinear problem in two spatial dimensions
\begin{equation}
\begin{aligned}
\label{eq:burger_fem}
u_t + \bm a\cdot\nabla\left(\frac{u^2}{2}\right) & = \nabla\cdot(\epsilon\nabla u), && (x,y) \in \Omega,\\
u & = u_0(x,y),&&(x,y) \in \Omega, t = 0.
\end{aligned}
\end{equation}
Inserting $\bm f = \left(\frac{u^2}{2}, \frac{u^2}{2}\right)^T$ to \eqref{eq:skew_symmetric_precise} gives $\alpha_1=\alpha_2=\frac{2}{3}$. From \eqref{eq:advection_diffusion_fem_bc}, a set of stable boundary condition reads
\begin{align}
\label{eq:burger_fem_bc}
\frac{1}{3}u(\bm a\cdot\mathbf{n}u-\abs{\bm a\cdot\mathbf{n}u})-\epsilon \mathbf{n}\cdot\nabla u = g(x,y,t), \quad(x,y) \in \partial\Omega, t > 0,
\end{align}
where $\bm a = (a_1, a_2)^T>0$ is a constant vector and $\epsilon$ is a real constant.
\subsubsection*{The FD--FD coupling}
A FD--FD semi-discretization of the coupling problem \eqref{eq:advection_diffusion_coupling_continuous} under the governing equation \eqref{eq:burger_fem} is given by
\begin{equation}
\label{eq:burger_sbp_sat}
\begin{aligned}
\bm u_t &= -\frac{1}{3}a_1{\bm U}\mathbf{D}_x\bm u -\frac{1}{3}a_1\mathbf{D}_x{\bm U}\bm u - \frac{1}{3}a_2{\bm U}\mathbf{D}_y\bm u -\frac{1}{3}a_2\mathbf{D}_y{\bm U}\bm u + \epsilon \mathbf{D}_{2x}\bm u + \epsilon \mathbf{D}_{2y}\bm u + \mbox{SAT}_L+\mbox{SAT}_{IL},\\
\bm v_t &= -\frac{1}{3}a_1\bm V\mathbf{D}_x\bm v -\frac{1}{3}a_1\mathbf{D}_x\bm V\bm v - \frac{1}{3}a_2\bm V\mathbf{D}_y\bm v -\frac{1}{3}a_2\mathbf{D}_y\bm V\bm v + \epsilon \mathbf{D}_{2x}\bm v + \epsilon \mathbf{D}_{2y}\bm v + \mbox{SAT}_R+\mbox{SAT}_{IR},
\end{aligned}
\end{equation}
where the interface SATs are
\begin{equation}
\label{eq:burger_sat_interface}
\begin{aligned}
\mbox{SAT}_{IL} & = \frac{1}{6} ({\bm H_x^L}^{-1} \bm e_n)\otimes (({\bm U}\bm u)_I-\bm{I}_{R2L}(\bm V\bm v)_I)+\frac{1}{6} ({\bm H_x^L}^{-1} \bm e_n)\otimes \left[{\bm U_I}(\bm u_I-\bm{I}_{R2L}\bm v_I)\right]\\
&+ \delta_L ({\bm H_x^L}^{-1} \bm e_n) \otimes ((\mathbf{D}_xu)_I - \bm{I}_{R2L}(\mathbf{D}_x\bm v)_I) + \sigma_L ({\bm H_x^L}^{-1}\bm d_n^T)\otimes(\bm u_I-\bm{I}_{R2L}\bm v_I),\\
\mbox{SAT}_{IR} & = -\frac{1}{6}({\bm H_x^R}^{-1} \bm e_1) \otimes ((\bm V\bm v)_I-\bm{I}_{L2R}({\bm U}\bm u)_I)-\frac{1}{6}({\bm H_x^R}^{-1} \bm e_1) \otimes \left[\overline{\bm v_I}(\bm v_I-\bm{I}_{L2R}\bm u_I)\right]\\
&+\delta_R ({\bm H_x^R}^{-1} \bm e_1) \otimes ((\mathbf{D}_x\bm v)_I-\bm{I}_{L2R}(\mathbf{D}_x\bm u)_I)+ \sigma_R ({\bm H_x^R}^{-1} \bm d_1^T) \otimes (\bm v_I-\bm{I}_{L2R}\bm u_I),
\end{aligned}
\end{equation}
where $\delta_L=-\sigma_R, \sigma_L =-\delta_R, \delta_L+\sigma_L = -1$.
\subsubsection*{The FD--FE coupling}
A FD--FE semi-discretization of the coupling problem \eqref{eq:advection_diffusion_coupling_continuous} under the governing equation \eqref{eq:burger_fem} is given by
\begin{equation}
\label{eq:burger_fe_sbp_sat}
\begin{aligned}
\bm u_t &= -\frac{1}{3}a_1{\bm U}\mathbf{D}_x\bm u -\frac{1}{3}a_1\mathbf{D}_x{\bm U}\bm u - \frac{1}{3}a_2{\bm U}\mathbf{D}_y\bm u -\frac{1}{3}a_2\mathbf{D}_y{\bm U}\bm u + \epsilon \mathbf{D}_{2x}\bm u + \epsilon \mathbf{D}_{2y}\bm u + \mbox{SAT}_L+\mbox{SAT}_{IL},\\
\bm v_t &= -\mathbf{M}^{-1}\mathbf{C}(\bm v)\bm v-\epsilon \mathbf{M}^{-1}\mathbf{A}\bm v+\frac{a_2}{3}\mathbf{M}^{-1}[\mathscr{R}_M^{N}(\bm v)-\mathscr{R}_M^{N}(\abs{\bm v})]\bm v\\
&\hspace{0.5cm}+\frac{a_1}{3}\mathbf{M}^{-1}[\mathscr{R}_M^{E}(\bm v)-\mathscr{R}_M^{E}(\abs{\bm v})]\bm v+\frac{a_2}{3}\mathbf{M}^{-1}[-\mathscr{R}_M^{S}(\bm v)-\mathscr{R}_M^{S}(\abs{\bm v})]\bm v - \bm r +\mbox{SAT}_{IR},
\end{aligned}
\end{equation}
where $\mathscr{R}_M^N,\mathscr{R}_M^E,\mathscr{R}_M^S$ are obtained by limiting the integration of $\mathscr{R}_M$ to $\Gamma_N,\Gamma_E,\Gamma_S$, respectively. The coupling penalty terms are given by
\begin{equation}
\label{eq:burger_fe_interface}
\begin{aligned}
\mbox{SAT}_{IL} & = \frac{1}{6} (H_x^{-1} \bm e_n)\otimes (({\bm U}\bm u)_I-\bm{I}_{R2L}\bm L_W\bm V\bm v)+\frac{1}{6} (H_x^{-1} \bm e_n)\otimes \left[{\bm U_I}(\bm u_I-\bm{I}_{R2L}\bm L_W\bm v)\right]\\
&+ \delta_L\epsilon(H_x^{-1} \bm e_n) \otimes ((\mathbf{D}_x\bm u)_I - \bm{I}_{R2L}\mathbf{D}_x^W\bm v)+\sigma_L\epsilon(H_x^{-1}\bm d_n^T)\otimes(\bm u_I-\bm{I}_{R2L}\bm L_W\bm v),\\
\mbox{SAT}_{IR} & = -\frac{1}{6}\mathbf{M}^{-1}\mathbf{R}_M^W(\bm V\bm v-\bm L_W^T\bm{I}_{L2R}({\bm U}\bm u)_I)-\frac{1}{6}\mathbf{M}^{-1}\bm V\mathbf{R}_M^W\bm L_W^T(\bm L_W\bm v-\bm{I}_{L2R}\bm u_I)\\
&+\delta_R\epsilon\mathbf{M}^{-1}\mathbf{R}_M^W\bm L_W^T(\mathbf{D}_x^W\bm v-\bm{I}_{L2R}(\mathbf{D}_x\bm u)_I) +\sigma_R\epsilon\mathbf{M}^{-1}(-\mathbf{R}^W_C)^T\bm L_W^T(\bm L_W\bm v - \bm{I}_{L2R}\bm u_I),
\end{aligned}
\end{equation}
where $\delta_L=-\sigma_R, \sigma_L =-\delta_R, \delta_L+\sigma_L = -1$.
\subsubsection*{Convergence study}
An analytic solution of \eqref{eq:burger_fem} subject to the boundary condition \eqref{eq:burger_fem_bc} is
\[
\cblue{u(x,y,t) = \frac{c}{a_1^2+a_2^2} - b \tanh\left(b\frac{a_1(x-x_0)+a_2(y-y_0)-ct}{2\epsilon}\right), \quad t > 0,}
\]
where $b$ and $c$ are arbitrary constants. For the following convergence study, we choose $b = 1, c = 2$. The semi-discrete scheme \eqref{eq:burger_fe_sbp_sat} is integrated in time using the standard Runge-Kutta 4. The time step size is chosen to be $\Delta t = \beta h^2, \beta = 0.1$. The domain setup is as in Figure \ref{fig:coupling-interface}, where $\Omega_L=[-2,0]\times[-1,1]$, and $\Omega_R=[0,2]\times[-1,1]$. The problem settings are: $\cblue{\bm a = (1,1)^T}, \epsilon = 0.1,x_0=-1,\cblue{y_0=-1}$. The solution is examined at $t = 1$. A convergence study is shown in Tables \ref{table:nonlinear_convergence_fd_fe} and \ref{table:nonlinear_convergence_fd_fe_nonmatching}. The notations in Tables \ref{table:nonlinear_convergence_fd_fe} and \ref{table:nonlinear_convergence_fd_fe_nonmatching} are same to the ones used in Table \ref{table:2by2_multiblock_test1}. In both tables, we can observe the expected convergence rate of second-order. It can also be seen that the use of higher-order FD operators slightly lowers the error and improves convergence rates.
\begin{table}[pos=h]
\centering
\caption{Convergence rates for the case of matching interface.}
\label{table:nonlinear_convergence_fd_fe}
\cblue{
\begin{tabular}{c|*4c|*4c}
\hline
\multirow{2}{*}{$m$ (FD,FE)} & \multicolumn{4}{c|}{Second-order FD} & \multicolumn{4}{c}{Fourth-order FD} \\ \cline{2-9}
& \scalebox{.8}{$l_2e(\bm u_{FD})$} & \scalebox{.8}{$l_2e(\bm u_{FE})$} & \scalebox{.8}{$l_2e(\bm u)$} & $Q$ & \scalebox{.8}{$l_2e(\bm u_{FD})$} & \scalebox{.8}{$l_2e(\bm u_{FE})$} & \scalebox{.8}{$l_2e(\bm u)$} & $Q$\\ \hline
21 & 2.91E-02 & 1.83E-02 & 4.73E-02 & -- & 1.80E-02 & 1.73E-02 & 3.53E-02 & -- \\
31 & 1.31E-02 & 9.19E-03 & 2.23E-02 & 1.85 & 7.64E-03 & 8.81E-03 & 1.64E-02 & 1.89 \\
41 & 7.43E-03 & 5.50E-03 & 1.29E-02 & 1.90 & 4.22E-03 & 5.30E-03 & 9.52E-03 & 1.89 \\
51 & 4.77E-03 & 3.64E-03 & 8.42E-03 & 1.91 & 2.67E-03 & 3.52E-03 & 6.19E-03 & 1.93 \\ \hline
\end{tabular}}
\end{table}
\begin{table}[pos=h]
\centering
\caption{Convergence rates for the case of 2 to 1 mesh ratio nonconformal coupling.}
\label{table:nonlinear_convergence_fd_fe_nonmatching}
\cblue{
\begin{tabular}{c|c|*4c|*4c}
\hline
\multirow{2}{*}{$m$ (FD)} & \multirow{2}{*}{$r_{\max}$ (FE)} & \multicolumn{4}{c|}{Second-order FD} & \multicolumn{4}{c}{Fourth-order FD} \\ \cline{3-10}
& & \scalebox{.8}{$l_2e(\bm u_{FD})$} & \scalebox{.8}{$l_2e(\bm u_{FE})$} & \scalebox{.8}{$l_2e(\bm u)$} & $Q$ & \scalebox{.8}{$l_2e(\bm u_{FD})$} & \scalebox{.8}{$l_2e(\bm u_{FE})$} & \scalebox{.8}{$l_2e(\bm u)$} & $Q$\\ \hline
11 & D/21 & 1.10E-01 & 2.14E-02 & 1.31E-01 & -- & 6.44E-02 & 2.11E-02 & 8.54E-02 & -- \\
21 & D/41 & 2.49E-02 & 5.93E-03 & 3.09E-02 & 2.08 & 8.44E-03 & 5.59E-03 & 1.40E-02 & 2.61 \\
31 & D/61 & 1.07E-02 & 2.75E-03 & 1.34E-02 & 2.06 & 3.11E-03 & 2.61E-03 & 5.72E-03 & 2.21 \\ \hline
\end{tabular}}
\end{table}
\section{Conclusion}
\label{sec:conclusion}
\cpurple{The multiblock coupling of FD schemes and FE schemes in this paper is presented within the SBP framework. The technique uses a SAT technique to impose the interface continuity weakly. SBP-preserving interpolation operators resolve the nonconforming interface distribution.} Accuracy, stability, and conservation are investigated on general scalar conservation laws. Numerical results are shown for a linear hyperbolic/parabolic problem and a nonlinear problem.\\
Future works may concern extensions for curvilinear grids, optimal time-stepping, an option for upwind operators \cite{Lundgren2020}, high-order finite elements or a solution to the case when block corners are unmatching. The work can be structurally extended for nonlinear systems.
\section{Introduction}
The Elsevier cas-sc class is based on the
standard article class and supports almost all of the functionality of
that class. In addition, it features commands and options to format the
\begin{itemize} \item document style \item baselineskip \item front
matter \item keywords and MSC codes \item theorems, definitions and
proofs \item lables of enumerations \item citation style and labeling.
\end{itemize}
This class depends on the following packages
for its proper functioning:
\begin{enumerate}
\itemsep=0pt
\item {natbib.sty} for citation processing;
\item {geometry.sty} for margin settings;
\item {fleqn.clo} for left aligned equations;
\item {graphicx.sty} for graphics inclusion;
\item {hyperref.sty} optional packages if hyperlinking is
required in the document;
\end{enumerate}
All the above packages are part of any
standard \LaTeX{} installation.
Therefore, the users need not be
bothered about downloading any extra packages.
\section{Installation}
The package is available at author resources page at Elsevier
(\url{http://www.elsevier.com/locate/latex}).
The class may be moved or copied to a place, usually,
\verb+$TEXMF/tex/latex/elsevier/+,
or a folder which will be read
by \LaTeX{} during document compilation. The \TeX{} file
database needs updation after moving/copying class file. Usually,
we use commands like \verb+mktexlsr+ or \verb+texhash+ depending
upon the distribution and operating system.
\section{Front matter}
The author names and affiliations could be formatted in two ways:
\begin{enumerate}[(1)]
\item Group the authors per affiliation.
\item Use footnotes to indicate the affiliations.
\end{enumerate}
See the front matter of this document for examples.
You are recommended to conform your choice to the journal you
are submitting to.
\section{Bibliography styles}
There are various bibliography styles available. You can select the
style of your choice in the preamble of this document. These styles are
Elsevier styles based on standard styles like Harvard and Vancouver.
Please use Bib\TeX\ to generate your bibliography and include DOIs
whenever available.
Here are two sample references:
See \citet{Fortunato2010}. Also refer \citet{Fortunato2010,NewmanGirvan2004}.
More citations are here \citep{Fortunato2010,Vehlowetal2013}.
\section{Floats}
{Figures} may be included using the command, \verb+\includegraphics+ in
combination with or without its several options to further control
graphic. \verb+\includegraphics+ is provided by {graphic[s,x].sty}
which is part of any standard \LaTeX{} distribution.
{graphicx.sty} is loaded by default. \LaTeX{} accepts figures in
the postscript format while pdf\LaTeX{} accepts {*.pdf},
{*.mps} (metapost), {*.jpg} and {*.png} formats.
pdf\LaTeX{} does not accept graphic files in the postscript format.
\begin{figure}
\centering
\includegraphics[scale=.75]{figs/Fig1.pdf}
\caption{The evanescent light - $1S$ quadrupole coupling
($g_{1,l}$) scaled to the bulk exciton-photon coupling
($g_{1,2}$). The size parameter $kr_{0}$ is denoted as $x$ and
the \PMS is placed directly on the cuprous oxide sample ($\delta
r=0$, See also Table \protect\ref{tbl1}).}
\label{FIG:1}
\end{figure}
The \verb+table+ environment is handy for marking up tabular
material. If users want to use {multirow.sty},
{array.sty}, etc., to fine control/enhance the tables, they
are welcome to load any package of their choice and
{cas-sc.cls} will work in combination with all loaded
packages.
\begin{table}[width=.9\linewidth,cols=4,pos=h]
\caption{This is a test caption. This is a test caption. This is a test
caption. This is a test caption.}\label{tbl1}
\begin{tabular*}{\tblwidth}{@{} LLLL@{} }
\toprule
Col 1 & Col 2 & Col 3 & Col4\\
\midrule
12345 & 12345 & 123 & 12345 \\
12345 & 12345 & 123 & 12345 \\
12345 & 12345 & 123 & 12345 \\
12345 & 12345 & 123 & 12345 \\
12345 & 12345 & 123 & 12345 \\
\bottomrule
\end{tabular*}
\end{table}
\section[Theorem and ...]{Theorem and theorem like environments}
{cas-sc.cls} provides a few shortcuts to format theorems and
theorem-like environments with ease. In all commands the options that
are used with the \verb+\newtheorem+ command will work exactly in the same
manner. {cas-sc.cls} provides three commands to format theorem or
theorem-like environments:
\begin{verbatim}
\newtheorem{theorem}{Theorem}
\newtheorem{lemma}[theorem]{Lemma}
\newdefinition{rmk}{Remark}
\newproof{pf}{Proof}
\newproof{pot}{Proof of Theorem \ref{thm2}}
\end{verbatim}
The \verb+\newtheorem+ command formats a
theorem in \LaTeX's default style with italicized font, bold font
for theorem heading and theorem number at the right hand side of the
theorem heading. It also optionally accepts an argument which
will be printed as an extra heading in parentheses.
\begin{verbatim}
\begin{theorem}
For system (8), consensus can be achieved with
$\|T_{\omega z}$ ...
\begin{eqnarray}\label{10}
....
\end{eqnarray}
\end{theorem}
\end{verbatim}
\newtheorem{theorem}{Theorem}
\begin{theorem}
For system (8), consensus can be achieved with
$\|T_{\omega z}$ ...
\begin{eqnarray}\label{10}
....
\end{eqnarray}
\end{theorem}
The \verb+\newdefinition+ command is the same in
all respects as its \verb+\newtheorem+ counterpart except that
the font shape is roman instead of italic. Both
\verb+\newdefinition+ and \verb+\newtheorem+ commands
automatically define counters for the environments defined.
The \verb+\newproof+ command defines proof environments with
upright font shape. No counters are defined.
\section[Enumerated ...]{Enumerated and Itemized Lists}
{cas-sc.cls} provides an extended list processing macros
which makes the usage a bit more user friendly than the default
\LaTeX{} list macros. With an optional argument to the
\verb+\begin{enumerate}+ command, you can change the list counter
type and its attributes.
\begin{verbatim}
\begin{enumerate}[1.]
\item The enumerate environment starts with an optional
argument `1.', so that the item counter will be suffixed
by a period.
\item You can use `a)' for alphabetical counter and '(i)' for
roman counter.
\begin{enumerate}[a)]
\item Another level of list with alphabetical counter.
\item One more item before we start another.
\item One more item before we start another.
\item One more item before we start another.
\item One more item before we start another.
\end{verbatim}
Further, the enhanced list environment allows one to prefix a
string like `step' to all the item numbers.
\begin{verbatim}
\begin{enumerate}[Step 1.]
\item This is the first step of the example list.
\item Obviously this is the second step.
\item The final step to wind up this example.
\end{enumerate}
\end{verbatim}
\section{Cross-references}
In electronic publications, articles may be internally
hyperlinked. Hyperlinks are generated from proper
cross-references in the article. For example, the words
\textcolor{black!80}{Fig.~1} will never be more than simple text,
whereas the proper cross-reference \verb+\ref{tiger}+ may be
turned into a hyperlink to the figure itself:
\textcolor{blue}{Fig.~1}. In the same way,
the words \textcolor{blue}{Ref.~[1]} will fail to turn into a
hyperlink; the proper cross-reference is \verb+\cite{Knuth96}+.
Cross-referencing is possible in \LaTeX{} for sections,
subsections, formulae, figures, tables, and literature
references.
\section{Bibliography}
Two bibliographic style files (\verb+*.bst+) are provided ---
{model1-num-names.bst} and {model2-names.bst} --- the first one can be
used for the numbered scheme. This can also be used for the numbered
with new options of {natbib.sty}. The second one is for the author year
scheme. When you use model2-names.bst, the citation commands will be
like \verb+\citep+, \verb+\citet+, \verb+\citealt+ etc. However when
you use model1-num-names.bst, you may use only \verb+\cite+ command.
\verb+thebibliography+ environment. Each reference is a
\verb+\bibitem+ and each \verb+\bibitem+ is identified by a label,
by which it can be cited in the text:
\noindent In connection with cross-referencing and
possible future hyperlinking it is not a good idea to collect
more that one literature item in one \verb+\bibitem+. The
so-called Harvard or author-year style of referencing is enabled
by the \LaTeX{} package {natbib}. With this package the
literature can be cited as follows:
\begin{enumerate}[\textbullet]
\item Parenthetical: \verb+\citep{WB96}+ produces (Wettig \& Brown, 1996).
\item Textual: \verb+\citet{ESG96}+ produces Elson et al. (1996).
\item An affix and part of a reference:
\verb+\citep[e.g.][Ch. 2]{Gea97}+ produces (e.g. Governato et
al., 1997, Ch. 2).
\end{enumerate}
In the numbered scheme of citation, \verb+\cite{<label>}+ is used,
since \verb+\citep+ or \verb+\citet+ has no relevance in the numbered
scheme. {natbib} package is loaded by {cas-sc} with
\verb+numbers+ as default option. You can change this to author-year
or harvard scheme by adding option \verb+authoryear+ in the class
loading command. If you want to use more options of the {natbib}
package, you can do so with the \verb+\biboptions+ command. For
details of various options of the {natbib} package, please take a
look at the {natbib} documentation, which is part of any standard
\LaTeX{} installation.
\section{Introduction}
The Elsevier cas-dc class is based on the
standard article class and supports almost all of the functionality of
that class. In addition, it features commands and options to format the
\begin{itemize} \item document style \item baselineskip \item front
matter \item keywords and MSC codes \item theorems, definitions and
proofs \item lables of enumerations \item citation style and labeling.
\end{itemize}
This class depends on the following packages
for its proper functioning:
\begin{enumerate}
\itemsep=0pt
\item {natbib.sty} for citation processing;
\item {geometry.sty} for margin settings;
\item {fleqn.clo} for left aligned equations;
\item {graphicx.sty} for graphics inclusion;
\item {hyperref.sty} optional packages if hyperlinking is
required in the document;
\end{enumerate}
All the above packages are part of any
standard \LaTeX{} installation.
Therefore, the users need not be
bothered about downloading any extra packages.
\section{Installation}
The package is available at author resources page at Elsevier
(\url{http://www.elsevier.com/locate/latex}).
The class may be moved or copied to a place, usually,\linebreak
\verb+$TEXMF/tex/latex/elsevier/+,
or a folder which will be read
by \LaTeX{} during document compilation. The \TeX{} file
database needs updation after moving/copying class file. Usually,
we use commands like \verb+mktexlsr+ or \verb+texhash+ depending
upon the distribution and operating system.
\section{Front matter}
The author names and affiliations could be formatted in two ways:
\begin{enumerate}[(1)]
\item Group the authors per affiliation.
\item Use footnotes to indicate the affiliations.
\end{enumerate}
See the front matter of this document for examples.
You are recommended to conform your choice to the journal you
are submitting to.
\section{Bibliography styles}
There are various bibliography styles available. You can select the
style of your choice in the preamble of this document. These styles are
Elsevier styles based on standard styles like Harvard and Vancouver.
Please use Bib\TeX\ to generate your bibliography and include DOIs
whenever available.
Here are two sample references:
\cite{Fortunato2010}
\cite{Fortunato2010,NewmanGirvan2004}
\cite{Fortunato2010,Vehlowetal2013}
\section{Floats}
{Figures} may be included using the command,\linebreak
\verb+\includegraphics+ in
combination with or without its several options to further control
graphic. \verb+\includegraphics+ is provided by {graphic[s,x].sty}
which is part of any standard \LaTeX{} distribution.
{graphicx.sty} is loaded by default. \LaTeX{} accepts figures in
the postscript format while pdf\LaTeX{} accepts {*.pdf},
{*.mps} (metapost), {*.jpg} and {*.png} formats.
pdf\LaTeX{} does not accept graphic files in the postscript format.
\begin{figure}
\centering
\includegraphics[scale=.75]{figs/Fig1.pdf}
\caption{The evanescent light - $1S$ quadrupole coupling
($g_{1,l}$) scaled to the bulk exciton-photon coupling
($g_{1,2}$). The size parameter $kr_{0}$ is denoted as $x$ and
the \PMS is placed directly on the cuprous oxide sample ($\delta
r=0$, See also Table \protect\ref{tbl1}).}
\label{FIG:1}
\end{figure}
The \verb+table+ environment is handy for marking up tabular
material. If users want to use {multirow.sty},
{array.sty}, etc., to fine control/enhance the tables, they
are welcome to load any package of their choice and
{cas-dc.cls} will work in combination with all loaded
packages.
\begin{table}[width=.9\linewidth,cols=4,pos=h]
\caption{This is a test caption. This is a test caption. This is a test
caption. This is a test caption.}\label{tbl1}
\begin{tabular*}{\tblwidth}{@{} LLLL@{} }
\toprule
Col 1 & Col 2 & Col 3 & Col4\\
\midrule
12345 & 12345 & 123 & 12345 \\
12345 & 12345 & 123 & 12345 \\
12345 & 12345 & 123 & 12345 \\
12345 & 12345 & 123 & 12345 \\
12345 & 12345 & 123 & 12345 \\
\bottomrule
\end{tabular*}
\end{table}
\section[Theorem and ...]{Theorem and theorem like environments}
{cas-dc.cls} provides a few shortcuts to format theorems and
theorem-like environments with ease. In all commands the options that
are used with the \verb+\newtheorem+ command will work exactly in the same
manner. {cas-dc.cls} provides three commands to format theorem or
theorem-like environments:
\begin{verbatim}
\newtheorem{theorem}{Theorem}
\newtheorem{lemma}[theorem]{Lemma}
\newdefinition{rmk}{Remark}
\newproof{pf}{Proof}
\newproof{pot}{Proof of Theorem \ref{thm2}}
\end{verbatim}
The \verb+\newtheorem+ command formats a
theorem in \LaTeX's default style with italicized font, bold font
for theorem heading and theorem number at the right hand side of the
theorem heading. It also optionally accepts an argument which
will be printed as an extra heading in parentheses.
\begin{verbatim}
\begin{theorem}
For system (8), consensus can be achieved with
$\|T_{\omega z}$ ...
\begin{eqnarray}\label{10}
....
\end{eqnarray}
\end{theorem}
\end{verbatim}
\newtheorem{theorem}{Theorem}
\begin{theorem}
For system (8), consensus can be achieved with
$\|T_{\omega z}$ ...
\begin{eqnarray}\label{10}
....
\end{eqnarray}
\end{theorem}
The \verb+\newdefinition+ command is the same in
all respects as its \verb+\newtheorem+ counterpart except that
the font shape is roman instead of italic. Both
\verb+\newdefinition+ and \verb+\newtheorem+ commands
automatically define counters for the environments defined.
The \verb+\newproof+ command defines proof environments with
upright font shape. No counters are defined.
\section[Enumerated ...]{Enumerated and Itemized Lists}
{cas-dc.cls} provides an extended list processing macros
which makes the usage a bit more user friendly than the default
\LaTeX{} list macros. With an optional argument to the
\verb+\begin{enumerate}+ command, you can change the list counter
type and its attributes.
\begin{verbatim}
\begin{enumerate}[1.]
\item The enumerate environment starts with an optional
argument `1.', so that the item counter will be suffixed
by a period.
\item You can use `a)' for alphabetical counter and '(i)'
for roman counter.
\begin{enumerate}[a)]
\item Another level of list with alphabetical counter.
\item One more item before we start another.
\item One more item before we start another.
\item One more item before we start another.
\item One more item before we start another.
\end{verbatim}
Further, the enhanced list environment allows one to prefix a
string like `step' to all the item numbers.
\begin{verbatim}
\begin{enumerate}[Step 1.]
\item This is the first step of the example list.
\item Obviously this is the second step.
\item The final step to wind up this example.
\end{enumerate}
\end{verbatim}
\section{Cross-references}
In electronic publications, articles may be internally
hyperlinked. Hyperlinks are generated from proper
cross-references in the article. For example, the words
\textcolor{black!80}{Fig.~1} will never be more than simple text,
whereas the proper cross-reference \verb+\ref{tiger}+ may be
turned into a hyperlink to the figure itself:
\textcolor{blue}{Fig.~1}. In the same way,
the words \textcolor{blue}{Ref.~[1]} will fail to turn into a
hyperlink; the proper cross-reference is \verb+\cite{Knuth96}+.
Cross-referencing is possible in \LaTeX{} for sections,
subsections, formulae, figures, tables, and literature
references.
\section{Bibliography}
Two bibliographic style files (\verb+*.bst+) are provided ---
{model1-num-names.bst} and {model2-names.bst} --- the first one can be
used for the numbered scheme. This can also be used for the numbered
with new options of {natbib.sty}. The second one is for the author year
scheme. When you use model2-names.bst, the citation commands will be
like \verb+\citep+, \verb+\citet+, \verb+\citealt+ etc. However when
you use model1-num-names.bst, you may use only \verb+\cite+ command.
\verb+thebibliography+ environment. Each reference is a\linebreak
\verb+\bibitem+ and each \verb+\bibitem+ is identified by a label,
by which it can be cited in the text:
\noindent In connection with cross-referencing and
possible future hyperlinking it is not a good idea to collect
more that one literature item in one \verb+\bibitem+. The
so-called Harvard or author-year style of referencing is enabled
by the \LaTeX{} package {natbib}. With this package the
literature can be cited as follows:
\begin{enumerate}[\textbullet]
\item Parenthetical: \verb+\citep{WB96}+ produces (Wettig \& Brown, 1996).
\item Textual: \verb+\citet{ESG96}+ produces Elson et al. (1996).
\item An affix and part of a reference:\break
\verb+\citep[e.g.][Ch. 2]{Gea97}+ produces (e.g. Governato et
al., 1997, Ch. 2).
\end{enumerate}
In the numbered scheme of citation, \verb+\cite{<label>}+ is used,
since \verb+\citep+ or \verb+\citet+ has no relevance in the numbered
scheme. {natbib} package is loaded by {cas-dc} with
\verb+numbers+ as default option. You can change this to author-year
or harvard scheme by adding option \verb+authoryear+ in the class
loading command. If you want to use more options of the {natbib}
package, you can do so with the \verb+\biboptions+ command. For
details of various options of the {natbib} package, please take a
look at the {natbib} documentation, which is part of any standard
\LaTeX{} installation.
\section{Introduction}
Two classfiles namely \file{cas-sc.cls} and \file{cas-dc.cls} were
written for typesetting articles submitted in journals of Elsevier's
Complex Article Service (CAS) workflow.
\subsection{Usage}
\begin{enumerate}
\item \file{cas-sc.cls} for single column journals.
\begin{vquote}
\documentclass[<options>]{cas-sc}
\end{vquote}
\item \file{cas-dc.cls} for single column journals.
\begin{vquote}
\documentclass[<options>]{cas-dc}
\end{vquote}
\end{enumerate}
and have an option longmktitle to handle long front matter.
\section{Front matter}
\begin{vquote}
\title [mode = title]{This is a specimen $a_b$ title}
\tnotemark[1,2]
\tnotetext[1]{This document is the results of the research
project funded by the National Science Foundation.}
\tnotetext[2]{The second title footnote which is a longer text
matter to fill through the whole text width and overflow into
another line in the footnotes area of the first page.}
\author[1,3]{CV Radhakrishnan}[type=editor,
auid=000,bioid=1,
prefix=Sir,
role=Researcher,
orcid=0000-0001-7511-2910]
\cormark[1]
\fnmark[1]
\ead{<EMAIL>}
\ead[url]{www.cvr.cc, <EMAIL>}
\end{vquote}
\begin{vquote}
\credit{Conceptualization of this study, Methodology,
Software}
\address[1]{Elsevier B.V., Radarweg 29, 1043 NX Amsterdam,
The Netherlands}
\author[2,4]{Han Theh Thanh}[style=chinese]
\author[2,3]{CV Rajagopal}[%
role=Co-ordinator,
suffix=Jr,
]
\fnmark[2]
\ead{<EMAIL>}
\ead[URL]{www.sayahna.org}
\credit{Data curation, Writing - Original draft preparation}
\address[2]{Sayahna Foundation, Jagathy, Trivandrum 695014,
India}
\author[1,3]{Rishi T.}
\cormark[2]
\fnmark[1,3]
\ead{<EMAIL>}
\ead[URL]{www.stmdocs.in}
\address[3]{STM Document Engineering Pvt Ltd., Mepukada,
Malayinkil, Trivandrum 695571, India}
\cortext[cor1]{Corresponding author}
\cortext[cor2]{Principal corresponding author}
\fntext[fn1]{This is the first author footnote. but is common
to third author as well.}
\fntext[fn2]{Another author footnote, this is a very long
footnote and it should be a really long footnote. But this
footnote is not yet sufficiently long enough to make two lines
of footnote text.}
\end{vquote}
\begin{vquote}
\nonumnote{This note has no numbers. In this work we
demonstrate $a_b$ the formation Y\_1 of a new type of
polariton on the interface between a cuprous oxide slab
and a polystyrene micro-sphere placed on the slab.
}
\begin{abstract}[S U M M A R Y]
This template helps you to create a properly formatted
\LaTeX\ manuscript.
\noindent\texttt{\textbackslash begin{abstract}} \dots
\texttt{\textbackslash end{abstract}} and
\verb+\begin{keyword}+ \verb+...+ \verb+\end{keyword}+
which contain the abstract and keywords respectively.
Each keyword shall be separated by a \verb+\sep+ command.
\end{abstract}
\begin{keywords}
quadrupole exciton \sep polariton \sep \WGM \sep \BEC
\end{keywords}
\maketitle
\end{vquote}
\begin{figure}
\includegraphics[width=\textwidth]{sc-sample.pdf}
\caption{Single column output (classfile: cas-sc.cls).}
\end{figure}
\begin{figure}
\includegraphics[width=\textwidth]{dc-sample.pdf}
\caption{Double column output (classfile: cas-dc.cls).}
\end{figure}
\subsection{Title}
\verb+\title+ command have the below options:
\begin{enumerate}
\item \verb+title:+ Document title
\item \verb+alt:+ Alternate title
\item \verb+sub:+ Sub title
\item \verb+trans:+ Translated title
\item \verb+transsub:+ Translated sub title
\end{enumerate}
\begin{vquote}
\title[mode=title]{This is a title}
\title[mode=alt]{This is a alternate title}
\title[mode=sub]{This is a sub title}
\title[mode=trans]{This is a translated title}
\title[mode=transsub]{This is a translated sub title}
\end{vquote}
\subsection{Author}
\verb+\author+ command have the below options:
\begin{enumerate}
\item \verb+auid:+ Author id
\item \verb+bioid:+ Biography id
\item \verb+alt:+ Alternate author
\item \verb+style:+ Style of author name chinese
\item \verb+prefix:+ Prefix Sir
\item \verb+suffix:+ Suffix
\item \verb+degree:+ Degree
\item \verb+role:+ Role
\item \verb+orcid:+ ORCID
\item \verb+collab:+ Collaboration
\item \verb+anon:+ Anonymous author
\item \verb+deceased:+ Deceased author
\item \verb+twitter:+ Twitter account
\item \verb+facebook:+ Facebook account
\item \verb+linkedin:+ LinkedIn account
\item \verb+plus:+ Google plus account
\item \verb+gplus:+ Google plus account
\end{enumerate}
\begin{vquote}
\author[1,3]{Author Name}[type=editor,
auid=000,bioid=1,
prefix=Sir,
role=Researcher,
orcid=0000-0001-7511-2910,
facebook=<facebook id>,
twitter=<twitter id>,
linkedin=<linkedin id>,
gplus=<gplus id>]
\end{vquote}
\subsection{Various Marks in the Front Matter}
The front matter becomes complicated due to various kinds
of notes and marks to the title and author names. Marks in
the title will be denoted by a star ($\star$) mark;
footnotes are denoted by super scripted Arabic numerals,
corresponding author by of an Conformal asterisk (*) mark.
\subsubsection{Title marks}
Title mark can be entered by the command, \verb+\tnotemark[<num>]+
and the corresponding text can be entered with the command
\verb+\tnotetext[<num>]+ \verb+{<text>}+. An example will be:
\begin{vquote}
\title[mode=title]{Leveraging social media news to predict
stock index movement using RNN-boost}
\tnotemark[1,2]
\tnotetext[1]{This document is the results of the research
project funded by the National Science Foundation.}
\tnotetext[2]{The second title footnote which is a longer
text matter to fill through the whole text width and
overflow into another line in the footnotes area of
the first page.}
\end{vquote}
\verb+\tnotetext+ and \verb+\tnotemark+ can be anywhere in
the front matter, but shall be before \verb+\maketitle+ command.
\subsubsection{Author marks}
Author names can have many kinds of marks and notes:
\begin{vquote}
footnote mark : \fnmark[<num>]
footnote text : \fntext[<num>]{<text>}
affiliation mark : \author[<num>]
email : \ead{<emailid>}
url : \ead[url]{<url>}
corresponding author mark : \cormark[<num>]
corresponding author text : \cortext[<num>]{<text>}
\end{vquote}
\subsubsection{Other marks}
At times, authors want footnotes which leave no marks in
the author names. The note text shall be listed as part of
the front matter notes. Class files provides
\verb+\nonumnote+ for this purpose. The usage
\begin{vquote}
\nonumnote{<text>}
\end{vquote}
\noindent and should be entered anywhere before the \verb+\maketitle+
command for this to take effect.
\subsection{Abstract and Keywords}
Abstract shall be entered in an environment that starts
with \verb+\begin{abstract}+ and ends with
\verb+\end{abstract}+. Longer abstracts spanning more than
one page is also possible in Class file even in double
column mode. We need to invoke longmktitle option in the
class loading line for this to happen smoothly.
The key words are enclosed in a \verb+{keyword}+
environment.
\begin{vquote}
\begin{abstract}
This is a abstract. \lipsum[3]
\end{abstract}
\begin{keywords}
First keyword \sep Second keyword \sep Third
keyword \sep Fourth keyword
\end{keywords}
\end{vquote}
\section{Main Matter}
\subsection{Tables}
\subsubsection{Normal tables}
\begin{vquote}
\begin{table}
\caption{This is a test caption.}
\begin{tabular*}{\tblwidth}{@{} LLLL@{} }
\toprule
Col 1 & Col 2\\
\midrule
12345 & 12345\\
12345 & 12345\\
12345 & 12345\\
\bottomrule
\end{tabular*}
\end{table}
\end{vquote}
\subsubsection{Span tables}
\begin{vquote}
\begin{table*}[width=.9\textwidth,cols=4,pos=h]
\caption{This is a test caption.}
\begin{tabular*}{\tblwidth}{@{} LLLLLL@{} }
\toprule
Col 1 & Col 2 & Col 3 & Col4 & Col5 & Col6 & Col7\\
\midrule
12345 & 12345 & 123 & 12345 & 123 & 12345 & 123 \\
12345 & 12345 & 123 & 12345 & 123 & 12345 & 123 \\
12345 & 12345 & 123 & 12345 & 123 & 12345 & 123 \\
\bottomrule
\end{tabular*}
\end{table*}
\end{vquote}
\subsection{Figures}
\subsubsection{Normal figures}
\begin{vquote}
\begin{figure}
\centering
\includegraphics[scale=.75]{Fig1.pdf}
\caption{The evanescent light - $1S$ quadrupole coupling
($g_{1,l}$) scaled to the bulk exciton-photon coupling
($g_{1,2}$). The size parameter $kr_{0}$ is denoted as $x$ and
the \PMS is placed directly on the cuprous oxide sample ($\delta
r=0$, See also Fig. \protect\ref{FIG:2}).}
\label{FIG:1}
\end{figure}
\end{vquote}
\subsubsection{Span figures}
\begin{vquote}
\begin{figure*}
\centering
\includegraphics[width=\textwidth,height=2in]{Fig2.pdf}
\caption{Schematic of formation of the evanescent polariton on
linear chain of \PMS. The actual dispersion is determined by
the ratio of two coupling parameters such as exciton-\WGM
coupling and \WGM-\WGM coupling between the microspheres.}
\label{FIG:2}
\end{figure*}\end{vquote}
\subsection{Theorem and theorem like environments}
CAS class file provides a few hooks to format theorems and
theorem like environments with ease. All commands the
options that are used with \verb+\newtheorem+ command will work
exactly in the same manner. Class file provides three
commands to format theorem or theorem like environments:
\begin{enumerate}
\item \verb+\newtheorem+ command formats a theorem in
\LaTeX's default style with italicized font for theorem
statement, bold weight for theorem heading and theorem
number typeset at the right of theorem heading. It also
optionally accepts an argument which will be printed as an
extra heading in parentheses. Here is an example coding and
output:
\begin{vquote}
\newtheorem{theorem}{Theorem}
\begin{theorem}\label{thm}
The \WGM evanescent field penetration depth into the
cuprous oxide adjacent crystal is much larger than the
\QE radius:
\begin{equation*}
\lambda_{1S}/2 \pi \left({\epsilon_{Cu2O}-1}
\right)^{1/2} = 414 \mbox{ \AA} \gg a_B = 4.6
\mbox{ \AA}
\end{equation*}
\end{theorem}
\end{vquote}
\item \verb+\newdefinition+ command does exactly the same
thing as with except that the body font is up-shape instead
of italic. See the example below:
\begin{vquote}
\newdefinition{definition}{Definition}
\begin{definition}
The bulk and evanescent polaritons in cuprous oxide
are formed through the quadrupole part of the light-matter
interaction:
\begin{equation*}
H_{int} = \frac{i e }{m \omega_{1S}} {\bf E}_{i,s}
\cdot {\bf p}
\end{equation*}
\end{definition}
\end{vquote}
\item \verb+\newproof+ command helps to define proof and
custom proof environments without counters as provided in
the example code. Given below is an example of proof of
theorem kind.
\begin{vquote}
\newproof{pot}{Proof of Theorem \ref{thm}}
\begin{pot}
The photon part of the polariton trapped inside the \PMS
moves as it would move in a micro-cavity of the effective
modal volume $V \ll 4 \pi r_{0}^{3} /3$. Consequently, it
can escape through the evanescent field. This evanescent
field essentially has a quantum origin and is due to
tunneling through the potential caused by dielectric
mismatch on the \PMS surface. Therefore, we define the
\emph{evanescent} polariton (\EP) as an evanescent light -
\QE coherent superposition.
\end{pot}
\end{vquote}
\end{enumerate}
\subsection{Enumerated and Itemized Lists}
CAS class files provides an extended list processing macros
which makes the usage a bit more user friendly than the
default LaTeX list macros. With an optional argument to the
\verb+\begin{enumerate}+ command, you can change the list
counter type and its attributes. You can see the coding and
typeset copy.
\begin{vquote}
\begin{enumerate}[1.]
\item The enumerate environment starts with an optional
argument `1.' so that the item counter will be suffixed
by a period as in the optional argument.
\item If you provide a closing parenthesis to the number in the
optional argument, the output will have closing
parenthesis for all the item counters.
\item You can use `(a)' for alphabetical counter and `(i)' for
roman counter.
\begin{enumerate}[a)]
\item Another level of list with alphabetical counter.
\item One more item before we start another.
\begin{enumerate}[(i)]
\item This item has roman numeral counter.
\end{vquote}
\begin{vquote}
\item Another one before we close the third level.
\end{enumerate}
\item Third item in second level.
\end{enumerate}
\item All list items conclude with this step.
\end{enumerate}
\section{Biography}
\verb+\bio+ command have the below options:
\begin{enumerate}
\item \verb+width:+ Width of the author photo (default is 1in).
\item \verb+pos:+ Position of author photo.
\end{enumerate}
\begin{vquote}
\bio[width=10mm,pos=l]{tuglogo.jpg}
\textbf{Another Biography:}
Recent experimental \cite{HARA:2005} and theoretical
\cite{DEYCH:2006} studies have shown that the \WGM can travel
along the chain as "heavy photons". Therefore the \WGM
acquires the spatial dispersion, and the evanescent
quadrupole polariton has the form (See Fig.\ref{FIG:3}):
\endbio
\end{vquote}
\section[CRediT...]{CRediT authorship contribution statement}
Give the authorship contribution after each author as
\begin{vquote}
\credit{Conceptualization of this study, Methodology,
Software}
\end{vquote}
To print the details use \verb+\printcredits+
\begin{vquote}
\author[1,3]{V. {{\=A}}nand Rawat}[auid=000,
bioid=1,
prefix=Sir,
role=Researcher,
orcid=0000-0001-7511-2910]
\end{vquote}
\begin{vquote}
\cormark[1]
\fnmark[1]
\ead{<EMAIL>}
\ead[url]{www.cvr.cc, www.tug.org.in}
\credit{Conceptualization of this study, Methodology,
Software}
\address[1]{Indian \TeX{} Users Group, Trivandrum 695014,
India}
\author[2,4]{Han Theh Thanh}[style=chinese]
\author[2,3]{T. Rishi Nair}[role=Co-ordinator,
suffix=Jr]
\fnmark[2]
\ead{<EMAIL>}
\ead[URL]{www.sayahna.org}
\credit{Data curation, Writing - Original draft preparation}
. . .
. . .
. . .
\printcredits
\end{vquote}
\section{Bibliography}
For CAS categories, two reference models are recommended.
They are \file{model1-num-names.bst} and \file{model2-names.bst}.
Former will format the reference list and their citations according to
numbered scheme whereas the latter will format according name-date or
author-year style. Authors are requested to choose any one of these
according to the journal style. You may download these from
The above bsts are available in the following location for you to
download:
\url{https://support.stmdocs.in/wiki/index.php?title=Model-wise_bibliographic_style_files}
\hfill $\Box$
\end{document}
|
\section{Introduction}
Global-local counting conjectures play a major role in modern representation theory of finite groups. Amongst them are the McKay Conjecture \cite{McK72} and its blockwise version, known as the Alperin-McKay Conjecture \cite{Alp76}, the Alperin Weight Conjecture \cite{Alp87} and a series of conjectures proposed by Dade in \cite{Dad92}, \cite{Dad94} and \cite{Dad97} that imply all the above mentioned conjectures. Dade's aim was to find a version of his conjecture strong enough to hold for every finite group if proved for all nonabelian simple groups. Unfortunately, such a reduction theorem has never been published. The first step towards the solution of the global-local conjectures has been achieved by Isaacs, Malle and Navarro in \cite{Isa-Mal-Nav07} where the McKay Conjecture was reduced to a stronger statement for simple groups. Inspired by this result, other reduction theorems have been proved (see \cite{Nav-Tie11}, \cite{Spa13II}, \cite{Spa13I}, \cite{Spa17} and \cite{Nav-Spa-Val}). However, contrary to Dade's philosophy, all the reduction theorems appeared so far reduce a certain statement for arbitrary finite groups to a much stronger statement for quasisimple groups.
Although these stronger statements, known as inductive conditions, have been originally thought for (quasi)simple groups, they can be stated for arbitrary finite groups. Then, going back to Dade's plan, by proving the inductive condition for simple groups it should be possible to obtain, not only the original conjecture, but even the inductive condition itself for every finite group. This was done in \cite{Nav-Spa14I} for the Alperin--McKay Conjecture.
In \cite{Spa17} Sp\"ath introduced the Character Triple Conjecture and showed that Dade's Projective Conjecture holds for every finite group if her conjecture holds for all quasisimple groups. Therefore Sp\"ath's conjecture plays the role of inductive condition for Dade's Projective Conjecture. Following \cite{Nav-Spa14I}, we would like to show that the Character Triple Conjecture holds for every finite group if it holds for all quasisimple groups. To prove such a reduction theorem, it is necessary to study the structure of a minimal counterexample. As for the above mentioned reductions, the first step in this direction is to show that such a counterexample cannot be $p$-solvable. This result was in preparation when the paper \cite{Spa17} was published. Unfortunately, the project was later abandoned and no proof has appeared. With permission of Sp\"ath, we take over the project and prove the Character Triple Conjecture for $p$-solvable groups in this paper.
More precisely, let $G$ be a finite group and fix a prime $p$. For any $d\geq 0$, we denote by ${\rm Irr}^d(G)$ the set of irreducible characters $\chi\in{\rm Irr}(G)$ with $p$-defect equal to $d$. Let $\mathfrak{N}(G)$ be the set of $p$-chains of $G$ starting with ${\mathbf{O}}_p(G)$. If $\d=\{D_0<D_1<\dots<D_n\}\in\mathfrak{N}(G)$, we denote by $|\d|$ the integer $n$, called the length of $\d$. This yields a partition of $\mathfrak{N}(G)$ into the set $\mathfrak{N}(G)_+$ of $p$-chains of even length and the set $\mathfrak{N}(G)_-$ of $p$-chains of odd length. Notice that $G$ acts by conjugation on the set of $p$-chains and let $G_\d$ be the stabilizer in $G$ of the chain $\d$. For $\epsilon\in\{+,-\}$ and $B$ a $p$-block of $G$, define
$${\mathcal{C}}^d(B)_\epsilon:=\left\lbrace(\d,\vartheta)\enspace\middle|\enspace \d\in\mathfrak{N}(G)_\epsilon,\vartheta\in{\rm Irr}^d(G_\d),{\rm bl}(\vartheta)^G=B\right\rbrace,$$
where ${\rm bl}(\vartheta)$ is the unique block of $G_\d$ containing $\vartheta$ and ${\rm bl}(\vartheta)^G$ is the block of $G$ obtain via Brauer's induction (this is defined by \cite[Lemma 3.2]{Kno-Rob89}). We denote by $\overline{(\d,\vartheta)}$ the $G$-orbit of $(\d,\vartheta)\in{\mathcal{C}}^d(B)_\epsilon$ and by ${\mathcal{C}}^d(B)_\epsilon/G$ the set of $G$-orbits. Now, our main result can be stated as follows.
\begin{theoA}
\label{thm:Main theorem}
Let $G$ be a finite $p$-solvable group with ${\mathbf{O}}_p(G)\leq {\mathbf{Z}}(G)$ and consider a $p$-block $B$ of $G$ with noncentral defect groups. Suppose that $G\unlhd A$ and denote by $A_B$ the stabilizer of $B$ in $A$. Then, for every $d\geq 0$, there exists an $A_B$-equivariant bijection
$$\Omega:{\mathcal{C}}^d(B)_+/G \to {\mathcal{C}}^d(B)_-/G$$
such that
$$\left(A_{\d,\vartheta},G_\d,\vartheta\right)\iso{G}\left(A_{\e,\chi},G_{\e},\chi\right),$$
in the sense of Definition \ref{def:N-block isomorphic character triples}, for every $(\d,\vartheta)\in{\mathcal{C}}^d(B)_+$ and $(\e,\chi)\in\Omega(\overline{(\d,\vartheta)})$.
\end{theoA}
Recall that for $\chi\in{\rm Irr}(G)$, the $p$-residue of $\chi$ is the nonnegative integer $r(\chi):=|G|_{p'}/\chi(1)_{p'}$. Following ideas of Isaacs and Navarro \cite{Isa-Nav02}, we include the $p$-residue of characters into the picture. However, we do not consider more technical refinements involving Galois automorphisms (see \cite{Nav04}) or $p$-local Schur indices (see \cite{Tur08II}) as done by Turull in \cite{Tur17} for Dade's Conjecture.
\begin{theoA}
\label{thm:Main theorem residues}
There exists a bijection $\Omega$ satisfying the conditions of Theorem \ref{thm:Main theorem} and such that
$$r(\vartheta)\equiv \pm r(\chi)\pmod{p},$$
for every $(\d,\vartheta)\in{\mathcal{C}}^d(B)_+$ and some $(\e,\chi)\in\Omega(\overline{(\d,\vartheta)})$.
\end{theoA}
As a corollary to our results, we show that Dade's Extended Projective Conjecture \cite[4.10]{Dad97}, with the Isaacs-Navarro refinement, holds for every $p$-solvable group.
\begin{corA}
\label{cor:Dade's Extended Projective Conjecture}
Dade's Extended Projective Conjecture with the Isaacs-Navarro refinement holds for every $p$-solvable group.
\end{corA}
\begin{proof}
This follows from Theorem \ref{thm:Main theorem residues} and \cite[Proposition 6.4]{Spa17}.
\end{proof}
The paper is structured as follows: in the next section, we establish some notation, we give the main definitions used in the paper and we prove some preliminary results. In the third section, we construct an equivariant defect preserving character bijection lying above the Glauberman correspondence that is well behaved with respect to $N$-block isomorphic character triples (see Definition \ref{def:N-block isomorphic character triples}). In the fourth section, we show how to construct $N$-block isomorphic character triples by using the Fong correspondence. Finally in the last section, we use the previously obtained results to prove Theorem \ref{thm:Main theorem} and Theorem \ref{thm:Main theorem residues}. This is done by inspecting the structure of a minimal counterexample.
\section{Preliminaries and notation}
We use standard notation from representation theory of finite groups as in \cite{Isa76}, \cite{Nag-Tsu89} and \cite{Nav98}. All groups considered in the sequel are assumed to be finite. For notational convenience, whenever necessary we denote the normalizer ${\mathbf{N}}_G(H)$ simply by $G_H$, for every $H\leq G$.
Let ${\rm Irr}(G)$ the set of ordinary irreducible characters. If $H\leq G$ and $\chi\in{\rm Irr}(G)$, then ${\rm Irr}(\chi_H)$ is the set of irreducible characters of $H$ which appear as constituents of the restricted character $\chi_H$. Moreover, for $\vartheta\in{\rm Irr}(H)$, we denote by ${\rm Irr}(G\mid \vartheta)$ the set of irreducible constituents of the induced character $\vartheta^G$. For $N\unlhd G$ and $\vartheta\in{\rm Irr}(N)$ we denote by $G_\vartheta$ the stabilizer of $\vartheta$ in $G$ and by ${\rm Irr}_G(N)$ the set of $G$-invariant irreducible characters of $N$. Then $(G,N,\vartheta)$ is a character triple if $N\unlhd G$ and $\vartheta\in{\rm Irr}_G(N)$.
Fix a prime $p$. For $\chi\in{\rm Irr}(G)$, there exist unique nonnegative integers $d(\chi)$ and $r(\chi)$, called respectively the $p$-defect and the $p$-residue of $\chi$, such that $r(\chi)p^{d(\chi)}=|G|/\chi(1)$ with $r(\chi)$ coprime to $p$. For any $d\geq 0$, we denote by ${\rm Irr}^d(G)$ the set of irreducible characters $\chi\in{\rm Irr}(G)$ such that $d(\chi)=d$. We denote by ${\rm Bl}(G)$ the set of $p$-blocks of $G$ and by ${\rm Bl}(G\mid b)$ the set of all blocks of $G$ covering $b$, where $N\unlhd G$ and $b\in{\rm Bl}(N)$. Let $G_b$ be the stabilizer of $b$ in $G$. If $H\leq G$ and $b\in{\rm Bl}(H)$, then $b^G$ is, when it is defined, the block obtained via Brauer induction. For $\chi\in{\rm Irr}(G)$, the block of $G$ that contains $\chi$ is ${\rm bl}(\chi)$. Let $\delta(B)$ be the set of defect groups of the block $B$ and $d(B)$ be its defect. If $D$ is a $p$-subgroup of $G$, then ${\rm Bl}(G\mid D)$ is the set of blocks of $G$ with $D\in\delta(B)$.
For the notion of projective representation, we refer to \cite{Nag-Tsu89}, \cite{Nav18} and \cite{Spa18}. We denote by ${\rm Proj}(G\mid \alpha)$ the set of projective representations of $G$ with factor set $\alpha$. Moreover, if $N\unlhd G$, then we can consider a representation of $G/N$ as a representation of $G$ that is constant on $N$-cosets by the usual inflation process. If $\mathcal{P}$ is a projective representation associated with a character triple $(G,N,\vartheta)$ (see \cite[Definition 1.7]{Spa18} and \cite[Definition 5.2]{Nav18}), then $\mathcal{P}$ yields a central extension of $G$ (see \cite[Theorem 5.6]{Nav18} and \cite[Theorem 1.12]{Spa18}). This is a standard construction and we will make use of it without further comment.
In \cite[Definition 6.3]{Spa17} a new equivalence relation on character triples was introduced. This will be of fundamental importance in what follows and, for completeness, we include the definition.
\begin{defin}[$N$-block isomorphic character triples]
\label{def:N-block isomorphic character triples}
Let $(H_1,M_1,\vartheta_1)$ and $(H_2,M_2,\vartheta_2)$ be two character triples and let $N$ be a group. We say that the two character triples are $N$-block isomorphic, and write
$$\left(H_1,M_1,\vartheta_1\right)\iso{N}\left(H_2,M_2,\vartheta_2\right),$$
if the following conditions are satisfied:
\begin{enumerate}
\item $N\unlhd NH_1=NH_2=:G$, $M_1=H_1\cap N$ and $M_2=H_2\cap N$. We denote the canonical isomorphisms by $l_i:H_i/M_i\to G/N$ and by $i:=l_2^{-1}\circ l_1:H_1/M_1\to H_2/M_2$;
\item For $i=1,2$, there exists a defect group $D_i\in\delta({\rm bl}(\vartheta_i))$ such that ${\mathbf{C}}_G(D_i)\leq H_i$. In particular ${\mathbf{C}}_G(N)\leq H_1\cap H_2$;
\item For $i=1,2$, there exists a projective representation $\mathcal{P}_i\in{\rm Proj}(H_i\mid \alpha_i)$ associated with $(H_i,M_i,\vartheta_i)$ such that $\alpha_1(x,y)=\alpha_2(i(x),i(y))$, for every $x,y\in H_1/M_1$, and with the property that $\mathcal{P}_{1,{\mathbf{C}}_G(N)}$ and $\mathcal{P}_{2,{\mathbf{C}}_G(N)}$ are associated with the same scalar function (see the comments preceding \cite[Definition 2.7]{Spa18} and \cite[Definition 10.14]{Nav18}). In this case, there exists a strong isomorphism of character triple $(i,\sigma):(H_1,M_1,\vartheta_1)\to (H_2,M_2,\vartheta_2)$ (see \cite[Problem 11.13]{Isa76}) given by
\begin{align*}
\sigma_{J_1}:{\rm Irr}(J_1\mid \vartheta_1)&\to {\rm Irr}(J_2\mid \vartheta_2)
\\
{\rm tr}(\mathcal{Q}_{J_1}\otimes \mathcal{P}_{1,J_1})&\mapsto {\rm tr}(\mathcal{Q}_{J_2}\otimes \mathcal{P}_{2,J_2}),
\end{align*}
for every $N\leq J\leq G$, where $J_i:=J\cap H_i$ and $\mathcal{Q}\in{\rm Proj}(J/N\mid \alpha^{-1}_{J\times J})$. Here $\alpha$ is the factor set of $G/N$ corresponding to $\alpha_i$ via the isomorphism $l_i:H_i/M_i\to G/N$ (see \cite[Theorem 3.3]{Spa17});
\item For $N\leq J\leq G$, we have
$${\rm bl}(\psi)^J={\rm bl}(\sigma_{J_1}(\psi))^J$$
for every $\psi\in{\rm Irr}(J_1\mid \vartheta_1)$. Observe in this situation that block induction is well defined (see \cite[Lemma 3.5]{Spa17}).
\end{enumerate}
In this situation, we say that $(\mathcal{P}_1,\mathcal{P}_2)$ is associated with $(H_1,M_1,\vartheta_1)\iso{N}(H_2,M_2,\vartheta_2)$, or that $(H_1,M_1,\vartheta_1)\iso{N}(H_2,M_2,\vartheta_2)$ is given by $(\mathcal{P}_1,\mathcal{P}_2)$. Whenever we want to specify the pair $(\mathcal{P}_1,\mathcal{P}_2)$ we write $\sigma^{(\mathcal{P}_1,\mathcal{P}_2)}_{J_1}$ instead of simply $\sigma_{J_1}$.
\end{defin}
The above definition gives a relation on character triples that extends the relation $\geq_b$ introduced in \cite[Definition 4.2]{Spa18} (see also \cite{Nav-Spa14I}). In fact, observe that $(G,N,\chi)\geq_b (H,M,\vartheta)$ if and only if $(G,N,\chi)\iso{N}(H,M,\vartheta)$. On the other hand, the notion of block isomorphism of character triples given in \cite[Definition 3.6]{Nav-Spa14I} is slightly different, and in some sense more restrictive, from the relation $\geq_b$ (see \cite[Remark 4.3 (c)]{Spa18}). For this reason, we will not refer to results of \cite{Nav-Spa14I} which involve directly block isomorphism of character triples.
Having defined $N$-block isomorphic character triples, we can now introduce Sp\"ath's Character Triple Conjecture (\cite[Conjecture 6.3]{Spa17}). For a central $p$-subgroup $Z$ of $G$, we denote by $\mathfrak{N}(G,Z)$ the set of normal $p$-chains of $G$ starting with $Z$. These are the chains $\d=\{Z=D_0<D_1<\dots<D_n\}$ of $p$-subgroups of $G$ with the property that each $D_i$ is normal in the largest subgroup $D_n$. We denote by $|\d|$ the integer $n$, called the length of $\d$. The set $\mathfrak{N}(G,Z)$ is partitioned into the set $\mathfrak{N}(G,Z)_+$ of $p$-chains of even length, and the set $\mathfrak{N}(G,Z)_-$ of $p$-chains of odd length. The group $G$ acts by conjugation on $\mathfrak{N}(G,Z)$ and we denote by $G_\d=\bigcap_i{\mathbf{N}}_G(D_i)$ the stabilizer in $G$ of the chain $\d$. Finally, let $B$ be a block of $G$ and, for $\epsilon\in\{+,-\}$ and $d\geq 0$, define ${\mathcal{C}}^d(B,Z)_\epsilon$ to be the set of pairs $(\d,\vartheta)$ with $\d\in\mathfrak{N}(G,Z)_\epsilon$ and $\vartheta\in{\rm Irr}^d(G_\d)$ satisfying ${\rm bl}(\vartheta)^G=B$. Again, the group $G$ acts on ${\mathcal{C}}^d(B,Z)_\epsilon$. We denote by $\overline{(\d,\vartheta)}$ the $G$-orbit of $(\d,\vartheta)\in{\mathcal{C}}^d(B,Z)_\epsilon$ and by ${\mathcal{C}}^d(B,Z)_\epsilon/G$ the set of $G$-orbits.
\begin{conj}[Sp\"ath's Character Triple Conjecture]
\label{conj:Character Triple Conjecture}
Let $G$ be a finite group, $Z\leq {\mathbf{Z}}(G)$ be a $p$-subgroup and consider $B\in{\rm Bl}(G)$ with defect groups strictly larger than $Z$. Suppose that $G\unlhd A$. Then, for every $d\geq 0$, there exists an ${\mathbf{N}}_A(Z)_B$-equivariant bijection
$$\Omega:{\mathcal{C}}^d(B,Z)_+/G \to {\mathcal{C}}^d(B,Z)_-/G$$
such that
$$\left(A_{\d,\vartheta},G_\d,\vartheta\right)\iso{G}\left(A_{\e,\chi},G_{\e},\chi\right),$$
for every $(\d,\vartheta)\in{\mathcal{C}}^d(B,Z)_+$ and some $(\e,\chi)\in\Omega(\overline{(\d,\vartheta)})$.
\end{conj}
By \cite[Lemma 3.8 (c)]{Spa17}, the above statement on character triples does not depend on the choice of $(\e,\chi)\in\Omega(\overline{(\d,\vartheta)})$ nor on the representative $(\d,\vartheta)$ of $\overline{(\d,\vartheta)}$. We will make use of this fact without further reference.
As shown in the following lemma, it is no loss of generality to assume ${\mathbf{O}}_p(G)\leq {\mathbf{Z}}(G)$ and consider $p$-chains with initial term ${\mathbf{O}}_p(G)$. This is an adaptation of a well known result (see, for instance, \cite[Theorem 9.16]{Nav18}).
\begin{lem}
\label{lem:Knorr-Robinson for Character Triple Conjecture}
Conjecture \ref{conj:Character Triple Conjecture} holds whenever $Z<{\mathbf{O}}_p(G)$.
\end{lem}
\begin{proof}
Consider $\d\in\mathfrak{N}(G,Z)$ with $\d=\{D_0<D_1\dots<D_n\}$. If ${\mathbf{O}}_p(G)\nleq D_n$, then define $\d^*$ to be the $p$-chain obtained by adding ${\mathbf{O}}_p(G)D_n$ to $\d$. Assume ${\mathbf{O}}_p(G)\leq D_n$ and let $k$ be the unique nonnegative integer such that ${\mathbf{O}}_p(G)\leq D_k$ and ${\mathbf{O}}_p(G)\nleq D_{k-1}$. If ${\mathbf{O}}_p(G)D_{k-1}=D_k$, then we define $\d^*$ by deleting the term $D_k$ from $\d$. If ${\mathbf{O}}_p(G)D_{k-1}<D_k$, then we define $\d^*$ by adding the term ${\mathbf{O}}_p(G)D_{k-1}$ to $\d$. This defines a self-inverse ${\mathbf{N}}_A(Z)$-equivariant bijection $^*:\mathfrak{N}(G,Z)\to \mathfrak{N}(G,Z)$ such that $|\d|=|\d^*|\pm 1$. In particular $G_\d=G_{\d^*}$ and we define $\Omega(\overline{(\d,\vartheta)}):=\overline{(\d^*,\vartheta)}$, for every $(\d,\vartheta)\in{\mathcal{C}}^d(B,Z)_+$.
\end{proof}
\subsection{A consequence of the Harris--Kn\"orr theorem}
Next, we collect some consequences of the Harris--Kn\"orr theorem that will be used in the sequel. The reader should notice that Corollary \ref{cor:Harris-Knorr} and Corollary \ref{cor:Harris-Knorr and p-chains} below can also be deduced from \cite[Theorem 4.1]{Tur17}. However, we present here an elementary argument.
\begin{lem}
\label{lem:Harris-Knorr}
Let $N\unlhd G$ and $P$ be a $p$-subgroup of $N$. Consider a block $b\in{\rm Bl}(N\mid P)$ and its Brauer first main correspondent $b'\in{\rm Bl}({\mathbf{N}}_N(P)\mid P)$. Let $B'\in{\rm Bl}({\mathbf{N}}_G(P))$ and set $B:=(B')^G$. Then $B'$ covers $b'$ if and only if $B$ covers $b$.
\end{lem}
\begin{proof}
The result follows from the proof of the Harris--Kn\"orr theorem \cite{Har-Kno85}.
\end{proof}
If $P$ is a $p$-group acting via automorphisms on a $p'$-group $N$, we denote by $f_P:{\rm Irr}_P(N)\to{\rm Irr}({\mathbf{C}}_N(P))$ the $P$-Glauberman correspondence (see \cite[Chapter 13]{Isa76} and \cite[\S 2.3]{Nav18}).
\begin{cor}
\label{cor:Harris-Knorr}
Let $N$ be a normal $p'$-subgroup of $G$ and $P$ be a $p$-subgroup of $G$. Consider $\mu\in{\rm Irr}_P(N)$ and set $\mu':=f_P(\mu)\in{\rm Irr}({\mathbf{C}}_N(P))$. If $B'\in{\rm Bl}({\mathbf{N}}_G(P))$, then $B'$ covers ${\rm bl}(\mu')$ if and only if $(B')^{N{\mathbf{N}}_G(P)}$ covers ${\rm bl}(\mu)$. Moreover, if $\mu$ is $G$-invariant, then $B'$ covers ${\rm bl}(\mu')$ if and only if $(B')^{G}$ covers ${\rm bl}(\mu)$.
\end{cor}
\begin{proof}
Let $b'$ be the unique block of ${\mathbf{N}}_{NP}(P)$ that covers ${\rm bl}(\mu')$, $b$ the unique block of $NP$ that covers ${\rm bl}(\mu)$ (see \cite[Corollary 9.6]{Nav98}) and notice that $b$ and $b'$ are Brauer first main correspondents over $P$. Now, ${{\rm Bl}(N{\mathbf{N}}_G(P)\mid {\rm bl}(\mu))}={{\rm Bl}(N{\mathbf{N}}_G(P)\mid b)}$ and ${{\rm Bl}({\mathbf{N}}_G(P)\mid {\rm bl}(\mu'))}={{\rm Bl}({\mathbf{N}}_G(P)\mid b')}$ and applying Lemma \ref{lem:Harris-Knorr} it follows that $B'$ covers ${\rm bl}(\mu')$ if and only if $(B')^{N{\mathbf{N}}_G(P)}$ covers ${\rm bl}(\mu)$. Moreover, if $\mu$ is $G$-invariant, then ${\rm bl}(\mu)$ is covered by $(B')^{N{\mathbf{N}}_G(P)}$ if and only if it is covered by $(B')^G$.
\end{proof}
With the same argument, we obtain a version of the above corollary for normal $p$-chains.
\begin{cor}
\label{cor:Harris-Knorr and p-chains}
Let $N$ be a normal $p'$-subgroup of $G$ and $\d$ be a normal $p$-chain of $G$ with last term $P$. Consider $\mu\in{\rm Irr}_P(N)$ and set $\mu':=f_P(\mu)\in{\rm Irr}({\mathbf{C}}_N(P))$. If $B'\in{\rm Bl}(G_\d)$, then $B'$ covers ${\rm bl}(\mu')$ if and only if $(B')^{NG_\d}$ covers ${\rm bl}(\mu)$. Moreover, if $\mu$ is $G$-invariant, then $B'$ covers ${\rm bl}(\mu')$ if and only if $(B')^G$ covers ${\rm bl}(\mu)$.
\end{cor}
\begin{proof}
The proof of Corollary \ref{cor:Harris-Knorr} applies with minor changes.
\end{proof}
\subsection{Construction of $N$-block isomorphic character triples}
We prove some useful results that can be used to construct $N$-block isomorphic character triples. First, we give a version of \cite[Theorem 3.14]{Nav-Spa14I} for our situation. This proposition allows to obtain new $N$-block isomorphic character triples involving irreducibly induced characters. This is the case, for instance, when we apply the Fong--Reynolds correspondence or the Clifford correspondence. Before proving this result, we need an easy lemma.
\begin{lem}
\label{lem:Irreducible induction}
Let $N\unlhd G$ and $\vartheta\in{\rm Irr}(N)$. If $\vartheta^G\in{\rm Irr}(G)$, then ${\mathbf{C}}_G(N)\leq N$.
\end{lem}
\begin{proof}
Set $H:=N{\mathbf{C}}_G(N)$ and observe that $\psi:=\vartheta^H\in{\rm Irr}(H)$. Since $\vartheta$ is $H$-invariant we have $\psi_N=e\vartheta$ with $e=|H:N|$. However $e=[\psi_N,\vartheta]=[\psi,\psi]=1$ and therefore ${\mathbf{C}}_G(N)\leq N$.
\end{proof}
\begin{prop}
\label{prop:Irreducible induction and double relation block}
Let $N\unlhd G$ and $G_0\leq G$. For $i=1,2$, consider $H_i\leq G$ such that $G=NH_i$ and set $M_i:=N\cap H_i$, $H_{0,i}:=G_0\cap H_i$, $M_{0,i}:=G_0\cap M_i$ and $N_0:=G_0\cap N\unlhd G_0$. Suppose that $G=G_0N$, that $H_i=H_{0,i}M_i$ and that $\varphi_i:=(\varphi_{0,i})^{M_i}\in{\rm Irr}(M_i)$, for some $\varphi_{0,i}\in{\rm Irr}(M_{0,i})$. If $(H_{0,1},M_{0,1},\varphi_{0,1})\iso{N_0}(H_{0,2},M_{0,2},\varphi_{0,2})$, there exists a defect group $D_i\in \delta({\rm bl}(\varphi_i))$ such that ${\mathbf{C}}_G(D_i)\leq H_i$, and induction ${\rm Ind}_{J_{0,i}}^{J_i}:{\rm Irr}(J_{0,i}\mid \varphi_{0,i})\to {\rm Irr}(J_i\mid \varphi_i)$ defines a bijection for every $N\leq J\leq G$, where $J_i:=J\cap H_i$ and $J_{0,i}:=J\cap H_{0,i}$, then $(H_1,M_1,\varphi_1)\iso{N}(H_2,M_2,\varphi_2)$.
\end{prop}
\begin{proof}
Assume $(H_{0,1},M_{0,1},\varphi_{0,1})\iso{N_0}(H_{0,2},M_{0,2},\varphi_{0,2})$ via $(\mathcal{P}_{0,1},\mathcal{P}_{0,2})$ and let $\alpha_{0,i}$ be the factor set of $\mathcal{P}_{0,i}$. Consider the canonical isomorphisms $l_{0,i}:H_{0,i}/M_{0,i}\to G_0/N_0$ and $l_i:H_i/M_i\to G/N$ and set $i_0:=l_{0,2}^{-1}\circ l_{0,1}$ and $i=l_2^{-1}\circ l_1$. If $j:G/N\to G_0/N_0$ and $j_i:H_i/M_i\to H_{0,i}/M_{0,i}$ are the canonical isomorphisms, then we have a commutative diagram
\begin{center}
\begin{tikzcd}
H_1/M_1\arrow[r, "l_1"]\arrow[rr, bend left={21pt}, "i"]\arrow[d,swap, "j_1"] &G/N\arrow[d, "j"] & H_2/M_2\arrow[l, swap, "l_2"]\arrow[d, "j_2"]
\\
H_{0,1}/M_{0,1}\arrow[r, swap, "l_{0,1}"]\arrow[rr, swap, bend right={21pt}, "i_0"] &G_0/N_0 & H_{0,2}/M_{0,2}\arrow[l, "l_{0,2}"]
\end{tikzcd}
\end{center}
As in \cite[Theorem 3.14]{Nav-Spa14I}, we consider the projective representation $\mathcal{P}_i:=(\mathcal{P}_{0,i})^{H_i}\in {{\rm Proj}(H_i\mid \alpha_i)}$ defined as follows: let $\{t_{i,1},\dots, t_{i,n}\}$ be a $H_i$-transversal for $H_{0,i}$ contained in $M_i$, where $n:={|G:G_0|}={|H_i:H_{0,i}|}$. For every $x\in H_i$, let
\begin{align*}
\mathcal{P}_{i,j,k}(x):=\begin{cases}
\mathcal{P}_{0,i}(t_{i,j}^{-1}xt_{i,k}), &\text{ if }t_{i,j}^{-1}xt_{i,k}\in H_{0,i}
\\
0,&\text{ otherwise }
\end{cases}
\end{align*}
and define
$$\mathcal{P}_i(x):=\begin{pmatrix}
\mathcal{P}_{i,1,1}(x) &\dots &\mathcal{P}_{i,1,n}(x)
\\
\vdots &&\vdots
\\
\mathcal{P}_{i,n,1}(x) &\dots &\mathcal{P}_{i,n,n}(x)
\end{pmatrix}.$$
Then, $\mathcal{P}_i$ is a projective representation of $H_i$ associated with $\varphi_i=\varphi_{0,i}^{M_i}$ with factor set $\alpha_i$ satisfying $\alpha_i(x,y)=\alpha_{0,i}(j_i(x),j_i(y))$, for all $x,y\in H_i/M_i$. Since
$$\alpha_{0,1}(j_1(x),j_1(y))=\alpha_{0,2}(i_0(j_1(x)),i_0(j_1(y))),$$
we conclude that $\alpha_1(x,y)=\alpha_2(i(x),i(y))$, for all $x,y\in H_1/M_1$.
We claim that ${\mathbf{C}}_{H_i}(M_i)\leq G_0$. In this case, since ${\mathbf{C}}_G(N)\leq {\mathbf{C}}_G(D_i)\leq H_i$, we deduce ${\mathbf{C}}_G(N)\leq {\mathbf{C}}_{G_0}(N_0)$. To prove the claim, fix $x\in {\mathbf{C}}_{H_i}(M_i)$, set $J_i:=\langle M_i,x \rangle$ and $J_{0,i}:=G_0\cap J_i$ and let $\varphi_{i,x}$ be an extension of $\varphi_i$ to $J_i$. Since ${\rm Ind}_{J_{0,i}}^{J_i}:{\rm Irr}(J_{0,i}\mid \varphi_{0,i})\to {\rm Irr}(J_i\mid \varphi_i)$ is a bijection, we can find an irreducible character $\varphi_{0,i,x}\in{\rm Irr}(J_{0,i}\mid \varphi_{0,i})$ such that $\varphi_{0,i,x}^{J_i}=\varphi_{i,x}$. By Lemma \ref{lem:Irreducible induction} we conclude that $x\in {\mathbf{C}}_{J_i}(J_{0,i})\leq J_{0,i}\leq G_0$. This proves the claim, hence ${\mathbf{C}}_G(N)\leq{\mathbf{C}}_{G_0}(N_0)$. Now, since $\mathcal{P}_{0,1,{\mathbf{C}}_{G_0}(N_0)}$ and $\mathcal{P}_{0,2,{\mathbf{C}}_{G_0}(N_0)}$ are associated with the same scalar function and $[t_{i,j},{\mathbf{C}}_G(N)]=1$ for every $i=1,2$ and $j=1,\dots, n$, then the same is true for $\mathcal{P}_{1,{\mathbf{C}}_G(N)}$ and $\mathcal{P}_{2,{\mathbf{C}}_G(N)}$.
Next, fix $N\leq J\leq G$, set $J_0:=J\cap G_0$, $J_i:=J\cap H_i$ and $J_{0,i}:=J\cap H_{0,i}$, and consider the bijections given by the character triple isomorphisms induced by $(\mathcal{P}_{0,1},\mathcal{P}_{0,2})$ and $(\mathcal{P}_1,\mathcal{P}_2)$:
\begin{align*}
\sigma_{0,J_{0,1}}:{\rm Irr}(J_{0,1}\mid \varphi_{0,1}) &\to {\rm Irr}(J_{0,2}\mid \varphi_{0,2})
\\
{\rm tr}(\mathcal{Q}_{0,J_{0,1}}\otimes \mathcal{P}_{0,1,J_{0,1}}) &\mapsto {\rm tr}(\mathcal{Q}_{0,J_{0,2}}\otimes \mathcal{P}_{0,2,J_{0,2}})
\end{align*}
where $\mathcal{Q}_0\in{\rm Proj}(J_0/N_0)$, and
\begin{align*}
\sigma_{J_1}:{\rm Irr}(J_1\mid \varphi_1) &\to {\rm Irr}(J_2\mid \varphi_2)
\\
{\rm tr}(\mathcal{Q}_{J_1}\otimes \mathcal{P}_{1,J_1}) &\mapsto {\rm tr}(\mathcal{Q}_{J_2}\otimes \mathcal{P}_{2,J_2})
\end{align*}
where $Q\in{\rm Proj}(J/N)$. Observe that $\sigma_{J_1}\left(\psi_0^{J_1}\right)=\left(\sigma_{0,J_{0,1}}(\psi_0)\right)^{J_2}$, for all $\psi_0\in {\rm Irr}(J_{0,1}\mid \varphi_{0,1})$. Let $\psi\in{\rm Irr}(J_1\mid\varphi_1)$ and write $\psi=\psi_0^{J_1}$, for some $\psi_0\in{\rm Irr}(J_{0,1}\mid \varphi_{0,1})$. Since by hypothesis ${\rm bl}(\psi_0)^{J_0}={\rm bl}(\sigma_{0,J_{0,1}}(\psi_0))^{J_0}$, we conclude that ${\rm bl}(\psi)^J={\rm bl}(\sigma_{J_1}(\psi))^J$.
\end{proof}
Whenever we have a pair of $N$-isomorphic character triples, there is an induced strong isomorphism of character triples with some special properties. In the following lemma we describe some of these special features.
\begin{lem}
\label{lem:Bijections induced by triple isomorphisms}
Let $(H_1,M_1,\vartheta_1)\iso{N}(H_2,M_2,\vartheta_2)$ given by $(\mathcal{P}_1,\mathcal{P}_2)$ and, for $N\leq J\leq G=NH_i$, consider the bijection $\sigma_{J_1}^{(\mathcal{P}_1,\mathcal{P}_2)}:{\rm Irr}(J_1\mid \vartheta_1)\to{\rm Irr}(J_2\mid \vartheta_2)$, where $J_i:=J\cap H_i$. Let $\psi_1\in{\rm Irr}(J_1\mid \vartheta_1)$ and $\psi_2:=\sigma_{J_1}^{(\mathcal{P}_1,\mathcal{P}_2)}(\psi_1)$. Then:
\begin{enumerate}
\item there exists $\wh{\mathcal{Q}}\in{\rm Proj}(JH_{i,J,\psi_i}/N)$ such that $\wh{\mathcal{Q}}_{i,J_i}\otimes \mathcal{P}_{i,J_i}$ affords $\psi_i$, where $\wh{\mathcal{Q}}_i:=\mathcal{Q}_{H_{i,J,\psi_i}}$;
\item if $\mathcal{D}_i:=\wh{\mathcal{Q}}_i\otimes \mathcal{P}_{i}$, then
$$(H_{1,J,\psi_1},J_1,\psi_1)\iso{J}(H_{2,J,\psi_2},J_2,\psi_2)$$
via $(\mathcal{D}_1,\mathcal{D}_2)$. Moreover
$$\sigma_{K_1}^{(\mathcal{D}_1,\mathcal{D}_2)}(\chi_1)=\sigma_{K_1}^{(\mathcal{P}_1,\mathcal{P}_2)}(\chi_1),$$
for every $J\leq K\leq JH_{i,J,\psi_i}$ and $\chi\in{\rm Irr}(K_1\mid \psi_1)\subseteq {\rm Irr}(K_1\mid \vartheta_1)$, where $K_i:=K\cap H_i$;
\item $d(\psi_1)-d(\psi_2)=d(\vartheta_1)-d(\vartheta_2)$.
\end{enumerate}
\end{lem}
\begin{proof}
First, as $JH_{1,J}=G_J=JH_{2,J}$, we may assume $J\unlhd G$. Moreover, since $(i,\sigma)$ is a strong isomorphism of character triples, we know that
$$\sigma_{J_1}(\psi_1)^{x_2}=\sigma_{J_1^{x_1}}(\psi_1^{x_1})=\sigma_{J_1}(\psi_1^{x_1}),$$
for every $x_1\in H_1$ and $x_2\in H_2$ such that $i(M_1x_1)=M_2x_2$. In particular $i(H_{1,\psi_1}/M_1)=H_{2,\psi_2}/M_2$ and so $JH_{1,\psi_1}=JH_{2,\psi_2}$. Therefore, we may assume $H_i=H_{i,\psi_i}$.
By \cite[Theorem 3.3]{Spa17}, there exists $\mathcal{Q}\in{\rm Proj}(J/N\mid \alpha^{-1}_{J\times J})$ such that $\psi_i$ is afforded by $\mathcal{Q}_{J_i}\otimes \mathcal{P}_{i,J_i}$. By \cite[Theorem 5.5]{Nav18}, there exists $\mathcal{D}_1\in{\rm Proj}(H_1)$ such that $\mathcal{D}_{1,J_1}=\mathcal{Q}_{J_1}\otimes \mathcal{P}_{1,J_1}$. Arguing as in \cite[p. 707]{Nav-Spa14I}, relying on the proof of \cite[Theorem 8.16]{Nav98} we can find $\wh{\mathcal{Q}}_1\in{\rm Proj}(H_1)$ such that
$$\mathcal{D}_1=\wh{\mathcal{Q}}_1\otimes \mathcal{P}_1$$
and $\wh{\mathcal{Q}}_{1,J_1}=\mathcal{Q}_{J_1}$. Since $N\leq {\rm Ker}(\mathcal{Q})$, we deduce that $M_1\leq {\rm Ker}(\mathcal{Q}_{J_1})\leq {\rm Ker}(\wh{\mathcal{Q}}_1)$. Now $\wh{\mathcal{Q}}_1\in{\rm Proj}(H_1/M_1)$ and, using the isomorphism $H_1/M_1\simeq G/N\simeq H_2/M_2$, we define $\wh{\mathcal{Q}}\in{\rm Proj}(G/N)$ and $\wh{\mathcal{Q}}_2\in {\rm Proj}(H_2/M_2)$. This proves (i). Set
$$\mathcal{D}_2:=\wh{\mathcal{Q}}_2\otimes \mathcal{P}_2.$$
We claim that $(H_1,J_1,\psi_1)\iso{J}(H_2,J_2,\psi_2)$ via $(\mathcal{D}_1,\mathcal{D}_2)$. Clearly the condition on the factor sets is satisfied. Moreover, since $\psi_i$ lies over $\vartheta_i$, we can find $D_i\in\delta({\rm bl}(\psi_i))$ and $Q_i\in\delta({\rm bl}(\vartheta_i))$ with $Q_i\leq D_i$ and ${\mathbf{C}}_G(Q_i)\leq H_i$. It follows that ${\mathbf{C}}_G(D_i)\leq H_i$. To conclude, we need to check the condition on block induction for $\sigma^{(\mathcal{D}_1,\mathcal{D}_2)}$. It's enough to show that $\sigma_{K_1}^{(\mathcal{D}_1,\mathcal{D}_2)}$ coincides with $\sigma_{K_1}^{(\mathcal{P}_1,\mathcal{P}_2)}$ on ${\rm Irr}(K_1\mid \psi_1)$, for every $J\leq K\leq G$, where $K_i:=K\cap H_i$. Consider $\chi_i\in{\rm Irr}(K_1\mid \psi_1)$ and let $\mathcal{R}\in{\rm Proj}(K/J)$ such that $\chi_1={\rm tr}(\mathcal{R}_{K_1}\otimes \mathcal{D}_{1,K_1})$. Then
\begin{align*}
\sigma_{K_1}^{(\mathcal{D}_1,\mathcal{D}_2)}(\chi_1)&={\rm tr}(\mathcal{R}_{K_2}\otimes \mathcal{D}_{2,K_2})
\\
&={\rm tr}(\mathcal{R}_{K_2}\otimes \wh{Q}_{K_2}\otimes \mathcal{P}_{2,K_2})
\\
&=\sigma_{K_1}^{(\mathcal{P}_1,\mathcal{P}_2)}({\rm tr}(\mathcal{R}_{K_1}\otimes \wh{\mathcal{Q}}_{K_1}\otimes \mathcal{P}_{1,K_1}))
\\
&=\sigma_{K_1}^{(\mathcal{P}_1,\mathcal{P}_2)}({\rm tr}\left(\mathcal{R}_{K_1}\otimes \mathcal{D}_{1,K_1}\right))
\\
&=\sigma_{K_1}^{(\mathcal{P}_1,\mathcal{P}_2)}(\chi_1).
\end{align*}
and the proof of (ii) is complete. To conclude, since $\psi_1(1)/\vartheta_1(1)=\psi_2(1)/\vartheta_2(1)$ by \cite[Lemma 11.24]{Isa76} and $|J:J_i|=|N:M_i|$, it follows that
$$p^{d(\psi_1)-d(\psi_2)}=\dfrac{|J_1|_p\psi_2(1)_p}{|J_2|_p\psi_1(1)_p}=\dfrac{|M_1|_p\vartheta_2(1)_p}{|M_2|_p\vartheta_1(1)_p}=p^{d(\vartheta_1)-d(\vartheta_2)}.$$
This finishes the proof.
\end{proof}
Given a defect preserving equivariant bijection respecting $N$-block isomorphic character triples, we show how to obtain another bijection over the given one that satisfies similar properties. For $N\unlhd G$ and $\mathcal{S}\subseteq {\rm Irr}(N)$, we denote by ${\rm Irr}(G\mid \mathcal{S})$ the set of $\chi\in{\rm Irr}(G)$ that lies over some $\vartheta\in\mathcal{S}$.
\begin{prop}
\label{prop:Constructing bijections over bijections}
Let $K\unlhd A$, $A_0\leq A$ with $A=KA_0$ and, for every subgroup $X\leq A$, set $X_0:=X\cap A_0$. Consider $A_0$-stable subsets of characters $\mathcal{S}\subseteq {\rm Irr}(K)$ and $\mathcal{S}_0\subseteq {\rm Irr}(K_0)$. Assume there exists an $A_0$-equivariant bijection
$$\Psi:\mathcal{S}\to\mathcal{S}_0$$
such that
$$\left(A_\vartheta,K,\vartheta\right)\iso{K}\left(A_{0,\vartheta},K_0,\Psi(\vartheta)\right)$$
and
$${\mathbf{C}}_A(D)\leq A_0,$$
for every $\vartheta\in\mathcal{S}$ and some defect group $D$ of ${\rm bl}(\Psi(\vartheta))$. Then, for every $K\leq J\leq A$, there exists an $A_{0,J}$-equivariant bijection
$$\Phi_J:{\rm Irr}(J\mid \mathcal{S})\to{\rm Irr}(J_0\mid \mathcal{S}_0)$$
such that
$$\left(A_{J,\chi},J,\chi\right)\iso{J}\left(A_{0,J,\chi},J_0, \Phi_J(\chi)\right)$$
and
$${\mathbf{C}}_A(Q)\leq A_0,$$
for every $\chi\in{\rm Irr}(J\mid \mathcal{S})$ and some defect group $Q$ of ${\rm bl}(\Phi_J(\chi))$. Moreover $\Psi$ preserves the defect of characters if and only if so does $\Phi_J$.
\end{prop}
\begin{proof}
Consider an ${\mathbf{N}}_{A_0}(J)$-transversal $\mathbb{S}$ in $\mathcal{S}$ and define $\mathbb{S}_0:=\{\Psi(\vartheta)\mid \vartheta\in\mathbb{S}\}$. Since $\Psi$ is $A_0$-equivariant, it follows that $\mathbb{S}_0$ is an ${\mathbf{N}}_{A_0}(J)$-transversal in $\mathcal{S}_0$. For every $\vartheta\in\mathbb{S}$, with $\vartheta_0:=\Psi(\vartheta)\in\mathbb{S}_0$, we fix a pair of projective representations $(\mathcal{P}^{(\vartheta)},\mathcal{P}^{(\vartheta_0)}_0)$ giving $(A_\vartheta,K,\vartheta)\iso{K}(A_{0,\vartheta},K_0,\vartheta_0)$. Now, let $\mathbb{T}$ be an ${\mathbf{N}}_{A_0}(J)$-transversal in ${\rm Irr}(J\mid \mathcal{S})$ such that every character $\chi\in\mathbb{T}$ lies above a character $\vartheta\in\mathbb{S}$ (this can be done by the choice of $\mathbb{S}$). Moreover, as $A=KA_0$, we have $J=KJ_0$ and therefore every $\chi\in\mathbb{T}$ lies over a unique $\vartheta\in\mathbb{S}$ by Clifford's theorem.
For $\chi\in\mathbb{T}$ lying over $\vartheta\in\mathbb{S}$, let $\varphi\in{\rm Irr}(J_\vartheta\mid \vartheta)$ be the Clifford correspondent of $\chi$ over $\vartheta$. Set $\vartheta_0:=\Psi(\vartheta)\in\mathbb{S}_0$ and consider the ${\mathbf{N}}_{A_0}(J)_\vartheta$-equivariant bijection $\sigma_{J_\vartheta}:{\rm Irr}(J_\vartheta\mid \vartheta)\to{\rm Irr}(J_{0,\vartheta}\mid \vartheta_0)$ induced by our choice of projective representations $(\mathcal{P}^{(\vartheta)},\mathcal{P}_0^{(\vartheta_0)})$. Let $\varphi_0:=\sigma_{J_\vartheta}(\varphi)$. Since $\Psi$ is $A_0$-equivariant, we deduce that $J_{0,\vartheta}=J_{0,\vartheta_0}$ and therefore $\Phi_J(\chi):=\varphi^{J_0}$ is irreducible by the Clifford correspondence. Then, we define
$$\Phi_J\left(\chi^x\right):=\Phi_J(\chi)^x,$$
for every $\chi\in\mathbb{T}$ and $x\in {\mathbf{N}}_{A_0}(J)$. This defines an ${\mathbf{N}}_{A_0}(J)$-equivariant bijection $\Psi:{\rm Irr}(J\mid \mathcal{S})\to{\rm Irr}(J_0\mid \mathcal{S}_0)$. Furthermore, using Lemma \ref{lem:Bijections induced by triple isomorphisms} it's clear that $\Psi$ preserves the defect of characters if and only if so does $\Phi_J$.
Next, using the fact that $\left(A_\vartheta,K,\vartheta\right)\iso{K}\left(A_{0,\vartheta},K_0,\vartheta_0\right)$ together with Lemma \ref{lem:Bijections induced by triple isomorphisms}, we have
$$\left(A_{\vartheta,J_\vartheta,\psi},J_\vartheta,\psi\right)\iso{J_\vartheta}\left(A_{0,\vartheta,J_\vartheta,\psi},J_{0,\vartheta},\psi_0\right)$$
and, because $A_{\vartheta,J}\leq A_{\vartheta,J_{\vartheta}}$, we obtain
\begin{equation}
\label{eq:Constructing bijections over bijections 1}
\left(A_{\vartheta,J,\psi},J_\vartheta,\psi\right)\iso{J_\vartheta}\left(A_{0,\vartheta,J,\psi},J_{0,\vartheta},\psi_0\right).
\end{equation}
By hypothesis there exists a defect group $D$ of ${\rm bl}(\vartheta_0)$ such that ${\mathbf{C}}_A(D)\leq A_0$. Since ${\rm bl}(\chi_0)$ covers ${\rm bl}(\vartheta_0)$ we can find a defect group $Q$ of ${\rm bl}(\chi_0)$ such that $D\leq Q$. It follows that ${\mathbf{C}}_A(Q)\leq {\mathbf{C}}_A(D)\leq A_0$. Finally, we obtain
$$\left(A_{J,\chi},J,\chi\right)\iso{J}\left(A_{0,J,\chi},J_0, \Phi_J(\chi)\right)$$
by applying Proposition \ref{prop:Irreducible induction and double relation block} together with \eqref{eq:Constructing bijections over bijections 1}.
\end{proof}
We end this section with an elementary but useful observation. Suppose to have $N$-block isomorphic character triples and that $N\leq \wh{N}$. Under certain assumptions, it's possible to deduce that those character triples are in fact $\wh{N}$-block isomorphic.
\begin{lem}
\label{lem:Relation on triple wrt different groups}
Let $(H_1,M_1,\vartheta_1)\iso{N}(H_2,M_2,\vartheta_2)$ with $H_iN=G$. Suppose that $G\leq \wh{G}$ and let $N\leq \wh{N}\unlhd \wh{G}$ with $\wh{G}=G\wh{N}$ and $N=G\cap \wh{N}$. If ${\mathbf{C}}_{\wh{G}}(D_i)\leq G$ for some $D_i\in\delta({\rm bl}(\vartheta_i))$, then $(H_1,M_1,\vartheta_1)\iso{\wh{N}}(H_2,M_2,\vartheta_2)$.
\end{lem}
\begin{proof}
This follows directly from Definition \ref{def:N-block isomorphic character triples}.
\end{proof}
\section{$N$-block isomorphic character triples and Glauberman correspondence}
The aim of this section is to prove Theorem \ref{thm:Really above the Glauberman correspondence with block relation with chains} which will be one of the main ingredients in the final proof. To prove this result, we need to extend the bijection given in \cite[Theorem 5.13]{Nav-Spa14I} to characters of positive height. This is done in Proposition \ref{prop:Above the Glauberman correspondence with block relation} for the case where the $D$-correspondence coincides with the Glauberman correspondence. Moreover, we obtain a canonical bijection.
Let $N\unlhd G$ and $\vartheta\in{\rm Irr}_G(N)$ such that $(o(\vartheta)\vartheta(1),|G:N|)=1$. We denote by $\vartheta^\diamond$ the canonical extension of $\vartheta$ to $G$, i.e. the unique extension of $\vartheta$ to $G$ such that $(o(\vartheta^\diamond),|G:N|)=1$ (see \cite[Corollary 8.16]{Isa76}). To prove Proposition \ref{prop:Above the Glauberman correspondence with block relation}, in addition to the argument developed in \cite[\S 5]{Nav-Spa14I}, we need the following result on the extendibility of the canonical extension. In what follows we will often use the following easy fact: if $H\leq G$ and $\chi\in{\rm Irr}(G)$ such that $\chi_H\in{\rm Irr}(H)$, then $o(\chi_H)$ divides $o(\chi)$.
\begin{lem}
\label{lem:Canonical extension}
Let $N$ be a normal $p'$-subgroup of $G$ and $P$ a $p$-subgroup of $G$ such that $K:=NP\unlhd G$. Consider $\mu\in{\rm Irr}_G(N)$ and let $\mu^\diamond\in{\rm Irr}_G(K)$. Then $\mu$ extends to $G$ if and only if $\mu^\diamond$ extends to $G$.
\end{lem}
\begin{proof}
One implication is trivial. Notice that $\mu^\diamond$ is $G$-invariant since $\mu$ is $G$-invariant. Assume that $\mu$ has an extension $\chi\in{\rm Irr}(G)$. We have to show that $\mu^\diamond$ extends to $H$, for every $H/K\in{\rm Syl}_q(G/K)$ and every prime $q$. If $q=p$, then $\mu$ has a canonical extension to $H$, which is also an extension of $\mu^\diamond$.
Assume $q\neq p$ and consider $\lambda\in{\rm Irr}(K/N)$ such that $\mu^\diamond=\lambda\chi_K$. Notice that, as $\mu^\diamond$ and $\chi_K$ are $G$-invariant, the character $\lambda$ is $G$-invariant. Since $K/N$ is a $p$-group and $H/K$ is a $q$-group, we deduce that $\lambda$ has a canonical extension $\lambda^\diamond$ to $H$. Then $\lambda^\diamond\chi_H$ is an extension of $\mu^\diamond$. This concludes the proof.
\end{proof}
\begin{hyp}
\label{hyp:Glauberman section}
Let $N$ be a normal $p'$-subgroup of $A$ and $P$ be a $p$-subgroup of $A$ such that $K:=NP\unlhd A$. Consider $\mu\in{\rm Irr}_A(N)$ and its Glauberman correspondent $f_P(\mu)\in{\rm Irr}_{A_P}(N_P)$. Let $\mu^\diamond\in{\rm Irr}_A(K)$ and $f_P(\mu)^\diamond\in{\rm Irr}_{A_P}(K_P)$ be the canonical extensions respectively of $\mu$ and $f_P(\mu)$.
\end{hyp}
Now, our aim is to obtain an adaptation of \cite[Proposition 5.12]{Nav-Spa14I} that includes canonical extensions (see Lemma \ref{lem:Canonical extension and values} below). This is done by proceeding as in \cite[\S 5]{Nav-Spa14I} and using Lemma \ref{lem:Canonical extension}.
\begin{lem}
\label{lem:Canonical extension and irreducible constituents}
Assume Hypothesis \ref{hyp:Glauberman section} and let $C$ be an abelian normal subgroup of $A$ with $C\leq {\mathbf{C}}_A(K)$. Suppose that $\mu^\diamond$ has an extension $\wt{\mu}$ to $A$. Then there exists an extension $\widetilde{f_P(\mu)}$ of $f_P(\mu)^\diamond$ to $A_P$ such that
$${\rm Irr}\left(\wt{\mu}_C\right)={\rm Irr}\left(\widetilde{f_P(\mu)}_C\right).$$
\end{lem}
\begin{proof}
Write $C_p:={\mathbf{O}}_p(C)$ and $C_{p'}:={\mathbf{O}}_{p'}(C)$ and set $\kappa:=\wt{\mu}_{NC_{p'}}$. Let $\kappa^\diamond$ be the canonical extension of $\kappa$ to $KC$. Since $\kappa$ extends to $A$, there exists an extension $\wt{\kappa}$ of $\kappa^\diamond$ to $A$ by Lemma \ref{lem:Canonical extension}. Observe that $\kappa^\diamond$ extends $\mu^\diamond$ and so does $\wt{\kappa}$. Now, by \cite[Corollary 6.17]{Isa76}, there exists a linear character $\eta\in{\rm Irr}(A/K)$ such that $\wt{\mu}=\wt{\kappa}\eta$. Let $\lambda$ and $\lambda_1$ be the unique irreducible constituent respectively of $\wt{\mu}_C$ and of $\wt{\kappa}_C$. Then $\lambda=\lambda_1\eta_C$. Next, consider the $P$-Glauberman correspondent $f_P(\kappa)\in{\rm Irr}((NC_{p'})_P)$ of $\kappa$ and let $f_P(\kappa)^\diamond$ be its canonical extension to $(KC)_P$. Using \cite[Theorem 6.5]{Tur08I} and \cite[Theorem 7.12]{Tur09}, as $\kappa$ extends to $A$, we conclude that $f_P(\kappa)$ extends to $A_P$. By Lemma \ref{lem:Canonical extension} there exists an extension $\wt{f_P(\kappa)}$ of $f_P(\kappa)^\diamond$ to $A_P$. As before, notice that $\wt{f_P(\kappa)}$ is an extension of $f_P(\mu)^\diamond$. Define $\wt{f_P(\mu)}:=\wt{f_P(\kappa)}\eta_{A_P}$. Since $K_P\leq {\rm Ker}(\eta_{A_P})$, it follows that $\wt{f_P(\mu)}$ is an extension of $f_P(\mu)^\diamond$. If $\lambda'$ and $\lambda_1'$ are the unique irreducible constituents respectively of $\wt{f_P(\mu)}_C$ and $\wt{f_P(\kappa)}_C$, then $\lambda'=\lambda_1'\eta_C$. Therefore, in order to conclude, it is enough to show that $\lambda_1=\lambda_1'$. Write $\lambda_1=\lambda_{1,p}\times \lambda_{1,p'}$ and $\lambda_1'=\lambda_{1,p}'\times \lambda_{1,p'}'$, with $\lambda_{1,p},\lambda_{1,p}'\in{\rm Irr}(C_p)$ and $\lambda_{1,p'},\lambda_{1,p'}'\in{\rm Irr}(C_{p'})$. First, because $f_P(\kappa)$ is an irreducible constituent of $\kappa_{NC_{p'}}$ and $C_{p'}\leq {\mathbf{Z}}(NC_{p'})$, it follows that
$${\rm Irr}\left(\wt{\kappa}_{C_{p'}}\right)={\rm Irr}\left(\kappa_{C_p'}\right)={\rm Irr}\left(f_P(\kappa)_{C_{p'}}\right)={\rm Irr}\left(\wt{f_P(\kappa)}_{C_{p'}}\right)$$
and therefore $\lambda_{1,p'}=\lambda_{1,p'}'$. Observe that $\wt{\kappa}_{N\times C_p}=(\kappa^\diamond)_{N\times C_p}=\mu\times \lambda_{1,p}$. Since $p$ does not divide $o(\kappa^\diamond)$, it follows that $p$ does not divide $o(\mu\times \lambda_p)$. In particular $(p,o(\lambda_p))=1$ and therefore $\lambda_{1,p}=1_{C_p}$. By the same argument, we obtain $\lambda_{1,p}'=1_{C_p}$. This shows that $\lambda_1=\lambda_1'$ and the proof is complete.
\end{proof}
Next, we extend Lemma \ref{lem:Canonical extension and irreducible constituents} to the case where $C$ is not necessarily abelian.
\begin{cor}
\label{cor:Canonical extension and irreducible constituents}
Assume Hypothesis \ref{hyp:Glauberman section} and suppose that $\mu^\diamond$ has an extension $\wt{\mu}$ to $A$. Then there exists an extension $\widetilde{f_P(\mu)}$ of $f_P(\mu)^\diamond$ to $A_P$ such that
$${\rm Irr}\left(\wt{\mu}_{{\mathbf{C}}_A(K)}\right)={\rm Irr}\left(\widetilde{f_P(\mu)}_{{\mathbf{C}}_A(K)}\right).$$
\end{cor}
\begin{proof}
Set $C:={\mathbf{C}}_A(K)$, $C':=[C,C]$ and $\overline{A}:=A/C'$. Since $\wt{\mu}_K$ is irreducible, as remarked before \cite[Definition 2.7]{Spa18}, we have $C\leq {\mathbf{Z}}(\wt{\mu})$ and \cite[Lemma 2.27]{Isa76} implies that $\wt{\mu}_C=\mu(1)\lambda$, for some linear character $\lambda\in{\rm Irr}(C)$. In particular $C'\leq {\rm Ker}(\lambda)\leq {\rm Ker}(\wt{\mu})$. It follows that $C'\cap K$ is contained in ${\rm Ker}(\mu^\diamond)$ and ${\rm Ker}(f_P(\mu)^\diamond)$ while $C'\cap N$ is contained in ${\rm Ker}(\mu)$ and ${\rm Ker}(f_P(\mu))$. Via the canonical isomorphism $\overline{N}\simeq N/C'\cap N$, we can identify $\mu$ with a character $\overline{\mu}$ of $\overline{N}$. Similarly we can consider $\overline{\mu^\diamond}$ as a character of $\overline{K}$, $\overline{f_P(\mu)}$ as a character of $\overline{N_P}$ and $\overline{f_P(\mu)^\diamond}$ as a character of $\overline{K_P}$. Notice that $\overline{A_P}=\overline{A}_{\overline{P}}$, $\overline{K_P}=\overline{K}_{\overline{P}}$ and $\overline{N_P}=\overline{N}_{\overline{P}}$. By \cite[Lemma 5.10]{Nav-Spa14I} the character $\overline{f_P(\mu)}$ coincides with the $\overline{P}$-Glauberman correspondent $f_{\overline{P}}(\overline{\mu})$ of $\overline{\mu}$. Moreover $\overline{\mu^\diamond}$ and $\overline{f_P(\mu)^\diamond}$ are the canonical extensions of $\overline{\mu}$ and of $\overline{f_P(\mu)}$. Applying Lemma \ref{lem:Canonical extension and irreducible constituents}, we find an extension $\psi$ of $\overline{f_P(\mu)^\diamond}$ to $\overline{A}_{\overline{P}}$ such that ${\rm Irr}\left(\overline{\wt{\mu}}_{\overline{C}}\right)={\rm Irr}\left(\psi_{\overline{C}}\right)$, where $\overline{\wt{\mu}}$ is the character of $\overline{A}$ corresponding to $\wt{\mu}$ via inflation. Now the inflation $\widetilde{f_P(\mu)}\in{\rm Irr}(A_P)$ of $\psi$ satisfies the required hypothesis.
\end{proof}
Recall that, if $\mathbf{R}$ is the ring of algebraic integers and $\mathbf{S}$ is the localization of $\mathbf{R}$ at some maximal ideal containing $p\mathbf{R}$, then $^*:\mathbf{S}\to\mathbb{F}$ denotes the canonical epimorphism, where $\mathbb{F}$ is the residue field of characteristic $p$ (see \cite[Chapter 2]{Nav98} for details).
\begin{lem}
\label{lem:Canonical extension and values}
Assume Hypothesis \ref{hyp:Glauberman section}. If $\mu^\diamond$ extends to $\wt{\mu}\in{\rm Irr}(A)$, then there exists an extension $\wt{f_P(\mu)}$ of $f_P(\mu)^\diamond$ to $A_P$ such that
$${\rm Irr}\left(\wt{\mu}_{{\mathbf{C}}_A(K)}\right)={\rm Irr}\left(\wt{f_P(\mu)}_{{\mathbf{C}}_A(K)}\right)$$
and
$$\wt{\mu}(x)^*=e\wt{f_P(\mu)}(x)^*,$$
for every $p$-regular $x\in A$ with $P\in {\rm Syl}_p({\mathbf{C}}_K(x))$, where $e:=[\mu_{N_P},f_P(\mu)]$.
\end{lem}
\begin{proof}
By Lemma \ref{lem:Canonical extension and irreducible constituents} there exists an extension $\chi$ of $f_P(\mu)^\diamond$ that satisfies the first condition. In order to conclude, it is enough to find a linear character $\wt{\xi}\in{\rm Irr}(A_P/{\mathbf{C}}_A(K)K_P)$ such that $\widetilde{f_P(\mu)}:=\wt{\xi}\cdot\chi$ satisfies the second condition.
First, we construct the linear character $\wt{\xi}$. Let $x$ be a $p$-regular element of ${\mathbf{C}}_A(P)K_P$, set $N^{(x)}:=N\langle x\rangle$, $K^{(x)}:=K\langle x \rangle$ and observe that $(N^{(x)})_P=(N_p)^{(x)}:=K_P\langle x\rangle$ and $(K^{(x)})_P=(K_p)^{(x)}:=K_P\langle x\rangle$.
\begin{center}
\begin{tikzpicture}[x=1.5cm, y=1cm]
\node(N) at (0,0) {$N$};
\node(K) at (0,1) {$K$};
\node(Nx) at (1,1) {$N^{(x)}$};
\node(Kx) at (1,2) {$K^{(x)}$};
\node(NP) at (1,-1) {$N_P$};
\node(KP) at (1,0) {$K_P$};
\node(NPx) at (2,0) {$N_P^{(x)}$};
\node(KPx) at (2,1) {$K_P^{(x)}$};
\path[-]
(N) edge node {} (K)
(N) edge node {} (NP)
(N) edge node {} (Nx)
(K) edge node {} (KP)
(K) edge node {} (Kx)
(Nx) edge node {} (Kx)
(Nx) edge node {} (NPx)
(NP) edge node {} (KP)
(NP) edge node {} (NPx)
(KPx) edge node {} (Kx)
(KPx) edge node {} (KP)
(KPx) edge node {} (NPx);
\end{tikzpicture}
\end{center}
Since $x$ is $p$-regular, the subgroup $N^{(x)}$ has order coprime to $p$ and we can consider the $P$-Glauberman correspondent $f_P(\wt{\mu}_{N^{(x)}})$ of $\wt{\mu}_{N^{(x)}}$. Moreover $f_P(\wt{\mu}_{N^{(x)}})_{N_P}=f_P(\mu)$ by \cite[Theorem A]{Isa-Nav91}. Now, if $f_P(\wt{\mu}_{N^{(x)}})^\diamond$ is the canonical extension of $f_P(\wt{\mu}_{N^{(x)}})$ to $K_P^{(x)}$, then $(f_P(\wt{\mu}_{N^{(x)}})^\diamond)_{K_P}=f_P(\mu)^\diamond$. Since $\chi_{K_P^{(x)}}$ is another extension of $f_P(\mu)^\diamond$ to $K_P^{(x)}$, it follows that there exists a unique linear character $\xi^{(x)}\in{\rm Irr}(K_P^{(x)}/K_P)$ such that $\xi^{(x)}\chi_{K_P^{(x)}}=f_P(\wt{\mu}_{N^{(x)}})^\diamond$. We define the map
\begin{align*}
\xi:{\mathbf{C}}_A(P)K_P&\to \mathbb{C}
\\
x &\mapsto \xi^{(x_{p'})}(x_{p'}).
\end{align*}
We claim that $\xi$ is a linear character of ${\mathbf{C}}_A(P)K_P$ with an extension $\wt{\xi}$ to $A_P$. To show that $\xi$ is an irreducible character we apply \cite[Corollary 8.12]{Isa76}. Clearly $\xi(1)=1$. Next, in order to show that $\xi$ is a class function we check that $\xi^{(x^n)}=(\xi^{(x)})^n$, for every $n\in A_P$ and every $p$-regular $x\in{\mathbf{C}}_A(P)K_P$. If this is the case, then
$$\xi(x^n)=\xi^{((x^n)_{p'})}((x^n)_{p'})=\xi^{((x_{p'})^n)}((x_{p'})^n)=(\xi^{(x_{p'})})^n((x_{p'})^n)=\xi^{(x_{p'})}(x_{p'})=\xi(x),$$
for every $x\in {\mathbf{C}}_A(P)K_P$ and $n\in A_P$. In particular $\xi$ is a class function. To prove the claim, just notice that $(\chi_{K_P^{(x)}})^n=\chi_{K_P^{(x^n)}}$ and that $(f_P(\wt{\mu}_{N^{(x)}})^\diamond)^n$ is the canonical extension of $f_P(\wt{\mu}_{N^{(x)}})^n=f_P(\wt{\mu}_{N^{(x^n)}})$, for every $n\in A_P$ and every $p$-regular $x\in{\mathbf{C}}_A(P)K_P$. Next, since $\xi^{(x)}=\xi^{(x^{-1})}$ for every $p$-regular $x\in{\mathbf{C}}_A(P)K_P$, we deduce that $\xi(x^{-1})=\xi^{-1}(x)$, for every $x\in {\mathbf{C}}_A(P)K_P$, and therefore $[\xi,\xi]=1$. Finally, fix $S\times T\leq {\mathbf{C}}_A(P)K_P$ with $S$ a $p$-group and $T$ a $p'$-group. Observe that $\xi_S=1_S$. On the other hand $\chi_{K_PT}$ and $f_P(\wt{\mu}_{NT})^\diamond$ are both extensions of $f_P(\mu)^\diamond$ and we can find a linear character $\lambda\in{\rm Irr}(K_PT/K_P)$ such that $\lambda\chi_{K_PT}=f_P(\wt{\mu}_{NT})^\diamond$. Moreover, for every $x\in T$, we have $\left(f_P(\wt{\mu}_{NT})^\diamond\right)_{K_P^{(x)}}=f_P(\wt{\mu}_{N^{(x)}})^\diamond$ and therefore $\xi_T=\lambda_T$. It follows that $\xi_{S\times T}\in\mathbb{Z}{\rm Irr}(S\times T)$ and hence $\xi$ is a linear character by \cite[Corollary 8.12]{Isa76}.
Next, we show that $\xi$ extends to $A_P$. To do so, we use \cite[Theorem 6.26]{Isa76}. Let $q$ be a prime dividing $o(\xi)$ and consider $S_q/{\mathbf{C}}_A(P)K_P\in{\rm Syl}_q(A_P/{\mathbf{C}}_A(P)K_P)$. Noticing that every $p$-element $x$ of ${\mathbf{C}}_A(P)K_P$ is contained in ${\rm Ker}(\xi)$, we deduce that $p$ does not divide $|{\mathbf{C}}_A(P)K_P:{\rm Ker}(\xi)|$ and hence $q\neq p$. Let $Q\in{\rm Syl}_q(A_P/N_P)$ such that $S_q={\mathbf{C}}_A(P)K_PQ$ and define $Q_1:=Q\cap {\mathbf{C}}_A(P)K_P$ and $\xi_1:=\xi_{Q_1}$. By \cite[Lemma 4.1]{Spa10}, we deduce that $\xi$ extends to $A_P$ if and only if $\xi_{Q_1}$ extends to $Q$. We are going to check the latter condition. Because $Q_1\leq A_P$ we deduce that $NQ_1$ is a $P$-invariant $p'$-group and that $(NQ_1)_P=N_PQ_1=Q_1$. We also have $KQ_1=(NQ_1)\rtimes P$ and $(KQ_1)_P=K_PQ_1=Q_1P$. Now we can consider the $P$-Glauberman correspondent $f_P(\wt{\mu}_{NQ_1})$ and its canonical extension $f_P(\wt{\mu}_{NQ_1})^\diamond$ to $Q_1P$. By \cite[Theorem A]{Isa-Nav91} we have $f_P(\wt{\mu}_{NQ_1})_{N_P}=f_P(\mu)$ and so $(f_P(\wt{\mu}_{NQ_1})^\diamond)_{K_P}=f_P(\mu)^\diamond$. Using Lemma \ref{lem:Canonical extension}, we obtain an extension $\psi$ of $f_P(\wt{\mu}_{NQ_1})^\diamond$ to $(KQ)_P=K_PQ$. By Gallagher's theorem there exists a unique linear character $\nu\in{\rm Irr}(K_PQ/K_P)$ such that $\chi_{K_PQ}\cdot\nu=\psi$. Finally, for every $x\in Q_1$, we have
\begin{align*}
\xi^{(x)}\chi_{K_P^{(x)}}&=f_P(\wt{\mu}_{N^{(x)}})^\diamond=(f_P(\wt{\mu}_{NQ_1})_{N_P^{(x)}})^\diamond
\\
&=(f_P(\wt{\mu}_{NQ_1})^\diamond)_{K_P^{(x)}}=(\psi_{PQ_1})_{K_P^{(x)}}
\\
&=\psi_{K_P^{(x)}}=\chi_{K_P^{(x)}}\nu_{K_P^{(x)}}
\end{align*}
and it follows that $\nu_{Q_1}=\xi_1$. This shows that $\nu_Q$ is an extension of $\xi_1$ to $Q$ and therefore $\xi$ extends to $S_q$. We conclude that $\xi$ has an extension $\wt{\xi}$ to $A_P$.
Define $\widetilde{f_P(\mu)}:=\wt{\xi}\chi$. By \cite[Theorem 2.6]{Nav-Spa14II} we deduce that
$$\wt{\mu}(x)^*=\wt{\mu}_{N^{(x)}}(x)^*=ef_P(\wt{\mu}_{N^{(x)}})(x)^*=e(\xi(x)\chi(x))^*=e\widetilde{f_P(\mu)}(x)^*,$$
for every $p$-regular $x\in E$ such that $P\in{\rm Syl}_p({\mathbf{C}}_K(x))$.
It remains to show that $C:={\mathbf{C}}_A(K)$ is contained in the kernel of $\xi$. First, observe that ${\rm Ker}(\xi)$ contains $C':=[C,C]\leq {\rm Ker}(\xi_C)$. Moreover, ${\rm Ker}(\xi)$ contains every $p$-element of $C$. Since $C/C'$ is abelian it's enough to show that every $p$-regular element $x$ of $C$ lies in ${\rm Ker}(\xi)$. By the Alperin argument we know that $B:={\rm bl}(\wt{\mu}_{K^{(x)}})$ and $B':={\rm bl}(f_P(\wt{\mu}_{N^{(x)}})^\diamond)$ are Brauer correspondents with $B$ covering $b:={\rm bl}(\wt{\mu}_{N^{(x)}})=\{\wt{\mu}_{N^{(x)}}\}$ and $B'$ covering $b':={\rm bl}(f_P(\wt{\mu}_{N^{(x)}}))=\{f_P(\wt{\mu}_{N^{(x)}})\}$. According to \cite[Theorem 4.14]{Nav98} it follows that $\lambda_B=\lambda_{B'}\circ {\rm Br}_P$. Since $x\in{\mathbf{C}}_A(K)$, we have $x\leq {\mathbf{Z}}(K^{(x)})$ and hence
$$\lambda_B(x)=\lambda_B\left(\left(x^{K^{(x)}}\right)^+\right)=\lambda_{B'}\left(\left(x^{K^{(x)}}\cap {\mathbf{C}}_{K^{(x)}}(P)\right)^+\right)=\lambda_{B'}(x).$$
By \cite[Theorem 9.5]{Nav98}, we conclude that
$$\left(\frac{\wt{\mu}(x)}{\wt{\mu}(1)}\right)^*=\lambda_B(x)=\lambda_{B'}(x)=\left(\frac{f_P(\wt{\mu}_{N^{(x)}})(x)}{f_P(\wt{\mu}_{N^{(x)}})(1)}\right)^*=\left(\frac{\widetilde{f_P(\mu)}(x)}{\widetilde{f_P(\mu)}(1)}\right)^*.$$
As ${\rm Irr}(\wt{\mu}_C)={\rm Irr}(\chi_C)$ and $x$ is $p$-regular, we obtain
$$\dfrac{\chi(x)}{\chi(1)}=\dfrac{\wt{\mu}(x)}{\wt{\mu}(1)}=\dfrac{\widetilde{f_P(\mu)}(x)}{\widetilde{f_P(\mu)}(1)}$$
and, in particular, $\xi(x)=1$. This concludes the proof.
\end{proof}
The following result extends the bijection given in \cite[Theorem 5.13]{Nav-Spa14I} (in the case where $K$ of \cite[Theorem 5.13]{Nav-Spa14I} is a $p'$-group) to characters of positive height. In this particular situation we obtain a canonical bijection.
\begin{prop}
\label{prop:Above the Glauberman correspondence with block relation}
Assume Hypothesis \ref{hyp:Glauberman section}. Then there exists a canonical $A_P$-equivariant bijection
\begin{align*}
\Psi_{\mu,P}:{\rm Irr}(K\mid \mu)&\to{\rm Irr}(K_P\mid f_P(\mu))
\\
\mu^\diamond\nu&\mapsto f_P(\mu)^\diamond\nu_{K_P},
\end{align*}
for every $\nu\in {\rm Irr}(K/N)$. Moreover $\Psi_{\mu,P}$ preserves the defect of characters and
$$(A_\vartheta, K, \vartheta)\iso{K}(A_{P,\vartheta}, K_P, \Psi_{\mu,P}(\vartheta))$$
and
$${\mathbf{C}}_{A}(D)\leq A_P,$$
for every $\vartheta\in{\rm Irr}(K\mid \mu)$ and some defect group $D$ of $\hspace{2pt}{\rm bl}(\Psi_{\mu,P}(\vartheta))$.
\end{prop}
\begin{proof}
Since $K=N\rtimes P$ and $K_P=N_P\times P$ are $p$-nilpotent groups with Sylow $p$-subgroup $P$, $\mu$ is $K$-invariant and $f_P(\mu)$ is $K_P$-invariant, we have ${\rm Irr}(K\mid \mu)={\rm Irr}({\rm bl}(\mu^\diamond))=\{\mu^\diamond\nu\mid \nu\in{\rm Irr}(K/N)\}$ and ${\rm Irr}(K_P\mid f_P(\mu))={\rm Irr}({\rm bl}(f_P(\mu)^\diamond))=\{f_P(\mu)^\diamond\nu\mid \nu\in{\rm Irr}(K_P/N_P)\}$. Thus, we obtain a defect preserving $A_P$-quivariant bijection by setting
$$\Psi_{\mu,P}(\mu^\diamond\nu):=f_P(\mu)^\diamond\nu_{K_P},$$
for every $\nu\in{\rm Irr}(K/N)$.
\begin{comment}
Since ${\rm bl}(\mu^\diamond)$ and ${\rm bl}(f_P(\mu)^\diamond)$ are Brauer first main correspondents over $P$,
\end{comment}
Furthermore, as $P$ is a common defect group of the two blocks ${\rm bl}(\vartheta)$ and ${\rm bl}(\Psi_{\mu,P}(\vartheta))$, the condition on defect groups is satisfied.
Consider $\mathcal{P}\in{\rm Proj}(A\mid \alpha)$ a projective representation associated with $\mu^\diamond$ and observe that $\mathcal{P}$ is also associated with $\mu$. Let $\wh{A}$ be the central extension of $A$ defined by $\mathcal{P}$ and $\epsilon:\wh{A}\to A$ be the map given by $\epsilon(x,s):=x$, for every $x\in A$ and $s\in S$, with kernel $S:=\langle\alpha(x,y)\mid x,y\in A\rangle$. For $H\leq A$, set $\wh{H}:=\epsilon^{-1}(H)$. Since $\alpha$ is constant on $K$-cosets and $\alpha(1,1)=1$, the set $H_0:=\{(h,1)\mid h\in H\}$ is a subgroup of $\wh{A}$, whenever $H\leq K$. In this case let $\vartheta_0\in{\rm Irr}(H_0)$ be the character corresponding to $\vartheta\in{\rm Irr}(H)$ via the isomorphism $\epsilon_{H_0}:H_0\to H$. Moreover $\wh{H}=H_0\times S$ and we define $\wh{\vartheta}:=\vartheta_0\times 1_S\in{\rm Irr}(\wh{H})$. Notice that $(\mu^\diamond)_0\in{\rm Irr}(K_0)$ is the canonical extension of $\mu_0$ and that $\wh{\mu^\diamond}\in{\rm Irr}(\wh{K})$ is the canonical extension of $\wh{\mu}$. Furthermore $f_P(\mu)_0=f_{P_0}(\mu_0)$ and $(f_P(\mu)^\diamond)_0$ is its canonical extension. As no confusion arise, we just write $\mu^\diamond_0$ (resp. $f_P(\mu)^\diamond_0$) instead of $(\mu^\diamond)_0=(\mu_0)^\diamond$ (resp. $(f_P(\mu)^\diamond)_0=f_{P_0}(\mu_0)^\diamond$).
Recall that the map defined by $\wh{\mathcal{P}}(x,s):=s\mathcal{P}(x)$, for all $(x,s)\in \wh{A}$, is an irreducible representation of $\wh{A}$ affording an extension $\tau$ of $\mu^\diamond_0$. Set $S_p:={\mathbf{O}}_p(S)$, $S_{p'}:={\mathbf{O}}_{p'}(S)$, $M:=N_0\times S_{p'}$, $Q:=P_0\times S_p$, and notice that $\wh{K}=M\rtimes Q$, $M_Q=(N_P)_0\times S_{p'}$ and $\wh{K}_Q=\wh{K_P}$. Let $\varphi:=\tau_M\in{\rm Irr}_{\wh{A}}(M)$ and consider its canonical extension $\varphi^\diamond\in{\rm Irr}(\wh{K})$. By Lemma \ref{lem:Canonical extension}, there exists an extension $\wt{\varphi}$ of $\varphi^\diamond$ to $\wh{A}$. Since $\wt{\varphi}_{K_0}$ is an extension of $\mu_0$ with $o(\wt{\varphi}_{K_0})$ dividing $o(\varphi^\diamond)$, we deduce that $\wt{\varphi}_{K_0}=\mu^\diamond_0$. Now, if $\wh{\mathcal{R}}$ is an irreducible representation of $\wh{A}$ affording $\wt{\varphi}$, then $\mathcal{R}(x):=\wh{\mathcal{R}}(x,1)$ defines a projective representation of $A$ associated with $\mu^\diamond$. Replacing $\mathcal{P}$ with $\mathcal{R}$, we may assume that $\tau$ extends $\varphi^\diamond$.
Now, Lemma \ref{lem:Canonical extension and values} yields an extension $\wt{f_Q(\varphi)}$ of $f_Q(\varphi)^\diamond$ to $\wh{A}_Q=\wh{A_P}$ such that
\begin{equation}
\label{eq:Canonical bijection consituents}
{\rm Irr}\left(\wt{\varphi}_{{\mathbf{C}}_{\wh{A}}(\wh{K})}\right)={\rm Irr}\left(\wt{f_Q(\varphi)}_{{\mathbf{C}}_{\wh{A}}(\wh{K})}\right)
\end{equation}
and
\begin{equation}
\label{eq:Canonical bijection values}
\wt{\varphi}(x)^*=e\wt{f_Q(\varphi)}(x)^*,
\end{equation}
for every $p$-regular $x\in \wh{A}$ such that $Q\in{\rm Syl}_p({\mathbf{C}}_{\wh{K}}(x))$, where $e:=[\varphi_{M_Q},f_Q(\varphi)]$. Observe that, by \cite[Theorem A]{Isa-Nav91} and using the fact that $S_p\leq {\mathbf{Z}}(\wh{A})$, we have $\wt{f_Q(\varphi)}_{(N_P)_0}=f_{P_0}(\mu_0)$ and $\wt{f_Q(\varphi)}_{(K_P)_0}=f_{P_0}(\mu_0)^\diamond$.
Let $\wh{\mathcal{P}}'$ be an irreducible representation of $\wh{A_P}$ affording $\wt{f_Q(\varphi)}$ and consider the projective representation $\mathcal{P}'$ of $A_P$ defined by $\mathcal{P}'(x):=\wh{\mathcal{P}}'(x,1)$, for every $x\in A_P$. Notice that $\mathcal{P}'$ is associated with $f_P(\mu)^\diamond$ and that its factor set coincides with $\alpha_{A_P\times A_P}$. Furthermore, as ${\mathbf{C}}_{\wh{A}}(\wh{K})=\wh{{\mathbf{C}}_A(K)}$ and by \eqref{eq:Canonical bijection consituents}, we deduce that $\mathcal{P}_{{\mathbf{C}}_A(K)}$ and $\mathcal{P}'_{{\mathbf{C}}_A(K)}$ are associated with the same scalar function
\begin{comment}
This shows that
$$\left(A,K,\mu^\diamond\right)\geq_c\left(A_P, K_P, f_P(\mu)^\diamond\right).$$
\end{comment}
Next, let $\vartheta=\mu^\diamond\nu\in{\rm Irr}(K\mid \mu)$, with $\nu\in{\rm Irr}(K/N)$, and observe that $A_\vartheta=A_\nu$. Let $\mathcal{Q}$ be a projective representation of $A_\nu$ associated with $\nu$ and notice that $\mathcal{Q}_{A_{P,\nu}}$ is a projective representation of $A_{P,\nu}$ associated with $\nu_{K_P}$. It follows that $\mathcal{S}:=\mathcal{P}_{A_\nu}\otimes \mathcal{Q}$ is a projective representation of $A_\nu$ associated with $\vartheta$, while $\mathcal{S}':=\mathcal{P}'_{A_{P,\nu}}\otimes \mathcal{Q}_{A_{P,\nu}}$ is a projective representation of $A_{P,\nu}$ associated with $\Psi_{\mu,P}(\vartheta)=f_P(\mu)^\diamond\nu_{K_P}$. We claim that $(A_\vartheta, K,\vartheta)\iso{K} (A_{P,\vartheta}, K_P, \Psi_{\mu,P}(\vartheta))$ via $(\mathcal{S},\mathcal{S}')$. By the previous paragraph, one can easily check that the group theoretical conditions hold, that the factor set of $\mathcal{S}'$ coincides with the restriction of the factor set of $\mathcal{S}$ and that $\mathcal{S}_{{\mathbf{C}}_{A_\nu}(K)}$ and $\mathcal{S}'_{{\mathbf{C}}_{A_\nu}(K)}$ are associated with the same scalar function. To conclude, it remains to check the condition on block induction. By the proof of \cite[Theorem 4.4]{Nav-Spa14I} it's enough to show that
$$\left(\frac{|K|_{p'}{\rm tr}(\mathcal{S}(x))}{p^{{\rm ht}(\vartheta)}\vartheta(1)_{p'}}\right)^*=\left(\frac{|K_P|_{p'}{\rm tr}(\mathcal{S}'(x))}{p^{{\rm ht}(\Psi_{\mu,P}(\vartheta))}\Psi_{\mu,P}(\vartheta)(1)_{p'}}\right)^*,$$
for every $p$-regular $x\in A_{P,\vartheta}$ such that $P\in{\rm Syl}_p({\mathbf{C}}_K(x))$. Fix a $p$-regular element $x\in A_{P,\vartheta}$ with $P\in{\rm Syl}_p({\mathbf{C}}_K(x))$. Then $Q\in{\rm Syl}_p({\mathbf{C}}_{\wh{K}}(x,1))$ and \eqref{eq:Canonical bijection values} implies
$${\rm tr}\left(\mathcal{S}(x)\right)^*=\wt{\varphi}(x,1)^*{\rm tr}(\mathcal{Q}(x))^*=\left(e\wt{f_Q(\varphi)}(x,1)\right)^*{\rm tr}(\mathcal{Q}(x))^*=e^*{\rm tr}\left(\mathcal{S}'(x)\right)^*.$$
As $e=[\mu_{N_P},f_P(\mu)]$ and by \cite[Theorem 5.2 (b)]{Nav-Spa14I}, we obtain
$$\vartheta(1)_{p'}=\mu(1)\equiv [\mu_{N_P},f_P(\mu)]|N:N_P|f_P(\mu)(1)\equiv e\frac{|K|_{p'}}{|K_P|_{p'}}\Psi_{\mu,P}(\vartheta)(1)_{p'}\pmod{p}$$
and therefore
$$\left(\frac{|K|_{p'}{\rm tr}(\mathcal{S}(x))}{p^{{\rm ht}(\vartheta)}\vartheta(1)_{p'}}\right)^*=\left(\frac{e|K|_{p'}{\rm tr}(\mathcal{S}'(x))}{\nu(1)\vartheta(1)_{p'}}\right)^*=\left(\frac{|K_P|_{p'}{\rm tr}(\mathcal{S}'(x))}{p^{{\rm ht}(\Psi_{\mu,P}(\vartheta))}\Psi_{\mu,P}(\vartheta)(1)_{p'}}\right)^*.$$
Now the proof is complete.
\end{proof}
As a consequence, applying Proposition \ref{prop:Constructing bijections over bijections} and Proposition \ref{prop:Above the Glauberman correspondence with block relation}, for $N\leq J\leq A$ we obtain a defect preserving $A_{P,J}$-equivariant bijection
$$\Phi:{\rm Irr}(J\mid \mu)\to {\rm Irr}(J_P\mid f_P(\mu))$$
such that
$$\left(A_{J,\chi}, J, \chi\right)\iso{J}\left(A_{J,P,\chi},J_P,\Phi(\chi)\right),$$
for every $\chi\in{\rm Irr}(J\mid \mu)$. Finally, we obtain the main result of this section by considering a normal $p$-chain $\d$ with last term $P$ and $J=NG_\d$.
\begin{theo}
\label{thm:Really above the Glauberman correspondence with block relation with chains}
Let $N\leq G\unlhd A$, with $N\unlhd A$ a $p'$-subgroup, and consider a normal $p$-chain $\d$ of $G$ with final term $P$. Let $\mu\in{\rm Irr}_A(N)$ and $f_P(\mu)\in{\rm Irr}(N_P)$ be its $P$-Glauberman correspondent. Then there exists a defect preserving $A_\d$-equivariant bijection
$$\Phi_{\mu,\d}:{\rm Irr}(NG_\d\mid \mu)\to{\rm Irr}(G_\d\mid f_P(\mu))$$
such that
$$\left(NA_{\d,\chi},NG_\d,\chi\right)\iso{G}\left(A_{\d,\chi},G_\d,\Phi_{\mu,\d}(\chi)\right),$$
for every $\chi\in{\rm Irr}(NG_\d\mid \mu)$.
\end{theo}
\begin{proof}
Let $K:=NP$ and observe that, without loss of generality, we may assume $K\unlhd A$. Now the result follows from Proposition \ref{prop:Constructing bijections over bijections}, Proposition \ref{prop:Above the Glauberman correspondence with block relation} and Lemma \ref{lem:Relation on triple wrt different groups}.
\end{proof}
\section{$N$-block isomorphic character triples and Fong correspondence}
\label{sec:Fong}
In this section, we show that the Fong correspondence \cite{Fon61} can be used to construct $N$-block isomorphic character triples. For completeness, we state the Fong correspondence in the form we need.
\begin{hyp}
\label{hyp:Fong correspondence}
Let $N$ be a normal $p'$-subgroup of $A$ and consider $\mu\in{\rm Irr}_A(N)$. Let $\mathcal{P}\in{{\rm Proj}(A\mid \alpha)}$ be a projective representation associated with $(A,N,\mu)$ such that $\alpha(x,y)^{|N|^2}=1$, for every $x,y\in A$ (see \cite[Theorem 3.5.7]{Nag-Tsu89}), and denote by $\wh{A}$ the $p'$-central extension of $A$ by $S:=\langle\alpha(x,y)\mid x,y\in A\rangle$ defined by $\mathcal{P}$ (see \cite[Section 5.3]{Nav18}). Let $\epsilon:\wh{A}\to A$ be the epimorphism given by $\epsilon(x,s):=x$, for every $x\in A$ and $s\in S$, and consider $N_0:=\{(n,1)\mid n\in N\}\unlhd \wh{A}$. For every $X\leq A$, set $\wh{X}:=\epsilon^{-1}(X)$ and $\wt{X}:=\wh{X}N_0/N_0$. Consider the irreducible representation $\wh{\mathcal{P}}$ of $\wh{A}$ defined by $\wh{\mathcal{P}}(x,s):=s\mathcal{P}(x)$, for every $x\in A$ and $s\in S$, and denote its character by $\tau$. Let $\wh{\lambda}\in{\rm Irr}(\wh{N})$ be the linear character defined by $\wh{\lambda}(n,s):=s^{-1}$, for every $n\in N$ and $s\in S$, and set $\wh{\mu}:=\mu_0\times 1_S\in{\rm Irr}(\wh{N})$, where $\mu_0$ correspond to $\mu$ via the isomorphism $N\simeq N_0$. Notice that $\tau$ extends $\wh{\mu}\wh{\lambda}^{-1}$. Finally, denote by $\wt{\mu}$ the character $\wh{\lambda}$ viewed as a character of $\wt{N}=\wh{N}/N_0$, that is $\wt{\mu}(N_0(n,s))=s^{-1}$, for every $n\in N$ and $s\in S$.
\end{hyp}
\begin{theo}[Fong]
\label{thm:Fong correspondence}
Assume Hypothesis \ref{hyp:Fong correspondence}. If $N\leq H\leq A$, then:
\begin{enumerate}
\item $\wt{H}$ is a $p'$-central extension of $H/N$ by the central $p'$-subgroup $\wt{N}\simeq S$;
\item There exists a bijection
\begin{align*}
{\rm Bl}(H\mid {\rm bl}(\mu))&\to {\rm Bl}(\wt{H}\mid {\rm bl}(\wt{\mu}))
\\
B &\mapsto \wt{B}
\end{align*}
\item Let $D\in\delta(B)$ and consider $Q\in{\rm Syl}_p(\wh{D})$ so that $\wh{D}=Q\times S$. Then $QN_0/N_0\in\delta(\wt{B})$. In particular $B$ and $\wt{B}$ have isomorphic defect groups;
\item For every $B\in{\rm Bl}(H\mid {\rm bl}(\mu))$ corresponding to $\wt{B}\in{\rm Bl}(\wt{H}\mid {\rm bl}(\wt{\mu}))$ via the bijection in (ii), there exists a defect preserving bijection
\begin{align*}
{\rm Irr}(B)&\to {\rm Irr}(\wt{B})
\\
\psi &\mapsto \wt{\psi}
\end{align*}
such that, if $\wh{\psi}$ is the inflation to $\wh{H}$ of the character of $\wh{H}/S\simeq H$ corresponding to $\psi$ and $\wt{\psi}'$ is the inflation to $\wh{H}$ of $\wt{\psi}$, then $\wh{\psi}=\tau_{\wh{H}}\wt{\psi}'$;
\item For $\wh{x}\in \wh{A}$ set $x:=\epsilon(\wh{x})$ and $\wt{x}:=N_0\wh{x}$. Then $\wt{\psi^x}=\left(\wt{\psi}\right)^{\wt{x}}$ and $\wt{B^x}=\left(\wt{B}\right)^{\wt{x}}$, for every $B\in{{\rm Bl}(G\mid {\rm bl}(\mu))}$ and $\psi\in{\rm Irr}(B)$.
\end{enumerate}
\end{theo}
\begin{proof}
Consider $\psi\in{\rm Irr}(H\mid \mu)$ afforded by $\mathfrak{X}$. We just show how to construct $\wt{\psi}$. By \cite[Theorem 10.11]{Nav18}, there exists an irreducible projective representation $\mathcal{Q}\in{\rm Proj}(H/N\mid \alpha_{H\times H}^{-1})$ such that $\mathfrak{X}=\mathcal{Q}\otimes \mathcal{P}_H$ unique up to similarity. Now, $\wh{\mathcal{Q}}(x,s):=\mathcal{Q}(x)s^{-1}$, for every $x\in H$ and $s\in S$, defines an irreducible linear representation of $\wh{H}$ with $N_0\leq {\rm Ker}(\wh{\mathcal{Q}})$ and whose character lies over $\wh{\lambda}$. If we consider the inflation $\wh{\mathfrak{X}}$ to $\wh{H}$ of the representation of $\wh{H}/S\simeq H$ corresponding to $\mathfrak{X}$, that is $\wh{\mathfrak{X}}(\wh{x}):=\mathfrak{X}(\epsilon(\wh{x}))$, for every $\wh{x}\in \wh{H}$, then $\wh{\mathfrak{X}}=\wh{\mathcal{Q}}\otimes \wh{\mathcal{P}}_{\wh{H}}$. Define $\wt{\mathfrak{X}}$ to be the irreducible representation of $\wt{H}=\wh{H}/N_0$ whose inflation is $\wh{\mathcal{Q}}$, and let $\wt{\psi}$ be the character afforded by $\wt{\mathfrak{X}}$. Then $\wt{\psi}\in{\rm Irr}(\wt{A}\mid \wt{\mu})$ and, if $\wh{\psi}$ is the inflation to $\wh{H}$ of the character of $\wh{H}/S\simeq H$ corresponding to $\psi$ and $\wt{\psi}'$ is the inflation of $\wt{\psi}$ to $\wh{H}$, then $\wh{\psi}=\tau_{\wh{H}}\wt{\psi}'$. The result follows from \cite{Fon61}. The description of defect groups is a consequence of the proof of \cite[2C]{Fon61}. To conclude, for $\wh{x}\in \wh{A}$, set $x:=\epsilon(\wh{x})$ and $\wt{x}:=N_0\wh{x}$. Then $\wh{\psi^x}=(\wh{\psi})^{\wh{x}}=(\wt{\psi}'\tau_{\wh{H}})^{\wh{x}}=(\wt{\psi}')^{\wh{x}}\tau_{\wh{H}}=(\wt{\psi}^{\wt{x}})'\tau_{\wh{H}}$, where $\wh{\psi^x}$ is the inflation to $\wh{H}$ of the character of $\wh{H}/S\simeq H$ corresponding to $\psi^x$ and $(\wt{\psi}^{\wt{x}})'$ is the inflation of $\wt{\psi}^{\wt{x}}$ to $\wh{H}$. Thus $(\wt{\psi})^{\wt{x}}$ coincides with $\wt{\psi^x}$ the Fong correspondent of $\psi^x$. In particular, since $\wt{\psi^x}\in {\rm Irr}(\wt{B^x})$ and $(\wt{\psi})^{\wt{x}}\in{\rm Irr}(\wt{B}^{\wt{x}})$, we conclude that $\wt{B^x}=\wt{B}^{\wt{x}}$.
\end{proof}
In the situation of Theorem \ref{thm:Fong correspondence}, we refer to $\wt{B}$ as the Fong correspondent of $B$ and to $\wt{\psi}$ as the Fong correspondent of $\psi$. An important feature of the Fong correspondence is that it is compatible with block induction.
\begin{prop}
\label{prop:Fong correspondence and block induction}
Assume Hypothesis \ref{hyp:Fong correspondence} and let $N\leq X\leq Y\leq A$. Let $b\in{\rm Bl}(X\mid {\rm bl}(\mu))$ with Fong correspondent $\wt{b}\in{\rm Bl}(\wt{X}\mid {\rm bl}(\wt{\mu}))$ and suppose that the induced blocks $b^Y$ and $(\wt{b})^{\wt{Y}}$ are defined. Then $\wt{b^Y}=(\wt{b})^{\wt{Y}}$.
\end{prop}
\begin{proof}
This result has been shown in \cite{Rob00}. It can also be deduced from \cite[Theorem 14.3]{Dad94}.
\end{proof}
For $x\in G$, we denote by ${\mathfrak{Cl}}_G(x)$ the $G$-conjugacy class of $x$. Moreover, for any subset $K$ of $G$, we denote by $K^+$ the sum of its elements in the group algebra of $G$.
\begin{theo}
\label{thm:Fong correspondence and character triple relations}
Assume Hypothesis \ref{hyp:Fong correspondence}. For $i=1,2$, consider $N\leq L_i\unlhd H_i\leq A$ and a $H_i$-invariant $\psi_i\in{\rm Irr}(L_i\mid \mu)$. Notice that $\wt{L}_i\unlhd \wt{H}_i$ and that the Fong correspondent $\wt{\psi}_i\in{\rm Irr}(\wt{L}_i\mid \wt{\mu})$ is $\wt{H}_i$-invariant. Let $L_i\leq G\unlhd A$ and assume
$$\left(\wt{H}_1,\wt{L}_1,\wt{\psi}_1\right)\iso{\wt{G}}\left(\wt{H}_2,\wt{L}_2,\wt{\psi}_2\right).$$
Then
$$\left(H_1,L_1,\psi_1\right)\iso{G}\left(H_2,L_2,\psi_2\right).
$$
\end{theo}
\begin{proof}
The group theoretical conditions are clearly satisfied and without loss of generality we may assume $A=GH_i$, $\wh{A}=\wh{G}\wh{H}_i$ and $\wt{A}=\wt{G}\wt{H}_i$. Consider $B_i:={\rm bl}(\psi_i)$ and its Fong correspondent $\wt{B}_i={\rm bl}(\wt{\psi}_i)$. By hypothesis, there exists a defect group $D_i\in\delta(\wt{B}_i)$ such that ${\mathbf{C}}_{\wt{A}}(D_i)\leq \wt{H}_i$. Furthermore, by Theorem \ref{thm:Fong correspondence} (iii) we can find a defect group $P_i\in\delta(B_i)$ such that, if $Q_i\in{\rm Syl}_p(\wh{P}_i)$, then $D_i=Q_iN_0/N_0$. In particular
\begin{equation}
\label{eq:Fong 1}
{\mathbf{C}}_{\wh{A}}(Q_i)\leq \wh{H}_i
\end{equation}
and, noticing that
$$\epsilon\left({\mathbf{C}}_{\wh{A}}(Q_i)\right)={\mathbf{C}}_A(P_i),$$
we obtain ${\mathbf{C}}_A(P_i)\leq H_i$. Fix projective representations $(\wt{\mathcal{R}}_1,\wt{\mathcal{R}}_2)$ associated with $\left(\wt{H}_1,\wt{L}_1,\wt{\psi}_1\right)\iso{\wt{G}}\left(\wt{H}_2,\wt{L}_2,\wt{\psi}_2\right)$ and let $\wt{\alpha}_i$ be the factor set of $\wt{\mathcal{R}}_i$. Consider a projective representation $\mathcal{R}_i\in{\rm Proj}(H_i\mid \alpha_i)$ associated with $\psi_i$ and define the projective representation $\wh{\mathcal{R}}_i\in{\rm Proj}(\wh{H}_i\mid \wh{\alpha}_i)$ given by
$$\wh{\mathcal{R}}_i(h):=\mathcal{R}_i(\epsilon(h)),$$
for every $h\in \wh{H}_i$. Notice that $\wh{\alpha}_i(h,k)=\alpha_i(\epsilon(h),\epsilon(k))$, for all $h,k\in \wh{H}_i$, and that $\wh{\mathcal{R}}_i$ is associated with $\wh{\psi}_i$. Let $\wt{\mathcal{R}}_i'\in{\rm Proj}(\wh{H}_i\mid \wt{\alpha}'_i)$ be the projective representation defined by
$$\wt{\mathcal{R}}_i'(h):=\wt{\mathcal{R}}_i(N_0h),$$
for every $h\in \wh{H}_i$. Clearly $\wt{\alpha}_i'(h,k)=\wt{\alpha}_i(N_0h,N_0k)$, for all $h,k\in \wh{H}_i$, and $\wt{\mathcal{R}}_i'$ is associated with $\wt{\psi}_i'$. As $\wh{\mathcal{R}}_i$ and $\wh{\mathcal{P}}_{\wh{H}_i}\otimes \wt{\mathcal{R}}_i'$ are a projective representations of $\wh{H}_i$ associated with $\tau_{\wh{L}_i}\wt{\psi}_i'=\wh{\psi}_i$, there exists a map $\wh{\xi}_i:\wh{H}_i/\wh{L}_i\to\mathbb{C^\times}$ such that $\wh{\xi}_i\wh{\mathcal{R}}_i=\wh{\mathcal{P}}_{\wh{H}_i}\otimes \wt{\mathcal{R}}_i'$. Let $\xi_i:H_i/L_i\to \mathbb{C}^\times$ corresponds to $\wh{\xi}_i$ via the isomorphism $H_i/L_i\simeq \wh{H}_i/\wh{L}_i$. Replacing $\mathcal{R}_i$ with $\xi_i\mathcal{R}_i$, we may assume
\begin{equation}
\label{eq:Fong 2}
\wh{\mathcal{R}}_i=\wh{\mathcal{P}}_{\wh{H}_i}\otimes \wt{\mathcal{R}}_i'.
\end{equation}
Now, as the factor sets $\wt{\alpha}_1$ and $\wt{\alpha}_2$ coincide under the isomorphism $\wt{H}_1/\wt{L}_1\simeq \wt{H}_2/\wt{L}_2$, we deduce that $\alpha_1$ and $\alpha_2$ coincide under the isomorphism $H_1/L_1\simeq H_2/L_2$. By hypothesis $\wt{\mathcal{R}}_1$ and $\wt{\mathcal{R}}_2$ define the same scalar function on ${\mathbf{C}}_{\wt{A}}(\wt{G})$. As ${\mathbf{C}}_{\wh{A}}(\wh{G})N_0/N_0\leq {\mathbf{C}}_{\wt{A}}(\wt{G})$ and ${\mathbf{C}}_{\wh{A}}(\wh{G})\leq \wh{H}_1\cap \wh{H}_2$ by \eqref{eq:Fong 1}, the scalar functions defined by $\wt{\mathcal{R}}'_1$ and $\wt{\mathcal{R}}'_2$ on ${\mathbf{C}}_{\wh{A}}(\wh{G})$ coincide. Now $\wh{\mathcal{R}}_{1,{\mathbf{C}}_{\wh{A}}(\wh{G})}$ and $\wh{\mathcal{R}}_{2,{\mathbf{C}}_{\wh{A}}(\wh{G})}$ are associated with the same scalar function and, since $\epsilon({\mathbf{C}}_{\wh{A}}(\wh{G}))={\mathbf{C}}_A(G)$ (see \cite[Theorem 4.1 (d)]{Nav-Spa14I}), the same is true for $\mathcal{R}_{1,{\mathbf{C}}_A(G)}$ and $\mathcal{R}_{2,{\mathbf{C}}_A(G)}$.
Next, consider $G\leq J\leq A$ and set $J_i:=J\cap H_i$. Notice that, if $\chi\in{\rm Irr}(J_1\mid \psi_1)$, then Theorem \ref{thm:Fong correspondence} (iv) implies that $\wt{\chi}\in{\rm Irr}(\wt{J}_1\mid \wt{\psi}_1)$. Write $\chi={\rm tr}(\mathcal{Q}_{J_1}\otimes \mathcal{R}_{1,J_1})$, for some $\mathcal{Q}\in{\rm Proj}(J/G)$. If we set $\wh{Q}(x):=\mathcal{Q}(\epsilon(x))$ for every $x\in \wh{J}$, then \eqref{eq:Fong 2} implies
\begin{align*}
\wh{\chi}_1&={\rm tr}\left(\wh{\mathcal{Q}}_{\wh{J}_1}\otimes \wh{\mathcal{R}}_{1,\wh{J}_1}\right)
\\
&={\rm tr}\left(\wh{\mathcal{Q}}_{\wh{J}_1}\otimes \wt{\mathcal{R}}'_{1,\wh{J}_1} \otimes \wh{\mathcal{P}}_{\wh{J}_1}\right)
\end{align*}
and therefore $\wt{\chi}'={\rm tr}(\wh{\mathcal{Q}}_{\wh{J}_1}\otimes \wt{\mathcal{R}}'_{1,\wh{J}_1})$. Now, let $\wt{\mathcal{Q}}\in{\rm Proj}(\wt{J}/\wt{G})$ correspond to $\wh{\mathcal{Q}}$ via the isomorphism $\wt{J}/\wt{G}\simeq \wh{J}/\wh{G}$ and observe that the Fong correspondent of $\chi$ can be written as $\wt{\chi}={\rm tr}(\wt{\mathcal{Q}}_{\wt{J}_1}\otimes \wt{\mathcal{R}}_{1,\wt{J}_1})$. By definition $\wt{\sigma}_{\wt{J}_1}(\wt{\chi})={\rm tr}(\wt{\mathcal{Q}}_{\wt{J}_2}\otimes \wt{\mathcal{R}}_{2,\wt{J}_2})$ so that its inflation $\wt{\sigma}_{\wt{J}_1}(\wt{\chi})'={\rm tr}(\wh{\mathcal{Q}}_{\wh{J}_2}\otimes \wt{\mathcal{R}}'_{2,\wh{J}_2})$. By Theorem \ref{thm:Fong correspondence} (iv) and \eqref{eq:Fong 2} we obtain
\begin{align*}
\tau_{\wh{J}_2}\wt{\sigma}_{\wt{J}_1}(\wt{\chi})'&={\rm tr}\left(\wh{\mathcal{P}}_{\wh{J}_2}\otimes\wh{\mathcal{Q}}_{\wh{J}_2}\otimes \wt{\mathcal{R}}'_{2,\wh{J}_2}\right)
\\
&={\rm tr}\left(\wh{\mathcal{Q}}_{\wh{J}_2}\otimes \wh{\mathcal{R}}_{2,\wh{J}_2}\right)
\\
&=\wh{\sigma_{J_1}(\chi)}
\\
&=\tau_{\wh{J}_2}\wt{\sigma_{J_1}(\chi)}'
\end{align*}
and therefore
$$\wt{\sigma_{J_1}(\chi)}=\wt{\sigma}_{\wt{J}_1}\left(\wt{\chi}\right).$$
Since by hypothesis ${\rm bl}(\wt{\sigma_{J_1}(\chi)})^{\wt{J}}={\rm bl}(\wt{\chi})^{\wt{J}}$, we conclude from Proposition \ref{prop:Fong correspondence and block induction} that ${\rm bl}(\sigma_{J_1}(\chi))^J={\rm bl}(\chi)^J$. This completes the proof.
\end{proof}
From now on we consider $N\leq G\unlhd A$. Since $\wt{N}$ is a central $p'$-subgroup of $\wt{G}$, for every $p$-subgroup $P$ of $G$ we have a decomposition $\wt{P}=\wt{N}\times {\mathbf{O}}_p(\wt{P})$. We write $\wt{P}_p:={\mathbf{O}}_p(\wt{P})$. Mapping $P$ to $\wt{P}_p$ induces a length preserving bijection
\begin{align}
\mathfrak{N}(G,Z)/G &\to\mathfrak{N}(\wt{G},\wt{Z}_p)/\wt{G} \label{eq:Bijection on chains}
\\
\d &\mapsto \wt{\d} \nonumber
\end{align}
which commutes with the action of $A$ and $\wt{A}$. In particular, observe that $\wt{NG_\d}=\wt{G}_{\wt{\d}}$. Using Theorem \ref{thm:Really above the Glauberman correspondence with block relation with chains}, Theorem \ref{thm:Fong correspondence} and Theorem \ref{thm:Fong correspondence and character triple relations} we obtain the following corollaries.
\begin{cor}
\label{cor:Fong correspondence and Glauberman bijection}
Assume Hypothesis \ref{hyp:Fong correspondence} and let $N\leq G\unlhd A$. Consider a normal $p$-chain $\d$ of $G$ with final term $P$ and let $f_P(\mu)\in{\rm Irr}(N_P)$ be the $P$-Glauberman correspondent of $\mu$. Then there exists a defect preserving bijection
$$\Gamma_{\mu,\d}:{\rm Irr}\left(G_\d\enspace\middle|\enspace f_P(\mu)\right)\to{\rm Irr}\left(\wt{G}_{\wt{\d}}\enspace\middle|\enspace \wt{\mu}\right)$$
commuting with the action of $A$ and $\wt{A}$.
\end{cor}
\begin{proof}
This follows immediately by Theorem \ref{thm:Really above the Glauberman correspondence with block relation with chains} and Theorem \ref{thm:Fong correspondence}.
\end{proof}
The bijections described in the previous corollary are compatible with the relation $\iso{N}$ from Definition \ref{def:N-block isomorphic character triples}.
\begin{cor}
\label{cor:Fong correspondence and Glauberman bijection with character triple relations}
Assume Hypothesis \ref{hyp:Fong correspondence} and let $N\leq G\unlhd A$. Consider normal $p$-chains $\d$ and $\e$ of $G$ with final term respectively $P$ and $Q$ and let $\Gamma_{\mu,\d}$ and $\Gamma_{\mu,\e}$ be the corresponding bijections given by Corollary \ref{cor:Fong correspondence and Glauberman bijection}. Let $\vartheta\in{\rm Irr}(G_\d\mid f_P(\mu))$ and $\chi\in{\rm Irr}(G_\e\mid f_Q(\mu))$ and suppose that
$$\left(\wt{A}_{\wt{\d},\Gamma_{\mu,\d}(\vartheta)},\wt{G}_{\wt{\d}},\Gamma_{\mu,\d}(\vartheta)\right)\iso{\wt{G}}\left(\wt{A}_{\wt{\e},\Gamma_{\mu,\e}(\chi)},\wt{G}_{\wt{\e}},\Gamma_{\mu,\e}(\chi)\right).$$
Then
$$\left(A_{\d,\vartheta},G_\d,\vartheta\right)\iso{G}\left(A_{\e,\chi},G_\e,\chi\right).$$
\end{cor}
\begin{proof}
This is a consequence of Theorem \ref{thm:Really above the Glauberman correspondence with block relation with chains}, Theorem \ref{thm:Fong correspondence and character triple relations} and Corollary \ref{cor:Fong correspondence and Glauberman bijection}.
\end{proof}
The result that we are actually going to need in the final proof is the following. This is obtained by putting together all the results obtained so far.
\begin{cor}
\label{cor:CTC for p-solvable in the stable case}
Assume Hypothesis \ref{hyp:Fong correspondence} and let $N\leq G\unlhd A$. Let $Z$ be a central $p$-subgroup of $G$ and consider a block $B\in{\rm Bl}(G\mid {\rm bl}(\mu))$ whose defect groups are larger than $Z$. Then the Fong correspondent $\wt{B}\in{\rm Bl}(\wt{G})$ has defect groups larger than $\wt{Z}_p$ and there exists a bijection
$$\Delta:{\mathcal{C}}^d(B,Z)/G\to{\mathcal{C}}^d(\wt{B},\wt{Z}_p)/\wt{G}$$
that preserves the length of the $p$-chains, commutes with the ation of $A$ and $\wt{A}$ and such that, if
$$\left(\wt{A}_{(\wt{\d},\wt{\vartheta})}, \wt{G}_{\wt{\d}}, \wt{\vartheta}\right)\iso{\wt{G}}\left(\wt{A}_{(\wt{\e},\wt{\chi})}, \wt{G}_{\wt{\d}}, \wt{\chi}\right),$$
then
$$\left(A_{(\d,\vartheta)},G_{\d},\vartheta\right)\iso{G}\left(A_{(\e,\chi)},G_{\e},\chi\right),$$
for every $(\d,\vartheta), (\e,\chi)\in{\mathcal{C}}^d(B,Z)$, $(\wt{\d},\wt{\vartheta})\in\Delta(\overline{(\d,\vartheta)})$ and $(\wt{\e},\wt{\chi})\in\Delta(\overline{(\e,\chi)})$.
\end{cor}
\begin{proof}
Let $\d\in\mathfrak{N}(G,Z)$ with last term $P$ and consider $\wt{\d}\in\mathfrak{N}(\wt{G},\wt{Z}_p)$. If $\vartheta\in{\rm Irr}(G_\d)$ and ${\rm bl}(\vartheta)^G=B$, then $\vartheta$ lies over $f_P(\mu)$ by Corollary \ref{cor:Harris-Knorr and p-chains}. Now, there exists a unique $\psi\in{\rm Irr}(NG_\d\mid \mu)$ such that $\vartheta=\Phi_{\mu,\d}(\psi)$ and $\Gamma_{\mu,\d}(\vartheta)=\wt{\psi}$ is the Fong correspondent of $\psi$. By Theorem \ref{thm:Really above the Glauberman correspondence with block relation with chains}, we know that ${\rm bl}(\vartheta)^{NG_\d}={\rm bl}(\psi)$, hence ${\rm bl}(\vartheta)^G=B$ if and only if ${\rm bl}(\psi)^G=B$. Furthermore, by Proposition \ref{prop:Fong correspondence and block induction} it follows that ${\rm bl}(\psi)^G=B$ if and only if ${\rm bl}(\wt{\psi})^{\wt{G}}=\wt{B}$. This shows that the set of characters of $G_\d$ whose block induces to $B$ is mapped via $\Gamma_{\mu,\d}$ to the set of characters of $\wt{G}_{\wt{\d}}$ whose block induces to $\wt{B}$. We define
$$\Delta\left(\overline{(\d,\vartheta)}\right):=\overline{\left(\wt{\d},\Gamma_{\mu,\d}(\vartheta)\right)},$$
for every $(\d,\vartheta)\in{\mathcal{C}}^d(B,Z)$. By \eqref{eq:Bijection on chains}, Corollary \ref{cor:Fong correspondence and Glauberman bijection} and Corollary \ref{cor:Fong correspondence and Glauberman bijection with character triple relations} we conclude that $\Delta$ is a bijection with the required properties.
\end{proof}
\section{Structure of a minimal counterexample}
In this section, we finally prove Sp\"ath's Character Triple Conjecture for $p$-solvable groups. Our proof is inspired by the argument developed in \cite{Rob00}. As in Robinson's work, what we are actually going to show is that a minimal counterexample $G$ to Conjecture \ref{conj:Character Triple Conjecture} satisfies ${\mathbf{O}}_p(G){\mathbf{O}}_{p'}(G)\leq {\mathbf{Z}}(G)$. Since the conjecture trivially holds for abelian groups, Theorem \ref{thm:Main theorem} will then follow as a corollary of (the proof of) Theorem \ref{thm:Minimal counterexample structure}. In this paper we consider subpairs in the sense of \cite{Ols82}, i.e. pairs $(P,b_P)$, where $P$ is a $p$-subgroup of $G$ and $b_P\in{\rm Bl}(P{\mathbf{C}}_G(P))$.
\begin{prop}
\label{prop:Minimal counterexample structure, unstable}
Assume that $G\unlhd A$ is a minimal counterexample to Conjecture \ref{conj:Character Triple Conjecture} with respect to $|G:{\mathbf{Z}}(G)|$ first and then to $|A|$ and consider $Z\leq {\mathbf{Z}}(G)$, $B\in{\rm Bl}(G)$ and $d\geq 0$ for which the conjecture fails to hold. Then every block $b\in{\rm Bl}({\mathbf{O}}_{p'}(G))$ covered by $B$ is $A$-invariant.
\end{prop}
\begin{proof}
Set $N:={\mathbf{O}}_{p'}(G)$ and fix a block ${\rm bl}(\mu)\in{\rm Bl}(N)$ covered by $B$. For every subgroup $H\leq A$, set $H^\vee:=H_\mu$. Let $B^\vee\in{\rm Bl}(G^\vee\mid {\rm bl}(\mu))$ be the Fong--Reynolds correspondent of $B$ over ${\rm bl}(\mu)$ \cite[Theorem 9.14]{Nav98}. Since $B$ and $B^\vee$ have a common defect group $D\leq G^\vee$, by \cite[Theorem 3.10]{Alp-Bro79} or \cite[Theorem 2.1]{Ols82} we can find a $B^\vee$-Sylow subpair $(D,b_D^\vee)$ such that $(D,b_D)$ is a $B$-Sylow subpair, where $b_D:=(b_D^\vee)^{D{\mathbf{C}}_G(D)}$. Notice that, by using Corollary \ref{cor:Harris-Knorr} together with the theory of subpairs, the block $b_D^\vee$ covers ${\rm bl}(f_D(\mu))$. Hence $b_D$ covers ${\rm bl}(f_D(\mu))$. More generally, the block $b_Q$ covers ${\rm bl}(f_Q(\mu))$, for every $B$-subpair $(Q,b_Q)\leq (D,b_D)$. Using this observation, we can construct an $A$-transversal $\mathbb{T}$ in ${\mathcal{C}}^d(B,Z)$ such that $P\leq G^\vee$ and $\vartheta\in{\rm Irr}(G_\d\mid f_P(\mu))$, for every $(\d,\vartheta)\in\mathbb{T}$ with $P$ the last term of $\d$.
Consider $(\d,\vartheta)\in\mathbb{T}$ and let $P$ be the last term of $\d$. Notice that $G_\d^\vee=G_{\d,f_P(\mu)}$ and let $\vartheta^\vee\in{\rm Irr}(G^\vee_\d\mid f_P(\mu))$ be the Clifford correspondent of $\vartheta$ over $f_P(\mu)$. As $A=GA^\vee$, we obtain an $A^\vee$-equivariant bijection
$$\Upsilon:{\mathcal{C}}^d(B,Z)/G\to {\mathcal{C}}^d(B^\vee,Z)/G^\vee$$
by defining $\Upsilon(\overline{(\d,\vartheta)}^y):=\overline{(\d,\vartheta^\vee)}^y$, for every $(\d,\vartheta)\in\mathbb{T}$ and $y\in A^\vee$. Since ${|G^\vee:{\mathbf{Z}}(G^\vee)|}\leq {|G:{\mathbf{Z}}(G)|}$, if $\mu$ is not $A$-invariant, then there exists an $A^\vee$-equivariant bijection
$$\Omega^\vee:\mathcal{C}^d(B^\vee,Z)_+/G^\vee\to \mathcal{C}^d(B^\vee,Z)_-/G^\vee$$
such that
$$\left(A^\vee_{(\d,\vartheta^\vee)},G^\vee_{\d},\vartheta^\vee\right)\iso{G^\vee}\left(A^\vee_{(\e,\chi^\vee)}, G^\vee_{\e},\chi^\vee\right),$$
for every $(\d,\vartheta^\vee)\in{\mathcal{C}}^d(B^\vee,Z)_+$ and $(\e,\chi^\vee)\in\Omega^\vee(\overline{(\d,\vartheta^\vee)})$. Combining $\Omega^\vee$ with $\Upsilon$ and applying Proposition \ref{prop:Irreducible induction and double relation block}, we obtain an $A$-equivariant bijection
$$\Omega:{\mathcal{C}}^d(B,Z)_+/G \to {\mathcal{C}}^d(B,Z)_-/G$$
such that
$$\left(A_{(\d,\vartheta)},G_\d,\vartheta\right)\iso{G}\left(A_{(\e,\chi)},G_{\e},\chi\right),$$
for every $(\d,\vartheta)\in{\mathcal{C}}^d(B,Z)_+$ and $(\e,\chi)\in\Omega(\overline{(\d,\vartheta)})$. This is a contradiction and therefore $\mu$ must be $A$-invariant.
\end{proof}
\begin{theo}
\label{thm:Minimal counterexample structure}
Assume that $G\unlhd A$ is a minimal counterexample to Conjecture \ref{conj:Character Triple Conjecture} with respect to $|G:{\mathbf{Z}}(G)|$ first and then to $|A|$ and consider $Z\leq {\mathbf{Z}}(G)$, $B\in{\rm Bl}(G)$ and $d\geq 0$ for which the conjecture fails to hold. Then ${\mathbf{O}}_p(G){\mathbf{O}}_{p'}(G)\leq {\mathbf{Z}}(G)$.
\end{theo}
\begin{proof}
Set $N:={\mathbf{O}}_{p'}(G)$ and fix a block ${\rm bl}(\mu)\in{\rm Bl}(N)$ covered by $B$. Notice that by Lemma \ref{lem:Knorr-Robinson for Character Triple Conjecture} we must have $Z={\mathbf{O}}_p(G)$. Thus it's enough to show that $N$ is contained in the center. By Proposition \ref{prop:Minimal counterexample structure, unstable}, we know that $\mu$ is $A$-invariant and therefore we can apply the results obtained in Section \ref{sec:Fong}. Let $\wt{B}\in{\rm Bl}(\wt{G})$ be the Fong correspondent of $B$. Since $\wt{N}\leq {\mathbf{Z}}(\wt{G})$, if $N\nleq {\mathbf{Z}}(G)$, then ${|\wt{G}:{\mathbf{Z}}(\wt{G})|}\leq {|\wt{G}:\wt{N{\mathbf{Z}}(G)}|}={|G:N{\mathbf{Z}}(G)|}<{|G:{\mathbf{Z}}(G)|}$ and we obtain an $\wt{A}$-equivariant bijection
$$\wt{\Omega}:{\mathcal{C}}^d(\wt{B},\wt{Z}_p)_+/\wt{G}\to {\mathcal{C}}^d(\wt{B},\wt{Z}_p)_-/\wt{G}$$
such that
$$\left(\wt{A}_{(\wt{\d},\wt{\vartheta})}, \wt{G}_{\wt{\d}}, \wt{\vartheta}\right)\iso{\wt{G}}\left(\wt{A}_{(\wt{\e},\wt{\chi})}, \wt{G}_{\wt{\d}}, \wt{\chi}\right),$$
for every $(\wt{\d},\wt{\vartheta})\in{\mathcal{C}}^d(\wt{B},\wt{Z}_p)_+$ and $(\wt{\e},\wt{\chi})\in\wt{\Omega}(\overline{(\wt{\d},\wt{\vartheta})})$. Combining $\wt{\Omega}$ with the bijection $\Delta$ given by Corollary \ref{cor:CTC for p-solvable in the stable case}, we obtain an $A$-equivariant bijection
$$\Omega:\mathcal{C}^d(B,Z)_+/G\to\mathcal{C}^d(B,Z)_-/G$$
such that
$$\left(A_{(\d,\vartheta)},G_\d,\vartheta\right)\iso{G}\left(A_{(\e,\chi)},G_{\e},\chi\right),$$
for every $(\d,\vartheta)\in\mathcal{C}^d(B,Z)_+$ and $(\e,\chi)\in\Omega(\overline{(\d,\vartheta)})$. This contradiction shows that $N$ must be contained in the center of $G$.
\end{proof}
Next, we consider the residue of characters. We are going to obtain Theorem \ref{thm:Main theorem residues} as a consequence of an analogous study of a minimal counterexample. We need the following result whose proof can be deduced by Glesser's paper \cite{Gle07}.
\begin{lem}
\label{lem:Residues}
Let $\Gamma_{\mu,\d}$ be the bijection of Corollary \ref{cor:Fong correspondence and Glauberman bijection}. Then
$$r\left(\Gamma_{\mu,\d}(\vartheta)\right)|N|\equiv \pm\mu(1)r(\vartheta)|\wt{N}|\pmod{p},$$
for every $\vartheta\in{\rm Irr}(G_\d\mid f_P(\mu))$.
\end{lem}
\begin{proof}
This follows from similar computations as the ones in the proofs of \cite[Corollary 3.4 and Theorem 3.8]{Gle07}.
\end{proof}
For completeness, we state the Isaacs-Navarro refinement of the Character Triple Conjecture.
\begin{conj}[Isaacs-Navarro refinement of the Character Triple Conjecture]
\label{conj:Character Triple Conjecture with residues}
There exists a bijection $\Omega$ as in Conjecture \ref{conj:Character Triple Conjecture} such that
$$r(\vartheta)\equiv\pm r(\chi)\pmod{p},$$
for every $(\d,\vartheta)\in{\mathcal{C}}^d(B,Z)_+$ and $(\e,\chi)\in\Omega(\overline{(\d,\vartheta)})$.
\end{conj}
Finally, using the proof of Theorem \ref{thm:Minimal counterexample structure} and Lemma \ref{lem:Residues} we obtain a similar structure theorem for a minimal counterexample of Conjecture \ref{conj:Character Triple Conjecture with residues}.
\begin{theo}
\label{thm:Minimal counterexample structure with residues}
Assume that $G\unlhd A$ is a minimal counterexample to Conjecture \ref{conj:Character Triple Conjecture with residues} with respect to $|G:{\mathbf{Z}}(G)|$ first and then to $|A|$ and consider $Z\leq {\mathbf{Z}}(G)$, $B\in{\rm Bl}(G)$ and $d\geq 0$ for which the conjecture fails to hold. Then ${\mathbf{O}}_p(G){\mathbf{O}}_{p'}(G)\leq {\mathbf{Z}}(G)$.
\end{theo}
\begin{proof}
Set $N:={\mathbf{O}}_{p'}(G)$ and fix a block ${\rm bl}(\mu)\in{\rm Bl}(N)$ covered by $B$. By the proof of Lemma \ref{lem:Knorr-Robinson for Character Triple Conjecture} we know that $Z={\mathbf{O}}_p(G)$ and it's enough to show that $N\leq {\mathbf{Z}}(G)$. Proceeding as in the proof of Proposition \ref{prop:Minimal counterexample structure, unstable} and noticing that induction of characters preserves the residue of characters, we deduce that $\mu$ must be $A$-invariant. Then, using Lemma \ref{lem:Residues} and adapting the the proof of Theorem \ref{thm:Minimal counterexample structure}, we obtain $N\leq {\mathbf{Z}}(G)$.
\end{proof}
\bibliographystyle{alpha}
|
\section{Introduction}
The advancements in the development of reliable and low-cost sensors, capable of measuring different structural response quantities (e.g. accelerations, displacements, strains, temperatures, loads, etc.) have led to vast scientific and practical developments in the field of Structural Health Monitoring (SHM) over the last four decades \cite{Worden_introduction}. Techniques for processing the raw measurement data and obtaining indicators of structural ``health" have been made readily available \cite{Limongelli}. However, despite the advancements in the field, SHM still remains predominantly applied within the research community \cite{Worden_book} and has not yet translated to extensive application on real-world structures and infrastructure systems. One main reason for this is that the effect and the potential benefit from the use of SHM systems can only be appraised on the basis of the decisions that are triggered by monitoring data. Key open-ended questions include \cite{VOI_ID}: How can information obtained from an SHM system provide optimal decision support? What is the Value of Information (VoI) from SHM systems? How can it be maximized?
Preposterior Bayesian decision analysis can be employed as a formal framework for quantifying the VoI \cite{Raiffa}, which adequately incorporates the uncertainties related to the structural performance and the associated costs, the monitoring measurements, etc. A VoI analysis provides the necessary mathematical framework for quantifying the benefit of an SHM system prior to its installation. In the civil and infrastructure engineering context, the computation of the VoI has been considered mainly related to optimal inspection planning for deteriorating structural systems \cite{Straub_Faber, Jesus, Vereecken}. Recent works \cite{Pozzi, Zonta, Straub_VOI, Thons, Konakli, Andriotis, Zhang, Iannacone} use the VoI concept in an attempt to quantify the value of SHM on idealized structural systems within a Bayesian framework. All works to date, however, adopt rather simplified assumptions regarding the type of information offered by the SHM system. They thus largely rely on hypothetical likelihood functions or observation models, which render these demonstrations, although insightful, not easily transferable to realistic applications. A first attempt towards modeling the entire SHM process and the monitoring information has been made by the authors in \cite{Kamariotis}, which is formalized and extended herein.
Installation of a continuous monitoring system on a structure allows for continuous measurement of the dynamic response of the structure (e.g. accelerations, strain, etc.). In an in-operation regime, a precise measurement of the acting loads, which are usually distributed along a system (e.g. wind, traffic), is a challenging task. Output-only operational modal analysis (OMA) \cite{SSI, Au_OMA} techniques have been developed to alleviate the burden of the absence of acting load measurements. Using an OMA procedure one can identify the system eigenfrequencies and mode shapes of typical structures excited by unmeasured ambient (broadband) loads. This is beneficial, since the operation of the structure is not obstructed, as it would be in the case of forced vibration testing.
Further to data acquisition and system identification, model updating forms a popular subsequent step toward modeling the system performance on the basis of the monitoring information. This process is also referred to as the process of establishing a digital twin via model updating \cite{Wright}. Bayesian model updating (BMU) using identified modal data has proved successful in identifying damage on a global or local level within a structure \cite{Vanik, Papadimitriou, Beck_MCMC, Simoen, Yuen, Behmanesh}. These methods hold significant promise for application with actual full-scale structures \cite{Ntotsios, Moaveni, Argyris}. The vast majority of studies are focused on investigating how the BMU framework performs in detecting, localizing and quantifying different types of artificially created damage given some fixed set of modal data. A few recent studies are concerned with BMU using vibrational data obtained in a continuous fashion from SHM systems \cite{Behmanesh,Simoen_progressive, Ierimenti}. However, no studies are available that systematically quantify the benefit of BMU using continuous SHM data towards driving optimal informed maintenance decision making.
This work embeds a sequential implementation of the BMU framework within a preposterior Bayesian decision analysis, to quantify the VoI from long-term vibrational data obtained from an SHM system. We employ a numerical benchmark for continuous monitoring under operational variability \cite{bench19} to test and demonstrate the approach. The numerical benchmark serves as a tool to create continuous reference monitoring data from a two-span bridge system subject to different types (scour, corrosion) of deterioration at specific hotspots over its lifespan. The benchmark is used as a simulator for extracting dynamic response data, i.e. simulated measurements (accelerations), corresponding to a typical deployment of accelerometers on the structure. Acceleration measurements are provided as input to an output-only OMA algorithm, which identifies the system's modal characteristics. We implement Bayesian model and structural reliability updating methods in a sequential setting for incorporating the continuous OMA-identified modal data within a decision making framework. This proposed procedure follows the roadmap to quantifying the benefit of SHM presented in \cite{VOI_ID}. We employ a simple heuristic-based approach for the solution of the life-cycle optimization problem in the preposterior Bayesian decision analysis. The resulting optimal expected total life-cycle costs are computed in the preposterior case, and compared against the optimal expected total life-cycle costs obtained in the case of only prior knowledge, thus enabling the quantification of the VoI of SHM.
\section{VoI from SHM analysis}
\label{sec:VoI_workflow}
The monitoring of a structural system through deployment of an appropriately designed SHM system is a viable means to support decision-making related to infrastructure maintenance actions. But is gathering this information worth it? Preposterior Bayesian decision analysis provides the necessary formal mathematical framework for quantifying the VoI of an SHM system. A concise representation of a such an analysis with the use of an influence diagram (ID) has been introduced in \cite{VOI_ID}. An adaptation of this ID for the purposes of the VoI analysis that we propose and apply on a simulated SHM benchmark study in this paper is offered in Figure \ref{fig: framework}.
\begin{figure}[ht]
\centerline{
\includegraphics[width=0.9\textwidth]{figures/VoI_SHM_ID.pdf}
}
\caption{Influence diagram of the SHM process for a preposterior Bayesian decision analysis to quantify the VoI.}
\label{fig: framework}
\end{figure}
Influence diagrams build upon Bayesian networks (BN), which offer a concise graphical tool to model Bayesian probabilistic inference problems, and extend these through the addition of decision and utility nodes to model decision-making under uncertainty \cite{Jensen}.
In the ID of Figure \ref{fig: framework}, green oval nodes model uncertain parameters and models/processes related to the structural system, the orange square node models the decision on the SHM system, while the orange oval node models the monitoring data that is extracted via use of a specific SHM system. This data can be used to learn the structural condition via Bayesian updating to then inform the decision on maintenance/repair actions (red square node). Finally, the grey diamond-shaped nodes represent the different costs that enter into the process. The box [t+1] shows that this ID represents a decision process over the lifetime of the structure. The blue text bubbles introduce the different computational methods that are incorporated in the different parts of the process. The large number of these bubbles highlights the modeling and computational challenges associated with a full VoI analysis.
In this paper, for the first time in existing literature, we avoid overly simplifying assumptions in some parts of the modeling of the preposterior Bayesian decision analysis for quantifying the VoI from SHM, but we still model some parts of the process in a simplistic way. The main contribution lies in the modeling of the SHM data. As can be seen in Figure \ref{fig: framework}, we employ continuous SHM information over the lifetime of a deteriorating structural system in the form of acceleration time series, which are subsequently processed by an OMA procedure that identifies the modal eigenfrequencies and mode shapes. These SHM modal data are then used within a BMU procedure to sequentially identify the structural condition (see Section \ref{sec:Bayes}). The way in which the SHM data sets are sampled within a preposterior Bayesian decision analysis with the use of the benchmark structural model is described in detail in Section \ref{subsec: Synthetic}. We treat the modeling of the structural performance node of the ID, as well as the incorporation of the monitoring information within a reliability updating, in a realistic and computationally efficient approach (see Section \ref{sec: SR}). To provide a computationally viable solution to the VoI analysis, we adopt a rather simplified modeling of the action decision node, and we perform the life cycle optimization with the use of heuristics (see Section \ref{sec:LCC}).
The solution of the preposterior Bayesian decision analysis leads to monitoring-informed optimization of the repair action, which in turn leads to the computation of the optimal expected total life-cycle cost in the case of having an SHM system installed. If the adopted SHM strategy is to implement no SHM system, then life-cycle optimization is conducted on the basis of prior information only. By comparing the optimal expected total life-cycle costs in the prior and preposterior cases, the VoI is implicitly quantified as the difference between the two.
\section{Bayesian model updating}
\label{sec:Bayes}
In this section, the Bayesian model updating framework with the use of OMA-identified modal data is presented. The Bayesian formulation presented here corresponds to the state-of-the-art formulation \cite{Simoen, Yuen, Moaveni}.
\subsection{Bayesian formulation}
We consider deterioration that leads to local stiffness reductions. The random variables (RVs) describing the uncertainty within the employed deterioration models are $\boldsymbol{\theta} \in {\rm I\!R}^d$, with $d$ being the total number of RVs. The goal of the Bayesian inverse problem is to infer the deterioration model parameters $\boldsymbol{\theta}$ given noisy OMA-identified modal data. These are the modal eigenvalues $\widetilde{\lambda}_m = (2\pi \widetilde{f}_m)^2$, which can be identified quite accurately, and/or mode shape vector components $\boldsymbol{\widetilde{\Phi}}_{m} \in {\rm I\!R}^{N_{s}}$ at the $N_s$ degrees of freedom (DOF) that correspond to the sensor locations, where $m=1,...,N_m$ is the number of identified modes. An accurate identification of the mode shape displacements requires the deployment of a relatively large number of sensors. Conditional on a fairly good representation of the mode shape vector, one can then derive other modal characteristics, such as the mode shape curvatures $\boldsymbol{\widetilde{K}}_{m} \in {\rm I\!R}^{N_{s}}$, which are shown to be more sensitive to local damage \cite{Pandey}. If only the eigenvalue data is at hand, damage can be detected on a global level, while damage localization requires the existence of spatial information, in the form of mode shape (or mode shape curvature) data.
Consider a linear finite element (FE) model, which is parameterized through the parameters $\boldsymbol{\theta}$ of the deterioration models. The goal of the Bayesian probabilistic framework is to estimate the parameters $\boldsymbol{\theta}$, and their uncertainty, such that the FE model predicted modal eigenvalues $\lambda_{m}(\boldsymbol{\theta})$ and mode shapes $\boldsymbol{\Phi}_{m}(\boldsymbol{\theta})$, or mode shape curvatures $\boldsymbol{K}_{m}(\boldsymbol{\theta})$, best match the corresponding SHM modal data.
Using Bayes' theorem, the posterior probability density function $\pi_{\text{pos}}$ of the deterioration model parameters $\boldsymbol{\theta}$ given an identified modal data set $[\boldsymbol{\widetilde{\lambda}},\boldsymbol{\widetilde{\Phi}}]$ is computed via equation (\ref{Bayes}); it is proportional to the likelihood function $L(\boldsymbol{\theta}; \boldsymbol{\widetilde{\lambda}},\boldsymbol{\widetilde{\Phi}})$ multiplied with the prior PDF of the model parameters $\pi_{pr}(\boldsymbol{\theta})$. The proportionality constant is the so-called model evidence $Z$ and requires the solution of a $d$-dimensional integral, shown in equation (\ref{Evidence}).
\begin{equation}
\pi_{\text{pos}}(\boldsymbol{\theta} \mid \boldsymbol{\widetilde{\lambda}},\boldsymbol{\widetilde{\Phi}}) \propto L(\boldsymbol{\theta}; \boldsymbol{\widetilde{\lambda}},\boldsymbol{\widetilde{\Phi}}) \pi_{\text{pr}}(\boldsymbol{\theta})
\label{Bayes}
\end{equation}
\begin{equation}
Z= \int_{\Omega_{\boldsymbol{\theta}}}L(\boldsymbol{\theta}; \boldsymbol{\widetilde{\lambda}},\boldsymbol{\widetilde{\Phi}}) \pi_{\text{pr}}(\boldsymbol{\theta}) d\boldsymbol{\theta}
\label{Evidence}
\end{equation}
The model updating procedure contains significant uncertainties, which should be taken into account within the Bayesian framework. According to \cite{Simoen}, these are classified into i) measurement uncertainty, including random measurement noise and variance or bias errors induced in the SSI procedure, and ii) model uncertainty. In \cite{Behmanesh} the existence of inherent variability emerging from changing environmental conditions is highlighted. The combination of all the above uncertainties is called the total prediction error in literature \cite{Simoen, Behmanesh}. In order to construct the likelihood function, the eigenvalue and mode shape (similarly for mode shape curvature) prediction errors for a specific mode $m$ are defined as in equations (\ref{eigenvalue_error}) and (\ref{modeshape_error}).
\begin{equation}
\eta_{\lambda_{m}} = \widetilde{\lambda}_{m} - \lambda_{m}(\boldsymbol{\theta}) \in {\rm I\!R}
\label{eigenvalue_error}
\end{equation}
\begin{equation}
\boldsymbol{\eta}_{\boldsymbol{\Phi}_m} = \gamma _{m} \boldsymbol{\widetilde{\Phi}}_{m} - \boldsymbol{\Phi}_{m}(\boldsymbol{\theta}) \in {\rm I\!R}^{N_{s}}
\label{modeshape_error}
\end{equation}
where $\gamma_{m}$ is a normalization constant, which is computed as in equation (\ref{gamma}). $\boldsymbol{\Gamma}$ is a binary matrix for selecting the FE degrees of freedom that correspond to the sensor locations.
\begin{equation}
\gamma_{m} = \frac{{\boldsymbol{\widetilde{\Phi}}_{m}}^T \boldsymbol{\Gamma}\boldsymbol{\Phi}_{m}}{\left \| \boldsymbol{\widetilde{\Phi}}_{m} \right \|^2}
\label{gamma}
\end{equation}
The probabilistic model of the eigenvalue prediction error is a zero-mean Gaussian random variable with standard deviation assumed to be proportional to the measured eigenvalues:
\begin{equation}
\eta_{\lambda_{m}} \sim \mathcal{N}\left(0, c_{\lambda m}^2 \widetilde{\lambda}_{m}^2 \right)
\label{eigenvalue_error_std}
\end{equation}
All the $N_s$ mode shape prediction error components in the vector $\boldsymbol{\eta}_{\boldsymbol{\Phi}_m}$ are assigned a zero-mean Gaussian random variable with the same standard deviation, assumed proportional to the $L_2$-norm of the measured mode shape vector. A multivariate Gaussian distribution is used to model this error:
\begin{equation}
\begin{split}
&\boldsymbol{\eta}_{\boldsymbol{\Phi}_m} \sim \mathcal{N}(\boldsymbol{0}, \boldsymbol{\Sigma}_{\boldsymbol{\Phi}_{m}}) \\
\boldsymbol{\Sigma}_{\boldsymbol{\Phi}_{m}} &= \text{diag}\left(c_{\Phi m}^2 \left \| \gamma_{m} \boldsymbol{\widetilde{\Phi}}_{m} \right \|^2\right)
\end{split}
\end{equation}
The factors $c_{\lambda m}$ and $c_{\Phi m}$ can be regarded as assigned coefficients of variation, and their chosen values reflect the total prediction error. In practical applications, usually very little (if anything) is known about the structure or the magnitude of the total prediction error. At the same time, even if the assumption of an uncorrelated zero mean Gaussian model for the errors has computational advantages and can be justified by the maximum entropy principle, the choice of the magnitude of the factors $c_{\lambda m}$ and $c_{\Phi m}$ clearly affects the results of the Bayesian updating procedure. It appears that most published works do not properly justify this particular choice of the magnitude of the error.
Assuming statistical independence among the $N_m$ identified modes, the likelihood function for a given modal data set can be written as in equation (\ref{likelihood}).
\begin{equation}
L\left(\boldsymbol{\theta}; \boldsymbol{\widetilde{\lambda}}, \boldsymbol{\widetilde{\Phi}}\right)= \prod_{m=1}^{N_m}N\left(\eta_{\lambda_{m}}; 0, c_{\lambda m}^2 \widetilde{\lambda}_{m}^2 \right)
N(\boldsymbol{\eta}_{\boldsymbol{\Phi}_m} ; \boldsymbol{0}, \boldsymbol{\Sigma}_{\boldsymbol{\Phi}_{m}})
\label{likelihood}
\end{equation}
The benefit of SHM is that the sensors can provide data in a continuous fashion, therefore resulting in an abundance of measurements received almost continually. Assuming independence among $N_t$ modal data sets obtained at different time instances, the likelihood can be expressed as:
\begin{equation}
L\left(\boldsymbol{\theta}; \boldsymbol{\widetilde{\lambda}}_{1}...\boldsymbol{\widetilde{\lambda}}_{N_t},\boldsymbol{\widetilde{\Phi}}_{1}...\boldsymbol{\widetilde{\Phi}}_{N_t} \right) = \prod_{t=1}^{N_t}\prod_{m=1}^{N_m}N\left(\widetilde{\lambda}_{t_m}- \lambda_{t_m}(\boldsymbol{\theta}); 0, c_{\lambda m}^2 \widetilde{\lambda}_{t_m}^2\right)
N\left(\gamma _{t_m} \boldsymbol{\widetilde{\Phi}}_{t_m} - \boldsymbol{\Phi}_{t_m}(\boldsymbol{\theta}); \boldsymbol{0}, \boldsymbol{\Sigma}_{\boldsymbol{\Phi}_{t_m}} \right)
\label{time_likelihood}
\end{equation}
where the index $t_m$ indicates the modal data of mode $m$ identified at time instance $t$. The formulation in equation \eqref{time_likelihood} allows for sequential implementation of the Bayesian updating process.
At any time step $t_i$ when new data becomes available, the distribution of the parameters given all the data up to time $t_i$, $\pi_{\text{pos}}(\boldsymbol{\theta} \mid \boldsymbol{\widetilde{\lambda}}_{1:i},\boldsymbol{\widetilde{\Phi}}_{1:i})$ or the one step ahead predictive distributions for time $t_{i+1}$ can be obtained. The inclusion of data in a continuous fashion can increase the level of accuracy of the Bayesian model updating procedure. However, one should be aware that the assumption of independence in equation \eqref{time_likelihood} typically does not hold. This could be addressed by a hierarchical modeling of $\boldsymbol{\theta}$ \cite{ Behmanesh}.
\subsection{Solution methods}
\label{subsec: Bayesian solution}
The solution of the Bayesian updating problem in the general case involves the solution of the $d$-dimensional integral for the computation of the model evidence. Analytic solutions to this integral are available only in special cases, otherwise numerical integration or sampling methods are deployed. The two solution methods that we employ within this work are the Laplace asymptotic approximation and an adaptive Markov Chain Monte Carlo (MCMC) algorithm.
\subsubsection{Laplace approximation}
\label{subsubsec: Laplace}
A detailed presentation of this method can be found in \cite{Katafygiotis, Papadimitriou}. The main idea is that for globally identifiable cases \cite{Katafygiotis}, and for large enough number of experimental data, the posterior distribution can be approximated by a multivariate Gaussian distribution $N(\boldsymbol{\mu}, \boldsymbol{\Sigma})$. The mean vector $\boldsymbol{\mu}$ is set equal to the most probable value, or maximum aposteriori (MAP) estimate, of the parameter vector, which is obtained by minimizing the negative logposterior:
\begin{equation}
\boldsymbol{\mu} = \boldsymbol{\theta}_{MAP} = \underset{\boldsymbol{\theta}}{\argmin}(-\operatorname{ln}\pi_{\text{pos}}(\boldsymbol{\theta} \mid\boldsymbol{\widetilde{\lambda}},\boldsymbol{\widetilde{\Phi}})) = \underset{\boldsymbol{\theta}}{\argmin}(-\operatorname{ln}L(\boldsymbol{\theta}; \boldsymbol{\widetilde{\lambda}},\boldsymbol{\widetilde{\Phi}}) -\operatorname{ln}\pi_{\text{pr}}(\boldsymbol{\theta}))
\label{MAP}
\end{equation}
and the covariance matrix $\boldsymbol{\Sigma}$ is equal to the inverse of the Hessian of the log-posterior evaluated at the MAP estimate. When new data becomes available, the new posterior distribution has to be approximated. The MAP estimate of the previous time step is used as the initial point for the optimization at the current time step, to facilitate a faster convergence of the optimization algorithm.
\subsubsection{MCMC sampling}
For more accurate estimates of the posterior distributions than the one obtained by using the Laplace approximation, one can resort to MCMC sampling methods. Among the multiple available MCMC algorithms, here we employ the adaptive MCMC algorithm from \cite{Haario}, in which the adaptation is performed on the covariance matrix of the proposal PDF. Whenever new data becomes available, the MCMC algorithm has to be rerun to obtain the new posterior distribution. The posterior mean of the parameters estimated via MCMC at the previous time step is used as seed of the new Markov chain, which allows the chain to converge faster.
\section{Structural reliability of a deteriorating structural system and its updating}
\label{sec: SR}
Estimation of the structural reliability, and the use of vibrational data to update this, is instrumental for the framework that we are presenting here. A detailed review of the ideas presented in this section can be found in \cite{Melchers, Straub1}.
\subsection{Structural reliability analysis for a deteriorating structural system}
\label{subsec: SR_prior}
In its simplest form, a failure event at time $t$ can be described in terms of a structural system capacity $R(t)$ and a demand $S(t)$. Both $R$ and $S$ are random variables. With $D(\boldsymbol{\theta},t)$ we define a parametric stochastic deterioration model. Herein we assume that the structural capacity $R(t)$ can be separated from the demand $S(t)$, and that the capacity is deterministic and known for a given deterioration $D(\boldsymbol{\theta},t)$, hence we write $R\left(D(\boldsymbol{\theta},t)\right)$. More details on how this deterministic curve can be obtained for specific cases are given in Section 6, which contains the numerical examples. Therefore, at a time $t$ the structural capacity includes the effect of the deterioration process. The uncertain demand acting on the structure is here modeled by the distribution of the maximum load in a one-year time interval. The cumulative distribution function (CDF) of this distribution is denoted $F_{s_{max}}$. Such a modeling choice simplifies the estimation of the structural reliability, as will be made clear in what follows, which is vital within a computationally expensive VoI analysis framework.
We discretize time in annual intervals $j=1,..,T$, where the $j$-th interval represents $t \in (t_{j-1}, t_j]$. For the type of problems that we are considering, the time-variant reliability problem can be replaced by a series of time-invariant reliability problems \cite{Straub1}. $F_j^*$ is defined as the event of failure in interval $(t_{j-1}, t_j]$. For a given value of the deterioration model parameters $\boldsymbol{\theta}$ and time $t_j$, the capacity $R\left(D(\boldsymbol{\theta},t_j)\right)$ is fixed, and the conditional interval probability of failure is defined as:
\begin{equation}
\text{Pr}(F_j^* \mid \boldsymbol{\theta}, t_j)= 1 -F_{s_{max}}\left(R\left(D(\boldsymbol{\theta},t_j)\right)\right)
\end{equation}
We define $\text{Pr}[F(t_i)] = \text{Pr}(F_1^*\cup F_2^*\cup...F_i^*)$ as the accumulated probability of failure up to time $t_i$.
One can compute $Pr[F(t_i)]$ through the conditional interval probabilities $Pr(F_j^*|R(\boldsymbol{\theta}, t_j))$ as:
\begin{equation}
\text{Pr}[F(t_i)\mid \boldsymbol{\theta}] =1-\prod_{j=1}^{i}[1 - \text{Pr}(F_j^*\mid \boldsymbol{\theta}, t_j)]
\label{conditional_accumulated}
\end{equation}
Following the total probability theorem, the unconditional accumulated probability of failure is:
\begin{equation}
\text{Pr}[F(t_i)] = \int_{\Omega_{\boldsymbol{\theta}}}\text{Pr}[F(t_i)\mid \boldsymbol{\theta}] \pi_{\text{pr}}(\boldsymbol{\theta})d\boldsymbol{\theta}
\label{accumulated}
\end{equation}
The solution to the above integral is approximated using Monte Carlo simulation (MCS). We draw samples from the prior distribution $\pi_{\text{pr}}(\boldsymbol{\theta}) $ of the uncertain deterioration model parameters and the integral in (\ref{accumulated}) is approximated by:
\begin{equation}
\text{Pr}[F(t_i)] \approx \frac{1}{n_{\text{MCS}}} \sum_{k=1}^{n_{\text{MCS}}} \text{Pr}[F(t_i)\mid \boldsymbol{\theta}^{(k)}]
\label{MCS}
\end{equation}
Having computed the probabilities $Pr[F(t_i)]$, one can compute the hazard function $h(t_i)$ for the different time intervals $t_i$, which expresses the failure rate of the structure conditional on survival up to time $t_{i-1}$:
\begin{equation}
h(t_i) = \frac{\text{Pr}[F(t_i)] - \text{Pr}[F(t_{i-1})]}{1 - \text{Pr}[F(t_{i-1})]}
\label{hazard_prior}
\end{equation}
\subsection{Structural reliability updating using SHM modal data}
\label{subsec:SR_updating}
The goal of SHM is to identify structural damage. Monitoring data can be employed in order to identify the parameters $\boldsymbol{\theta}$ of the deterioration models and obtain their posterior distribution, as shown in Section \ref{sec:Bayes}. Consequently this leads to the updating of the accumulated probability of failure at time $t_i$, which can now be conditioned on data $\boldsymbol{Z}_{1:i-1}$ obtained up to time $t_{i-1}$.
\begin{equation}
\text{Pr}[F(t_i)\mid \boldsymbol{Z}_{1:i-1}] = \text{Pr}(F_1^*\cup F_2^*\cup...F_i^*\mid \boldsymbol{Z}_{1:i-1})
\label{conditional}
\end{equation}
The accumulated probability of failure up to time $t_i$ conditional on modal data obtained up to time $t_{i-1}$ is:
\begin{equation}
\text{Pr}[F(t_i)\mid \boldsymbol{Z}_{1:i-1}] =\\
\int_{\Omega_{\boldsymbol{\theta}}}\text{Pr}[F(t_i)\mid \boldsymbol{\theta}] \pi_{\text{pos}}(\boldsymbol{\theta}| \boldsymbol{\widetilde{\lambda}}_{1:i-1}, \boldsymbol{\widetilde{\Phi}}_{1:i-1})d\boldsymbol{\theta}
\label{conditional_integral}
\end{equation}
In (\ref{conditional_integral}), one needs to integrate over the posterior distribution of the parameters $\boldsymbol{\theta}$. As described in Section (\ref{subsec: Bayesian solution}), two different methods for obtaining samples from this posterior distribution at each time step are implemented. In the case that an adaptive MCMC algorithm is used, at every step of the sequential updating we obtain the desired posterior distribution of the parameters in the form of correlated MCMC samples. In the case that the posterior distributions are approximated by multivariate Gaussian distributions using the Laplace approximation, independent posterior samples can be drawn from this approximate posterior density. Using $n_{\text{pos}}$ samples $\boldsymbol{\theta}^{(k)}$ from either MCMC or the asymptotic approximation , the integral in equation (\ref{conditional_integral}) can be approximated:
\begin{equation}
\text{Pr}[F(t_i)\mid \boldsymbol{Z}_{1:i-1}]
\approx \frac{1}{n_{\text{pos}}} \sum_{k=1}^{n_{\text{pos}}} \text{Pr}[F(t_i)\mid \boldsymbol{\theta}^{(k)}]
\label{accumulated_posterior}
\end{equation}
The hazard function conditional on the monitoring data can then be obtained as:
\begin{equation}
h(t_i \mid \boldsymbol{Z}_{1:i-1}) = \frac{\text{Pr}[F(t_i) \mid \boldsymbol{Z}_{1:i-1}] - \text{Pr}[F(t_{i-1}) \mid \boldsymbol{Z}_{1:i-1}]}{1 - \text{Pr}[F(t_{i-1})\mid \boldsymbol{Z}_{1:i-1}]}
\label{hazard_posterior}
\end{equation}
\section{Life-cycle cost with SHM}
\label{sec:LCC}
\subsection{Life-cycle optimization based on heuristics}
The VoI is the difference in life-cycle cost between the cases with and without SHM system. To calculate the life-cycle cost we optimize the maintenance strategy. A strategy $S$ is a set of policies that determine which action to take at any time step $t_i$, conditional on all the information at hand up to that time \cite{Jensen}, \cite{Elizabeth}. One may define policies based on simple decision rules, also called heuristics, which may emerge from basic engineering understanding.
A detailed presentation of the use of heuristics in optimal inspection and maintenance planning can be found in \cite{Jesus, Elizabeth}. With the use of heuristics, the space of solutions to the decision problem is drastically reduced, but the problem is solved only approximately. Here, we utilize a simple heuristic for maintenance decisions. The simple heuristic chosen in this work is the following: Perform a repair action whenever the estimate of the hazard function (the conditional failure rate) is larger than a predefined threshold $h_{thres}$. The use of the hazard function as a decision criteria for condition assessment and maintenance planning is a popular choice in literature \cite{Elingwood}. The parameter $w = h_{thres}$ describing the heuristic is a parameter of the strategy $S$. For simplicity, we assume herein that performing a repair action results in replacing the damaged components and bringing them back to the initial state, and that no failure will occur once a repair action has been performed. In this way, after a repair action, the computation of the total life cycle cost stops. This modeling choice is simplifying, but does allow for a viable computation of the VoI herein.
The total life-cycle cost $C_{\text{tot}}$ is here taken as the total cost of maintenance and the risk of failure costs over the lifetime of the structure. The initial cost is not included in $C_{\text{tot}}$, because it is the same with or without SHM, therefore it cancels out when caclulating the VoI.
With the use of heuristics, solving the decision problem boils down to finding the optimal value of the heuristic parameter $w$ which minimizes the expected total cost, i.e. to the solution of the optimization problem:
\begin{equation}
w^* = \underset{w}{\argmin} \boldsymbol{\text{E}}[C_{\text{tot}} \mid w]
\label{optimization}
\end{equation}
\subsection{Computation of the expected total life-cycle cost in the prior case}
\label{subsec: Prior_LCC}
In the prior case, where only the prior deterioration model is available, the expectation in equation (\ref{optimization}) is with respect to the system state, i.e. the deterioration model parameters $\boldsymbol{\theta}$.
The total cost of maintenance and risk is the sum of the repair costs and the risk of failure costs over the lifetime of the bridge, $C_{\text{tot}}(w, \boldsymbol{\theta}) = C_{\text{R}}(w) + C_{\text{F}}(w, \boldsymbol{\theta})$, therefore the expected total life-cycle cost for a given heuristic parameter $w$ is:
\begin{equation}
\boldsymbol{\text{E}}_{\boldsymbol{\theta}}[C_{\text{tot}} \mid w] = \boldsymbol{\text{E}}_{\boldsymbol{\theta}}[C_{\text{R}}(w) \mid w] +
\boldsymbol{\text{E}}_{\boldsymbol{\theta}}[C_{\text{F}}(w, \boldsymbol{\theta}) \mid w]
\label{cost_breakdown}
\end{equation}
The first part of the right hand side of equation (\ref{cost_breakdown}) can be computed in the following way. We draw samples $\boldsymbol{\theta}^{(k)}, k=1,..,n_{\text{MCS}}$, from the prior distribution $\pi_{\text{pr}}(\boldsymbol{\theta})$ and use them to compute the accumulated probability of failure via equation (\ref{MCS}), and subsequently compute the hazard function with equation (\ref{hazard_prior}). When the hazard function exceeds the threshold, i.e. when $h(t_i) \geq w$, then we define $t_{\text{repair}}(w) = t_{i-1}$ as the time that the repair takes place. The time of repair is thus a function of our chosen heuristic. Hence the expected total cost of repair over the lifetime is given as:
\begin{equation}
\boldsymbol{\text{E}}_{\boldsymbol{\theta}}[C_{\text{R}}(w) \mid w] = \hat{c}_R\gamma(t_{\text{repair}}(w))
\label{cost_repair}
\end{equation}
where $\hat{c}_R$ is the fixed cost of the repair, and $\gamma(t) = \frac{1}{(1+r)^t}$ is the discounting function, $r$ being the annually compounded discount rate.
The risk of failure over the lifetime can be computed via MCS, using the samples $\boldsymbol{\theta}^{(k)}$, $k=1,..,n_{\text{MCS}}$, that were drawn from the prior distribution $\pi_{\text{pr}}(\boldsymbol{\theta})$, with the following formula:
\begin{equation}
\boldsymbol{\text{E}}_{\boldsymbol{\theta}}[C_{\text{F}}(w, \boldsymbol{\theta}) \mid w] \approx \frac{1}{n_{\text{MCS}}} \sum_{k=1}^{n_{\text{MCS}}} C_{\text{F}}(w, \boldsymbol{\theta}^{(k)})
\label{MC_for_risk}
\end{equation}
where:
\begin{equation}
C_{\text{F}}(w, \boldsymbol{\theta}^{(k)}) = \sum_{i=1}^{t_{repair}(w)}\hat{c}_F\gamma(t_i) \{\text{Pr}[F(t_i)\mid \boldsymbol{\theta}^{(k)}]- \text{Pr}[F(t_{i-1})\mid \boldsymbol{\theta}^{(k)}]\}
\label{cost_failure}
\end{equation}
and $\hat{c}_F$ is the fixed cost of the failure event.
Following the solution of the optimization problem in (\ref{optimization}), the expected total life-cycle cost associated with the optimal decision in the prior case without any monitoring data is $\boldsymbol{\text{E}}_{\boldsymbol{\theta}}[C_{\text{tot}} \mid w_0^*]$.
\subsection{Computation of the expected total life-cycle cost in the preposterior case}
\label{subsec: Preposterior_cost}
The goal of a preposterior analysis is to act as a decision tool on whether collecting SHM data is beneficial, and to quantify the VoI of an SHM system, prior to its installation. Therefore this type of analysis is performed before any actual SHM data are obtained. Instead, the SHM monitoring data histories must be sampled over the lifetime from the prior distribution of the uncertain deterioration model parameters $\boldsymbol{\theta}$, as will be explained shortly. A sampled monitoring data history vector $\boldsymbol{Z}= [\boldsymbol{Z}_1,...,\boldsymbol{Z}_{n_T}]$ contains the OMA identified modal data at fixed time instances over the structure lifetime. The rate at which this data is sampled should be chosen on the basis of the problem at hand. In this case, we explore a slow evolving deterioration process, and further ignore dependence on temperature effects. Therefore, for this investigation we employ one set of OMA-identified modal data per year. Under the currently assumed independence on environmental and operational conditions (EOCs) this sparse assumption is further justified from the observation that increasing the number of data sets used in the BMU process seemingly decreases the parameter estimation uncertainty, albeit not properly reflecting the full variability of the updated parameters \cite{Behmanesh, Vanik}.
In a preposterior analysis, the expectation in equation (\ref{optimization}) is operating over both the system state $\boldsymbol{\theta}$ and on the monitoring outcomes $\boldsymbol{Z}$.
\begin{equation}
\boldsymbol{\text{E}}_{\boldsymbol{\theta}, \boldsymbol{Z}}[C_{\text{tot}} \mid w] = \int_{\Omega_{\boldsymbol{\theta}}} \int_{\Omega_{\boldsymbol{Z}}} C_{\text{tot}}(w, \boldsymbol{\theta}, \boldsymbol{z})f_{\boldsymbol{\Theta}, \boldsymbol{Z}}(\boldsymbol{\theta},\boldsymbol{z} )d\boldsymbol{z}d\boldsymbol{\theta}
\label{posterior_case}
\end{equation}
The total cost of maintenance and risk is again the sum of the repair cost and the risk of failure cost over the lifetime of the structure, which now both depend also on the monitoring outcomes $\boldsymbol{Z}$, $C_{\text{tot}}(w, \boldsymbol{\theta}, \boldsymbol{Z}) = C_{\text{R}}(w, \boldsymbol{Z}) + C_{\text{F}}(w, \boldsymbol{\theta}, \boldsymbol{Z})$.
The integral in equation (\ref{posterior_case}) is computed with crude MCS. We draw samples from the uncertain deterioration model parameters $\boldsymbol{\theta}$, which correspond to a deterioration history over the lifetime, as given by the deterioration model equation $D(\boldsymbol{\theta},t)$. For each of these histories, we generate noisy acceleration measurements every year, feed them into an SSI algorithm, and obtain one vector of monitoring modal data $\boldsymbol{Z}$ (one identified modal data set per year). In this way we are jointly sampling the system state space and monitoring data space, and equation (\ref{posterior_case}) is approximated as:
\begin{equation}
\boldsymbol{\text{E}}_{\boldsymbol{\theta}, \boldsymbol{Z}}[C_{\text{tot}} \mid w] =\frac{1}{n_{MCS}}\sum_{k = 1}^{n_{MCS}}[C_{\text{R}}(w, \boldsymbol{z}^{(k)}) + C_{\text{F}}(w, \boldsymbol{\theta}^{(k)}, \boldsymbol{z}^{(k)})]
\label{post_case_MCS}
\end{equation}
For each of the sampled system states and corresponding monitoring data, we compute the updated hazard rate as given by equation (\ref{hazard_posterior}), and when $h(t_i \mid \boldsymbol{z}^{(k)}_{1:i-1}) \geq w$, then $t_{\text{repair}}(w, \boldsymbol{z}^{(k)}) = t_{i-1}$.
The cost of repair is:
\begin{equation}
C_{\text{R}}(w, \boldsymbol{z}^{(k)}) = \hat{c}_R\gamma(t_{\text{repair}}(w,\boldsymbol{z}^{(k)}))
\label{cost_repair_post}
\end{equation}
The risk of failure is:
\begin{equation}
C_{\text{F}}(w, \boldsymbol{\theta}^{(k)},\boldsymbol{z}^{(k)} ) = \sum_{i=1}^{t_\text{repair}(w, \boldsymbol{z}^{(k)})}\hat{c}_F\gamma(t_i) \{\text{Pr}[F(t_i)\mid \boldsymbol{\theta}^{(k)}]-
\text{Pr}[F(t_{i-1})\mid \boldsymbol{\theta}^{(k)}]\}
\label{cost_failure_post}
\end{equation}
Comparing equations (\ref{cost_failure}) and (\ref{cost_failure_post}) it is evident that adoption of the same samples of $\boldsymbol{\theta}$ in both prior and preposterior analysis, leads to an identical estimate of the risk of failure for the two analyses up to the time of the repair. The only difference between prior and preposterior case is the resulting $t_\text{repair}(w, \boldsymbol{z}^{(k)})$.
Solving equation (\ref{optimization}), we obtain the optimal expected total life-cycle cost given the monitoring data, $\boldsymbol{\text{E}}_{\boldsymbol{\theta, Z}}[C_{\text{tot}} \mid w_{mon}^*]$.
\subsection{Summary of the proposed methodology to calculate the VoI}
\label{subsec: VOI}
The proposed procedure for the VoI analysis consists of the following steps:
\begin{enumerate}
\item Choose a prior stochastic deterioration model describing the structural condition over the lifetime of the structure. Define a decision analysis time discretization, maintenance/repair actions, costs of actions and cost of failure event. Choose a heuristic parameter $w$ (threshold on hazard rate) for a heuristic-based solution of the decision problem.
\item Draw Monte Carlo samples $\boldsymbol{\theta}$ of the stochastic deterioration model parameters.
\item Perform a prior decision analysis:
\begin{itemize}
\item Use the prior $\boldsymbol{\theta}$ samples to estimate the lifetime accumulated probability of failure $Pr[F(t_i)]$ and the corresponding hazard rate $h(t_i)$.
\item Solve the LCC optimization problem to obtain the optimal value of the heuristic parameter $w_0^*$ and the corresponding optimal $t_{repair}$. Obtain the optimal expected LCC in the prior case: $\boldsymbol{\text{E}}_{\boldsymbol{\theta}}[C_{\text{tot}}(\boldsymbol{\theta},w) \mid w_0^*]$.
\end{itemize}
\item Perform a preposterior decision analysis:
\begin{itemize}
\item For each individual prior sample $\boldsymbol{\theta}$ realization and given value of the heuristic parameter $w$ do the following:
\begin{enumerate}[label=(\alph*)]
\item Sample the corresponding noisy acceleration time series data for every year over the lifetime of the structure. Feed the accelerations into the SSI algorithm to identify the structure's modal data vectors $\boldsymbol{Z}$.
\item Perform a posterior Bayesian analysis: BMU to sequentially learn the posterior distributions of $\boldsymbol{\theta}$ and subsequently obtain an updated estimate of the accumulated probability of failure $\text{Pr}[F(t_i)\mid \boldsymbol{Z}_{1:i-1}]$ and the hazard rate $h(t_i \mid \boldsymbol{Z}_{1:i-1})$.
\item Find the time to perform the repair action for this specific deterioration and monitoring data realization, conditional on a value of the heuristic parameter $w$.
\end{enumerate}
\item Solve the LCC optimization problem to obtain the optimal value of the heuristic parameter $w_{mon}^*$ which minimizes the expected LCC in the preposterior case $\boldsymbol{\text{E}}_{\boldsymbol{\theta, Z}}[C_{\text{tot}}(\boldsymbol{\theta},\boldsymbol{Z},w) \mid w_{mon}^*]$.
\end{itemize}
\item Compute the VoI.
\begin{equation}
VoI= \boldsymbol{\text{E}}_{\boldsymbol{\theta}}[C_{\text{tot}}(\boldsymbol{\theta},w) \mid w_0^*] - \boldsymbol{\text{E}}_{\boldsymbol{\theta, Z}}[C_{\text{tot}}(\boldsymbol{\theta},\boldsymbol{Z},w) \mid w_{mon}^*]
\label{VOI}
\end{equation}
\end{enumerate}
\subsection{Value of Partial Perfect Information}
\label{subsec: PI}
The case of partial perfect information is related to a hypothetical situation, in which the SHM system provides perfect information on the condition of the structure. This means that there is no uncertainty on the parameters $\boldsymbol{\theta}$ of the deterioration model, and the optimal decision is found conditional on this perfect knowledge of $\boldsymbol{\theta}$. Because the SHM system is not able to provide any information about the load acting on the structure, which here is modeled by an uncertain Gumbel random variable, therefore one uses the term ``partial''.
Estimation of the value of partial perfect information is given by:
\begin{equation}
VPPI= \underset{w}{\min} \boldsymbol{\text{E}}_{\boldsymbol{\theta}}[C_{\text{tot}}(\boldsymbol{\theta},w)] - \boldsymbol{\text{E}}_{\boldsymbol{\theta}}\{\underset{w}{\min}[C_{\text{tot}}(\boldsymbol{\theta},w)\mid \boldsymbol{\theta}]\}
\label{VPPI}
\end{equation}
The left hand side of equation (\ref{VPPI}) is the optimal expected total life-cycle cost in the prior case, exactly as presented in Section \ref{subsec: Prior_LCC}. On the right hand side, first the optimal heuristic is found conditional on exact knowledge of $\boldsymbol{\theta}$, then the expected value of the total life-cycle costs associated with optimal decisions is computed. This quantity corresponds to the value of information that one would obtain in the case of perfect monitoring and perfect decision making with the chosen heuristic.
The VPPI provides an upper limit on the value that the VoI can obtain. Since it can be computed much easier than the VoI, the VPPI can provide a first estimate on the maximum investment that should be made for SHM systems. Therefore, we motivate the idea that a VPPI computation should always be performed first.
\section{Numerical investigations}
\label{sec: Numerical_Investigations}
\subsection{Numerical benchmark: Continuously monitored bridge system subject to deterioration}
We consider the two-span bridge model of Figure~\ref{f:benchmark}, with its reference behavior \cite{bench19} simulated by a FE model of isoparametric plane stress quadrilateral elements. This benchmark structure has been developed as part of the TU1402 COST Action and serves for verification of analysis methods and tools for SHM. 200 elements are employed to mesh the $x$ direction, and 6 elements are assumed per height ($y$ direction). The beam dimensions form configurable parameters of the benchmark and are set as: height $h$ = 0.6m, width $w$ = 0.1m, while the lengths are $L_1$ = 12m for
the first span and $L_2$ = 13m for the second span. A linear elastic material with Young's modulus $E$ = 30GPa, Poisson ratio $\nu$ = 0.2, and material density $\rho$ = 2000 kg/m\textsuperscript{3} is assigned. Elastic boundaries in both directions are assumed for all three support points, in the form of translational springs with $K_x$ = 10\textsuperscript{8} N/m and $K_y$ = 10\textsuperscript{7} N/m.
It is assumed that the simulated two-span bridge is continuously monitored using a set of sensors measuring vertical acceleration, whose locations correspond to predefined FE nodes. A distributed Gaussian white noise excitation $F(x)$ is used as the load acting on the bridge, to simulate the unknown ambient excitation. A dynamic time history analysis of the model, for a given realization of the load, results in the measured vertical acceleration signals at the assigned sensor locations.
\begin{figure}[ht]
\centerline{
\includegraphics[width=\textwidth]{figures/benchmark_4.pdf}
}
\caption{Benchmark model}
\label{f:benchmark}
\end{figure}
\subsection{Deterioration modeling}
A prior model describing structural deterioration is a prerequisite for a VoI analysis. A detailed presentation of probabilistic deterioration models for life-cycle performance assessment of structures can be found in \cite{Elingwood, Frangopol, Biondini}. For time-dependent reliability assessment purposes, the use of simple empirical models, which are still flexible enough to model different kinds of deterioration mechanisms, can be adopted \cite{Elingwood}.
Within this work, we use a simple rate equation of the form:
\begin{equation}
D(t) = A t^B
\label{deterioration}
\end{equation}
to model structural deterioration, where $D(t)$ is the unit-less deterioration parameter (loss of stiffness) entering in the assumed damage model, and $A, B$ are random variables driving the uncertainty in this model. Parameter $A$ models the deterioration rate, while parameter $B$ is related to the nonlinearity effect in terms of a power law in time. We consider herein the following two case studies related to structural deterioration of the bridge structure.
\subsubsection{Bridge system subject to scour}
We assume that the middle elastic support (pier) of the bridge structure is subjected to gradual deterioration, simulating the case of scour \cite{Prendergast}. Damage is introduced as a progressive reduction of the stiffness in $y$-direction of the spring $K_y^{(2)}$ at the middle elastic support of the bridge (Figure \ref{f:benchmark}). The evolution of the stiffness reduction of the vertical spring support over the lifespan of the bridge is described by employing the damage model of equation (\ref{stifness_reduction}), where $K_{y,0}^{(2)}$ is the initial undamaged value, and $D(t)$ is the stiffness reduction described by equation (\ref{deterioration}). We consider a lifespan of $T$=50 years for the structure. The uncertain parameters of the deterioration model are summarized in Table \ref{table:Parameters_scour}. The mean and coefficient of variation of the parameters $A$ and $B$ are chosen to reflect a significant a-priori uncertainty. They result in a 10\% probability that $D(t=50)>9$ at the end of the lifespan.
\begin{equation}
K_y^{(2)}(t) = \frac{K_{y,0}^{(2)}}{(1+D(t))} = \frac{K_{y,0}^{(2)}}{(1+At^{B})}
\label{stifness_reduction}
\end{equation}
\begin{table}[!ht]
\caption{Parameters of the stochastic deterioration model for scour.}
\footnotesize
\centering
\begin{tabular}{cccc}\hline
Parameter&Distribution&Mean&CV\\\hline
A&Lognormal&7.955$\times$10\textsuperscript{-4}&0.5\\
B&Normal&2.0&0.15\\\hline
\end{tabular}
\label{table:Parameters_scour}
\end{table}
\subsubsection{Bridge system subject to corrosion deterioration}
As a second separate case study, we assume that the bridge structure is subjected to gradual deterioration from corrosion in the middle of both midspans (elements in black in Figure \ref{f:benchmark}). At both locations, damage is introduced as a progressive reduction of the stiffness at the bottom 2 elements of the FE mesh. For the deterioration hotspots at the left and right midspans, the evolution of the elements' stiffness reduction over the lifespan of the bridge is described by employing the damage model of equation (\ref{Youngs_modulus_reduction_left}). $E^{(0)}$ is the initial undamaged value of the Young's modulus, and $D_1(t)$, $D_2(t)$ are the deterioration models (reduction of stiffness) employed for each location, as described by equation (\ref{deterioration}). The random variables of the deterioration models are summarized in Table \ref{table:Parameters}. According to \cite{Elingwood}, for this simple empirical deterioration model, a value of $B$=0.5 corresponds to diffusion-controlled damage processes. Therefore the mean values of $B_1$ and $B_2$ have been chosen equal to 0.5. The mean and coefficient of variation of the four uncertain parameters are chosen so that they result in a 1\% probability that $D(t=50)>$9 at the end of the lifespan.
\begin{equation}
E_j(t) = \frac{E^{(0)}}{(1+D_j(t))} = \frac{E^{(0)}}{(1+A_jt^{B_j})}, j = 1, 2
\label{Youngs_modulus_reduction_left}
\end{equation}
\begin{table}[ht]
\caption{Parameters of the stochastic deterioration model for corrosion.}
\footnotesize
\centering
\begin{tabular}{cccc}\hline
Parameters&Distribution&Mean&CV\\\hline
$A_1, A_2$&Lognormal&0.506&0.4\\
$B_1, B_2$&Normal&0.5&0.15\\\hline
\end{tabular}
\label{table:Parameters}
\end{table}
\subsection{Synthetic monitoring data creation}
\label{subsec: Synthetic}
For the purpose of the VoI analysis framework presented in this paper, for every deterioration time instance at which we want to simulate a monitoring data set obtained from the deployed SHM system, the corresponding stiffness reduction is implemented in the FE benchmark model, a dynamic time history analysis is run and the ``true" vertical acceleration signals $\ddot{x}$ at the sensor locations (FE nodes) are obtained. The noise-free acceleration time series data set is contaminated with Gaussian white noise of 2\% root mean square noise-to-signal ratio, simulating a sensor measurement error. Subsequently the noisy accelerations $\tilde{\ddot{x}}$ are fed into an output-only operational modal analysis (OMA) scheme. Specifically, the stochastic subspace identification (SSI) \cite{SSI} algorithm is used to identify a set of the lower eigenvalues (squares of natural frequencies) and mode shapes. The data creation process can be seen in Figure \ref{f:synthetic}.
\begin{figure}
\centerline{
\includegraphics[width=\textwidth]{figures/synthetic_data_creation.pdf}
}
\caption{Process for generating the SHM data}
\label{f:synthetic}
\end{figure}
\subsection{Continuous Bayesian model updating}
\label{subsec: BMU results_chapter_6}
Initially we demonstrate how the Bayesian framework performs in learning the parameters of the deterioration model on the basis of availability of the SHM modal data. In this work, the model predicting the eigenvalues and mode shapes for the Bayesian updating process is the same FE model as the one described in Section \ref{subsec: Synthetic} for the creation of the noise-contaminated synthetic data. Despite addition of artificial noise, adoption of the same model constitutes a so-called inverse crime \cite{inverse_crime}. This is the a built in feature of standard preposterior analysis.
We draw samples of the deterioration parameters, defining the evolution of sample deterioration curves. For each of these deterioration curves we create one monitoring history, i.e., we generate one set of OMA-identified modal data every year over the fifty years of the lifetime. In this simple example, the structural properties are not assumed influenced by environmental (temperature, humidity) and operation (non stationary effects due to traffic) variability. For this reason, we assume it suffices to utilize one estimate of the modal properties set per year. Using this data, we employ the sequential Bayesian deterioration model updating framework of Section \ref{sec:Bayes}.
The sequential Bayesian analysis framework requires a substantial number of evaluations of the likelihood function, implying multiple forward runs of the FE model. Within a VoI framework, Bayesian analysis must be performed numerous times. For this reason a VoI analysis can quickly become intractable. To enable the VoI analysis, we employ simple surrogate models to replace the structural FE model, which are described in the following two subsections.
\subsubsection{Bridge system subject to scour deterioration - Global damage identification}
In this assumed damage scenario we are interested in identifying damage in a global scale, for which use of the OMA-identified eigenvalue data alone may be sufficient. The benefit is that eigenvalue data can be successfully identified from an OMA procedure, even when only a rather small number of accelerometers are employed on the structure. The sensor placement that we assume here is the one corresponding to Figure \ref{f:benchmark_3}, with twelve employed sensors. This configuration is selected on the basis of engineering judgment, when seeking to identify the type of damage (local stiffness reduction) considered herein. Using the SSI algorithm, we identify the lower $N_m=6$ modes, which we then use for the updating.
\begin{figure}
\centerline{
\includegraphics[width=\textwidth]{figures/benchmark_3.pdf}
}
\caption{Bridge system subject to scour damage}
\label{f:benchmark_3}
\end{figure}
We employ a surrogate model to replace the structural FE model for facilitating the Bayesian updating procedure. To this end, we create a fine uniform grid of values for $D(t)$; for each of these, we execute a modal analysis using the FE model and store the output eigenvalues. Eventually, we replace the modal analysis run of the structural FE model with a simple nearest neighbor lookup in the precomputed database.
For illustrating the data sampling and updating process, we assume a scenario where the underlying ``true" deterioration model corresponds to parameters values $A^*$=9.85$\times$10\textsuperscript{-4} and $B^*$ = 2.28. The ``true" deterioration curve can be seen in black in all the subfigures of Figure \ref{f:updating_scour}.
Figure \ref{f:parameters_learnt_scour} demonstrates how the distribution of the deterioration model parameters is updated by comparing the prior PDF of $A$ and $B$ with the posterior PDF of $A$ and $B$ at year 25 and year 50. For this analysis, both factors $c_{\lambda m}$ and $c_{\Phi m}$ are assumed equal to $0.02$, i.e. we assume that the total prediction error causes up to two percent deviation on the nominal model predicted values. 5000 MCMC samples are used for the Bayesian analysis at each time step. The posterior PDFs are given via a kernel density estimation using the 5000 posterior MCMC samples of the parameters. It is observed that using one SHM data set per year, the uncertainty in the deterioration model parameters gradually decreases, the PDFs become narrower and peak around the underlying ``true'' values for which the data was created.
Figure \ref{f:updating_scour} contains the following: The mean estimated deterioration model together with its $90\%$ credible interval in the prior case, obtained via a MCS from the prior distribution of the uncertain parameters, is plotted in the left panel in green. In red we plot the posterior predictive mean models together with their $90\%$ credible intervals, which are estimated with posterior MCMC samples using monitoring data up to the three different time instances. For example in the second column, we use the monitoring data of the first ten years to obtain the posterior distribution $\pi_{\text{pos}}(\boldsymbol{\theta} \mid \boldsymbol{\widetilde{\lambda}}_{1:10},\boldsymbol{\widetilde{\Phi}}_{1:10})$, and then we use the posterior MCMC samples to predict the evolution of the deterioration model over the structural lifetime. We observe that already the data obtained during the first few years of the deterioration process (up to year 10) help in shifting the mean posterior model towards the underlying ``true'' deterioration curve, however the posterior uncertainty in the estimation is still relatively large. The posterior uncertainty is reduced significantly as more SHM modal data become gradually available (year 25, year 50).
\begin{figure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_a_25.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_b_25.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_a_50.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_b_50.pdf}
\end{subfigure}
\caption{Prior PDF and posterior PDF at years 25 and 50 for deterioration model parameters ($c_{\lambda m}=c_{\Phi m}=0.02$)}
\label{f:parameters_learnt_scour}
\end{figure}
\begin{figure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/prior_model_scour.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/year_10_updating.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/year_25_updating.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/year_50_updating.pdf}
\end{subfigure}
\caption{Sequential Bayesian learning of the scour deterioration model}
\label{f:updating_scour}
\end{figure}
\subsubsection{Bridge system subject corrosion deterioration - Damage detection and localization}
\label{subsubsec: Updating_corrosion}
\begin{figure}
\centerline{
\includegraphics[width=\textwidth]{figures/benchmark_2.pdf}
}
\caption{Bridge system subject to corrosion damage in two locations}
\label{f:benchmark_2}
\end{figure}
In the assumed scenario with two potential corrosion damage locations, the employed Bayesian model updating framework should be able to both detect and localize damage. Therefore both eigenvalue as well as mode shape displacement data should become available. As discussed in Section \ref{sec:Bayes}, a relatively large number of sensors is required for an accurate measurement and representation of the mode shape displacements. The sensor placement that we assume here is the one corresponding to Figure \ref{f:benchmark_2}, with 24 equally distributed accelerometers. By using a finite difference scheme, we can also obtain the mode shape curvatures, which are used instead of the mode shapes in the likelihood function, which seems to enhance the localization capabilities of the framework. Also in this case we identify the lower $N_m$=6 modes.
For defining a surrogate model, we create a two-dimensional grid of values for $D_1(t), D_2(t)$, and for each of the grid points we run a modal analysis with the FE model, and we store the output eigenvalues and mode shape vectors. Eventually we employ the following surrogates: For each of the eigenvalues, we fit a two-dimensional polynomial regression response surface model. For the mode shape displacement vector data, we replace the run of the structural FE model with a simple nearest neighbor lookup in the precomputed two-dimensional database.
For illustration purposes, we draw one sample $\boldsymbol{\theta}^*$, which corresponds to the underlying ``true" deterioration parameter values $A_1^*=0.65$, $B_1^*=0.55$, $A_2^*=0.42$ and $B_2^*=0.48$ (``true" deterioration curves can be seen in black in all the subfigures of Figure \ref{f:updating}).
Figures \ref{f:parameters_learnt_1}, \ref{f:parameters_learnt_2}, \ref{f:parameters_learnt_3} demonstrate how the distribution of the deterioration models' parameters is updated, by comparing the prior PDFs with the posterior PDFs at three different time instances. Both factors $c_{\lambda m}$ and $c_{\Phi m}$ in the likelihood function are assumed equal to $0.02$. In Figure \ref{f:updating}, we compare the underlying ``true" deterioration model with the deterioration model estimated using MCS in the prior case, and with the ones estimated with 5000 posterior MCMC samples at three different time instances.
\begin{figure}[!ht]
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_a1_10.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_b1_10.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_a2_10.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_b2_10.pdf}
\end{subfigure}
\caption{Prior PDF and posterior PDF at year 10 for deterioration models parameters ($c_{\lambda m}=c_{\Phi m}=0.02$)}
\label{f:parameters_learnt_1}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_a1_25.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_b1_25.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_a2_25.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_b2_25.pdf}
\end{subfigure}
\caption{Prior PDF and posterior PDF at year 25 for deterioration models parameters ($c_{\lambda m}=c_{\Phi m}=0.02$)}
\label{f:parameters_learnt_2}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_a1_50.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_b1_50.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_a2_50.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_b2_50.pdf}
\end{subfigure}
\caption{Prior PDF and posterior PDF at year 50 for deterioration models parameters ($c_{\lambda m}=c_{\Phi m}=0.02$)}
\label{f:parameters_learnt_3}
\end{figure}
Section \ref{sec:Bayes} discusses the fact that quite often the choice of the magnitude of factors $c_{\lambda m}$ and $c_{\Phi m}$ for constructing the likelihood function can be arbitrary, since usually very little is known about the magnitude of the total prediction error. Figure \ref{f:parameters_learnt_cov_5} attempts to demonstrate how crucial this choice can be for the results of the Bayesian updating, by performing it additionally for $c_{\lambda m}=c_{\Phi m}=0.05$. Comparing Figure \ref{f:parameters_learnt_cov_5} to Figure \ref{f:parameters_learnt_3} (both at year 50), it can be clearly observed that the posterior distribution of the deterioration model parameters that one learns is significantly affected by the choice of these factors.
\begin{figure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/prior_D1.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/year_10_updating_D1.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/year_25_updating_D1.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/year_50_updating_D1.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/prior_D2.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/year_10_updating_D2.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/year_25_updating_D2.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/year_50_updating_D2.pdf}
\end{subfigure}
\caption{Sequential Bayesian learning of the two corrosion deterioration models}
\label{f:updating}
\end{figure}
\begin{figure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_a1_50_cov_5.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_b1_50_cov_5.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_a2_50_cov_5.pdf}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/posterior_b2_50_cov_5.pdf}
\end{subfigure}
\caption{Prior PDF and posterior PDF at year 50 for deterioration models parameters ($c_{\lambda m}=c_{\Phi m}=0.05$)}
\label{f:parameters_learnt_cov_5}
\end{figure}
\subsection{Time-dependent structural reliability and its updating using monitoring data}
The uncertain demand acting on the structure is modeled by the maximum load in a one-year time interval with a Gumbel distribution (left subfigure of Figure \ref{f:capacity_Pr_F_scour}). The parameters of the Gumbel distribution are chosen such that the probability of failure in the initial undamaged state is equal to $10^{-6}$ and the coefficient of variation of the random load is $20\%$.
\subsubsection{Bridge system subject to scour deterioration}
The deterministic capacity curve $R(D(\boldsymbol{\theta},t))$ of the damaged structure for any realization of the scour deterioration $D(\boldsymbol{\theta},t)$ can be seen in the middle panel of Figure \ref{f:capacity_Pr_F_scour}. To determine this curve, we consider that when scour damage occurs in the middle support, the critical quantity that increases is the normal stress at the middle of the second, slightly longer, midspan. We create a fine one-dimensional grid of possible values as input for $D(\boldsymbol{\theta},t)$, for each of those we run a static analysis of our model, and we evaluate the loss of load bearing capacity of the structure relative to the initial undamaged state.
\begin{figure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/Gumbel.pdf}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/capacity_curve.pdf}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.\linewidth]{figures/prior_PrF_hazard_rate_scour.pdf}
\end{subfigure}
\caption{left: CDF of the Gumbel distribution for the load with location $a_n = 0.0509$, scale $b_n= 0.297$. middle: Structural capacity in function of scour deterioration, right: Time-dependent structural reliability curves estimated with the prior model in the scour deterioration case.}
\label{f:capacity_Pr_F_scour}
\end{figure}
In the right panel of Figure \ref{f:capacity_Pr_F_scour} we plot the time-dependent accumulated probability of failure and the hazard function in the prior case, together with the 95\% credible intervals, estimated using $10^4$ prior samples. Because of the skewness of the assumed prior deterioration model, the mean estimated curves are not contained within the 95\% credible intervals.
\subsubsection{Bridge system subject to corrosion deterioration}
When damage (stiffness reduction) occurs in the elements at the bottom of each midspan, the quantity that increases critically are the normal stresses at the top of each midspan. We create a two-dimensional grid of possible values of the two corrosion deteriorations $D_1$ and $D_2$. For each of those possible combinations, we run a static analysis with our model, and we evaluate the loss of load bearing capacity of the bridge structure relative to the undamaged state. Eventually we fit a two-dimensional polynomial regression response surface curve that describes $R(\boldsymbol{D}(\boldsymbol{\theta},t))$; it can be seen in the left panel of Figure \ref{f:SR_corrosion_example}.
As presented in Section \ref{subsec:SR_updating}, learning the parameters of the deterioration models, and the reduction of the uncertainty in their estimation through the sequential acquisition of SHM modal data, affects the estimation of the time-dependent structural reliability. In Figure \ref{f:SR_corrosion_example}, we plot in green the accumulated probability of failure and the hazard rate of the bridge structure in the case of using the prior deterioration model, and we compare it with the red plots of the accumulated probability of failure and the hazard rate conditional on the continuous monitoring data (1 data set per year), which correspond to the underlying ``true'' deterioration models described by $A_1^*=0.65$, $B_1^*=0.55$, $A_2^*=0.42$ and $B_2^*=0.48$. The prior estimates are obtained with 5000 Monte Carlo samples following equations (\ref{MCS}), (\ref{hazard_prior}). The posterior estimates are obtained via equations (\ref{accumulated_posterior}), (\ref{hazard_posterior}) using 5000 MCMC samples at each time step. The $95\%$ credible intervals are computed using the Monte Carlo prior samples in the prior case, and the MCMC posterior samples in the posterior case. It is observed that the uncertainty in the estimation of the structural reliability is reduced in the posterior case. This reduction of the uncertainty and the updated estimate of the structural reliability form the basis for the VoI analysis.
\begin{figure}[ht]
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=0.90\linewidth]{figures/capacity_response_surface_2.pdf}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.05\linewidth]{figures/prior_filtering_PrF_5000_MCMC.pdf}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=1.05\linewidth]{figures/CORRECT_hazard_post_prior_5000_MCMC.pdf}
\end{subfigure}
\caption{left: Polynomial regression response surface for structural capacity in function of corrosion deterioration, right: time dependent structural reliability curves in the prior/posterior corrosion case.}
\label{f:SR_corrosion_example}
\end{figure}
\subsection{VoI analysis}
\label{subsubsec: VOI_results}
The VoI is computed with equation (\ref{VOI}) following the Bayesian preposterior decision analysis framework presented in Section \ref{sec:LCC}. The expected total life-cycle costs, $\boldsymbol{\text{E}}_{\boldsymbol{\theta}}[C_{\text{tot}} \mid w]$ in the prior case and $\boldsymbol{\text{E}}_{\boldsymbol{\theta}, \boldsymbol{Z}}[C_{\text{tot}}|w]$ in the preposterior case, are both computed with MCS. In the preposterior case, as already explained in Section \ref{subsec: Preposterior_cost}, the system state space $\boldsymbol{\theta}$ and the monitoring data space $\boldsymbol{Z}$ are jointly sampled.
For each full history of modal data, one sequential Bayesian posterior analysis has to be performed, which is a costly procedure by itself. It is clear that such an analysis can be very computationally expensive, therefore some considerations on the available computational budget, and how to distribute it, have to be made in advance. The computational cost of the VoI analysis is approximately proportional to the number of MCS samples used in the expected life-cycle cost computation and the necessary corresponding synthetic monitoring data creation, and by the computational cost of the employed method for performing the sequential Bayesian updating.
For our investigation we assume $\hat{c}_F = 10^7$\euro{}, and for the repair cost $\hat{c}_R$ we investigate different ratios $\frac{\hat{c}_R}{\hat{c}_F}=[10^{-1},10^{-2},10^{-3}]$, and for each of those we calculate the VoI. The discount rate is taken as $r=2\%$.
The solution to the stochastic life-cycle optimization problem of equation (\ref{optimization}) is performed through an exhaustive search among a large discrete set of values of the heuristic parameter (the threshold at which a repair is performed).
\subsubsection{VoI results for bridge system subject to scour deterioration}
For this example, we draw 1000 samples of $\boldsymbol{\theta}$, which are used in both prior and preposterior analysis. In the preposterior case, for each $\boldsymbol{\theta}$ sample we create one continuous set of identified modal data $\boldsymbol{Z}_{1:50}$. For the 1000 different sequential Bayesian analyses that have to be performed, we employ the adaptive MCMC algorithm. For the estimation of the different posterior accumulated probabilities of failure in equation (\ref{accumulated_posterior}), 2000 posterior MCMC samples are used.
Tables \ref{table:LCC_scour} and \ref{table:LCC_SHM_scour} summarize the results of the life cycle optimization, documenting the optimal value of the heuristic parameter $w^*$, and the optimal expected total life cycle costs that correspond to $w^*$. Table \ref{table:LCC_scour} documents also the optimal time for a repair action in the prior case. This is not documented in Table \ref{table:LCC_SHM_scour}, since in the preposterior case there is not one single optimal $t_{repair}$ value, but $t_{repair}$ varies for each sample $\boldsymbol{\theta}$ and the corresponding monitoring history.
Table \ref{table:VOI_scour} documents the resulting VoI values that we obtain with the 1000 samples via equation (\ref{VOI}) for the three different cost ratios, while Table \ref{table:VPPI_scour} reports the VPPI values obtained via equation (\ref{VPPI}), related to the hypothetical case when we learn perfectly the condition of the structure from the SHM system. We also include the CV of the mean VoI, VPPI estimates, which quantifies the uncertainty in the estimates obtained via MCS. In cost ratio cases for which the optimal action in the prior case is not to perform any repair, the VoI estimate has a quite large variability. This is because the samples in the preposterior analysis that lead to a different optimal $t_{repair}$ than in the prior case are only a few, which is an indication that a larger number of Monte Carlo samples or more efficient sampling techniques (e.g. importance sampling) should be used to reduce the variance. It is important to take into account that equation (\ref{VPPI}) for computing the VPPI can easily be solved even for a very large number of MC samples, which would reduce the variability of the estimate shown here.
For all the cost ratio cases, the VoI is positive, which indicates a potential benefit of installing an SHM system on the deteriorating bridge structure. It is interesting to compare the obtained VoI values to the VPPI values. We observe that in this example the VoI from SHM extracted via Bayesian model updating is close to optimal, as it provides almost the full VPPI value.
\begin{table}
\caption{Results of preposterior Bayesian decision analysis for the scour example}
\begin{subtable}{.5\textwidth}
\footnotesize
\centering
\caption{Life-cycle optimization in the prior case.}
\begin{tabular}{cccc}
$\frac{\hat{c}_R}{\hat{c}_F}$&$w_0^*$&$\boldsymbol{\text{E}}[C_{\text{tot}}|w_0^*]$&$t_{repair}$\\\hline
$10^{-1}$&$\ge2\times10^{-3}$&45395&no repair\\
$10^{-2}$&$\ge2\times10^{-3}$&45395&no repair\\
$10^{-3}$&2$\times10^{-5}$&5924 &year 31\\\hline
\end{tabular}
\label{table:LCC_scour}
\end{subtable}
\begin{subtable}{.5\textwidth}
\footnotesize
\centering
\caption{Life-cycle optimization in the preposterior case.}
\begin{tabular}{cccc}
$\frac{\hat{c}_R}{\hat{c}_F}$&$w_{mon}^*$&$\boldsymbol{\text{E}}[C_{\text{tot}}|w_{mon}^*]$\\\hline
$10^{-1}$&2.1$\times10^{-2}$&12552\\
$10^{-2}$&1.2$\times10^{-3}$&3125\\
$10^{-3}$&9.9$\times10^{-5}$&1109\\\hline
\end{tabular}
\label{table:LCC_SHM_scour}
\end{subtable}
\begin{subtable}{.5\textwidth}
\footnotesize
\centering
\caption{Value of information (VoI)}
\begin{tabular}{ccc}
$\frac{\hat{c}_R}{\hat{c}_F}$&VoI (CV)\\\hline
$10^{-1}$&32843 (0.34)\\
$10^{-2}$&42270 (0.30)\\
$10^{-3}$&4815 (0.02)\\\hline
\end{tabular}
\label{table:VOI_scour}
\end{subtable}
\begin{subtable}{.5\textwidth}
\footnotesize
\centering
\caption{Value of partial perfect information (VPPI)}
\begin{tabular}{ccc}
$\frac{\hat{c}_R}{\hat{c}_F}$&VPPI (CV)\\\hline
$10^{-1}$&35013 (0.23)\\
$10^{-2}$&42717 (0.21)\\
$10^{-3}$&4918 (0.02)\\\hline
\end{tabular}
\label{table:VPPI_scour}
\end{subtable}
\end{table}
\subsubsection{VoI results for bridge system subject to corrosion deterioration}
\label{subsubsec: VoI_results_corrosion}
For this second example, we draw 2000 samples of $\boldsymbol{\theta}$, which are used in both prior and preposterior analysis. In the preposterior case, for each $\boldsymbol{\theta}$ sample we create one continuous set of identified modal data $\boldsymbol{Z}_{1:50}$. For the 2000 different sequential Bayesian analyses that have to be performed, we employ the Laplace approximation method of Section \ref{subsubsec: Laplace} for the solution, which introduces an approximation error in the posterior solution, especially in the initial years, when the data set is not so large, yet is computationally much faster than an MCMC solution. For the estimation of the posterior accumulated probability of failure in equation (\ref{accumulated_posterior}), 10000 samples are drawn from the approximate multivariate Gaussian posterior distribution.
The computed VoI and VPPI estimates can be seen in Table \ref{table:VOI_corrosion}. We observe that the VoI is 0 in the case when the costs have a ratio $\frac{\hat{c}_R}{\hat{c}_F}=10^{-1}$, which means that one does not get any benefit from the data obtained from the SHM system. This is related to the fact that, for this cost ratio, the optimal decision is to not perform a repair action in the lifespan of the bridge, in both the prior and all the preposterior samples, since at all time steps the cost of a repair is much larger than the the risk of failure cost. For the cost ratio $\frac{\hat{c}_R}{\hat{c}_F}=10^{-2}$, we observe that the VoI from SHM extracted via Bayesian model updating is not optimal, as it does not provide the full VPPI value, but $51\%$ of this value, while for the cost ratio $\frac{\hat{c}_R}{\hat{c}_F}=10^{-3}$ it only provides $22\%$ of the VPPI value.
\begin{table}
\caption{Results of preposterior Bayesian decision analysis for the corrosion example}
\begin{subtable}{.5\textwidth}
\footnotesize
\centering
\caption{Life-cycle optimization in the prior case.}
\begin{tabular}{cccc}
$\frac{\hat{c}_R}{\hat{c}_F}$&$w_0^*$&$\boldsymbol{\text{E}}[C_{\text{tot}}|w_0^*]$&$t_{repair}$\\\hline
$10^{-1}$&$\ge2.8\times10^{-4}$&$26792$&no repair\\
$10^{-2}$&$\ge2.8\times10^{-4}$&$26792$&no repair\\
$10^{-3}$&$2\times10^{-5}$&$9308$&year $8$\\\hline
\end{tabular}
\label{table:LCC}
\end{subtable}
\begin{subtable}{.5\textwidth}
\footnotesize
\centering
\caption{Life-cycle optimization in the preposterior case.}
\begin{tabular}{cccc}
$\frac{\hat{c}_R}{\hat{c}_F}$&$w_{mon}^*$&$\boldsymbol{\text{E}}[C_{\text{tot}}|w_{mon}^*]$\\\hline
$10^{-1}$&$\ge9\times10^{-3}$&$26792$\\
$10^{-2}$&$7.5\times10^{-4}$&$25334$\\
$10^{-3}$&$1.83\times10^{-5}$&$9200$\\\hline
\end{tabular}
\label{table:LCC_SHM}
\end{subtable}
\begin{subtable}{.5\textwidth}
\footnotesize
\centering
\caption{Value of information (VoI)}
\begin{tabular}{ccc}
$\frac{\hat{c}_R}{\hat{c}_F}$&VoI (CV)\\\hline
$10^{-1}$&0\\
$10^{-2}$&1458 (0.42)\\
$10^{-3}$&108 (0.08)\\\hline
\end{tabular}
\label{table:VOI}
\end{subtable}
\begin{subtable}{.5\textwidth}
\footnotesize
\centering
\caption{Value of partial perfect information (VPPI)}
\begin{tabular}{ccc}
$\frac{\hat{c}_R}{\hat{c}_F}$&VPPI (CV)\\\hline
$10^{-1}$& 132 (0.30)\\
$10^{-2}$& 2871 (0.22)\\
$10^{-3}$& 497 (0.05)\\\hline
\end{tabular}
\label{table:VPPI}
\end{subtable}
\label{table:VOI_corrosion}
\end{table}
At a glance, one could claim by looking at the VPPI values, that for both constructed examples, the upper limit to how much value the SHM information can have for supporting the single repair decision is small, since this value does not contain the cost of the SHM system itself. In principle, if the cost of obtaining this information, i.e. the cost of the SHM system (installation, maintenance etc.), is higher than the VPPI, then no further investigation into a VoI analysis would make sense. The resulting VPPI and VoI estimates for these simplified example case studies are affected by the fact that only a single repair action case is explored.
\subsubsection{VoI results - Sensor placement study}
The purpose of this section is to demonstrate that the presented VoI analysis can be employed as a formal decision analysis tool for performing various parametric studies related to different choices in designing the SHM system and performing the Bayesian model updating procedure.
One critical choice when designing an SHM system is the number and position of the sensors to be employed on the structure. One could employ the proposed VoI analysis to perform optimal sensor placement studies for a deteriorating structural system. Each sensor arrangement choice will result in a VoI value, and the choice which leads to the highest VoI would be the preferred one.
Herein we demonstrate this with the use of the second example of the bridge system subject to corrosion deterioration at two locations. For the decision problem, we now fix the cost of failure to $\hat{c}_F = 10^7$\euro{} and the cost of repair to $\hat{c}_R = 3.5\times10^4$\euro{}. We consider the following two different arrangements of the sensors: i) 24 uniformly distributed accelerometers along the structure, ii) 12 uniformly distributed accelerometers along the structure. In both cases the VoI analysis is performed by drawing 1000 samples of $\boldsymbol{\theta}$.
It becomes evident that in the case that the structure is subjected to deterioration at two different damage locations, the number of sensors and consequently the quality of the mode shape displacement or curvature information that one obtains clearly affects the BMU results and therefore leads to a notable difference in the heuristic-based life-cycle optimization and the VoI result that we obtain.
\begin{table}[ht]
\caption{Parametric study for the effect of the number of sensors on VoI result}
\begin{subtable}{1\textwidth}
\footnotesize
\centering
\caption{Life-cycle optimization in the prior case.}
\begin{tabular}{cc}
$w_0^*$&$t_{repair}^*$\\\hline
$6.1\times10^{-5}$&21\\\hline
\end{tabular}
\vspace{5pt}
\label{table:LCC_prior_parametric}
\end{subtable}
\begin{subtable}{.5\textwidth}
\footnotesize
\centering
\caption{Life-cycle optimization in the preposterior case.}
\begin{tabular}{cc}
sensors&$w_{mon}^*$\\\hline
24&$1\times10^{-4}$\\
12&$3.2\times10^{-4}$\\\hline
\end{tabular}
\label{table:LCC_opt_parametric}
\end{subtable}
\begin{subtable}{.5\textwidth}
\footnotesize
\centering
\caption{Effect of number of sensors on the VoI}
\begin{tabular}{c|cccc}
VPPI (CV)&sensors&VoI (CV)&$\frac{\text{VoI}}{\text{VPPI}}$\\\hline
7681 (2.6\%)&24&4614 (5.3\%)&60\%\\
&12&2711 (15\%)&35\%\\\hline
\end{tabular}
\label{table:VOI_parametric}
\end{subtable}
\label{table:VOI_corrosion_parameter_study}
\end{table}
\section{Concluding remarks}
This paper investigates the quantification of the VoI yielded via adoption of SHM systems acting in long-term prognostic mode for cases of deterioration. A preposterior Bayesian decision analysis for quantifying the VoI, specifically tailored for application on an employed numerical benchmark structural model is presented. The modeling of the acquired SHM data is done in a realistic way, following a state-of-the-art operational modal analysis procedure. The data is used within a Bayesian model updating framework, implemented in a sequential setting, to continuously update the uncertain structural condition, which subsequently leads to the updating of the estimate of the structural reliability. A heuristic-based solution to the simplified decision problem is provided for finding the optimal time to perform a single repair action, which might be needed during the lifetime of the structure. We discuss specific computational aspects of a VoI calculation. The VoI analysis requires the integration over the monitoring data, which are here modeled in a realistic way, adding an extra computationally expensive layer in the analysis. In addition to the VoI solution, an upper limit to the VoI through the value of partial perfect information is also provided, related to hypothetical situations of perfect knowledge on the system condition. It should be noted that the resulting VoI estimates are affected by the fact that only a single repair action case is explored. In the present exemplary analysis, we do not take into account dependence on varying environmental effects (e.g dependence on temperature). The VoI analysis and results presented herein focus on demonstrating, for the first time, a VoI analysis on the full SHM chain, from data acquisition to utilization of a structural model for the purpose of the updating and reliability calculation.
\section*{Declaration of competing interest}
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
\section*{Acknowledgments}
The authors would like to gratefully acknowledge the support of the TUM Institute for Advanced Study through the Hans Fischer Fellowship.
|
\section{Introduction}
Neural Machine Translation (NMT) has been shown to be very sensitive to noise \citep{belinkov2018synthetic,michel-neubig-2018-mtnt,ebrahimi2018adversarial}, with even small perturbations in the inputs often leading to mistranslations. To improve the robustness of NMT models, current research mostly focuses on adapting the model to noisy texts via methods such as fine-tuning \citep{michel-neubig-2018-mtnt,alam-anastasopoulos-2020-fine}, noise-injection \cite{belinkov2018synthetic,cheng-etal-2018-towards,karpukhin-etal-2019-training}, and data augmentation through back-translation \cite{berard-etal-2019-naver,vaibhav-etal-2019-improving,li-specia-2019-improving}, etc. In these approaches, the translation model is trained or fine-tuned on the noisy data so that it can learn from the noise. However, methods using extra context to help translate noisy texts have not been investigated.
\begin{figure}[hbt]
\centering
\includegraphics[width=.75\columnwidth]{figure/dense_att_ex2.pdf}
\caption{As showed by \citet{caglayan-etal-2019-probing}, multimodality can help translate unknown words, but fail when there is noise in the input. The misspelled word ``song'' is correctly translated as ``enfant'' (child) when it is replaced with an unknown token, but translated literally as ``chanson'' (song) otherwise.}
\label{fig:ex_att}
\end{figure}
Studies in Multimodal Machine Translation (MMT) have shown that visual information improves translation quality when the textual context is incomplete \citep{caglayan-etal-2019-probing,imankulova-EtAl:2020:WMT,caglayan-etal-2020-simultaneous}. However, as exemplified by \citet{caglayan-etal-2019-probing} (Figure \ref{fig:ex_att}), an MMT model trained on clean data was not able to handle noise. When the word ``son'' was misspelled as ``song'', the model disregarded the visual information and used the literal translation ``chanson''. The MMT model attended to the relevant region in the image and generated the intended translation ``enfant'' only when the noise was masked by a placeholder in the input, imitating an out-of-vocabulary (OOV) example.
Given that the visual modality has been shown to help predict unknown words,
we investigate whether adding multimodal information to adaption-based methods would further improve translation robustness. To answer this question, we build MMT models in conjunction with noise injection techniques and investigate their behaviour during training and inference on both noisy and clean data. To further improve robustness, we extend the current \textit{adversarial training} method (i.e., training NMT models on noisy texts) and propose an \textit{error correction training} method. In addition to training the model with noise-injected source sentences and their clean translation counterparts, we introduce error correction as an auxiliary task and add a separate decoder to the model, which is used to denoise the source sentence.\footnote{Codes are available at \url{https://github.com/Nickeilf/Visual-Cues-Error-Correction}} Our {\bf main contributions} can be summarized as:
\begin{itemize}
\item To the best of our knowledge, this is the first work
combining adversarial training with \textit{multimodal} NMT to improve translation robustness. We evaluate robustness on three types of noise that mimic errors commonly introduced by humans. Systematic experiments reveal that multimodality can improve model performance on both known and unseen noise.
\item We propose an \textit{error correction training} method for translation by introducing denoising as an auxiliary task. We show that the robustness of both NMT and MMT models is improved with this method.
\item We demonstrate that the model using visual features also learns to correct grammatical errors more accurately, indicating the potential for multimodal monolingual error correction.
\end{itemize}
The paper is organised as follows: In Section \ref{sec:bg}, we present the background and related work. In Section \ref{sec:method}, we introduce the types of noise injected and the error correction training method. In Section \ref{sec:exp}, we describe our experiment settings, with experiment results in Section \ref{sec:result}, and further analysis in Section \ref{sec:analysis}.
\section{Background and Related Work}
\label{sec:bg}
\paragraph{Robust NMT}
Although NMT models can achieve high performance on clean data, they are very brittle to non-standard inputs, such as noisy texts \citep{belinkov2018synthetic}. Different types of noisy data have been proposed to test translation robustness, e.g. synthetic word perturbations \citep{belinkov2018synthetic}, grammatical errors \citep{anastasopoulos-etal-2019-neural}, and user-generated texts from social platform \cite{michel-neubig-2018-mtnt,li-EtAl:2019:WMT1,specia-EtAl:2020:WMT1}.
The most common approach to improve translation robustness is to train the model on noisy data, which is referred to as adversarial training. Since parallel data with noisy source sentences and clean translations is difficult to obtain, the clean training data is often injected with different types of artificial noise, e.g. random word perturbations like character insertion/deletion/substitution \citep{belinkov2018synthetic,karpukhin-etal-2019-training,passban2020revisiting,xu2021addressing}, noise generated via back-translation \citep{berard-etal-2019-naver,vaibhav-etal-2019-improving,li-specia-2019-improving}, and adversarial examples generated by white-box generator model
\citep{cheng-etal-2018-towards,cheng-etal-2019-robust,cheng-etal-2020-advaug}. Even though this method has been shown to improve NMT performance on noisy data, the types of noise used thus far are not common in real data. For example, it would be highly unlikely for human authors to misspell the word ``robust'' as ``zobust'', but such random transformations are used when synthesizing noisy training data for MT. In addition, back-translation paraphrases the texts to introduce noise, however such noise is less realistic as human-generated errors, which include mispellings and grammatical errors. In adversarial approaches for other NLP tasks, \citet{ribeiro-etal-2020-beyond} and \citet{ma2019nlpaug} introduce various methods to inject both artificial and realistic noise. Inspired by these work, we focus on three types of noise that are commonly generated by humans in real texts and experiment with these for the translation task.
\paragraph{MMT} Multimodal machine translation extends the framework of NMT by incorporating extra modalities, e.g. image \citep{specia-etal-2016-shared} or audio \citep{sulubacak2020multimodal}. In our case, the extra modality is given as visual features from an image network to complement the textual context. In standard MMT, these features can be fused with the textual representation by simple operations such as concatenation \citep{caglayan2016multimodal}, hidden states initialization \citep{calixto-liu-2017-incorporating}, or via attention mechanisms \citep{libovicky-helcl-2017-attention,calixto-etal-2016-dcu,calixto-etal-2017-doubly,yao-wan-2020-multimodal} and latent variables \citep{calixto-etal-2019-latent}.
Recent research has shown that the extra modality helps translation, especially when the input is incomplete \citep{caglayan-etal-2019-probing,caglayan-etal-2020-simultaneous,imankulova-EtAl:2020:WMT} or ambiguous \citep{ive-etal-2019-distilling,DBLP:journals/corr/abs-1908-01665}. \citet{wu2019transformerbased} hinted at the possibility of multimodality helping NMT in dealing with natural noise stemming from the speech recognition system used as a first step in their pipeline approach to speech translations from videos. Their results, however, were inconclusive.
\begin{table*}[htb]
\centering
\begin{tabular}{ll}
\toprule
\textbf{clean} & a pink flower is starting to bloom .\\
\textbf{edit-distance} & a pink flower is staring to \red{loom} .\\
\textbf{homophone} & a pink \red{flour} is starting to bloom .\\
\textbf{keyboard} & a pink flower is \red{starring} to bloom .\\
\toprule
\end{tabular}
\caption{An example of noise injected to the clean text. The noisy substitutes are marked in red.}
\label{tab:noise_example}
\end{table*}
\citet{salesky2021robust} investigate the robustness of open-vocabulary translation by representing texts as images followed by optical character recognition to cover some cases of noise such as misspellings. This is an interesting but orthogonal area of research since no external visual information is used.
Therefore, it remains an open question whether MMT can perform better than NMT on noisy texts, and whether multimodality can be complementary rather than redundant to previous text-based robustness techniques. The work by \citet{caglayan-etal-2019-probing} is the closest to our approach, however they focused mainly on identifying \textit{when} the visual information is helpful. As such, they only performed experiments comparing NMT and MMT in the presence of unknown words consisting of placeholders used to mask
out words in the source sentence. In contrast, we focus on multimodal models for realistic noise that includes in- and out-of-vocabulary words, such as misspellings or correctly-spelled words used in an incorrect context.
\section{Methods}
\label{sec:method}
In this section, we introduce our methods to improve and evaluate the robustness of NMT and MMT models. In Section \ref{sec:noise_injection}, we describe three techniques to inject realistic noise into training and test data. In Section \ref{sec:error_correction}, we introduce our error correction training method.
\subsection{Noise Injection}
\label{sec:noise_injection}
In previous work on noise injection, the perturbations are often arbitrary, which would result in unrealistic noise. To simulate the natural noise in real situations, we add constraints to the random perturbations. We select three constrained noise injection methods that can be applied to both training and test data, with each method simulating one type of human-generated errors:
\paragraph{Edit distance} A word is randomly replaced with another word in the vocabulary where the edit distance between the two words is less than two characters. The edit-distance noise simulates the occurrence of confusable spellings (e.g. sat vs seat) and also some grammatical errors (e.g. horse vs horses).
\paragraph{Homophones} A word is randomly replaced with another word that shares the same pronunciation. We use the CMU Pronouncing Dictionary\footnote{\url{http://www.speech.cs.cmu.edu/cgi-bin/cmudict?in=C+M+U+Dictionary}} to transform words into phonemes and find noisy substitutes with the same pronunciation. This simulates errors made by applications such as automatic speech recognition, or by non-native speakers.
\paragraph{Keyboard \citep{belinkov2018synthetic}} A character in a word is randomly replaced with an adjacent key on the standard QWERTY keyboard. The keyboard noise simulates the real-life typos when users accidentally press wrong keys while typing.
\\\\
Table \ref{tab:noise_example} shows examples of the three types of noise we experimented with. The edit distance and homophone noise types are applied on the word level, while the keyboard noise is on the character level.
Word-level noise is more likely to break the sentence context even though the noisy substitutes are correctly spelled words. On the contrary, character-level noise is likely to introduce misspelled words and increase the out-of-vocabulary (OOV) rate.
When constructing the noisy training or test sets, we sample from the three types of noise following a uniform distribution, where to each sentence we apply only one type of noise. To avoid substituting words not carrying much contextual information (e.g. articles and punctuations)
, we only perturb words with more than two characters. The noise level is controlled by the hyperparameter $\mathit{n}$, which defines the maximum number of words replaced with noisy counterparts per sentence. The noise injection procedure can be characterized as: given a source sentence $\mathbf{x} = [x_1, x_2, ..., x_M]$ and a target translation $\mathbf{y} = [y_1, y_2, ..., y_N]$, noise will be injected to the clean source sentence $\mathbf{x}$ to obtain its noisy variant $\mathbf{x'} = [x_1, ..., x'_{a_i} ,..., x_M]$, where $a_i$ is the position of the noisy substitutes ($i=\{1,2,...,n\}$).
\begin{figure*}
\centering
\includegraphics[width=0.35\textwidth]{figure/nmt_cor.pdf}
\includegraphics[width=0.334\textwidth]{figure/mmt_cor.pdf}
\caption{Illustration of the joint training of machine translation and error correction for NMT and MMT models. Solid lines: translation flow. Dotted lines: error correction flow. Left: NMT with error correction training. Right: MMT with error correction training.}
\label{fig:model}
\end{figure*}
\subsection{Error Correction Training}
\label{sec:error_correction}
We introduce error correction \citep{ng-etal-2014-conll,yuan-briscoe-2016-grammatical} as an auxiliary task to help improve the robustness against noisy inputs. For that, we add a second decoder to the MT architecture, which is only used for the error correction task.
During training, the noisy sentence $\mathbf{x'}$ is encoded by the encoder, which is shared between the translation and correction tasks, into hidden states $\mathbf{h'}$. The hidden state representation is then fed to both decoders. The translation decoder aims to generate a correct translation $\mathbf{y}$ while the correction decoder aims to recover the original source sentence $\mathbf{x}$. This method is also compatible with the MMT model, where the error correction decoder will use both visual and textual hidden states to recover the clean source sentences. Figure \ref{fig:model} gives an illustration of the model architecture.
Compared to the standard MT model, the version with error correction training (which we refer to as \textit{NMT-cor} and \textit{MMT-cor} hereinafter) maximizes both the probability of generating correct translations $P(\mathbf{y}|\mathbf{x'};\bm{\theta_{mt}})$ and the probability of recovering the clean source sentences $P(\mathbf{x}|\mathbf{x'};\bm{\theta_{cor}})$.
\begin{equation}
\begin{aligned}
P(\mathbf{y}|\mathbf{x'};\bm{\theta_{mt}}) &= \prod_{t=1}^{N}P(y_t|\mathbf{y}_{1:t-1},\mathbf{x'};\bm{\theta_{mt}}) \\
P(\mathbf{x}|\mathbf{x'};\bm{\theta_{cor}}) &= \prod_{t=1}^{M}P(x_t|\mathbf{x}_{1:t-1},\mathbf{x'};\bm{\theta_{cor}}) \\
\end{aligned}
\end{equation}
The $\bm{\theta_{mt}}$ represents parameters for the translation component and the $\bm{\theta_{cor}}$ represents parameters for the error correction component, with $\bm{\theta_{mt}} = \{\bm{\theta_{enc}}, \bm{\theta_{mt\_dec}}\}, \bm{\theta_{cor}} = \{\bm{\theta_{enc}}, \bm{\theta_{cor\_dec}}\}$. Our hypothesis is that the auxiliary task of error correction may help the encoder with a noise-invariant representation, which would indirectly improve the translation of noisy sentences. During training, we jointly optimize the sum of the translation loss and the error correction loss, as is shown in Equation \ref{eq:loss}:
\begin{equation}
\label{eq:loss}
\begin{aligned}
\mathcal{L}_{mt}(\boldsymbol{\theta}_{mt}) &= \frac{1}{|\mathbf{D}|}\sum_{(\mathbf{x'},\mathbf{y})\in\mathbf{D}}-\log P(\mathbf{y}|\mathbf{x'};\boldsymbol{\theta}_{mt}) \\
\mathcal{L}_{cor}(\boldsymbol{\theta}_{cor}) &= \frac{1}{|\mathbf{D}|}\sum_{(\mathbf{x'},\mathbf{x})\in\mathbf{D}}-\log P(\mathbf{x}|\mathbf{x'};\boldsymbol{\theta}_{cor}) \\
\mathcal{L}(\boldsymbol{\theta}) &= \mathcal{L}_{mt}(\boldsymbol{\theta}_{mt}) + \boldsymbol{\lambda}\mathcal{L}_{cor}(\boldsymbol{\theta}_{cor})
\end{aligned}
\end{equation}
where $\boldsymbol{\lambda} \ge 0$ is the factor that controls the weight of the error correction loss, and $\mathbf{D}$ represents the noise-injected data consisting of triples in the form of $(\mathbf{x}, \mathbf{x'}, \mathbf{y})$.
\section{Experiments}
\label{sec:exp}
\subsection{Datasets}
We experiment with the Multi30K dataset \citep{elliott-etal-2016-multi30k}, using both the En-Fr and En-De language pairs. This is the standard dataset for MMT and has been used in all open challenges on the topic \citep{specia-etal_WMT:2016,elliott-EtAl:2017:WMT,barrault-etal_WMT:2018}. Following \citet{caglayan-etal-2019-probing}, we use both the \textit{train} and \textit{valid} splits as our training set. The \textit{test2016-flickr} set is used as our development set for checkpoint selection. For evaluation, we test the models on both \textit{test2017-flickr} and \textit{test2017-mscoco} sets \citep{elliott-etal-2017-findings}. We use a word-level vocabulary and build vocabularies for the original source and target languages, as well as the vocabulary on noisy source texts.\footnote{Therefore there is no OOV word in the noisy training data, but the test data might still contain OOV words -- noisy or not -- with respect to the training.} We use the pre-processed data in Multi30K, which is lowercased, normalized, and tokenized with Moses \cite{koehn-etal-2007-moses}. We also performed experiments using a subword-level vocabulary (BPE), which led to further improvements, but the trend in the results is the same (see Appendix \ref{sec:subword}).
Following \citet{caglayan-etal-2020-simultaneous}, we use the ``bottom-up-top-down'' (BUTD) features \citep{butd} extracted from a pre-trained Faster R-CNN ResNet-101 object detector. Each image is represented as 36 pooled feature vectors $V \in \mathbb{R}^{36\times2048}$, with each vector representing a local object region.
\subsection{Models}
\paragraph{NMT and MMT Models}
Our baseline NMT model is the standard Transformer model \citep{attention-is-all-you-need},
with 6 layers for both the encoder and the decoder. The hidden state size is 512 while the feed-forward dimension is 1024. The number of attention heads is set to 4. Dropout (0.3) is applied to both self/cross-attention and the position-wise feed-forward layer, and Pre-norm \cite{nguyen2019transformers} is applied to boost convergence. Our baseline MMT model follows the same architecture and hyperparameters as the baseline NMT model, except for the multimodal components. We use the serial multimodal cross-attention \citep{libovicky-etal-2018-input}, where an extra cross-attention sublayer is appended in the decoder layer to perform attention over the visual features. We also experiment with GRU models \citep{cho-etal-2014-learning}, following the hyperparameter settings of \citet{caglayan-etal-2019-probing}. Due to space restrictions, we include the detailed results with GRU models in Appendix \ref{sec:gru_results}. The GRU results display the same trend as the experimental results using Transformer models.
\paragraph{Error Correction Models}
The error correction NMT/MMT models adopt the same encoder and decoder as the baseline NMT/MMT models, except for a second decoder added for error correction training. During training, we compute the cross-entropy loss for translation, as well as for error correction in the correction-based models. In these models, the two losses are summed and optimized jointly on the same batch.
We found the best $\lambda$ value ($\lambda \in \{0.2, 0.2, 0.4, 0.4, 0.8\}$) for different levels of noise (number of noisy words $n \in \{1, 2, 4, 6, 10 \}$) during hyperparameter tuning. See Appendix \ref{sec:lambda} for more details.
\paragraph{Training and Evaluation}
We use ADAM \cite{kingma2017adam} as the optimizer and adopt the \textit{noam} learning rate scheduler \cite{attention-is-all-you-need} with a warm-up of 8000 steps. The training batch size is 64. Models are evaluated using the METEOR score \citep{denkowski-lavie-2014-meteor}, which is the main metric for multimodal machine translation \cite{barrault-etal_WMT:2018}.
For the evaluation of error correction, we use ERRANT \citep{bryant-etal-2017-automatic} to compute the $F_{0.5}$ score. During evaluation, we select the checkpoint with the best performance on the development set and generate the translation and correction using beam search of size 12. All models are implemented using \textit{nmtpytorch}\footnote{\url{https://github.com/lium-lst/nmtpytorch}} and \textit{pysimt}\footnote{\url{https://github.com/ImperialNLP/pysimt}}. Each model is run with three random seeds and the average results are reported. Each run takes approximately 2 hours to train on an RTX 2080 Ti GPU.
\section{Results}
\label{sec:result}
\subsection{Testing for Robustness to Noise}
We first evaluate the robustness of standard NMT and MMT models \textbf{trained on clean data} by \textbf{testing on the noise-injected data}. This setting represents regular models that are not specifically adapted to noise. Figure \ref{fig:clean_train} presents the change in METEOR ($\Delta$METEOR) between standard MMT and NMT models tested on data with different noise levels. The $\Delta$METEOR is consistently above 0 for both test sets in the two language pairs. As the noise level increases, the difference between NMT and MMT models is larger, showing that the visual information in the MMT model leads to predictions that are more robust to noise.
\begin{figure}[htb]
\centering
\includegraphics[width=.95\linewidth]{figure/clean_train_delta.pdf}
\caption{Performance gain from multimodality on different test sets when models are \textit{trained on clean data} but \textit{tested on noisy data} ($\Delta$METEOR = MMT - NMT).}
\label{fig:clean_train}
\end{figure}
\begin{table*}[htb]
\centering
\begin{tabular}{l|c c c c c c | c c c c c c}
\toprule
& \multicolumn{6}{c}{\textbf{flickr2017}} & \multicolumn{6}{|c}{\textbf{mscoco2017}} \\
& clean & $n$=1 & $n$=2 & $n$=4 & $n$=6 & $n$=10 & clean & $n$=1 & $n$=2 & $n$=4 & $n$=6 & $n$=10 \\
\midrule
\multicolumn{11}{@{}l}{\textit{en-fr}} \\
NMT & 70.6 & 64.2 & 60.2 & 55.2 & 51.8 & 49.4 & 64.2 & 58.3 & 54.3 & 48.8 & 45.7 & 43.2 \\
MMT & 70.9 & 64.7 & 61.0 & 56.8 & 53.7 & 51.1 & 64.4 & 59.3 & 55.4 & 50.1 & 47.8 & 45.2 \\
NMT-cor & --- & 64.9 & 61.6 & 57.4 & 54.7 & 55.0 & --- & 59.2 & 55.2 & 51.4 & 48.0 & 47.2 \\
MMT-cor & --- & \bf 65.2 & \bf 62.2 & \bf 59.0 & \bf 56.7 & \bf 55.5 & --- & \bf 59.6 & \bf 56.4 & \bf 52.4 & \bf 50.0 & \bf 48.9 \\
\midrule
\multicolumn{11}{@{}l}{\textit{en-de}} \\
NMT & 52.3 & 47.2 & 44.3 & 40.2 & 38.4 & 36.7 & 47.5 & 43.5 & 40.2 & 36.8 & 34.0 & 32.5 \\
MMT & 52.6 & 47.7 & 45.2 & 41.3 & 39.3 & 37.6 & 47.7 & 43.9 & 41.0 & 37.9 & 35.1 & 33.9 \\
NMT-cor & --- & 47.9 & 45.6 & 42.9 & 41.4 & 41.1 & --- & \bf 44.2 & 41.9 & 38.4 & 36.8 & 36.2 \\
MMT-cor & --- & \bf 48.0 & \bf 46.1 & \bf 43.5 & \bf 42.5 & \bf 41.8 & --- & 43.9 & \bf 42.3 & \bf 39.7 & \bf 38.2 & \bf 37.4 \\
\bottomrule
\end{tabular}
\caption{Results in METEOR scores of models trained and tested on different levels of noisy data. The train and test data are injected with the same proportion of noise. $n$ indicates the max number of noisy words in the train/test set. *-cor indicates the models with error correction training.}
\label{tab:mix_noise}
\end{table*}
\subsection{Training for Robustness to Noise}
To test models for their ability to adapt to noisy data, we \textbf{train models on data with added noise}, sampling from the three types of noise in Section~\ref{sec:noise_injection} and \textbf{test them on noisy test data}, with noise added in the same fashion. METEOR score results are shown in Table \ref{tab:mix_noise}.
The training on noisy data is equivalent to the ``adversarial training'' experiments in previous studies \citep{belinkov2018synthetic,karpukhin-etal-2019-training}. In this setting, a text-only NMT model still suffers from significant performance degradation as the number of noisy words grows, for example dropping from 70.6 METEOR on clean test data to 49.4 under the noisiest setting for en-fr on flickr2017. A drop is also observed for the MMT model, however it is smaller for both language pairs and test sets. As $n$ becomes larger, the gain from the visual context is more obvious, showing that additional context in the form of image features is increasingly important for translation when the quality of the textual input is degraded.
With the addition of the error correction training, both NMT and MMT models further improve their performance, with NMT-cor even outperforming the base MMT model. The MMT-cor model performs better than both NMT-cor and base MMT models, demonstrating that the improvements from error correction and visual cues are complementary. Similar to the benefit from visual features, the difference between models with and without error correction training becomes larger when the noise level increases.
In addition to the performance on noisy texts, another important aspect when measuring robustness is to evaluate whether the performance of the models on clean data is harmed when the model is adapted to the noisy data. Following \citet{karpukhin-etal-2019-training}, we {\bf train models on a mixture of noisy and clean data (0.5/0.5)} and {\bf test them on clean (original) data}.
Table \ref{tab:clean_drop} shows the performance drop on the clean Flickr2017 En-Fr test set, compared to the baseline NMT model trained with clean data only.
\begin{table}[htb]
\centering
\begin{tabular}{l | c c c c c}
\toprule
$n$ = & 1 & 2 & 4 & 6 & 10 \\
\midrule
NMT & $\downarrow$0.2 & $\downarrow$1.0 & $\downarrow$1.4 & $\downarrow$2.0 & $\downarrow$2.3 \\
MMT & $\downarrow$0.2 & $\downarrow$0.7 & $\downarrow$1.7 & $\downarrow$2.1 & $\downarrow$2.4 \\
MMT-cor & $\downarrow$0.0 & $\downarrow$0.4 & $\downarrow$0.9 & $\downarrow$1.7 & $\downarrow$2.1 \\
\bottomrule
\end{tabular}
\caption{Performance drop (the lower the better) on clean Flickr2017 En-Fr test set when models are \textit{trained on mixed data}, compared to baseline NMT model (70.6 METEOR) trained on clean data.}
\label{tab:clean_drop}
\end{table}
The trend is same for models on the other datasets/language pairs: the larger the proportion of noise in the training data, the higher the performance drop on the clean test set.
However, the largest drop in METEOR is only 2.4, showing that mixing clean and noisy training data is a good strategy.\footnote{In additional experiments, we found that models trained on entirely noisy data show much more severe performance drops as \textit{$n$} becomes larger -- see Appendix \ref{sec:train_noise_test_clean}.}
Both MMT and MMT-cor show a similar performance drop to the base NMT model, which indicates that the use of visual context and error correction training does not harm performance on clean texts.
The corresponding results for Table \ref{tab:mix_noise} and \ref{tab:clean_drop} with GRU models can be found in Appendix \ref{sec:gru_results}, showing a similar benefit when using multimodal information and error correction training.
\section{Analysis}
\label{sec:analysis}
\paragraph{Robustness on Unseen Noise}
Since in realistic applications the noise distribution at test time is unknown, we evaluate models using different noise proportions and types at training and test time. For the former, we test the {\em same model} ($n$=4) on various test sets created with different values of $\textit{n}$. For the latter, we test the same model ($n$=4) on the test set where words are randomly replaced with unknown tokens (i.e. ``[UNK]'') to simulate unseen noise (noisy words from different corpora or domains, e.g. new emojis).
Table~\ref{tab:unk} shows results for both cases.
\begin{table}[htb]
\centering
\begin{tabular}{l | c c c c}
\toprule
\textit{$n$} = & 1 & 2 & 6 & 10 \\
\midrule
NMT & 62.5 & 59.3 & 51.6 & 49.2 \\
MMT & 62.9 & 60.1 & 52.8 & 51.0 \\
MMT-cor & {\bf 64.1} & {\bf 62.0} & {\bf 55.5} & {\bf 53.8} \\
\midrule
UNK= & 1 & 2 & 3 & 4 \\
\midrule
NMT & 55.5 & 46.7 & 38.7 & 31.6 \\
MMT & 57.0 & 48.8 & 41.2 & 34.8 \\
MMT-cor & {\bf 57.9} & {\bf 49.9} & {\bf 42.6} & {\bf 36.1} \\
\bottomrule
\end{tabular}
\caption{Performance of NMT and MMT models trained noisy data with $n$=4 but tested on data with different noise proportion and noise types. All models are tested on Flickr2017 En-Fr.}
\label{tab:unk}
\end{table}
The overall trend is similar to the case when the train/test noise are the same: models with visual information and error correction training achieve better performance. The METEOR score of train/test noise proportion mismatch is close to the score in Table~\ref{tab:mix_noise} under the same noise proportio
, showing that the models are robust to unknown noise distributions. As for the evaluation on unknown noise types, the MMT model outperforms the NMT model, which indicates the better ability of the MMT model to handle unseen noise.
\begin{table*}[htb]
\small
\renewcommand{\arraystretch}{0.95}
\centering
\begin{tabular}{cl@{}}
\toprule
\MR{10}{*}{\includegraphics[height=2.5cm]{figure/example1.jpg}} &
\texttt{SRC}: women are playing lacrosse with an orange ball . \\
& \texttt{NSY}: women [art] playing lacrosse with an [strange] ball . \\
&\texttt{NMT}: des femmes jouent au lacrosse avec une balle \false{étrange} . \\
& \texttt{NMT}\textsubscript{cor}: des femmes jouent au lacrosse avec une boule \false{étrange} . \\
& \phantom{\texttt{NMT}\textsubscript{cor}:} \textit{(women are playing lacrosse with a strange ball.)} \\
& \texttt{MMT}\textsubscript{cor}: des femmes jouent \true{à la} lacrosse avec une balle \true{orange} . \\
& \texttt{REF}: des femmes jouent \true{à la} crosse avec une balle \true{orange} . \\
& \phantom{\texttt{NMT}\textsubscript{cor}:} \textit{(women are playing lacrosse with an orange ball .)} \\
& \texttt{COR-NMT}: women \true{are} playing lacrosse with an \false{old} ball . \\
& \texttt{COR-MMT}: women \true{are} playing lacrosse with an \true{orange} ball . \\
\midrule
\MR{10}{*}{\includegraphics[height=2.5cm]{figure/example2.jpg}} &
\texttt{SRC}: a man with his bicycle selling his products on a street \\
& \texttt{NSY}: a [kan] with his [bicycld] selling his products on a street \\
& \texttt{NMT}: un homme avec son \false{casque} vendant ses produits dans une rue \\
& \phantom{\texttt{NMT}:} \textit{(a man with his helmet selling his products on a street)} \\
& \texttt{NMT}\textsubscript{cor}: un homme avec son \true{vélo} vendant ses produits dans une rue \\
& \texttt{MMT}\textsubscript{cor}: un homme avec son \true{vélo} vendant ses produits dans une rue \\
& \texttt{REF:} un homme avec son \true{vélo} vendant ses produits dans une rue \\
& \phantom{\texttt{NMT}\textsubscript{cor}:} \textit{(a man with his bicycle selling his products on a street)} \\
& \texttt{COR-NMT}: a man with his \true{bicycle} selling his products on a street \\
& \texttt{COR-MMT}: a man with his \true{bicycle} selling his products on a street \\
\bottomrule
\end{tabular}
\caption{Qualitative examples for both translation and error correction, where noise is indicated by the words in square brackets. Underlined and bold words highlight the bad and good lexical choices, respectively. \texttt{NSY}: noisy sentence. \texttt{COR-*}: corrected sentence (output from the error correction decoder).}
\label{tbl:imgcomp}
\end{table*}
\paragraph{Visual Sensitivity}
To further probe the effect of the visual information on MMT and MMT-cor models, we apply the \textit{incongruent decoding} evaluation approach \citep{elliott-2018-adversarial,caglayan-etal-2019-probing} by feeding the multimodal models with incorrect visual features at test time, i.e. features taken from a different test sample. The expectation is that the multimodal model will suffer due to the incorrect visual context, performing worse compared to using the correct visual features. Figure \ref{fig:incongruent} shows the performance gap between congruent decoding and incongruent decoding.
\begin{figure}
\centering
\includegraphics[width=.95\linewidth]{figure/tf-incongruent-delta.pdf}
\caption{Performance gap in METEOR score between congruent decoding and incongruent decoding ($\Delta$METEOR = congruent - incongruent).}
\label{fig:incongruent}
\end{figure}
The $\Delta$METEOR is always positive for both MMT and MMT-cor models, and this difference is amplified with a larger noise ratio in the test data, reaching up to 7.2 METEOR scores when $n$=10. We note that the $\Delta$METEOR for the MMT-cor model is similar to the MMT model, but slightly lower, indicating that the error correction training helps the model recover from incorrect image features to a small extent on noisier data.
\paragraph{Error Correction Quality}
To understand whether visual information can also benefit error correction, we compute the span-based correction $F_{0.5}$ score as commonly used in the Grammatical Error Correction task \citep{dahlmeier-ng-2012-better}. The $<$noisy, corrected$>$ and $<$noisy, clean$>$ pairs are first transformed into two lists of edits, where adding/replacing/deleting a word at any position counts as one edit. The evaluation is then performed by calculating the precision/recall/F0.5 between these edit sets.
We report the results in Table \ref{tab:error_correction} for both NMT-cor and MMT-cor models trained on different values of \textit{$n$}. The MMT-cor model outperforms the NMT-cor model, with an improvement of up to +1.7 and +2.6 $F_{0.5}$ on the two test sets. This improvement indicates that visual features can also be beneficial for error correction performance, showing a potential for the task of multimodal error correction, which has yet to be explored.
\begin{table}[htb]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{l | c c c | c c c}
\toprule
& \multicolumn{3}{c|}{flickr2017} & \multicolumn{3}{c}{mscoco2017} \\
& Prec & Rec & $F_{0.5}$ & Prec & Rec & $F_{0.5}$ \\
\midrule
\multicolumn{7}{@{}l}{\textit{$n$}=1} \\
NMT-cor & 41.9 & 52.5 & 43.7 & 45.1 & 51.4 & 46.2 \\
MMT-cor & \bf43.3 & \bf54.0 & \bf45.1 & \bf46.5 & \bf53.8 & \bf47.8 \\
\midrule
\multicolumn{7}{@{}l}{\textit{$n$}=2} \\
NMT-cor & 56.7 & 62.2 & 57.7 & \bf53.2 & 56.2 & \bf53.8 \\
MMT-cor & \bf57.0 & \bf63.3 & \bf58.1 & 52.9 & \bf56.4 & 53.6 \\
\midrule
\multicolumn{7}{@{}l}{\textit{n=4}} \\
NMT-cor & 66.6 & 69.1 & 67.1 & 65.7 & 66.7 & 65.9 \\
MMT-cor & \bf67.7 & \bf71.5 & \bf68.5 & \bf66.1 & \bf67.6 & \bf66.4 \\
\midrule
\multicolumn{7}{@{}l}{\textit{$n$}=6} \\
NMT-cor & 68.7 & 70.0 & 69.0 & 67.0 & 66.1 & 66.8 \\
MMT-cor & \bf70.4 & \bf71.8 & \bf70.7 & \bf68.3 & \bf67.6 & \bf68.2\\
\midrule
\multicolumn{7}{@{}l}{\textit{$n$}=10} \\
NMT-cor & 72.5 & 73.2 & 72.6 & 67.1 & 66.4 & 67.0 \\
MMT-cor & \bf73.9 & \bf74.5 & \bf74.1 & \bf69.8 & \bf68.6 & \bf69.6\\
\bottomrule
\end{tabular}
}
\caption{Error Correction score in $F_{0.5}$ for both NMT-cor and MMT-cor models.}
\label{tab:error_correction}
\end{table}
\begin{figure*}[htb]
\centering
\includegraphics[width=\textwidth]{figure/attention_translation.pdf}
\caption{Attention map of the MMT-cor system on input texts and visual features when generating the translation from noisy input with the target decoder.}
\label{fig:trans_attention}
\end{figure*}
\paragraph{Qualitative Examples}
We provide two qualitative examples of the visual features and error correction training helping the model handle input noise in Table \ref{tbl:imgcomp}
(see Appendix \ref{sec:qualitative_examples} for more examples). In the first example, the source sentence is injected with the ``edit-distance'' noise, with ``are'' and ``orange'' replaced with ``art'' and ``strange'' respectively. Both NMT and NMT-cor models fail to include ``orange'' in the translation, as it is difficult to recover from this error without visual information, while the MMT-cor model is able to generate the correct output. The source sentence in the second example is injected the ``keyboard'' noise, with ``man'' replaced with ``kan'' and ``bicycle'' replaced with ``bicycld''. Although the training data is injected with the same types of noise, the NMT model fails to translate correctly. The reason might be that ``bicycle'' has multiple noisy variants, such as ``bicycld'', ``bocycle'', etc., so the NMT model can hardly learn a strong relationship between ``bicycld'' and ``vélo'' (translation of ``bicycle''). However, the NMT-cor model could relate ``bicycld'' with ``bicycle'', which helps to predict the correct translation ``vélo''.
In Figure \ref{fig:trans_attention}, we also present the attention map of the MMT-cor system when generating the translation. The input is injected with noise by substituting ``sit'' with ``sheet'', and ``wine'' with ``wire''. When generating ``sont assises'' (are sitting), although the attention on the input text still mainly focuses on the noisy word ``sheet'' (with a small proportion focusing on the preposition ``at''), the visual attention is able to focus on the people in the image; therefore, the model obtains the correct information from the visual input and is able to generate the correct translation. Similarly, the model generates ``vin'' (wine) by attending to the glasses in the images and is not distracted by the noisy input word ``wire''.
The attention map for the example when generating the error correction output can be found in Figure \ref{fig:cor_attention} in the Appendix
\section{Conclusions}
\label{sec:conclusion}
In this paper we propose to explore visual cues in order to improve model robustness to noise in machine translation. We combine adversarial training on artificially generated noisy examples with visually-informed multimodal machine translation. By training multimodal models on noisy data, we show that the extra visual context can improve translation robustness on both known and unseen noise. We also propose a novel error correction training method, jointly optimizing the translation model with an auxiliary objective for correcting input errors, which we show can further improve the robustness of both text-only and multimodal translation models. Future work in this area could investigate the integration of further modalities, such as audio in the speech translation setting. In addition to translation, we found that the model using visual features can also help correct errors in the source language. This opens up a promising direction for multimodal monolingual error correction, a task not yet explored.
\section*{Acknowledgements}
Lucia Specia and Zhenhao Li received support from MultiMT project (H2020 ERC Starting Grant No. 678017) and the Air Force Office of Scientific Research (under award number FA8655-20-1-7006).
|
\section{Introduction}
We are interested in solving boundary value problem
{\setlength\arraycolsep{2pt}
\begin{eqnarray}
L_2 y & = & f \ {\rm on}\ (a,b), \label{boundvalue} \\
y(a) & = & \alpha, \ y(b)=\beta, \nonumber
\end{eqnarray}}%
where $L_2$ is differential operator of order $2$, which can be represented in product formulation
\begin{equation} \label{productdifop}
L_2 y:=p_1 D(p_2 D(p_3 y))
\end{equation}
for some $p_1,p_2,p_3$ of constant sign,
with $D$ denoting ordinary derivative, like in \cite{bosnerroginanonuniformtension,bosnerroginasingular,bosnerad}.
For example, each linear second order differential operator
$$
L_2y = y''+p_1 y'+p_2 y
$$
with non constant coefficients, can be represented in the form
$$
L_2 y=(py')'+q y,
$$
where $p(t):=e^{\int_a^tp_1(s)\,ds}$ and $q(t):=p_2(t)p(t)$. By \cite{coppeldisconjugacy}, there exists $\ell>0$ such that on each segment with length smaller than $\ell$, we can find an $u\in\ker{(L_2)}$ without zeros. In that case we have
\begin{equation} \label{productop}
L_2 y=u^{-1}D\bigg(p u^2 D\left(\frac{y}{u}\right)\bigg).
\end{equation}
If our segment $[a,b]$ is such that $b-a<\ell$, we can transform the linear equation in the product form.
Also, if the equation in (\ref{boundvalue}) is linear with constant coefficients, we can use an another approach to do the same thing, which says that there exist a \dword{critical length} $\ell$ (see \cite{carnicermainarpenacritical,carnicerpenacritlengthcycloid,beccaricasciolamazurecritical}) which determines that the kernel of the operator $L_2$ on some interval $[a,b]$ is an Extended Chebyshev space, if the length of $[a,b]$ is less than $\ell$. In that case, it is possible to transform the operator $L_2$ in the product representation like in (\ref{productdifop}) (see \cite{coppeldisconjugacy,mazurefinding}).
Specially, we are aiming here on singular and singularly perturbed boundary value problems.
One way of solving such a problem are the collocation methods \cite{russell1,deBoor:1,ahlbergito,kadalbajofitted,marusicrogina,marusiccollocation,bosnerroginasingular,bosnerad}, in which we are searching for $s\in\mathcal{S}$, where $\mathcal{S}$ is, usually, some spline space, such that
$$
L_2s(\tau_i)=f(\tau_i), \quad i=1,\dots,n,
$$
for some choice of collocation points $\tau_i$, $i=1,\dots,n$, $n:=\dim{(\mathcal{S})}$. In the collocation method, $L_2 s$ is actually interpolated, so next to the interpolation, the idea is to replace it, with some other kind of approximation, to be precise, with the approximation by CCC--Schoenberg operators \cite{bosnerroginaschoenberg} (see also \cite{mazurechebschoenberg}). Something similar for the polynomial splines was done in \cite{FOUCHER20093455}.
The paper is divided in five sections. In Section \ref{sec1} we gathered all facts and theorems we need to define our methods: starting from defining CCC--space and CCC--splines, followed by CCC--Schoenberg operator, and finally interpolation. In the next section we introduce the collocation and quasi-collocation method and give their error bounds. We also suggest two ways of calculating these approximations, and then test them in Section \ref{secexample}.
We end with the conclusion.
\section{Preliminaries} \label{sec1}
The main tool for our method are splines. To define a spline space, at the beginning we have to determine where the pieces of a spline are coming from.
\subsection{CCC--space}
First, we specify the space that we consider. For more details and proofs check \cite{bosnerroginaschoenberg,bosnerphd}.
\begin{definition} \label{defCCC}
Let $u_1$ be a bounded positive function on the interval $[a,b]$, and let there be
given a \dword{measure vector} $d\mbox{\boldmath $\sigma$}:=(d\sigma_2,\ldots
d\sigma_{k})^{\rm T}$, where $\sigma_2,\dots,\sigma_k$ are bounded, continuous and strictly increasing functions on $[a,b]$. Let
{\setlength\arraycolsep{2pt}
\begin{eqnarray}
u_2(x) & = & u_1(x)\int_{a}^x d\sigma_{2}(\tau_{2}), \nonumber \\
& \vdots & \label{CCCs} \\
u_{k}(x) & = & u_1(x)\int_{a}^x d\sigma_{2}(\tau_{2}) \int_{a}^{\tau_2} \ldots \int_{a}^{\tau_{k-1}}
d\sigma_{k}(\tau_{k}), \nonumber
\end{eqnarray}}%
for $x\in[a,b]$. We call $U_k=\{u_i\}_1^k$ a \dword{Canonical Complete Chebyshev (CCC)--system}.
\end{definition}
In the most general case, it is enough to assume that $\sigma_i$ are only right continuous~\cite{schumaker1}. However, in the
sequel we shall use measures that posses densities with respect to the Lebesgue measure (possibly not smooth, or not even continuous), and henceforth assume that
$\sigma_i$ are continuous. As a consequence, every $u_i$, $i=1,\dots,k$ from Definition \ref{defCCC} is continuous. These densities do not have to be strictly positive: they can posses some final number of zeros, and they can even have some poles (see for examples \cite{bosnerroginavariable,bosnerroginasingular}), as long as their distribution function are bounded and strictly increasing.
If for the density $w_i$ of the measure $d\sigma_i$, we have $w_i>0$ and $w_i\in C^{k-i}([a,b])$, and this is true for each $i=2,\dots,k$, then this CCC--system is also an \dword{Extended Complete Chebyshev (ECC)--system} \cite{schumaker1}.
The $i^{\rm th}$ \dword{reduced system} is defined to be a Chebyshev system
corresponding to the \dword{reduced measure vector}, that is
$$
d\mbox{\boldmath $\sigma$}^{(i)} := (d\sigma_{i+2}, \ldots, d\sigma_k)^
{\rm T}
,\quad i=1,\ldots k-2.
$$
We write $U_k^{(i)}:=\{u_{i,j}\}_{j=1}^{k-i}$ for $ i=1,2, \dots ,k-1$, where
{\setlength\arraycolsep{2pt}
\begin{eqnarray*}
u_{i,1}(x) & = & 1, \\
u_{i,2}(x) & = & \int_{a}^x d\sigma_{i+2}(\tau_{i+2}), \\
& \vdots & \\
u_{i,k-i}(x) & = & \int_a^x d\sigma_{i+2}(\tau_{i+2})\ldots \int_a^{\tau_{k-1}}
d\sigma_{k}(\tau_{k}).
\end{eqnarray*}}%
\begin{definition} \label{CCCspace}
The linear space
$\mathcal{S}(k,d\mbox{\boldmath $\sigma$},u_1):={\rm span}\{U_k\}$ is called a \dword{CCC--space} if $U_k$ is a CCC--system.
If $u_1\equiv 1$ then $\mathcal{S}(k,d\mbox{\boldmath $\sigma$}):=\mathcal{S}(k,d\mbox{\boldmath $\sigma$},u_1)$.
\end{definition}
Also, if $U_k$ is an ECC--system, then the space form Definition \ref{CCCspace} is called an \dword{ECC--space}.
\begin{definition} \label{genderiv}
Assuming that each of the functions $\sigma_2$, \dots ,$\sigma_k$ have been extended
to an interval to the left of $a$ and to the right of $b$, \textit{i.e}. to an interval $[c,d]\supset
[a,b]$ with $c<a$, $d>b$,
then for
$f\in\mathcal{S}(k,d\mbox{\boldmath $\sigma$},u_1)$ and $x\in[a,b]$ we define \dword{generalized derivatives} as
$$
L_j:=D_{j}\cdots D_{1}D_0, \quad j=0,\ldots ,k-1,
$$
where
{\setlength\arraycolsep{2pt}
\begin{eqnarray}
D_{0}f(x) & := & \frac{f(x)}{u_1(x)}, \nonumber \\
D_{j}f(x) & := & \lim_{\delta \rightarrow 0^+}{\frac{f(x+\delta)-f(x)}
{\sigma_{j+1}(x+\delta)-\sigma_{j+1}(x)}},\ j=1,\ldots ,k-1. \label{defDj}
\end{eqnarray}}%
\end{definition}
If each $\sigma_i$ posses a density, let us say, $w_i$, {\it i.e.} $d\sigma_i=w_i\,d\lambda$, then
\begin{equation} \label{ECCder}
D_i=\frac{D}{w_{i+1}}, \quad i=2,\dots,n,
\end{equation}
where $D$ stands for the ordinary derivative.
Also, as we mentioned before, $\frac{1}{w_{i+1}}$ does not have to be strictly positive or continuous, and as a consequence we allow functions $p_1$ and $p_2$ in (\ref{productdifop}) to have some zeros or poles.
For an arbitrary measure $d\bar{\sigma}_{k+1}$, we can also define
$$
D_kf(x) := \lim_{\delta \rightarrow 0^+}{\frac{f(x+\delta)-f(x)}
{\bar{\sigma}_{k+1}(x+\delta)-\bar{\sigma}_{k+1}(x)}},
$$
for $x\in[a,b]$, and $L_{k}:=D_{k}\cdots D_{1}D_{0}$. Now it is obvious that $\ker{(L_k)}=\mathcal{S}(k,d\mbox{\boldmath $\sigma$},u_1)$.
Then we have
$$
L_ju_i=\left \{ \begin{array}{ll}
0, & i=1,2, \dots ,j \\
u_{j,i-j}, & i=j+1, \dots ,k,
\end{array} \right. \qquad j=0,1, \dots ,k.
$$
The generalized derivatives can also be calculated for a function from a more general space of functions, like $C[a,b]$,
for which the required limits of the form (\ref{defDj}) exist.
\begin{theorem}
CCC--space $\mathcal{S}(k,d\mbox{\boldmath $\sigma$},u_1)$ is $k$--dimensional linear space, with basis $\{u_1, u_2,$ $ \dots ,u_k\}$, and $L_{j}$
is linear operator
$\mathcal{S}(k,d\mbox{\boldmath $\sigma$},u_1) \rightarrow \mathcal{S}(k-j,d\mbox{\boldmath $\sigma$}^{(j)}):={\rm span}\{U_k^{(j)}\}$ for $j=1, \dots
,k-1$.
\end{theorem}
\subsection{CCC--splines}
For a partition
$\Delta= \{x_{i}\}_{i=0}^{d+1}$ of an interval $[a,b]$,
given \dword{multiplicity vector} $\boldsymbol{m}=(m_{1},\ldots,$ $m_{d})$,
($0 < m_i \leqslant k$), and $M:=\sum_{i=1}^{d} m_i$, we shall denote by
$\mbox{\boldmath $T$}:=\{t_{1} \ldots t_{2k+M}\}$ an \dword{extended partition}
in the usual way:
$$
t_1 = \cdots = t_k = x_0=a,
$$
\begin{equation} \label{extendpart}
(t_{k+1},\dots,t_{k+M}) = (x_1^{[m_1]},\dots,x_d^{[m_d]}),
\end{equation}
$$
b=x_{d+1} = t_{k+M+1} = \cdots = t_{2k+M},
$$
where $x^{[m]}$ stands for $x$ repeated $m$ times.
\begin{definition}
Let $\mathcal{S}(k,d\mbox{\boldmath $\sigma$},u_1)$ be a CCC--space. A set $\mathcal{S}(k, d\mbox{\boldmath $\sigma$},u_1, \mbox{\boldmath $T$})$ of bound\-ed
functions on $[a,b]$ such that:
\begin{itemize}
\item[(i)] for each $s\in\mathcal{S}(k, d\mbox{\boldmath $\sigma$},u_1, \mbox{\boldmath $T$})$ and each $i=0,
\dots ,d$, there exists $s_i\in\mathcal{S}(k,d\mbox{\boldmath $\sigma$},u_1)$ such that
$s|_{[x_i,x_{i+1}]}=s_i|_{[x_i,x_{i+1}]}$,
\item[(ii)]
$L_js_{i-1}(x_i)=L_js_i(x_i)$ for $j=0, \dots ,k-m_i-1$, $i=1, \dots ,d$,
\end{itemize}
is called the \dword{space of Canonical Complete Chebyshev (CCC)--splines on extended partition $\mbox{\boldmath $T$}$}. If $u_1\equiv 1$, then $\mathcal{S}(k, d\mbox{\boldmath $\sigma$}, \mbox{\boldmath $T$}):=\mathcal{S}(k, d\mbox{\boldmath $\sigma$},u_1, \mbox{\boldmath $T$})$.
\end{definition}
It is clear that $\mathcal{S}(k, d\mbox{\boldmath $\sigma$},u_1, \mbox{\boldmath $T$})$ is a linear space, and its dimension is equal to
$$
n:=\dim{\mathcal{S}(k, d\mbox{\boldmath $\sigma$},u_1, \mbox{\boldmath $T$})}=k+M.
$$
For the most convenient basis of the CCC--spline space $\mathcal{S}(k, d\mbox{\boldmath $\sigma$},u_1, \mbox{\boldmath $T$})$ we use the
\dword{B-splines} $T_j^k$, $j=1,\dots,n$
(see~\cite{bosnerroginaschoenberg,schumaker2,schumaker1}).
The CCC--B-splines posses the usual properties, listed in the following theorem.
\begin{theorem} \label{bsplprop}
Let $\mbox{\boldmath $T$}=\{t_i\}_{i=1}^{n+k}$ be an extended partition of $[a,b]$, and let $t_i<t_{i+k}$, $i=1,2,\dots,n$, then
{\setlength\arraycolsep{2pt}
\begin{eqnarray*}
T_i^k(x) & = & 0, \quad x<t_i\ {\rm and}\ x>t_{i+k},\ i=1,2,\dots,n, \nonumber \\
T_i^k(x) & > & 0, \quad t_i<x<t_{i+k},\ i=1,2,\dots,n, \nonumber \\
\sum_{i=1}^n T_i^k(x) & = & u_1, \quad x\in[a,b]. \label{partitonunity}
\end{eqnarray*}}%
\end{theorem}
\begin{remark} \label{partofunity}
Direct consequence of Theorem \ref{bsplprop} is that for $u_1\equiv 1$ the B-splines make partition of unity:
$$
\sum_{i=1}^n T_i^k(x) = 1
$$
for all $x\in[a,b]$.
\end{remark}
Let the B-splines associated with the $i^{\rm th}$
reduced space be denoted as
$T_j^{k-i}$, $i=0,\cdots,k-1$, and let
\begin{equation} \label{integralbspl}
C_j^{k-i}:=\int_{t_j}^{t_{j+k-i}} T_j^{k-i}(t)d\sigma_{i+1}(t).
\end{equation}
It is known that the de\thinspace{}Boor--Cox type recurrence does not exist for the general
Chebyshev splines~\cite{schumaker4}, but in many cases we can use the \dword{derivative formula} (\ref{ederivative})
instead~\cite{roginasingular,roginaderivative,bosnerphd}.
\begin{theorem}[Derivative formula] \label{derivativethm}
Let $L_{1}$ be the first generalized derivative with respect to CCC--space
$\mathcal{S}(k,d\mbox{\boldmath $\sigma$},u_1)$, and let the multiplicity vector $\boldsymbol{m}=(m_1,\dots ,m_l)$ satisfy
$m_i\leqslant k$ for $i=1, \dots, l$. Then for $x\in [a,b]$ and $i=1, \dots, n$,
the following derivative formula holds:
\begin{equation}
L_1T_i^{k}(x) = \frac{T_i^{k-1}(x)}
{C_i^{k-1}}-\frac{T_{i+1}^{k-1}(x)}{C_{i+1}^{k-1}}.
\label{ederivative}
\end{equation}
\end{theorem}
The formula can also be used to inductively define B-splines~\cite{bister}, although for stable calculation of CCC--splines,
the algorithms based on knot insertion are recommended
\cite{bosnerphd,roginafourth,roginabosnerlower,bosnerweighted,bosnerroginanonuniformtension,bosnerad,bosnerroginasingular,bosnerroginavariable}.
The difference between the spaces $\mathcal{S}(k, d\mbox{\boldmath $\sigma$},u_1, \mbox{\boldmath $T$})$ and $\mathcal{S}(k, d\mbox{\boldmath $\sigma$}, \mbox{\boldmath $T$})$ is that the functions from the first one are equal to the functions from the latter multiplied by $u_1$. The same is true for their B-splines.
\begin{theorem} \label{u1jed1}
The B-splines $T_i^k\in\mathcal{S}(k, d\mbox{\boldmath $\sigma$},u_1, \mbox{\boldmath $T$})$ , $\bar{T}_i^k\in\mathcal{S}(k, d\mbox{\boldmath $\sigma$}, \mbox{\boldmath $T$})$ satisfy
$$
T_i^k(x) = u_1(x)\cdot\bar{T}_i^k(x), \qquad i=1,\dots,n,
$$
for all $x\in[a,b]$.
\end{theorem}
\begin{remark} \label{bsompu1}
From the perspective of Theorem \ref{u1jed1}, from now on we can assume, without loss of generality, that $u_1\equiv 1$.
\end{remark}
\subsection{Marsden's identity} \label{sec2}
In the polynomial spline theory, Greville abscis\ae, or knot averages, $\{t_i^*\}_{i=1}^n$
are known as de\thinspace{}Boor points of the identity function, since
$$
x=\sum_{i=1}^n t_i^*B_i^k(x), \qquad t_i^*:=\frac{t_{i+1}+\cdots+t_{i+k-1}}{k-1},
$$
where $\{B_i^k\}_{i=1}^n$ are polynomial B-splines of order $k$ associated with the extended partition
$\{t_i\}_{i=1}^{k+n}$.
In the Chebyshev setting, the function $u_2$ takes over the role of the identity function, and it can be shown that
de\thinspace{}Boor points of $u_2$ can be obtained as the convex combination of $u_2(t_{i+1}),\dots,u_2(t_{i+k-1})$. By Theorem 18 from \cite{bosnerroginaschoenberg} we know how to represent each of the functions $u_j$ as a linear combination of the B-splines
$$
u_j(x)=\sum_{i=1}^{n} \eta_{i,j}^k T_i^k(x), \qquad j=1,2,\dots,k,
$$
and in particular, we know how to calculate $\eta_{i,2}^k$. Since the function $u_2$ is strictly increasing, $(\eta_{i,2}^k)_{i=1}^n$ is strictly increasing, as well as their originals, {\it i.e} the \dword{CCC--Greville points}
$(\xi_{i,2}^k)_{i=1}^n$ ($\eta_{i,2}^k=u_j(\xi_{i,2}^k)$). Even more, we have that $\xi_{i,j}^k\in[t_{i+1},t_{i+k-1}]$.
\subsection{CCC--Schoenberg operators}
In the polynomial case, Greville abscis\ae{}
define classical Schoenberg operator applied to a function $f\in C[a,b]$
$$
S[f]:=\sum_{i=1}^n f(t_i^*)B_i^k,
$$
where $\{B_i^k\}_{i=1}^k$ are as before.
When working with CCC--splines, as we mentioned, the function $u_2$ takes over the role of the identity function, and $(\xi_{i,2}^k)_{i=1}^n$ of the Greville abscis\ae{}. But, the generalization of the Schoenberg operator goes in two directions: first, by replacing the polynomial setting by CCC one, and second, the Greville abscis\ae{} can be replaced by a more general set of points than just $(\xi_{i,2}^k)_{i=1}^n$. For proofs and details see \cite{bosnerroginaschoenberg}.
\begin{definition} \label{Schoenbergdef}
A linear operator $S$ on $C[a,b]$ is called \dword{a CCC--Schoenberg operator based on the CCC--spline space
$\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$})$} if
\begin{itemize}
\item[(i)] it is of the form
\begin{equation} \label{defeqshoenberg}
S[f]:=\sum_{i=1}^n f(\zeta_i)T_i^k
\end{equation}
for all $f\in C[a,b]$, where the nodes $\zeta_1,\dots,\zeta_n$ satisfy $a=\zeta_1<\zeta_2<\cdots<\zeta_{n-1}<\zeta_n=b$,
\item[(ii)] it reproduces a two-dimensional CCC--space $\mathcal{S}_2\subset\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$})$ in the sense
$$
S[f]=f,\quad {\rm for}\ \forall f\in\mathcal{S}_2.
$$
\end{itemize}
\end{definition}
Because the B-splines satisfy the partition of unity, the operator $S$ defined by (\ref{defeqshoenberg}) must reproduce
the constants. Also, since $\mathcal{S}_2$ is a CCC--space, there exists a continuous, strictly increasing distribution function $\bar{\sigma}_2$ such that
$\mathcal{S}_2=\mathcal{S}(2,d\bar{\mbox{\boldmath $\sigma$}}_2)$, with $d\bar{\mbox{\boldmath $\sigma$}}_2:=(d\bar{\sigma}_2)^{\rm T}$. That means that actually $\mathcal{S}_2={\rm span}\{1,\bar{\sigma}_2\}\subset
\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$})$, and $\bar{\sigma}_2=\sum_{i=1}^n \nu_iT_i^k$ for some $\nu_1,\dots,\nu_n$.
The following theorem determines which strictly increasing splines can be reproduced by a CCC--Schoenberg operator.
\begin{theorem} \label{strictdeboor}
For a given continuous spline $s:=\sum_{i=1}^n s_iT_i^k\in\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$})$, the following two properties are equivalent:
\begin{itemize}
\item[(i)] the spline $s$ is strictly increasing on $[a,b]$ and it is reproduced by a unique CCC--Schoenberg
operator based on $\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$})$;
\item[(ii)] the de\thinspace{}Boor points $s_i$, $i=1,\dots,n$, of $s$ form a strictly increasing sequence.
\end{itemize}
\end{theorem}
\begin{remark}
\begin{itemize}
\item[(i)] The connection between $(s_i)_{i=1}^n$ from Theorem \ref{strictdeboor} and $(\zeta_i)_{i=1}^n$ from (\ref{defeqshoenberg}) is simple:
$$
s(\zeta_i)=s_i, \qquad i=1,\dots,n.
$$
\item[(ii)] It can be shown that $\zeta_i\in[t_{i+1},t_{i+k-1}]$.
\end{itemize}
\end{remark}
Let $\mathcal{S}_2={\rm span}\{1,\bar{\sigma}_2\}$, where $\bar{\sigma}_2\in\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$})$ satisfies the property (ii) of Theorem
\ref{strictdeboor}, $\bar{\sigma}_2\in C[a,b]$, and let the CCC--Schoenberg operator
$$
S[f]:=\sum_{i=1}^n f(\zeta_i)T_i^k
$$
reproduces the CCC--space $\mathcal{S}_2$. The question that arises is how well $S[f]$ approximates a function $f$
satisfying some mild conditions.
Let us for $f\in C[a,b]$ define
$$
\bar{L}_1f(x):=\lim_{\delta\to 0^{+}}\frac{f(x+\delta)-f(x)}{\bar{\sigma}_2(x+\delta)-\bar{\sigma}_2(x)},
$$
and similarly as for dual derivatives
(see the (3) from \cite{bosnerroginaschoenberg}), derivative
$$
\bar{L}_1^- f(x):=\lim_{\delta\to 0^{+}}\frac{f(x)-f(x-\delta)}{\bar{\sigma}_2(x)-\bar{\sigma}_2(x-\delta)}
$$
as well as
$$
\bar{L}_2:=\bar{D}_2\bar{L}_1, \qquad \bar{L}_2^-:=\bar{D}_2^{-}\bar{L}_1^-,
$$
where
$$
\bar{D}_2 f(x):=\lim_{\delta\to 0^{+}}\frac{f(x+\delta)-f(x)}{\bar{\sigma}_3(x+\delta)-\bar{\sigma}_3(x)}, \qquad
\bar{D}_2^- f(x):=\lim_{\delta\to 0^{+}}\frac{f(x)-f(x-\delta)}{\bar{\sigma}_3(x)-\bar{\sigma}_3(x-\delta)},
$$
for an arbitrary $d\bar{\sigma}_3$.
Finally, let $\Lambda_{(d\bar{\sigma}_2,d\bar{\sigma}_3)}([a,b])$ be the space of continuous
function on $[a,b]$ such that
\begin{itemize}
\item[(i)] $\bar{L}_1f(x)=\bar{L}_1^-f(x)$, for $x\in[a,b]$,
\item[(ii)] $\bar{L}_2f$ and $\bar{L}_2^-f$ are bounded and integrable on $[a,b]$ with respect to the $d\bar{\sigma}_3$.
\end{itemize}
If we have (\ref{ECCder}), and if the distribution function $\bar{\sigma}_2$ is smooth, as we mentioned in \cite{bosnerroginaschoenberg}, it should not be a big problem to check whether $f\in\Lambda_{(d\bar{\sigma}_2,d\bar{\sigma}_3)}([a,b])$.
Let
$$
\|f\|_\infty:=\max_{a\leqslant x\leqslant b}{|f(x)|},
$$
then for the approximation by CCC--Schoenberg operator to a function we can state the
following theorem.
\begin{theorem} \label{thmvdquadr}
Let $f\in\Lambda_{(d\bar{\sigma}_2,d\bar{\sigma}_3)}([a,b])$, where $\bar{\sigma}_2\in\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$})$ is continuous and satisfies (ii) of Theorem
\ref{strictdeboor}, while
$d\bar{\sigma}_3$ is arbitrary. Let $S$ be the CCC--Schoenberg operator which reproduces $\mathcal{S}_2={\rm span}\{1,\bar{\sigma}_2\}$.
Then
$$
\|f-S[f]\|_\infty\leqslant C \bar{h}_{(d\bar{\sigma}_2,d\bar{\sigma}_3,\mbox{\boldmath $T$})}^2
$$
where
$$
\bar{h}_{(d\bar{\sigma}_2,d\bar{\sigma}_3,\mbox{\boldmath $T$})}:=\max_{0\leqslant i\leqslant d}\max{\{\bar{\sigma}_2(x_{i+1})-\bar{\sigma}_2(x_i),\bar{\sigma}_3
(x_{i+1})-\bar{\sigma}_3(x_i)\}},
$$
and $C$ does not depend on $\mbox{\boldmath $T$}$.
\end{theorem}
\subsection{Interpolation}
Although, the interpolation problem in Chebyshev setting is well known, like for an example in \cite{mulbachinterpol,schumaker1,kochlycheinterpol,kayumovmazure,kvasovkimweight,CARNICER2014interpol,mazureinterpvsdesign}, we will pay some more attention to it, especially to the error bounds.
We follow the ideas from \cite{deboor1}. At the beginning, we will assume that all multiplicities of the interior knots are less then $k$ ($m_i<k$, $i=1,\dots,d$), so that $\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$})\subset C([a,b])$.
We observe the following interpolation problem. For the function $f\in C([a,b])$, and the given interpolating points $(\tau_i)_{i=1}^n$, we search for a spline
$$
s(x)=\sum_{i=1}^n c_iT_i^k(x),
$$
$s\in\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$})$, such that $s(\tau_i)=f(\tau_i)$. The \dword{interpolation operator} $I$ is such that $I[f]:=s$.
Let us denote with $A$ the interpolation matrix $A:=[T_j^k(\tau_i)]_{i,j}$. The same proof of the Schoenberg--Whitney Theorem for the polynomial case, works also for any CCC--spline space, so the matrix $A$ is regular if $T_i^k(\tau_i)\ne0$, $i=1,\dots,n$. Also, because of the support property of the CCC--B-splines, the interpolation matrix is banded, and because of Theorem 16 and 17 from \cite{bosnerroginaschoenberg}, $A$ is totally positive.
Now, we would like to give some error bounds of the interpolating spline. Let $f\in C([a,b])$, and we start with the fact that $I[g]=g$ for any $g\in\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$})$, and therefore
{\setlength\arraycolsep{2pt}
\begin{eqnarray*}
\|f-I[f]\|_\infty & \leqslant & \|f-g\|_\infty+\|I[f-g]\|_\infty\leqslant\|f-g\|_\infty+\|I\|_\infty\|f-g\|_\infty \nonumber \\
& = & (1+\|I\|_\infty)\|f-g\|_\infty.
\end{eqnarray*}}%
By choosing the closest spline from $\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$})$ to function $f$, we get
\begin{equation} \label{interpolgrocj1}
\|f-I[f]\|_\infty \leqslant (1+\|I\|_\infty)\,{\rm dist}\,(f,\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$})).
\end{equation}
The calculation of bounds for ${\rm dist}\,(f,\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$}))$ follows mostly the derivation from \cite{deboor1}, with one exception, so we will concentrate more on this part.
Let the operator $B$ be defined with
$$
B[g]:=\sum_{i=1}^n g(\tau_i)T_i^k,
$$
for some $\tau_1\leqslant\cdots\leqslant\tau_n$ and any continuous function $g$. This operator reproduces constants. Then for $x\in[t_j,t_{j+1}\rangle$
we have
{\setlength\arraycolsep{2pt}
\begin{eqnarray*}
|f(x)-B[f](x)| & \leqslant & \sum_{i=j+1-k}^j |f(x)-f(\tau_i)|T_i^k(x) \\
& \leqslant & \max_{j-k+1\leqslant i\leqslant j}{\{|f(x)-f(\tau_i)|\}},
\end{eqnarray*}}%
then, by choosing the same $(\tau_i)_{i=1}^n$ as in \cite{deboor1},
we get, as before,
{\setlength\arraycolsep{2pt}
\begin{eqnarray}
\max_{j-k+1\leqslant i\leqslant j}\!\! & \{ & \!\!\!|f(x) - f(\tau_i)|\} \nonumber \\
& \leqslant & \max\{|f(y)-f(z)|: y,z\in[t_{j+1-\frac{k}{2}},t_{j+1}]\ {\rm or} \nonumber \\
& & y,z\in[t_j,t_{j+\frac{k}{2}}]\} \label{disttoder} \\
& \leqslant & \left\lfloor \frac{k+1}{2} \right\rfloor\omega(f,\mbox{\boldmath $h$}_1), \nonumber
\end{eqnarray}}%
where
$$
\mbox{\boldmath $h$}_1:=\max_{0\leqslant i\leqslant d}{\{x_{i+1}-x_i\}},
$$
and $\omega$ is modulus of continuity. Therefore it is again
\begin{equation} \label{distmodul1}
{\rm dist}\,(f,\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$}))\leqslant{\rm const}\, \omega(f,\mbox{\boldmath $h$}_1).
\end{equation}
But, if the function $f$ has piecewise continuous $L_1f$, then we proceed differently. By the generalized Taylor expansion from Theorem 8 \cite{bosnerroginaschoenberg}, and the Mean value theorem for Lebesgue--Stieljes integrals (see \cite{burrill}) which says that there exists $d_{y,z}$, with $\min_{[y,z]}{|L_1f|}\leqslant d_{y,z}\leqslant\max_{[y,z]}{|L_1f|}$,
such that from (\ref{disttoder}) it follows
{\setlength\arraycolsep{2pt}
\begin{eqnarray*}
\max_{j-k+1\leqslant i\leqslant j}\!\! & \{ & \!\!\!|f(x) - f(\tau_i)|\} \nonumber \\
& = & \max{\left\{\left|\int_y^z L_1f(\tau_2)d\sigma_2(\tau_2)\right| :y,z\in[t_{j+1-\frac{k}{2}},t_{j+1}]\ {\rm or}\ y,z\in[t_j,t_{j+\frac{k}{2}}]\right\}} \nonumber \\
& \leqslant & \max{\left\{\left|\int_y^z |L_1f(\tau_2)|d\sigma_2(\tau_2)\right| :y,z\in[t_{j+1-\frac{k}{2}},t_{j+1}]\ {\rm or}\ y,z\in[t_j,t_{j+\frac{k}{2}}]\right\}} \nonumber \\
& = & \max{\left\{d_{y,z}\left|\int_y^z d\sigma_2(\tau_2)\right| :y,z\in[t_{j+1-\frac{k}{2}},t_{j+1}]\ {\rm or}\ y,z\in[t_j,t_{j+\frac{k}{2}}]\right\}} \nonumber \\
& \leqslant & \|L_1f\|_\infty
\max{\left\{\left|\int_y^z d\sigma_2(\tau_2)\right| :y,z\in[t_{j+1-\frac{k}{2}},t_{j+1}]\ {\rm or}\ y,z\in[t_j,t_{j+\frac{k}{2}}]\right\}} \nonumber \\
& = & \|L_1f\|_\infty \max{\{\sigma_2(t_{j+1})-\sigma_2(t_{j+1-\frac{k}{2}}),\sigma_2(t_{j+\frac{k}{2}})-\sigma_2(t_j)\}} \nonumber \\
& \leqslant & \|L_1f\|_\infty \left\lfloor \frac{k+1}{2} \right\rfloor\mbox{\boldmath $h$}_2,
\end{eqnarray*}}%
since $\sigma_2$ is strictly increasing, and with
$$
\mbox{\boldmath $h$}_2:=\max_{0\leqslant i\leqslant d}{\{\sigma_2(x_{i+1})-\sigma_2(x_i)\}}.
$$
All together, we get
\begin{equation} \label{nejednl1g}
{\rm dist}\,(f,\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$}))\leqslant{\rm const}\, \mbox{\boldmath $h$}_2 \|L_1f\|_\infty.
\end{equation}
Now,
$$
{\rm dist}\,(f,\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$}))={\rm dist}\,(f-s,\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$})), \quad \forall s\in\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$}),
$$
and from (\ref{distmodul1}) we get
$$
{\rm dist}\,(f,\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$}))\leqslant{\rm const}\, \omega(f-s,\mbox{\boldmath $h$}_1), \quad \forall s\in\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$}).
$$
If $f$ and $s$ have piecewise continuous first generalized derivatives, then from (\ref{nejednl1g}) follows
$$
{\rm dist}\,(f,\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$}))\leqslant{\rm const}\,\mbox{\boldmath $h$}_2\|L_1f-L_1s\|,
$$
for all $s\in\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$})$. Again, we can choose $L_1s$ to be closest to the $L_1f$ in $\mathcal{S}(k-1,d\mbox{\boldmath $\sigma$}^{(1)},\mbox{\boldmath $T$})$, to obtain the estimate
$$
{\rm dist}\,(f,\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$}))\leqslant{\rm const}\,\mbox{\boldmath $h$}_2\,{\rm dist}\,(L_1f,\mathcal{S}(k-1,d\mbox{\boldmath $\sigma$}^{(1)},\mbox{\boldmath $T$})).
$$
In case $L_1f$ is continuous, we could use (\ref{distmodul1}) again to get
$$
{\rm dist}\,(f,\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$}))\leqslant{\rm const}\,\mbox{\boldmath $h$}_2\omega(L_1f,\mbox{\boldmath $h$}_1).
$$
If $f$ has even continuous second generalized derivative, and if $k>2$, we can proceed in the same way to achieve
$$
{\rm dist}\,(f,\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$}))\leqslant{\rm const}\,\mbox{\boldmath $h$}_2\mbox{\boldmath $h$}_3\omega(L_2f,\mbox{\boldmath $h$}_1),
$$
with generally
$$
\mbox{\boldmath $h$}_j:=\max_{0\leqslant i\leqslant d}{\{\sigma_j(x_{i+1})-\sigma_j(x_i)\}}, \quad j=2,\dots,k.
$$
Let us just remark that all of the ``const'' in previous inequalities are not necessarily the same.
That way we have proven the Jackson type theorem for CCC--splines.
\begin{theorem} \label{tmjackson}
For $j=0,\dots,k-1$, there exists constant {\rm const}, which depends on $k$ and $j$, such that for all $f$ with continuous $L_if$, $i=0,\dots,j$ on $[a,b]$
$$
{\rm dist}\,(f,\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$}))\leqslant{\rm const}\,\mbox{\boldmath $h$}_2\cdots\mbox{\boldmath $h$}_{j+1}\omega(L_jf,\mbox{\boldmath $h$}_1),
$$
and for $j=k-1$ if $f$ has also continuous $L_kf$ for some arbitrary measure $d\bar{\sigma}_{k+1}$, then
$$
{\rm dist}\,(f,\mathcal{S}(k,d\mbox{\boldmath $\sigma$},\mbox{\boldmath $T$}))\leqslant{\rm const}\mbox{\boldmath $h$}_2\cdots\mbox{\boldmath $h$}_k\bar{\mbox{\boldmath $h$}}_{k+1}\|L_kf\|_\infty,
$$
with
$$
\bar{\mbox{\boldmath $h$}}_{k+1}:=\max_{0\leqslant i\leqslant d}{\{\bar{\sigma}_{k+1}(x_{i+1})-\bar{\sigma}_{k+1}(x_i)\}}.
$$
\end{theorem}
Now we go back to (\ref{interpolgrocj1}), where
we can also say a little bit more about
$$
\|I\|_\infty:=\max{\left\{ \frac{\|I[g]\|_\infty}{\|g\|_\infty}:g\in C([a,b])\backslash \{0\} \right\}}.
$$
Let say that our interpolating spline of function $g$ is
$$
s(x):=I[g](x)=\sum_{i=1}^nc_iT_i^k(x),
$$
with $c:=[c_1,\dots,c_n]^{\rm T}=A^{-1}d$, where $d:=[g(\tau_1),\dots,g(\tau_n)]^{\rm T}$. Then
{\setlength\arraycolsep{2pt}
\begin{eqnarray*}
\|I[g]\|_\infty & = & \|\sum_{i=1}^nc_iT_i^k\|_\infty \leqslant \max_{x\in[a,b]}\sum_{i=1}^n|c_i|T_i^k(x) \\
& \leqslant & \|c\|_\infty\sum_{i=1}^n T_i^k(x) = \|c\|_\infty = \|A^{-1}d\|_\infty \\
& \leqslant & \|A^{-1}\|_\infty\|d\|_\infty\leqslant \|A^{-1}\|_\infty\|g\|_\infty,
\end{eqnarray*}}%
for all $g\in C([a,b])\backslash \{0\}$, so
$$
\|I\|_\infty \leqslant \|A^{-1}\|_\infty.
$$
Since $\|A\|_\infty=1$, because the B-splines make partition of unity, for the condition number of $A$ we have $\kappa_\infty(A)=\|A^{-1}\|_\infty$, and
$$
\|I\|_\infty \leqslant \kappa_\infty(A).
$$
Like with polynomial splines, $\kappa_\infty(A)$ depends on the choice of the interpolation points $(\tau_i)_{i=1}^n$, but if we choose them like the nodes of the CCC--Schoenberg operator, to fulfil $\tau_i\in[t_{i+1},t_{i+k-1}]$, and not too close to each other, like for an example the CCC--Greville points, we expect $\kappa_\infty(A)$ to be decent.
\section{Collocation and quasi-collocation}
Now we are back to the problem (\ref{boundvalue}), with $f\in C[a,b]$, but we will specify the operator $L_2$ with
\begin{equation} \label{defL2}
L_2=D_2 D_1 D_0
\end{equation}
where $D_0$, $D_1$ and $D_2$ are defined as in (\ref{defDj}) for some $\sigma_2$, $\sigma_3$ and $u_1$, and from now on, we cancel the constraint $u_1\equiv 1$. Let $\mathcal{S}_2:=\ker{(L_2)}={\rm span}\,{\{u_1,u_2\}}$ with $u_2$ as in (\ref{CCCs}).
For initial value problems, we already have the Green's function, given by Theorem 7 from \cite{bosnerroginaschoenberg}, the same we use to define the B-splines themselves.
In our case, as for the linear differential operator, we can construct the Green's function also for the boundary value problems (like for tension splines in \cite{marusicroginasharp}).
\begin{theorem} \label{Gaussbv}
The solution to the problem (\ref{boundvalue}), with (\ref{defL2}), is of the form
\begin{equation} \label{solutionGreeen}
y(x)=u(x)+u_1(x)\int_a^b G(x,\tau)f(\tau)\,d\sigma_3(\tau),
\end{equation}
for $x\in[a,b]$, where $u\in\mathcal{S}_2$ is such that $u(a)=\alpha$ and $u(b)=\beta$,
and $G$ is the generalized Green's function:
$$
G(x,y)=\left\{ \begin{array}{ll}
\displaystyle -\frac{\int_a^y d\sigma_2 \int_x^b d\sigma_2}{\int_a^b d\sigma_2}, & a\leqslant y\leqslant x\leqslant b \\[1em]
\displaystyle -\frac{\int_a^x d\sigma_2 \int_y^b d\sigma_2}{\int_a^b d\sigma_2}, & a\leqslant x\leqslant y\leqslant b
\end{array} \right. .
$$
\end{theorem}
\begin{proof}
The unique function $u$ from (\ref{solutionGreeen}) always exists by Remark 1 from \cite{bosnerroginaschoenberg}.
First, we want to prove that $y$ satisfies the differential equation. Since
{\setlength\arraycolsep{2pt}
\begin{eqnarray}
y(x) & = & u(x)-u_1(x)\frac{\int_x^b d\sigma_2}{\int_a^b d\sigma_2} \int_a^x f(\tau)\,d\sigma_3(\tau) \int_a^\tau d\sigma_2 \nonumber \\
& & -u_1(x)\frac{\int_a^x d\sigma_2}{\int_a^b d\sigma_2} \int_x^b f(\tau)\,d\sigma_3(\tau) \int_\tau^b d\sigma_2, \label{yxexpr}
\end{eqnarray}}%
and since the same rule for the generalized derivative of the product of functions is valid, as for the ordinary derivative, we have
{\setlength\arraycolsep{2pt}
\begin{eqnarray*}
L_2 y(x) & = & \frac{1}{\int_a^b d\sigma_2} D_2\bigg[-\left(D_1\int_x^b d\sigma_2 \right)\cdot \int_a^x f(\tau)d\sigma_3(\tau) \int_a^\tau d\sigma_2 \\
& & -\int_x^b d\sigma_2 \cdot D_1\left(\int_a^x f(\tau)d\sigma_3(\tau) \int_a^\tau d\sigma_2 \right) \\
& & -\left(D_1\int_a^x d\sigma_2\right)\cdot\int_x^b f(\tau)\,d\sigma_3(\tau) \int_\tau^b d\sigma_2 \\
& & -\int_a^x d\sigma_2 \cdot D_1\left(\int_x^b f(\tau)\,d\sigma_3(\tau) \int_\tau^b d\sigma_2 \right) \bigg] \\
& = & \frac{1}{\int_a^b d\sigma_2} D_2\bigg[\int_a^x f(\tau)d\sigma_3(\tau) \int_a^\tau d\sigma_2 \\
& & -\int_x^b d\sigma_2 \cdot D_1\left(\int_a^x f(\tau)d\sigma_3(\tau) \int_a^\tau d\sigma_2 \right) -\int_x^b f(\tau)\,d\sigma_3(\tau) \int_\tau^b d\sigma_2 \\
& & -\int_a^x d\sigma_2 \cdot D_1\left(\int_x^b f(\tau)\,d\sigma_3(\tau) \int_\tau^b d\sigma_2 \right)\bigg].
\end{eqnarray*}}%
Because by the Mean value theorem for Lebesgue--Stieltjes integrals
$$
D \int_a^x g(\tau)\, d\sigma(\tau) = \lim_{\delta \to 0^+}\frac{\int_x^{x+\delta} g(\tau)\, d\sigma(\tau)}{\int_x^{x+\delta}d\sigma(\tau)} = \lim_{\delta \to 0^+}\frac{g(\xi_x) \int_x^{x+\delta} d\sigma(\tau)}{\int_x^{x+\delta}d\sigma(\tau)} = g(x)
$$
for any measure $\sigma$, any continuous function $g$ for which the above integrals exist, and some $\xi_x\in(x,x+\delta)$, it follows that
{\setlength\arraycolsep{2pt}
\begin{eqnarray}
L_2 y(x) & = & \frac{1}{\int_a^b d\sigma_2} \bigg[f(x)\int_a^x d\sigma_2 \\
& & -\left(D_2\int_x^b d\sigma_2\right)\cdot D_1\left(\int_a^x f(\tau)d\sigma_3(\tau) \int_a^\tau d\sigma_2 \right) \nonumber \\
& & -\int_x^b d\sigma_2 \cdot D_2D_1\left(\int_a^x f(\tau)d\sigma_3(\tau) \int_a^\tau d\sigma_2 \right)+f(x)\int_x^b d\sigma_2 \nonumber \\
& & -\left(D_2\int_a^x d\sigma_2\right) \cdot D_1\left(\int_x^b f(\tau)\,d\sigma_3(\tau) \int_\tau^bd\sigma_2\right) \label{l2yexpr} \\
& & -\int_a^x d\sigma_2 \cdot D_2D_1\left(\int_x^b f(\tau)\,d\sigma_3(\tau) \int_\tau^bd\sigma_2\right)\bigg]. \nonumber
\end{eqnarray}}%
According to the definition of the generalized derivatives, it is trivial to see that
$$
D_1g(x)=D_2g(x)\cdot D_1\sigma_3(x) \quad {\rm and} \quad D_2g(x)=D_1g(x)\cdot D_2\sigma_2(x),
$$
which yields
{\setlength\arraycolsep{2pt}
\begin{eqnarray}
\left(D_2\int_x^b d\sigma_2\right)\!\!\! & \cdot & \!\!\!D_1\left(\int_a^x f(\tau)d\sigma_3(\tau) \int_a^\tau d\sigma_2 \right) \nonumber \\
& + & \left(D_2\int_a^x d\sigma_2\right) \cdot D_1\left(\int_x^b f(\tau)\,d\sigma_3(\tau) \int_\tau^b d\sigma_2\right) \nonumber \\
& = & -D_2\sigma_2(x)\cdot D_2\left(\int_a^x f(\tau)d\sigma_3(\tau) \int_a^\tau d\sigma_2 \right)\cdot D_1\sigma_3(x) \nonumber \\
& & +D_2\sigma_2(x)\cdot D_2\left(\int_x^b f(\tau)\,d\sigma_3(\tau) \int_\tau^b d\sigma_2\right)\cdot D_1\sigma_3(x) \label{l2y1} \\
& = & -D_2\sigma_2(x)\cdot f(x)\int_a^x d\sigma_2\cdot D_1\sigma_3(x) \nonumber \\
& & -D_2\sigma_2(x)\cdot f(x)\int_x^bd\sigma_2\cdot D_1\sigma_3(x) \nonumber \\
& = & -D_2\sigma_2(x)\cdot D_1\sigma_3(x)\cdot f(x)\int_a^b d\sigma_2, \nonumber
\end{eqnarray}}%
and in the same way it can be shown that
{\setlength\arraycolsep{1pt}
\begin{eqnarray}
\int_x^b d\sigma_2 & \cdot & D_2D_1\left(\int_a^x f(\tau)d\sigma_3(\tau) \int_a^\tau d\sigma_2 \right) + \int_a^x d\sigma_2 \label{l2y2} \\
& \cdot & D_2D_1\left(\int_x^b f(\tau)\,d\sigma_3(\tau) \int_\tau^bd\sigma_2\right)
= D_2\sigma_2(x)\cdot D_1\sigma_3(x)\cdot f(x)\int_a^b d\sigma_2. \nonumber
\end{eqnarray}}%
When we put (\ref{l2y1}) and (\ref{l2y2}) back in (\ref{l2yexpr}), we get
$$
L_2y(x)=f(x).
$$
Trivially we can show that $y$ expressed by (\ref{yxexpr}) satisfies the boundary conditions.
\end{proof}
As we mentioned before, the collocation method corresponds to the interpolation problem of the function $f$, so we seek the spline $s\in\mathcal{S}$ from some spline space $\mathcal{S}$, such that
\begin{equation} \label{coloceq}
L_2s=I[f],
\end{equation}
but we offer also an another approach,
to find $s\in\mathcal{S}$ such that
\begin{equation} \label{quasicoloceq}
L_2s = S[f],
\end{equation}
where both, the interpolation operator and the CCC--Schoenberg operator, are associated with the second reduced system.
To be precise, we take the measures $d\sigma_2$ and $d\sigma_3$, which define the operator $L_2$, as the first two measures in the measure vector $d\mbox{\boldmath $\sigma$}$ associated with $\mathcal{S}$, and we add to them at least one more arbitrary measure. That way, if $\mathcal{S}(k,d\mbox{\boldmath $\sigma$},u_1)$ denotes the underlying CCC--space, the CCC--space associated with the second reduced system $\mathcal{S}(k-2,d\mbox{\boldmath $\sigma$}^{(2)})$ is at least two-dimensional. We will discus the possible choice of this added measures a little bit later.
For the spline space, we take the extend partition $\mbox{\boldmath $T$}$ with all interior knots of multiplicity one, to assure the splines space associated with the second reduced space is at least continuous. Now, our spline space is $\mathcal{S}:=\mathcal{S}(k,d\mbox{\boldmath $\sigma$},u_1,\mbox{\boldmath $T$})$, so the operators $I$ in (\ref{coloceq}) and $S$ in (\ref{quasicoloceq}) are from the spline space $\mathcal{S}^{(2)}:=\mathcal{S}(k-2,d\mbox{\boldmath $\sigma$}^{(2)},\mathbf{1},\mbox{\boldmath $T$})$.
Mostly we use the CCC--Greville points ($\xi_{i,2}^k$) for the CCC--Schoenberg operator, since we know how to calculate them, and although this calculation is complex, for the quasi-collocation we need these Greville points for the spline space of order $k-2$, which is usually not very high. To calculate them, for an example, for order 3 or 4 is often not that complicated. We suggest also to use them as collocation points.
The next question is how good these two approximations are. The answer gives the next theorem.
\begin{theorem} \label{bvquadconv}
Let $f$ from (\ref{boundvalue}), $\bar{\sigma}_2$ and $\bar{\sigma}_3$ satisfy the conditions of Theorem \ref{thmvdquadr}, only for the spline space $\mathcal{S}^{(2)}$ instead. Then for the the spline $s\in\mathcal{S}$ for which (\ref{quasicoloceq}) holds, we have
\begin{equation} \label{bvquadconveq}
\|s-y\|_\infty\leqslant \bar{C}_S \bar{h}_{(d\bar{\sigma}_2,d\bar{\sigma}_3,\mbox{\boldmath $T$})}^2,
\end{equation}
where $\bar{C}_S$ does not depend on $\mbox{\boldmath $T$}$. If $f$ is such that $L_{2,i}f$ is continuous on $[a,b]$ for $i=0,1,\dots,k-2$, where the generalized derivatives $L_{2,i}$ are associated with the second reduced measure vector $d\mbox{\boldmath $\sigma$}^{(2)}=(d\sigma_4,\dots,d\sigma_k)^{\rm T}$,
then $s\in\mathcal{S}$ for which (\ref{coloceq}) holds satisfies
\begin{equation} \label{bvcolocconverg}
\|s-y\|_\infty\leqslant \bar{C}_I \mbox{\boldmath $h$}_4\cdots\bar{\mbox{\boldmath $h$}}_{k+1}\|L_{2,k-2}f\|_\infty,
\end{equation}
for an arbitrary measure $d\bar{\sigma}_{k+1}$, where $\bar{C}_I$ does not depend on $\mbox{\boldmath $T$}$.
\end{theorem}
\begin{proof}
Let the operator $Q$ be equal to either $I$ or $S$.
According to Theorem \ref{Gaussbv}, for both, the exact solution and the approximation, we have
{\setlength\arraycolsep{2pt}
\begin{eqnarray}
y(x) & = & u(x)+u_1(x)\int_a^b G(x,\tau)f(\tau)\,d\sigma_3(\tau) \nonumber \\
s(x) & = & u(x)+u_1(x)\int_a^b G(x,\tau)L_2s(\tau)\,d\sigma_3(\tau) \nonumber \\
& = & u(x)+u_1(x)\int_a^b G(x,\tau)Q[f](\tau)\,d\sigma_3(\tau) \label{sxGreen}
\end{eqnarray}
for some $x\in[a,b]$. Therefore
{\setlength\arraycolsep{2pt}
\begin{eqnarray*}
|s(x)-y(x)| & = & \left|u_1(x)\int_a^b G(x,\tau)\big(Q[f](\tau)-f(\tau)\big)d\sigma_3(\tau)\right| \\
& \leqslant & |u_1(x)|\,|Q[f](\xi)-f(\xi)|\,\left|\int_a^b G(x,\tau)\,d\sigma_3(\tau)\right| \\
& \leqslant & \|u_1\|_\infty \|Q[f]-f\|_\infty \left| \int_a^x G(x,\tau)\,d\sigma_3(\tau)+\int_x^b G(x,\tau)\,d\sigma_3(\tau)\right| \\
& \leqslant & 2\int_a^b d\sigma_2\int_a^b d\sigma_3 \, \|u_1\|_\infty \|Q[f]-f\|_\infty,
\end{eqnarray*}}%
where we used the Mean value theorem for Lebesgue--Stieltjes integrals again.
Theorem \ref{thmvdquadr} gives then (\ref{bvquadconveq}), while (\ref{interpolgrocj1}) and Theorem \ref{tmjackson}, the inequality (\ref{bvcolocconverg}).
\end{proof}
\begin{remark} \label{remarkcolqcol}
Actually, we could generalize Theorem \ref{bvquadconv} to any operator $Q$ for which we know the error bounds.
If we use the CCC--Greville points for the Schoenberg operator, the error bound is $O(\mbox{\boldmath $h$}_4\mbox{\boldmath $h$}_5)$.
According to Theorem \ref{tmjackson}, if the function $f$ is less smooth, concerning generalized derivatives, the error bound of the collocation method would be of lower order (``order'' here means the number of factors $\mbox{\boldmath $h$}_j$), but the same would be true for the quasi-collocation method if the second generalized derivative of $f$ in $\mathcal{S}_2$ is not at least piecewise continuous.
\end{remark}
Now we can say how to choose the added measures. They should be chosen to reduce the error of function $f$, or its bounds, of the approximation either by interpolation or the Schoenberg operator, as we will see in examples. So the question of the choice of the added measures becomes, more or less, the question how some spline space is good for approximating some given function.
\subsection{Calculation of the approximation} \label{calcaprox}
In this subsection we will offer two ways for calculating the approximation $s$. In the first way, we explicitly calculate the de\thinspace{}Boor points, {\it i.e.} the coefficients of $s$ expressed by the B-splines.
Let
\begin{equation} \label{sviadbpoints}
s(x)=\sum_{i=1}^n a_i T_i^k(x)
\end{equation}
for some $x\in[a,b]$, then by Theorem \ref{derivativethm}
{\setlength\arraycolsep{2pt}
\begin{eqnarray}
L_1s(x) & = & \sum_{i=2}^n d_i T_i^{k-1}(x) = \sum_{i=2}^n \frac{a_i-a_{i-1}}{C_i^{k-1}} T_i^{k-1}(x), \label{l1seq} \\
L_2s(x) & = & \sum_{i=3}^n f_i T_i^{k-2}(x) = \sum_{i=3}^n \frac{d_i-d_{i-1}}{C_i^{k-2}} T_i^{k-2}(x). \label{l2seq}
\end{eqnarray}}%
If we want that $s$ satisfies (\ref{coloceq}) or (\ref{quasicoloceq}), then either we calculate $(f_i)_{i=3}^n$ by solving the interpolation linear system, or $f_i:=f(\zeta_i)$ with $\zeta_i$ from Definition \ref{Schoenbergdef}. Either way, we know the coefficients $f_i$, $i=3,\dots,n$. Since $f_i=\frac{d_i-d_{i-1}}{C_i^{k-2}}$, it follows that
{\setlength\arraycolsep{2pt}
\begin{eqnarray*}
d_i & = & d_{i-1}+f_i C_i^{k-2} \\
& = & d_2+\sum_{j=3}^i f_jC_j^{k-2}
\end{eqnarray*}}%
for $i=3,\dots,n$, and in the same way
{\setlength\arraycolsep{2pt}
\begin{eqnarray}
a_i & = & a_{i-1}+d_iC_i^{k-1} \nonumber \\
& = & a_1+\sum_{\ell=2}^i d_\ell C_\ell^{k-1} \nonumber \\
& = & a_1+\sum_{\ell=2}^iC_\ell^{k-1}\left(d_2+\sum_{j=3}^\ell f_jC_j^{k-2}\right) \label{aiexp}
\end{eqnarray}
for $i=2,\dots,n$. That way, the only unknowns are $a_1$ and $d_2$, which can be found from the boundary values. From $s(a)=\alpha$ we get
$$
a_1=\frac{\alpha}{u_1(a)},
$$
and from $s(b)=\beta$, $a_n=\frac{\beta}{u_1(b)}$. The expression (\ref{aiexp}) gives
$$
a_n = \frac{\alpha}{u_1(a)} +d_2\underbrace{\sum_{\ell=2}^n C_\ell^{k-1}}_{=:A_1}+\underbrace{\sum_{\ell=2}^n C_\ell^{k-1}\sum_{j=3}^n f_jC_j^{k-2}}_{=:A_2} = \frac{\beta}{u_1(b)}
$$
which finally leads to
$$
d_2=\frac{\frac{\beta}{u_1(b)}-\frac{\alpha}{u_1(a)}-A_2}{A_1}.
$$
The alternative way to calculate the coefficients $\{a_i\}_{i=1}^n$ is to formulate the linear system from (\ref{l1seq}) and (\ref{l2seq}):
$$
\frac{\frac{a_i-a_{i-1}}{C_i^{k-1}}-\frac{a_{i-1}-a_{i-2}}{C_{i-1}^{k-1}}}{C_i^{k-2}}=f_i,
$$
for $i=3,\dots n$, together with the boundary conditions. Of course, the collocation coefficients $\{a_i\}_{i=1}^n$ can also be calculated in the usual way, by forming the collocation matrix.
We suggest also another approach, which calculates the value of the spline $s$ (\ref{sviadbpoints}) at some point directly, without explicitly knowing the de\thinspace{}Boor points $a_i$, and which is based on the Green's function and (\ref{sxGreen}). The idea is to find a numerical integration (if it is possible) which can calculate the integrals in (\ref{sxGreen}) (having in mind the expression (\ref{yxexpr})) exactly on each subinterval of the extended partition $\mbox{\boldmath $T$}$. This way produces a more stable algorithm, than the first one, but also since the calculation of the B-splines, or the evaluation of the spline, can be expensive, that way we only need the B-splines of two order less. The same idea can be applied also to the initial value problems, since as we mentioned before, the solution of it can be represented as in (\ref{solutionGreeen}).
\section{Examples} \label{secexample}
Bellow we illustrate both kind of approximations, and comment several issues mentioned before.
\begin{example}{1} For the first example we chose the singular boundary value problem
{\setlength\arraycolsep{2pt}
\begin{eqnarray}
D(\sqrt{x}Dy(x)) & = & f(x), \label{examplesingulareq}\\
y(0)=y(1) & = & 0, \nonumber
\end{eqnarray}}%
with two choices of the function $f$:
\begin{itemize}
\item[a)]
$f(x)=x^2+x^\frac{5}{2}$,
\item[b)] $\displaystyle f(x)=\frac{e^{100}(1-e^{-100x})}{e^{100}-1}$.
\end{itemize}
For more details and motivation for this problem, see \cite{bosnerroginasingular}.
So, $u_1\equiv 1$, and the first two measures are: $d\sigma_2(\tau_2)=\frac{d\tau_2}{\sqrt{\tau_2}}$ and $d\sigma_3(\tau_3)=d\tau_3$, and we append them with the same two measures, getting
$d\mbox{\boldmath $\sigma$}=(\frac{d\tau_2}{\sqrt{\tau_2}}, d\tau_3,$ $\frac{d\tau_4}{\sqrt{\tau_4}}, d\tau_5)^{\rm T}$. That way, our CCC--space is $\mathcal{S}(k,d\mbox{\boldmath $\sigma$})={\rm span}\,\{1,x^{\frac{1}{2}},x^{\frac{3}{2}},$ $x^2,x^3\}$. The CCC--space associated to the second reduced system is $\mathcal{S}(k-2,d\mbox{\boldmath $\sigma$}^{(2)})={\rm span}\,\{1,x^{\frac{1}{2}},x^{\frac{3}{2}}\}$, and all about how to calculate the associated singular splines can be found in \cite{bosnerroginasingular}.
We will calculate the collocation and quasi-collocation approximation via Green's function and Gaussian formula.
In our case, (\ref{sxGreen}) is equal to
{\setlength\arraycolsep{2pt}
\begin{eqnarray}
s(x) & = & \int_0^1 G(x,t)Q[f](t)\,dt \nonumber \\
& = & -2 \Big( (1-\sqrt{x})\int_0^x Q[f](t)\sqrt{t}\,dt+\sqrt{x}\int_x^1 Q[f](t)(1-\sqrt{t})\,dt \Big) \label{singintz} \\
& = & -4 \Big( (1-\sqrt{x})\int_0^{\sqrt{x}} Q[f](z^2)z^2\,dz+\sqrt{x}\int_{\sqrt{x}}^1 Q[f](z^2)(1-z)z\,dz \Big), \nonumber
\end{eqnarray}}%
by substitution $t=z^2$. Both integrals in the last equation are divided on the integrals over the subintervals, to be precise, if $x\in[t_i,t_{i+1})$, then the first integral is expressed as
$$
\int_0^{\sqrt{x}} = \sum_{j=k}^{i-1} \int_{\sqrt{t_j}}^{\sqrt{t_{j+1}}} + \int_{\sqrt{t_i}}^{\sqrt{x}},
$$
and analogously the second integral.
Since, after the substitution, the function being integrated by every such integral is polynomial of degree 5, we apply the appropriate Gauss-Legendre integration formula to calculate the integrals in (\ref{singintz}) exactly.
For both, the collocation and quasi-collocation, we use the CCC--Greville points, which are in $\mathcal{S}^{(2)}$ equal to
$$
\tau_i=\zeta_i=\left(\frac{2}{3}\frac{t_{i+1}+\sqrt{t_{i+1}}\sqrt{t_{i+2}}+t_{i+2}}{\sqrt{t_{i+1}}+\sqrt{t_{i+2}}}\right)^2.
$$
We will consider now each of the proposed functions $f$ separately.
\begin{itemize}
\item[a)] Table \ref{tablgressingular} brings us maximal errors of both approximations for several equidistant partitions, where $n$ is dimension of the underlying spline space. We can clearly see the difference between the approximation by collocation, which is of order 3, in the sense of Theorem \ref{bvquadconv}, and the one by quasi-collocation of order 2.
We also checked the estimation of the condition number of the interpolation matrix by the Lapack routine \texttt{dgbcon} for all partitions, and they are all around $\kappa_\infty(A)\approx 2.41$.
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
$n$ & quasi-collocation & collocation \\
\hline
24 & 0.48545E-03 & 0.69229E-06 \\
44 & 0.12130E-03 & 0.64411E-07 \\
84 & 0.30320E-04 & 0.59423E-08 \\
164 & 0.75792E-05 & 0.54331E-09 \\
324 & 0.18948E-05 & 0.49382E-10 \\
644 & 0.47368E-06 & 0.44657E-11 \\
1284 & 0.11842E-06 & 0.40257E-12 \\
2564 & 0.29605E-07 & 0.36190E-13 \\
5124 & 0.74012E-08 & 0.32283E-14 \\
10244 & 0.18503E-08 & 0.54123E-15 \\
20484 & 0.46258E-09 & 0.45797E-15 \\
40964 & 0.11565E-09 & 0.10825E-14 \\
81924 & 0.28911E-10 & 0.10825E-14 \\
163844 & 0.72278E-11 & 0.11796E-14 \\
\hline
\end{tabular}
\caption{The maximal errors of quasi-collocation and collocation with singular splines for the case a).} \label{tablgressingular}
\end{center}
\end{table}
\item[b)] In this example we will first observe just one equidistant partition, which gives the spline space of dimension $n=15$, and the errors made by both approximations in Figure \ref{slikasinggres}. We can notice that they are, unlikely to the case a), of the same order of magnitude.
\begin{figure}[h!]
\centering
\includegraphics[width=90mm]{slike_gr.pdf}
\caption{The errors of quasi-collocation and collocation with singular splines for the case b) and $n=15$.} \label{slikasinggres}
\end{figure}
The reason for it is that the second and third generalized derivatives of $f$ are unbounded around zero, so the interpolation of $f$ isn't that good, as we can see in Figure \ref{slikasingf}. The interpolation of $f$ oscillates, while with the approximation by CCC--Schoenberg operator, thanks to its shape preserving properties (see \cite{bosnercrnkovicskificimage,mazurechebschoenberg}), this is not the case.
Therefore, the collocation method is not that much more precise than quasi-collocation, as in case a), what we can see in Table \ref{tablgressingularb} (see Remark \ref{remarkcolqcol}).
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
$n$ & quasi-collocation & collocation \\
\hline
24 & 0.10289E-02 & 0.33555E-03 \\
44 & 0.39974E-03 & 0.30239E-04 \\
84 & 0.10698E-03 & 0.18377E-04 \\
164 & 0.22180E-04 & 0.40525E-05 \\
324 & 0.41036E-05 & 0.69207E-06 \\
644 & 0.81481E-06 & 0.11379E-06 \\
1284 & 0.27664E-06 & 0.19011E-07 \\
2564 & 0.81434E-07 & 0.32334E-08 \\
5124 & 0.22471E-07 & 0.56262E-09 \\
10244 & 0.59872E-08 & 0.97335E-10 \\
20484 & 0.15618E-08 & 0.17194E-10 \\
40964 & 0.40189E-09 & 0.30431E-11 \\
81924 & 0.10250E-09 & 0.53792E-12 \\
163844 & 0.25981E-10 & 0.95002E-13 \\
\hline
\end{tabular}
\caption{The maximal errors of quasi-collocation and collocation with singular splines for the case b).} \label{tablgressingularb}
\end{center}
\end{table}
For this function $f$ we could maybe make some better choice for the added measures,
but we wanted to present a case when the collocation method is not much better than quasi-collocation.
\begin{figure}[h!]
\centering
\includegraphics[width=90mm]{slike_f.pdf}
\caption{The function $f$ for the case b) and its approximations $S[f]$ and $I[f]$.} \label{slikasingf}
\end{figure}
\end{itemize}
\end{example}
\begin{example}{2}
Next, we solve the boundary value problem
{\setlength\arraycolsep{2pt}
\begin{eqnarray*}
D^2 y(x)-p^2 y(x) & = & \frac{1}{\cosh{(p x)}} D\left( \cosh{(px)}^2 D\frac{y}{\cosh{(p x)}} \right) = x^2, \\
y(0)=y(1) & = & 0, \nonumber
\end{eqnarray*}}%
with $p>0$, where we used (\ref{productop}) with $u(x)=\cosh{(p x)}$. This equation for large parameter $p$ is singularly perturbed, giving the solution with one boundary layer, therefore the polynomial splines are, generally, not the proper choice for solving this problem.
Here $u_1(x)=\cosh{(px)}$, and we add one more Lebesgue measure to get the spline space of order 4, with the measure vector $d\mbox{\boldmath $\sigma$}=(\frac{d\tau_2}{\cosh{(p\tau_2)}^2},\cosh{(p\tau_3)}\,d\tau_3,$ $d\,\tau_4)^{\rm T}$. The algorithms for calculating with such, so called, tension splines, can be found in \cite{bosnerroginatension,bosnerroginanonuniformtension}, although dealing with different CCC--systems, the resulting spline spaces for uniform parameters are the same.
In this example, the space $\mathcal{S}^{(2)}$ is of order $2$, so the interpolation and the Schoenberg operator give the same approximation ($Q=S=I$). Also, $\mathcal{S}^{(2)}$ is actually the polynomial spline space, but nevertheless, our main goal is to compare the two ways of calculation of the approximation from Subsection \ref{calcaprox}. So,
$s_1$ is the approximation of the exact solution $y$, where the de\thinspace{}Boor points are calculated exactly, and $s_2$ is obtained via Green's function and Gaussian formula.
Now, $s_2$ is
{\setlength\arraycolsep{2pt}
\begin{eqnarray*}
s_2(x) & = & u_1(x)\int_0^1 G(x,t)Q[f](t)\cosh{(pt)}\,dt \\
& = & -\frac{1}{p\sinh{p}} \left( \sinh{(p(1-x))}\int_0^x Q[f](t)\sinh{(pt)}\,dt \right. \\
& & \left. +\sinh{(px)}\int_x^1 Q[f](t)\sinh{(p(1-t))}\,dt \right),
\end{eqnarray*}}%
and the integrals over the subintervals can be calculated by using the integral formula that is exact for functions $\cosh{(qx)}$, $x\cosh{(qx)}$, $\sinh{(qx)}$ and $x\sinh{(qx)}$ for some $q>0$, over $[0,1]$, since $Q[f]$ is a linear polynomial spline.
Also, because $\mathcal{S}^{(2)}$ is the space of polynomial splines, we can numerically verify the order of convergence from Theorem \ref{bvquadconv} in the usual way, denoting them as $p_1$ and $p_2$ for $s_1$ and $s_2$, respectively.
For $p=10$, the errors and the numerical orders are given in Table \ref{tenserrtblep10}, while for $p=100$ in Table \ref{tenserrtblep100}. The errors for $p=10$ are similar for both calculation algorithms,
for $p=100$, the first algorithm can not keep pace with the second one,
and finally for $p=1000$ the first algorithm can not be used at all, so we present only the data for $s_2$ in Table \ref{tenserrtblep1000}.
When the tension splines of order $4$ are concerned, the instability of the first algorithm grows with the growth of the tension parameter $p$, due to the fact that the integrals $C_i^2$ are exponentially growing with $p$, and even more, for the fixed $p$, $C_i^2$ are growing exponentially with $i$, possibly causing overflow, while $C_i^3$ are going smaller with $p$ and $i$ even faster. That way we deal with very big or small values, which vary a lot in their magnitude, obviously causing the instability.
This behaviour is because the density of the measures $d\sigma_2=\frac{d\tau_2}{\cosh{(p\tau_2)}^2}$ and $d\sigma_3=\cosh{(p\tau_3)}\,d\tau_3$, is in one case fast exponentially decaying, while in another case exponentially growing.
This example is chosen to show the advantage of the second algorithm.
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
$n$ & $\|s_1-y\|_\infty$ & $p_1$ & $\|s_2-y\|_\infty$ & $p_2$ \\
\hline
23 & 0.412413376460648E-05 & & 0.412413376454793E-05 & \\
43 & 0.102848556711096E-05 & 2.003570 & 0.102848556711487E-05 & 2.003570 \\
83 & 0.256960984222447E-06 & 2.000900 & 0.256960984225049E-06 & 2.000900 \\
163 & 0.642302149141127E-07 & 2.000225 & 0.642302149119443E-07 & 2.000225 \\
323 & 0.160569241188825E-07 & 2.000057 & 0.160569241284235E-07 & 2.000057 \\
643 & 0.401419225810884E-08 & 2.000014 & 0.401419227935920E-08 & 2.000014 \\
1283 & 0.100354564719005E-08 & 2.000003 & 0.100354565412894E-08 & 2.000003 \\
2563 & 0.250886219785307E-09 & 2.000001 & 0.250886259683947E-09 & 2.000001 \\
5123 & 0.627215159150485E-10 & 2.000001 & 0.627215601504971E-10 & 2.000000 \\
10243 & 0.156804530297705E-10 & 1.999993 & 0.156804100953645E-10 & 1.999998 \\
20483 & 0.392006956409507E-11 & 2.000016 & 0.392012854469326E-11 & 1.999990 \\
40963 & 0.980084503138245E-12 & 1.999901 & 0.980073661116521E-12 & 1.999939 \\
81923 & 0.244914765551441E-12 & 2.000626 & 0.245122932368558E-12 & 1.999385 \\
163843 & 0.612101515307106E-13 & 2.000437 & 0.613597714305136E-13 & 1.998140 \\
\hline
\end{tabular}
\end{center}
\caption{The errors and numerical orders of $s_1$ and $s_2$ for tension splines of order $4$ and $p=10$.} \label{tenserrtblep10}
\end{table}
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
$n$ & $\|s_1-y\|_\infty$ & $p_1$ & $\|s_2-y\|_\infty$ & $p_2$ \\
\hline
23 & 0.903501233085530E-04 & & 0.507641834927993E-07 & \\
43 & 0.903501233085530E-04 & 0.000000 & 0.112310617536170E-07 & 2.176317 \\
83 & 0.444711676796353E+20 & -78.703618 & 0.266114078252636E-08 & 2.077378 \\
163 & 0.903501233085530E-04 & 78.703618 & 0.654711583612886E-09 & 2.023113 \\
323 & 0.903501233085530E-04 & 0.000000 & 0.162991554037177E-09 & 2.006062 \\
643 & 0.903501233085530E-04 & 0.000000 & 0.407045782035943E-10 & 2.001534 \\
1283 & 0.903501233085530E-04 & 0.000000 & 0.101734313112450E-10 & 2.000385 \\
2563 & 0.903501233085530E-04 & 0.000000 & 0.254318830263719E-11 & 2.000096 \\
5123 & 0.903501233085530E-04 & 0.000000 & 0.635786697811627E-12 & 2.000024 \\
10243 & 0.903501233085530E-04 & 0.000000 & 0.158946640571589E-12 & 2.000000 \\
20483 & 0.903501233085530E-04 & 0.000000 & 0.397372971116736E-13 & 1.999977 \\
40963 & 0.444711676796353E+20 & -78.703618 & 0.993564226118432E-14 & 1.999809 \\
81923 & 0.903501233085530E-04 & 78.703618 & 0.248550637536871E-14 & 1.999073 \\
163843 & 0.903501233085530E-04 & 0.000000 & 0.623240066326136E-15 & 1.995680 \\
\hline
\end{tabular}
\end{center}
\caption{The errors and numerical orders of $s_1$ and $s_2$ for tension splines of order $4$ and $p=100$.} \label{tenserrtblep100}
\end{table}
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
$n$ & $\|s_2-y\|_\infty$ & $p_2$ \\
\hline
23 & 0.623000000005656E-09 & \\
43 & 0.154240187269918E-09 & 2.014053 \\
83 & 0.371083218788035E-10 & 2.055364 \\
163 & 0.831544592344676E-11 & 2.157877 \\
323 & 0.181163400130717E-11 & 2.198502 \\
643 & 0.420507116972695E-12 & 2.107089 \\
1283 & 0.102616142194353E-12 & 2.034873 \\
2563 & 0.254876882039150E-13 & 2.009385 \\
5123 & 0.636138175827156E-14 & 2.002388 \\
10243 & 0.158969110661614E-14 & 2.000594 \\
20483 & 0.397412082863075E-15 & 2.000039 \\
40963 & 0.993848109210704E-16 & 1.999538 \\
81923 & 0.248615816722162E-16 & 1.999107 \\
163843 & 0.622611567879273E-17 & 1.997514 \\
\hline
\end{tabular}
\end{center}
\caption{The errors and numerical orders of $s_2$ for tension splines of order $4$ and $p=1000$.} \label{tenserrtblep1000}
\end{table}
\end{example}
\begin{example}{3}
Our last example is a very simple one:
{\setlength\arraycolsep{2pt}
\begin{eqnarray*}
D^2 y(x) & = & \frac{\sinh{(px)}}{\sinh{p}}px, \\
y(0) & = & y(1) = 0, \nonumber
\end{eqnarray*}}%
with $p>0$, which is also a singularly perturbed equation,
and for a large parameter $p$ the solution, as well as the right hand size $f$, has one boundary layer. In this case, $u_1\equiv 1$, and both measures, $d\sigma_2$ and $d\sigma_3$ are Lebesgue measures. The next question is how to chose the remaining measures. Because of the boundary layer of the solution, the simplest choice of just adding some more Lebesgue measures is not the best one, since, as we mentioned before,
polynomial splines do not catch that layer well (especially if the number of knots isn't too big). So, we take the following measure vector
$$
d\mbox{\boldmath $\sigma$}=\left( d\tau_2,d\tau_3,\cosh{(p\tau_4)}\,d\tau_4,\frac{d\tau_5}{\cosh^2(p\tau_5)}\right)^{\rm T},
$$
which gives us the tension splines of order $5$ (piecewisely spanned by $\{1,x,x^2, \cosh{(px)},$ $\sinh{(px)}\}$), which is a reasonable choice, concerning the right hand side of the equation, and the boundary layer. We use
the second approximation by integrals $s_2$ for quasi-collocation, because for that algorithm we only need to calculate the B-splines of order $3$ from the space $\mathcal{S}^{(2)}$.
In this case the approximations by quasi-collocation and collocation are almost the same, because $\mathcal{S}^{(2)}$ is almost the space of polynomial splines of order 1, so we will skip the collocation.
Then, we compare $s_2$ with the polynomial splines of order 5 $s_{sp}$ and $s_{ip}$, obtained by the first algorithm. The approximation $s_{sp}$ is achieved by quasi-collocation, and $s_{ip}$ by collocation.
Here
{\setlength\arraycolsep{2pt}
\begin{eqnarray}
s_2(x) & = & \int_0^1 G(x,t)S[f](t)\,dt \nonumber \\
& = & - (1-x)\int_0^x S[f](t)\,dt -x \int_x^1 S[f](t)(1-t)\,dt, \label{s2tens5int}
\end{eqnarray}}%
and $S[f]$ is piecewisely spanned by $\{1,\cosh{(px)},\sinh{(px)}\}$. It is possible to find stable integration formulas to calculate integrals in (\ref{s2tens5int}) exactly. For the points $(\zeta_i)_{i=1}^n$ which define the CCC--Schoenberg operator in $\mathcal{S}^{(2)}$, we chose the Greville abscis\ae{}, where $u_{2,2}(x)=\frac{\sinh{px}}{p}$, so
$$
\zeta_i=\frac{1}{p}\arcsinh{\left(\frac{\sinh{\frac{p(t_{i+1}+t_{i+2})}{2}}}{\cosh{\frac{p(t_{i+2}-t_{i+1})}{2}}}\right)}.
$$
We compare these three approximation for $p=10000$ in Table \ref{tens5errtblep10000}.
The stagnation in errors of $s_2$ for smaller $n$ is most probably due a specific shape of the solution, which looks almost as a linear function, with a very steep boundary layer. The maximum of the exact solution is around $10^{-5}$, so we see from that table, that the polynomial splines only for large $n$ come a little bit closer to the solution, and even interpolation of $f$, which is of order 3, doesn't help a lot. Also, in the same table we see the connection between large error of $f$ and large error of $y$.
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
$n$ & $\|s_2-y\|_\infty$ & $\|L_2s_2-f\|_\infty$ & $\|s_{sp}-y\|_\infty$ & $\|L_2s_{sp}-f\|_\infty$ \\
\hline
24 & 0.200E-07 & 0.368E+00 & 0.200E+01 & 0.974E+04 \\
44 & 0.200E-07 & 0.368E+00 & 0.510E+00 & 0.954E+04 \\
84 & 0.200E-07 & 0.368E+00 & 0.129E+00 & 0.919E+04 \\
164 & 0.200E-07 & 0.368E+00 & 0.323E-01 & 0.861E+04 \\
324 & 0.200E-07 & 0.368E+00 & 0.801E-02 & 0.768E+04 \\
644 & 0.200E-07 & 0.368E+00 & 0.194E-02 & 0.628E+04 \\
1284 & 0.182E-07 & 0.364E+00 & 0.469E-03 & 0.437E+04 \\
2564 & 0.890E-08 & 0.287E+00 & 0.142E-03 & 0.246E+04 \\
5124 & 0.324E-08 & 0.133E+00 & 0.431E-04 & 0.117E+04 \\
10244 & 0.105E-08 & 0.478E-01 & 0.116E-04 & 0.501E+03 \\
20484 & 0.288E-09 & 0.175E-01 & 0.296E-05 & 0.186E+03 \\
40964 & 0.738E-10 & 0.564E-02 & 0.743E-06 & 0.578E+02 \\
81924 & 0.186E-10 & 0.151E-02 & 0.186E-06 & 0.152E+02 \\
163844 & 0.465E-11 & 0.384E-03 & 0.465E-07 & 0.383E+01 \\
\hline
\end{tabular} \\[1em]
\begin{tabular}{|c|c|c|}
\hline
$n$ & $\|s_{ip}-y\|_\infty$ & $\|L_2s_{ip}-f\|_\infty$ \\
\hline
24 & 0.120E+01 & 0.965E+04 \\
44 & 0.306E+00 & 0.938E+04 \\
84 & 0.776E-01 & 0.893E+04 \\
164 & 0.196E-01 & 0.819E+04 \\
324 & 0.499E-02 & 0.702E+04 \\
644 & 0.132E-02 & 0.531E+04 \\
1284 & 0.340E-03 & 0.312E+04 \\
2564 & 0.547E-04 & 0.121E+04 \\
5124 & 0.452E-05 & 0.333E+03 \\
10244 & 0.318E-06 & 0.610E+02 \\
20484 & 0.205E-07 & 0.873E+01 \\
40964 & 0.129E-08 & 0.113E+01 \\
81924 & 0.808E-10 & 0.864E-01 \\
163844 & 0.505E-11 & 0.135E-01 \\
\hline
\end{tabular}
\end{center}
\caption{The errors of $s_s$, $s_{sp}$ and $s_{ip}$, together with the errors of the approximations of $f$ for $p=10000$.} \label{tens5errtblep10000}
\end{table}
To illustrate such big differences, we concentrate more closely on the case when $n=24$. First, we observe the approximations of the function $f$, {\it i.e.} $S[f]$ from the spaces of tension and polynomial splines, and $I[f]$ from polynomial spline space. As we can see on Figure \ref{slikatensf}, the tension spline approximation is almost exactly above the function $f$, which is almost everywhere equal to zero, except that very steep boundary layer on the right, while the polynomial ones are quite far away, and the interpolation is oscillating a lot again.
\begin{figure}[h!] \label{slikatensf}
\centering
\includegraphics[width=90mm]{slike_f_tension.pdf}
\caption{The function $f$ and its approximations.}
\end{figure}
For the approximation of the solution, the situation is even worse, the polynomial splines are completely wrong and way out of scale (the approximation by polynomial quasi-collocation is at least closer in shape), as we can see on Figures \ref{slikatensu1} and \ref{slikatensu2}.
\begin{figure}[h!]
\centering
\includegraphics[width=90mm]{slike_u_tension1.pdf}
\caption{The solution and all three approximations.} \label{slikatensu1}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=90mm]{slike_u_tension2.pdf}
\caption{The solution and the tension spline approximation.} \label{slikatensu2}
\end{figure}
Therefore, we can conclude, that the simplest choice for added measures, {\it i.e.} the Lebesgue measures, doesn't have to be always good, and as we mentioned before,
the better approximation of $f$ suggests that the chosen measures will lead to the smaller error of the final solution.
\end{example}
\section{Conclusion}
In this paper we demonstrated one more possible use of the CCC--Schoenberg operators, in solving boundary value problems, with emphasize on the singular and singularly perturbed ODE, and compare it with the collocation method. We proposed the algorithm for calculating both of them, based on the integral of the Green's function, which uses only the B-splines of two orders less. They are, of course, simpler to calculate than the B-splines from the original spline space.
For the mostly used spline spaces, it should not be a problem to find integration formulas, which allow us to calculate the required integrals exactly. When we can numerically stably calculate the B-splines from the original space, the first algorithm based on calculation of de\thinspace{}Boor points, can also be used without a problem, in cases where the integrals $C_i^{k-1}$ and $C_i^{k-2}$ in (\ref{l1seq}) and (\ref{l2seq}) are ``decent''.
We also show in our examples the cases when the quasi-collocation can almost be as good as the collocation method.
Therefore, in our opinion, this quasi-collocation method could be a useful tool for solving these specific problems.
\bibliographystyle{spmpsci}
|
\section{Introduction}\label{intro}
We analyse the controllability of the interacting particle system of $N$ agents on the plane, governed by second-order dynamics
\begin{equation}\label{syst}
\begin{split}
\dot x_i(t) &= v_i(t) \\
\dot v_i(t) &= (\alpha-\beta \vert v_i(t)\vert^2 ) v_i(t) - \frac{1}{N} \sum_{\substack{j=1\\ j\neq i}}^N \nabla W(x_i(t)-x_j(t)) + u_i(t)\,,
\end{split}
\end{equation}
where $x_i(t)\in\mathbb{R}^2$ (resp., $v_i(t)\in\mathbb{R}^2$) is the position (resp., the velocity) of the $i^\textrm{th}$ agent.
In this model, the term $ (\alpha-\beta \vert v_i\vert^2 ) v_i$, where $\alpha\geq0$ and $\beta>0$ are fixed, represents a \emph{self-propulsion} force, while
\begin{equation}\label{syst2b}
F_i(x) = \frac{1}{N} \sum_{\substack{j=1\\ j\neq i}}^N \nabla W(x_i-x_j) , \qquad F=(F_1,\ldots,F_N)^\top\,,
\end{equation}
expresses an \emph{attraction-repulsion} force through the pairwise interaction potential $W$. The control $u=(u_1,\ldots,u_N)$, with $u_i(t)\in\mathbb{R}^2$, is subject to the constraint $\Vert u(t)\Vert\leq M$ for almost every $t\in\mathbb{R}$, where $M>0$ is fixed. Here, $\vert\cdot\vert$ is the Euclidean norm in $\mathbb{R}^2$ and $\Vert\cdot\Vert$ is the $\infty$-norm in $(\mathbb{R}^2)^{N}$ or $(\mathbb{R}^2)^{2N}$ associated to $|\cdot|$, i.e.,
$$
\Vert v\Vert=\max_{i=1,\ldots N} |v_i|,\qquad
\Vert (x,v)\Vert=\max_{i=1,\ldots N} |x_i|+\max_{i=1,\ldots N} |v_i|.
$$
{Solutions of the control system \eqref{syst} need to be interpreted in the Caratheodory sense, see, e.g., \cite{BressanPiccoli,Sontag}. Existence and uniqueness of the solution is classical, provided that $u(t)\in L^\infty([0,T];\mathbb{R}^N)$: this condition will always be satisfied, as we deal with bounded controls.}
Models of the form \eqref{syst} are particular examples of agent-based models (ABMs). ABMs appear in biology, mathematics, physics, and engineering in order to describe the motion of a collection of $N$ individual entities at the microscopic scale interacting through simple rules. These types of models have been proposed to describe the flocking of birds \cite{Camazine,lukeman,science1999}, the schooling of fish \cite{Aoki,alethea,Bjorn,Hemelrijk:Hildenbrandt,Huth:Wissel}, and swarms of bacteria \cite{KochWhite}, among others. We refer to the surveys \cite{CFTV,survey} for more general models in the area of interacting particle systems in collective behavior.
Model \eqref{syst} was introduced in \cite{Levine:2000} and extensively studied in \cite{bertozzi,Dorsogna2,Dorsogna} giving a detailed description of patterns and stability properties of particular solutions through numerical experiments. The role of the self-propulsion term of strength $\alpha>0$ versus friction of strength $\beta >0$ is to fix a typical cruise speed for agents. In fact, in the absence of interactions $W=0$ (no potential) and $u=0$ (no control) in \eqref{syst}, except for the unstable equilibrium $v=0$, all trajectories converge to $\vert v_i\vert^2=\frac{\alpha}{\beta}$ along heteroclinic orbits. These terms will actually promote the appearance of particular solutions : \emph{flock} and \emph{mill} solutions (defined below). Even more complicated solutions as double-mills have been studied in the literature \cite{Dorsogna3}.
We assume throughout the article that the interaction potential is radial $W(x)=U(\vert x\vert)$, with $U$ of class $C^2$ except possibly at the origin, and that the interactions are negligible for large distances $\lim_{r\to+\infty} U'(r)=0$. Typical potentials used in previous works are Morse potentials of the form $U(r) =- C_A e^{-r/\ell_A} + C_R e^{-r/\ell_R}$, the index A standing for ``attractive" and the index R for ``repulsive". As shown in \cite{bertozzi, Dorsogna}, the interesting region is when $\ell=\frac{\ell_R}{\ell_A}<1$ and $C=\frac{C_R}{C_A}>1$. In this case, the derivative $U'$ of the potential is such that $\vert U'\vert$ is bounded, $U'$ is positive up to some $r_0>0,$ and is then negative and converging to $0$ as $r\rightarrow +\infty$. We will refer to this kind of potentials as bounded repulsive-attractive potentials. Other potentials of interest are power-law potentials \cite{KSUB,BCLR} given by $U(r)=\frac{\vert r\vert^a}{a}-\frac{\vert r\vert^b}{b}$ with {$a>b>0$}, for which we have $U'(0)=-\infty$, thus avoiding collisions due to an increasing repulsion whenever two particles get closer. We will refer to this kind of potentials as unbounded repulsive-attractive potentials.
Another family of ABMs of interest arises when introducing alignment mechanisms in the modeling. A basic example of this family is the Cucker-Smale model introduced in \cite{CuckerSmale1,CuckerSmale2} and further developed in \cite{CFRT,HaLiu,HaTa,mt}, among others. The main phenomenon in those models is the emergence of alignment, i.e., consensus in velocity. Imposing consensus in velocity has also been analysed from the point of view of control \cite{BFK15,CFPT2013,CFPT2015,HK18}. These consensus models also have applications in swarm robotics \cite{CKPP19,CKP21}, social and pedestrian dynamics \cite{ACFK17,ABCK,KZ20,Suttida2} where control theory is applied with different regulation objectives expressed in both ad-hoc and optimal control designs \cite{albi2021momentdriven,ALBI201886,BAILO20181,Suttida1}.
Despite the simplicity of the model \eqref{syst}, a striking phenomenon regarding the long-time asymptotics of solutions occurs. There are several stable self-organized patterns that emerge from these dynamics depending on the initial data even for the same parameter values and interaction potentials \cite{Dorsogna2,Dorsogna}. More precisely, flock and mill solutions, which are relevant examples of self-organized configurations for the swarming model, appear asymptotically. Flock and mill solutions are not equilibria in the classical sense (with $\dot x_i=\dot v_i=0$ for all $i=1,\ldots,N$), but are rather solutions with specific invariance properties inherited from \eqref{syst}. Flock solutions describe configurations with agents moving by uniform translation: a \emph{flock} is a trajectory $(x(t),v(t))=( x^*+t v^*, v^*)$ in which {$x^*=(x^*_1,\ldots, x^*_N)$ is a vector of $N$ positions and $v^*=( v^*_1,\ldots, v^*_N)$ is a vector of $N$ identical velocities $v^*_1=\ldots=v^*_N$ that moreover satisfy $|v^*_i|=\sqrt{\frac{\alpha}{\beta}}$}. A \emph{flock ring} is a flock in which the position of agents $x_i$ are equally distributed on a circle with a certain radius $R$, i.e., {$x^*=(x^*_1,\ldots, x^*_N)$ is of the form $x^*_i=c+\hat x_i$, where $c\in \mathbb{R}^2$ is the center of the circle and $\hat x_1,\ldots,\hat x_N\in \mathbb{R}^2$ for $i=1,\ldots, N$ are equispaced points on a circle of radius $R$ centered at zero, i.e.
\begin{equation} \label{e-hatx}
\hat x_i= R \, \mathcal{R}_\theta \begin{pmatrix}\cos\left(\frac{2\pi i}{N}\right) \\[2mm] \sin\left(\frac{2\pi}{N i}\right)\end{pmatrix},\quad \mathcal{R}_\theta=\begin{pmatrix} \cos\theta& -\sin\theta\\ \sin\theta & \phantom{-}\cos\theta \end{pmatrix}.
\end{equation}
for some angle $\theta$.}
A \emph{mill} for \eqref{syst} corresponds to $N$ agents rotating with a constant angular velocity $\omega$ with respect to a center $c$, i.e.,
$(x_i(t),v_i(t))=(\mathcal{R}_{\omega t}( x^*_i-c)+c,\mathcal{R}_{\omega t} v^*)$
for some $(x^*,v^*)\in\mathbb{R}^{4N}$. A \emph{mill ring} is a mill in which the position of agents $x_i$ are moreover equally distributed on a circle with a certain radius $R$, i.e., { $x^*_i=c+\hat x_i$ with $\hat x_i$ given by \eqref{e-hatx}}. As a consequence, velocities satisfy
$$
v_i^*=\frac{1}{R}\sqrt{\frac{\alpha}{\beta}} \, \hat x_i^\perp
\qquad\textrm{with}\qquad
\hat x_i^\perp=\mathcal{R}_{\pi/2} \hat x_i .
$$
{The linear stability properties of flock and mill rings have been studied in detail in \cite{bertozzi,KSUB}. The nonlinear stability analysis of general flocks and mills has been fully analysed in \cite{ABCV,CHM}. As a consequence, in the vicinity of certain flocks/mills, it can be shown that the system self-organizes towards a flock/mill. The precise notion of vicinity is referred to the mentioned literature.} Notice that the space positions of flock solutions to \eqref{syst} correspond to stationary states of the first order model
\begin{equation}\label{eq:firstorder}
\dot{x}_i=-\sum_{\substack{j=1\\j\neq i}}^N \nabla
W\left(x_j-x_i\right), \quad i=1,\dots, N.
\end{equation}
Flock shapes, stationary states for \eqref{eq:firstorder}, for different potentials can have many different shapes even for biologically motivated potentials \cite{KSUB,bertozzi,BUKB,CHM2}, and their regularity heavily depends on the repulsion strength at the origin, see \cite{BCLR,BCLR2,CDM}. Characterizing all possible mill and flock profiles for a given potential is equivalent to characterizing all possible stationary states of the first order system \eqref{eq:firstorder} or related equations. This difficult problem has not been solved except in very particular choices of the parameters for the power-law potentials. It can be shown that for a repulsive-attractive potential there is a unique flock and mill ring solution and that the radius is characterized uniquely by balances of the relevant forces: attraction, repulsion and centrifugal forces \cite{BCLR,ABCV}. However, showing that they are the unique flock or mill solution is a challenging problem. It is possible to find potentials for which stable mills exist and they are not rings by numerical experiments. Moreover, some compactly supported potentials generically allow the existence of \emph{flock and mill clusters}, i.e., clusters of particles in which each cluster is a mill, or different flocks in separate directions. For the specific case of flock and mill rings, the radius $R$ is characterized by being a solution to
\begin{equation}\label{e-radius}
\sum_{p=1}^{N-1}\sin\Pt{\frac{p\pi}N}\tilde U'\Pt{2R\sin\Pt{\frac{p\pi}N}}=0\,,
\end{equation}
where $\tilde U(r)=U(r)-\omega^2\frac{r^2}2$, see \cite{bertozzi,ABCV,CHM}. Flock solutions correspond to $\omega=0$.
Our main goal in this work is to show that constructive controls can be designed to steer the system from any initial data to these various self-organized configurations in interacting particle systems of collective behavior. The main strategy is to prove that the system \eqref{syst} enjoys interesting controllability properties, such as being able to: steer the system from any initial condition to some/any flock and/or mill; keep the system close to a flock or a mill with an appropriate feedback law; pass from a flock to a mill or conversely. The control $u$ is assumed to satisfy the constraint $\|u\|\leq M$ with the minimal threshold $M>0$ being clearly identified.
The rest of the paper is structured as follows. Section \ref{mainsec} explains in details the main results and the main novelty of our approach: mixing different control techniques with a deep understanding of the heteroclinic connections in these models. In Section \ref{sec_num}, we recall some known stability properties of flocks and mills, together with proposing different feedback designs for transitioning to flock and mill configurations. The control design is guided by controllability results, local stability properties and heteroclinic connections, and is enriched by the use of numerical optimal controls. {Even if the main contribution of the paper consists of the results in Section \ref{mainsec}, which characterize the controlled transition between different flocks and mills, as the proofs indicate, these transitions are constructed in steps as a concatenation of different feedback controls. Section \ref{sec_num} goes beyond the existence of these controls and provides concrete constructions for such feedback laws, making use of both stability properties of the system and optimal control elements.} In Sections \ref{sec_proof_mainthm} and \ref{sec_proof_thm2}, we provide a complete and detailed proof of Theorems \ref{mainthm} and \ref{mainthm_variant}.
\section{Main results}\label{mainsec}
Our two main results, Theorem \ref{mainthm} in Section \ref{sec_bounded_interactions}, and Theorem \ref{mainthm_variant} in Section \ref{sec_unbounded_interactions}, deal with bounded and unbounded interaction potentials (as defined in the introduction), respectively.
For each statement, we present a brief sketch of our control strategies with a full proof in the last sections.
\subsection{First main result: bounded interactions}\label{sec_bounded_interactions}
In our first main result for bounded repulsive-attractive potentials, we assume that our potential is radial, of class $C^2$, with bounded interactions, i.e., $\sup_{r\geq 0} U'(r)<+\infty$ and negligible interactions at $\infty$, i.e., $\lim_{r\to+\infty} U'(r)=0$. Since the potential is $C^2$ at the origin and radial then $U'(0)=0$. This ensures classical well-posedness of \eqref{syst} (see Remark \ref{rem_potential} further).
\begin{theorem}\label{mainthm}[Control for bounded interactions]
Let $U:[0,+\infty)\rightarrow\mathbb{R}$ such that $W(x)=U(|x|)$ is of class $C^2$ satisfying $U'(0)=0$, $\lim_{r\to+\infty} U'(r)=0$ and:
\begin{enumerate}[label=$\bf (U_{\arabic*})$]
\item\label{U1} $C^2$-boundedness: there exist a positive constant $C$ such that $|U(r)|+|U'(r)|+|U''(r)|<C$ for every $r\in[0,+\infty)$.
\end{enumerate}
If the upper bound $M$ for the control action is such that
\begin{equation}\label{hyp_M}
M> M_{\alpha,\beta}=\sqrt\frac{4\alpha^{3}}{27\beta}
\end{equation}
then, given any $\bar v\in\mathbb{R}^2$ such that $\vert\bar v\vert = \sqrt{\frac{\alpha}{\beta}}$, {there exists $\bar x=(\bar x_1,\ldots,\bar x_N)\in \mathbb{R}^{2N}$ such that} the control system can be steered, in sufficiently large time and with a feedback control, from any initial condition to any neighborhood of the flock $(\bar x_1+t\bar v,\bar v),(\bar x_2+t\bar v,\bar v),\ldots,(\bar x_N+t\bar v,\bar v)$ with a control satisfying $\|u\|\leq M$.
\noindent Denoting by $M_F=\sup_{r>0} \vert U'(r)\vert$. Under the stronger assumption
\begin{equation}\label{hyp_M_stronger}
M > \max \left( M_{\alpha,\beta} , \ M_F \right) ,
\end{equation}
the control system can be steered, in sufficiently large time and with a feedback control, from any initial condition to any neighborhood of any flock, flock ring, mill, mill ring.
\end{theorem}
\begin{remark}\label{rem_potential}
The result can be generalized to potentials satisfying $U'(0)<0$, i.e., repulsive at $0$ but nonsmooth at $0$: for instance, in the case of the widely used Morse potentials \cite{Dorsogna} in the biologically reasonable, one has bounded interactions up to $0$ even if the potential is not $C^2$ at the origin. In Theorem \ref{mainthm} above, we have assumed that $U'(0)=0$ to ensure { that $\nabla W(x_i-x_j)$ is Lipschitz with respect to $x$, hence existence and uniqueness of solutions to \eqref{syst} in the Caratheodory sense is guaranteed (see \cite{BressanPiccoli,Sontag})}. In the more general case where $U'(0)<0$, collisions may occur in finite time and thus well-posedness is not a priori ensured. {Anyway, it is always possible to slightly modify our feedback controls in order to avoid the origin, and thus avoid problematic points for local well-posedness, so essentially we can always assume without loss of generality that the radial potential is of class $C^2$ at the origin if it has bounded interactions, i.e., if $U'(0)$ is finite or $M_F<\infty$.}
\end{remark}
\begin{remark}[The role of the assumptions]
Condition \eqref{hyp_M} means that the control can counteract the natural tendency of the system to stabilize $|v|$ to $\sqrt\frac\alpha\beta$.
The value $M_{\alpha,\beta}$ corresponds to the maximum of the function $s\to \alpha s-\beta s^3$ on the half line $s\geq 0$, which is attained at $v=\sqrt{\frac{\alpha}{3\beta}}$. Under the condition \eqref{hyp_M_stronger}, one can moreover counteract the potential interactions, which allows us to design controls steering the system to any mill. Actually, when the potential allows the existence of cluster mills {(i.e., separated groups of agents, each of them forming a mill)}, it is even possible to steer the system to any such cluster mill configuration.
\end{remark}
The control strategy to achieve these various objectives can be made explicit and we will provide it in a \emph{feedback} form, making it particularly convenient to implement in practice. It can even be provided in a \emph{componentwise sparse feedback} form, provided that $M$ is large enough. Hereafter, we explain the controllability strategy, by sketching the proof of Theorem \ref{mainthm} (full detail of the proof is given in Section \ref{sec_proof_mainthm}).
Given any $\varepsilon>0$, we set
$$
\Omega_\varepsilon = \{ (x,v)\in\mathbb{R}^{4N}\ \mid\ v=0,\ \|F(x)\|\leq \varepsilon \} .
$$
When $u=0$, the set $\Omega_0$ (for $\varepsilon=0$) is invariant under the dynamics. The $x$-components of each point of $\Omega_0$ are indeed the flocks of the uncontrolled system. Note that, for $\varepsilon>0$ small, $\Omega_\varepsilon$ consists both of topological neighborhoods of $\Omega_0$ in the $x$-variable and of some components ``at infinity'', where $F$ is small because $\lim_{r\to+\infty} U'(r)=0$.
\medskip
\paragraph{\bf Step 1: reaching $\Omega_\varepsilon$} The first step of our strategy consists of steering the system from its given (arbitrary) initial condition to the set $\Omega_\varepsilon$, {where $\varepsilon>0$ is a fixed parameter. It needs to be chosen sufficiently small to ensure ``good approximations'' of the nonlinear dynamics \eqref{syst} by the linearized one}. With this goal, we use the Jurdjevic-Quinn method \cite{JQ}, which is a very powerful approach in control design to derive feedback controls which moreover enjoy instantaneous optimality properties (see \cite{CFPT2013,CFPT2015} as well as \cite{CPRT2,CPRT_A2017} for its application to Cucker-Smale multi-agent models). This is done by differentiating with respect to time an appropriate Lyapunov-type functional, then choosing adequately the feedback control, and finally using arguments close to the LaSalle invariance principle.
\medskip
\subparagraph{\bf Step 1.1: Jurdjevic-Quinn stabilization}
Defining the Lyapunov functional
$$
V = \frac{1}{2} \sum_{i=1}^N \vert v_i\vert^2 + \frac{1}{2N} \sum_{\substack{i,j=1 \\ i\neq j}}^N W(x_i-x_j) ,
$$
along the trajectories we have
$$
\dot V = \sum_{i=1}^N (\alpha-\beta \vert v_i\vert^2 ) \vert v_i \vert^2 + \langle v_i, u_i\rangle\,,
$$
which leads us to define appropriate feedback controls $u_i$ making $V$ decrease: roughly speaking, we would like to take $u_i=-M\frac{v_i}{\vert v_i\vert}$ when $\vert v_i\vert\leq\sqrt\frac{\alpha}{\beta}$ and $0$ otherwise. Then, using arguments similar to those used in the LaSalle invariance principle, we obtain convergence to $\Omega_\varepsilon$.
Although this gives the main idea, the complete argument in Section \ref{sec_proof_step1} is not so easy. The main difficulty is that $V$ is not proper (i.e., $V$ is not infinite at infinity), even taking the quotient with respect to translations. This does not ensure that the system asymptotically reaches a neighborhood $\Omega_0$ as in the classical LaSalle principle, as it can also converge to a component at infinity.
\medskip
\subparagraph{\bf Step 1.2: Reaching $\Omega_\varepsilon$ in finite time}
Once $(x,v)$ is sufficiently close to $\Omega_\varepsilon$ at time $T$, both velocities and forces are small. We then apply the control
$$
u_i=-(\alpha -\beta |v_i|^2)v_i+F_i(x)-\eta\frac{v_i}{|v_i|}
$$
with $\eta>0$ small to steer each component of $v(T)$ such that $v_i\left(T+\tfrac{|v_i(T)|}{\eta}\right)=0$. Note that $\vert u_i\vert$ remains small because we are near $\Omega_\varepsilon$. By applying this control to each agent, we reach $\Omega_\varepsilon$ in finite time.
\medskip
\paragraph{\bf Step 2: Local controllability near $\Omega_\varepsilon$}
At the end of Step 1, the system is in $\Omega_\varepsilon$. Since $v=0$ and $F(x)\simeq 0$ there, we can set
\begin{equation}\label{controlnear}
u_i=-(\alpha-\beta \vert v_i\vert^2 ) v_i + F_i(x)+w_i
\end{equation}
and consider $w$ as a new control, subject to the constraint $\Vert w\Vert\leq M/2$, and thus focus on the very simple control system
\begin{equation}\label{very_simple_control_system}
\dot x_i = v_i , \qquad \dot v_i = w_i ,
\end{equation}
near $v=0$. It is obvious to generate feedback controls $w$, satisfying $\Vert w\Vert \leq M/2$, steering this simplified control system from any point in $\Omega_\varepsilon$ to any other point in a local neighborhood. Note that $\|u\|\leq M$, since the agents remain near $\Omega_\varepsilon$.
Notice that we cannot assure at this stage that two disconnnected components of $\Omega_\varepsilon$ can be joined by controls of this form, since the size of the control $M$ is disconnected from the size of the force, due to the interaction potential $U$. In order to choose particular spatial configurations such as flocks or mill rings we control the interaction potential in the next steps by increasing the value for the constraint $\|u\|\leq M$ to $M>M_F$.
\medskip
\paragraph{\bf Step 3: reaching flocks and mills}
\subparagraph{\bf Step 3.1: Reaching flocks} Here, we are still under the assumption $M> M_{\alpha,\beta}$. It follows from Step 2 that we can steer the control system to a point near $\Omega_\varepsilon$ which is such that $F_i(x)\simeq 0$ and $v_i=\nu\bar v$ for every $i\in\{1,\ldots,N\}$ for some $\bar v$ (that we can choose arbitrarily) such that $\vert\bar v\vert=\sqrt\frac{\alpha}{\beta}$ and some $\nu>0$ small. In other words, we preliminary place the system in a configuration in which all components $v_i$ are small and equal. Note that, in this preliminary step, we are not free to choose the spatial components $x_i$ where we want such as a predetermined flock profile. This will be done in Step 3.3 below, once we assume the size of our control $M$ may overcome the total maximal interaction force.
After having set $v_i=\nu\bar v$, we keep the control active but small to counteract interaction forces ($u_i=F_i\simeq 0$) and we let the system evolve. We have $v_i(t)=v_j(t)$ for all $i,j$ and thus $x_i(t)-x_j(t)$ remains constant. Each variable $v_i$ evolves according to $\dot v_i = (\alpha-\beta\vert v_i\vert^2)v_i$, starting at the initial value $\nu\bar v$, and hence $v_i(t)\rightarrow\bar v$ as $t\rightarrow+\infty$. This is a motion along an heteroclinic orbit, see Section \ref{sec_proof_step3}.
Note that, along this trajectory, in case of unstability the motion can be stabilized by using a feedback control: one linearizes the system along the nominal trajectory, and then correct errors by feedback.
\medskip
\subparagraph{\bf Step 3.2: Passing from a flock velocity to another flock velocity} We now have the control system in a flock and we may want to steer the system to a flock with the same relative positions and a different velocity, without necessarily starting the whole procedure from scratch (i.e., achieve Steps 1, 2, 3.1 again).
This can be done by {using the technique of} quasi-static deformation (see \cite{coron-trelat04,coron-trelat06} and see Section \ref{sec_proof_step3} for details) as follows.
{Let us consider the initial flock $(\bar x(t)=\bar x(0)+\bar v^0 t,\bar v^0)$ meaning that all velocities are equal to $\bar v^0$ with $\vert\bar v^0\vert=\sqrt{\frac{\alpha}{\beta}}$. Assuming that the system is near this initial flock, we want to steer it to (or near to) the target flock $(\bar x(t)=\bar x(0)+\bar v^1 t,\bar v^1)$ with $\vert\bar v^1\vert=\sqrt{\frac{\alpha}{\beta}}$. The idea is to deform the flock sufficiently slowly in time in order to be able to compensate for the small errors by an appropriate feedback law. Recall that, at any given flock, we have $(\alpha-\beta \vert v_i\vert^2 ) v_i=0$ and $F_i(x)=0$.
Let us consider a continuous path $\tau\in[0,1]\mapsto\bar v(\tau)$ satisfying $|\bar v(\tau)|=\sqrt\frac\alpha\beta$ for every $\tau \in[0,1]$ and such that $\bar v(0)=\bar v^0$, the initial velocity of the flock, and $\bar v(1)=\bar v^1$, the velocity of the target flock. Follow the path slowly-in-time, by taking $\tau=\kappa t$ with $\kappa>0$ small enough. Now, along this path, the simplified control system is not autonomous linear but is anyway a \emph{slowly-varying} linear control system, obviously satisfying the Kalman controllability condition. Then, by pole-shifting, it is possible to design feedback controls, tracking this path and thus steering the control system, in time $1/\varepsilon$, to any point of a neighborhood of $\bar v_1$. More details are given in Section \ref{sec_proof_step3}.
Note that stabilization by classical pole-shifting under the Kalman controllability condition may fail in general for non-autonomous linear control systems but remains valid if the matrices of the system are sufficiently slowly varying in time (see \cite[Chap. 9.6]{khalil}).
To summarize, by quasi-static deformation, we can bend the motion of the flock by moving slowly the value of $\bar v$ and thus steer the system to another flock. See a numerical example in Figure \ref{quasi} below.
\medskip
\subparagraph{\bf Step 3.3: Reaching mills} We now make the additional assumption that $M> M_F$.
At the end of Step 1, the system is in $\Omega_\varepsilon$. Since $v=0$ there, we can again take the control \eqref{controlnear} and consider $w$ as a new control, but now, in contrast to the previous steps, since $F(x)$ will not remain small, $w$ is subject to the constraint $\Vert w\Vert\leq \eta$ for some $\eta>0$ {sufficiently small to ensure the constraint $\|u\|\leq M$}. We can however still focus on the simplified control system \eqref{very_simple_control_system} near $v=0$. It is obvious to generate feedback controls $w$, satisfying $\Vert w\Vert \leq \eta$, steering this simplified control system from any point with $v\simeq 0$ to { a configuration in which all agents are equidistributed on a circle (eventually promoting a mill ring). More precisely, this means that} all $x_i$ are placed along a circle of radius $R_\textrm{mill}$ (where $R_\textrm{mill}$ is a value of a possible mill) and all speeds are given by $v_i = \nu x_i^\perp$ for some $\nu>0$ small. This can be done either by quasi-static deformation as before, or by optimal control. Afterwards, we choose the control
\begin{equation*
u_i=F_i-\frac{|v_i|^2}{R_\textrm{mill}} \frac{x_i}{|x_i|}
\end{equation*}
to ensure that each agent undergoes the correct centripetal force {(i.e., $\dot v_i=-\frac{|v_i|^2}{R_\textrm{mill}} \frac{x_i}{|x_i|}$)} and moves along the circle of radius $R_\textrm{mill}$. We let the system evolve and observe that $|x_i(t)-x_j(t)|$ remains constant for each pair $i,j$, as in Step 3.1. Moreover, $v_i$ evolves according to
$$
\dot v_i = (\alpha-\beta\vert v_i\vert^2)v_i-\frac{|v_i|^2}{R_\textrm{mill}} \frac{x_i}{|x_i|}
$$
along an heteroclinic orbit, and $v_i(t)-\sqrt\frac{\alpha}{\beta} \, x_i(t)^\perp \rightarrow 0$ as $t\rightarrow+\infty$, i.e., we have convergence to a mill. Such statements can be easily checked in polar coordinates.
Note that we have steered the system to a mill of which we can choose the center. Moreover, we can even steer the system to some arbitrary mill clusters: to do that, it suffices to choose appropriate mill clusters that are sufficiently far one from each other.
\medskip
\subparagraph{\bf Step 3.4: Reaching flock rings} To reach a flock ring, that is, a flock where all agents are equidistributed along a circle of radius $R_\textrm{mill}$, we first proceed as in Step 3.3, except that the target velocity $v_i = \nu x_i^\perp$ is replaced by $v_i = \nu \bar v$ for some $\bar v$ such that $\vert\bar v\vert=1$ (the desired direction of the flock). We then follow Step 3.1.
\medskip
\subparagraph{\bf Step 3.5: Passing from any flock or mill to any other} By the same procedure as in Step 3.2, it is now clear that we can pass from any flock (or flock ring) or mill to any other, without having to restart the whole procedure at Step 1.
\medskip
\paragraph{\bf Step 4: Sparsification}
The feedback controls defined above are not componentwise sparse, in the sense that, at any instant of time, several (usually all) components of the control are active. In order to keep a minimal amount of intervention at any instant of time, the notion of componentwise sparse control has been introduced in \cite{CFPT2013, CFPT2015} (see also \cite{PRT2015} for the corresponding notion in infinite dimension), meaning that, at any fixed time, at most one component of the control can be active. This notion models the action of one leader on a group of agents, like a single dog acting on a flock of sheep.
It has been shown in \cite{CPRT_A2017} how to design, for dissipative control systems, a componentwise sparse feedback control, starting from any feedback control. This can be done, for instance, by applying an averaging procedure. This is however at the unavoidable price of requiring that $M>N M_{\alpha,\beta}$.
Anyway, this ``sparsification" procedure is general enough to produce sparse feedback controls, steering the control system \eqref{syst} from any initial condition to any flock or mill, in sufficiently large time.
\medskip
We conclude with several remarks on extensions of these results.
\begin{remark}[Regularity of the control] All along the proof, we use controls that are either in a feedback form $u_i=u_i(x_i,v_i)$ and or in an open-loop form $u_i=u_i(t)$. In particular, feedback controls appear for stabilization in \eqref{e-controlO} and are always Lipschitz with respect to $(x_i,v_i)$, ensuring existence and uniqueness of Caratheodory solutions of \eqref{syst} (see \cite{BressanPiccoli,Sontag}). Both open-loop and feedback controls can be smoothened and thus can be chosen as $C^\infty$ functions, with no major difference with respect to the results provided here. Indeed, they can be smoothened even with an arbitrarily small increase of the $C^1$ norm, ensuring that the constraint \eqref{hyp_M_stronger} is always satisfied. \FR{This can be seen as a consequence of the fact that, for systems with Lipschitz vector fields, $L^1$ convergence of controls implies convergence of trajectories, see, e.g., \cite{BressanPiccoli,trelatbook2005}. Indeed a Lipschitz control $u(t,x)$ can always be approximated in $L^1$ by a sequence of $C^\infty$ functions $u^n(t,x)$ satisfying $\|u^n-u\|_{L^1}\to 0$ and the associated solution $(x^n,v^n)$ converges to the solution $(x,v)$.}
\end{remark}
\begin{remark}[On exact controllability]
In Theorem \ref{mainthm}, we have established asymptotic feedback controllability to flocks or mills. Since the linearized system around any flock or mill satisfies the Kalman condition (we are here in finite dimension), it follows that the system \eqref{syst} is locally controllable around flocks and mills. Therefore, as soon as the agents are close enough to a flock (or a mill), one can always design an open-loop control steering the system in finite time exactly to the flock (or to the mill). In other words, in the framework of Theorem \ref{mainthm}:
\begin{itemize}
\item Under \eqref{hyp_M}, the system \eqref{syst} can be steered in large time to any flock.
\item Under \eqref{hyp_M_stronger}, the system \eqref{syst} can be steered in large time to flocks or mills.
\end{itemize}
This is a global controllability result to flocks and mills.
\end{remark}
\begin{remark}[On sharpness of the assumptions and black hole phenomenon]
In Theorem \ref{mainthm}, we have assumed that the value of $M$ is large enough. If $M$ is too small, then it may happen that we do not have a sufficiently strong control to achieve our objectives. Given that, for certain potentials, some of the flocks or mills seem to be strongly attractive, when $M$ is too small we may even fall in the ``black hole phenomenon" (see \cite{PPT2019}). This simply means that the attraction power of a given mill would be too strong to be countered by the control: one then cannot escape from such a basin of attraction. This consideration shows that our assumption on $M$ is, in some sense, unavoidable. However, it is likely that, for some specific classes of potentials, the assumption can be weakened.
\end{remark}
\begin{remark}[Mean-field limit]\label{rem:mf}
All the results in Theorem \ref{mainthm} related to feeback controls are still valid for smooth solutions of the mean-field partial differential equation of Vlasov-type obtained as the formal limit $N\to+\infty$ of \eqref{syst} (see \cite{CPRT2,PRT2015}). This is due to the fact that feedback controls are Lipschitz functions of the state (see \cite{PR18}). The singular character of the flock and mill solutions as solutions of the partial differential equation is not a difficulty, since smooth solutions exist globally and only concentrate in the velocity variable as $t\to+\infty$. We do not provide any details.
\end{remark}
\subsection{Second main result: unbounded interactions}\label{sec_unbounded_interactions}
In Theorem \ref{mainthm}, we have assumed that $\vert U'\vert$ is bounded. However, such an assumption does not involve the case of a potential that explodes near $r=0$, i.e., satisfying $U(0)=-\infty$. Such potentials are often considered in collective behavior models, since they reflect the fact that two agents cannot meet: such an assumption rules out shocks.
In this case, we can refine Theorem \ref{mainthm} as follows.
\begin{theorem}\label{mainthm_variant}
Let $U:(0,+\infty)\rightarrow\mathbb{R}$ generating a radial potential $W(x)=U(|x|)$ of class $C^2$ except at the origin, $\lim_{r\to+\infty} U'(r)=0$ satisfying:
\begin{enumerate}[label=$\bf (U_{\arabic*})$]
\setcounter{enumi}{1}
\item\label{U2} Repulsiveness at $0$: $\lim_{r\to 0}U'(r)=-\infty$ and for every $R>0$ there exist a positive constant $C(R)$ such that $|U(r)|+|U'(r)|+|U''(r)|<C(R)$ for every $r\in[R,+\infty)$.
\end{enumerate}
\begin{enumerate}[leftmargin=*]
\item If $M> M_{\alpha,\beta}$,
then the control system \eqref{syst} can be steered, in sufficiently large time and with a feedback control, to any neighborhood of any flock.
\item Set $\tilde M_F=\sup_{r>0} U'(r)$ and $\tilde M_N=\sup\{ |U'(r)|\,: \, r>2\sin(\tfrac{\pi}N) \bar R\}$. Under the stronger assumption $M> \max\left(M_{\alpha,\beta} + \tilde M_F, \tilde M_N\right)$
for some $\bar R>0$, the control system \eqref{syst} can be steered, in sufficiently large time and with a feedback control, from any initial condition to any neighborhood of any flock ring with radius larger than $\bar R$.
\item Under the stronger assumption $
M> \max\left(M_{\alpha,\beta} + \tilde M_F,\tilde M_N+\frac{\alpha}{\beta \bar R}\right)
$,
for some $\bar R>0$, the control system \eqref{syst} can be steered to any neighborhood of any mill ring with radius larger than $\bar R$.
\end{enumerate}
\end{theorem}
The proof of the first statement is nearly identical to Steps 1, 2 and 3.1 of the proof of Theorem \ref{mainthm}. Few more estimates are needed to ensure convergence of velocities and forces to $0$ (see details in Section \ref{s-21}).
To prove the two last statements, the differences are more significant. Hereafter, we give a brief sketch of the strategy. Full details are given in Section \ref{s-22}.
In contrast to Theorem \ref{mainthm}, and since the potential is now infinite at $0$, to prove Theorem \ref{mainthm_variant} the main idea is now to a priori blow-up the group of agents, i.e., by steering them far from each other. Afterwards, we place all of them along an adequate configuration to let them converge to a flock or a mill. The first step is done by creating a fictitious potential, killing the attractive part of the initial potential $U$ and adding a small repulsive part.
\medskip
\paragraph{\bf Step 1: Blow-up}
In contrast to Theorem \ref{mainthm}, the set $\Omega_0$ may now be empty (for instance, take the potential $U(r)=1/r$), but in the strategy that we develop below, this is actually an advantage, because in this case the Jurdjevic-Quinn strategy steers at least one of the $N$ particles sufficiently far from all others.
\medskip
\subparagraph{\bf Step 1.1: Fictitious purely radial potential}
We apply the control
\begin{equation}\label{control_fictitious}
u_i=\frac{1}{N}\sum_{\substack{j=1\\ j\neq i}}^N \left( \nabla W (x_i-x_j)-\nabla\tilde W (x_i-x_j) \right) +w_i = F_i(x) - \tilde F_i(x) + w_i
\end{equation}
which amounts to replacing in \eqref{syst} the radial potential $U$ with a new potential $\tilde U$, that we choose purely repulsive. This means that we replace the potential $W(x)=U(|x|)$ with $\tilde W(x)=\tilde U(|x|)$. Hence, by choosing this control, we cancel the attractive part of the potential $U$ and we add a small repulsive part. The constraint $\|u\|\leq M$ is satisfied if $\|w\|\leq \tilde M$ with
\begin{equation}\label{tildeM}
\tilde M=M- \sup_{r>0} |\tilde U'(r)-U'(r)|.
\end{equation}
In such a way, the control system \eqref{syst} becomes
\begin{equation}\label{syst2}
\begin{split}
\dot x_i(t) &= v_i(t) \\
\dot v_i(t) &= (\alpha-\beta \vert v_i(t)\vert^2 ) v_i(t) - \tilde F_i(x(t)) + w_i(t)
\end{split}
\end{equation}
where { $\tilde F$ is the force associated to the potential $\tilde W$} and $w$ is the new control, subject to the constraint $\|w(t)\|\leq \tilde M$ for almost every $t$.
\medskip
\subparagraph{\bf Step 1.2: Blowing-up all agents}
We apply Step 1 of Theorem \ref{mainthm} (Jurdjevic-Quinn stabilization procedure) to the modified control system \eqref{syst2}, in order to steer it to the set
$$
\tilde \Omega_\varepsilon = \{ (x,v)\in\mathbb{R}^{4N}\ \mid\ v=0,\ \|\tilde F(x)\|<\varepsilon \} .
$$
By choosing $\varepsilon>0$ sufficiently small, we thus obtain that one of the agents is moved arbitrarily far from all others, since all forces are repulsive. Moreover, by a simple geometric observation, one can prove that this agent is also far from the convex hull of all other agents, i.e., it can be genuinely moved away from all other $N-1$ agents.
Repeating then the same strategy to all agents, one by one, we ultimately obtain that all particles can be moved far away one from each other.
\medskip
\paragraph{\bf Step 2: Circular equidistributed configuration} Since all agents are arbitrarily far, we have $F(x)\simeq 0$ by the assumption $\lim_{r\to+\infty} U'(r)=0$. We can then steer all agents to a circular equidistributed configuration of large radius. The only technical detail is that we need to ensure that all particles keep being sufficiently far one from each other, so that $F(x)\simeq 0$. Details are given in Section \ref{s-22}.
\medskip
\paragraph{\bf Step 3: Converging to the desired target} We finally steer the configuration to the desired one as follows: we first give the desired initial impulse to the velocity variable to have either a flock ring or a mill ring with a large radius (as in the proof of Theorem \ref{mainthm}, Step 3.1 for flock rings and Step 3.3 for mill rings).
We then reduce the radius to the desired $r\geq \bar R$ by applying a suitable central force. If the target is a flock ring, we can counteract the interaction forces thanks to the assumption $M>\sup\{ |U'(r)|\ \mid\ r>2\sin(\tfrac{\pi}N) \bar R\}$. If the target is a mill ring, we also need to counteract the centripetal force, which is possible thanks to the assumption $M>\sup\{ |U'(r)|+\frac{\alpha}{\beta \bar R} \ \mid\ r>2\sin(\tfrac{\pi}N) \bar R \}$.
\section{Local stability and feedback control design}\label{sec_num}
In this section, we recall some known stability properties of flocks and mills established in \cite{ABCV,bertozzi,CHM}. These properties are useful for control design since under conditions of local stability of the corresponding flock or mill solutions, our controls can be be switched off once in a neighborhood of the profiles and the free dynamics of the system will self-regulate towards the desired limiting state. We show how to use the strategies depicted in the previous section to build feedback controls that stabilize to the flock or mill solutions.
\subsection{Local stability of flock manifolds} {We now consider the system \eqref{syst} with no control and we study its stability properties.} First remark that \eqref{syst} with $u\equiv 0$ is obviously invariant under translations and rotations in space, and thus, if we find a flock solution, this particular solution gives rise to infinitely many flock solutions via these invariances. Moreover, once we have a flock solution the direction of the translational velocity of the flock can be freely chosen. It is then natural to deal with flock solutions seen as a manifold of configurations, that we describe in the following
\begin{definition}\label{d-flockmill} Let $(x^*,v^*)$ be an initial configuration such that the corresponding solution of \eqref{syst} is a flock. The \emph{flock manifold} associated to $x^*$ is
$$\mathcal{F}(x^*)=\left\{
\begin{pmatrix} x\\ v \end{pmatrix}\in\mathbb{R}^{4N}
\ \mid\ x\in RT(x^*),\, v=\mathbf{1}_N\otimes \bar v,\, \bar v\in \mathbb{R}^2,\, |\bar v|=\sqrt{\frac{\alpha}{\beta}}
\right\}
$$
where
$
RT(x^*)=\left\{ \mathbf{1}_N\otimes b+\Pt{\mathrm{Id}_N\otimes \mathcal{R}_\theta} x^*
\ \mid\ \theta\in[0,2\pi),\ b\in \mathbb{R}^2\right\}
$ is the family of states obtained by rotations and translations from $x^*$.
{Here, the symbol $M\otimes K$ denotes the Kronecker product: given a matrix of dimension $n\times m$ and a matrix of dimension $p\times r$, the matrix $M\otimes K$ has dimension $np\times mr$ and is built by replacing each element $m_{ij}$ in $M$ with the matrix $m_{ij} K$.}
\end{definition}
The invariance of trajectories plays a crucial role in the study of solutions of \eqref{syst}. In particular, the concept of stability needs to be adapted to the fact that solutions do not converge to a precise point: instead, they converge to the manifold in the sense of the distance, while they go to infinity in the $x$-variable. For this reason, we adapt the classical definition of asymptotic convergence to this setting, by considering neighborhoods invariantly defined based on the metric. We denote by
$
d((x,v),A)=\underset{a\in A}{\inf}\|(x,v)-a\|
$
the distance to the set $A\subset \mathbb{R}^{4N}$.
\begin{definition}\label{d-invnei}
A manifold $A$ is locally asymptotically stable if there exists a $\varepsilon_0$-neighborhood $V_0$ of $A$ such that for each $\varepsilon$-neighborhood $V\subset V_0$ of $A$, there exists a $\delta$-neighborhood $U$ of $A$ such that $(x(0),v(0))\in U$ implies both $(x(t),v(t))=(x_1(t),\ldots,x_N(t),v_1(t),\ldots,v_N(t))\in V$ and $\underset{t\to \infty}{\lim}d((x(t),v(t)),A)=0.$
\end{definition}
The asymptotic stability of the flock manifold was analysed in \cite{CHM}. There, it was shown that it is intimately related to the stability of the linearized system of the associated first-order system \eqref{eq:firstorder}: $\dot x_i=F_i(x)$ with $F_i$ given by \eqref{syst2b}, around a stationary configuration $\hat x$ given by $\dot h = G(\hat x) h$ with
\begin{equation*
G_{ij}(\hat x)=\begin{cases}
-\sum_{k\neq i} \mathrm{Hess}\, W(\hat x_i- \hat x_k)&\mbox{~~ for ~} i=j,\\
\mathrm{Hess}\, W(\hat x_i- \hat x_j) &\mbox{~~ for ~} i\neq j.
\end{cases}
\end{equation*}
Under certain assumptions on the eigenvalues and eigenspaces of $G(\hat x)$, we have a linear system $\dot h = G(\hat x) h$ with a zero eigenvalue of multiplicity $4$, and all other eigenvalues have a negative real part. We refer to \cite[Theorem 1]{CHM} for the exact assumptions needed for local asymptotic stability around the flocking manifold.
\begin{proposition} Let $\hat x$ be a stationary state for \eqref{eq:firstorder}. Then, the manifold $\mathcal{F}(\hat x)$ is locally asymptotically stable for all configurations $x^*$ under certain assumptions on the the eigenvalues and eigenspaces of $G(\hat x)$, in the following sense: any small enough perturbation in the variables $(x,v)$ of the flock solution $z^*\in \mathcal{F}(\hat x)$ associated to $x^*$ under the dynamics \eqref{syst} exponentially converges to $\mathcal{F}(\hat x)$.
\end{proposition}
This attractivity property of $\mathcal{F}(\hat x)$ is natural, in view of the fact that local perturbations in $(x,v)$ might introduce rotations and translations, that is the flock solutions are stable under small perturbations leading to another flock with a small deviation in their direction. In the following, we illustrate the use of these stability properties in conjunction with Theorem \ref{mainthm} to provide an effective feedback design for stabilization towards flock solutions.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.33\textwidth]{F1t1.png}
\includegraphics[width=0.33\textwidth]{E1f1unc.png}\hfill
\includegraphics[width=0.33\textwidth]{E1F1.png}
\\
\includegraphics[width=0.33\textwidth]{t4f.png}\hfill
\includegraphics[width=0.33\textwidth]{t20.png}\hfill
\includegraphics[width=0.33\textwidth]{t200.png}\\
\includegraphics[width=0.33\textwidth]{t4c.png}\hfill
\includegraphics[width=0.33\textwidth]{t20c.png}\hfill
\includegraphics[width=0.33\textwidth]{t200c.png}
\caption{Jurdjevic-Quinn stabilization for the quasi-Morse potential \eqref{qmpot} with $C=0.6, p=1.5$, $l=0.5,\alpha=2,\beta=1.5, N=200$. From left to right. Top: initial state, uncontrolled and controlled energy evolution. Middle: uncontrolled evolution at $t=4,20,200$. Bottom: controlled evolution at $t=4,20,200$. The Jurdjevic-Quinn feedback law stabilizes towards a neighborhood of $\Omega_{\epsilon}$.}\label{step11}
\end{figure}
\subsection{Flock control with a quasi-Morse potential}
Given the potential
\begin{align}\label{qmpot}
W(x)=U(|x|)=V(|x|)-CV(|x|/l)\,,\qquad V(r)=-e^{-r^p/p}\,,
\end{align}
it is known from \cite{CHM2} that repulsive-attractive forces lead to a locally stable flock manifold with a particular spatial configuration. Figure \ref{step11} provides an illustration of free and controlled dynamics for this system. Given a random configuration of particles at rest (top left), the free dynamics evolve towards a ring formation which grows in time, weakening the influence of the potential (middle row). By implementing the Jurdjevic-Quinn feedback control from Theorem \ref{mainthm}, Step 1.1, the evolution is controlled towards a bounded configuration at rest (bottom left): it corresponds to the locally stable uncontrolled profile found in \cite{CHM2}. The uncontrolled dynamics then converge towards a state where $\|F(x)\|=0$ due to expansion, while $\|v(x)\|$ remains constant because of self-propulsion (top, middle). Instead, the Jurdjevic-Quinn stabilization ensures decay on both quantities (top, right), generating a configuration that can be subsequently stabilized towards different flocks (bottom middle, right).
In Figure \ref{quasi}, we illustrate the quasi-static deformation between different flocks pointing towards different directions using the strategy of Step 3.2 of the previous section. Given a time horizon $T$, initial and terminal velocities $v_0$ and $v_T$, characterized by angles $\theta_0$ and $\theta_T$ respectively, and magnitude $\sqrt{\frac{\alpha}{\beta}}$, we transition from $v_0$ to $v_T$ through the action of the linear, time-dependent, feedback control
\begin{align}\label{quasiflock}
u_i(v_i,t)=-M(v_i-\mathcal{R}_{\theta(t)}v_0)\,,\quad \theta(t)=\theta_0+\frac{t}{T}(\theta_T-\theta_0)\,,
\end{align}
where $\mathcal{R}_{\theta}$ is the rotation matrix.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.33\textwidth]{quasistaticinit.png}\hfill
\includegraphics[width=0.33\textwidth]{quasistaticfinal.png}
\includegraphics[width=0.33\textwidth]{quasiangles.png}
\caption{Quasi-static deformation for flock transition with a quasi-Morse potential \eqref{qmpot} with $C=0.6$, $p=1.5, l=0.5,\alpha=2,\beta=1.5, N=200$. Left: initial flock. Middle: terminal flock. Right: average angle evolution. The use of the quasi-static feedback law \eqref{quasiflock} connects different flocks.}\label{quasi}
\end{figure}
\subsection{Local stability of mill rings} \label{s-millstab} We now turn our attention to the study of the local behavior of mill rings. We introduce adapted coordinates, and study the asymptotic stability for solutions exhibiting rotational symmetries only. It is then useful to introduce the time-varying orthonormal frame in $\mathbb{R}^{2N}$
$$
e_i^r=\left(\cos\left(\frac{2\pi i}N+\omega t\right),\sin\left(\frac{2\pi i}N+\omega t\right)\right),\qquad e_i^\perp=(e_i^r)^\perp,
$$
with $e^\perp$ denoting the rotated vector by $\pi/2$ of $e\in\mathbb{R}^2$, and $\omega$ the angular velocity of the mill. Without loss of generality, we consider a mill ring of $N$ agents rotating around the point $(0,0)$. By rearranging indices, we assume that positions and velocities satisfy
$$
x_i(t)=R e^r_i,\qquad v_i(t)=\sqrt\frac{\alpha}{\beta} e^\perp_i.
$$
for $i=1,\ldots, N$. It is clear that $\dot e^r_i=\omega e^\perp_i$ and $\dot e^\perp_i=-\omega e^r_i$. The mill radius $R$ is given by the solution of \eqref{e-radius}. We now consider a perturbation around such a mill solution, that we write as
$$
x_i(t)=(R+r_i(t))\mathcal{R}_{\theta_i(t)}e^r_i,\qquad v_i(t)=\left(\sqrt\frac{\alpha}{\beta}+w_i(t)\right) \mathcal{R}_{\tau_i(t)}e^\perp_i,
$$
{ by writing:
\begin{itemize}
\item $x_i$ in polar coordinates, where $r_i,\theta_i$ are perturbations of the radius and angle variables with respect to $R, \mathrm{arg}(e^r_i)$;
\item $v_i$ in polar coordinates, where $w_i,\tau_i$ are perturbations of the radius and angle variables with respect to $\sqrt\frac{\alpha}{\beta}, \mathrm{arg}(e^\perp_i)$.
\end{itemize}
Here, $\mathrm{arg}(v)$ is the argument of the nonzero vector $v$, i.e., the angle in its polar coordinates.}
A straightforward computation shows that the system \eqref{syst} can be written in term of the new variables $r_i,\theta_i,w_i,\tau_i$.
A general result of stability around such solutions seems out of reach (see some results for a first-order system with a similar structure in \cite{bertozzi}). We instead briefly investigate the local stability of solutions with rotational symmetries, i.e., solutions that satisfy
$
r_i=r_j, \theta_i=\theta_j, w_i=w_j,\tau_i=\tau_j$ with $i,j=1,\ldots, N.
$
Notice that due to the rotational symmetry, the last equation becomes $\dot \tau_i=-\omega$. By dropping the index for variables $r_i,w_i$ and introducing the variable $\gamma=\theta_i-\tau_i$, the system is reduced to
\begin{equation}\label{e-mill-lin}
\begin{split}
\dot r&=\left(\sqrt\frac{\alpha}{\beta}+w \right)\sin(\gamma)\,,\qquad
\dot \gamma=\left(\frac{\sqrt\frac{\alpha}{\beta}+w}{R+r}-\frac{\phi(r)}{\sqrt\frac{\alpha}{\beta}+w}\right)\cos(\gamma)\,,\\
\dot w&=-\left(2\sqrt{\alpha\beta}w+\beta w^2\right)\left(\sqrt\frac{\alpha}{\beta}+w \right)-\phi(r)\sin(\gamma)\,,
\end{split}
\end{equation}
where
\begin{eqnarray*}
\phi(r)&=&(1,0)\cdot \frac1N\sum_{j=1}^{N-1}\nabla W\left((R+r)((1,0)-\left(\cos\left(\frac{2\pi j}N\right),\sin\left(\frac{2\pi j}N\right)\right)\right)\\
&=&\frac1N\sum_{j=1}^{N-1}\sin\left(\frac{\pi j}N\right)U'\left((R+r)\sin\left(\frac{\pi j}N\right)\right)\,.
\end{eqnarray*}
Note that, since \eqref{e-radius} is satisfied, we have
$
\frac{\sqrt\frac{\alpha}{\beta}}{R}=\frac{\phi(0)}{\sqrt\frac{\alpha}{\beta}}=\omega.
$
In particular, the trajectory $(r(t),\gamma(t),w(t))=(0,0,0)$ is a solution of \eqref{e-mill-lin}. This shows invariance of the mill solution under a same translation in all variables $\alpha_i,\beta_i$. One can easily study linear stability properties for \eqref{e-mill-lin}: the linearized system is given by
\begin{equation*}
\begin{pmatrix}
\dot r\\\dot \gamma\\\dot w
\end{pmatrix}=A \begin{pmatrix}
r\\ \gamma\\ w
\end{pmatrix} .
\mbox{~~~with~~~}A=
\begin{pmatrix}
0 & \sqrt\frac{\alpha}{\beta} & 0\\
-\frac{\omega}R -\frac{\phi'(0)}{\sqrt\frac{\beta}{\alpha}} & 0& \frac{2}{R}\\
0 & -\phi(0) & -2\alpha
\end{pmatrix}
\end{equation*}
{One can compute the characteristic polynomial $\lambda^3+a_2\lambda^2+a_1\lambda^1+a_0$ of $A$, by identifying coefficients with principal minors of $A$. We have
$$
a_2=-\mathrm{Tr}(A)=2\alpha,\qquad
a_1=\left(\frac{\omega}R +\frac{\phi'(0)}{\sqrt\frac{\beta}{\alpha}}\right)\sqrt\frac{\alpha}{\beta}+\phi(0)\frac2R ,
$$
$$
a_0=-\mathrm{det}(A)=\left(\frac{\omega}R +\frac{\phi'(0)}{\sqrt\frac{\beta}{\alpha}}\right)\sqrt\frac{\alpha}{\beta}(2\alpha) .
$$
Recall a special case of the Routh–Hurwitz criterion for third-order monic polynomials (see \cite{trelatbook2005}): the linear stability property is ensured if and only if $a_2,a_0,a_2a_1-a_0>0$. Then, recalling that $\alpha,\beta,\phi(0)>0$, the linear stability property is ensured by the condition $\frac{\omega}R +\sqrt\frac{\alpha}{\beta}\phi'(0)>0.$}
This condition is always satisfied for power-law potentials of the form
\begin{align}\label{plaw}
U(s)=\frac{|s|^a}a - \frac{|s|^b}{b},\qquad a>b>0 ,
\end{align}
since we have in this case
$$
\phi'(0)=\sqrt\frac{\alpha}{\beta}\omega(a-b)R^b\sum_{j=1}^{N-1}\sin\left(\frac{\pi j}N\right).
$$
This linear stability analysis shows that the equilibrium solution $(r(t),\gamma(t),w(t))=(0,0,0)$, i.e., the mill ring solution is a locally asymptotically stable equilibrium point to \eqref{e-mill-lin}. As a consequence, the mill ring solution is locally asymptotically stable for perturbations keeping the rotational symmetry of \eqref{syst}. In the following, we explore the interplay between stability of mill rings and control design for power law potentials.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.49\textwidth]{free13d.png}\hfill
\includegraphics[width=0.49\textwidth]{free1en.png}\\
\includegraphics[width=0.49\textwidth]{R0n.png}\hfill
\includegraphics[width=0.49\textwidth]{G0n.png}
\caption{Mill ring stability in the power-law potential \eqref{plaw} with $a=4,b=1,\alpha=10, \beta=3$, $N=200$. Top left: an initial ring configuration converges to mill ring solution (sample agent trajectory in grey). Top right: energy evolution of the swarm towards the mill. Bottom left: evolution of the distance to the stable mill radius $R$ for mill configurations departing from an initial radius $R_0$. Bottom right: evolution of the distance to stable mill radius for rings of radius $R$, initial velocity rotated from the tangential velocity by $\gamma_0$. The evolution to the stable mill configuration is robust to perturbations.}
\label{mill1}
\end{figure}
\subsection{Controlling mills for a power-law potential}
Although the local stability analysis is enough for our purposes of controlling the system \eqref{syst} towards mill ring solutions, we observe that for power-law potentials \eqref{plaw}, the mill ring is globally asymptotically stable for solutions with rotational symmetry. Figure \ref{mill1} (top row) shows the evolution of a particular solution, not a small perturbation, with rotational symmetry, converging towards a stable mill with radius given by \eqref{e-radius}, similar to a nonlinear damped oscillator. The second row further illustrates the stability of the mill ring by considering the evolution of two types of perturbations. On the left, the initial configuration is a mill ring with radius different from the stable solution. On the right, the initial configuration is a ring of stable radius, however the tangential velocities are shifted by an angle $\gamma_0$. In both plots, the vertical axis represents the distance with respect to the stable radius, and we can observe that for both types of perturbations the uncontrolled dynamics stabilize towards the mill.
Feedback controls can be used to induce or accelerate convergence to mill ring solutions. For example, given a stable flock ring configuration, the system can be stabilized towards a mill through the action of the feedback law
\begin{equation*}
u_i(v_i(t))=-M\left(v_i-\sqrt{\frac{\alpha}{\beta}}\frac{x_i^\perp}{|x_i^\perp|}\right)\,.
\end{equation*}
A design alternative is to resort to instantaneous controls \cite{APZ14}, which can be interpreted as feedback laws in the same spirit of model predictive control strategies. We synthesize a feedback control by solving
\begin{equation}\label{instfeed}
\underset{u\in[-1,1]^2}{\min} \sum\limits_{i=1}^N \left|v_i-\sqrt{\frac{\alpha}{\beta}}\frac{x_i^\perp}{|x_i^{\perp}|^2}\right|+(|x_i-x_m|^2-R_m^2)^2+\lambda_1|u|+\lambda_2|u|^2\,,\quad \lambda_1,\lambda_2>0\,,
\end{equation}
where $x_m$ corresponds to the center of mass of the swarm, $R_m$ is the desired mill radius, and $(x_i,v_i)$ is the future state of the system after a small control horizon $\Delta t$. We consider $\ell_1$ and $\ell_2$-norm control penalties to induce sparsification in time. For the sake of real-time computability, this optimization is reduced to a single control signal $u\in\mathbb{R}^2$, which enters the dynamics through an incremental rotation of $\frac{2\pi}{N}$. This control differs from the signal that would be obtained optimizing each $u_i$ separately, however it still succeeds in stabilizing around the mill ring solution, as shown in Figure \ref{instcont}, presumably due to the large basin of attraction surrounding the mill ring solution.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.33\textwidth]{randtomillt0.png}\hfill
\includegraphics[width=0.33\textwidth]{randtomillt10.png}\hfill
\includegraphics[width=0.33\textwidth]{randtomillt40.png}\\
\includegraphics[width=0.49\textwidth]{randtomill.png}\hfill
\includegraphics[width=0.49\textwidth]{randtomillc.png}\\
\caption{Stabilization towards a mill ring for the power-law potential \eqref{plaw} with $a=4,b=1$, $\alpha=10, \beta=3$, $N=20$, using the feedback \eqref{instfeed}. Top (left to right): swarm at $t=0,10,40$. Bottom left: evolution of the configuration radius towards the stable mill. Bottom right: control signal.}\label{instcont}
\end{figure}
We apply a similar idea to control a stable mill towards a flocking configuration of different radius. In this case we compute one control variable per agent by solving
\begin{equation}\label{instfeed2}
\underset{u\in[-1,1]^{2N}}{\min} \sum\limits_{i=1}^N |v_i-\bar v|^2+(|x_i-x_m|^2-R_f^2)^2+\lambda|u_i|^2\,,\quad \lambda>0\,,
\end{equation}
where $\bar v,R_f$ are the desired flocking velocity and radius, respectively. Figure \ref{mill2flock} illustrates the transition from the milling to the flocking regime.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.33\textwidth]{t0.png}\hfill
\includegraphics[width=0.33\textwidth]{t4.png}\hfill
\includegraphics[width=0.33\textwidth]{t40.png}\\
\includegraphics[width=0.49\textwidth]{agent10.png}\hfill
\includegraphics[width=0.49\textwidth]{mill2flocke.png}
\caption{Controlled transition from mill to flock for the power-law potential ($a=4,b=1,\alpha=10,\beta=3,N=20$), using the feedback \eqref{instfeed2}. Top: swarm at $t=0,4,40$. Bottom left: a sample control signal. Bottom right: evolution of the swarm radius, from a stable mill to a stable flock. }\label{mill2flock}
\end{figure}
\section{Proof of Theorem \ref{mainthm}}\label{sec_proof_mainthm}
This section is devoted to provide the full detail of the proof of Theorem \ref{mainthm}, sketched in Section \ref{sec_bounded_interactions}.
\subsection{Proof of Step 1}\label{sec_proof_step1}
In this first step, the goal is to steer the control system from any initial point to a point in $\Omega_\varepsilon$.
\subsubsection*{Step 1.1. Jurdjevic-Quinn stabilization}
Let $
\gamma>\max\left(1,\frac1M \sqrt{\frac{\alpha^3}{\beta}}\right)
$
be fixed. We apply the feedback control
\begin{equation}
u_i(v_i)=
\begin{cases}
0& \text{if }\, |v_i|\geq 2\gamma\sqrt{\frac{\alpha}{\beta}},\\
-M\frac{v_i}{|v_i|}\left(2-\frac{|v_i|}{\gamma\sqrt{\frac{\alpha}{\beta}}}\right)& \text{if }\, |v_i|\in\left(\gamma\sqrt{\frac{\alpha}{\beta}},2\gamma\sqrt{\frac{\alpha}{\beta}}\right),\\
-M\frac{v_i}{|v_i|}& \text{if }\, |v_i|\in\left[\frac{ \sqrt{\frac{\alpha}{\beta}}}{\gamma},\gamma\sqrt{\frac{\alpha}{\beta}}\right],\\
-M\frac{\gamma}{\sqrt{\frac{\alpha}{\beta}}} v_i & \text{if }\, |v_i|<\frac{ \sqrt{\frac{\alpha}{\beta}}}{\gamma}.
\end{cases}\label{e-controlO}
\end{equation}
Since the control is Lipschitz with respect to the $(x,v)$ variables, we have existence and uniqueness of solutions of \eqref{syst} for a fixed initial condition. Note that the control law satisfies the constraint $|u_i|\leq M$ by construction. Setting the total energy of the system as
\begin{equation*
V(t)=\frac12 \sum_{i=1}^N |v_i(t)|^2+{\frac1{2N}} \sum_{\substack{i,j=1\\ i\neq j}}^N W(x_i(t)-x_j(t)) ,
\end{equation*}
using \eqref{syst}, we have
\begin{equation*}
\dot V =\sum_{i=1}^N v_i\cdot \dot v_i+{\frac1{2N}} \sum_{\substack{i,j=1\\ i\neq j}}^N \nabla W(x_i-x_j)(\dot x_i-\dot x_j) =
\sum_{i=1}^N \left((\alpha-\beta|v_i|^2)|v_i|^2+v_i\cdot u_i\right).
\end{equation*}
Notice that by skew-symmetry of $\nabla W(x_i-x_j)$, we have
$$
\sum_{\substack{i,j=1\\ i\neq j}}^N \nabla W(x_i-x_j)\cdot v_j=-\sum_{\substack{i,j=1\\ i\neq j}}^N \nabla W(x_i-x_j)\cdot v_i\,.
$$
Given $a_1=\frac{1}{\gamma} \sqrt{\frac{\alpha}{\beta}}$ and $a_2=\gamma\sqrt{\frac{\alpha}{\beta}}$, we split the indices $i=1,\ldots, N$ in the sets
$$
I_{\infty}(t)=\left\{i: |v_i(t)|>a_2\right\},\, I_{1}(t)=\left\{i: |v_i(t)|\in\left[a_1,a_2\right]\right\}\mbox{ and}\,
I_{0}(t)=\left\{i:|v_i(t)|<a_1\right\}.
$$
Apply the control law \eqref{e-controlO} and notice that for every $i\in I_\infty(t)$ we obtain
$$
\left\{v_i\cdot u_i=0 \quad\textrm{or} \quad v_i\cdot u_i=-M|v_i|\left(2-\frac{|v_i|}{\gamma\sqrt{\frac{\alpha}{\beta}}}\right)<0\right\} \quad\mbox{and}\quad (\alpha-\beta|v_i|^2)|v_i|^2<0.
$$
We deduce
\begin{multline*}
\dot V \leq \sum_{i\in I_\infty(t)}(\alpha-\beta|v_i|^2)|v_i|^2+\sum_{i\in I_1(t)}|v_i|(\alpha|v_i|-\beta|v_i|^3-M)\\
+\sum_{i\in I_0(t)}\left(\alpha-\beta|v_i|^2-M\frac{\gamma}{\sqrt{\frac{\alpha}{\beta}}}\right)|v_i|^2.
\end{multline*}
The maximum of the function $v\to \alpha v-\beta v^3$ over $v\geq 0$ is $\sqrt\frac{4\alpha^{3}}{27\beta} <M$. We have $\alpha|v_i|-\beta|v_i|^3-M<0$ and $\alpha-M\frac{\gamma}{\sqrt{\frac{\alpha}{\beta}}}<0$ by the choice of $\gamma$. This implies that
\begin{equation}\label{e-stimaV}
\dot V\leq \sum_{i\in I_\infty(t)}(\alpha-\beta|v_i|^2)|v_i|^2-\sum_{i\in I_1(t)\cup I_0(t)}\beta|v_i|^4.
\end{equation}
The right-hand side is then nonpositive being the sum of two nonpositive terms, hence $\dot V\leq 0$.
Note that we cannot directly apply the LaSalle invariance principle to the system, because we cannot ensure boundedness of the trajectories and the functional $V$ is not proper.
Nevertheless, $V$ is bounded below, since both $|v|$ and $W$ are bounded below, as a consequence of the corresponding assumption on $U$. Since $V(t)$ is decreasing, hence bounded above, both $v_i(t)$ and $W(x_i(t)-x_j(t))$ are bounded. Thanks to the assumption \ref{U1}, $\nabla W$ is bounded. This implies that $\dot v_i(t)$ is bounded too.
\begin{lemma}\label{l-v0}
{The system \eqref{syst} with control law \eqref{e-controlO} satisfies} $\lim_{t\to +\infty} v_i(t)=0$ for $i=1,\ldots, N$.
\end{lemma}
\begin{proof}
By contradiction, if this is not the case, there exists an index $i$ and a sequence of times $t_k\to\infty$ such that $|v_i(t_k)|>C$. Since $\dot v_i$ is bounded, this implies that there exists a uniform $\tau$ such that $|v_i(t)|>C/2$ for all $t\in(t_k-\tau,t_k+\tau)$. By using this property in \eqref{e-stimaV} discarding the first term in the right-hand side, we infer that $V(t)\to -\infty$. This raises a contradiction.
\end{proof}
\begin{lemma}\label{l-F0}
{The system \eqref{syst} with control law \eqref{e-controlO} satisfies}
$\lim_{t\to +\infty} F_i(x(t))=0$ for $i=1,\ldots, N$.
\end{lemma}
\begin{proof}
Since the $v_i$s are bounded, then the functions $x_i(t)-x_j(t)$ are Lipschitz. Since \ref{U1} holds, then both $\nabla W(x_i(t)-x_j(t))$ and its derivative are bounded; hence functions $\nabla W(x_i(t)-x_j(t))$ are Lipschitz too, with a Lipschitz constant that we denote with $L$.
Assume now, by contradiction, that there exists an index $i$ such that $F_i(x(t))$ does not converge to 0. Thus, there exists a sequence of times $t_k\to\infty$ such that
\begin{enumerate}
\item either $F_i(x(t_k))\to \bar F$ for some non-zero vector $\bar F$;
\item or $|F_i(x(t_k))|\to +\infty$.
\end{enumerate}
In the first case, for each $\varepsilon>0$ there exists an index $K>0$ such that $\|F_i(x(t))-\bar F\|<\varepsilon+L(t-t_k)$ for all $k>K$ and $t>t_k$. Recalling that $\lim_{t\to +\infty} v_i(t)=0$, take now $\eta>0$ sufficiently small and $k$ sufficiently large to have both $|v_i(t)|<\eta$ and $|\alpha v_i(t)-\beta v_i(t)|\cdot |v_i(t)|^2<\eta$ for all $t>t_k$. This implies $|\dot v_i(t_k+\tau)-\bar F|<2\eta+L\tau$ for all $\tau>0$. Since $|v_i(t_k)|<\eta$, then $|v_i(t_k+\tau)-\bar F\tau|<\eta+2\eta\tau+L\frac{\tau^2}2$ for $\tau>0$. Fix $\tau>0$ sufficiently small to have $L\frac{\tau^2}2<|\bar F|\frac{\tau}2$ and observe that this implies $|v_i(t_k+\tau)|>|\bar F|\frac{\tau}2-\eta-2\eta\tau$. Let $\eta\to 0$ and note that this implies $\lim_{k\to\infty}v_i(t_k+\tau)\neq 0$. This raises a contradiction.
The second case is similar: consider the unit vectors $\frac{F_i(x(t))}{|F_i(x(t))|}$, that admit a converging subsequence (that we do not relabel) to an unitary vector $\bar F$. Following computations of the previous case, we have $|v_i(t_k+\tau)|>|F_i(x(t_k))| \frac{\tau}2-\eta-2\eta\tau$, which does not converge to $0$ for $\eta\to 0$ and $k\to\infty$. This raises a contradiction.
\end{proof}
Finally, let us choose a time $T_0$ at which we stop the control strategy \eqref{e-controlO}. This choice is driven by correctly initializing the next step. Let $\varepsilon'>0$ be a constant to be chosen later. Since both $v_i(t)$ and $F_i(x(t))$ converge to $0$, we choose a time $T_0$ at which $|v_i(T_0)|<\varepsilon'$ and $|F_i(x(T_0))|<\varepsilon'$ for $i=1,\ldots,N$.
\subsubsection*{Step 1.2. Reaching $\Omega_\varepsilon$ in finite time}
We now steer each $v_i$ exactly to zero. {We first define the trajectory for each $v_i$: this in turn gives the trajectory of the $x_i$ by integration, and the control $u_i$ by identification in the second equation of \eqref{syst}}. Choose $T_{1,i}=T_0+\tfrac1{\varepsilon'}|v_i(T_0)|$ and define
$$
v_i(t)=\begin{cases}
v_i(T_0)-\varepsilon'(t-T_0)\frac{v_i(T_0)}{|v_i(T_0)|}&\mbox{~~for~}t\in[T_0,T_{1,i}],\\
0&\mbox{~~for~}t>T_{1,i}.
\end{cases}
$$
Then choose $T_1=\max(T_{1,i})$ as the final time of the strategy.
A direct computation shows that
\begin{equation}\label{e-xi12}
|x_i(t)-x_i(T_0)|\leq \int_{T_0}^{T_{i,1}}|v_i(T_0)|-\varepsilon'(t-T_0)\,dt\leq \frac{|v_i(T_0)|^2}{2\varepsilon'}<\frac{\varepsilon'}{2}
\end{equation}
for every $t\in[T_0,T_1]$.
{Since $\nabla W(x_i-x_j)$ is $L$-Lipschitz continuous and bounded, as recalled in Lemma \ref{l-F0}, we infer that
$
\|F_i(T_0+t)\|\leq \|F_i(T_0)\|+L\tfrac{\varepsilon'}{2}<(1+\tfrac{L}2)\varepsilon'.
$
Notice that $u$ bounded implies that all solutions of (1.1) are Lipschitz with respect to time. This implies that there exists $\bar x$ (that we cannot choose) such that} $(x(T_1),v(T_1))=(\bar x,0)\in\Omega_\varepsilon$ by imposing $\varepsilon'\leq \varepsilon/(1+\tfrac{L}2)$.
By a simple estimate in the second equation of \eqref{syst}, the control satisfies
$$
|u_i|\leq |\dot v_i|+(\alpha-\beta |v_i|^2)|v_i|+|F_i|< \varepsilon'+\alpha \varepsilon'+(1+\tfrac{L}2)\varepsilon'\leq M,
$$
by imposing $\varepsilon'\leq M/(2+\alpha+\tfrac{L}2)$.
Summing up, choosing $\varepsilon'=\min\left(\frac\varepsilon{1+\tfrac{L}2},\frac{M}{2+\alpha+\tfrac{L}2}\right)$, all conditions are satisfied.
\subsection{Proof of Step 2}\label{sec_proof_step2}
Step 2 is obvious, the control system \eqref{very_simple_control_system} being straightforward to control. We do not provide any detail. \FR{Given a connected neighborhood $\mathcal{N}$ of $(\bar x,0)$ inside $\Omega_\varepsilon$, we are then able to steer the system to any chosen point in $\mathcal{N}$.}
\subsection{Proof of Step 3}\label{sec_proof_step3}
\subsubsection*{Step 3.1: Reaching flocks}
{Fix a unit vector $\bar v$ and note that $\Omega_\varepsilon$ open implies that there exists $\delta>0$ such that the configuration $(\bar x, \delta \bar v)$ belongs to the neighborhood $\mathcal{N}$ of $(\bar x, 0)$ given at Step 2}. Then, we can steer the system from $(\bar x,0)$ to $(\bar x, \delta \bar v)$ at a time $T_2>T_1$ with a control $u=\bar u+z$ satisfying $\|u\|<2\varepsilon<M$, again by a local controllability argument.
We then choose the controls $u_i=F_i$ on the time interval $[T_2,+\infty)$. The velocity variables are then the solutions of
\begin{equation}\label{e-vfree}
\dot v_i(t)=(\alpha -\beta|v_i(t)|^2)v_i(t),\qquad v_i(T_2)= \tfrac{\delta}2 \bar v.
\end{equation}
They all coincide at each time, i.e., $v_i(t)=v_j(t)$, hence relative positions are all constant with respect to time, i.e $x_i(t)-x_j(t)=\bar x_i-\bar x_j$. This in turn implies that interaction forces keep being constant with respect to time, thus $\|u_i(t)\|=\|F_i(t)\|=\|F_i(T_2)\|<\varepsilon$, since $(x(T_2),v(T_2))\in\Omega_\varepsilon$.
Moreover, all velocities converge to $\sqrt\frac\alpha\beta \bar v$ for $t\to +\infty$, since they solve \eqref{e-vfree}. This implies that the system converges to an $\varepsilon$-flock, as stated.
\begin{remark}
The motion of the velocity variables $v_i(t)$, solutions of \eqref{e-vfree}, exactly follows \emph{heteroclinic trajectories}, in the sense that $v_i(t)$ passes from (a neighborhood of) the unstable equilibrium $0$ to the asymptotically stable family of equilibria $\left\{\|v\|=\sqrt\frac\alpha\beta\right\}$. The existence of such heteroclinic trajectories is certainly one of the main interesting features of the dynamics of \eqref{syst}, promoting convergence to flocks or mills.
\end{remark}
\subsubsection*{Step 3.2: Passing from a flock to another flock}
Assume that the system is at (or near) a flock of velocity $\bar v^0$. We want to steer the system to another flock, of velocity $\bar v^1$. Along the motion, the relative positions $x_i-x_j$ will remain constant.
The strategy that we use here is by \emph{quasi-static deformation}. Take a continuous path $\tau\in[0,1]\mapsto\bar v(\tau)$ such that $\bar v(0)=\bar v^0$ and $\bar v(1)=\bar v^1$, satisfying $\|\bar v(\tau)\|=\sqrt\frac\alpha\beta$ for every $\tau\in[0,1]$, e.g., the shortest arc on the circle.
{Given any fixed $\tau\in[0,1]$, the flock under consideration is $(\bar x(t) = \bar x(0)+\bar v(\tau) t, \bar v(\tau))$.}
Since relative positions do not change, forces $F_i$ do not change and then, for each $\tau$, we have a flock of velocity $\bar v(\tau)$.
Of course, the corresponding path of flocks, parametrized by $\tau\in[0,1]$ is \emph{not} a solution of \eqref{syst}. It is rather to be thought of as a path of equilibrium points for the dynamics \eqref{syst}.
Following the idea of \cite{coron-trelat04,coron-trelat06}, we track this path, in large time, by designing appropriate feedback controls.
{To this aim, for any given fixed $\tau\in[0,1]$, we linearize the control system \eqref{syst} at the corresponding flock: we set
$$
x_i(t) = \bar x_i(0)+\bar v(\tau)t+\delta x_i(t), \qquad v_i(t)=\bar v(\tau)+\delta v(t) .
$$
Plugging in \eqref{syst}, using that $F_i(\bar x(t))=0$ and that $\vert\bar v(\tau)\vert=\sqrt{\frac{\alpha}{\beta}}$, we get, at the first order, the linear system
$$
\delta\dot x_i(t) = \delta v_i(t),\qquad \delta \dot v_i(t) = 2\langle\bar v(\tau),\delta v(t)\rangle\bar v(\tau)+dF_i(\bar x_i(0)+\bar v(\tau)t).\delta x(t)+u_i(t) .
$$
We make a change of control by setting $\delta u_i(t) = -2\langle\bar v(\tau),\delta v(t)\rangle\bar v(\tau)-dF_i(\bar x_i(0)+\bar v(\tau)t).\delta x(t)+w_i(t)$, thus obtaining the very simple control system
$$
\delta\dot x_i(t) = \delta v_i(t),\qquad \delta \dot v_i(t) = w_i(t),
$$
i.e., we recover the system \eqref{very_simple_control_system}.
}
{This has been done for $\tau$ fixed. Now, the idea is to perform the above deformation slowly in time, by setting $\tau=\varepsilon t$, for some $\varepsilon>0$ small enough, and thus $t\in[0,1/\varepsilon]$, and compensate for the errors by designing an adequate feedback control.}
Of course, for every fixed value of $\tau$, the above control system is linear autonomous, of the form $\dot X(t) = AX(t)+Bu(t)$ for some matrices $A$ and $B$. It obviously satisfies the {Kalman controllability condition} and is thus controllable and also feedback stabilizable (for instance by standard pole shifting, see, e.g., \cite{LeeMarkus,Sontag,trelatbook2005}). But now, along the path of flocks {that we want to track slowly in time}, we do not have anymore a linear autonomous control system, but a linear \emph{instationary} control system, of the form $\dot X(t) = A(\varepsilon t)X(t)+B(\varepsilon t)u(t)$ for some matrices $A(\varepsilon t)$ and $B(\varepsilon t)$ depending on time {but varying slowly in time}. For every $\tau$, the pair $(A(\tau),B(\tau))$ still satisfies the Kalman condition. For linear instationary control systems the Kalman condition is not sufficient to ensure controllability nor stabilizability properties (see counterexamples, e.g., in \cite{khalil,Sontag,trelatbook2005}). But here, we follow the path slowly in time: by setting $\tau=\varepsilon t$, the abovementioned control system takes the form $\dot X(t) = A(\varepsilon t)X(t)+B(\varepsilon t)u(t)$ and is therefore a \emph{slowly-varying} (in time) linear control system, satisfying the Kalman condition. As explained in detail in \cite{coron-trelat04}, and according to an argument of \cite[Chap. 9.6]{khalil}, if $\varepsilon>0$ is small enough then the Kalman condition is still sufficient to ensure that this slowly-varying linear control system can be feedback stabilized by usual pole shifting, with a feedback control of the form $u(t) = K(\varepsilon t) X(t)$. Note anyway that such a feedback is also slowly varing in time, so is not a ``pure" feedback. One may want to obtain a feedback, not depending on time, but defined piecewise in time. This is possible by slightly modifying the above definition of the feedback. The resulting \emph{staircase method} has been used, e.g., in \cite{PoucholTrelatZuazua}.
Eventually, such feedback controls make it possible to track the path of flocks and thus steer the control system, in time $1/\varepsilon$, to any point of a neighborhood of $\bar v_1$.
If one moreover aims to choose precise $x$-positions (keeping anyway the same relative positions as those of the initial flock), it is sufficient to observe that all such configurations differ from a translation vector $X$. Therefore, it suffices to use a quasi-static deformation on the positions as well.
\subsubsection*{Proof of Steps 3.3, 3.4, 3.5: reaching mills, flock rings}
The strategy, described in Section \ref{sec_bounded_interactions}, is similar to what has been described above, and we thus do not give any detail.
\section{Proof of Theorem \ref{mainthm_variant}}\label{sec_proof_thm2}
\subsection{Proof of the first statement} \label{s-21}
The proof of the first statement is almost identical to the proof of Theorem \ref{mainthm}, with the following differences:
\begin{itemize}
\item In Step 1.1, we follow the proof until $\dot V\leq 0$, due to \eqref{e-stimaV}. Since $V$ is bounded below, both $v_i$ and $W_i(x_i-x_j)$ are bounded below. Now, using the assumptions \ref{U2}, boundedness of $W$ implies that $x_i(t)-x_j(t)$ is bounded away from $0$ by a constant $2\ell>0$, hence $\nabla W( x_i(t)-x_j(t))$ is bounded. This in turn implies that $\dot v_i(t)$ is bounded as well. One can then prove Lemmas \ref{l-v0} and \ref{l-F0} in this case too.
\item In Step 1.2, we observe that \eqref{e-xi12} ensures that $|x_i(t)-x_j(t)|>\ell$, provided $\varepsilon'<\sqrt{2\ell}$. We then use \ref{U2} to ensure that $\nabla W$ is $L$-Lipschitz continuous for $|x_i(t)-x_j(t)|>\ell${, for some $L>0$}. We now choose $\varepsilon'=\min(\sqrt{2\ell},\tfrac{2}{L},\tfrac\eps2)$ to ensure that $(x(T_1),v(T_1))\in\Omega_\varepsilon$.
\end{itemize}
\subsection{Proof of the second statement}\label{s-22}
We follow the sketch of the proof given in Section \ref{sec_unbounded_interactions}.
\subsubsection{Proof of Step 1}
\paragraph{\bf Step 1.1: Fictitious purely radial potential}
The proof is based on the method of ``artificial potential field'', which is widely used in robotics (see, e.g., \cite[Chap. 7]{spong}). Replace the potential $U$ with a purely repulsive potential $\tilde U$, that is \FR{chosen as follows. Define $\eta=\frac14(M-(M_{\alpha,\beta}+\tilde M_F))>0$ and choose $R_0>0$ such that $|U'(r)|<\eta$ for all $r\geq R_0$. This is possible because $\lim_{r\to+\infty}U'(r)=0$. Take now $\phi:(0,+\infty)\to\mathbb{R}$ a $C^\infty$ function with bounded $C^1$ derivative, satisfying}
\FR{\begin{itemize}
\item $\phi(r)=1$ for $r\in (0,R_0]$ and $\phi(r)=0$ for $r\in[R_0+1,+\infty)$;
\item $\phi$ is decreasing for $r\in[R_0,R_0+1]$.
\end{itemize}
Define
$$\tilde U'(r)=\phi(r)(U'(r)-\tilde M_F)-\eta\frac{1}{1+r^2}$$}
\FR{and $\tilde U(r)=\int_1^r \tilde U'(r)$. Then $\tilde U$ is $C^2$, satisfies $\lim_{r\to+\infty} \tilde U'(r)=0$ and \ref{U2}. More crucially, $\tilde U'(r)<0$ for all times, i.e., $\tilde U$ is purely repulsive. Moreover, $|\tilde U'(r)-U'(r)|\leq \tilde M_F+\eta$ for $r\in (0,R_0)$, and $|\tilde U'(r)-U'(r)|\leq |\phi(r)-1| |U'(r)|+\phi(r)\tilde M_F+\eta\leq \eta+\tilde M_F+\eta$ for $r\in [R_0,+\infty)$. Then
\begin{equation}\label{e-U'}
|\tilde U'(r)-U'(r)|\leq \tilde M_F+2\eta.
\end{equation}}
\FR{Define now $\tilde W(x)=\tilde U(|r|)$ and choose the control $u_i=F_i(x)-\tilde F_i(x) +w_i$ (see \eqref{control_fictitious}), so that the new control system is \eqref{syst2}, with the new controls $w_i$ satisfying $\|w_i\|\leq \tilde M$ with $\tilde M$ defined by \eqref{tildeM}. Note that \eqref{e-U'} implies
\begin{equation}\label{e-w-const}
\tilde M\geq M-(\tilde M_F-2\eta)=\tilde M_{\alpha,\beta}+\eta ,
\end{equation}
i.e., the system \eqref{syst2} satisfies the assumptions of the first statement of Theorem \ref{mainthm}. This will be used in the next step.}
\paragraph{\bf Step 1.2: Blowing-up all agents}
Fix $\varepsilon>0$ to be chosen later. Applying Step 1 of Theorem \ref{mainthm} to the control system \eqref{syst2}, we can steer it to $\tilde \Omega_\varepsilon = \{ (x,v)\ \mid\ v=0,\ \|\tilde F(x)\|<\varepsilon \}$ in finite time $T_0$. The crucial observation here is that \FR{$\tilde U'(r)\leq -\eta\frac{1}{1+r^2}$}, i.e., all forces are purely repulsive. We now study the configuration $(x(T_0),v(T_0))=(\bar x,0)\in \tilde \Omega_\varepsilon$. Since it is a configuration of $N$ agents in the plane, the convex closure of positions $(x_1,\ldots, x_N)$ is a polygon of $n\leq N$ vertices, in which at least one of the internal angles is smaller than $\frac{n-2}{n}\pi$, thus smaller than $\frac{N-2}{N}\pi$.
By relabelling indices, we assume that $x_N$ is one of those vertices. By a simple geometrical observation, all interaction forces point outwards of the polygon (see Figure \ref{f-angle1}).
%
\begin{figure}[h]
\centerline{\input{angle}}
\caption{The outer vector $z_N$.}\label{f-angle1}
\end{figure}
%
More precisely, consider $z_N$ to be the unit vector in the direction of the outer angle bisector and note that each component of the force $\tilde F_{Nj}=\frac{1}{N} \nabla \tilde W(\bar x_N-\bar x_j)$ satisfies
$$
\tilde F_{Nj}\cdot z_N\geq \frac1N |\tilde U' (|\bar x_N-\bar x_j|)|\cos(\tfrac{N-2}{2N} \pi).
$$
This in turn implies that
\begin{eqnarray}\label{e-TF}
\|\tilde F_N\|&\geq& \tilde F_N\cdot z_N \geq \tfrac1N \sup_{j\neq N} |\tilde U' (|\bar x_N-\bar x_j|)|\cos(\tfrac{N-2}{2N} \pi)\\
&\geq&\tfrac{1}N \eta \frac{1}{1+\inf_{j\neq N} |\bar x_N-\bar x_j|^2}\cos(\tfrac{N-2}{2N} \pi).\nonumber
\end{eqnarray}
\FR{Recall that the original potential $U(r)$ satisfies $\lim_{r\to+\infty}U'(r)=0$. Then, there exists a sufficiently large distance $\mathcal{L}>1$ such that $|U'(r)|<\eta$ for every $r>\mathcal{L}$. Take $\mathcal{L}$ satisfying this condition and define $\varepsilon=\frac1N \eta\frac1{1+\mathcal{L}^2} \cos(\tfrac{N-2}{2N} \pi)$.
The condition $(\bar x,0)\in \tilde \Omega_\varepsilon$ implies $\|\tilde F_N\|<\varepsilon$. This in turn implies that $|x_N-x_j|>\mathcal{L}$ for all $x\neq N$, due to \eqref{e-TF}}.
Since the distance between $x_N$ and the other agents is larger than $\mathcal{L}$, then the components $\tilde F_{Nj}$ and $\tilde F_{jN}$ of the forces are \FR{smaller than $\varepsilon$}. Hence, all configurations $((\bar x_1,\ldots, \bar x_{N-1},\bar x_N +\tau_N z_N),0)$ with $\tau_N>0$ belong to $\tilde \Omega_{\varepsilon}$. By quasi-static deformation, we can steer $\bar x_N$ arbitrarily far from the other $\bar x_i$'s along the direction $z_N$, by choosing $\tau_N$ sufficiently large.
{In this quasi-static deformation, the path of steady-states is $((\bar x(\tau)= (\bar x_1,\ldots, \bar x_{N-1},\bar x_N +\tau_N z_N), \bar v(\tau)=0)$. Each of them is a flock. For any given fixed $\tau\in[0,1]$, we linearize the control system \eqref{syst} at the corresponding flock: similarly as in Section \ref{sec_proof_step3}, we arrive at a linear system of the form $\delta\dot x_i(t) = \delta v_i(t)$, $\delta \dot v_i(t) = dF_i(\bar x_i(0)+\bar v(\tau)t).\delta x(t)+u_i(t)$, and then of the form \eqref{very_simple_control_system} by changing the control. We do not give more details since the procedure is the same as in Section \ref{sec_proof_step3}.
}
We next apply the same strategy to the remaining $N-1$ agents and we steer one of them away from all others, while keeping $x_N$ \FR{further than $\mathcal{L}$} due to the choice of $\tau_N$. We repeat the procedure to the remaining $N-2$ agents, while keeping both $x_{N-1},x_N$ \FR{further than $\mathcal{L}$}, and so forth. In finite time, we are able to steer all agents to a configuration $(\bar x,0)\in \tilde \Omega_\varepsilon$ with $|\bar x_i-\bar x_j|$ \FR{larger than $\mathcal{L}$}.
\subsubsection{Proof of Step 2} Since all agents are far one from each other, we have \FR{$|\tilde F_i(x)|<N\varepsilon<\eta$} for $i=1,\ldots,N$. \FR{This means that we can again change the control $w_i$ into $z_i=w_i-F_i(x)$, and that the constraint on the control satisfies $|z_i|\leq M-(\tilde M_F-2\eta)-\eta=\tilde M_{\alpha,\beta}+\eta$, due to \eqref{e-w-const}.}
We can now steer all agents to a circular equidistributed configuration of large radius $R$, again by quasi-static deformation, as follows.
Choose a point $x^*$ of the plane, that does not belong to any of the lines passing through $(\bar x_i,\bar x_j)$ and apply a coordinate translation to have $x^*=0$. {Consider the half-line starting at $0$ and passing through $\bar x_i$, and define $\tilde x_i$ as the point on the half-line at distance
R>\frac{\alpha}{\beta \varepsilon}
$ from 0}. We want to steer each particle $x_i(t)$ starting at $\bar x_i$ to such $\tilde x_i$. The crucial observation is that each pairwise distance $|x_i(t)-x_j(t)|$ needs to be kept larger \FR{than $\mathcal{L}$} to ensure that \FR{$|\tilde F_i(x)|<\eta$} along the motion.
Notice that each angle $\widehat{\bar x_i x^* \bar x_j}$ is nonzero for $i\neq j$ by the choice of $x^*=0$, hence there exists a minimal angle $\theta>0$. Consider now one of the indices $i$ realizing the maximal distance $|\bar x_i|$ (that we assume to be the index $1$) and move it along the quasi-static trajectory $x_1(\tau)=\bar x_1 +\tau(\tilde x_1-\bar x_1)$. Since $\bar x_1$ was chosen to realize the maximal distance, for each $j\neq 1$ the triangle with vertices $0,\bar x_j, x_1(\tau)$ has an internal angle $\alpha_j$ in $\bar x_j$ that is increasing with respect to time, hence the distance $|\bar x_j-x_1(\tau)|$ is increasing too (see Figure \ref{f-angle2}, left).
\begin{figure}[h]
\centerline{\input{angle2}}
\caption{Left: Moving $x_1(\tau)$ increases the distance. Right: the minimum distance when moving $x_2(\tau)$ is realized by the right triangle.}
\label{f-angle2}
\end{figure}
We now choose one of the indices $i=2,\ldots N$ realizing the maximal distance $|x^*-\bar x_i|$, that we assume to be the index $2$, and move it along the quasi-static trajectory $x_2(\tau)=x^*+\bar x_2 +\tau(\tilde x_2-\bar x_2)$
{(and corresponding velocity $\bar v(\tau)=0$).}
It is then clear that the distance $|\bar x_j-x_2(\tau)|$ is increasing for all $j=3,\ldots N$, due to the same observation as above. Instead, the distance $|\tilde x_1-x_2(\tau)|$ can eventually decrease, up to the minimum that is realized when the triangle with vertices $x^*, x_2(\tau), \tilde x_1$ is right in $x_2(\tau)$, see Figure \ref{f-angle2}, right. Such a minimal distance is thus larger than $R\sin(\theta)$, where $\theta$ is the minimal angle given above. By choosing $R>\FR{\mathcal{L}}/\sin(\theta)$, we are ensured that the minimal distance is greater than $\FR{\mathcal{L}}$.
Repeat the same construction for the indices $3$ to $N$ and hence steer all agents to a circle of radius $R$. By contruction, the pairwise distance is larger than $\FR{\mathcal{L}}$. Rearrange indices on the circle to have $\tilde x_i=R(\cos(\alpha_i),\sin(\alpha_i))$ with $0< \alpha_1<\alpha_2<\ldots \leq \alpha_N\leq2\pi$. Consider now the target equidistributed configuration with radius $R$: we simultaneously steer each $\tilde x_i$ to $\hat x_i=R(\cos(\tfrac{2i\pi}{N}),\sin(\tfrac{2i\pi}{N}))$ by using again a quasi-static deformation along the path
$$
x_i(\tau)=R(\cos((1-\tau)\alpha_i +\tau \tfrac{2i\pi}{N}),\sin((1-\tau)\alpha_i +\tau \tfrac{2i\pi}{N}))
$$
{(and corresponding velocity $\bar v(\tau)=0$).}
It is easy to verify that the simultaneous displacement along the circle ensures that the minimal distance $|x_i(\tau)-x_j(\tau)|$ is realized either at the beginning or at the end of the deformation, hence in all cases $|x_i(\tau)-x_j(\tau)|>\FR{\mathcal{L}}$.
\subsubsection{Proof of Step 3}
At the end of Step 3, we have steered the system to a circular equidistributed configuration $(\hat x,0)$ of large radius $R>\bar R$ in $\tilde \Omega_\varepsilon$. Since pairwise distances are arbitrarily large, we also have $(\hat x,0)\in\Omega_{\varepsilon'}$ for an arbitrarily small $\varepsilon'>0$, i.e., we can go back to the original system \eqref{syst} with the potential $W$. Since the system \eqref{syst} is locally controllable around $(\hat x,0)$, we can control it to a desired configuration, as follows.
\medskip
\paragraph{\bf Reaching a flock ring}
{We first steer the system to $(\hat x,\delta \bar v)$ with $\bar v$ the desired unitary velocity direction and $\delta>0$ sufficiently small: this can be achieved exactly by local controllability.} We then let the system evolve by choosing $u_i=F_i$, which ensures that all velocities satisfy $v_i(t)=v_j(t)$, hence $x_i(t)-x_j(t)$ keeps being constant, hence $F_i\simeq 0$ along the motion. This also ensures that all velocities $v_i$ converge to $\sqrt\frac\alpha\beta \bar v$, as in Step 3.1 in the proof of Theorem \ref{mainthm}.
We next reduce the flock radius. Note that each flock ring of radius $r$ can be realized as a trajectory of the system \eqref{syst}, provided that the control can be chosen as $u_i=F_i$. In the case of the flock ring, we have
$
|x_i(t)-x_j(t)|\geq |x_i(t)-x_{i+1}(t)|=2\sin(\tfrac{\pi}N) \bar R
$
and thus
$$
\|F_i\|\leq \sup_{r>2\sin(\tfrac{\pi}N) \bar R} |U'(r)| .
$$
The condition $\|u_i\|>\sup_{r>2\sin(\tfrac{\pi}N) \bar R} |U'(r)|$ ensures that each flock ring of radius $r\geq R$ can indeed be realized as a trajectory of the system. Hence, by a quasi-static deformation, we can steer the system from a flock of radius $r_1$ to a flock of radius $r_2$ whenever $r_1,r_2\geq \bar R$.
\medskip
\paragraph{\bf Reaching a mill ring}
We first steer the system to $(\hat x_i,\hat v_i)$ with $\hat v_i=\varepsilon x_i^\perp$ for some small $\varepsilon>0$, by local controllability around $(\hat x_i,0)\in\Omega_\varepsilon$. We choose
$
u_i=F_i-\frac{|v_i|^2}{R} \frac{x_i}{|x_i|}
$
to ensure that each agent undergoes the correct centripetal force and moves along the circle of radius $R$. Note that $\|u\|\leq \varepsilon+\frac{\alpha}{\beta R}\leq 2\varepsilon$ due to $R>\frac{\alpha}{\beta \varepsilon}$, hence the control is arbitrarily small, as in Step 3.3 in the proof of Theorem \ref{mainthm}.
Similarly to the flock ring, we next reduce the radius to a chosen $r\geq \bar R$. Note that each mill ring with such a radius is a trajectory of \eqref{syst}, provided that $u_i=F_i-\frac{|v_i|^2}{r} \frac{x_i}{|x_i|}$. By symmetry of the configuration, both $F_i$ and $-\frac{|v_i|^2}{r} \frac{x_i}{|x_i|}$ are radial forces. Hence
$$
\|u_i\|\leq \|F_i\|+\frac{\alpha}{\beta r}\leq \sup_{r>2\sin(\tfrac{\pi}N) \bar R} |U'(r)|+\frac{\alpha}{\beta \bar R}\leq M,
$$
which ensures that any mill of radius $r\geq \bar R$ can be reached. Using a quasi-static deformation, we can steer the system from any mill ring of radius $r_1$ to any mill ring of radius $r_2$, whenever $r_1,r_2\geq \bar R$.
\section{Concluding remarks}
In this work, we have studied the controlled transition to and between flocks and mills for second-order swarming systems. We have shown that, through a combination of stability properties of the dynamics, the choice of suitable Lyapunov functions and the use of quasi-static deformations, it is possible to construct feedback laws steering the dynamics towards flocking and milling regimes. We have provided an effective optimisation-based synthesis of instantaneous controls guiding the dynamics to different basins of attraction, where self-organization achieves the desired configuration.
The present work opens different research perspectives. So far, we have restricted our construction to instantaneous optimal feedback laws, which are fast to compute, but lack anticipation properties which are fundamental to truly benefit from the self-organization features of the dynamics. In particular, the use of a finite horizon optimal control, along with a choice of a sparse or total-variation control penalty, might induce a more parsimonious control action, acting over a reduced number of agents and time instances, to steer the swarm towards a self-organization regime before switching off.
Another challenge appears as the number of agents in the swarm increases, a natural scenario in agent-based models. Then, the computation of finite-horizon feedback laws even if theoretically possible, see Remark \ref{rem:mf}, becomes prohibitively expensive, and it is necessary to resort to open-loop controls embedded into a model predictive control scheme. An alternative to this problem is to consider a mean-field approximation of the dynamics, working with a density of agents $\rho=\rho(x,v,t)$ instead of the microscopic state of the swarm, and reformulating the control design at this level. However, the synthesis of optimal control laws for transitioning between mean-field flocks and mills is an open problem and may be the subject of future research.
\bibliographystyle{abbrv}
|
\section{Introduction}
Let $G$ be a finite, simple, undirected, and connected graph with vertex set $V(G)$ and edge set $E(G)$. For any two vertices $x, y\in V(G)$, let $d(x,y)$ denote the minimum number of edges connecting the vertices $x$ and $y$ in $G$. For $v\in V(G)$, the \emph{open neighborhood} of $v$ is $N(v)=\{u\in V(G): uv\in E(G)\}$. The \emph{degree} of a vertex $v\in V(G)$ is $|N(v)|$; a \emph{leaf} is a vertex of degree one, and a \emph{major vertex} is a vertex of degree at least three. The \emph{complement} of $G$, denoted by $\overline{G}$, is the graph whose vertex set is $V(G)$ and $xy\in E(\overline{G})$ if and only if $xy\not\in E(G)$ for any distinct $x,y\in V(G)$. We denote by $P_n$, $C_n$, $K_n$ and $K_{t, n-t}$, respectively, the path, the cycle, the complete graph and the complete bi-partite graph on $n$ vertices.
A vertex $z \in V(G)$ \emph{resolves} a pair of vertices $x$ and $y$ in $G$ if $d(x,z) \neq d(y,z)$. For two distinct vertices $x,y \in V(G)$, let $R_v\{x,y\}=\{z \in V(G): d(x,z) \neq d(y,z)\}$. A vertex subset $S \subseteq V(G)$ is a \emph{(vertex) resolving set} of $G$ if $|S\cap R_v\{x,y\}| \ge 1$ for every pair of distinct vertices $x,y \in V(G)$, and the \emph{metric dimension} $\dim(G)$ of $G$ is the minimum cardinality among all resolving sets of $G$. For an ordered set of distinct vertices $U=\{u_1, \ldots, u_k\} \subseteq V(G)$, the distance vector of a vertex $x\in V(G)$ with respect to $U$ is $\textnormal{code}_U(x)=(d(x, u_1), \ldots, d(x, u_k))$. Metric dimension, introduced by Slater~\cite{slater} and by Harary and Melter~\cite{harary}, has applications in robot navigation~\cite{tree2}, sonar~\cite{slater} and combinational optimization~\cite{sebo}, to name a few. It is noted in~\cite{NP} that determining the metric dimension of a general graph is an NP-hard problem.
For any vertex $v\in V(G)$ and any edge $e=xy\in E(G)$, let $d(e,v)=\min\{d(x,v), d(y,v)\}$. For any distinct edges $e_1, e_2 \in E(G)$, let $R_e\{e_1, e_2\}=\{v\in V(G): d(v,e_1) \neq d(v, e_2)\}$. A vertex subset $S \subseteq V(G)$ is an \emph{edge resolving set} of $G$ if $|S \cap R_e\{e_1, e_2\}| \ge 1$ for every pair of distinct edges $e_1, e_2\in E(G)$, and the \emph{edge (metric) dimension} $\textnormal{edim}(G)$ of $G$ is the minimum cardinality among all edge resolving sets of $G$. Kelenc et al.~\cite{edim} introduced and initiated the study of edge dimension, and it is stated in~\cite{edim} that determining the edge dimension of a general graph is an NP-complete problem. It is easy to see that, for any connected graph $G$ of order $n\ge3$, $1\le \textnormal{edim}(G)\le n-1$ (see~\cite{edim}); for characterization of graphs $G$ with $\textnormal{edim}(G)=n-1$ and $\textnormal{edim}(G)=n-2$, respectively, see~\cite{nina} and~\cite{geneson}. For an ordered set of distinct vertices $U=\{u_1, \ldots, u_k\} \subseteq V(G)$, the distance vector of an edge $e\in E(G)$ with respect to $U$ is $\textnormal{code}_U(e)=(d(e, u_1), \ldots, d(e, u_k))$.
The fractionalization of various graph parameters has been extensively studied (see~\cite{fractionalization}). Currie and Oellermann~\cite{oellermann} defined fractional metric dimension as the optimal solution to a linear programming problem by relaxing a condition of the integer programming problem for metric dimension. Arumugam and Mathew~\cite{fdim} officially studied the fractional metric dimension of graphs. For a function $g$ defined on $V(G)$ and for $U\subseteq V(G)$, let $g(U)=\sum_{s\in U} g(s)$. A real-valued function $g:V(G)\rightarrow[0,1]$ is a \emph{(vertex) resolving function} of $G$ if $g(R_v\{x,y\}) \ge 1$ for every pair of distinct vertices $x,y\in V(G)$, and the \emph{fractional metric dimension} $\dim_f(G)$ of $G$ is $\min\{g(V(G)): g \mbox{ is a resolving function of } G\}$. Notice that $\dim_f(G)$ reduces to $\dim(G)$ if the codomain of resolving functions is restricted to $\{0,1\}$.
Analogous to resolving function and fractional metric dimension, we introduce edge resolving function and fractional edge (metric) dimension as follows. A real-valued function $g: V(G) \rightarrow [0,1]$ is an \emph{edge resolving function} of $G$ if $g(R_e\{e_1, e_2\})\ge1$ for every pair of distinct edges $e_1, e_2\in E(G)$, and the \emph{fractional edge dimension} $\textnormal{edim}_f(G)$ of $G$ is $\min\{g(V(G)): g \mbox{ is an edge resolving function of }G\}$. Notice that $\textnormal{edim}_f(G)$ reduces to $\textnormal{edim}(G)$ if the codomain of edge resolving functions is restricted to $\{0,1\}$.
In this paper, we introduce and study the fractional edge dimension of graphs. We obtain some general results on the edge dimension and fractional edge dimension of graphs. We also determine the fractional edge dimension of some graph classes.
The paper is organized as follows. In Section~\ref{sec_general}, we observe that $1\le\textnormal{edim}_f(G)\le \frac{|V(G)|}{2}$ for any connected graph $G$ of order at least three. We characterize connected graphs $G$ satisfying $\textnormal{edim}_f(G)=1$, and we show that $\dim_f(G)=\frac{|V(G)|}{2}$ implies $\textnormal{edim}_f(G)=\frac{|V(G)|}{2}$, but not vice versa; we show that there exists a family of graphs $H$ with $\textnormal{edim}_f(H)=\frac{|V(H)|}{2}>\dim_f(H)$ such that $\frac{\textnormal{edim}_f(H)}{\dim_f(H)}$ can be arbitrarily large. We show that, for an edge resolving set $S$ of $G$, $\{\textnormal{code}_S(e): e\in E(G)\}$ does not uniquely determine $G$, i.e., there exist two non-isomorphic graphs on the same vertex set with the same edge metric coordinates with respect to the same edge resolving set. We show that there exist graphs $G$ and $H$ with $H \subset G$ such that both $\textnormal{edim}(H)-\textnormal{edim}(G)$ and $\textnormal{edim}_f(H)-\textnormal{edim}_f(G)$ can be arbitrarily large. We also consider the relation between planarity of a graph $G$ and $\textnormal{edim}(G)=2$: we show that $\textnormal{edim}(G)=2$ implies $G$ contains neither $K_5$ nor $K_{3,3}$ as a subgraph, while there exists a non-planar graph $H$ with $\textnormal{edim}(H)=2$. In Section~\ref{sec_graphs}, we determine $\textnormal{edim}_f(G)$ when $G$ is a tree, a cycle, the Petersen graph, a wheel graph, a complete multi-partite graph and a grid graph, respectively.
\section{General results on edge dimension and fractional edge dimension}\label{sec_general}
In this section, we obtain some general results on edge dimension and fractional edge dimension. We begin with some terminology and useful observations. Two vertices $u,w\in V(G)$ are called \emph{twins} if $N(u)-\{w\}=N(w)-\{u\}$; notice that a vertex is its own twin. Hernando et al.~\cite{Hernando} observed that the twin relation is an equivalence relation and that an equivalence class under it, called a \emph{twin equivalence class}, induces either a clique or an independent set. We note that, for distinct twins $x$ and $y$ of $G$, if $z\in N(x) \cap N(y)$, then $R_e\{zx, zy\}=\{x,y\}$.
\begin{observation}\label{obs_twin}
Let $x$ and $y$ be distinct members of the same twin equivalence class of a graph $G$.
\begin{itemize}
\item[(a)] \emph{\cite{Hernando}} For any resolving set $R$ of $G$, $R \cap\{x,y\} \neq\emptyset$.
\item[(b)] For any edge resolving set $S$ of $G$, $S \cap\{x,y\} \neq\emptyset$.
\item[(c)] \emph{\cite{yi}} For any resolving function $g$ of $G$, $g(x)+g(y)\ge1$.
\item[(d)] For any edge resolving function $h$ of $G$, $h(x)+h(y)\ge1$.
\end{itemize}
\end{observation}
\begin{observation}\label{obs_frac}
Let $G$ be a connected graph of order at least three. Then
\begin{itemize}
\item[(a)] \emph{\cite{fdim}} $\dim_f(G) \le \dim(G)$;
\item[(b)] $\textnormal{edim}_f(G) \le \textnormal{edim}(G)$.
\end{itemize}
\end{observation}
First, we show that there exist two non-isomorphic graphs on the same vertex set with the same edge metric coordinates with respect to the same edge resolving set. Seb\"{o} and Tannier~\cite{sebo} observed that, for a minimum resolving set $S$ of a graph $G$, the vectors $\{\textnormal{code}_S(v): v\in V(G)\}$ may not uniquely determine $G$ (see Figure~\ref{fig_unique}(a)). Similarly, we show that there exist two non-isomorphic graphs $H_1$ and $H_2$ with $V(H_1)=V(H_2)$ and $\{\textnormal{code}_S(e): e\in E(H_1)\}=\{\textnormal{code}_S(e'): e'\in E(H_2)\}$, where $S$ is a common minimum edge resolving set for $H_1$ and $H_2$; see Figure~\ref{fig_unique}(b).
\begin{figure}[ht]
\centering
\begin{tikzpicture}[scale=.7, transform shape]
\node [draw, fill=black, shape=circle, scale=.8] (a1) at (0,0) {};
\node [draw, shape=circle, scale=.8] (a2) at (0,1.5) {};
\node [draw, shape=circle, scale=.8] (a3) at (0,3) {};
\node [draw, fill=black, shape=circle, scale=.8] (a4) at (1.5,0) {};
\node [draw, shape=circle, scale=.8] (a5) at (1.5,1.5) {};
\node [draw, shape=circle, scale=.8] (a6) at (1.5,3) {};
\node [draw, fill=black, shape=circle, scale=.8] (a11) at (4,0) {};
\node [draw, shape=circle, scale=.8] (a22) at (4,1.5) {};
\node [draw, shape=circle, scale=.8] (a33) at (4,3) {};
\node [draw, fill=black, shape=circle, scale=.8] (a44) at (5.5,0) {};
\node [draw, shape=circle, scale=.8] (a55) at (5.5,1.5) {};
\node [draw, shape=circle, scale=.8] (a66) at (5.5,3) {};
\draw(a1)--(a2)--(a3);\draw(a4)--(a5)--(a6);\draw(a2)--(a5);\draw(a11)--(a22)--(a33)--(a66)--(a55)--(a44);\draw(a22)--(a55);
\node [scale=1] at (-0.6,3) {$(2,3)$};
\node [scale=1] at (-0.6,1.5) {$(1,2)$};
\node [scale=1] at (-0.6,0) {$(0,3)$};
\node [scale=1] at (2.1,3) {$(3,2)$};
\node [scale=1] at (2.1,1.5) {$(2,1)$};
\node [scale=1] at (2.1,0) {$(3,0)$};
\node [scale=1] at (3.4,3) {$(2,3)$};
\node [scale=1] at (3.4,1.5) {$(1,2)$};
\node [scale=1] at (3.4,0) {$(0,3)$};
\node [scale=1] at (6.1,3) {$(3,2)$};
\node [scale=1] at (6.1,1.5) {$(2,1)$};
\node [scale=1] at (6.1,0) {$(3,0)$};
\node [scale=1.3] at (0.75,-0.75) {$G_1$};
\node [scale=1.3] at (4.75,-0.75) {$G_2$};
\node [draw, fill=black, shape=circle, scale=.8] (b1) at (10,3) {};
\node [draw, shape=circle, scale=.8] (b2) at (9,2.2) {};
\node [draw, fill=black, shape=circle, scale=.8] (b3) at (9,0.9) {};
\node [draw, shape=circle, scale=.8] (b4) at (10,0) {};
\node [draw, shape=circle, scale=.8] (b5) at (11,0.9) {};
\node [draw, shape=circle, scale=.8] (b6) at (11,2.2) {};
\node [draw, fill=black, shape=circle, scale=.8] (b11) at (14.5,3) {};
\node [draw, shape=circle, scale=.8] (b22) at (13.5,2.2) {};
\node [draw, fill=black, shape=circle, scale=.8] (b33) at (13.5,0.9) {};
\node [draw, shape=circle, scale=.8] (b44) at (14.5,0) {};
\node [draw, shape=circle, scale=.8] (b55) at (15.5,0.9) {};
\node [draw, shape=circle, scale=.8] (b66) at (15.5,2.2) {};
\draw(b1)--(b2)--(b3)--(b4)--(b5)--(b6)--(b1);\draw(b2)--(b6);\draw(b11)--(b22)--(b33)--(b44)--(b55)--(b66)--(b11);\draw(b22)--(b55);
\node [scale=1] at (9.1,2.8) {$(0,1)$};
\node [scale=1] at (10.9,2.8) {$(0,2)$};
\node [scale=1] at (10,1.9) {$(1,1)$};
\node [scale=1] at (8.5,1.55) {$(1,0)$};
\node [scale=1] at (11.5,1.55) {$(1,2)$};
\node [scale=1] at (9,0.3) {$(2,0)$};
\node [scale=1] at (11,0.3) {$(2,1)$};
\node [scale=1] at (13.6,2.8) {$(0,1)$};
\node [scale=1] at (15.4,2.8) {$(0,2)$};
\node [scale=1] at (14.55,1.95) {$(1,1)$};
\node [scale=1] at (13,1.55) {$(1,0)$};
\node [scale=1] at (16,1.55) {$(1,2)$};
\node [scale=1] at (13.5,0.3) {$(2,0)$};
\node [scale=1] at (15.5,0.3) {$(2,1)$};
\node [scale=1.3] at (10,-0.75) {$H_1$};
\node [scale=1.3] at (14.5,-0.75) {$H_2$};
\node [scale=1.3] at (2.75,-1.5) {\textbf{(a)}};
\node [scale=1.3] at (12.25,-1.5) {\textbf{(b)}};
\end{tikzpicture}
\caption{\small{(a)~\cite{sebo} Two non-isomorphic graphs $G_1$ and $G_2$ with $V(G_1)=V(G_2)$ and $\{\textnormal{code}_S(v):v\in V(G_1)\}=\{\textnormal{code}_S(w):w\in V(G_2)\}$ on the common minimum resolving set $S$, comprised of the solid vertices, for $G_1$ and $G_2$; (b) Two non-isomorphic graphs $H_1$ and $H_2$ with $V(H_1)=V(H_2)$ and $\{\textnormal{code}_S(e): e\in E(H_1)\}=\{\textnormal{code}_S(e'):e'\in E(H_2)\}$ on the common minimum edge resolving set $S$, comprised of the solid vertices, for $H_1$ and $H_2$.}}\label{fig_unique}
\end{figure}
Second, we examine the relation between $\textnormal{edim}(G)=2$ and the planarity of $G$. We recall some terminology. A graph is \emph{planar} if it can be drawn in a plane without edge crossing. For two graphs $G$ and $H$, $H$ is called a \emph{minor} of $G$ if $H$ can be obtained from $G$ by vertex deletion, edge deletion, or edge contraction. We recall the following results.
\begin{theorem}\emph{\cite{wagner}}\label{graph_minor}
A graph $G$ is planar if and only if neither $K_5$ nor $K_{3,3}$ is a minor of $G$.
\end{theorem}
\begin{theorem}\emph{\cite{tree2}}\label{dim2_planarity}
\begin{itemize}
\item[(a)] A graph $G$ with $\dim(G)=2$ cannot have $K_5$ or $K_{3,3}$ as a subgraph.
\item[(b)] There exists a non-planar graph $G$ with $\dim(G)=2$.
\end{itemize}
\end{theorem}
Analogous to Theorem~\ref{dim2_planarity}, we show that $\textnormal{edim}(G)=2$ implies $G$ contains neither $K_5$ nor $K_{3,3}$ as a subgraph, while there exists a non-planar graph $H$ with $\textnormal{edim}(H)=2$.
\begin{theorem}
If $G$ is a graph with $\textnormal{edim}(G)=2$, then $G$ contains neither $K_5$ nor $K_{3,3}$ as a subgraph.
\end{theorem}
\begin{proof}
Let $S=\{w_1, w_2\} \subset V(G)$ be an edge resolving set of $G$ with $|S|=2$.
First, suppose $G$ contains a clique with 5 vertices $v_1, v_2, v_3, v_4, v_5$. By relabeling the vertices of $G$ if necessary, let $d(w_1, v_1)=\min\{d(w_1, v_i): 1 \le i \le 5\}$. Then $\textnormal{code}_{\{w_1\}}(v_1v_2)=\textnormal{code}_{\{w_1\}}(v_1v_3)=\textnormal{code}_{\{w_1\}}(v_1v_4)=\textnormal{code}_{\{w_1\}}(v_1v_5)$. Since $v_1,v_2,v_3,v_4,v_5$ are all adjacent, there exist two edges, say $v_1v_i$ and $v_1v_j$ for distinct $i,j\in\{2,3,4,5\}$, with $\textnormal{code}_S(v_1v_i)=\textnormal{code}_S(v_1v_j)$, contradicting the assumption that $S$ is an edge resolving set of $G$. So, if $\textnormal{edim}(G)=2$, then $G$ does not contain $K_5$ as a subgraph.
Second, suppose $G$ contains $K_{3,3}$ as a subgraph; let $V(K_{3,3})=V_1\cup V_2$ such that each vertex in $V_1=\{u_1, u_2, u_3\}$ is adjacent to each vertex in $V_2=\{v_1, v_2, v_3\}$. Note that, for each $i,j\in\{1,2,3\}$, $\textnormal{code}_{\{w_1\}}(u_iv_j) \in \{\alpha, \alpha+1\}$ and $\textnormal{code}_{\{w_2\}}(u_iv_j) \in \{\beta, \beta+1\}$ for some non-negative integers $\alpha$ and $\beta$; thus, there are only four possible values for $\textnormal{code}_S(e)$ for $e\in E(K_{3,3})$. Since $|E(K_{3,3})|=9$, there exist two distinct edges, say $e_1$ and $e_2$, with $\textnormal{code}_S(e_1)=\textnormal{code}_S(e_2)$, contradicting the assumption that $S$ is an edge resolving set of $G$. So, if $\textnormal{edim}(G)=2$, then $G$ does not contain $K_{3,3}$ as a subgraph.~\hfill
\end{proof}
\begin{theorem}
There exists a non-planar graph $G$ with $\textnormal{edim}(G)=2$.
\end{theorem}
\begin{proof}
Let $G$ be the graph given in Figure~\ref{fig_edim2_nonplanar}. Since $G$ contains $K_{3,3}$ as a minor, $G$ is not planar by Theorem~\ref{graph_minor}. Next, we show that $S=\{x_1, y_4\}$ forms an edge resolving set of $G$. Note the following: $\textnormal{code}_S(u_1x_1)=(0,2)$, $\textnormal{code}_S(x_1x_2)=(0,3)$, $\textnormal{code}_S(x_2v_1)=(1,4)$, $\textnormal{code}_S(u_1v_2)=(1,1)$, $\textnormal{code}_S(u_1v_3)=(1,2)$, $\textnormal{code}_S(u_2v_1)=(2,4)$, $\textnormal{code}_S(u_2y_1)=(3,3)$, $\textnormal{code}_S(y_1y_2)=(4,2)$, $\textnormal{code}_S(y_2y_3)=(4,1)$, $\textnormal{code}_S(y_3y_4)=(3,0)$, $\textnormal{code}_S(y_4v_2)=(2,0)$, $\textnormal{code}_S(u_2v_3)=(2,3)$, $\textnormal{code}_S(u_3z_1)=(3,2)$, $\textnormal{code}_S(z_1z_2)=(4,3)$, $\textnormal{code}_S(z_2z_3)=(3,4)$, $\textnormal{code}_S(z_3v_1)=(2,5)$, $\textnormal{code}_S(u_3v_2)=(2,1)$ and $\textnormal{code}_S(u_3v_3)=(2,2)$. So, $S$ is an edge resolving set of $G$ with $|S|=2$, and thus $\textnormal{edim}(G) \le 2$. Since a set consisting of one vertex fails to form an edge resolving set of $G$, $\textnormal{edim}(G)=2$.~\hfill
\end{proof}
\begin{figure}[ht]
\centering
\begin{tikzpicture}[scale=.7, transform shape]
\node [draw, shape=circle, scale=.8] (u1) at (0,6) {};
\node [draw, shape=circle, scale=.8] (u2) at (0,3) {};
\node [draw, shape=circle, scale=.8] (u3) at (0,0) {};
\node [draw, shape=circle, scale=.8] (v1) at (10,6) {};
\node [draw, shape=circle, scale=.8] (v2) at (10,3) {};
\node [draw, shape=circle, scale=.8] (v3) at (10,0) {};
\node [draw, fill=black, shape=circle, scale=.8] (u11) at (3.3,6) {};
\node [draw, shape=circle, scale=.8] (u12) at (6.6,6) {};
\node [draw, shape=circle, scale=.8] (u21) at (2,3) {};
\node [draw, shape=circle, scale=.8] (u22) at (4,3) {};
\node [draw, shape=circle, scale=.8] (u23) at (6,3) {};
\node [draw, fill=black, shape=circle, scale=.8] (u24) at (8,3) {};
\node [draw, shape=circle, scale=.8] (u31) at (1.3,0.8) {};
\node [draw, shape=circle, scale=.8] (u32) at (2.5,1.5) {};
\node [draw, shape=circle, scale=.8] (u33) at (7.5,4.5) {};
\draw(u1)--(u11)--(u12)--(v1)--(u2)--(v3);\draw(v2)--(u1)--(v3)--(u3)--(v2);\draw(u2)--(u21)--(u22)--(u23)--(u24)--(v2);\draw(u3)--(u31)--(u32)--(u33)--(v1);
\node [scale=1.2] at (-0.4,6) {$u_1$};
\node [scale=1.2] at (-0.4,3) {$u_2$};
\node [scale=1.2] at (-0.4,0) {$u_3$};
\node [scale=1.2] at (10.4,6) {$v_1$};
\node [scale=1.2] at (10.4,3) {$v_2$};
\node [scale=1.2] at (10.4,0) {$v_3$};
\node [scale=1.1] at (3.3,6.35) {$x_1$};
\node [scale=1.1] at (6.6,6.35) {$x_2$};
\node [scale=1.1] at (2,3.35) {$y_1$};
\node [scale=1.1] at (4,3.35) {$y_2$};
\node [scale=1.1] at (6,3.35) {$y_3$};
\node [scale=1.1] at (8,3.35) {$y_4$};
\node [scale=1.1] at (1.3, 1.15) {$z_1$};
\node [scale=1.1] at (2.5,1.85) {$z_2$};
\node [scale=1.1] at (7.6,4.85) {$z_3$};
\end{tikzpicture}
\caption{\small{A non-planar graph $G$ with $\textnormal{edim}(G)=2$, where $\{x_1, y_4\}$ is an edge resolving set of $G$}.}\label{fig_edim2_nonplanar}
\end{figure}
Third, we show that there exist graphs $G$ and $H$ with $H \subset G$ such that both $\textnormal{edim}(H)-\textnormal{edim}(G)$ and $\textnormal{edim}_f(H)-\textnormal{edim}_f(G)$ can be arbitrarily large; along the way, we show that $\dim_f(H)-\dim_f(G)$ can be arbitrarily large. We begin by recalling the characterization result of graphs $G$ satisfying $\dim_f(G)=\frac{|V(G)|}{2}$ that are used in proving our results.
\begin{theorem}\emph{\cite{fdim, frac_kang}}\label{fracdim_bounds}
Let $G$ be a connected graph of order $n\ge2$. Then $1 \le \dim_f(G)\le \frac{n}{2}$, and $\dim_f(G)=\frac{n}{2}$ if and only if there exists a bijection $\phi:V(G) \rightarrow V(G)$ such that $\phi(v)\neq v$ and $|R\{v, \phi(v)\}|=2$ for all $v\in V(G)$.
\end{theorem}
For an explicit characterization of graphs $G$ satisfying $\dim_f(G)=\frac{|V(G)|}{2}$, we recall the following construction from~\cite{fracdim2}. Let $\mathcal{K}=\{K_{s}: s \ge 2\}$ and $\overline{\mathcal{K}}=\{\overline{K}_t: t \ge 2\}$. Let $H[\mathcal{K} \cup \overline{\mathcal{K}}]$ be the family of graphs obtained from a connected graph $H$ by replacing each vertex $u_i \in V(H)$ by a graph $H_i \in \mathcal{K} \cup \overline{\mathcal{K}}$, and each vertex in $H_i$ is adjacent to each vertex in $H_j$ if and only if $u_iu_j \in E(H)$.
\begin{theorem}~\emph{\cite{fracdim2}}\label{n_2}
Let $G$ be a connected graph of order at least two. Then $\dim_f(G)=\frac{|V(G)|}{2}$ if and only if $G \in H[\mathcal{K} \cup \overline{\mathcal{K}}]$ for some connected graph $H$.
\end{theorem}
Now, we recall that metric dimension is not a monotone parameter on subgraph inclusion (see~\cite{lineG}). We recall the following stronger result from~\cite{broadcast}.
\begin{theorem}\emph{\cite{broadcast}}
There exist connected graphs $G$ and $H$ such that $H \subset G$ and $\frac{\dim(H)}{\dim(G)}$ can be arbitrarily large.
\end{theorem}
Following~\cite{broadcast}, for $m \ge 3$, let $H_m=K_{\frac{m(m+1)}{2}}$; let $V(H_m)$ be partitioned into $V_1, V_2, \ldots, V_m$ such that $V_i=\{w_{i,1}, w_{i,2}, \ldots, w_{i,i}\}$ with $|V_i|=i$, where $i \in \{1,2,\ldots,m\}$. Let $G_m$ be the graph obtained from $H_m$ and $m$ isolated vertices $u_1,u_2, \ldots, u_m$ such that, for each $i \in \{1,2,\ldots, m\}$, $u_i$ is joined by an edge to each vertex of $V_i \cup (\cup_{j=i+1}^{m}\{w_{j,i}\})$. See Figure~\ref{fig_subgraph} when $m=4$. Note that $H_m\subset G_m$, $\dim_f(H_m) =\frac{m(m+1)}{4}$ by Theorem~\ref{n_2}, and $\dim_f(G_m) \le m$ by Observation~\ref{obs_frac}(a) since $\{u_1, u_2, \ldots, u_m\}$ forms a resolving set of $G_m$ (see~\cite{broadcast}). So, $\frac{\dim_f(H_m)}{\dim_f(G_m)} \ge \frac{m+1}{4} \rightarrow \infty$ as $m\rightarrow \infty$. Thus, we have the following corollary.
\begin{figure}[ht]
\centering
\begin{tikzpicture}[scale=.7, transform shape]
\node [draw, shape=circle, scale=.8] (0) at (0, 0) {};
\node [draw, shape=circle, scale=.8] (1) at (1.5, 0) {};
\node [draw, shape=circle, scale=.8] (2) at (3, 0) {};
\node [draw, shape=circle, scale=.8] (3) at (4.5, 0) {};
\node [draw, shape=circle, scale=.8] (4) at (6, 0) {};
\node [draw, shape=circle, scale=.8] (5) at (7.5, 0) {};
\node [draw, shape=circle, scale=.8] (6) at (9, 0) {};
\node [draw, shape=circle, scale=.8] (7) at (10.5, 0) {};
\node [draw, shape=circle, scale=.8] (8) at (12, 0) {};
\node [draw, shape=circle, scale=.8] (9) at (13.5, 0) {};
\node [draw, shape=circle, scale=.8] (a) at (0, -3) {};
\node [draw, shape=circle, scale=.8] (b) at (2.25, -3) {};
\node [draw, shape=circle, scale=.8] (c) at (6, -3) {};
\node [draw, shape=circle, scale=.8] (d) at (11.25, -3) {};
\draw[thick, dotted](-0.5,-0.6) rectangle (0.5,0.7);
\draw[thick, dotted](1,-0.6) rectangle (3.5,0.7);
\draw[thick, dotted](4,-0.6) rectangle (8,0.7);
\draw[thick, dotted](8.5,-0.6) rectangle (14,0.7);
\node [scale=1] at (0.1,0.3) {$w_{1,1}$};
\node [scale=1] at (1.6,0.3) {$w_{2,1}$};
\node [scale=1] at (3.1,0.3) {$w_{2,2}$};
\node [scale=1] at (4.6,0.3) {$w_{3,1}$};
\node [scale=1] at (6.1,0.3) {$w_{3,2}$};
\node [scale=1] at (7.6,0.3) {$w_{3,3}$};
\node [scale=1] at (9.1,0.3) {$w_{4,1}$};
\node [scale=1] at (10.6,0.3) {$w_{4,2}$};
\node [scale=1] at (12.1,0.3) {$w_{4,3}$};
\node [scale=1] at (13.6,0.3) {$w_{4,4}$};
\node [scale=1.1] at (0,-3.4) {$u_{1}$};
\node [scale=1.1] at (2.25,-3.4) {$u_{2}$};
\node [scale=1.1] at (6,-3.4) {$u_{3}$};
\node [scale=1.1] at (11.25,-3.4) {$u_{4}$};
\node [scale=1.1] at (0,1.15) {$V_1$};
\node [scale=1.1] at (2.25,1.15) {$V_2$};
\node [scale=1.1] at (6,1.15) {$V_3$};
\node [scale=1.1] at (11.25,1.15) {$V_4$};
\node [scale=1.1] at (15,1.6) {$H_4=K_{10}$};
\node [scale=1.5] at (-3.5,-1) {$G_4$:};
\draw[thick,dashed] (6.75,0.3) ellipse (9.3cm and 1.7cm);
\draw(0)--(a)--(1);\draw(3)--(a)--(6);\draw(1)--(b)--(2);\draw(4)--(b)--(7);\draw(3)--(c)--(4);\draw(5)--(c)--(8);\draw(6)--(d)--(7);\draw(8)--(d)--(9);
\end{tikzpicture}
\caption{\small \cite{broadcast} Graphs $G_m$ and $H_m$ with $H_m \subset G_m$ such that $\frac{\dim(H_m)}{\dim(G_m)}$ can be arbitrarily large.}\label{fig_subgraph}
\end{figure}
\begin{corollary}
There exist connected graphs $G$ and $H$ such that $H \subset G$ and $\frac{\dim_f(H)}{\dim_f(G)}$ can be arbitrarily large.
\end{corollary}
Next, we show that edge dimension is not a monotone parameter on subgraph inclusion.
\begin{lemma}\label{edim_subgraph}
There exist connected graphs $G$ and $H$ such that $H \subset G$ and $\textnormal{edim}(G) <\textnormal{edim}(H)$.
\end{lemma}
\begin{proof}
Let $H=K_{4,2}$ and $G$ be the graphs in Figure~\ref{fig_subgraph_edim}; notice that $H \subset G$. Then $\textnormal{edim}(H)=\textnormal{edim}(K_{4,2})=4$ (see~\cite{edim}) and $\textnormal{edim}(G) \le 3$ since $\{a,b,c\}$ forms an edge resolving set for $G$; thus, $\textnormal{edim}(G) <\textnormal{edim}(H)$.
\end{proof}
\begin{figure}[ht]
\centering
\begin{tikzpicture}[scale=.7, transform shape]
\node [draw, shape=circle, scale=.8] (1) at (0.2,6) {};
\node [draw, shape=circle, scale=.8] (2) at (0.2,4) {};
\node [draw, shape=circle, scale=.8] (3) at (0.2,2) {};
\node [draw, shape=circle, scale=.8] (4) at (0.2,0) {};
\node [draw, shape=circle, scale=.8] (a) at (5, 4) {};
\node [draw, shape=circle, scale=.8] (b) at (5, 2) {};
\node [draw, fill=black, shape=circle, scale=.8] (0) at (-0.5, 1) {};
\node [draw, shape=circle, scale=.8] (00) at (-0.5, -1.5) {};
\node [draw, fill=black, shape=circle, scale=.8] (01) at (3.5, -1.5) {};
\node [draw, fill=black, shape=circle, scale=.8] (02) at (-2, 1) {};
\node [draw, shape=circle, scale=.8] (03) at (-1.5, 3) {};
\draw[thick, dotted](-0.1,-0.25) rectangle (5.3,6.3);
\node [scale=1.1] at (2.65,6.7) {$H=K_{4,2}$};
\node [scale=1.5] at (-4.5,2.5) {$G$:};
\node [scale=1.2] at (-2.3,1) {$a$};
\node [scale=1.2] at (-0.8,1) {$b$};
\node [scale=1.2] at (3.85,-1.5) {$c$};
\node [scale=1] at (-1.4,3.65) {$(1,3,2)$};
\node [scale=1] at (-2.4,2.2) {$(0,3,2)$};
\node [scale=1] at (-2,-0.3) {$(0,2,1)$};
\node [scale=1] at (1.35,-1.8) {$(1,2,0)$};
\node [scale=1] at (0.55,-0.75) {$(1,1,1)$};
\node [scale=1] at (4.55,-0.7) {$(2,2,0)$};
\node [scale=1] at (-0.75,1.6) {$(3,0,2)$};
\node [scale=1] at (-0.75,0.35) {$(2,0,2)$};
\node [scale=1] at (2.1,5.7) {$(3,2,2)$};
\node [scale=1] at (0.65,5) {$(3,2,1)$};
\node [scale=1] at (1,4.2) {$(2,2,2)$};
\node [scale=1] at (0.75,3.35) {$(2,2,1)$};
\node [scale=1] at (0.75,2.65) {$(3,1,2)$};
\node [scale=1] at (1,1.75) {$(3,1,1)$};
\node [scale=1] at (0.75,1.05) {$(2,1,2)$};
\node [scale=1] at (2,0.35) {$(2,1,1)$};
\draw(a)--(1)--(b);\draw(a)--(2)--(b);\draw(a)--(3)--(b);\draw(a)--(4)--(b);\draw(3)--(0)--(4);\draw(2)--(03)--(02)--(00)--(01)--(b);\draw(00)--(4);
\end{tikzpicture}
\caption{\small Graphs $G$ and $H$ such that $H\subset G$ and $\textnormal{edim}(G) <\textnormal{edim}(H)$, where the 3-vector next to each edge $e\in E(G)$ is $\textnormal{code}_{\{a,b,c\}}(e)$.}\label{fig_subgraph_edim}
\end{figure}
Based on the construction of graphs in Figure~\ref{fig_subgraph_edim}, we show the existence of graphs $G$ and $H$ such that $H \subset G$ and both $\textnormal{edim}(H)-\textnormal{edim}(G)$ and $\textnormal{edim}_f(H)-\textnormal{edim}_f(G)$ can be arbitrarily large.
\begin{proposition}
There exist graphs $G$ and $H$ such that $H \subset G$ and both $\textnormal{edim}(H)-\textnormal{edim}(G)$ and $\textnormal{edim}_f(H)-\textnormal{edim}_f(G)$ can be arbitrarily large.
\end{proposition}
\begin{proof}
Let $k\ge2$. Let $H_k$ be a graph with vertex set $V(H_k)=\cup_{i=1}^{3k}\{x_i, y_i\}$ and edge set $E(H_k)=\cup_{i=1}^{3k-1}\{x_ix_{i+1}, y_iy_{i+1}, x_iy_{i+1}, y_ix_{i+1}\}$. We note that $x_i$ and $y_i$ are twins in $H_k$, where $i\in\{1,2,\ldots, 3k\}$. Let $G_k$ be the graph obtained from the disjoint union of $H_k$, $K_1$, and $k$ copies of $P_4$ that is given by $a_i, b_i, c_i, d_i$ for $i\in\{1,2,\ldots, k\}$ as follows: (i) the vertex $z\in V(K_1)$ is adjacent to $x_1$ and $y_1$ in $G_k$; (ii) $y_{3j+1}$ is adjacent to $b_{j+1}$, $y_{3j+2}$ is adjacent to $a_{j+1}$, and $y_{3j+3}$ is adjacent to $d_{j+1}$ in $G_k$, where $j\in\{0,1, \ldots, k-1\}$. See Figure~\ref{fig_subgraph_edim_large} for graphs $G_3$ and $H_3$.
First, we show that $\textnormal{edim}(H_k)=3k=\textnormal{edim}_f(H_k)$. Note that $\textnormal{edim}(H_k)\ge 3k$ by Observation~\ref{obs_twin}(b) and $\textnormal{edim}_f(H_k)\ge 3k$ by Observation~\ref{obs_twin}(d). Since $\cup_{i=1}^{3k}\{x_i\}$ forms an edge resolving set of $H_k$, $\textnormal{edim}_f(H_k) \le \textnormal{edim}(H_k) \le 3k$ by observation~\ref{obs_frac}(b).
Second, we show that $\textnormal{edim}_f(G_k) \le\textnormal{edim}(G_k) \le 1+2k$. Since $\{z\} \cup(\cup_{i=1}^{k} \{a_i, c_i\})$ forms an edge resolving set of $G_k$, $\textnormal{edim}_f(G_k) \le\textnormal{edim}(G_k) \le 2k+1$ by Observation~\ref{obs_frac}(b).
Thus, $\textnormal{edim}(H_k)-\textnormal{edim}(G_k) \ge 3k-(1+2k)=k-1 \rightarrow \infty$ and $\textnormal{edim}_f(H_k)-\textnormal{edim}_f(G_k) \ge k-1 \rightarrow \infty$ as $k \rightarrow \infty$.~\hfill
\end{proof}
\begin{figure}[ht]
\centering
\begin{tikzpicture}[scale=.65, transform shape]
\node [draw, fill=black, shape=circle, scale=.8] (0) at (-1,1) {};
\node [draw, shape=circle, scale=.8] (11) at (0,2) {};
\node [draw, shape=circle, scale=.8] (12) at (0,0) {};
\node [draw, shape=circle, scale=.8] (21) at (2,2) {};
\node [draw, shape=circle, scale=.8] (22) at (2,0) {};
\node [draw, shape=circle, scale=.8] (31) at (4,2) {};
\node [draw, shape=circle, scale=.8] (32) at (4,0) {};
\node [draw, shape=circle, scale=.8] (a) at (2,-2) {};
\node [draw, fill=black, shape=circle, scale=.8] (a1) at (2,-1) {};
\node [draw, shape=circle, scale=.8] (a21) at (4,-0.8) {};
\node [draw, fill=black, shape=circle, scale=.8] (a22) at (4,-1.6) {};
\node [draw, shape=circle, scale=.8] (41) at (6,2) {};
\node [draw, shape=circle, scale=.8] (42) at (6,0) {};
\node [draw, shape=circle, scale=.8] (51) at (8,2) {};
\node [draw, shape=circle, scale=.8] (52) at (8,0) {};
\node [draw, shape=circle, scale=.8] (61) at (10,2) {};
\node [draw, shape=circle, scale=.8] (62) at (10,0) {};
\node [draw, shape=circle, scale=.8] (b) at (8,-2) {};
\node [draw, fill=black, shape=circle, scale=.8] (b1) at (8,-1) {};
\node [draw, shape=circle, scale=.8] (b21) at (10,-0.8) {};
\node [draw, fill=black, shape=circle, scale=.8] (b22) at (10,-1.6) {};
\node [draw, shape=circle, scale=.8] (71) at (12,2) {};
\node [draw, shape=circle, scale=.8] (72) at (12,0) {};
\node [draw, shape=circle, scale=.8] (81) at (14,2) {};
\node [draw, shape=circle, scale=.8] (82) at (14,0) {};
\node [draw, shape=circle, scale=.8] (91) at (16,2) {};
\node [draw, shape=circle, scale=.8] (92) at (16,0) {};
\node [draw, shape=circle, scale=.8] (c) at (14,-2) {};
\node [draw, fill=black, shape=circle, scale=.8] (c1) at (14,-1) {};
\node [draw, shape=circle, scale=.8] (c21) at (16,-0.8) {};
\node [draw, fill=black, shape=circle, scale=.8] (c22) at (16,-1.6) {};
\draw[thick, dotted](-0.3,-0.3) rectangle (16.3,2.3);
\node [scale=1.3] at (15,2.7) {$H_3$};
\node [scale=1.5] at (-2.5,1) {$G_3$:};
\draw(11)--(0)--(12);
\draw(21)--(11)--(22);\draw(31)--(21)--(32);\draw(41)--(31)--(42);\draw(51)--(41)--(52);\draw(61)--(51)--(62);\draw(71)--(61)--(72);\draw(81)--(71)--(82);\draw(91)--(81)--(92);
\draw(21)--(12)--(22);\draw(31)--(22)--(32);\draw(41)--(32)--(42);\draw(51)--(42)--(52);\draw(61)--(52)--(62);\draw(71)--(62)--(72);\draw(81)--(72)--(82);\draw(91)--(82)--(92);
\draw(12)--(a);\draw(22)--(a1)--(a)--(a22)--(a21)--(32);
\draw(42)--(b);\draw(52)--(b1)--(b)--(b22)--(b21)--(62);
\draw(72)--(c);\draw(82)--(c1)--(c)--(c22)--(c21)--(92);
\node [scale=1.3] at (-1.35,1) {$z$};
\node [scale=1.1] at (0,1.65) {$x_1$};
\node [scale=1.1] at (2,1.65) {$x_2$};
\node [scale=1.1] at (4,1.65) {$x_3$};
\node [scale=1.1] at (6,1.65) {$x_4$};
\node [scale=1.1] at (8,1.65) {$x_5$};
\node [scale=1.1] at (10,1.65) {$x_6$};
\node [scale=1.1] at (12,1.65) {$x_7$};
\node [scale=1.1] at (14,1.65) {$x_8$};
\node [scale=1.1] at (16,1.65) {$x_9$};
\node [scale=1.1] at (0,0.35) {$y_1$};
\node [scale=1.1] at (2,0.35) {$y_2$};
\node [scale=1.1] at (4,0.35) {$y_3$};
\node [scale=1.1] at (6,0.35) {$y_4$};
\node [scale=1.1] at (8,0.35) {$y_5$};
\node [scale=1.1] at (10,0.35) {$y_6$};
\node [scale=1.1] at (12,0.35) {$y_7$};
\node [scale=1.1] at (14,0.35) {$y_8$};
\node [scale=1.1] at (16,0.35) {$y_9$};
\node [scale=1.1] at (2.4,-1) {$a_1$};
\node [scale=1.1] at (2.05,-2.4) {$b_1$};
\node [scale=1.1] at (4.05,-1.95) {$c_1$};
\node [scale=1.1] at (4.4,-0.8) {$d_1$};
\node [scale=1.1] at (8.4,-1) {$a_2$};
\node [scale=1.1] at (8.05,-2.4) {$b_2$};
\node [scale=1.1] at (10.05,-1.95) {$c_2$};
\node [scale=1.1] at (10.4,-0.8) {$d_2$};
\node [scale=1.1] at (14.4,-1) {$a_3$};
\node [scale=1.1] at (14.05,-2.4) {$b_3$};
\node [scale=1.1] at (16.05,-1.95) {$c_3$};
\node [scale=1.1] at (16.4,-0.8) {$d_3$};
\end{tikzpicture}
\caption{\small Graphs $G_k$ and $H_k$ such that $H_k\subset G_k$ and both $\textnormal{edim}(H_k)-\textnormal{edim}(G_k)$ and $\textnormal{edim}_f(H_k)-\textnormal{edim}_f(G_k)$ can be arbitrarily large, where $k\ge2$.}\label{fig_subgraph_edim_large}
\end{figure}
\begin{question}
Are there connected graphs $G$ and $H$ with $H \subset G$ such that both $\frac{\textnormal{edim}(H)}{\textnormal{edim}(G)}$ and $\frac{\textnormal{edim}_f(H)}{\textnormal{edim}_f(G)}$ can be arbitrarily large?
\end{question}
We conclude this section with some general results on fractional edge dimension. For any connected graph $G$ of order at least three, it is easy to see that $1\le \textnormal{edim}_f(G) \le \frac{|V(G)|}{2}$. We characterize graphs $G$ satisfying $\textnormal{edim}_f(G)=1$. We show that $\dim_f(G)=\frac{|V(G)|}{2}$ implies $\textnormal{edim}_f(G)=\frac{|V(G)|}{2}$, while there exist graphs $G$ with $\textnormal{edim}_f(G)=\frac{|V(G)|}{2}>\dim_f(G)$; in fact, there exists a family of graphs $G$ with $\textnormal{edim}_f(G)=\frac{|V(G)|}{2}$ such that $\frac{\textnormal{edim}_f(G)}{\dim_f(G)}$ can be arbitrarily large.
\begin{proposition}\label{edim_frac_bounds}
If $G$ is a connected graph of order $n\ge 3$, then $1\le \textnormal{edim}_f(G)\le \frac{n}{2}$.
\end{proposition}
\begin{proof}
Let $G$ be a connected graph of order $n\ge 3$. The lower bound follows from the definition of fractional edge dimension. Next, we prove the upper bound. Let $g: V(G) \rightarrow [0,1]$ be a function defined by $g(v)=\frac{1}{2}$ for each $v\in V(G)$; then $g(V(G))=\frac{n}{2}$. It suffices to show that $g$ is an edge resolving function of $G$. Let $e_1=ab$ and $e_2=cd$ be any distinct edges in $G$. If $\{a,b\} \cap \{c,d\} =\emptyset$, then $R_e\{e_1, e_2\} \supseteq \{a,b,c,d\}$. If $\{a,b\} \cap \{c,d\} \neq \emptyset$, say $a=c$, by relabeling the vertices of $G$ if necessary, then $b\neq d$ and $R_e\{e_1, e_2\} \supseteq \{b,d\}$. In each case, $|R_e\{e_1, e_2\}|\ge 2$ and $g(R_e\{e_1, e_2\}) \ge 2 (\frac{1}{2})=1$. So, $g$ is an edge resolving function of $G$, and thus $\textnormal{edim}_f(G) \le g(V(G))=\frac{n}{2}$.~\hfill
\end{proof}
Next, we characterize graphs achieving the lower bound of Proposition~\ref{edim_frac_bounds}. We recall the following result.
\begin{proposition}\emph{\cite{frac_sdim}}\label{dim_frac_1}
For any connected graph $G$ of order $n \ge 2$, $\dim_f(G)=1$ if and only if $G=P_n$.
\end{proposition}
In proving Proposition~\ref{edim_frac_1}, we use the same technique used for Proposition~\ref{dim_frac_1} in~\cite{frac_sdim} by adjusting to edge resolving function.
\begin{proposition}\label{edim_frac_1}
For any connected graph $G$ of order $n \ge 3$, $\textnormal{edim}_f(G)=1$ if and only if $G=P_n$.
\end{proposition}
\begin{proof}
$(\Leftarrow)$ Let $P_n$ be given by $u_1, u_2, \ldots, u_n$, where $n\ge 3$, and let $g: V(P_n) \rightarrow [0,1]$ be a function defined by $g(u_1)=\frac{1}{2}=g(u_n)$ and $g(v)=0$ for each $v\in V(P_n)-\{u_1, u_n\}$. For any distinct edges $e_1, e_2 \in E(P_n)$, $R_e\{e_1, e_2\} \supseteq \{u_1, u_n\}$; thus, $g(R_e\{e_1, e_2\}) \ge g(u_1)+g(u_n)=1$. So, $g$ is an edge resolving function of $P_n$ with $g(V(P_n))=1$, and thus $\textnormal{edim}_f(P_n) \le 1$. By Proposition~\ref{edim_frac_bounds}, $\textnormal{edim}_f(P_n)=1$.
($\Rightarrow$) Let $\mathcal{C}=\displaystyle\bigcap_{e_1, e_2 \in E(G)} R_e\{e_1, e_2\}$, where the intersection is taken over all pairs of distinct edges of $G$; notice that $\mathcal{C} \neq \emptyset$ if and only if $G=P_n$ for some $n$. Let $\textnormal{edim}_f(G)=1$. Assume, to the contrary, that $G\neq P_n$ for any $n$. Let $e_1, e_2, e_3,e_4\in E(G)$ such that $e_1 \neq e_2$ and $e_3\neq e_4$. Let $A=R_e\{e_1, e_2\}$, $B=R_e\{e_3, e_4\}$, $A'=A-(A\cap B)$, $B'=B-(A\cap B)$ and let $g:V(G)\rightarrow [0,1]$ be an edge resolving function of $G$ with $g(V(G))=1$. Since $g$ is an edge resolving function of $G$, we have the following:
\begin{equation}\label{eq1}
g(A')+g(A\cap B) \ge 1 \mbox{ and } g(B')+g(A\cap B)\ge 1 \tag{$\star$}.
\end{equation}
By the assumption that $\textnormal{edim}_f(G)=1$, we have
\begin{equation}\label{eq2}
g(A')+g(A\cap B)+g(B')=1 \tag{$\dagger$}.
\end{equation}
Now, from~(\ref{eq1}) and~(\ref{eq2}), $g(A')=0=g(B')$. Since $e_1, e_2, e_3, e_4$ are arbitrary, $g$ is zero except on $\mathcal{C}$. Since $G\neq P_n$, $\mathcal{C}=\emptyset$ and $g(V(G))=0$; this contradicts the assumption that $g(V(G))=1$.~\hfill
\end{proof}
Next, we show that, for any connected graph $G$ of order at least three, $\dim_f(G)=\frac{|V(G)|}{2}$ implies $\textnormal{edim}_f(G)=\frac{|V(G)|}{2}$, but not vice versa.
\begin{theorem}\label{edim_frac_n2}
Let $G$ be a connected graph of order $n\ge 3$.
\begin{itemize}
\item[(a)] If $\dim_f(G)=\frac{n}{2}$, then $\textnormal{edim}_f(G)=\frac{n}{2}$.
\item[(b)] There exists a graph $G$ satisfying $\textnormal{edim}_f(G)=\frac{n}{2}>\dim_f(G)$. Moreover, there exists a family of graphs $G$ with $\textnormal{edim}_f(G)=\frac{n}{2}>\dim_f(G)$ such that $\frac{\textnormal{edim}_f(G)}{\dim_f(G)}$ can be arbitrarily large.
\end{itemize}
\end{theorem}
\begin{proof}
Let $G$ be a connected graph of order $n\ge 3$.
\medskip
(a) Suppose $\dim_f(G)=\frac{n}{2}$. By Theorem~\ref{n_2}, $G\in H[\mathcal{K} \cup \overline{\mathcal{K}}]$ for some connected graph $H$. Let $g: V(G)\rightarrow[0,1]$ be any edge resolving function of $G$; we show that $g(V(G))\ge \frac{n}{2}$. First, let $G=K_n$ with $V(G)=\{u_1, u_2, \ldots, u_n\}$. By Observation~\ref{obs_twin}(d), $g(u_i)+g(u_j)\ge1$ for any distinct $i,j\in\{1,2,\ldots, n\}$. By summing over the ${n \choose 2}$ inequalities, we have $(n-1) \sum_{i=1}^{n} g(u_i) \ge {n \choose 2}$, and thus $g(V(G))=\sum_{i=1}^{n}g(u_i)\ge \frac{n}{2}$. Second, let $G\neq K_n$, and let $\mathcal{C}=\{Q_1, Q_2, \ldots, Q_t\}$ be the collection of twin equivalence classes of $G$, where $t\ge2$; then, by Theorem~\ref{n_2}, $n=\sum_{i=1}^{t}|Q_i|$ and $|Q_i|\ge 2$ for each $i\in\{1,2,\ldots, t\}$. Again, by Observation~\ref{obs_twin}(d), $g(Q_i) \ge \frac{|Q_i|}{2}$ for each $i\in \{1,2, \ldots, t\}$. By summing over the $t$ inequalities, we have $g(V(G))=\sum_{i=1}^{t}g(Q_i)\ge \sum_{i=1}^{t} \frac{|Q_i|}{2}=\frac{n}{2}$. So, in each case, $g(V(G))\ge \frac{n}{2}$, and thus $\textnormal{edim}_f(G)\ge \frac{n}{2}$. Since $\textnormal{edim}_f(G)\le \frac{n}{2}$ by Proposition~\ref{edim_frac_bounds}, $\textnormal{edim}_f(G)=\frac{n}{2}$.
\medskip
(b) If $G=K_{a_1, a_2, \ldots, a_k}$ is a complete $k$-partite graph of order $n=\sum_{i=1}^{k}a_i\ge5$ such that $k\ge 3$ and $a_i=1$ for exactly one $i\in\{1,2,\ldots, k\}$, then $\textnormal{edim}_f(G)=\frac{n}{2}>\frac{n-1}{2}=\dim_f(G)$ (see Theorem~\ref{dim_frac_graph}(e) and Proposition~\ref{edim_frac_kpartite}).
\medskip
Next, we show the existence of graphs $G$ such that $\textnormal{edim}_f(G)=\frac{n}{2}$ and $\frac{\textnormal{edim}_f(G)}{\dim_f(G)}$ can be arbitrarily large. We recall the family of graphs $G$ constructed in~\cite{nina} to show that $\frac{\textnormal{edim}(G)}{\dim(G)}$ can be arbitrarily large. For each integer $k\ge 3$ and the set $X=\{0,1,\ldots,k-1\}$, let $G_k$ be a graph of order $k+2^k$ such that $V(G_k)=A \cup B$, where $|A|=k$ and $|B|=2^k$. Let $A=\{a_0, a_1, \ldots, a_{k-1}\}$ and $B=\{b_S:S \subseteq X\}$. The edge set is specified as follows: (i) each of the sets $A$ and $B$ induces a clique in $G_k$; (ii) indexing the elements of $B$ by subsets of $X$, let $b_S\in B$ be adjacent to $a_i\in A$ if $i\in S$ for $S \subseteq X$; (iii) there are no other edges. See Figure~\ref{fig_edim_big} for the graph $G_3$. Let $k\ge 3$ be an integer.
First, we show that $\dim_f(G_k) =k$. Let $g:V(G_k) \rightarrow [0,1]$ be any resolving function of $G_k$. Note that, for each $i\in\{0,1,\ldots, k-1\}$, $R_v\{b_{\{i\}}, b_{\{i, i+1\}}\}=\{b_{\{i\}}, b_{\{i, i+1\}}, a_{i+1}\}$ and $g(b_{\{i\}})+g(b_{\{i,i+1\}})+g(a_{i+1})\ge 1$, where the subscript is taken modulo $k$. By summing over the $k$ inequalities, we have $g(A)+\sum_{i=0}^{k-1}(g(b_{\{i\}})+g(b_{\{i, i+1\}}))\ge k$, where the subscript is taken modulo $k$. So, $g(V(G_k)) \ge k$, and hence $\dim_f(G_k)\ge k$. On the other hand, $\dim_f(G_k) \le \dim(G_k)=k$ by Observation~\ref{obs_frac}(a) and the fact that $\dim(G_k)=k$ as shown in~\cite{nina}.
Second, we show that $\textnormal{edim}_f(G_k)=\frac{k+2^k}{2}$. Since $\textnormal{edim}_f(G_k) \le \frac{k+2^k}{2}=\frac{|V(G_k)|}{2}$ by Poposition~\ref{edim_frac_bounds}, it suffices to show that $\textnormal{edim}_f(G_k) \ge \frac{k+2^k}{2}$. Let $h:V(G_k) \rightarrow [0,1]$ be any edge resolving function of $G_k$. Note that, for each $i\in\{0,1,\ldots, k-1\}$, $R_e\{b_{\{i, i+1\}}a_i, b_{\{i, i+1\}}a_{i+1}\}=\{a_i, a_{i+1}\}$ and $h(a_i)+h(a_{i+1})\ge 1$, where the subscript is taken modulo $k$. By summing over the $k$ inequalities, we have $2h(A)\ge k$, i.e., $h(A) \ge \frac{k}{2}$ ($\blacktriangle$). Since $R_e\{a_0b_{\{0\}}, a_0b_X\}=\{b_{\{0\}}, b_X\}$, we have $h(b_{\{0\}})+h(b_X)\ge 1$ ($\blacktriangleleft$). We also note that, for any distinct $Y,Z \subsetneq X$ with $Y\neq\{0\}$ and $Z \neq\{0\}$, $R_e\{b_Xb_Y, b_X b_Z\}=\{b_Y, b_Z\}$ and $h(b_Y)+h(b_Z)\ge 1$. By summing over the ${2^k-2 \choose 2}$ inequalities, we have $(2^k-3)[(\sum_{S\subseteq X} h(b_S))-(h(b_{\{0\}})+h(b_X))] \ge {2^k-2 \choose 2}$, i.e., $(\sum_{S\subseteq X} h(b_S))-(h(b_{\{0\}})+h(b_X)) \ge \frac{2^{k}-2}{2}$ ($\blacktriangleright$). By summing over the three inequalities ($\blacktriangle$), ($\blacktriangleleft$) and ($\blacktriangleright$), we have $h(V(G_k))=h(A)+h(B) \ge\frac{k}{2}+\frac{2^k}{2}$; thus, $\textnormal{edim}_f(G_k)\ge \frac{k+2^k}{2}$.
Therefore, $\textnormal{edim}_f(G_k)=\frac{|V(G_k)|}{2}$ and $\frac{\textnormal{edim}_f(G_k)}{\dim_f(G_k)}=\frac{k+2^k}{2k} \rightarrow \infty$ as $k \rightarrow \infty$.~\hfill
\end{proof}
\begin{figure}[ht]
\centering
\begin{tikzpicture}[scale=.67, transform shape]
\node [draw, shape=circle, scale=.8] (a1) at (-1, 5) {};
\node [draw, shape=circle, scale=.8] (a2) at (0, 3) {};
\node [draw, shape=circle, scale=.8] (a3) at (-1, 1) {};
\node [draw, shape=circle, scale=.8] (b1) at (4, 6) {};
\node [draw, shape=circle, scale=.8] (b2) at (4, 5) {};
\node [draw, shape=circle, scale=.8] (b3) at (4, 4) {};
\node [draw, shape=circle, scale=.8] (b4) at (4, 3) {};
\node [draw, shape=circle, scale=.8] (b5) at (4, 2) {};
\node [draw, shape=circle, scale=.8] (b6) at (4, 1) {};
\node [draw, shape=circle, scale=.8] (b7) at (4, 0) {};
\node [draw, shape=circle, scale=.8] (c) at (8.3, 3) {};
\node [scale=1.3] at (-1.4,5) {$a_0$};
\node [scale=1.3] at (-0.4,3) {$a_1$};
\node [scale=1.3] at (-1.4,1) {$a_2$};
\node [scale=1.3] at (4.3, 6.3) {$b_{\{0\}}$};
\node [scale=1.3] at (4.3, 5.3) {$b_{\{1\}}$};
\node [scale=1.3] at (4.3,4.3) {$b_{\{2\}}$};
\node [scale=1.3] at (4.4, 3.4) {$b_{\{0,1\}}$};
\node [scale=1.3] at (4.3, 1.6) {$b_{\{0,2\}}$};
\node [scale=1.3] at (4.3,0.6) {$b_{\{1,2\}}$};
\node [scale=1.3] at (4.4,-0.4) {$b_{\{0,1,2\}}$};
\node [scale=1.3] at (8.3,3.45) {$b_{\emptyset}$};
\node [scale=1.3] at (9.2,3) {$K_8$};
\draw(a1)--(a2)--(a3)--(a1);
\draw(b1)--(c);\draw(b2)--(c);\draw(b3)--(c);\draw(b4)--(c);\draw(b5)--(c);\draw(b6)--(c);\draw(b7)--(c);
\draw(b1)--(a1)--(b4);\draw(b5)--(a1)--(b7);\draw(b2)--(a2);\draw(a2)--(b4);\draw(b6)--(a2)--(b7);\draw(b3)--(a3)--(b5);\draw(b6)--(a3)--(b7);
\draw[thick, dashed] (3.5,-1) rectangle (8.7,7);
\end{tikzpicture}
\caption{\small The graph $G_3$.}\label{fig_edim_big}
\end{figure}
\begin{question}
Is there a connected graph $G$ such that $\frac{\dim_f(G)}{\textnormal{edim}_f(G)}$ can be arbitrarily large?
\end{question}
\begin{question}
Can we characterize connected graphs $G$ for which $\textnormal{edim}_f(G)=\frac{|V(G)|}{2}$ hold?
\end{question}
\section{The fractional edge dimension of some graph classes}\label{sec_graphs}
In this section, we determine $\textnormal{edim}_f(G)$ when $G$ is a tree, a cycle, the Petersen graph, a wheel graph, a complete multi-partite graph and a grid graph (also known as the Cartesian product of two paths), respectively.
We recall some terminology. Fix a tree $T$. A leaf $\ell$ is called a \emph{terminal vertex} of a major vertex $v$ if $d(\ell,v) < d(\ell,w)$ for every other major vertex $w$ in $T$. The terminal degree, $ter(v)$, of a major vertex $v$ is the number of terminal vertices of $v$ in $T$, and an \emph{exterior major vertex} is a major vertex that has positive terminal degree. Let $M(T)$ be the set of exterior major vertices of $T$. Let $M_1(T)=\{w\in M(T): ter(w)=1\}$ and $M_2(T)=\{w\in M(T): ter(w) \ge 2\}$; then $M(T)=M_1(T) \cup M_2(T)$. Let $\sigma(T)$ denote the number of leaves of $T$, $ex(T)$ the number of exterior major vertices of $T$, and let $ex_1(T)=|M_1(T)|$. For each $v\in M(T)$, let $T_v$ be the subtree of $T$ induced by $v$ and all vertices belonging to the paths joining $v$ with its terminal vertices, and let $L(v)$ be the set of terminal vertices of $v$ in $T$.
First, we recall the fractional metric dimension of some graph classes.
\begin{theorem}\label{dim_frac_graph}
\begin{itemize}
\item[(a)] \emph{\cite{yi}} For any non-trivial tree $T$, $\dim_f (T)=\frac{1}{2}(\sigma(T)-ex_1(T))$.
\item[(b)] \emph{\cite{fdim}} For $n\ge 3$, $\dim_f(C_n)=\left\{
\begin{array}{ll}
\frac{n}{n-1} & \mbox{ if $n$ is odd},\\
\frac{n}{n-2} & \mbox{ if $n$ is even}.
\end{array}\right.$
\item[(c)] \emph{\cite{fdim}} For the Petersen graph $\mathcal{P}$, $\dim_f(\mathcal{P})=\frac{5}{3}$.
\item[(d)] \emph{\cite{fdim}} For the wheel graph $W_n=K_1+C_{n-1}$ of order $n\ge 4$, $\dim_f(W_n)=\left\{
\begin{array}{ll}
2 & \mbox{ if } n\in\{4,5\},\\
\frac{3}{2} & \mbox{ if } n=6,\\
\frac{n-1}{4} & \mbox{ if } n \ge 7.
\end{array}\right.$
\item[(e)] \emph{\cite{yi}} For $k\ge2$, let $G=K_{a_1, a_2, \ldots, a_k}$ be a complete $k$-partite graph of order $n=\sum_{i=1}^{k}a_i$, and let $s$ be the number of partite sets of $G$ consisting of exactly one element. Then
\begin{equation*}
\dim_f(G)=\left\{
\begin{array}{ll}
\frac{n-1}{2} & \mbox{ if } s=1,\\
\frac{n}{2} & \mbox{ otherwise}.
\end{array}\right.
\end{equation*}
\item[(f)] \emph{\cite{fdim}} For $s,t\ge 2$, $\dim_f(P_s \square P_t)=2$, where $P_s \square P_t$ denotes the Cartesian product of $P_s$ and $P_t$.
\end{itemize}
\end{theorem}
Next, we determine $\textnormal{edim}_f(T)$ for a tree $T$; we show that $\textnormal{edim}_f(T)=\dim_f(T)$. We begin with the following useful lemma, which can be obtained from Lemma 18 of~\cite{frac_kdim} by adjusting the statement and its proof for two distinct edges and edge resolving sets.
\begin{lemma}\label{lemma_tree2}
Let $T$ be a tree with $ex(T) \ge 2$. For $w\in M_2(T)$, let $e_1\in E(T_w)$ and $e_2\in E(T)-E(T_w)$. Then either $R_e\{e_1,e_2\} \supseteq V(T_w)$ or $R_e\{e_1,e_2\} \supseteq V(T_{w'})$ for some $w'\in M_2(T)-\{w\}$.
\end{lemma}
\begin{proposition}\label{edim_frac_tree}
For any tree $T$ of order at least three, $\textnormal{edim}_f(T)=\frac{1}{2}(\sigma(T)-ex_1(T))$.
\end{proposition}
\begin{proof}
Let $T$ be a tree of order at least three. If $ex(T)=0$, then $T$ is a path and $\textnormal{edim}_f(T)=1=\frac{1}{2}(\sigma(T)-ex_1(T))$ by Proposition~\ref{edim_frac_1}. So, suppose $ex(T)\ge 1$; then $M_2(T) \neq \emptyset$.
\medskip
First, we show that $\textnormal{edim}_f(T) \ge \frac{1}{2}(\sigma(T)-ex_1(T))$. Let $g:V(T)\rightarrow [0,1]$ be any edge resolving function of $T$. Fix $v\in M_2(T)$ with $ter(v)=a \ge2$. Let $N(v)=\{s_1, s_2, \ldots, s_a\}$ and $L(v)=\{\ell_1, \ell_2, \ldots, \ell_a\}$ such that $s_i$ lies on the $v-\ell_i$ path, where $i\in\{1,2,\ldots, a\}$. For each $i\in\{1,2,\ldots, a\}$, let $P^i$ denote the $s_i-\ell_i$ path. Note that, for any distinct $i,j\in\{1,2,\ldots, a\}$, $R_e\{vs_i, vs_j\} =V(P^i) \cup V(P^j)$; thus, $g(V(P^i))+g(V(P^j)) \ge 1$. By summing over the ${a \choose 2}$ inequalities, we have $(a-1) \sum_{i=1}^{a}g(V(P^i))\ge {a \choose 2}$, which implies $g(V(T_v)) \ge \sum_{i=1}^{a}g(V(P^i)) \ge \frac{a}{2}$. So,
$$g(V(T)) \ge \!\!\!\sum_{v\in M_2(T)} \!\!\!g(V(T_v)) \ge\! \!\!\sum_{v\in M_2(T)} \!\!\!\frac{ter(v)}{2}=\frac{1}{2}\left(\sum_{v\in M(T)} \!\!\!ter(v)-\!\!\!\!\!\sum_{w\in M_1(T)} \!\!\!ter(w)\right)=\frac{1}{2}(\sigma(T)-ex_1(T)),$$
and thus $\textnormal{edim}_f(T) \ge \frac{1}{2}(\sigma(T)-ex_1(T))$.
\medskip
Next, we show that $\textnormal{edim}_f(T) \le \frac{1}{2}(\sigma(T)-ex_1(T))$. For $x\in V(T)$, let $h:V(T) \rightarrow [0,1]$ be a function defined by
\begin{equation*}
h(x)=\left\{
\begin{array}{ll}
\frac{1}{2} & \mbox{ if $x$ is a terminal vertex of an exterior major vertex $w\in M_2(T)$},\\
0 & \mbox{ otherwise.}
\end{array}\right.
\end{equation*}
Notice that $h(V(T))=\frac{1}{2}(\sigma(T)-ex_1(T))$. It suffices to show that $h$ is an edge resolving function of $T$. Let $e_1, e_2 \in E(T)$ with $e_1\neq e_2$. We consider three cases: (1) $e_1,e_2\in E(T_w)$ for some $w\in M_2(T)$; (2) $e_1\in E(T_w)$ and $e_2\not\in E(T_w)$ for some $w\in M_2(T)$; (3) $e_1,e_2\in E(T)-\cup_{w\in M_2(T)}E(T_w)$. In case~(1), there exist distinct terminal vertices, say $\ell$ and $\ell'$, of $w$ such that both $e_1$ and $e_2$ lie on the $\ell-\ell'$ path in $T$; thus, $R_e\{e_1, e_2\} \supseteq \{\ell, \ell'\}$ and $h(R_e\{e_1, e_2\}) \ge h(\ell)+h(\ell')=1$. In case~(2), by Lemma~\ref{lemma_tree2}, either $R_e\{e_1,e_2\} \supseteq V(T_w)$ or $R_e\{e_1,e_2\} \supseteq V(T_{w'})$ for some $w'\in M_2(T)-\{w\}$; thus, $h(R_e\{e_1, e_2\}) \ge \min\{h(V(T_w)), h(V(T_{w'}))\}\ge1$ since $w,w'\in M_2(T)$. So, we consider case~(3). Note that $e_1\in E(T_y)$ for some $y\in M_1(T)$ or $e_1\not\in E(T_z)$ for any $z\in M(T)$; similarly, $e_2\in E(T_{y'})$ for some $y'\in M_1(T)$ or $e_2\not\in E(T_{z'})$ for any $z'\in M(T)$. If $\{e_1, e_2\} \subseteq E(T_v)$ for some $v\in M_1(T)$, then $d(v, e_1)\neq d(v, e_2)$, and there exist distinct $v', v''\in M_2(T)$ such that $v$ lies on the $v'-v''$ path in $T$ and $R_e\{e_1, e_2\} \supseteq V(T_{v'})\cup V(T_{v''})$; thus $h(R_e\{e_1, e_2\})\ge h(V(T_{v'}))+h(V(T_{v''}))\ge 2$. If $\{e_1, e_2\} \not\subseteq E(T_v)$ for any $v\in M(T)$, then there exist distinct $w_1, w_2\in M_2(T)$ such that both $e_1$ (or $y$) and $e_2$ (or $y'$) lie on the $w_1-w_2$ path in $T$; then $d(w_1, e_1)=d(w_1, e_2)$ and $d(w_2, e_1)=d(w_2, e_2)$ imply $e_1=e_2$, contradicting the assumption. So, $R_e\{e_1, e_2\} \supseteq V(T_{w_1})$ or $R_e\{e_1, e_2\} \supseteq V(T_{w_2})$, and thus $h(R_e\{e_1, e_2\}) \ge \min\{h(V(T_{w_1})), h(V(T_{w_2}))\}\ge 1$.~\hfill
\end{proof}
Next, we determine $\textnormal{edim}_f(C_n)$ for $n\ge 3$; we show that $\textnormal{edim}_f(C_n)=\dim_f(C_n)$.
\begin{proposition}
For $n\ge 3$, $\textnormal{edim}_f(C_n)=\left\{
\begin{array}{ll}
\frac{n}{n-1} & \mbox{ if $n$ is odd},\\
\frac{n}{n-2} & \mbox{ if $n$ is even}.
\end{array}\right.$
\end{proposition}
\begin{proof}
For $n\ge 3$, let $C_n$ be given by $u_0, u_1, \ldots, u_{n-1}, u_0$. Let $g:V(C_n) \rightarrow [0,1]$ be any edge resolving function of $C_n$. For each $i\in\{0, 1, \ldots, n-1\}$, we have the following: (i) if $n$ is odd, then $R_e\{u_iu_{i+1}, u_{i+1}u_{i+2}\}=V(C_n)-\{u_{i+1}\}$ and $g(V(C_n))-g(u_{i+1}) \ge 1$, where the subscript is taken modulo $n$; (ii) if $n$ is even, then $R_e\{u_iu_{i+1}, u_{i+1}u_{i+2}\}=V(C_n)-\{u_{i+1}, u_{i+1+\frac{n}{2}}\}$ and $g(V(C_n))-g(u_{i+1})-g(u_{i+1+\frac{n}{2}})\ge 1$, where the subscript is taken modulo $n$. In each case, by summing over the $n$ inequalities, we have $(n-1)g(V(C_n))\ge n$ for an odd $n$, and $(n-2)g(V(C_n))\ge n$ for an even $n$. Thus, $\textnormal{edim}_f(C_n) \ge \frac{n}{n-1}$ if $n$ is odd, and $\textnormal{edim}_f(C_n) \ge \frac{n}{n-2}$ if $n$ is even.
Now, let $h_0$ and $h_1$ be functions defined on $V(C_n)$ as follows: (i) if $n$ is odd, let $h_1(v)=\frac{1}{n-1}$ for each $v\in V(C_n)$; (ii) if $n$ is even, let $h_0(v)=\frac{1}{n-2}$ for each $v\in V(C_n)$. If $n$ is odd, $h_1$ is an edge resolving function of $C_n$ since $|R_e\{e_1, e_2\}| \ge n-1$ for any distinct $e_1, e_2\in E(C_n)$; thus, $\textnormal{edim}_f(C_n) \le h_1(V(C_n))=\frac{n}{n-1}$. If $n$ is even, $h_0$ is an edge resolving function of $C_n$ since $|R_e\{e_1, e_2\}| \ge n-2$ for any distinct $e_1, e_2\in E(C_n)$; thus, $\textnormal{edim}_f(C_n) \le h_0(V(C_n))=\frac{n}{n-2}$.~\hfill
\end{proof}
Next, for the Petersen graph $\mathcal{P}$, we show that $\textnormal{edim}_f(\mathcal{P})=\frac{5}{2}>\dim_f(\mathcal{P})$.
\begin{proposition}\label{edim_frac_petersen}
For the Petersen graph $\mathcal{P}$, $\textnormal{edim}_f(\mathcal{P})=\frac{5}{2}$.
\end{proposition}
\begin{figure}[ht]
\centering
\begin{tikzpicture}[scale=.8, transform shape]
\node [draw, shape=circle, scale=.8] (1) at (0,2) {};
\node [draw, shape=circle, scale=.8] (2) at (-1.9, 0.65) {};
\node [draw, shape=circle, scale=.8] (3) at (-1.36, -1.45) {};
\node [draw, shape=circle, scale=.8] (4) at (1.36, -1.45) {};
\node [draw, shape=circle, scale=.8] (5) at (1.9, 0.65) {};
\node [draw, shape=circle, scale=.8] (11) at (0,0.9) {};
\node [draw, shape=circle, scale=.8] (22) at (-0.85, 0.3) {};
\node [draw, shape=circle, scale=.8] (33) at (-0.64, -0.65) {};
\node [draw, shape=circle, scale=.8] (44) at (0.64, -0.65) {};
\node [draw, shape=circle, scale=.8] (55) at (0.85, 0.3) {};
\node [scale=1] at (0,2.35) {$u_0$};
\node [scale=1] at (-2.3,0.7) {$u_1$};
\node [scale=1] at (-1.75,-1.5) {$u_2$};
\node [scale=1] at (1.75,-1.5) {$u_3$};
\node [scale=1] at (2.3,0.7) {$u_4$};
\node [scale=1] at (0.38,1) {$w_0$};
\node [scale=1] at (-0.9,0) {$w_1$};
\node [scale=1] at (-0.6,-1) {$w_2$};
\node [scale=1] at (0.6,-1) {$w_3$};
\node [scale=1] at (1,0) {$w_4$};
\draw(1)--(2)--(3)--(4)--(5)--(1);\draw(11)--(33)--(55)--(22)--(44)--(11);\draw(1)--(11);\draw(2)--(22);\draw(3)--(33); \draw(4)--(44);\draw(5)--(55);
\end{tikzpicture}
\caption{Labeling of the Petersen graph.}\label{fig_petersen}
\end{figure}
\begin{proof}
Let the vertices of the Petersen graph $\mathcal{P}$ be labeled as in Figure~\ref{fig_petersen}.
First, we show that $\textnormal{edim}_f(\mathcal{P})\ge \frac{5}{2}$. Let $g:V(\mathcal{P}) \rightarrow [0,1]$ be any edge resolving function of $\mathcal{P}$. For each $i\in\{0,1,2,3,4\}$, we have $R_e\{u_iu_{i+1}, w_{i+2}w_{i+4}\}=\{u_i, u_{i+1}, w_{i+2}, w_{i+4}\}$, and thus $g(u_i)+g(u_{i+1})+g(w_{i+2})+g(w_{i+4}) \ge 1$, where the subscript is taken modulo $5$. By summing over the five inequalities, we have $2g(V(\mathcal{P}))\ge 5$, i.e., $g(V(\mathcal{P}))\ge \frac{5}{2}$. So, $\textnormal{edim}_f(\mathcal{P})\ge \frac{5}{2}$.
Second, we show that $\textnormal{edim}_f(\mathcal{P})\le \frac{5}{2}$. Let $h:V(G) \rightarrow [0,1]$ be a function defined by $h(v)=\frac{1}{4}$ for each $v\in V(\mathcal{P})$; then $h(V(\mathcal{P}))=\frac{10}{4}=\frac{5}{2}$. It suffices to show that $h$ is an edge resolving function of $\mathcal{P}$. Let $e_1$ and $e_2$ be distinct edges in $\mathcal{P}$. If $e_1$ and $e_2$ are not adjacent in $\mathcal{P}$, then, clearly, $|R_e\{e_1, e_2\}|\ge 4$. So, suppose $e_1$ and $e_2$ are adjacent in $\mathcal{P}$. Since $\mathcal{P}$ is edge-transitive (see~\cite{petersen}), we may assume that $e_1=u_0u_1$; then $e_2\in\{u_1u_2, u_1w_1, u_0u_4, u_0w_0 \}$. If $e_2=u_1u_2$, then $R_e\{e_1,e_2\}=\{u_0, u_2, u_3, u_4, w_0, w_2\}$ with $|R_e\{e_1,e_2\}|=6$; similarly, for each $e_2\in\{u_1w_1, u_0u_4, u_0w_0 \}$, we have $|R_e\{e_1,e_2\}|=6$. So, for any distinct edges $e_1, e_2\in E(\mathcal{P})$, $|R_e\{e_1, e_2\}|\ge 4$ and $h(R_e\{e_1, e_2\}) \ge 4(\frac{1}{4})=1$; thus, $h$ is an edge resoling function of $\mathcal{P}$. So, $\textnormal{edim}_f(\mathcal{P}) \le h(V(\mathcal{P}))=\frac{5}{2}$.~\hfill
\end{proof}
Next, we determine $\textnormal{edim}_f(W_n)$ for the wheel graph $W_n$ of order $n\ge4$; we show that $\textnormal{edim}_f(W_n)>\dim_f(W_n)$ for $n\ge 5$.
\begin{proposition}
Let $W_{n}=K_1+C_{n-1}$ be the wheel graph of order $n \ge 4$. Then
\begin{equation*}
\textnormal{edim}_f(W_{n})=\left\{
\begin{array}{ll}
\frac{n}{2} & \mbox{ if } n \in\{4,5\},\\
\frac{n-1}{2} & \mbox{ if } n\ge6.
\end{array}\right.
\end{equation*}
\end{proposition}
\begin{proof}
Let $n\ge 4$. Let $C_{n-1}$ be given by $u_0, u_1, \ldots, u_{n-2}, u_0$, and let $W_n$ be the wheel graph obtained from disjoint union of $C_{n-1}$ and $K_1$ by joining an edge between the vertex $v$ of $K_1$ and each vertex of $C_{n-1}$. Let $g: V(W_n) \rightarrow [0,1]$ be any edge resolving function of $W_n$.
First, suppose $n=4$. Then $\textnormal{edim}_f(W_4)=2$ by Theorem~\ref{dim_frac_graph}(d) and Theorem~\ref{edim_frac_n2}(a).
Second, suppose $n=5$. For each $i\in\{0,1,2,3\}$, we have the following: (i) $R_e\{vu_i, vu_{i+1}\}=\{u_i, u_{i+1}\}$ and $g(u_{i})+g(u_{i+1}) \ge 1$, where the subscript is taken modulo $4$; (ii) $R_e\{u_iv, u_iu_{i+1}\}=\{v, u_{i+1}\}$ and $g(v)+g(u_{i+1}) \ge 1$, where the subscript is taken modulo $4$. By summing over the four inequalities of (i) and (ii), respectively, we have $2 \sum_{i=0}^{3} g(u_i)\ge 4$ ($\spadesuit$) and $4g(v)+\sum_{i=0}^{3} g(u_i) \ge 4$ ($\maltese$). Now, if we multiply $\frac{3}{2}$ to the inequality ($\spadesuit$) and add it to the inequality ($\maltese$), then we have $4 g(V(W_5)) \ge 10$, i.e., $g(V(W_5))\ge \frac{5}{2}$; thus, $\textnormal{edim}_f(W_5) \ge \frac{5}{2}$. By Proposition~\ref{edim_frac_bounds}, $\textnormal{edim}_f(W_5)=\frac{5}{2}$.
Next, suppose $n\ge 6$. Note that, for each $i\in\{0,1,\ldots, n-2\}$, $R_e\{vu_i, vu_{i+1}\}=\{u_i, u_{i+1}\}$ and thus $g(u_i)+g(u_{i+1}) \ge 1$, where the subscript is taken modulo $(n-1)$. By summing over the $(n-1)$ inequalities, we have $2g(V(W_n))\ge n-1$; thus, $g(V(W_n)) \ge \frac{n-1}{2}$, which implies $\textnormal{edim}_f(W_n) \ge \frac{n-1}{2}$. Now, let $h:V(G) \rightarrow [0,1]$ be a function defined by $h(v)=0$ and $h(u_i)=\frac{1}{2}$ for each $i \in \{0,1,\ldots, n-2\}$; then $h(V(W_n))=\frac{n-1}{2}$. We show that $h$ is an edge resolving function of $W_n$. Let $e_1, e_2\in E(W_n)$ with $e_1\neq e_2$. If $e_1$ and $e_2$ are not adjacent in $W_n$, then $|R_e\{e_1, e_2\}| \ge 4$. If $e_1$ and $e_2$ are adjacent in $W_n$, then $R_e\{vu_i, vu_j\} \supseteq \{u_i, u_j\}$, $R_e\{u_iu_{i+1}, u_iu_{i-1}\} \supseteq \{u_{i+1}, u_{i-1}\}$, and $R_e\{u_iv, u_iu_{i+1}\} \supseteq \{v, u_{i+1}, u_{i+3}\}$, where the subscript is taken modulo $(n-1)$. In each case, $|R_e\{e_1, e_2\} \cap V(C_{n-1})| \ge 2$, and thus $h(R_e\{e_1, e_2\}) \ge 2(\frac{1}{2})=1$. So, $h$ is an edge resolving function of $W_n$; thus, $\textnormal{edim}_f(W_n) \le h(V(W_n))=\frac{n-1}{2}$. Therefore, $\textnormal{edim}_f(W_n)=\frac{n-1}{2}$ for $n\ge 6$.~\hfill
\end{proof}
Next, we determine $\textnormal{edim}_f(G)$ when $G=K_{a_1,a_2, \ldots, a_k}$ is a complete $k$-partite graph of order $n=\sum_{i=1}^{k}a_i\ge3$; we show that $\textnormal{edim}_f(G) \ge \dim_f(G)$, where $\textnormal{edim}_f(G) > \dim_f(G)$ if and only if $k\ge 3$ and $a_i=1$ for exactly one $i\in\{1,2,\ldots, k\}$.
\begin{proposition}\label{edim_frac_kpartite}
For $k\ge 2$, let $G$ be a complete $k$-partite graph of order $n\ge 3$. Then
\begin{equation}
\textnormal{edim}_f(G)=\left\{
\begin{array}{ll}
\frac{n-1}{2} & \mbox{ if $G=K_{1, n-1}$},\\
\frac{n}{2} & \mbox{ otherwise}.
\end{array}\right.
\end{equation}
\end{proposition}
\begin{proof}
For $k \ge 2$, let $G=K_{a_1, a_2,\ldots, a_{k}}$ be a complete $k$-partite graph of order $n=\sum_{i=1}^{k}a_i\ge 3$, and let $s$ be the number of partite sets of $V(G)$ consisting of exactly one element. Let $V(G)$ be partitioned into $V_1, V_2, \ldots, V_{k}$ with $|V_i|=a_i$ for each $i \in \{1,2,\ldots, k\}$; further, let $a_1 \le a_2 \le \ldots \le a_{k}$. We consider two cases.
\textbf{Case 1: $s \neq 1$.} In this case, $\dim_f(G)=\frac{n}{2}$ by Theorem~\ref{dim_frac_graph}(e). By Theorem~\ref{edim_frac_n2}(a), $\textnormal{edim}_f(G)=\frac{n}{2}$.
\textbf{Case 2: $s=1$.} First, suppose $k=2$; then $G=K_{1,n-1}$. By Proposition~\ref{edim_frac_tree}, $\textnormal{edim}_f(K_{1,n-1})=\frac{n-1}{2}$.
Second, suppose $k\ge 3$; we show that $\textnormal{edim}_f(G)=\frac{n}{2}$. Let $g:V(G) \rightarrow [0,1]$ be any edge resolving function of $G$. Let $V_1=\{v\}$, $V_2=\{x_1, x_2, \ldots, x_{\alpha}\}$ and $V_3=\{z_1, z_2, \ldots, z_{\beta}\}$, where $\alpha=a_2\ge2$ and $\beta=a_3\ge2$. We note the following: (1) for each $i\in\{1,2,\ldots, \alpha\}$, $R_e\{z_1v, z_1x_i\}=\{v, x_i\}$ and $g(v)+g(x_i)\ge1$; (2) for any distinct $i,j\in\{1,2,\ldots,\alpha\}$, $R_e\{z_1x_i, z_1x_j\}=\{x_i,x_j\}$ and $g(x_i)+g(x_j)\ge 1$. So, we have $\alpha \cdot g(v)+\sum_{i=1}^{\alpha}g(x_i)\ge \alpha$ by summing over the $\alpha$ inequalities from (1), and $(\alpha-1)\sum_{i=1}^{\alpha}g(x_i) \ge {\alpha \choose 2}$ by summing over the ${\alpha \choose 2}$ inequalities from (2); thus, $\alpha(g(v)+\sum_{i=1}^{\alpha}g(x_i)) \ge \alpha+{\alpha \choose 2}$, i.e., $g(V_1)+g(V_2)=g(v)+\sum_{i=1}^{\alpha}g(x_i)\ge \frac{1+\alpha}{2}=\frac{a_1+a_2}{2}$. We also note that, for any distinct $i,j\in \{1,2,\ldots, \beta\}$, $g(z_i)+g(z_j)\ge 1$ by Observation~\ref{obs_twin}(d). By summing over the ${\beta \choose 2}$ inequalities, we have $(\beta-1)\sum_{i=1}^{\beta} g(z_i)\ge {\beta \choose 2}$, i.e., $g(V_3)=\sum_{i=1}^{\beta}g(z_i) \ge \frac{\beta}{2}=\frac{a_3}{2}$. Similarly, if $k\ge4$, then we have $g(V_i)\ge \frac{a_i}{2}$ for each $i\in\{4,\ldots, k\}$. Thus, $g(V(G))=\sum_{i=1}^{k}g(V_i)\ge\sum_{i=1}^{k}\frac{a_i}{2}=\frac{n}{2}$, and hence $\textnormal{edim}_f(G)\ge \frac{n}{2}$. Since $\textnormal{edim}_f(G)\le\frac{n}{2}$ by Proposition~\ref{edim_frac_bounds}, we have $\textnormal{edim}_f(G)=\frac{n}{2}$ for $k\ge 3$.~\hfill
\end{proof}
Next, for the Cartesian product of two paths $P_s \square P_t$, where $s,t\ge2$, we show that $\textnormal{edim}_f(P_s\square P_t)=2=\dim_f(P_s \square P_t)$.
\begin{proposition}\emph{\cite{edim}}\label{edim_grid}
For $s,t \ge 2$, $\textnormal{edim}(P_s \square P_t)=2$.
\end{proposition}
\begin{proposition}
For $s,t \ge 2$, $\textnormal{edim}_f(P_s \square P_t)=2$.
\end{proposition}
\begin{proof}
Let $s,t\ge 2$. Let the vertices of $P_s \square P_t$ be labeled as in Figure~\ref{fig_grid}, and let $W=\{u_{i,j}: i \in \{1,s\} \mbox{ or } j \in \{1,t\}\} \subseteq V(P_s \square P_t)$. We note the following: (i) $R_e\{u_{1,1}u_{2,1}, u_{1,1}u_{1,2}\}=\{u_{i,1}: 2 \le i \le s\} \cup\{u_{1,j}: 2 \le j \le t\}$; (ii) $R_e\{u_{1,t}u_{2,t},u_{1,t}u_{1,t-1}\}=\{u_{i,t}: 2 \le i \le s\} \cup \{u_{1,j}: 1 \le j \le t-1\}$; (iii) $R_e\{u_{s,1}u_{s-1,1}, u_{s,1}u_{s,2}\}=\{u_{i,1}: 1 \le i \le s-1\} \cup \{u_{s,j}: 2 \le j\le t\}$; (iv) $R_e\{u_{s,t}u_{s-1,t}, u_{s,t}u_{s,t-1}\}=\{u_{i,t}: 1 \le i \le s-1\} \cup \{u_{s,j}: 1 \le j \le t-1\}$. If $g: V(P_s \square P_t) \rightarrow [0,1]$ is any egde resolving function of $P_s \square P_t$, then $\sum_{i=2}^{s}g(u_{i,1})+\sum_{j=2}^{t}g(u_{1,j}) \ge 1$, $\sum_{i=2}^{s} g(u_{i,t})+\sum_{j=1}^{t-1}g(u_{1,j}) \ge1$, $\sum_{i=1}^{s-1} g(u_{i,1})+\sum_{j=2}^{t} g(u_{s,j})\ge 1$ and $\sum_{i=1}^{s-1}g(u_{i,t})+\sum_{j=1}^{t-1}g(u_{s,j}) \ge 1$. By summing over the four inequalities, we have $2g(W)\ge 4$. So, $g(V(P_s \square P_t)) \ge g(W)=2$, and thus $\textnormal{edim}_f(P_s \square P_t) \ge 2$. Since $\textnormal{edim}_f(P_s \square P_t) \le 2$ by Observation~\ref{obs_frac}(b) and Proposition~\ref{edim_grid}, we have $\textnormal{edim}_f(P_s \square P_t)=2$.~\hfill
\end{proof}
\begin{figure}[ht]
\centering
\begin{tikzpicture}[scale=.65, transform shape]
\node [draw, shape=circle] (4) at (0,4.5) {};
\node [draw, shape=circle] (3) at (0,3) {};
\node [draw, shape=circle] (2) at (0,1.5) {};
\node [draw, shape=circle] (1) at (0,0) {};
\node [draw, shape=circle] (44) at (1.5,4.5) {};
\node [draw, shape=circle] (33) at (1.5,3) {};
\node [draw, shape=circle] (22) at (1.5,1.5) {};
\node [draw, shape=circle] (11) at (1.5,0) {};
\node [draw, shape=circle] (444) at (3,4.5) {};
\node [draw, shape=circle] (333) at (3,3) {};
\node [draw, shape=circle] (222) at (3,1.5) {};
\node [draw, shape=circle] (111) at (3,0) {};
\node [draw, shape=circle] (4444) at (4.5,4.5) {};
\node [draw, shape=circle] (3333) at (4.5,3) {};
\node [draw, shape=circle] (2222) at (4.5,1.5) {};
\node [draw, shape=circle] (1111) at (4.5,0) {};
\node [draw, shape=circle] (44444) at (6,4.5) {};
\node [draw, shape=circle] (33333) at (6,3) {};
\node [draw, shape=circle] (22222) at (6,1.5) {};
\node [draw, shape=circle] (11111) at (6,0) {};
\node [draw, shape=circle] (444444) at (7.5,4.5) {};
\node [draw, shape=circle] (333333) at (7.5,3) {};
\node [draw, shape=circle] (222222) at (7.5,1.5) {};
\node [draw, shape=circle] (111111) at (7.5,0) {};
\draw(1)--(2)--(3)--(4);
\draw(11)--(22)--(33)--(44);
\draw(111)--(222)--(333)--(444);
\draw(1111)--(2222)--(3333)--(4444);
\draw(11111)--(22222)--(33333)--(44444);
\draw(111111)--(222222)--(333333)--(444444);
\draw(1)--(11)--(111)--(1111)--(11111)--(111111);
\draw(2)--(22)--(222)--(2222)--(22222)--(222222);
\draw(3)--(33)--(333)--(3333)--(33333)--(333333);
\draw(4)--(44)--(444)--(4444)--(44444)--(444444);
\node [scale=1.3] at (-0.7,0) {$u_{1,1}$};
\node [scale=1.3] at (-0.7,1.5) {$u_{1,2}$};
\node [scale=1.3] at (-0.7,3) {$u_{1,3}$};
\node [scale=1.3] at (-0.7,4.5) {$u_{1,4}$};
\node [scale=1.3] at (1.5,-0.4) {$u_{2,1}$};
\node [scale=1.3] at (2.05,1.75) {$u_{2,2}$};
\node [scale=1.3] at (2.05,3.25) {$u_{2,3}$};
\node [scale=1.3] at (1.5,4.9) {$u_{2,4}$};
\node [scale=1.3] at (3,-0.4) {$u_{3,1}$};
\node [scale=1.3] at (3.55,1.75) {$u_{3,2}$};
\node [scale=1.3] at (3.55,3.25) {$u_{3,3}$};
\node [scale=1.3] at (3,4.9) {$u_{3,4}$};
\node [scale=1.3] at (4.5,-0.4) {$u_{4,1}$};
\node [scale=1.3] at (5.05,1.75) {$u_{4,2}$};
\node [scale=1.3] at (5.05,3.25) {$u_{4,3}$};
\node [scale=1.3] at (4.5,4.9) {$u_{4,4}$};
\node [scale=1.3] at (6,-0.4) {$u_{5,1}$};
\node [scale=1.3] at (6.55,1.75) {$u_{5,2}$};
\node [scale=1.3] at (6.55,3.25) {$u_{5,3}$};
\node [scale=1.3] at (6,4.9) {$u_{5,4}$};
\node [scale=1.3] at (8.2,0) {$u_{6,1}$};
\node [scale=1.3] at (8.2,1.5) {$u_{6,2}$};
\node [scale=1.3] at (8.2,3) {$u_{6,3}$};
\node [scale=1.3] at (8.2,4.5) {$u_{6,4}$};
\end{tikzpicture}
\caption{Labeling of $P_6 \square P_4$.}\label{fig_grid}
\end{figure}
We conclude this section with a few open problems and a remark showing that $\textnormal{edim}(G)-\textnormal{edim}_f(G)$ can be arbitrarily large.
\begin{question}
Are there graphs $G$ satisfying $\dim(G)=\textnormal{edim}(G)$ and $\dim_f(G)\neq \textnormal{edim}_f(G)$? Are there graphs $H$ satisfying $\dim(H)\neq\textnormal{edim}(H)$ and $\dim_f(H)=\textnormal{edim}_f(H)$?
\end{question}
\begin{question}
Does $\dim(G)>\textnormal{edim}(G)$ imply $\dim_f(G) \ge \textnormal{edim}_f(G)$? Similarly, does $\dim(G)<\textnormal{edim}(G)$ imply $\dim_f(G) \le \textnormal{edim}_f(G)$?
\end{question}
\begin{remark}
There exists a graph $G$ such that $\textnormal{edim}(G)-\textnormal{edim}_f(G)$ can be arbitrarily large. For $n\ge 3$, $\textnormal{edim}(K_n)=n-1$ (see~\cite{edim}) and $\textnormal{edim}_f(K_n)=\frac{n}{2}$ by Proposition~\ref{edim_frac_kpartite}; thus, $\textnormal{edim}(K_n)-\textnormal{edim}_f(K_n)=n-1-\frac{n}{2}=\frac{n-2}{2} \rightarrow \infty$ as $n \rightarrow \infty$. For another example, let $T$ be a tree with exterior major vertices $v_1, v_2, \ldots, v_k$ such that $ter(v_i)=\alpha\ge 3$ for each $i\in\{1,2,\ldots, k\}$, where $k\ge1$. Then $\textnormal{edim}(T)=\dim(T)=(\alpha-1)k$ (see~\cite{edim}) and $\textnormal{edim}_f(T)=\dim_f(T)=(\frac{\alpha}{2})k$ by Proposition~\ref{edim_frac_tree}; thus, $\textnormal{edim}(T)-\textnormal{edim}_f(T)=(\frac{\alpha-2}{2})k \rightarrow \infty$ when $\alpha \rightarrow \infty$ or $k \rightarrow \infty$.
\end{remark}
|
\section{Introduction}
\label{sec:introduction}
Methods for embedding networks
have become ubiquitous tools for analysis \cite{goyal2018graph,cai2018comprehensive}.
These methods exploit the geometrical representation of networks in a feature space,
which can be used in tasks such as compression \cite{wang2016structural}, clustering \cite{white2005spectral}, visualization \cite{pearson1901liii}, link prediction \cite{wang2016structural}, node classification \cite{tang2015line}, or scaling of a latent property that is determinant in the structure of networks.
Examples of this latter application can be found in the use of multidimensional scaling of networks for the retrieval of latent features in music for artist recommendation \cite{platt2004fast}, or in the use of Expectation Maximization algorithms for the estimation of ideological positions of voters, legislators, or online users \cite{imai2016fast}.
A growing domain of research has sought to exploit these
scaling methods to infer ideological feature spaces in which to embed users of social networks.
Taking inspiration in ideological scaling of voting data, pioneered by Poole et al. in the 1980s \cite{poole1985spatial}, and developed in the 1990s into the widely-used NOMINATE method \cite{poole1991patterns}, some adaptations to social network data have been proposed. The first such adaptation was proposed by Bond et al. \cite{bond2015quantifying}, then popularized by Barbera in the mid- to late 2010s \cite{barbera2015birds,barbera2015tweeting}.
These new scaling methods have been successful in identifying latent features in social networks related to the ideology of users.
They are, however, often limited to small subsets of the whole network of users.
The reason for this may be found in computational limitations in computing scaling for large networks, but mostly in the need for interpretability for the found features.
On Twitter, traditionally, an ideological scaling is limited to a sub-graph of parliamentarians and their followers, using the embedded features of the first ones to provide an interpretation for the found features \cite{barbera2015understanding}.
This raises the question:
How to compute features for portions of the network outside this \textit{seed} sub-graph?
The setting of this problem is similar to that of problems such as latent feature or link prediction, that accounts for a wealth of works to be discussed in Section~\ref{sec:related_work}.
But crucially, it deviates from these settings in that a minority of nodes have \textit{known} features (estimated via scaling), while those of a comparatively large set of nodes must be estimated from this initial seed set.
This marks a difference with statistical learning methods, where a flexible model is learned on a majority of nodes, and then applied to a minority of nodes with missing values.
The setting of ideological scaling invites the use of less flexible models, including strong assumptions about the link between ideology and the structure of the network.
This article takes inspiration in methods for information propagation in social networks and message coherence in telecommunications networks to propose a method for the propagation of scaled ideological features, from a seed set of nodes, to larger parts of a network.
While this problem arises in the case of ideological scaling in social networks in particular, it can also suit a more general class of problems in which n-dimensional features are only known for a subset of nodes in a network.
One may consider that latent ideological features are defined only for users that follow members of parliament (MPs).
However, our method supposes that a larger set of nodes could be positioned in the same latent feature space.
This underlying hypothesis is leveraged by our method exploiting two strong assumptions: 1) the ideology of a node is similar to that of its neighbors if this neighborhood is ideologically coherent, and 2) the ideology of a node is similar to other structurally similar nodes, that follow the same users for example, if these followed users are ideologically coherent.
After providing a review of the relevant related work in Section~\ref{sec:related_work}, we lay out the definitions and the notation needed for the treatment of the problem of coherent feature propagation in networks in Section~\ref{sec:preliminaries}.
Using this framework, we define the proposed propagation methods in Section~\ref{sec:value_propagation_methods}.
To analyze these methods, we develop an application case based on the ideology scaling on Twitter.
Section~\ref{sec:illustrative_case_data} presents the case of a seed set of Twitter accounts, followers of MPs in France, with known ideological scaling computed using procedures well-established in the literature.
Finally, in Section~\ref{sec:numerical_results}, we devise and execute experimental protocols to assess 1) the prediction capabilities using both of our methods following the two hypotheses in estimating ideological features, and 2) the limits of the estimation in terms of the trade-off between precision of the estimation, and its reach as parts of the whole network.
\section{Related Work}
\label{sec:related_work}
The methods proposed in this article are based in the propagation of features, from nodes for which they are known, to nodes for which they are not. This is mainly related to three different domains of research here reviewed. We intentionally omit the --vast-- scientific literature regarding the identification of important and central nodes in networks, outside the scope of these selected domains.
\subsubsection{Latent ideological features in social networks}
The first category of related works include those that compute latent features in social networks.
While several such methods exist, this article is concerned by those that leverage the interpretability of a subset of nodes to extract meaningful feature spaces.
This stands in contrast to works that produce feature embeddings on which to perform statistical learning such as deep learning, without necessarily providing an interpretation for the embedded features (cf. \cite{yan2006graph}).
A numerous family of methods, connected with link prediction, learn latent features in social networks, but specifically in relation with the probability of existence of a link, to be used in the propagation of local features into other parts of the time-evolving network \cite{heaukulani2013dynamic,wang2016structural}, or the retrieval of missing features for some nodes \cite{tang2015line}.
The most relevant works for this article, are those that use multidimensional scaling to extract ideological features \cite{barbera2015birds,barbera2015tweeting}.
Other works seek to predict ideology for large networks, linking prediction to homophily, but use ideological labels and not a --continuous-- ideological feature space \cite{colleoni2014echo,himelboim2016valence,xiao2020timme}.
\subsubsection{Propagation in social networks}
A second category of works pertinent for this study concerns the propagation in social networks, accounting for numerous results in a long tradition, but organized here in propagation of \textit{information} and of \textit{properties}.
Propagation of information counts important cases such as the spread of rumors and misinformation \cite{bao2013new}, with connection to epidemiological models \cite{sahneh2014competitive,chakrabarti2008epidemic}.
This setting is fundamentally different from the one of this article in that it considers the network as the support for the flow of messages that have specific positions in time and space (as opposed to permanent latent features or properties).
However, there are possible connections between network structure and homophily on the one hand, and structure in networks and flow of information on the other \cite{aral2009distinguishing,mcpherson2001birds}.
This connection is also related, for example, to types of users in rumor propagation cascades in social networks \cite{friggeri2014rumor}.
Finally, propagation in social networks has connections with label propagation on graphs, used for example in heuristic procedures in community detection \cite{zhu2002learning,cheng2010modeling}. However, again, this differs from the setting of this article in that a label can be equated to discrete classification, whereas features can provide (n-dimensional) continuous indices for nodes.
\subsubsection{Node coherence}
Node coherence is a concept that can be encountered in domains such as neuroscience \cite{thagard1998coherence}, artificial intelligence \cite{mackworth1977consistency}, or communications networks \cite{larsson2004large}, and often addresses the degree of dissimilarity of signals coming from different sources.
This concept is of particular importance in multi-path networks.
Depending on the type of network, coherence can be taken into account for node properties rather than relayed messages.
The most relevant domain of works along this line, is found in node coherence in opinion dynamics, where an opinion is often modeled as a continuous feature \cite{edmonds2012modeling}.
In this domain, the concept of opinion coherence is sometimes considered in dynamic models \cite{rodriguez2016collective}.
In opinion dynamics, however, the focus is put in measuring different indices for node coherence to then be used in simulations to study evolution of features in systems.
\section{Preliminaries}
\label{sec:preliminaries}
This section proposes the notation and definitions required to treat the concept of node and neighborhood coherence in networks, before proposing feature propagation methods in Section~\ref{sec:value_propagation_methods}.
\subsection{The universe network}
\label{sec:universe_network}
Let us consider a large \textit{universe network} as a directed graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$, for some non-empty set $\mathcal{V}$ with directed edges $\mathcal{E}\subseteq\mathcal{V}\times\mathcal{V}$.
In social networks, nodes typically represent user accounts, and directed edges represent a relation of \textit{following} between them.
Information flows \textit{downstream}, contrary to the direction of the edges: if a user shares information, it will be received by her followers and not by her \textit{followees}\footnote{Followees are also called \textit{friends} on Twitter}.
\begin{figure}[h!]
\centering
\includegraphics[width=0.5\columnwidth]{Directions.pdf}
\caption{Convention used for \textit{upstream} and \textit{downstream} directions on directed social networks, where edges indicate following/friendship.}
\label{fig:information_flow}
\end{figure}
For a node $v\in\mathcal{V}$, we consider the set of its \textit{downstream} or in-neighbors $\Uinneighbors{v}=\left\{ u \in \mathcal{V} : (u,v)\in\mathcal{E} \right\}$ and its \textit{upstream} or out-neighbors $\Uoutneighbors{v}=\left\{ u \in \mathcal{V} : (v,u)\in\mathcal{E} \right\}$. Abusing notation, we define the \textit{downstream} neighborhood of a set of nodes $V\subset\mathcal{V}$ as $\Uinneighbors{V} = \left\{ u \in \mathcal{V} : \exists v\in V \left( (u,v)\in \mathcal{E} \right) \right\}$,
and its \textit{upstream} neighborhood as $\Uoutneighbors{V} = \left\{ u \in \mathcal{V} : \exists v \in V \left( (v,u)\in \mathcal{E} \right) \right\}$.
It is worth noticing that, in general, we cannot assure $V\cap n_d (V)=\varnothing$, for $d \in\{\text{up},\text{down}\}$.
Given a direction $d$, we will denote its opposite direction with a bar, as $\overline{d}$.
\subsection{Features of nodes and estimation error}
\label{subsec:nodes_features}
Let us suppose that we can attribute ${N}$-dimensional features for some limited set of nodes $V\subset\mathcal{V}$.
When available for a node $v\in\mathcal{V}$, we denote their known ${N}$-dimensional features by $e(v)\in\mathbb{R}^{N}$, and its $i$-th component by $e_i (v)$.
When initially unavailable for a node $v\in\mathcal{V}$, ${N}$-dimensional features may be attributed through feature propagation from nodes for which they are known.
Methods for doing so are at the core of this article, and will be proposed in Section~\ref{sec:value_propagation_methods}.
We denote estimated features for a node $v\in\mathcal{V}$ by $\hat{e}(v)\in\mathbb{R}^{N}$, and by $\hat{e}_i(v)$ its $i$-th component.
Whenever we can estimate features $\hat{e}(v)$ through propagation methods for a node $v\in\mathcal{V}$ for which we also know its \textit{true} features $e(v)$, we can compute the estimation error $\error{\hat{e}(v)}{e(v)}$.
We measure this estimation error as the $p$-norm in the feature space:
\begin{equation}
\error{\hat{e}(v)}{e(v)}=\|\hat{e}(v)-e(v) \|_p.
\label{eq:estimation_error}
\end{equation}
For a set of nodes $V$, the mean estimation error will be denoted by $E(V)=(1/|V|)\sum\nolimits_{v\in V} \error{\hat{e}(v)}{e(v)}$.
\subsection{Node coherence}
\label{subsec:node_coherence}
Node coherence is a key concept in this article, and seeks to capture the degree of similarity of a set of nodes in the feature space.
Given $V\subset\mathcal{V}$ for which features are known or estimated, we denote by $I(V)$ the incoherence of $V$.
Several incoherence metrics are possible, of which we propose to use the standard deviation of the distance to the centroid of $V$ in the feature space:
\begin{equation}
I(V) = \sqrt{\frac{1}{|V|}\sum\limits_{v\in V}\| e(v)-c_V\|^2_p },
\label{eq:set_incoherence}
\end{equation}
\noindent{}where $c_V = (1/|V|)\sum\nolimits_{v\in V} e(v)$ is the centroid of $V$.
We prefer the definition of Equation \eqref{eq:set_incoherence} over other intuitive common metrics such as entropy \cite{onesto2019relating}: a set $V$ can have minimal entropy and high values of deviation $I(V)$ if the underlying probability distribution of $V$ is multi-modal in the feature space.
\subsection{$\varepsilon$-coherent neighborhoods of a set of nodes}
Whenever we have a set of nodes $V\subset\mathcal{V}$, we can consider the up- or downstream neighborhoods of nodes that are also coherent: i.e., nodes that are linked to a set, and whose incoherence with respect to this set is below a given threshold $\varepsilon$.
We formalize this notion, for both directions, as the \textit{$\varepsilon$-coherent upstream} and \textit{downstream neighborhood} of set $V$:
\begin{equation*}
n^\varepsilon_d(V) = \left\{ u\in n_d(V) : I(n_{\overline{d}}(u)\cap V)\leq \varepsilon \right\}
\end{equation*}
\noindent{}for $d\in\{\text{up},\text{down}\}$, and $\overline{d}$ the direction opposite to $d$.
\begin{figure}[h!]
\centering
\includegraphics[width=0.85\columnwidth]{CoherentNeighborhoods.pdf}
\caption{Illustration of the construction of a $\varepsilon$-coherent downstream neighborhood $n^\varepsilon_{\text{down}}(V)$ of a set $V$ embedded in a 1-dimensional feature space.}
\label{fig:coherent_neighborhood}
\end{figure}
\section{Propagation Methods for Latent Features in Social Networks}
\label{sec:value_propagation_methods}
We consider two different approaches for propagating ideological features in a larger set of the network: Method A) directed sequences of $\varepsilon$-coherent neighborhoods, and Method B) sequences of projections using $\varepsilon$-coherent neighborhoods.
Both families of methods are based on the concept of coherent neighborhoods, but motivated by distinct intuitions.
Method A assumes that the features attached to nodes are ``diffusing'' along the social network formed by follower/followee relationships.
This method is rooted in the larger homophily hypothesis ``that birds of a feather flock together'' \cite{mcpherson2001birds}, meaning that two connected users are likely to demonstrate homophilic behavior, sharing a similar ideological position.
Method B has different premises.
Now, if unknown, ideological features of a node are estimated using those of other nodes that occupy a similar position in the network.
This method is based on the structural equivalence hypothesis, which posits that two nodes sharing the same neighborhood are similar \cite{sailer1978structural}.
The approach of this second family of methods is, for example, at the core collaborative filtering approaches in algorithmic recommendation.
Here we will follow the same principle, estimating the ideological features of a user on Twitter, as the aggregation of features of users following, or being followed by the same users.
\subsection{Directed sequences of $\varepsilon$-coherent neighborhoods}
\label{subsec:definition_method_1}
This method generates two sequences of sets of nodes: one of coherent nodes $\{V_i\}_{i\geq 0}$, for which we estimate the ideology, and one of incoherent nodes $\{\overline{V}_i\}_{i\geq 0}$, that we avoid using in ideology estimation.
Let us consider a seed set $V_0\subset\mathcal{V}$ for which features $e(v)$ for $v\in V_0$ are known,
and a direction $d\in\{\text{up},\text{down}\}$. Starting at $V_0$, a directed sequence of $\varepsilon$-coherent sets of nodes $V_0,V_1,V_2,\ldots$ is computed as $V_{i+1} = V_i \cup \Delta V_{i}$ for $i=0,1,2,\ldots$, with
\begin{equation}
\Delta V_{i} = \left\{v\in n^\varepsilon_{d} \left( V_{i} \right) : v\notin \left(V_i \cup \overline{V}_i \right) \right\},
\label{eq:directed_coherent_sequence}
\end{equation}
\noindent{}where $\overline{V}_{i+1} = \overline{V}_i \cup \Delta \overline{V}_{i}$ for $i=0,1,2,\ldots$, with $\overline{V}_0=\varnothing$ and
\begin{equation}
\Delta \overline{V}_i = \left\{v\in n_{d}\left( V_{i} \right) : v\notin \left( \overline{V}_i \cup \Delta V_i \right) \right\}.
\label{eq:directed_incoherent_sequence}
\end{equation}
By definition, $\{V_i\}_{i\geq0}$ is an incremental ($V_{i}\subset V_{i+1}$) by disjoint additions ($V_i\cap\Delta V_i=\varnothing$) sequence of $\varepsilon$-coherent neighbors.
Sequence $\{\overline{V}_i\}_{i\geq0}$ is also increasing ($\overline{V}_{i}\subset \overline{V}_{i+1}$) by disjoint additions ($\overline{V}_i\cap\Delta \overline{V}_i=\varnothing$), but of incoherent neighbors.
Nodes deemed incoherent are stored to avoid coherence collisions: an incoherent node cannot become coherent by virtue of new nodes who had been attributed features by propagation.
Whenever new $\varepsilon$-coherent neighbors $\Delta V_i$ are discovered at the $i$-th step, the features of their nodes are estimated as
\begin{equation}
\hat{e}(v) =\frac{1}{\left|n_{\overline{d}}(v)\cap V_i\right|} \sum\limits_{u\in n_{\overline{d}}(v)\cap V_i} \hat{e}(u), \;\text{ for }v\in\Delta V_i,
\end{equation}
\noindent{}setting $\hat{e}(v)=e(v)$ for $v\in V_0$.
\begin{figure}[h!]
\centering
\includegraphics[width=0.75\columnwidth]{MethodA.pdf}
\caption{Schematic representation of the computation of $\Delta V_i$ and $\Delta \overline{V}_i$ on the $i$-th step of an upstream directed sequence of $\varepsilon$-coherent neighborhoods using \textbf{Method A}. In the upstream variant, users from $V_i$ follow users from $\Delta V_i$.}
\label{fig:diagram_methodA}
\end{figure}
\subsection{Sequences of projections using $\varepsilon$-coherent neighborhoods}
\label{subsec:definition_method_2}
Inspired by the way in which features are computed in multi-dimensional scaling for users depending who they follow, we propose a second family of methods, Method B, for feature propagation using $\varepsilon$-coherent neighborhoods.
Most ideology scaling methods exploit the collective structural relations between nodes that are important for the latent features of a network (typically politicians) and their followers \cite{barbera2015birds,barbera2015tweeting}.
While intuitive, the first family of methods proposed in Section~\ref{subsec:definition_method_1} overlooks this aspect.
This can have disadvantages: for example, some nodes can attract followers without themselves following many nodes coherently. This could arguably be the case of Twitter accounts of media outlets, institutions, or public figures.
Accordingly, we propose a second family of methods centered around the notion of structural similarity: nodes that are similar because they follow (upstream) or are followed (downstream) by the same users, hereafter called \textit{pivots}.
As with the previous method, Method B generates two sequences of sets of nodes: one of coherent nodes $\{V_i\}_{i\geq 0}$, for which we estimate the ideology, and one of incoherent nodes $\{\overline{V}_i\}_{i\geq 0}$ that we now avoid using as pivots.
Let us consider a seed set $V_0\subset\mathcal{V}$ for which features are known and a direction $d\in\{\text{up},\text{down}\}$.
Starting at $V_0$, a sequence of projected $\varepsilon$-coherent sets of nodes $V_0,V_1,V_2,\ldots$ is computed as $V_{i+1}=V_i\cup\Delta V_{i}$ for $i=0,1,2,\ldots$, for which we consider the set $P^\varepsilon_i$ of $\varepsilon$-coherent pivots:
\begin{equation}
P^\varepsilon_i = n^\varepsilon_d (V_i)\backslash{}\overline{V}_i,
\end{equation}
\noindent{}where $\overline{V}_{i+1}=\overline{V}_i \cup \Delta\overline{V}_i$ for $i=0,1,2,\ldots$, with $\overline{V}_0=\varnothing$, and
\begin{equation}
\Delta \overline{V}_i = \left\{v\in n_{d}(V_i): v\notin\left( n^\varepsilon_{d}(V_i)\cup \overline{V}_i \right) \right\}.
\end{equation}
As with the previous family of Method A, the sets $\overline{V}_i$ store the nodes deemed incoherent and that cannot be used, but now as pivots.
The sets $P^\varepsilon_i$ are used at each iteration to compute additions $\Delta V_i$ according to coherent structural similarity:
\begin{equation}
\Delta V_i = \left\{v\in n^\varepsilon_{\overline{d}}\left( P^\varepsilon_i \right): v \notin V_i \right\}.
\end{equation}
Again by definition, sequences $\{V_i\}_{i\geq0}$ and $\{\overline{V}_i\}_{i\geq0}$ are incremental by disjoint additions.
In contrast with the first family of Method A from Section~\ref{subsec:definition_method_1}, now it is the coherence of the pivot nodes in sets $P^\varepsilon_i$ that is assured.
Whenever new neighbors $\Delta V_i$ are discovered at the $i$-th iteration, their features are estimated as
\begin{equation}
\hat{e}(v) =\frac{1}{\left|C_i(v,P^\varepsilon_i,V_i)\right|} \sum\limits_{u\in C_i(v,P^\varepsilon_i,V_i)} \hat{e}(u), \;\text{ for }v\in\Delta V_i,
\end{equation}
\noindent{}where $C_i(v,P^\varepsilon_i,V_i)$ is the set of co-neighbors of $v$ in $V_i$ through pivot $P^\varepsilon_i$:
\begin{equation}
C_i(v,P^\varepsilon_i,V_i)= V_i \cap n_{\overline{d}}\left( n_{d}(v) \cap P^\varepsilon_i \right) .
\label{eq:coneighbors}
\end{equation}
\begin{figure}[h!]
\centering
\includegraphics[width=0.7\columnwidth]{MethodB.pdf}
\caption{Schematic representation of the computation of $\Delta V_i$ and $\Delta \overline{V}_i$ using pivot $P^\varepsilon_i$ on the $i$-th step of an upstream (co-followers-based) sequence of projected $\varepsilon$-coherent neighborhoods using \textbf{Method B}. In the upstream variant, users from $V_i$ and $\Delta V_i$ co-follow users from pivot set $P^\varepsilon_i$.}
\label{fig:diagram_methodB}
\end{figure}
\section{French Political Twitter Dataset}
\label{sec:illustrative_case_data}
In order to analyze the feature propagation methods described in the previous section on real data, here we present a pertinent part of the Twitter network\footnote{In conformity with the General Data Protection Regulation 2016/679, the project called \textit{Ideology scaling Twitter France}, whose data we exploit, has been declared the 19 Mars 2020 at the registry of data processing at the \textit{Fondation Nationale de Sciences Politiques} (Sciences Po), and respects Twitter's data use policies.} and an ideological scaling producing an interpretable ideological feature space.
\subsection{French MPs and their followers}
Our data collection starts with the set $\mathcal{P}$ of the $831$ (out of 925) French MPs present on Twitter\footnote{The list of French MPs present on social networks is provided by the French chambers of parliament at \url{http://www2.assemblee-nationale.fr/deputes/liste/reseaux-sociaux}, for deputies, and \url{http://www.senat.fr/espace_presse/actualites/201402/les_senateurs_sur_twitter.html} for senators.}, belonging to 10 main different parties or groups/alliances of parties.
We then proceeded to collect all the followers of the accounts of MPs in $\mathcal{P}$.
This collection was conducted on May 2019, and resulted in the constitution of the set $\mathcal{F}=n_\text{down}(\mathcal{P})$ of followers of $\mathcal{P}$, which amounts to $|\mathcal{F}|=\setFsize$ unique Twitter accounts.
\subsection{An ideological inference procedure via multidimensional scaling}
We follow the methodology described by Barbera et al. \cite{barbera2015tweeting} for inferring ideological features through multidimensional scaling of the sub-graph of the MPs and their followers.
After removing from this sub-graph the followers that follow less than 3 MPs, and then removing users that had a repeated set of followed MPs (to assure full rank of the adjacency matrix), we obtained a set $\tilde{\mathcal{F}}$ of $368.831$ accounts.
We represent this sub-graph as a $\{0,1\}^{|\tilde{\mathcal{F}}|\times|\mathcal{P}|}$ adjacency matrix , where a relation of following is encoded with the value 1, and its absence with the value 0.
Next, we produce a reduced-dimensionality representation of these $368.831$ observations using a Correspondence Analysis (CA) \cite{greenacre2017correspondence}.
The first 2 principal components of the reduced-dimensionality space, PC1 and PC2, explain 0,97\%{} and 1,07\%{} of the inertia.
This might seem as little explanation of the variability in the observed ways in which users from $\tilde{\mathcal{F}}$ follow MPs.
However, the experience in several countries has revealed that the first components offer useful interpretations in the Twitter MPs' network.
Examples include the UK, Spain, Italy, and Netherlands \cite{barbera2015birds}).
Similar results have also been found in France \cite{briatte2015recovering,cointet2020color}.
In accordance with analyses made for the cases of other countries, those of France suggest feature interpretation where the first component is identified with left-right ideology.
\subsection{Interpreting the ideological feature space }
To provide and interpretation of PC1 and PC2 for our sets $\mathcal{P}$ and $\tilde{\mathcal{F}}$, we consider exogenous variables for the parties of the MPs.
We use the 2019 Chapel Hill Expert Survey (CHES) data \cite{chesdata2019}: an estimation on ideology and policy position of European parties, made by 421 political scientists specializing in political parties.
Out of the 10 political parties identified for accounts in $\mathcal{P}$, 8 are also present in the CHES data.
We computed the correlations between the positions of MPs on PC1 and PC2 of the feature space with all the 51 criteria were CHES data produces estimation for their political parties.
This allowed to identify the two most relevant criteria related to our axes.
Fig.~\ref{fig:boxplot_embedded_values_french_MPs} shows the ordering of these eight parties according to the two different criteria: 1) parties' economic views, from left to right, and 2) parties' attitudes towards European integration, from opposed to favorable. For each party in Fig.~\ref{fig:boxplot_embedded_values_french_MPs}, we provide the embedded features of PC1 per MP for criterion 1 (left-right ideology), and the embedded features of PC2 per MP for criterion 2 (attitudes towards European integration).
While it is not the objective of this article,
Figure~\ref{fig:boxplot_embedded_values_french_MPs} illustrates the potential applicability of the multidimensional scaling procedure.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{boxplots_ranking.pdf}
\caption{Distribution of the embedded ideological features PC1 and PC2 for the French MPs on Twitter per party. Parties are ordered according to the two criteria using the 2019 Chapel Hill Expert Survey data: left-right economics for feature PC1 (top), and attitudes toward European integration for feature PC2 (bottom).}
\label{fig:boxplot_embedded_values_french_MPs}
\end{figure}
Under this interpretation, the French Twitter network of MPs and their followers (the sets $\mathcal{P}$ and $\tilde{\mathcal{F}}$) may be represented in a 2-dimensional latent ideological feature space.
In this feature space, the first dimension (PC1) provides an index for the concept of \textit{left} and \textit{right} positions, and the second dimension (PC2) provides an index for attitudes towards European integration, which is in line with observations that have been made in the case of other countries \footnote{Research results in several countries have also identified this second axis as related with issues dividing public along pro- and anti-establishment positions \cite{schmidt2020measuring,chiche2000espace}.}.
Fig.~\ref{fig:2d_histogram_embedded_seed_bipartite} illustrates the positions of the sets $\mathcal{P}$ of parliamentarians and $\tilde{\mathcal{F}}$ of their followers in
this bidimensional ideological feature space.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{emb_space_PbarA_hex_2.pdf}
\caption{Position of $\mathcal{P}{}$ French MPs in the first two embedded dimensions, PC1 (left to right ideology) and PC2 (pro- and against European integration), histogram of the positions of their followers $\tilde{\mathcal{F}}{}$, and positions of sampled accounts in set $\mathcal{A}$.}
\label{fig:2d_histogram_embedded_seed_bipartite}
\end{figure}
\subsection{Datasets for experimentation}
We are now concerned with the problem of using the set $\tilde{\mathcal{F}}$ with $368.831$ accounts with known features to establish datasets for the testing and evaluation.
We cannot set $V_0=\tilde{\mathcal{F}}$ for testing and evaluation.
This is because when producing the next set $V_1$ in the sequence (with either method) the new added nodes would not have known \textit{true} features $e(v)$ (as computed with ideological scaling) to test the estimated ones $\hat{e}(v)$, for $v\in \Delta V_0$.
To circumvent this difficulty, we take a subset $\mathcal{A}\subset\tilde{\mathcal{F}}$, and then collect its upstream and downstream neighborhoods to use in the described methods, allowing for some elements of these new neighborhoods to be also in $\tilde{\mathcal{F}}$.
To account for the possible specificities of the different regions of the ideological feature, we sample $4.483$\footnote{Originally, we sampled 5.000 nodes, but the later collection of followers and followees for 517 of these nodes was not possible because these accounts were either closed between collections, or private and information could not be collected.} nodes uniformly in space from $\tilde{\mathcal{F}}$ to produce set $\mathcal{A}$ (see Fig.~\ref{fig:2d_histogram_embedded_seed_bipartite}).
Had we not sampled with spatial uniformity, the sampling would have produced a majority of nodes near the origin of the ideological feature space (see the concentration of nodes near the origin in Fig.~\ref{fig:2d_histogram_embedded_seed_bipartite}).
Next, we collect the followees/friends of $\mathcal{A}$ as $\mathcal{B} = n_{\text{up}}(\mathcal{A})$ (obtaining $|\mathcal{B}|=1.304.812$), and the followers of set $\mathcal{A}$ as $\mathcal{C} = n_\text{down}(\mathcal{A})$ (obtaining $|\mathcal{C}|=5.528.716$).
This sub-sampling and collection operations achieve sets for our test evaluations that are such that $|\mathcal{B} \cap \tilde{\mathcal{F}}| = 138.424$, and $|\mathcal{C} \cap \tilde{\mathcal{F}}| = 231.035$.
\section{Numerical Experiments}
\label{sec:numerical_results}
In this section we analyze some results of the methods proposed in Section~\ref{sec:value_propagation_methods} for estimating ideological features on Twitter data.
For both methods, we center the analysis around the set $\mathcal{A}$ built for this purpose, and described in the previous section.
We will be interested in analyzing, for different values of coherence $\varepsilon$, 1) the accuracy (the degree of error in estimating ideological features), and 2) the coverage (the size of the disjoint additions with estimated features) of the proposed methods.
The main aspect we seek to investigate in this section is the relation between the coherence threshold parameter $\varepsilon$ and the trade-off between estimation error and size of a disjoint addition $\Delta V_i$ at a step $i$.
Throughout this section, we compute error $E$ (cf. Equation~\eqref{eq:estimation_error}) and coherence $I$ (cf. Equation~\eqref{eq:set_incoherence}) using $p=2$.
\subsection{Directed sequences of $\varepsilon$-coherent neighborhoods}
Because we have collected the upstream and downstream neighborhoods of set $\mathcal{A}$ (sets $\mathcal{B}$ and $\mathcal{C}$), we can readily compute the first step for Method A in both directions.
Figure~\ref{fig:1st_iterations_directed_sequence} reports the error $E(\Delta V_0)$ and the size of $\Delta V_0$ for the first step ($i=0$).
\begin{figure}[h!]
\centering
\includegraphics[width=\columnwidth]{sequence_directed_1st_iteration_error_inset_coverage.pdf}
\caption{Mean estimation error $E(\Delta V_0)$ and size of $\Delta V_0$ for Method A in the upstream and downstream directions, using $V_0=\mathcal{A}$.}
\label{fig:1st_iterations_directed_sequence}
\end{figure}
While intuitive, the estimation of ideological features through directed propagation of Method A in either direction performs poorly.
Imposing higher levels of coherence (lower thresholds $\varepsilon$ for incoherence) does improve the mean estimation error of nodes in
$\Delta V_0$.
However, this improvement is marginal and, most importantly, limited.
Upstream neighbors (friends/followees) of users in $\mathcal{A}$ that are followed by highly coherent users, have their ideology estimated with a mean error no lower than $E(\Delta V_0)=$0.6, which amounts to 12.5\%{} of the diameter of $\tilde{\mathcal{F}}$ in the feature space.
Similarly, downstream neighbors (followers) of users in $\mathcal{A}$ that follow highly coherent users, have their ideology estimated with a mean error no lower than $E(\Delta V_0)=$0.575 (12.0\%{} of the diameter of $\tilde{\mathcal{F}}$).
Had set $\Delta V_1$ be computed (upstream or downstream), its ideological features would be estimated only from those already --poorly-- estimated for set $\Delta V_0$ (disjoint additions, assured by Equation~\eqref{eq:directed_coherent_sequence}, imply that $n_{d}(V_0) \cap \Delta V_1 = \varnothing$).
\subsection{Sequences of projections using $\varepsilon$-coherent neighborhoods}
The setting of Method B is similar to that of collaborative filtering (CF) \cite{schafer2007collaborative} in Recommender Systems.
In user-based CF, the similarity of users is computed according to structural similarity on a bipartite graph: users are similar if they have chosen similar items.
Items can then be proposed to a user among those already chosen by similar users (structural similarity is related to preference).
In the setting of our Method B, users are deemed similar if they follow, or are followed by (depending if the selected direction is upstream or downstream) similar users of a so-called pivot set for a predetermined coherence $\varepsilon$.
Taking on known evaluation protocols for CF \cite{bobadilla2012collaborative}, we propose a method for assessing the accuracy and coverage of Method B.
Given a positive integer $K$, we perform a $K$-fold bipartite cross-validation. We divide $\mathcal{A}$ in $K$ parts, taking one as $\mathcal{A}^K_\text{test}$ and the rest as $\mathcal{A}^K_\text{train}$.
We set $V_0 = \mathcal{A}^K_\text{train}$, we compute $P^\varepsilon_0$ using $\mathcal{B}$ and $\mathcal{C}$ for upstream and downstream directions, and then set $\mathcal{R}=\mathcal{A}^K_\text{test} \cap \Delta V_0$.
To analyze the coverage, we examine the quantity $\text{cov}(\mathcal{R}) = |\mathcal{R}|/|\mathcal{A}^K_\text{test}|$, ranging from 0 (no coverage) to 1 (total coverage).
We choose $K=20$ and report the results for varying values of incoherence $\varepsilon$, providing, for the ensemble of the 20 folds, the median, the maximum, and the minimum value in Fig.~\ref{fig:1st_iteration_projected_sequence}.
\begin{figure}[h!]
\centering
\includegraphics[width=\columnwidth]{sequence_projected_1st_iteration_error_insets.pdf}
\caption{Estimation error and coverage of set $\mathcal{R}=\mathcal{A}^K_\text{test} \cap \Delta V_0$ and $P_0$ for the Method B in a $K$-fold cross-validation retrieval of ideologies of $\mathcal{A}$ ($K=$20). The distribution of errors and coverage over the folds is reported with the median, the minimum, and the maximum values .}
\label{fig:1st_iteration_projected_sequence}
\end{figure}
Results obtained with Method B, reported in Fig.~\ref{fig:1st_iteration_projected_sequence}, are more satisfactory than those obtained with Method A.
Thresholds for incoherence below $\varepsilon=1$ (approximately) already allow for less error in estimation in comparison to Method A.
In fact, under our setting, estimation errors for Method B can be made as low as (approximately) 0.1 in the feature space by sacrificing coverage of new nodes added in the following step of the sequence.
With a small value of threshold for $\epsilon$ (lower than 0.1) the estimation error is around 0.1 in distance in the feature space (2.1\%{} of the diameter of $\tilde{\mathcal{F}}$).
This trade-off in coverage does not come at great expense, as suggested by our proposed metric $\text{cov}(\mathcal{R})$: at least 50\%{} of nodes left in the test set $\mathcal{A}^K_{\text{test}}$ can be recovered, independently of the threshold for incoherence.
The amount of nodes in the set of pivots $P^\varepsilon_0$ is that of $\Delta V_0$ for Method A, and follows intuition in that, when forcing less incoherence, less nodes are available for the search of co-neighbors (cf. Equation~\eqref{eq:coneighbors}).
While it was possible to evaluate the quality of Method A precisely, the same evaluation would have required the collection of the downstream neighbors of $\mathcal{B}$ (i.e. $n_{\text{down}}\left( n_{\text{up}}(\mathcal{A})\right)$) and the upstream neighbors of $\mathcal{C}$ (i.e. $n_{\text{up}}\left( n_{\text{down}}(\mathcal{A})\right)$), which would amount to a mixed-direction snowball-sampling of the Twitter network from $\mathcal{A}$.
Thus the need for a different evaluation protocol for both methods.
While there is a slight risk of over-estimating the coverage capacity of Method B using our testing protocol (due to the fact that $\mathcal{R}$ is restricted to $\mathcal{A}$) it has the great advantage of allowing for experimentation with large sets: the (up- and downstream) neighbors of $\mathcal{A}$, i.e., the total population for feature propagation, account for nearly 6.5M users.
\section{Conclusions}
\label{sec:conclusions}
We proposed two methods (A and B) for the propagation of latent ideological features on Twitter.
These methods use propagation for the estimation of ideology of users, represented in a continuous n-dimensional features space.
Method A is based on the homophily hypothesis: if users are connected (one follows the other) their ideology is similar.
Method B is based on the structural similarity hypothesis: if users are connected to the same neighborhoods their ideology is similar.
In these methods, we proposed the consideration and formalization of these notions when depending on ideological coherence of neighbors, which we model with a single parameter.
To analyze our methods, we collected Twitter data using the accounts of French MPs.
We were able to produce a 2-dimensional ideological embedding for a sub-graph of nearly 370K users. With the help of the Chapel Hill Expert Survey, we validate the interpretation of the two emerging dimensions as related to left-right ideology, and attitudes toward European integration.
Collecting again neighborhoods of some of these users we used our methods to propagate these ideological features with predefined degrees of accuracy within a potential population of nearly 6.5M users.
To total number of users for which ideology is propagated, depends on the error to be accepted, which is determined by a coherence threshold $\varepsilon$.
The protocols established for evaluation allow us to analyze the relation between estimation accuracy for ideology in propagation, and the coverage of the method (i.e., the number of users to which we can propagate).
The trade-off between these two competing and desirable properties was modulated by our coherence parameter.
Analyzing Methods A and B, we find that, when coherence is considered, the ideology of a user on Twitter is better estimated using other users that are structurally similar (Method B), than using other users that might be directly connected (Method A).
These results lend support to the structural similarity hypothesis over the homophily hypothesis.
Even more: while the difference is small, the ideology of a user is better estimated using users that follow the same accounts, than using users that are followed the same Twitter accounts.
The formalism used to treat coherence in social networks, and the methods proposed for propagation, can be extended to any other network were directed edges model the fact a user can receive information from another.
On Facebook, for example, the fact that user $a$ and user $b$ are \textit{friends}, could be represented by two directed edges between both users, and in different directions.
In other networks, such as Instagram or YouTube (using the \textit{subscription} relation) the formalism and methods can be directly applied.
\section*{Acknowledgments}
We thank Justin Clarke and Benjamin Ooghe-Tabanou for their help in the collection of Twitter data.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Neural networks have seen great success empirically, which has inspired a large amount of theoretical work. However, neural networks lack rigorous mathematical foundations to explain their success at the moment. Optimization, among other aspects of deep learning theory, has received considerable attention in recent years. One challenge in deep learning is to avoid gradient descent being stuck at poor local minima. Thus, analyzing the optimization landscape of neural networks has been a major subject of study. In fact, almost all non-linear real-valued neural networks are shown to have poor local minima, which includes neural networks with {\it quadratic}, {\it tanh}, {\it sigmoid}, {\it arctan}, ReLU, ELU, and SELU activations \cite{yun2018small} . Even under the over-parametrization condition, real-valued neural networks have strict poor local minima \cite{9194023}. In this paper, we demonstrate that neural networks with complex-valued weights have a completely different optimization landscape. For shallow neural networks with {\it quadratic} activations, complex-valued weights eliminate all poor local minima.
We focus on the loss surface of shallow neural networks with {\it quadratic} activations, which have been analyzed frequently in the real-valued case \cite{NIPS2019_9111,SJL19,pmlr-v80-du18a,mannelli2020optimization,gamarnik2020stationary} and were shown to have infinitely many spurious local minima \cite{yun2018small}. However, with complex-valued weights, we prove that there are no spurious local minima. In addition, we provide some specific examples of poor local minima in the real-valued case and demonstrate how they become saddle points in the complex-valued neural networks (CVNNs). Since complex analytic functions have superior landscapes, which will be discussed later, we conjecture that the \enquote{no spurious local minima} result also holds true for deeper complex networks and complex networks with other analytic activations such as {\it tanh} and {\it sigmoid}. In comparison, we prove that spurious local minima exist in complex neural networks with a non-analytic activation $\mathbb{C}$ReLU.
This is the first work analyzing the loss surface of non-linear CVNNs. Along the way, we develop a novel set of tools and techniques for analyzing the optimization landscape of CVNNs, which may be useful in other contexts.
\subsection{Motivation}
The minimum modulus principle (MMP) is a fundamental result in complex analysis which roughly states that \enquote{complex analytic functions have no spurious local minima} with respect to the modulus. In Appendix we plot nine common analytic functions on different fields $\mathbb{R}$ and $\mathbb{C}$. It can be noted that poor local minima and local minima not achieving a value of zero in the real-valued case become saddle points when functions are complex-valued. Therefore, we conjecture that all CVNNs with analytic activations have a superior optimization landscapes. Note that fully-connected multi-layer perceptrons with analytic activation functions are analytic with respect to either the input or each weight matrix. Although generally the loss function of a CVNN is not analytic because the loss is calculated by summing the modulus, the MMP still suggests the potential for better optimization landscapes in CVNNs. Non-analytic functions not only lose the MMP but are also non-differentiable. Thus, analyzing the optimization landscape of CVNNs is not trivial and is not connected with the MMP.
\subsection{Related Work}
The analysis of optimization landscape of neural networks began with linear neural networks. Baldi et al. \cite{BPK89, BL12} proved the \enquote{no spurious local minima} result for both real-valued and complex-valued shallow linear neural networks. This was later translated to deep linear neural networks by Kawaguchi \cite{NIPS2016_6112}. Work on the optimization of linear networks from other perspectives can be found elsewhere \cite{hardt2018identity,lu2017depth,pmlr-v80-laurent18a,zhang2020depth}. Although optimizing linear networks being a non-convex problem have a nice landscape, the representation power of linear network is limited as they can only fit linearly separable data.
For non-linear networks, all demonstrations of \enquote{no spurious local minima} required unrealistic assumptions or an extremely restricted parameter space. For instance, independence between weights was assumed for deep ReLU networks \cite{NIPS2016_6112}. For shallow ReLU networks, the result in \cite{wu2018no} only holds for two hidden units networks and weight vectors must to be unit-normed and orthogonal. For shallow quadratic networks, \cite{NIPS2019_9111} assumed Gaussian feature vectors; while \cite{SJL19} assumed the weight vector connecting the hidden layer and the output node must contain at least d positive entries and d negative entries. Without these unrealistic assumptions, all common non-linear real-valued networks were shown to have poor local minima \cite{yun2018small}.
The optimization landscape of neural networks under over-parametrization has also been studied \cite{pmlr-v70-nguyen17a,nguyen2018on, li2021benefit,lederer2021optimization}. Unfortunately, it has been shown that neural networks with over-parametrization can still have strict poor local minima \cite{Sharifnassab2020Bounds, 9194023}, and utilizing the set of measure zero only eliminated non-strict poor local minima, i.e. sub-optimal basins.
Some attempts from the algorithm dynamics have been made, showing that gradient descent can converge to global minima under certain conditions \cite{du2019gradient, pmlr-v97-du19c}. However, the networks in the analysis are mainly ultra-wide networks.
Finally, by analyzing the algorithm dynamics \cite{bubeck2020network} and \cite{pmlr-v32-andoni14} observed that complex-valued weights are more likely to escape from poor local minima in shallow neural networks with polynomial activations, but unfortunately they did not show that there are provably no poor local minima. However, our results draw such a conclusion for the quadratic case.
\section{Main Results}
We analyze shallow neural networks with {\it quadratic} activation functions over the fields $\mathbb{R}$ and $\mathbb{C}$, with the network structure and loss function defined as below.
Let the dataset be $\{(\textbf{x}_i, y_i)\}$ for $i=1,2,\dots, n$ with $\textbf{x}_i \in \mathbb{F}^d$ and $y_i \in \mathbb{F}$. We consider one hidden layer complex-valued neural networks with the {\it quadratic} activation $\psi$. It is in the form of
$$
\textbf{x} \mapsto \textbf{v}^T \psi(\textbf{Wx})
$$
where $\textbf{W} \in \mathbb{F}^{k\times d}, \textbf{v} \in \mathbb{F}^k$, ${v}_i \neq {0}$, \text{and } $k \geq d$. The training loss as a function of the weights ($\textbf{W}$, $\textbf{v}$) is
\begin{align*}
\mathcal{L}(\textbf{W},\textbf{v}) &= \frac{1}{2n}\sum^n_{i=1}\parallel y_i - \textbf{v}^T\psi(\textbf{W}\textbf{x}_i)\parallel^2.
\end{align*}
It is noted that $\parallel y_i - \textbf{v}^T\psi(\textbf{W}\textbf{x}_i)\parallel^2 = (y_i - \textbf{v}^T\psi(\textbf{W}\textbf{x}_i))^2$ when $\mathbb{F} = \mathbb{R}$.
We prove that one hidden layer CVNNs with {\it quadratic} activations have no spurious local minima.
\begin{theorem} Assume the dataset, loss function, and training model are defined as above with $\mathbb{F} = \mathbb{C}$. Then the training loss as a function of the weight $(\text{\normalfont \textbf{W}}$, $\textbf{\normalfont \textbf{v}})$ has no spurious local minima, i.e. all local minima are global.
\end{theorem}
Lemma 2.2 shows that Theorem 2.1 does not hold when $\mathbb{F}=\mathbb{R}$, i.e. in real-valued networks poor local minima exist. This provides evidence for a substantial difference between real-valued and complex-valued networks.
We require only the mild assumptions that $v_i \neq {0}$ and $k \geq d$. Note that if ${v}_i ={0}$ is required to achieve the global minimum, then the $i^{th}$ row of $\textbf{W}$ will be trained to be a zero row vector since there is no restriction on $\textbf{W}$. $k \geq d$ implies that the number of hidden nodes is larger than the number of input nodes, which is a commonly-used setting. In contrast, the assumption of mild over-parametrization in other works requires the number of hidden nodes to be larger than the number of input samples, i.e. $k \geq n$, which is unrealistic. Moreover, when over-parametrization is not mild, for example, in neural tangent kernel (NTK), the assumption of $k \geq O(n^4)$ was made.
We use the semi-definite property of the Wirtinger Hessian matrix to prove Theorem 2.1. Firstly, we derive a simplified expression of the Wirtinger Hessian multiplied by an arbitrary direction $\textbf{U}$. By leveraging the properties of complex numbers, we show that every local minimum ($\textbf{W}$, $\textbf{v}$) must satisfy Lemma 4.2, which makes it to be a global minimum. A full proof is in Section 4 and the appendix.
Now we provide a stark evidence of the distinction between complex-valued and real-valued networks. Under the exact same architecture and conditions as defined above, we provide examples of poor local minima when weights are real, and demonstrate that poor local minima become saddle points when weights are complex.
Lemma 2.2 describes the existence of local minima in shallow real networks as demonstrated in \cite{yun2018small}. Note that the $b_i$ terms in there can be ignored without affecting the result.
\begin{lemma}[Theorem 2, Corollary 3 in \cite{yun2018small}] Let the loss function $\mathcal{L}(\textbf{\normalfont \textbf{W}},\textbf{\normalfont \textbf{v}}) $ and the network structure be defined as above. Consider the dataset
$$
\textbf{X} =
\begin{bmatrix}
\textbf{x}_{1} & \textbf{x}_{2} & \textbf{x}_{3}
\end{bmatrix}
=
\begin{bmatrix}
1 & 0 & \frac{1}{2} \\
0 & 1 & \frac{1}{2}
\end{bmatrix} ,
\textbf{Y} =
\begin{bmatrix}
{y}_{1} & {y}_{2} & {y}_{3}
\end{bmatrix}
=
\begin{bmatrix}
0 & 0 & 1
\end{bmatrix},
$$
and the weight
$$
\bar{\textbf{\normalfont \textbf{W}} }=
\begin{bmatrix}
\bar{\textbf{w} }_{1} \\
\bar{\textbf{w} }_{2}
\end{bmatrix}
=
\begin{bmatrix}
1 & 1 \\
1 & 1
\end{bmatrix},
\bar{\textbf{\normalfont \textbf{v}}} =
\begin{bmatrix}
\bar{{v} }_{1} & \bar{{v} }_{2} \\
\end{bmatrix}
=
\begin{bmatrix}
\frac{1}{6} & \frac{1}{6}
\end{bmatrix}.
$$
When $\mathbb{F} = \mathbb{R}$ the weight is a poor local minimum of the loss function.
\end{lemma}
Notice that $\bar{{v}}_i \neq {0}$ and $k=2=d$. Now we analyze $(\bar{\textbf{W} }, \bar{\textbf{v}})$ in two different networks and in two ways. On one hand, we show that the Hessians at the point are different in real networks and in complex networks. On the other hand, we prove that there is a point with strictly less loss in an arbitrarily small neighborhood of $(\bar{\textbf{W} }, \bar{\textbf{v}})$ in the complex network.
\begin{lemma} Assume the dataset, loss function, and training model are defined as above. At $\bar{\textbf{\normalfont \textbf{W}} }$, the Hessian has no negative eigenvalue when $\mathbb{F} = \mathbb{R}$ and has both positive and negative eigenvalues when $\mathbb{F} = \mathbb{C}$.
\end{lemma}
By analyzing a specific critical point, Lemma 2.3 shows us that complex-valued weights can turn poor local minima into saddle points. As an alternative way to illustrate the insight, we provide Lemma 2.4. The proofs are provided in Appendix.
\begin{lemma} Assume the dataset, loss function, and training model are defined as above. When $\mathbb{F} = \mathbb{C}$, within an arbitrarily small neighborhood of $(\bar{\textbf{\normalfont \textbf{W}}},\bar{\textbf{\normalfont \textbf{v}}})$ there is a point $(\hat{\textbf{\normalfont \textbf{W}}},\hat{\textbf{\normalfont \textbf{v}}})$ such that $\mathcal{L}(\hat{\textbf{\normalfont \textbf{W}}},\hat{\textbf{\normalfont \textbf{v}}}) < \mathcal{L}(\bar{\textbf{\normalfont \textbf{W}}},\bar{\textbf{\normalfont \textbf{v}}}) $.
\end{lemma}
To further differentiate between analytic activated networks and non-analytic activated networks, we explore the optimization landscape of $\mathbb{C}$ReLU activated CVNNs. $\mathbb{C}$ReLU is the complex counterpart of ReLU as defined in Appendix. Unlike {\it quadratic}, {\it tanh}, and {\it sigmoid}, $\mathbb{C}$ReLU is not analytic.
\begin{theorem} Assume the dataset is $\{(\textbf{x}_i, y_i)\}$ for $i=1,2,\dots, n$ with $\textbf{x}_i \in \mathbb{R}^d$ and $y_i \in \mathbb{R}$.
The training model we consider is one hidden layer complex-valued neural networks with $\mathbb{C}$\text{\normalfont ReLU} activation. It is in the form of
$$
\textbf{x} \mapsto \textbf{v}^T \mathbb{C}\text{\normalfont ReLU}(\textbf{\normalfont \textbf{W}} \textbf{x})
$$
with $\textbf{\normalfont \textbf{W}} \in \mathbb{C}^{k\times d},\text{ and } \textbf{\normalfont \textbf{v}} \in \mathbb{C}^k$. Suppose the dataset is real-valued, $x_i$'s are distinct, the hidden layer has a width of at least 2, and cannot be fitted linearly, then the loss function
\begin{align*}
\mathcal{L}(\textbf{\normalfont \textbf{\normalfont \textbf{W}}}, \textbf{\normalfont \textbf{\normalfont \textbf{v}}}) &= \frac{1}{2n}\sum^n_{i=1}\parallel y_i - \textbf{\normalfont \textbf{v}}^T \mathbb{C}\text{\normalfont ReLU}(\textbf{\normalfont \textbf{W}} \textbf{x})\parallel^2
\end{align*}
has infinitely many spurious local minima.
\end{theorem}
Note that the conditions are slightly different than before, and the dataset is assumed to be real-valued. If poor local minima exist for real-valued datasets, they exist for complex-valued datasets. The assumptions of distinct and non-linearly separable training points and $k\geq 2$ are also mild. The proof builds on previous work by \cite{yun2018small} with some modifications to accommodate the complex case. First, we construct a point that is as good as the global minimum of a linear model. Then, we prove the point is a local minimum. The rest of the proof is the same as in the proof of Theorem 1 in \cite{yun2018small}. The authors prove that that the local minimum is spurious by showing the existence of a point with strictly less loss. The proof is provided in Appendix due to space constraints.
\section{Preliminaries}
This section provides an introduction to complex analysis and Wirtinger calculus. More definitions and lemmas are provided in Appendix.
\subsection{Notations and useful identities}
Let $\mathbb{R}$ and $\mathbb{C}$ denote the real and complex fields respectively. They share many common properties as number fields. Note that $\mathbb{R} \subseteq \mathbb{C}$ and $\mathbb{R}^{m \times n} \subseteq \mathbb{C}^{m \times n}$. Let $z = z_1 + iz_2 \in \mathbb{C}$, we use $||z || = \sqrt{z_1^2 + z_2^2} \in \mathbb{R}$ to denote its modulus and $z^* = z_1 - iz_2$ to denote its conjugate. $\mathcal{R}(\textbf{z})$ and $\mathcal{I}(\textbf{z})$ are used to denote the real and imaginary part of a complex vector $\textbf{z} \in \mathbb{C}^n$. For $\textbf{M} \in \mathbb{C}^{m \times n}$, $\textbf{M}^T$ and $\textbf{M}^*$ are the transpose and conjugate transpose of $\textbf{M}$, $\textbf{M}^C$ denotes the matrix whose entries are conjugates of entries in $\textbf{M}$, Null$(\textbf{M}) := \{\textbf{v} \ |\ \textbf{M}\textbf{v}=0 \}$ denotes the null space, and vec$(\textbf{M})$ denotes the vectorization of $\textbf{M}$. For $z \in \mathbb{C}$ and $\textbf{M} \in \mathbb{C}^{m \times n}$, we have $z^* = z^C$ and $\textbf{M}^* = (\textbf{M}^C)^T$.
\subsection{Complex functions}
A complex function $f: \mathbb{C} \mapsto \mathbb{C}$ is given by $f(z) = u(z) + iv(z)$. We can also think of $f$ as $f: \mathbb{R}^2 \mapsto \mathbb{R}^2$ where $f(x,y) = (u(x,y),v(x,y))$. A complex-valued multivariate function $f : \mathbb{C}^n \mapsto \mathbb{C}$ is given by $f(\textbf{z}) = u(\textbf{z}) + iv(\textbf{z})$ where $u(\textbf{z}), v(\textbf{z}) \in \mathbb{R}$. The most prominent function used in this work is a real-valued function with complex-valued matrix input $f : \mathbb{C}^{m\times n} \mapsto \mathbb{R}$.
A complex function is analytic if it is differentiable at every point and in it's neighborhood in the domain. An analytic function must satisfy the Cauchy Riemann equations (CRE). See Appendix for the definitions of differentiable, analytic, CRE, and more. We mention that a non-constant real-valued complex function does not satisfy the CRE and is thus not analytic.
Recall that the loss function
$$
\mathcal{L}(\textbf{W}) = \frac{1}{2n}\sum^n_{i=1}\parallel y_i - \textbf{v}^T\psi(\textbf{W}\textbf{x}_i)\parallel^2
$$
is not analytic. However, $y_i - \textbf{v}^T\psi(\textbf{W}\textbf{x}_i)$ is analytic for each $i$, which indicates that their derivatives are well defined.
\subsection{Wirtinger calculus}
Since $\mathcal{L}(\textbf{W})$ is not differentiable in the traditional sense, we require a new way of calculating the complex gradient. Many non-differentiable complex functions are differentiable in the real sense if we treat $ \mathbb{C}^n $ as $\mathbb{R}^{2n}$. $\mathcal{L}(\textbf{W})$ is one of them. Wirtinger calculus, also known as the $\mathbb{C}\mathbb{R}$-Calculus, provides a neat way of deriving the derivatives. For a differentiable complex function, Wirtinger derivatives are the same as the traditional derivatives. Using Wirtinger calculus provides not only defined derivatives that reflect a function's gradient, but also have meaningful results for the first and second derivatives. Note that critical points, positive semi-definite Hessian, and Taylor expansions all have their counterparts in Wirtinger calculus. Due to space constraints, an extended introduction of Wirtinger calculus can be found in Appendix. A comprehensive introduction can be found in \cite{KreutzDelgado2005TheCG} and \cite{BP10}.
\section{CVNNs without Poor Local Minima}
We now provide the proof sketch of Theorem 2.1. The proof partially builds on \cite{SJL19} and is completed in the appendix. We first demonstrate how to derive the Wirtinger derivatives of the loss function with respect to the weight matrix. Secondly, we state two important lemmas where the proofs are in the appendix. Then we show that the proof of Theorem 2.1 is mainly consisted of three steps. Lastly, we provide proof sketches for Lemma 2.3 and Lemma 2.4.
\subsection{Derivative calculations}
Firstly we observe that $\mathcal{L}(\textbf{W})$ is a function which maps complex input to real output, i.e. $\mathcal{L}(\textbf{W}): \mathbb{C}^{k\times d} \mapsto \mathbb{R}$, and it is not differentiable because conjugate functions do not satisfy the CRE (Cauchy-Riemann Equation). Now let $\mathcal{L}(\textbf{W}) = \frac{1}{2n} \sum^n_{i=1} \mathcal{L}_i(\textbf{W})^*\mathcal{L}_i(\textbf{W})$ where $\mathcal{L}_i(\textbf{W}): \mathbb{C}^{k\times d} \mapsto \mathbb{C}$ given by $\mathcal{L}_i(\textbf{W}) = y_i - \textbf{v}^T\psi(\textbf{W}\textbf{x}_i)$. Observe that $\mathcal{L}_i(\textbf{W})$ is complex differentiable in the traditional sense and thus $\mathcal{L}_i(\textbf{W})$ has well-defined first and second derivatives. For a fixed $i$, we let $\mathcal{G}_i(\textbf{W}) = \mathcal{L}_i(\textbf{W})^*\mathcal{L}_i(\textbf{W})$. We now show how to calculate the derivatives of $\mathcal{L}_i(\textbf{W})$ and $\mathcal{G}_i(\textbf{W})$. The derivatives of $\mathcal{L}(\textbf{W})$ follow easily by linearity.
\subsubsection{Derivative calculations of $\mathcal{L}_i(\textbf{W})$}
Now we have
$$
\mathcal{L}_i(\textbf{W}) = \textbf{v}^T\psi(\textbf{W}\textbf{x}_i) - y_i
$$
which is complex differentiable. The first derivative with respect to the $q$-th row of $\textbf{W}$ is denoted by $\nabla_{\textbf{w}_q}\mathcal{L}_i(\textbf{W})$ and we have
$$
\nabla_{\textbf{w}_q}\mathcal{L}_i(\textbf{W}) = v_q \psi^{'}(\langle \textbf{w}_q, \textbf{x}_i \rangle)\textbf{x}_i
$$
$$
\nabla_{\textbf{W}}\mathcal{L}_i(\textbf{W}) = \textbf{D}_{\textbf{v}} \psi^{'}( \textbf{W} \textbf{x}_i )\textbf{x}_i^T
$$
where $ \textbf{D}_{\textbf{v}} =$ diag$(v_1,\dots,v_k)$. And the second derivatives
$$
\frac{\partial^2}{\partial\textbf{w}^2_p} \mathcal{L}_i(\textbf{W}) = v_p \psi^{''}(\langle \textbf{w}_p, \textbf{x}_i \rangle)\textbf{x}_i\textbf{x}_i^T
$$
$$
\frac{\partial^2}{\partial\textbf{w}_p\textbf{w}_q} \mathcal{L}_i(\textbf{W}) = 0
$$
for $p \neq q$.
\subsubsection{Derivative calculations of $\mathcal{G}_i(\textbf{W})$}
By the product rule of Wirtinger calculus, we have
$$
\nabla_{\textbf{W}} \mathcal{G}_i(\textbf{W}) =\nabla_{\textbf{W}} (\mathcal{L}_i^* \mathcal{L}_i)(\textbf{W}) = \nabla_{\textbf{W}} \mathcal{L}_i^*(\textbf{W}) \mathcal{L}_i(\textbf{W}) + \nabla_{\textbf{W}} \mathcal{L}_i(\textbf{W}) \mathcal{L}_i^*(\textbf{W}) = \nabla_{\textbf{W}} \mathcal{L}_i(\textbf{W}) \mathcal{L}_i^*(\textbf{W}).
$$
$ \nabla_{\textbf{W}} \mathcal{L}_i^*(\textbf{W}) \mathcal{L}_i(\textbf{W}) = 0$ since $\mathcal{L}_i^*$ is conjugate-complex differentiable. Similarly,
$$
\nabla_{\textbf{W}^C} \mathcal{G}_i(\textbf{W}) =\nabla_{\textbf{W}^C} (\mathcal{L}_i^* \mathcal{L}_i)(\textbf{W}) = \nabla_{\textbf{W}^C} \mathcal{L}_i^*(\textbf{W}) \mathcal{L}_i(\textbf{W}) + \nabla_{\textbf{W}^C} \mathcal{L}_i(\textbf{W}) \mathcal{L}_i^*(\textbf{W}) = \nabla_{\textbf{W}^C} \mathcal{L}_i^*(\textbf{W}) \mathcal{L}_i(\textbf{W}).
$$
Based on that we have the second derivatives
$$
\nabla^2_{\textbf{W}} \mathcal{G}_i(\textbf{W}) = \nabla^2_{\textbf{W}} \mathcal{L}_i(\textbf{W}) \mathcal{L}_i^*(\textbf{W})
$$
$$
\nabla^2_{\textbf{W}^C} \mathcal{G}_i(\textbf{W}) = \nabla^2_{\textbf{W}^C} \mathcal{L}_i^*(\textbf{W}) \mathcal{L}_i(\textbf{W})
$$
$$ \nabla_{\textbf{w}_p}\nabla_{\textbf{w}_p^C} \mathcal{G}_i(\textbf{W}) = v_p^*v_p \psi^{'}(\langle \textbf{w}_p, \textbf{x}_i \rangle)^*\psi^{'}(\langle \textbf{w}_p, \textbf{x}_i \rangle) \textbf{x}_i^C \textbf{x}_i^T$$
$$ \nabla_{\textbf{w}_q}\nabla_{\textbf{w}_p^C} \mathcal{G}_i(\textbf{W}) = v_p^*v_q \psi^{'}(\langle \textbf{w}_p, \textbf{x}_i \rangle)^*\psi^{'}(\langle \textbf{w}_q, \textbf{x}_i \rangle) \textbf{x}_i^C \textbf{x}_i^T $$
$$ \nabla_{\textbf{w}_p^C}\nabla_{\textbf{w}_p} \mathcal{G}_i(\textbf{W}) = v_p^*v_p \psi^{'}(\langle \textbf{w}_p, \textbf{x}_i \rangle)^*\psi^{'}(\langle \textbf{w}_p, \textbf{x}_i \rangle) \textbf{x}_i \textbf{x}_i^*$$
$$ \nabla_{\textbf{w}_q^C}\nabla_{\textbf{w}_p} \mathcal{G}_i(\textbf{W}) = v_pv_q^* \psi^{'}(\langle \textbf{w}_p, \textbf{x}_i \rangle)\psi^{'}(\langle \textbf{w}_q, \textbf{x}_i \rangle)^* \textbf{x}_i \textbf{x}_i^*. $$
Notice that $\nabla_{\textbf{W}}\nabla_{\textbf{W}^C} \mathcal{G}_i(\textbf{W})$ and $\nabla_{\textbf{W}^C}\nabla_{\textbf{W}} \mathcal{G}_i(\textbf{W})$ are $kd \times kd$ matrices.
\subsection{Technical lemmas}
We provide some important lemmas before proving the theorem. The proofs can be found in the appendix.
\begin{lemma} If $\widetilde{\textbf{\normalfont \textbf{W}}}$ is a local minimum of $\mathcal{L}(\widetilde{\textbf{\normalfont \textbf{W}}})$,
\begin{align*}
0 \leq
(\textbf{h}^*,\textbf{h}^T)\cdot \widetilde{\nabla}^2 \mathcal{L}(\widetilde{\textbf{\normalfont \textbf{W}}}) \cdot
\begin{pmatrix}
\textbf{h} \\
\textbf{h}^C
\end{pmatrix} \in \mathbb{R}
\end{align*}
where
\begin{align*}
\widetilde{\nabla}^2 \mathcal{L}(\widetilde{\textbf{\normalfont \textbf{W}}}) =
\begin{pmatrix}
\nabla_{\textbf{\normalfont \textbf{W}}}\nabla_{\textbf{\normalfont \textbf{W}}^C} \mathcal{L}(\widetilde{\textbf{\normalfont \textbf{W}}}) & \nabla^2_{\textbf{W}^C} \mathcal{L}(\widetilde{\textbf{\normalfont \textbf{W}}})\\
\nabla^2_{\textbf{\normalfont \textbf{W}}} \mathcal{L}(\widetilde{\textbf{\normalfont \textbf{W}}}) & \nabla_{\textbf{\normalfont \textbf{W}}^C}\nabla_{\textbf{\normalfont \textbf{W}}} \mathcal{L}(\widetilde{\textbf{\normalfont \textbf{W}}})
\end{pmatrix}
\end{align*}
for all $\textbf{h} \in \mathbb{C}^{kd}$.
\end{lemma}
\begin{lemma} Any point $\widetilde{\textbf{\normalfont \textbf{W}}} \in \mathbb{C}^{k \times d}$ obeying
$$
\frac{1}{2n} \sum^n_{i=1}(\textbf{x}^T_i \widetilde{\textbf{\normalfont \textbf{W}}}^T \text{\normalfont diag(\textbf{v})} \widetilde{\textbf{\normalfont \textbf{W}}}\textbf{x}_i - y_i)^*\textbf{x}_i\textbf{x}_i^T =\frac{1}{2n}\sum_{i=1}^n\mathcal{L}_i^*(\widetilde{\textbf{\normalfont \textbf{W}}})\textbf{x}_i\textbf{x}_i^T= 0
$$
is a global optimum of the loss function
\begin{align*}
\mathcal{L}(\textbf{\normalfont \textbf{W}}) &= \frac{1}{2n}\sum^n_{i=1}\parallel \textbf{x}^T_i \textbf{\normalfont \textbf{W}}^T \text{\normalfont diag(\textbf{v})} \textbf{\normalfont \textbf{W}}\textbf{x}_i - y_i\parallel^2\\
&= \frac{1}{2n}\sum^n_{i=1}\parallel y_i - \textbf{\normalfont \textbf{v}}^T\psi(\textbf{\normalfont \textbf{W}}\textbf{x}_i)\parallel^2.
\end{align*}
\end{lemma}
\subsection{Proof sketch of Theorem 2.1}
\noindent \textbf{Step 1:} Let $(\textbf{W}, \textbf{v})$ be a local minimum. We first illustrate that \enquote{a global minimum of $\mathcal{L}$ with respect to $\textbf{W}$} is equivalent to saying \enquote{a global minimum of $\mathcal{L}$ with respect to $(\textbf{W}, \textbf{v})$}. The main reason is that unlike real-valued quadratic functions, which map $\mathbb{R}$ to $\mathbb{R}^+$, complex-valued quadratic functions map $\mathbb{C}$ to $\mathbb{C}$. Therefore any non-zero value of $v_i$ can be achieved equivalently by training the $i^{th}$ row of $\textbf{W}$ despite equipped with {\it quadratic} activations.
\noindent \textbf{Step 2:} Next we have two cases, $\textbf{W}$ being full rank and being rank-deficient. The full rank case is relatively easy. Thus we focus on the rank-deficient case. By some calculations on the matrices, we derive an expression for the Wirtinger Hessian multiplied by an arbitrary direction $\textbf{U}=\textbf{a}\textbf{b}^T$, $\mathcal{H}= \mathcal{R}(\frac{1}{n}(\textbf{a}^T\textbf{D}_{\textbf{v}}\textbf{a})\textbf{b}^T\left(\sum_{i=1}^n\mathcal{L}_i^*(\textbf{W})\textbf{x}_i\textbf{x}_i^T\right)\textbf{b})$ where $\textbf{b}$ is an arbitrary vector. We argue that $\textbf{a}^T\textbf{D}_{\textbf{v}}\textbf{a}$ can be assumed non-zero here. By Lemma 4.1 we have $\mathcal{H} \geq 0$.
\noindent \textbf{Step 3:} Now we make use of $\textbf{b}$ and the properties of complex numbers to prove $\sum_{i=1}^n\mathcal{L}_i^*(\textbf{W})\textbf{x}_i\textbf{x}_i^T = 0$, which concludes the proof by Lemma 4.2. Note that $\mathcal{H} = \mathcal{R}(z_1 \cdot z_2)$ where $z_1, z_2 \in \mathbb{C}$. We prove by contradiction. Suppose $\sum_{i=1}^n\mathcal{L}_i^*(\textbf{W})\textbf{x}_i\textbf{x}_i^T \neq 0$, we can show that $\mathcal{H} < 0$, which contradicts to Lemma 4.1. The main strategy is that we can \enquote{manipulate} the positive and negative signs of $z_2$, and observe that $\mathcal{H}$ is the real part of a complex number multiplied by $z_2$. For example, if $z_1 = 2+ 3i$, then taking $z_2 =1+4i$ makes $\mathcal{R}(z_1 \cdot z_2) = \mathcal{R}(-10+11i) = -10 < 0$.
\subsection{Complex-valued weights turn local minima into saddle points}
In this section, we provide proof sketches for Lemma 2.3 and Lemma 2.4.
Recall that the weight
$$
(\bar{\textbf{\normalfont \textbf{W}} }, \bar{\textbf{\normalfont \textbf{v}}})=
(
\begin{bmatrix}
1 & 1 \\
1 & 1
\end{bmatrix},
\begin{bmatrix}
\frac{1}{6} & \frac{1}{6}
\end{bmatrix})
$$
is a local minimum of the real-valued network for the given dataset. In the proof of Lemma 2.3, we derive the Hessian matrices at $\bar{\textbf{\normalfont \textbf{W}} }$ for real-valued and complex-valued networks and calculate their eigenvalues. We denote the Hessian matrix by $\mathcal{H}^{\mathbb{R}}_{\bar{\textbf{W}}}$ for the real-valued case and $\mathcal{H}^{\mathbb{C}}_{\bar{\textbf{W}}}$ for the complex-valued case. Note that $\mathbb{C}^{n}$ can be treated as $\mathbb{R}^{2n}$ in some sense. Therefore $\mathcal{H}^{\mathbb{C}}_{\bar{\textbf{W}}}$ is four times the size of $\mathcal{H}^{\mathbb{R}}_{\bar{\textbf{W}}}$. By similar derivative calculations as before and substituting the weight and dataset given in Lemma 2.2 into the expression, we have
$$
\mathcal{H}^{\mathbb{R}}_{\bar{\textbf{W}}} =
\begin{bmatrix}
\frac{7}{108} & \frac{-1}{108} & \frac{5}{108} & \frac{1}{108} \\
\frac{-1}{108} & \frac{7}{108} & \frac{1}{108} & \frac{5}{108} \\
\frac{5}{108} & \frac{1}{108} & \frac{7}{108} & \frac{-1}{108} \\
\frac{1}{108} & \frac{5}{108} & \frac{-1}{108} & \frac{7}{108}
\end{bmatrix}
$$
which has no negative eigenvalues, and
$$
\mathcal{H}^{\mathbb{C}}_{\bar{\textbf{W}}} =
\begin{bmatrix}
\mathcal{H}_1 & \mathcal{H}_1& \mathcal{H}_2 & \mathcal{H}_3 \\
\mathcal{H}_1 & \mathcal{H}_1 & \mathcal{H}_3 & \mathcal{H}_2 \\
\mathcal{H}_2 & \mathcal{H}_3 & \mathcal{H}_1 & \mathcal{H}_1 \\
\mathcal{H}_3 & \mathcal{H}_2 & \mathcal{H}_1 & \mathcal{H}_1
\end{bmatrix}
$$
where
$$
\mathcal{H}_1 =
\begin{bmatrix}
\frac{5}{216} & \frac{1}{216} \\
\frac{1}{216} & \frac{5}{216}
\end{bmatrix},
\mathcal{H}_2 =
\begin{bmatrix}
\frac{1}{108} & -\frac{1}{108} \\
-\frac{1}{108} & \frac{1}{108}
\end{bmatrix}
$$
and $\mathcal{H}_3$ is a zero matrix.
It can be verified that $\mathcal{H}^{\mathbb{C}}_{\bar{\textbf{W}}}$ has both negative and positive eigenvalues. This illustrates that complex-valued weights turn local minima into saddle points.
As an alternative proof, in the proof of Lemma 2.4, we make the following permutation on $\bar{\textbf{\normalfont \textbf{W}} }$,
$$
\hat{\textbf{W}} =
\begin{bmatrix}
1-\frac{1}{10^N} & 1+ \frac{i}{10^N} \\
1-\frac{1}{10^N} & 1+ \frac{i}{10^N}
\end{bmatrix}
$$ for an arbitrarily large $N\in \mathbb{N}^+$,
and by simple calculations we show that $ \mathcal{L} ( \hat{\textbf{W}},\bar{\textbf{v}}) < \frac{1}{9} = \mathcal{L}(\bar{\textbf{W}},\bar{\textbf{v}})$.
\bibliographystyle{alpha}
|
\section{Introduction}
\label{sec:introduction}
\subsection{Similarity and conceptual problems in modern physics}
\label{sub:similarity_and_conceptual_problems_in_modern_physics}
In Henri Poincar\'e's magistral \emph{Science and Method} \citep{poincare2003science} we're invited to image a universe like our own that is identical in every respect except that it is a thousand times larger. Such a universe, he argues, would be indiscernible from our own because any reference standard that could be used to measure the length of a body would have grown in exact proportion the length of that body, and therefore any measurement using this standard would be unaffected. He calls such a transformation a \emph{similarity} in reference to the same geometric symmetry discussed as early as Euclid.
Poincar\'e's argument for empirical indiscernibility under similarity can be applied more generally to measurements of temporal duration. If duration is measured using relative changes within a system --- say by recording the number of oscillations of the pendulum of a clock --- then, by Poincar\'e's argument, the transformed system should be indiscernible from the original, and therefore no change in duration can be measured.
But compared to length, representations of duration are made considerably more subtle by dynamical considerations. Useful standards of duration are defined by their ability to bring the dynamical laws into a particular form. For example, absolute time in Newtonian mechanics is defined in terms of inertial motion, which in turn is determined by the form of the dynamical laws. The precise way that a similarity transformation acts on the representations of a dynamical system is therefore affected by how the temporal standards depend on the laws. This requires a more comprehensive approach.
The most important consideration in defining such an approach is the way in which the form of the laws affects the transformation properties of the velocities. In Lagrangian theories, length and time standards are related by the conventions used to define the unit of Hamilton's principal function. Heuristically speaking, this unit defines a standard of angular momentum that can be used to convert lengths to velocities using some convention for inertial mass. We will call similarity transformations that take into account the dynamical considerations involved in defining velocity \emph{dynamical similarities} because of their relation to the transformations of the same name first studied in the context of the Kepler problem (see Section~\ref{sub:DS as subsystem} for the explicit connection).
The primary goal of this paper is to give a detailed analysis of dynamical similarity. We will show that, while often recognised in different guises, dynamical similarity is an under-appreciated symmetry in many physical systems with significant implications for major conceptual problems in modern physics. To achieve this goal, we need a philosophical framework for understanding symmetry that is general enough to handle the peculiar features of dynamical similarity. In developing such a framework, we accomplish several noteworthy secondary goals. In particular, we give an account of symmetry that identifies surplus structure\footnote{ In this paper it will be convenient to define surplus structure as eliminable structure according to one of the senses of surplus structure discussed in \cite{nguyen2020surplus,bradley2020representational}. This is because, in the examples we will discuss, dynamically similar surplus structure will be explicitly eliminable. It is, however, possible to rephrase our discussion of dynamical similarity by taking surplus structure to be a representational redundancy. This will be shown in future work. } using the freely specifiable initial data of a theory required for empirical adequacy. We then use this account to resolve certain tensions between epistemic and dynamic definitions of symmetry in various well-studied cases including the Galileo ship and the Kepler problem.
We will then apply our framework to a cosmological setting and find that dynamical similarity relates surplus structure. We will then review previous technical results\footnote{ Mostly notably those developed in \cite{Koslowski:2016hds} and \cite{Sloan:2019wrz}. } that show how many kinds of singularities in general relativity --- including those of certain important big bang and black hole solutions --- are removed when dynamically-similar structure is taken to be surplus. Finally, we will see that a novel explanation for the arrow of time arises for such systems.
Central to our arguments is our analysis of symmetry. We will develop a principle, which we will call the \emph{Principle of Essential and Sufficient Autonomy (PESA)}, that we will use to identify the observable structures of a theory. While the PESA can be applied to general symmetries, it will be introduced here in order to accommodate certain non-standard formal properties of dynamical similarity that prevent an analysis in terms of standard techniques. The PESA gives a clear prescription for formally characterising the models that a particular theory has hypothesised to be empirically identical.\footnote{ We use the word `model' in the philosopher's sense as the representation of a possibility. This usually translates, in physics jargon, to an `history' or sometimes a `solution' of a theory. } If such a theory is found to be empirically adequate, then the PESA suggests a strong norm for classifying as surplus any structures that vary between identified models and as observable any structures that remain unchanged. Such an identification of models can also be motived by the Principle of the Identity of Indiscernibles (PII) since the relevant models agree with respect to all empirically accessible properties. The two-step application of hypothesising observable structure using the PESA and then checking for empirical adequacy sidesteps tensions between dynamic and epistemic definitions of symmetry discussed at length in the literature.\footnote{ See for example \cite{vanFraassen:sym_and_surplus,roberts2008puzzle,healey2009perfect,belot:sym_and_equiv,dasgupta2016symmetry,pittphilsci16622,belot2018fifty} and the more extended discussions in Section~\ref{sub:dynamics_versus_epistemology}. } Our proposal separates the formal properties of symmetry (using the PESA) from contingent facts about the world (by checking for empirical adequacy). More speficially, the PESA gives precise dynamical criteria that can be used to distinguish observable and surplus structure by considering the number of independent input data required to solve the evolution equations. In this way our novel proposal represents a distinct alternative to category-theoretic (cf. \cite{weatherall2016understanding,nguyen2020surplus,bradley2020representational}) definitions of surplus structure or other recent proposals for analysing symmetry (cf. \cite{CAULTON2015153,dewar2019sophistication,read_martins_sophistry}) that do not provide specific dynamical criteria that can be applied to dynamical similarity. To test our formalism, we will apply it to some well-studied physical examples\footnote{ These cover many of the prominent situations discussed in \cite{kosso2000empirical,brading2004gauge,greaves2014empirical,belot:sym_and_equiv,pittphilsci16623} and \cite{saunders2013rethinking,knox2014newtonian}. See Section~\ref{sec:assessing_indiscernibility_examples} for details. }. We reproduce orthodox expectations regarding the epistemic considerations of the symmetries of those systems. Passing these tests give us confidence that our proposal can be reliably applied to the novel cosmological applications of dynamical similarity considerer in this paper.
After establishing the reliability of our framework we will turn attention to dynamical similarity. Following and expanding upon \cite{Sloan:2018lim}, we will give a general definition of dynamical similarity, illustrate how it is importantly different from other symmetries in physics, and develop a general scheme for eliminating surplus structure due to dynamical similarity in different contexts. In applying the PESA in light of the empirical adequacy of different physical theories, we will find that when dynamical similarities act on approximately (but not completely) isolated subsystems of the universe, the removal of dynamically similar structure is \emph{not} justified. This applies to many well-studied cases in physics including the so-called Runge--Lenz symmetries of the Kepler problem (see \cite{belot:sym_and_equiv}). Here, our analysis is consistent with standard interpretations. But when dynamical similarities act on the universe, our formalism prescribes that the removal of dynamically similar structure \emph{is} justified. This applies, in particular, to modern theories of cosmology.
To be more concrete about the implications of removing dynamically similar structure, we will develop a conceptual framework, leveraged on previous formal results, in which pathologies in the evolution equations of certain cosmological systems will be seen to be due to the surplus structure introduced by dynamical similarity. These pathologies can then be removed if the symmetry is understood as relating indiscernible states of the universe. This leads to a form of singularity resolution in a classical theory of gravity. Moreover, mathematical features unique to dynamical similarity will be shown to define certain privileged structures --- namely \emph{attractors} and so called \emph{Janus points} (to be defined in Section~\ref{sub:friction_janus_points_and_attractors}) --- on the space of indiscernible models of cosmological theories. In addition, the dynamics on the space of discernible states will be seen to have a fundamentally \emph{frictional} character. The privileged structures and the friction-like behaviour establish a preferred arrow of time for observers that find themselves in a state near an attractor. This suggests a novel solution to the problem of the arrow of time similar to that proposed in \cite{Barbour_2014} and \cite{barbour:janus}. Finally, the central role played by the standard of angular momentum in Lagrangian theories with dynamical similarity suggests exciting implications for the foundations of statistical and quantum mechanics. Removing the ``extra mathematical hooks'' of dynamical similarity thus has significant implications for many empirical and conceptual problems in modern physics.
\subsection{The dynamics of similarity}
\label{sub:dyn of sim}
In the previous section we described how a similarity transformation should act on the representations of a dynamical system. We highlighted the subtle role of the dynamical laws in determining the transformation properties of the velocities. We will now make these comments more precise by giving a formal definition of dynamical similarity. This definition will provide the starting point for the analysis of the remainder of the paper.
Our definition must involve a statement of the laws themselves because these define the convenient temporal standards for the theory. It will thus be helpful in our analysis to restrict to a general class of dynamical systems called \emph{Lagrangian systems}. A Lagrangian system is a system whose dynamically possible models (DPMs), $\gamma_\text{DMP}$, are those curves on the state space of the theory that are stationary points of an action functional, $S[\gamma]$, such that $\delta S[\gamma]|_{\gamma_\text{DPM}} = 0$. Given such a system, we define a \emph{dynamical similarity} as a transformation that rescales the action functional by a constant:\footnote{ More specifically, dynamical similarities will be maps on certain representational structures (see Section~\ref{sub:general_definition}) of the theory that induce a pullback of $S$ that is a rescaling of it. }
\begin{equation}\label{eq:DS gen}
S[\gamma] \to c S[\gamma']\,.
\end{equation}
Under this definition, if $\gamma$ is a DPM then $\gamma'$ is also a DPM because \eqref{eq:DS gen} preserves the stationary condition of $S[\gamma]$. Dynamical similarities, when they exist, therefore map DPMs to DPMs, and consequently preserve the condition for satisfaction of the laws.
To understand how such transformations match the intuitions of the previous section, consider that the units of $S$ are those of angular momentum, and therefore that the effect of a dynamical similarity is to actively rescale the global standard of angular momentum for the system.\footnote{ Throughout this paper, we will use an \emph{active} notion of dynamical similarity whereby the dynamical similarities will be thought of as rescaling the standard of angular momentum of the system. } Different choices of $c$ therefore represent different ratios between the standard of length and the standard of velocity. Using these transformations, one can then determine how to rescale the standards of length and time to produce a similarity transformation that preserves the condition for satisfaction of the laws. In Section~\ref{sub:general_definition}, we will see more explicitly how to do this in general (e.g., Equation~\ref{eq:local S}) and in Sections~\ref{eq:kepler universal} we treat two examples that help to build intuition for dynamical similarity and the situations in which it should be expected to have important implications for physics.
We end these introductory remarks with two important comments that will figure prominently in our discussions below.
First, dynamical similarities are somewhat unconventional symmetries in that they do not preserve the formal relationships between position and momentum.\footnote{ I.e., dynamical similarities do not preserve symplectic structure. These features will be expanded upon in Section~\ref{sub:a_non_symplectic_symmetry}.} This distinguishes them from other well-known gauge symmetries of physics such as those of electromagnetism, the standard model of particle physics and general relativity. Because of these differences, dynamical similarities cannot be treated using standard physics tools that assume symmetries of a standard form.\footnote{ E.g., Noether's theorems and Dirac's algorithm. The PESA is compared with Dirac's algorithm in Section~\ref{sub:the_pesa}. } The PESA, however, will apply to dynamical similarity by construction. As we will see, these peculiarities are what leads to so many of the interesting and noteworthy features of dynamical similarity.
Second, as we will see in Section~\ref{sub:DS as subsystem}, well-known applications of dynamical similarity concern subsystems of the universe and take dynamical similarities to transform between \emph{discernible} models of a theory. As an example, dynamical similarities transform between distinct planetary orbits consistent with Kepler's third law. It is therefore important that our proposal reproduce these intuitions while also justifying our novel treatment of cosmology. We will see in Sections~\ref{sub:DS as subsystem} and \ref{sub:friedmann_lema^itre_robertson_walker_cosmology} that this is indeed the case.
Third, standard practice in cosmology recognises that \emph{similarity} transformations relate indiscernible states of the world. It is an uncontroversial fact that the observational data in cosmology are independent of the convention used for fixing spatial size.\footnote{ To be more precise, the observational data is independent of the initial value of the scalar factor. } What is not generally recognised, however, are the \emph{dynamical} implications of this fact. In particular, the surplus structure associated with \emph{dynamical} similarity is usually retained within the formalism. The remarkable consequences for cosmology that we will explore in this paper are a direct consequence of explicitly treating dynamical similarity as one would any symmetry that relates indiscernible states of the world.
\subsection{Prospectus}
\label{sub:prospectus}
In this paper we develop a general framework for symmetries that distinguishes the observable and surplus structures of a theory by using the minimal freely specifiable initial data for the theory that is necessary to achieve empirical adequacy. This framework is then applied to well-studied examples including Galilean invariance and the symmetries of the Kepler problem. We find that our framework gives a precise dynamical criterion for identifying the observables of those systems, and that those observables agree with epistemic expectations in terms of a standard universal--subsystem distinction. We then apply our framework to dynamical similarity. First we give a general definition of dynamical similarity. Then we show, with the help of some previous results, how the dynamics of our observables leads to singularity resolution and the emergence of an arrow of time in cosmology.
Our general procedure will be developed in Section~\ref{sec:PESA}. First, we will set up a framework for symmetries in Section~\ref{sub:a_framework_for_symmetry} that distinguishes two kinds of symmetry: \emph{universal symmetries} that act on the universe as a whole and \emph{subsystem symmetries} that act on isolated subsystems. Knowing which context --- universal or subsystem --- applies to any given theory is a subtle matter. We will propose the PESA, which we define in Section~\ref{sub:the_pesa}, to provide clear criteria for identifying the surplus structure of a theory, and therefore clearly distinguishing between these contexts.
We will then test our proposal in Section~\ref{sec:assessing_indiscernibility_examples}. First we explore the well-studied Galilean boosts as examples of universal (Section~\ref{sec:universal Gal}) and subsystem (Section~\ref{sec:Gal ship}) symmetries. In Section~\ref{sub:DS as subsystem} we then explore the Kepler problem, and see that Kepler's third law is consistent with dynamical similarity being a subsystem symmetry. In all these examples, the PESA is seen to reproduce orthodox views. In Section~\ref{eq:kepler universal} we develop intuition for dynamical similarity as a universal symmetry by considering hypothetical models related to the Kepler problem.
The remaining sections of the paper will be devoted to giving a more precise definition of dynamical similarity, which expands upon the formalism presented in \cite{Sloan:2018lim}, and giving some interesting applications to cosmology. We will highlight the main representational features of dynamical similarity in Section~\ref{sub:general_definition}, and outline a general procedure for writing a dynamically-similar description of Lagrangian systems in Section~\ref{sub:geometric_structure}. In Section~\ref{sub:friction_janus_points_and_attractors} we will then describe how some of the interesting features resulting from dynamical similarity arise.
Finally, in Section~\ref{sec:applications} we review some results of \cite{Koslowski:2016hds} and \cite{Sloan:2019wrz} and show how our framework can be used to address several important conceptual problems in modern physics. In Section~\ref{sub:cosmo probs}, we show how removing the surplus structure due to dynamical similarity can resolve the initial singularity (i.e., the \emph{big bang}) in an important class of cosmological theories. And in Section~\ref{sub:sing res in GR} we describe how these results may generalise to a larger class of singularities in general relativity. As a final application in Section~\ref{sub:aot} we discuss how dynamical similarity leads to a new proposal from solving the problem of the arrow of time. In Section~\ref{sec:conclusions} we conclude by summarising and discussing some potential applications to the foundations of thermodynamics and quantum mechanics.
\section{The Principle of Essential and Sufficient Autonomy}
\label{sec:PESA}
\epigraph{Methods for removing excess structure are much more than mopping up procedures. They are not something merely to be done \emph{after} our representations have been crafted, like portraitists erasing stray pencil marks, or sculptors removing extra clay. Methods for removing excess structure are the very heart of theorizing.}{ \cite[Ch. 23, p.390]{vanFraassen:sym_and_surplus}. }
\subsection{The very heart of theorizing}
\label{sub:a_framework_for_symmetry}
In this section we will introduce a powerful normative principle, central to our analysis, that we will call the \emph{Principle of Essential and Sufficient Autonomy (PESA)}. This principle identifies the representational structures of a theory that are observable, and distinguishes these from those that are surplus. It can thus be used to identify and remove excess structure. The PESA is closely related to what Barbour has called \emph{Poincar\'e's principle}\footnote{ The principle was introduced in \cite{Barbour_Bertotti}. See Section~5 of \cite{barbour2010definition} for the historical connection with Poincar\'e. } but differs importantly in the way that it evades certain ontological subtleties that we will outline below.
As we will see, the virtues of the PESA are not necessary obvious when contemplating a theory purely at the formal level. Rather, the PESA's appeal is most apparent during the process of theory-construction. Within this process, the PESA is an extremely useful conceptual and formal tool for developing and refining a particular theory.
In \cite{vanFraassen:sym_and_surplus}, theory-construction is described as a two-stage process:
\begin{quote}
There are two stages in theory construction. The first is to generate a set of models rich enough to embed the phenomena, the second is to attempt to simplify those models by exposing and eliminating excess structure. Continuing in this way the structure of the models is pared down, being careful not to jeopardize their capacity to embed the phenomena. [p.390]
\end{quote}
We find this characterisation compelling because it highlights the delicate balance that must be achieved between wanting to reduce the amount of surplus structure and retaining the ability to describe the phenomena. The PESA aims to find this balance by imposing precise mathematical criteria. But before we can state the principle explicitly, it will first be necessary to address some known problems related to defining symmetry.
\subsubsection{Dynamics versus epistemology}
\label{sub:dynamics_versus_epistemology}
The broad topic of `symmetry' has been extensively discussed in the philosophical literature. An important question in these discussions is what, if any, metaphysical or epistemological conclusions can be drawn from symmetries. Perhaps unsurprisingly, the answers provided to this question depend upon the way that symmetries are defined. Defining symmetry involves some well-known difficulties that originate from a tension between the dynamical and epistemic aspects of symmetry. This tension can be used to divide definitions of symmetry into those that make use of the dynamical properties of a theory or those that make use of epistemic considerations.\footnote{ An alternative way to characterise the difference between `dynamic' and `epistemic' definitions of symmetry is in terms of `theoretical' versus `empirical' symmetries as defined in \cite{healey2009perfect}. }
\emph{Dynamic} definitions of symmetry define symmetries as structure-preserving transformations between a theory's DPMs. Such definitions have proved useful, for example, when the dynamic symmetries are compared with the space-time symmetries of a theory according to the analysis of \cite[\S3.4]{earman1989world}. In such approaches the formal invariances of the underlying mathematical structures of a theory are believed to support reliable inferences about the ontological properties of the world.
\emph{Epistemic} definitions of symmetry define symmetries as transformations between a theory's indiscernible models. In such definitions, it is an epistemic consideration --- namely, indiscernibility --- that is used as the primary criterion for defining a symmetry transformation. Such attempts are exemplified by the approach of \cite{dasgupta2016symmetry} linking good notions of symmetry to undetectability.
While some authors (e.g., \cite{pittphilsci16622}) have framed the tension between dynamic and epistemic definitions as a dialectic between opposing camps, it is easy to find authors who find difficulties with both views. These difficulties arise through the existence of examples where dynamic and epistemic considerations lead to incompatible conclusions. Using such examples, \citet[\S1]{belot:sym_and_equiv} shows how the ``unhappy'' combination of dynamic and epistemic considerations leads to a ``recipe for disaster'' when trying to provide a precise definition of symmetry. Similar examples can be found in \cite{roberts2008puzzle}, where the tension is framed as a ``puzzle.''
A vivid way to illustrate the origin of this tension is to consider systems where symmetries are seen to have \emph{direct empirical consequences}. Examples where this occurs are given in \cite{kosso2000empirical} and further details have been discussed in \cite{brading2004gauge,greaves2014empirical,friederich2015symmetry,pittphilsci16981}. In such cases, symmetry transformations that act uniformly on the representations of a subsystem can fit the dynamical definition of a symmetry --- in that they preserve the condition for satisfaction of the laws --- but not the epistemic definition --- in that they relate empirically distinguishable states.
Perhaps the most well-studied example is that of \emph{Galileo's ship}, which will be discussed in detail in Section~\ref{sec:Gal ship}. In this example, a ship is moving almost frictionlessly with respect to a shore. Galileo boost transformations of the ship that leave the shore fixed preserve the condition of satisfaction of the laws but relate \emph{discernible} states of the system. We will have much more to say about symmetries of this kind later. But for now we note that they are exemplars for illustrating how dynamic definitions of symmetry do not always agree with purely epistemic definitions.
To sidestep difficulties of this kind, we will attempt to avoid metaphysical considerations as much as possible in our analysis of symmetry. Instead, we will focus on the ability, or inability, of a theory to describe phenomena. We will try to draw a clear separation between the formal definition of a theory and the contingent fact of whether that theory can usefully describe the world. By insisting on this separation, we can refrain from making ontological commitments based on the purely formal properties of a theory that will put dynamical definitions of symmetry in conflict with epistemic considerations.
Because of this, we will not be able to draw strong metaphysical conclusions from fromal definitions of symmetry alone. But this does not mean that nothing useful can be learned from our analysis of symmetry. Quite to the contrary, we will be able to give precise dynamical criteria, in terms of the amount of independently specifiable initial data of a theory, that will exactly indicate the empirical structure that a particular theory hypothesises. Moreover, these criteria will be used to naturally distinguish different theoretical contexts where the formal structures of a theory bear different hypothesised relations to the phenomena. This leads to a rich view of symmetry that can be usefully applied to a multitude of different physical systems.
\subsubsection{Kinds of symmetry}
\label{sub:contexts_of_symmetry}
Given the considerations of the previous section, we will find it useful, on the one hand, to define symmetries in a more `dynamic' tradition in terms of formal transformations that preserve certain representational structures of a theory. At a minimum, we will require that a symmetry preserve a theory's DPMs. However, symmetries can, and usually do, preserve additional structures. When they do we can characterise symmetries according to those structures (e.g., `space-time symmetries' preserve certain space-time structure).\footnote{ A framework for analysing symmetry in this way for the purposes of understanding temporal symmetry in quantum gravity will be presented in \cite{time_regained}. This is in the same spirit as category-theoretic definitions of symmetry in terms of structure-preserving isomorphisms such as those discussed in \cite{weatherall2016understanding,nguyen2020surplus,bradley2020representational}. }
We will also find it useful, on the other hand, to appeal to a more `epistemic' tradition by distinguishing different theoretical contexts that characterise the relationship between the formally defined symmetries of a theory and the empirical properties of the world. In this way, we will allow for symmetries to be transformations both between surplus structure and certain measurable quantities. Guided by the examples discussed above where symmetries are seen to have direct empirical consequences, we will define two kinds of symmetry that will distinguish different theoretical contexts where symmetries will relate both discernible and indiscernible states.
To make our language more precise, we will first gives some definitions. We will call \emph{appearances} structures in the world that can be described by experimental and measurement reports.\footnote{ Although our language roughly follows \cite[p.64]{van1980scientific}, the distinction we wish to draw between `surplus' and `observable' structure should not be confused with the more subtle and complex observable-unobservable distinction.}. We then call \emph{observable structure}, or simply \emph{observables}, the (suitably interpreted) representational substructures of the models of a theory that are candidates for describing observable phenomena. In particular, the observables of a theory are those structures that a theory hypothesises can be mapped bijectively to appearances. Conversely, we will call \emph{surplus structure} any representational structures in the theory that cannot be mapped bijectively to appearances. Finally, we say that a theory is \emph{empirically adequate} if the theory has one model whose observable structures are isomorphic to all appearances.
We pause briefly to clarify a potential ambiguity in our terminology. We have decided to adopt the term `observable' to refer to the representational structures that a particular theory hypothesises to be isomorphic to the appearances when the theory is empirically adequate. This is importantly \emph{not} the sense of `observable' used by Van~Fraassen to denote structures discernible by the senses. Instead, it is the sense of `observable' used by practising physicists when describing the gauge-invariant representational structures of a theory. As we will see in detail in Section~\ref{sub:the_pesa} below, the notion of observable we have in mind will be determined explicitly by the form of the laws of the theory by the way in which those laws are well-posed. As such, our usage is perhaps best understood as a generalisation of the notion of \emph{Dirac observable}, which is made in terms of canonical constraints implied by the action principle of the theory. This point is emphasised in Section~1.5.2 of the standard text of \cite{henneaux1992quantization}:\footnote{ Note that they refer to \emph{Dirac observables} as \emph{classical observables}. }
\begin{quote}
Indeed, no information other than the action principle was required in the determination and classification of the constraints. Thus, the action itself enables one to decide what are the observables.
\end{quote}
Our definition is in this spirit but does not rely on canonical constraints and, as we will see, will be applicable to more general contexts such as that of dynamical similarity.
With this usage in mind, we can use our language to define a \emph{universal symmetry} as a symmetry that leaves invariant all the observable structures of a theory. Canonical examples of universal symmetries are the bulk diffeomorphism symmetries in general relativity and gauge symmetries of classical electromagnetism and Yang--Mills theory.\footnote{ When gauge-conserving boundary conditions have been imposed (cf. \cite{belot2018fifty}). } It is important to emphasise that this definition is independent of whether a theory with a universal symmetry is empirically adequate or not. We can then define a \emph{universal context} in which the empirical adequacy of a theory with a universal symmetry is being investigated. It is only once this context has been specified that different prescriptive ontological commitments, pertaining for example to surplus structure, can be assessed.
We define a \emph{subsystem symmetry} as a symmetry that leaves invariant some, but not all, observable structures of a theory.\footnote{ This definition matches the formal definition of a \emph{regional symmetry} as defined in \cite{pittphilsci16981}, but where the structures that shift the region are observable in the theory in question. We defer to that paper for a more technical definition. } The subsystem in question is simply the set of observable structures on which the symmetry acts non-trivially. A canonical example of a subsystem symmetry is the boost symmetry of a Galileo ship. We can then define a \emph{subsystem context} in which the empirical adequacy of a theory with a subsystem symmetry is being investigated. We note that the terms `universal' and `subsystem' do not reference the domain of applicability of the theory itself. Thus, universal symmetries can arise in theories that are themselves not considered to be universal (e.g., the $U(1)$ invariance of electromagnetism).
\subsubsection{Two problems for a theory}
\label{sub:two_deficiencies_of_a_theory}
In this section, we will try to motivate the PESA by highlighting two important ways in which a theory can be problematic in being able to describe phenomena. The PESA will then be designed to identify these deficiencies. This identification will be a subtle task in the cosmological examples we will consider later. To begin, we must first be clear about what we will understand a theory to be.
We will understand a theory to broadly constitute two things: i) a statement of its DPMs by a specification of its basic mathematical structures and laws, and ii) a collections of hypotheses about how these models are related to structures in the world. These hypotheses include assumptions about measurement procedures, and must provide maps between observables and appearances.\footnote{ In the \emph{DEKI} account of representation developed in \cite{frigg2020modelling}, these maps correspond to a \emph{key} for how interpreted representations (i.e., our `observables') are used to exemplify properties that can be imputed onto the target system. } In this way a theory, on our view, must itself specify all of its observables. Given this way of understanding a theory, one \emph{first} hypothesises about which quantities are observables, and \emph{then} checks for empirically adequacy as a test of these hypotheses. Because of this, whether a particular symmetry is universal or subsystem could depend simply on which structures the theory identifies as observable.
There are two ways that a theory's observables can be inadequate in describing the phenomena of the world. In the first case, a theory can have too \emph{few} observables to describe all the appearances. Such a theory cannot be empirically adequate and must either be rejected or extended. This situation is usually straightforward to diagnose because there will be phenomena that the theory can't describe.
In the second case, a theory can have too \emph{many} observables than are strictly necessary for describing the appearances. Here, if the system is deterministic then the dynamics of the observables are underdetermined by the phenomena. The evolution equations of the theory therefore cannot be solved in terms of the observables in question.
It might seem that such underdetermination would be obvious to diagnose. But because it is often possible to fix the extraneous observables by making various auxiliary assumptions (often implicitly), this second case is usually more subtle to diagnose than the first. In this case, a theory may be believed to be empirically adequate --- i.e., it may be believed to have an isomorphism between observables and appearances --- even though it posits unnecessary observables, and therefore the map between observables and appearances is in fact many-to-one. Such a misdiagnosis can lead to faulty conceptualising that misinterprets surplus for observable structure. We will see that this is precisely what happens with dynamically-similar structure in the applications of Section~\ref{sec:applications}. In those cases, the PESA will be our primary guide towards deciding the appropriate set of observables for describing the phenomena.
\subsection{The PESA}
\label{sub:the_pesa}
We have now compiled the necessary ingredients to state the PESA. Consider a theory whose representations involve a general state $\psi(t,b^I(t))$ at time $t$ that is completely specified by the time-dependent structures $b^I(t)$, which form a generating set for some algebra $\mathcal B$. Consider then that the time evolution of the state is determined from the $b^I(t)$ according to equations of the form
\begin{equation}\label{eq:red ev}
\dot b^I(t) = f^I(b^I, t)\,,
\end{equation}
for some functions $f^I$ for which we won't specify any significant restrictions. This form of the laws is very general as any higher order system can be re-written as a first order system by adding structures to the set $B = \{b^I\}$. In general, however, the equations \eqref{eq:red ev} will not be well-posed in terms of initial data that can be determined by the appearances alone. By `well-posed' we mean that a specification of arbitrary $b^i(t_0)$ at some time $t_0$ uniquely determines the solution $b^i(t)$ at some later time $t$ for all $t>t_0$ relevant to the theory. The PESA makes use of this fact to give a dynamic definition of a theory's observables that can be used to identify surplus structure.
Let $\mathcal P$ denote the space of all possible distinct appearances at a time $t$ hypothesised by a theory. The PESA states that the \emph{observable algebra} $\mathcal A$ of the system should be generated by the subset $A = \{a^i\} \subseteq B$ such that the restriction of \eqref{eq:red ev} to $A$, i.e.,
\begin{equation}\label{eq:gen ev}
\dot a^i(t) = \tilde f^i(a^i, t)\,,
\end{equation}
is well posed and that, for the maps $\mathcal O$ of the form $\mathcal O: \mathcal A \to \mathcal P$,\footnote{ The representational structures $a^i$ are assumed to be equipped with a suitable interpretation. When this is done, the maps $\mathcal O$ can be roughly thought of as providing a \emph{key} as defined in the DEKI account of representation developed in \cite{frigg2020modelling}. }
\begin{enumerate}[(i)]
\item The equations \eqref{eq:gen ev} are \emph{autonomous} in $A$.\label{crit:autonomy}
\item The maps $\mathcal O:\mathcal A \to \mathcal P$ are onto; i.e., the algebra $\mathcal A$ has \emph{sufficient} structure for representing $\mathcal P$.\label{crit:sufficiency}
\item The maps $\mathcal O$ are one-to-one; i.e., the algebra $\mathcal A$ has a \emph{necessary} (or \emph{essential}) amount structure for representing $\mathcal P$.\label{crit:necessity}
\end{enumerate}
Observables are then defined to be elements of $\mathcal A$. Note that the PESA consists of two non-trivial steps. First, requiring that the evolution of the system be reducible to \emph{well-posed} equations of the form \eqref{eq:gen ev}. And second, requiring that the criteria (i)-(iii) are satisfied for these evolution equations. Focus on both of these steps is what leads to the novelty of the PESA.
The criterion \eqref{crit:autonomy} implies that the time $t$ be uniquely specifiable in terms of a generating set $A$ of the observable algebra, and that the states of a theory can be specified without referencing an external time parameter. The criterion \eqref{crit:sufficiency} requires that the appearances do not underdetermine the initial data required to solve the evolution equations while the criterion \eqref{crit:necessity} ensures that the appearances do not overdetermine these initial data.
The autonomy requirement is motivated by the observation that the PESA has prescriptive force only when the dynamical laws of the theory in question depend on a closed set of structures. If the laws were to depend on additional external structures, then the PESA would be incapable of assessing whether these external structures are observable or not. The autonomy requirement therefore demands that one consider all structures --- particularly the evolution parameter(s) --- involved in solving the dynamical equations when determining the observable algebra of a theory.\footnote{ The autonomy requirement is also very helpful for treating dynamical similarities since these usually rescale the evolution parameter, and so the invariant system is usually expressed as an autonomous set of differential equations.}
The two criteria of sufficiency and necessity are directly motivated by the two ways discussed in the previous section that a theory can be inadequate in accounting for the phenomena: namely that the theory can either have too few or too many observables. These criteria are also clearly motivated by the two stages of theory-construction highlighted in the quote from \cite{vanFraassen:sym_and_surplus} at the beginning of Section~\ref{sub:a_framework_for_symmetry}. The PESA is simply a way to formalise these intuitions into precise mathematical criteria that link the well-posedness of the evolution equations of a theory to its observables. In practice, the value of the PESA is in emphasising that one should explicitly test for each criterion --- autonomy, sufficiency and necessity --- separately.
To clarify how the principle should be used in practice, under the PESA \emph{all} structures that must be specified in order to solve a computational algorithm for determining the future state of a system should be part of the generating set of the observable algebra. This includes initial or boundary data for the dynamical equations. But it also includes \emph{all} other parameters of the theory such as particle masses and couplings constants. In this way, certain structures can easily be under-counted when constructing the observable algebra.
Conversely, the generating set of the observable algebra can also be over-counted because it is not always obvious that a structure can be removed without compromising the ability of a theory to define the maps $\mathcal O$. For instance, in a theory where gauge-fixing conditions are used to give a unique evolution to the equations of a system, the map between gauge-fixed structures and appearances is not one-to-one because many gauge-fixing conditions can be used to describe the same phenomena. The PESA would therefore prescribe against identifying gauge-fixed and observable structures because of the necessity criterion~\eqref{crit:necessity}. We will see important examples in Section~\ref{sub:sing res in GR} where certain structures can be removed from the observable algebra of general relativity without affecting the predictions for cosmology. In such cases, the PESA prescribes that the theory should be expressible in terms of the smallest possible autonomous algebra.
The PESA can be usefully compared and contrasted with standard techniques for analysing surplus structure in the physics literature. A standard technique that is similar to the PESA, called the \emph{Dirac algorithm} after the procedure developed in \cite{dirac2001lectures} (see \cite[\S1.5]{henneaux1992quantization} for a modern treatment), also defines an observable algebra using the formal properties of the evolution equations. The observable algebra in question defines the \emph{Dirac observables} as representational structures in terms of which the evolution equations are well-posed.\footnote{ More specifically, the Dirac observables are those functions on the phase space of a Hamiltonian representation of a theory that commute with all the first class constraints of the theory.\label{ftn:Dirac obs} } The Dirac observables, however, are defined using specific formal structures that exist only for certain kinds of theories (see footnote~\ref{ftn:Dirac obs}). Moreover, even in cases where the Dirac observables are formally defined, there is disagreement about their correct physical interpretation when it comes to time evolution in parametrization-invariant systems.\footnote{ See critical discussions in \cite{barbour2008constraints,pons2005dirac,pons2010observables,kuchar1991problem,kuchavr2011time,pitts2014change,gryb2016schrodinger}. } As we will see in Section~\ref{sub:a_non_symplectic_symmetry}, theories where dynamically similar structure is surplus cannot be analysed using such techniques because the formal properties of dynamical similarity violate the conditions under which the Dirac algorithm can be applied. The PESA (or something like it) is therefore required for an analysis of dynamical similarity and may also be useful in understanding the interpretation of Dirac observables with regards time reparametrization invariance.
It should be emphasised the necessity criterion~\eqref{crit:necessity} is not simply an Occamist norm. If a theory identifies observables that are unnecessary, then that theory may provide false explanations or predictions that rely on the behaviour of the unnecessary observables. We note, however, that the PESA aims makes no metaphysical claims about the explanatory role of surplus and observable structures. Instead it highlights the fact that these play different formal roles in finding well-posed evolution equations. Moreover, the PESA does not demand an \emph{explicit} characterisation of the observable algebra. In many theories, such as general relativity and Yang--Mills theory, no such explicit construction is available. In such cases an \emph{implicit} characterisation of the observable algebra, using for example dynamical constraints, that formally identifies the maps $\mathcal O$ is sufficient. Our formalism is thus compatible with the different notions of surplus structure discussed in \cite{dewar2019sophistication,nguyen2020surplus,bradley2020representational}.
In the universal context, the autonomy requirement removes any dependence of the evolution of a theory's state on temporal structures that do not exist within the theory. In the subsystem context, if there is a temporal structure external to the subsystem in question, then this structure must exist somewhere in the observable algebra of the theory.
\subsubsection{Poincar\'e observables}
\label{sub:poincare_observables}
The observable algebra selected by the PESA is similar to a proposal made at least as early as Poincar\'e, and which has been re-emphasised in \cite{Barbour_Bertotti} in the form of \emph{Poincar\'e's principle}. One of the key ways in which the PESA differs from these earlier proposals is that, in separating a theory's formal symmetries from contingent facts, the PESA is agnostic to metaphysical claims about the world. The principle is therefore neither committed to Poincar\'e's instrumentalism nor to Barbour's relationalism. Importantly, the PESA is \emph{not} committed to a strong relational ontology even though it also takes inspiration from Poincar\'e's writings. Since Barbour has linked Poincar\'e's principle to his form of relationalism, we find it useful to take a moment to illustrate how Poincar\'e's writings are compatible with the PESA independently of relational assumptions.
In Chapter 7 of \emph{Science and Hypothesis} \citep{poincare_foundation}, Poincar\'e studies the case of relative motion and in particular is considering the ontological status of the absolute orientation of a Keplerian system in absolute space. He is particularly concerned with the role of angular momentum effects in such a system if it is treated as isolated, where ``thick clouds hide the stars from men who cannot observe them, and even are ignorant of their existence.''[p.109] In this context, he asks: ``How will those men know that the earth turns round?'' In our language, he is considering the role of rotational symmetry in the universal context.
He notices that the interpretation of angular momentum in this system is ambiguous and can be treated in at least two ways: on the one hand as a constant of motion (which he calls an `accidental constant') in an experimentally undetectable absolute space, and on the other hand as a constant of nature (which he calls an `essential constant') encoded in the relational laws of the epistemically accessible universe. In the end, he concludes that choosing between the two amounts ``only the choice of hypotheses.'' He emphasises that what matters are the precise amount of data required to describe the phenomena: [p.114]
\begin{quote}
``Provided that the future indications of our instruments can only depend on the indications which they have given us, or that they might have formerly given us [\textit{autonomy}], such is all we want [\textit{sufficiency}], and with these conditions we may rest satisfied [\textit{necessity}].'' [Italicised words added for emphasis.]
\end{quote}
In the quote above, we've indicated how the criteria of autonomy, sufficiency and necessity, which are required by the PESA, appear to play an essential role in Poincar\'e's analysis. Moreover, he focuses on the total number of independent data required to find the future state of the system, which is how we defined the observable algebra above. For these reasons, we will refer to the elements of the observable algebra specified by the PESA simply as \emph{Poincar\'e observables} in analogy to the \emph{Dirac observables} commonly defined in the physics literature on gauge theories.
\subsection{The PESA and empirical adequacy}
\label{sub:the_pesa_and_indiscernibility}
It is clear that Poincar\'e had strong empiricist intuitions in mind when emphasising the role of what we have called Poincar\'e observables. While he admits that the conventional interpretation of universal angular momentum in terms of changes of orientation in absolute space is ``the most convenient solution for the geometer'', he adds that ``it is not the most satisfactory for the philosopher, because this orientation \emph{does not exist}'' [emphasis added].
While we will not find it necessary to endorse Poincar\'e's metaphysics, we do think that something can be usefully gained by considering the Poincar\'e observables. When a theory is found to be empirically adequate, we will take the view that \emph{any transformations that leave the Poincar\'e observables invariant relate states that are empirically identified.} When a theory is empirically adequate, there is an isomorphism between the Poincar\'e observables of one of a theory's models and the appearances. Thus, only changes in the Poincar\'e observables can lead to empirically discernible states. States where the Poincar\'e observables take the same value are therefore identical in all their observable properties. In this way, our view is also partially motivated by the Principle of the Identity of Indiscernibles (PII).
We can compare the PESA when applied to an empirically adequate theory to standard treatments of symmetry given in the philosophical literature. We restrict our discussion to a proposal made in \cite{CAULTON2015153} because of its strong similarities to our own proposal. Caulton begins by defining what he calls an \emph{analytic} symmetry. In our terminology, an analytic symmetry corresponds roughly to a universal symmetry as we have defined it.\footnote{ Some notable differences: there is no requirement in \cite{CAULTON2015153} that a symmetry relate DPMs and structures are assumed to take specific values (i.e., they are treated as \emph{quantities}). } The proposal is then described in two phases:
\begin{quote}
During the first phase we set up representational links between the theory and the observable portion of the physical world, under the assumption that the theory is empirically adequate (or similar). In the second phase, we maximise the theory’s analytic symmetries, taking advantage of the representational links forged in the first phase so as not to compromise empirical adequacy. [\S 4]
\end{quote}
This procedure is clearly very close to the one we endorse in this paper. Once the appropriate representational links have been set up, maximising a theory's analytic symmetries is analogous to the necessity criterion~\eqref{crit:necessity}. The sufficiency criterion~\eqref{crit:sufficiency} is automatically satisfied when the theory is empirically adequate. Thus, the main difference between the proposals is our emphasis on having an observable algebra in terms of which the dynamical system can be expressed as a set of well-posed and autonomous evolution equations. More specifically, the PESA links observable structure to those structures that provide the necessary and sufficient input data for the evolution equations. Such a requirement does not exist within Caulton's proposal. It is inspired by the Dirac algorithm but applicable to more general transformations such as dynamical similarity.
The two-step process of specifying observables using the PESA and then checking for empirical adequacy will be our main methodology for analysing dynamical similarity. Using examples we will now see that this process is consistent with the standard treatment of gauge symmetries in physics.
\section{Assessing (in)discernibility in familiar examples}
\label{sec:assessing_indiscernibility_examples}
\subsection{Galilean boosts}
\label{sub:galilean_boosts}
We will now illustrate the power of the PESA by applying it to some well-studied examples of symmetry in the literature and show that it eliminates tensions between dynamic and epistemic considerations. The first example we will consider is the that of the Galilean boost symmetries. These act on a Newtonian system by transforming the velocities, $\dot{\vec q}_I$, of rigid bodies (or point particles) by a constant vector $\vec a$:
\begin{equation}\label{eq:Gal boosts}
\dot{\vec q}_I\to \dot{\vec q}_I + \vec a\,,
\end{equation}
where $I$ ranges over the total number of bodies in the system. The Lagrangian for the theory is
\begin{equation}\label{eq:Gal L}
\ensuremath{\mathcal {L}} = \sum_I \frac{m_I}2 \lf( \dot{\vec q}_I \ensuremath{\right})^2 - V(\vec q_I)\,,
\end{equation}
where $m_I$ are the masses of the bodies and $V(\vec q_I)$ is the potential energy function invariant under Euclidean transformations of the centre-of-mass positions $\vec q_I$ of the bodies.
\subsubsection{Barbour--Berttoti theory: boosts as a universal symmetry}\label{sec:universal Gal}
To understand the sense in which the Galilean boost symmetries \eqref{eq:Gal boosts} can be thought of as universal symmetries of Newtonian mechanics, we first recall some well-studied facts about Newtonian mechanics. Newton was able to produce an empirically adequate theory of planetary motion by assuming that isolated systems are described by indiscernible states related by Galilean boosts acting on the system. This fact is stated and proved in Corollary~V of Newton's \emph{Principia} \citep{newton1999principia}.\footnote{ Corollary~VI states that this is also the case for isolated systems undergoing \emph{arbitrary} linear accelerations. }
An efficient way to understand this invariance in a modern language is to follow a procedure developed in \cite{Barbour_Bertotti} where an isolated Newtonian system is actively boosted by introducing auxiliary fields $\vec b(t)$ that are allowed to be arbitrary functions of time. The boosted Lagrangian involves a transformation of the velocities of the form $\dot{\vec{q}} \to \dot{\vec{q}} + \vec b(t)$ and is therefore given by
\begin{equation}
\ensuremath{\mathcal {L}} _\text{BB} = \sum_I \frac{m_I}2 \lf( \dot{\vec q}_I + \vec b \ensuremath{\right})^2 - V(\vec q_I)\,.
\end{equation}
The Euler--Lagrange equations for $\vec b$ say that the system must be invariant under arbitrary infinitesimal boosts. These equations imply the vanishing of the total linear momentum of the system:
\begin{equation}\label{eq:BB cond}
\vec P_\text{tot} = \sum_I \vec p_I = 0\,,
\end{equation}
where $\vec p_I = \diby{ \ensuremath{\mathcal {L}} _\text{BB}}{\dot{\vec q}^I}$ are the generalised momenta of the system.
This condition can be satisfied by going to a reduced description where the centre-of-mass position (and velocity) has been removed from the system. Such a description can be obtained, for instance, by defining the centre-of-mass variables $\vec Q_I = \vec q_I - \vec q_\text{cm}$, where $\vec q_\text{cm} = \sum_I m_I \vec q_I$, and then eliminating one such variable in terms of the others (the choice is arbitrary). In such a reduced description, the variable $\vec q_\text{cm}$ no longer appears. This indicates that, in BB theory, changing the centre-of-mass position (and velocity), even in a time dependent way, has no effect on the variables satisfying the dynamical constraints \eqref{eq:BB cond}. Importantly, satisfaction of these constraints \emph{reduces} the number of independently specifiable initial data for the system by $6$: the centre-of-mass position and velocity, which are underdetermined by the equations of motion of the BB system. The largest, well-posed autonomous system that satisfies \eqref{eq:BB cond} is therefore $6$ dimensions smaller than the original Newtonian system.
Moreover, when the dynamical constraints \eqref{eq:BB cond} are satisfied, it is easy to see that the original Lagrangian transforms in such a way that
\begin{equation}\label{eq:symp Gal}
S \to S + \phi\,,
\end{equation}
for some constant $\phi$, and therefore that Galilean boosts preserve the stationary points of the action. They are thus a symmetry of the theory by our definition.
The PESA can now be used to identify observables and provide clarity on the status of global boosts in Newtonian mechanics. For isolated systems, the remarkable success of Newtonian mechanics illustrates that coordinates $\vec Q^I$ (with one eliminated) can be used to write an empirically adequate theory of planetary motion. Moreover, Corollary V, or equivalently the analysis of Barbour and Berttoti outlined above, shows that the necessary and sufficient autonomous system required to predict future states from initial states does \emph{not} require a specification of the centre-of-mass velocity of the system, which is arbitrary. In other words, the smaller, well-posed posed system satisfying \eqref{eq:BB cond} is nevertheless sufficient for describing the appearances of Newtonian mechanics. Thus, the PESA prescribes that the centre-of-mass velocity is \emph{not} a Poincar\'e observable because it is not necessary \eqref{crit:necessity}.\footnote{ Autonomy can be achieved in all the examples studied in Section~\ref{sec:assessing_indiscernibility_examples} by expressing the equations of motion using \emph{Jacobi's} principle following the analysis of \cite{Barbour_Bertotti}. Doing this does not change the quantitative predictions of any of these theories. } Moreover, because global Galilean boosts leave the Poincar\'e observables invariant, they are universal symmetries of Newtonian mechanics by our definition. This agrees with modern orthodoxy regarding the role of global boosts in Newtonian mechanics as expressed for instance in \cite{saunders2013rethinking} and \cite{knox2014newtonian}.
\subsubsection{Galileo's ship: boosts as a subsystem symmetry}\label{sec:Gal ship}
In the previous section, we focused attention on the case where the boosts act on isolated systems. We now turn attention to the more familiar context in which the boosts act uniformly on approximately isolated subsystems of the universe but leave the rest of the system's structures invariant.
This context is nicely illustrated by the \emph{Galileo ship} scenario. In this scenario, a ship is floating frictionlessly (or nearly so) on water in such a way that it has uniform velocity relative to a shore. Galileo famously remarked that if you are below deck and all the windows of the ship are closed, then it would not be possible to detect observables differences between situations where the ship was moving with different uniform velocities relative to the shore \citep[pp. 186-187 (Second Day)]{galileo_dialogues}:
\begin{quote}
``...have the ship proceed with any speed you like, so long as the motion is uniform and not fluctuating this way and that. You will discover not the least change in all the effects named, nor could you tell from any of them whether the ship was moving or standing still.''
\end{quote}
The `effects named' in the above quote are examples given by Galileo --- such as the flight path of butterflies or the trajectories of fish swimming in a bowl. These phenomena can be represented as rigid bodies in the Newtonian system described by the Lagrangian \eqref{eq:Gal L}. The ship and all the bodies in it form a subsystem of the universe that behaves, for all practical purposes, as if it is isolated from the rest of the system.
What is different from this context and that treated in the previous section is that here there are means to detect relative motion between the ship and the shore. As Galileo observers, if one climbs above the deck of the ship ``more or less noticeable differences would be seen in some of the effects noted.'' In this case, empirical adequacy cannot be achieved if the centre-of-mass velocity of the ship is allowed to vary arbitrarily: changing the velocity of the ship will obviously lead to discernible changes in the ship's motion relative to the shore.
Unlike in the universal case, empirical adequacy can only be achieved if the relative velocity of the ship and shore is taken to be a Poincar\'e observable. No BB-like constraint of the form \eqref{eq:BB cond} can be used to reduce the system to a smaller, well-posed system that is empirically adequate. This implies that the Galilean boosts of Galileo's ship are subsystem symmetries by our definition: they act invariantly on the ship and the shore but not invariantly on the relative velocity of the ship and shore. In this case, the the PESA prescribes that states states where the ship is boosted relative to the shore should \emph{not} be empirically identical. In other words, boosts of this kind should have empirical consequences. The dynamical criteria of the PESA and the empirical adequacy of Newtonian mechanics (in the energy regime in question) therefore reproduce the standard expectations.
\subsection{Kepler's third law}
\label{sub:kepler_s_third_law}
\subsubsection{Dynamical similarity as a subsystem symmetry}
\label{sub:DS as subsystem}
In the previous section we showed how the PESA can be used in conjunction with empirical adequacy to reproduce orthodox views regarding the status of boost symmetries in Newtonian mechanics. In this section, we will showcase the power of the PESA by applying it the less familiar symmetry of dynamical similarity. As we will see in Section~\ref{sub:a_non_symplectic_symmetry}, dynamical similarities have formal features that make them inapplicable to standard symmetry treatments such as Noether's theorems or the Dirac constraint algorithm.\footnote{ Although there are generalisations of Noether's first theorem that apply to contact systems \cite{bravetti2021geometric}, and therefore may be applicable to the case of dynamical similarity. } But because the PESA applies to any theory whose laws can be expressed as dynamical equations, it can be applied to any general system that has a dynamical similarity. We begin our analysis with the well-studied context of dynamical similarity arising in the Kepler problem to illustrate how the PESA can be applied more generally. We will show explicitly at the end of this section that the dynamical similarities that act in the Kepler problem exactly fit our definition of a subsystem symmetry.\footnote{ Incidentally, this makes dynamical similarity one of the oldest known dynamical symmetries.} It is helpful to note that the dynamical similarity transformations discussed here are simply the large group transformations of the infinitesimal \emph{Runge--Lenz} symmetries described cryptically in \cite{belot:sym_and_equiv} and \cite{pittphilsci16623} (to see this compare the transformations of this section with equations~(62) of \cite{prince1981lie}).
To see the relationship between dynamical similarity and Kepler's third law, consider two bodies co-orbiting under an inverse square law of attraction. It is well-known (see for example \cite[\S3.7]{goldstein2000classical}) that the trajectories of these bodies are described by conic sections, which in turn are defined by their semi-major axis and eccentricity. We denote the separation of the bodies by $r$. In the context we are interested in this section, which is also the context in which Kepler's laws where originally proposed, the two-body system is further described by the angle $\theta$ made by a line between them to a fixed orientation with reference to some idealized distant stars. For simplicity we can absorb all constants such as the masses and the gravitational coupling into the single constant $C$ such that the Lagrangian for the system takes the form:
\begin{equation}
\ensuremath{\mathcal {L}} = T-V = \frac{1}{2} \left(\frac{\de r}{\de t}\right)^2+ \frac{r^2}{2} \left(\frac{\de \theta}{\de t}\right)^2 + \frac{C}{r} \,.
\label{eq:KeplerLagrangian}
\end{equation}
We will now define a transformation $D$ that acts on both $r$ and $t$ while leaving $\theta$ and $C$ fixed in such a way that the Lagrangian is rescaled. Let us pick a positive real number $\beta$. If we let $D: r\rightarrow \beta r$ and $D: t\rightarrow \beta^\alpha t$ we see that if we choose $\alpha=3/2$ then the Lagrangian is rescaled under $D: \ensuremath{\mathcal {L}} \rightarrow \beta^{-1} \ensuremath{\mathcal {L}} $. Because $D$ rescales $ \ensuremath{\mathcal {L}} $ and therefore the action, it is a dynamical similarity by our definition. This in turn implies that $D$ preserves the stationary points of the theory, and is therefore a symmetry.
An immediate consequence of $D$ being a symmetry of the Kepler problem is consistency with Kepler's third law, which can be stated as follows:
\begin{quote}
The square of a planet's orbital period is proportional to the cube of the length of the semi-major axis of its orbit.
\end{quote}
In particular, this means that for every closed solution to the two body problem with a semi-major axis length $l$ and period $T$, there exists a solution with semi-major axis $\beta l$ whose orbital period is $\beta^{3/2} T$ for all positive real numbers $\beta$. This follows trivially from the action of $D$ because of the scaling properties of length, via the transformation of $r$, and duration, via the transformation properties of $t$.
Because $D$ acts equally on points along a trajectory, both semi-major and semi-minor axes are equally rescaled, and hence any solution is mapped onto one that has the same eccentricity.\footnote{Note that Kepler's third law applies also to solutions with difference eccentricities and is, therefore, a stronger claim that what can be proved with dynamical similarity alone.} Under this transformation the total energy of the system, $H=T+V$, transforms non-trivially but surfaces of constant energy are merely rescaled: $D:H \rightarrow \beta^{-1}H$, and therefore their shape is preserved. Likewise the angular momentum is also rescaled: $J=r^2 \dot{\theta} \rightarrow \sqrt{\beta}J$, as is expected from the fact that the action, and therefore the standard of angular momentum, is rescaled in the same way.
We now turn to the question of whether the Kepler problem is formulated in the universal or the subsystem context. If the two bodies are taken to represent an idealisation of a planet-sun system in our solar system, then a transformation that changes the energy and angular momentum of this two-body system is clearly intended to represent a discernible change. This is because the transformation $D$ is intended to act uniformly on one particular planet-sun system while keeping the motion of other planets and the distant stars fixed.
The PESA therefore prescribes, by the sufficiency criterion \eqref{crit:sufficiency}, that in an empirically adequate theory $\beta$ should be a Poincar\'e observable. To determine $\beta$ the value of the rescaled semi-major axis and orbital period can be compared with the value of the semi-major axes and orbital periods of the remaining planets in the solar system. Since $D$ acts invariantly on the observables of the two-body subsystem but does not leave invariant the Poincar\'e observable $\beta$, $D$ fits our definition of a subsystem symmetry of the Newtonian $N$-body problem. Kepler's laws are therefore clearly applicable to the subsystem context. As in the case of Galileo's ship, dynamical similarities in the Kepler problem can be seen to imply direct empirical consequences. Once again, the PESA and the empirical adequacy of Kepler's laws has reproduced orthodox expectations.
\subsubsection{Dynamical similarity as a universal symmetry}
\label{eq:kepler universal}
In this section we imagine a hypothetical thought experiment where dynamical similarity is being studied for the Kepler problem in the universal context. This is similar to the hypothetical world considered by Poincar\'e in Chapter 7 of \emph{Science and Hypothesis} that we described in Section~\ref{sub:poincare_observables}, but adapted to dynamical similarity rather than to rotation. In this thought experiment, we imagine that there is no other matter content in the universe aside from two self-gravitating bodies and a collection of infinitely distant stars that have fixed relative configurations amongst themselves on the celestial sphere i.e., they are literally dots on the celestial sphere. These stars provide a reference frame for determining the orientation of a line connecting the two bodies expressed as an angle, $\theta$, between this line and one of the distant stars. While $\theta$ can be empirically determined using the system of distant stars as a reference, in this world the distance $r$ between the two bodies cannot be compared to any other length.
On the other hand, the \emph{relative} instantaneous rate of change, $\frac 1 r \frac{\de r}{\de \theta}$, that uses the dimensionless angle $\theta$ as a clock can be empirically accessible because the \emph{relative} change of the size of the system can be instantaneously compared with rate of change of $\theta$. This is a possible observable quantity that can be represented in a dynamical system where instantaneous change is empirically accessible. Because there is one internal clock and two dimensionless velocities in terms of this clock, we expect an autonomous system of three variables to be required for empirical adequacy in this hypothetical world.
Our task now is to find a representation of this autonomous system. We start with a representation of the system in terms of the configuration variables $r$ and $\theta$ and their velocities as measured by some auxiliary clock $t$. We take the Lagrangian for the system to be that of the standard Kepler Lagrangian \eqref{eq:KeplerLagrangian}. The empirical inaccessibility of $r$ suggests that it is surplus, and that it should be removed. The subtlety, as always, lies in removing the dependence of $t$ on the length standard used to define $r$. Because this standard determines the unit of the action, we simply need to find a set of variables for the entire isolated system that is invariant under the dynamical similarity transformation $D$ defined in the previous section. In other words, we are looking for a representation of the system where dynamical similarity is treated as a universal symmetry.
A general procedure for reducing an isolated dynamical system to an autonomous system invariant under dynamical similarity was developed in \cite{Sloan:2018lim}. An outline of this procedure is given in Section~\ref{sub:geometric_structure} with a brief technical description in Footnote~\ref{ftn:geo contact}. The output of the procedure is a particular (non-unique) representation of an algebra of invariants for the system and a projection of the dynamical equations generated by the Lagrangian onto this space of invariants. For the purposes of this section, we will simply give one such representation. In Appendix~\ref{sec:homo V} we show how the representation used in this section can be generalised to an arbitrary homogeneous potential.
It is straightforward to explicitly verify that the base elements $a=\sqrt{r} \dot{r} = \frac{2}{3} \frac{\de}{\de t} \left(r^{3/2}\right) $, $b=r^{3/2} \dot{\theta}$ and $\theta$ form an algebra invariant under $D$.\footnote{In this representation of dynamical similarity the constant $C$ is taken to be invariant under $D$. This set of invariants can be constructed from the symplectic 2-form and the action of $D$ using $\rho = r^{1/2}$ (see Footnote~\ref{ftn:geo contact} for the definition of $\rho$ and a brief description of this construction).} We can construct this set by determining the transformation properties of the velocities and multiplying by the appropriate power of $r$ that leaves the combination invariant. In terms of this algebra, it is a short calculation to show that the evolution equations of \eqref{eq:KeplerLagrangian} reduce to the autonomous system of equations:\footnote{ The first order form of this system highlights the number of independent initial data and the conditions for integrability --- both of which are integral to the PESA. }
\begin{align}\label{eq:reduced eoms}
\frac{\de b} {\de \theta} &= \frac{a}{2} & \frac{\de a}{\de \theta} = \frac{a^2 - 2C}{2b} + b\,,
\end{align}
where $\theta$ has been used as an internal clock. This dynamical system is therefore equivalent to the original dynamical system defined by \eqref{eq:KeplerLagrangian} but with all standards of length removed (including those appearing in temporal intervals). This three-dimensional reduced system is autonomous and has the necessary and sufficient number of base elements for achieving empirical adequacy in this hypothetical world. The PESA therefore prescribes that $a$, $b$ and $\theta$ should be Poincar\'e observables for this theory. It is important to note that the system defined by \eqref{eq:reduced eoms} is not a standard Hamiltonian system (we will see why we expect this more generally in Section~\ref{sub:a_non_symplectic_symmetry}), and therefore standard tools for analysing symmetries, such as Noether's theorems and the Dirac algorithm, are not applicable. It is a distinct advantage of the PESA that it can provide a concrete specification of the observable algebra for this system.
Clearly it is difficult to have clear intuitions for what phenomena could be observed in such a hypothetical world. Adding only a bit of structure, however, allows us to see how the PESA does lead to intuitive results, and greatly illuminates how more conventional physics is embedded in the formalism of dynamical similarity in the universal context.
The difficulty in the previous example is that the orbital radius $r$ and period $T$ rely implicitly on spatial and temporal standards, and are therefore not invariant under dynamical similarity. If, however, we make use of spatial and temporal standards \emph{within} the system, then a more intuitive picture emerges.
We can add such structures to our hypothetical system by considering an idealisation of the Earth-Sun system where the Earth is taken to be a uniform sphere\footnote{ This idealisation serves only to fix the factor of proportionality between $R^2$ and the momentum of inertia. It can therefore be straightforwardly generalised.} of radius $R$ that has a constant rotational velocity $\dot\phi$ about its axis of rotation. The angle $\phi$ can be determined by marking a fixed position on the Earth and taking the angle between this and a reference star on the rotational plane. The Lagrangian for this system is:
\begin{equation}
\ensuremath{\mathcal {L}} _\text{ext} = \frac{ \dot r^2}2 + \frac {r^2 \dot \theta^2} 2 + \frac{R^2 \dot \phi^2} 5 + \frac C r\,,
\end{equation}
where again we have absorbed units of mass and factors of $G$ into the constant $C$.
The equation of motion for $R$ tells us immediately that $R$ is a constant of the motion --- as expected for the radius of the Earth. We can therefore use it to set a natural standard for length in the system. By multiplying by the appropriate powers of $R$, we can easily construct an invariant notion of radial distance between the Earth-Sun system by forming the ratio: $\gamma = \frac r R$ and an invariant momentum using $p = R^{1/2} \dot r = R^{3/2}\dot\gamma$. Noting that $\theta$ and $\phi$ are already invariant, we can construct the remaining invariants by rescaling the canonical momenta appropriately by $R$:\footnote{This set of invariants can be constructed by using $\rho = R^{1/2}$. See Footnote~\ref{ftn:geo contact} for details.}
\begin{align}
p_\theta &= \frac{r^2 \dot \theta}{R^{1/2}} = \gamma^2 R^{3/2} \dot \theta &
p_\phi &= R^{3/2} \dot \phi\,.
\end{align}
In terms of these variables, time derivatives always appear in the equations of motion with the pre-factor $R^{3/2}$, which cancels the transformation of $t$ under $D$ by construction.
We can obtain a scale-free autonomous system by switching to an internal clock. For this system, a natural choice is a rescaling of $\phi$ such that $R^{3/2} \frac{\de \phi}{\de t_\phi} = 1$, where we define $t_\phi \propto \phi$. Such a choice is always available for this system because $\dot \phi$ and $R$ are constants of motion of the original system. Physically, the clock $t_\phi$ can be interpreted as reading out a sidereal day. In terms of $t_\phi$, a short calculation shows that the equations of motion for the system reduce to:
\begin{align}
p &= \gamma' & p' &= - \frac C {\gamma^2} \\
\theta' &= \gamma^2 p_\theta & p_\theta' &= 0\,,
\end{align}
where primes indicate derivatives with respect to $t_\phi$. We thus obtain the equations of motion for the original Kepler system \eqref{eq:KeplerLagrangian} but with length standards given by the radius of the Earth and temporal standards given by the sidereal day.
This example illustrates that the PESA reproduces our intuitions for a closed dynamical system with internally defined standards for length and duration. This provides more evidence that the PESA is a reliable symmetry principle for when we apply it in a novel way to the cosmological applications of Section~\ref{sec:applications}. The dynamical similarity $D$ acts on representations of length in such a way that the privileged rod, in this case of radius of the Earth, gets rescaled in exactly the same way as the distances being measured, in this case the radial distance between the Sun and the Earth. Similarly, representations of time intervals are also rescaled so that the number of sidereal days in a sidereal year is fixed. One obtains the standard Kepler system precisely because the rods and clocks used to describe the autonomous motion are dynamically isolated from the rest of the system.
Let us make a couple important comments to help build intuition for the role of dynamical similarity in physical systems. The case treated above is an example of a more general result that a conservative system can always be approximately recovered from a dynamically similar one when the internal rods and clocks used to describe the motion are approximately dynamically decoupled from the system itself. In the Kepler system, this also provides a vivid illustration of the difference between the universal and subsystem contexts: a subsystem transformation that rescales the orbital radius and period but keeps the length $R$ and and sidereal time $t_\phi$ fixed will produce an empirically distinct model of the system. This is one way (and perhaps an historically accurate way) to understand Kepler's third law for fixed eccentricity. In the universal context one simply investigates indiscernibility under different representations of the same internal clocks and rods.
What makes the standard Kepler picture work is therefore the dynamical isolation of the internal clocks and rods. This, however, is necessarily an idealisation because in order to gain knowledge of the value of any clock or rod, an observer must interact with it --- even if only very weakly. In practice, because there is no gravitational screening, no clock or rod can ever be perfectly isolated from the system it is describing. In the Earth-Sun system, the sidereal day is affected by, among other things, the tidal drag between the Earth and the Moon, and is therefore not a dynamically isolated clock. Indeed, this is an example of a much more general discussion about the role and limitations of idealised inertial clocks dating back at least as far as Mach.\footnote{ For a list of notable discussions of these points see \cite{barbour2001discovery,barbour:mach_before_mach,mach1907science,pfister2014ludwig}. } Dynamically similarity becomes particularly important in such cases because the system can no longer be idealised by a conservative system and instead exhibits friction-like behaviour. As we will see in Section~\ref{sub:cosmo probs}, nowhere is this more important than in the cosmological case. There, the expansion of the universe affects all clocks in the universe and implies that, on cosmological time-scales, the system is non-conservative. As we will see, this leads to significant implications for several conceptual problems in cosmology.
\section{Dynamical similarity}
\label{sec:dynamical_similarity}
In this section we give a general definition of dynamical similarity for Lagrangian systems and a brief outline of some of the geometric structures that underlie the framework. We will see that a key mathematical tool in our analysis will be the odd-dimensional cousin of symplectic geometry called \emph{contact geometry}, and that conservative Hamiltonian systems exhibiting dynamical similarity will generically have a description in terms of \emph{contact Hamiltonian systems}. For more details on the link between dynamical similarity and contact Hamiltonian mechanics, see \cite{Sloan:2018lim}. For a more details regarding the mathematical construction of contact Hamiltonian systems see Appendix~4 of \cite{arnol2013mathematical} and \cite{2019JMP....60j2902D,Bravetti_2017}. \cite{Bravetti:2018rts} gives a recent review with applications to thermodynamics and \cite{geiges2008introduction} gives a detailed list of proofs of all relevant theorems.
\subsection{General definition}
\label{sub:general_definition}
We are primarily interested in understanding the action of dynamical similarity in Lagrangian systems in the universal context. In this context, states related by dynamical similarity are thought to be empirically indiscernible, and a mathematical procedure for identifying such states warranted. We will now give such a procedure by defining equivalence classes of DPMs under dynamical similarity. The equivalence relations thus obtained defines a projection from the original system to a smaller autonomous system that is invariant under dynamical similarity. In this way, the procedure is analogous to the reduction of a system by a gauge transformation as is familiar to standard gauge theory. We will not pursue a complete geometric construction here, but will outline a coordinate-based approach, motivated by a powerful theorem, that we will illustrate in the applications of Section~\ref{sec:applications}.
\subsubsection{A non-symplectic symmetry}
\label{sub:a_non_symplectic_symmetry}
We begin by highlighting the way in which dynamical similarities are importantly different from most standard gauge symmetries studies in the literature. This will also allow us to give a more precise definition of dynamical similarity. Consider a general Lagrangian system described by the action functional
\begin{equation}
S[\gamma] = \int_\gamma \de t\, \ensuremath{\mathcal {L}} (q, \dot q) \,,
\end{equation}
where $ \ensuremath{\mathcal {L}} (q, \dot q)$ is the Lagrange density of the system; $(q, \dot q)$ are generalized coordinates and their velocities providing coordinates for the tangent bundle, $T\mathcal C$, over the configurations space $\mathcal C$, $t$ is the time parameter, and $\gamma$ are the kinematically possible models of the theory. As stated in the introductory Section~\ref{sub:dyn of sim}, the DPMs for this system satisfy
\begin{equation}\label{eq:eoms}
\delta S[\gamma]\big\rvert_{\gamma_\text{DPM}} = 0\,.
\end{equation}
Now consider the general class of transformations $ \ensuremath{D} $ that act on the basic structures of a theory\footnote{ We will be more explicit about what these structures are in Section~\ref{sub:the_action_of_dynamical_similsubarity}. } such that the pullback of the action $S$ by $ \ensuremath{D} $ is
\begin{equation}\label{eq:gen dyn sym}
\ensuremath{D} ^* S \to c S + \Phi\,,
\end{equation}
for some constant $\Phi$ and some non-zero positive constant $c$. The transformation $ \ensuremath{D} $ is the most general transformation that preserves the DPMs of the theory.
An important way to characterise the different symmetries of $ \ensuremath{D} $ can be made using the Hamiltonian formalism. This formalism exists when one can write the Lagrangian in the form $ \ensuremath{\mathcal {L}} \de t = p \de q - H \de t$, where the term $\theta = p \de q $ is called the \emph{symplectic potential}. The symplectic potential can be thought of as providing the relationship between the generalised momenta $p$ and the velocities $\dot q$ via the definition $p = \diby{ \ensuremath{\mathcal {L}} }{\dot q}$, and is therefore central to the formation of geometric structures on phase space. An example of such a structure is the natural volume element on phase space called the \emph{Liouville measure}, which can be conveniently written in the language of differential forms or the coordinates $(q,p)$ as
\begin{equation}\label{eq:Liouville}
\mu_L(R) = \int_R (\de\theta)^N = \int_R \de^N p\, \de^N q\,,
\end{equation}
where $N$ is the dimension of $\mathcal C$ and $\de$ is the exterior derivative operator on phase space. An important requirement for the Hamiltonian to exist is therefore that the volume-form defined by $\mu_L(R)$ be non-degenerate.
To understand the significance of the symplectic potential, consider what happens when the theory undergoes a transformation that shifts it by an exact differential $\theta \to \theta + \de \phi$. Under such a transformation, $ \ensuremath{\mathcal {L}} \de t \to \ensuremath{\mathcal {L}} \de t + \de \phi$, where $\phi$ is an arbitrary phase space function, and the action is shifted by a constant: $S[\gamma] \to S[\gamma] + \phi\bigr\rvert_{t_0}^{t_1}$.\footnote{ Note that we therefore have $\Phi = \phi\bigr\rvert_{t_0}^{t_1}$. } This case thus corresponds to a symmetry $ \ensuremath{D} $ in which $c = 1$. Motivated by the fact that the exterior derivative operator on phase space satisfies $\de^2 = 0$, we can define the \emph{symplectic 2-form} $\omega = \de \theta$ that characterises \emph{symplectic structure}. Symmetries for which $c = 1$ preserve the symplectic 2-form, and we will therefore refer to them as \emph{symplectic symmetries}.\footnote{ Symplectic symmetries are often referred to as \emph{divergence} symmetries (see, for example, \cite{belot:sym_and_equiv} and \cite{olver2000applications}) because they shift the Lagrangian by an exact form. }
It is difficult to overstate the importance of the symplectic structure defined by $\omega$ in formulating the Hamiltonian mechanics of dynamical systems. We've already mentioned how the natural notion of volume on phase space can be defined in these terms. Perhaps even more important is the fact that the dynamical equations themselves in the form of Hamilton's equations are also expressed in terms of the inverse of $\omega$. Even the quantum formalism is motivated by the preservation of the anti-commutation relation between $q$ and $p$ that arise through the definition of $\omega$. It is thus not surprising that symplectic symmetries play an important role in physics and are the subject of most of literature on this topic. A simple example of symplectic symmetries is the Galilean boost symmetries of Section~\ref{sub:galilean_boosts} --- as is illustrated by Equation~\ref{eq:symp Gal}. Indeed, using techniques similar to those used in that section it is straightforward to show that all the Galilean symmetries are symplectic symmetries of Newtonian mechanics.
Dynamical similarities, on the other hand, correspond to the case where $c \neq 1$ according to the definition given in Section~\ref{sub:dyn of sim}, and are therefore non-symplectic symmetries. It is rather remarkable that the physics and philosophy literature has focused almost entirely on symplectic symmetries while almost ignoring the obvious generalisation to $c \neq 1$.\footnote{ An exception being the \emph{scaling} symmetries discussed in \cite{belot:sym_and_equiv}, which can be shown to be dynamical similarities. } In fact, all the interesting features of dynamical similarity that will be introduced in Section~\ref{sub:friction_janus_points_and_attractors} and studied in the applications of Section~\ref{sec:applications} are due to the non-symplectic nature of the dynamical similarities. Moreover, it is because of the non-symplectic character of dynamical similarity that standard symmetry techniques do not apply, and why an analysis in terms of the PESA is necessary.
\subsubsection{The action of dynamical similarity}
\label{sub:the_action_of_dynamical_similsubarity}
To understand how dynamical similarities act on the basic structures of a theory, first recall that the action can be written in terms of the Hamiltonian as
\begin{equation}\label{eq:gen S}
S = \int \de t \lf( \dot q \cdot p - \ensuremath{\mathcal {H}} \ensuremath{\right})\,.
\end{equation}
In order for the pullback of $S$ to have the form \eqref{eq:gen dyn sym} under $ \ensuremath{D} $, we must have that the phase space quantities below transform as
\begin{align}
\ensuremath{D} ^* p\de q &= c p \de q + \de \phi \\
\ensuremath{D} ^* \ensuremath{\mathcal {H}} \de t &= c \ensuremath{\mathcal {H}} \de t + \de \phi\,.
\end{align}
We further restrict $ \ensuremath{D} $ to be such that it has a group action on $\Gamma$. Under this restriction, all symmetry transformations of $ \ensuremath{D} $ depend only on upon the instantaneous state of the system (i.e., the values of $q$ and $p$), the choice of Hamiltonian $ \ensuremath{\mathcal {H}} $ and the time parameter $t$. Moreover, such a symmetry will map the structures of a theory to structures in the same theory. The group action of $ \ensuremath{D} $ on $\Gamma$ can be conveniently expressed in terms of the infinitesimal representations:\footnote{ For simplicity in this paper we work with explicit expressions in terms of Darboux coordinates on $\Gamma$. For the more geometrically inclined reader, these definitions are equivalent to requiring that the tangents, $v_D$, to the orbits of $D$ in $\Gamma$ are defined implicitly as the unique solution to: $\iota_{v_D} \omega = \theta + \de\phi$, where $\omega$ and $\theta$ are the symplectic 2-form and symplectic potential respectively and $\iota$ denotes the interior product on the exterior algebra of $\Gamma$. }
\begin{subequations}\label{eq:local S}
\begin{align}
\delta_ \ensuremath{D} q &= n q + \pb{q}{\phi} & \delta_ \ensuremath{D} p &= (1-n) p + \pb p \phi \label{eq:local pq}\\
\delta_ \ensuremath{D} \ensuremath{\mathcal {H}} &= \Lambda \ensuremath{\mathcal {H}} + \pb \ensuremath{\mathcal {H}} \phi & \delta_ \ensuremath{D} t &= (1- \Lambda) t \label{eq:local Ht} \,,
\end{align}
\end{subequations}
for the arbitrary real numbers $n$ and $\Lambda$, and where the \emph{Poisson bracket} is defined by $\pb f g = \diby f q \diby g p - \diby g q \diby f p$.\footnote{ Note that the equations \eqref{eq:local S} are partial differential equations for $\phi$ on phase space. We are not aware of any general existence theorems for these equations, but we can show that solutions exist for all the examples treated in this paper. }
The first line \eqref{eq:local pq} indicates that the dynamical similarity $ \ensuremath{D} $ rescales the coordinates $q$ and $p$ of phase space in such a way that the combination $q \cdot p$, which has the units of the action (or angular momentum), gets rescaled by a constant. The transformation $D$ therefore relates representations of the system with different conventions for the unit of angular momentum --- in line with our general expectations for a dynamical similarity. The extra Poisson bracket terms involving $\phi$ in the first line reflect the freedom to absorb units differently into either the momentum variables $p$ or the configuration variables $q$. The function $\phi$ encodes all the different possibilities for doing this. For example, $n$ can be redefined arbitrarily to $n\to n + m$ by shifting $\phi$ by $m p\cdot q$.\footnote{ This freedom reflects the ability to redefine a dynamical similarity up to a canonical transformation. In many concrete problems it is useful to make use of this freedom to find the appropriate representation of dynamical similarity for the system in question. }
One important consequence of dynamical similarities being non-symplectic symmetries is that they do not preserve the Liouville measure $\mu_L$. This can easily be seen by noting that the transformations \eqref{eq:local pq} rescale the coordinates $q$ and $p$ in such a way that the integrand of \eqref{eq:Liouville} is necessarily rescaled. This means that the natural notion of volume on phase space is not invariant under dynamical similarity. Some of the consequences of this will be explored in Section~\ref{sub:friction_janus_points_and_attractors} and studied explicitly in the applications of Section~\ref{sec:applications}.
In general the Hamiltonian, $ \ensuremath{\mathcal {H}} (q,p,k_i)$, will be a function both of the phase space variables $(q,p)$ and of certain coupling constants $k_i$, which often carry units themselves. If we assume for simplicity that these couplings are invariant under $ \ensuremath{D} $,\footnote{We have already stated that this assumption is not particularly well-motivated but will forgo a more detailed analysis of the more general case for future work.} then $ \ensuremath{\mathcal {H}} $ will transform in a fixed way once a choice for $\phi$ is made in the first line. The second line \eqref{eq:local Ht} should therefore be regarded either as a consistency requirement with the first line or as an independent equation that fixes $\Lambda$ in terms of $\phi$ and $n$. We will see in concrete examples, including the all-important cosmological case of Section~\ref{sub:friedmann_lema^itre_robertson_walker_cosmology}, that satisfying \eqref{eq:local Ht} can lead to highly restrictive conditions on the possible choices of $\phi$, $n$ and $\Lambda$.
\subsubsection{The subsystem context}
\label{sub:the_sub_system_context}
While most of this section is concerned primary with treating dynamical similarity in the universal context, it can also be useful to understand the considerations that are relevant for dynamical similarity in the subsystem context. In general, if a theory has a subsystem symmetry (recall our definition in Section~\ref{sub:contexts_of_symmetry}) then the Lagrange density of a theory can be decomposed as
\begin{equation}
\ensuremath{\mathcal {L}} = \ensuremath{\mathcal {L}} _\text{sub} + \ensuremath{\mathcal {L}} _\text{inv} + \ensuremath{\mathcal {L}} _\text{int}\,.
\end{equation}
In this decomposition, the term $ \ensuremath{\mathcal {L}} _\text{inv}$ only depends upon the structures that are invariant under the symmetry, the term $ \ensuremath{\mathcal {L}} _\text{sub}$ only depends upon the structures on which the subsystem symmetry acts, and the interaction term $L_\text{int}$ involves the interaction of these two structures.
The invariance property of $ \ensuremath{\mathcal {L}} _\text{inv}$, the transformation properties of $ \ensuremath{\mathcal {L}} _\text{sub}$ and the linearity of the variations $\delta$ imply
\begin{equation}
\ensuremath{D} ^*\delta S \to \ensuremath{D} ^*\delta S_\text{inv} + \ensuremath{D} ^*\delta S_\text{sub} + \ensuremath{D} ^*\delta S_\text{int} = \ensuremath{D} ^*\delta S_\text{int}\,,
\end{equation}
for any value of $c$ and $\phi$. We therefore find that a sufficient condition for a system to have an exact (or approximate) subsystem symmetry is that the interaction term, $S_\text{int}$, between the subsystem structures and the invariant structures be exactly (or approximately) zero. Moreover, this condition does not depend on whether the symmetry is symplectic or not.\footnote{Thus, dynamically similar systems can display \emph{subsystem-recursivity} as defined in \cite{pittphilsci16623}. Note however that a dynamical similarity that acts on a subsystem is not a dynamical similarity of the entire system even if it is a symmetry.}
The universal/subsystem distinction is thus indifferent to the symplectic properties of a symmetry. Intuitions about subsystems gained by studying standard gauge symmetries can therefore remain more-or-less intact when studying dynamical similarity. This can be seen explicitly in the Galilean ship example of Section~\ref{sec:Gal ship} and the Keplerian example of Section~\ref{sub:DS as subsystem}. In both examples, it is the dynamical isolation of the subsystem that permits a subsystem symmetry.
\subsection{Contact geometry and dynamically similar evolution}
\label{sub:geometric_structure}
In this section we will see how to construct a system of dynamically-similar invariants for a conservative Lagrangian system. The invariant description will no longer be conservative and will display friction-like behaviour. To see this, let us begin with Hamilton's equations, which specify the DPMs for the dynamical system with the action \eqref{eq:gen S} in terms of curves on phase space that satisfy:
\begin{align}\label{eq: conservative ham}
\dot q &= \frac{\partial \ensuremath{\mathcal {H}} }{\partial p} = \pb q \ensuremath{\mathcal {H}} & \dot p &= -\frac{\partial \ensuremath{\mathcal {H}} }{\partial q} = \pb p \ensuremath{\mathcal {H}} \,,
\end{align}
where the Hamiltonian $ \ensuremath{\mathcal {H}} $ has no explicit dependence on $t$. It is straightforward to show that the evolution equations preserve the value of the Hamiltonian, which is interpreted as the total energy $E$ of the system, and that the system is therefore conservative.
As we have defined them in Section~\ref{sub:the_action_of_dynamical_similsubarity}, dynamical similarities have a one-dimensional group action on phase space parametrised by the group parameter $c$. We therefore expect $D$ to define one-dimensional orbits on the phase space $\Gamma$, and that the space of dynamically-similar invariants should be odd-dimensional with one dimension less than $\Gamma$. For this reason, \emph{contact geometries}, which are odd-dimensional, are an ideal tool for representing such systems. We will now briefly review these geometries. See the references cited at the beginning of Section~\ref{sec:dynamical_similarity} for more details and proofs.
Contact geometries are constructed to be such that vector fields on them can always be divided into a vertical subspace, which can loosely be thought of as spanning the odd dimension, and an horizontal subspace, which is even dimensional and symplectic.\footnote{ The vertical subspace is simply the kernel of $\de \eta$ (with $\eta$ defined by Equation~\ref{eq:eta darboux}) and the horizontal subspace by the kernel of $\eta$. } According to a theorem by Darboux, a canonical choice of local coordinates $(Q, P, A)$ can always be found such that $(Q,P)$ are canonically conjugate coordinates along the horizontal directions and $A$ is a coordinate along the vertical direction. The geometry of a contact manifold is specified by a \emph{contact 1-form} $\eta$ that can be written in these canonical coordinates as
\begin{equation}\label{eq:eta darboux}
\eta = P\de Q - \de A
\end{equation}
in analogy to the symplectic potential $\theta = p \de q$ of a phase space. In the same way that the geometry of a symplectic manifold is specified by the symplectic 2-form, the geometry of a contact manifold is specified by the contact 1-form. Moreover, just as the symplectic 2-form is required to be such that the Liouville measure is non-zero, the contact 1-form is required to be such that the volume-form
\begin{equation}\label{eq:contact vol}
\mu_c = \eta \wedge (\de \eta)^{N-1}
\end{equation}
on the contact manifold is non-degenerate. In fact, this non-degeneracy requirement is a necessary and sufficient condition for $\eta$ to be a contact 1-form.
The first goal is to find a maximal set of independent phase space functions that are invariant under dynamical similarity. Once this is a achieved, the challenge is to find an autonomous set of evolution equations that will evolve the invariant quantities in the same way as Hamilton's equations \eqref{eq: conservative ham}. In this way, the new system will evolve identically to the original system as far as any invariant quantities are concerned. It is in finding such evolution equations that contact geometry is useful. Towards this end, it will be useful to construct a contact 1-form directly in terms of these invariants.
A simple way to construct the required invariant 1-form is to follow the procedure used in Section~\ref{sub:kepler_s_third_law} and Appendix~\ref{sec:homo V}, which mirrors the general procedure developed in \cite{Sloan:2018lim}. In the examples considered in those sections, we found a function on phase space, which in general we can call $\rho$, that scales in the same way as the Lagrangian under $D$. Then we use this function to rescale all other quantities on $\Gamma$ in a way that the combination is invariant under $D$. For example, an invariant can be constructed from the original Hamiltonian by taking the combination
\begin{equation}\label{eq:contact ham gen}
\ensuremath{\mathcal {H}} ^c = \frac{ \ensuremath{\mathcal {H}} } { \rho^\Lambda }\,,
\end{equation}
where $\Lambda$ is determined by the choice made in \eqref{eq:local Ht}. Using a bit of ingenuity, or the geometrically inspired procedure of Footnote~\ref{ftn:geo contact}, one can obtain a full set of invariants from the original phase space functions $q$ and $p$ by multiplying by powers of $\rho$ such that $\eta$ takes the standard form \eqref{eq:eta darboux}. The difficulty, of course, is not in finding an invariant set, but one in which a contact form $\eta$ can be constructed that can lead to a simple set of evolution equations.
To construct such evolution equations we are guided by contact geometry. On a contact manifold, if one is given some function $ \ensuremath{\mathcal {H}} ^c$, which we will call a \emph{contact Hamiltonian}, then it is always possible to construct a \emph{contact vector field} using $\eta$ that is generated by $ \ensuremath{\mathcal {H}} ^c$.\footnote{ More concretely, the invariant contact form can be constructed from some phase space function $\rho$ satisfying $ \ensuremath{\mathfrak {L}} _D \rho = \rho$ using $\eta = \frac{\iota_D \omega }\rho$. The contact Hamiltonian is related to the projection, $p_H$, of the Hamilton vector field of the Hamiltonian function $v_H$ onto the orthogonal compliment of $D$ via $ \ensuremath{\mathcal {H}} ^c = \iota_{p_H} \eta$. See \cite{Bravetti_2017} or \cite{2019JMP....60j2902D} for more details on how to construct the contact Hamiltonian.\label{ftn:geo contact}} This procedure is analogous to how one can use the usual Hamiltonian function to generate Hamilton's equations using the symplectic 2-form in standard Hamiltonian mechanics. Using this procedure, one can pick a contact Hamiltonian $ \ensuremath{\mathcal {H}} ^c$ such that the evolutions equations are equivalent to \eqref{eq: conservative ham}. It turns out that a good choice for $ \ensuremath{\mathcal {H}} ^c$ is given by \eqref{eq:contact ham gen} and the equations of motion take the form:
\begin{align}\label{eq:contact ham}
\dot Q &= \frac{\partial \ensuremath{\mathcal {H}} ^c}{\partial P} & \dot P &= -\frac{\partial \ensuremath{\mathcal {H}} ^c}{\partial Q} - P \frac{\partial \ensuremath{\mathcal {H}} ^c}{\partial A} & \dot A &= P \frac{\partial \ensuremath{\mathcal {H}} ^c}{\partial P} - \ensuremath{\mathcal {H}} ^c \,.
\end{align}
While the explicit form of the equations depends on the choice of $\rho$, it can be shown\footnote{ Since $\rho$ is defined by $ \ensuremath{\mathfrak {L}} _D \rho = \rho$, any other function $\rho'$ must be related to $\rho$ by $\rho' = \alpha \rho$, where $\alpha$ is an invariant: $ \ensuremath{\mathfrak {L}} _D \alpha = 0$. The contact form $\eta'$ defined via the procedure in Footnote~\ref{ftn:geo contact} can then be related to $\eta$ by performing a contact transformation. The projection $p_H$, and therefore $ \ensuremath{\mathcal {H}} ^c$, is then also invariant up to a contact transformation. } that different choices of $\rho$ lead to an equivalent set of equations of motion. We are thus lead to a unique autonomous system of equations (up to coordinates transformations that preserve the contact form) in terms of a set of invariants of $D$.\footnote{Recently, it has been emphasised in \cite{Sloan:2020taf} that the equations of motion \eqref{eq:contact ham} can be obtained by a variational principle first proposed by Herglotz \citep{herglotz,guenther1996herglotz} in which the action only involves invariants of $D$.} While the system \eqref{eq:contact ham} is explicitly non-symplectic, it can nevertheless be cast as a theory that satisfies evolutions equations of the form \eqref{eq:gen ev}. The PESA is therefore straightforwardly applicable to dynamically similar systems.
\subsection{Friction, Janus points and attractors}
\label{sub:friction_janus_points_and_attractors}
In this section we study some of the consequences of the invariant evolution equations \eqref{eq:contact ham}. Because the space of invariants is a contact manifold and not a symplectic manifold, the equations of motion \eqref{eq:contact ham} cannot be written in the simple Poisson-bracket form \eqref{eq: conservative ham}. Instead, friction-like terms linear in the momenta $P$ appear in the equations for $\dot P$ and $\dot A$. These arise as advertised because of the non-symplectic character of $D$. Unlike the symplectic equations \eqref{eq: conservative ham}, the friction-like equations \eqref{eq:contact ham} are not invariant under the usual time-reversal operation. We discuss the potential implications of this for addressing the problem of the arrow of time in Section~\ref{sub:aot}.
One way to understand the friction-like behaviour of the system physically is to consider that the contribution of the total energy of the system due to $\rho$ is not included in the invariant system because $\rho$ is used to set the standard of length, and is therefore not itself empirically accessible. Removing $\rho$ from the system therefore removes its contribution to the total energy in the redundant representation. The energy of the remaining system must then compensate for this, and this compensation appears in the friction-like terms of \eqref{eq:contact ham}. In a genuine frictional system, the energy losses due to friction are interpreted as being dissipated into degrees of freedom that are epistemically inaccessible for practical reasons. In contrast, in dynamically similar systems changes in energy can either be losses or gains depending on the dynamics of $\rho$, and these changes are due to the impossibility of being able to observe of the overall scale of the system. We thus use the word `friction-like' to describe the evolution equations \eqref{eq:contact ham} since the analogy to friction holds only at the formal level.
We now investigate some of the consequences of this friction-like behaviour. As already discussed in Section~\ref{sub:the_action_of_dynamical_similsubarity}, the Liouville measure --- i.e., the natural notion of volume on phase space --- is rescaled $\mu_L(R) \to c^N \mu_L(R)$ under $ \ensuremath{D} $. This means that the volume of a region $R$ in phase space calculated with the Liouville measure will attribute a volume to states related by $ \ensuremath{D} $ as if they were ontologically distinct. In the universal context, the necessity criterion \eqref{crit:necessity} of the PESA for empirically adequate theories implies that, when the theory is empirically adequate, states related by $ \ensuremath{D} $ should be treated as identical. The Liouville measure therefore introduces a distinction without a difference and is no longer the appropriate measure to use for assigning volumes. What then are the properties of measures that are compatible with the PESA in the universal context?
To answer this question, we must determine how the evolution of a measure projects onto the invariant space. To be concrete, let us consider the natural volume form $\mu_c(R)$ on a contact manifold defined in \eqref{eq:contact vol} using $\eta$. In the original Hamiltonian system, the Liouville measure is particularly interesting because it is preserved under the evolution equations \eqref{eq: conservative ham}. The dynamically similar measure $\mu_c(R)$, however, is not. To see this, we use the equations \eqref{eq:contact ham} to evolve $\mu_c(R)$. A short calculation (see \cite{Bravetti_2017}) leads to the result:
\begin{equation}\label{eq:focusing}
\dot \mu_c(R) = - N \diby{ \ensuremath{\mathcal {H}} ^c}{A} \mu_c (R)\,.
\end{equation}
Thus, when the frictional term in $\dot P$ of \eqref{eq:contact ham} is not zero --- i.e., when $\diby{ \ensuremath{\mathcal {H}} ^c} A \neq 0$ --- then the natural volume form on the invariant space evolves according to \eqref{eq:focusing}. In particular, when $\diby{ \ensuremath{\mathcal {H}} ^c} A > 0$ we have standard friction-like behaviour with the measure shrinking in time. In contrast, when $\diby{ \ensuremath{\mathcal {H}} ^c} A < 0$, the measure is growing in time. This shrinking and growing can be understood as a kind of focusing (or defocusing) of the dynamical trajectories on the invariant space.
The measure-focusing we have just described has important consequences for many kinds of scientific inferences made in a theory. When evaluating a particular piece of evidence --- for instance in the context of a Bayesian update of some prior --- it is necessary to know the likelihood of that evidence given the theory. These likelihoods are probabilities that a particular state is observed given a particular theoretical model of the system, and are therefore represented as (normalisable) measures on the state space of that theory. But in the universal context, the appropriate measures to use are measures on the invariant state space. As we have just seen, such projections can have very different behaviour when compared to measures on the redundant space. These considerations can therefore have a dramatic effect on any inference made within a theory. In cosmology, this becomes particularly relevant as we will explore more fully in Section~\ref{sub:cosmo probs}.
\paragraph*{Janus Points and attractors}
To end this section, we consider two kinds of privileged structures on the state space of a theory described by the friction-like evolution equations \eqref{eq:contact ham}. Before doing so we note that, because of the relation \eqref{eq:focusing}, the quantity $\diby{ \ensuremath{\mathcal {H}} ^c}{A}$ at a particular point along a dynamical trajectory gives the rate of expansion (or contraction) of the measure $\mu_c$ in an infinitesimal region around that point. Using this interpretation, we consider points along a dynamical trajectory at which $\diby{ \ensuremath{\mathcal {H}} ^c}{A}$, and therefore $\dot \mu_c$ (in an infinitesimal region around that point), changes sign. When this occurs, the canonical measure goes from focusing to defocusing or vice versa. In \cite{Barbour_2014,barbour2013gravitational,barbour:janus}, it is argued that the direction in which the focusing of the measure occurs can be taken to define an arrow of time for observers in systems moving along the trajectories in the focused regions. They call points where such an inflection occurs \emph{Janus} points, in reference to the two-faced Roman god of transitions and time, because different local arrows of time emerge on either side of such points.
Janus points come in two different kinds. The first kind, which we call \emph{type~1}, occurs when $\diby{ \ensuremath{\mathcal {H}} ^c}{A} = 0$ but where $\ddot\mu_c \neq 0$. In this case, the measure has an extremum at the Janus point but this extremum is not a fixed point of the measure. These conditions ensure that two different local arrows of time will emerge on either side of the Janus point. An example of a type-1 Janus point is found in the Newtonian $N$-body gravitational system with zero total energy and zero angular momentum treated in \cite{Barbour_2014}.\footnote{ Note that these authors do not distinguish between type-1 and 2 Janus points, which is a language we have introduced in this text. } In these systems a unique type-1 Janus point occurs for all solutions when the moment of inertia of the system is minimised.
The only other way for $\dot \mu_c$ to change sign when moving through a state space point $x$ along a dynamical trajectory is if it runs through infinity at $x$. More precisely, we define a \emph{type-2} Janus point as a state space point $x$ along a dynamical trajectory such that $s \diby{ \ensuremath{\mathcal {H}} ^c}{A}(x^-) \to \infty $ and $-s \diby{ \ensuremath{\mathcal {H}} ^c}{A}(x^+) \to \infty $, for $s \in \{ -1 , +1 \}$ and where $x^-$ approaches $x$ from below and $x^+$ exits $x$ from above. An important example of a type-2 Janus point can be shown to occur in symmetry-reduced models of general relativity at the big bang. These models are convenient for describing general relativity near a singularity and will be studied explicitly in Section~\ref{sub:sing res in GR}. There, the invariant system will be seen to evolve smoothly at the Janus point, resolving the usual big bang singularity.
A second class of privileged state occurs when there is a region $R$ in the invariant space that is such that the asymptotic dynamical flow of the measure converges: $\lim_{t\to \infty} \dot \mu_c(R)\to 0$. In this case, the region $R$ is a trapped attractive surface or \emph{attractor}.\footnote{ See, for example, \cite{hurley1982attractors} for a technical definition. } When such a region exists, there is a basin of attraction $B(R)$ where states evolve dynamically in such a way that the measure is always shrinking as $R$ is approached. The converse; i.e., a region away from which the measure is always growing; is called a \emph{repellor}. Attractors in particular are important in the study of dynamical systems as they signal the emergence of stable structures that may be used as records (these stable structures may themselves not be static). We thus find that the friction-like behaviour of dynamically similar systems entails a rich set of non-trivial structures --- Janus points, attractors and repellors --- that don't exist in conservative systems. We now study how these structures appear in important applications of general relativity.
\section{Dynamical similarity in the universe}
\label{sec:applications}
In this section we apply our analysis of symmetry in the context of modern cosmology. We argue that standard practice in cosmology is consistent with the universal context of dynamical similarity. We use the PESA to construct a set of Poincar\'e observables for cosmology and show that these observables are necessary and sufficient for producing an empirical adequate theory of the background cosmic geometry. The PESA then implies that cosmological models that agree on their Poincar\'e observables should represent the same physical state.
We then apply this analysis to a more general class of homogeneous geometries that are believed to approximate the generic behaviour of general relativistic space-times near a singularity. Remarkably, the autonomous evolution of the Poincar\'e observables for these systems is well-behaved at the singularity. This suggests a general procedure for resolving singularities in general relativity using dynamical similarity. We also use the measure focusing in these systems to argue for a new proposal for an explanation of the arrow of time in terms of Janus points and attractors.
\subsection{Conceptual problems in cosmology}\label{sub:cosmo probs}
Cosmological solutions to Einstein's equations are found by applying the so-called \emph{cosmological principle}. This principle roughly states that on sufficiently large scales the properties of the universe are the same for all observers. This implies certain symmetries among the largest structures of the observable universe. Most typically, these symmetries include homogeneity and isotropy. Homogeneity singles out a privileged notion of simultaneity, and this can be used to demarcate standards of length and duration. The intuitions developed in the previous sections can therefore be imported to the relativistic case.
Many conceptual problems in cosmology involve looking for explanations that increase the credence of a certain hypothesis as compared to some `typical' class of counterfactuals. For example, inflation is said to solve the flatness problem because, given the assumptions of inflation, the observed relative flatness of the current universe is, it is claimed, significantly more likely than if inflation had not happened. Other conceptual problems in cosmology, such as the \emph{horizon} or the \emph{cosmological constant} problems, have a similar logical structure but applied to observations regarding correlations between apparently causally disconnected regions of space-time and the value of cosmic acceleration respectively.
What is necessary to formulate all of these problems is a measure on the space of DPMs of the theory. This measure can be used to compute likelihoods within the theory to quantitatively evaluate explanatory claims. In particular, a measure for computing the likelihood of inflation given some matter field dynamics is essential for evaluating any claims about inflation. But as we saw in Section~\ref{sub:friction_janus_points_and_attractors}, dynamical similarity implies that the natural measure is time-dependent. The resulting measure focusing therefore significantly affects any explanatory inference made in cosmology if dynamical similarity is found to be a universal symmetry.
To illustrate these implications, we apply our formalism to homogeneous and isotropic cosmologies and compute the measure focusing in the natural measure of those theories. We find that the measure focusing due to dynamical similarity leads to surprising insights regarding the nature of well-known measure ambiguities for such systems.
\subsubsection{Dynamical similarity in Friedmann--Lema\^itre--Robertson--Walker cosmology}
\label{sub:friedmann_lema^itre_robertson_walker_cosmology}
Let us now apply our formalism to a theory that can describe the first-order behaviour of our observable universe after the onset of inflation. The theory in question is flat Friedmann--Lema\^itre--Robertson--Walker (FLRW) cosmology. This theory assumes a flat, homogeneous and isotropic spatial geometry coupled to scalar matter fields $\phi_i(t)$ with an arbitrary potential $V(\phi)$. A complete mathematical treatment of dynamical similarity in this system is given in \citet{Sloan:2019wrz}.
An Ansatz for the metric of the flat FLRW theory can be expressed in terms of the volume of (a fiducial cell within) the spatial manifold, $v$:
\begin{equation}
\de s^2 = -\de t^2 + v^{2/3} \left(\de x^2 + \de y^2 + \de z^2\right)\,.
\end{equation}
Readers may be more familiar with this metric expressed in terms of a scale factor $a=v^{1/3}$. We have chosen to use the volume instead of the scale factor in this case as it is more convenient for constructing the invariant algebra and equations of motion.
The Lagrangian for this theory is the Einstein-Hilbert action plus the usual scalar field action $S=\int \sqrt{-g} \lf[R - \frac 1 2 \nabla \phi \cdot \nabla \phi - V(\phi)\ensuremath{\right}]$ restricted to homogeneous and isotropic fields. Rendered in terms of our choice of variables this leads (after removing a boundary term) to the Lagrangian density
\begin{equation}\label{eq:FLRW L}
\ensuremath{\mathcal {L}} = v \left(-\frac{2}{3} \frac{\dot{v}^2}{v^2} + \sum_i \frac{\dot{\phi}^2_i}{2} - V(\vec{\phi}) \right)\,.
\end{equation}
It is straightforward to find a representation of the dynamical similarity transformation, $D$, for this system as it acts only on the volume and not on $t$. Thus, $D: v \rightarrow \alpha v$, which suggests using the spatial volume $v$ as the standard of length for the system. That $v$ is not an invariant of the original system of variables and can thus be used to set a length standard for the system is somewhat unsurprising, and relates to a well-known feature of flat FLRW cosmology that solutions are independent of an overall choice of scale factor (or volume).
Following the procedure described in Section~\ref{sub:geometric_structure} and using $\rho = v$, we find a simple representation of the Poincar\'e observables to be the Hubble expansion parameter $H=\frac{\dot{v}}{3v}$, the scalar fields $\phi_i$ and their invariant velocities $\dot \phi_i$.\footnote{ A representative of the symplectic potential for the FLRW theory is $\theta_\text{FLRW} = - v \de H + \pi_\phi^i \de \phi_i$, where $\pi_\phi^i = v \dot \phi^i$. Using $\rho = v$, we find an invariant contact form: $\eta_\text{FLRW} = - \de H + \dot\phi^i \de \phi_i $, which leads to the invariant algebra above. The contact Hamiltonian is then $ \ensuremath{\mathcal {H}} ^c_\text{FLRW} = \ensuremath{\mathcal {H}} _\text{FLRW}/v $ and the contact equations of motion \eqref{eq:contact ham} can be shown to lead to the equations \eqref{Klein-Gordon}. Note that $ \ensuremath{\mathcal {H}} _\text{FLRW}$ vanishes on dynamical trajectories, which leads to some simplifications.\label{ftn:FLRW geom} } Given the invariance of $t$ and $\phi$, the only non-trivial invariant is the Hubble parameter $H$. This encodes relative changes of scale in the universe and is observable in terms of the red-shifting of light emitted from distant galaxies. Thus although the overall scale (expressed in terms of volume or scale factor) cannot be determined, its relative change (i.e. the Hubble parameter) can.
To understand how this relative change is manifested in a dynamically similar terms, we can examine the dynamical equations of the theory. In terms of the Poincar\'e observables listed above, the contact equations \eqref{eq:contact ham} for this system can readily be computed (see Footnote~\ref{ftn:FLRW geom} for more details). They are found to be expressible as the well-known Klein-Gordon and Friedmann equations of standard cosmology:
\begin{equation}
\label{Klein-Gordon} \ddot{\phi_i} + 3H \dot{\phi_i} + \frac{\partial V}{\partial \phi_i}=0 \quad H^2 = \frac{8\pi}{3}\left(\sum_i \frac{\dot{\phi}_i^2}{2} + V(\vec{\phi})\right)\,.
\end{equation}
In \cite{Sloan:2020taf} these were obtained using a variational principle due to Herglotz which makes no reference to the volume $v$. The set of invariants described by the dynamical system above can be expressed as an autonomous system, for example, by using one of the scalar fields as an internal clock. Once expressed in this way, the FLRW theory has been written as a set of autonomous equations invariant under dynamical similarity.
The equations we have obtained are the same equations used to described the evolution of the first order background fields in inflationary cosmology. Given that inflationary theory is empirically adequate and, in particular, that a specification of the value of $v$ is not required in inflationary theory for empirical adequacy, the necessity criterion \eqref{crit:necessity} of the PESA and empirical adequacy suggest that \emph{dynamical similarity is a universal symmetry of inflationary theory}. Put differently, current best practice in cosmology is operating in the universal context of dynamical similarity. In this context, representations of states differing by a dynamical similarity transformation, and therefore by a choice of $v$, should be physically identified. While the gauge character of $v$ is widely acknowledged by cosmologists, many of the consequences of treating dynamical similarity as a universal symmetry discussed in Section~\ref{sub:friction_janus_points_and_attractors} are not. We consider some of these consequences now.
\subsubsection{Measures focusing, Janus points and attractors in FLRW}
\label{sub:measure fosusing in FLRW}
The most important characteristic of the scalar field equations~\eqref{Klein-Gordon} is the friction-like character of the term linear in $\dot\phi_i$. To understand the effects of this term, consider choosing a quadratic potential for $V(\phi)$. In this case, the equations are mathematically identical to the equations of motion of a set of harmonic oscillators subject to a frictional term proportional to $H$. In a representation in terms of $v$, the term linear in $\dot \phi$ is seen as due to the expansion of universe via the (re-scaled) velocity $H$ of $v$. This can be used to model the red-shift phenomena often associated with expansion. In the invariant picture, the red-shift is described in terms of a damping of the frequency of the scalar field oscillations. Universal expansion in this context is therefore accounted for in the scale-free system as a friction-like term in \eqref{Klein-Gordon}. Moreover, it is only through this frictional term that the $\phi_i$ fields interact: were the Hubble parameter zero, there would be no coupling between the fields at all. The frictional term is therefore responsible for much of the non-trivial behaviour of the system.
To understand the effect of this frictional term in FLRW cosmology, let us examine its implication towards the construction of measures. Using our choice of variables, the scalar fields and their momenta are invariants of $D$. Since these form a canonical pair, the only non-symplectic element of the invariant algebra is the Hubble parameter $H$. As such the contact form can be represented as $\eta=-dH + \dot{\phi}\, \de\phi$ (see Footnote~\ref{ftn:FLRW geom} for more details). As discussed in Section~\ref{sub:friction_janus_points_and_attractors}, the natural notion of volume on the invariant space is $\mu_c = \eta \wedge d\eta^{N}$, where $N$ is the number of scalar fields. For simplicity of exposition we will restrict ourselves to the case of a single field.
Our goal is to use $\mu_c$ to compute a measure on the space of DPMs of the theory. To do this, we must first remember that $\mu_c$ is a measure on the invariant state space, and not the invariant space of models. We can construct the space of models by choosing a clock internal to the system and then use this clock to define a surface of initial values. Such a surface will then parametrise the space of DPMs of the theory because each choice of initial state with the same value of the internal clock is an initial condition for a different DPM. We can then define a measure on the space of DPMs by restricting $\mu_c$ to this surface. As a final step, we use the fact that along a DPM the system obeys the Friedman equation (i.e., the second equation of \eqref{Klein-Gordon}). If we use a surface of constant red-shift $H = H_0$ to define an internal clock and solve $\dot\phi$ using the Friedman equation, we can define the following measure on the space of DPMs for the single-field FLRW theory:
\begin{equation}\label{eq:GHS measure}
\Omega_{DS} = \sqrt{\frac{3H^2_0}{4\pi} - V(\phi)} d\phi\,.
\end{equation}
This measure is very similar to that first developed in \cite{Gibbons:1986xk} (GHS), and later used in \cite{Gibbons:2006pa} to argue that inflation is fine-tuned. The key difference is that the GHS measure was applied to the space of DPMs of the original symplectic system and therefore included $v$: $\Omega_{GHS}=\Omega_{DS} \wedge \de v$. The difficulty with this measure is that the range of $v$ is unbounded, and therefore any measure defined in this way is subject to regularisation ambiguities. These regularisation ambiguities can be seen through the $H_0$ dependence of $\Omega_{DS}$: the measure depends on the choice of initial value surface unlike the Liouville measure of symplectic systems. Ironically, the authors note that physical processes such as inflation can be defined without reference to $v$ but do not identify dynamical similarity as an explicit symmetry of the system. More importantly, they do not connect the time dependence of the measure to the non-symplectic character of the gauge symmetry inherent to this system.
When examining the FLRW system as a friction-like system resulting from a non-symplectic symmetry, we can examine the effects of measure focusing discussed in Section~\ref{sub:friction_janus_points_and_attractors}. Applying equation~\eqref{eq:focusing} to this case, we find the simple result:
\begin{equation}\label{eq:vol weight}
\dot \eta = \frac{3H}{4\pi} \eta \rightarrow \eta_2 = \eta_1 Exp \left[ \frac{3}{4\pi} \int_{t_1}^{t_2} H dt \right] \rightarrow \eta_2 = \frac{v_2}{v_1} \eta_1 \,.
\end{equation}
Because the natural dynamically similar measure is constructed from the contact form, dynamical similarity implies a natural `volume weighting' to measures \cite{Sloan:2015bha} (where by `volume' we mean spatial volume of a fiducial cell). Thus, solutions which undergo the most expansion between two surfaces of constant red-shift will be those on which the measure focusses \citep{Sloan:2016nnx}. An immediate consequence of this is that de~Sitter expansion is an attractive fixed point of the dynamics since space begins to expand exponentially near such a point. Moreover, near the big bang at $v \to 0$ the measure diverges and therefore obeys a necessary condition for a type-2 Janus point as defined in Section~\ref{sub:friction_janus_points_and_attractors}. Indeed, in Section~\ref{sub:sing res in GR} below we will see that the evolution of the Poincar\'e observables is well-defined at $v = 0$, and that this does indeed correspond to a type-2 Janus point.
The focusing of the measure when using red-shift as an internal clock is apparent from the $H_0$ dependence of $\Omega_{DS}$ in \eqref{eq:GHS measure}. The ambiguity in choosing a measure on the space of models in FLRW cosmology can then be understood in terms of the time-dependence of the measure due to dynamical similarity. Because there is no preferred instant to define an initial value surface, the time variation of the measure immediately implies a kind of ambiguity. This ambiguity, however, does not extend to the amount of measure-focusing that occurs between two surfaces of constant red-shift, which is fixed by \eqref{eq:vol weight}. Being close to de~Sitter expansion, as is true in the current epoch, thus implies that a significant amount of inflation must have occurred in the past. This lends credence to the hypothesis of an inflationary scenario, as was first pointed out in \cite{Corichi:2013kua}.
Note that the measure-focusing implied by dynamical similarity could have significant implications for many other conceptual problems in cosmology. For example, we will see in Section~\ref{sub:aot} that the time-dependence of the measure can be used to provide an explanation for the arrow of time in terms of the Janus point and attractor of the FLRW theory. Other conceptual problems such as the flatness problem or the horizon problem, which depend upon measures on the space of DPMs in the presence of inhomogeneities, could also be affected. We leave those considerations to future work.
\subsection{Singularity resolution in general relativity}\label{sub:sing res in GR}
Singularities occur in general relativity when the evolution equations become non-prescriptive. The precise mathematical characterisation of a singularity can differ depending on the kind of pathology that one wants to emphasise. In one common usage, a space-time point is said to be singular if some physically significant curvature invariant (e.g., the Ricci scalar) is divergent there. Singularities of this kind, although significant in some regard, are usually not seen as fundamental threats to the theory. A more worrisome notion of singularity occurs when the equations of motion can no longer uniquely determine the evolution of the space-time geometry along time-like geodesics within a finite proper-time. When this occurs, the space-time is called \emph{geodesically incomplete} and general relativity is often said to predict its own demise.
It was proven in the famous Hawking--Penrose theorems of \cite{hawking1970singularities,Hawking:1976ra} that geodesic incompleteness is an ubiquitous feature of cosmological space-times in general relativity. The theorems state that if expansion is occurring and the matter content of the universe obeys certain causally motivated conditions,\footnote{ More specifically, if a finite region along a slicing of the space-time is backwardly trapped and the matter content obeys the null energy condition.} then the space-time will be geodesically incomplete in the past. If the conditions in question are satisfied everywhere on a space-like hyper-surface, then all time-like trajectories that intersect this hyper-surface will have started in a singular region often regarded as a \emph{big bang}.
What is significant about these theorems is that they involve a failure of the continuity of the evolution equations of the system at this singular region. In the cosmological case, they are examples of an application of the Picard--Lindel\"of theorems,\footnote{ See \cite{lindelof1894application} for the original work and \cite[\S2.9, pp 43-44]{apostol2012modular} for a modern exposition. } which give conditions under which a series of differential equations uniquely determine a solution. Broadly speaking, they show that in a set of (possibly coupled) first order ordinary differential equations, $\dot{x_i} = f_i (x_j)$ there exists a unique solution whenever the functions $f_i$ are Lipschitz continuous in the $x_j$. In the Friedmann equation of FLRW cosmology, for example, this continuity condition is violated as $H \rightarrow \infty$ and $v \rightarrow 0$, and leads to the big bang.
This is where the PESA becomes especially relevant. The PESA demands the essential and sufficient autonomous system required to achieve empirical adequacy. But as we have shown in Section~\ref{sub:measure fosusing in FLRW} above, empirical adequacy can be achieved even when the variable $v$ has been removed. The standard FLRW system does include $v$ and is therefore sufficient but not necessary for producing an empirically adequate theory. One can then immediately ask whether the failure of Lipschitz continuity persists when $v$ is removed from the system? The answer is a resounding NO: the dynamical equations of the invariant system are fully prescriptive through the big bang! Instead, the big bang is replaced by a type-2 Janus point as advertised in Section~\ref{sub:measure fosusing in FLRW}.
This perhaps surprising result was worked out for the FLRW case in \cite{Sloan:2019wrz} and, for more general homogeneous space-times in \cite{Koslowski:2016hds}. To illustrate these results here, we will focus on the algebraically simple case of two free scalar fields evolving on an FLRW background. The simplicity of the example however does not limit the generality of the result. For full details of the more general case, we refer the reader to the referenced papers.
\subsubsection{Singularity resolution in FLRW space-time}\label{sub:FLRW sing res}
To construct an algebra of Poincar\'e observables for the FLRW system, we begin by defining a new set of variables by performing a compactification of the scalar fields to a 2-sphere using the substitution: $\phi_1 = |\tan \beta| \cos \alpha$ and $\phi_2 = |\tan \beta| \sin \alpha$. The singularity of GR lies at $\beta \rightarrow \pi/2$, at which point both $\phi_1$ and $\phi_2$ tend to infinity. The compactification can be seen as taking the $\phi_1 - \phi_2$ plane and placing a sphere tangent to the origin. The point $\phi_1,\phi_2$ is projected onto the sphere by drawing a line from the centre of the sphere to $(\phi_1, \phi_2)$. The variable $\alpha$ corresponds to the longitude on the sphere at which the point intersects the sphere, and $\beta$ the angle this line makes with a line from the origin of the plane to the centre of the sphere as shown in Figure~\ref{Projection}.
\begin{figure}[h]
\includegraphics[width=\linewidth]{Flavian2.pdf}
\caption{The projection from the $\phi_1 - \phi_2$ plane onto the sphere in $\alpha,\beta$. The singularity of general relativity is at $\beta=\pi/2$ which is a regular point of the equations of motion on the sphere.}
\label{Projection}
\end{figure}
One can then examine the equation of motion for $\alpha$ in terms of $\beta$. We note here that $\dot{\phi}_1$ and $\dot{\phi}_2$ also tend to infinity at the singularity. We could consider compactifying them in the same way as we did the positions $\phi_1,\phi_2$, onto a second sphere $(\theta,\chi)$. However, were we to do so, we would quickly see that at the singularity, $\theta \rightarrow \alpha$. By examining the equations of motion for the scalar fields, we see that although $\theta-\alpha$ tends to zero, $\lambda = \frac{\sin(\alpha-\theta)}{\cos \beta}$ remains non-zero and finite. This leads us to express the dynamics of our system in terms of $\alpha$ and $\lambda$:
\begin{equation}
\frac{\de \alpha}{\de \beta} = \frac{\lambda}{\sin \beta \sqrt{1-\lambda^2 \cos^2 \beta}} \quad \frac{\de \lambda}{\de \beta} = \frac{\lambda}{\cos \beta} \left(\sin \beta - \frac{1}{\sin \beta} \right)\,.
\end{equation}
Remarkably these equations remain finite as we approach the singularity; in the limit as $\beta \rightarrow \pi/2$, we see $\frac{\de \alpha}{\de \beta} \rightarrow \lambda$ and $\frac{\de \lambda}{\de \beta} \rightarrow 0$. Because of this, they satisfy the conditions of the Picard--Lindel\"of theorem, and thus have a unique solution about this point. In fact the system of equations can be explicitly integrated. The solutions are geodesics on the sphere. This is the expected geodesic motion on the $(\phi_1,\phi_2)$ plane in the absence of potentials. Since this autonomous system remains well-posed at the point corresponding to the singularity of general relativity, we can use it to continue the solutions beyond this point and thus achieved evolution through the big bang. This evolution, however, goes through a type-2 Janus point as the measure has a divergence at $H\to \infty$ (i.e., at the point corresponding to the big bang in the redundant representation) according to \eqref{eq:vol weight}. Thus, the big bang is replaced by a smooth type-2 Janus point in the dynamical system prescribed by the PESA and the requirement of empirical adequacy.
\subsubsection{General singularity resolution}
An obvious question regarding the results of the previous section is whether they can be generalised to more generic space-times. After all, the power of the Hawking--Penrose singularity theorems is that they apply to generic classes of physically plausible solutions of the Einstein equations. However, a conjecture by Belinkskii, Khalatnikov and Lifschitz\footnote{ See \cite{Belinsky:1970ew,Uggla:2003fp,Ashtekar:2011ck,Belinski:2017fas}.} suggests that the generic behaviour of a physically plausible class of solutions of general relativity near singularities can be described by homogeneous but anisotropic space-times. Such space-times go by the name of \emph{Bianchi models} in which space-time is split into the product of a three-dimensional homogeneous spatial manifold and a one dimensional temporal manifold valued on the real line. The Bianchi models are classified into eleven types (nine labelled by Roman numerals with two subtypes) according to the structure of their spatial symmetries. One can therefore dramatically extend the generality of our results by applying our formalism to the Bianchi models. This has been achieved in \cite{Koslowski:2016hds}, where a rational reconstruction of the analysis suggests that the PESA can be used to produce a singularity-free evolution in terms of a dynamically similar algebra of Poincar\'e observables. We review these results now.
It is convenient to express the Bianchi space-times in terms of \emph{Misner variables}, which are covariant vector fields $\sigma_i$ on the spatial manifold. In terms of these variables, the metric is
\begin{equation}
\de s^2 = -\de t^2 + v^{2/3} \left( \exp \left(-\frac{q_1}{\sqrt{6}} - \frac{q_2}{2} \right) \sigma_1^2+ \exp \left(-\frac{q_1}{\sqrt{6}} + \frac{q_2}{2} \right) \sigma_1^2 + \exp \left(-\frac{\sqrt{2}}{3} q_2 \right) \sigma_3^2 \right)\,,
\end{equation}
where the normalizations of $q_1$ and $q_2$ have been chosen so that when expressed as a Lagrangian theory, their conjugate momenta appear as scalar fields subjected to a potential. This potential is determined by the symmetries of the spatial manifold, but it always takes a standard form as a function of the Misner variables multiplied by the volume raised to the one-third power; i.e., $v^{\frac{1}{3}} V_s (q_1, q_2)$. This homogeneity in $v$ is central in establishing the dynamical similarity of these models because it allows re-scalings of the volume to be understood independently of the remaining variables. Up to a boundary term which removes the second derivatives of $v$ from the equations, the Lagrangian for these models can be written as
\begin{equation}
\ensuremath{\mathcal {L}} = v \left(-\frac{2}{3}\frac{\dot{v}^2}{v^2} + \frac{\dot{q_1}^2}{2} + \frac{\dot{q_2}^2}{2} - v^{\frac{1}{3}} V_s (q_1, q_2) \right) \,.
\end{equation}
The Lagrangian is very similar to that for the FLRW models \eqref{eq:FLRW L}, though instead of matter fields we have anisotropies, and the coupling to their potential has a different dependence on the volume.
Using the volume scaling of these fields, we can find a representation for the dynamical similarity transformation for this model as $D: t \rightarrow \alpha t, v \rightarrow \alpha^3 v$. This has an intuitive geometrical interpretation: it rescales each of the four dimensions of space-time equally $\de s^2 \rightarrow \alpha^2 \de s^2$.
From the form of $D$ we can find the autonomous algebra, $ \ensuremath{\mathcal {A}} $, of Poincar\'e observables following the general procedure outlined in Section~\ref{sub:geometric_structure}. One such representation that is also dynamically closed is given by (using $\rho = 2 v^{2/3}$)
\begin{equation}
\ensuremath{\mathcal {A}} = \lf\{q_1, \, q_2,\, \Psi_1= v^{\frac{1}{3}}\dot{q}_1,\, \Psi_2= v^{\frac{1}{3}}\dot{q}_2, \, \Phi = -\frac{2\dot{v}}{v^{\frac{2}{3}}} \ensuremath{\right}\}\,.
\end{equation}
The equations of motion \eqref{eq:contact ham} take the form
\begin{equation}
\dot{q_i} = v^{-\frac{1}{3}} \Psi_i \quad \dot{\Psi_i} = v^{-\frac{1}{3}} \left(\frac{\partial V_s}{\partial q_i} - \frac{\Phi \Psi_i}{3} \right) \quad \dot{\Phi} = v^{-\frac{1}{3}} \left(\frac{\Psi_1^2}{2} + \frac{\Psi_2^2}{2} + \frac{\Phi^2}{6} \right)\,.
\end{equation}
As in our previous examples, we see that an autonomous system independent of $v$ (which we've chosen to set the length standard for this system) can be constructed by using any of the above invariants as internal clocks. This is of particular importance in addressing the issue of singularities. The equation of motions for the system, when written in the original set of variables, has terms that are proportional to $v^{-1}$. Hence the dynamical equations are not Lipschitz continuous at $v=0$, and the evolution becomes singular.
Rather remarkably, in the autonomous system, this problem is bypassed. The invariant equations remain Lipschitz continuous throughout the evolution as they are independent of the problematic variable $v$ and hence can be evolved beyond the apparent singularity. This property has been shown to hold in several systems \citep{Koslowski:2016hds}. In Bianchi systems we can compactify $\Psi_1, \Psi_2$ on a sphere such that $\Psi_1 = |\tan \beta| \cos \alpha$ and $\Psi_2 = |\tan \beta| \sin \alpha$ following the procedure used for the FLRW theory of Section~\ref{sub:FLRW sing res}. Doing so we see that the singularity of GR is at $\beta = \pi/2$. Incredibly, this turns out to be a regular point of the equations of motion of the invariant system.
\subsubsection{Singularity resolution in black holes?}
A natural question to ask at this point is whether the elimination of singular behaviour seen in the previous sections also occurs in the case of black holes. While we will not claim here to provide a definite answer to this question, we will highlight some on-going investigations using idealised models that suggest that dynamical similarity may have important implications for the physics of black holes.
The interior of a black hole corresponds closely to a Kantowski--Sachs cosmology, which has the spatial topology of a 2-sphere crossed with a real direction. Under interchange of the time and radial directions this correspondence is made exact, and so we can examine the evolution of the interior of a black hole towards the singularity by examining the evolution of a Kantowski--Sachs space-time towards its big bang. The Kantowski--Sachs space-times can be represented in a way very similar to the Bianchi models presented above. Applying similar machinery, one can find that an autonomous system invariant under dynamical similarity can be constructed and has a well-defined evolution through the point in the dynamics that is singular in general relativity \citep{sloan-mercati:BHres}.
As this construction produces a well-defined dynamical system when the general relativistic dynamics becomes singular, the question arises as to how this result can be squared with the singularity theorems of Hawking and Penrose. The key to understanding this is to note that the essential and sufficient autonomous system that is well-defined through the general relativistic singularity is not itself well-represented by a Lorentzian space-time. At the point of the singularity, the spatial manifold is found to change its handedness.\footnote{ See \cite{Koslowski:2016hds} for details. } This means that the sign of the determinant of the spatial metric switches sign precisely when the system passes the singular point.
This is fatal for a description of the system in terms of space-time. The geodesic equation requires the determinant of the metric to have a well-defined sign, and therefore breaks down at this point. When this happens, the space-time representation of the system becomes geodesically incomplete in accordance with the Hawking--Penrose singularity theorem. However, in the universal context, a change in the handedness of space leads to an indiscernible change in the system since the internal relations between events remain unchanged.\footnote{ For a lengthy discussion of this point in relation to a classic discussion initiated by Kant, see \cite{pooley2003handedness}. } Geodesic completeness of this kind in therefore not sufficient to ensure that the empirically relevant observables of the system are badly behaved. Indeed, the Poincar\'e observables for this system do have well-defined evolution through the point at which parity is changed, and this is all that is required to describe the physical properties of this system. Of course, it remains to be seen whether such a description of the system can ever be found to be empirically adequate. There are many open questions left to resolve. Nevertheless, the Hawking--Penrose singularity theorems are ineffective at signalling a breakdown of determinism in the dynamically similar theory.
\begin{comment}
***Old section on measures****
A measure assigns a positive real number to a set of possible observations made in a theory. Measures on models play an array of roles in physical theories. They are the basis of statistical mechanics, with entropy being a count of the number of microstates compatible with a given macrostate. Quantum mechanics has as its basic entity a wavefunction which is a distribution across a set of observables. Probabilities in physics are often couched in the language of measures, with likelihoods being described as the relative sizes of set of solutions under a given measure.
In Hamiltonian terms, our measures are formed on a $2n$ dimensional phase-space, $\Gamma$. Any volume form on phase-space is proportional to the Liouville form, given by taking powers of the symplectic two-form to create a $2n$-form: $\Omega_L = \omega^n = (\de P_x \wedge \de x)^n $. We then pull this volume form back to the space of equivalence class members.
Applying the PII, we avoid multiply counting solutions by symmetry reducing the map of the space of solutions to the space of distinguishable solutions. For example, when working with the Kepler problem, there is a symmetry under translation of the system, corresponding to a freedom to pick the origin of coordinates. Systems which differ only by an overall translation thus form an equivalence class. We typically fix this symmetry by placing the centre of mass of the system at the origin of the coordinates. We count solutions by equivalence class under symmetries eliminating an infinite choice. In this way each symmetry usually is used to remove two degrees of freedom from the system. Reducing the system by fixing translation invariance in the Kepler example we fix both the position and the momentum of the centre of mass (usually to zero). To reduce the measure on the original space we pull back the measure to surfaces which gauge fix. Since dynamical similarities are not gauge fixed in the same way, we cannot pull back to a constant surface to gauge fix in the same way. In the Kepler problem this would be like fixing $r$ for all time rather than at an instant, which is clearly not compatible with the equations of motion. Hence when reducing by dynamical similarity only a single degree of freedom is removed from the system, and so we arrive at an odd dimensional space.
In the cause of standard canonical transformations such as translations, the symmetry transformations are symplectomorphisms and thus Liouville's theorem ensures that the volume form $\overleftarrow{\Omega}_L$ is preserved under time evolution. However, since dynamical similarities necessarily rescale the symplectic structure this is no longer the case; any volume form on a space of intrinsically distinguishable models will not be time invariant. Liouville's theorem continues to hold in the original space, however since the orbit of the dynamical similarity is unbounded (i.e. we can rescale by any positive real number) this space will be have infinite volume. To arrive at finite measures, typically cut-offs are used to place limits on the interval. These cut-offs however are not time independent; different fractions of the solution space will lie within the bounds at different times. The set of solutions that leave the interval defined by the cut-off may depend on the state of the system, and hence the cut-off weighs solutions differently at differing times, and hence appears at an attractor\cite{Corichi:2013kua}.
When working with the space of distinguishable solutions, the counterpart to the symplectic geometry used in phase-space is a contact geometry (CITE). It has been shown by one of us that the dynamics of this space can be obtained by a geometrical procedure from the phase-space of the original system \cite{Sloan:2018lim}. An alternate way to see the effect of fixing our choice of gauge is that we take interior products with the generators of symmetries. Following the same process we again see that we arrive at an odd dimensional measure on the space of distinguishable solutions. This measure is formed from the contact form on the odd dimensional space, and is the counterpart to the symplectic measure. The one key distinction is that the volume form constructed from the contact form is not time independent\cite{2020arXiv200615123B}. This focussing/dispersion of a measure is a hallmark of frictional systems. One way to understand this is to consider that in the Lagrangian system before we have reduced by dynamical similarity, we have a conserved measure. If we pick a set of solutions and calculate the volume that they occupy on phase-space, this will be conserved. However, during evolution, this set of solutions may contract or expand along the indistinguishable direction, and thus to conserve volume should expand or contract on the distinguishable directions respectively. Hence any measure made on the lower dimensional distinguishable space will undergo focussing. The counterpart of the Liouville measure is given $\Omega_C = \eta \wedge d\eta^n$, where $\eta$ is the contact form, the odd-dimensional counterpart to a symplectic potential. In Darboux coordinates, $\eta = -dA + p dq$, where $A$ is the frictional direction - the conjugate direction to the one removed by dynamical similarity. Unlike the symplectic potential, the contact form is not preserved by time evolution, but in fact evolves according to
\begin{equation}
\dot \eta = - \frac{\partial \ensuremath{\mathcal {H}} ^c}{\partial A} \eta
\end{equation}
As the measure undergoes focussing, there are special points on the evolution of any solution at which the measure will be at a (local) minimum or maximum. These are the points at which the system has found an attractor or repeller. From the evolution equation of the measure on shape space, we see that at these points there is no `friction-like' term acting, so the system is locally symplectic at this point. Equivalently, one can think of these as being the points on the evolution of the system where no friction acts. In the Kepler example they correspond to the perihelion and aphelion of the orbits. In cosmological systems they correspond to maximum expansion of the system and the singularity. They are points of qualitative symmetry along a solution; the measure will either increase in both the past and future of these points, or decrease in both. As such these represent what are known as the `Janus points' (CITE JULIAN ET AL).
\end{comment}
\subsection{The problem of the arrow of time}\label{sub:aot}
At a macroscopic level, processes occur in a way that defines a significant gradient in time that is persistent across the observable universe. No supernova has been seen to implode into a star while a hot cup of coffee cools on a desk. Astrophysical processes and earthly thermodynamic systems alike seem to unfold in a consistently time-asymmetric way. This monotonic gradient in time defines a persistent \emph{arrow of time} across the known universe. Certain quantitative properties of a system --- most commonly the entropy --- can be used to give quantitative meaning to this gradient, and therefore can be used to define time's arrow. More generally, \cite{price2002boltzmann} defines an arrow of time in terms of the following criteria:
\begin{quote}
``(i) ... an asymmetry of physical processes in time, not with an asymmetry in time itself; (ii) that the objective asymmetry concerned comprises a monotonic \emph{gradient}, rather than an increase or decrease; (iii) that the asymmetry in nature is a matter of numerical imbalance between temporal mirror images, not of literal reversibility; and (iv) that if need be the term `entropy' itself is to be thought of as a kind of variable place-holder for the relevant properties of a vast list of actual physical asymmetries.''[Original emphasis.]
\end{quote}
A problem occurs when attempting to account for the observed arrows of time in terms of the microscopic laws of physics. These laws do not exhibit a significant amount of time asymmetry --- certainly nothing like the amount required to explain the enormous temporal gradients observed across our universe.
A prominent approach to this problem is to postulate a \emph{Past Hypothesis} \citep{albert2009time,goldstein2001boltzmann} whereby the earliest known states of the universe are postulated to be highly atypical. According to the arguments of the Past Hypothesis, if the past states are highly atypical, then the present state is highly likely be vastly more typical because of a tendency of many dynamical systems to evolve towards more typical states.\footnote{See \cite{frigg2009typicality} for a critique of the dynamical aspects of this argument.} Thus, the Past Hypothesis is said to provide an explanation for the existence of a significant gradient in time, which in this case is quantified by the typicality of the state of the early universe.\footnote{ This notion typicality is usually expressed in terms of some version of the Boltzmann entropy of the system.} The Past Hypothesis is thus said to provide an explanation for the arrow of time in terms of fundamentally time-symmetric laws.
What is immediately apparent from the line of reasoning outlined above is that the Past Hypothesis relies very strongly on having an unambiguous measure, stable in time, that can be used to determine the typicality of a particular state of the world. Ambiguity in the measure is problematic because different choices of measure could disagree on the significance or direction of the gradient. Time dependence in the measure is problematic because a time-varying measure could itself be the source of time-asymmetry, and thus undermine the explanatory force of the Past Hypothesis.\footnote{ See \cite{gryb2020new} for more details on about this argument as well as \cite{Curiel:2015oea} and \cite{earman2006past} on ambiguity issues regarding the measure in the Past Hypothesis. }
The arguments of Section~\ref{sub:friction_janus_points_and_attractors}, however, suggest that the natural measures of cosmology, and indeed of any system where dynamical similarity is a universal symmetry, will evolve in time. Moreover, because of Equation~\eqref{eq:vol weight}, the amount of measure focusing between two instants is proportional to the amount of expansion that takes place in the universe between those two instants. This leads to an enormous amount of measure focusing ($\sim 85$ orders of magnitude) between the present epoch and the onset of inflation. Thus, when dynamical similarity is taken into account, the dynamics of the measure itself already provides a gradient in time that is big enough to account for a significant numerical imbalance, and thereby define a cosmological arrow of time according to Price's definition above. Because this arrow of time results from the dynamical equations of the invariant theory, we have a dynamical explanation of a cosmological arrow that does not involve the use of a Past Hypothesis. The dynamical nature of this explanation can be seen from the fact that the friction-like terms of Equations~\ref{eq:contact ham} are not invariant under the time reversal operation $t \to -t$.
It's important to note however that the mere existence of friction-like terms in the measure dynamics is not sufficient to provide an adequate explanation of the arrow of time. One might worry that some form of time-symmetry is encoded in the fact that for every DPM in cosmology there is a matching DPM that is representationally its time-reversed image. Moreover, as we already noted, the significant amount of time-asymmetry observed in our universe must be accounted for. These two issues can be addressed by a scenario first proposed in \cite{Barbour_2014} and expanded upon in \cite{barbour:janus}. We will give a slightly modified account here.
In our account, the space of DPMs of a cosmological theory should contain a dense set of DPMs characterised by a unique Janus point that is in the attractive basin of some attractor for the system. In such a scenario, an observer described by a state in a DPM near the attractor will see an arrow of time pointing from the Janus point to the attractor. The divergent amount of measure-focusing that occurs near the attractor could thus explain the enormous time-asymmetry seen by such observers, and therefore provide the basis for a cosmological arrow of time of the form required by Price.
Moreover, a symmetry of the space of DPMs about the Janus points can then explain how such arrows of time can nevertheless emerge when the system is embedded into a time-symmetric symplectic system. Because the arrows of time are assigned to observes near attractors, two separate arrows of time can then emerge for observers along dynamical trajectories pointing towards different attractors on opposite sides of a Janus point. As we have seen, such a scenario occurs in the cosmological models studied in Section~\ref{sub:cosmo probs} where the generic solutions were seen to exhibit type-2 Janus points in the basins of de~Sitter attractors.
Because Janus points are defined by a dynamical condition on the state space, they need only lie in the attractive basin of the relevant attractor, and can thus belong to a dense set of trajectories which may well be typical according to some choice of measure. Thus, the existence of a Janus point is not a Past Hypothesis itself because Janus points can (and do in the examples explored in \cite{Barbour_2014}) belong to a typical set of trajectories.
An open question remains to show that the cosmological arrow of time defined by measure focusing can be used to recover the local arrows of time (e.g., the thermodynamic and perceptual arrows, etc) familiar to our everyday experience of the world. This is the subject of on-going investigations.
\section{Conclusions}
\label{sec:conclusions}
\subsection{Dynamical similarity and cosmology}
\label{sub:dynamical_similarity_and_cosmology}
We have studied several different aspects of dynamical similarity and have shown that there are good reasons to believe that this symmetry has important implications for cosmology. We found that removing the ``extra mathematical hooks'' on which similarity hangs reveals a well-posed system with a smooth Janus point replacing a violent big bang. Attractors resulting from friction-like equations lead to an arrow of time for observers, like us, nearing such attractors. When the universe is stripped of its dynamically similar structure, it appears simultaneously less pathological and more comprehensible.
But are we justified in treating as surplus what seems like vital representational structure in some of our most successful theories? We believe so. To address this question, we have developed a conceptual framework for identifying surplus structure in a theory. The principle we proposed --- the PESA --- prescribes that the observable structure of a theory be the structure that must be specified in the form of input data in order to uniquely solve the evolution equations. When applied to cosmology, the PESA combined with the empirical adequacy of the standard model of cosmology implies that dynamically similar structure is indeed surplus.
The implications for treating dynamical similar structure as surplus are profound and, perhaps, counter-intuitive. It was thus helpful for us to build intuition for dynamical similarity by studying explicit examples. Perhaps the oldest known example of dynamical similarity is illustrated in the Kepler problem, which was studied in Section~\ref{sub:kepler_s_third_law}. What we saw there was that it is important to distinguish between situations in which a symmetry should be seen to relate empirically discernible and indiscernible states. Here too, examples were illustrative. By studying the Galilean boost symmetries in different scenarios, we were able to see that empirically discernible states are related by symmetries only when such symmetries act uniformly on subsystems while leaving the remaining system invariant. This intuition was then applied to the Kepler problem. In every example studied, the PESA was found to reproduce orthodox views of symmetry when the theory under study was found to be empirically adequate.
This, we believe, gives us confidence that the PESA is a good guide for determining surplus structure in a theory. Our main motivation for introducing the PESA, however, was to apply it to dynamical similarity in the cosmological setting, where we've obtained novel results regarding the interpretation of cosmological solutions. The crucial element of the PESA in regards to cosmology is the criterion of necessity. This criterion can often be overlooked because a theory can successfully describe the phenomena even if it has more structure than is strictly necessary. But to describe cosmological phenomena, our analysis shows that dynamically similar structure is simply not necessary. No external scale is required to evolve the system: the evolution of the Poincar\'e observables we defined in Section~\ref{sub:friedmann_lema^itre_robertson_walker_cosmology} exactly match their evolution in general relativity except at the big bang. There, it is only surplus structure that ruins the well-posedness of the evolution equations. When this happens, the ``mental structures that we attach to the phenomena'' \citep{vanFraassen:sym_and_surplus} have the potential to mislead. Stripping ourselves of dynamically similar structure removes the initial singularity and suggests an alternative explanation for a cosmological arrow of time. In cosmology, we may therefore have been mislead to believe in a violent beginning and time-symmetric laws.
\subsection{Further implications of dynamical similarity}
\label{sub:further_implications_of_dynamical_similarity}
As we saw in Section~\ref{sec:dynamical_similarity} dynamical similarity is a very general symmetry of any Lagrangian system in both the universal and subsystem contexts. This suggests that dynamical similarity may have many more applications to physical systems beyond the limited number of systems studied in this paper. When the system is conservative, we saw that the mathematics of dynamical similarity is generically described by contact, rather than symplectic, geometry. But while contact geometry is less familiar to most physicists than symplectic geometry, contact geometry is nevertheless known to have a rich mathematical structure.
One potential application of dynamical similarity is in the foundation of thermodynamics and the identification of certain thermodynamical quantities to apparently conservative systems. Recently, contact systems with equations of motion of the form \eqref{eq:contact ham} have been shown to display strong formal analogies to thermodynamic systems \citep{Bravetti:2015kpo,Bravetti:2018rts}. Natural geometric structures on contact manifolds can be used to define thermodynamic quantities such as equilibrium states and notions of entropy by taking advantage of measure focusing and attractors. Using these structures, it is then possible to describe thermodynamic processes such as approach-to-equilibrium and entropy-production. In the universal context, dynamical similarity could then be used to write the evolution equations of apparently conservative systems in terms of quantities that have formal analogies to thermodynamic quantities like entropy. Such an analysis may be extremely valuable in studying gravitational systems --- particularly in cosmology. More generally, dynamical similarity and contact geometry may prove to be a valuable way to gain a deeper understanding of the known thermodynamic properties of general relativistic systems such as black holes.
A final important consideration is the implications of dynamical similarity to quantum mechanics. Because dynamical similarities alter the symplectic structure by rescaling the unit of angular momentum, dynamical similarities will have important implications for the commutation relations of quantum mechanics. In the universal context, Planck's constant $\hbar$ must be determined relative to other standards of angular momentum within the system. At the representational level, this should imply a dynamical similarity. It is unclear what the precise consequences of this would be in a universal quantum system, but it is clear that more research into such questions would be extremely valuable. At the classical level, dynamically similar systems display friction-like dissipative behaviour. At the quantum level, this may generalise to non-unitary evolution equations, perhaps in the form of a Lindblad master equation. The implications of this to the foundations of quantum mechanics, particularly with regard to the role of decoherence, should be further investigated.
\section*{Acknowledgements}
\label{sec:acknowledgements}
We would like to acknowledge many stimulating interactions with Julian Barbour, Flavio Mercati and Tim Koslowski regarding the role of dynamical similarity in the universe and Alessandro Bravetti and Connor Jackman regarding the mathematical aspects of dynamical similarity. We'd also like to acknowledge discussions with Simon Friederich about the arrow of time as well as Karim Th\'ebault and Henrique Gomes about the nature of symmetry. SG's research was funded by a Young Academy Groningen Scholarship.
|
\section{Introduction}\label{sec1}
\setcounter{equation}{0}
It is almost impossible to cite all the contributes of the Laplace transform:
\begin{eqnarray*}
\begin{array}{c}
{\cal L}(f) := \displaystyle \int_0^\infty \exp^{-1}(s\,t) f(t) \, dt = F(s) \,,
\end{array}
\end{eqnarray*}
to the solution of differential equations \cite{gh-os, wid}.
Indeed, the Fourier and the Laplace transforms are the most useful tools in
Analysis and Mathematical Physics.\\
Actually these transforms are nothing but functions acting in function spaces,
so that it is quite obvious that many others transforms can be defined
similar to them.
The Bell polynomials \cite{be} have been applied in many different fields of mathematics.
In order to avoid unuseful repetitions, which would be classified as plagiarism by the
modern \emph{artificial deficiency}, we limit ourselves to recall the articles
\cite{be-r, c-r, dc-r, n-r3}.
Some generalized forms of Bell polynomials already appeared
in literature (see e.g. \cite{fu, r-s}). The multivariare case was also considered
in \cite{b-n-r, n-r1, n-r, f-d-d-b}. Connections with number theory have been examined
in \cite{n-r4, qi}. See also \cite{w-w}.
To the author's knowledge a connection of the two before mentioned topics has not been
considered in literature. This is the subject of the present article which the proposal
to introduce a wide extension of the Laplace tranform by using the Blissard umbral calculus.
Computational techniques in case of analytic functions are also given in the last
sections.
The obtained results, although formal, since they are based on an umbral approach,
allow to consider infinite many other transforms which can be computed by essentially
algebraic methods.
\section{Recalling the Bell polynomials}\label{sec2}
Considering the $n$-times differentiable functions $x=g(t)$ and $y=f(x)$, defined in
given intervals of the real axis, the composite function $\Phi(t):= f(g(t))$, can
be differentiated with respect to $t$, up to the $n$th order, by using the chain rule.\\
\noindent We use the notations:
$$
\Phi_m:= D_t^m \Phi(t), \qquad f_h:= D_x^hf(x)|_{x=g(t)}, \qquad
g_k:= D_t^k g(t).
$$
Then the $n$th derivative of $\Phi(t)$ is represented by
$$
\Phi_n = Y_n(f_1,g_1;f_2,g_2;\dots;f_n,g_n),
$$
where $Y_n$ denotes the $n$th Bell polynomial.\\
\noindent The first few Bell polynomials are:
\begin{eqnarray}
\label{1}
\begin{array}{l}
Y_1(f_1,g_1)=f_1 g_1 \\
Y_2(f_1,g_1;f_2,g_2)=f_1g_2+f_2g_1^2 \\
Y_3(f_1,g_1;f_2,g_2;f_3,g_3)=f_1g_3+f_2(3g_2g_1)+
f_3g_1^3 \\
\dots
\end{array}
\end{eqnarray}
\noindent Further examples can be found in \cite{rio}, p. 49,
where a recursion formula and the explicit expression given by
the Fa\`a di Bruno formula is also recalled.\\
\noindent A proof of the Fa\`a di Bruno formula based on the
{\em umbral calculus} is given in \cite{rom} and \cite{r-r}.
However, the Fa\`a di Bruno is not convenient by the computational point
of view, owing the higher computational complexity with respect to the
recursion.\\
\noindent The traditional form of the Bell polynomials \cite{co} is given by:
\begin{eqnarray}
\label{2}
\begin{array}{c}
Y_n(f_1,g_1;f_2,g_2;\dots;f_n,g_n) = \displaystyle \sum_{k=1}^n B_{n,k}(g_1,g_2,\dots ,g_{n-k+1}) f_k \,,
\end{array}
\end{eqnarray}
where the $B_{n,k}$ satisfy the recursion \cite{co}:
\begin{eqnarray}
\label{3}
B_{n,k}(g_1,g_2,\dots ,g_{n-k+1}) =
\displaystyle \sum_{h=0}^{n-k} \ {{n-1} \choose {h}} \
B_{n-h-1,k-1}(g_1,g_2,\dots ,,g_{n-h-k+1})\, g_{h+1} \,.
\end{eqnarray}
The $B_{n,k}$ functions for any $k=1,2,\dots,n$ are polynomials
in the $g_1,g_2,\dots,g_n$ variables homogeneous of degree $k$ and
{\em isobaric} of weight $n$ (i.e. they are linear combinations of
monomials $g_1^{k_1}g_2^{k_2}\cdots g_n^{k_n}$
whose weight is constantly given by $k_1+2k_2+\ldots +nk_n =n$), so that
\begin{eqnarray}
\label{4}
B_{n,k}(\alpha \beta \, g_1, \alpha \beta^2 g_2,\dots , \alpha \beta^{n-k+1} g_{n-k+1})
= \alpha^k \beta^n B_{n,k}(g_1,g_2,\dots ,g_{n-k+1}) \,,
\end{eqnarray}
and
\begin{eqnarray}
\label{5}
Y_n(f_1, \beta \, g_1;f_2, \beta^2 g_2;\dots;f_n, \beta^n g_n) = \beta^n \,
Y_n(f_1,g_1;f_2,g_2;\dots;f_n,g_n) \,.
\end{eqnarray}
\section{The Blissard problem}\label{sec3}
John Blissard (1803-1875) published in 1861-1862 papers \cite{bli} introducing a symbolic method
showing that some sequences of numbers $\{b_k\}$ can be substituted by powers $\{b^k\}$ so as to
obtain valid formulas. The Bernoulli numbers was shown to be a first example of such a sequence.
The Blissard symbolic method at present is called the \emph{umbral calculus}, a term coined by
J.J. Sylvester.
The modern version of the umbral calculus \cite{r-r, rom} considers the umbral algebra, as the
algebra of linear functionals on the vector space of polynomials, with the product defined by a
binomial type formula. An extensive bibliogaphy of the subject can be found in \cite{dib}.
The so called Blissard problem is described as follows \cite{rio}.\\
Given the formal power series
\begin{eqnarray}
\label{6}
e^{a t} = \sum_{k=0}^\infty \frac{a^k t^k}{k!} =
\sum_{k=0}^\infty \frac{a_k t^k}{k!},
\end{eqnarray}
associted to the sequence $a = \{ a_k \}$, where
\begin{eqnarray}
\label{7}
a^k := a_k, \qquad \forall k \geq 0, \qquad a_0:= 1,
\end{eqnarray}
the solution of the equation
\begin{eqnarray}
\label{8}
e^{at} e^{bt} = 1
\end{eqnarray}
with respect to the unknown sequence $b = \{ b_n \}$, is given by
\begin{eqnarray}
\label{9}
\left\{
\begin{array}{l}
b_0 := 1 , \, \\
\phantom{\rule{1 pt}{26 pt}}
b_n = Y_n ( -1!, a_1; 2!, a_2; -3!, a_3; \dots; (-1)^n n!, a_n), \quad (\forall \, n>0),
\end{array}
\right.
\end{eqnarray}
where $Y_n$ is the $n$th Bell polynomial \cite{rio}.
\section{A first extension of the Laplace transform}\label{sec4}
In \cite{d-r} the Laguerre-type exponentials has been defined, for every integer $r \geq 1$,
according to the equation:
\begin{eqnarray}
\label{10}
\begin{array}{c}
e_r(x):= \displaystyle \sum_{k=0}^\infty \frac{x^k}{(k!)^{r+1}} \, .
\end{array}
\end{eqnarray}
Obviously, for $r=0$, it results: $e_0(x) \equiv \exp(x)$.\\
Actually these functions are a particular case of the Le Roy functions \cite{ler},
and more generally of the generalized Mittag–Leffler functions defined in \cite{g-p},
and deeply studied in \cite{g-r-m}.
A comparison among the functions $[e_r(t)]^{-1}$, $(r=1,2)$,
$\exp(-t)$ and the limit value $\displaystyle \lim_{r \rightarrow + \infty} [e_r(t)]^{-1} = 1/(1+t)$
shows that for every $r\geq 1$ the functions $[e_r(t)]^{-1} \in L^1(0, +\infty)$, while
the limit value $1/(1+t)$ does not satisfy this condition.
Consider the following transforms:
\begin{eqnarray}
\label{11}
\begin{array}{c}
{\cal L}_{1}(f) := \displaystyle \int_0^\infty [e_1(s\,t)]^{-1} f(t) \, dt = \displaystyle \int_0^\infty
\left[ \displaystyle \sum_{k=0}^\infty \frac{{(s\,t)}^k}{(k!)^2} \right]^{-1} f(t) \, dt = F_1(s) \,,
\end{array}
\end{eqnarray}
\noindent and in general:
\begin{eqnarray}
\label{12}
\begin{array}{c}
{\cal L}_{r}(f) := \displaystyle \int_0^\infty [e_r(s\,t)]^{-1} f(t) \, dt = \displaystyle \int_0^\infty
\left[ \displaystyle \sum_{k=0}^\infty \frac{{(s\,t)}^k}{(k!)^r} \right]^{-1} f(t) \, dt = F_r(s) \, .
\end{array}
\end{eqnarray}
As in the ordinary Laplace transform, the integrals in equations (\ref{11})-(\ref{12}) exist
for all real numbers $\displaystyle \mathrm{Re}(s)>a$, where the constant $a$, called the convergence
abscissa, depends on the function $f$ and determines the region of convergence. \\
Note that the increasing behaviour of the Laguerre-type exponentials, in the interval $(0, +\infty)$
is lower with respect to the ordinary exponential, so that for any fixed $f$, we can choose,
at least, the same convergence abscissa of the ordinary Laplace transform.
For every $r \geq 1$, an approximation of the transform (\ref{12}) is obtained by using
the truncated Laguerre-type exponential of order $r$, putting for a fixed integer $n$:
\begin{eqnarray}
\label{13}
\begin{array}{c}
\displaystyle \int_0^\infty \left[ \displaystyle \sum_{k=0}^n \frac{{(s\,t)}^k}{(k!)^r} \right]^{-1} f(t) \, dt
= F_r^{[n]}(s) \,.
\end{array}
\end{eqnarray}
Since for $r=0$ the Laguerre-type exponentials give back the ordinary
exponential function, then, in this case, equation (\ref{11}) reduces to the ordinary Laplace
transform, that is it results:
\begin{eqnarray}
\label{14}
\begin{array}{c}
{\cal L}_{0}(f) = {\cal L}(f) := \displaystyle \int_0^\infty \exp^{-1}(s\,t) f(t) \, dt = F_0(s) \, .
\end{array}
\end{eqnarray}
\subsection{The inversion formula}
The Laguerre-type exponentials are monotonic increasing functions in the interval
$(0, +\infty)$, so that can be inverted in the same interval. Therefore, we can \emph{conjecture}
that the transform (\ref{12}) (in particular (\ref{11})), admits the inversion formula:
\begin{eqnarray}
\label{15}
\begin{array}{c}
{\cal L}^{-1}_{r}(f)(t) := \displaystyle \frac{1}{2 \pi {\rm i}} \displaystyle \lim_{\tau \rightarrow \infty}
\int_{\gamma - {\rm i} \, \tau}^{\gamma + {\rm i} \, \tau} [e_r(s\,t)] F_r(s) \, ds \,,
\end{array}
\end{eqnarray}
where $\gamma$ is a real number so that the contour path of integration is in the
region of convergence of $F_r(s)$. It should also be possible to transform the contour
into a closed curve, allowing the use of the residue theorem.
However, the proof of the equation (\ref {15}) is not easy to carry out, since it would be
necessary to introduce an extension of the Fourier transform based on Laguerre exponentials,
a topic still far from being obtained.
\vspace{ .25cm}
\section{The isomorphism \ ${\cal T}_s$ and its iterations}\label{sec5}
In previous articles (see e.g. \cite{r-t}), it was shown that there exist a differential
isomorphism ${\cal T}:= {\cal T}_s$, acting into the space
${\cal A}:= {\cal A}_s$ of analytic functions of the
variable $s$, by means of the correspondence:
\begin{eqnarray*}
D_s:= D \equiv \frac{d}{ds} \; \rightarrow \; {\hat D}_L:= D_s s D_s ; \qquad s
\cdot \; \rightarrow \; {\hat D}_s^{-1}, \qquad
\end{eqnarray*}
where
\begin{eqnarray*}
{\hat D}_s^{-n}F(s):= \frac{1}{(n-1)!}\int_0^s (s -
\xi)^{n-1}F(\xi) d\xi \,.
\end{eqnarray*}
The isomorphism ${\cal T}:= {\cal T}_s$ can be iterated
producing a set of generalized Laguerre derivatives as follows.
According to the results in \cite{r-t} we put, for every integer $m \geq 1$,
\begin{eqnarray*}
{\cal T}_s^{m-1}{\hat D}_L = {\cal T}_s^{m-1} (DsD) = DsDsD \cdots
sD =: {\hat D}_{mL},
\end{eqnarray*}
where the last operator contains $s+1$ ordinary derivatives,
denoted by $D \equiv D_s$.
The action of ${\cal T}_s$, on powers, and
consequently on all functions belonging to \ ${\cal A}:= {\cal
A}_s$ is as follows:
\begin{eqnarray*}
{\hat D}_s^{-n} (1)= \frac{s^n}{n!} \ ,
\end{eqnarray*}
and, by induction:
\begin{eqnarray*}
{\cal T}_s^{m-1} {\hat D}_s^{-1}(1) = {\hat D}_{{\cal
T}_s^{m-1}}^{-1} (1) \quad \Rightarrow \quad {\hat D}_{{\cal
T}_s^{m-1}}^{-n} (1) = \frac{s^n}{(n!)^s} \ . \quad
\nonumber
\end{eqnarray*}
Note that the Laguerre-type exponentials are obtained, acting with these
iterated isomorphisms on the classical exponential, since:
\begin{eqnarray*}
{\cal T}_s^m (e^s) = \sum_{k=0}^\infty \frac{{\cal T}_s (s^k)
}{(k!)^m} = \sum_{k=0}^\infty \frac{s^k}{(k!)^{m+1}} = e_m(s) \,.
\end{eqnarray*}
It has been shown in a number of articles \cite{b-d-r, be-br-ri, br-ce-ri, ce-ge-ri},
that new sets of special functions, namely the Laguerre-type special functions, can be
introduced and some of their applications have been considered in \cite{br-ri, da-he-ri,
dea-ri, m-r3}.
\subsection{Computation via the isomorphisms ${\cal T}_s$}
Acting with the isomorphism ${\cal T}_s$ on both sides of equation (\ref{14}), we find:
\begin{eqnarray*}
\begin{array}{c}
\displaystyle \int_0^\infty {\cal T}_s [\exp^{-1}(s\,t)]\, f(t) \, dt = {\cal T}_s [F_0(s)] \,,
\end{array}
\end{eqnarray*}
that is
\begin{eqnarray*}
\begin{array}{c}
\displaystyle \int_0^\infty [e_1(s\,t)]^{-1} f(t) \, dt = {\cal T}_s [F_0(s)] \,,
\end{array}
\end{eqnarray*}
so that, comparing this result with equation (\ref{11}), we find:
\begin{eqnarray*}
\begin{array}{c}
{\cal T}_s [F_0(s)] = F_1(s) \,.
\end{array}
\end{eqnarray*}
Of course this equation can be generalized starting from (\ref{12}), obtaining:
\begin{eqnarray*}
\begin{array}{c}
\displaystyle \int_0^\infty {\cal T}_s [e_r(s\,t)]^{-1} f(t) \, dt = \displaystyle \int_0^\infty
\displaystyle \sum_{k=0}^\infty [e_{r+1}(s\,t)]^{-1} f(t) \, dt = {\cal T}_s F_r(s) \,,
\end{array}
\end{eqnarray*}
and therefore
\begin{eqnarray*}
\begin{array}{c}
{\cal T}_s [F_r(s)] = F_{r+1}(s) \,.
\end{array}
\end{eqnarray*}
\section{A more general extension of the Laplace transform}\label{sec6}
A further extension of the transforms (\ref{11})-(\ref{12}) is as follows.\\
Given the sequence $a := \{ a_k \} = (1, a_1, a_2, a_3, \dots )$, we consider the function:
\begin{eqnarray}
\label{16}
\begin{array}{c}
\displaystyle \frac{1}{1 + a_1 t + a_2 \frac{t^2}{2!} + a_3 \frac{t^3}{3!} + \dots } \, \quad (t\geq 0).
\end{array}
\end{eqnarray}
When $a_k = 1/(k!)^{r}$ the function (\ref{10}) is recovered, and for $r=0$ we find
again $\exp(-t)$.
Note that the functions (\ref{16}) are complete monotonic functions decreasing from
the initial value $1$, at $t=0$, and vanishing at infinity.
Therefore, according to the umbral method, we put by definition:
\begin{eqnarray}
\label{17}
\begin{array}{c}
{\cal L}_a(f):= \displaystyle \int_0^\infty \frac{f(t)}{\displaystyle \sum_{k=0}^\infty \frac{a_k (s\,t)^k}{k!}}\, dt =
\displaystyle \int_0^\infty \frac{f(t)}{\displaystyle \sum_{k=0}^\infty \frac{a^k (s\,t)^k}{k!}}\, dt =
F_a(s)
\end{array}
\end{eqnarray}
Recalling the Blissard problem, the solution of the the umbral equation
\begin{eqnarray}
\label{18}
\begin{array}{c}
\displaystyle \frac{1}{\displaystyle \sum_{k=0}^\infty \frac{a^k (s\,t)^k}{k!}} =
\displaystyle \sum_{k=0}^\infty \frac{b^k (s\,t)^k}{k!}\,
\end{array}
\end{eqnarray}
that is
\begin{eqnarray}
\label{19}
\begin{array}{c}
\exp[a (s\,t)] \exp[b (s\,t)] = 1 \,,
\end{array}
\end{eqnarray}
is given by equation (\ref{9}).
Therefore, the generalized Laplace transform (\ref{13}) writes:
\begin{eqnarray}
\label{20}
\begin{array}{c}
{\cal L}_a(f) := \displaystyle \int_0^\infty f(t) \left[ 1+ \displaystyle \sum_{k=1}^\infty
Y_k (-1!, a_1; 2!, a_2; \dots ; (-1)^k k!, a_k) \frac{(s\,t)^k}{k!} \right] \, dt = F_a(s) \,.
\end{array}
\end{eqnarray}
By using equation (\ref{2}), equation (\ref{20}) becomes
\begin{eqnarray}
\label{21}
\begin{array}{c}
{\cal L}_a(f) = \displaystyle \int_0^\infty f(t) \left[ 1+ \displaystyle \sum_{k=1}^\infty \sum_{h=1}^k (-1)^h h! \,
B_{k,h}(a_1, a_2, \dots, a_{k-h+1}) \displaystyle \frac{(s\,t)^k}{k!} \right] \, dt = \\
\phantom{\rule{1 pt}{30 pt}}
= \displaystyle \int_0^\infty f(t) \left[ 1+ \displaystyle \sum_{k=1}^\infty \sum_{h=1}^k (-1)^h h! \,
B_{k,h}(a_1, a_2, \dots, a_{k-h+1}) \, \displaystyle \frac{(s\,t)^k}{k!} \right] \, dt = F_a(s) \,.
\end{array}
\end{eqnarray}
It is convenient to introduce the definition
\begin{eqnarray}
\label{22}
\begin{array}{c}
C_k(a):= \displaystyle \sum_{h=1}^k (-1)^h h! \, B_{k,h}(a_1, a_2, \dots, a_{k-h+1}) \,, \quad C_0(a):= 1 \,,
\end{array}
\end{eqnarray}
so that equation (\ref{20}) writes
\begin{eqnarray}
\label{23}
\begin{array}{c}
{\cal L}_a(f) = \displaystyle \int_0^\infty f(t) \displaystyle \sum_{k=0}^\infty C_k(a) \, \displaystyle \frac{(s\,t)^k}{k!} \, dt = F_a(s) \,.
\end{array}
\end{eqnarray}
Note that the ordinary Laplace transform corresponds to the sequence:
$a = (1,1,\dots, 1, \dots)$, that is $a_k \equiv 1\,, \forall k\geq 1$. Therefore, it results:
\begin{eqnarray}
\label{24}
{\cal L}_{(1,1,\dots, 1, \dots)}(f) \equiv {\cal L}(f) \,.
\end{eqnarray}
In this case, we find:
\begin{eqnarray*}
B_{k,h} (1,1,\dots ,1) = S(k,h) \,,
\end{eqnarray*}
where $S(k,h)$ are the Stirling numbers of the second kind \cite{co}.\\
Then
\begin{eqnarray*}
C_k (1,1,\dots ,1) = \displaystyle \sum_{h=1}^k (-1)^h h! \, S(k,h) \,, \quad C_0(a):= 1 \,.
\end{eqnarray*}
Recalling the known identity
\begin{eqnarray*}
\displaystyle \sum_{h=1}^k (-1)^{k-h} h! \, S(k,h) = 1 \,,
\end{eqnarray*}
we find
\begin{eqnarray*}
C_k (1,1,\dots ,1) = (-1)^k \,, \quad \forall k \geq 0 \,,
\end{eqnarray*}
so that the ordinary expression of the Laplace transform is recovered.
\vspace{ .25cm}
A number of sums defining the coefficients $C_k(a)$ corresponding to different sequences
$a=\{a_k \}$ can be found in \cite{qi}, however, in what follows, we will assume the
fundamental hypothesis that preservs the property of the ordinary Laplace transform:\\
\textbf{HP.} \emph{For every fixed} $s$ \emph{in the region of convergence, the power series }
$\sum_{k=0}^\infty C_k(a) \, (s\,t)^k/k!$, \emph{in equation} (\ref{23}) \emph{has an exponential
decay to zero when} $t \rightarrow \infty$.\\
In this framework, another possibility is to assume $a_k = k!$. In this case equation (\ref{16}) becomes:
\begin{eqnarray}
\label{25}
\begin{array}{c}
\displaystyle \frac{1}{1 + t + t^2 + t^3 + \dots } \, \quad (t\geq 0).
\end{array}
\end{eqnarray}
The truncation of the geometric series at the denominator in equation (\ref{25}) produces
graphs corresponding to the sequences
$$(1,1,0,0,0,\dots), \ (1,1,1,0,0,0,\dots), \ (1,1,1,1,0,0,0,\dots).$$
The decreasing character of the corresponding graphs increases as the number of units increase.
\begin{rem} {\rm
Increasing the values of the sequence $\{a_k \}$ in equation (\ref{16}), the corresponding
graphs exibhit a more fast decreasing character.
Then the transforms coresponding to the relevant truncations can be limiteded to a small
interval of the type $[0,L]$, as the values of the function (\ref{16}) become
negligible outside this interval.}
\end{rem}
\section{Properties}\label{sec7}
From the definition (\ref{23}) the following properties are derived:\\
$\bullet$ \textbf{Linearity}
\begin{eqnarray}
\label{26}
\begin{array}{c}
{\cal L}_a (A f_1 + B f_2 ) = A\,{\cal L}_a (f_1) + B\,{\cal L}_a (f_2) \,.
\end{array}
\end{eqnarray}
$\bullet$ \textbf{Homothetic property}\\
Putting: $x\,a :=(x a_1, x^2 a_2, \dots, x^n a_n , \dots)$, by using the isobaric property (\ref{5})
of the Bell polynomials, it results:
\begin{eqnarray}
\label{27}
\begin{array}{c}
{\cal L}_{x\,a}(f) = F_a(x\,s) \,.
\end{array}
\end{eqnarray}
This can be interpreted as an homothety between the space of the $a$ parametes and that of the
variable $s$.
$\bullet$ \textbf{Scaling property}
\begin{eqnarray}
\label{28}
\begin{array}{c}
{\cal L}_{a}(f(d\, t)) = \displaystyle \frac{1}{d} F_a\left(\frac{s}{d}\right) \,.
\end{array}
\end{eqnarray}
\textbf{Proof.} From equation (\ref{23}) we find:
\begin{eqnarray*}
\begin{array}{c}
F_a \left(\frac{s}{b}\right) = \displaystyle \int_0^\infty f(t) \displaystyle \sum_{k=0}^\infty C_k(a) \,
\left(\frac{s}{b}\right)^k \frac{t^k}{k!} \, dt \,,
\end{array}
\end{eqnarray*}
changing variable, putting $t = b\,x$, it results
\begin{eqnarray*}
\begin{array}{c}
F_a \left(\frac{s}{b}\right) = b \,\displaystyle \int_0^\infty f(b\, x) \displaystyle \sum_{k=0}^\infty C_k(a) \,
\frac{(s\, x)^k}{k!}\, dx = {\cal L}_{a}(f(b\, x)) \,,
\end{array}
\end{eqnarray*}
that is equation (\ref{28}), up to the change of name of the variable $t$.
$\bullet$ \textbf{Action on the derivative}
\begin{eqnarray}
\label{29}
\begin{array}{c}
{\cal L}_a(f') = \displaystyle \int_0^\infty f'(t) \displaystyle \sum_{k=0}^\infty C_k(a) \, \frac{(s\, t)^k}{k!} \, dt = \\
\phantom{\rule{1 pt}{26 pt}}
= -\, s \displaystyle \int_0^\infty f(t) \displaystyle \sum_{k=0}^\infty C_{k+1}(a) \, \frac{(s\, t)^k}{k!} \, dt
- f(0) \,.
\end{array}
\end{eqnarray}
\textbf{Proof.} It is sufficient to integrate by parts and to use the above HP.
\section{Computational techniques}\label{sec8}
According to the above definitions, it is possible to prove the theorems
\begin{thm}
Let $f(t)$ be an analytic function on the real axis.
Using the Taylor expansion of the function $f(t)$, centered at the origin:
\begin{eqnarray}
\label{30}
\begin{array}{c}
f(t) = \displaystyle \sum_{k=0}^\infty c_k \frac{t^k}{k!} \,,
\end{array}
\end{eqnarray}
equation \emph{(\ref{23})} writes:
\begin{eqnarray}
\label{31}
\begin{array}{c}
{\cal L}_a(f) = \displaystyle \int_0^\infty \displaystyle \sum_{n=0}^\infty \displaystyle \sum_{k=0}^n {n \choose k}
c_{n-k} \, C_k(a) \, \frac{(s\,t)^k}{k!} \, dt = F_a(s) \,.
\end{array}
\end{eqnarray}
\end{thm}
\textbf{Proof. }- In fact, from equations (\ref{23})-(\ref{30}), by using
the Cauchy product of power series, we find
\begin{eqnarray*}
{\cal L}_a(f) = \displaystyle \int_0^\infty \sum_{k=0}^\infty c_k \frac{t^k}{k!}
\displaystyle \sum_{k=0}^\infty \, C_k(a) \frac{(st)^k}{k!} \, dt = \displaystyle \int_0^\infty
\displaystyle \sum_{n=0}^\infty \displaystyle \sum_{k=0}^n {n \choose k}
c_{n-k}\, C_k(a) \frac{(s\,t)^k}{k!}\, dt \,,
\end{eqnarray*}
that is the result.
\begin{thm}
Let $f(t)$ be a function expressed by the Laurent expansion:
\begin{eqnarray}
\label{32}
\begin{array}{c}
f(t) = \displaystyle \sum_{k=0}^\infty c_k \frac{t^{-k}}{k!} \,,
\end{array}
\end{eqnarray}
then, equation \emph{(\ref{23})} writes:
\begin{eqnarray}
\label{33}
\begin{array}{c}
{\cal L}_a(f) = \displaystyle \int_0^\infty \displaystyle \sum_{n=0}^\infty \displaystyle \sum_{k=0}^n {n \choose k}
c_{n-k} \, C_k(a) \, \frac{ s^k \, t^{-n+2k}}{n!}\, dt
= F_a(s) \,.
\end{array}
\end{eqnarray}
\end{thm}
\textbf{Proof. }- In fact, considering the Cauchy product:
\begin{eqnarray}
\label{34}
\begin{array}{c}
\displaystyle \sum_{k=0}^\infty c_k x^{-k} \sum_{k=0}^\infty a_k x^k = \sum_{n=0}^\infty
\sum_{k=0}^n c_{n-k} x^{-n+k} a_k x^k = \sum_{n=0}^\infty \sum_{k=0}^n a_k c_{n-k} x^{-n+2k} \,.
\end{array}
\end{eqnarray}
from equations (\ref{23})-(\ref{32}), we find
\begin{eqnarray*}
\begin{array}{c}
{\cal L}_a(f) = \displaystyle \int_0^\infty \sum_{k=0}^\infty c_k \frac{t^{-k}}{k!}
\displaystyle \sum_{k=0}^\infty C_k(a) \, \frac{(s\,t)^k}{k!} \, dt =
\displaystyle \int_0^\infty \displaystyle \sum_{n=0}^\infty \displaystyle \sum_{k=0}^n \frac{c_{n-k}}{(n-k)!} \frac{C_k(a)}{k!}
\, s^k t^{-n+2k} \, dt \,,
\end{array}
\end{eqnarray*}
that is the result.
\section{A general isomorphism \ ${\cal T}_s(a)$}\label{sec9}
The results of Section 5 suggests the possibility to introduce a more
general isomorphisms. The isomorphism ${\cal T}_s$, defined in Section 5
is determined by the sequence $a^{[1]}:= (1, 1/2!, 1/3!, \dots )\,$.
Now, given a sequence of nonvanishing real numbers
$a:= (a_1, a_2, a_3, \dots )$, $(a_k \neq 0, \forall k)$, we can define a correspondence
acting into the space ${\cal A}:= {\cal A}_s$ of analytic functions of the $s$ variable
by means of the position:
\begin{equation*}
{\cal T}_s(a) s^n = a_n s^n \,.
\end{equation*}
In particular, the isomorphism ${\cal T}_s$ is recovered, since it results:
${\cal T}_s(a^{[1]}) \equiv {\cal T}_s$.\\
Even if this isomorphism is not derived form a differential operator, it is still
possible to apply it to the generalized Laplace transform (\ref{23}), obtaining the
equation:
\begin{equation*}
{\cal T}_s(a) [F_a(s)] = \displaystyle \int_0^\infty f(t) \displaystyle \sum_{k=0}^\infty C_k(a)
\, a_k \, (s\,t)^k \, dt \,.
\end{equation*}
\section{Conclusion}\label{sec10}
\noindent It has been shown that, by exploiting Laguerre-type exponentials, it is possible
to introduce generalized forms of the Laplace transform that it is supposed to be applied
in the treatment of differential equations that use the Laguerre derivative instead of the
ordinary one. For these transforms it was also possible to deduce the transformed functions
by means of a differential isomorphism studied in previous articles.
The particular form of Laguerre exponentials also suggested a wider extension of the Laplace
transform, which is associated with a sequence of numbers denoted by the umbral symbol $a$.
This extension, through the solution of the Blissard problem, which uses Bell's polynomials
in a natural way, has made it possible to formally define a whole class of transforms, each
of which is associated with a fixed sequence. Some fundamental calculation rules have been
demonstrated for all the generalized transformations considered.
Numerous problems remain open, first of all the existence and the analytical proof of the
inverse transformation, which should be based on the extension of the Fourier transform to
the Laguerrian case, with all the problems related to the study of a completely new Fourier
type analysis.\\
\vspace{ .25cm}
\noindent \textbf{Acknowledgment} - I want to thank Prof. Dr. Francesco Mainardi for let me know
his work on the generalized Mittag-Leffler functions and their applications in the Laplace
transform theory.
\vspace{ .5cm}
\noindent \textbf{\large Compliance with ethical standards}
\noindent \textbf{Funding} \ {This research received no external funding.}
\noindent \textbf{Conflict of interest} \ The author declares that he has
not received funds from any institution.
\vspace{ .25cm}
|
\section{Introduction \label{sec:intro}}
One of the spin-$\frac{1}{2}$ doubly charmed baryons, $\Xi_{cc}^{+}$, was first reported by the SELEX Collaboration~\cite{Mattson:2002vu}. However, the following studies performed by the FOCUS~\cite{Ratti:2003ez}, Belle~\cite{Chistov:2006zj}, and BaBar~\cite{Aubert:2006qw} Collaborations found no evidence on its existence. In 2017, the LHCb Collaboration observed another spin-$\frac{1}{2}$ doubly charmed baryon $\Xi_{cc}^{++}$ in the decay mode $\Xi_{cc}^{++}\to\Lambda_c^+K^-\pi^+\pi^+$~\cite{Aaij:2017ueg}. The latest and more accurate measurement of its mass is $m_{\Xi_{cc}^{++}}=3621.55\pm0.23\pm0.30$~MeV~\cite{Aaij:2019uaz}. Since then, the properties, decay and production mechanisms of doubly charmed baryons have been extensively studied theoretically~(see Refs.~\cite{Chen:2016spr,Liu:2019zoy} and references cited therein).
The masses of the $\frac{3}{2}$ doubly charmed baryons and their spin-$\frac{1}{2}$ counterparts are related to the $D^*$ and $D$ masses via the so-called heavy anti-quark
di-quark symmetry (HADS)~\cite{Hu:2005gf}, i.e.,
\begin{eqnarray}
&&m_{\Xi_{cc}^*}-m_{\Xi_{cc}}=\frac{3}{4}(m_{D^*}-m_D),\\
&&m_{\Omega_{cc}^*}-m_{\Omega_{cc}}=\frac{3}{4}(m_{D_s^*}-m_{D_s}).
\end{eqnarray}
These relations seem to be broken at the level of 25\% based on various theoretical and lattice QCD studies~\cite{Padmanath:2015jea,Chen:2017kxr,Alexandrou:2017xwd,Mathur:2018rwu}. Future discovery of the doubly charmed spin-$\frac{3}{2}$ baryons by the LHCb and Belle II experiments~\cite{Cerri:2018ypt,Kou:2018nap} will contribute tremendously to our understanding of the heavy quark symmetries as well as the non perturbative strong interaction.
The magnetic moments of baryons play a vital role in understanding their internal structure. For the spin-$\frac{1}{2}$ doubly charmed baryons, their magnetic moments have been systematically investigated in the heavy-baryon (HB) chiral perturbation theory (ChPT)~\cite{Li:2017cfz,Li:2020uok}, the extended on-mass shell (EOMS) BChPT~\cite{Liu:2018euh,Blin:2018pmj}, and the light-cone QCD sum rule (LCSR)~\cite{Ozdem:2018uue} after the discovery of $\Xi_{cc}^{++}$.
Up to now, the magnetic moments of spin-$\frac{3}{2}$ doubly charmed baryons have also been examined in a variety of phenomenological models~\cite{Lichtenberg:1976fi,Bose:1980vy,Bernotas:2012nz,Albertus:2006ya,Patel:2007gx,Dhir:2009ax,Sharma:2010vv}, the LCSR~\cite{Ozdem:2019zis}, the HB ChPT~\cite{Meng:2017dni}, and lattice QCD simulations~\cite{Can:2015exa}. It should be stressed that the lattice QCD study~\cite{Can:2015exa} only calculated the magnetic moment of $\Omega_{cc}^{*+}$ for an unphysical $m_\pi\approx156~{\rm MeV}$.
Compared to other phenomenological models, ChPT~\cite{Weinberg:1978kz,Gasser:1984gg,Gasser:1987rb,Scherer:2002tk} provides a systematic expansion of physical observables (magnetic moments in the present case) order by order. The chiral order ${n_\chi}$ is defined as $n_\chi=4L-2N_M-N_B+\sum_kkV_k$ for a given Feynman diagram with $L$ loops, $N_M~(N_B)$ internal meson~(baryon) propagators, and $V_k$ vertices from $k^{\rm th}$ order Lagrangians. In the one-baryon sector, because of the large nonzero baryon masses in the chiral limit, one needs to develop a power counting scheme different from that used for the mesonic sector. Over the years, three approaches have been developed and extensively studied, i.e., the HB~\cite{Jenkins:1990jv,Bernard:1995dp}, the infrared (IR)~\cite{Becher:1999he}, and the EOMS~\cite{Fuchs:2003qc} schemes. A brief summary and comparison of these three approaches can be found in Ref.~\cite{Geng:2013xn}.
The EOMS scheme has been successfully applied to study the magnetic moments of baryons~\cite{Geng:2009ys,Geng:2008mf,Geng:2009hh,Xiao:2018rvd,Liu:2018euh,Xiabng:2018qsd,Blin:2018pmj} in the past two decades, and it was shown that a better description of lattice QCD quark-mass dependent data can be achieved compared to the other two alternatives. In Refs.~\cite{Liu:2018euh,Blin:2018pmj}, the magnetic moments of the spin-$\frac{1}{2}$ doubly charmed baryons were studied in the EOMS BChPT. It was shown that the lattice QCD data of Ref.~\cite{Can:2013tna} can be described quite well. On the other hand, the extrapolated magnetic moments at the physical point are quite different from some of the phenomenological model predictions, which remain to be tested by future experimental measurements.
In the present work, we study the magnetic moments of the spin-$\frac{3}{2}$ doubly charmed baryons up to the next-to-leading order (NLO) in covariant baryon chiral perturbation theory (BChPT) with the extended-on-mass shell (EOMS) renormalization scheme. In addition to predicting the physical magnetic moments, we study their light-quark mass dependence such that they can be used to extrapolate future lattice QCD simulations to the physical point. The tree-level leading order contributions will be estimated by the quark model due to lack of experimental or lattice QCD data. The two low energy constants (LECs) $C$ and $H$ controlling the loop contributions are determined in two different ways: the quark model and the lattice QCD simulation supplemented by the quark model. In addition, to gain more insights into heavy quark spin symmetry, we compare the magnetic moments of the spin-$\frac{3}{2}$ doubly charmed baryons with those of their spin-$\frac{1}{2}$ counterparts, particularly, their light-quark mass dependence and the loop contributions via virtual spin-$\frac{3}{2}$ and -$\frac{1}{2}$ baryons.
This work is organized as follows. In Sec.~II, we introduce the electromagnetic form factors of the spin-$\frac{3}{2}$ baryons, provide the effective Lagrangians, and calculate the pertinent Feynman diagrams up to the next-to-leading order. In Sec.~III, we fix the relevant low energy constants with the help of the quark model and the lattice QCD simulation, predict the light-quark mass dependence of the magnetic moments, compare with previous studies, and examine the loop contributions. A short summary is given in Sec. IV.
\section{Theoretical formalism}
\label{sec:TH}
The matrix elements of the electromagnetic current $J_\mu$ for a spin-$\frac{3}{2}$ doubly charmed baryon can be parameterized as follows~\cite{Nozawa:1990gt,Geng:2009ys}:
\begin{eqnarray}
\langle T(p_f)|J_\mu|T(p_i)\rangle=&&-\bar{u}_\alpha(p_f)\left\{\left[\gamma_\mu F_1(\tau)+\frac{i\sigma_{\mu\nu}q^\nu}{2m_T}F_2(\tau)\right]g^{\alpha\beta}\right.\nonumber\\
&&\left.+\left[\gamma_\mu F_3(\tau)+\frac{i\sigma_{\mu\nu}q^\nu}{2m_T}F_4(\tau)\right]\frac{q^\alpha q^\beta}{4m_T^2}\right\}u_\beta(p_i),
\end{eqnarray}
where $\bar{u}_\alpha(p_f)$ and $u_\beta(p_i)$ are the Rarita-Schwinger~(RS) spinors~\cite{Rarita:1941mf}, $m_T$ is the doubly charmed baryon mass, and $\tau=-\frac{q^2}{4m_T^2}$. The four-momentum transfer is defined as $q=p_f-p_i$. The electric monopole~$G_{E0}$, quadrupole~$G_{E2}$, magnetic
dipole~$G_{M1}$, and octupole~$G_{M3}$ form factors can be expressed in terms of the four electromagnetic form factors $F_i's~(i=1,2,3,4)$, respectively,
\begin{eqnarray}
&&G_{E0}(\tau)=\left[F_1(\tau)-\tau F_2(\tau)\right]+\frac{2}{3}\tau G_{E2}(\tau),\nonumber\\
&&G_{E2}(\tau)=\left[F_1(\tau)-\tau F_2(\tau)\right]-\frac{1}{2}(1+\tau)\left[F_3(\tau)-\tau F_4(\tau)\right],\nonumber\\
&&G_{M1}(\tau)=\left[F_1(\tau)+F_2(\tau)\right]+\frac{4}{5}\tau G_{M3}(\tau),\nonumber\\
&&G_{M3}(\tau)=\left[F_1(\tau)+F_2(\tau)\right]-\frac{1}{2}(1+\tau)\left[F_3(\tau)+F_4(\tau)\right].
\end{eqnarray}
At $q^2=0$, $Q_T=eG_{E0}(0)=eF_1(0)$ is the charge of the doubly charmed baryon, $\kappa_T=\frac{e}{2m_T}G_{M1}(0)=\frac{e}{2m_T}F_2(0)$ is the so-called anomalous magnetic moment, and the magnetic moment is defined as $\mu_T=\frac{m_N}{m_T}(\kappa_T+Q_T)$, where $m_N=940$ MeV is the nucleon mass. For the convenience of comparison with the magnetic moments of the doubly charmed baryons $\mu_T$ obtained in other approaches~\cite{Lichtenberg:1976fi,Bose:1980vy,Bernotas:2012nz,Oh:1991ws,Albertus:2006ya,Patel:2007gx,Dhir:2009ax,Sharma:2010vv,
Meng:2017dni,Ozdem:2019zis}, we take the nuclear magneton $\mu_N$ as the units of $\mu_T$ in this work.
\begin{figure}[h!]
\centering
\includegraphics[width=5cm]{Figa.png}\\
\includegraphics[width=5cm]{Figb.png}~~~\includegraphics[width=5.1cm]{Figc.png}\\
\includegraphics[width=5cm]{Figd.png}~~~\includegraphics[width=5.1cm]{Fige.png}
\caption{Feynman diagrams contributing to the spin-$\frac{3}{2}$ doubly charmed baryon magnetic moments up to NLO. Diagram (a) contributes at LO, diagrams (b)-(e) with a photon coupling to an intermediate meson or baryon contribute at NLO. The heavy~(light) solid, dashed, and wiggly lines denote spin-$\frac{3}{2}(\frac{1}{2})$ doubly charmed baryons, Goldstone bosons, and photons, respectively. The heavy dots represent the ${\cal O}(p^2)$ vertices.}\label{Figdiagram}
\end{figure}
In Fig.~\ref{Figdiagram}, diagrams~(a) and (b)-(e) contribute at ${\cal O}(p^2)$ and ${\cal O}(p^3)$, respectively. The leading order contribution to the magnetic moments $\mu_T$ is provided by the following Lagrangian:
\begin{eqnarray}
{\cal L}_{TT}^{(2)}=\frac{-ib_1^{tt}}{2m_T}\bar{T}^\mu\hat{F}_{\mu\nu}^+T^\nu+\frac{-ib_2^{tt}}{2m_T}\bar{T}^\mu T^\nu\mbox{\rm Tr}(F_{\mu\nu}^+),\label{eq:LagTree}
\end{eqnarray}
where the superscript in the Lagrangian stands for the chiral order, $b_1^{tt}$ and $b_2^{tt}$ are LECs, $\hat{F}_{\mu\nu}^+=F_{\mu\nu}^+-\frac{1}{3}{\rm Tr}(F_{\mu\nu}^+)$, $F_{\mu\nu}^+=|e|(u^\dag Q_HF_{\mu\nu}u+uQ_HF_{\mu\nu}u^\dag)$, $F_{\mu\nu}=\partial_\mu A_\nu-\partial_\nu A_\mu$, $Q_H={\rm diag}(2,1,1)$ is the charge operator of the doubly charmed baryon, $u={\rm exp}[i\Phi/2F_\phi]$ with the unimodular matrix containing the pseudoscalar
nonet $\Phi$, and $F_\phi$ is the pseudoscalar meson decay constant. In principle, one can use either the chiral limit value or the physical value for the decay constant. In the present work, similar to Ref.~\cite{Meng:2017dni}, we choose to use the physical values, which are different for $\pi$, $K$, and $\eta$, i.e., $F_\pi=92.4~{\rm MeV}$, $F_K=1.22F_\pi$, and $F_\eta=1.3F_\pi$. In the present work, we denote the spin-$\frac{3}{2}$ and -$\frac{1}{2}$ doubly charmed baryons by $T^\mu$ and $B$, respectively,
\begin{eqnarray}
T^\mu=\left(
\begin{array}{c}
\Xi_{cc}^{*++}\\
\Xi_{cc}^{*+}\\
\Omega_{cc}^{*+}
\end{array}
\right),\qquad B=\left(
\begin{array}{c}
\Xi_{cc}^{++}\\
\Xi_{cc}^{+}\\
\Omega_{cc}^{+}
\end{array}
\right).
\end{eqnarray}
The loop diagrams contributing at NLO are determined by the lowest order Lagrangians ${\cal L}_B^{(1)}$+${\cal L}_T^{(1)}$+${\cal L}_{TB}^{(1)}$+${\cal L}_{BB}^{(1)}$+${\cal L}_M^{(2)}$, which are,
\begin{eqnarray}
{\cal L}_B^{(1)}&=&\bar{B}(i\slashed{D}-m_B)B,\nonumber\\
{\cal L}_T^{(1)}&=&\bar{T}^\mu\left[-g_{\mu\nu}(i\slashed{D}-m_T)+i(\gamma_\mu D_\nu+\gamma_\nu D_\mu)-\gamma_\mu(i\slashed{D}+m_T)\gamma_\nu\right] T^\nu,\nonumber\\
{\cal L}_{TB}^{(1)}&=&\frac{iC}{2m_TF_\phi}\left(\partial^\alpha\bar{T}^\mu\right)\gamma_{\alpha\mu\nu}B\partial^\nu\Phi+\mbox{H.c.},\nonumber\\
{\cal L}_{TT}^{(1)}&=&\frac{iH}{2m_TF_\phi}\bar{T}^\mu\gamma_{\mu\nu\rho\sigma}\gamma_5\left(\partial^\rho T^\nu\right)\partial^\sigma\Phi,\nonumber\\
{\cal L}_M^{(2)}&=&\frac{F_\phi^2}{4}{\rm Tr}[\nabla_\mu U(\nabla^\mu U)^\dag],
\end{eqnarray}
with
\begin{eqnarray}
&&D_\mu B=\partial_\mu B+\Gamma_\mu,\nonumber\\
&&\Gamma_\mu=\frac{1}{2}(u^\dag\partial_\mu u+u\partial_\mu u^\dag)-\frac{i}{2}(u^\dag v_\mu u+uv_\mu u^\dag)=-ieQ_HA_\mu,\nonumber\\
&&u_\mu=i(u^\dag\partial_\mu u-u\partial_\mu u^\dag)+(u^\dag v_\mu u-uv_\nu u^\dag),\nonumber\\
&&U=u^2=e^{\frac{i\Phi}{F_\phi}},\qquad\nabla_\mu U=\partial_\mu U+ieA_\mu[Q_l,U],
\end{eqnarray}
where $\gamma^{\mu\nu\alpha\beta}=\frac{1}{2}\left[\gamma^{\mu\nu\alpha},\gamma^\beta\right]$, $\gamma^{\mu\nu\alpha}=\frac{1}{2}\left\{\gamma^{\mu\nu},\gamma^{\alpha}\right\}$, $\gamma^{\mu\nu}=\frac{1}{2}\left[\gamma^\mu,\gamma^\nu\right]$, $v_\mu$ stands for the vector source, and the charge matrix for the light $u,d,s$ quarks is $Q_l={\rm diag}(2/3,-1/3,-1/3)$. Note that for the Lagrangians of $\bar{T}B\Phi$ and $\bar{T}T\Phi$, we use the ``consistent'' coupling scheme introduced in Refs.~\cite{ Pascalutsa:2006up,Pascalutsa:1999zz,Pascalutsa:2000kd}. It has been applied to study the magnetic moments of octet baryons~\cite{Geng:2009hh}, decuplet baryons~\cite{Geng:2009ys}, and singly charmed baryons~\cite{Xiabng:2018qsd}, which can provide a proper description of the experimental/lattice QCD data and converges relatively faster.
The leading order tree-level contributions to the magnetic moments of the doubly charmed baryons $T^\mu$ can be obtained from Eq.~(\ref{eq:LagTree}) as follows:
\begin{eqnarray}
\kappa_T^{(2)}=\alpha b_1^{tt}+\beta b_2^{tt},\label{TreeMM}
\end{eqnarray}
with the values of $\alpha$ and $\beta$ listed in Table~\ref{eq:TreeCCs}.
\begin{table}[h!]
\caption{\label{eq:TreeCCs}Coefficients of the tree-level contributions in Eq.~(\ref{TreeMM}).}
\begin{center}
\begin{tabular}{cccc}
\hline
\hline
~~~~~~ & ~~~~~~$\Xi_{cc}^{*++}$~~~~~~ & ~~~~~~$\Xi_{cc}^{*+}$~~~~~~ & ~~~~~~$\Omega_{cc}^{*+}$~~~~~~\\
\hline
$\alpha$ & $\frac{4}{3}$ & $-\frac{2}{3}$ & $-\frac{2}{3}$\\
\hline
$\beta$ & $8$ & $8$ & $8$\\
\hline
\hline
\end{tabular}
\end{center}
\end{table}
The loop diagrams of Figs.~\ref{Figdiagram} (b), (c), (d), and (e) contribute to the magnetic moments at ${\cal O}(p^3)$, which are written as,
\begin{eqnarray}
\kappa_T^{(3)}&=&\sum_{\phi=\pi,K}\frac{H^2}{F_\phi^2}\xi_{T\phi}^{(3,b)}H_T^{(b)}(m_\phi)
+\sum_{\phi=\pi,K}\frac{C^2}{F_\phi^2}\xi_{T\phi,\delta}^{(3,c)}H_T^{(c)}(\delta,m_\phi)\nonumber\\
&&+\sum_{\phi=\pi,K,\eta}\frac{H^2}{F_\phi^2}\xi_{T\phi}^{(3,d)}H_T^{(d)}(m_\phi)
+\sum_{\phi=\pi,K,\eta}\frac{C^2}{F_\phi^2}\xi_{T\phi,\delta}^{(3,e)}H_T^{(e)}(\delta,m_\phi),\label{LoopMM}
\end{eqnarray}
with the coefficients $\xi_{T\phi,\delta_i}^{(3;b,c,d,e)}$ tabulated in Table~\ref{tab:LoopCCs}. Here, $\delta=m_T-m_B$ is the mass difference between the spin-$\frac{3}{2}$ and spin-$\frac{1}{2}$ doubly charmed baryons. The loop functions $H_T^{(b,d)}(m_\phi)$ and $H_T^{(c,e)}(\delta,m_\phi)$ correspond to $\frac{16\pi^2}{m_T^2}\left(H_2^{(g,h)}-H_{\rm PC}^{(g,h)}\right)$ and $\frac{16\pi^2}{m_T^2}\left(H_2^{(d,e)}-H_{\rm PC}^{(d,e)}\right)$, which can be found in the Appendix of Ref.~\cite{Geng:2009ys}. The power counting breaking~(PCB) terms $H_{\rm PC}^{(g,h)}$ and $H_{\rm PC}^{(d,e)}$ are determined by expanding $m_\phi$ up to ${\cal O}(p^0)$. Note that the pertinent loop functions are regularized with the $\widetilde{MS}$ scheme.
\begin{table}[h!]
\caption{\label{tab:LoopCCs}Coefficients of the loop contributions in Eq.~(\ref{LoopMM}) for the $T^\mu$ states.}
\begin{center}
\begin{tabular}{cccc}
\hline
\hline
~~~~~~ & ~~~~~~$\Xi_{cc}^{*++}$~~~~~~ & ~~~~~~$\Xi_{cc}^{*+}$~~~~~~ & ~~~~~~$\Omega_{cc}^{*+}$~~~~~~\\
\hline
$\xi_{T\pi}^{(3,b)}$ & $\frac{1}{2}$ & $-\frac{1}{2}$ & $0$\\
\hline
$\xi_{TK}^{(3,b)}$ & $\frac{1}{2}$ & $0$ & $-\frac{1}{2}$\\
\hline
$\xi_{T\pi,\delta}^{(3,c)}$ & $\frac{1}{2}$ & $-\frac{1}{2}$ & $0$\\
\hline
$\xi_{TK,\delta}^{(3,c)}$ & $\frac{1}{2}$ & $0$ & $-\frac{1}{2}$\\
\hline
$\xi_{T\pi}^{(3,d)}$ & $1$ & $\frac{5}{4}$ & $0$\\
\hline
$\xi_{TK}^{(3,d)}$ & $\frac{1}{2}$ & $\frac{1}{2}$ & $\frac{3}{2}$\\
\hline
$\xi_{T\eta}^{(3,d)}$ & $\frac{1}{6}$ & $\frac{1}{12}$ & $\frac{1}{3}$\\
\hline
$\xi_{T\pi,\delta}^{(3,e)}$ & $1$ & $\frac{5}{4}$ & $0$\\
\hline
$\xi_{TK,\delta}^{(3,e)}$ & $\frac{1}{2}$ & $\frac{1}{2}$ & $\frac{3}{2}$\\
\hline
$\xi_{T\eta,\delta}^{(3,e)}$ & $\frac{1}{6}$ & $\frac{1}{12}$ & $\frac{1}{3}$\\
\hline
\hline
\end{tabular}
\end{center}
\end{table}
In addition, we can obtain the HB counterparts of the loop functions by performing $1/m_T$ expansions for the loop functions obtained in the EOMS scheme, which turn out to agree with those of Ref.~\cite{Meng:2017dni}.
\section{Results and discussions}
\label{sec:Results}
In the following numerical analysis, we take the masses of the spin-$\frac{1}{2}$ doubly charmed baryons to be $m_B=m_{\Xi_{cc}^{+}}=3.62~{\rm GeV}$~\cite{Zyla:2020zbs}. Here, we do not consider the mass difference among the doubly charmed baryon triplet, in other words, we neglect SU(3) symmetry breakings in the masses, which are of higher chiral order. On the other hand, although the spin-$\frac{3}{2}$ doubly charmed baryons have not been observed, many theoretical studies~\cite{Bagan:1992za,Roncaglia:1995az,SilvestreBrac:1996bg,Ebert:1996ec,Tong:1999qs,Gershtein:2000nx,Kiselev:2001fw,Kiselev:2002iy,
Narodetskii:2001bq,Ebert:2002ig,Mathur:2002ce,Albertus:2006ya,Martynenko:2007je,Tang:2011fv,Roberts:2007ni,Valcarce:2008dr,Karliner:2014gca,
Shah:2016vmd,Chen:2016spr,Hu:2005gf,Lu:2017meb,Xiao:2017udy,Patel:2007gx,Alexandrou:2012xk,Namekawa:2013vu,Kiselev:2017eic,Migura:2006ep,
Chiu:2005zc,Vijande:2004at,Flynn:2003vz,Lewis:2001iz,Itoh:2000um,Bahtiyar:2018vub,Bahtiyar:2020uuj} have shown that the mass splitting $\delta$ between the spin-$\frac{3}{2}$ and spin-$\frac{1}{2}$ doubly charmed baryons is in the range of a few dozens MeV to 100 MeV. In this work, we choose $\delta=100~{\rm MeV}$. Nevertheless, we have performed calculations with different values for $\delta$ and found that the magnetic moments are not very sensitive to $\delta$, consistent with the study of Ref.~\cite{Meng:2017dni}. The mass difference $\delta$ vanishes in the heavy quark mass limit. For the masses of the pseudoscalar mesons, we use the PDG values~\cite{Zyla:2020zbs}. In addition, we fix the renormalization scale at $\mu=1$~GeV. We have checked that the magnetic moments in both the EOMS BChPT and HB ChPT are almost independent of $\mu$.
We do not have experimental or adequate lattice QCD data to determine the LECs $b_1^{tt}$ and $b_2^{tt}$. As a result, we turn to the quark model, where the magnetic moments can be calculated from the sum of the magnetic moments of the three constituent quarks, i.e.,
\begin{eqnarray}
\mu_q=\frac{e_q}{2m_q}=Q_q\frac{m_N}{m_q},\qquad (q=u,d,s,c)\label{quarkmodelMM}
\end{eqnarray}
where $\mu_q$ is in units of the nuclear magneton $\mu_N$, $Q_q$ stands for the quark charge, and $m_q$ is the constituent quark mass. In this work, we take the constituent quark masses from Ref.~\cite{Lichtenberg:1976fi}, which are $m_u=m_d=336$~MeV, $m_s=540$~MeV, and $m_c=1660$~MeV~\footnote{In Ref.~\cite{Meng:2017dni}, the authors estimated the LO magnetic moments of the spin-$\frac{3}{2}$ doubly charmed baryons using two sets of constituent quark masses~\cite{Lichtenberg:1976fi,Karliner:2014gca} and found that the so-obtained results are consistent with each other.}. Using Eq.~(\ref{quarkmodelMM}), one can easily obtain the LO contributions to the magnetic moments of the spin-$\frac{3}{2}$ doubly charmed baryons as follows:
\begin{eqnarray}
&&\mu_{\Xi_{cc}^{*++}}^{(2)}=\frac{m_N}{m_{\Xi_{cc}^{*++}}}(2+\kappa_{\Xi_{cc}^{*++}}^{(2)})=2\mu_c+\mu_u=2.61,\nonumber\\
&&\mu_{\Xi_{cc}^{*+}}^{(2)}=\frac{m_N}{m_{\Xi_{cc}^{*+}}}(1+\kappa_{\Xi_{cc}^{*+}}^{(2)})=2\mu_c+\mu_d=-0.18,\nonumber\\
&&\mu_{\Omega_{cc}^{*+}}^{(2)}=\frac{m_N}{m_{\Omega_{cc}^{*+}}}(1+\kappa_{\Omega_{cc}^{*+}}^{(2)})=2\mu_c+\mu_s=0.17.\label{LOMMnum}
\end{eqnarray}
Note that, the quark model predictions break the SU(3) flavor symmetry, because of the use of different quark masses. While in the leading order chiral Lagrangian of Eq.~(\ref{eq:LagTree}) which respects SU(3) symmetry, only two LECs are available. As a result, we could not fix the two leading order LECs by the quark model. In this work, we simply use the quark model predictions as the leading order results for the EOMS BChPT.
The results in Eq.~(\ref{LOMMnum}) show that the LO magnetic moments of $\Xi_{cc}^{*+}$ and $\Omega_{cc}^{*+}$ are relatively small. The reason for this is that the contributions of one light and two heavy quarks cancel each other due to the opposite quark charge. The three quark charges in $\Xi_{cc}^{*++}$ are all positive, and therefore the LO magnetic moment of $\Xi_{cc}^{*++}$ is the largest. This point has also been noted in Ref.~\cite{Meng:2017dni}.
For the contributions of the loop diagrams, we determine the relevant LECs $C$ and $H$ in two ways. In the first case, the two LECs are related to the axial-vector coupling of the nucleon $g_A$ with the help of the quark model~\cite{Li:2017cfz,Li:2017pxa}, i.e., $C=-\frac{2\sqrt{3}}{5}g_A$, $H=-\frac{3}{5}g_A$, and $g_A=1.267$. In the second scenario, we define a common factor $\rho$ to re-scale the LECs $C$ and $H$ simultaneously. In this way, we have $C=-\frac{2\sqrt{3}}{5}g_A\cdot\rho$ and $H=-\frac{3}{5}g_A\cdot\rho$. The parameter $\rho$ can be determined by the lattice QCD result for the magnetic moment of $\Omega_{cc}^{*+}$~\cite{Can:2015exa} with the leading order contributions determined by the quark model.
\subsection{The magnetic moments of spin-$\frac{3}{2}$ charmed baryons in case 1}
\begin{figure}[h!]
\centering
\includegraphics[width=10cm]{MMpi2.png}\\
\caption{Magnetic moments of the spin-$\frac{3}{2}$ charmed baryons as a function of $m_\pi^2$ in case 1. The solid and dashed lines in red represent the results of the EOMS BChPT and HB ChPT, respectively.}\label{FigMMpi2}
\end{figure}
In this case, the tree and loop level contribution to the magnetic moments of the spin-$\frac{3}{2}$ doubly charmed baryons are estimated by the quark model, as explained above. In Fig.~\ref{FigMMpi2}, we plot the magnetic moments of the spin-$\frac{3}{2}$ doubly charmed baryons $\mu_T$ as a function of $m_\pi^2$. It is shown that there is some distinct difference between the EOMS BChPT and HB ChPT results. In general, the light quark mass dependence is milder in the EOMS scheme than in the HB scheme. It should be noted that in Ref.~\cite{Liu:2018euh} the HB and EOMS results are almost identical. This is because for the spin-$\frac{1}{2}$ doubly charmed baryons, the loop contributions are much suppressed because of the small coupling $g_a=0.08$ determined by the lattice QCD data, $g_a=0.2$ by the heavy antiquark diquark symmetry, and $g_a=0.25$ by the quark model~\cite{Liu:2018euh}. For the spin-$\frac{3}{2}$ doubly charmed baryons, the two LECs $H=-0.76$ and $C=-0.88$ are more than three times larger, resulting in a loop contribution of more than 10 times larger. Therefore, the relatively large LECs $H$ and $C$ are responsible for the large difference between the EOMS BChPT and HB ChPT results. The difference for the $\Omega_{cc}^{*+}$ baryon is the most notable.
\begin{table}[h!]
\caption{\label{tab:loopDecomposition}Decomposition of the loop contributions to the magnetic moments of spin-$\frac{3}{2}$ doubly charmed baryons in case 1. The subscript $T$ and $B$ denote the loop diagrams with intermediate $T^\mu$ and $B$ states at ${\cal O}(p^3)$, respectively.}
\begin{center}
\begin{tabular}{ccccccccccc}
\hline
\hline
\multirow{2}{0.5cm} & & \multicolumn{4}{c}{Case 1: EOMS} & & \multicolumn{4}{c}{Case 1: HB}\\
\cline{3-6}\cline{8-11}
& & ~~~~${\cal O}(p^2)$~~~~ & ~~~~${\cal O}(p^3)_T$~~~~ & ~~~~${\cal O}(p^3)_B$~~~~ & ~~~~$\mu_{\rm tot}$~~~~ & & ~~~~${\cal O}(p^2)$~~~~ & ~~~~${\cal O}(p^3)_T$~~~~ & ~~~~${\cal O}(p^3)_B$~~~~ & ~~~~$\mu_{\rm tot}$~~~~\\
\hline
\multicolumn{2}{c}{$\mu_{\Xi_{cc}^{*++}}$} & $2.61$ & $0.27$ & $0.62$ & $3.50$ & & $2.61$ & $0.39$ &$0.51$ & $3.51$\\
\hline
\multicolumn{2}{c}{$\mu_{\Xi_{cc}^{*+}}$} & $-0.18$ & $-0.08$ & $-0.13$ & $-0.39$ & & $-0.18$ & $-0.11$ & $0.02$ & $-0.27$\\
\hline
\multicolumn{2}{c}{$\mu_{\Omega_{cc}^{*+}}$} & $0.17$ & $-0.15$ & $-0.39$ & $-0.37$ & & $0.17$ & $-0.27$ & $-0.54$ & $-0.64$\\
\hline
\hline
\end{tabular}
\end{center}
\end{table}
In Table~\ref{tab:loopDecomposition}, we decompose the loop contributions to the magnetic moments $\mu_T$ into those from the intermediate $T^\mu$ and $B$ states, respectively. For the spin-$\frac{3}{2}$ doubly charmed baryons, it should be noted that the absolute contributions of the intermediate $B$ baryons are more than those of intermediate $T^\mu$ baryons and their contributions are of the same sign in the EOMS BChPT and HB ChPT except for the $\Xi_{cc}^{*+}$ baryon in the HB ChPT. It indicates that loop corrections are important and non-negligible. Interestingly, for $\Omega_{cc}^{*+}$ the loop correction is much larger than the tree-level contribution, which can be attributed to the larger $H$ and $C$ predicted by the quark model, as also noted in Ref.~\cite{Meng:2017dni}. It should be mentioned that in the study of the spin-$\frac{1}{2}$ doubly charmed baryons~\cite{Liu:2018euh} the authors found that the coupling $g_a$ determined by the lattice QCD data~\cite{Can:2013tna} is much smaller than the one predicted by the quark model. Similar thing could happen here, as shown below. In a recent work on spin-$\frac{1}{2}$ doubly charmed baryons in the HB ChPT~\cite{Li:2020uok}, it is shown that the loop contributions of intermediate $T^\mu$ and $B$ cancel each other, which is opposite to the case of spin-$\frac{3}{2}$ baryons. This points to the importance of loop corrections again for spin-$\frac{3}{2}$ states.
\subsection{The magnetic moment of spin-$\frac{3}{2}$ charmed baryons in case 2}
\begin{figure}[h!]
\centering
\includegraphics[width=10cm]{MMpi2LQCD.png}\\
\caption{Magnetic moments of the spin-$\frac{3}{2}$ charmed baryons as a function of $m_\pi^2$ in case 2. The data point in blue stands for the lattice QCD unphysical value at $m_\pi\approx156~{\rm MeV}$. The solid and dashed lines in red represent the results of the EOMS BChPT and HB ChPT, respectively.}\label{FigMMpi2LQCD}
\end{figure}
In a lattice QCD simulation~\cite{Can:2015exa}, Can et al. studied the magnetic moment of $\Omega_{cc}^{*+}$ for a pion mass of $\approx156~{\rm MeV}$ and they obtained $\mu_{\Omega_{cc}^{*+}}=0.000(10)$, which is rather different from the predictions of the HB ChPT and EOMS BChPT shown in the previous section. However, keeping in mind that the BChPT predictions rely on the quark model inputs, both at tree level and one-loop level. As there are four unknow LECs in the BChPT, we could not determine all of them using the lattice QCD magnetic moment for $\Omega_{cc}^{*+}$. Therefore, in the present work, we tentatively assume that the quark model predicted magnetic moments are resonable, which are taken as the leading order ChPT results, and use the only lattice QCD datum to fix the strength of the loop functions, i.e., $C$ and $H$. As the one lattice QCD magnetic moment can not fix two LECs, we further assume that they are re-scaled by a same factor, namely, we multiply both LECs $C$ and $H$ by a common factor $\rho$, leading to effectively only one unknown parameter $\rho$ which can be determined by reproducing the single lattice QCD magnetic moment for $\Omega_{cc}^{*+}$. This way, we obtain $\rho_{\rm EOMS}=0.563(17)$ and $\rho_{\rm HB}=0.455(13)$ in the EOMS BChPT and HB ChPT, respectively. As a result, the LECs $C$ and $H$ are almost reduced by half compared to the predictions of the quark model. Using them, we replot the magnetic moments of the spin-$\frac{3}{2}$ doubly charmed baryons $\mu_T$ as a function of $m_\pi^2$. As shown in Fig.~\ref{FigMMpi2LQCD}, the HB and EOMS results are not much different, which is similar to the case of spin-$\frac{1}{2}$ doubly charmed baryons~\cite{Liu:2018euh}. This indicates that the quark model predictions for $C$ and $H$ might be overestimated, which should be verified by future experimental or more lattice QCD data.
\begin{table}[h!]
\caption{\label{tab:loopDecompositionLQCD}Decomposition of the loop contributions to the magnetic moments of spin-$\frac{3}{2}$ doubly charmed baryons in case 2. The subscript $T$ and $B$ denote the loop diagrams with intermediate $T^\mu$ and $B$ states at ${\cal O}(p^3)$, respectively.}
\begin{center}
\begin{tabular}{ccccccccccc}
\hline
\hline
\multirow{2}{0.5cm} & & \multicolumn{4}{c}{Case 2: EOMS} & & \multicolumn{4}{c}{Case 2: HB}\\
\cline{3-6}\cline{8-11}
& & ~~~~${\cal O}(p^2)$~~~~ & ~~~~${\cal O}(p^3)_T$~~~~ & ~~~~${\cal O}(p^3)_B$~~~~ & ~~~~$\mu_{\rm tot}$~~~~ & & ~~~~${\cal O}(p^2)$~~~~ & ~~~~${\cal O}(p^3)_T$~~~~ & ~~~~${\cal O}(p^3)_B$~~~~ & ~~~~$\mu_{\rm tot}$~~~~\\
\hline
\multicolumn{2}{c}{$\mu_{\Xi_{cc}^{*++}}$} & $2.61$ & $0.08$ & $0.20$ & $2.89$ & & $2.61$ & $0.08$ &$0.11$ & $2.80$\\
\hline
\multicolumn{2}{c}{$\mu_{\Xi_{cc}^{*+}}$} & $-0.18$ & $-0.03$ & $-0.04$ & $-0.25$ & & $-0.18$ & $-0.02$ & $0.00$ & $-0.20$\\
\hline
\multicolumn{2}{c}{$\mu_{\Omega_{cc}^{*+}}$} & $0.17$ & $-0.04$ & $-0.12$ & $0.001$ & & $0.17$ & $-0.05$ & $-0.11$ & $0.001$\\
\hline
\hline
\end{tabular}
\end{center}
\end{table}
Next, we also decompose the loop contributions to the magnetic moments $\mu_T$. As can be seen in Table~\ref{tab:loopDecompositionLQCD}, the predictions of the magnetic moments for all the spin-$\frac{3}{2}$ states are smaller in absolute value than those obtained by the quark model in case 1 due to the reduced loop contributions. Especially, the puzzling feature that the loop correction for $\Omega_{cc}^{*+}$ is much larger than the tree-level contribution in case 1 has disappeared. Nevertheless, the contributions of the intermediate $B$ baryons and $T^\mu$ baryons still add coherently in the EOMS BChPT and HB ChPT, indicating the importance of the loop contributions.
\begin{figure}[h!]
\centering
\includegraphics[width=10cm]{FigMMdiff.png}\\
\caption{Magnetic moments of the spin-$\frac{3}{2}$ charmed baryons obtained in different approaches. The solid stars denote
the results predicted in the EOMS BChPT in two different ways. The others are taken from the quark model~\cite{Lichtenberg:1976fi} (D. B. Lichtenber, 77), the bag model~\cite{Bose:1980vy,Bernotas:2012nz} (S. K. Bose et al., 80 and A. Bernotas et al., 12), the Skymion model~\cite{Oh:1991ws} (Y. s. Oh et al., 91), non-relativistic quark model~\cite{Albertus:2006ya} (C. Albertus et al., 07), hyper central model~\cite{Patel:2007gx} (B. Patel et al., 08), effective mass and screened charge scheme~\cite{Dhir:2009ax} (R. Dhir et al., 09), the chiral constituent quark model~\cite{Sharma:2010vv} (N. Sharma et al., 10), the HB ChPT~\cite{Meng:2017dni} (HB ChPT, 17), and the light-cone QCD sum rule (LCSR)~\cite{Ozdem:2019zis} (LCSR, 20).}\label{FigMMdiff}
\end{figure}
In Fig.~\ref{FigMMdiff} we compare our predicted magnetic moments with those obtained by other approaches. The predicted magnetic moments for the three spin-$\frac{3}{2}$ states using the EOMS BChPT in case 2 are: $\mu_{\Xi_{cc}^{*++}}=2.891(16)$, $\mu_{\Xi_{cc}^{*+}}=-0.248(4)$, and $\mu_{\Omega_{cc}^{*+}}=0.001(10)$. The predictions in case 1 can be found in Table.~\ref{tab:loopDecomposition}. The uncertainties in case 2 originate from the factor $\rho_{\rm EOMS}$ determined by the lattice QCD data. One finds that the magnetic moments predicted in case 2 are closer to the results of other approaches. The large differences of magnetic moments predicted in case 1 and case 2 may indicate the inconsistency between the quark model and the lattice QCD simulations. Hopefully, future lattice QCD or experimental studies can clarify the interesting situation.
\subsection{Heavy quark symmetry and its breaking}
Here, we discuss the heavy quark symmetry and its breaking and their impact on the magnetic moments of the $T^\mu$ baryons at ${\cal O}(p^3)$. In the heavy quark mass limit, i.e., $\delta=0$, the loop contributions of intermediate $B$ baryons are exactly twice as much as those of intermediate $T^\mu$ baryons. This can be checked by setting $\delta=0$ in the HB ChPT loop functions. Compared to the HB ChPT, loop functions in the EOMS BChPT include relativistic corrections which break the heavy quark symmetry. For the diagrams with a photon attached to an intermediate meson, the loop contributions for $\delta=0$ can be expressed as:
\begin{eqnarray}
&&H^2\cdot H_T^{(b)}(m_\phi)=\frac{9}{25}g_A^2m_T\cdot\left(\frac{m_\pi}{12\pi}m_\phi-\frac{36\log\left(\frac{m_T^2}{m_\phi^2}\right)-49}{144\pi^2 m_T}m_\phi^2+{\cal O}\left(\frac{1}{m_T^2}\right)\right),\\
&&C^2\cdot H_T^{(c)}(0,m_\phi)=\frac{12}{25}g_A^2m_T\cdot\left(\frac{m_\pi}{8\pi}m_\phi-\frac{27\log\left(\frac{m_T^2}{m_\phi^2}\right)-28}{96\pi^2 m_T}m_\phi^2+{\cal O}\left(\frac{1}{m_T^2}\right)\right).
\end{eqnarray}
It can be clearly seen that the contributions of intermediate $B$ baryons are twice as much as those of intermediate $T^\mu$ baryons at the order of $\left(\frac{1}{m_T}\right)^0$. However, at higher orders $\left(\frac{1}{m_T}\right)^n~(n\geq1)$ the relation is broken. For the
diagrams with a photon attached to an intermediate baryon, one can obtain the contribution for $\delta=0$ as follows:
\begin{eqnarray}
&&H^2\cdot H_T^{(d)}(m_\phi)=\frac{9}{25}g_A^2m_T\cdot\left(-\frac{21\log\left(\frac{m_T^2}{m_\phi^2}\right)-3\log\left(\frac{m_\phi^2}{\mu^2}\right)-167}{864\pi^2 m_T}m_\phi^2+{\cal O}\left(\frac{1}{m_T^2}\right)\right),\nonumber\\
&&C^2\cdot H_T^{(e)}(0,m_\phi)=\frac{12}{25}g_A^2m_T\cdot\left(\frac{5}{48\pi^2m_T}m_\phi^2+{\cal O}\left(\frac{1}{m_T^2}\right)\right).\label{BpHQS}
\end{eqnarray}
From Eq.~(\ref{BpHQS}), we find that the contributions start at $\frac{1}{m_T}$. These higher order relativistic corrections break the heavy quark symmetry relation as well. Note that the contribution of the diagram with a photon attached to an intermediate baryon vanishes at ${\cal O}(p^3)$ in the HB ChPT and the $m_T$ in the above equations outside the brackets is from the baryon field normalization.
\section{Summary}
We studied the magnetic moments of the spin-$\frac{3}{2}$ doubly charmed baryons in the covariant baryon chiral perturbation theory (BChPT) up to the next-to-leading order. To recover the power counting, we adopted the extended-on-mass-shell renormalization (EOMS) scheme. Due to lack of experimental and lattice QCD data, the leading order contributions to the magnetic moments are fixed by the quark model. The other two low energy constants $C$ and $H$ which appear in the loop contributions are determined by two ways: the quark model (case 1) and the lattice QCD data supplemented by the quark model (case 2).
To facilitate future lattice QCD simulations, we predicted the magnetic moments of the spin-$\frac{3}{2}$ doubly charmed baryons as a function of $m_\pi^2$. In case 1, the larger couplings $C$ and $H$~(compared to the case of the spin-$\frac{1}{2}$ doubly charmed baryons) resulted in large differences between the EOMS BChPT and HB ChPT results. The predicted magnetic moment for $\Omega_{cc}^{*+}$ is found to differ from that predicted by the lattice QCD simulations for $m_\pi\approx156~{\rm MeV}$. The discrepancy can be reconciled by suppressing the loop contributions. We show that if the quark model predicted $H$ and $C$ is reduced by half, one can obtain a magnetic moment for $\Omega_{cc}^{*+}$ consistent with that of the lattice QCD. In this case, the quark mass dependencies of the magnetic moments are much smaller than those of case 1. It should be noted that our predicted magnetic moments in case 2 for the three spin-$\frac{3}{2}$ states are closer to those of the majority of other approaches.
An interesting discovery is that for the magnetic moments of the spin-$\frac{3}{2}$ doubly charmed baryons, the contributions at ${\cal O}(p^3)$ of intermediate $T^\mu$ and $B$ baryons add coherently, instead destructively, as for the spin-$\frac{1}{2}$ doubly charmed baryons. Meanwhile, in the heavy quark mass limit, the intermediate $B$ contribution is twice that of the $T^\mu$ contribution. In reality, this relation is broken by the finite heavy quark masses and relativistic corrections. The coherent interference of the loop contributions of intermediate $T^\mu$ and $B$ baryons indicates that the loop contributions are relatively more important, compared to the case of spin-$\frac{1}{2}$ doubly charmed baryons.
\section{Acknowledgements}
We thank Dr. Kadir Utku Can
for reminding us of Ref.~\cite{Can:2015exa}. R. X. S. thanks Lu Meng for useful discussions. This work was partly supported by the National Natural Science Foundation of China (NSFC) under Grants No. 11975041, No.11735003, No.11961141004, and the Academic Excellence Foundation of BUAA for Ph.D. Students.
|
\section{Introduction}\label{intro}
\subsection{Model and motivations.}\label{subs:model_motiv}
The goal of this article is to study the evolutionary dynamics of a phenotypically structured population in an environment which varies with a linear trend but in an oscillatory manner. We study the following non-local parabolic equation
\begin{equation}
\label{GeneralCase_ntilde}
\left\{
\begin{array}{l}
\partial_t\widetilde{n}-\sigma\partial_{xx}\widetilde{n}=\widetilde{n}[a({e(t)},x-\widetilde{c}t)-\widetilde{\rho}(t)],\quad (t,x)\in[0,+\infty)\times\mathbb{R},\\
\widetilde{\rho}(t)=\displaystyle\int_{\mathbb{R}}\widetilde{n}(t,x)dx,\\
\widetilde{n}(t=0,x)=\widetilde{n}_0(x).
\end{array}
\right.
\end{equation}
This equation models the dynamics of a population which is structured by a phenotypic trait $x\in\mathbb{R}$. Here, $\widetilde{n}$ corresponds to the density of individuals with trait $x$. We denote by $a({e(t)},x-\widetilde{c}t)$ the intrinsic growth rate of an individual with trait $x$ at time $t$. The term $-\widetilde{c}t$ has been introduced to consider {a shifting of the fitness landscape} with a linear trend. {The function $e(t):\mathbb{R}_+\to E$, represents the environmental state at time $t$ (for instance, the temperature) and is assumed to be periodic, with $E$ corresponding to the set of the states of the environment, for instance an interval corresponding to the possible temperatures. The variation of the environmental state $e$ may have an impact on the optimal trait (the trait that maximizes $a(e,x)$) or other parameters of selection as for instance the pressure of selection (corresponding to the curvature of $a(e,x)$ around its maximum point, see Section \ref{Bio_example} for some examples).} The term $\widetilde{\rho}$ which corresponds to the total size of the population represents a competition term. Here, we assume indeed a uniform competition between all the individuals. The diffusion term models the mutations, with $\sigma$ the mutation rate.
{A natural motivation to study such type of problem is the fact that many natural populations are subject both to a directional change of the phenotypic optimum and fluctuations of the environment (\cite{chevin-et-al}). Such fluctuations may be periodic, due for instance to seasonal effects, or stochastic due to the random change of the environment.
Here we consider a deterministic growth rate that varies with a linear trend but in an oscillatory manner. Our study provides also insights for the case of random environments since it is based on some homogenization techniques that could also be used to study the random case. However, the study of the random fluctuations would still need considerable work and is out of the scope of this article. }
Will the population be able to adapt to the environmental change? Is there a {maximal} speed above which the population will get extinct? How is such maximal speed modified due to the fluctuations?
\subsection{Related works.}\label{subs:related_works}
{The impact of changing environments on the evolution of quantitative traits has been studied using closely related quantitative genetics models in the biological literature (see for instance \cite{ML.WG.AW:91,ML.RL:93,RB.LM:95,RL.SS:96,MK.SM:14}). In these works one usually assumes that the growth rate $a$ has a particular form of quadratic type and that the environmental change has only an impact on the optimal trait. However, the environmental variations may also modify other parameters of selection, as for instance the pressure of selection \cite{JG.BT.HD:20}. Finally, the works studying a periodic environment, consider only a particular sinusoidal form of periodic variation \cite{ML.WG.AW:91,RL.SS:96}.}
\noindent
Models closely related to \eqref{GeneralCase_ntilde}, but with a local reaction term and no fluctuation, have been widely studied (see for instance \cite{Berestycki-et-al, Berestycki_Rossi1, Berestycki_Rossi2,Berestycki_Fang}). Such models are introduced to study dynamics of populations structured by a space variable neglecting evolution. It is shown in particular that there exists a critical speed of environment change $c^*$, such that the population survives if and only if the environment change occurs with a speed less than $c^*$. We also refer to \cite{Bouhours2016} where an integro-difference model has been studied for the spatial dynamics of a population in the case of a randomly changing environment. Moreover, in \cite{Raoul_Berest_Alfaro}, both spatial and evolutionary dynamics of a population in an environment with linearly moving optimum has been studied. While in the present work, we don't include any spatial structure, we take into account oscillatory change of environment in addition to a change with linear trend.
\\
\noindent
The evolutionary dynamics of structured populations under periodic fluctuations of the environment has been recently studied by \cite{Souganidis, Lorenzi-Desvillettes, Mirrahimi&Figueroa_pprint, Almeida}. The works in \cite{Lorenzi-Desvillettes, Almeida} are focused on the study of a particular form of growth rate $a$ and in particular some semi-explicit solutions to such equations are provided. In \cite{Souganidis,Mirrahimi&Figueroa_pprint} some asymptotic analysis of such equations for general growth rates are provided. The present article is closely related to \cite{Mirrahimi&Figueroa_pprint} where a periodically evolving environment was considered without the linear trend. The presence of such linear trend of environment change leads to new difficulties in the asymptotic analysis. Moreover, we go further than the results in \cite{Mirrahimi&Figueroa_pprint} and provide an asymptotic expansion for the average size of the population in terms of the mutation rate. Such expansion is closely related to an asymptotic expansion of the Floquet eigenvalue for the linear problem. Furthermore in a very recent work \cite{Carrere_Nadin_pprint} the authors study a closely related model, but without the linear change of the environment, and study the impact of the different parameters of the model on the final population size.\\
In this article, we use an asymptotic approach based on Hamilton-Jacobi equations with constraint. This approach has been developed during the last decade to study the asymptotic solutions of selection-mutation equations, assuming small effect of the mutations. Such equations have the property that their solution concentrate as Dirac masses on the fittest traits. There is a large literature on this approach. We refer to \cite{DJMP,Perthame_Barles,MAA} for the establishment of the basis of this approach for homogeneous environments.
\subsection{Mathematical assumptions}\label{subs:math_assumptions}
To introduce our assumptions, we first define
$$
\overline{a}(y)=\displaystyle\frac{1}{T}\int_0^Ta({e(t)},y)dt.
$$
{We then assume that $e:\mathbb{R}_+\to E$, a periodic function, and $a\in L^\infty(E,C^3(\mathbb{R}))$ are such that:
\begin{equation}\tag{H1}
\label{a_W3inf}
e(t)=e(t+T),\ \forall\ t\in\mathbb{R}_+,\quad\text{and}\ \exists\ d_0>0: \Vert a(e,\cdot)\Vert_{L^{\infty}(\mathbb{R})}\leq d_0\quad\forall\ e\in E,
\end{equation}}
and that the averaged function $\overline{a}$ attains its maximum and
\begin{equation}\tag{H2a}
\label{max_a}
\max_{x\in\mathbb{R}} \ \overline{a}(x)>0,
\end{equation}
which means that there exist at least some traits with strictly positive average growth rate.\\
Moreover, for some of our results (Theorem \ref{Unique_Identif} and Theorem \ref{Asymp_Expansion}) we assume that this maximum is attained at a single point $x_m$; that is
\begin{equation}\tag{H2b}
\label{x_m}
\exists! \ x_m: \max_{x\in\mathbb{R}} \overline{a}(x)=\overline{a}(x_m),
\end{equation}
and also
\begin{equation}\tag{H3}
\label{x_var}
\exists! \ \overline{x}\leq x_m; \ \overline{a}(\overline{x})+\frac{\widetilde{c}^2}{4\sigma}=\overline{a}(x_m).
\end{equation}
{Let us explain the role of the trait $\overline x$ in our results. We will show in Theorem \ref{Unique_Identif} that, as the mutation rate $\sigma$ vanishes and when the speed of the environmental change $\widetilde c$ is not too high, the phenotypic density $\widetilde n$ concentrates around a single trait which moves linearly with the same speed $\widetilde c$. The population density concentrates indeed around the trait $\overline x+\widetilde c t$ and follows in this way the optimum of the average environment, that is $x_m+\widetilde c t$, with a constant lag.\\}
Finally, we make the following assumption on the initial data:
\begin{equation}\tag{H4}
\label{n0_exp}
0\leq {\widetilde{n}_0(x)}\leq e^{C_1-C_2|x|}, \quad \forall x\in\mathbb{R},
\end{equation}
which indicates that the initial density of individuals with large traits is exponentially small.\\
\subsection{Preliminary results.}\label{subs:pre_rsults}
To avoid the shift in the growth rate $a$, we transform our problem with a change of variables. We introduce indeed $n(t,x)=\widetilde{n}(t,x+\widetilde{c}t)$ which satisfies:
\begin{equation}
\label{Case_n}
\left\{
\begin{array}{l}
\partial_tn-\widetilde{c}\partial_xn-\sigma\partial_{xx}n=n[a({e(t)},x)-\rho(t)],\quad (t,x)\in[0,+\infty)\times\mathbb{R},\\
\rho(t)=\displaystyle\int_{\mathbb{R}}n(t,x)dx,\\
n(t=0,x)=\widetilde{n}_0(x).
\end{array}
\right.
\end{equation}
\noindent
Next, we introduce the linearized problem associated to \eqref{Case_n}. Let $m(t,x)=n(t,x)e^{\int_0^t\rho(s)ds}$, for $n$ the solution of \eqref{Case_n}, then $m$ satisfies
\begin{equation}
\label{modelLap_in_m}
\left\{\begin{array}{cr}
\partial_t m-\widetilde{c}\partial_x m-\sigma\partial_{xx}m=a({e(t)},x)m,& (t,x)\in[0,+\infty)\times\mathbb{R},\\
m(t=0,x)=\widetilde{n}_0(x), & x\in\mathbb{R}.\\
\end{array}
\right.
\end{equation}
We also introduce the corresponding parabolic eigenvalue problem as follows
\begin{equation}
\label{Model_EigenvalueR}
\left\{\begin{array}{cr}
\partial_tp_c-\widetilde{c}\partial_xp_c-\sigma\partial_{xx}p_c-a({e(t)},x)p_c=\lambda_{\widetilde{c},\sigma} p_c,& (t,x)\in[0,+\infty)\times\mathbb{R},\\
0<p_c; \ p_c(t,x)=p_c(t+T,x), & (t,x)\in[0,+\infty)\times\mathbb{R}.\\
\end{array}
\right.
\end{equation}
For better legibility, we omit the tilde in the index of $p_c$, while we still refer to the problem with constant $\widetilde{c}$.
We also define the eigenvalue problem in the bounded domain $[-R,R]$, for some $R>0$,
\begin{equation}
\label{Model_EigenvalueBR}
\left\{\begin{array}{cr}
\partial_tp_R-\widetilde{c}\partial_xp_R-\sigma\partial_{xx}p_R-a({e(t)},x)p_R=\lambda_R p_R,& (t,x)\in[0,+\infty)\times [-R,R],\\
p_R=0, & (t,x)\in[0,+\infty)\times\{-R,R\},\\
0<p_R;\ p_R(t,x)=p_R(t+T,x), & (t,x)\in[0,+\infty)\times[-R,R].
\end{array}
\right.
\end{equation}
It is known that problem \eqref{Model_EigenvalueBR} has a unique eigenpair $(\lambda_R,p_R)$ with $p_R$ a strictly positive eigenfunction such that $\Vert p_R(0,\cdot)\Vert_{L^\infty([-R,R])}=1$, (see \cite{hess}). Another fundamental result (see for instance \cite{huska08}), for our purpose is that the function $R\mapsto\lambda_R$ is decreasing and $\lambda_R\rightarrow\lambda_{\widetilde{c},\sigma}$ as $R\rightarrow+\infty$.\\
To announce our first result we introduce another assumption. We assume that $a$ takes small values at infinity in the following sense: there exist positive constants $\delta$ and $R_0$ such that
\begin{equation}\tag {Hc}
\label{a_lambda_neg}
a({e},x)+\lambda_{\widetilde{c},\sigma}\leq-\delta, \quad \forall {e\in E} \text{ and }|x|\geq R_0.
\end{equation}
\begin{proposition}
\label{Prop1}
Assume \eqref{a_W3inf}, \eqref{n0_exp} and \eqref{a_lambda_neg}. Then for problem \eqref{Model_EigenvalueR} there exists a unique generalized principal eigenfunction $p_c$ associated to $\lambda_{\widetilde{c},\sigma}$, with $\Vert p_c(0,\cdot)\Vert_{L^\infty(\mathbb{R})}=1$. Moreover, we have $p_c=\displaystyle\lim_{R\rightarrow\infty}p_R$ and
\begin{equation}
\label{bound_p}
p_c(t,x)\leq \Vert p_c\Vert_{L^\infty} e^{-\nu(|x|-R_0)},\qquad\forall(t,x)\in[0,+\infty)\times\mathbb{R},
\end{equation}
for $\nu= -\frac{\widetilde{c}}{2\sg}+\sqrt{\frac{\delta}{\sg}+\frac{1}{2}\left(\frac{\widetilde{c}}{\sg}\right)^2}$.\\
Finally, the eigenfunction $p_c(t,x)$ is exponentially stable, in the following sense; there exists $\alpha>0$ such that:
\begin{equation}
\label{Exp_Conv_m}
\Vert m(t,x)e^{t\lambda_{\widetilde{c},\sigma}}-\alpha p_c(t,x)\Vert_{L^{\infty}(\mathbb{R})}\rightarrow0\quad\text{exponentially fast as}\ t\rightarrow\infty.
\end{equation}
\end{proposition}
The proof of this proposition is based on the results in \cite{Mirrahimi&Figueroa_pprint}.\\
We next define the $T-$periodic functions $Q_c(t)$ and $P_c(t,x)$ as follows:
\begin{equation}
\label{PyQ}
Q_c(t)=\dfrac{\int_{\mathbb{R}}a({e(t)},x)p_c(t,x)dx}{\int_{\mathbb{R}}p_c(t,x)dx},\quad P_c(t,x)=\dfrac{p_c(t,x)}{\int_{\mathbb{R}}p_c(t,x)dx},
\end{equation}
and we recall a result proved in \cite{Mirrahimi&Figueroa_pprint}.
\begin{proposition}
\label{prop2}
There exists a unique periodic solution $\widehat{\rho}(t)$ to the problem
\begin{equation}
\label{Rho_equation}
\left\{\begin{array}{l}
\dfrac{d\widehat{\rho}}{dt}=\widehat{\rho}\left[Q_c(t)-\widehat{\rho}\right],\quad t\in(0,T),\\
\widehat{\rho}(0)=\widehat{\rho}(T),
\end{array}\right.
\end{equation}
if and only if $\displaystyle\int_0^TQ_c(t)dt>0$. Moreover this solution can be explicitly expressed as follows:
\begin{equation}
\label{Exp_rho_var}
\widehat{\rho}(t)=\frac{1-\exp\left[-\displaystyle\int_0^{T}Q_c(s)ds\right]}{\exp\left[-\displaystyle\int_0^{T}Q_c(s)ds\right]\displaystyle\int_t^{t+T}\exp\left[\displaystyle\int_t^s Q_c(\theta)d\theta\right]ds}.
\end{equation}
\end{proposition}
\subsection{The main results and the plan of the paper.}\label{subs:main_results}
{We are interested in} determining conditions on the environment shift speed $\widetilde{c}$ which leads to extinction or survival of the population. In the case of the population survival we then try to characterize asymptotically the population density considering small effect of the mutations. \\
\noindent
To present our result on the survival criterion, we define the "critical speed".
\begin{definition}
We define the critical speed $\widetilde{c}^*_\sigma$ as follows
\begin{equation}
\label{Critical_c}
\widetilde{c}^*_\sigma=\left\{
\begin{array}{lr}
2\sqrt{-\sg\lambda_{0,\sigma}},&\text{if}\ \lambda_{0,\sigma}<0,\\
0,& \text{otherwise},
\end{array}
\right.
\end{equation}
where $\lambda_{0,\sigma}$ corresponds to the principal eigenvalue introduced by Proposition \ref{Prop1}, in the case $c=0$.
\end{definition}
The next result shows that $\widetilde{c}^*_\sigma$ is indeed a critical speed of {environmental} change above which the population goes extinct. \\
\begin{proposition}{{(Long time behavior)}}\\
\label{ConvLambda}
Let $n(t,x)$ be the solution of \eqref{Case_n}. Assume \eqref{a_W3inf}, \eqref{max_a}, \eqref{n0_exp} and \eqref{a_lambda_neg}. Then the following statements hold:
\begin{itemize}
\item[(i)] if $\widetilde{c}\geq \widetilde{c}^*_\sigma$, then the population will go extinct, i.e.
$\rho(t)\rightarrow0,$ as $t\rightarrow\infty$,
\item[(ii)] if $\widetilde{c}<\widetilde{c}^*_\sigma$, then $\vert\rho(t)-\widehat{\rho}(t)\vert\rightarrow0$, as $t\rightarrow\infty$, with $\widehat{\rho}$ the unique solution of \eqref{Rho_equation}.
\item[(iii)] Moreover, $\quad\left\Vert\dfrac{n(t,x)}{\rho(t)}-P_c(t,x)\right\Vert_{L^\infty}\longrightarrow0$, as $t\rightarrow\infty$. Consequently we have, as $t\rightarrow\infty$:
\begin{equation}
\label{convPerio}
\Vert n(t,\cdot)-\widehat{\rho}(t)P_c(t,\cdot)\Vert_{L^\infty}\rightarrow 0,\:\mathrm{if}\:\widetilde{c}<\widetilde{c}^*_\sigma\quad\mathrm{and}\quad\Vert n\Vert_{L^\infty}\rightarrow0,\:\mathrm{if}\:\widetilde{c}\geq \widetilde{c}^*_\sigma.
\end{equation}
\end{itemize}
\end{proposition}
\begin{remark}
Note that if $\lambda_{0,\sigma} \geq 0$, then $\widetilde{c}^*_{\sigma} = 0$, which means that the population goes extinct even without {environmental} linear change, that is $\widetilde{c} = 0$.
\end{remark}
Proposition \ref{ConvLambda} allows to relate extinction/survival of the population to the {environmental} change speed and shows that if the change goes "too fast" the population will not be able to follow the environment change and will get extinct. However, if the change speed is "moderate" the phenotypic density $n$ converges to the periodic function $n_c(t,x)=\widehat{\rho}(t)P_c(t,x)$, which is in fact the unique periodic solution of \eqref{Case_n}.\\
\noindent
Next, we are interested in describing this periodic solution $n_c$, asymptotically as the effect of mutations is small. To this end, with
a change of notation, we take $\sigma = \e^2$ and $\widetilde{c}=\e c$, and we study asymptotically the solution $(n_{\e c},\widehat{\rho}_{\e c})$ as $\e$ vanishes. {For better legibility, we also define $c^*_{\e}:=\dfrac{\widetilde{c}^*_{\e^2}}{\e}$ where $\widetilde{c}^*_{\e^2}$ stands for the critical speed $\widetilde{c}^*_\sigma$ with $\sigma=\e^2$.} Note that, in view of Proposition \ref{ConvLambda}, to provide an asymptotic analysis considering $\sigma=\e^2$ small, a rescaling of the {environmental} shift speed as $\widetilde{c}=\e c$ is necessary {(see also Theorem \ref{Asymp_Expansion}). The population can tolerate only {an environmental} shift with small speed if the mutations have small effect.}\\
In order to keep the notation simpler we denote $(n_{\e c},\widehat{\rho}_{c\e })=(n_{\e},\rho_{\e})$, which is the unique periodic solution of the problem:
\begin{equation}
\label{n_eps}
\left\{
\begin{array}{l}
\partial_tn_\e-\e c\partial_xn_\e-\e^2\partial_{xx}n_\e=n_\e[a({e(t)},x)-\rho_\e(t)],\quad (t,x)\in[0,+\infty)\times\mathbb{R},\\
\rho_\e(t)=\displaystyle\int_\mathbb{R} n_\e(t,x)dx,\\
n_\e(0,x)=n_\e(T,x).
\end{array}
\right.
\end{equation}
To study asymptotically this problem we perform a Hopf-Cole transformation (or WKB ansatz), i.e we consider
\begin{equation}
\label{HopfCole}
n_\varepsilon=\frac{1}{\sqrt{2\pi\varepsilon}}\exp{\left(\frac{\psi_\varepsilon}{\varepsilon}\right)}.
\end{equation}
This change of variable comes from the fact that with such rescaling the solution $n_\e$ will naturally have this form. While we expect that $n_\e$ tends to a Dirac mass, as $\e\to0$, $\psi_\e$ will have a non singular limit. From this transformation \eqref{HopfCole} we deduce that $\psi_\e$ solves:
\begin{equation}
\label{Eq_psi_eps}
\frac{1}{\e}\p_t\psi_\e-\e\p_{xx}\psi_\e=\left\vert\p_x\psi_\e+\frac{c}{2}\right\vert^2+a(\textcolor{black}{e(t)},x)-\frac{c^2}{4}-\rho_\e(t),\quad(t,x)\in[0,+\infty)\times\mathbb{R}.
\end{equation}
Here is our first main result:
\begin{theorem}{{(Asymptotic behavior)}}\\
\label{Conv_HJ}
Assume \eqref{a_W3inf}, \eqref{max_a} and \eqref{a_lambda_neg} and also that {$c<\displaystyle\liminf_{\e\to0} c^*_{\e}$}. Then the following statements hold:
\begin{itemize}
\item[(i)] As $\e\rightarrow0$, we have $\Vert\rho_\e(t)-\widetilde{\varrho}(t)\Vert_{L^\infty}\rightarrow0$,
with $\widetilde{\varrho}(t)$ {a $T-$periodic function. }
\item[(ii)] Moreover, as $\e\rightarrow0$, $\psi_\e(t,x)$ converges locally uniformly to a function $\psi(x)\in C(\mathbb{R})$, a viscosity solution to the following equation:
\begin{equation}
\label{LimitEq}
\left\{\begin{array}{rcl}
\displaystyle-\left|\p_x \psi+\frac{c}{2}\right|^2&=&\overline{a}(x)-\overline{\rho}-\frac{c^2}{4},\quad x\in\mathbb{R},\\
\displaystyle\max_{x\in\mathbb{R}}\psi(x)&=&0,\\
-A_1\vert x\vert^2-\frac{c}{2}x-A_2&\leq \psi\leq& c_1-c_2\vert x\vert,
\end{array}
\right.
\end{equation}
with
$$
\bar{\rho}=\int_0^T\widetilde{\varrho}(t)dt,
$$
for some positive constants $A_1,A_2, c_1$ and $c_2=-\frac{c}{2}+\sqrt{\delta+\frac{c^2}{2}}$.\\
\end{itemize}
\end{theorem}
The above theorem is closely related to Theorem 4 in \cite{Mirrahimi&Figueroa_pprint}. A new difficulty comes from the drift term. To deal with the drift term we use a Liouville transformation (see for instance \cite{Berestycki_Rossi1, Berestycki_Rossi2}) that allows us to transform the problem to a parabolic problem without drift.\\
\noindent
{To present our next result,} let us consider the eigenproblem \eqref{Model_EigenvalueR} for $\sigma=\e^2$ and $\widetilde{c}=c\e$, that is:
\begin{equation}
\label{Eigenvalue_lambda_ceps}
\left\{\begin{array}{cr}
\partial_tp_{c\e}-\e c\partial_xp_{c\e}-\e^2\partial_{xx}p_{c\e}-a({e(t)},x)p_{c\e}=p_{c\e}\lambda_{c,\e} ,& (t,x)\in[0,+\infty)\times\mathbb{R},\\
0<p_{c\e}; \ p_{c\e}(t,x)=p_{c\e}(t+T,x), & (t,x)\in[0,+\infty)\times\mathbb{R}
\end{array}
\right.
\end{equation}
Here we denote $\lambda_{c,\e}$ the eigenvalue $\lambda_{\widetilde{c},\sigma}$ with $\sigma=\e^2$ and $\widetilde{c}=c\e$ for better legibility.
\begin{theorem}{{(Uniqueness and identification of the solution)}}\\
\label{Unique_Identif}
Let $\lambda_{c,\e}$ be the principal eigenvalue of problem \eqref{Eigenvalue_lambda_ceps} and assume \eqref{a_W3inf}, \eqref{x_m}, \eqref{x_var} and \eqref{a_lambda_neg}. Assume in addition that {$c<\displaystyle\liminf_{\e\to0} c^*_{\e}$}, then the following statements hold:
\begin{itemize}
\item[(i)]{Let $\overline{\rho}_{\e}=\frac{1}{T}\int_0^T\rho_{\e}(t)dt$, then $\overline{\rho}_{\e}=-\lambda_{c,\e}$.}
\item[(ii)] The viscosity solution of \eqref{LimitEq} is unique and it is indeed a classical solution given by
\begin{equation}
\label{Exp_Sol_psi}
\psi(x)=\frac{c}{2}(\overline{x}-x)+ \int_{\overline{x}}^{x_m}\sqrt{\overline{a}(x_m)-\overline{a}(y)}dy-\left\vert \int_{x_m}^{x}\sqrt{\overline{a}(x_m)-\overline{a}(y)}dy\right\vert.
\end{equation}
where $\overline{x}<x_m$ is given in \eqref{x_var}. {Moreover, as $\e\to 0$, $\overline \rho_\e$ converges to $\overline \rho=\overline a(\overline x)$.}
\item[(iii)] Furthermore, let $n_\varepsilon$ solve \eqref{n_eps}, then
\begin{equation}
\label{Dirac_mass_conv}
n_\e(t,x)-\widetilde{\varrho}(t)\delta(x-\overline{x})\rightharpoonup 0,\quad\text{as } \e\to0,
\end{equation}
point wise in time, weakly in $x$ in the sense of measures, {with $\widetilde{\varrho}$ the unique periodic solution of the following equation
\begin{equation}
\label{eq_varrho}
\left\{\begin{array}{l}
\dfrac{d\widetilde{\varrho}}{dt}=\widetilde{\varrho}\left[a({e(t)},\bar x)-\widetilde{\varrho}\right],\quad t\in(0,T),\\
\widetilde{\varrho}(0)=\widetilde{\varrho}(T).
\end{array}\right.
\end{equation}}
\end{itemize}
\end{theorem}
\begin{remark}
The statement $(iii)$ in Theorem \ref{Unique_Identif} implies, for the solution $\widetilde{n}_\e$ to the initial problem \eqref{GeneralCase_ntilde} with $\sigma=\e^2$ and $\widetilde c=c\e$, that
\begin{equation}
\label{Dirac_mass_n_tilde}
\widetilde{n}_\e(t,x)-\widetilde{\varrho}(t)\delta(x-\overline{x}-ct)\rightharpoonup 0, \quad \text{ as } \e\to 0,
\end{equation}
point wise in time, weakly in $x$ in the sense of measures. This implies that the phenotypic density of the population concentrates on a {dominant trait which follows the optimal trait with the same speed but with a constant lag $x_m-\overline x$.}
\end{remark}
\\
{Note that while in \cite{Mirrahimi&Figueroa_pprint} the uniqueness of the viscosity solution to the corresponding Hamilton-Jacobi equation with constraint was immediate, here to prove the uniqueness of the viscosity solution more work is required. In particular, in order to prove such result the constraint is not enough and we use also the bounds on $\psi$, given in \eqref{LimitEq}. More precisely we introduce a new function $u(x)=\psi(x)+\frac{c}{2}x$ which solves
\begin{equation}
\label{Comp_Pb_u}
\left\{
\begin{array}{c}
-\vert \p_x u\vert^2=\overline{a}(x)-\overline{\rho}-\frac{c^2}{4},\quad x\in\mathbb{R},\\\\
{\max_{x\in \mathbb{R}} u(x)-\f{c}{2}x=0,}\\\\
-A_1\vert x\vert^2-A_2\leq u(x)\leq c_1-c_2\vert x\vert+\frac{c}{2}x,
\end{array}
\right.
\end{equation}
where the constants $A_1$, $A_2$, $c_1$, $c_2$ are the same as in \eqref{LimitEq}.\\
The main idea comes from the fact that any viscosity solution to a Hamilton-Jacobi equation of type \eqref{Comp_Pb_u} but in a bounded domain $\Omega$ can be uniquely determined by its values on the boundary points of $\Omega$ and by its values at the maximum points of the RHS of the Hamilton-Jacobi equation \cite{Lions}. \\
\noindent
{Finally, in our last result we provide an asymptotic expansion} for the Floquet eigenvalue which leads to an asymptotic expansion for the critical speed $c^*_\e$ and the average size of the population $\bar \rho_\e$.
\begin{theorem}{(Asymptotic expansions)}\\
\label{Asymp_Expansion}
Let $\lambda_{c,\e}$ be the principal eigenvalue of problem \eqref{Eigenvalue_lambda_ceps} and assume \eqref{a_W3inf}, \eqref{x_m} and \eqref{a_lambda_neg}. Assume in addition that {$c<\displaystyle\liminf_{\e\to0} c^*_{\e}$}, then the following asymptotic expansions hold
\begin{equation}
\label{Eigenvalue_Approx}
\overline{\rho}_{\e}=-\lambda_{c,\e}=\overline{a}(x_m)-\frac{c^2}{4}-\e \sqrt{-\overline{a}_{xx}(x_m)/2}+o(\e),
\end{equation}
\begin{equation}
\label{Crit_Speed_Approx}
c^*_{\e}=2\sqrt{\overline{a}(x_m)}-\e {\sqrt{-\f{\overline{a}_{xx}(x_m)}{2\, \overline a(x_m)}}}+o(\e).
\end{equation}
\end{theorem}
Note that the expansion for the Floquet eigenvalue is indeed related to the harmonic approximation of the {ground state energy} of the Schr\"odinger operator \cite{Helffer}. However, here we have a parabolic, non self-adjoint operator}.\\
{In Section \ref{Bio_example} we study an illuminating biological example and show thanks to the above result that the fluctuations of the environment may help the population to follow the {environmental} shift.}\\
\noindent
The paper is organized as follows: in Section \ref{Long_time} we deal with the long time study of the problem and prove the preliminary results Proposition \ref{Prop1} and Proposition \ref{ConvLambda}.
Next in Section \ref{Asymp_Analysis} we provide an asymptotic analysis of the problem considering small effect of mutations and we prove Theorem \ref{Conv_HJ}. {In Section \ref{Uniqueness}, we obtain the uniqueness of the viscosity solution to \eqref{LimitEq} and prove Theorem \ref{Unique_Identif}.} Section \ref{Eigen_Approx} is devoted to {the approximations} of the principal eigenvalue (average size of the population) and the critical speed, given in Theorem \ref{Asymp_Expansion}. {In Section \ref{Bio_example} we study a biological example and discuss the effect of the fluctuations on the critical speed of survival and on the phenotypic distribution of the population. Finally}, in Appendix A and B, we provide some technical results and computations.
\section{The convergence in long time}
\label{Long_time}
In this section we provide the proofs of Proposition \ref{Prop1} and Proposition \ref{ConvLambda}. To this end, we make a change of variable which allows us to transform the problem into a parabolic equation without the drift term.\\
Let $m(t, x)$ satisfy the linearized problem \eqref{modelLap_in_m}, we denote $\mathcal{P}_0$ and $\mathcal{P}_c$ the linear operators associated to problem \eqref{modelLap_in_m}, for $\widetilde{c}=0$ and $\widetilde{c}>0$ respectively, that is:
\begin{equation}
\label{P0&Pc}
\mathcal{P}_0\omega:=\p_t\omega-\sg \p_{xx}\omega-a(\textcolor{black}{e(t)},x)\omega,\qquad
\mathcal{P}_c\omega:=\p_t\omega-\widetilde{c}\p_x\omega-\sg \p_{xx}\omega-a(\textcolor{black}{e(t)},x)\omega.
\end{equation}
In Subsection \ref{Liouville}, we introduce the Liouville transformation and provide a relation between $\mathcal{P}_0$ and $\mathcal{P}_c$ which allows us to obtain a relationship between $\widetilde{c}$ and $\lambda_{\widetilde{c},\sigma}$. Next in Subsection \eqref{Proof_Prop1} and \eqref{Proof_Prop4} we provide the proofs of Proposition \ref{Prop1} and Proposition \ref{ConvLambda} respectively.
\subsection{Liouville transformation.}
\label{Liouville}
Here, we reduce the parabolic equation \eqref{Case_n} to a parabolic problem without the drift term via a Liouville transformation (see for instance \cite{Berestycki_Rossi1, Berestycki_Rossi2} where this transformation is used for an elliptic problem).\\
Let $M(t,x)$ be given by
\begin{equation}
\label{Liouville_transf}
M(t,x):=m(t,x)e^{\frac{\widetilde{c}}{2\sigma}x} ,
\end{equation}
for $m(t,x)$ the solution of the linearized problem \eqref{modelLap_in_m}, then $M$ satisfies:
\begin{equation}
\label{Liouville_Eq}
\p_tM-\sg \p_{xx}M=\Big[a(\textcolor{black}{e(t)},x)-\frac{\widetilde{c}^2}{4\sg}\Big]M.
\end{equation}
We denote $\mathcal{\widetilde{P}}$ the linear operator associated to the above equation, i.e.
$$
\mathcal{\widetilde{P}}\omega:=\p_t\omega-\sg \p_{xx}\omega-a_c(\textcolor{black}{e(t)},x)\omega,
$$
where $a_c(\textcolor{black}{e(t)},x)=\left[a(\textcolor{black}{e(t)},x)-\frac{\widetilde{c}^2}{4\sg}\right]$.\\
We establish in the next lemma the relation between the principal eigenvalues associated to the operators $\mathcal{P}_0$, $\mathcal{P}_c$ and $\mathcal{\widetilde{P}}$.
\begin{lemma}
\label{lambda_c}
Let $\lambda(\mathcal{P},\mathbb{D})$ denote the principal eigenvalue of the operator $\mathcal{P}$ in the domain $\mathbb{D}$, it holds
$$
\lambda_{\widetilde{c},\sigma}=\lambda\left(\mathcal{P}_c,\mathbb{R}_+\times\mathbb{R}\right)=\lambda\left(\mathcal{\widetilde{P}},\mathbb{R}_+\times\mathbb{R}\right).
$$
Moreover, let $\lambda_{0,\sigma}=\lambda(\mathcal{P}_0,\mathbb{R}_+\times\mathbb{R})$, then
$\lambda_{\widetilde{c},\sigma}=\lambda_{0,\sigma}+\frac{\widetilde{c}^2}{4\sg}$.
\end{lemma}
\begin{proof}
The proof follows from the definition of the eigenfunction and eigenvalue and the fact that
$$
\mathcal{\widetilde{P}}\omega=\left(\mathcal{P}_c\left(\omega e^{-\frac{\widetilde{c}}{2\sg}x}\right)\right)e^{\frac{\widetilde{c}}{2\sg}x}.
$$
\end{proof}
\subsection{Proof of Proposition \ref{Prop1}.}
\label{Proof_Prop1}
Proposition \ref{Prop1} can be proved following similar arguments as in the proof of Lemma 6 in \cite{Mirrahimi&Figueroa_pprint}. Note that the argument in \cite{Mirrahimi&Figueroa_pprint} is based on an exponential separation result for linear parabolic equations in \cite{huska08} that holds for general linear operators of the form
$$
\omega_t=L(t,x)\omega,\quad\text{in}\ [0,+\infty)\times\mathbb{R},
$$
with $L(t,x)$ being any time-dependent second-order elliptic operator in non-divergence form, i.e:
$$
L(t,x)\omega=a_{ij}(t,x)\p_i\p_j \omega + B_i(t,x)\p_i \omega + A(t,x)\omega,
$$
where the functions $B_i, A\in L^\infty(\mathbb{R}_+\times\mathbb{R})$ and $a_{ij}$ satisfies
$$
a_{ij}(t,x)\xi_i\xi_j\geq \alpha_0|\xi|^2,\quad (t,x)\in \mathbb{R}_+\times\mathbb{R},
$$
(see Section 9 in \cite{huska08} for more details).\\
Here, we only provide the proof of the inequality \eqref{bound_p} which is also obtained by an {adaptation} of the proof of Lemma 6 in \cite{Mirrahimi&Figueroa_pprint}.
Let $\widetilde{a}_c(\textcolor{black}{e(t)},x)=a_c(\textcolor{black}{e(t)},x)+\lambda_{\widetilde{c},\sigma}$ then $p_c$ is a positive periodic solution of the following equation:
\begin{equation}
\label{a_tilda}
\partial_t p_c-\widetilde{c}\p_x p_c-\sigma\p_{xx} p_c=p_c\widetilde{a}_c(\textcolor{black}{e(t)},x),\quad\mathrm{in}\;\mathbb{R}\times \mathbb{R}.
\end{equation}
Note that we have defined $p_c$ in $(-\infty, 0]$ by periodic prolongation. We denote $\Vert p_c\Vert_{L^\infty(\mathbb{R}\times\mathbb{R})}=\Gamma$ and define:
$$
\zeta(t,x)=\Gamma e^{-\delta(t-t_0)}+\Gamma e^{-\nu(|x|-R_0)},
$$
for some $\nu$ to be found later and $\delta$, $R_0$ given in \eqref{a_lambda_neg}. One can verify that
$$
\Gamma\leq \zeta(t,x)\quad\mathrm{if}\;|x|=R_0\;\mathrm{or}\;t=t_0.
$$
Furthermore if $|x|>R_0$ or $t>t_0$ evaluating in \eqref{a_tilda} shows:
$$
\begin{array}{rcl}
\partial_t\zeta-\widetilde{c}\p_x\zeta-\sigma\p_{xx}\zeta-\zeta\widetilde{a}_c(\textcolor{black}{e(t)},x)&=&\Gamma e^{-\delta(t-t_0)}(-\delta-\widetilde{a}_c(\textcolor{black}{e(t)},x))\\
&& +\Gamma e^{-\nu(|x|-R_0)}\left(\widetilde{c}\nu\frac{x}{|x|}-\sigma\nu^2-\widetilde{a}_c(\textcolor{black}{e(t)},x)\right)\geq 0,
\end{array}
$$
since $\widetilde{a}_c(\textcolor{black}{e(t)},x)+\frac{\widetilde{c}^2}{4\sigma}=a(\textcolor{black}{e(t)},x)+\lambda_{\widetilde{c},\sigma}\leq -\delta$ thanks to assumption \eqref{a_lambda_neg} and choosing $\nu$ conveniently such that the inequality holds. Indeed, since $-1\leq\frac{x}{|x|}\leq 1$, we have:
$$
\widetilde{c}\nu\frac{x}{|x|}-\sigma\nu^2-\widetilde{a}_c(\textcolor{black}{e(t)},x)\geq -\widetilde{c}\nu-\sg\nu^2+\delta+\frac{\widetilde{c}^2}{4\sg}\geq 0
$$
for
$$
\frac{-\widetilde{c}-\sqrt{4\delta\sg+2\widetilde{c}^2}}{2\sg}\leq\nu\leq \frac{-\widetilde{c}+\sqrt{4\delta\sg+2\widetilde{c}^2}}{2\sg}.
$$
Thus $\zeta$ is a supersolution of \eqref{a_tilda} on:
$$
\Lambda_0=\{(t,x)\in (t_0,\infty)\times \mathbb{R}\;;|x|>R_0\},
$$
which dominates $p_c$ on the parabolic boundary of $\Lambda_0$. Applying the maximum principle to $\zeta-p_c$, we obtain
$$
p_c(t,x)\leq \Gamma e^{-\delta(t-t_0)}+\Gamma e^{-\nu(|x|-R_0)},\qquad |x|\geq R_0,\;t\in(t_0,\infty).
$$
Taking the limit $t_0\rightarrow-\infty$ yields
$$
p_c(t,x)\leq \Gamma e^{-\nu(|x|-R_0)},\qquad |x|\geq R_0,\;t< +\infty,
$$
in particular, for $\nu=\dfrac{-\widetilde{c}+\sqrt{4\delta\sg+2\widetilde{c}^2}}{2\sg}$. We conclude that $p_c$ satisfies \eqref{bound_p}.
\vbox{\hrule height0.6pt\hbox{%
\vrule height1.3ex width0.6pt\hskip0.8ex
\vrule width0.6pt}\hrule height0.6pt
}
\subsection{Proof of Proposition \ref{ConvLambda}.}
\label{Proof_Prop4}
The proof of Proposition \ref{ConvLambda}, is closely related to the proof of Proposition 2 in \cite{Mirrahimi&Figueroa_pprint} but we need to verify two properties before applying the arguments in \cite{Mirrahimi&Figueroa_pprint}. To this end we prove the following lemmas. The rest of the proof follows from the arguments in \cite{Mirrahimi&Figueroa_pprint}.
\begin{lemma}
\label{Relation_lb_c}
Let $\lambda_{\widetilde{c},\sigma}$ be the principal eigenvalue of problem \eqref{Model_EigenvalueR}. Then, $
\lambda_{\widetilde{c},\sigma}<0$ if and only if $\widetilde{c}<\widetilde{c}^*_\sigma.$
\end{lemma}
\begin{proof}
Follows directly from the definition of $\widetilde{c}^*_\sigma$.
\end{proof}\\
\begin{lemma}
\label{bound_n}
Assume \eqref{a_W3inf} and \eqref{n0_exp} and let $C_3=C_2(\sigma C_2+\widetilde{c})+d_0$ then the solution $n(t,x)$ to equation \eqref{Case_n} satisfies:
$$
n(t,x)\leq \exp\left(C_1-C_2|x|+C_3t\right),\quad\forall(t,x)\in(0,+\infty)\times\mathbb{R}.
$$
\end{lemma}
\begin{proof}
We argue by a comparison principle argument. Define the function
$$\bar n(t,x)=\exp\left(C_1-C_2|x|+C_3t\right).$$
We prove that $n\leq \bar n$. One can verify that for $C_3$ defined as in the formulation of the Lemma, we have the following inequality a.e:
$$
\begin{array}{c}
\partial_t \bar n -\widetilde{c}\p_x \bar n-\sigma\p_{xx} \bar n-\left[a(\textcolor{black}{e(t)},x)-\rho(t)\right] \bar n = \\
e^{\left(C_1-C_2|x|+C_3t\right)}\left[C_3-\sigma
C_2^2+C_2\frac{cx}{|x|}-a(\textcolor{black}{e(t)},x)+\rho(t)\right]
\geq 0.
\end{array}
$$
Moreover, we have for $t=0$, $n(0,x)\leq \bar{n}(0,x)$ thanks to assumption \eqref{n0_exp}. We can then apply a maximum principle to $d(t,x)=\bar n(t,x)-n(t,x)$, in the class of $L^2$ functions, and we conclude that:
$$
0 \leq d(t,x)\Rightarrow n(t,x)\leq \bar n(t,x),\quad\forall(t,x)\in(0,+\infty)\times\mathbb{R}.
$$
\end{proof}
\section{Regularity estimates}
\label{Asymp_Analysis}
In this section, we prove Theorem \ref{Conv_HJ}. To this end we first provide some uniform bounds for $\rho_\e(t)$. Then, in Subsection \ref{Regularity}, we prove that $\psi$ is locally uniformly bounded, Lipschitz continuous with respect to $x$
and locally equicontinuous in time. Finally in the last subsection we conclude the proof of Theorem \ref{Conv_HJ} by letting $\e$ go to zero and describing the limits of $\psi_\e$ and $\rho_\e$.
\subsection{Uniform bounds for $\rho_\e$.}\label{subs:unif_bounds_rhoe} We have the following result on $\rho_\e$.
\begin{proposition}
\label{bounds_rhoe}
Assume \eqref{a_W3inf}, \eqref{a_lambda_neg} and let $\widetilde{c}=\e c$ with {$c<\displaystyle\liminf_{\e\to0} c^*_{\e}$}. Then for all $0<\e\leq \e_0$, there exist positive constants $\rho_m$ and $\rho_M$ such that:
\begin{equation}
\label{Rho_bouded}
0 < \rho_m \leq \rho_\e(t) \leq \rho_M,\quad \forall t\geq 0.
\end{equation}
\end{proposition}
The proof of this result follows similar arguments as in \cite{Mirrahimi&Figueroa_pprint}. For the convenience of the reader, we provide this proof in Appendix A.
\subsection{Regularity results for $\psi_\e$.}
\label{Regularity}
In this subsection we prove some regularity estimates on $\psi_\e$ which give the basis to prove the convergence of $\psi_\varepsilon$ and $\rho_\varepsilon$ as $\varepsilon\rightarrow 0$ in Subsection \ref{limit_equation}.
We claim the following Theorem.\\
\begin{theorem}
\label{regularity_theo}
Assume \eqref{a_W3inf}, \eqref{max_a} and \eqref{a_lambda_neg}. Let $\psi_\e$ be a $T-$periodic solution to \eqref{Eq_psi_eps}. Then the following items hold:
\begin{itemize}
\item[(i)] The sequence $(\psi_\e)_\e$ is locally uniformly bounded; i.e.
\begin{equation}
\label{Bounds_psi_eps}
-A_1\vert x\vert^2-\frac{c}{2}x-A_2\leq \psi_\e\leq c_1-c_2\vert x\vert, \quad \forall (t,x)\in\mathbb{R}_+\times\mathbb{R},
\end{equation}
for some positive constants $A_1,A_2, c_1$ and $c_2=-\frac{c}{2}+\sqrt{\delta+\frac{c^2}{2}}$.
\item[(ii)] Moreover, the sequence $(\phi_\e=\sqrt{2c_1-\psi_\e})_\e$, is uniformly Lipschitz continuous with respect to $x$ in $(0,+\infty)\times\mathbb{R}$.
\item[(iii)] Also, $(\psi_\e)_\e$ is locally equicontinuous in time in $[0,T]\times \mathbb{R}$ and satisfies
\begin{equation}
\label{Equi_u}
|\psi_\e(t,x)-\psi_\e(s,x)|\to0\quad\mathrm{as}\:\e\to 0, \quad \forall\ 0\leq s\leq t\leq T.
\end{equation}
\end{itemize}
\end{theorem}
In the next subsections we provide the proof of the lower bound in \eqref{Bounds_psi_eps} and the uniform Lipschitz continuity of $\phi_\e$. The proof of the other properties can be obtained by an adaptation of the arguments in \cite{Mirrahimi&Figueroa_pprint}. For the convenience of the reader we provide them in Appendix A.
\subsubsection{Lower bound for $\psi_\e$.}\label{subs:lower_bound_psie}
To obtain the lower bound for $\psi_\e$ we use the bounds for $a$ in \eqref{a_W3inf} and for $\rho_\varepsilon$ in \eqref{Rho_bouded} and we obtain for $D_0=d_0+\rho_M$
$$
\partial_t n_\e-c\e\partial_x n_\e-\e^2\partial_{xx} n_\e\geq -D_0n_\e.
$$
Let $n_\e^*$ be the solution of the following Cauchy problem
$$
\left\{
\begin{array}{l}
\partial_t n_\e^*-c\e\partial_x n^*_\e-\e^2\partial_{xx} n_\e^*+D_0 n_\e^*=0,\\
n_\e^*(0,x)=n_\e^0,
\end{array}
\right.
$$
we define $N_\e^*$ analogously to \eqref{Liouville_transf} by the Liouville transformation of $n_\e^*$ as follows
$$
N_\e^*(t,x):=n_\e^*(t,x)e^{\frac{c}{2\e}x}.
$$
Then, $N_\e^*$ solves the heat equation
$$
\left\{
\begin{array}{l}
\partial_t N_\e^*-\e^2\partial_{xx} N_\e^*+D_1N_\e^*=0,\\
N_\e^*(0,x)=n_\varepsilon^0(x)e^{\frac{c}{2\e}x},
\end{array}
\right.
$$
for $D_1=D_0+\frac{c^2}{4}$. The solution to the latter equation is given explicitly by the Heat Kernel $K$,
$$
N_\e^*(t,x)=e^{-D_1t}\left(N_\e^*(0,y)\ast K\right)=\frac{e^{-D_1t}}{\e\sqrt{4\pi t}}\int_{\mathbb{R}}N_\e^*(0,y)e^{-\frac{|x-y|^2}{4t \e^2}}dy,\quad t>0.
$$
Note that $N_\e^*(0,x)$ from its definition can be written as follows
\begin{equation}
\label{N_e*_pce}
N_\e^*(0,x):=\frac{p_{c\e}(0,x)}{\int_\mathbb{R} p_{c\e}(0,x)dx}\rho_\e(0)e^{\frac{c}{2\e}x}.
\end{equation}
We recall that $p_{c\e}$ is uniquely determined once $\Vert p_{c\e}(0,x)\Vert_{L^\infty(\mathbb{R})}=1$ is fixed. Then, one can choose $x_\e$ such that $ p_{c\e}(0,x_\e)=1$. From an elliptic-type Harnack inequality in a bounded domain we can obtain
\begin{equation}
\label{BoundInf_p}
p_{c\e}(t_0,x_\e)\leq \sup_{y\in B(x_\e,\e)}p_{c\e}(t_0,y)\leq C p_{c\e}(t_0,x),\quad \forall(t_0,x)\in[\delta_0,2T]\times B(x_\e,\e),
\end{equation}
where $\delta_0$ is such that $0<\delta_0<T$ and $C$ is a positive constant depending on $\delta_0$ and $d_0$ (we refer to Appendix A-Proof of upper bound, for more details on this inequality). We then use the $T-$periodicity of $p_{c\e}$ to conclude that the last inequality is satisfied {for all} $t\in[0,T]$.\\
From \eqref{bound_p}, \eqref{N_e*_pce} and \eqref{BoundInf_p} we deduce that
$$
\e^{-1}D_2e^{-\frac{D_3}{\e}+\frac{c}{2\e}x}\leq \rho_m \dfrac{p_\e(0,x)e^{\frac{c}{2\e}x}}{\int_{\mathbb{R}}p_\e(0,x)dx}\leq N^*_\e(0,x),\quad\forall x\in B(x_\e,\e),
$$
for some positive constants $D_2$ and $D_3$ depending on $\Vert p_\e\Vert_{L^\infty}$, $\rho_m$, $\delta$, and the constants of hypothesis \eqref{a_lambda_neg}. Then, for all $(t,x)\in(0,+\infty)\times\mathbb{R}$
$$
\begin{array}{rcl}
N_\e^*(t,x)
&\geq &\displaystyle\frac{D_2}{\e^{2}\sqrt{4\pi t}}e^{-\frac{D_3+\e D_1 t}{\e}}\int_{B(x_\e,\e)}e^{\frac{c}{2\e}y}e^{-\frac{|x-y|^2}{4t \e^2}}dy\\
&\geq &
\displaystyle\frac{D_2|B(x_\e,\e)|}{\e^2\sqrt{4\pi t}}\exp\left\{-\frac{|x|^2+(|x_\e|+\e)^2}{2 t\e^2}+\frac{c}{2}\left(\frac{x_\e}{\e}-1\right)-\frac{D_3+D_1t\e}{\e}\right\}.
\end{array}
$$
By the definition of $n_\e^*$ and the comparison principle we obtain that $n_\e^*(t,x)\leq n_\e(t,x)$ and hence
$$
n_\e(t,x)\geq
\displaystyle\frac{D_2|B(x_\e,\e)|}{\e^2\sqrt{4\pi t}}\exp\left\{-\frac{|x|^2+(|x_\e|+\e)^2}{2 t\e^2}+\frac{c}{2}\left(\frac{x_\e-x}{\e}-1\right)-\frac{D_3+D_1t\e}{\e}\right\}.
$$
This, together with the definition of $\psi_\varepsilon$, implies that
$$
\e\log\left(\frac{D_2|B(x_\e,\e)|}{\e^{2}\sqrt{4\pi t}}\right)-\frac{|x|^2+(|x_\e|+\e)^2}{2t\e}+\frac{c}{2}(x_\e-x-\e)-(D_3+D_1t\e)\leq \psi_\e(t,x),\quad\forall t\geq 0.
$$
In particular, we obtain that $\forall t\in\left[1,1+\e T\right]$.
$$
\e\log\left(\frac{D_2|B(x_\e,\e)|}{\e^{3/2}\sqrt{4\pi t}}\right)-\frac{|x|^2+(|x_\e|+\e)^2}{2t}+\frac{c}{2}(x_\e-x-\e)-(D_3+D_1t)\leq \psi_\e\left(\frac{t}{\e},x\right)$$
Note that $x_\e$ is uniformly bounded in $\e$ thanks to \eqref{bound_p}. Then we can conclude by using the periodicity of $\psi_\e$. We obtain a quadratic lower bound for $\psi_\varepsilon$ for all $t\geq0$; that is, there exist $A_1$, $A_2\geq 0$ and $\e_0$ small enough such that for all $\e\leq\e_0$,
\begin{equation}
\label{Low_Bound_psi_e}
-A_1|x|^2-\frac{c}{2}x-A_2\leq \psi_\e(t,x),\quad\forall t\geq 0.
\end{equation}
\subsubsection{Lipschitz bounds.}\label{subsubs:lipschitz}
In this section we prove the Lipschitz bounds for $\phi_\e$. To this end we use a Bernstein type method closely related to the one used in \cite{MAA,Mirrahimi&Figueroa_pprint}. Let $\phi_\e = \sqrt{2c_1-\psi_\e}$, for $c_1$ given by \eqref{Bounds_psi_eps}, then $\phi_\e$ satisfies
$$
\frac{1}{\e}\partial_t\phi_\e -c\partial_x\phi_\e -\e\partial_{xx}\phi_\e -\left(\frac{\e}{\phi_\e }-2\phi_\e \right)|\partial_x\phi_\e |^2=\frac{a(\textcolor{black}{e(t)},x)-\rho_\e(t)}{-2\phi_\e }.
$$
Define $\Phi_\e=\p_x \phi_\e$, which is also $T-$periodic. We differentiate the above equation with respect to $x$ and multiply by $\frac{\Phi_\e }{|\Phi_\e |}$, i.e.,
$$
\begin{array}{rcl}
\frac{1}{\e}\partial_t|\Phi_\e |-c\partial_x|\Phi_\e |-\e\partial_{xx}|\Phi_\e |-2\left(\frac{\e}{\phi_\e }-2\phi_\e \right)\Phi_\e \cdot\partial_x|\Phi_\e |+\left(\frac{\e}{\phi_\e ^2}+2\right)|\Phi_\e |^3 &&\\\\
\leq
\dfrac{\left(a(\textcolor{black}{e(t)},x)-\rho_\e(t)\right)|\Phi_\e |}{2\phi_\e ^2}-\dfrac{\partial_x a\cdot \Phi_\e }{2\phi_\e |\Phi_\e |}. &&
\end{array}
$$
From \eqref{Bounds_psi_eps} we deduce that
$$
\sqrt{c_1}\leq \phi_\e \leq \sqrt{A_1|x|^2+\frac{c}{2}x+A_3},\qquad\forall\ t\geq 0,\ x\in\mathbb{R},
$$
for $A_3=A_2+2c_1$. It follows that
$$
\left|2\left(\frac{\e}{\phi_\e }-2\phi_\e \right)\right|\leq A_4|x|+A_5,
$$
for some positive constants $A_4$ and $A_5$. From here, we deduce for $\vartheta$ large enough
\begin{equation}
\label{wp_eps_theta}
\frac{1}{\e}\partial_t|\Phi_\e |-c\partial_x|\Phi_\e |-\e\partial_{xx}|\Phi_\e |-\big(A_4|x|+A_5\big)\big\vert \Phi_\e \cdot\partial_x|\Phi_\e |\big\vert+2\left(|\Phi_\e |-\vartheta\right)^3\leq0.
\end{equation}
Let $T_M>2T$ and $A_6$ to be chosen later, define now, for $(t,x)\in\Big(0,\frac{T_M}{\varepsilon}\Big]\times [-R,R]$
$$
\Theta_\e(t,x)=\frac{1}{2\sqrt{t\varepsilon}}+\frac{A_6R^2}{R^2-|x|^2}+\vartheta.
$$
We next verify that $\Theta_\e$ is a strict supersolution of \eqref{wp_eps_theta} in $\Big(0,\frac{T_M}{\varepsilon}\Big]\times [-R,R]$. To this end we compute
$$
\partial_t \Theta_\e=-\frac{1}{4t\sqrt{t\e}},\quad \partial_x \Theta_\e=\frac{2A_6R^2x}{(R^2-|x|^2)^2},\quad \partial_{xx} \Theta_\e=\frac{2A_6R^2}{(R^2-|x|^2)^2}+\frac{8A_6R^2|x|^2}{(R^2-|x|^2)^3},
$$
and then replace in \eqref{wp_eps_theta} to obtain
$$
\begin{array}{l}
\frac{1}{\e}\partial_t\Theta_\e-c\partial_x\Theta_\e-\e\partial_{xx} \Theta_\e-\big(A_4|x|+A_5\big)|\Theta_\e\cdot\partial_x \Theta_\e|+2\left(\Theta_\e-\vartheta\right)^3\\\\
=-\frac{1}{4\e t\sqrt{\e t}}-\frac{2cA_6R^2x}{(R^2-|x|^2)^2}-\e\Big[\frac{2A_6R^2}{(R^2-|x|^2)^2}+\frac{8A_6R^2|x|^2}{(R^2-|x|^2)^3}\Big]\\
-\big(A_4|x|+A_5\big)\left(\frac{1}{2\sqrt{\e t}}+\frac{A_6R^2}{R^2-|x|^2}+\vartheta\right)\frac{2A_6R^2|x|}{(R^2-|x|^2)^2}+2\left(\frac{1}{2\sqrt{\e t}}+\frac{A_6R^2}{R^2-|x|^2}\right)^3\\\\
\geq -\e \Big[\frac{2A_6R^2d}{(R^2-|x|^2)^2}+\frac{8A_6R^4}{(R^2-|x|^2)^3}\Big]
-\big(A_4R+A_5\big)\left(\frac{1}{2\sqrt{\e t}}+\frac{A_6R^2}{R^2-|x|^2}+\vartheta\right)\frac{2A_6R^3}{(R^2-|x|^2)^2}\\\\
\quad+\frac{3A_6R^2}{R^2-|x|^2}\Big(\frac{1}{2t\e}+\frac{A_6R^2}{\sqrt{\e t}(R^2-|x|^2)}\Big)+\frac{2A_6R^3}{(R^2-|x|^2)^2}\left(\frac{A_6^2R^3}{R^2-|x^2|}-c\right),
\end{array}
$$
where, for the inequality, we have used that $|x|\leq R$. \\
One can verify that the RHS of the above inequality is strictly positive for $R > 1$, $\varepsilon\leq 1$, and $A_6>>\sqrt{T_M}$. Therefore, $\Theta_\e$ is a strict supersolution of \eqref{wp_eps_theta} in $\Big(0,\frac{T_M}{\e}\Big] \times [-R,R]$ and for $\varepsilon\leq 1$.\\
We next prove that
$$
|\Phi_\e (t,x)|\leq \Theta_\e(t,x)\quad\mathrm{in}\;\Big(0,\frac{T_M}{\e}\Big]\times [-R,R].
$$
To this end, we notice that $\Theta_\e(t, x)$ goes to $+\infty$ as $|x|\rightarrow R$ or as $t\rightarrow0$. Therefore, $|\Phi_\e |(t, x) - \Theta_\e(t, x)$ attains its maximum at an interior point of $\Big(0,\frac{T_M}{\varepsilon}\Big]\times [-R,R]$. We choose $t_{\rm max} \leq \frac{T_M}{\e}$ the smallest time such that the maximum of $|\Phi_\e |(t, x)-\Theta_\e(t, x)$ in the set $(0,t_{\rm max}]\times [-R,R]$ is equal to 0. If such $t_{\rm max}$ does not exist, we are done.\\\\
Let $x_{\rm max}$ be such that $|\Phi_\e |(t, x)- \Theta_\e(t, x)\leq |\Phi_\e |(t_{\rm max}, x_{\rm max})- \Theta_\e(t_{\rm max}, x_{\rm max})= 0$ for all $(t, x) \in(0, t_{\rm max})\times [-R,R]$. At such point, we have
$$
0\leq\partial_t\big(|\Phi_\e |- \Theta_\e\big)(t_{\rm max}, x_{\rm max}),\quad 0\leq -\p_{xx}\big( |\Phi_\e |- \Theta_\e\big)(t_{\rm max}, x_{\rm max}),$$
$$ |\Phi_\e |(t_{\rm max}, x_{\rm max})\p_x |\Phi_\e |(t_{\rm max}, x_{\rm max})= \Theta_\e(t_{\rm max}, x_{\rm max})\p_x \Theta_\e(t_{\rm max}, x_{\rm max}).
$$
Combining the above properties with the facts that $|\Phi_\e |$ and $\Theta_\e$ are respectively sub- and strict super-solution of \eqref{wp_eps_theta}, we obtain that
$$
(|\Phi_\e |(t_{\rm max}, x_{\rm max})-\vartheta)^3-(\Theta_\e(t_{\rm max}, x_{\rm max})-\vartheta)^3<0\Rightarrow |\Phi_\e |(t_{\rm max}, x_{\rm max})<\Theta_\e(t_{\rm max}, x_{\rm max}),
$$
which is in contradiction with the choice of $(t_{\rm max}, x_{\rm max})$. We deduce, then that
$$
|\Phi_\e (t,x)|\leq \frac{1}{2\sqrt{\e t}}+\frac{A_6R^2}{R^2-|x|^2}+\vartheta\quad\mathrm{for}\;(t,x)\in\Big(0,\frac{T_M}{\e}\Big] \times [-R,R],\;\forall\;R>1.
$$
We note that for $\e<\e_0$ small enough we have $\frac{T_M}{\e}>\frac{2T}{\e}>\frac{T}{\e}+T>\frac{T}{\e}$. Letting $R\rightarrow\infty$ we deduce that
$$
|\Phi_\e (t,x)|\leq \frac{1}{2\sqrt{\e t}}+A_6+\vartheta\leq \frac{1}{2\sqrt{T}}+A_6+\vartheta\quad\mathrm{for}\;(t,x)\in\left[\frac{T}{\e},\frac{T}{\e}+T \right] \times \mathbb{R}.
$$
Finally we use the periodicity of $\Phi_\e $ to extend the result for all $t\in[0,+\infty)$ and rewriting the result in terms of $\phi_\e$ we obtain for some positive constant $A_7$
\begin{equation}
\label{Lip_phi}
|\partial_x\phi_\e|\leq A_7,\quad\text{ in }[0,+\infty)\times\mathbb{R}.
\end{equation}
\subsection{Derivation of the Hamilton-Jacobi equation with constraint.}
\label{limit_equation}
In this section we derive the Hamilton-Jacobi equation with constraint \eqref{LimitEq} using the regularity estimates in Theorem \ref{regularity_theo}.
\subsubsection{Convergence along subsequences of $\psi_\e$ and $\rho_\e$.}\label{subsubs:conv_psie_rhoe}
According to section \ref{Regularity}, $\{\psi_\e\}$ is locally uniformly bounded and equicontinuous, so by the Arzela-Ascoli Theorem after extraction of a subsequence, $\psi_\e(t,x)$ converges locally uniformly to a continuous function $\psi(t,x)$. Moreover from \eqref{Equi_u}, we obtain that $\psi$ does not depend on $t$, i.e $\psi(t,x)=\psi(x)$.\\
Furthermore, from the uniform bounds on $\rho_\e$ in \eqref{Rho_bouded} we obtain that $|\frac{d\rho_\varepsilon}{dt}|$ is also bounded. Then we apply the Arzela-Ascoli Theorem to guarantee the locally uniform convergence along subsequences of $\rho_\e(t)$, to a function $\widetilde{\varrho}(t)$ as $\e\rightarrow 0$.
\subsubsection{The Hamilton-Jacobi equation with constraint.}
\label{Eq_psi}
Here we use a perturbed test function argument (see for instance \cite{EvansHomog}), in order to prove that, $\psi(x)=\lim_{\e\to0}\psi(t,x)$ is in fact a viscosity solution of the following Hamilton-Jacobi equation.
\begin{equation}
\label{HJ_psi}
-\left|\p_x \psi+\frac{c}{2}\right|^2=\overline{a}(x)-\overline{\rho}-\frac{c^2}{4},
\end{equation}
where $\overline{\rho}=\frac{1}{T}\int_0^T\widetilde{\varrho}(t)dt$.
We prove that $\psi$ is a viscosity sub-solution and one can use the same type of argument to prove that it is also a super-solution.\\\\
Let us define the auxiliary ``cell problem":
\begin{equation}
\label{CellPb}
\left\{\begin{array}{cr}
\partial_t \phi=a(\textcolor{black}{e(t)},x)-\widetilde{\varrho}(t)-\overline{a}(x)+\overline{\rho},&(t,x)\in[0,+\infty)\times\mathbb{R}^d,\\
\phi(0,x)=0,\\
\phi:\;T-periodic.
\end{array}
\right.
\end{equation}
This equation has a unique smooth solution, that we can explicitly write:
$$
\phi(t,x)=-t(\overline{a}(x)-\overline{\rho})+\int_0^t(a(\textcolor{black}{e(t)},x)-\widetilde{\varrho}(t))dt.
$$
\\
Let $\varphi\in C^\infty(\mathbb{R})$ be a test function and assume that $\psi-\varphi$ has a strict local maximum at some point $x_0\in\mathbb{R}$, with $\psi(x_0)=\varphi(x_0)$. We must prove:
\begin{equation}
\label{CondViscSub}
-\left|\p_x \varphi(x_0)+\frac{c}{2}\right|^2-\overline{a}(x_0)+\frac{c^2}{4}+\overline{\rho}\leq 0.
\end{equation}
We define the perturbed test function $\Psi_\e(t,x)=\varphi(x)+\e \phi(t,x)$, such that $\psi_\e-\Psi_\e$ attains a local maximum at some point $(t_\e,x_\e)$. We note that $\Psi_\e$ converges locally uniformly to $\varphi$ as $\e\rightarrow0$ since $\phi$ is locally bounded by definition, and hence one can choose $x_\e$ such that $x_\e\to x_0$ as $\e\to 0$, (see Lemma 2.2 in \cite{GuyBarles}). Then $\Psi_\e$ satisfies:
$$
\frac{1}{\e}\partial_t \Psi_\e(t_\e,x_\e)-\e\p_{xx} \Psi_\e(t_\e,x_\e)-\left|\p_x \Psi_\e(t_\e,x_\e)+\frac{c}{2}\right|^2-a(\textcolor{black}{e(t_\e)},x_\e)+\frac{c^2}{4}+\rho_\e(t_\e)\leq 0,
$$
since $\psi_\e$ is a solution of \eqref{Eq_psi_eps}. The above line gives:
$$
\begin{array}{c}
\partial_t \phi(t_\e,x_\e)-\e\p_{xx} \varphi(x_\e)-\e^2\p_{xx} \phi(t_\e,x_\e)-\left|\p_x \varphi(x_\e)+\e\p_x \phi(t_\e,x_\e)+\frac{c}{2}\right|^2
\\
-a(\textcolor{black}{e(t_\e)},x_\e)+\frac{c^2}{4}+\rho_\e(t_\e)\leq 0.
\end{array}
$$
Using \eqref{CellPb}, this last equation becomes:
\begin{equation}
\label{EqLim}
\begin{array}{c}
-\e\p_{xx} \varphi(x_\e)-\e^2\p_{xx} \phi(t_\e,x_\e)-\left|\p_x \varphi(x_\e)+\e\p_x \phi(t_\e,x_\e)+\frac{c}{2}\right|^2
\\
+(\rho_\e-\widetilde{\varrho})(t_\e)-\overline{a}(x_\e)+\overline{\rho}+\frac{c^2}{4}\leq 0.
\end{array}
\end{equation}
Next we pass to the limit as $\e\rightarrow0$. We know from Subsection 3.3.1 that $\rho_\e\rightarrow\widetilde{\varrho}$ locally uniformly as $\e\rightarrow0$. Moreover $\phi$ is smooth with locally bounded derivatives with respect to $x$, thanks to its definition. Using these arguments and letting $\e\rightarrow0$ in \eqref{EqLim} we obtain \eqref{CondViscSub} which implies that $\psi$ is a viscosity sub-solution of \eqref{HJ_psi}.\\\\
Furthermore, note that $\psi$ is also bounded from above, by taking the limit as $\e\rightarrow0$ in \eqref{Bounds_psi_eps}, i.e.,
\begin{equation}
\label{Bouds_psi}
\psi(x)\leq c_1-c_2|x|,
\end{equation}
and attains its maximum. We claim that
$$
\max_{x\in\mathbb{R}}\psi(x)=0.
$$
Indeed, from the upper bound for $\rho_\e$ in \eqref{Rho_bouded}, the definition of $\psi_\e$ in \eqref{HopfCole} and the continuity of $\psi$, we obtain that $\psi(x)\leq0$. Moreover, from the locally uniform convergence of $\psi_\e$ to $\psi$, as $\e\to0$, and \eqref{Bounds_psi_eps} we deduce that $\max_{x\in\mathbb{R}} \psi(x)<0$ implies that $\psi_\e(x)<-\beta$, for all $x\in\mathbb{R}$ and $\e\leq\e_0$ and some positive constant $\beta$. This is in contradiction with the fact that $\rho_\e$ is bounded by below by a positive constant $\rho_m$ (we refer to section 4.3 of \cite{Mirrahimi&Figueroa_pprint} for more details). \\
\section{Uniqueness of the viscosity solution to (\ref{LimitEq}) and explicit identification}
\label{Uniqueness}
In this section we {provide the proof of Theorem \ref{Unique_Identif}.} To this end, we first derive an equivalent Hamilton-Jacobi equation to \eqref{LimitEq} by mean of the Liouville transformation and prove some properties of the eigenvalue $\lambda_{c,\e}$. We then prove the uniqueness of the viscosity solution to such equivalent equation. This allows us to establish the uniqueness of the solution to \eqref{LimitEq} and to identify it explicitly. {Finally, we provide the proof of the convergence of $n_\e$ to the Dirac mass given by \eqref{Dirac_mass_conv}.}
\subsection{Derivation of an equivalent Hamilton-Jacobi equation.}\label{subs:equiv_HJ-eqtn}
In this subsection, we define a new function
\begin{equation}
\label{psi_u}
u(x) := \psi(x)+\frac{c}{2}x,
\end{equation}
which solves the following Hamilton-Jacobi equation in the viscosity sense
\begin{equation}
\label{HJ_u}
-|\p_xu|^2=\overline{a}(x)-\overline{\rho}-\frac{c^2}{4}.
\end{equation}
Note that the transformation \eqref{psi_u} is indeed analogous to the Liouville transformation presented in Section \ref{Liouville}.\\
{We} have the following boundedness result for $u$.
\begin{lemma}
\label{lem:Upp_bound_u}
{The function $u(x)$, defined by \eqref{psi_u}, is locally bounded} and satisfies
\begin{equation}
\label{Upp_bound_u}
-A_1|x|^2-A_2\leq u(x)\leq c_1-c_2|x|+\frac{c}{2}x,\quad \forall\ x\in\mathbb{R},
\end{equation}
where the constants $A_1$, $A_2$, $c_1$ and $c_2$ are given in \eqref{Bounds_psi_eps}.
\end{lemma}
\begin{proof}
From Subsection \ref{Regularity} we got uniform bounds for $\psi_\e$ in \eqref{Bounds_psi_eps}, which lead to bounds on $\psi$. That is
$$
-A_1|x|^2-\frac{c}{2}x-A_2\leq\psi(x)\leq c_1-c_2|x|.
$$
Then, the bounds \eqref{Upp_bound_u} follow directly from the definition of $u(x)$ in \eqref{psi_u}.
\end{proof}\\\\
Therefore, we conclude that the function $u$ satisfies \eqref{Comp_Pb_u}. {In Subsection \ref{sec:uniq_u} we will prove a uniqueness result for \eqref{Comp_Pb_u} which will imply the uniqueness of $\psi$, the solution to \eqref{LimitEq}.}
\subsection{Some properties of the eigenvalue $\lambda_{c,\e}$.}
\label{ConvEigenvalue}
In this subsection we prove Theorem \ref{Unique_Identif}-(i). {We also establish that $\lambda_{c,\e}$ is uniformly bounded above and below by negative constants and derive some properties of the limit, along subsequences, of $\lambda_{c,\e}$.} \\
\noindent
From the equation \eqref{Eigenvalue_lambda_ceps} we can integrate in $\mathbb{R}$, divide by $\int_\mathbb{R} p_{c\e}(t,x)dx$ and integrate again in $t\in[0,T]$ and obtain
\begin{equation}
\label{Q_eps&lambda_eps}
\lambda_{c,\e}=-\dfrac{1}{T}\displaystyle\int_0^TQ_{c\e}(t)dt.
\end{equation}
where $Q_{c\e}(t)$ is defined analogously to \eqref{PyQ} from the periodic eigenfunction $p_{c\e}$. {We next use the relationship between the solution $n_\e$ to \eqref{n_eps} and the eigenfunction $p_{c\e}$ to {obtain the first claim of Theorem \ref{Unique_Identif}.} Indeed, from equation \eqref{n_eps} after an integration in $x\in\mathbb{R}$ we obtain:
$$
\frac{d \rho_\e(t)}{d t}=\int_\mathbb{R} n_\e(t,x)a(\textcolor{black}{e(t)},x)dx- \rho^2_\e(t).
$$
We divide by $\rho_\e(t)$ and use the relation between $n_\e$ and $p_{c\e}$ inside of the integral, that is:
$$
\rho_\e(t)+ \frac{d}{dt}\ln\rho_\e(t)=\frac{\int_\mathbb{R} p_{c\e}(t,x)a(\textcolor{black}{e(t)},x)dx}{\int_\mathbb{R} p_{c\e}(t,y)dy}.
$$
Note that the RHS is exactly $Q_{c\e}$. We then integrate in $[0,T]$ and using \eqref{Q_eps&lambda_eps} and the $T-$periodicity of $\rho_\e$ we deduce that
\begin{equation}
\label{rho_e_lamnda_e}
\overline{\rho}_\e=-\lambda_{c,\e},
\end{equation}
}
\noindent
{
We next prove that $\lambda_{c,\e}$ is uniformly bounded above and below by negative constants. Combining \ref{Q_eps&lambda_eps} and \eqref{a_W3inf} we obtain that
$$
-d_0\leq \lambda_{c,\e}.
$$
Moreover, since we are in the case {$c<\displaystyle\liminf_{\e\to0} c^*_{\e}$}, we can find a positive constant $\tau$ such that for every $\e\leq\e_0$, with $\e_0$ small enough we have $c< c^*_{\e}-\tau$. Then from the definition of $c^*_{\e}$ we deduce {that}
$$
c<2\sqrt{-\lambda_{0,\e^2}}-\tau=2\sqrt{-\lambda_{c,\e}+\frac{c^2}{4}}-\tau,
$$
which leads to
$$
\lambda_{c,\e}<-\frac{c\tau}{2}-\frac{\tau^2}{4},
$$
and hence, for $\lambda_m =\frac{c\tau}{2}+\frac{\tau^2}{4}$ we obtain
\begin{equation}
\label{lambda_m}
\lambda_{c,\e}\leq -\lambda_m<0.
\end{equation}
Thus $(\lambda_{c,\e})_\e$ is uniformly bounded {above and below by negative constants}. This implies that we can extract a subsequence, still called $\lambda_{c,\e}$, which converges as $\e\to 0$ to some negative value $\lambda_1$. Moreover passing to the limit as $\e\to0$ in assumption \eqref{a_lambda_neg} we obtain, for all such limit values $\lambda_1$,
\begin{equation}
\label{a_neg_lambda_limit}
\overline{a}(x)\leq -\delta-\lambda_1,\quad \forall|x|\geq R_0.
\end{equation}
Note that passing to the limit of $\overline\rho_\e$ as $\e\to 0$ along the same subsequence, we obtain that
\begin{equation}
\label{rho_lambda1}
\overline{\rho}=-\lambda_1.
\end{equation}
}
\subsection{Uniqueness and explicit formula for $u(x)$.}
\label{sec:uniq_u}
In this subsection we prove the uniqueness of the viscosity solution of the Hamilton-Jacobi equation \eqref{Comp_Pb_u}. To this end we consider the Hamilton-Jacobi equation as follows
\begin{equation}
\label{MyH}
-\vert \p_x u\vert^2=h(x),\quad x\in\Omega,
\end{equation}
where $h\in C^1(\mathbb{R})$. Note that this corresponds to our problem for $h(x)=\bar h(x):= \overline{a}(x)-\overline{\rho}-\frac{c^2}{4}$. \\
We divide the proof of the uniqueness result into several steps. We first prove that, in the case where $\Omega$ is an open bounded domain and $h<0$ in $\Omega$, a viscosity solution to \eqref{MyH} can be uniquely determined {by} its values on the boundary of $\Omega$. We then use this property and \eqref{Upp_bound_u} to prove that in our problem it is not possible that $\bar h(x)<0$ for all $x\in\mathbb{R}$. We prove indeed that $\max \bar h(x)=0$ and this maximum is attained only at the point $x_m$. Finally we use these properties to conclude that $u$ is indeed uniquely determined by an explicit formula.
\textbf{Step 1: If $h<0$ and $\Omega$ is bounded then the viscosity solution to \eqref{MyH} is uniquely determined by its values on the boundary of $\Omega$.} Suppose that $h(x)<0$, for every $x\in\Omega$. For this problem we obtain uniqueness of the viscosity solution thanks to {a representation formula for the function $u$. Indeed, for $\Omega$ bounded we define $L(x,y)$ as follows}
\begin{equation}
\label{Def_L}
\begin{array}{rcl}
L(x,y)&=&\sup\left\{\int_0^{T_0}-\sqrt{-h(\xi(t))}dt/ (T_0,\xi)\text{ such that }\xi(0)=x,\ \xi(T_0)=y, \right. \\
&&\left. \xi(t)\in\Omega, \forall t\in[0,T_0], \left\vert\frac{d\xi}{dt}\right\vert\leq 1\text{ a.e in }[0,T_0]\right\},
\end{array}
\end{equation}
{and in \cite{Lions}-Chapter 5 the following is proved.}
\begin{proposition}
\label{Uniq_bounded_domain}
Assume that $h(x)<0$, $\forall x\in\Omega$, {with $\Omega$ a bounded domain}. The function
$$
u=\displaystyle\inf_{y\in\p\Omega}[\varphi(y)+L(x,y)],
$$
is the unique viscosity solution of
$$
\vert D u\vert=\sqrt{-h(x)} \text{ in } \Omega;\quad u=\varphi \text{ on } \p\Omega.
$$
\end{proposition}
\textbf{Step 2: $\max_{x\in\mathbb{R}} \bar h(x)=\bar h(x_m)=0$ and the maximum is only attained at this point.}
We assume in the contrary that $\max_{x\in\mathbb{R}}\bar h(x)<0$. We consider $\Omega=B_{R'}=(-R',R')$ for $R'>0$, to be chosen later. According to step 1, we can express the value of the viscosity solution of \eqref{MyH} at the point $0$, for $h(x)=\bar h(x)$, as follows:
{$$
u(0)=\max \left\{u(-R')-\left\vert\int_{-R'}^0\sqrt{-\bar h(y)}dy\right\vert; u(R')-\left\vert\int^{R'}_0\sqrt{-\bar h(y)}dy\right\vert\right\}.
$$}
Note that thanks to \eqref{a_neg_lambda_limit} and \eqref{rho_lambda1}, we obtain that
$$
\sqrt{\delta+\frac{c^2}{4}}\leq \sqrt{-\bar h(y)},\quad \forall \vert y\vert\geq R_0.
$$
We deduce that, for all $R'>R_0$,
$$
\int_{-R'}^0\sqrt{-\bar h(y)}dy\geq \sqrt{\delta+\frac{c^2}{4}}\,(R'-R_0),\quad
\text{ and }\quad
\int_0^{R'}\sqrt{-\bar h(y)}dy\geq \sqrt{\delta+\frac{c^2}{4}}\,(R'-R_0).
$$
Next we combine the above inequalities with the third line of \eqref{Comp_Pb_u} to obtain
$$
\begin{array}{rl}
u(0)\leq\max &\left\{c_1-R'\sqrt{\delta+\frac{c^2}{2}}-\left(R'-R_0\right)\sqrt{\delta+\frac{c^2}{4}};\right.\
\\
& \left. \; \; c_1+cR'-R'\sqrt{\delta+\frac{c^2}{2}}-\left(R'-R_0\right)\sqrt{\delta+\frac{c^2}{4}}\right\},
\end{array}
$$
for $c_1$ given in \eqref{Comp_Pb_u}. This implies that, taking $R'$ arbitrarily large, $u(0)$ is arbitrarily small which is a contradiction. Therefore the assumption on $\bar h(x)$ of being strictly negative in $\Omega$ is false. \\
We have proved that $\bar h(x)$ vanishes at some point $x\in\mathbb{R}$. Note also from \eqref{HJ_u} that
$$
\bar{a}(x)-\bar{\rho}-\frac{c^2}{4}\leq0,
$$
and $\max_{x\in\mathbb{R}} \bar h(x)$ is attained at the unique maximum point of $\bar a$, which is $x_m$.\\
\textbf{Step 3: Identification of $u$ in $\mathbb{R}$.} We now prove that the solution $u$ is uniquely determined by its value at the maximum point of $\bar h(x)$. That is, for all $x\in \mathbb{R}$\\
\begin{equation}
\label{u_x_m}
u(x)=-\left\vert\displaystyle\int_{x_m}^x\sqrt{-\bar h(y)}dy\right\vert+u(x_m).
\end{equation}
To this end we choose $0<R$, and $0<R'$ such that $R<<R'$ and we consider the domain $\overline{B_{R'}}=[-R',x_m]\cup[x_m,R']$. Note that $\bar h<0$ in the sets $(-R',x_m)$ and $(x_m,R')$. We can thus apply Proposition \ref{Uniq_bounded_domain} in these domains:
$$
u(x)=\max \left\{u(-R')-\left\vert\int_{-R'}^{x}\sqrt{-\bar h(y)}dy\right\vert; u(x_m)-\left\vert\displaystyle\int_{x_m}^x\sqrt{-\bar h(y)}dy\right\vert . \right\}, \quad \forall x\in (-R,x_m),
$$
$$
u(x)=\max \left\{ u(R')-\left\vert\int_{R'}^{x}\sqrt{-\bar h(y)}dy\right\vert ; u(x_m)-\left\vert\displaystyle\int_{x_m}^x\sqrt{-\bar h(y)}dy\right\vert . \right\}, \quad \forall x\in (x_m,R),
$$
We next prove the following inequalities for $R'$ large enough,
$$
\left\vert\displaystyle\int_{x_m}^x\sqrt{-\bar h(y)}dy\right\vert-u(x_m)\leq
\left\vert\int_{-R'}^x\sqrt{-\bar h(y)}dy\right\vert-u(-R'),\quad \forall x\in (-R,x_m),
$$
$$
\left\vert\displaystyle\int_{x_m}^x\sqrt{-\bar h(y)}dy\right\vert-u(x_m)\leq
\left\vert \int^{x}_{R'}\sqrt{-\bar h(y)}dy\right\vert-u(R'), \quad \forall x\in (x_m,R).
$$
and combine them with the above lines to obtain \eqref{u_x_m} for all $x\in[-R,R]$. Since $R$ is arbitrary we thus obtain \eqref{u_x_m}. \\
Suppose that $x_m<x<R$. We prove the second inequality (the first one follows from an analogous argument). We claim that, for $R'$ large enough
\begin{equation}
\label{Sup_Ineq}
-\displaystyle\int_{x_m}^x\sqrt{-\bar h(y)}dy+u(x_m)+\int_{x}^{R'}\sqrt{-\bar h(y)}dy-u(R')\geq 0.
\end{equation}
Indeed, for $x\in [x_m,R]$ we have
\begin{equation}
\label{xm_x}
-\int_{x_m}^{x}\sqrt{-\bar h(y)}dy\geq -\int_{x_m}^{R}\sqrt{-\bar h(y)}dy.
\end{equation}
Moreover, from the upper bound for $u$ in \eqref{Upp_bound_u} we obtain for all $x\in B_{R'}(x_m)$,
\begin{equation}
\label{u_R'}
u(x)\leq c_1-c_2|x|+\frac{c}{2}x, \quad \Rightarrow u(R')\leq c_1+\frac{c}{2}R'.
\end{equation}
Furthermore, following similar arguments as in the previous step we obtain that:
\begin{equation}
\label{h_delta_c}
\sqrt{-\bar h(y)}\geq\sqrt{\delta+\frac{c^2}{4}},\quad \forall y\in(R_0,R').
\end{equation}
Finally, putting together \eqref{xm_x}, \eqref{u_R'} and \eqref{h_delta_c} we obtain:
$$
\begin{array}{l}
-\displaystyle\int_{x_m}^x\sqrt{-\bar h(y)}dy+u(x_m)+\int_{x}^{R'}\sqrt{-\bar h(y)}dy-u(R')
\\
\geq -\displaystyle\int_{x_m}^R\sqrt{-\bar h(y)}dy+u(x_m)+(R'-R_0)\sqrt{\delta+\frac{c^2}{4}}-\frac{c}{2}R'-c_1
\\ \geq0,
\end{array}
$$
for $R'$ large enough.\\
\textbf{Step 4: Uniqueness of $u$.}
We finally determine the value of $u$ at $x_m$ which leads to the uniqueness and an explicit formula for $u$ thanks to \eqref{u_x_m}.\\
Replacing the value of $\overline h$ in \eqref{u_x_m}, we obtain
\begin{equation}
\label{Solution_u}
u(x)=u(x_m)-\left\vert \displaystyle\int_{x_m}^{x}\sqrt{\overline{\rho}+\frac{c^2}{4}-\overline{a}(y)}dy\right\vert,\quad \forall x\in\mathbb{R}.
\end{equation}
This directly implies that $u$ is in fact a classical solution for $x\in\mathbb{R}$ which attains its maximum at $x=x_m$. We also recall the second property in \eqref{Comp_Pb_u}, that is
$$
\max_{x\in \mathbb{R}} u(x)-\f{c}{2}x=0.
$$
We denote the set of maximum points of $u(x)-\f{c}{2}x$ by $X^*$, i.e
$$
X^*:=\{x^* \in \mathbb{R} \text{ such that $u(x^*)-\f{c}{2}x^*=0$}\}.
$$
Let $x^*\in X^*$, we evaluate the above formula of $u$ at $x^*$ in order to obtain an expression for $u(x_m)$. This implies
\begin{equation}
\label{Solution_psi}
u(x)=\frac{c}{2} x^* +\left\vert \int_{x_m}^{x^*}\sqrt{\overline{a}(x_m)-\overline{a}(y)}dy\right\vert-\left\vert \int_{x_m}^{x}\sqrt{\overline{a}(x_m)-\overline{a}(y)}dy\right\vert,\quad \forall x\in\mathbb{R}.
\end{equation}
Moreover, we have
$$
u(x_m)-\f{c}{2}x_m \leq u(x^*)-\f{c}{2}x^*=0,
$$
which implies that
$$
\frac{c}{2}(x^*-x_m)\leq 0,
$$
and hence $x^*\leq x_m$. Note also that we have
$\overline{a}(x_m)=\overline{\rho}+\frac{c^2}{4}$ from step 2
and $\bar a(x^*)=\bar \rho$ thanks to \eqref{LimitEq}. Combining these properties with assumption \eqref{x_var} it follows that $x^*=\overline{x}$, and hence $u$ is uniquely determined. As a consequence we obtain the explicit formula \eqref{Exp_Sol_psi} for $\psi(x)$. This ends the proof of Theorem \ref{Unique_Identif}-(ii).
\subsection{Convergence to the Dirac mass.}\label{subs:conv_dirac_mass}
We deal in this subsection with the result for the convergence of $n_\varepsilon$, that is Theorem \ref{Unique_Identif}-(iii).\\
Call $f_\e(t,x)=\dfrac{n_\e(t,x)}{\rho_\e(t)}$, then $f_\e$ is uniformly bounded in $L^\infty(\mathbb{R}_+,L^1(\mathbb{R}))$. Next, we fix $t\geq0$, and we prove that $f_\e(t,\cdot)$, converges, along subsequences, to a measure, as follows
$$
f_\varepsilon(t,\cdot)\rightharpoonup\delta(\cdot-\bar{x})\quad\mathrm{as}\quad \varepsilon\rightarrow0,
$$
weakly in the sense of measures.\\
Indeed, we already know that
$$
\max_{x\in\mathbb{R}}\psi(x)=\psi(\bar{x})=0\quad \text{ and }\quad \psi(x)\leq c_1-c_2\vert x\vert,
$$
for $\bar x$ given in Theorem \ref{Unique_Identif}. This implies that for any $\zeta>0$, there exists $\beta>0$ such that $\psi(x)\leq -\beta$ for every $x\in\mathbb{R}\setminus [\bar x-\zeta,\bar x+\zeta]$. \\
We denote $\mathcal{O}=\mathbb{R}\setminus [\bar{x}-\zeta,\bar{x}+\zeta]$ and choose $\chi\in C_c(\mathcal{O})$, such that $\mathrm{supp}\; \chi\subset \mathcal{K}$, for some compact set $\mathcal{K}$, then it follows that
$$
\left|\int_{\mathcal{O}}f_\e(t,x)\chi(x)dx\right|\leq \frac{1}{\rho_m}\int_{\mathcal{O}}e^{\frac{\psi_\e(t,x)}{\e}}|\chi(x)|dx\leq \frac{1}{\rho_m}\int_{\mathcal{K}}e^{{\frac{\psi_\e(t,x)}{\e}}}|\chi(x)|dx.
$$
From the locally uniform convergence of $\psi_\e$, to $ \psi(x)$, we obtain that there exists $\e_0>0$ such that $\forall\e<\e_0$, $\psi_\e(t,x)\leq-\frac{\beta}{2}$, $\forall x\in\mathcal{K}$, and hence
$$
\int_{\mathcal{K}}e^{\frac{\psi_\e(t,x)}{\e}}|\chi(x)| dx\leq\int_{\mathcal{K}}e^{-\frac{\beta}{2\e}}|\chi(x)|dx\rightarrow0\quad\mathrm{as}\;\e\rightarrow0,
$$
since $\chi$ is bounded in $\mathcal{K}.$ Therefore, thanks to the uniform $L^1$ bound of $f_\e$, we obtain that $f_\e$ converges weakly in the sense of measures and along subsequences to $\mu\delta(x-\bar{x})$ as $\e\to0$. Then to prove that in fact, $\mu=1$ we can proceed as in Section 4.3 in \cite{Mirrahimi&Figueroa_pprint}.
\subsection{Identification of the limit of $\rho_\varepsilon$.}\label{identif_rhoe}
In order to identify the limit of $\rho_\varepsilon$ we first write thanks to Proposition \ref{ConvLambda} and Proposition \ref{prop2} the following explicit expression for $\rho_\e$:
\begin{equation}
\label{Exp_rho_eps}
\rho_\e(t)=\frac{1-\exp\left[-\displaystyle\int_0^{T}Q_{c\e}(s)ds\right]}{\exp\left[-\displaystyle\int_0^{T}Q_{c\e}(s)ds\right]\displaystyle\int_t^{t+T}\exp\left[\displaystyle\int_t^s Q_{c\e}(\theta)d\theta\right]ds},
\end{equation}
where $Q_{c\e}$ is defined analogously to \eqref{PyQ}, using the periodic eigenfunction $p_{c\e}$ of problem \eqref{Eigenvalue_lambda_ceps}. \\
We then compute the limit of $Q_{c\e}$ as $\e\to0$. We know that $p_{c\e}(t,x)=\dfrac{n_\e(t,x)}{\rho_\e(t)}\displaystyle\int_{\mathbb{R}}p_{c\e}(t,y)dy$. {Replacing $p_{c\e}$ by this quantity in the formula for $Q_{c\e}$} we obtain
$$
\begin{array}{rl}
Q_{c\e}(t)&=\dfrac{\displaystyle\int_{\mathbb{R}}a(e(t),x)p_{c\e}(t,x)dx}{\displaystyle\int_{\mathbb{R}}p_{c\e}(t,x)dx}
=\dfrac{\displaystyle\int_{\mathbb{R}}a(\textcolor{black}{e(t)},x)\dfrac{n_\e(t,x)}{\rho_\e(t)}\int_{\mathbb{R}}p_{c\e}(t,y)dy dx}{\displaystyle\int_{\mathbb{R}}p_{c\e}(t,x)dx}\\
&=\dfrac{\displaystyle\int_{\mathbb{R}}a(e(t),x)n_\e(t,x)dx}{\rho_\e(t)}.
\end{array}
$$
{From the previous subsection} we deduce that:
$$
\lim_{\e\rightarrow0}Q_{c\e}(t)=\lim_{\e\rightarrow0}\int_{\mathbb{R}}f_\e(t,x)a(\textcolor{black}{e(t)},x)dx=a(\textcolor{black}{e(t)},\bar x).
$$
Finally we can pass to the limit in the expression \eqref{Exp_rho_eps} for $\rho_\e$, to obtain the following explicit formula for $\tilde{\varrho}$
\begin{equation}
\label{Exp_rho}
\widetilde{\varrho}(t)=\frac{1-\exp\left[-\displaystyle\int_0^{T}a(s,\bar x)ds\right]}{\exp\left[-\displaystyle\int_0^{T}a(s,\bar x)ds\right]\displaystyle\int_t^{t+T}\exp\left[\displaystyle\int_{t}^s a(\theta,\bar x)d\theta\right]ds},
\end{equation}
which is in fact the unique periodic solution of the equation \eqref{eq_varrho} {thanks to Proposition \ref{prop2}}. Therefore using the convergence result for $\rho_\varepsilon$ we deduce finally \eqref{Dirac_mass_conv} and this ends the proof of Theorem \ref{Unique_Identif}.
\section{Approximations of the eigenvalue}
\label{Eigen_Approx}
In this section we prove Theorem \ref{Asymp_Expansion}, i.e., the asymptotic expansions \eqref{Eigenvalue_Approx} and \eqref{Crit_Speed_Approx}. Note that the first equality in \eqref{Eigenvalue_Approx} has been already obtained in Section \ref{ConvEigenvalue}. \\
To this end we develop an asymptotic approximation {of order $\e$ of the eigenvalue $\lambda_{c,\e}$ given by the eigenvalue problem \eqref{Eigenvalue_lambda_ceps}}.
To obtain such asymptotic expansion we construct an approximate eigenfunction
$\widetilde{p}_\e$ corresponding to an approximate eigenvalue $\widetilde{\lambda}_\e$ which solves an equation close to \eqref{Eigenvalue_lambda_ceps}. We then use Proposition \ref{Prop1} to prove that $\widetilde{\lambda}_\e$ approximates $\lambda_{c,\e}$ with an error of order $\e^2$. \\
To construct an approximate eigenfunction, we first try to approximate $w_\e$, obtained from the Hopf-Cole transformation of $p_{c\e}$ as follows:
\begin{equation}
\label{p_ce_e=w_e}
p_{c\e}(t,x)=\frac{1}{\sqrt{2\pi\e}}e^{\frac{w_\e(t,x)}{\e}}.
\end{equation}
One can verify that $w_\e$ solves:
\begin{equation}
\label{v_eps}
\frac{1}{\e}\partial_tw_{\e}-\e\partial_{xx}w_{\e}=\left\vert\partial_x w_\e+\frac{c}{2}\right\vert^2+a(\textcolor{black}{e(t)},x)+\lambda_{c,\e}-\frac{c^2}{4}.
\end{equation}
We can obtain similar bounds for $w_\e$ as for $\psi_\e$, which guarantee the convergence along subsequences of $w_\e$ to certain function $w=w(x)$, which is in fact the limit of the whole sequence $w_\e$, and satisfies the following Hamilton-Jacobi equation in the viscosity sense
\begin{equation}
\label{HJ_w}
-\left\vert\partial_x w+\frac{c}{2}\right\vert^2=\overline{a}(x)+\lambda_1-\frac{c^2}{4}.
\end{equation}
\begin{remark}
Note that in order to obtain the limit equation \eqref{HJ_w} we can argue exactly as for $\psi_\e$ in Section \ref{Eq_psi}, by a "perturbed test function" argument, (see also \cite{Mirrahimi&Figueroa_pprint}).
\end{remark}
{Note that, thanks to theorems \ref{Conv_HJ} and \ref{Unique_Identif}, $\lambda_1=-\overline \rho$ and $ \psi(x)$ is a solution to \eqref{HJ_w}.} We then, write (formally)
\begin{equation}
\label{Formal_exp_phi_lambda}
w_\e(t,x)=\psi(x)+\e \phi(t,x)+\e^2\omega(t,x)+o(\e^2)\quad\text{and} \quad \lambda_{c,\e}={-\overline \rho}+\e\lambda_2+o(\e),
\end{equation}
for some $T-$periodic functions $\phi$ and $\omega$, and we construct the following approximated eigenpair:
\begin{equation}
\label{Def_approx_eigenpair}
\widetilde{\psi_\e}=\psi+\e \phi\quad\text{and} \quad \widetilde{\lambda}_{\e}={-\overline \rho}+\e\lambda_2.
\end{equation}
We then substitute this pair $(\widetilde{\psi_\e}, \widetilde{\lambda}_{\e})$ into \eqref{v_eps} and obtain:
\begin{equation}
\label{Eq_Approx_eigenpair}
\p_t \phi-c\psi_x-c\e\p_x\phi-\e\psi_{xx}-\e^2\p_{xx} \phi=|\psi_x +\e\p_x \phi|^2+a(\textcolor{black}{e(t)},x) {-\overline \rho}+\e\lambda_2+o(\e),
\end{equation}
where the notations $\psi_x$ and $\psi_{xx}$ correspond respectively to the first and second derivative of $\psi$.\\
Regrouping {the terms with} similar powers of $\e$ we obtain the following system for $\phi$,
\begin{equation}
\label{System_approx}
\left\{
\begin{array}{l}
\displaystyle\p_t \phi=\left|\psi_x +\frac{c}{2}\right|^2+a(\textcolor{black}{e(t)},x)-\frac{c^2}{4} {-\overline \rho},\\\\
\displaystyle-\psi_{xx}=[2\psi_x +c]\frac{1}{T}\int_0^T\p_x \phi(t,x)dt+\lambda_2.
\end{array}
\right.
\end{equation}
We remark that the previous system has a unique solution $\phi$ up to addition by a constant. Indeed, from equation \eqref{LimitEq} we obtain
$$
\displaystyle\p_t \phi=a(\textcolor{black}{e(t)},x)-\bar{a}(x).
$$
Integrating in $[0,t]$ leads to
$$
\phi(t,x)=\phi(0,x)+\int_0^ta(\tau,x)d\tau-t\bar{a}(x),
$$
and the value of $\phi(0,x)$ can be obtained from the second equation in \eqref{System_approx} once we fix $\phi(0,x_m)$.
Note that here we use the fact that $2\psi_x + c$ vanishes only at the point $x_m$. \\
We now define $\widetilde{p}_\e(t,x):=\frac{1}{\sqrt{2\pi\e}}e^{\frac{\widetilde{\psi}_\e(t,x)}{\e}}$, and use the system \eqref{System_approx}, to obtain the equality:
\begin{equation}
\label{App_p_eps_tilde}
\mathcal{P_\e}\widetilde{p}_\e-\widetilde{\lambda}_\e\widetilde{p}_\e=-\e^2\left(|\p_x \phi|^2+\p_{xx}\phi\right)\widetilde{p}_\e,
\end{equation}
for $\mathcal{P_\e}$ the following parabolic operator
$$
\mathcal{P_\e}p=\p_tp-c\e\p_xp-\e^2\p_{xx}p-a(\textcolor{black}{e(t)},x)p.
$$
We denote
$$
\lambda_{\e^+}=\widetilde{\lambda}_\e+\e^2K, \quad\text{ and }\quad \lambda_{\e^-}=\widetilde{\lambda}_\e-\e^2K,
$$
with
\begin{equation}
\label{Def_K}
K=\Vert \vert\p_x\phi\vert^2+\p_{xx}\phi\Vert_{L^\infty},
\end{equation}
where the well definition of $K$ is guaranteed by the next lemma which is proved in the next subsection.
\begin{lemma}
\label{Borne_operat}
The constant $K$ given in \eqref{Def_K} is well defined. Moreover the function $\phi$ computed above solves \eqref{System_approx} with
{$\lambda_2=\sqrt{-\bar a_{xx}(x_m)/2}$}.
\end{lemma}
{We then deduce from \eqref{App_p_eps_tilde} that}
$$
\widetilde{p}_\e\lambda_{\e^-}\leq \p_t \widetilde{p}_\e-c\e \p_x \widetilde{p}_\e -\e^2\p_{xx}\widetilde{p}_\e-a(\textcolor{black}{e(t)},x)\widetilde{p}_\e\leq\widetilde{p}_\e\lambda_{\e^+}.
$$
{We next} define the functions
$$
\overline{q}_\e(t,x)=\widetilde{p}_\e(t,x)e^{- t\lambda_{\e^-}},\quad \underline{q}_\e(t,x)=\widetilde{p}_\e(t,x)e^{-t \lambda_{\e^+}}.
$$
One can verify that $\overline{q}_\e$ and $\underline{q}_\e$ are super- and sub-solution of the linear problem \eqref{modelLap_in_m} with $\sigma=\e^2$ and $\widetilde{c}=c\e$, that is
$$
\partial_t \underline{q}_\e-c\e\p_x \underline{q}_\e-\e^2\p_{xx} \underline{q}_\e \leq \underline{q}_\e a(\textcolor{black}{e(t)},x),
$$
$$
\partial_t \overline{q}_\e-c\e\p_x \overline{q}_\e-\e^2\p_{xx} \overline{q}_\e \geq \overline{q}_\e a(\textcolor{black}{e(t)},x).
$$
We then apply a Comparison Principle and obtain that the solution $q_\e(t,x)$ to the following linear problem
\begin{equation}
\label{linear_eps}
\left\{
\begin{array}{ll}
\partial_t q_\e-c\e\p_x q_\e-\e^2\p_{xx} q_\e = q_\e a(\textcolor{black}{e(t)},x),\\
q_\e(0,x)=\widetilde{p}_\e(0,x),
\end{array}
\right.
\end{equation}
satisfies
$$
\underline{q}_\e(t,x)\leq q_\e(t,x)\leq \overline{q}_\e(t,x),\quad\forall(t,x)\in\mathbb{R}_+\times \mathbb{R}.
$$
From the proof of Proposition 1 in Section 2.2 (see equation \eqref{Exp_Conv_m}), applied to the case $\sigma=\e^2$ and $c=c\e$ we know that $q_\e$
converges exponentially fast as $t\to +\infty$ to the periodic eigenfunction in \eqref{Eigenvalue_lambda_ceps}, (see also \cite{huska08}); that is, we can write for some positive constants $\alpha$ and $\beta$,
\begin{equation}
\label{Conv_linear_eps}
\Vert q_\e e^{t\lambda_{c,\e}}-\alpha p_{c\e}\Vert_{L^\infty}\leq e^{-\beta t}.
\end{equation}
We recall that $q_\e e^{t\lambda_{c,\e}}$ can indeed be written as
$$
q_\e e^{t\lambda_{c,\e}}=\widetilde{q}_{\e,1}+\widetilde{q}_{\e,2},
$$
with $\widetilde{q}_{\e,1}(t,\cdot)\in \text{span}\{p_{c\e}(t,\cdot)\}$, $\widetilde{q}_{\e,2}\to 0$ exponentially fast and
$$
\int_\mathbb{R} \widetilde{q}_{\e,2}(t,x)p_{c\e}^*(t,x)dx=0,
$$
where $p_{c\e}^*$ is the principal eigenfunction to the adjoint problem
\begin{equation}
\label{Pb_adjoint}
-\p_t p_{c\e}^*+c\e\p_x p_{c\e}^*-\e^2\p_{xx}p_{c\e}^*=(a(\textcolor{black}{e(t)},x)+\lambda_{c,\e})p_{c\e}^*,
\end{equation}
(see Theorem 2.2 {in} \cite{huska08} and the proof of Lemma 6 {in} \cite{Mirrahimi&Figueroa_pprint}). The positivity of $\alpha$ is then derived from the fact that $q_\e(0,x)$ and $p_{c\e}^*$ are positive functions. \\
On the one hand equation \eqref{Conv_linear_eps} implies that,
$$
0\leq {\widetilde{p}_\e} e^{(-\lambda_{\e^+}+\lambda_{c,\e})t}\leq \alpha p_{c\e}+e^{-\beta t}.
$$
Since $p_{c\e}$ and $\widetilde{p}_\e$ are time-periodic functions, then necessarily
$$
\lambda_{c,\e}-\lambda_{\e^+}\leq 0,
$$
otherwise we get a contradiction as $t\to+\infty$. Therefore\
\begin{equation}
\label{Left_eigenvalue}
\lambda_{c,\e}-\widetilde{\lambda}_\e\leq K \e^2,
\end{equation}
where $K$ is defined in \eqref{Def_K}.\\
On the other hand, from \eqref{Conv_linear_eps} we obtain
$$
{\widetilde{p}_\e}e^{(-\lambda_{\e^-}+\lambda_{c,\e})t}\geq \alpha p_{c\e}-e^{-\beta t}.
$$
Note that if $\lambda_{c,\e}-\lambda_{\e^-}\leq 0$ we obtain from the $T-$periodicity of the eigenfunctions, as $t\to+\infty$, that $p_{c\e}\leq 0$, which is also a contradiction. {We deduce} that
$$
\lambda_{c,\e}-\lambda_{\e^-}\geq 0.
$$
Therefore we have
\begin{equation}
\label{Right_eigenvalue}
\lambda_{c,\e}-\widetilde{\lambda}_\e\geq -K\e^2.
\end{equation}
Combining both inequalities \eqref{Left_eigenvalue} and \eqref{Right_eigenvalue} we write
\begin{equation}
\label{Borne_eigenvalues}
\big \vert \lambda_{c,\e}-({-\overline \rho}+\e\lambda_2)\big \vert\leq K\e^2,
\end{equation}
which leads thanks to {Theorem \ref{Unique_Identif}-(ii) and} Lemma \ref{Borne_operat} to an approximation for the eigenvalue of order $\e^2$ as follows:
$$
\lambda_{c,\e}=-\overline{a}(x_m)+\frac{c^2}{4}+\e \sqrt{-\bar{a}_{xx}(x_m)/2}+o(\e).
$$
The approximation \eqref{Crit_Speed_Approx} for the critical speed $c_\e^*$ can be derived from the above approximation and \eqref{Critical_c}. Indeed, from \eqref{Critical_c} and the definition of $c_\e^*$ we obtain
$$
c_\e^*=2\sqrt{\bar a(x_m)-\e\sqrt{-\frac{\bar a_{xx}(x_m)}{2}}+o(\e)}=2\sqrt{\bar a(x_m)}-\e {\sqrt{-\frac{\bar a_{xx}(x_m)}{2\,\bar a(x_m)}}}+o(\e).
$$
\subsection{Boundedness of $K$.} In this subsection we prove Lemma \ref{Borne_operat}. We provide the proof in several steps.\\
\begin{proof}(\textbf{Proof of Lemma \ref{Borne_operat}})
\subparagraph{{Step 1:} $\vert \p_x \phi\vert$ is bounded.}
An integration in $[0,T]$ of the first equation in \eqref{System_approx} gives us the
already known equation for $\psi$ in {\eqref{HJ_psi}. This allows us to rewrite the equation} as follows:
\begin{equation}
\label{Deriv_x_phi}
\p_t\phi=a(\textcolor{black}{e(t)},x)-\overline{a}(x)\Rightarrow \p_x \phi(t,x)=\p_x \phi(0,x)+\int_0^t a_x(\textcolor{black}{e(\tau)},x)d\tau-t\overline{a}_x(x),
\end{equation}
where $a_x$ and $\overline{a}_x$ denote the derivatives with respect to $x$ of $a(\textcolor{black}{e(t)},x)$ and $\overline{a}(x)$ respectively. This implies that in order to bound $\p_x\phi$ we just need to bound the derivative of $\phi$ at point $t=0$ since $a(\textcolor{black}{e(t)},x)\in L^\infty(\mathbb{R}_+,C^3(\mathbb{R}))$.\\
Then from the second equation in \eqref{System_approx} we obtain:
\begin{equation}
\label{Deriv_phi_def}
\frac{1}{T}\int_0^T\p_x\phi(t,x)dt=\dfrac{-\psi_{xx}(x)+\psi_{xx}(x_m)}{2\psi_x(x)+c},
\end{equation}
if the last formula is well defined, i.e., if all the derivatives exist.\\
Note that, an integration for $t\in[0,T]$ in the equation \eqref{Deriv_x_phi} leads to, (after dividing by $T$)
$$
\frac{1}{T}\int_0^T\p_x \phi(t,x)dt=\p_x \phi(0,x)+\frac{1}{T}\int_0^T\int_0^t a_x(\textcolor{black}{e(\tau)},x)d\tau dt-\frac{T}{2}\overline{a}_x(x),
$$
since $\p_x \phi(0,x)$ does not depend on $t$. We then deduce from the last formula and \eqref{Deriv_phi_def}
\begin{equation}
\label{Deriv_phi_0}
\p_x\phi(0,x)=\dfrac{-\psi_{xx}(x)+\psi_{xx}(x_m)}{2\psi_x(x)+c}+G(x),
\end{equation}
where
\begin{equation}
\label{G(x)}
G(x)=-\frac{1}{T}\int_0^T\int_0^t a_x(\textcolor{black}{e(\tau)},x)d\tau dt+\frac{T}{2}\overline{a}_x(x),
\end{equation}
is a regular function. We next prove that the derivatives involved in \eqref{Deriv_phi_def} exist. To this end we claim the following technical result.
\begin{lemma}
\label{Diff_psi}
The function $\psi(x)$ is twice differentiable for every $x\in\mathbb{R}$ and
\begin{equation}
\label{Deriv2_psi}
\psi_{xx}(x)=\left\{
\begin{array}{lr}
-\dfrac{\overline{a}_x(x)}{2\sqrt{\overline{a}(x_m)-\overline{a}(x)}},& x<x_m,\\\\
-\sqrt{-\overline{a}_{xx}(x_m)/2},& x=x_m,\\\\
\dfrac{\overline{a}_x(x)}{2\sqrt{\overline{a}(x_m)-\overline{a}(x)}},& x>x_m.
\end{array}
\right.
\end{equation}
\end{lemma}
\begin{proof}(\textbf{Proof of Lemma \ref{Diff_psi}})\\
Indeed, from the explicit formula \eqref{Exp_Sol_psi} we differentiate and obtain:
\begin{equation}
\label{Deriv_psi}
\psi_x(x)=\left\{
\begin{array}{lr}
-\frac{c}{2}+\sqrt{\overline{a}(x_m)-\overline{a}(x)},& x<x_m,\\
-\frac{c}{2}, & x=x_m,\\
-\frac{c}{2}-\sqrt{\overline{a}(x_m)-\overline{a}(x)},& x>x_m.
\end{array}
\right.
\end{equation}
We next compute
$$
\lim_{x\to x_m^+}\frac{\psi_x(x)-\psi_x(x_m)}{x-x_m}=\lim_{x\to x_m^+}\frac{-\sqrt{\bar{a}(x_m)-\bar a(x)}}{x-x_m}=\lim_{x\to x_m^+}\frac{-\sqrt{f(x)}}{x-x_m},
$$
where we have denoted $f(x)=\overline{a}(x_m)-\overline{a}(x)$. We write a Taylor expansion of $f$ around $x=x_m$, i.e.:
$$
f(x)=-\frac{1}{2}\bar a_{xx}(x_m)(x-x_m)^2-\frac{1}{6}\bar a_{xxx}(x_m)(x-x_m)^3+o((x-x_m)^3),
$$
since $f(x_m)=0$ and $x_m$ is a maximum point. It implies that:
$$
\lim_{x\to x_m^+}\frac{\psi_x(x)-\psi_x(x_m)}{x-x_m}=\lim_{x\to x_m^+}\dfrac{\bar a_x(x)}{2\sqrt{\bar a(x_m)-\bar a(x)}}=-\sqrt{-\overline{a}_{xx}(x_m)/2}.
$$
Note that $x_m$ being a maximum point, $\overline{a}(x_m)\geq\overline{a}(x)$, $\forall x\in\mathbb{R}$ and $\overline{a}_{xx}(x_m)\leq 0$. Following similar arguments one can prove that
$$
\lim_{x\to x_m^-}\frac{\psi_x(x)-\psi_x(x_m)}{x-x_m}=\lim_{x\to x_m^-}\dfrac{-\bar a_x(x)}{2\sqrt{\bar a(x_m)-\bar a(x)}}=-\sqrt{-\overline{a}_{xx}(x_m)/2}.
$$
\end{proof}
We pursue with the proof of Lemma \ref{Borne_operat}.\\
By substituting the derivatives of $\psi$ in \eqref{Deriv_phi_def} we obtain for every $x\neq x_m$:
{
\begin{equation}
\label{Deriv_phi}
\p_x\phi(0,x)=G(x)+\left\{
\begin{array}{lr}
\dfrac{\overline{a}_x(x)-\sqrt{-2\overline{a}_{xx}(x_m)(\overline{a}(x_m)-\overline{a}(x))}}{4(\overline{a}(x_m)-\overline{a}(x))},& x<x_m,\\\\
\dfrac{\overline{a}_x(x)+\sqrt{-2\overline{a}_{xx}(x_m)(\overline{a}(x_m)-\overline{a}(x))}}{4(\overline{a}(x_m)-\overline{a}(x))},& x>x_m.
\end{array}
\right.
\end{equation}
\\\\
We can bound $\p_x\phi(0,x)$ near to $x=x_m$. We write the limits as $x\to x_m$ in \eqref{Deriv_phi} in terms of $f$ and its derivatives $A=-\overline{a}_{xx}(x_m)/2$ and $B=-\overline{a}_{xxx}(x_m)/6$, and compute:
\begin{equation}
\label{Limit_d_phi}
\displaystyle\lim_{x\to x_m^{\mp}}\p_x\phi(0,x)=\lim_{x\to x_m^{\mp}} G(x)+\lim_{x\to x_m^{\mp}} \frac{- f'(x)\mp 2\sqrt{A}\sqrt{f(x)}}{4 f(x)},
\end{equation}
if both limits exist. Note that from the definition of $G$ in \eqref{G(x)} {we deduce that the first limit in \eqref{Limit_d_phi} exists and is equal to $G(x_m)$}. We, then, only need to compute the second one to guarantee the existence of $\p_x\phi(0,x_m)$. We compute both lateral limits separately:
$$
\begin{array}{l}
\lim_{x\to x_m^{-}} \dfrac{- f'(x)-2\sqrt{A}\sqrt{f(x)}}{4 f(x)}\\\\
=\lim_{x\to x_m^{-}}\frac{-[2A(x-x_m) +3B( x-x_m)^2+o((x-x_m)^2)]- 2A\vert x-x_m\vert\sqrt{1+\frac{B}{A}( x-x_m)+o(( x-x_m) )}}{4\left(A(x-x_m)^2+B(x-x_m)^3+o((x-x_m)^3\right))}\\\\
=\lim_{x\to x_m^{-}}\frac{-2A(x-x_m) -3B( x-x_m)^2+o((x-x_m)^2)+ 2A(x-x_m)(1+\frac{B}{2A}( x-x_m)+o(( x-x_m) )}{4\left(A(x-x_m)^2+B(x-x_m)^3+o((x-x_m)^3\right))}\\\\
=-\dfrac{B}{2A}.
\end{array}
$$
Following similar arguments one can prove that
$$
\displaystyle\lim_{x\to x_m^{+}} \dfrac{- f'(x)-2\sqrt{A}\sqrt{f(x)}}{4 f(x)}=-\dfrac{B}{2A}.
$$
From this last computation and formula \eqref{Deriv_phi} we deduce that $\p_x\phi(t,x)$ is bounded for every $(t,x)\in\mathbb{R}_+\times\mathbb{R}$ and
$$
\p_x\phi(0,x_m)=G(x_m)-\frac{\overline{a}_{xxx}(x_m)}{6\overline{a}_{xx}(x_m)}.
$$
}
\subparagraph{{Step 2:} $\vert \p_{xx} \phi\vert$ is bounded.} {Again in order to bound $\p_{xx} \phi(t,x)$ we can bound $\p_{xx} \phi(0,x)$ according to formula \eqref{Deriv_x_phi}. Note that far from $x_m$ this derivative exists and it is bounded because of the regularity of $a(\textcolor{black}{e(t)},x)$. To verify the boundedness near of $x_m$ we follow the same arguments as above for the first derivative, that is, we denote $f(x)=\bar a(x_m)-\bar a(x)$ as before and we compute
\begin{equation}
\begin{array}{rcl}
\label{limit_d2_phi}
\displaystyle\lim_{x\to x_m^{\mp}}\frac{\p_x\phi(0,x)-\p_x\phi(0,x_m)}{x-x_m}&=&\displaystyle\lim_{x\to x_m^{\mp}}\frac{G(x)-G(x_m)}{x-x_m} \\
&+& \displaystyle\lim_{x\to x_m^{\mp}} \frac{- f'(x)\mp 2\sqrt{A}\sqrt{f(x)}+\frac{2B}{A}f(x)}{4 f(x)(x-x_m)},
\end{array}
\end{equation}
if both limits in the RHS exist and are bounded. {Note that the first limit in the RHS of \eqref{limit_d2_phi} exist and is equal to $G'(x_m)$ because of the definition of $G$ in \eqref{G(x)} and the regularity of $a(\textcolor{black}{e(t)},x)$.}
Moreover, using the Taylor expansion for $f(x)$ around $x=x_m$ the terms in the numerator of \eqref{limit_d2_phi} can be developed as follows
$$\begin{array}{rcl}
f'(x)&=&2A(x-x_m) +3B( x-x_m)^2+O((x-x_m)^3),\\\\
2\sqrt{A}\sqrt{f(x)}&=& 2A\vert x-x_m\vert\sqrt{1+\frac{B}{A}(x-x_m)+O((x-x_m))^2}\\
&=&2A\vert x-x_m\vert\big(1+\frac{B}{2A}(x-x_m)+O((x-x_m)^2)\big),\\\\
\frac{2B}{A}f(x)&=&2B(x-x_m)^2+O((x-x_m)^3).
\end{array}
$$
We substitute into \eqref{limit_d2_phi} and it holds that the terms remaining in the numerator are of order $(x-x_m)^3$. Indeed,
$$
\begin{array}{l}
\displaystyle\lim_{x\to x_m^-} \frac{ - f'(x)- 2\sqrt{A}\sqrt{f(x)}+\frac{2B}{A}f(x)}{4 f(x)(x-x_m)}\\\\
=\displaystyle\lim_{x\to x_m^-} \left[\frac{-2A(x-x_m) -3B( x-x_m)^2+2A(x-x_m)(1+\frac{B}{2A}(x-x_m))}{4A(x-x_m)^3+4B(x-x_m)^4+O((x-x_m)^4)}\right.\\\\
\qquad \displaystyle\left.+\frac{2B(x-x_m)^2+\frac{2B^2}{A}(x-x_m)^3+O((x-x_m)^3)}{4A(x-x_m)^3+4B(x-x_m)^4+O((x-x_m)^4)}\right] \\\\
=\dfrac{B^2}{2A^2},
\end{array}
$$
and by an analogous procedure we can obtain
$$
\displaystyle\lim_{x\to x_m^+} \frac{ - f'(x)+ 2\sqrt{A}\sqrt{f(x)}+\frac{2B}{A}f(x)}{4 f(x)(x-x_m)}=\dfrac{B^2}{2A^2}.
$$
We then conclude that the second derivative of $\phi$ at point $(0,x_m)$ is bounded and
$$
\p_{xx}\phi(0,x_m)=G'(x_m)+\frac{\overline{a}_{xxx}^2(x_m)}{18\ \overline{a}_{xx}^2(x_m)}.
$$
}
\subparagraph{{Step 3:} $\lambda_2=\sqrt{-\bar a_{xx}(x_m)/2}$.}
We next evaluate the second equation in \eqref{System_approx} at $x=x_m$ to obtain $\lambda_2=\sqrt{-\bar a_{xx}(x_m)/2}$.
\end{proof}
\section{An illustrating biological example}
\label{Bio_example}
In this section we discuss the effect of the periodic fluctuations on the critical speed of survival and the {phenotypic distribution of the population} for the following particular growth rate
\begin{equation}
\label{a_no_clim}
\textcolor{black}{a(\textcolor{black}{e},x)=r-g(e)(x-\theta(e))^2,}
\end{equation}
\textcolor{black}{where $r$ is a positive constant corresponding to the maximal growth rate. The positive function $g$
represents {the pressure of selection} and the function $\theta$ represents the optimal trait, both being functions of the environmental state $e$. As above, we assume that $e(t):\mathbb{R}_+\to E$ is a periodic function with period $T=1$.}\\
\noindent
We compute the mean of $a(\textcolor{black}{e(t)},x)$
$$
\overline{a}(x)=\int_0^1a(\textcolor{black}{e(t)},x)dt=r-x^2\bar g+2xg_1-g_2,
$$
where
\begin{equation}
\label{g}
\bar{g}=\int_0^1g(\textcolor{black}{e(t)})dt,\quad g_1=\int_0^1g(\textcolor{black}{e(t)})\theta(\textcolor{black}{e(t)})dt,\quad g_2=\int_0^1g(\textcolor{black}{e(t)})\theta^2(\textcolor{black}{e(t)})dt,
\end{equation}
and we observe that the maximum of $\overline{a}(x)$ is attained at $x_m=\frac{g_1}{\bar{g}}$, with
$$
\bar a(x_m)=r+\frac{g_1^2}{\bar{g}}-g_2.
$$
\textcolor{black}{In what follows, we try to characterize the phenotypic density $n_\e$, the solution to \eqref{n_eps}.}
\textcolor{black}{From} {Theorem} \ref{Unique_Identif}-(ii) we obtain that $\psi(x)$ the solution of the Hamilton-Jacobi equation \eqref{LimitEq} attains its maximum at
$$
\bar x=x_m-\frac{c}{2\sqrt{\bar{g}}}=\frac{g_1}{\bar{g}}-\frac{c}{2\sqrt{\bar{g}}}.
$$
Let $\psi(x)$ be given by \eqref{Exp_Sol_psi}, then for this specific growth rate it can be written as follows
$$
\begin{array}{rcl}
\psi(x)&=&\displaystyle\frac{c}{2}\left(x_m-\frac{c}{2\sqrt{\bar g}}-x\right)+\int_{x_m-\frac{c}{2\sqrt{\bar g}}}^{x_m}\sqrt{\bar{g}(y-x_m)^2}dy-\left|\int_{x_m}^x\sqrt{\bar{g}(y-x_m)^2}dy\right|\\
&=&-\displaystyle\frac{\sqrt{\bar g}}{2}\left(x+\frac{c}{2\sqrt{\bar g}}-\frac{g_1}{\bar{g}}\right)^2\\
&=&-\displaystyle\frac{\sqrt{\bar g}}{2}\left(x-\bar{x}\right)^2.
\end{array}
$$
Moreover, the asymptotic expansions in Theorem \ref{Asymp_Expansion} {imply that}
{$$
\overline{\rho}_\e= r+\frac{g_1^2}{\bar{g}}-g_2-\dfrac{c^2}{4}-\e\sqrt{\bar{g}}+o(\e),\quad
c_\e^*= 2\sqrt{r+\dfrac{g_1^2}{\bar{g}}-g_2}-{\e\sqrt{\frac{\bar{g}}{r+\dfrac{g_1^2}{\bar{g}}-g_2}}}+o(\e).
$$}
Furthermore, following the arguments in \cite{Mirrahimi&Figueroa_pprint}-Section 5, we can also obtain an approximation of order $\e$ for the phenotypic mean $\mu_\e$ and the variance $\sigma^2_\e$ of the population's distribution, that is:
$$
\begin{array}{rcl}
\mu_\e(t) &=&\displaystyle\frac{1}{\rho_\e(t)}\int_{\mathbb{R}}x\ n_\e(t,x)dx =\frac{g_1}{\bar{g}}-\frac{c}{2\sqrt{\bar{g}}} +\e D(t)+o(\e),
\\\\
\sigma^2_\e &=&\displaystyle\frac{1}{\rho_\e(t)}\int_{\mathbb{R}}(x-\mu_\e)^2n_\e(t,x)dx = \frac{\e}{\sqrt{\bar{g}}}+o(\e),
\end{array}
$$
where $D(t)=\p_x \phi(\bar x,t)$ for $\phi$ the solution of the system \eqref{System_approx}. {We refer the readers to the Appendix B for more details on the derivation of the moments}.\\
One can verify that for this growth rate we have
$$
D(t)=-c\sqrt{\bar{g}}\left(t-\frac{1}{2}\right)+2\int_0^1\int_0^\tau g(\textcolor{black}{e(s)})(\bar{x}-\theta(\textcolor{black}{e(s)}))dsd\tau-2\int_0^tg(\textcolor{black}{e(s)})(\bar{x}-\theta(\textcolor{black}{e(s)}))ds.
$$
Note that the phenotypic mean is $1-$periodic since $D(0)=D(1)$. Moreover $\langle \mu_\e(t)\rangle=\frac{g_1}{\bar{g}}-\frac{c}{2\sqrt{\bar{g}}} {+o(\e)}$ since $\int_0^1D(t)dt=0$.\\\\
We are now interested in comparing these quantities with the case where there is no fluctuations. To do {so} we first consider a case where $\textcolor{black}{g(e)}=g>0$ is constant and then a case where {$\theta$} is constant.\\
\begin{itemize}
\item[Case 1.] $\textcolor{black}{g(e)}=g$ constant. Note that, in such a case $g_1=g\bar\theta$ and $g_2=g\int_0^1\theta^2(\textcolor{black}{e(t)})dt$ with $\bar \theta=\int_0^1\theta(\textcolor{black}{e(t)})dt$. We compute
\begin{equation}
\label{moments-thetap}
\begin{array}{l}
\overline{\rho}_{\e,g(e)=g}= r+g\left[\bar{\theta}^2-\int_0^1\theta^2(\textcolor{black}{e(t)})dt\right]-\dfrac{c^2}{4}-\e\sqrt{g}+o(\e), \\\\
\big\langle \mu_{\e,g(e)=g}(t)\big\rangle=\bar \theta-\frac{c}{2\sqrt{g}}+o(\e), \\\\
c_{\e,g(e)=g}^*= 2\sqrt{r+g\left[\bar{\theta}^2-\int_0^1\theta^2(e(t))dt\right]}-\e\sqrt{\dfrac{g}{r+g\left[\bar{\theta}^2-\int_0^1\theta^2(\textcolor{black}{e(t)})dt\right]}}+o(\e).
\end{array}
\end{equation}
We compare then, the sub-cases where \textcolor{black}{$e$} is constant or periodic.
\begin{itemize}
\item[a)] If $e(t)$ is a $1-$periodic function then, $\bar \theta^2<\int_0^1\theta^2(e(t))dt$ and we obtain
$$
\begin{array}{c}
\overline{\rho}_{\e,p}< r-\dfrac{c^2}{4}-\e\sqrt{g}+o(\e),\quad \big\langle \mu_{\e,p}(t)\big\rangle=\bar \theta-\dfrac{c}{2\sqrt{g}}+o(\e), \\
c_{\e,p}^*< 2\sqrt{r}-\e\sqrt{\dfrac g r}+o(\e).
\end{array}
$$
\item[b)] If $\textcolor{black}{e(t)\equiv \overline e}$ is constant, typically equal to the averaged state of the periodic environment above, (so that \textcolor{black}{$\theta\equiv\theta({\overline e})=:\theta_{\overline e}$}) we obtain in particular that $\bar \theta^2=\int_0^1\theta^2(\textcolor{black}{\overline e})dt$ and hence
$$
\begin{array}{c}
\overline{\rho}_{\e,c}= r-\dfrac{c^2}{4}-\e\sqrt{g} +o(\e),\quad \big\langle \mu_{\e,c}(t)\big\rangle=\theta_{\overline e}-\dfrac{c}{2\sqrt{g}}+o(\e),\\
c_{\e,c}^*= 2\sqrt{r}-\e\sqrt{\dfrac g r}+o(\e).
\end{array}
$$
\end{itemize}
Thus, by keeping the pressure of selection constant, we deduce that, {for $\e$ small,}
$$
\overline{\rho}_{\e,p}\leq \overline{\rho}_{\e,c}\quad \text{ and } \quad c_{\e,p}^*\leq c_{\e,c}^*.
$$
This means that having an oscillating optimal trait is not beneficial for the population, in the sense that the mean total size of the population decreases with respect to the case with a constant optimal trait and the critical speed which leads the population to extinct is smaller in the periodic case. Note also from \eqref{moments-thetap} that in the periodic case the mean population size is reduced by the product of the selection pressure and the variance of the optimal trait (that is $\int_0^1\theta^2(\textcolor{black}{e(t)})dt-\bar{\theta}^2$).\\
\item[Case 2.] $\theta(\textcolor{black}{e(t)})=\theta$ constant. Note that, in such a case $g_1=\bar g\theta$ and $g_2=\bar g\theta^2$. We compute
$$
\begin{array}{c}
\overline{\rho}_{\e,\theta(e(t))=\theta}= r-\dfrac{c^2}{4}-\e\sqrt{\bar g} +o(\e),
\quad \big\langle \mu_{\e,\theta(e(t))=\theta}(t)\big\rangle= \theta-\dfrac{c}{2\sqrt{\bar g}} +o(\e),\\
c_{\e,\theta(e(t))=\theta}^*= 2\sqrt{r}-\e\sqrt{\dfrac{\bar g}{r}} +o(\e).
\end{array}
$$
We compare then, the sub-cases where \textcolor{black}{$e$} is constant or periodic.
\begin{itemize}
\item[a)] If $\textcolor{black}{e(t)}$ is a $1-$periodic function then we obtain
$$
\begin{array}{c}
\overline{\rho}_{\e,p}= r-\dfrac{c^2}{4}-\e\sqrt{\bar g} +o(\e),\quad \big\langle \mu_{\e,p}(t)\big\rangle= \theta-\dfrac{c}{2\sqrt{\bar g}}+o(\e), \\
\displaystyle c_{\e,p}^*= 2\sqrt{r}-\e\sqrt{\dfrac{\bar g}{r}}+o(\e).
\end{array}
$$
\item[b)] If $\textcolor{black}{e(t)\equiv \overline e}$ is constant, typically equal to the averaged state of the periodic environment above, (\textcolor{black}{so that $g\equiv g({\overline e})=:g_{\overline e}$}), we obtain
$$
\begin{array}{c}
\overline{\rho}_{\e,c}= r-\dfrac{c^2}{4}-\e\sqrt{g_{\overline e}} +o(\e),\quad \big\langle \mu_{\e,c}(t)\big\rangle=\theta-\dfrac{c}{2\sqrt{g_{\overline e}}}+o(\e),\\
c_{\e,c}^*= \displaystyle 2\sqrt{r}-\e\sqrt{\f{g_{\overline e}}{r}}+o(\e).
\end{array}
$$
\end{itemize}
If we choose an oscillating selection pressure function $\textcolor{black}{g(e)}$ which satisfies:
\begin{equation}
\label{Cond_g}
\bar g < \textcolor{black}{g(\overline e)},
\end{equation}
which holds for instance if $g(\cdot)$ is a concave function, then we obtain that
$$
\overline{\rho}_{\e,c} < \overline{\rho}_{\e,p}\quad \text{ and } \quad c_{\e,c}^* < c_{\e,p}^*.
$$
This means that the mean total size of the population increases with respect to the case with a constant \textcolor{black}{environmental state}. Moreover, the critical speed above which the population goes extinct is larger in the periodic case. This means that the periodic fluctuations can help the population to follow the environment change. \\
Note that the condition \eqref{Cond_g} imposed to \textcolor{black}{$g(e(t))$} is the opposite to the one imposed in \cite{Mirrahimi&Figueroa_pprint} (equation 51 of Section 6.2), leading to more performant populations. There, it was proved that in presence of the mutations and while the fluctuations act on the pressure of the selection (that is with a similar growth rate, however with $c=0$ and under the condition $\bar g>g(\textcolor{black}{\overline e })$), a fluctuating environment can select for a population with smaller variance and in this way lead to more performant populations. What is beneficial in a (in average) constant environment may indeed be disadvantageous in a changing environment. \\
Note also that in the present example under condition \eqref{Cond_g}, we have
$$
\Big\vert\big\langle \mu_{\e,p}(t)\big\rangle-\theta\Big\vert > \Big\vert \big\langle \mu_{\e,c}(t)\big\rangle-\theta\Big\vert.
$$
This means that even if the population can follow the \textcolor{black}{environmental} change in a better way by considering a fluctuating environment, this population is \textcolor{black}{less adapted}.
\end{itemize}
|
\section{Introduction}
The continuum Gaussian free field (GFF) is a generalisation of Brownian motion to higher dimensions, which is defined on balls as follows.
\begin{definition}[Gaussian free field]
Let $d \geq 1$ and $\mathbb{B} \subseteq \mathbb{R}^d$ denote the open unit ball.
Then for $a\in \mathbb{R}^d$ and $r>0$ the $d$-dimensional zero boundary continuum GFF in $a+r\mathbb{B}$ is the centred Gaussian process $(h^{a+r\mathbb{B}},f)_{f\in C_c^\infty(\mathbb{R}^d)}$ whose covariance is given by
$$\mathbb{E}((h^{a+r\mathbb{B}},f)(h^{a+r\mathbb{B}},g)) = \int\int_{(\mathbb{R}^d)^2} f(z)G^{a+r\mathbb{B}}(z,w)g(w)\, dz dw;\quad f,g\in C_c^\infty(\mathbb{R}^d),$$
where $G^{a+r\mathbb{B}}$ denotes the zero boundary Green's function for the Laplacian in $a+r\mathbb{B}$ .
\end{definition}
In dimension $d = 1$, this corresponds to a slightly awkward definition of the Brownian bridge on intervals {$(a-r,a+r)$}: in this case the GFF can actually be defined as an a.s.\ continuous function, i.e. indexed instead by points. In $d \geq 2$ the GFF does not make sense as a pointwise defined function, thus the more general definition. However, one can still heuristically see the continuum GFF as a Gaussian height function parameterised by a domain $D \subseteq \mathbb{R}^d$ \cite{SheffieldGFF}. The Gaussian free field has played many roles in probability and mathematical physics since the 1970s: it is the stationary solution of the stochastic heat equation; is used to describe both free and interactive Euclidean Quantum field theories; it appears as a corrector in stochastic homogenisation; to name just a few. Recently, the $2d$ GFF has played a crucial role in the probabilistic study of statistical physics models, Schramm-Loewner evolution, Liouville quantum gravity and Liouville field theory \cite{Dub, MS3, DS11, DKRV}. It is both the proven and the conjectured scaling limit of several natural discrete height functions \cite{NS, Kenyon_GFF, RiderVirag}
There are many characterisations of Brownian motion that single it out among $1d$ stochastic processes and it is similarly natural to also ask how the GFF in higher dimensions can be characterised. Recently, the $2d$ continuum GFF was characterised as the only conformally invariant field, satisfying a domain Markov property and certain minimal moment and continuity assumptions \cite{BPR18,BPR20}.
This short note provides a first characterisation of the $d-$dimensional continuum Gaussian free field for $d \geq 3$, answering Question 6.1 in \cite{BPR18}. It also provides a new characterisation of the GFF in $d= 2$ that relaxes the assumption of conformal invariance and {thereby generalises the main result of \cite{BPR18} via a new proof}. Our characterisation is based only on {scaling} and the domain Markov property and in particular, we do not need to assume rotational invariance.
More precisely, but still somewhat informally, we prove that, under certain mild moment and continuity assumptions, the $d$-dimensional Gaussian free field is the only stochastic process in $d \geq 2$ that is translation invariant, exhibits a certain scaling, and satisfies the following domain Markov property - for each ball, the field inside can be written as a sum of an independent scaled copy of the original field plus the ``harmonic extension" of its behavior on the boundary. We prove the theorem by identifying the covariance structure as the Green's kernel (lighter step), and then proving Gaussianity (a bit more involved). The two arguments are quite independent of each other (although we use the knowledge of the covariance kernel to simplify some arguments later on), but both make strong use of the domain Markov property. {In particular, our proof of Gaussianity via spherical harmonics stresses the nice interplay between the GFF and harmonic functions, and could be of independent interest.}
This characterisation could potentially be helpful when identifying the GFF as a continuum scaling limit, since proving conformal invariance or even rotational invariance from the lattice is known to be very difficult.\footnote{In fact, let us thank here H. Duminil-Copin and V. Tassion for asking us whether only rotational invariance should suffice for the characterisation of the 2$d$ continuum GFF. This discussion happened in a nice and friendly workshop in Fribourg, organised by I. Manolescu, who we would hereby like to thank too. We would finally like to thank N. Berestycki, J.C. Mourrat and G. Ray for many interesting and fruitful discussions on this topic.} However, the key step in this case would be proving this concrete form of domain Markov property in the limit, which admittedly is probably not easy at all - it would in some sense already mean a certain link of the covariance kernel to random walks and Brownian motion.
\begin{theorem}[Characterisation of the $d$-dimensional GFF]\label{thm_main}
Let $d \geq 2$ and $\mathbb{B} \subseteq \mathbb{R}^d$ denote the open unit ball. Suppose that for every $a\in \mathbb{R}^d, r>0$ $(h^{a+r\mathbb{B}},f)_{f\in C_c^\infty(\mathbb{R}^d)}$ is a centered stochastic process indexed by
$C_c^\infty(\mathbb{R}^d)$.
If
the collection $\{h^{a+r\mathbb{B}}\; ;\; a\in \mathbb{R}^d, r>0\}$ satisfies the following conditions, then for some $c$ independent of $a$ and $r$, each $h^{a+r\mathbb{B}}$ has the law of $c$ times a $d-$dimensional zero boundary Gaussian free field in $a+r\mathbb{B}$.
\begin{enumerate}
\item\label{lin} \textbf{Linearity} $(h^\mathbb{B},f)$ is a.s.\ linear in $f$.
\item\label{sup} \textbf{Zero outside of $\mathbb{B}$} For any $f$ with $\mathrm{supp}(f)\subset \mathbb{R}^d\setminus \mathbb{B}$ $(h^\mathbb{B},f)=0$ a.s.
\item\label{inv} \textbf{Translation and scaling} The law of $h^{a+r\mathbb{B}}$ is the rescaled image of $h^\mathbb{B}$ under translation by $a$ and scaling by $r$: $$(h^{a+r\mathbb{B}},f(r\cdot+a))=r^{1+\frac{d}{2}}(h^\mathbb{B},f(\cdot)).$$ Note that if $h$ were a function and $(
\cdot,
\cdot)$ was the $L^2$ inner product, this would be equivalent to $h^{a+r\mathbb{B}}(a+r\cdot)=r^{\frac{2-d}{2}}h^\mathbb{B}(\cdot)$, exhibiting half the scaling factor of $G^\mathbb{B}$.
\item\label{dmp} \textbf{Domain Markov property} Suppose that $a+r\mathbb{B} \subseteq \mathbb{B}$. Then
we can write $$(h^\mathbb{B},f)_{f\in C_c^\infty(\mathbb{R}^d)}=(h^{a+r\mathbb{B}}_\mathbb{B},f)_{f\in C_c^\infty(\mathbb{R}^d)}+(\varphi^{a+r\mathbb{B}}_\mathbb{B},f)_{f\in C_c^\infty(\mathbb{R}^d)}$$ where the two summands are independent,
$h^{a+r\mathbb{B}}$ is equal in law to $h^{a+r\mathbb{B}}$, and $\varphi_\mathbb{B}^{a+r\mathbb{B}}$ is a stochastic process that a.s.\ corresponds to integrating against a harmonic function when restricted to $a+r\mathbb{B}$.
\item\label{mom} \textbf{Existence of fourth moments} We have $\mathbb{E}((h^\mathbb{B},f)^4)<\infty$ for all $f\in
C_c^\infty(\mathbb{R}^d)$.
\item\label{K2c} \textbf{Continuity of the covariance functional} The covariance $K_2(f,g):=\mathbb{E}((h^\mathbb{B},f)(h^\mathbb{B},g))$, is a continuous bilinear form on $C_c^\infty(D)$.
\item\label{zbc} \textbf{Zero boundary conditions} For any sequence $(f_n)_{n\ge 0}$ of smooth positive functions with $\int f_n$ uniformly bounded, $d(\mathrm{supp}(f_n),0)\to 1$ as $n\to \infty$ and $\sup_n \sup_{r<1}\sup_{x,y\in \partial(r\mathbb{B})} |f_n(x)/f_n(y)|<\infty$, we have that $K_2(f_n,f_n)\to 0$
\end{enumerate}
\end{theorem}
We have not striven for most general technical assumptions, but rather have tried to keep the proofs light and self-contained. However, we believe that several assumptions can most likely be relaxed.
\begin{itemize}
\item First, the moment assumption can be probably relaxed using methods of \cite{BPR20}.
\item Also, although using spherical harmonics is key in our proofs, one could make this work for other reasonable domains, e.g. with smooth boundary. This follows rather directly from our techniques if the domain Markov property is assumed for both the domain shape and balls, but probably variations of this are also possible.
\item We use the same form of domain Markov property (DMP) as in \cite{BPR18}, but only for balls. The DMP plays a key role in our argument - it is used both in showing that the covariance kernel is harmonic off the diagonal, and in proving Gaussianity. It seems difficult to relax the harmonicity condition for the boundary data extension in the current approach, but it would be very interesting to determine if this is possible. On the other hand, as in the case of Brownian motion, one could easily envisage replacing the condition of having an exact (scaled) copy of the field with a more relaxed, martingale type of condition.
\item It might seem a bit surprising that we are not using rotational invariance, however, the reason is that the DMP already implies rather easily that the covariance is a harmonic function and this is a very strong property. The exact scaling could similarly be relaxed, at least to some extent.
\item Finally, let us comment on the exact form of the zero boundary condition. In \cite{BPR18} this convergence was only asked for rotationally symmetric functions. We need the slightly generalised form (that is implied for example by conformal invariance and the rotationally invariant form) basically in two places: to determine that the covariance is the Green's kernel and to obtain the uniqueness of the domain Markov property. Notice that our condition is still less restricting than asking zero boundary conditions locally.
\end{itemize}
We will now present the argument in three sections. First, we discuss some immediate consequences of the assumptions (along similar lines to \cite{BPR18}, so we will keep this brief). In Section \ref{sec:cov}, using the DMP, scaling and translation invariance, we show how to deduce that the covariance kernel is the Green's kernel (\cref{lem:cov}). Finally, the main part of the paper is Section \ref{sec:gauss}, where we prove Gaussianity (\cref{prop:gauss}) - we do this using solely the DMP, and a decomposition of the underlying functional space using spherical harmonics. Theorem \ref{thm_main} is an immediate consequence of \cref{lem:cov} and \cref{prop:gauss}. Sections \ref{sec:cov} and \ref{sec:gauss} can be read quite independently of each other, although we use the identification of the covariance kernel to simplify some arguments in the latter.
\begin{definition}[Scaling function]\label{def:s}
In what follows we write $s(r)$ for the function on $(0,\infty)$ defined by $-\log r$ when $d=2$ and $r^{2-d}$ when $d\ge 3$.
\end{definition}
\section{Immediate consequences}\label{sec:immediate_consequences}
{Here we discuss some immediate properties of our assumptions. The section is self-contained, but we remain brief, as similar properties have been shown in detail in the 2$d$ case in \cite{BPR18}.}
\paragraph{The domain Markov decomposition is unique.} Indeed, suppose that for some $a,r$ we had two decompositions as in \ref{dmp} of \cref{thm_main}: $$h^\mathbb{B}=h_\mathbb{B}^{a+r\mathbb{B}}+\varphi_\mathbb{B}^{a+r\mathbb{B}}=\tilde{h}_\mathbb{B}^{a+r\mathbb{B}}+\tilde{\varphi}_\mathbb{B}^{a+r\mathbb{B}}.$$
Then for any $z\in a+r\mathbb{B}$, by harmonicity, there exists a sequence $(f_n)_{n\ge 0}$ of functions satisfying the zero boundary condition assumption \ref{zbc} from \cref{thm_main}, and such that $(\varphi,f_n(a+r\cdot))=\varphi(z)$ for all $n$ and any harmonic $\varphi$ in $a+r\mathbb{B}$. The zero boundary condition then implies that $(h_\mathbb{B}^{a+r\mathbb{B}}-\tilde{h}_\mathbb{B}^{a+r\mathbb{B}},f)\to 0$ a.s.\ along a subsequence as $n\to \infty$. Thus, it must be the case that $\varphi_\mathbb{B}^{a+r\mathbb{B}}(z)=\tilde{\varphi}_\mathbb{B}^{a+r\mathbb{B}}(z)$ a.s. Applying this for a dense collection of $z$ and using harmonicity of $\varphi_\mathbb{B}^{a+r\mathbb{B}},\tilde{\varphi}_\mathbb{B}^{a+r\mathbb{B}}$ proves the uniqueness.
We will use the following consequences of this uniqueness repeatedly:
\begin{itemize}
\item if $a+r\mathbb{B} \subset a'+r'\mathbb{B}\subset \mathbb{B}$ then \begin{equation}\label{eq:it_dmp}\varphi_\mathbb{B}^{a+r\mathbb{B}}-\varphi_\mathbb{B}^{a'+r'\mathbb{B}} \text{ is independent of } \varphi_{\mathbb{B}}^{a'+r'\mathbb{B}} \text{ and equal in law to } (r/r')^{\frac{2-d}{2}}\varphi_\mathbb{B}^{(a-a')+(r/r')\mathbb{B}};\end{equation}
\item
we can apply the domain Markov property in several balls at once. More precisely, if $B_1,\cdots, B_n$ are $n$ balls, and $h^\mathbb{B}=h^i+\varphi^i$ in each ball, then \begin{equation}
\label{eq:multidmp}
\varphi:=h^\mathbb{B} - \sum_{i=1}^n h^i \text{ is a.s.\ harmonic in } \cup_{i}B_i
\end{equation} and $h^i$ is independent of $\{\varphi,(h^j)_{j\ne i}\}$ for each $i$. Indeed, for any fixed $i$ we can write $\varphi=\varphi^i-\sum_{j\ne i}h^j$ and since $h^j$ is zero in $B_i$ for all $j\ne i$, $\varphi=\varphi^i$ is harmonic in $B_i$. But also for every $j\ne i$, $h^i$ is zero in $B_j$ and therefore $h^j$ is measurable with respect to $\varphi^i$. So the collection $\{\varphi,(h^j)_{j\ne i}\}=\{\varphi^i-\sum_{j\ne i} h^j,(h^j)_{j\ne i}\}$ is $\varphi^i$-measurable and therefore independent of $h^i$.
\end{itemize}
\paragraph{The 2-point function.} For $z_1,z_2\in (1-\varepsilon)\mathbb{B}$ with $|z_1-z_2|>2\varepsilon$, define the harmonic functions $\varphi_\mathbb{B}^{z_1+\varepsilon\mathbb{B}}, \varphi_\mathbb{B}^{z_2+\varepsilon\mathbb{B}}$ according to the domain Markov decomposition of $h^\mathbb{B}$ in $z_1+\varepsilon\mathbb{B},z_2+\varepsilon\mathbb{B}$ respectively. Then by \eqref{eq:it_dmp} and \eqref{eq:multidmp}, the quantity
\[k_2(z_1,z_2)=\mathbb{E}(\varphi_\mathbb{B}^{z_1+\varepsilon\mathbb{B}}(z_1)\varphi_\mathbb{B}^{z_2+\varepsilon\mathbb{B}}(z_2))\] is well-defined, i.e., it does not depend on $\varepsilon > 0$ satisfying the above conditions. Note that, by harmonicity, we can alternatively write \begin{equation}\label{k:conv}k(z_1,z_2)=\mathbb{E}((h^\mathbb{B},\eta_{z_1})(h^\mathbb{B},\eta_{z_2}))\end{equation} for any smooth functions { $\eta_{z_1}=\eta_{z_1}^\varepsilon,\eta_{z_2}=\eta_{z_2}^\varepsilon$ }of mass one, that are supported in $z_1+\varepsilon\mathbb{B},z_2+\varepsilon\mathbb{B}$ and rotationally symmetric about $z_1,z_2$ respectively.
For every $\delta > 0$, the following crude upper bound for $z_1\ne z_2\in (1-\delta)\mathbb{B}$ is rather direct with $C = C(\delta)$:
\begin{equation}\label{eq:var_bounds} |k_2(z_1,z_2)|\le C(1+s(|z_1-z_2|)). \end{equation}
To justify this, observe that by Cauchy--Schwarz and the domain Markov property, it suffices to show that for any $z\in (1-\delta)\mathbb{B}$ and $\varepsilon < \delta$
\begin{equation}\label{eq:circlebound} \mathbb{E}(\varphi_\mathbb{B}^{z+\varepsilon\mathbb{B}}(z)^2)\le C(1+s(\varepsilon)).\end{equation}
Indeed, as long as $d(z_i,\partial \mathbb{B})>|z_1-z_2|/2$ for $i=1,2$ we can set $\varepsilon=|z_1-z_2|/2$ in the definition of $k_2(z_1,z_2)$; otherwise we crudely set $\varepsilon = \delta/2$ (leading to the dependence of $C$ on $\delta$).
Let us first consider the case $z=0$. If $2^{-n}\le \varepsilon\le 2^{-(n-1)}$, then the domain Markov property implies that $\mathbb{E}(\varphi_\mathbb{B}^{\varepsilon\mathbb{B}}(0)^2)\le \mathbb{E}(\varphi_\mathbb{B}^{2^{-n}\mathbb{B}}(0)^2)$. Further, by the domain Markov property again and also scaling, the right hand side can be written as the sum $\sum_{m=0}^{n-1} 2^{m(d-2)/2} X_m$, with the $X_m$ i.i.d.\ each having the law of $\varphi_{\mathbb{B}}^{0.5\mathbb{B}}(0)$. Adding up the variances gives the desired bound.
When $z\ne 0$ with $|z|=r<1$, let $r' = 1/(1+r)\in (1/2,1)$, $\varepsilon' = \varepsilon/(1+r)$ and $z'=-r'z$.
By applying the Markov property for $h_\mathbb{B}^{z'+r'\mathbb{B}}$ in $\varepsilon'\mathbb{B}$ we can write $$\varphi_\mathbb{B}^{\varepsilon'\mathbb{B}}(0) = \varphi_\mathbb{B}^{z' +r'\mathbb{B}}(0) + \tilde{\varphi}(0)$$
where the summands are independent and by translation invariance and scaling, $\tilde{\varphi}(0)$ has the law of $(r')^{2-d}$ times $\varphi_\mathbb{B}^{z+\varepsilon\mathbb{B}}(z)$. But since the variance of the summands add up, the variance of $\tilde{\varphi}(0)$ is no greater than the variance of $\varphi_\mathbb{B}^{\varepsilon'\mathbb{B}}(0)$. Hence the claim follows from the case $z = 0$.
\medskip
From here, the assumption of continuity for $K_2$ shows that $k_2$ is the covariance function of the field, in the sense that for any $f_1,f_2 \in C_c^\infty(\mathbb{B})$, we have
\begin{equation}\label{Kk}K_2(f_1,f_2)=\iint_{\mathbb{B}^2} f_1(z_1)f_2(z_2)k_2(z_1,z_2) \, dz_1 dz_2,\end{equation}
where the right-hand side is well defined as the limit
\begin{equation}\label{eq:Kklimit}\lim_{a\to 0}\iint_{\mathbb{B}^2\setminus\{|z_1-z_2|\le a\}} f_1(z_1)f_2(z_2)k_2(z_1,z_2) \, dz_1 dz_2\end{equation}
(see just below for a proof). In other words, we show that there is no ``diagonal contribution'' of $K_2$, and so it does not matter than $k_2$ is only defined off the diagonal. This already rules out, for example, the possibility that the field is a white noise or one of its derivatives.\footnote{We would like to thank A. Sep\'{u}lveda for raising the need for extra clarity on this point.}
\paragraph{The diagonal contribution of the 2-point function is negligible.}
Let us now justify the above statements. {For each $0<\varepsilon<d(z_1,\partial \mathbb{B})\wedge d(z_2,\partial \mathbb{B})$, let $\eta_{z_1}^\varepsilon,\eta_{z_2}^\varepsilon$ be unit-mass radially symmetric mollifiers in $\varepsilon$ balls around $z_1,z_2$, as in \eqref{k:conv}.}
Then if $f_1,f_2\in C_c^\infty(\mathbb{B})$,
the functions $z\mapsto \int_{\mathbb{B}} f_i(z_i)\eta_{z_i}^\varepsilon(z) \, dz_i$ converge to $f_i$ in $C_c^\infty(\mathbb{B})$ for $i=1,2$, and since $K_2$ is assumed to be a continuous bilinear form on $C_c^\infty(\mathbb{B})$, this means that that $$K_2(f_1,f_2)=\lim_{\varepsilon\to 0} \iint_{\mathbb{B}^2} f_1(z_1)f_2(z_2) K_2(\eta_{z_1}^\varepsilon,\eta_{z_2}^\varepsilon) \, dz_1 dz_2.$$
Now, we can break up the right-hand side of the above as
$$\lim_{\varepsilon\to 0} \left(\iint_{|z_1-z_2|>2\varepsilon}f_1(z_1)f_2(z_2)K_2(\eta_{z_1}^\varepsilon, \eta_{z_2}^\varepsilon)\, dz_1 dz_2 + \iint_{|z_1-z_2|\le 2\varepsilon}f_1(z_1)f_2(z_2)K_2(\eta_{z_1}^\varepsilon, \eta_{z_2}^\varepsilon) dz_1 dz_2\right)$$
where by
\eqref{k:conv}, we have $$\iint_{|z_1-z_2|>2\varepsilon}f_1(z_1)f_2(z_2)K_2(\eta_{z_1}^\varepsilon, \eta_{z_2}^\varepsilon)\, dz_1 dz_2 = \iint_{|z_1-z_2|>2\varepsilon}f_1(z_1)f_2(z_2)k_2(z_1,z_2)\, dz_1 dz_2$$ for each $\varepsilon.$ On the other hand, $K_2(\eta_{z_1}^\varepsilon, \eta_{z_2}^\varepsilon)^2\le {\mathbb{E}((h^\mathbb{B},\eta_{z_1}^\varepsilon)^2)\mathbb{E}((h^\mathbb{B},\eta_{z_1}^\varepsilon)^2)}$ for each $\varepsilon$ by Cauchy-Schwarz, and by \eqref{eq:circlebound}, the domain Markov property and scaling, we can bound $\mathbb{E}((h^\mathbb{B},\eta_{z}^\varepsilon)^2)=\mathbb{E}(\varphi_\mathbb{B}^{2\varepsilon\mathbb{B}}(z)^2)+\mathbb{E}((h_\mathbb{B}^{2\varepsilon\mathbb{B}},\eta_z^\varepsilon)^2)\le C(1+s(\varepsilon))$ for some constant $C$. Note that this constant $C$ does not depend on $\varepsilon$, and the bound holds for all $z$ in the compact supports of $f_1$ and $f_2$ simultaneously. This implies that $$\limsup_{\varepsilon\to 0} \iint_{|z_1-z_2|\le 2\varepsilon}f_1(z_1)f_2(z_2)K_2(\eta_{z_1}^\varepsilon, \eta_{z_2}^\varepsilon) dz_1 dz_2=0$$ and hence that $$\lim_{\varepsilon\to 0} \iint_{|z_1-z_2|>2\varepsilon}f_1(z_1)f_2(z_2)k_2(z_1,z_2)\, dz_1 dz_2 \text{ exists and is equal to } K_2(f_1,f_2)$$ as required. A similar argument appears in \cite[Proof of Lemma 2.18]{BPR18}.
\paragraph{The 4-point function.}
Similarly, due to \eqref{eq:multidmp}, for $z_1,\dots, z_4\in \mathbb{B}$ with $\min_{j\ne i}{|z_i-z_j|}>2a_i$ and $d(z_i,\partial \mathbb{B})>a_i$ for each $i$, $k_4(z_1,\dots, z_4)=\mathbb{E}(\prod \varphi_\mathbb{B}^{z_i+a_i\mathbb{B}}(z_i))$ is well-defined, i.e., it does not depend on the choice of $a_i$ satisfying the above conditions. Again we can write this as $\mathbb{E}(\prod_{i=1}^4(h^\mathbb{B},\eta_{z_i}))$ for any smooth functions $(\eta_{z_i})_{1\le i \le 4}$, with $\eta_{z_i}$ having mass one, being rotationally symmetric about $z_i$ and supported in $z_i+a_i\mathbb{B}$ for $i=1,2,3,4$.
Using the same argument as in the 2-point case, one can now bound $|k_4(z_1, \dots, z_4)|$. Indeed, by Holder's inequality it suffices this time to obtain bounds on the fourth moments of $\varphi_\mathbb{B}^{z+\varepsilon\mathbb{B}}(z)$ for $\varepsilon>0$. This can be done very similarly to the 2-point bound. One first treats $z = 0$ using the i.i.d nature of increments - the only change here is that when expanding $\mathbb{E}((\sum_{m=0}^{n-1} \alpha_m X_m)^4)$ with independent $X_m$s one must consider both $\sum \alpha_m^4 \mathbb{E}(X_m^4)$ and $\sum_{m\ne l} \alpha_m^2\alpha_l^2 \mathbb{E}(X_m^2)\mathbb{E}(X_l^2)$ - and then transports this bound to general $z$ using translation invariance and scaling. We omit the details and give the final bound: there are constants $C(\delta)>0$ such that for any distinct $z_1,z_2,z_3,z_4\in (1-\delta)\mathbb{B}$ \begin{equation}\label{eq:moment_bounds} |k_4(z_1,\dots, z_4)|^4\leq C(\delta)\prod_{i=1}^4 (1+\max_{j\ne i} s(|z_i-z_j|)^2).\end{equation}
\paragraph{Spherical averages.}
For $z\in \mathbb{B}$ and $\varepsilon<d(z,\partial \mathbb{B})$ we define $h_\varepsilon(z)=\varphi_\mathbb{B}^{z+\varepsilon\mathbb{B}}(z)$ to be the $\varepsilon$-\emph{spherical average} of $h^\mathbb{B}$ around $z$.
{
The spherical average admits the following natural approximations. Write $\rho_n^{\varepsilon}$ for a sequence of smooth test functions with total mass one that are rotationally symmetric about $z$ and supported in the annular region $z+\{(1-2^{-2n})\varepsilon\mathbb{B}\setminus (1-2^{-n})\varepsilon\mathbb{B}\}$ for each $n$. Then $\mathbb{E}((h^\mathbb{B},\rho_n^{\varepsilon})^2)=\mathbb{E}(h_\varepsilon(z)^2)+\mathbb{E}((h_\mathbb{B}^{z+\varepsilon\mathbb{B}},\rho_n^{\varepsilon})^2)$ for each $n$ by the Markov property, where the second term on the right-hand side goes to $0$ as $n\to \infty$ by the zero boundary condition assumption. It therefore follows that $$\mathbb{E}(h_\varepsilon(z)^2)=\lim_{n\to \infty} \mathbb{E}((h^\mathbb{B},\rho_n^{\varepsilon})^2)$$
and moreover this convergence is uniform for, say, $\varepsilon \in (\delta,1)$.
Using the same construction as above with $z=0$ and $\varepsilon=r_n$ converging to $1$ as $n\to\infty$, we also see that $\lim_{n\to \infty} \mathbb{E}(h_{r_n}(0)^2)=\lim_{n\to \infty} \mathbb{E}(h^\mathbb{B},\rho_n^{r_n})^2)$, and this final limit is $0$, again by the zero boundary condition. Since $\mathbb{E}(h_r(0)^2)$ is decreasing as $r\uparrow 1$, this implies that}
\begin{equation}\label{eq:circav0bc}
\mathbb{E}(h_r(0)^2)\searrow 0 \text{ as } r\uparrow 1.
\end{equation}
\section{Covariance is the Green function}\label{sec:cov}
Let us start by showing that scaling and translation invariance together with the domain Markov property already imply that the covariance kernel is the Green's function:
\begin{prop}\label{lem:cov}
The function $k_2(x,y)$ (defined for $x\ne y)$ is a positive multiple of the zero boundary Green's function. In particular, by \eqref{Kk}, this implies that the bilinear form $K_2$ is a multiple of the map $(f_1,f_2)\mapsto \iint_{\mathbb{B}^2} G^\mathbb{B}(z_1,z_2)f_1(z_1)f_2(z_2) \, dz_1 dz_2$.
\end{prop}
\begin{proof}
Write $G^\mathbb{B}$ for the zero boundary Green's function in $\mathbb{B}$.
We are going to use the following characterisation of $G^\mathbb{B}$ (see for example \cite[Lemma 3.7]{WPgff}).
\begin{itemize}
\item[$\star$] Suppose that for $y\in \mathbb{B}$, $k_y(x)$ is a harmonic function defined in in $\mathbb{B}\setminus \{y\}$, such that $k_y(x)-bs(|x-y|)$ is bounded in a neighbourhood of $y$ for some $b > 0$ and such that $(k_y,f_n
\to 0$ as $n\to \infty$ for any sequence of functions $f_n$ as in our zero boundary condition \ref{zbc} of \cref{thm_main}.\footnote{Note that for any fixed $y$, the support of $f_n$ will not intersect $y$ and so $(k_y,f_n)$ makes perfect sense as the integral of $k_y$ against $f_n$ in $\mathbb{B}$.}\footnote{The proof in \cite{WPgff} works exactly the same if we use this ``zero boundary condition'' for $k_y$ rather than a pointwise zero boundary condition.} Then $k_y(x)=bG^\mathbb{B}(x,y)$ for all $x\ne y$; $x,y\in \mathbb{B}$.
\end{itemize}
Let us first see that for fixed $y\in \mathbb{B}$ the function $k_y(\cdot):=k_2(y,\cdot)$ is harmonic away from $\{y\}$. Indeed, by the harmonicity of the function appearing in the domain Markov decomposition it is easy to see that for any $\eta>0$ and $x\in \mathbb{B}$ such that $|x-y|\wedge d(x,\partial \mathbb{B})>\eta$:
\begin{equation} \label{eqn:mvpk2} k_2(x,y)=\int k_2(w,y) \rho_{x}^\eta(dw)\end{equation}
where $\rho_{x}^\eta$ is uniform measure $\partial(x+\eta\mathbb{B})$. Note that the support of $\rho_x^\eta$ does not include $y$ by assumption, so the integral against $k_2(\cdot, y)$ is perfectly well defined. The proof of this is verbatim that given in \cite[Lemma 2.9]{BPR18}, which applies directly to all $d\ge 2$. This implies that $k_y$ is indeed harmonic in $\mathbb{B}\setminus \{y\}$. Note that in particular we have continuity in this region, which we did not assume a priori.
We now check the boundary condition for $k_y$: if $f_n$ are a sequence of functions as in \ref{zbc} of \cref{thm_main}, we have $(k_y,f_n)=\int k_2(y,x)f_n(x) \, dx=\int k_2(w,x) f_n(x) \, dx \rho_y^\delta(dw)$ for some $\delta(y)>0$ and all $n$ large enough. By Cauchy--Schwarz we see that the right-hand side is bounded by the square root of $\mathbb{E}(h_\delta(y)^2)\mathbb{E}((h^\mathbb{B},f_n)^2)$ which converges to $0$ by the zero boundary condition \ref{zbc}.
Now, notice that by \eqref{eq:var_bounds}, for any $\delta > 0$ there are some constants $c_1, c_2 > 0$ such that $w\mapsto k_2(0,w) + c_1 s(|w|)+c_2$ is a positive harmonic function in $(1-\delta)\mathbb{B} \setminus \{0\}$. Thus by B\^ocher's theorem \cite[Chapter III]{axlerharm}, we conclude that {there is some harmonic function $\nu(w): (1-\delta)\mathbb{B} \to \mathbb{R}$ such that in $(1-\delta)\mathbb{B} \setminus \{0\}$ we can write $k_2(0,w) + c_1 s(|w|)+c_2 = c_3 s(|w|) + \nu(w)$.} In particular, there is some $b > 0$ such that $k_2(0,w) - bs(|w|)$ is harmonic and bounded in $(1-2\delta)\mathbb{B}\setminus\{0\}$ and thus can be extended to a harmonic function on $(1-2\delta)\mathbb{B}$. Note that $b$ must be positive, since $\int k_2(0,z)\rho_0^{|w|}(dz)=\mathbb{E}(h_{|w|}(0)^2)$ is positive and increasing to $\infty$ as $w \downarrow 0$ by the domain Markov property.
From here, we return to a fixed $y\in \mathbb{B}$. We choose $\varepsilon>0$ such that $d(y,\partial \mathbb{B})>2\varepsilon$ and write $h^\mathbb{B}|_{y+2\varepsilon\mathbb{B}}=h_\mathbb{B}^{y+2\varepsilon\mathbb{B}}+\varphi_\mathbb{B}^{y+2\varepsilon\mathbb{B}}$. Then for $x\in y+\varepsilon\mathbb{B}$ we have $k_2(x,y)=\mathbb{E}((h^\mathbb{B},\eta_{x})(h^\mathbb{B},\eta_{y}))$ where $\eta_x$ and $\eta_y$ are smooth functions with mass one, radially symmetric around $x,y$ and supported in small non-intersecting balls around $x$ and $y$ respectively. Using the decomposition and harmonicity of $\varphi_\mathbb{B}^{y+2\varepsilon\mathbb{B}}$ we have $$k_2(x,y)=\mathbb{E}((h^\mathbb{B}_{y+2\varepsilon\mathbb{B}},\eta_x)(h^\mathbb{B}_{y+2\varepsilon\mathbb{B}},\eta_y))+\mathbb{E}(\varphi_\mathbb{B}^{y+2\varepsilon\mathbb{B}}(x)\varphi_\mathbb{B}^{y+2\varepsilon\mathbb{B}}(y))$$
where the second term on the right-hand side is bounded by Cauchy-Schwarz and \eqref{eq:var_bounds}. The first term is equal to $bs(|x-y|)+O(1)$ in some neighbourhood of $y$
by the assumption \ref{inv} on the distribution of $h^{y+2\varepsilon\mathbb{B}}_\mathbb{B}$, and the previous paragraph.
We have therefore shown that $x\mapsto k_2(x,y)$ satisfies the condition $\star$ and is therefore equal to a multiple (which must actually be $b$) of $G^\mathbb{B}(x,y)$.
Since $y\in \mathbb{B}$ was arbitrary we are done.
\end{proof}
\begin{cor}\label{cor}
For any $f\in H^{-1}(\mathbb{B})$, take a sequence of smooth functions $(f_n)_{n\ge 0}$ such that $G^\mathbb{B}(f-f_n,f-f_n)\to 0$. Then we may define $(h^\mathbb{B},f)$ to be the $L^2$ limit of $(h^\mathbb{B},f_n)$ as $n\to \infty$. The limit does not depend on the approximation.
\end{cor}
In what follows, we will therefore use the notation $(h^\mathbb{B}, f)$ for $f\in H^{-1}(\mathbb{B})$ without further justification.
\section{Gaussianity}\label{sec:gauss}
It now remains to argue that the field is Gaussian. We do this via a decomposition of the field into a sequence of radial processes. We show using the domain Markov property that each of these processes is Gaussian, and that moreover, the whole sequence is jointly Gaussian.
First, we will see a bound for the 4-point function that is reminiscent of Wick's theorem: this will help us deduce continuity of our processes. In the second subsection, we do the basic case - the case of spherical averages. Finally, we extend this to a wider range of processes, obtained from so called spherical harmonics. The Gaussianity of the field $(h^\mathbb{B}, f)_{f \in C_c^\infty(\mathbb{R}^d)}$ is proved in this final subsection.
\subsection{Weak Gaussianity in terms of the 4-point function}
To prove continuity of our radial processes, we need the following bound that is implied by our assumption on existence of fourth moments. Notice that this can be seen as establishing a very weak form of Wick's theorem, i.e. getting us closer to Gaussianity. Recall that for $r>0$, $h_r(0)=\varphi_\mathbb{B}^{r \mathbb{B}}(0)$ is the spherical average at radius $r$ around the origin, and converges to $0$ in probability as $r\uparrow 1$.
\begin{lemma}\label{lem:4mtbnd} For $r\in (0,1)$ we have that
\begin{equation}\label{4mom_av} \mathbb{E}(h_r(0)^4)=\iint_{\partial(r\mathbb{B})^4} k_4(z_1,z_2,z_3,z_4) \prod_{i=1}^4 \rho_0^r(dz_i) \end{equation}
where the right-hand side is well defined as the limit when $a\to 0$ of $\iint_{|z_i-z_j|>a \, \forall i,j} k_4(z_1,z_2,z_3,z_4) \prod \rho_0^r(dz_i).$ Moreover, for some constant $c(d)$ and some $\eta\in[0,1)$:
\begin{equation}\label{K4bound} |k_4(z_1,z_2,z_3,z_4)|\le c(d) \delta^{-\eta} g(z_1,z_2,z_3,z_4)\end{equation} for all $\delta\in (0,1]$ and distinct $z_1,\dots, z_4 \in \partial (1-\delta)\mathbb{B}$, where $$g(z_1,z_2,z_3,z_4)= G^\mathbb{B}(z_1,z_2)G^\mathbb{B}(z_3,z_4)+G^\mathbb{B}(z_1,z_3)G^\mathbb{B}(z_2,z_4)+G^\mathbb{B}(z_1,z_4)G^\mathbb{B}(z_2,z_3)$$
{is the four-point function for the $d-$dimensional zero boundary GFF in $\mathbb{B}$.}
\end{lemma}
Before proving the lemma, let us show that for any {$0<r<1-\eta$},
and for some constant $C(\eta)$ depending only on $\eta$,
\begin{equation}\label{eq:k4sphere_crude}
\limsup_{a>0}\iint_{\partial(r\mathbb{B})^4\setminus\{A_a\}} |k_4(z_1,z_2,z_3,z_4)| \prod_{i=1}^4 \rho_0^r(dz_i) \le C(\eta) s(r)^2,
\end{equation}
where $A_a=\{z_1,z_2,z_3,z_4\in \mathbb{B}^4: d(z_i,z_j)\le a \text{ for some } i\ne j\}$.
Indeed, by symmetry it suffices to bound the integral over the region where $\min(|z_i-z_j|)=|z_1-z_2|$. However, on this region, by \eqref{eq:moment_bounds}, we can bound $|k_4|$ above by an $\eta$-dependent constant times
$$ (1+s(|z_1-z_2|))\sum_{i\ne 3} \sqrt{1+s(|z_i-z_3|)} \sum_{j\ne 4} \sqrt{1+s(|z_j-z_4|)}.$$ Then expanding the product of the two sums, we can show the desired bound for the integral of each term separately (uniformly in $a$), using Cauchy--Schwarz in the integral over $z_3$ and $z_4$ for all terms, except $s(|z_3-z_4|)s(|z_1-z_4|)$ which can be integrated directly. Here we are using the fact that for any fixed $z_1\in \partial (r\mathbb{B})$, we have that $\int_{\partial (r\mathbb{B})} s(|z_1-z|){\rho_0^r(dz)}=O(s(r))$.
In particular, applying dominated convergence,
this shows that the integral in \eqref{4mom_av} is well-defined and finite.
The expression for the 4-point function now also follows from dominated convergence:
\begin{proof}[Proof of \eqref{4mom_av}]
Fix $r\in (0,1)$ and for each $n\ge 1$ partition the sphere $\partial((r-2^{-n})\mathbb{B})$ into regions each having diameter no larger than $2^{-n}$. For each of these regions we can then choose a ball of maximal radius centred at a point in the region, so that the ball intersected with the sphere lies inside the region, but within distance $2^{-2n}$ from the boundary of the region. This produces a sequence $\{z_{i,n}\}_{1\le i\le m_n}$ and radii $
\{r_{i,n}\}_{1\le i \le m_n}$ (all less than $2^{-(n+1)}$) such that the balls $z_{i,n}+r_{i,n}\mathbb{B}$ do not intersect.
We can now set $\nu_{i,n}$ for each $i,n$ to be a smooth mollifier supported on $z_{i,n}+r_{i,n}\mathbb{B}$, that is radially symmetric about $z_{i,n}$ and has total mass one. Using that $k_2=bG^\mathbb{B}$, we see that $m_n^{-1}\sum_{i=1}^n (h^\mathbb{B},\nu_{i,n})\to h_r(0)$ in $L^2$ as $n\to \infty$. On the other hand, the definition of $k_4$ gives that $k_4(z_{i,n},z_{j,n},z_{k,n},z_{l,n})=\mathbb{E}((h^\mathbb{B},\nu_{i,n})(h^\mathbb{B},\nu_{j,n})(h^\mathbb{B},\nu_{k,n})(h^\mathbb{B},\nu_{l,n}))$ for any distinct $1\le i,j,k,l\le m_n$ and every $n$. Dominated convergence using \eqref{eq:k4sphere_crude} then allows us to conclude.
\end{proof}
The Wick-type of bound is slightly trickier:
\begin{proof}[Proof of \eqref{K4bound}] Fix $z_1,z_2,z_3,z_4$ distinct and for each $j$ write $a_j=\min_{i\ne j} d(z_i,z_j)/2$.
Denote $B_j:=(z_j+a_j\mathbb{B})\cap \mathbb{B}$ for each $j$ so that the $B_j$ do not intersect. Write $\rho_j$ for uniform measure on $\partial(z_j+a_j\mathbb{B})$ and $\tilde{\rho}_j$ for harmonic measure on $\partial B_j\setminus \partial \mathbb{B}$, seen from $z_j$.
\begin{claim}\label{claim:4mt} We can write
\begin{equation}\label{eq:k4rewrite}k_4(z_1,z_2,z_3,z_4)=\mathbb{E}((h^\mathbb{B},\tilde{\rho}_1)(h^\mathbb{B},\tilde{\rho}_2)(h^\mathbb{B},\tilde{\rho}_3)(h^\mathbb{B},\tilde{\rho}_4)).\end{equation}
\end{claim}
\noindent This claim could be checked straightforwardly if we could apply the domain Markov property inside the regions $B_1$ to $B_4$, so it should not be too surprising. However, since the Markov property has only been assumed for balls, we will have to do a little bit of work to show this. This will be carried out shortly, but let us first see how \eqref{K4bound} follows.
First, notice that using the domain Markov property we can write $$\mathbb{E}((h^{2\mathbb{B}},\tilde\rho_j)^4)=\mathbb{E}((\varphi^{\mathbb{B}}_{2\mathbb{B}},\tilde{\rho}_j)^4)+6\mathbb{E}((\varphi^{\mathbb{B}}_{2\mathbb{B}},\tilde{\rho}_j)^2)\mathbb{E}((h^\mathbb{B}_{2\mathbb{B}},\tilde{\rho}_j)^2)+\mathbb{E}((h^\mathbb{B}_{2\mathbb{B}},\tilde{\rho}_j)^4)$$ and hence bound $\mathbb{E}((h^\mathbb{B},\tilde{\rho}_j)^4)\le \mathbb{E}((h^{2\mathbb{B}},\tilde{\rho}_j)^4)$ for each $j$ . By the same dominated convergence argument as for \eqref{K4bound}, and writing $k_4^{2\mathbb{B}}$ for the four-point function of $h^{2\mathbb{B}}$, we see that the latter is equal to $\iint_{\partial (z_j+a_j\mathbb{B})} |k_4^{2\mathbb{B}}(x,y,z,w)| \tilde{\rho_1}(dx)\dots \tilde{\rho_4}(dw)$. But then by \eqref{eq:k4sphere_crude} and using that each $z_j+a_j\mathbb{B}$ is far from the boundary of $2\mathbb{B}$ we see that this is less than some constant, not depending on $\delta$, times
\begin{equation}\label{density_sup}\sup_{\partial B_j\setminus \partial \mathbb{B}} \|\frac{d\tilde{\rho}_j}{d\rho_j}\| \; s(a_j)^2.
\end{equation}In the case that $a_j>\delta$, $\sup_{\partial B_j\setminus \partial \mathbb{B}} \|d\tilde{\rho}_j/d\rho_j\|$ can be bounded above by the probability that $d$-dimensional Brownian motion on the half space $\{(x_1,\dots, x_d): x_1>0\}$, started from $(\delta/a_j)$ reaches the boundary of the unit sphere before hitting the hyperplane $\{x_1=0\}$. This is bounded above by a constant times $(\delta/a_j)$ \cite[Theorem 4.4]{Burdzy}. When $\delta>a_j$ we have {$\sup_{\partial B_j\setminus \partial \mathbb{B}} \|d\tilde{\rho}_j/d\rho_j\|=1$}. So overall, we obtain the bound (where from now on $a\lesssim b$ means $a\le C b$ for some constant $C$ depending on the dimension):
$$ \mathbb{E}((h^\mathbb{B},\tilde{\rho}_j)^4)^{1/4}\lesssim (\frac{\delta}{a_j}\wedge 1)\sqrt{s(a_j)}. $$
Thus applying Cauchy--Schwarz to \eqref{eq:k4rewrite} we see that
$$k_4(z_1,z_2,z_3,z_4)\lesssim \prod_{j=1}^4 (\frac{\delta}{a_j}\wedge 1) \sqrt{s(a_j)}.$$
On the other hand, we can lower bound $g(z_1,z_2,z_3,z_4)$ using the explicit expression $$G^\mathbb{B}(x,y)=s(|x-y|)-s(|x||y-\tilde{x}|) \quad ; \quad \tilde{x}=|x|^{-2}x.$$
In particular, when $|x|=|y|=1-\delta=r$, we have that $|x|^2|y-\tilde{x}|^2|x-y|^{-2}=(1+|x-y|^{-2}\delta^2(2-\delta)^2)$. This implies that, on the region $|x-y|>\delta$, we have $ G^\mathbb{B}(x,y)\gtrsim \delta^2|x-y|^{-2}$ for $d=2$ and $G^\mathbb{B}(x,y) \gtrsim |x-y|^{2-d}\delta^2|x-y|^{-2}$ for $d\ge 3$. On the region $|x-y|<\delta$ we have $G^\mathbb{B}(x,y)\gtrsim s(|x-y|)$.
Without loss of generality we can now assume that $a_1\ge a_2\ge a_3\ge a_4$.
Combining the above lower bounds for $G^\mathbb{B}$ and the definition of $g$ with the upper bounds for $k_4$, we obtain \eqref{K4bound} under the condition that $a_2>a_1/10$. Note that the $\delta^{-\eta}$ correction is only actually needed when the dimension $d=2$.
When $a_2 \leq a_1/10$, i.e. one point is considerably further than the rest, our bounds do not suffice - this is because we haven't taken properly care of cancellations occurring in the {third moment, when three points are together}. Let us do that now.
Notice that when $a_2 \leq a_1/10$ we have $B_2,B_3,B_4\subset z_2+a_1\mathbb{B}$; we write $\hat\rho_i$ for harmonic measure seen from $z_2,z_3,z_4$ on $\partial(z_2+a_1\mathbb{B})\setminus \partial\mathbb{B}$. We need an extension of \cref{claim:4mt}, that first separates the three points, and looks at the occurring cancellations:
\begin{claim}
In the case that $a_2<a_1/10$ we can further write $k_4(z_1,z_2,z_3,z_4)$ as
\begin{equation}\label{eq:k4rewrite2}\mathbb{E}((h^\mathbb{B},\tilde{\rho}_1)\prod_{i\ne 1} (h^\mathbb{B},\hat\rho_i))+\sum_{\sigma\in S({2,3,4})} \mathbb{E}((h^\mathbb{B},\tilde{\rho}_1)(h^\mathbb{B},\hat\rho_{\sigma(2)}))\mathbb{E}((h^\mathbb{B},\hat\rho_{\sigma(3)}-\tilde{\rho}_{\sigma(3)})(h^\mathbb{B},\hat\rho_{\sigma(4)}-\tilde{\rho}_{\sigma(4)})).\end{equation}
\end{claim}
Again, we postpone the proof of the claim and first see how it implies \eqref{K4bound}.
From the same Cauchy--Schwarz argument and bounds as in $a_2>a_1/10$ case (noting also that $d(z_j,\partial (z_2+a_1\mathbb{B}))>a_1/2$ for $j=3,4$), we can deduce that the first term in \eqref{eq:k4rewrite2} is $\lesssim \delta^{-\eta} g(z_1,z_2,z_3,z_4)$ for some $\eta\in [0,1)$. Again the $\delta^{-\eta}$ correction is only needed when $d=2$.
To deal with the latter terms in \eqref{eq:k4rewrite2}, we use the fact that the covariance of $h^\mathbb{B}$ is equal to $bG^\mathbb{B}$. Indeed, using the explicit expression for $G^\mathbb{B}$ we see that the latter term is given
by $$4b^2\sum_{\sigma\in S({2,3,4})}G^\mathbb{B}(z_1,z_{\sigma(2)})G^{(z_2+a_1\mathbb{B})\cap \mathbb{B}}(z_{\sigma(3)},z_{\sigma(4)}),$$
and since $G^{D'}\le G^D$ for $D'\subset D$, \eqref{K4bound} then follows in the case $a_2 \leq a_1/10$ too. \medskip
{It now remains to prove the claims.}
\begin{proof}[Proof of Claims:]
We start with \eqref{eq:k4rewrite}. As mentioned above, we cannot apply the Markov property directly inside the regions $B_j$. Instead we will work in $2\mathbb{B}$ and use the equivalent of \eqref{eq:k4rewrite} for the field $h^{2\mathbb{B}}$, together with the Markov decomposition $h^{2\mathbb{B}}=h^\mathbb{B}_{2\mathbb{B}}+\varphi^\mathbb{B}_{2\mathbb{B}}$. In this decomposition, the two summands are independent, $h^{\mathbb{B}}_{2\mathbb{B}}$ has the law of $h^\mathbb{B}$ and $\varphi_{2\mathbb{B}}^\mathbb{B}$ is harmonic inside $\mathbb{B}$.
For each $i = 1 \dots 4$, let $\eta_i$ be a smooth function supported in $B_i$, rotationally symmetric about $z_i$, such that $k_4(z_1,z_2,z_3,z_4)=\mathbb{E}(\prod_i (h^\mathbb{B},\eta_i))$. Let us also denote by $\rho_i'$ the harmonic measure on $\partial B_i$ (now including the part on $\partial \mathbb{B}$) seen from $z_i$
By the Markov decomposition for the field $h^{2\mathbb{B}}$ inside $\{z_1+a_1\mathbb{B},\dots, z_4+a_4\mathbb{B}\}$, as in \eqref{eq:multidmp} with $n=4$, we have that {$\mathbb{E}(\prod_{i=1}^4 (h^{2\mathbb{B}},\eta_i)) = \mathbb{E}(\prod_{i=1}^4 (h^{2\mathbb{B}},\rho_i')).$} We use the domain Markov decomposition $h^{2\mathbb{B}}=h^\mathbb{B}_{2\mathbb{B}}+\varphi^\mathbb{B}_{2\mathbb{B}}$ to rewrite this as
$$\mathbb{E}(\prod_{i=1}^4 (h^\mathbb{B}_{2\mathbb{B}}+\varphi^\mathbb{B}_{2\mathbb{B}},\eta_i)) = \mathbb{E}(\prod_{i=1}^4 (h^\mathbb{B}_{2\mathbb{B}}+\varphi^\mathbb{B}_{2\mathbb{B}},\rho_i')).$$
Opening the brackets and using the independence of $h^\mathbb{B}_{2\mathbb{B}}$ and $\varphi^\mathbb{B}_{2\mathbb{B}}$ we can write this further as
$$\sum_{S \cup S' = \{1,2,3,4\}} \mathbb{E}(\prod_{i\in S} (h^\mathbb{B}_{2\mathbb{B}},\eta_i))\mathbb{E}(\prod_{j\in S'} (\varphi^\mathbb{B}_{2\mathbb{B}},\eta_j)) = \sum_{S \cup S' = \{1,2,3,4\}} \mathbb{E}(\prod_{i\in S} (h^\mathbb{B}_{2\mathbb{B}},\rho_i'))\mathbb{E}(\prod_{j\in S'} (\varphi^\mathbb{B}_{2\mathbb{B}},\rho_j')).$$
As $h^{\mathbb{B}}_{2\mathbb{B}}$ has the law of $h^\mathbb{B}$, to prove that $\mathbb{E}(\prod_{i=1}^4 (h^{\mathbb{B}},\eta_i)) = \mathbb{E}(\prod_{i=1}^4 (h^{\mathbb{B}},\rho_i'))$ it suffices to show that all terms with $|S| \neq 4$ cancel out. Since $(h^\mathbb{B},\tilde{\rho}_i)=(h^\mathbb{B},\rho_i')$ a.s.\ for each $i$, this proves the claim.
To show the cancellation, first observe that as $\varphi^\mathbb{B}_{2\mathbb{B}}$ is harmonic inside each $B_j$, we have that $(\varphi^\mathbb{B}_{2\mathbb{B}},\eta_j) = (\varphi^\mathbb{B}_{2\mathbb{B}},\rho_j')$ for every $j = 1 \dots 4$ and thus the terms with $|S|' = 4$ cancel out. Also, both $h^\mathbb{B}_{2\mathbb{B}}$ and $\varphi^\mathbb{B}_{2\mathbb{B}}$ are mean zero, so all terms with $|S| = 1$ or $|S|' = 1$ also cancel out. We are left to consider the cases when $|S| = 2$ and $|S'| = 2$.
But we already know that $$\mathbb{E}(\prod_{j\in S'} (\varphi^\mathbb{B}_{2\mathbb{B}},\eta_j)) = \mathbb{E}(\prod_{j\in S'} (\varphi^\mathbb{B}_{2\mathbb{B}},\rho_j')),$$ and thus it remains to just verify that for $i \neq j$, we have that $\mathbb{E}((h^\mathbb{B}_{2\mathbb{B}},\eta_i) (h^\mathbb{B}_{2\mathbb{B}},\eta_j)) = \mathbb{E}((h^\mathbb{B}_{2\mathbb{B}},\rho_i') (h^\mathbb{B}_{2\mathbb{B}},\rho_j'))$. This can be however verified directly via the fact that the covariance is a multiple of the Green's function and all $\eta_i, \eta_j, \rho_i',\rho_j'$ have disjoint support.
The proof of \eqref{eq:k4rewrite2} given \eqref{eq:k4rewrite} follows from the same argument, using the domain Markov decomposition of $h^{2\mathbb{B}}$ inside $z_1+a_1\mathbb{B}$ and $z_2+a_1\mathbb{B}$. We omit the details.
\end{proof}
\end{proof}
\subsection{Gaussianity of spherical averages}
We now show that the $r$-spherical average process around $0$ with a varying radius is a Gaussian process.
\begin{lemma}\label{lem:ca_bm}
$(h_{r}(0))_{r\in(0,1)}$ is a Gaussian process.
\end{lemma}
To prove this, we will use the fact that any continuous stochastic process (indexed by positive time) with independent increments, is Gaussian. The fact that $h_r(0)$ has a continuous modification in $r$
comes from \cref{lem:4mtbnd} of the previous subsection.
\begin{proof}[Proof of \cref{lem:ca_bm}]
Since the variance of this process increases as $r\downarrow 0$, it is natural to parameterise the time so that the process starts at $r = 1$. Thus we set $X_t = h_{1-r}(0)$. Then $X_t\to 0$ in probability as $t\downarrow 0$ and $X$ has independent increments by the domain Markov property (more precisely \eqref{eq:it_dmp}). Thus, to show that it is a Gaussian process, by \cite[Theorem 11.4]{Kallenberg}, it suffices to prove that it admits a continuous modification in $t$.
We apply Kolmogorov's continuity criterion to show this. By the domain Markov property and scaling assumption it suffices to control the behavior at time $0$, i.e. it is enough to show that for some $C\in (0,\infty)$, some $\eta<1$ and all $\delta\in (0,1)$:
\begin{equation}\label{eqn:required_4mom}
\mathbb{E}(X_\delta^4)\le C\delta^{2-\eta}.
\end{equation}
For this we use \eqref{K4bound}. We obtain that
$$\mathbb{E}(X_\delta^4)=\iint_{\partial((1-\delta)\mathbb{B})^4} k_4(z_1,z_2,z_3,z_4) \prod_{i=1}^4 \rho_0^{1-\delta}(dz_i) \lesssim \delta^{-\eta}\iint_{\partial((1-\delta)\mathbb{B})^4} g(z_1,z_2,z_3,z_4) \prod_{i=1}^4 \rho_0^{1-\delta}(dz_i).$$
Now recalling that $g(z_1,z_2,z_3,z_4)$ is the four-point function for the zero boundary GFF in $\mathbb{B}$, we see that the integral on the right hand side is the fourth moment of the spherical average of the GFF at radius $1-\delta$. But this spherical average is a centred Gaussian with variance $-\log(1-\delta)$ when $d=2$ and $1-(1-\delta)^{2-d}$ when $d>2$ - see \cite[equation (13)]{WPgff}. Since these are both of order $\delta$ as $\delta\to 0$ the fourth moment is of order $\delta^2$ by Gaussianity {of the GFF itself}.
\end{proof}
\subsection{Gaussianity in the general case}
\noindent In what follows, we will often use the co-ordinates $z=(r,\overline{\theta})=(|z|,z/|z|)$ for a point in $\mathbb{R}^d, d\ge 2$.
{We will generalise {the case of spherical averages} to a wider class of processes, stemming from so called spherical harmonics. The interest comes from the following classical theorem (see e.g. \cite[Chapter IV]{SteinWeiss}).
\begin{theorem}[Expansion using spherical harmonics]\label{thm:sph}
In each $d \geq 2$, there is a collection of smooth functions $\psi_{n,j}(\overline{\theta}): \partial \mathbb{B} \to \mathbb{R}$ with $n \in \mathbb{N}_0, M_n \in \mathbb{N}$ and $j \in \{1, \dots, M_n\} $ such that
\begin{enumerate}
\item The functions $\psi_{n,j}(\overline{\theta})$ form an orthonormal basis of $L^2(\partial \mathbb{B})$;
\item For every $n \in \mathbb{N}_0$ and $j \in \{1, \dots, M_n\} $, we have that $(r,\overline{\theta})\mapsto r^n\psi_{n,j}(\overline{\theta})$ is harmonic in $\mathbb{B}$;
\item For each $n \in \mathbb{N}_0$, one can find radially symmetric functions $f_{n,i}(r):[0,1] \to \mathbb{R}$ with $i \in \mathbb{N}_0$ such that $(e_{n,j,i} = f_{n,i}\psi_{n,j})_{n \in \mathbb{N}_0, j \in \{1, \dots, M_n\},i \in \mathbb{N}}$ form an orthonormal basis of $L^2(\mathbb{B})$.
\end{enumerate}
\end{theorem}
\begin{rmk}
In fact one can write out a specific collection of such functions using Legendre polynomials and Bessel functions and choose $e_{n,j,i}$ to be eigenfunctions of $\Delta$. This is not necessary here.
For example, in $d = 2$, one has $M_0=1$ and $M_n = 2$ for $n\ge 1$, with the $\psi$-s given by the usual Fourier series on the circle. That is, $$\{\{e_{0,1,k}\}_{k\ge 1},\{e_{n,1,k},e_{n,2,k}\}_{n\ge 1,k\ge 1}\}:=\{\{J_0(\alpha_{0,k})\}_{k\ge 1}, \{J_n(\alpha_{n,k}r) \sin(n\theta), J_n(\alpha_{n,k}r) \cos(n\theta)\}_{n\ge 1, k\ge 1}\}$$ form an orthonormal basis of $L^2(\mathbb{B})$, where $(J_n)_{n\ge 0}$ are the Bessel functions and $\alpha_{n,k}$ are the zeroes of $J_n$ for each $n$.
\end{rmk}
Using these notations, the main result of this section can be stated as follows:
\begin{prop}[Gaussianity]\label{prop:gauss}
The random variables $(h^\mathbb{B},e_{n,j,i})_{n \in \mathbb{N}, j \in \{1, \dots, M_n\},i \in \mathbb{N}}$ are jointly Gaussian. In particular $(h^\mathbb{B}, f)_{f \in C_c^\infty(\mathbb{R}^d)}$ is a Gaussian process.
\end{prop}
To prove \cref{prop:gauss} we will choose appropriate radial functions $g_{n,j}$ for which we can verify that $h^\mathbb{B}$ tested against $ g_{n,j}(r)\psi_{n,j}(\overline{\theta})$ on each sphere at radius $r$ is a Gaussian process in $r\in(0,1)$. The key observation is the following. For $r\in (0,1)$ and a smooth function $\psi: \partial \mathbb{B} \to \mathbb{R}$ let $\nu^\psi_r$ be the signed measure defined by the condition that
\begin{equation}\label{eq:av}
\nu^\psi_r(\phi)=\int_{\partial \mathbb{B}} \psi(\overline{\theta}) \phi(r,\overline{\theta}) \rho_0^1(d\overline{\theta})
\end{equation}
for all functions $\phi$ such that $\phi(r,\overline{\theta})\in L^1(\partial \mathbb{B},\rho_0^1(d\overline{\theta}))$, where as before $\rho_0^1$ is uniform measure on $\partial \mathbb{B}$.
\begin{lemma}\label{lem:sin_const}
Suppose that $\varphi$ is a harmonic function in $\mathbb{B}$. Suppose also that $\psi(\overline{\theta})$ is a smooth function defined on $\partial \mathbb{B}$ such that $(r,\overline{\theta})\mapsto \psi(\overline{\theta})r^{n}$ is harmonic in $\mathbb{B}$. Then $r^{-n}\nu^\psi_r(\varphi)$ is constant {as a function of $r$ on $(0,1)$}.
\end{lemma}
\begin{proof}
Let us fix $r_0\in (0,1)$: we will show that $(\frac{d}{dr})[r^{-n}\nu_{r}^\psi(\varphi)]|_{r=r_0}=0$, which implies the result. By the second Green's identity applied in $r_0\mathbb{B}$ to the harmonic functions $\varphi$ and $(r,\overline{\theta})\mapsto r^n\psi$, we can write
$$\int_{\partial (r_0\mathbb{B})}(\varphi\frac{d}{dr}[r^n\psi] - r^n\psi\frac{d}{dr}\varphi) = \int_{r_0 \mathbb{B}} (\varphi \Delta[ r^n\psi] - r^n\psi \Delta \varphi),$$
where we are integrating against the standard volume measure on $r_0\mathbb{B}$ on the right-hand side, and the induced measure on its boundary on the left, which is a multiple of uniform measure.
Now the right-hand side is zero as both $r^n\psi$ and $\varphi$ are harmonic by assumption. Thus we deduce that
$$n\int_{\partial \mathbb{B}}\varphi(r_0,\overline{\theta})\psi(\overline{\theta}) \, \rho_0^1(d\overline{\theta})= r_0\int_{\partial \mathbb{B}}\psi(\overline{\theta})\frac{d}{dr}\varphi(r,\overline{\theta})\big|_{r=r_0} \, \rho_0^1(d\overline{\theta}).$$
It follows that
$$\left.\frac{d}{dr}[r^{-n}\nu^\psi_r(\varphi)]\right|_{r=r_0} = -nr_0^{-n-1}\int_{\partial \mathbb{B}}\varphi(\overline{\theta}) \psi(r_0,\overline{\theta})+ r_0^{-n}\int_{\partial \mathbb{B}}\psi(\overline{\theta})\frac{d}{dr}\varphi(r,\overline\theta)\big|_{r=r_0} = 0.$$
\end{proof}
Notice that all the functions $\psi_{n,j}$ of \cref{thm:sph} satisfy the conditions for the function $\psi$ in this lemma. A similar proof to the spherical average case now implies that any finite linear combination of the functions $(r^{-n}\nu^{\psi_{n,j}}_r)_{n \geq 0, j \in \{1, \dots, M_n\}}$ tested against $h^\mathbb{B}$ gives rise to a Gaussian process.
\begin{lemma}\label{lem:gp} The process $(A_r)_{r\in (0,1)}$ defined by
\begin{equation}\label{eq:gp}
A_r := \sum_{i=1}^k a_i (h^\mathbb{B},r^{-n_i}\nu^{\psi_{n_i,j_i}}_r)
\end{equation} is Gaussian for any $k\ge 0$, $n_i\in \mathbb{N}_0,j_i \in \{1, \dots, M_{n_i}\}$ and $(a_1,\dots, a_k)$ real.
In particular, for any fixed radius $r \in (0,1)$, we have that $(h^\mathbb{B}, \nu^{\psi_{n,j}}_r)_{n \in \mathbb{N}_0, j \in \{1, \dots, M_n\}}$ is jointly Gaussian.
\end{lemma}
}
\begin{proof}
First let us fix $a_0,\dots, a_k$ in $\mathbb{R}$. We again parametrise the process from large radii towards small radii, i.e. let's set $X_t = A_{1-t}$. We first argue using \cref{lem:sin_const} that $X_t$ has independent increments.
For clarity, let us first show this for $\tilde X_t = (h^\mathbb{B},(1-t)^{-n_i}\nu^{\psi_{n_i,j_i}}_{1-t})$. Indeed, for each $i \in \{1, \dots, k\}$ and each $s < r < 1$ we can write by domain Markov property
$$(h^\mathbb{B},s^{-n_i}\nu^{\psi_{n_i,j_i}}_s) = (h^{r\mathbb{B}},s^{-n_i}\nu^{\psi_{n_i,j_i}}_s) + (\varphi^{r\mathbb{B}},s^{-n_i}\nu^{\psi_{n_i,j_i}}_s).$$
As already mentioned, $\psi_{n_i,j_i}$ satisfies the conditions of \cref{lem:sin_const}, and thus
$$(\varphi^{r\mathbb{B}},s^{-n_i}\nu^{\psi_{n_i,j_i}}_s) = \lim_{u\uparrow r}(\varphi^{u\mathbb{B}},u^{-n_i}\nu^{\psi_{n_i,j_i}}_u)=(h^\mathbb{B},r^{-n_i}\nu_r^{\psi_{n_i,j_i}})$$
where the last equality follows since $(h^{r\mathbb{B}},u^{-n_i}\nu^{\psi_{n_i,j_i}}_u)$ converges to $0$ as $u\uparrow r$ in $L^2$, by a direct calculation using the Green's function.
As $h^{r\mathbb{B}}$ and $\varphi^{r\mathbb{B}}$ are independent, we conclude that {$\tilde X_t = (h^\mathbb{B},(1-t)^{-n_i}\nu^{\psi_{n_i,j_i}}_{1-t})$ has independent increments.} {The same} argument can be directly applied to get the claim for $X_t$.
{Now, $X_t$ is also centred and square integrable, converging to $0$ in probability as $t\downarrow 0$}, by definition and assumptions on $h^\mathbb{B}$. Moreover, {using the fact that each $\psi_{n,j}$ is bounded on $\partial \mathbb{B}$, we can use the 4th moment bound \eqref{K4bound} to apply Kolmogorov's criteria similarly to the case of the spherical average and obtain that $X_t$ possesses an a.s. continuous modification.} This implies that the process is Gaussian by \cite[Theorem 11.4]{Kallenberg}.
For the final claim, let us fix $r\in (0,1)$. Then $(h^\mathbb{B}, \nu^{\psi_{n,j}}_r)_{n \geq 0, j \in \{1, \dots, M_n\}}$ is jointly Gaussian, since for any $k\ge 1$, $b_1,\dots, b_k\in \mathbb{R}$, $n_i\in \mathbb{N}_0$ and $j_i\in \{1,\dots, M_{n_i}\}$, we can apply the above with $a_i=r^{n_i}b_i$ for each $i$ to see that $\sum_i b_i (h^\mathbb{B},\nu_r^{\psi_{n_i,j_i}})$ is a Gaussian random variable.
\end{proof}
There is one more step required to deduce \cref{prop:gauss}: to extend the joint Gaussianity to varying radii.
\begin{lemma}
For any $k \geq 1$, and any $(r_i,n_i,j_i)_{1\le i \le k}$ with $r_i\in (0,1), n_i\in \mathbb{N}_0$ and $j_i\in \{1,\dots, M_{n_i}\}$ for each $i$, the vector $$\left((h^\mathbb{B}, \nu^{\psi_{n_1,j_1}}_{r_1}),\dots, (h^\mathbb{B}, \nu^{\psi_{n_k,j_k}}_{r_k})\right)$$ is Gaussian.
In particular, we have that $(h^\mathbb{B}, \nu^{\psi_{n,j}}_r)_{r \in (0,1), n \geq 0, j \in \{1, \dots, M_n\}}$ is jointly Gaussian.
\end{lemma}
\begin{proof}
The second statement is an immediate consequence of the first. So, let us fix $k,(r_i,n_i,j_i)_{i\le k}$ as in the statement.
For now, let us assume that $r_1>\dots>r_k>0$ (we come back to the general case at the end of the proof). We iterate the domain Markov property inside each ball $r_i\mathbb{B}$ as follows: we write $h^\mathbb{B}=h^0$, then $h^0=h^1+\varphi^1$ by applying the Markov property inside $r_1\mathbb{B}$, then $h^1=h^2+\varphi^2$ inside $r_2\mathbb{B}$ etc.\ so that $(\varphi^1,\dots, \varphi^k)$ are mutually independent by \eqref{eq:it_dmp}. Shortening the notation $\nu_r^{\psi_{n_i,j_i}}$ to $\nu_r^i$ for each $i=1,\dots, k$, it suffices to show that \begin{equation}\label{dmp_jointcircles}\left((h^0,\nu_{r_1}^1),(\varphi^1,\nu_{r_2}^2),(h^1,\nu_{r_2}^2),\dots, (\varphi^{k-1},\nu_{r_k}^k),(h^{k-1},\nu_{r_k}^k)\right)\end{equation} is a Gaussian vector - indeed, we can write each $(h^\mathbb{B}, \nu_{r_i}^i) = \sum_{j =1}^i (\varphi^j, \nu_{r_i}^i)$. {However, as in the proof \cref{lem:gp}, it follows from \cref{lem:sin_const}}
that $$(h^{j-1},\nu_{r_j}^{\psi_{n,i}})=(\varphi^j,\nu_{r_j}^{\psi_{n,i}})=(\frac{r_{j+1}}{r_j})^{n}(\varphi^j,\nu_{r_{j+1}}^{\psi_{n,i}})$$ for any $\psi_{n,i}$ and any $j$.
Hence, \eqref{dmp_jointcircles} can be rewritten as
\begin{equation*}\big((\varphi^1,\nu_{r_1}^1),(\frac{r_2}{r_1})^{n_2}(\varphi^1,\nu_{r_1}^2),(\varphi^2,\nu_{r_2}^2),(\frac{r_3}{r_2})^{n_3}(\varphi^2,\nu_{r_2}^3),\dots,(\varphi^{k-1},\nu_{r_{k-1}}^{k-1}),(\frac{r_k}{r_{k-1}})^{n_k}(\varphi^{k-1},\nu_{r_{k-1}}^k),(\varphi^k,\nu_{r_k}^k)\big).\end{equation*}
{But now $(\varphi^k,\nu_{r_k}^k)$ is Gaussian, and each pair $((\varphi^i,\nu_{r_i}^i),(\frac{r_{i+1}}{r_i})^{n_{i+1}}(\varphi^i,\nu_{r_i}^{i+1}))$ with $i = 1 \dots k -1$ is jointly Gaussian by \cref{lem:gp}. Moreover, this singleton, and each pair in the list are independent of one another by construction.} Thus we indeed have a Gaussian vector.
Finally, if the $r_i$'s are not distinct, the same argument still works: the ``pairs'' just mentioned will simply be larger tuples, concluding the proof of the lemma.
\end{proof}
{We are now ready to prove \cref{prop:gauss}.}
\begin{proof}[Proof of \cref{prop:gauss}]
Consider the basis {$(e_{n,j,i} = f_{n,i}\psi_{n,j})_{n \in \mathbb{N}_0, j \in \{1, \dots, M_n\},i \in \mathbb{N}}$} of $L^2(\mathbb{B})$ given by \cref{thm:sph}. The previous lemma implies that
\begin{equation}\label{eq:joint_manycircles}
\{(h^\mathbb{B}, f_{n,i}(r)\nu_r^{\psi_{n,j}})\}_{r \in (0,1), n \in \mathbb{N}_0, j \in \{1, \dots, M_n\}, i \in \mathbb{N}}
\end{equation}
is a Gaussian process. In particular,
\begin{equation}\label{eq:finally_gaussian}
{(h^\mathbb{B},f_{n,i}\psi_{n,j})_{n \in \mathbb{N}_0, j \in \{1, \dots, M_n\}, i \in \mathbb{N}}}
\end{equation}
is also a Gaussian process. Indeed, the random variables in \eqref{eq:finally_gaussian} exist by \cref{cor}, and we obtain the Gaussianity since they can be defined as almost sure limits of weighted sums of elements in the collection \eqref{eq:joint_manycircles}.
The fact that $e_{n,j,i}$ form a basis of $L^2(\mathbb{B})$, together with linearity of $(h^\mathbb{B}, f)$ and that $h^\mathbb{B}$ is zero outside of $\mathbb{B}$, now implies that $(h^\mathbb{B},f)_{f \in C_c^\infty(\mathbb{R}^d)}$ is a Gaussian process.
\end{proof}
\bibliographystyle{alpha}
|
\section{Introduction}
\label{section1}
Graph is a general description of data and their relations, which can model a variety of scenarios, from biomedical networks to citation networks, or other things defined by relations.
There are a variety of graph-based applications in the machine learning field, such as link prediction~\cite{SDNE}, node classification~\cite{GAE, social}, and recommendation~\cite{deepwalk, node2vec}.
As an important part of graph analysis, graph representation learning is the basis for graph data to be processed by conventional machine learning methods.
In recent years, a great deal of work has greatly promoted the development of graph representation learning.
Most graph representation learning methods can be broadly classified into two categories: proximity-based methods and neural network based methods.
Proximity-based methods, such as DeepWalk~\cite{deepwalk}, node2vec~\cite{node2vec}, and LINE~\cite{LINE}, focus on extracting the patterns from the graph data.
These methods tend to learn the node representation with various order proximity preserving by constructing the node paths through random walk or local neighborhood. Besides,~\cite{qiu2018network} demonstrated that matrix factorization methods~\cite{grarep} are also equivalent to proximity-based approaches.
While neural network-based methods usually apply graph convolution to aggregate neighborhood information~\cite{GCN, GAT, Graphsage}, or use autoencoder to merge structure and node features~\cite{SDNE, GAE, ARGA}.
\begin{figure}[t]
\centering
\includegraphics[width=3.5in]{figure/expression.pdf}
\caption{A simple example in a citation network. A paper $u$ for visual question answering (VQA) may simultaneously have citation relationships with papers in the fields of ML, NLP, and CV because of the related researches it involves. Through differentiated aggregation, the learned separable representation by graph disentanglement can reveal more about the latent semantics among different neighborhoods.
}
\label{example}
\end{figure}
Significant progress in graph representation learning has been made recently, especially the emergence of graph convolutional networks (GCNs)~\cite{survey}. Nevertheless, most of the existing methods learn node representation by treating local neighborhood or node path as a whole, ignoring the subtle differences between node neighbors. Actually, such practices make model prone to learning over-smoothed representations when network goes deeper~\cite{powerful}.
In general, a real-world graph is relatively complex and formed by the interaction of a variety of latent factors.
In other words, links between a node and its neighbors may be caused by different components, which is why nuances are presented between node neighbors. However, these nuances are often overlooked by existing methods, limiting the expressiveness of learned node representations. Instead, we would like to consider the distinction of node neighbors and disentangle the node representation towards discovering those distinct latent components with richer information, i.e., graph disentanglement.
As an example shown in Fig.~\ref{example}, unlike traditional undifferentiated aggregation of neighboring nodes by an aggregation function, graph disentanglement aims to learn the node representation that consists of some distinct latent components with each containing a certain semantic.
Thus, how to disentangle the node representation into multiple components with distinct latent semantics is of great significance for graph representation learning, which will also be helpful for improving the robustness and interpretability of graph neural network.
In the field of image analysis and understanding, disentanglement learning has been successfully applied to a variety of tasks, such as foreground separation and visual reasoning~\cite{drit,van2019disentangled}.
However, less focus has been put on disentangled representation learning of graph~\cite{DisenGCN, IPGDN, FactorGCN}.
Different from those holistic methods like GCN~\cite{GCN}, ~\cite{DisenGCN} and~\cite{IPGDN} have put focus on finding the difference of latent semantic among node neighborhood information, and utilized a neighborhood routing mechanism to infer the latent factors by iteratively analyzing the clusters of component spaces formed by nodes and their neighbors, thereby learning the disentangled components of node as representation.
Despite of the essential exploration by ~\cite{DisenGCN} and~\cite{IPGDN}, there are still some key issues that should be addressed:
\textbf{(i)} the existing methods attempt to perform micro-disentanglement by projecting node's neighborhoods into specific multiple component
spaces, but they lack consideration for the correlations of them within the component. Inevitably, it will cause the aggregated representation to be none representative of the latent semantics associated with the component;
\textbf{(ii)} In addition to the micro-disentanglement at the node level, the components should also be macro-separable at the component distribution level, while DisenGCN neglects to restrict interdependence among components; \textbf{(iii)} The disentangled representations should be inversely further reveal the intrinsic graph structure on the basis of the given initial adjacency matrix.
To address the above key issues, we propose in this paper an \textbf{\underline{a}}dversarial \textbf{\underline{d}}isentangled \textbf{\underline{g}}raph \textbf{\underline{c}}onvolutional \textbf{\underline{n}}etwork for graph disentanglement, named as ADGCN.
Specifically, to infer multiple latent factors that originate from the complex links among graph nodes, we consider simultaneously the intra-component correlation and inter-component separability to achieve component-specific aggregation for the micro-disentanglement of the node.
Furthermore, adversarial learning is adopted to restrict the interdependence among components for the macro-disentanglement of the graph.
Besides, the learned disentangled representation is also used to explore the underlying structure of the graph progressively. To achieve efficient graph refinement, we propose a diversity-preserving node sampling method to guided the local structure aware graph refinement, and further to spread to the overall graph.
We summarize the main contributions as follows:
\begin{itemize}
\item We propose a novel \underline{A}dversarial \underline{D}isentangled \underline{G}raph \underline{C}onvolutional \underline{N}etwork (ADGCN), for disentangled graph representation learning, thus to unveil the associated multiple latent components.
\item
To achieve micro-disentanglement for each node, a component-specific aggregation with dynamic assignment mechanism is proposed, which ensures the compactness of the component space and allocates the neighbors of the target node to the corresponding component space to realize micro-disentanglement.
\item
For the purpose of implementing macro-disentanglement at the component distribution level, we propose an adversarial regularizer to improve the separability among component distributions, thus learning the more distinct components.
\item
To refine the global graph structure in a progressive way, we propose a diversity-preserving node sampling approach to perform local structure aware refinement. Hence, the latent global topological graph stucture can be well reveiled to benefit graph disentanglement.
\end{itemize}
\begin{figure*}[t]
\centering
\includegraphics[width=6.1in]{figure/network.pdf}
\caption{Architecture overview of a single-layer ADGCN. $A^{(t)}$ represent the refined adjacency at epoch $t$.
$\mathbf{X}$ and $\mathbf{W}_{k}$ denote the node raw feature and the projection matrix to $k$-th component space respectively.
$\{\mathbf{h}^{k}_{u}\}^{K}_{k=1}$ represent the components of node $u$ and $\mathbf{h}^{(t)}_{u}$ as the output denotes the disentangled representations at epoch $t$, where $\mathbf{h}^{(t)}_{u}$ is used to update $A^{(t)}$ to $A^{(t+1)}$ through sample-based graph refinement.
'$+$' and '$-$' represent the real and fake samples that are sampled from $\mathbf{h}^{(t)}_{u}$ and $\mathbf{c}^{(t)}_{u}$, respectively.}
\label{ADGCN}
\end{figure*}
\section{Related Work}
\subsection{Graph convolutional networks}
Based on graph signal processing theory, some early GNNs~\cite{bruna2013,defferrard2016} implement graph Fourier transformation~\cite{shuman2013} in the spectral domain, i.e., spectral convolution.
To reduce the computational cost, the linear filter instead of the spectral graph convolution was applied in the later work GCN~\cite{GCN}, which is just similar to the message passing in~\cite{MPNN}.
In addition, the direct graph convolution in spatial domain was also investigated~\cite{Graphsage, kearnes2016molecular,GAT}.
However, these methods are commonly plagued with over-smoothing problem when deeper model is involved.~\cite{powerful} systematically analyzed the causes of the problem in GCNs and points out that the propagation process is a special form of Laplace smoothing. In addition, the expressiveness of GCNs was also further discussed by ~\cite{morris2019weisfeiler}.
Recently, many works have been proposed to overcome these problems of GCNs in the aspect of improving propagation methods or aggregation functions. JK-Net~\cite{JK-Net} adopted jump connection and layer aggregation mechanism to adaptively merge features aggregated from different hop subgraphs for each node. Inspired by PageRank~\cite{PageRank}, APPNP~\cite{PageRank} proposed to separate the transformation from the propagation. Then, \cite{AP-GCN} and \cite{DAGNN} proposed different adaptive propagation schemes based on APPNP to capture information from large receptive fields. However, most of the existing methods ignore the complex genesis of a graph, which is influenced by multiple factors. Their direct aggregation of neighboring nodes induces further entanglement of different factors, degrading the expressiveness of the model.
Essentially GAT~\cite{GAT} is also a holistic approach, although it learns weights for different neighbors.
Currently, there are still less studies on disentangled representation learning of graph data. DisenGCN~\cite{DisenGCN} first studied the problem of graph disentanglement at node level. On this basis, IPGDN~\cite{IPGDN} enhanced the independence constraint of graph disentanglement by using HSIC~\cite{HSIC}. Besides, \cite{FactorGCN} proposed graph-level disentangling for graph classification. In contrast to \cite{DisenGCN} and \cite{IPGDN}, ADGCN not only consider the disentanglement of individual nodes, but also impose constraints at the component distribution level.
\subsection{Adversarial learning on graph}
As a powerful generative model, Generative adversarial network has been proposed by Goodfellow et al.~\cite{GAN}. Then, ~\cite{molgan} and~\cite{netgan} successfully applied adversarial learning to handle graph data for graph generation task.
Besides, The success of~\cite{BiGAN,ALI} shows that GAN and adversarial learning also have excellent capability for data understanding. Hence, some studies have also applied adversarial learning to the field of graph representation learning.
\cite{graphgan} and \cite{ProGAN} generated node pairs to approximate the linkage distribution, enhancing the structure-preserving capability of the node representation. And \cite{asine} also proposed to sample sub-graph to learn node embedding for signed network based on the adversarial learning. Another common practice, as in ARGA~\cite{ARGA} and DBGAN~\cite{DBGAN}, is to utilize adversarial learning as regularizer to constrain the representation distribution of nodes, thereby improving the generalization ability of the representations. AIDW~\cite{AIDW} adopted similar form to introduce adversarial learning to~\cite{deepwalk}.
In addition, DGI~\cite{DGI} constructed degraded graphs for adversarial learning with the original graph.
~\cite{lu2020} maintained consistency of node representation between multiple graphs through adversarial learning.
Different from the above works, we apply an adversarial regularizer to control the separability between component distributions for graph disentanglement.
\section{Methodology}
\subsection{Problem Formulation and Overview of Framework}
To begin with some definitions, an undirected graph is given as $\mathcal{G} = (V,E)$, where $V$ consists of the set of nodes with $|V| = n$, and $E$ is the set of edges among nodes. To represent the link strength (similarity) among graph nodes, we use $\mathbf{A}\in R^{n\times n}$ to denote the adjacency matrix\footnote{For a graph with hard link, we have $\mathbf{A}_{i,j}$ = 1 if there exists an edge $(i,j) \in E$, otherwise $\mathbf{A}_{i,j}= 0$.}, and $\mathbf{D}=diag([\mathbf{d}_{u}]_{u=1,\cdots,n})$ to denote the corresponding degree matrix with $d_{u}=\sum_{i=1}^{n}\mathbf{A}_{u,i}$. Meanwhile, let $\mathbf{X}=[\mathbf{x}_{1},\cdots,\mathbf{x}_{n}]\in\mathbb{R}^{n \times f}$ denote the node feature matrix with $\mathbf{x}_{u}\in\mathbb{R}^{f}$, $u=1,\cdots,n$, representing the raw feature of the $u$-th node, and $\mathbf{H}=[\mathbf{h}_{1},\cdots,\mathbf{h}_{n}]\in\mathbb{R}^{n \times d}$ the node representation matrix.
In addition,
$\mathbf{N}(u)$ is defined as a node subset containing neighbors of node $u$ and $\tilde{\mathbf{N}}(u) = \mathbf{N}(u) \cup \{u\}$.
\paragraph{Micro-Disentanglement}
Assuming that the graph $\mathcal{G}$ is formed by $K$ latent factors corresponding to $K$ component spaces. For \underline{a specific node $u$}, we want to decompose its latent representation $\mathbf{h}_{u} \in \mathbf{H}$ into $K$ component representations $\{\mathbf{h}^{1}_{u}, \mathbf{h}^{2}_{u}, \cdots, \mathbf{h}^{K}_{u}\}$, where $\mathbf{h}^{k}_{u} \in \mathbb{R}^{\frac{d}{K}}$ corresponds to the $k$-th factor.
\paragraph{Macro-Disentanglement}
On the basis of micro-disentanglement, for \underline{all nodes $V$} in the given graph, we still expect the component representations lying on different component spaces are far away from each other, i.e., the difference between any two component distributions should be evident.
\paragraph{Overview of Framework} The overall illustration of the proposed ADGCN model is shown in Fig.~\ref{ADGCN}.
For a specific node $u$ and its neighbors in the given graph, the initial embeddings $\{\mathbf{c}^{k}_{i}\}^{K}_{k=1}$ are first obtained from $\mathbf{x}_{i}$ through projection matrices $\mathbf{W} = \{\mathbf{W}_{1}, \mathbf{W}_{2}, \cdots, \mathbf{W}_{K}\}$, where $i \in \tilde{\mathbf{N}}(u)$.
Then, through the component-specific aggregation, $\{\mathbf{c}^{k}_{i}\}_{i \in \tilde{\mathbf{N}}(u)}$ are aggregated into $\mathbf{h}^{k}_{u}$ in the $k$-th component space, and the concatenation of $\{\mathbf{h}^{k}_{u}\}^{K}_{k=1}$ is $\mathbf{h}_{u}=\textrm{concat}(\mathbf{h}^{1}_{u},\mathbf{h}^{2}_{u},\cdots,\mathbf{h}^{K}_{u})$.
For adversarial learning, $\mathbf{c}^{k}_{u}$ and $\mathbf{h}^{k}_{u}$ are input into the discriminator $D$ where $\mathbf{c}^{k}_{u}$ and $\mathbf{h}^{k}_{u}$ are regarded as a \textit{"fake"} sample and a \textit{"real"} sample respectively. In addition, $\mathbf{h}_{u}$ is used for graph refining to update $\mathbf{A}^{(t-1)}$ to $\mathbf{A}^{(t)}$.
\subsection{Component-Specific Aggregation for Micro-Disentanglement}
\label{MDCL}
\begin{figure}[t]
\centering
\includegraphics[width=3.2in]{figure/new_micro.pdf}
\caption{Illustration of component-specific aggregation based on dynamic assignment.
The proposed dynamic assignment iteratively allocates the neighbor nodes to different component spaces according to the hubness score $p^{k}_{u \to v}$. Meanwhile, according to the authority score $q^{k}_{v \to u}$ of node $v$ in each component space, $\{\mathbf{c}^{k}_{i}\}_{i \in \tilde{\mathbf{N}}(u)}$ are aggregated into $\tilde{\mathbf{h}}^{k}_{u}$. Here, only three component spaces are considered.}
\label{micro_disen}
\end{figure}
Based on an important assumption, i.e., the links between nodes are formed by the interaction of different latent factors, our DBGAN learn disentangled components by aggregating the different neighborhood information of the node to achieve the micro-disentanglement.
Intuitively, if node $u$ and its neighbor node $v$ are similar in the $k$-th component space, the link between the two nodes is likely caused by the factor $k$.
Furthermore, for a subset $\mathbf{N}_{k}(u) \subseteq \mathbf{N}(u)$, if all nodes in $\mathbf{N}_{k}(u)$ are connected to node $u$ due to factor $k$, these nodes should also be similar and form a cluster with $u$ in the $k$-th component space.
It is worth noting that the above two points correspond to the first-order proximity and the second-order proximity respectively in the topological structure formed by the factor $k$, respectively~\cite{LINE}. Based on these two points, DisenGCN~\cite{DisenGCN} proposed a neighborhood routing mechanism, which was also followed by~\cite{IPGDN}. However, neighborhood routing mechanism merely considers the restriction on the separability of inter-component and constrains that each neighbor should belong to only one component space, ignoring multi-relations among nodes.
Motivated by the observations mentioned above, we propose a dynamic assignment mechanism as shown in Fig.~\ref{micro_disen} to achieve component-specific aggregation for micro disentanglement.
Particularly, to characterize the interactive relationship between node $u$ and its neighbor node $v\in \mathbf{N}(u)$ in each component space, we define the \textbf{\textit{hubness score}} $p^{k}_{u \to v}$ and the \textbf{\textit{authority score}} $q^{k}_{v\to u}$ of node $u$ and $v$ in the $k$-th component space as follows:
\begin{equation}
p^{k}_{u \to v} = \frac{\mathrm{exp}(s^{k}_{uv})}{\sum_{k=1}^{K}\mathrm{exp}(s^{k}_{uv})}
\label{p}
\end{equation}
\begin{equation}
q^{k}_{v\to u} = \frac{\mathrm{exp}(s^{k}_{uv})}{\sum\limits_{v\in \mathbf{N}(u)}\mathrm{exp}(s^{k}_{uv})}
\label{q}
\end{equation}
where $s^{k}_{uv} = \cos(\mathbf{h}^{k}_{u}, \mathbf{c}^{k}_{v})$ denotes the similarity between the target node $u$ and its neighbor $v \in \mathbf{N}(u)$ in the $k$-th component space\footnote{To avoid numerical instability caused by inconsistent norms between different component spaces, $l_{2}$-normalization is adopted for $h^{k}_{u}$ and $c^{k}_{v}$ to facilitate the calculation of similarity.}.
In fact, as a bidirectional information interaction between node $u$ and its neighbor $v$, the hubness score $p^{k}_{u \to v}$ means the probability that $v$ is connected to $u$ by factor $k$, and the authority score $q^{k}_{v\to u}$ refers to the aggregation importance of $v$ for $u$ in the $k$-th component space. In particular, for a specific neighbor node $v$, the higher the hubness score $p^{k}_{u \to v}$, the more effective information between $v$ and $u$ have about factor $k$, therefore $v$ will tend to be allocated to the $k$-th component space for the aggregation of $u$. Meanwhile, a high authority score $q^{k}_{v\to u}$ means a high similarity between $u$ and $v$ in the $k$-th component space.
Based on $p^{k}_{u \to v}$ and $q^{k}_{v\to u}$, the proposed dynamic assignment mechanism can be implemented in an iterative way to obtain the aggregated representation $\mathbf{h}_u^k$ by:
\begin{equation}
{\mathbf{h}^{k}_{u}}= \underbrace{\tilde{\mathbf{h}}^{k}_{u} + \alpha \sum_{v \in \mathbf{N}(u)}q^{k}_{v\to u}\mathbf{c}^{k}_{v}}_{intra-component} + \underbrace{\beta \sum_{v \in \mathbf{N}(u)}p^{k}_{u \to v}\mathbf{c}^{k}_{v}}_{inter-component}
\label{h}
\end{equation}
where $\tilde{\mathbf{h}}^{k}_{u}$ denotes the aggregated $\mathbf{h}^{k}_{u}$ in previous iteration, $\alpha$ and $\beta$ are trade-off coefficients. The intra-component part with $q^{k}_{v\to u}$ ensures that $\tilde{h^{k}_{u}}$ tends to be aggregated into the densest area in $k$-th component space, and the inter-component part with $p^{k}_{u \to v}$ constrains the separability between $\{\mathbf{h}^{k}_{u}\}^{K}_{k=1}$. Meanwhile, different from the restricted single-relation constraint in~\cite{DisenGCN,IPGDN}, the co-aggregation of intra-component and inter-component also allows for the existence of multiple relationships between nodes.
To make the proposed ADGCN be capable of inductive learning, we take the similar way as in~\cite{Graphsage} by randomly sampling the neighboring nodes of $u$ to get $\mathbf{N}(u)$ with fixed size.
It should be noted that the average sampling is used for unweighted graphs, and probability sampling based on edge weights is applied for weighted graphs.
The pseudocode of the component-specific node representation aggregation is summarized in \textbf{Alg}. \ref{algorithm1}.
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\begin{algorithm}[t]
\caption{Procedure of the component-specific aggregation for Micro-disentanglement}
\small
\label{algorithm1}
\begin{algorithmic}[1]
\Require $\left \{ \mathbf{x}_i \in \mathbb{R}^f\right \} $: the set of node feature vectors, $i \in \tilde{\mathbf{N}}\left ( u \right )$, projection matrices $\mathbf{W} = \{\mathbf{W}_{k}\}_{k=1}^{K}$;
\Ensure $\mathbf{h}_u$, $\mathbf{c}_u$;
\renewcommand{\algorithmicensure}{\textbf{Hyber-paramters:}}
\Ensure $K$: number of component.
$\tilde{T}$: iterations of dynamic assignment.
$\alpha, \beta$: trade-off coefficients.
\For{$i\in \tilde{\mathbf{N}}\left( u \right)$}
\For{$k=1$ to $K$}
\State $\mathbf{c}^{k}_i \leftarrow \sigma \left ( \mathbf{W}_k \mathbf{x}_i \right )$
\State $\mathbf{c}^{k}_i \leftarrow \frac{\mathbf{c}^k_i}{\left \| \mathbf{c}^k_i \right \|_{2}} $
\EndFor
\EndFor
\State initialize $\tilde{\mathbf{h}}^{k}_{u} = \mathbf{c}^k_u$
\For{iteration $\tilde{t}=1$ to $\tilde{T}$}
\For{$v \in \mathbf{N} \left( u \right)$}
\State $s^k_v(u) \leftarrow \cos(\mathbf{h}^{k}_{u}, \mathbf{c}^{k}_{v})$, $\forall k = 1, \cdots, K$.
\State Calculate $p^k_{u\to v}$ and $q^k_{v \to u}$ by Eq.$\left( \ref{p} \right)$ and Eq.$\left( \ref{q} \right)$
\EndFor
\For{$k=1$ to $K$ do}
\State $ {\mathbf{h}^{k}_{u}}= \tilde{\mathbf{h}}^{k}_{u} + \alpha \sum_{v \in \mathbf{N}(u)}q^{k}_{v\to u}\mathbf{c}^{k}_{v} + \beta \sum_{v \in \mathbf{N}(u)}p^{k}_{u \to v}\mathbf{c}^{k}_{v}$
\State $\mathbf{h}^{k}_{u} =\frac{{\mathbf{h}}^{k}_{u}}{\left \| {\mathbf{h}}^{k}_{u} \right \|_{2}} $
\State $\tilde{\mathbf{h}}^{k}_{u} \leftarrow \mathbf{h}^{k}_{u}$
\EndFor
\EndFor
\State $\mathbf{c}_{u} \leftarrow $ $\textrm{concat}(\mathbf{c}^{1}_{u},\mathbf{c}^{2}_{u},\cdots,\mathbf{c}^{K}_{u} )$
\State $\mathbf{h}_{u} \leftarrow \textrm{concat}(\mathbf{c}^{1}_{u},\mathbf{c}^{2}_{u},\cdots,\mathbf{c}^{K}_{u} )$
\end{algorithmic}
\end{algorithm}
\subsection{Adversarial Learning for Macro-Disentanglement}
\begin{figure}[!t]
\centering
\includegraphics[width=3in]{figure/macro.pdf}
\caption{The iterative adversarial training tends to progressively improves the separability among component distributions.}
\label{macro_disen}
\end{figure}
In essence, the goal of disentangled representation learning is to make a specific component be only sensitive to the variation in the corresponding factor, while not affected by the one in other factors~\cite{RDCM}.
In practice, in addition to the micro-disentanglement mentioned above, there should exist only weak correlation, or strong separability in other words, among component distributions corresponding to different factors.
To attain the above objective, we propose an adversarial regularizer to achieve macro-disentanglement at component distribution level by explicitly imposing constraints on the interdependence among components.
Specifically, we introduce a discriminator $D$ with two branches $D_{src}$ and $D_{cls}$.
The discrimination branch $D_{src}$ treats the disentangled component representation $\mathbf{h}^{k}_{u}$ as the "real" sample, and the projected representation $\mathbf{c}^{k}_{u}$ as the "fake" sample.
The auxiliary classification branch $D_{src}$ is added to incorporate factor information into the discriminator, which allows a single discriminator to control multiple factors. Thus, the discriminator $D$ not only discriminates whether the input sample is a disentangled component or not, but also distinguishes which component distribution it comes from.
Here, $\mathbf{h}^{k}_{u}$ and $\mathbf{c}^{k}_{u}$ are taken as the "real" sample and the "fake" sample for disentanglement, respectively. In order to confuse the discriminator $D$ as much as possible, $\{\mathbf{W}_{k}\}^{K}_{k=1}$ will try to project $\mathbf{x}_{u}$ onto the distributions of the disentangled components $\{\mathbf{h}^{k}_{u}\}^{K}_{k=1}$, , where $\{\mathbf{W}_{k}\}^{K}_{k=1}$ can be regarded as the generator $G$ in GAN. As shown in Fig.~\ref{macro_disen}, it means that the distributions of $\{\mathbf{c}^{k}_{u}\}^{K}_{k=1}$ will become mutually independent as training progresses, i.e., $\mathbf{c}^{k}_{u}$ will obey the distribution in the specific $k$-th component space and be independent of each other.
Therefore, similar to ACGAN~\cite{ACGAN}, the total loss for macro disentanglement is composed of the adversarial loss and the component classification loss, and given by:
\begin{equation}
\mathcal{L}_{adv}=\mathbb{E}_{\mathbf{h}}[\mathrm{log}D_{src}(\mathbf{h})]+\mathbb{E}_{\mathbf{c}}[\mathrm{log}(1-D_{src}(\mathbf{c}))]
\end{equation}
\begin{equation}
\mathcal{L}_{cls}=\mathbb{E}_{k,\mathbf{h}}[-\mathrm{log}D_{cls}(k|\mathbf{h})]+\mathbb{E}_{k,\mathbf{c}}[-\mathrm{log}(D_{cls}(k|\mathbf{c}))]
\end{equation}
In essence, $D$ tries to maximize $\mathcal{L}_{adv}$ to distinguish between $\{\mathbf{c}^{k}_{u}\}^{K}_{k=1}$ and $\{\mathbf{h}^{k}_{u}\}^{K}_{k=1}$, while $\{\mathbf{W}_{k}\}^{K}_{k=1}$ tries to confuse $D$ by minimizing $\mathcal{L}_{adv}$ to fit $\{\mathbf{c}^{k}_{u}\}^{K}_{k=1}$ to the distributions of $\{\mathbf{h}^{k}_{u}\}^{K}_{k=1}$. By this way, it can also accelerate the convergence of Algorithm~\ref{algorithm1}.
Furthermore, by using auxiliary classifier to minimize $\mathcal{L}_{cls}$, $D$ can learn to classify $\mathbf{c}^{k}_{\cdot}$ and $\mathbf{h}^{k}_{\cdot}$ to the corresponding $k$-th component space. And $\{\mathbf{W}_{k}\}^{K}_{k=1}$ also tries to minimize $\mathcal{L}_{cls}$ since $\mathbf{h}^{k}_{u}$ is derived from $\{\mathbf{c}^{k}_{i}\}_{i \in \tilde{\mathbf{N}}(u)}$, it will be beneficial of boosting the separability among component spaces and making D to be with stronger discrimination ability. Thereby, we can capture the multiple disentangled component distributions with weak correlation to each other to achieve macro-disentanglement.
Since $\mathbf{h}^{k}_{u}$ and $\{\mathbf{c}^{k}_{i}\}_{i \in \tilde{\mathbf{N}}(u)}$ are dynamically correlated, the conventional GAN training method~\cite{GAN} is not suitable for such case. Thus, an iterative adversarial training is applied to progressively improve the separability among component spaces via an EM-like formulation given by \textbf{Alg}. \ref{algorithm3}. In fact, the process of aggregating $\{\mathbf{c}_{i}\}^{n}_{i=1}$ into $\{\mathbf{h}_{i}\}^{n}_{i=1}$ is equivalent to the E-step, where $\{\mathbf{h}_{i}\}^{n}_{i=1}$ can be seen as achieving the `expectation'. Then, the separability between components is expected to be maximized by means of adversarial learning, i.e., the `maximization' of the M-step.
\subsection{Local Structure Aware Graph Refinement Based on Diversity-preserving Node Sampling}
\iffalse
\fi
\begin{algorithm}[t]
\caption{Local structure aware graph refinement in the $t$-th epoch}
\label{algorithm2}
\begin{algorithmic}[1]
\Require $\mathcal{G} = (V, E) $, $\mathbf{H}^{(t-1)}=\left[\mathbf{h}_1^{(t-1)},\cdots \mathbf{h}_n^{(t-1)}\right]$, $\mathbf{A}^{(0)}$;
\Ensure $\mathbf{A}^{(t)}$;
\renewcommand{\algorithmicensure}{\textbf{Hyber-paramters:}}
\Ensure: $m$, $K$, $\gamma$
\State Initialization: $J_s = \phi$
/* \textit{Degree-based Node Sampling with Diversity Preserving} */
\For{$u\in V$}
\State $\mathbf{d}_u \leftarrow \sum_{v}^{|V|} \mathbf{A}^{(0)}_{u,v}$
\State $p_{s}(u) \leftarrow \frac{\mathbf{d}_u}{\sum_{i=1}^{n} \mathbf{d}_{i}}$
\EndFor
\For{$1$ to $m$}
\State Take a sampling on node $u_s$ from $V$ according to $p_{s}(u_s)$
\While{$v \in \tilde{\mathbf{N}}\left (u_s \right )$}
\State Calculate $\Delta_{u_{s},v} = \| \mathbf{h}_{u_s}^{(t-1)} - \mathbf{h}_{v}^{(t-1)} \|_2 $ %
\State $p_{s}(v) \leftarrow p_{s}(v) \cdot f(\Delta_{u_{s},v}) $
\EndWhile
\State $J_s \leftarrow J_s \cup \{u_s\}$
\EndFor
/* \textit{Graph Refinement} */
\While{sample node $u_s \in J_s$}
\For{$k=1$ to $K$}
\State Select $\mathcal{k}$ nearest neighbors of $\mathbf{h}^{(t-1),k}_{u_s}$ as ${\mathbf{N}'}\left (u_s \right )$
\State $ \tilde{\mathbf{A}}^{(t-1),k}_{u_s,{\mathbf{N}'}\left (u_s \right )} \leftarrow \cos (\mathbf{h}^{(t-1),k}_{u_s}, \mathbf{h}^{(t-1),k}_{{\mathbf{N}'}\left (u_s \right )})$
\EndFor
\EndWhile
\State $\tilde{\mathbf{A}}^{(t-1)} \leftarrow \textrm{maxpooling}(\tilde{\mathbf{A}}^{(t-1),1},\tilde{\mathbf{A}}^{(t-1),2}, \cdots, \tilde{\mathbf{A}}^{(t-1),K})$
\State Obtain the refined $\mathbf{A}^{(t)}$ by Eq.(\ref{refining}).
\end{algorithmic}
\label{refinement}
\end{algorithm}
\begin{algorithm}[t]
\caption{EM-like procedure of the proposed ADGCN}
\label{algorithm3}
\begin{algorithmic}[1]
\Require
$\mathcal{G} = \{V, E\} $, $\mathbf{X}$, $A$.
\Ensure $\mathbf{H}$: the node disentangled representation matrix of the given graph
\renewcommand{\algorithmicensure}{\textbf{Hyber-paramters:}}
\Ensure: $K$, $\tilde{T}$, $T$, $m$, $\alpha$, $\beta$, $\gamma$, $\lambda$, $\eta$
\State Initialization:
$\mathbf{W}=\{\mathbf{W}_k\}_{k=1}^{K}$: projection matrix
$\theta_{D}$: parameters of the discriminator $D$.
\For{epoch $t = 1$ to $T$}
\State \textbf{E-step:}
Obtain $\left \{ \mathbf{h}_i^{(t-1)} \right \}_{i=1}^n$ and $\left \{ \mathbf{c}_i ^{(t-1)}\right \}_{i=1}^n$ through \textbf{Alg.}~\ref{algorithm1}.
Obtain $A^{(t)}$ through \textbf{Alg.}~\ref{algorithm2} for graph refinement.
\State \textbf{M-step:}
Sample the "real" components from $\left \{ \mathbf{h}_i^{(t-1)} \right \}_{i=1}^n$.
Sample the "fake" components from $\left \{ \mathbf{c}_i^{(t-1)} \right \}_{i=1}^n$.
Update $\mathbf{W}_k$,$k=1,\cdots,K$ by minimizing Eq. (\ref{loss_G}).
Update $\theta_{D}$ by minimizing Eq. (\ref{loss_D}).
\EndFor
\Return $\mathbf{H}=\left[\mathbf{h}_1^{(T-1)},\cdots,\mathbf{h}_n^{(T-1)}\right]$.
\end{algorithmic}
\end{algorithm}
\begin{figure*}[t]
\centering
\includegraphics[width=7in]{figure/graph_refine.pdf}
\caption{Illustration of the diversity-preserving graph refinement. Since the initially given adjacency matrix $\mathbf{A}$ is extremely sparse, we have marked the existing edges $\mathbf{A}_{ij}$ using a circle for more clear presentation.}
\label{graph_refine}
\end{figure*}
For the real-world graph data, the complex relationship between nodes is often represented as a hard binary link (i.e., 0 or 1). Obviously,
it is a discrete and simplified form of continuous relationship between nodes, which seriously limits the expressibility of the learned node representation. On the other hand, the node representation obtained in the embedding space can in turn be used to reveal the intrinsic relationship between nodes. To better characterize the node relationships and further facilitate the learning of node representation, an intuitive way is to refine the originally given graph structure with the embedded node representations~\cite{IDGL}. However, such global refinement of the relationships among all nodes without distinction will inevitably lead to some noisy edges, which may further confuse the training of the node representation learning model. In addition, it also has scalability problems on large graphs. To address these issues, we propose a local structure aware graph refinement as shown in Fig.~\ref{graph_refine} to progressively reveal the latent relationships of nodes, thus to achieve efficient and robust graph refinement.
The key of achieving local structure aware graph refinement is how to obtain a diversified core subset $J_s \subset V$ for local refinement, so as to maintain the global coverage of the graph as much as possible. Determinant Point Process (DPP)~\cite{DPP} is a classical algorithm for diversity-preserving sample selection.
However, the high computational cost of DPP is somewhat the opposite of what we intended. Hence, for local structure aware graph refinement, we propose an degree-based node sampling method with good diversity preserving to serve as an efficient approximation of DPP sampling on graph data.
Considering the characteristics of the graph structure, the normalized degree centrality indicating the importance of a node is used as the sampling probability $p_s(\cdot)$ for node selection, that is to say, we have $p_s(u)=\frac{\mathbf{d}_{u}}{\sum_{i=1}^{n}\mathbf{d}_{i}}$ for $u$-th node. Let $J_{s}\subset{V}$, $|J_s| = m$, denote the core set to be selected for local aware graph refinement.
To realize the diversity of the selected core set $J_s$, we propose a sequential node selection method by the means of soft-deflation. Specifically, for a node $u_s\in J_s$ that has been selected, a soft-deflation is applied to its each neighbour $v \in \tilde{\mathbf{N}}\left (u_s \right )$, i.e., we will have $p_s(v) \cdot f(\Delta_{u_{s},v} )\rightarrow{p_s(v)}$, where $f(\cdot)\in [0,1)$ is the soft-deflation function and $\Delta_{u_{s},v}$ denotes the distance between $u_s$ and its neighbor $v$.
Here, we follow the definition of $f(\cdot)$ in~\cite{DPP-landmark} and consider $f(\Delta) = \sin^2(\Delta/\tau)$, where $\tau$ is the scaling factor to keep $\Delta/\tau$ ranging in $[0, \pi/2)$.
Through the soft-deflation in sequential node selection, we can ensure that, for each node in the associated neighborhood node set $\Psi(J_s) = \bigcup_{u_s \in J_s}\tilde{\mathbf{N}}(u_s)$, it will be further selected into the core set with less probability in the subsequent node selection. By this way, it means that the finally obtained core set $J_s$ can achieve the maximum coverage of the graph to a certain extent.
Then, with the assistance of the diversity-preserving core subset $J_s$, the local structure aware graph refinement can be carried out as follows:
\begin{align}
\mathbf{A}^{(t)}_{u_s,{\mathbf{N}}\left (u_s \right )}= (1 - \gamma) \mathbf{A}^{(0)}_{u_s,{\mathbf{N}}\left (u_s \right )} + \gamma \tilde{\mathbf{A}}^{(t-1)}_{u_s, {\mathbf{N}}\left (u_s \right )}
\label{refining}
\end{align}
where $u_s \in J_s$, $\mathbf{A}^{(0)}$ represents the initial adjacency matrix and $\tilde{\mathbf{A}}^{(t-1)}$ denotes the adjacency refining matrix in the $t$-th epoch, and $\gamma$ is a trade-off coefficient.
As mentioned above, the link relationship between two nodes may come from a certain component space.
To capture the underlying topological structure of graph, we first obtain $K$ component adjacency matrices $\{\tilde{\mathbf{A}}^{(t-1),k}\}_{k=1}^{K}$ based on the associated node representations $\{\mathbf{h}^{(t-1),k}_{u}\}_{u \in V}$ in $t$-th epoch, and then an element-wise max pooling is applied to form a component sensitive adjacency matrix $\tilde{\mathbf{A}}^{(t-1)} = \mathbf{max pooling}(\tilde{\mathbf{A}}^{(t-1),1},\tilde{\mathbf{A}}^{(t-1),2}, \cdots, \tilde{\mathbf{A}}^{(t-1),K})$. Essentially, the adopted max pooling plays a role of integrating the specific structures of multiple latent components into an unified topological one, thus to boost the link characteristics of nodes.
As a matter of fact, it is not difficult to find from Eq.(\ref{refining}) that the local aware refinement guided by core set $J_s$ can spread to the overall graph, thus achieving global refinement of graph structure. Meanwhile, it can not only prevent the introduction of noisy edges in the refinement of graph, but also help to learn more robust disentangled representations progressively with the progress of training. The detail about local aware graph refinement with diversity-preserving node sampling is illustrated in \textbf{Alg.}~\ref{refinement}.
\iffalse
To capture the underlying topological structure of graph in the corresponding $k$-th component space, an adjacency matrix $\tilde{A}^{k}$ with K-NN mode is constructed based on the associated node representations $\{\mathbf{h}^{k}_{u}\}_{u \in V}$.
Given $K$ component adjacency matrices $\{\tilde{A}^{1},
\tilde{A}^{2}, \cdots, \tilde{A}^{K}\}$, the following element-wise max pooling is applied to form a component sensitive adjacency matrix $\tilde{A}^{(t-1)} = \mathbf{max}(\tilde{A}^{1},\tilde{A}^{2}, \cdots, \tilde{A}^{k})$.
Essentially, the adopted max pooling plays a role of integrating the multiple latent component specific structures into an unified topological one, thus to enhance the link characterization of nodes.
Furthermore, by combining the original adjacency matrix $A$ and the component sensitive adjacency matrix $\tilde{A}^{(t-1)}$, progressively updated by fusing with the component sensitive adjacency matrix $\tilde{A}^{(t-1)}$ as follows:
\begin{align}
A^{(t)}_{ij}= (1 - \gamma) A_{ij} + \gamma \tilde{A}_{ij}^{(t-1)}
\label{AdjUpdate}
\end{align}
where $\gamma \in (0,1)$ is a trade-off coefficient. With the above, the graph structure can be refined to $A^{(t)}$ and will participate in the training of the next epoch.
The graph refinement process described above requires calculating the similarity of all node pairs, which greatly increase complexity for both computational time and memory consumption, limiting the scalability on large graphs. More importantly, excessive modification of the original structure may be caused by noise nodes and the incomplete representations. Local structure aware graph refinement is necessary for coping with these issues mentioned above, which can progressively update the global graph structure with less computation over the process of iteration, while avoiding over-modification. The key of achieving local structure aware graph refinement is how to select a core subset of nodes to preserve global diversity for local refinement. Determinant Point Process (DPP)~\cite{DPP} is a classical algorithm for diversity-preserving sample selection. However, the high computational costs of DPP is contrary to part of the intention for node sampling. Thus, for diversity-preserving local structure aware graph refinement, we propose a degree-based diversity preserving node sampling method as an efficient approximation of DPP sampling on graph data as shown in Fig.~\ref{graph_refine}.
Considering the characteristics of the graph structure, we use the normalized degree centrality, which indicates the importance of a node, as the initialized probability of node selection. Unlike the original DPP where the size of the sample subset could not be set, we set the number of sampling iteration to $m$ and one node $v_i$ is selected with probability $p_{s}(i)$ at each iteration. Inspired by~\cite{DPP-landmark}, we perform a soft-deflation to neighbor nodes of the sampled node $v_i$ to reduce the probability of being sampled, thus ensuring the diversity of the sampling node set. Specifically, once the node $v_i$ was selected, the distances $\{\Delta_j\}_{v_j\in \tilde{\mathbf{N}}\left (v_i \right )}$ of the selected node $v_i$ to its neighbor nodes can be calculated. For the neighbor node $v_j\in \tilde{\mathbf{N}}\left (v_i \right )$, $p_{s}(j)$ is updated by a soft-deflation function $f(\Delta_j)$. Hence, for those nodes close to $v_i$, they will be selected with lower probability in the subsequent sampling process. Following the definition of $f(\cdot)$ in~\cite{DPP-landmark}, here we consider $f(\Delta_j) = \sin^2(\Delta_j/\tau)$. Therefore a core subset $J_s \subseteq V$ with fixed size $m$ is obtained. Finally, we just need to select the top $\mathbf{k}$ similar pairs for each node $v_i \in J_s$ to construct the local structure aware adjacency matrix $\tilde{A}^{k}$ of $k$-component space and then perform the remaining steps of graph refinement. As the iteration proceeds, the local structure aware graph refinement guided by the core subset will spread to the global structure, thus refining the global structure. The complete graph refinement process is shown in Algorithm.~\ref{refinement}.
\fi
\section{Optimization}
We mainly conduct experimental evaluations on node classification task, and the last layer of ADGCN is designed as a full-connected layer to output the predicted class label of nodes. Besides, ReLU is used to for activation following the transformation by the projection matrix $\mathbf{W}=\{{\mathbf{W}_{k}}\}_{k=1}^{K}$. In view of node classification tasks and macro-disentanglement adversarial regularizer, the overall loss of ADGCN is as follows:
\begin{equation}
\mathcal{L} = \mathcal{L}_{t} + \lambda (\mathcal{L}_{adv} + \eta \mathcal{L}_{cls})
\label{loss_G}
\end{equation}
\begin{equation}
\mathcal{L}_{D} = -\mathcal{L}_{adv} + \eta \mathcal{L}_{cls}
\label{loss_D}
\end{equation}
where $\mathcal{L}_{t}$ denotes the task-aware loss, $\lambda$ and $\eta$ are hyper-parameters of the adversarial regularizer. For the semi-supervised node classification task, $\mathcal{L}_{t}$ is the cross-entropy loss. The same as~\cite{DisenGCN}, we use the multi-label soft margin loss as $\mathcal{L}_{t}$ for the multi-label node classification task, which can be defined as follows:
\begin{align}
\mathcal{L}_{t} = -\frac{1}{C}\sum ^{C}_{c=1} &[\mathbf{y}_{u}(c)\cdot \log [g(\hat{\mathbf{y}}_{u}(c))] \nonumber
\\ &+(1-\mathbf{y}_{u}(c))\cdot \log[g(-\hat{\mathbf{y}}_{u}(c))]]
\end{align}
where $g(\cdot)$ denotes the $\mathbf{sigmoid}$ function, $\mathbf{y}_{u}(c)$ and $\hat{\mathbf{y}}_{u}(c)$ denote the true label and the predicted label of node $u$, respectively. And we optimize the proposed model with Adam optimizer~\cite{Adam}.
\section{Experimental Results and Analysis}
\subsection{Evaluation Setup}
\subsubsection{Datasets}
\begin{table*}[t]
\centering
\caption{Statistics of the used datasets.}
\begin{tabular}{lcrrrrrc}
\hline
Dataset & Type & \multicolumn{1}{c}{Nodes} & \multicolumn{1}{c}{Edges} & \multicolumn{1}{c}{Classes} & \multicolumn{1}{c}{Features} & \multicolumn{1}{c}{Edge Density} & Multi-label \\ \hline
Citeseer~\cite{Citeseer} & Citation network & 3327 & 4552 & 6 & 3703 & 0.0008 & No \\
Cora~\cite{Cora} & Citation network & 2708 & 5278 & 7 & 1433 & 0.0014 & No \\
Pubmed~\cite{Pubmed} & Citation network & 19717 & 44324 & 3 & 500 & 0.0002 & No \\
Coauther CS~\cite{pitfall} & Co-authorship network & 18333 & 81894 & 15 & 6805 & 0.0005 & No \\
Amazon Computers~\cite{pitfall} & Co-purchase network & 13381 & 245778 & 10 & 767 & 0.0027 & No \\
Amazon Photo~\cite{pitfall} & Co-purchase network & 7487 & 119043 & 8 & 745 & 0.0042 & No \\ \hline
Wikipedia~\cite{POS_tagger} & Word co-occurrence & 4777 & 184812 & 40 & - & 0.0161 & Yes \\
Blogcatalog~\cite{blog} & Social network & 10312 & 333983 & 39 & - & 0.0062 & Yes \\ \hline
\end{tabular}
\label{table::dataset}
\end{table*}
In the following experiment, we adopt eight real-world graph datasets and the statistics about them are summarized in Table \ref{table::dataset}.
On the evaluation of the semi-supervised node classification and clustering, three citation networks Cora~\cite{Cora}, Citeseer~\cite{Citeseer}, and Pubmed~\cite{Pubmed} are used, whose links and labels represent citations and research fields, respectively. In addition, we take three co-authorship and co-purchase graphs, i.e., Coauthor CS~\cite{pitfall}, Amazon Computers~\cite{pitfall}, and Amazon Photo~\cite{pitfall}, for performance evaluation on the task of semi-supervised node classification.
For the case of multi-label node classification, Wikipedia~\cite{node2vec} and Blogdatalog~\cite{blog} are used.
Wikipedia is a co-occurrence network whose labels represent the Part-of-Speech (POS) tags.
Blogdatalog is a social network, whose labels and links represent user interests and social relations between users respectively. Since these two datasets do not contain node features, we instead use the rows of the adjacency matrix as node features.
\subsubsection{Comparison methods}
To evaluate the performance of ADGCN, we select several baselines for comparison, especially those that have achieved the
state-of-the-art results recently. Specifically, we mainly consider the following types of baselines: (i) general baselines,including MLP, Label Propagation (LabelProp)~\cite{lp}, Planetoid~\cite{planetoid}; (ii) spectral convolution-based methods, including ChebNet~\cite{ChebNet}, GCN~\cite{GCN}; (iii) spatial convolution-based methods, including GAT~\cite{GAT}, MoNet~\cite{MoNet}, GraphSAGE~\cite{Graphsage}, APPNP~\cite{PPNP}, SGC~\cite{SGC}, AP-GCN~\cite{AP-GCN}, ARMA~\cite{ARMA}, and IDGL~\cite{IDGL}. Besides, we also compare the most related state-of-the-art works in graph representation learning, i.e., DisenGCN~\cite{DisenGCN} and IPGDN~\cite{IPGDN}. Due to the good performance on multi-label node classification, three probabilistic models DeepWalk~\cite{deepwalk}, LINE~\cite{LINE}, and node2vec~\cite{node2vec} are also used for comparison.
\subsubsection{Implementation details}
To detail the implementation, we define $\Delta d = \frac{d}{K}$ as the dimension of node components, $K^{(l)}$ as the number of components in the $l$-th layer, respectively. And $\tilde{T}$ denotes the iterations of the dynamic assignment in the component-specific aggregation layer.
In semi-supervised node classification,
we fix $K^{(l)}$ to $K=5$ for the adopted six datasets and the dimension of $\mathbf{h}_{u}$ is set to be $K \cdot \Delta d$. For multi-label node classification task, due to the large number of tags, we follow the setting in DisenGCN. Specifically, the number $K$ of the intermediate layers is set to gradually decrease to learn the hierarchical disentangled representation, and we use skip-connection to preserve the representation of different levels. As for the output dimension of the first layer, it is set to $K^{(1)} \cdot \Delta d = 128$, where $K^{(1)} \in \{4,8,\cdots,32\}$, $K^{(l)}-K^{(l+1)}=\Delta K \in \{0,2,4\}$, and $\Delta d$ is fixed to $\frac{128}{K^{(1)}}$. As hyper-parameters, both $K^{(1)}$ and $\Delta K$ are searched through hyperopt~\cite{hyperopt}. For macro-disentangled adversarial learning in the above two tasks, both $\lambda$ and $\eta$ are set to 1. In addition, we tune the other hyper-parameters of ADGCN using hyperopt. For each hyper-parameter combination, we run 200 epochs and choose the best combination on the validation set. Then we report the averaged results over 30 times of execution on each dataset.
\begin{table*}[]
\centering
\caption{ Results on citation datasets with both fixed and random splits in terms of semi-supervised node classification accuracy ($\%$).}
\begin{tabular}{lcccccc}
\hline
\multirow{2}{*}{Models} & \multicolumn{2}{c}{Cora} & \multicolumn{2}{c}{CiteSeer} & \multicolumn{2}{c}{PubMed} \\ \cline{2-7}
& Fixed & Random & Fixed & Random & Fixed & Random \\ \hline
MLP & $61.6\pm0.6$ & $59.8\pm2.4$ & $61.0\pm1.0$ & $58.8\pm2.2$ & $74.2\pm0.7$ & $70.1\pm2.4$ \\
ChebNet~\cite{ChebNet} & $80.5\pm1.1$ & $76.8\pm2.5$ & $69.6\pm1.4$ & $67.5\pm2.0$ & $78.1\pm0.6$ & $75.3\pm2.5$ \\
GCN~\cite{GCN} & $81.3\pm0.8$ & $79.1\pm1.8$ & $71.1\pm0.7$ & $68.2\pm1.6$ & $78.8\pm0.6$ & $77.1\pm2.7$ \\
GAT~\cite{GAT} & $83.1\pm0.4$ & $80.8\pm1.6$ & $70.8\pm0.5$ & $68.9\pm1.7$ & $79.1\pm0.4$ & $77.8\pm2.1$ \\
APPNP~\cite{PPNP} & $83.3\pm0.5$ & {\color{red}$\textbf{81.9}\pm1.4$} & $71.8\pm0.4$ & $69.8\pm1.7$ & $80.1\pm0.2$ & $79.5\pm2.2$ \\
SGC~\cite{SGC} & $81.7\pm0.1$ & $80.4\pm1.8$ & $71.3\pm0.2$ & $68.7\pm2.1$ & $78.9\pm0.1$ & $76.8\pm2.6$ \\
DisenGCN~\cite{DisenGCN} & $83.1\pm1.1$ & $80.3\pm2.0$ & $73.6\pm0.8$ & $70.1\pm1.8$ & $79.7\pm0.7$ & $78.7\pm1.7$ \\
IPGDN~\cite{IPGDN} & $84.1$ & - & $74.0$ & - & {\color{red}$\textbf{81.2}$} & - \\
AP-GCN~\cite{AP-GCN} & - & $79.8\pm3.2$ & - & $69.2\pm3.0$ & - & {\color{blue}$\textbf{79.2}\pm1.6$} \\ \hline
\textbf{ADGCN} & {\color{blue}$\textbf{84.3}\pm0.6$} & $80.8\pm1.7$ & {\color{blue}$\textbf{74.3}\pm0.7$} & {\color{blue}$\textbf{70.4}\pm1.5$} & {\color{red}$\textbf{81.2}\pm0.5$} & {\color{blue}$\textbf{79.2}\pm1.7$} \\
\textbf{ADGCN-R} & {\color{red}$\textbf{84.5}\pm0.7$} & {\color{blue}$\textbf{81.0}\pm2.0$} & {\color{red}$\textbf{74.5}\pm0.8$}& {\color{red}$\textbf{71.0}\pm1.4$} & {\color{blue}$\textbf{80.4}\pm0.4$} & {\color{red}$\textbf{79.9}\pm1.3$} \\ \hline
\end{tabular}
\label{citation}
\end{table*}
\begin{table}[]
\caption{ Results on the other three datasets with random splits in terms of semi-supervised node classification accuracy ($\%$).}
\begin{tabular}{lccc}
\hline
Models &
\begin{tabular}[c]{@{}c@{}}Coauthor \\ CS\end{tabular} &
\begin{tabular}[c]{@{}c@{}}Amazon \\ Computers\end{tabular} &
\begin{tabular}[c]{@{}c@{}}Amazon\\ Photo\end{tabular} \\ \hline
MLP & $88.3\pm0.7$ & $44.9\pm5.8$ & $69.6\pm3.8$ \\
LabelProp~\cite{lp} & $73.6\pm3.9$ & $70.8\pm8.1$ & $72.6\pm11.1$ \\
LabelProp NL~\cite{lp} & $76.7\pm1.4$ & $75.0\pm2.9$ & $83.9\pm2.7$ \\
GCN~\cite{GCN} & $91.1\pm0.5$ & $82.6\pm2.4$ & $91.2\pm1.2$ \\
GAT~\cite{GAT} & $90.5\pm0.6$ & $78.0\pm19.0$ & $85.7\pm20.3$ \\
MoNet~\cite{MoNet} & $90.8\pm0.6$ & $83.5\pm2.2$ & $91.2\pm1.3$ \\
GraphSAGE-mean~\cite{Graphsage} & $91.3\pm2.8$ & $82.4\pm1.8$ & $91.4\pm1.3$ \\
GraphSAGE-maxpool~\cite{Graphsage} & $85.0\pm1.1$ & - & $90.4\pm1.3$ \\
GraphSAGE-meanpool~\cite{Graphsage} & $89.6\pm0.9$ & $79.9\pm2.3$ & $90.7\pm1.6$ \\
APPNP~\cite{PPNP} & {\color{blue}$\textbf{93.2}\pm0.1$}& $80.1\pm0.3$ & $89.3\pm0.2$ \\
DisenGCN~\cite{DisenGCN} & $92.9\pm 0.2$ & $84.6\pm0.9$ & {\color{blue}$91.7\pm0.7$} \\
ARMA~\cite{ARMA} & $92.4\pm0.3$ & $80.7\pm0.3$ & $89.4\pm0.3$ \\ \hline
ADGCN & $92.9 \pm 0.3$ & {\color{blue}$\textbf{84.7}\pm1.1$} &{\color{blue}$\textbf{91.7}\pm0.4$} \\
ADGCN-R & {\color{red}$\textbf{93.3}\pm0.4$}&{\color{red}$\textbf{85.2}\pm0.7$}& {\color{red}$\textbf{92.1}\pm0.3$} \\ \hline
\end{tabular}
\label{co-network}
\end{table}
\subsection{Semi-Supervised Node Classification}
In view of semi-supervised node classification on Cora, Citeseer, and Pubmed datasets, we use the same fixed dataset split as GCN with each dataset containing only 20 labeled nodes for each class, 500 nodes for validation, and 1000 nodes for test. Meanwhile, we also conduct random split as~\cite{PPNP}, which uses 20 labeled nodes per class as the training set, 500 nodes as the validation set, and 1000 nodes as the test set for all models. For Coauthor CS, Amazon Computers, and Amazon Photo, the random split is also adopted, that is, 20 labeled nodes per class are taken as the training set, 30 labeled nodes per class nodes as the validation set, and the rest nodes as the test set.
The comparison results are shown in Table \ref{citation} and Table \ref{co-network}, respectively, in which ADGCN-R denotes ADGCN with graph refinement.
Compared with the holistic approaches such as GAT and GCN, DisenGCN, IPGDN, and ADGCN achieve better performance, which can be largely due to the disengagement analysis of the graph structure. In particular, ADGCN outperforms over DisenGCN by about $0.8\%$, $1.0\%$, and $0.9\%$ on Cora, Citeseer, and Pubmed datasets, respectively. Besides, ADGCN-R also achieves the best performance on three co-author/purchase datasets.
This just verifies that ADGCN can further boost the performance and the separability among component distributions, which benefits from the combination of macro-disentanglement and micro-disentanglement.
\subsection{Multi-Label Node Classification}
Following the experiment settings in~\cite{node2vec} for multi-label node classification, we report the performance of each method by increasing the number of nodes labeled for training from 10 $\%$ to 80 $\%$ of the total number of nodes $\left | V \right | $, and the other nodes are divided into two sets for verification and testing.
We used two indicators, Macro-F1 and Micro-F1, to evaluate the performance of each model, and the performances on the two datasets are shown in Fig.\ref{multi-classificaion}. It can be observed that our ADGCN performs the best compared to all baselines including DisenGCN, especially on Wikipedia dataset. Meanwhile, in terms of Macro-F1, graph disentanglement methods achieve higher score than other methods. It indicates that graph disentanglement can facilitate the understanding of the relationship between nodes, so as to be more robust to graph data with class imbalance.
\begin{figure*}[t]
\subfigure[Macro-F1($\%$), Wikipedia]{
\centering
\includegraphics[width=1.7in]{fig/macro_f1_wiki.png}
}%
\subfigure[Micro-F1($\%$), Wikipedia]{
\centering
\includegraphics[width=1.7in]{fig/micro_f1_wiki.png}
}%
\centering
\subfigure[Macro-F1($\%$), Blogdatalog]{
\centering
\includegraphics[width=1.7in]{fig/macro_f1_blog.png}
}%
\subfigure[Micro-F1($\%$), Blogdatalog]{
\centering
\includegraphics[width=1.7in]{fig/micro_f1_blog.png}
}%
\centering
\caption{Macro-F1 and Micro-F1 scores on the multi-label classification tasks.}
\centering
\label{multi-classificaion}
\end{figure*}
\begin{figure*}[t]
\subfigure[Original node feature $\mathbf{X}$]{
\centering
\includegraphics[width=1.5in]{comVis/feat.pdf}
}%
\subfigure[Learned representation $\mathbf{H}$]{
\centering
\includegraphics[width=1.5in]{comVis/emb.pdf}
}%
\subfigure[1st component space]{
\centering
\includegraphics[width=1.5in]{comVis/k0.pdf}
}%
\centering
\subfigure[2nd component space]{
\centering
\includegraphics[width=1.5in]{comVis/k1.pdf}
}%
\\
\subfigure[3rd component space]{
\centering
\includegraphics[width=1.5in]{comVis/k2.pdf}
}%
\subfigure[4th component space]{
\centering
\includegraphics[width=1.5in]{comVis/k3.pdf}
}%
\subfigure[5th component space]{
\centering
\includegraphics[width=1.5in]{comVis/k4.pdf}
}%
\centering
\caption{Visualization of the node representation similarity matrix on Cora dataset. (a) the similarity matrix of node feature $\mathbf{X}$, (b) the similarity matrix of learned node representations. (c), (d), (e), (f), and (g) are the similarity matrix of node representations in different component spaces. We use the red and blue box markers to show the entangled relationship of nodes in different component spaces.}
\centering
\label{component visualization}
\end{figure*}
\subsection{Clustering Analysis}
For the node clustering task, we use the K-means~\cite{Kmeans} to cluster the learned representations as in~\cite{IPGDN}. The setting of the three citation datasets is the same as in semi-supervised node classification task with fixed split.
The results of the node clustering are shown in Table~\ref{node_clustering}. It can be observed that the performance of graph disentangled methods show significant advantage over other compared methods. Moreover, our ADGCN achieves the best performance compared to DisenGCN and IPGDN.
It means that the proposed macro and micro disentanglement are greatly beneficial to exploring the inherent similarity between nodes. Meanwhile, the excellent performance of these graph disentanglement based methods also demonstrates that graph disentanglement can indeed provide benefits in terms of graph clustering and community structure detection.
\begin{table*}[t]
\caption{Node clustering results(\%).}
\centering
\begin{tabular}{clcccccccc}
\toprule
\textbf{Datasets} & \textbf{Metrics} & \textbf{DeepWalk}\cite{deepwalk} & \textbf{Planetoid}\cite{planetoid} & \textbf{ChebNet}\cite{ChebNet} & \textbf{GCN}\cite{GCN} & \textbf{GAT}\cite{GAT} & \textbf{DisenGCN}\cite{DisenGCN}& \textbf{IPGDN}\cite{IPGDN} & \textbf{ADGCN}\\
\midrule
\multirow{3}*{Cora}
& ACC & 62.2 & 67.2 & 71.9 & 73.5 & 75.2 & 75.5 & {\color{blue}76.1} & {\color{red}\textbf{81.0}} \\
& NMI & 50.3 & 52.0 & 49.8 & 51.7 & 57.0 & 58.4 & {\color{blue}59.2} & {\color{red}\textbf{62.3}} \\
& ARI &40.8 &40.5 &42.4 &48.9 &54.1 & 60.4 & {\color{blue}61.0} & {\color{red}\textbf{62.1}}\\ \hline
\multirow{3}*{Citeseer}
& ACC & 52.1 & 61.0 & 65.0 & 67.7 & 68.0 & 68.2 & {\color{blue}68.9} & {\color{red}\textbf{69.3}}\\
& NMI & 30.5 & 41.2 & 42.6 & 42.8 & 43.1 & 43.7 & {\color{blue}44.3} & {\color{red}\textbf{46.6}}\\
& ARI &20.6 &22.1 &41.5 &42.8 &43.6 & 42.5 & {\color{blue}43.0} & {\color{red}\textbf{47.6}}\\ \hline
\multirow{3}*{Pubmed}
& ACC & 65.2 & 64.6 & 75.2 & 75.6 & 76.3 & 77.0 & {\color{blue}77.8} & {\color{red}\textbf{78.4}}\\
& NMI & 29.6 & 32.5 & 35.6 & 35.0 & 35.0 & 36.1 & {\color{blue}37.0} & {\color{red}\textbf{39.2}}\\
& ARI &36.6 &33.9 &38.6 &40.9 &41.4 & 41.6 & {\color{blue}42.0} & {\color{red}\textbf{44.1}}\\
\bottomrule
\end{tabular}
\label{node_clustering}
\end{table*}
\subsection{Analysis of Disentanglement}
\begin{figure}[t]
\centering
\subfigure[DisenGCN]{
\centering
\includegraphics[width=1.5in]{comVis/GAD_DisenGCN.pdf}
}%
\subfigure[ADGCN]{
\centering
\includegraphics[width=1.5in]{comVis/GAD_our.pdf}
}%
\caption{The component confusion matrix $C$ obtained by DisenGCN and ADGCN on Cora dataset. Here, the group average similarity score is adopted to measure the correlation between different component distributions. The smaller the score, the weaker the correlation between the two component distributions.}
\label{macro_disentanglement}
\end{figure}
To intuitively show the disentanglement achieved by ADGCN, we visualize the similarity matrix of Cora dataset in both the original feature and the learned representation spaces.
In contrast to the original feature in Fig.~\ref{component visualization}(a), the similarity matrix based on the learned representation has shown explicitly block effect as we can see from
Fig.~\ref{component visualization}(b)
. It indicates that the representation learned by our model is capabel of effectively capturing the intra-class similarity and inter-class differences.
More importantly, to reveal the intrinsic entanglement between different classes of nodes, we visualize the similarity matrices of the component level, thus showing the correlations between nodes in different component spaces.
From Fig.~\ref{component visualization} (c-g), it can be observed that each component does reflect certain entangled information among different classes, and the complex entanglements between different two classes can be well unveiled by the different combinations of components.
Taking the first three classes (marked in red box) in Fig.~\ref{component visualization} as an example, class 1 and class 2 are significantly entangled but both can be well distinguished from class 3 in the 1st component space. These three classes also exhibit other different entanglement relationships in the 2nd and 3rd component spaces, respectively. Similarly, the three classes marked by the blue box also show various entanglement relations in different component spaces.
As shown in Fig.~\ref{macro_disentanglement}, we also present the component confusion matrix $C$ to illustrate the correlations among the distributions in different component spaces. Here, we use the group average similarity to compute $C$. Specifically, the correlation score between the $i$-th and $j$-th component distributions is given by $C_{ij} = \frac{1}{|V|^2}\sum_{u,v \in V}\cos(\mathbf{h}^i_u,\mathbf{h}^j_v )$. The smaller the score, the weaker the correlation between the two component distributions. It's obvious that the representations in different component spaces learned by our ADGCN have weaker correlations compared to DisenGCN~\cite{DisenGCN}, which means that ADGCN indeed achieves macro disentanglement and also validates the effectiveness of the proposed adversarial regularizer for macro-disentanglement.
\subsection{Evaluation on Progressively Refined Graph Structure}
\label{gsa}
\begin{table}[]
\centering
\caption{The semi-supervised node classification accuracies of different graph on three datasets.}
\begin{tabular}{cccc}
\hline
\multirow{2}{*}{Graph} & \multicolumn{3}{c}{Datasets} \\ \cline{2-4}
& Cora & Citeseer & Pubmed \\ \hline
$\mathbf{A}$ & 68.0 & 45.3 & 63.0 \\ \hline
$\mathbf{A}_{k-NN}$ & 74.9 & 57.5 & 69.4 \\ \hline
$\mathbf{A}^{(T)}$ & \textbf{77.3} & \textbf{63.5} & \textbf{72.6} \\ \hline
\end{tabular}
\label{three_graph}
\end{table}
As we know, the adjacency matrix reflects the structure information of graph. Therefore, to verify the effectiveness of the local aware graph refinement based on diversity preserving node sampling, the label propagation~\cite{lp} is applied to evaluate the ability of the refined adjacency matrix to characterize node relationships in graph. The performance on the two datasets is shown in Table.\ref{three_graph}, where $\mathbf{A}_{k-NN}$ and $\mathbf{A}^{(T)}$ denote the adjacency matrix constructed by the learned representation $H^{(T)}$ and the final adjacency matrix refined through progressive graph refinement, respectively. As we can see that $\mathbf{A}^{(T)}$ performs better than both $\mathbf{A}$ and $\mathbf{A}_{k-NN}$ by about $9.3\%$ and $2.4\%$, respectively, on Cora dataset. The results show that the graph refinement strategy can effectively boost the node representation and the graph structure collaboratively.
\begin{figure}[t]
\centering
\subfigure[Cora]{
\centering
\includegraphics[width=1.5in]{AblationStudy/Cora_new.pdf}
}%
\subfigure[Citeseer]{
\centering
\includegraphics[width=1.5in]{AblationStudy/Citeseer_new.pdf}
}%
\caption{Effectiveness evaluation of Micro-D, Macro-D, and graph refinement on two datasets, where w/o Micro-D and w/o Marco-D represents ADGCN without micro disentanglement and ADGCN without macro disentanglement, respectively. We use the results of semi-supervised node classification to verify the effectiveness of each part.}
\label{ablation_study}
\end{figure}
\subsection{Ablation Study}
Through the semi-supervised node classification on Cora and Citeseer datasets with random split, we validate the effectiveness of different parts in our models. As shown in the Fig.\ref{ablation_study}, the proposed macro disentanglement, micro disentanglement, and diversity-presevering graph refinement are equally important to learn latent node representation. Here, the baseline model refers to the GCN with multi-component. Specifically, both ADGCN w/o Macro-D and ADGCN w/o Micro-D achieve better performance than the baseline, verifying the usefulness of the two settings. More importantly, the performance of the combination of Macro-D and Micro-D (i.e., ADGCN) verifies that both Macro-D and Micro-D are mutually reinforcing. Besides, it's can be seen that the adoption of the progressive graph refinement has also provided a further performance improvement.
\subsection{Hyper-parameter Sensitivity Analysis}
\label{HA}
\begin{figure}[t]
\centering
\includegraphics[width=3in]{fig/hyper.png}
\caption{Hyper-parameter sensitivity on Cora dataset with fixed split.}
\label{hyper}
\end{figure}
We finally study the effect of several hyper-parameters, including the number of layers, the iterations of dynamic assignment $\tilde{T}$, and the dimension of component, on Cora dataset with fixed split.
The results in Fig.\ref{hyper} demonstrate that: (i) with the number of layers increasing, the performance of our model tends to be stable without over-smoothing; (ii) due to the application of the dynamic assignment mechanism, ADGCN can efficiently complete the aggregation in the component space and achieve better performance compared to DisenGCN.
\section{Conclusion}
In this paper, we propose an ADGCN model that takes both macro-disentanglement and micro-disentanglement into account for graph representation learning.
By micro-disentanglement, it can achieve component-specific aggregation of inter-component and intra-component through the dynamic assignment of node neighborhood.
Furthermore, we innovatively introduce the macro-disentanglement adversarial regularizer to explicitly constrain the interdependence between components.
Meanwhile, we also propose a local structure aware graph refinement based on diversity-preserving node sampling. With the training going on, the global graph structure can be progressively refined to reveal the latent relationships of nodes. Our ongoing research work will further explore the interpretability of graph disentanglement and use it for more realistic graph generation and explainable applications.
\iffalse
\appendices
\section{Proof of the First Zonklar Equation}
Appendix one text goes here.
\section{}
Appendix two text goes here.
\section*{Acknowledgment}
The authors would like to thank...
\fi
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
{
\bibliographystyle{IEEEtran}
|
\section{\label{sec:level1}First-level heading:\protect\\ The line
Stand-off magnetometry, in which most of the experimental apparatus can be remote from the place where the magnetic field is measured, provides opportunity for remote geophysical research or hazardous environment monitoring~\cite{Fu2020Sensitive}.
Stand-off magnetometry based on resonant scattering from atoms, in particular, laser-guide-star (LGS)-based magnetometry~\cite{Higbie2011,Kane2018,Pedreros2018,Fan2019remote,Pedreros2018b}, where fluorescence-collection efficiency for a typical telescope is less than $10^{-10}$ or even lower, is often limited by the scarce number of detected photons. The problem could be solved if the excited atoms were forced to radiate towards the source in the form of laser-like emission. This is, indeed, possible using the technique of mirrorless lasing, a method to generate backward directional emission with the help of so called backward amplified spontaneous emission (ASE)~\cite{Thompson2014Pulsed,Akulshin2018}, which was demonstrated with atomic vapors confined to vapor cells. The question of the possibility to ``scale'' the laboratory results to on-sky experiments remains open and is a subject of ongoing investigation~\cite{Akulshin2018,hickson2020amplified}.
To apply mirrorless lasing to stand-off magnetometry, the central question is how to extract magnetic information from the backward emission.
It was previously shown that the threshold of mirrorless lasing and intensity of directional radiation are sensitive to the magnetic field in the interaction region~\cite{Akulshin2018}. However, since the intensity of directional radiation is also a complex nonlinear function of many experimental parameters such as laser-beam intensity, laser frequency, atomic vapor density, shape and dimensions of the excitation region, etc., determining the strength of magnetic fields from the intensity of directional light requires detailed calibration of this dependence that is practically impossible in the case of remote detection.
In this work, we suggested and implemented a technique of mirrorless-lasing-based stand-off magnetometry, which is absolute and calibration-free, and conducted successful proof-of-principle laboratory experiments with a sodium vapor cell.
The method is based on the principle of "free" evolution of the ground state atomic spins under the influence of the magnetic field~\cite{Budker2007Optical}.
To implement this, we perform the measurement in three stages: 1. Optical pumping of the ground-state spin polarization; 2. Free evolution of the atomic spins ``in the dark''; 3. Read out via mirrorless lasing, the intensity of which is related to the ground-state spin polarization.
As the mirrorless lasing process requires high-intensity light to excite the atoms and generate population inversion between specific excited states, the ground-state spin polarization is strongly affected by this light. To reconstruct the full evolution of spin polarization, we repeat this measurement and combine the results with different evolution times. The resulting spin-precession signal oscillating at the Larmor frequency enables scalar magnetometry. Considering the fact that the returned emission does not require a retroreflector \cite{patton2012remotely}, our method is promising for adjustment-free, long-baseline remote magnetometry.
\begin{figure*}
\centering
\includegraphics[width=2\columnwidth]{Fig1.eps}
\caption{\textbf{Schematic of stand-off magnetometer.}
\textbf{(a)} Energy levels of sodium atoms and excitation scheme. Pump beam is a left circularly polarized 589 nm laser exciting the 3$S_{1/2}$ to 3$P_{3/2}$ transition, which generates atomic spin polarization in the 3$S_{1/2}$ state; the probe beam is a combination of right circularly polarized 589 nm and 569 nm lights, which excites atoms to 4$D_{5/2}$ and therefore produces population inversion between 4$P_{3/2}$ and 4$S_{1/2}$, thus enabling gain for 2.21 $\mu$m spontaneous emission propagating backwards and generating a laser-like emission, or mirrorless lasing.
\textbf{(b)} Experimental setup. EOM, electro-optic modulator driven by 1.713 GHz signal to generate a sideband for hyperfine repumping; AOM1, acousto-optic modulator
used to pulse the pump beam; AOM2, acousto-optic modulator
used to pulse the probe beam; BE, beam expander; $\lambda/2$, half-wave plate; $\lambda/4$, quater-wave plate; PBS, polarizing beam splitter; BS, beam splitter; L, a convex lens; GP, glass plate for reflecting the mirrorless-lasing emission; GM, gold coated mirror; PD, photodiode; DSO, digital storage oscilloscope; PC, personal computer.
\textbf{(c)} An example of the timing diagram. The figures at the top and the middle are the driving signals of AOM1 and AOM2 respectively, and the figure on the bottom represents the light pulses sent into the vapor cell.
}\label{Fig_1}
\end{figure*}
The experimental arrangement is shown in Fig.~\ref{Fig_1}. Light from a Coherent CR-699 dye laser tuned to the sodium D$_2$-line ($3\text{S}_{1/2} \rightarrow 3\text{P}_{3/2}$ transition), see Fig.~\ref{Fig_1}(a), passes through an electro-optic modulator (EOM), see Fig.~\ref{Fig_1}(b), to produce sidebands used for hyperfine repumping. The beam then passes through an acousto-optic modulator (AOM1) that can switch the light power between the zeroth and first diffraction orders. The light from the first diffraction order is steered into a buffer-gas-free sodium vapor cell (3~cm diameter and 10~cm length) to generate atomic spin polarization in the ground state. The zeroth diffraction order beam is combined with the light from another dye laser (Coherent CR-899) tuned to 569 nm.
The combined beam is passed through a second AOM (AOM2) which is used a switch to control the optical excitation of the polarized ground state at a specific time. The probe beam excites ground-state atoms up to the $4\text{D}_{5/2}$ state enabling mirrorless lasing from population inverted transitions. Mirrorless-lasing emission at 2.21 $\mu$m is detected in the backward direction~\cite{Akulshin2018} with an InGaAs photodetector (Thorlabs PDA10D2, 25 MHz bandwidth) placed 1.5~m away from the center of the vapor cell.
The timing diagram is depicted in Fig.~\ref{Fig_1}(c). The driving signals to AOM1 and AOM2 are pulsed at the same frequency. Each period starts with a pulse of several $\mu$s duration sent to AOM1, which generates a pump-beam pulse to polarize the sodium atoms. After the end of the pump-beam pulse, the spin polarization evolves freely. After some time, a shorter pulse is sent to AOM2, generating a probe-beam pulse and a subsequent mirrorless-lasing emission. As the mirrorless lasing signal is a pulse train repeated at the modulation frequency, we accumulate the signal in a digital storage oscilloscope and then analyze the data to extract the pulse amplitude, which reflects the state of the spin polarization.
\begin{figure*}
\centering
\includegraphics[width=2\columnwidth]{Fig2-V10.eps}
\caption{
\textbf{(a)} Free precession signal with 42.5~${\mu}$T bias magnetic field applied along the $x$ direction. \textbf{(b)} Blue points: amplitude spectrum of the free precession signal shown in (a); red line: a Lorentz fitting. \textbf{(c)} Blue points: measured magnetic fields at different $B_x$ applied, which are derived from the central frequencies of the Lorentz fitting, as shown in (b); red line: a fitting of the measured magnetic fields, which suggests a 1.0(0.4)~$\mu$T residual magnetic field along -$x$ direction and a 8.7(1.2)~$\mu$T residual magnetic field in the $y-o-z$ plane in the magnetic shield; gray dashed line: a line that assumes no residual magnetic field in the magnetic shield, so that the total magnetic strength $|B|$ equals to the applied $|B_x|$.
}\label{Fig_2}
\end{figure*}
The energy levels of sodium atoms and the excitation scheme are shown in Fig.~\ref{Fig_1}(a). The pump beam at 589 nm is resonant with the sodium D2 transition (from $2\text{S}_{1/2},\ F=2$ to $3\text{P}_{3/2},\ F'=1,2,3$). It is left circularly polarized such that it polarizes the sodium ground state spin. Probing with mirroless lasing is similar to that in Ref.~\cite{Akulshin2018}, in which the atoms in the $3\text{S}_{1/2}$ ground state were pumped into the $3\text{P}_{3/2}$ state with right circularly polarized 589~nm light and some of them were further pumped to the $4\text{D}_{5/2}$ state with right circularly polarized 569~nm light. Atoms in the $4\text{D}_{5/2}$ state decay into $4\text{P}_{3/2}$ which results in population inversion between $4\text{P}_{3/2}$ and $4\text{S}_{1/2}$. Such population inversion provides gain for 2.21
~$\mu$m spontaneous emission which propagates back to the photodetector. It is important for our technique that the intensity of the backward emission depends on the polarization of the ground state.
A typical spin-evolution signal detected with mirrorless lasing, Fig.~\ref{Fig_2}(a), shows an oscillating decay pattern. For this measurement, we have a modulation period of 20~$\mu$s, a pump pulse width of 5~$\mu$s and a probe pulse width of 0.4~$\mu$s. The peak power of the pump beam is 173~mW, and the peak power of the 589~nm and 569~nm components of the probe beam is 48~mW and 49~mW, respectively. The repumping sideband of 589 nm laser takes about 20$\%$ of the total power. The waist dimensions of the pump and probe beams at the center of the sodium cell are about 1.5$\times$3~mm$^2$ and 1$\times$1~mm$^2$, respectively. The vapor cell is heated to $\sim$169$^\circ$C. We scan the evolution time of the spin polarization from 0~$\mu$s to 14~$\mu$s with a 0.1~$\mu$s step.
The oscillating frequency of the evolution signal is directly determined by the Larmor frequency of sodium ground state and the decay rate is determined by the transit time of sodium atoms across the pump beam.
The signal starts from a lowest point. This is because the pump and probe beams are of opposite polarizations and the atoms are prepared in an unfavored state for mirrorless lasing during the pump stage. Then, as the spin polarization precess, the mirrorless lasing signal increases and reaches a peak when the polarization precesses for about 180$^\circ$.
After that, the oscillating amplitude decays as the optically pumped atoms escape from the pump region and the averaged polarization decreases. When the evolution time increases to about 12~$\mu$s, the signal reaches a steady state, corresponding to no remaining polarization before detection.
The amplitude spectrum of this evolution signal is shown in Fig.\,\ref{Fig_2}(b), in which there is a peak at 0.30\,MHz, corresponding to the Larmor frequency at 43~$\mu$T (the gyromagnetic ratio of $2\text{S}_{1/2}\ F=2$ level is $\approx$7\,Hz/nT). The amplitude spectrum deviates from the Lorentz fitting for frequency higher than 0.6 MHz, which may be due to the contributions from high-order harmonics, as there may be some nonlinear dependence of the mirrorless lasing signal on the atomic polarization.
To test the magnetic-field dependence of this peak frequency, we measure the evolution signal at different magnetic fields $B_x$ and use the peak frequency from the Lorentz fitting to estimate the magnetic field strength, with the results shown in Fig.\,\ref{Fig_2}(c). According to a fitting represented by the red line, there is a 1.0(0.4)~$\mu$T residual magnetic field along the -$x$ direction and a 8.7(1.2) ~$\mu$T residual magnetic field in the $y-o-z$ plane in the magnetic shield. Such magnetic-field dependence of this peak frequencies suggests that this method could be used for magnetometry.
In conclusion, we have demonstrated scalar stand-off magnetometry based on 2.21~$\mu$m mirrorless lasing with a buffer-gas-free sodium vapor cell. By operating in three stages as optical pumping, dark evolution and mirrorless lasing detection, we observed free-precession signals of sodium ground-state polarization. The complex signal may have some vector information in it, which can be investigated in the future. Compared with conventional fluorescence detection, the detection of backward directional emission in this work is expected to dramatically improve the efficiency of light collection and thus the magnetometry sensitivity. Vapor cells filled with buffer gas such as He or Ne are expected to exhibit much longer coherence times as the atoms would spend more time in the light-beam area, which is beneficial to high-sensitivity magnetometry. Our method can potentially also be extended to other mirrorless lasing schemes or other atoms. Ongoing work is exploring the feasibility of such schemes for remote geophysical magnetometry~\cite{Pedreros2018,Pedreros2018b}.
\section*{Acknowledgement}
The authors are grateful to Aram Papoyan for helpful discussions, as well as to the European Southern Observatory for the loan of a dye-laser system. This work was supported in part by the DFG Project ID 390831469: EXC 2118 (PRISMA+ Cluster of Excellence). DB also received support from the European Research Council (ERC) under the European Union Horizon 2020 Research and Innovation Program (grant agreement No. 695405) and from the DFG Reinhart Koselleck Project. RZ acknowledges support from the China
Scholarship Council (CSC) enabling his research at the
Helmholtz-Institut Mainz.
FPB received support from the European Union's Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No 893150.
|
\section{Introduction}
The fundamental theorem of calculus states that
$\int_a^b \partial_x f(x)\,dx = f(b) - f(a)$, where $f$ is some smooth function defined on the compact interval $[a,b]$. This theorem is critical in many applications, including computational sciences \cite{israel-theorem,thomson-theorem}. For example, if the function $f$ is strongly oscillatory, a numerical quadrature on the left-hand side would require many points and much computation to obtain accurate results. Nevertheless, the fundamental theorem guarantees that the positive and negative derivatives of this oscillatory function largely cancel each other out. Indeed, one can simply compute the right-hand side directly, without truncation error.
As a generalization of the fundamental theorem, consider the integral of $\partial_x f$ over the same domain under some Lebesgue measure $m$, which is an antiderivative of the density function $\rho$ (a.k.a. the Radon-Nikodym derivative \cite{nagy-density}), i.e., $dm(x) = \rho(x)\,dx$. In the classical version of the theorem, as mentioned in the first paragraph, the density $\rho$ is constant and equals $1/(b-a)$ everywhere on the domain. If this is not the case, however, the integration by parts of $\partial_x f$ involves the derivative of $\rho$,
\begin{equation}
\label{eqn:intro1}
\int_a^b \partial_x f(x)\,dm(x) = f\rho\Big|_a^b - \int_a^b f(x)\,\partial_x\rho(x)\,dx = f\rho\Big|_a^b - \int_a^b f(x)\,\frac{\partial_x\rho}{\rho}(x)\,dm(x)
\end{equation}
The integral in Eq. \ref{eqn:intro1} can be approximated using a Monte Carlo integration scheme if a set of realizations of $x$, $\{x^{1},x^{2},...,x^{N}\}$, distributed according to $m$, is given. However, if $f$ is a strongly-oscillatory function with large magnitude, the Monte Carlo method applied directly to the integral on the left-hand side (LHS) of Eq. \ref{eqn:intro1} would require a large amount of data to obtain an approximation with a reasonably small error \cite{olver-oscillatory,makri-oscillatory}. Alternatively, one can consider the right-hand side (RHS) of the same equation, which requires the function $f$ itself, not its derivative. Assuming the density $\rho$ is a well-behaved function, the variance of the integrand on the RHS is significantly smaller and, therefore, remarkably less data is needed to obtain an accurate result. However, extra computational effort must be put to evaluate $\partial_x\rho/\rho = \partial_x\log\rho$. The computation of that function, which we denote by $g$ and call it the {\em density gradient}, is the main focus of this paper.
Lebesgue integrals involving functions with high fluctuations are critical in the field of sensitivity analysis of chaotic dynamical systems. Ruelle \cite{ruelle-original,ruelle-corrections} derived a closed-form expression, known as the {\em linear response} formula, for the parametric derivative of the mean of a quantity of interest $J$. The linear response formula includes Lebesgue integrals of directional derivatives of a strongly oscillatory $J$ over the manifold of a chaotic system. A regularized version of Ruelle's formula, known as the space-split sensitivity (S3), was obtained through the integration by parts of the original formulation \cite{chandramoorthy-s3}. The S3 algorithm was successfully applied in various low-dimensional systems in the computation \cite{sliwiak-1d} and assessment of existence \cite{sliwiak-differentiability} of parametric derivatives of statistical quantities describing chaos. The crux of the computation of the regularized Ruelle's formula is the {\em SRB density gradient}, defined as a directional derivative of the logarithm of the SRB density \cite{young-srb,crimmins-srb} along the unstable manifold. While an efficient numerical procedure for the approximation of the SRB density gradient specialized to systems with one-dimensional unstable manifolds is available \cite{chandramoorthy-s3,sliwiak-1d,sliwiak-differentiability, chandramoorthy-clv}, we still lack a generalizable algorithm applicable to arbitrary higher-dimensional chaotic systems.
The main purpose of this work is to derive a general formula for the density gradient $g$, defined on a differentiable $m$-dimensional manifold $M$ immersed in the Euclidean space $\mathbb{R}^n$, $m\leq n$. In our analysis, we parameterize $M$ using the chart $x(\xi):\mathbb{R}^m\to \mathbb{R}^n$. Here, the $g$ function is an $m$-element vector, where the $i$-th component equals a directional derivative of $\log\rho$, in the direction of a unit vector $s_{i}$, i.e., $g_{i} = \partial_{s_i}\rho/\rho = (\nabla_x\rho\cdot s_i)/\rho$. The scalar function $\rho$ is the density implied by $x(\xi)$. Without loss of generality, we assume that the $i$-th directional derivative is computed along the isoparametric line in the direction of increasing $i$-th component of $\xi$. Analogously to Eq. \ref{eqn:intro1}, the Lebesgue integral of the directional derivative of $J$ over $M$ with measure $m$ can be written using $g_i$,
\begin{equation}
\label{eqn:intro2}
\int_{M} \nabla_{x}J(x)\cdot s_{i}(x)\,dm(x) = - \int_{M} J(x)\,g_{i}(x)\,dm(x),
\end{equation}
where $J$ is assumed to vanish on the boundary of $M$. For the reasons indicated above, it is computationally efficient to apply the Monte Carlo method to the RHS of Eq. \ref{eqn:intro2}. Analogous integration by parts is required to regularize the linear response \cite{chandramoorthy-s3}. Thus, the derivation of a computable expression for the density gradient defined on higher-dimensional smooth manifolds is a milestone in constructing algorithms for differentiating SRB measures. In addition, an explicit formula for $g$ might serve as a valuable tool in general numerical procedures involving integrals over geometrically complex domains.
The structure of this paper is the following. First, in Section \ref{sec:1D}, we derive a computable expression for the density gradient defined on one-dimensional manifolds (straight lines and curves). We also demonstrate a numerical example of Monte Carlo integration of a highly-oscillatory function, and show the advantage of using the density gradient in computing integrals of this type. In Section \ref{sec:general}, we extend all the concepts introduced in Section \ref{sec:1D} to higher-dimensional manifolds. Section \ref{sec:recursion} focuses on a recursive algorithm for the density gradient defined on a sequence of evolving manifolds under a differentiable map $\varphi$. Sections \ref{sec:1D}-\ref{sec:recursion} include examples of $x(\xi)$ defined by popular dynamical systems, as well as numerical results validating the derived expressions. Finally, Section \ref{sec:conclusion} concludes the paper.
\section{Computing $g$ on one-dimensional manifolds}\label{sec:1D}
In this section, we focus on the computation of the density gradient $g$ in the simplest topological setting. In particular, we consider one-dimensional manifolds, which can be described using a single parameter $\xi\in[0,1]$. That manifold is a curve $\mathcal{C}$ immersed in the Euclidean $\mathbb{R}^n$ space. We assume there exists a one-to-one map $x(\xi)\in\mathcal{C}\subset \mathbb{R}^n$, which is at least twice differentiable with respect to $\xi$, i.e., $x(\xi)\in C^2[0,1]$. In this case, the density gradient function is a scalar quantity defined as a directional derivative along $\mathcal{C}$ of logarithmic density, $g = \partial_s\log\rho$, where $\rho:\mathcal{C}\to[0,1]$ is a density function implied by $x(\xi)$. If we think of $\xi$ as a realization of the random variable uniformly distributed in $[0,1]$, then $x(\xi)$ is in fact the inverse cumulative distribution function (inverse CDF, a.k.a. the quantile function). Intuitively, $x(\xi)$ tells us that $100\xi\;\%$ of all points mapped from the uniformly distributed set are located on the curve segment between $x(0)$ and $x(\xi)$. On the other hand, the density function $\rho$ indicates the density of points mapped on $\mathcal{C}$ per unit curve length. Therefore, $\rho$ is counter-proportional to the magnitude of the first derivative of $x(\xi)$.
In the following three subsections, we analytically derive the expression for $g$ in terms of the inverse CDF $x(\xi)$ for simple line manifolds, $n=1$ (Section \ref{sec:lines}), and general curves, $n\geq 1$ (Section \ref{sec:curves}), and demonstrate its importance in a numerical integration experiment (Section \ref{sec:integral}). We illustrate all relevant concepts using a certain $x(\xi)$ associated with the Van der Pol equation,
\begin{equation}
\label{eqn:van-der-pol}
\frac{d^2u}{dt^2} = 2(1-u^2)\frac{du}{dt} - u, \;\;u(0) = -a,\;\;\frac{du}{dt}(0) = 0,
\end{equation}
which describes the coordinates of a 2D non-conservative oscillator with non-linear dumping \cite{ginoux-vanderpol}. In our numerical examples, we choose $a = 2.0199$, in which case the solution $[u(t), du/dt(t)]^T$ approximately lies on the limit cycle with period $T = 2T_{1/2}\approx 7.638$ and $u(t)\in[-a,a]$ for all $t\geq 0$. Figure \ref{fig:van-der-pol} illustrates the limit cycle of Eq. \ref{eqn:van-der-pol}, which has been computed using the second-order Runge-Kutta (midpoint) method with time step $\Delta t = 0.0001$.
\begin{figure}
\centering
\includegraphics[width=1\textwidth]{figures/vdp_cycle.png}
\caption{Trajectory of the Van der Pol oscillator (Eq. \ref{eqn:van-der-pol}). The red dot represents the initial condition, as well as the solution after time $T$, while the blue dot indicates the solution after time $T_{1/2}$. The vertical dashed lines correspond to $u = -a$ and $u = a$ (boundaries of the range of $u$). At the green dots, the solution satisfies $du/dt+d^3u/dt^3 = 0$, while the zero acceleration state, $d^2u/dt^2=0$, is represented by orange dots.}
\label{fig:van-der-pol}
\end{figure}
\subsection{Lines: $\mathcal{C}\subset\mathbb{R}$}\label{sec:lines}
We start from the simplest case, i.e., when $\mathcal{C}$ is a bounded line segment in $\mathbb{R}$, between $a$ and $b$. The corresponding inverse CDF $x(\xi)$ differentiably maps $[0,1]$ to $[a,b]$ and is related to the density function by the following expression,
\begin{equation}
\label{eqn:line1}
\xi(x) = \int_{a}^{x(\xi)} \rho(y)\;dy \;\;\;\forall x \in [a,b].
\end{equation}
Since $\xi\in[0,1]$, $\rho(x)$ is in fact the probability density function (PDF) corresponding to the CDF $\xi(x)$, which satisfies $d\xi=\rho(x)\; dx$. Using the inverse function theorem, which asserts $f'(f^{-1}(c)) = 1/[(f^{-1})'(c)]$ for any differentiable one-to-one function $f$ at any $c$ such that $(f^{-1})'(c)\neq0$, we conclude that
\begin{equation}
\label{eqn:line2}
\frac{dx}{d\xi}(\xi)\;\rho(x(\xi)) = 1.
\end{equation}
Eq. \ref{eqn:line2} indicates that at any point $x(\xi)$ on the manifold, the product of the PDF and derivative of the inverse CDF is constant. Thus, by differentiating Eq. \ref{eqn:line2} with respect to $\xi$ and reshuffling terms, we obtain a direct expression for $g$ at each point on the manifold,
\begin{equation}
\label{eqn:line3}
g(x(\xi)) = \partial_{x}\log\rho(x(\xi)) = \frac{\partial_x\rho(x(\xi))}{\rho(x(\xi))} = -\frac{\frac{d^2 x}{d\xi^2}(\xi)}{\Big(\frac{d x}{d\xi}(\xi)\Big)^2}.
\end{equation}
To illustrate these functions and their relation, we will consider the solution to Eq. \ref{eqn:van-der-pol}, $u(t)$, for $t\in[0,T_{1/2}]$, where $T_{1/2} \approx 3.819$. Based on Figure \ref{fig:van-der-pol}, it is evident that $u(t)$ is a one-to-one smooth function and $du/dt \geq 0$ in that time interval. In fact, we can apply the linear transformation $t\to \xi$ to notice that
\begin{equation}
\label{eqn:line4}
x(\xi) = u\left(\xi T_{1/2}\right)
\end{equation}
is a representation of the inverse CDF. Next, we compute the first and second derivative of Eq. \ref{eqn:line4} with respect to $\xi$ and plug them to Eq. \ref{eqn:line3} to obtain the following formula for $g$ along the trajectory,
\begin{equation}
\label{eqn:line5}
g(u(t)) = -\frac{\frac{d^2u}{dt}(t)}{\left(\frac{du}{dt}(t)\right)^2} \stackrel{\text{Eq. \ref{eqn:van-der-pol}}}{=} -\frac{2(1-u^2(t))\frac{du}{dt}(t) - u(t)}{\left(\frac{du}{dt}(t)\right)^2}.
\end{equation}
We observe that the density gradient is invariant to any linear change of variables, i.e., when $d\xi/dt$ is constant. Given a numerical solution to Eq. \ref{eqn:van-der-pol}, the density can be directly computed from $\rho(u(t)) = (T_{1/2}\; du/dt (u(t)))^{-1}$, which follows from Eq. \ref{eqn:line2}, whereas the density gradient function can be evaluated using Eq. \ref{eqn:line5}.
Figure \ref{fig:line} illustrates the inverse CDF $x(\xi)$, defined by Eq. \ref{eqn:line4}, as well as the corresponding density and density gradient. We clearly observe that both $\rho$ and $g$ are undefined at the endpoints, i.e., at $\xi = 0$ and $\xi = 1$, which is a consequence of zero slope of $x(\xi)$. Moreover, the larger the rate of change of $x$, the smaller the value of $\rho$, which confirms our previous intuitive explanation of the density function. We also notice that the density gradient is zero at the point corresponding to a local extremum of $\rho$ and the inflection point of $x(\xi)$.
\begin{figure}
\includegraphics[width=1.\textwidth]{figures/vdp_half_xrg.png}
\caption{The inverse CDF function $x(\xi)$ defined by the solution to the Van der Pol equation, such that $x(\xi(t)) = u(t)$ for all $t\in [0, T_{1/2}]$ (red), and the corresponding density (blue) and density gradient function (green). We used data presented in Figure \ref{fig:van-der-pol} to compute all the three functions.}
\label{fig:line}
\end{figure}
\subsection{Approximating integrals of a highly oscillatory function}\label{sec:integral}
We now demonstrate the use of the density gradient function in the numerical computation of a highly oscillatory function. Consider the following Lebesgue integral,
\begin{equation}
\label{eqn:line-integral}
I = \int_{-a}^{a} \partial_xf(x)\;d\xi(x),
\end{equation}
where $\xi(x)$ denotes a Lebesgue measure defined by Eq. \ref{eqn:line1}, while $f$ is a function whose first derivative is integrable and bounded. Certainly, it is assumed the above integral converges. Indeed, a sufficient condition for the convergence of $I$ in this case is Lebesgue-integrability of the density gradient with respect to the density $\rho$ \cite{sliwiak-differentiability}, i.e., $g\in L^1(\rho)$. However, the necessary and sufficient condition imposes extra requirements for the $f$ function itself, i.e., $\partial_x f\in L^1(\rho)$ or, equivalently, $\partial_x f\,\rho \in L^1[-a,a]$. In our experiment, the function $f$ has the following form,
\begin{equation}
\label{eqn:line-function}
f(x) = \left((x-a)(x+a)\sin(Kx^2)\right)^2,
\end{equation}
with some positive number $K$. We use Eq. \ref{eqn:line1} to rewrite the above integral, and then integrate it by parts. There exist a few scenarios when the resulting boundary term vanishes. One option is that the product $\partial_x f\,\rho$ is periodic and integrable on $[-a,a]$. Another possibility is when both $\partial_x f$ and $\rho$ are bounded and at least one of them vanishes at the domain boundaries. In any case, two new versions of $I$, alternative to the original form (in Eq. \ref{eqn:line-integral}), are available,
\begin{equation}
\label{eqn:line-integral2}
\int_{-a}^{a} \partial_xf(x)\;\rho(x)\;dx = I =
- \int_{-a}^{a} f(x)\;g(x)\;d\xi(x).
\end{equation}
To numerically approximate the integral $I$, we apply three distinct approaches. The integral in Eq. \ref{eqn:line-integral} and the RHS of Eq. \ref{eqn:line-integral2} can be estimated using a Monte Carlo method, which requires generating a random sequence $\{x^1,x^2,...,x^N\}$ distributed according to the measure $\xi$. If such a sequence is available, then the integral of any Lebesgue-integrable function $h(x)$ can be approximated as follows,
\begin{equation}
\label{eqn:line-integral3}
\int_{-a}^{a} h(x)\;d\xi(x) \approx \frac{1}{N}\sum_{i=1}^N h(x^{i}),
\end{equation}
since $\xi\in[0,1]$. Finally, the integral on the LHS of Eq. \ref{eqn:line-integral2} is evaluated using a standard trapezoidal rule with a uniform $N$-element grid for $x$ between $-a$ and $a$. In the calculation, we allot
\begin{equation}
\label{eqn:line-integral4}
x^{i}=u\left(\frac{i-1}{N-1}T_{1/2}\right).
\end{equation}
It can be numerically verified that for this particular choice of the sequence, $g\notin L^1(\rho)$, but $\partial_x f\in L^1(\rho)$. It means means that the integral $I$ converges despite the blow-up of $\rho$ and $g$ at the boundaries of $[-a,a]$. To assess the Lebesgue-integrability of these functions, we applied the procedure described in Section 4 of \cite{sliwiak-differentiability}. This algorithm approximates the slope of the distribution tail of any function in the logarithmic scale.
In order to compare the performance of these three integration methods, we proceed as follows. First, we generate the sequence $\{x^1,x^2,...,x^{N}\}$, $N = 10^5$ (time step is chosen such that $\Delta t = T_{1/2}/(N-1)$) and, using Eq. \ref{eqn:line2} and Eq. \ref{eqn:line5}, we directly evaluate $\rho$ and $g$ at all points from that sequence. Subsequently, both the density and density gradient functions are linearly interpolated everywhere between $-a$ and $a$. We use these interpolators to approximate the two functions at any point of a uniform grid (trapezoidal rule) or sequence defined by Eq. \ref{eqn:line-integral4} (Monte Carlo) for an arbitrary value of $N$. If $K$ is sufficiently small, then the approximation error of the trapezoidal rule is expected to be upperbounded by $\mathcal{O}(1/N)$, because the integrand, $\partial_x f\,\rho$, is Lebesgue-integrable \cite{cruz-traprule}. According to the Nyquist-Shannon sampling theorem, however, the discrete representation of the integrand may not be captured properly if $K$ is very large, in which case the trapezoidal rule's error decays as in a typical Monte Carlo method. Figure \ref{fig:convergence} shows the behavior of the relative error of the approximation of $I$ obtained using these three methods. The error is computed with respect to the reference solution obtained through the trapezoidal rule using $N=10^8$ points.
\begin{figure}
\includegraphics[width=0.49\textwidth]{figures/convergence_10.png}
\includegraphics[width=0.49\textwidth]{figures/convergence_100000.png}
\caption{Relative error of the approximation of $I$ for $K = 10$ (left) and $K = 100000$ (right) obtained using three methods: Monte Carlo integration applied to Eq. \ref{eqn:line-integral} (blue), Monte Carlo integration applied to the RHS of Eq. \ref{eqn:line-integral2} (red), and trapezoidal rule applied to the LHS of Eq. \ref{eqn:line-integral2} (green). Black and orange dashed lines are reference lines representing functions proportional to $N^{-1/2}$ and $N^{-1}$, respectively. In each of these plots, we computed the relative error with respect to the approximation of $I$ obtained using the trapezoidal with $N = 10^8$ samples.}
\label{fig:convergence}
\end{figure}
We observe that for a moderately-oscillatory integrand ($K = 10$), the relative error of the trapezoidal rule (green curve) decays as $\mathcal{O}(1/N)$, which confirms the theoretical estimates. In this case, the performance of both the Monte Carlo approximations (blue and red curves) does not differ much from the trapezoidal rule's. The Monte Carlo approximation clearly converges to a solution slightly different than the reference solution, which is a consequence of the fact the latter was generated using the trapezoidal rule for a linearly interpolated function. This example indicates that there is no reason to perform integration by parts and compute $g$ to approximate integrals of low- or moderately-oscillatory functions. The right-hand side plot of Figure \ref{fig:convergence} corresponds to a different scenario, i.e., when $f$ is highly-oscillatory ($K=10^5$). The error now decays $\mathcal{O}(1/\sqrt{N})$ at $N\in[10^1,5\cdot10^5]$, regardless of the integration method. The trapezoidal rule requires almost $10^6$ samples to guarantee satisfactory accuracy. Note $\partial_x f$ has a magnitude proportional to $K$, and thus the variance of the sequence $\{\partial_x f(x^1),\partial_x f(x^2),...,\partial_x f(x^N)\}$ is of the order of $K^2$. Therefore, the Monte Carlo approach applied to Eq. \ref{eqn:line-integral} requires $\mathcal{O}(10^{10})$ samples to secure error of the order of 1. A similar error can be achieved if we perform integration by parts and compute $g$ and generate only $\mathcal{O}(1)$ samples, since the variance is reduced $10^{10}$ times.
In conclusion, the computational cost of the Monte Carlo method can be dramatically reduced using the generalized fundamental calculus theorem. In case of the $f$ function, the regularization of the integral in Eq. \ref{eqn:line-integral} may decrease the cost even $K^2$ times. This result is significant specifically in the context of strongly fluctuating functions.
\subsection{General curves: $\mathcal{C}\subset \mathbb{R}^{n}$}\label{sec:curves}
We extend the concepts introduced in Section \ref{sec:lines} to the case in which $x(\xi)$ differentiably maps $[0,1]$ to $\mathcal{C}\subset \mathbb{R}^{n}$, where $n$ is some positive integer. Geometrically, $x(\xi)$ represents a curve embedded in the $n$-dimensional Euclidean space. The measure $\xi(x)$ can now be expressed as an integral of the density, $\rho: \mathcal{C}\to [0,1]$, along $\mathcal{C}$ with respect to the arc length $s$,
\begin{equation}
\label{eqn:curve1}
\xi(x) = \int_{\mathcal{C}[x(0),\,x(\xi)]} \rho(x)\;ds,
\end{equation}
where $\mathcal{C}[x(0),x(\xi)]$ denotes a segment of $\mathcal{C}$ between the points indicated in the square bracket. Due to the parameterization $x(\xi)$, the length of the curve $\mathcal{C}$ equals $\int_{\mathcal{C}}ds$, while the arc length differential $ds$ is related to $d\xi$ by $ds = \|dx/d\xi\|\;d\xi$. Using this relation and Eq. \ref{eqn:curve1}, we obtain the following identity,
\begin{equation}
\label{eqn:curve2}
\rho(x(\xi))\;\left\|\frac{dx}{d\xi}(\xi)\right\| = 1.
\end{equation}
We now differentiate Eq. \ref{eqn:curve2} with respect to $\xi$, apply the chain rule and reshuffle terms,
\begin{equation}
\label{eqn:curve3}
g(x(\xi))= \partial_s\log(\rho(x(\xi))) =\frac{\partial_{s}\rho}{\rho}(x(\xi)) = -\frac{\frac{d x}{d\xi }(\xi)\cdot \frac{d^2 x}{d\xi^2}(\xi)}{\| \frac{d x}{d\xi }(\xi)\|^3},
\end{equation}
where $\partial_s$ denotes the directional derivative along the curve $\mathcal{C}$ in the direction of increasing $\xi$. Note the expression for $g$ in Eq. \ref{eqn:curve3} reduces to Eq. \ref{eqn:line3} if $x(\xi)$ represents a line manifold, i.e., $\mathcal{C}\subset\mathbb{R}^1$.
As an example, we re-consider the Van der Pol oscillator (Eq. \ref{eqn:van-der-pol}). This time, however, $x(\xi)$ represents a curve embedded in $\mathbb{R}^2$. In particular, $x(\xi)$ describes a two-dimensional loop such that
\begin{equation}
\label{eqn:curve4}
x(\xi) = \begin{bmatrix}u(\xi T)\\\frac{du}{dt}(\xi T)\end{bmatrix}
\end{equation}
(see Figure \ref{fig:van-der-pol} for an illustration of the loop). If a numerical solution to Eq. \ref{eqn:van-der-pol} is available, one can combine Eq. \ref{eqn:curve2} with Eq. \ref{eqn:curve4} to directly evaluate the density function. Similarly, by plugging Eq. \ref{eqn:curve4} to Eq. \ref{eqn:curve3}, it is possible to compute the density gradient, analogously to the procedure described in Section \ref{sec:lines}. Consequently, on the RHS of Eq. \ref{eqn:curve3}, $dx/d\xi$ can be replaced with $du/dt$, and $d^2x/d\xi^2$ with $d^2u/dt^2$. We can do so because the density gradient is invariant to any linear transformation of variables.
Figure \ref{fig:curve-density} illustrates the density function $\rho$, as well as the length of the curve segment $ \mathcal{C}[x(0),x(\xi)]$, versus the parameter $\xi$. We observe $\rho$ is large if the slope of the length function is small, and vice versa, which is analogous to the $x\--\rho$ relation in Figure \ref{fig:line}. In this particular case, $\rho(\xi)$ is clearly a periodic function with period 0.5. This property is manifested in Figure \ref{fig:van-der-pol}. Indeed, one can notice the relation between $du/dt$ and $u$ at $t\in[0,T_{1/2}]$ is the same as $-du/dt$ and $-u$ at $t\in[T_{1/2},T]$, where $T_{1/2}$ corresponds to $\xi = 0.5$.
Figure \ref{fig:curve-gradient} shows the density gradient $g$ computed using two distict ways: through a direct evaluation via Eq. \ref{eqn:curve3} and a finite difference scheme (see the caption for more details). The two approaches provide visibly identical solutions, which confirms the correctness of Eq. \ref{eqn:curve3}. Clearly, the density gradient inherits the periodic behavior of $\rho$. We notice that the density gradient vanishes if the numerator of Eq. \ref{eqn:curve3} is zero, which can happen if $d^2u/dt^2 = 0$ (at the two orange dots in Figure \ref{fig:van-der-pol}) and/or $du/dt + d^3u/dt^3 = 0$ (at the six green dots in Figure \ref{fig:van-der-pol}). These two cases coincide with the local extrema of the density function (i.e., $d\rho/d\xi = 0$ if at least one of these equations is satisfied). However, zero density gradient does not imply the inflection point ($d^2u/dt^2 = 0$), in contrast to the line manifold case (see Section \ref{sec:lines}).
\begin{figure}
\includegraphics[width=1.05\textwidth]{figures/vdp_dens_length.png}
\caption{The density $\rho$ (blue) and length of the curve segment $\mathcal{C}[x(0),x(\xi)]$ (red) associated with the map $x(\xi)$ defined by Eq. \ref{eqn:curve4}. The former is computed using the analytical expression in Eq. \ref{eqn:curve2}, while the latter is approximated by summing the length of consecutive linear segments connecting the points in the sequence $\{x(0), x(\Delta t/T), x(2\Delta t/T),...,x(\xi)\}$, obtained in the numerical integration of Eq. \ref{eqn:van-der-pol}.}
\label{fig:curve-density}
\end{figure}
\begin{figure}
\includegraphics[width=1.\textwidth]{figures/vdp_g.png}
\caption{The density gradient function $g$ computed directly (using Eq. \ref{eqn:curve3}) and through a finite difference method. In the latter approach, we note $\partial_s \rho = \partial_{\xi}\rho / \partial_{\xi} s$. Both the numerator and denominator is approximated using the central finite difference scheme on a uniform grid using data presented in Figure \ref{fig:curve-density}.}
\label{fig:curve-gradient}
\end{figure}
\section{Computing $g$ on general smooth manifolds}\label{sec:general}
The purpose of this section is to generalize the concept of the density gradient and derive a formula for $g$ defined on higher-dimensional manifolds. Here, we consider a smooth invertible map $x(\xi):U\to M$, where $U\subset \mathbb{R}^m$, $M\subset \mathbb{R}^n$, $m\leq n$, $x = [x_1,...,x_n]^T$ and $\xi = [\xi_1,...,\xi_m]^T$. $U$ is an $m$-orthotope (hyperrectangle), which is defined as the Cartesian product of $m$ 1D line manifolds (i.e., intervals of the real line). We no longer assume that these elementary sets only involve numbers between 0 and 1. $M$ is an oriented differentiable manifold, whose shape is defined by the chart $x$. For example, if $m=2$ and $n=3$, then $M$ represents a smooth surface. The density gradient $g$ is now defined as a directional gradient of the logarithm of the density function $\rho:M\to[0,1]$ implied by the chart $x(\xi)$. In particular, $g = \nabla_s\log\rho$, $\nabla_s := [\partial_{s_1}, \partial_{s_2}, ..., \partial_{s_m}]^T$, where $\partial_{s_i}$, $i=1,...,m$, denote directional derivatives along the corresponding {\em isoparametric curves}. The $i$-th component of $g$ is the rate of change of $\log\rho$ along the curve whose preimage involves vectors $\xi\in U$ with constant all coordinates except $\xi_i$. If $\log\rho$ is differentiable with respect to all the coordinates of $x$ and $\nabla_x :=[\partial_{x_1}, \partial_{x_2}, ..., \partial_{x_n}]^T$, then $\partial_{s_{i}} \log\rho = \nabla_x\log\rho\cdot s_i$, where $s_{i}$ denotes the unit vector that is tangent to the corresponding isoparametric curve and points in the direction of increasing $\xi_i$. In Section \ref{sec:general-formula}, we derive a generic formula for $g$, while Section \ref{sec:general-example} provides a specific example of a two-dimensional smooth manifold embedded in $\mathbb{R}^3$ (with $m=2$ and $n=3$).
\subsection{Derivation of the general formula}\label{sec:general-formula}
Recall $x(\xi):U\to M$ is an invertible and differentiable map, where $U\subset R^m$, $M\subset R^n$, and $m\leq n$, while $\rho(x):M\to [0,1]$ is the density function implied by that chart. Let $\omega(x)$ be the natural volume form defined on $M$. Therefore, the Lebesgue measure $m$ of any subset $V\subset U$, mapped by $x$ to $N\subset{M}$, equals
\begin{equation}
m(V) = \int_{N}\rho(x)\;d\omega(x),
\end{equation}
which implies that the volume element $dm$ defined on $U$ can be expressed in terms of $\rho$ and the volume element defined on $M$, at every point $x(\xi)$,
\begin{equation}
\label{eqn:general-measure}
dm = d\xi_1\wedge d\xi_2\wedge ... \wedge d\xi_m = \rho(x)\;d\omega(x) = \rho(x)\;dx_1\wedge dx_2\wedge ... \wedge dx_n.
\end{equation}
The wedge symbol ($\wedge$) denotes the exterior product, while $d\xi_{i}$, $i=1,...,m$ and $dx_{i}$, $i=1,...,n$ represent covectors (1-forms) associated with the corresponding coordinate directions. Intuitively, these 1-forms measure small displacements in the direction of one coordinate. The volume element on $M$, $d\omega$, can be expressed in terms of $\xi$,
\begin{equation}
\label{eqn:general-substitution}
d\omega(x(\xi)) = \sqrt{\det C(x(\xi))}\;d\xi_1\wedge d\xi_2\wedge ... \wedge d\xi_m,
\end{equation}
where $C$ represents the $m\times m$ metric tensor of the coordinate transformation $\xi \to x$, defined as
\begin{equation}
\label{eqn:general-metric}
C(x(\xi)) = [\nabla_{\xi} x(\xi)]^T\;\nabla_{\xi} x(\xi),
\end{equation}
or, componentwise,
\begin{equation}
C_{ij}(x(\xi)) = \partial_{\xi_i} x(\xi) \cdot \partial_{\xi_j} x(\xi).
\end{equation}
The vector gradient $\nabla_{\xi} x(\xi)$ is represented by an $n\times m$ matrix, in which the $j$-th column contains the derivative of $x$ with respect to $\xi_j$, i.e., $[\nabla_{\xi}x(\xi)]_{ij} = \partial_{\xi_{j}}x_{i}(\xi)$.
Combining Eq. \ref{eqn:general-measure} and \ref{eqn:general-substitution}, we conclude that the relation between the density function $\rho$ and metric tensor $C$, at any point $x(\xi)\in{M}$, can be written in the following way,
\begin{equation}
\label{eqn:general1}
\rho(x(\xi))\;\sqrt{\det C(x(\xi))} = 1,
\end{equation}
which is a generalization of Eq. \ref{eqn:curve2}.
Let us now QR-factorize the vector gradient $\nabla_{\xi} x(\xi)$,
\begin{equation}
\label{eqn:general-qr}
\nabla_{\xi} x(\xi) = Q(x(\xi))\;R(x(\xi)),
\end{equation}
where $Q$ is an $n\times m$ matrix, whose columns form an orthonormal basis for the column space of $\nabla_{\xi} x(\xi)$, while $R$ is an $m\times m$ upper-triangular matrix. Note $Q^TQ = I$ everywhere on $M$. Using this property, we immediately notice that $C = R^T R$ and, therefore, Eq. \ref{eqn:general1} reduces to
\begin{equation}
\label{eqn:general2}
\rho(x(\xi))\;|\det R(x(\xi))| = 1.
\end{equation}
For any invertible matrix $A(s)$, which depends on a scalar $s$, the following indentity is true,
\begin{equation}
\label{eqn:general-indentity}
\frac{\partial \det A(s)}{\partial s} = \det A\; \mathrm{tr}\left(A^{-1}(s)\;\frac{\partial A(s)}{\partial s}\right).
\end{equation}
Differentiating Eq. \ref{eqn:general2} with respect to $\xi_{i}$, applying chain rule and Eq. \ref{eqn:general-indentity}, we obtain the following expression for the $i$-th component of the density gradient,
\begin{equation}
\label{eqn:general-g1}
g_i(x(\xi)) = \frac{\partial_{s_i}\rho(x(\xi))}{\rho(x(\xi))} = - \frac{\partial_{s_i} \det R(x(\xi))}{\det R(x(\xi))} = - \frac{\partial_{\xi_i} \det R(x(\xi))}{\det R(x(\xi))\|\partial_{\xi_i}x(\xi)\|}.
\end{equation}
Eq. \ref{eqn:general-g1} is computationally inconvenient, as it involves evaluating the determinant of $R$ and its directional derivative. Our goal is to rewrite the RHS of that equation such that only first and second parametric derivatives of $x(\xi)$, as well as $Q$ and $R$ factors, are involved.
Since $R$ is an upper-triangular matrix, we notice that
\begin{equation}
\label{eqn:general-R}
\frac{\partial \det R}{\det R} = \frac{\partial \left(\prod_{k=1}^m R_{kk}\right)}{\prod_{k=1}^m R_{kk}} = \sum_{k=1}^{m}\frac{(\partial R)_{kk}}{R_{kk}} = \mathrm{tr}(\partial R\;R^{-1}).
\end{equation}
Now, differentiating Eq. \ref{eqn:general-qr} with respect to $\xi_i$, and then left- and right-multiplying the resulting expression by $Q^T$ and $R^{-1}$, respectively, we obtain
\begin{equation}
\label{eqn:general-aux1}
Q^T(x(\xi)) \;\partial_{\xi_i}\nabla_{\xi}x(\xi)\;R^{-1}(x(\xi)) = Q^T(x(\xi))\;\partial_{\xi_i}Q(x(\xi)) + \partial_{\xi_i}R(x(\xi))\;R^{-1}(x(\xi)).
\end{equation}
Note that since $Q^TQ = I$, then $Q^T\;\partial_{\xi_i} Q$ is anti-symmetric, which means its trace vanishes. Therefore, the following equality
\begin{equation}
\label{eqn:general-treq}
\mathrm{tr}\left(Q^T(x(\xi)) \;\partial_{\xi_i}\nabla_{\xi}x(\xi)\;R^{-1}(x(\xi))\right) = \mathrm{tr}\left(\partial_{\xi_i}R(x(\xi))\;R^{-1}(x(\xi))\right)
\end{equation}
holds everywhere on $M$. Finally, by combining Eq. \ref{eqn:general-g1}, \ref{eqn:general-R} and \ref{eqn:general-treq}, we obtain the general formula for $g_i$,
\begin{equation}
\label{eqn:general-g2}
g_i(x(\xi)) = \partial_{s_{i}}\log\rho(x(\xi)) = -\frac{\mathrm{tr}\left(Q^T(x(\xi)) \;\partial_{\xi_i}\nabla_{\xi}x(\xi)\;R^{-1}(x(\xi))\right)}{\|\partial_{\xi_i}x(\xi)\|},
\end{equation}
which holds everywhere on $M$ for $i = 1,...,m$. Using Einstein's summation convention, Eq. \ref{eqn:general-g2} can be rewritten to
\begin{equation}
\label{eqn:general-g2-einstein}
g_i(x(\xi)) = -\frac{q_j(x(\xi))\cdot\partial_{
\xi_i} \partial_{\xi_k} x(\xi)\;R_{kj}^{-1}(x(\xi))}{\|\partial_{\xi_i}x(\xi)\|},
\end{equation}
where $q_{j}(x(\xi))$ denotes the $j$-th column of $Q(x(\xi))$. Thus, to directly compute the density gradient at any point on a manifold, all first and second derivatives of the chart $x(\xi)$ must be found. In addition, QR factorization of the vector gradient $\nabla_\xi x$ and inversion of the $R$ matrix must be performed. In practice, inverting the triangular matrix $R$ means solving a linear system using the backward substitution method, which requires $\mathcal{O}(m^2)$ operations. Note Eq. \ref{eqn:general-g2-einstein} reduces to Eq. \ref{eqn:curve3} if $m=1$. In the following section, we present an example illustrating some of these quantities. Although Eq. \ref{eqn:general-g2-einstein} is a formula for the derivative in the direction of a isoparametric curve, we can compute derivatives of $\log\rho$ in an arbitrary direction using the distributive law of the dot product.
\subsection{Example: a surface manifold}\label{sec:general-example}
As an example of a surface manifold (with $m=2$ and $n=3$), let us consider $x(\xi) = u(\xi) = [u_1(\xi),u_2(\xi),u_3(\xi)]^T$, where $\xi= [c,t]^T$, $-5 \leq c \leq 5$, $0 \leq t \leq 0.4$, $\left.u(\xi)\right|_{t=0} = [c,c,28]^T$, and $\partial_t u(\xi) = f(u(\xi))$, where $f$ is defined as follows,
\begin{equation}
\label{eqn:general-lorenz}
\begin{split}
& \partial_t u_1(\xi) = 10\;(u_2(\xi)-u_1(\xi)),\\& \partial_t u_2(\xi) = u_1(\xi)\,(28 - u_3(\xi)) - u_2(\xi), \\ & \partial_t u_3(\xi) = u_1(\xi)\,u_2(\xi) - \frac{8}{3}u_3(\xi).
\end{split}
\end{equation}
System \ref{eqn:general-lorenz} represents the Lorenz '63 oscillator, which is a mathematical model used for atmospheric convection \cite{lorenz-climate}. This system is known to exhibit chaotic behavior. However, we are interested in the solution in a short time interval, such that the trajectories do not intersect and the resulting surface is orientable. In particular, we compute $x(\xi)$ by numerically integrating System \ref{eqn:general-lorenz} in time for different values of $c\in[-5,5]$, using the second-order Runge-Kutta scheme with $\Delta t = 0.002$. There are two reasons we have chosen this particular $x(\xi)$. First, it serves as a perfect example of a problem, in which the smooth one-to-one solution, $x(\xi)$, cannot be found analytically. Thus, the computation of $g$ should be performed numerically using closed-form relations derived in Section \ref{sec:general-formula}. Second, the surface described by the chart $x(\xi)$ can be obtained as a evolution of 1D manifolds. This observation is utilized in Section \ref{sec:recursion}, where we derive expressions for evolving manifolds. To evaluate $\rho$ and $g$, we directly use Eq. \ref{eqn:general2} and Eq. \ref{eqn:general-g2}, respectively. To find these quantities, the vector gradient $\nabla_{\xi}x(\xi)=[\partial_c x(\xi), \partial_t x(\xi)]$, as well as the following second derivatives: $\partial_c^2 x(\xi), \partial_t^2 x(\xi), \partial_c\partial_t x(\xi)$, must be found at every point on the manifold. The time derivative, $\partial_t x(\xi) = f(x(\xi))$, is obtained automatically as we integrate System \ref{eqn:general-lorenz} in time. The second derivative of $x$ with respect to $t$ is obtained using the chain rule, $\partial_t^2 x(\xi) = \partial_t f(x(\xi)) = Df(x(\xi))\,f(x(\xi))$, where $Df$ denotes the Jacobian of System \ref{eqn:general-lorenz}. Thus, from the computational point of view, we need to solve a tangent equation to find $\partial_t x(\xi)$ at every point of the trajectory defined by System \ref{eqn:general-lorenz}. Using this approach, one can analogously find derivatives with respect to $c$. Let $v(\xi) = \partial_c x(\xi)$ and $w(\xi) = \partial^2_c x(\xi)$. Using the chain rule, we conclude that $\partial_t v(\xi) = Df(x(\xi))\,v(\xi)$, $\left.v(\xi)\right|_{t=0} = [1,1,0]^T$ and, by differentiating again, $\partial_t w(\xi) = D^2f(x(\xi))(w(\xi), w(\xi)) + Df(x(\xi))\,w(\xi)$, $\left.w(\xi)\right|_{t=0} = [0,0,0]^T$, where $D^2 f$ denotes the Hessian of $f$. Using Einstein's summation convention, the $i$-th component of the blinear form $D^2f(x(\xi))(w(\xi), w(\xi))$ can be written as $\partial_{x_k}\partial_{x_{l}}f_i\,w_k\,w_l$. Finally, the mixed derivative $\partial_c\partial_t x(\xi) = \partial_t v(\xi)$ is a byproduct of the numerical integration of the tangent equation for $v$. We solve all of these tangent equations using the same time integrator as the one mentioned above. Since $m=2$, the $2\times 2$ $R$ matrix is inverted analytically at every point on the trajectory.
In this case, the $U$ space, which is the domain (preimage) of $x$, is in fact a Cartesian product of $[-5,5]$ and $[0,0.4]$. The upper plot in Figure \ref{fig:general-mesh} graphically represents $U$, while the lower plot illustrates the $u_1\--u_3$ projection of $M$, obtained through the mapping $x(\xi)$. For completeness, in Figure \ref{fig:general-mesh-ext}, we also include the $u_1\--u_2$ and $u_2\--u_3$ projection of the deformed mesh. It is clear that the deformation is symmetric with respect to $c = 0$. We also observe that fibers (isoparametric lines) corresponding to larger values of $t$ are subject to greater stretching than those at smaller $t$. These features are reflected by the distribution of the density function $\rho$, plotted in Figure \ref{fig:general-density}. The smaller the area of each distorted quadrilateral of the mesh, the larger the value of the density function. Indeed, the smallest values of the density distribution are located around $t=0.4$. This region coincides with the most stretched quadrilaterals.
\begin{figure}
\includegraphics[width=1.0\textwidth]{figures/surface_reference.png}
\includegraphics[width=1.0\textwidth]{figures/surface_deformed_1_3.png}
\caption{Upper plot: a structured mesh representing the domain $U=\{(c, t)\,|\,c\in[-5,5],\, t\in[0,0.4]\}$. The black lines correspond to fixed values of $t$, while the red lines illustrate $\xi$ with a fixed value of $c$. The red and black dashed lines represent $c=5$ and $t=0.4$, while the red and black bold lines refer to $c = -2.5$ and $t=0.2$, respectively. Lower plot: $u_1\--u_3$ projection of the image of the structured mesh obtained through the mapping $x(\xi)$.}
\label{fig:general-mesh}
\end{figure}
\begin{figure}
\includegraphics[width=1.0\textwidth]{figures/surface_deformed_1_2.png}
\includegraphics[width=1.0\textwidth]{figures/surface_deformed_2_3.png}
\caption{Extension of Figure \ref{fig:general-mesh}. $u_1\--u_2$ (upper plot) and $u_2\--u_3$ (lower plot) projection of the image of the structured mesh obtained through the mapping $x(\xi)$.}
\label{fig:general-mesh-ext}
\end{figure}
\begin{figure}
\includegraphics[width=1.\textwidth]{figures/density_xz.png}
\caption{$u_1\--u_3$ projection of the density function $\rho$.}
\label{fig:general-density}
\end{figure}
Figure \ref{fig:general-density-gradient} shows the two components of the density gradient $g:=[g_c,g_t]^T = [\partial_{s_1}\log\rho,\partial_{s_2}\log\rho]^T $, corresponding respectively to the $c$- and $t$-direction. The distribution of $g_{c}$ is clearly symmetric with respect to the reflection points on the isoparametric line $c=0$, which is a manifestation of the fact the density is symmetric and directional derivative is computed in the direction of increasing $c$. The symmetry of $g_t$ is a direct consequence of the definition $g_t:=\partial_{s_2}\log\rho$, where $\log\rho$ itself is symmetric. Note the largest-in-magnitude values of $g_a$ concentrate around the boundaries of the range of $c$, i.e., at $c=\pm 5$ and, in case of $g_t$, around $u_1 = 0$. This reflects the fact the density gradient measures the relative rate of change of the density. In particular, its value becomes large if the rate of change of the density is large and/or the density itself is small. Figure \ref{fig:general-direct-fd} illustrates the density gradient along the bold isoparametric curves from Figure \ref{fig:general-mesh}, computed using Eq. \ref{eqn:general-g2-einstein} directly and through finite differencing. In case of both $g_c$ and $g_t$, we observe a good agreement between the solution computed directly and finite difference approximation, which validates our derivation of Eq. \ref{eqn:general-g2-einstein}.
\begin{figure}
\includegraphics[width=1.\textwidth]{figures/ga_xz.png}
\includegraphics[width=1.\textwidth]{figures/gt_xz.png}
\caption{$u_1\--u_3$ projection of the directional derivative of $\log\rho$, in the c-direction, $g_1:=g_c$ (upper plot), and $t$-direction, $g_2 := g_t$ (lower plot).}
\label{fig:general-density-gradient}
\end{figure}
\begin{figure}
\includegraphics[width=1.\textwidth]{figures/fd_vs_exact_a_t.png}
\caption{The first and second component of the density gradient function $g=[g_c,g_t]$, respectively at $t=0.2$ (black solid line on Figure \ref{fig:general-mesh}) and $c=-2.5$ (red solid line on Figure \ref{fig:general-mesh}), computed directly using Eq. \ref{eqn:general-g2-einstein} and through a finite difference method. In the latter approach, we note $\partial_{s_{i}} \rho = \partial_{\xi_{i}}\rho / \partial_{\xi_i} s_i$, $i = 1,2$, $\xi_1=c$, $\xi_2 = t$, where $s_i$ denotes the length of the isoparametric curve associated with $\xi_i$. Both the numerator and denominator is approximated using the central finite difference scheme on a uniform grid. The relation $s_{i}(\xi_i)$ is found in a way analogous to the one described in Section \ref{sec:curves}.}
\label{fig:general-direct-fd}
\end{figure}
\section{Recursive algorithm for $g$ along trajectories defined by diffeomorphism $\varphi$}\label{sec:recursion}
Using the results presented in Section \ref{sec:1D} and \ref{sec:general}, we now propose an iterative method for the density gradient along trajectories defined by a $C^2$ diffeomorphism $\varphi:M^k\to M^{k+1}$, $k\in\mathbb{Z}$, where both $M^k$ and $M^{k+1}$ represent differentiable manifolds of the same dimension embedded in $\mathbb{R}^n$, $n\in \mathbb{Z}^{+}$. Let us consider two different charts, $x^{k}(\xi)\in N^k\subset M^k$ and $x^{k+1}(\xi)\in N^{k+1}\subset M^{k+1}$, such that
\begin{equation}
\label{eqn:recursion-map}
x^{k+1}(\xi) = \varphi(x^k(\xi))
\end{equation}
for all $\xi\in V\subset{U}\subset\mathbb{R}^m$, $1\leq m\leq n$, $k\in\mathbb{Z}$. Let $\omega^k$ and $\omega^{k+1}$ be the natural volume forms in $M^k$ and $M^{k+1}$, respectively, where $\omega^{k+1}$ is the pushforward of $\omega^k$ under $\varphi$. Therefore, for all $k\in\mathbb{Z}$, the Lebesgue measure $m$ of the subspace $V$ can be expressed as follows,
\begin{equation}
\label{eqn:recursion-lebesgue}
m(V) = \int_{N^k}\rho^{k}(x)\;d\omega^k(x) = \int_{N^{k+1}}\rho^{k+1}(x)\;d\omega^{k+1}(x),
\end{equation}
where $\rho^k$ and $\rho^{k+1}$ are densities implied by $x^k(\xi)$ and $x^{k+1}(\xi)$, respectively. Following the procedure involving Eq. \ref{eqn:general-measure}-\ref{eqn:general-metric}, it is possible to find the relation between $\rho^k$, $\rho^{k+1}$, and the metric tensors of the two transformations: $\xi\to x^k$ and $\xi\to x^{k+1}$. Thus, by applying the chain rule, we find a relation between the parametric derivatives of $x^k(\xi)$ and $x^{k+1}(\xi)$, thanks to which a general recursive formula for the density gradient along the trajectory defined by $\varphi$ can be inferred. The $g^k$ function should be understood as the directional derivative of the (logarithmic) density implied by the chart $x^k(\xi)$. In Section \ref{sec:recursion-derivation}, we derive an iterative procedure for $g^k$, while Section \ref{sec:recursion-example} presents the use of the proposed algorithm by revisiting the Lorenz '63 oscillator. Throughout this section, repeated indices in the subscript of any term imply summation (Einstein's notation), unless otherwise stated.
\subsection{A generic recursive procedure for $g^k$}\label{sec:recursion-derivation}
As pointed out above, the first step is to find a relation between the parametric gradients of $x^k$ and $x^{k+1}$. Applying the definition of $\varphi$ from Eq. \ref{eqn:recursion-map} and the chain rule, we can expand $\nabla_{\xi}x^{k+1}$ in the following way,
\begin{equation}
\label{eqn:recursion-first-deriv}
\nabla_{\xi}x^{k+1}(\xi) = D\varphi(x^k(\xi))\;\nabla_{\xi}x^{k}(\xi),
\end{equation}
or, equivalently,
\begin{equation}
\label{eqn:recursion-first-deriv-comp}
\partial_{\xi_i}x^{k+1}(\xi) = D\varphi(x^k(\xi))\;\partial_{\xi_i}x^{k}(\xi),
\end{equation}
where $D\varphi$ denotes the $n \times n$ Jacobian matrix of $\varphi$, i.e., $(D\varphi)_{ij} = \partial_{x_j}\varphi_i$.
By differentiating Eq. \ref{eqn:recursion-first-deriv-comp} once more, with respect to $\xi_{j}$, we obtain
\begin{equation}
\label{eqn:recursion-second-deriv-comp}
\partial_{\xi_i}\partial_{\xi_j}x^{k+1}(\xi) = D^2\varphi(x^k(\xi))\left(\partial_{\xi_i}x^{k}(\xi),\partial_{\xi_j}x^{k}(\xi)\right) + D\varphi(x^k(\xi))\;\partial_{\xi_i}\partial_{\xi_j}x^{k}(\xi),
\end{equation}
where $D^2\varphi$ is the Hessian of $\varphi$, which is in fact a third-order $n\times n\times n$ tensor. Analogously to the example presented in Section \ref{sec:general-example}, the first term in the RHS of Eq. \ref{eqn:recursion-second-deriv-comp} is a bilinear form that outputs an $n$-element vector. In this case, the $i$-th component of that vector equals $\partial_{x_p}\partial_{x_q}\varphi_i(x^k(\xi))\;\partial_{\xi_i}x_p^k(\xi)\;\partial_{\xi_j}x_q^k(\xi)$.
In the second step, we directly use the formula for the density gradient derived in Section \ref{sec:general-formula}. Let $f^k:=f(x^k(\xi))$ be a shorthand notation for any function $f$ defined at $x^k(\xi)$, and $e_i(x^k(\xi)):= \partial_{\xi_i}x^k(\xi)$, $a_{ij}(x^k(\xi)) := \partial_{\xi_i}\partial_{\xi_j}x^{k}(\xi)$. Thus, by combining Eq. \ref{eqn:recursion-first-deriv-comp}, \ref{eqn:recursion-second-deriv-comp} with Eq. \ref{eqn:general-g2-einstein} derived for a generic chart $x(\xi)$, we conclude that
\begin{equation}
\label{eqn:recursion-g}
g_i^k = - \frac{(R^{-1}_{lj})^k}{\|e_i^k\|}\,q_j^k\cdot a_{il}^k ,
\end{equation}
\begin{equation}
\label{eqn:recursion-qr}
(\nabla_\xi x)^k = [e_1^k\,e_2^k\dotsm e^k_m] = Q^k\,R^k = [q_1^k\,q_2^k\dotsm q^k_m]\,R^k.
\end{equation}
\begin{equation}
\label{eqn:recursion-e}
e^{k+1}_{i} = D\varphi^k e^k_i,
\end{equation}
\begin{equation}
\label{eqn:recursion-a}
a^{k+1}_{ij} = D^2\varphi^k (e_i^k,e_j^k) + D\varphi^k a^k_{ij},
\end{equation}
hold for any $\xi\in{V}\subset U$.
To summarize, if a map $\varphi$ relating two consecutive points on the trajectory, $x_{k}(\xi)$ and $x_{k+1}(\xi)$, is available, then the density gradient at one point can be computed using information associated with the other point. In particular, according to Eq. \ref{eqn:recursion-g}, the $i$-th component of $g$ requires knowledge of $e_j$, $j = 1,...,m$ and $a_{pq}$, $p, q = 1,...,m$ at the same point. Thus, to compute one component of the density gradient at $x^k(\xi)$ for some $\xi$, we need to apply the recursion in Eq. \ref{eqn:recursion-e} $km$ times and, analogously, the recursion in Eq. \ref{eqn:recursion-a} $1/2\,k m^2$ times. The $1/2$ factor is a consequence of the fact that $a_{ij}(\xi) = a_{ji}(\xi)$ for any admissible $\xi$, because $x_{k}$ is assumed to be twice differentiable for any $k\in\mathbb{Z}$. In addition, at every step $k$, the QR factorization of $(\nabla_{\xi} x)^k = [e_1^k\,e_2^k\dotsm e^k_m]$ and inversion (either direct if $m$ is small or through solving a linear system) of the resulting $m\times m$ $R^k$ matrix must be performed. We assume $x^0(\xi)$ is given, from which we directly compute initial conditions for recursions in Eq. \ref{eqn:recursion-e} and \ref{eqn:recursion-a}.
The recursion involving Eq. \ref{eqn:recursion-g}-\ref{eqn:recursion-a} can be used to devise algorithms for differentiating the invariant, physical SRB measure $m_{\mathrm{SRB}}$, which is guaranteed to exist in uniformly hyperbolic systems. In general, $m_{\mathrm{SRB}}$ is not absolutely continuous everywhere on the manifold, but only conditional measures of $m_{\mathrm{SRB}}$ along unstable manifolds are absolutely continuous. The SRB density gradient $g_{\mathrm{SRB}}$, defined as a directional derivative of the conditional SRB density on the unstable manifold, is a byproduct of the integration by parts (analogous to Eq. \ref{eqn:intro2}), preceded by the disintegration of $m_{\mathrm{SRB}}$ \cite{chandramoorthy-s3,sliwiak-1d}. Thus, if a direction of the unstable manifold is given, the recursive formula presented in this section might be further developed to compute the SRB density gradient, defined on a manifold of any dimension, along a trajectory initiated at a $m_{\mathrm{SRB}}$-typical point.
\subsection{Example: evolution of a 1D manifold}\label{sec:recursion-example}
In this section, we demonstrate the application of the recursive scheme for the density gradient $g^k$. For this purpose, let us re-consider the Lorenz '63 oscillator, defined by System \ref{eqn:general-lorenz}. In particular, we define $\varphi$, such that it represents numerical time integration of System \ref{eqn:general-lorenz} for a period of $\Delta t$, i.e., $u(t+\Delta t) = \varphi(u(t))$ with $u(t)$ being the solution of the system at time $t$. Let us consider a 1D smooth manifold embedded in $\mathbb{R}^3$ described by the following chart $x^0(c) = [c,c,28]^T$, $-5\leq c\leq 5$. Note $x^0(c)$ coincides with the black solid boundary of the surface depicted in Figures \ref{fig:general-mesh}-\ref{fig:general-mesh-ext}. Now, by applying $\varphi$ recursively, the next step is to numerically compute a sequence of charts $\{x^0(c), x^1(c), x^2(c), ...\}$, where $x^{k+1}(c) = \varphi(x^{k}(c))$. Our aim is to compute $g^k = \partial_s\log\rho^k$, where $\rho^k$ is a density implied by the chart $x^k(c)$. The operator $\partial_s$ denotes a generic directional derivative along the curve in the direction of increasing $c$. The formulas derived in Section \ref{sec:recursion-derivation} give us all necessary tools to compute $g^k$ along the trajectory defined by $\varphi$. In this example, however, we consider the simplest case, $m=1$. Eq. \ref{eqn:recursion-g}-\ref{eqn:recursion-a} can be dramatically simplified, because $\nabla_{\xi} x = dx/dc$ is just a vector, and thus QR factorization is equivalent to normalizing that vector. Let $e = dx/dc = \|dx/dc\|\,q$ and $a = d^2x/dc^2$ and, therefore,
\begin{equation}
\label{eqn:recursion-g-1D}
g^k = - \frac{q^k\cdot a^k}{\|e^k\|^2},
\end{equation}
\begin{equation}
\label{eqn:recursion-e-1D}
e^{k+1} = D\varphi^k\,e^k,\;\;\;q^k = \frac{e^k}{\|e^k\|},
\end{equation}
\begin{equation}
\label{eqn:recursion-a-1D}
a^{k+1} = D^2\varphi^k(e^k,e^k) + D\varphi^k\,a^k.
\end{equation}
Note Eq. \ref{eqn:recursion-g-1D}-\ref{eqn:recursion-a-1D} can be derived directly using Eq. \ref{eqn:curve3} and the chain rule for parametric derivatives.
How does this example differ from the one presented in Section \ref{sec:general-example}? There, we used a chart $x_s(\xi):\mathbb{R}^2\to\mathbb{R}^3$, $\xi=[c,t]^T$, which defined a two-dimensional manifold. The rate of change of $x(\xi)$ in the $t$-direction was determined by the Lorenz '63 oscillator (System \ref{eqn:general-lorenz}). Here, using the iterative procedure, we generate a bunch of 1D manifolds $x^k(c)$. The evolution of these curves (in geometric sense) is determined by $\varphi$, which is in fact a discrete version of System \ref{eqn:general-lorenz}. Thus, if we generate infinitely many such curves and $\Delta t\to 0$, we effectively obtain the same surface as the one shown in Figure \ref{fig:general-mesh}. Intuitively, the density $\rho_s$ implied by $x_s(\xi)$ measures number of points mapped from a uniform distribution per unit surface area. Likewise, the density $\rho^k$, implied by the chart $x^k(c)$, measures number of points mapped from a uniform distribution per unit curve length. Since we use the same discretization scheme to integrate differential equations, the localization of points obtained in both the computation of surface from Section \ref{sec:general-example} and, here, evolution of curves is exactly the same. However, the density $\rho^k$ does not equal to the marginal distribution of $\rho_s$ at $t = k\,\Delta t$ (assuming uniform discretization of time). In case of the surface example, the value of the density function reflects the densification of points, mapped from a uniform distribution, in both the $t$ and $c$ directions. In the latter example, the density is determined only by the localization of points along the evolving curve. Figure \ref{fig:recursion-fd} illustrates the density gradient $g^k$ along the evolving curve, recorded at three different time steps $k$. We observe $g^k=0$ at $k=0$, which is a consequence of the choice of the uniformly distributed initial condition. Due to the symmetric geometry of $M^k$, defined by the Lorenz '63 oscillator at $t\in[0,0.4]$, the density gradient features symmetric behavior with respect to the origin of the $g^k(c)$-vs.-$c$ relation.
\begin{figure}
\centering
\includegraphics[width=1.\textwidth]{figures/iterative_fd_vs_exact.png}
\caption{Density gradient function computed using the recursion involving Eq. \ref{eqn:recursion-g-1D}-\ref{eqn:recursion-a-1D} at three different time steps $k = t/\Delta t$. The finite difference approximation is generated using the approach described in Section \ref{sec:curves}.}
\label{fig:recursion-fd}
\end{figure}
\section{Conclusions and future work}\label{sec:conclusion}
A Monte Carlo integration scheme applied to a highly-oscillatory function might be remarkably expensive. The computational cost, however, can be dramatically reduced by integrating the original formulation by parts. Such treatment gives rise to a new quantity, i.e., a directional derivative of the logarithm of the density implied by a chart describing the integration domain. The computation of that derivative, which we call the {\em density gradient}, requires knowledge of the first and second derivatives of the chart with respect to the domain parameterization. If the domain manifold evolves according to some diffeomorphism $\varphi$, the calculation of the density gradient along a trajectory requires solving a collection of first- and second-order tangent equations, involving both the Jacobian and Hessian of $\varphi$. The number of these equations is respectively proportional to $m$ and $m^2$, where $m$ is the dimension of the manifold.
The formulas derived in this paper is a major step toward constructing generalizable algorithms for {\em SRB density gradient}. This quantity plays a major role in the sensitivity analysis of uniformly hyperbolic systems, including many popular chaotic equations. Using the recursive formula for the density gradient along a trajectory defined by $\varphi$ and the definition of the SRB measure, one can potentially devise a trajectory-driven procedure for the SRB density gradient. This is in fact the subject of the authors' ongoing investigation.
\section*{Acknowledgments}
This work was supported by Air Force Office of Scientific Research Grant No. FA8650-19-C-2207 and U.S. Department of Energy Grant No. DE-FOA-0002068-0018.
\bibliographystyle{elsarticle-num-names}
|
\section{Introduction}
The field of random graphs was born in a series of papers of Erd\H{o}s and Rényi~\cite{ER2, ER1, ER3}. The paper~\cite{ER1} concentrates in particular on the existence of a giant component in the random graphs $\mathcal G(n,M)$ and $\mathcal G(n, p)$, that is, a connected component that contains a constant proportion of all $n$ vertices in the graph. In the $\mathcal G(n,M)$ model, $M$ edges are chosen among all $\binom{n}{2}$ pairs of vertices uniformly at random to form a random graph with exactly $M$ edges, while in the $\mathcal G(n,p)$ model every pair of vertices forms an edge with probability $p$ in the final graph independently from all other pairs (or equivalently, $G\in \mathcal G(n,p)$ is a random subgraph of the complete graph on $n$ vertices after $p$-percolation of its edges). In~\cite{ER1}, Erd\H{o}s and Rényi proved the following (by now very classical) result: for any $\varepsilon > 0$, if $M\le (1-\varepsilon) n/2$, then all connected components in the random graph $G\in \mathcal G(n,M)$ have $O(\log n)$ vertices asymptotically almost surely (a.a.s.), while if $M\ge (1+\varepsilon) n/2$, then the largest component in the random graph $G\in \mathcal G(n,M)$ contains $\Omega(n)$ vertices but the second largest contains only $O(\log n)$ vertices a.a.s. Later Bollob{\'a}s~\cite{Bol1} and {\L}uczak~\cite{Luc1} made a precise analysis of the more complicated regime when $M = n/2 + o(n)$ and exhibited a critical window around $M = n/2$ of width of order $\Theta(n^{2/3})$, in which a number of connected components with $\Theta(n^{2/3})$ vertices in each happen to coexist a.a.s. All results above have natural analogues for $\mathcal G(n,p)$. Aldous~\cite{Ald} later made a beautiful connection between the sizes of the connected components in the critical regime in $\mathcal G(n,p)$ and the zeros of a Brownian motion with a suitable drift.
In fact, percolation of finite graphs was considered in many particular cases. Another classical example is the \emph{hypercube in dimension $n$}, denoted by $H_n$. The graph $H_n$ has vertices $\{0,1\}^n$ and two vertices $u$ and $v$ are connected by an edge if they differ in exactly one entry. In~\cite{ES} Erd\H{o}s and Spencer showed that if $p\le (1-\varepsilon)/n$, then a.a.s. $p$-percolation of $H_n$ leaves a graph with largest component, containing at most $o(2^n)$ vertices, and they conjectured that a component with $\Omega(2^n)$ vertices is a.a.s. present if $p\ge (1+\varepsilon)/n$. This conjecture was confirmed by Ajtai, Koml{\'o}s and Szemer{\'e}di in~\cite{AKS}. A following series of papers of Bollob{\'a}s, Kohayakawa and {\L}uczak~\cite{BKL}, Borgs, Chayes, van der Hofstad, Slade and Spencer~\cite{BCVdHSS1, BCVdHSS2, BCVdHSS3}, van der Hofstad and Nachmias~\cite{vdHN}, and Hulshof and Nachmias~\cite{HN} provides a deep understanding of the critical percolation on the hypercube in high dimension.
To the best of our knowledge there were two attempts for generalising the sharp threshold phenomenon for the existence of a giant component for large families of finite graphs. Chung, Horn and Lu~\cite{CHL} showed the existence of a sharp threshold under several conditions involving the spectrum of the adjacency matrix of the base graph. Sadly, their conditions are not satisfied for the hypercube $H_n$, see~\cite{Chu}. Alon, Benjamini and Stacey~\cite{ABC} proved the existence of a sharp threshold in expanders of uniformly bounded degree. Here as well, although the hypercube $H_n$ has indeed good expansion properties~\cite{Til}, its degree goes to infinity with $n$.
Our goal in this paper is to generalise the existence of a sharp threshold for the appearance of a giant component for Cartesian products of graphs under two assumptions: on the maximum degrees and on the isoperimetric constants of the graphs in the product. In particular, our result ensures the existence of a sharp threshold for the appearance of a giant component for the Cartesian product of any sequence of $n$ connected graphs with uniformly bounded orders, the hypercube $H_n$ being a particular case of the latter. We believe it is worth making the connection with Joos~\cite{Joo2}, who studies the threshold probability for connectivity of percolated sparse graphs and, as a corollary, completely solves the problem for Cartesian powers of a graph $G$.
\subsection{Notation and terminology}
For every positive integer $n$, we denote by $[n]$ the set $\{1,2,\dots,n\}$. In this paper, for any three positive real numbers $a,b,c$, by $a/bc$ or $a/b\cdot c$ we mean $a/(bc)$.
For a graph $G$, the \emph{order} of $G$ is the cardinality of its vertex set $V(G)$, and the \emph{size} of $G$ is the cardinality of its edge set $E(G)$. For a vertex $v\in V(G)$, we denote by $\deg_G(v)$, or just $\deg(v)$, the degree of $v$ in $G$, and by $CC_G(v)$, or just $CC(v)$, the connected component of $v$ in $G$. Then, the \emph{average degree} of $G$ is defined by
\begin{equation*}
\overline{d}(G) = \dfrac{1}{|V(G)|}\sum_{v\in V(G)} \deg_G(v).
\end{equation*}
The maximum degree of a graph $G$ is denoted $\Delta(G)$, and the order of the largest connected component in $G$ is denoted $L_1(G)$. Finally, for any graph $G$ with $|V(G)|\ge 2$, the \emph{isoperimetric constant} of $G$ is given by
\begin{equation*}
i(G) = \min_{\substack{S\subseteq V(G);\\ 1\le |S|\le |V(G)|/2}} \dfrac{|\partial S|}{|S|},
\end{equation*}
where $\partial S = \partial_G S$ is the set of edges in $G$ between a vertex in $S$ and a vertex in $V(G)\setminus S$. Clearly if $L_1(G) < |V(G)|$, then $i(G) = 0$. For a set $S\subseteq V(G)$, we also denote by $N_G(S)$, or just $N(S)$, the set of vertices in $G$ at graph distance 1 from $S$ in $G$, and also $N_G[S]$, or simply $N[S]$, is defined as $S\cup N(S)$.
For any sequence of $n$ graphs $G_1, G_2, \dots, G_n$, the \emph{Cartesian product of $G_1, G_2,\dots, G_n$}, denoted by $G_1\square G_2\square \dots \square G_n$ or $\square_{i\in [n]} G_i$, is the graph with vertex set
\begin{equation*}
\{(v_1, v_2, \dots, v_n)\hspace{0.2em}|\hspace{0.2em} \forall i\in [n], v_i\in V(G_i)\}
\end{equation*}
and edge set
\begin{equation*}
\{(u_1, u_2, \dots, u_n) (v_1, v_2, \dots, v_n)\hspace{0.2em}|\hspace{0.2em} \exists i\in [n], \forall j\neq i, u_j=v_j\text{ and } u_iv_i\in E(G_i)\}.
\end{equation*}
A $p$-percolation of a graph $G$ is a random process in which every edge in $G$ is retained with probability $p$ and deleted with probability $1-p$, independently from all other edges. If an edge is retained, we say that it is \emph{open}, and if it is deleted, we say that it is \emph{closed}. The graph consisting of all open edges is a random subgraph of $G$, which we denote by $G_p$.
For a sequence of probability spaces $(\Omega_n, \mathcal F_n, \mathbb P_n)_{n\geq 1}$ and a sequence of events $(A_n)_{n\geq 1}$, where $A_n\in \mathcal F_n$ for every $n\geq 1$, we say that $(A_n)_{n\geq 1}$ happens \emph{asymptotically almost surely} or \emph{a.a.s.} if $\underset{n\to +\infty}{\lim}\mathbb P_n(A_n) = 1$. The sequence of events $(A_n)_{n\geq 1}$ itself is said to be \emph{asymptotically almost sure} or again \emph{a.a.s}.
Our main result, Theorem~\ref{main thm}, is of asymptotic nature. Therefore, below we use the well-known asymptotic notations $o, O, \Omega$ and $\Theta$. For two functions $f,g:\mathbb N\to \mathbb R^+$ we also write $f(n)\ll g(n)$ or $g(n)\gg f(n)$ if $f(n) = o(g(n))$. Moreover, if the limit variable is not $n$, we will indicate this using lower indices such as $O_x$.
\subsection{Our result}
Throughout the paper we fix two absolute constants $\gamma > 0$ and $C\in \mathbb N$ (that is, these constants do not depend on any other parameters in the sequel). Let $(G_{n,j})_{n\in \mathbb N, j\in [n]}$ be finite connected graphs with at least one edge such that, for every $n\in \mathbb N$ and $j\in [n]$:
\begin{enumerate}
\item\label{cn 1} $\Delta(G_{n,j})\le C$, and
\item\label{cn 2} $i(G_{n,j})\ge n^{-\gamma}$.
\end{enumerate}
Define $G_{[n]} = \square_{j\in [n]} G_{n,j}$. In the sequel we write $G_j$ for $G_{n,j}$, $G$ for $G_{[n]}$, and $\overline{d}$ for $\overline{d}(G_{[n]})$, hopefully taking enough care to ensure that no confusion arises due to this abuse of notation. We insist that we reserve the notation $G_p$ for the subgraph of $G$ after $p$-percolation, so $G_p$ is not part of $(G_j)_{j\in [n]}$. For any vertex $v\in V(G)$ we denote by $CC_p(v) = CC_{G_p}(v)$. Since a vertex $v = (v_1, \dots, v_n)\in V(G)$ has degree $\sum_{j=1}^n
\deg_{G_j}(v_j)$, we conclude that
\begin{equation}\label{eq 1}
\overline{d} = \sum_{j=1}^n \overline{d}(G_j).
\end{equation}
Now we present the main result of the paper.
\begin{theorem}\label{main thm}
Fix $\varepsilon\in (0,1)$. In the above setup:
\begin{enumerate}
\item[a)] if $p = (1-\varepsilon)/\overline{d}$, then a.a.s. $L_1(G_p)\le \exp\left(-\dfrac{\varepsilon^2 n}{9 C^2}\right) |V(G)|$, and
\item[b)] if $p = (1+\varepsilon)/\overline{d}$, then there is a positive constant $c_1 = c_1(\varepsilon, \gamma, C)$ such that a.a.s. $L_1(G_p)\ge c_1 |V(G)|$.
\end{enumerate}
\end{theorem}
\begin{remark}\label{rem 2.1}
One may replace the constant $C$ with a function $n^{\alpha}$ in condition~\ref{cn 1}, where $\alpha = \alpha(\gamma)$ is a positive constant, and Theorem~\ref{main thm} will still be valid. We present the proof only of the given more simplified version of Theorem~\ref{main thm} for two reasons: first, the idea of the proof is the same and this more general version would only make the exposition more technical, and second, we believe that even this more general framework does not fully explain the existence of a sharp threshold for the giant component problem for product graphs.
\end{remark}
Let us give a quick overview of the proof of Theorem~\ref{main thm}. The first point concerns the study of two subcritical exploration processes. The first one ensures an upper bound on the order of the union of all components, containing at least one vertex of ``high'' degree. The second process deals with the remaining ``low'' degree vertices conditionally on the edges, exposed during the first process, and is therefore directly dominated by a subcritical branching process. The proof of the second point is inspired by the special case of the hypercube, studied in~\cite{AKS}. It relies on consecutively constructing connected components with larger and larger polynomial orders via the technique of two-round exposure (or rather multi-round exposure in our case). Once the correct polynomial order is attained, we show by the same technique that the condition~\ref{cn 2} on the isoperimetric constants of the graphs in the product ensures that a constant proportion of the above components are merged together in $G_p$ a.a.s.
The paper is organised as follows. In Section~\ref{sec prelims} we introduce several preliminary results. Then, in Section~\ref{sec pt 1 proof} we prove Theorem~\ref{main thm}~a), and in Section~\ref{sec pt 2 proof} we prove Theorem~\ref{main thm}~b). Finally, Section~\ref{sec discussion} is dedicated to a discussion and a couple of open questions.
\section{Preliminaries}\label{sec prelims}
\subsection{Probabilistic preliminaries}
\textbf{Chernoff's inequality:} We first state a version of the famous Chernoff's inequality, see e.g. (\cite{JLR}, Theorem 2.1).
\begin{lemma}[\cite{JLR}, Theorem 2.1]\label{chernoff}
Let $X \sim \mathrm{Bin}(n,p)$ be a Binomial random variable with parameters $n$ and $p$. Then, for any $t \ge 0$ we have
\begin{eqnarray*}
\mathbb P( X \ge \mathbb E[X] + t ) &\le& \exp \left( - \frac {t^2}{2 (\mathbb E[X] + t/3)} \right), \text{ and}\\
\mathbb P( X \le \mathbb E[X] - t ) &\le& \exp \left( - \frac {t^2}{2 \mathbb E[X]} \right).
\end{eqnarray*}
\end{lemma}
\vspace{1em}
\noindent
\textbf{The bounded difference inequality:} The following well-known inequality is a simple consequence of the Azuma-Hoeffding martingale inequality, see~\cite{JLR} or also~\cite{Mac} for an improvement.
\begin{theorem}[The bounded difference inequality, see e.g.~\cite{JLR}]\label{azuma}
Consider a sequence $(X_i)_{i\in [n]}\in \prod_{i\in [n]} \Lambda_i$ of $n$ independent random variables. Fix a function $f:\prod_{i\in [n]} \Lambda_i\to \mathbb R$ and suppose that there exist $(C_i)_{i\in [n]}$ such that, for every $i\in [n]$, $(x_j)_{j\in [n]}\in \prod_{j\in [n]} \Lambda_j$ and $x'_i\in \Lambda_i$, we have
\begin{equation*}
|f(x_1, \dots, x_{i-1}, x_i, x_{i+1}, \dots, x_n) - f(x_1, \dots, x_{i-1}, x'_i, x_{i+1}, \dots, x_n)|\le C_i.
\end{equation*}
Then, for every $t\ge 0$,
\begin{equation*}
\mathbb P(|f(X_1,\dots,X_n) - \mathbb E[f(X_1,\dots,X_n)]|\ge t)\le 2\exp\left(- \dfrac{t^2}{2\sum_{i\in [n]} C^2_i}\right).
\end{equation*}
\end{theorem}
\vspace{1em}
\noindent
\textbf{The Bienaym\'e-Galton-Watson random tree:} By now a very well-known and studied model is the \emph{Bienaym\'e-Galton-Watson random tree}, or \emph{BGW tree}. Let $\nu$ be a probability distribution over $\mathbb N\cup \{0\}$ and let $X$ be a random variable with distribution $\nu$. The BGW tree with progeny distribution $\nu$ is constructed as follows. Starting from a vertex $v_0$ (the root), every vertex gives birth (just once) to a random number of children, distributed according to $\nu$ and independent from all other vertices in the tree. The BGW tree is \emph{subcritical} if $\mathbb E[X] < 1$, \emph{supercritical} if $\mathbb E[X] > 1$, and \emph{critical} otherwise. It is a basic fact in the theory of branching processes that a subcritical BGW tree is almost surely finite while a supercritical BGW tree has strictly positive probability to be infinite. The next lemma makes the first statement more precise by giving a probabilistic estimate on the size of a subcritical BGW tree, see e.g.~(\cite{Bla}, Theorem 2.3.2).
\begin{lemma}[\cite{Bla}, Theorem 2.3.2]\label{size lem}
Let $T$ be a subcritical BGW tree such that $\mathbb E[s^X] < +\infty$ for some $s > 1$. Define
\begin{equation*}
h_{\nu} = \sup_{\theta > 0} \left(\theta - \log\left(\mathbb E[\exp(\theta X)]\right)\right).
\end{equation*}
Then, for every $k\ge 1$ we have
\begin{equation*}
\mathbb P(|V(T)|\ge k)\le \exp(-kh_{\nu}).
\end{equation*}
\end{lemma}
Fix $\varepsilon\in (0,1)$. We will use the above result in the particular case when $X\sim \mathrm{Bin}(n, p)$ with $p = (1-\varepsilon)/n$. We have
\begin{equation*}
\mathbb E[\exp(\theta X)] = \sum_{i=0}^n \binom{n}{k} \exp(\theta k) p^k (1-p)^{n-k} = (1-p+\exp(\theta) p)^n.
\end{equation*}
Then, using that $\log(1+\theta) = \theta + O_{\theta}(\theta^2)$, as $n\to +\infty$ we have
\begin{align*}
h_X =\hspace{0.3em} &\sup_{\theta > 0} (\theta - n\log(1-p+\exp(\theta) p))\\
=\hspace{0.3em} &\sup_{\theta > 0} (\theta - n(-p+\exp(\theta) p + O(1/n^2)))\\
=\hspace{0.3em} &\sup_{\theta > 0} (\theta + 1 - \varepsilon - (1-\varepsilon)\exp(\theta) + O(1/n)).
\end{align*}
Since $\exp(\theta) = 1+\theta+O_{\theta}(\theta^2)$, the latter quantity tends to a constant $\phi = \phi(\varepsilon) > 0$ as $n\to +\infty$, where
\begin{equation}\label{eq phi}
\phi(\varepsilon) = \sup_{\theta > 0} (\theta + 1 - \varepsilon - (1-\varepsilon)\exp(\theta)).
\end{equation}
\begin{corollary}\label{size cor}
The BGW tree $T$ with progeny distribution $\mathrm{Bin}(n, (1-\varepsilon)/n)$ satisfies
\begin{equation*}
\mathbb P(|V(T)|\ge k)\le \exp(-(1+o(1))k\phi).
\end{equation*}
In particular, for every $\varepsilon \in (0,1)$ there is $k_0 = k_0(\varepsilon)\in \mathbb N$ such that for every $k\ge k_0$ we have
\begin{equation*}
\mathbb P(|V(T)|\ge k)\le \exp(-k\phi/2).
\end{equation*}
\end{corollary}
\subsection{Combinatorial preliminaries}
\textbf{The isoperimetric constant of a product graph:} Recall that $G$ is a graph, defined as a Cartesian product of the graphs $G_1, G_2, \dots, G_n$. The next result, due to Chung and Tetali~\cite{CT}, makes a connection between the isoperimetric constant of $G$ and the isoperimetric constants of $(G_k)_{k\in [n]}$, see also Tillich~\cite{Til} for a slight improvement.
\begin{theorem}[\cite{CT}, Theorem 2]\label{isoper thm}
\begin{equation*}
\frac{1}{2} \min_{k\in [n]} i(G_k)\le i(G)\le \min_{k\in [n]} i(G_k).
\end{equation*}
\end{theorem}
We directly deduce the following corollary.
\begin{corollary}\label{isoper cor}
Under condition~\ref{cn 2} on $(G_k)_{k\in [n]}$ we have $n^{-\gamma}/2\le i(G)$.
\end{corollary}
\vspace{1em}
\noindent
\textbf{The largest connected component and ``balanced'' empty cuts:} The following easy observation makes a connection between empty cuts in a graph and the size of the largest connected component.
\begin{observation}\label{ob discrete cont}
Fix $k\in \mathbb N$. Let $H$ be a graph with $h$ vertices and let $\mathcal C_1, \dots, \mathcal C_k$ be disjoint connected subgraphs of $H$ such that $\cup_{j\in [k]} V(\mathcal C_j) = V(H)$. Suppose that for any set $J\subseteq [k]$ such that $|\cup_{j\in J} V(\mathcal C_j)|\in [h/3, 2h/3]$, there exists an edge in $H$ between a vertex in $\cup_{j\in J} \mathcal C_j$ and a vertex in $\cup_{j\in [k]\setminus J} \mathcal C_j$. Then, there is a connected component of $H$ that contains more than $h/3$ vertices.
\end{observation}
\begin{proof}
We argue by contradiction. Suppose that $H$ contains $m$ connected components and each of them has order at most $h/3$. Then, consider the graphs $H_0 = \varnothing, H_1, H_2, \dots, H_m$, where for every $i\in [m]$ we define $H_{i}$ to be the union of $H_{i-1}$ and some connected component in the graph $H\setminus H_{i-1}$. Since $|V(H_m)| = h$ and for every $i\in [m]$, $|V(H_i)| - |V(H_{i-1})|\le h/3$, by discrete continuity there is $\ell\in [m]$ such that $|V(H_{\ell})|\in [h/3, 2h/3]$. This is a contradiction with the assumption in the statement for the family $\{\mathcal C_j: \mathcal C_j\cap H_{\ell} = \emptyset\}$. The observation is proved.
\end{proof}
\section{\texorpdfstring{Proof of Theorem~\ref{main thm}~a) -- the subcritical regime}{}}\label{sec pt 1 proof}
We begin with a proof of Theorem~\ref{main thm}~a). Our first step will be to estimate the number of vertices of $G$ of degree at least $(1+\varepsilon/2)\overline{d}$. For every $i\in [n]$, let $X_i$ be the degree of a uniformly chosen vertex in $G_i$. For every $i\in [n]$, define $S_i = X_1+X_2+\dots+X_i$. Since $\mathbb E[S_n] = \overline{d}$ by \eqref{eq 1} and for every $i\in [n]$ we have $\Delta(G_i)\le C$, we conclude by the bounded difference inequality (Theorem~\ref{azuma}) that
\begin{equation}\label{eq 2}
\mathbb P(|S_n - \overline{d}|\ge \varepsilon \overline{d}/2)\le 2\exp\left(-\dfrac{\varepsilon^2 \overline{d}^2/4}{2C^2 n}\right)\le 2\exp\left(-\dfrac{\varepsilon^2 n}{8C^2}\right).
\end{equation}
The second inequality comes from the fact that $\overline{d}\ge n$: indeed, by \eqref{eq 1} we have $\overline{d} = \sum_{i=1}^n \overline{d}(G_i)$, and every graph among $(G_i)_{i\in [n]}$ is connected and therefore its average degree is at least 1. Therefore, the number of vertices in $G$ with degree more that $(1+\varepsilon/2)\overline{d}$ is at most a $2\exp\left(-\frac{\varepsilon^2 n}{8C^2}\right)$-proportion of all vertices of $G$.
\begin{proof}[Proof of Theorem~\ref{main thm}~a)]
First, we prove that the number of vertices connected via a path in $G_p$ to a vertex of degree at least $(1+\varepsilon/2) \overline{d}$ in $G$ is at most $\exp\left(-\dfrac{\varepsilon^2 n}{9 C^2}\right) |V(G)|$ a.a.s. Indeed, let $U$ be the set of vertices of degree at least $(1+\varepsilon/2) \overline{d}$ in $G$. Then, by \eqref{eq 2} we have $|U|\le \exp\left(-\dfrac{\varepsilon^2 n}{8 C^2}\right) |V(G)|$. We consider the following stochastic process. Let $U_0 = N_{G_p}(U)$, and for every positive integer $k$ we inductively define $U_k = N_{G_p}(U_{k-1})\setminus (U\cup U_0\cup \dots \cup U_{k-1})$. Since for every set $V\subseteq V(G)\setminus U$ we have $|\partial V|\le (1+\varepsilon/2)\overline{d} |V|$, we have that, for every $k\ge 1$,
\begin{equation*}
\mathbb E[|U_k|\hspace{0.2em}|\hspace{0.2em}U_{k-1}]\le p|\partial U_{k-1}|\le p(1+\varepsilon/2)\overline{d} |U_{k-1}|\le (1-\varepsilon/2)|U_{k-1}|.
\end{equation*}
We conclude that for every $k\ge 1$,
\begin{equation*}
\mathbb E[|U_k|]\le (1-\varepsilon/2)^k \mathbb E[|U_0|]\le (1-\varepsilon/2)^k Cn |U|.
\end{equation*}
Thus, we get by Markov's inequality that
\begin{equation*}
\mathbb P(\exists k\ge 1, |U_k|\ge (1-\varepsilon/2)^{k/2} Cn^2 |U|)\le \sum_{k\ge 1} \mathbb P(|U_k|\ge (1-\varepsilon/2)^{k/2} Cn^2 |U|)\le \sum_{k\ge 1} \dfrac{(1-\varepsilon/2)^{k/2}}{n} = o(1).
\end{equation*}
We deduce that the union of all connected components of $G_p$, containing at least one vertex of $U$, contains at most $\sum_{k\ge 1} (1-\varepsilon)^{k/2} Cn^2 |U| = \Theta(n^2 |U|) = o(|V(G)|)$ vertices a.a.s.
Denote the set of vertices in all explored connected components by $\overline{U}$. Here, an edge of $G$ is explored if the fact that it is open or not has been revealed, and a connected component is explored if all edges it contains have been explored and are open, while all edges on its boundary have been explored and are closed. After exploring all connected components of $G_p$, containing at least one vertex in $U$, we are left with unexplored edges, incident only to vertices of degree less than $(1+\varepsilon/2) \overline{d}$ in $G$. We prove that in the remainder of $G_p$ there is a.a.s. no connected component of order more than $\lceil 4\log |V(G)|/\phi(\varepsilon/2)\rceil$, with $\phi$ defined in \eqref{eq phi}. Indeed, choose any vertex $v$ and start an exploration process of its connected component $CC_p(v)$ in $G_p$. Note that any vertex in $V(G)\setminus \overline{U}$ is incident to less than $(1+\varepsilon/2) \overline{d}$ unexplored edges. Thus, the number of edges in $CC_p(v)$ is stochastically dominated by the number of explored edges in a BGW tree $T$ with progeny distribution $\mathrm{Bin}(\lfloor(1+\varepsilon/2)\overline{d}\rfloor, p)$. Since $\overline{d}\to +\infty$ with $n$ and
\begin{equation*}
p = \dfrac{1-\varepsilon}{\overline{d}}\le \dfrac{1-\varepsilon/2}{(1+\varepsilon/2)\overline{d}}\le \dfrac{1-\varepsilon/2}{\lfloor(1+\varepsilon/2)\overline{d}\rfloor},
\end{equation*}
by Corollary~\ref{size cor} we get that for every $k\ge 1$ and for every $n$ large enough
\begin{equation*}
\mathbb P(|V(CC_p(v))|\ge k)\le \mathbb P(|V(T)|\ge k)\le \exp\left(-\dfrac{k\phi(\varepsilon/2)}{2}\right).
\end{equation*}
Choosing $k = k_0 := \lceil 4\log |V(G)|/\phi(\varepsilon/2)\rceil$, we get that with probability at most $1/|V(G)|^2$, $CC_p(v)$ contains at least $k_0$ vertices. A union bound over all vertices in $V(G)\setminus \overline{U}$ implies that with probability at most $1/|V(G)|$, the largest component in $G_p$, containing no vertex in $U$, is of order at least $k_0$. Since
\begin{equation*}
\exp\left(-\dfrac{\varepsilon^2 n}{9 C^2}\right) |V(G)|\ge 2^{-n/2} |V(G)|\ge \sqrt{|V(G)|}\gg \log |V(G)|,
\end{equation*}
the proof is finished.
\end{proof}
\section{\texorpdfstring{Proof of Theorem~\ref{main thm}~b) -- the supercritical regime}{}}\label{sec pt 2 proof}
The remainder of the paper will be directed towards proving Theorem~\ref{main thm}~b). The main technique, well-known under the name \emph{two-round exposure} or \emph{sprinkling}, has by now become a classical tool in the field of random graphs. It states that the graph $G_p$ may be realised as a union of two random graphs on the same vertex set $G_{p_1}$ and $G_{p_2}$, sampled independently from each other, where $(1-p_1)(1-p_2) = 1-p$. Indeed, the probability that an edge in $G$ does not appear in $G_p$ is $1-p$, while by independence the probability that an edge in $G$ does not appear in $G_{p_1}\cup G_{p_2}$ is $(1-p_1)(1-p_2)$. Moreover, in both $G_p$ and $G_{p_1}\cup G_{p_2}$, different edges appear independently from each other.
In our case, inspired by \cite{AKS}, we show that one may choose $p_1$ and $p_2$ appropriately so that a.a.s. $G_{p_1}$ consists of a number of connected components of order at least $\Omega(n^k)$ for some large enough positive integer $k$, which contain a constant proportion of all vertices of $G$. Then, at the second stage, we show that a.a.s. a constant proportion of all such component merge in a connected component of size $\Theta(|V(G)|)$.
In the sequel, $\deg(v)$ will refer to the degree of a vertex $v$ in $G$.
\begin{observation}\label{ob neighbours}
For some $i\in [n]$, let
\begin{equation*}
v_1 = (w_1, \dots, w_{i-1}, u_1, w_{i+1}, \dots, w_n)\text{ and }v_2 = (w_1, \dots, w_{i-1}, u_2, w_{i+1}, \dots, w_n)
\end{equation*}
be two vertices in $G$. Then, $|\deg(v_1) - \deg(v_2)|\le C-1$.
\end{observation}
\begin{proof}
We have $\deg(v_1) - \deg(v_2) = \deg_{G_i}(u_1) - \deg_{G_i}(u_2)$. The claim follows since the graph $G_i$ is connected and has maximum degree at most $C$.
\end{proof}
\begin{corollary}\label{cor neighbours}
If $C\ge 2$, two vertices $v_1$ and $v_2$ in $G$ are at graph distance at least $\dfrac{|\deg(v_1) - \deg(v_2)|}{C-1}$.
\end{corollary}
Let $D$ be the set of vertices of degree at most $(1-\varepsilon/2)\overline{d}$ in $G$.
\begin{lemma}\label{AKS lem 1}
Fix $\varepsilon\in (0, 0.1)$ and $p\ge (1+7\varepsilon/8)/\overline{d}$. There is a constant $c_1 = c_1(\varepsilon) > 0$ such that, for every large enough $n$, every vertex in $G$ of degree at least $(1-\varepsilon/4) \overline{d}$ participates in a connected component of $G_p$ of size at least $\varepsilon \overline{d}/4C$ with probability at least $c_1$.
\end{lemma}
\begin{proof}
Fix a vertex $v_0$ satisfying $\deg_G(v_0)\ge (1-\varepsilon/4) \overline{d}$ and start an exploration process of the connected component of $v_0$ in $G_p$ as follows. We divide the vertices of $G_p$ in several categories: \emph{active}, when the edges, incident to this vertex, have not been explored from the vertex itself, but it has been attained via a path from $v_0$, \emph{passive}, if the vertex was active before but the edges in its neighbourhood have been explored from it, and \emph{processed}, when the vertex is either active or passive. For example, in the beginning only the vertex $v_0$ is active and there are no passive vertices. A reformulation of the statement of the lemma is that, by starting an exploration process of $G_p$ from $v_0$, with probability at least $c_1$ at least $\varepsilon \overline{d}/4C$ vertices will be processed in the end. Start by exploring all edges in $G$, going out of $v_0$, and make all neighbours of $v_0$ in $G_p$ active. Then, make $v_0$ passive and find an active vertex $v_1$, if it exists. Then, explore all edges incident to $v_1$ in $G$ and make all neighbours of $v_1$ in $G_p$ that have not yet been processed active. Then, make $v_1$ passive and find an active vertex $v_2$, if it exists, etc. Continue with the exploration until either all or at least $\varepsilon \overline{d}/4C$ vertices in the connected component of $v_0$ in $G_p$ have been processed.\par
Fix an integer $n\ge 8C^2/\varepsilon$. If $C\ge 2$, by Corollary~\ref{cor neighbours} every vertex of degree at least $(1-\varepsilon/4) \overline{d}$ in $G$ is at distance at least $\varepsilon \overline{d}/4(C-1)\ge \varepsilon \overline{d}/4C + 1$ from $D$. The same holds if $C=1$ since $D = \emptyset$ then. Fix any integer $k\in [2,\varepsilon \overline{d}/4C]$ (this interval is non-empty since $\overline{d}\ge n\ge 8C/\varepsilon$). Under the assumption that at most $k$ vertices in $G_p$ have been made passive before exploring the neighbourhood of a particular active vertex $u$, at least
\begin{equation*}
\deg_G(u) - 1 - C - (C-1) (k-1)\ge \deg_G(u) - Ck\ge (1-\varepsilon/2) \overline{d} - \varepsilon \overline{d}/4\ge (1 - 3\varepsilon/4) \overline{d}
\end{equation*}
neighbours of $u$ have never been processed before (here, $1+C+(C-1)(k-1)\le Ck$ is an upper bound of the total number of processed vertices after $k$ steps, and $\deg_G(u)\ge (1-\varepsilon/2) \overline{d}$ since $u$ is at distance at most $k\le \varepsilon \overline{d}/4C$ from $v_0$ that satisfies $\deg_G(v_0)\ge (1-\varepsilon/4) \overline{d}$). Therefore, until the number of processed vertices is at most $\varepsilon \overline{d}/4C$, the number of edges of $G$, incident to the currently explored vertex $u$ and leading to vertices which have never been processed before, is at least $(1 - 3\varepsilon/4) \overline{d}$. We may conclude that the exploration of the connected component of $v_0$ in $G_p$, up to the moment of finding $\lceil \varepsilon \overline{d}/4C\rceil$ processed vertices, stochastically dominates the exploration of a BGW tree with progeny distribution $\mathrm{Bin}\left(\lceil(1-3\varepsilon/4)\overline{d}\rceil, \frac{1+7\varepsilon/8}{\overline{d}}\right)$. For every $\varepsilon\le 0.1$, the BGW tree with these parameters is supercritical since
\begin{equation*}
\lceil(1-3\varepsilon/4)\overline{d}\rceil\cdot \dfrac{1+7\varepsilon/8}{\overline{d}}\ge 1 + \dfrac{\varepsilon}{8} - \dfrac{21\varepsilon^2}{32} > 1,
\end{equation*}
and therefore it has probability $c_1 = c_1(\varepsilon) > 0$ to grow to infinity. Thus, with probability at least $c_1$, the exploration of $G_p$ from $v_0$ leads to at least $\varepsilon \overline{d}/4C$ processed vertices, which proves the lemma.
\end{proof}
Following~\cite{AKS}, we call a connected subgraph of $G_p$ a \emph{cell}. Note that a connected component of $G_p$ is a cell, but a cell does not have to be a connected component of $G_p$ itself. Fix the constant $c_1 = c_1(\varepsilon)$ given by Lemma~\ref{AKS lem 1}. We say that a vertex $v$ is a neighbour of a set of vertices $A$ in a graph $H$ if there is a vertex $u\in A$, which is a neighbour of $v$ in $H$. Let $P_p$ be the following property of a vertex $v$ of $G$:\\
\begin{center}
the vertex $v$ is a neighbour in $G$ to at least $c_1 \varepsilon n/64 C$ disjoint cells in $G_p$, each of order at least $\varepsilon n/8C$.
\end{center}
\begin{lemma}\label{AKS lem 2}
Fix $\varepsilon\in (0, 0.1)$ and $p = (1+\varepsilon)/\overline{d}$. There is a constant $c_2 = c_2(\varepsilon) > 0$ such that, for every large enough $n$, every vertex in $G$ of degree at least $(1-\varepsilon/8) \overline{d}$ has property $P_p$ with probability at least $1 - \exp(- c_2 n)$.
\end{lemma}
\begin{proof}
Fix a vertex $v$ satisfying $\deg_G(v)\ge (1-\varepsilon/8) \overline{d}$. Let $v = (v_1, v_2, \dots, v_n)$. For every $i\in [n]$, let $u_i$ be a neighbour of $v_i$ in $G_i$, and define
\begin{equation*}
H_i = \left(\underset{j\in [i-1]}{\square} v_j\right)\square\hspace{0.2em} u_i\hspace{0.2em}\square\left(\underset{k\in [n]\setminus [i]}{\square} G_k\right).
\end{equation*}
Thus, the graph $H_i$ is a conveniently chosen projection of $G$, which is isomorphic to $\underset{k\in [n]\setminus [i]}{\square} G_k$. Also, for every $i\in [n]$, let $\hat{v}_i = (v_1, \dots, v_{i-1}, u_i, v_{i+1}, \dots, v_n)$. Then, we claim that, for every $i\le i_{\max} := \lfloor\varepsilon n/16C\rfloor - 1$, the vertex $\hat{v}_i$ has degree at least $(1-\varepsilon/8) \overline{d} - C(i+1)\ge (1-\varepsilon/4) \overline{d}(H_i)$ in $H_i$. Indeed,
\begin{align*}
\left(1-\dfrac{\varepsilon}{8}\right)\overline{d} - C(i+1) & \ge\hspace{0.3em} \left(1-\dfrac{\varepsilon}{8}\right)\overline{d}(H_i) - C(i+1)\\
& \ge\hspace{0.3em} \left(1-\dfrac{\varepsilon}{8}\right)\overline{d}(H_i) - \dfrac{\varepsilon n}{16}\\
& \ge\hspace{0.3em} \left(1-\dfrac{\varepsilon}{4}\right)\overline{d}(H_i) + \dfrac{\varepsilon}{8}\left(\overline{d}(H_i) - \dfrac{n}{2}\right)\\
& \ge\hspace{0.3em} \left(1-\dfrac{\varepsilon}{4}\right)\overline{d}(H_i) + \dfrac{\varepsilon}{8}\left(n- i - \dfrac{n}{2}\right)\ge \left(1-\dfrac{\varepsilon}{4}\right)\overline{d}(H_i).
\end{align*}
Moreover, by the choice of $i$ we also have
\begin{equation*}
(1+\varepsilon) \overline{d}(H_i)\ge (1+\varepsilon) (\overline{d} - C(i+1))\ge (1+\varepsilon)\overline{d} - 2C(i+1)\ge \left(1+\dfrac{7 \varepsilon}{8}\right) \overline{d}.
\end{equation*}
Thus, $p = (1+\varepsilon)/\overline{d}\ge (1+7\varepsilon/8)/\overline{d}(H_i)$, and we may apply Lemma~\ref{AKS lem 1} to the vertex $\hat{v}_i$ in $H_i$ and deduce that the probability that $\hat{v}_i$ participates in a cell in $H_i$ of order at least $\varepsilon \overline{d}(H_i)/4C\ge \varepsilon n/8C$ is at least $c_1$. Since the graphs $(H_i)_{i\in [i_{\max}]}$ are disjoint, the events
\begin{equation*}
\left(A_i := \{\text{the connected component of } \hat{v}_i \text{ in } H_i \text{ contains at least } \varepsilon n/8C \text{ vertices}\}\right)_{i\in [i_{\max}]}
\end{equation*}
are independent and each of them happens with probability at least $c_1$. Thus, by Chernoff's inequality (Lemma~\ref{chernoff}) for $(\mathds 1_{A_i})_{i\in i_{\max}}$ with $t = \mathbb E[\sum_{i=1}^{i_{\max}} \mathds 1_{A_i}]/2\ge c_1 i_{\max}/2$, the vertex $v$ is incident to at least $c_1 i_{\max}/2\ge c_1 (\varepsilon n/32 C)/2 = c_1 \varepsilon n/64 C$ disjoint cells in its neighbourhood in $G$ with probability at least $1-\exp(-t/8)\ge 1 - \exp(-c_1 \varepsilon n/512 C)$. Thus, $c_2 = c_1 \varepsilon / 512 C$ satisfies our requirements, and the lemma is proved.
\end{proof}
Fix the constant $c_2 = c_2(\varepsilon) > 0$, given by Lemma~\ref{AKS lem 2}.
\begin{corollary}\label{AKM cor 2 bar}
Fix $\varepsilon\in (0,0.1)$ and $p = (1+\varepsilon)/\overline{d}$. Then, every vertex $v\in V(G)$ with $\deg_G(v)\ge (1-\varepsilon/16)\overline{d}$ satisfies the following property with probability at least $1 - \exp(-(c_2+o(1))n)$: $v$ is a neighbour in $G$ to at least $c_1\varepsilon n/64 C$ disjoint cells of $G_p$, each containing at least $\varepsilon n/17 C$ vertices with the property $P_p$.
\end{corollary}
\begin{proof}
Fix any vertex $v$ satisfying $\deg_G(v)\ge (1-\varepsilon/16)\overline{d}$. By Lemma~\ref{AKS lem 2} it has probability at least $1-\exp(-c_2n)$ to have property $P_p$. We condition on this event. Then, for every cell $\mathcal C$ among the first $\lceil c_1\varepsilon n/64 C\rceil$ disjoint neighbouring cells of size at least $\varepsilon n/8 C$, corresponding to $v$, put a label $\ell_v$ on the $\lceil\varepsilon n/17 C\rceil$ vertices of $\mathcal C$ that are closest to $v$ in the graph $G_p$ (if some set of vertices is at the same distance to $v$ in $G_p$, make an arbitrary choice which of them to label, if necessary). Thus, for every vertex $u$ which has received a label $\ell_v$ we have $d_G(u,v)\le d_{G_p}(u,v)\le \lceil\varepsilon n/17 C\rceil$. Moreover, by Corollary~\ref{cor neighbours} for every large enough $n$ we have $|\deg(u) - \deg(v)|\le C\lceil\varepsilon n/17 C\rceil\le \varepsilon n/16$ and so $\deg(u)\ge \deg(v) - \varepsilon n/16\ge \overline{d} - \varepsilon \overline{d}/16 - \varepsilon n/16 \ge (1 - \varepsilon/8) \overline{d}$.
Note that a total of at most $\lceil c_1\varepsilon n/64 C\rceil\cdot (\varepsilon n/16 C) = \Theta(n^2)$ vertices will receive the label $\ell_v$, and furthermore by Lemma~\ref{AKS lem 2} each of these vertices has property $P_p$ with probability at least $1 - \exp(-c_2 n)$. Then, conditionally on the event that $v$ has property $P_p$, any vertex $u$ with label $\ell_v$ has property $P_p$ with probability
\begin{align*}
& \mathbb P(u \text{ has }P_p\hspace{0.2em}|\hspace{0.2em} v\text{ has } P_p) = \dfrac{\mathbb P(u \text{ and } v \text{ have }P_p)}{\mathbb P(v\text{ has } P_p)}\ge 1 - 2\exp(-c_2 n).
\end{align*}
Thus, the vertex $v$ satisfies the property from the statement of the corollary with probability at least
\begin{equation*}
1 - \sum_{u \text{ has label }\ell_v} \mathbb P(u \text{ does not have }P_p\hspace{0.2em}|\hspace{0.2em} v\text{ has } P_p)\ge 1 - \Theta(n^2)\exp(-c_2 n) = 1 - \exp(-(c_2+o(1)) n).
\end{equation*}
The corollary is proved.
\end{proof}
With the help of Corollary~\ref{AKM cor 2 bar}, we are ready to improve on Lemma~\ref{AKS lem 2} by showing that every vertex of sufficiently high degree in $G$ has, with high probability, many neighbours in $G$, which participate in connected components of $G_p$ of order $\Omega(n^2)$. Denote $c'_1 = \min(c_1(\varepsilon/2), 1)$ and $c'_2 = c_2(\varepsilon/2)$.
\begin{lemma}\label{AKS lem 3}
Fix $\varepsilon\in (0,0.1)$ and $p = (1+\varepsilon)/\overline{d}$. There are constants $c_3 = c_3(\varepsilon) > 0$ and $c_4 = c_4(\varepsilon) > 0$ such that for every vertex $v$ satisfying $\deg_G(v)\ge (1-\varepsilon/32) \overline{d}$, the following property holds with probability at least $1 - \exp(-(c_4+o(1))n)$: $v$ is adjacent (in $G$) to at least $c_3 n$ vertices, participating in connected components of $G_p$ of order at least $c_3\varepsilon n^2/32 C$.
\end{lemma}
\begin{proof}
We use the technique of two-round exposure with $p_1 = (1 + \varepsilon/2)/\overline{d}$ and $p_2$ given by the equation $(1-p_1)(1-p_2) = (1-p)$. Since $\overline{d}\to +\infty$ with $n$, $p_2 = (\varepsilon/2+o(1))/\overline{d}$, so for every large enough $n$ we have $p_2\ge \varepsilon/4\overline{d}$.
Fix a vertex $v$ satisfying $\deg_G(v)\ge (1-\varepsilon/32)\overline{d}$. By Corollary~\ref{AKM cor 2 bar}, applied with $\varepsilon/2$ instead of $\varepsilon$, we get that with probability $1 - \exp(-(c'_2+o(1))n)$ the vertex $v$ is a neighbour (in $G$) to at least $c'_1\varepsilon n/128 C$ disjoint cells of $G_{p_1}$, each containing at least $\varepsilon n/34 C$ vertices with the property $P_{p_1}$. We condition on this event. Fix any such vertex $v$ and let $\mathcal C_1, \dots, \mathcal C_k$ be the cells, which correspond to $v$ in the above statement, with $k\ge c'_1\varepsilon n/128 C$.
Fix an arbitrary cell, say $\mathcal C_1$, and let $u_1, u_2, \dots, u_m$ be vertices in $\mathcal C_1$ which satisfy the property $P_{p_1}$, where $m = \lfloor \varepsilon n/34 C\rfloor$. Moreover, assume that for every vertex $u_i$ and every cell $\mathcal C$ in a fixed set of $\lceil c'_1\varepsilon n/128 C\rceil$ disjoint cells, which witnesses that $u_i$ satisfies the property $P_{p_1}$, $\mathcal C$ contains exactly $\lceil \varepsilon n/16 C\rceil$ vertices (clearly any connected graph $H$ contains a connected subgraph of any order between $1$ and $|V(H)|$). For every $i\in [m]$, we associate the above set of cells to the vertex $u_i$.
Now, we consider the independent percolation of the edges in $G$ with parameter $p_2$. This is our second round. We do the following exploration process. List $u_1, \dots, u_m$ in this order and start exploring their neighbourhoods one by one. If $u_1$ connects (during the second round percolation with parameter $p_2$) to a neighbouring cell of size $\lceil \varepsilon n/16 C\rceil$, which was associated to it, then name this cell $\mathcal C'_1$. Then, go to $u_2$. If $\mathcal C'_1$ was well defined, there are at most two cells $\mathcal C'$ among the ones, associated to $u_2$, such that $|\mathcal C'\cap \mathcal C'_1|\ge \lceil \varepsilon n/16 C\rceil/2$. Let $(w^2_j)_{j\in [2]}$ be the two vertices, which connect $u_2$ to neighbouring cells, associated to $u_2$ and with the largest intersection with $\mathcal C'_1$. Then, if $u_2$ connects to a neighbouring cell associated to it via an edge, different from $u_2w^2_1$ and $u_2w^2_2$, name this cell $\mathcal C'_2$ and go to $u_3$. Then, since $|\mathcal C'_1\cup \mathcal C'_2|\le 2\lceil \varepsilon n/16 C\rceil$, there are at most $4$ cells $\mathcal C'$, associated to $u_3$, for which $|\mathcal C'\cap (\mathcal C'_1\cup \mathcal C'_2)|\ge \lceil \varepsilon n/16 C\rceil/2$. Let $(w^3_j)_{j\in [4]}$ be the four vertices, which connect $u_3$ to neighbouring cells, associated to $u_3$ and with the largest possible intersection with $\mathcal C'_1\cup \mathcal C'_2$. Then, if $u_3$ connects to a neighbouring cell associated to it via an edge, different from $(u_3w^3_j)_{j\in [4]}$, name this cell $\mathcal C'_3$ and continue with $u_4$, ets.
Suppose that in the moment of exploring the neighbourhood of the vertex $u_i$ we came across the cells $\mathcal C'_{i_1}, \mathcal C'_{i_2}, \dots, \mathcal C'_{i_j}$ for some $1\le i_1 < \dots < i_j \le i-1$. Then,
\begin{equation*}
\bigg| \bigcup_{1\le s\le j} \mathcal C'_{i_s} \bigg| = \sum_{1\le s\le j} \bigg|\mathcal C'_{i_s}\setminus \left(\mathcal C'_{i_1}\cup \dots\cup \mathcal C'_{i_{s-1}}\right)\bigg|\ge \dfrac{j\varepsilon n}{32 C}.
\end{equation*}
Also, for every $i\le i_0 := \lfloor c'_1\varepsilon n/512 C\rfloor$ (by definition of $c'_1$ we have $i_0\le m$), there are at least $c'_1\varepsilon n/128 C - 2 c'_1\varepsilon n/512 C\ge c'_1\varepsilon n/256 C$ cells, associated to the vertex $u_i$, which do not intersect the union of cells $\mathcal C'_{i_1}, \mathcal C'_{i_2}, \dots, \mathcal C'_{i_j}$ in more than $\lceil \varepsilon n/16 C\rceil/2$ vertices. Thus, for every large enough $n$ and every $i\le i_0$, $u_i$ has probability at least $p_3 := p_2 c'_1\varepsilon n/256C\ge c'_1\varepsilon^2/1024 C^2$ to connect to a cell, which does not intersect the union of $\mathcal C'_{i_1}, \mathcal C'_{i_2}, \dots, \mathcal C'_{i_j}$ in more than $\lceil \varepsilon n/16 C\rceil/2$ vertices. We conclude that the indicator functions of the events
\begin{equation*}
\left(\left\{u_i \text{ connects to a cell } \mathcal C' \text{ associated to it and such that } |\mathcal C'\cap \left(\cup_{1\le \ell\le j} \mathcal C_{i_{\ell}}\right)| < \dfrac{\lceil \varepsilon n/16 C\rceil}{2}\right\}\right)_{i\in [i_0]}
\end{equation*}
stochastically dominate a family of i.i.d. random variables $(B_i)_{i\in [i_0]}$ with Bernoulli distribution with parameter $p_3$. By a direct application of Chernoff's inequality (Lemma~\ref{chernoff}) we conclude that for every large enough $n$ and $c_3 = c_3(\varepsilon) = (c'_1)^2\varepsilon^3/2^{21} C^3$
\begin{equation*}
\mathbb P\left(\sum_{i\in [i_0]} B_i\le \frac{p_3i_0}{2}\right)\le \exp\left(-\frac{p_3i_0}{8}\right)\le \exp\left(-\dfrac{(c'_1\varepsilon n/2)\cdot (c'_1 \varepsilon^2)}{8\cdot 512C\cdot 1024 C^2}\right) = \exp(-c_3 n/4).
\end{equation*}
Thus, for every large enough $n$ and every $\ell\in [k]$, the connected component of $\mathcal C_{\ell}$ in $G_p = G_{p_1}\cup G_{p_2}$ contains at least $(p_3 i_0/2)\cdot (\varepsilon n/32 C)\ge c_3\varepsilon n^2/32 C$ vertices with probability at least $1-\exp(-c_3 n/4)$. A union bound over all $k\le Cn$ cells shows that, for every large enough $n$ and for every vertex $v$ satisfying $\deg_G(v)\ge (1-\varepsilon/32) \overline{d}$ and with at least $c'_1\varepsilon n/128 C$ neighbouring cells of order $\lceil\varepsilon n/16 C\rceil$ in $G_{p_1}$, with probability at least $1 - Cn\exp(-c_3 n/4)$ each of the connected components of $\mathcal C_1, \mathcal C_2, \dots, \mathcal C_k$ in $G_p$ is of order at least $c_3\varepsilon n^2/32 C$.
Now, let $c_4 = \min(c'_2, c_3/4)$. Then, with probability at least $1 - \exp(-(c_4+o(1)) n)$, the vertex $v$ is incident to at least $c_3 n$ vertices, which participate in connected components of $G_p$ of size at least $c_3\varepsilon n^2/32 C$. The lemma is proved.
\end{proof}
Up to this moment, we ensured the existence of a large number of connected components of order at least $\Theta(n^2)$ in $G_p$. Recall that $\gamma$ is a positive constant such that, for every $j\in [n]$, $n^{-\gamma}\le i(G_j)$. If $\gamma\in (0,1)$, we are ready to complete the proof of Theorem~\ref{main thm}. However, for larger values we will need to show more. In the sequel we ensure that there are a lot of components of size $\Omega(n^{\gamma'+2})$ in $G_p$ for some $\gamma' > \gamma$. The aim of the next lemma is to iterate the procedure of Lemma~\ref{AKS lem 2}, Corollary~\ref{AKM cor 2 bar} and Lemma~\ref{AKS lem 3} to provide the a.a.s. existence of these larger connected components. Unlike the results we presented above, we will mostly rely on the asymptotic notations $\Theta$ and $\Omega$ in the proof of Lemma~\ref{AKS lemma plus} rather than give explicit constants to simplify the presentation, having in mind that very similar but more precise formulations of the claims below were already presented in detail.
\begin{lemma}\label{AKS lemma plus}
Fix any integer $k\ge 2$, any $\varepsilon\in (0, 0.1)$ and $p = (1+\varepsilon)/\overline{d}$. Then, there are positive constants $\beta_k\ge 32, C_k = C_k(\varepsilon), C'_k = C'_k(\varepsilon), C''_k = C''_k(\varepsilon)$ such that for every vertex $v$ satisfying $\deg_G(v)\ge (1-\varepsilon/\beta_k) \overline{d}$, the following property holds with probability at least $1 - \exp(-(C''_k+o(1))n)$: $v$ is adjacent (in $G$) to at least $C'_k n$ vertices, participating in connected components in $G_p$ of order at least $C_k n^k$.
\end{lemma}
\begin{proof}
We argue by induction. By Lemma~\ref{AKS lem 3} the statement is true for $k=2$ with parameters $\beta_2 = 32, C_2 = c_3\varepsilon/32C, C'_2 = c_3$ and $C''_2 = c_4$ for every $\varepsilon\in (0, 0.1)$.
Suppose that the statement is satisfied for some $k-1\ge 2$.
Fix $p_0 = (1+\varepsilon/4)/\overline{d}$, $p'_0 = (\varepsilon/4+o(1))/\overline{d}$ and $p_1 = (1+\varepsilon/2)/\overline{d}$ so that $(1-p_0)(1-p'_0) = 1 - p_1$. Moreover, for any vertex $v$ in $G$, denote by $P_{p,k}$ the following property:
\begin{center}
the vertex $v$ is a neighbour in $G$ of $\Omega(n)$ disjoint cells in $G_p$, each of order $\Omega(n^k)$.
\end{center}
The next claim is an analogue of Lemma~\ref{AKS lem 2}, so we give only the main points of the proof.
\begin{claim}\label{AKS claim 2}
Every vertex of degree at least $(1-\varepsilon/8\beta_{k-1})\overline{d}$ in $G$ has property $P_{p_1, k-1}$ with probability $1 - \exp(-\Omega(n))$.
\end{claim}
\begin{proof}
We follow the proof of Lemma~\ref{AKS lem 2}. Fix a vertex $v = (v_1, v_2, \dots, v_n)$ satisfying $\deg_G(v)\ge (1-\varepsilon/8\beta_{k-1})\overline{d}$. For every $i\in [n]$, let $u_i$ be a neighbour of $v_i$ in $G_i$. Denote $\hat{v}_i = (v_1, \dots, v_{i-1}, u_i, v_{i+1}, \dots, v_n)$ and
\begin{equation*}
H_i := \left(\underset{j\in [i-1]}{\square} v_j\right)\square\hspace{0.2em} u_i\hspace{0.2em}\square\left(\underset{k\in [n]\setminus [i]}{\square} G_k\right).
\end{equation*}
Then, for every $\varepsilon\in (0, 0.1)$ there exists a positive constant $\overline{c}_{k-1} = \overline{c}_{k-1}(\varepsilon)\le 1/2$ such that, for every $i\le \overline{c}_{k-1}n$, the degree of $\hat{v}_i$ in $H_i$ is at least $(1-\varepsilon/4\beta_{k-1})\overline{d}(H_i) = (1-(\varepsilon/4)/\beta_{k-1})\overline{d}(H_i)$. By the induction hypothesis, applied with $\varepsilon/4$, $p_0$, $\hat{v}_i$ and $H_i$ (which is isomorphic to the product of at least $n-i\ge n/2$ of the graphs $(G_j)_{j\in [n]}$), the vertex $\hat{v}_i$ is incident to at least $\Theta(n/2)$ vertices, participating in connected components in $H_{i,p_0}$ of order $\Omega((n/2)^{k-1})$ with probability $1 - \exp(-\Omega(n/2))$. (Note that although $\Omega(n) = \Omega(n/2)$ and $\Omega(n^{k-1}) = \Omega((n/2)^{k-1})$, we add the constants to indicate that the graph $H_i$ is a product of less that $n$, but at least $n/2$ graphs. When considered appropriate, similar implicit indications are given below as well).
It remains to notice that the graphs $(H_i)_{1\le i\le \overline{c}_{k-1}n}$ are disjoint and therefore the vertices $(\hat{v}_i)_{1\le i\le \overline{c}_{k-1}n}$ connect to cells of order $\Omega((n/2)^{k-1})$ in $(H_i)_{1\le i\le \overline{c}_{k-1}n}$ respectively at the second round percolation with parameter $p'_0$ independently and with probability $p'_0 \Omega(n/2) = \Omega(1)$. Thus, by Chernoff's inequality (Lemma~\ref{chernoff}) we deduce that the vertex $v$ has $\Omega(1) \cdot \overline{c}_{k-1} n/2 = \Omega(n)$ neighbours, which participate into disjoint cells of $G_{p_1} = G_{p_0}\cup G_{p'_0}$ of size $\Omega((n/2)^{k-1})$ with probability $1 - \exp(-\Omega(1) \cdot \overline{c}_{k-1} n/8) = 1 - \exp(-\Omega(n))$. The proof is completed.
\end{proof}
\begin{claim}\label{AKS claim 2 bar}
Every vertex $v$ with $\deg_G(v)\ge (1-\varepsilon/16\beta_{k-1})\overline{d}$ satisfies the following property with probability $1 - \exp(-\Omega(n))$: the vertex $v$ is a neighbour in $G$ to $\Omega(n)$ disjoint cells of $G_{p_1}$, each containing $\Omega(n)$ vertices with the property $P_{p_1, k-1}$.
\end{claim}
\begin{proof}
We follow the proof of Corollary~\ref{AKM cor 2 bar}.
Fix any vertex $v$ satisfying $\deg_G(v)\ge (1-\varepsilon/16\beta_{k-1})\overline{d}$. Since $\beta_{k-1}\ge 32$, by Lemma~\ref{AKS lem 2} $v$ has probability $1-\exp(-\Omega(n))$ to have property $P_{p_1}$. We condition on this event. Then, for every cell $\mathcal C$ among the $\Omega(n)$ disjoint neighbouring cells of order at least $\lceil\varepsilon n/16C\rceil$, corresponding to $v$, put a label $\ell_v$ on the $\lceil\varepsilon n/(16 C\beta_{k-1} + 1)\rceil$ vertices of $\mathcal C$ that are closest to $v$ in the graph $G_{p_1}$ (if some set of vertices is at the same distance to $v$ in $G_p$, make an arbitrary choice which of them to label, if necessary). Thus, for every vertex $u$ which has received a label $\ell_v$ we have $d_G(u,v)\le d_{G_{p_1}}(u,v)\le \lceil\varepsilon n/(16 C\beta_{k-1} + 1)\rceil$. Moreover, by Corollary~\ref{cor neighbours} for every large enough $n$ we have
$$|\deg(u) - \deg(v)|\le C\lceil\varepsilon n/(16 C\beta_{k-1} + 1)\rceil\le \varepsilon n/16 \beta_{k-1}$$
and so
$$\deg(u)\ge \deg(v) - \varepsilon n/16 \beta_{k-1}\ge \overline{d} - \varepsilon \overline{d}/16 \beta_{k-1} - \varepsilon n/16 \beta_{k-1} \ge (1 - \varepsilon/8 \beta_{k-1}) \overline{d}.$$
Note that a total of $O(n^2)$ vertices will receive the label $\ell_v$, and furthermore by Claim~\ref{AKS claim 2} each of these vertices has property $P_{p_1, k-1}$ with probability $1 - \exp(-\Omega(n))$. Then, conditionally on the event that $v$ has property $P_{p_1}$, any vertex $u$ with label $\ell_v$ has property $P_{p_1, k-1}$ with probability
\begin{align*}
& \mathbb P(u \text{ has }P_{p_1, k-1}\hspace{0.2em}|\hspace{0.2em} v\text{ has } P_{p_1}) = \dfrac{\mathbb P(u \text{ has } P_{p_1, k-1} \text{ and } v \text{ has }P_{p_1})}{\mathbb P(v\text{ has } P_{p_1})}\ge 1 - 2\exp(-\Omega(n)) = 1 - \exp(-\Omega(n)).
\end{align*}
Thus, the vertex $v$ satisfies the property from the statement of the claim with probability at least
\begin{equation*}
1 - \sum_{u \text{ has label }\ell_v} \mathbb P(u \text{ does not have }P_{p_1, k-1}\hspace{0.2em}|\hspace{0.2em} v\text{ has } P_{p_1})\ge 1 - O(n^2)\exp(-\Omega(n)) = 1 - \exp(-\Omega(n)).
\end{equation*}
The claim is proved.
\end{proof}
The finish the proof of the lemma, we follow the ideas of the proof of Lemma~\ref{AKS lem 3}. We use once again the technique of two-round exposure with $p_1 = (1 + \varepsilon/2)/\overline{d}$ and $p_2 = (\varepsilon/2+o(1))/\overline{d}$ such that $(1-p_1)(1-p_2) = (1-p)$.
Fix a vertex $v$ satisfying $\deg_G(v)\ge (1-\varepsilon/32 \beta_{k-1})\overline{d}$. By Claim~\ref{AKS claim 2 bar} we get that, for some positive constant $\hat{C}_{k-1} = \hat{C}_{k-1}(\varepsilon)$, with probability $1 - \exp(-\Omega(n))$ the vertex $v$ is a neighbour (in $G$) to at least $\hat{C}_{k-1} n$ disjoint cells of $G_{p_1}$, each containing $\Omega(n)$ vertices with the property $P_{p_1, k-1}$. Let us condition on this event, and let $\mathcal C_1, \dots, \mathcal C_t$ be the cells, which correspond to $v$ in the above statement, where $t$ is an integer satisfying $\hat{C}_{k-1} n\le t\le Cn$.
Fix an arbitrary cell among $\mathcal C_1, \dots, \mathcal C_t$, say $\mathcal C_1$, and let $u_1, u_2, \dots, u_m$ be vertices in $\mathcal C_1$ which satisfy the property $P_{p_1, k-1}$, where $m = \Omega(n)$. Moreover, assume that for every vertex $u_i$ and every cell $\mathcal C$ among a fixed set of $\Omega(n)$ disjoint cells, which witnesses that $u_i$ satisfies the property $P_{p_1, k-1}$, $\mathcal C$ contains exactly $\lceil \hat{C}'_{k-1} n^{k-1}\rceil$ vertices, where $\hat{C}'_{k-1}$ is a positive constant depending only on $k$ and $\varepsilon$. By the very same exploration procedure as in the proof of Lemma~\ref{AKS lem 3} we show that with probability $1 - \exp(-\Omega(n))$ the connected component of the cell $\mathcal C_1$ in $G_p = G_{p_1}\cup G_{p_2}$ contains at least $s = \Omega(n)$ cells $\mathcal C'_1, \mathcal C'_2, \dots, \mathcal C'_s$ such that, for every $i\in [s]$,
\begin{equation*}
|V(\mathcal C'_i\setminus \cup_{j\in [i-1]} \mathcal C'_j)|\ge \dfrac{\lceil \hat{C}'_{k-1} n^{k-1}\rceil}{2}.
\end{equation*}
Recall that $\mathcal C'_1, \mathcal C'_2, \dots, \mathcal C'_s$ are cells, associated to different vertices among $u_1, u_2, \dots, u_m$, and therefore these are not necessarily disjoint. Thus, bounding from below the number of vertices in every new cell that does not participate in the union of the previous ones is crucial to attain the lower bound on the size of the union.
Since the above reasoning applies to each of the cells $\mathcal C_1, \mathcal C_2, \dots, \mathcal C_t$, associated to $v$, and $t = \Theta(n)$, we conclude by union bound that $v$ satisfies property $P_{p,k}$ with probability $1 - \Theta(n)\exp(-\Omega(n)) = 1 - \exp(-\Omega(n))$, which finishes the proof (the constants $C_k, C'_k$ and $C''_k$ are hidden in the $\Omega$ notation but $\beta_k$ could be defined recursively by $\beta_k = 32\beta_{k-1}$, so in particular $\beta_k = 32^{k-1}$).
\end{proof}
We are ready to prove Theorem~\ref{main thm}~b). Fix $k = \lceil 1+\gamma\rceil + 3$. Up to now, we have ensured the existence of a number of cells in $G_p$, which contain at least $C_k n^k$ vertices. In the sequel, let $\hat{C}_k = C_k(\varepsilon/2), \hat{C}'_k = C'_k(\varepsilon/2)$ and $\hat{C}''_k = C''_k(\varepsilon/2)$.
\begin{proof}[Proof of Theorem~\ref{main thm}~b)]
It is sufficient to prove the claim for every $\varepsilon\in (0, 0.1)$. Once again, we consider two-round exposure of $G_p = G_{p_1}\cup G_{p_2}$ with $p_1 = (1+\varepsilon/2)/\overline{d}$ and $p_2 = (\varepsilon/2+o(1))/\overline{d}$. By Lemma~\ref{AKS lemma plus} and Markov's inequality with probability at least $1 - \exp(-(\hat{C}''_k/2+o(1))n)$ all but at most an $\exp(-\hat{C}''_k n/2)$-proportion of all vertices of degree at least $(1-(\varepsilon/2)/\beta_k)\overline{d}$ in $G$ have at least $\hat{C}'_k n$ neighbours, which participate in connected components of $G_{p_1}$ of order at least $\hat{C}_k n^k$. We condition on this event. Then, the number of edges, adjacent to vertices in connected components of $G_{p_1}$ of order at least $\hat{C}_k n^k$ is at least $(1+o(1)) \hat{C}'_k n |V(G)|/2$ (recall that asymptotically almost all vertices of $G$ have degree at least $(1-(\varepsilon/2)/\beta_k)\overline{d}$). Since every vertex has degree at most $C n$ in $G$, there are at least $(1+o(1)) \hat{C}'_k |V(G)|/2C$ vertices of $G$ in connected components of $G_{p_1}$ of order at least $\hat{C}_k n^k$.
We prove that the following property holds with probability $1 - \exp(-\Omega(|V(G)|/n^k))$:\\
\noindent
the vertices in all connected components in $G_{p_1}$ of order at least $\hat{C}_k n^k$ cannot be partitioned into two sets, $V_1$ and $V_2$, such that $||V_1|-|V_2||\le (|V_1|+|V_2|)/3$ (or equivalently $|V_1|/2\le |V_2|\le |V_1|$ up to symmetry considerations) so that there is no path in $G_p$ between $V_1$ and $V_2$.\\
\noindent
On the above event, by Observation~\ref{ob discrete cont} we may directly conclude that the largest connected component of $G_p$ contains at least $(|V_1|+|V_2|)/3\ge (1+o(1)) \hat{C}_k |V(G)|/6 C$ vertices, which would finish the proof of Theorem~\ref{main thm}~b).
Since the number of connected components of $G_{p_1}$ of order at least $\hat{C}_k n^k$ is $O(|V(G)|/n^k)$, there are $2^{O(|V(G)|/n^k)}$ ways to partition these components into two sets. We will be interested only in partitions $(V_1, V_2)$ such that $|V_1|/2\le |V_2|\le |V_1|$. Consider two cases:
\begin{enumerate}
\item $N_{G}[V_1]\cap N_{G}[V_2]\ge |V(G)|/n^{k-2}$, and
\item $N_{G}[V_1]\cap N_{G}[V_2] < |V(G)|/n^{k-2}$.
\end{enumerate}
In the first case, we know by our conditioning that $(1 + o(1))|V(G)|/n^{k-2}$ of the vertices in $N_{G}[V_1]\cap N_{G}[V_2]$ have at least $\hat{C}'_k n/2$ neighbours (in $G$) in either $V_1$ or $V_2$, or in both. Therefore, the probability that a fixed vertex in $N_{G}[V_1]\cap N_{G}[V_2]$ connects $V_1$ and $V_2$ at the second round percolation with parameter $p_2$ is at least $(1 - (1-p_2)^{\hat{C}'_k n/2})\cdot p_2 = \Theta(1/n)$. Moreover, the above events are independent for different vertices in $N_{G}[V_1]\cap N_{G}[V_2]$. Therefore, the probability that $V_1$ and $V_2$ do not get connected at the second round percolation with parameter $p_2$ is at most
\begin{equation*}
\left(1-\Theta(1/n)\right)^{(1+o(1))|V(G)|/n^{k-2}} = \exp\left(-\Theta(|V(G)|/n^{k-1})\right) \ll 2^{-O(|V(G)|/n^k)}.
\end{equation*}
In the second case, since the number of edges between $V_2$ and $N_G(V_2)$ is at least $i(G) |V_2|$ by assumption, by Corollary~\ref{isoper cor} there are at least $i(G) |V_2|/Cn \ge n^{-1-\gamma} |V_2|/2C$ vertices in $V(G)\setminus V_2$, adjacent to $V_2$ in $G$. But
$$|N_{G}[V_1]\cap N_{G}[V_2]| < |V(G)|/n^{k-2}\ll n^{-1-\gamma} |V_2|/2C,$$
so by our conditioning each of the $(1+o(1)) n^{-\gamma-1} |V_2|/2C$ vertices in $N_G[V_2]\setminus N_G[V_1]$ have at least $\hat{C}'_k n$ edges towards $V_2$ in $G$. On the other hand, since $|V_1|\ge |V_2|$ and $|V_1\cap N_G[V_2]| = o(|V_1|)$, we have by Corollary~\ref{isoper cor} that there are at least
$$i(G) \min(|N_G[V_2]\setminus N_G[V_1]|, (1+o(1))|V_1|) = \Omega(n^{-\gamma} |V(G)|)$$
edges, going out of $N_G[V_2]\setminus N_G[V_1]$. One may directly deduce that there are $\Omega(n^{-\gamma} |V(G)|)/Cn = \Omega(n^{-\gamma-1} |V(G)|)$ disjoint edges, which have one endvertex in $N_G[V_2]\setminus N_G[V_1]$ and one endvertex in $V(G)\setminus N_G[V_2]$. Since all but $\exp(-\Omega(n)) |V(G)|$ vertices have at least $\hat{C}'_k n$ edges towards $V_1\cup V_2$ by our conditioning, we deduce that there are $\Omega(n^{-\gamma - 1} |V(G)|)$ disjoint edges $uv$ in $G$ such that $u$ has at least $\hat{C}'_k n$ edges towards $V_1$ and $v$ has at least $\hat{C}'_k n$ edges towards $V_2$. We conclude that for any such edge $u$ and $v$ there is a path from $V_1$ through $u$ and $v$ towards $V_2$ with probability $(1 - (1-p_2)^{\hat{C}'_k n})\cdot p_2\cdot (1 - (1-p_2)^{\hat{C}'_k n}) = \Theta(1/n)$. Therefore, the probability that $V_1$ and $V_2$ do not get connected at the second round percolation with parameter $p_2$ is at most
\begin{equation*}
\left(1-\Theta(1/n)\right)^{\Omega(|V(G)|/n^{\gamma+1})} = \exp\left(-\Omega(|V(G)|/ n^{\gamma+2})\right) \ll 2^{-O(|V(G)|/n^k)}.
\end{equation*}
We conclude the proof of Theorem~\ref{main thm}~b) by a union bound over all $2^{O(|V(G)|/n^k)}$ partitions of the components of size at least $\hat{C}_k n^k$ in $G_{p_1}$.
\end{proof}
\section{Discussion and further questions}\label{sec discussion}
In this paper we proved that there is a sharp threshold for the existence of a giant component after percolation of the product graph $G = G_1\square \dots\square G_n$ under the assumptions that $\max_{j\in [n]} \Delta(G_j)$ is uniformly bounded from above by a constant and $\min_{j\in [n]} i(G_j)$ decays to zero at most polynomially fast. As Remark~\ref{rem 2.1} points out, at the price of a more technical exposition Theorem~\ref{main thm} may be generalised for graphs with slowly increasing degrees. Except for simplicity, we spared the details also because we believe that Theorem~\ref{main thm} may also be proved in an even more general setting.
To begin with, we were not able to find convincing counterexamples of the sharp threshold phenomenon without the maximum degrees assumption. In the proof of Theorem~\ref{main thm} presented above, this assumption was used in most of our lemmas.
\begin{question}
Can one prove an analogue of Theorem~\ref{main thm} without the assumption on the maximum degrees of $(G_j)_{j\in [n]}$?
\end{question}
Concerning the assumption on the decay of the isoperimetric constants, we show that it cannot be removed entirely. Consider the graph $G$ where $G_1 = G_2 = \dots = G_{n-1}$, each containing two vertices ($0$ and $1$) a single edge ($01$), and $G_n$ being a cycle of length $2^{2^n}$. Then, all vertices in $G$ will have degree $n+1$. Fix $p = 2/(n+1)$. Note that for any edge $uv$ of $G_n$ we have that the probability that each of the edges $((x,u)(x,v))_{x\in \{0,1\}^{n-1}}$ of $G$ disappears after $p$-percolation is $(1 - 2/(n+1))^{2^{n-1}} = \exp(-(1+o(1)) 2^n/(n+1))$. Thus, on average many of the sets of edges $((x,u)(x,v))_{x\in \{0,1\}^{n-1}}\subseteq E(G)$ for different edges $uv$ of $G_n$ disappear a.a.s. after $p$-percolation, so no giant component exists since for any two edges $u_1v_1$ and $u_2v_2$ of $G_n$, the edge set $\{(x,u_1)(x,v_1)\}_{x\in \{0,1\}^{n-1}}\cup \{(x,u_2)(x,v_2)\}_{x\in \{0,1\}^{n-1}}\subseteq E(G)$ forms a cut in $G$. Although somewhat trivial, this example leads to another logical question.
\begin{question}
Can one prove an analogue of Theorem~\ref{main thm} if $\min_{j\in [n]}i(G_j)$ decreases faster than a polynomial function of $n$?
\end{question}
Of course, graph products other than the Cartesian product exist as well. It might be interesting to study the appearance of a giant component with respect to them.
\begin{question}
Can one prove analogous results for other graph products?
\end{question}
\section{Acknowledgements}
The author would like to thank Dieter Mitsche, Guillem Perarnau and Ivailo Hartarsky for several useful remarks, and Felix Joos for turning my attention to the reference~\cite{Joo2}. I am also grateful to the two anonymous referees for a number of important comments and suggestions.
\bibliographystyle{plain}
|
\section{Introduction}
\label{S:1}
\indent According to the “International Vocabulary of Basic and General
Terms in Metrology" a measurement is defined as a “set of operations having the
object of determining a value of a quantity” \cite{dybkaer2011definitions}. Therefore, a result of a measurement is an estimate of the value of a measurand and should be accompanied by an uncertainty statement \cite{choi2003uncertainty}. The uncertainty of measurement is the doubt that exists about the result of any measurement. But for every measurement - even the most careful - there is always a margin of
doubt. It is a complicated subject, and still evolving. So there is a great need for a guide that provides clear, down-to-earth explanations, easy enough for non-expert readers \cite{bell2001beginner}. \textcolor{red}{While error is the difference between the measured value and the ‘true value’ of the quantity being measured.} The ‘true’ value can only be found from a very large set of measurements and thus any uncertainty whose value is not known is a measure of uncertainty. Hence, only an estimate of the standard deviation can be found from a moderate number of values and one standard deviation is an estimate of the uncertainty. The “Guide to the Expression of Uncertainty of Measurement” (GUM) provides the general rules for quantifying the uncertainty in a measurement \cite{bipm2008evaluation}. As a rule of thumb usually between 4 and 10 readings is sufficient to make a good quality measurement and its uncertainty. \\
\indent The uncertainty evaluation is an important part of a good quality measurement. It is only
complete if it is accompanied by a statement of the uncertainty. A measurand $Y$ is determined
from N other quantities $X_i$ using a functional relationship as follows
\begin{align}
Y = f(X_1, X_2, …, X_N)
\end{align}
where the quantities $X_1, X_2, …, X_N$ may themselves be viewed as measurands and may depend on other quantities like $X_i=f(x_i)$. Two approaches based on the evaluation method are used to estimate the standard
uncertainty: Type A evaluation is based on statistical treatment on a series of observations, and Type B evaluation is based on non-statistical one such as data from manufacturer’s specifications, previous experimental data, theoretical data, etc. Note that both Type A and Type B uncertainties can be due to a “random effect” as well as to a “systematic effect” in nature. Random effect is recognizable from the variations in the repeated observations, where uncertainty analysis of the observed data is governed by the rules of statistics. On the other hand, though the contribution of systematic effect appears in the repeated measurements, but we can learn nothing from these. Other methods are needed to estimate the uncertainties due to systematic effects.\\
\indent Inner shell ionization of an atom is a very interesting field of study for the theoreticians as well as experimentalists since last century. For a better understanding of the inner shell ionization process, a lot of work had been reported using different experimental pathways. In this endeavour, inner shell ionization induced by electron \cite{burhop1940inner,rester1966k,PhysRevA.11.1475,PhysRevLett.39.82,PhysRevA.77.042701,LIU20191,zhao2020measuring,tian2020measurements}, photon \cite{storm1967photon,VEIGELE197351,PhysRevA.31.2918,doi:10.1002/xrs.1300180503,doi:10.1002/xrs.1300230505,ertuugrul1996measurement}, proton \cite{PhysRev.165.66,garcia1970inner,PhysRevA.29.70,PhysRevA.35.2007,CHEN1989257,PhysRevA.42.261,MONTANARI2017236,MIRANDA2019,PEREZ201921} and varieties of heavy ions \cite{PhysRevLett.29.1361,ANDERSEN1976507,PhysRevA.30.2234,SARKAR199523,PhysRevA.30.2234,singh2000subshell,PhysRevA.73.032712,MSIMANGA201690,VERMA201830,GORLACHEV201831,chang2018multiple,oswal2018x,SILHADI2019158,OSWAL2020108809} had been reported. To explain the ionization processes, various theories like SCA \cite{KOCBACH1980281}, PWBA \cite{merzbacher1958handbuch}, ECPSSR \cite{brandt1979shell, brandt1981energy,lapicki2002status, PhysRevA.20.465,PhysRevA.23.1717}, ECUSAR \cite{lapicki2002status} and SLPA \cite{PhysRevA.67.062901} have been used. In a lot of studies, comparisons were made between the experimental cross-section data and corresponding theoretical predictions. Accuracy of experimental data is very important. Given uncertainty alongside the data gives the accuracy of such measurement. Finite uncertainty in the x-ray production cross-section data is indispensable due to the uncertainty in photo-peak area evaluation, efficiency of the detector, ion beam current, target thickness, etc. The ionization cross-sections can be extracted from the measured x-ray production cross-sections using suitable atomic parameters. Uncertainties in the x-ray production cross-sections and the used atomic parameters are propagated into the ionization cross-sections. Conversion of K shell x-ray production cross-section into the ionization cross-section requires only one atomic parameter, fluorescence yield \cite{merlet2004measurements}, hence evaluating the uncertainty is simple. In contrast, obtaining the L subshell ionization cross-section from the x-ray production cross-section data requires several atomic parameters including fluorescence yield, Coster-Kronig yield, the fraction of radiative transition rates, which in turn introduce the Type B uncertainty. Thus, the calculation of uncertainty is complex and significant discrepancies prevalent in literature. To examine the reason behind such large variation, we have made a thorough study of the propagation of uncertainties from the L x-ray production cross-sections to the L subshell ionization cross-section. Details will be seen in the next section, where the method of weighted uncertainty propagation rule has been applied. For example, we have taken results from recent work \citet{OSWAL2020108809} and assessed the quoted accuracy. More so, an attempt is put on suggesting the best possible way to perform the most accurate experiment. Probably, besides the L subshell ionization, this work will be a general guideline for uncertainty analysis of many other measurements relevant to physics or beyond.\\
\section{Present scenario of measurement uncertainty in L-shell ionization studies}
In the past, many experiments have studied the L-shell ionization in various targets bombarded by different projectiles in certain energy ranges and quoted the measurement uncertainties. A brief list is given in the table \ref{S:2}. Though the experiments have followed similar detection systems, but the measurement uncertainties vary in a wide range of 4-35\%. This scenario is a deterrent for improving the theoretical understanding.
\label{S:2}
\begin{landscape}
\begin{table}
\tiny
\caption{\label{tab:table1} Uncertainties quoted for L shell production and ionization cross sections in earlier experiments.}
\centering
\begin{tabular}{l l l l l l l l l l}
\hline
\textbf{Ref.} & \makecell{\textbf{Targets}}& \makecell{\textbf{Projectile, its energy}} & \makecell{\textbf{$\delta L_1$}, \textbf{$\delta L_2$}, \textbf{$\delta L_3$ (\%)}} & \makecell{\textbf{Remarks}}\\
\hline
\citet{PhysRevA.19.1930}& \makecell{W} & \makecell{electron, 11-40 keV} & \makecell{17, 15, 15} & \makecell{Nearly equal uncertainties\\ for all subshells}\\
\citet{palinkas1980subshell}& \makecell{Au, Pb, Bi} & \makecell{electron, 60-600 keV} & \makecell{22, 15, 11} & \makecell{7-10\% uncertainties for x-ray production \\ cross sections for all the three elements \\ }\\
\citet{reusch1986method}& \makecell{$29\leq Z \leq79$} & \makecell{electron, 50-200 keV} & \makecell{7-15, 6-15, 6-15}& \makecell{Nearly same uncertainties for all the subshells\\ and about 15\% for total L ionization cross section}\\
\citet{PhysRevA.37.106}& \makecell{W} & \makecell{electron, 12-40 keV}&\makecell{--} &\makecell{10\% uncertainty for $\sigma_3/\sigma_2$ and $\sigma_2/\sigma_1$}\\
\citet{llovet2014cross}& \makecell{$1\leq Z \leq99$}&\makecell{electrons and positron, \\up to 1 GeV}& \makecell{10-30, 10-30, 10-30}& \makecell{Similar uncertainties for all elements \\and all subshells}\\
\citet{barros2015ionization}& \makecell{Au}&\makecell{electron, 50-100 keV} &\makecell{22, 11, 12} & \makecell{Without inclusion of uncertainties \\ in atomic relaxation parameters}\\
\citet{barros2015ionization}& \makecell{Au} & \makecell{electron, 50-100 keV}& \makecell{23, 12, 13} &\makecell{With inclusion of uncertainty in atomic \\relaxation parameters \cite{rao1972atomic,kolbe2012subshell,scofield1974exchange,campbell1989interpolated}}\\
\citet{barros2015ionization}&\makecell{Au} &\makecell{electron, 50-100 keV} & \makecell{28, 16, 16} &\makecell{With inclusion of uncertainty in atomic \\ relaxation parameters as given in \cite{krause1979atomic,scofield1974relativistic,rao1972atomic}}\\
\citet{bernstein1954shell}& \makecell{Ta, Au, Pb, U} & \makecell{H, 1.5-4.25 MeV} &\makecell{--}& \makecell{20\% Uncertainty for all \\ x-ray production cross-sections}\\
\citet{PhysRevA.11.607}& \makecell{Ta, Au, Bi} & \makecell{H, 1 to 5.5 and \\ He, 1 to 11 MeV} & \makecell{22, 28, 21}& \makecell{Uncertainties for $L_\alpha$ is about \\ 15\%, for $L_{\gamma_{1+5}}$ and $L_{\gamma_{2+3+4}}$ are about 15-18\%}\\
\citet{PhysRevA.15.943}& \makecell{Au, Tl, Pb, Bi, Th, U}& \makecell{H, 0.5-3.5 MeV} &\makecell{--}& \makecell{7\% uncertainty for production cross section\\ and error bar for ionization cross sections \\given in figures only}\\
\citet{wheeler1979subshell}& \makecell{Pr, Sm, Tb, Ho, Yb} &\makecell{H, 150-400 keV}& \makecell{20, 20, 15} & \makecell{Same uncertainty for all the elements}\\
\citet{sokhi1984experimental}& \makecell{Different targets}&\makecell{H} & \makecell{6-20, 7-20, 6-26}& \makecell{Uncertainty differs with target}\\
\citet{miranda2014experimental}& \makecell{Ne to Am} & \makecell{H, 0 keV to 1 GeV} & \makecell{--}& \makecell{No regular trend seen in uncertainties}\\
\citet{semaniak1994a} & \makecell{Between La to Au}&\makecell{ $^{14}N$, 1.75-22.4 MeV}& \makecell{25-10, 25-15, 25-10} & \makecell{Uncertainties at low energy were \\10-25\% and 10\% at high energies}\\
\citet{Semaniak1995b}& \makecell{$72\leq Z \leq 90$}&\makecell{C and N, 0.4 to 1.8 MeV/amu} &\makecell{6-20, 5-12, 4-10}& \makecell{Uncertainties differs with target}\\
\citet{dhal1995subshell}& \makecell{Au, Bi}&\makecell{B, 4.8-8.8 MeV} & \makecell{10, 10, 10}& \makecell{Same for both the elements}\\
\citet{banas2002role}& \makecell{Au}&\makecell{O of 0.4-2.2 MeV/amu} &\makecell{15-30, 15-30, 15-30}& \makecell{10-20\% for all production cross-section}\\
\citet{pajek2003multiple}& \makecell{Au, Bi, Th, U}&\makecell{O, 6.4 to 70 MeV} & \makecell{--}& \makecell{Uncertainties for production cross-sections \\ of $L\alpha_{1,2}$ were 9\%, for $L_{\gamma_1}$ 13\% and \\ for $L_{\gamma_{2,3}}$ 10-25\% and\\ for subshell ionization cross section the \\error bars included in figure only}\\
\citet{fijal2008subshell}& \makecell{Au and Bi}&\makecell{S, 12.8-120 MeV} & \makecell{10-30, 10-30, 10-30}& \makecell{Uncertainties for production cross-sections \\ of $L\alpha_{1,2}$ were 7-15\%, for $L_{\gamma_1}$ 10-18\% and \\ for $L_{\gamma_{2,3}}$ 15-25\% for all elements}\\
\citet{OSWAL2020108809}& \makecell{Ta, Pt, Th, U}& \makecell{Si of 84-140 MeV} & \makecell{15-20 for Ta and 30-35 \\for other targets, 12-15, 12-15} & \makecell{10-12\% for the production cross sections,\\ uncertainties were evaluated \\ quite satisfactorily} \\
\hline
\end{tabular}
\end{table}
\end{landscape}
\begin{figure*}
\includegraphics[width=140mm,scale=01.0]{Spectrum_L_shell_Ta.jpg
\caption{\label{fig:1} L x-ray spectra of $_{73}{Ta}$ bombarded with the 84 MeV $^{28}{Si}$ ions. Deconvoluted x-ray lines due to different transitions are shown along with the background due to Compton scattering.}
\end{figure*}
\section{Measurement uncertainty evaluation method and results}
\indent A representative spectrum of emitted L x-rays induced by heavy particles is shown in Fig.~\ref{fig:1}. Thus, we determine the experimentally x-ray production cross-section for a particular x-ray peak ($\sigma_x^i$) using a relation
\begin{eqnarray}
\sigma_x^i= \frac{Y_x^i A Sin\theta}{N_A \epsilon n_p t \beta}\label{eq:2}
\end{eqnarray}
\noindent where $Y_x^i$ is the intensity of the $i$th x-ray peak, $A$ is the atomic weight of the target, $\theta$ is the angle between the incident ion beam and the target foil normal, $N_A$ is the Avogadro number, $n_p$ is the number of incident projectiles, $\epsilon$ is the effective efficiency of the x-ray detector, $t$ is the target thickness in $\mu$g/cm$^2$ and $\beta$ is a correction factor for the absorption of the emitted x-rays inside the target. Measurement uncertainty on the $\sigma_x^i$ is introduced due to uncertainties in different parameters used in Eqs.~(\ref{eq:2}). Let us consider an experiment \cite{OSWAL2020108809} to have ideas on the uncertainties involved, which are as follows: the photo peak area, $Y_x^i$, evaluation ($\leq1\% $ for the L$_\alpha$ x-ray peak and $\approx$ 3\% for the other peaks), the ion beam current ($\approx 5\% $), ${Sin\theta}$ ($\approx 1\% $) and the target thickness ($\approx$ 3\%). The uncertainty in the detector efficiency values, $\epsilon$, is $\approx$ 8\% in the energy region of current interest. The relative combined standard uncertainty equation for the x-ray production cross-section as given in Eqs.~(\ref{eq:2}) can be written including the relative combined variance and covariance as follows
\begin{equation}
\begin{aligned}
{(\frac{\delta\sigma^i_{x}}{\sigma^i_{x}})}^2= {(\frac{{\delta Y{_x^i}}}{Y{_x^i}})}^2+{(\frac{{\delta\epsilon}}{\epsilon})}^2+{(\frac{{\delta{n_p}}}{n_p})}^2+ {(\frac{{\delta{t}}}{t})}^2+{(\frac{{\delta{Sin\theta}}}{Sin\theta})}^2+{(\frac{{\delta{\beta}}}{\beta})}^2\\
+ 2~{\delta t}~{\delta \beta}~\frac{\delta \sigma^i_x}{\delta t}~\frac{\delta \sigma^i_x}{\delta\beta} \frac{1}{{\sigma^i_x}^2}
- 2~{\delta \epsilon}~{\delta Y{_x^i}}~\frac{\delta \sigma^i_x}{\delta\epsilon}~\frac{\delta \sigma^i_X}{\delta Y{_x^i}} \frac{1}{{\sigma^i_x}^2}
\label{eq:3}
\end{aligned}
\end{equation}
\begin{table}
\caption{\label{tab:table2}
X-ray production cross-sections with uncertainty for the elements of our interest (uncertainties are obtained using the uncertainties on various parameters mentioned in \cite{OSWAL2020108809} and given in terms of percentage).}
\centering
\begin{tabular}{l l l l}
\hline
\textbf{El.} & \textbf{$\sigma_{L_{\gamma_{2+3}}}$} & \textbf{$\sigma_{L_{\gamma_{1+5}}}$} & \textbf{$\sigma_{L_{\alpha_{1+2}}}$}\\
\hline
$_{73}{Ta}$ & 193 $\pm$ 12 & 290 $\pm$ 12 & 6614 $\pm$ 11\\
$_{78}{Pt}$ & 154 $\pm$ 12& 231 $\pm$ 12& 5542 $\pm$ 11 \\
$_{90}{Th}$ & 42 $\pm$ 12& 63 $\pm$ 12 & 1537 $\pm$ 11\\
$_{92}{U}$ & 28 $\pm$ 12& 42 $\pm$ 12& 1098 $\pm$ 11\\
\hline
\end{tabular}
\end{table}
{\noindent Last two terms in uncertainty arise due to the contributions coming from the covariance relationship of the detector efficiency ($\epsilon$) with the $i^{th}$ intensity peak ($Y_i^x$) and the absorption correction factor of the emitted x-rays inside the target ($\beta$) with the target thickness (t).
Simplifying these two terms by the equation (\ref{eq:2}) we can write the equation (\ref{eq:3}) as under
\begin{equation}
\begin{aligned}
{(\frac{\delta\sigma^i_{x}}{\sigma^i_{x}})}^2= {(\frac{{\delta Y{_x^i}}}{Y{_x^i}})}^2+{(\frac{{\delta\epsilon}}{\epsilon})}^2+{(\frac{{\delta{n_p}}}{n_p})}^2+ {(\frac{{\delta{t}}}{t})}^2+{(\frac{{\delta{Sin\theta}}}{Sin\theta})}^2+{(\frac{{\delta{\beta}}}{\beta})}^2\\+2~\frac{\delta \epsilon}{\epsilon}~\frac{{\delta Y{_x^i}}}{Y{_x^i}}+2~\frac{\delta t}{t}~\frac{\delta \beta}{\beta}
\label{eq:4}
\end{aligned}
\end{equation}
}
The absorption correction factor $\beta$ is written as
\begin{eqnarray}
\beta= \frac{1-\exp^{-\mu t}}{\mu t}
\label{eq:5}
\end{eqnarray}
and
\begin{eqnarray}
\frac{\delta \beta}{\beta}\approx -\mu t~ (\frac{\delta t}{t}+\frac{\delta\mu}{\mu}).
\label{eq:6}
\end{eqnarray}
Where, $\mu$ is for attenuation co-efficient inside the target. The unit of $\mu$ is $cm^2/g$ and that of $t$ is $\mu g/cm^2$; the value $\mu t$ is very small. For instance, $\mu$ of carbon for Fe $K_\alpha$ energy (6.4 keV) is 8.73 $cm^2/g$, $\frac{\delta\mu}{\mu}=0.5\%$ \cite{creagh1990problems} and if the target thickness is taken as 50 $\mu$g/cm$^2$ and $\frac{\delta t}{t}=3\%$, we get $\frac{\delta\beta}{\beta}=0.0013\%$. Thus, $\frac{\delta \beta}{\beta}$ is negligibly small and the terms containing $\delta\beta$ can be equated to zero. With this circumstances, the equation (\ref{eq:4}) can be written as
\begin{equation}
\begin{aligned}
{(\frac{\delta\sigma^i_{x}}{\sigma^i_{x}})}^2= {(\frac{{\delta Y{_x^i}}}{Y{_x^i}})}^2+{(\frac{{\delta\epsilon}}{\epsilon})}^2+{(\frac{{\delta{n_p}}}{n_p})}^2+ {(\frac{{\delta{t}}}{t})}^2+{(\frac{{\delta{Sin\theta}}}{Sin\theta})}^2+2~\frac{\delta \epsilon}{\epsilon}~\frac{{\delta Y{_x^i}}}{Y{_x^i}}
\label{eq:7}
\end{aligned}
\end{equation}
\\
\indent Here we show that the $\epsilon$ does not have any relation with any quantities present in equation (\ref{eq:2}). The detector efficiency ($\epsilon$) is written according to a semi empirical model \cite{mohanty2008comparison} as
\begin{equation}
\begin{aligned}
\epsilon=\frac{\Omega}{4\pi}exp(-\sum_1^n \mu_id_i)f_E(1-exp(-\mu_iD))
\label{eq:8}
\end{aligned}
\end{equation}
where $\Omega$ is the fractional solid angle subtended by the detector crystal at the source and i denotes the medium between the source and the beryllium window, the window itself, a possible ice layer, the gold electrode, the frontal crystal dead layer etc. The $d_i$ are the thickness of these absorbers and $\mu_i$ are their linear absorption coefficients taken from XCOM database \cite{berger2010xcom}. The term $f_E$ represents the escape peak correction factor. The $\mu_C$ is the photoelectric absorption coefficient of the detector crystal of thickness $D$. There is no quantity present in RHS of equation (\ref{eq:2}) is present in the above equation. Hence, $\epsilon$ does not have obvious covariance with another quantity. However, the magnitude of $Y^x_i$ is strongly affected by the value of $\epsilon$ and $\epsilon=f(E)$, $E$=energy of x-ray, hence, the $\epsilon$ is having relation with $Y^x_i$. Whereas $Y^x_i$ is directly proportional to the number of projectiles ($N_p$) as well as the target thickness ($t$). Hence, $\epsilon$ does have covariance with the $Y^x_i$ and no covariance with either $N_p$ and $t$.\\
\indent The measured L x-ray production cross-sections are given in Table~\ref{tab:table2}. To calculate the ionization cross-section from the x-ray production cross-section, we need appropriate relation between these two. It is a simple relation for the K shell case involving only fluorescence yield: $\sigma_K = \sigma_x\omega_K $, where $\sigma_K$ is the K shell ionization cross-section, $\sigma_x$ is the K x-ray production cross-section, $\omega_K$ is the K x-ray fluorescence yield \cite{tiracsouglu2008determination}. However, the ionization cross-sections of different L subshells \cite{singh2000subshell,rahangdale2016spectroscopic} are related to the x-ray production cross-sections through the following relations involving several atomic parameters like the fraction of radiative transition rates, fluorescence yields corresponding to $L_i$ subshells ($\omega_i$) and Coster-Kronig transition probability ($f_{ij}$) between two subshells $i$ and $j$ is written as follows
\begin{eqnarray}
\sigma_{L_1}= \frac{\sigma_{L_{\gamma_{2+3}}}}{{\omega_1}{S_{\gamma_{2+3,1}}}}.\label{eq:9}
\end{eqnarray}
\noindent \textcolor{red}{where} $\sigma_{L_1}$ is the L$_1$ subshell ionization cross-section, ${S_{\gamma_{2+3,1}}}$ is the fraction of radiative transition rates for ${L_{\gamma_{2+3}}}$ and ${\omega_1}$ is fluorescence yield of all the transitions terminating at L$_1$. Similarly, the L$_2$ subshell ionization cross-section $\sigma_{L_2}$ is given as
\begin{eqnarray}
\sigma_{L_2}= \frac{\sigma_{L_{\gamma_{1+5}}}}{{\omega_2}{S_{\gamma_{1+5,2}}}}-\sigma_{L_1}f_{12}.\label{eq:10}
\end{eqnarray}
\noindent Where ${S_{\gamma_{1+5,1}}}$ is the fraction of radiative transition rates for ${L_{\gamma_{1+5}}}$, ${\omega_2}$ is fluorescence yield of all the transitions terminating at L$_2$ and $f_{12}$ is the Coster-Kronig transition because of vacancy transfer from L$_1$ to L$_2$. Lastly, the L$_3$ subshell ionization cross-section $\sigma_{L_3}$ is given as
\begin{eqnarray}
\sigma_{L_3}= \frac{\sigma_{L_{\alpha_{1+2}}}}{{\omega_3}{S_{\alpha_{1+2,3}}}}-\sigma_{L_1}(f_{12}f_{23}+f_{13})-\sigma_{L_2}f_{23}.\label{eq:11}
\end{eqnarray}
\noindent Where ${S_{\alpha_{1+2,1}}}$ is the fraction of radiative transition rates, ${L_{\alpha_{1+2}}}$, ${\omega_3}$ is fluorescence yield of all the transitions terminating at L$_3$, $f_{13}$ is the Coster-Kronig transition because of vacancy transfer from L$_1$ to L$_3$, and $f_{23}$ is the Coster-Kronig transition because of vacancy transfer from L$_2$ to L$_3$.
\indent We can notice that obtaining ${\sigma_{L_1}}$ from Eqs.~(\ref{eq:9}) is simpler than finding ${\sigma_{L_{2,3}}}$ from Eqs.~(\ref{eq:10}) and Eqs.~(\ref{eq:11}), respectively, as Eqs.~(\ref{eq:8}) does not contain any Coster-Kronig transition probability. Whereas Eqs.~(\ref{eq:10}) and Eqs.~(\ref{eq:11}) contain one $f_{ij}$ and three $f_{ij}$s, respectively. The value of ${\omega}$’s and f’s are taken from Campbell et al. \cite{campbell2003fluorescence,campbell2009fluorescence} and the value of emission rates are taken from \citet{campbell1989interpolated} for obtaining the fraction of radiative transition rates. The emission rates of $L_1$, $L_2$ and $L_3$ subshells for the elements of our interest are given below in Table~\ref{tab:table3}, where, $L_1M_3$ etc., denote the origin of x-ray transition. For example, $L_1M_3$ is due to an electron jumping from $M_3$ to vacant $L_1$ subshell. $L_3O_{45}$ denotes a pair of emissions from $O_{45}$ to vacant $L_1$ subshell having line energies very close. For all the emission rate data, the uncertainty reported \cite{campbell1989interpolated} is 0.2\%. Using these emission rates, we can calculate the fraction of radiative transition rates when two or more lines are measured as a single line due to limited resolving power of detector used. Since $L_{\gamma_{2+3}}$ line is due to $L_{\gamma_{2}}$ ($L_1N_2$) and $L_{\gamma_{3}}$ ($L_1N_3$) transitions, so the fraction of radiative transition rates for $L_{\gamma_{2+3}}$ line is written as
\begin{eqnarray}
S_{\gamma_{2+3,1}}=\frac{L_1N_2+L_1N_3}{\sum_{i}{L_1X_i}}=\frac{L_1N_2}{\sum_{i}{L_1X_i}}+\frac{L_1N_3}{\sum_{i}{L_1X_i}}
\label{eq:12}
\end{eqnarray}
\noindent Now, uncertainty equation for the ${S_{\gamma_{2+3,1}}}$ can be written as
\begin{eqnarray}
{(\frac{{\delta{S_{\gamma_{2+3,1}}}}}{S_{\gamma_{2+3,1}}})}^2= {(\frac{{\delta{L_1N_2}}}{L_1N_2})}^2+{(\frac{{\delta{L_1N_3}}}{L_1N_3})}^2+2
{(\frac{{\sum_{i}\delta L_1X_i}}{{\sum_{i}L_{1}{X_i}}})}^2 \label{eq:13}
\end{eqnarray}
\noindent Similarly, we can calculate $S_{\gamma_{1+5,2}}$ and $S_{\alpha_{1+2,3}}$. Since $L_{\gamma{1+5}}$ line is due to $L_{\gamma{1}}$ $(L_2N_4)$ and $L_{\gamma{5}}$ $(L_2N_1)$ transitions so the fraction of radiative transition rates for for $L_{\gamma_{1+5}}$ line is
\begin{eqnarray}
S_{\gamma_{1+5,1}}
= \frac{L_2N_4+L_2N_1}{\sum_{i}{L_2Y_i}}\label{eq:14A}
=\frac{L_2N_4}{\sum_{i}{L_2Y_i}}+\frac{L_2N_1}{\sum_{i}{L_2Y_i}}.\label{eq:14B}
\end{eqnarray}
\noindent $L_{\alpha_{1+2}}$ line is due to $L_{\alpha_{1}}$($L_3M_4$) and $L_{\alpha_{2}}$($L_3M_5$) transitions so the fraction of radiative transition rates for $L_{\alpha_{1+2}}$ line is:
\begin{eqnarray}
S_{\alpha_{1+2,3}} = \frac{L_3M_4+L_3M_5}{\sum_{i}{L_3Z_i}} \label{eq:15A}
=\frac{L_3M_4}{\sum_{i}{L_3Z_i}}+\frac{L_3M_5}{\sum_{i}{L_3Z_i}}.
\label{eq:15B}
\end{eqnarray}
\noindent All calculated fraction of radiative transition rates for elements of our interest are given in Table~\ref{tab:table4}. Uncertainty equation for $S_{\gamma_{1+5,1}}$ and $S_{\alpha_{1+2,3}}$ can also be written in terms of Eqs.~(\ref{eq:11}). Emission rates and corresponding uncertainties are known \cite{campbell1989interpolated}, so we can calculate uncertainty in every fraction of radiative transition rates as shown in Table~\ref{tab:table4} for the elements of our interest.
\begin{table}
\caption{\label{tab:table3}
$L_1$, $L_2$ and $L_3$ emission rates for electric dipole transitions \cite{campbell1989interpolated} are listed with multiplication factor $10^{-2}$. All rates are primarily in $eV/\hbar$ ($=1.519\times$ ${10}^{15}$ $S^{-1}$), hence rates in the table are in unit of $1.519\times$ ${10}^{13}$ $S^{-1}$.}
\begin{tabular}{l l l l l l l l l l l}
\hline
\textbf{El.} & \textbf{$L_1L_3$} & \textbf{$L_1M_2$} & \textbf{$L_1M_3$} & \textbf{$L_1N_2$} & \textbf{$L_1N_3$} & \textbf{$L_1O_{23}$} & \textbf{$L_1P_{23}$}& & \\
\hline
$_{73}{Ta}$ & 1.446 & 24.65 & 31.54 & 6.176 & 8.386 & 2.189\\
$_{78}{Pt}$ & 2.636 & 34.94 & 41.27 & 9.023 & 11.54 & 3.561\\
$_{90}{Th}$ & 10.82 & 75.68 & 69.68 & 20.75 & 21.87 & l0.l0& 2.004 \\
$_{92}{U}$ & 13.62 & 85.59 & 74.82 & 23.63 & 23.93 & ll.56 & 2.146\\
\hline
\textbf{El.} & \textbf{$L_2M_1$} & \textbf{$L_2M_4$} & \textbf{$L_2N_1$} &\textbf{$L_2N_4$}&\textbf{$L_2O_1$} &\textbf{$L_2O_4$} & \textbf{$L_2P_1$}\\
\hline
$_{73}{Ta}$ &2.928 & 107.4& 0.7266& 21.02& 0.1256& 0.4713& 0.01584\\
$_{78}{Pt}$& 4.061& 148.0& l.038& 30.63 & 0.2039& 2.772& 0.01050\\
$_{90}{Th}$ & 8.374& 295.1& 2.283& 68.56 & 0.5646 & 13.35 & 0.1240\\
$_{92}{U}$& 9.375 & 328.0& 2.580 & 77.63 & 0.6538 & 15.59 & 0.1397\\
\hline
\textbf{El.} &\textbf{$L_3M_1$} & \textbf{$L_3M_4$} & \textbf{$L_3M_5$} & \textbf{$L_3N_1$} &\textbf{$L_3N_4$} & \textbf{$L_3N_5$} & \textbf{$L_3O_1$} & \textbf{$L_3O_{45}$} & \textbf{$L_3P_1$}\\
\hline
$_{73}{Ta}$ & 4.378 & 9.655 & 85.06 & 1.054 & 1.772 & 15.88 & 0.l809 & 0.3825 & 0.0227\\
$_{78}{Pt}$ & 6.377 & 13.11& 115.3 & 1.569 & 2.511 & 22.61 & 0.3051 & 2.176 & 0.01513\\
$_{90}{Th}$ & 14.67 & 25.02 & 219.5 & 3.773 & 5.l92 & 47.4O & 0.9264 & 9.983 & 0.2009\\
$_{92}{U}$& 16.72 & 27.58 & 241.8 & 4.314 & 5.796 & 53.13 & 1.076 & 11.50 & 0.2299\\
\hline
\end{tabular}
\end{table}
\begin{table}
\caption{\label{tab:table4} The fraction of radiative transition rates for the elements of our interest (uncertainties are given in percentage).}
\centering
\begin{tabular}{l l l l}
\hline
\textbf{El.}& \textbf{$S_{\gamma_{2+3,1}}$} & \textbf{$S_{\gamma_{1+5,2}}$} &\textbf{$S_{\alpha_{1+2,3}}$} \\
\hline
$_{73}{Ta}$& 0.1957 $\pm$ 1& 0.1639 $\pm$ 1& 0.8000 $\pm$ 1\\
$_{78}{Pt}$& 0.1997 $\pm$ 1& 0.1696 $\pm$ 1& 0.7831 $\pm$ 1\\
$_{90}{Th}$& 0.2020 $\pm$ 1 & 0.1824 $\pm$ 1& 0.7485 $\pm$ 1\\
$_{92}{U}$& 0.2021 $\pm$ 1 & 0.1848 $\pm$ 1& 0.7437 $\pm$ 1\\
\hline
\end{tabular}
\end{table}
Now, from Eqs.~(\ref{eq:8}), the uncertainty equation for $\sigma_{L_1}$ can be written as:
\begin{eqnarray}
{(\frac{{\delta\sigma_{L_1}}}{\sigma_{L_1}})}^2= {(\frac{{\delta\sigma_{L_{\gamma_{2+3}}}}}{\sigma_{L_{\gamma_{2+3}}}})}^2+{(\frac{{\delta\omega_{1}}}{\omega_1})}^2+{(\frac{{S_{\gamma_{2+3,1}}}}{S_{\gamma_{2+3,1}}})}^2 \label{eq:16}
\end{eqnarray}
\noindent Value of $\sigma_{L_{\gamma_{2+3}}}$ (barns/atom), $\omega_1$ and $S_{\gamma_{2+3,1}}$ with uncertainties, required for calculation of uncertainty of $\sigma_{L_1}$, are given in Table~\ref{tab:table5}. Uncertainty of {$\sigma_{L_{\gamma_{2+3,1}}}$} is taken from Table~\ref{tab:table2}, that of $\omega_1$ is taken from \cite{campbell2003fluorescence,campbell2009fluorescence} and that of ${S_{\gamma_{2+3,1}}}$ is taken from Table~\ref{tab:table4}. Values of the $\sigma_{L_1}$ along with the uncertainties are given in Table~\ref{tab:table5}.
\begin{table}
\caption{\label{tab:table5} $\sigma_{L_1}$ (barns/atom) along with uncertainties for various target atoms by Si beams at 84 MeV. Corresponding values of $\sigma_{L_{\gamma_{2+3}}}$ (barns/atom) \cite{OSWAL2020108809}, conversion parameters ${\omega_1}$ \cite{campbell2003fluorescence, campbell2009fluorescence} and {$S_{\gamma_{2+3,1}}$} are also given (uncertainties given in percentage).}
\centering
\begin{tabular}{l l l l l}
\hline
\textbf{El.} & \textbf{$\sigma_{L_{\gamma_{2+3}}}$}&\textbf{${\omega_1}$}& \textbf{$S_{\gamma_{2+3,1}}$} & \textbf{$\sigma_{L_1}$}\\ \hline
$_{73}{Ta}$ & 193 $\pm$ 12 & 0.144 $\pm$ 15 & 0.1957 $\pm$ 1 & 6858 $\pm$ 19\\
$_{78}{Pt}$ & 154 $\pm$ 12 & 0.114 $\pm$ 30 &0.1997 $\pm$ 1 & 6756 $\pm$ 32\\
$_{90}{Th}$ & 42 $\pm$ 12 & 0.159 $\pm$ 35 & 0.2020 $\pm$ 1 & 1308 $\pm$ 37\\
$_{92}{U}$ & 28 $\pm$ 12 & 0.168 $\pm$ 35 & 0.2021 $\pm$ 1 & 819 $\pm$ 37 \\
\hline
\end{tabular}
\end{table}
Now, Eqs.~(\ref{eq:10}) can be written as
\begin{eqnarray*}
\sigma_{L_2}= \sigma^A_{L_2}+\sigma^B_{L_2}
\end{eqnarray*}
\noindent where,
\begin{eqnarray*}
\sigma^A_{L_2}= \frac{\sigma_{L_{\gamma_{1+5}}}}{{\omega_2}{S_{\gamma_{1+5,2}}}}, \sigma^B_{L_2}=-\sigma_{L_1}f_{12}
\end{eqnarray*}
\noindent So, the uncertainty equation for ${\sigma_{L_2}}$ can be written as
\begin{eqnarray}
{(\frac{\delta\sigma_{L_2}}{\sigma_{L_2}})}^2 = {{(\frac{\sigma^A_{L_2}}{\sigma_{L_2}})}^2} {(\frac{\delta\sigma^A_{L_2}}{\sigma^A_{L_2}})}^2+{(\frac{\sigma^B_{L_2}}{\sigma_{L_2}}^2 )}{(\frac{\delta\sigma^B_{L_2}}{\sigma^B_{L_2}})}^2 \label{eq:17}
\end{eqnarray}
\noindent where,
\begin{gather}
{(\frac{\delta\sigma^A_{L_2}}{\sigma^A_{L_2}})}^2 = {(\frac{\delta\sigma_{L_{\gamma_{1+5}}}}{\sigma_{L_{\gamma_{1+5}}}})}^2+{(\frac{\delta\omega_{2}}{\omega_{2}})}^2+{(\frac{\delta{S_{\gamma_{1+5,2}}}}{S_{\gamma_{1+5,2}}})}^2 \tag{17.1}\label{eq:17.1}
\end{gather}
\noindent and
\begin{gather}
{(\frac{\delta\sigma^B_{L_2}}{\sigma^B_{L_2}})}^2 = {(\frac{\delta\sigma_{L_1}}{\sigma_{L_1}})}^2 + {(\frac{\delta{f_{12}}}{{f_{12}}})}^2 \tag{17.2}\label{eq:17.2}
\end{gather}
\noindent Values of $\sigma_{L_{\gamma_{1+5}}}$ (barns/atom), $\omega_2$, $S_{\gamma_{1+5,2}}$ and $f_{12}$ with uncertainties, required for calculation of uncertainty for $\sigma_{L_2}$ are given in Table~\ref{tab:table6}. $\sigma_{L_1}$ along with its uncertainty is taken from Table~\ref{tab:table5}. 10\% uncertainty in production cross-section ({$\sigma_{L_{\gamma_{1+5,2}}}$}) is taken, which is due to the uncertainty in the photo-peak area evaluation ($\approx4\%$), the ion beam current ($\approx5\%$) and the target thickness ($\approx3\%$) \cite{OSWAL2020108809}. The uncertainty in the effective efficiency values is 5-8\% in the energy region of current interest \cite{OSWAL2020108809}, highest value 8\% is taken. Uncertainty for $\omega_2$ is 5\% for all element of our interest \cite{campbell2003fluorescence,campbell2009fluorescence}. For $f_{12}$, uncertainty for $_{73}{Ta}$ is 20\%, for $_{78}{Pt}$ is 30-40\%, for $_{90}{Th}$ is 50-100\% and for $_{92}{U}$ is 50-100\% \cite{campbell2003fluorescence,campbell2009fluorescence}. $\delta S_{{\gamma_{1+5,2}}}$ is due to uncertainty in the emission rate, which is only 0.2\% \cite{campbell1989interpolated}. Values of $\sigma_{L_2}$ with uncertainties is given in Table~\ref{tab:table6}.
\begin{table}
\caption{\label{tab:table6} $\sigma_{L_2}$ (barns/atom) along with uncertainties for various target atoms by Si beams of 84 MeV. Corresponding values of $\sigma_{L_{\gamma_{1+5}}}$ (barns/atom) \cite{OSWAL2020108809}, conversion parameters ${\omega_2}$, $f_{12}$ \cite{campbell2003fluorescence, campbell2009fluorescence}) and {$S_{\gamma_{1+5, 2}}$} \cite{campbell1989interpolated} are also given (uncertainties given in percentage).}
\begin{tabular}{l l l l l l l l}
\hline
\textbf{Elements} & \textbf{$\sigma_{L_{\gamma_{1+5}}}$}& \textbf {${\omega_2}$} &\textbf{$S_{\gamma_{1+5,2}}$}& \textbf{${f_{12}}$} & \textbf{$\sigma_{L_2}$}\\
\hline
$_{73}{Ta}$& 290 $\pm$ 12& 0.28 $\pm$ 5 & 0.1639 $\pm$ 1&0.118 $\pm$ 20 & 5510 $\pm$ 16\\
$_{78}{Pt}$& 231 $\pm$ 12& 0.344 $\pm$ 5& 0.1696 $\pm$ 1&0.075 $\pm$ 30-40 & 3448 $\pm$ 17\\
$_{90}{Th}$& 63 $\pm$ 12& 0.503 $\pm$ 5& 0.1824 $\pm$ 1&0.040$\pm$ 50-100 & 635 $\pm$ 15\\
$_{92}{U}$& 42 $\pm$ 12& 0.506 $\pm$ 5& 0.1848 $\pm$ 1&0.035 $\pm$ 50-100 & 417 $\pm$ 15\\
\hline
\end{tabular}
\end{table}
Similarly, the Eqs.~(\ref{eq:11}) can be written as
\begin{eqnarray*}
\sigma_{L_3}= \sigma^C_{L_3}+ \sigma^D_{L_3}+ \sigma^E_{L_3}+ \sigma^F_{L_3}
\end{eqnarray*}
\noindent So the uncertainty equation can be written as
\begin{equation}
\begin{aligned}
{(\frac{\delta\sigma_{L_3}}{\sigma_{L_3}})}^2 = {(\frac{\sigma^C_{L_3}}{\sigma_{L_3}})}^2 {(\frac{\delta\sigma^C_{L_3}}{\sigma^C_{L_3}})}^2+{(\frac{\sigma^D_{L_3}}{\sigma_{L_3}} )}^2{(\frac{\delta\sigma^D_{L_3}}{\sigma^D_{L_3}})}^2+{(\frac{\sigma^E_{L_3}}{\sigma_{L_3}})}^2 {(\frac{\delta\sigma^E_{L_3}}{\sigma^E_{L_3}})}^2+\\{(\frac{\sigma^F_{L_3}}{\sigma_{L_3}} )}^2{(\frac{\delta\sigma^F_{L_3}}{\sigma^F_{L_3}})}^2\label{eq:18}
\end{aligned}
\end{equation}
\noindent where,
\begin{eqnarray*}
{\sigma^C_{L_3} = \frac{\sigma_{L_{\alpha_{1+2}}}}{{\omega_3}{S_{\alpha_{1+2,3}}}},
\sigma^D_{L_3} = -\sigma{L_1}(f_{12}f_{23}),
\sigma^E_{L_3}=-\sigma{L_1}f_{13},
\sigma^F_{L_3}=-\sigma_{L_2}f_{23}}
\end{eqnarray*}
\noindent and the corresponding uncertainty equations as follows
\begin{gather}
{(\frac{\delta\sigma^C_{L_3}}{\sigma^C_{L_3}})}^2={(\frac{\delta\sigma_{L_{\alpha_{1+2}}}}{\sigma_{L_{{\alpha_{1+2}}}}})}^2+{(\frac{\delta\omega_{3}}{\omega_{3}})}^2+{(\frac{\delta{S_{\alpha_{1+2,3}}}}{S_{\alpha_{1+2,3}}})}^2 \tag{18.1}\label{eq:18.1}
\end{gather}
\begin{gather}
{(\frac{\delta\sigma^D_{L_3}}{\sigma^D_{L_3}})}^2={(\frac{\delta\sigma_{L_1}}{\sigma_{L_1}})}^2 + {(\frac{\delta{f_{12}}}{{f_{12}}})}^2+{(\frac{\delta{f_{23}}}{{f_{23}}})}^2 \tag{18.2}\label{eq:18.2}
\end{gather}
\begin{gather}
{(\frac{\delta\sigma^E_{L_3}}{\sigma^E_{L_3}})}^2={(\frac{\delta\sigma_{L_1}}{\sigma_{L_1}})}^2 + {(\frac{\delta{f_{13}}}{{f_{13}}})}^2 \tag{18.3}\label{eq:18.3}
\end{gather}
\noindent and
\begin{gather}
{(\frac{\delta\sigma^F_{L_3}}{\sigma^F_{L_3}})}^2={(\frac{\delta\sigma_{L_2}}{\sigma_{L_2}})}^2 + {(\frac{\delta{f_{23}}}{{f_{23}}})}^2 \tag{18.4}\label{eq:18.4}
\end{gather}
\noindent Values of $\sigma_{L_{\alpha_{1+2}}}$, $\omega_3$, {$S_{\alpha_{1+2,3}}$}, $f_{23}$ and $f_{13}$, with uncertainties required to get the uncertainty in $\sigma_{L_3}$ are given in Table~\ref{tab:table7}. Values of $f_{12}$, $\sigma_{L_1}$ and $\sigma_{L_2}$ along with uncertainties are taken from Table~\ref{tab:table5} and Table~\ref{tab:table6}, respectively. The values of $\sigma_{L_3}$ along with the uncertainties are given in Table~\ref{tab:table7}.
\begin{table}
\caption{\label{tab:table7} $\sigma_{L_3}$ (barns/atom) along with uncertainties for various target atoms by Si beams of 84 MeV. Corresponding values of $\sigma_{L_{\alpha_{1+2}}}$ (barns/atom) \cite{OSWAL2020108809}, conversion parameters (${\omega_3}$, $f_{23}$, $f_{13}$ \cite{campbell2003fluorescence, campbell2009fluorescence}) and $S_{\gamma_{1+5, 2}}$ \cite{campbell1989interpolated} are also given (uncertainties are given in percentage).}
\begin{tabular}{l l l l l l l l}
\hline
\textbf{El.} & \textbf{$\sigma_{L_{\alpha_{1+2}}}$} &\textbf{${\omega_3}$} & $S_{\alpha_{1+2, 3}}$& \textbf{${f_{23}}$}& \textbf{${f_{13}}$} & \textbf{$\sigma_{L_3}$}\\
\hline
$_{73}{Ta}$& 6614 $\pm$ 11& 0.251 $\pm$ 5 & 0.8000 $\pm$ 1& 0.134 $\pm$ 10& 0.328 $\pm$ 15 & 29844 $\pm$ 13\\
$_{78}{Pt}$& 5542 $\pm$ 11& 0.303 $\pm$ 5 &0.7831 $\pm$ 1& 0.126 $\pm$ 10& 0.545 $\pm$ 20 & 19175 $\pm$ 16\\
$_{90}{Th}$ & 1537 $\pm$ 11 & 0.424 $\pm$ 5 & 0.7485 $\pm$ 1& 0.103 $\pm$ 10& 0.62 $\pm$ 15& 3962 $\pm$ 17\\
$_{92}{U}$ & 1099 $\pm$ 11& 0.444 $\pm$ 5&0.7437 $\pm$ 1& 0.14 $\pm$ 5& 0.62 $\pm$ 15& 2756 $\pm$ 16\\
\hline
\end{tabular}
\end{table}
\indent Lastly, total L-shell ionization ($\sigma_{L}$) and its uncertainty are given as under:
\begin{eqnarray}
\sigma_{L} = \sum_{i}\sigma_{L_i} = \sigma_{L_1} + \sigma_{L_2} + \sigma_{L_3}\label{eq:19}
\end{eqnarray}
\begin{eqnarray}
{(\frac{\delta\sigma_{L}}{\sigma_{L}})}^2={(\frac{\sigma_{L_1}}{\sigma_{L}})}^2{(\frac{\delta\sigma_{L_1}}{\sigma_{L_1}})}^2 + {(\frac{\sigma_{L_2}}{\sigma_{L}})}^2{(\frac{\delta\sigma_{L_2}}{\sigma_{L_2}})}^2 + {(\frac{\sigma_{L_3}}{\sigma_{L}})}^2{(\frac{\delta\sigma_{L_3}}{\sigma_{L_3}})}^2 \label{eq:20}
\end{eqnarray}
We have given $\sigma_{L}$ as a function of energies in Table~\ref{tab:table8}. Now, we know the measured $\sigma_{L_{1,2,3,L}}$ along with their correct uncertainties involved so that we can make use of these with right confidence.
\begin{landscape}
\begin{table}
\caption{\label{tab:table8} Values of total L shell ionization cross-section $\sigma_L$ (barns/atom) along with uncertainties of various target atoms by Si beams of different energies (uncertainties are given in percentage).}
\begin{tabular}{l l l l l l l l}
\hline
\textbf{El.} & \textbf{84 MeV}& \textbf{90 MeV}& \textbf{98 MeV} &\textbf{107 MeV} &\textbf{118 MeV}& \textbf{128 MeV}&\textbf{140 MeV} \\
\hline
$_{73}{Ta}$& 42213 $\pm$ 10& 53086 $\pm$10 &64543 $\pm$ 10 &76243 $\pm$ 10 &167575 $\pm$ 10 & 191674 $\pm$ 10 &198142 $\pm$ 10\\
$_{78}{Pt}$& 29378 $\pm$ 13 &30303 $\pm$ 13 &38547 $\pm$ 13 &49786 $\pm$ 13 &95815 $\pm$ 13 &99213 $\pm$ 14& 107667 $\pm$15\\
$_{90}{Th}$ & 5905 $\pm$ 14 & 7186 $\pm$ 14 & 9024 $\pm$ 14 & 13166 $\pm$ 14 & 21059 $\pm$ 14 &25499 $\pm$ 14 & 30782 $\pm$ 14\\
$_{92}{U}$ & 3992 $\pm$ 14 &4394 $\pm$ 14 &6173 $\pm$ 14 &7956 $\pm$ 14 &13377 $\pm$ 14 &15819 $\pm$ 14 &20089 $\pm$ 14\\
\hline
\end{tabular}
\end{table}
\end{landscape}
\begin{figure}
\includegraphics[width=140mm,scale=01.0]{Sig_L_2020_New.jpg
\caption{\label{fig:2} Total ionization cross-section vs bombarding energy of Si ion on various target Ta, Pt, Th and U. Measured data have been compared with ECUSAR \cite{lapicki2002status} and SLPA theory \cite{montanari2011collective}.}
\end{figure}
\section{Discussions on the salient results}
\indent The uncertainty in ${\sigma_{L_1}}$ as given in Table~\ref{tab:table5} for $_{73}{Ta}$ is about 19\%, where that of $_{78}{Pt}$ it is 32\%, and those of $_{90}{Th}$ and $_{92}{U}$ it is about 37\%; even though the uncertainty in ${\sigma_{L_{\gamma_{2+3}}}}$ is about 12\%. \textcolor{red}{It is thus clear from the results that the uncertainty in ${\sigma_{L_1}}$ is primarily dictated by the large uncertainty in ${\omega_1}$ as given in Table~\ref{tab:table5}.} In the case of ${\sigma_{L_2}}$ the uncertainties are between $15-17\%$ for elements $_{73}{Ta}$, $_{78}{Pt}$, $_{90}{Th}$ and $_{92}{U}$ (Table~\ref{tab:table6}). It means the uncertainties are quite close to the uncertainty found in ${\sigma_{L_{\gamma_{1+5}}}}$ 12\%. It happens because of small uncertainty (5\%) in ${\omega_2}$. One important point is to note here that large uncertainty in ${f_{12}}$ does not affect the overall uncertainty in ${\sigma_{L_2}}$ because of weighted uncertainty propagation rule. We can notice that the value of ${f_{12}}$ is rather more important than the associated uncertainty because of the weighted factor changes with its contribution.\\
\indent The above trend seen with ${\sigma_{L_2}}$ plays a similar role in propagating the uncertainty in ${\sigma_{L_3}}$. Here also, the uncertainties are between $13-17\%$ for all four elements $_{73}{Ta}$, $_{78}{Pt}$, $_{90}{Th}$ and $_{92}{U}$. Since the first term containing fluorescence yield $\omega_3$ and the fraction of radiative transition rates $S_{\alpha_{1+2,3}}$ along with production cross-section ${\sigma_{L_{\alpha_{1+2}}}}$ in Eqs.~(\ref{eq:10}) gives major contribution, the uncertainty associated with this part is the main source of uncertainty. Further, uncertainty in $\omega_3$ is not much (5\%) and that in $S_{\alpha_{1+2,3}}$ is only 1\%, hence, the uncertainty in ${\sigma_{L_3}}$ is closed to that in ${\sigma_{L_{\alpha_{1+2}}}}$. Though the uncertainty in ${f_{23}}$ and ${f_{13}}$ are about 10 and 15\%, respectively, still their share in uncertainty propagation is small because their contributions in Eqs.~(\ref{eq:10}) (part 2 and 3), are small. Let us now list the uncertainty message on the L subshell ionization cross-sections reported earlier in Table \ref{tab:table9}.\\
\textcolor{red}{\indent Uncertainty in all L subshell ionization cross-sections cannot be equal, because the uncertainty in ${\sigma_{L_1}}$ is always larger due to larger uncertainty in ${\omega_1}$ than that in $\omega_{2}$ and $\omega_{3}$. The uncertainty for $\sigma_{L_2}$ and $\sigma_{L_3}$ are nearly equal and highest uncertainty is with $\sigma_{L_1}$. These facts are well represented by the present work shown in Table \ref{tab:table9}. However, many cases shown in Table \ref{tab:table9} do not support these facts. Hence, only some measurements treat the uncertainty associated with the L subshell ionization quite reasonably. This scenario is a deterrent for improving the theoretical understanding of the concerned complex mechanism. To give a correct treatment one has to follow elaborate error analysis as demonstrated in the present work. Having done this, one can confidently test the theories and look for any new physical processes if occurred in the experiment.}
\begin{landscape}
\begin{table}
\tiny
\caption{\label{tab:table9} Comparison of uncertainties (\%) quoted for L shell ionization cross sections in the previous experiments and present study.}
\centering
\begin{tabular}{l l l l l l l}
\hline
\textbf{Ref.} & \makecell{Targets} & \makecell{\textbf{Uncertainty for}}& \makecell{\textbf{Projectile, its energy}} & \makecell{\textbf{Uncertainties (\%)}} & \makecell{\textbf{Remarks}}\\
\hline
\citet{palinkas1980subshell}& \makecell{Au, Pb, Bi} & \makecell{electron, 60-600 keV} &\makecell{$\sigma_{L_1}$, {$\sigma_{L_2}$}, {$\sigma_{L_3}$}} & \makecell{22, 15, 11} & \makecell{Fair treatment made}\\
\citet{PhysRevA.37.106} & \makecell{W} &\makecell{electron, 12-40 keV}&\makecell{$\sigma_3/\sigma_2$, $\sigma_2/\sigma_1$} &\makecell{10}& \makecell{{$\sigma_3/\sigma_2$} must have lower \\error than that in {$\sigma_2/\sigma_1$}}\\
\citet{reusch1986method}& \makecell{$29\leq Z \leq79$} & \makecell{electron, 50-200 keV} & \makecell{$\sigma_{L_1}$, {$\sigma_{L_2}$}, {$\sigma_{L_3}$}, $\sigma_L$} & \makecell{7-15, 6-15, 6-15, 15} & \makecell{Error in $\sigma_{L_1}$ is underestimated}\\
\citet{bernstein1954shell}& \makecell{Ta, Au, Pb, U} & \makecell{H, 1.5-4.25 MeV} &\makecell{$\sigma_{L_x}$}& \makecell{20} & \makecell{No uncertainty message on L \\subshell ionization cross sections}\\
\citet{PhysRevA.11.607}& \makecell{Ta, Au, Bi} &\makecell{H, 1 to 5.5 and \\ He, 1 to 11 MeV} &\makecell{$\sigma_{L_1}$, {$\sigma_{L_2}$}, {$\sigma_{L_3}$} }& \makecell{22, 28, 21}& \makecell{Uncertainty for $L_2$ cannot \\ be higher than $L_1$ }\\
\citet{semaniak1994a}& \makecell{Between La to Au}& \makecell{ $^{14}N$, 1.75-22.4 MeV}&\makecell{$\sigma_{L_1}$, {$\sigma_{L_2}$}, {$\sigma_{L_3}$} }& \makecell{10-25\% for low energies \\and 10\% for higher energies}& \makecell{All subshells cannot \\have same uncertainty}\\
\citet{Semaniak1995b}& \makecell{$72\leq Z \leq 90$} &\makecell{C and N, 0.4 to 1.8 MeV/amu}&\makecell{$\sigma_{L_1}$, {$\sigma_{L_2}$}, {$\sigma_{L_3}$} }& \makecell{6-20, 5-12, 4-10} & \makecell{Achieving uncertainty as\\ low as 4\% for $\sigma_{L_3}$ is doubtful}\\
\citet{dhal1995subshell}& \makecell{Au, Bi}& \makecell{B, 4.8-8.8 MeV}& \makecell{$\sigma_{L_1}$, {$\sigma_{L_2}$}, {$\sigma_{L_3}$}} & \makecell{10\% for both elements.}& \makecell{All subshells cannot \\have same uncertainty}\\
\citet{banas2002role}& \makecell{Au} & \makecell{O of 0.4-2.2 MeV/amu}& \makecell{$\sigma_{L_1}$, {$\sigma_{L_2}$}, {$\sigma_{L_3}$}}&\makecell{15-30\%
for all subshell} & \makecell{Large error for $\sigma_{L_1}$ and small \\for $\sigma_{L_2}$ and $\sigma_{L_3}$is correct}\\
\citet{fijal2008subshell}& \makecell{Au and Bi}& \makecell{S, 12.8-120 MeV} & \makecell{$\sigma_{L_1}$, {$\sigma_{L_2}$}, {$\sigma_{L_3}$}} &\makecell{10-30\% \\for all subshell}& \makecell{Large error for $\sigma_{L_1}$ and small \\for $\sigma_{L_2}$ and $\sigma_{L_3}$is correct}\\
\citet{OSWAL2020108809}&\makecell{Ta, Pt, Th, U}& \makecell{Si, 84-140 MeV}&\makecell{$\sigma_{L_1}$, {$\sigma_{L_2}$}, {$\sigma_{L_3}$}}& \makecell{15-20\% for Ta and\\ 30-35\% for others,\\ 12-15, 12-15}& \makecell{Good estimation}\\
Present work &\makecell{Ta, Pt, Th, U}& \makecell{Si, 84-140 MeV} & \makecell{$\sigma_{L}$} & \makecell{10\% for Ta, 13-15\% for \\Pt and 14\% for Th and U}& \makecell{correct treatment made}\\
Present work &\makecell{Ta}& \makecell{Si, 84 MeV} & \makecell{{$\sigma_{L1}$},{$\sigma_{L_2}$}, {$\sigma_{L_3}$}} & \makecell{19\%, 16\% and 13\% }& \makecell{Shown for a specific energy}\\
Present work &\makecell{Pt}& \makecell{Si, 84 MeV} & \makecell{{$\sigma_{L1}$},{$\sigma_{L_2}$}, {$\sigma_{L_3}$}} & \makecell{32\%, 17\% and 16\% }& \makecell{Shown for a specific energy}\\
Present work &\makecell{Th}& \makecell{Si, 84 MeV} & \makecell{{$\sigma_{L1}$},{$\sigma_{L_2}$}, {$\sigma_{L_3}$}} & \makecell{37\%, 15\% and 17\% }& \makecell{Shown for a specific energy}\\
Present work &\makecell{U}& \makecell{Si, 84 MeV} & \makecell{{$\sigma_{L1}$},{$\sigma_{L_2}$}, {$\sigma_{L_3}$}} & \makecell{37\%, 15\% and 16\% }& \makecell{Shown for a specific energy}\\
\hline
\end{tabular}
\end{table}
\end{landscape}
\indent The above methodology of error propagation can in principle be followed for future experiments. It can also be applied to earlier experiments if either data are still available for reanalysis or the production cross-section and its uncertainty are given in the concerned publications. However, the best available set of atomic parameters will be used to obtain L subshell ionization cross-sections from the L shell x-ray production cross-sections. \citet{barros2015ionization} have discussed this issue recently in great detail. The atomic parameters are the main source of uncertainty, especially the fluorescence yields. Till date, $\omega_2$ and $\omega_3$ are mostly known within 5\%, but $\omega_1$ is having large uncertainty in particular towards the heavier target side. This is the reason for the larger uncertainty in $\sigma_{L_1}$ in heavy atoms Th, U, etc. Though Coster-Kronig rates contain large uncertainty, but they do not affect much the L subshell cross-section because of their minor contribution (Eqs.~(\ref{eq:10})). Thanks to the theoretical advancements that the emission rates for the electric dipole transitions are known with almost no uncertainty (0.2\%). \\
\indent Plots of the total L shell ionization cross-section $\sigma_{L}$ along with uncertainty versus the projectile energy (Fig.~\ref{fig:2}) as given in Table~\ref{tab:table8} exhibit an unusual feature. The $\sigma_{L}$ behaves similarly, as the theoretical trend predicted by ECUSAR \cite{lapicki2002status} and SLPA \cite{montanari2011collective} up to a certain energy $\approx$ 100 MeV and takes a rise beyond this till it reaches a saturated value. Since we have dealt with the associated uncertainty in the measurement correctly, we can rely on the trend and infer on the concerned mechanism confidently. However, it is a nontrivial and requires a detail study that we keep for an interesting future investigation.\\
\section{Suggestion on making accurate L subshell cross section measurements}
\indent Of course on the experimental side, statistical uncertainty is an important issue for any measurements, in our case, it is L shell x-ray production cross-sections from which L subshell ionization cross-sections are derived. In this case, some other issues are more vital than the statistical uncertainty. For example, detector efficiency is the major source of uncertainty. Hence, final uncertainty on the L shell x-ray production cross-sections depend on the individual uncertainty on various measurements such as detector efficiency, counting statistics, number of incident ions and target thickness and target contamination. Let us discuss how the uncertainty of these experimental parameters can be kept minimum. We will make an attempt to find the best possible way of making the most accurate L shell x-ray production cross-section measurements below.\\
\indent First, we consider measurement of the detector efficiency of x-ray semiconductor detector used for x-ray production cross-sections. There are several efficiency measurements, for example, \citet{barfoot1984si,tribedi1992efficiency,mohanty2008comparison,kumar2017shell}. It is shown there that efficiency for a high purity germanium (HPGe) detector can be known within 5\% \cite{salgado2006determination}, for a Si(Li) detector within 6\% \cite{mohanty2008comparison}, and expected to get similar accuracy for a silicon drift detector (SDD) \cite{zhou2019multiple}. Hence, at this point, one can measure the detector efficiency within 5-6\%. Absolute efficiency of the detector depends not only on detector properties discussed but also on the correct position and solid angle called geometrical efficiency. To define it well, a collimator with a dimension smaller than the detector crystal size in front of the detector is normally used \cite{kumar2017shell}.\\
\indent Next, we discuss the measurement of the number of projectile-ions bombarded on the target. Since charge state of the heavy projectile-ions is changed to a good extent during the ion-solid collisions, charge collected in Faraday cage cannot give correct counting of incident particles. One must follow any one of the following techniques to have less than 1\% uncertainty: (i) elastically scattered projectiles from a heavy target like gold \cite{nandi2002lifetime}, keeping the whole scattering chamber electrically isolated for using it as a Faraday cup \cite{narvekar1992simple}, using two Faraday cups one in front of the target and another behind the target \cite{zhou2013k}. Target thickness, as well as target contamination, can be correctly measured with Rutherford back-scattering technique \cite{feldman1986fundamentals}.\\
\section{Crystal spectrometer vs semiconductor x-ray detector in measuring L subshell cross section}
\indent High-resolution studies using crystal spectrometers help us know preciously the characteristic line energies. However, such efforts do not improve the accuracy of the L subshell ionization because in high-resolution cases, we need only the emission rates, the fraction of radiative transition rates has no role over here. Since emission rates are very accurate (within 0.2 \%) and thus the fraction of radiative transition rates, as discussed above, does not introduce considerable uncertainty at all. Hence, experiments performed with crystal spectrometers, for example \cite{reusch1986method,PhysRevA.37.106,llovet2014cross}, reported exactly similar uncertainty message as the studies using semiconductor x-ray detectors do. \\
\section{Conclusion}
\indent We always gain a good understanding of any physical process by comparing the experimental data with the theoretical values. The comparison is meaningful only if the uncertainty in experiment and theory is accounted to the best possible level. However, it is seldom seen that the theoretical works talk any thing about the uncertainty in the calculation. Often, we can find experimental studies quote these uncertainties, however, most of the time the uncertainties are not estimated correctly, especially for a complex phenomenon. We have taken a test case through L subshell ionization of atoms by particle or photon impact. Experimentally, measured x-ray production cross-sections are used to obtain the ionization cross-sections, which are theoretically calculated. Furthermore the uncertainty of the x-ray production cross-section is mainly governed by statistics (Type A uncertainty) and detector efficiency (Type B uncertainty) but the ionization cross-section involves many atomic parameters (systematic uncertainty) having a wide uncertainty spectrum. Consequently, determining the measurement uncertainty in the L subshell ionization cross-section is always quite complex. We have thoroughly studied this issue in a systematic manner, where the rule of the weighted propagation of the uncertainty is introduced so that on every occasion, percentage uncertainty can be used. We notice that $\sigma_{L_1}$ suffers from the maximum uncertainty because of a large uncertainty in the best fluorescence parameters (systematic uncertainty) available till date relevant to L$_1$ (2s$_{1/2}$) subshell compared to those associated with the other two subshells L$_2$ (2p$_{1/2}$) and L$_3$ (2p$_{3/2}$). \textcolor{red}{Hence, a comparison between the theory and experiment would give higher importance to $L_2$ and $L_3$ subshell ionisation cross-sections because of smaller measurement uncertainty than that in $L_1$.}\\
\indent Sometimes the measured data show a certain departure from the usual trend indicating certain unexplored mechanism. Nevertheless, a lack of confidence in associated uncertainties forbids one to unfold the mystery. At such juncture, log scale is used so that the data may look smoother than when it is on a linear scale, and thus any excessive scatter or unusual trend can be less noticeable. In contrast, accurate uncertainty analysis gives us full confidence about the data, and thus one will not hesitate to draw any meaningful inference if any departure is seen with the measured data. Furthermore, proper uncertainty analysis teaches us the best possible care that can be taken for the improved measurements. Thus, we have suggested a novel way by which the reliable and accurate L subshell ionization cross-section data can be obtained. Finally, we believe this work provides a general overview to every researcher that the uncertainty aspect must be taken into utmost care for right advancement of science.\\
\section{Credit authorship contribution statement}
Shashank Singh: Methodology, calculation, preparation of Tables and figures, Soumya Chatterjee: Methodology, calculation, preparation of Tables and figures, D. Mitra: Formal analysis, methodology, validation, and T. Nandi: Conceptualization, methodology, supervision, validation, visualization, writing - original draft, writing - review and editing.
\section{Acknowledgments}
{One of the authors (S.S.) gratefully acknowledges his supervisors K.P. Singh, Mumtaz Oswal and B. R. Behera for their support to work with any other experts freely for widening his exposures. We gratefully acknowledge Adedamola David Aladese for checking the manuscript thoroughly and giving invaluable comments.}
\bibliographystyle{apsrev4-1}
|
\section*{Abstract} The flow past a fixed finite-length circular cylinder, the axis of which makes a nonzero angle with the incoming stream, is studied through fully-resolved simulations, from creeping-flow conditions to strongly inertial \color{black} regimes. The investigation focuses on the way the body aspect ratio $\cchi$ (defined as as the length-to-diameter ratio), the inclination angle $\theta$ with respect to the incoming flow and the Reynolds number $\text{Re}$ (based on the cylinder diameter) affect the flow structure past the body and therefore the hydrodynamic loads acting on it. \color{black} The configuration $\theta=0^\circ$ (where the cylinder is aligned with the flow) is first considered from creeping-flow conditions up to $\text{Re}=400$, with aspect ratios up to $20$ ($10$) for $\text{Re}\leq10$ ($\text{Re}\geq10$). In the low-to-moderate Reynolds number regime ($\text{Re}\lesssim5$),
influence or the aspect ratio, inclination (from $0^{\circ}$ to $30^{\circ}$), and inertial effects is examined by comparing numerical results for the axial and transverse force components and the spanwise torque with theoretical predictions based on the slender-body approximation, possibly incorporating finite-Reynolds-number corrections. Semiempirical models based on these predictions and incorporating finite-length and inertial corrections extracted from the numerical data are derived. \color{black}
For large enough Reynolds numbers ($\text{Re}\gtrsim10^2$), separation takes place along the upstream part of the lateral surface of the cylinder, deeply influencing the surface stress distribution.
Numerical results are used to build empirical models for the force components and the torque, valid for moderately inclined cylinders ($|\theta|\lesssim30^\circ$) of arbitrary aspect ratio up to $\text{Re}\approx300$ \color{black} and matching those obtained at low-to-moderate Reynolds number. \color{black}
\section{Introduction}
\label{intro}
The flow past rodlike cylindrical particles or cylindrical fibers with circular cross section is involved in many industrial and natural processes such as bubbling fluidized beds, pulp and paper making or the sedimentation of ice crystals in clouds. Despite the large number of studies devoted to the flow past a circular cylinder held perpendicular to the incoming flow in the laminar and transitional regimes, much less is known when the body is arbitrarily inclined or even aligned with this incoming flow. Three dimensionless parameters then govern the problem when the upstream flow is steady and uniform: the aspect ratio $\cchi=L/D$ where $L$ is the length of the cylinder and $D$ its diameter, the inclination angle $\theta$ (pitch or yaw angle) which is the angle between the cylinder axis and the incoming velocity, and the Reynolds number $\text{Re}=\frac{\rho UD}{\mu}$, where $U$ is the norm of the upstream velocity, and $\rho$ and $\mu$ are the fluid density and dynamic viscosity, respectively.\\
\indent Up to now, the flow past long rigid cylinders and the loads acting upon them have been investigated in two markedly different contexts. The first of them is the dynamics of dilute suspensions of slender particles, the theoretical study of which was pioneered by Batchelor \cite{batchelor1970} and Cox \cite{cox1970} in the creeping-flow limit. The corresponding results for the force and torque acting on an isolated fiber (see also \cite{tillett1970}), based on the slender-body theory, were extended to low-but-finite Reynolds numbers by Khayat and Cox \cite{khayat1989}. Since then, these various predictions have been extensively used in numerical simulations, be it to study the influence of hydrodynamic interactions and concentration on the sedimentation of a fiber suspension \citep{mackaplow1998,butler2002} or to reveal the dispersion properties of such suspensions in isotropic \citep{shin2005} or wall-bounded \citep{marchioli2010} turbulence; see \cite{voth2017} for a review.\\
\indent The second stream of investigations, often motivated by vortex-induced vibrations and applications to fluid-structure interactions, has focused on larger Reynolds numbers.
\citet{ramberg1983} studied experimentally the flow past long inclined cylinders with various end shapes for Reynolds numbers in the range $1.5\times10^2-10^3$. His experiments provide a qualitative map of the wake topology and shedding process as a function of the inclination angle.
In particular, they show that, unlike the classical vortex patterns observed when the cylinder is held perpendicular to the flow \citep{williamson1996}, the wake is dominated by a pair of counter-rotating vortices emanating from the ends when the inclination angle is small enough.
Numerical studies of the flow past an inclined cylinder in inertia-dominated regimes have also been reported \cite{vakil2009,pierson2019}. Based on the numerical data, these investigations proposed empirical expressions for the drag and lift forces valid throughout the considered range of $\cchi$ and $\text{Re}$ for arbitrary inclinations. Both studies examined the applicability and limitations of the so-called Independence Principle \citep{sears1948, ramberg1983,zdravkovich1997b}. This `principle', which states that the perpendicular force on a long cylinder depends solely on the normal velocity component of the incoming flow, was shown to apply only to large inclinations, $\theta\gtrsim45^\circ$. \citet{pierson2019} also considered the possibility to extend trigonometric relations valid under Stokes flow conditions to obtain the drag and lift forces at an arbitrary $\theta$ through simple linear combinations of the drag forces corresponding to the two extreme cases $\theta=0^\circ$ and $\theta=90^\circ$, an approach that has proved successful for prolate spheroids over a wide range of Reynolds numbers \citep{sanjeevi2017}.
\vspace{2mm}\\
\indent In this paper, we expand on available studies by considering the flow past finite-length cylinders with flat ends, from creeping flow conditions up to \color{black} $\text{Re}\gtrsim 300$ \color{black} for moderate inclinations, namely $0^\circ\leq\theta\leq30^\circ$. The cylinder aspect ratio is varied from $2$ to $20$ in the reference case $\theta=0^\circ$, and from $3$ to $7$ in inclined configurations. By considering this wide range of Reynolds number, we aim at bridging the gap between conditions typical of submillimeter-diameter fibers relevant for instance to papermaking (for which $D$ stands typically in the range $15-30\,\mu$m) and millimeter-diameter rodlike particules relevant to fluidized beds (which have aspect ratios typically in the range $2-10$). The reason why we concentrate on low-to-moderate inclinations is two-fold. First, as already mentioned, the current knowledge of the flow structure and drag variations with $\text{Re}$ over a slender circular cylinder aligned with the flow is still far from complete. To the best of our knowledge, no study has considered this configuration from creeping-flow conditions up to Reynolds of some hundreds which are easily reached in some of the applications mentioned above. Similar to the case of a finite-length cylinder held perpendicular to the flow, a more complete description of this reference configuration is mandatory to improve the predictions of the trajectories of sedimenting rodlike particles and fibers spanning all possible orientations with respect to their path. Then, with the same final objective in mind, an important question in inertia-dominated regimes is to understand how the flow structure and the loads on the body are affected by the loss of axial symmetry encountered as soon as the inclination angle becomes nonzero. In particular, given the three-dimensional nature of the flow past an inclined cylinder, it is not clear how far the trigonometric approach mentioned above can be used to predict realistically the loads acting on it, based only on results for the axisymmetric geometry corresponding to $\theta=0^\circ$ and the nearly two-dimensional geometry (for large $\chi$) corresponding to $\theta=90^\circ$. These are the main objectives of the present investigation.\\
\indent The paper is organized as follows. We first present the numerical approach in Sec. \ref{num} and, in appendix \ref{app:num}, provide a validation of this approach by comparing some of the results obtained at moderate Reynolds number with those of \cite{pierson2019}. In Sec. \ref{aligned}, we specifically examine the case of a finite-length cylinder aligned with the incoming flow. We first consider (Sec. \ref{creep}) low-to-moderate Reynolds numbers, for which we use numerical results to improve over available drag predictions provided by the slender-body approximation (in appendix \ref{app:slender}, we extend the available theoretical prediction in the creeping-flow limit by computing explicitly the next-order finite-aspect-ratio correction). Then (Sec. \ref{moder}), we consider Reynolds numbers in the range $20-400$ and use numerical predictions for the pressure and viscous contributions to the drag to obtain an empirical drag law valid whatever $\cchi$ throughout this range of $\text{Re}$. The inclined configuration is examined in \color{black} Sec. \ref{incl1} in the low-to-moderate Reynolds number range.
\color{black} We compare numerical findings with available theoretical predictions and make use of results established in Sec. \ref{creep} to provide semiempirical laws for the drag, lift and torque as a function of the control parameters (since the law for the transverse force involves the drag on a cylinder held perpendicular to the flow, we discuss slender-body predictions in this specific configuration and extend them empirically in appendix \ref{perps}). In Sec. \ref{incl2}, we proceed with the flow past an inclined cylinder in the moderate-to-large Reynolds number range $10\lesssim\text{Re}\lesssim300$. We analyze the structure of the steady non-axisymmetric flow and its connection with the observed, sometimes nonintuitive, variations of the loads with $\text{Re}$, $\theta$ and $\cchi$. We finally provide empirical fits capable of reproducing these complex variations throughout the explored range of parameters. \color{black} A summary of the main results and a discussion of some open issues are provided in Sec. \ref{Summ}
\section{Numerical methodology}
\label{num}
We consider the uniform incompressible steady flow of a Newtonian fluid past a finite-length circular cylinder. Computations are carried out with the JADIM code developed at IMFT. This code was used in the past to investigate various problems involved in the local dynamics of particle-laden and bubbly flows, among which the hydrodynamic forces acting on spheres in uniform or accelerated flows \citep{magnaudet1995}, the transition in the wake of spheres, disks and short cylinders \citep{fabre2008, auguste2010bif} and the path instabilities of freely-falling disks and light spheres \citep{auguste2013, auguste2018}. The code solves the three-dimensional unsteady Navier-Stokes equations using a finite volume discretization on a staggered grid. Centered schemes are used to discretize spatial derivatives in the momentum equation. Time advancement is achieved by combining a third-order Runge-Kutta algorithm for advective terms with a Crank-Nicolson scheme for viscous terms. The divergence-free condition is satisfied to machine accuracy at the end of each time step using a projection method. More details on the numerical methodology may be found in \cite{magnaudet1995} and \cite{calmet1997}.\\
\indent The present investigation makes use of a cylindrical computational domain with length $\mathcal{L}$ and radius $\mathcal{R}$ (Fig. \ref{fig:domain2D}). The length $\mathcal{L}$ may be decomposed as $\mathcal{L}=L_{up}+L+L_{down}$, where $L_{up}$ (resp. $L_{down}$) is the distance between the domain inlet (outlet) and the upstream (downstream) end of the cylinder. In what follows, for moderate-to large Reynolds numbers (say $\text{Re}>5$), we select $L_{up}=12D\cchi^{1/3}$ and $L_{down}=20D\cchi^{1/3}$. The reason why $L_{up}$ and $L_{down}$ are defined based on $D\cchi^{1/3}$, a length scale proportional to the diameter of the equivalent sphere, \textit{i.e.} the sphere with the same volume as the cylinder, is that this choice makes the size of the domain vary with the body aspect ratio while keeping the computational cost reasonable \citep{pierson2019}. With the above choice, $L_{down}$ is larger than $15D$ even for $\cchi=\mathcal{O}(1)$, which guarantees that the near wake is properly resolved for short-length cylinders \citep{auguste2010}. The radius of the numerical domain is chosen as $\mathcal{R}=0.5D + 20D\cchi^{1/3}(1+0.8\sin\theta)$ (Fig. \ref{fig:domain2D}). It increases with the inclination angle to make sure that the wake is correctly captured whatever the body inclination. \color{black} At low Reynolds number, care must be taken of artificial confinement effects inherent to the $1/\mathfrak{r}$-decay of the disturbance (with $\mathfrak{r}$ the distance to the body center). For this reason, in the low-to-moderate-$\text{Re}$ regime $\text{Re}\leq5$, we increase $L_{up}$ and $L_{down}$ by a factor of at least $2$, and increase $\mathcal{R}$ by a factor of at least $3$ compared to the above values. \color{black}
A uniform fluid velocity making an angle $\theta$ with the body axis is specified on the inlet plane (Fig. \ref{fig:domain2D}). A no-slip boundary condition is imposed on the body, while a non-reflecting boundary condition is imposed on both the outlet plane and the lateral boundary \citep{magnaudet1995}. We define a Cartesian coordinate system ($x,y,z$) centered at the body geometrical center, with $x$ parallel to the body symmetry axis and $z$ perpendicular to the plane containing the direction of the upstream flow and the body axis. In this coordinate system, the incoming velocity is $\mathbf{U}=(U\cos\theta,U\sin\theta,0))$.
\begin{figure}[H]
\centering
\begin{tikzpicture}[scale=1.0]
\draw (1.5,1) -- (3.5,1) -- (3.5,1.5) -- (1.5,1.5) -- (1.5,1) ;
\draw[red, <->] (1.5,1.25) -- (3.5,1.25) node [at end, above left, red] {$L$};
\draw[red, <->] (2.5,1)node [below, red] {$\frac{D}{2}$} -- (2.5,1.5) ;
\draw (-2,1) rectangle (7.5,3.5);
\draw[dashed] (1.25,1)--(1.25,1.75);
\draw[dashed] (1.25,1.75)--(3.75,1.75);
\draw[dashed] (3.75,1.75)--(3.75,1);
\draw[<->, thick] (-2,4)-- node[above]{$\mathcal{L}$} (7.5,4) ;
\draw[<->, thick] (9,1)-- node[right]{$\mathcal{R}$} (9,3.5) ;
\draw[<->, thick] (-2.0,1.25)--(1.5,1.25) node[at start, above right]{$L_{up}$};
\draw[<->, thick] (3.5,1.25)--(7.5,1.25) node[at end, above left]{$L_{down}$};
\draw[->, thick] (5,0)--(5.5,0) node[right]{$\bf{e_x}$};
\draw[->, thick] (5,0)--(5,0.5) node[right]{$\bf{e_y}$};
\draw[->, thick] (5,0)--(4.65,-0.28) node[left]{$\bf{e_z}$};
\draw[dashed] (-0.5,1,2)--(0.5,1,2);
\draw[->,black] (0.5,1,2) arc (0:23:1.1)node[right]{$\theta$};
\draw[->, thick] (-0.5,1,2)-- node[above]{$\mathbf{U}$} (0.5,1.5,2) ;
\draw (-2.5,2.25) node[]{Inlet};
\draw (3.5,3.7) node[]{Outlet};
\draw (8.3,2.25) node[]{Outlet};
\end{tikzpicture}
\caption{Scheme of the computational domain in an azimuthal plane (not to scale).}
\label{fig:domain2D}
\end{figure}
Simulations are performed with an axisymmetric cylindrical grid made involving regions with uniform and non-uniform cell distributions. In the cross-sectional plane $z=0$, a refined uniform distribution is used near the corners of the body, to properly capture the local flow (see \cite{auguste2013} for details). A slightly non-uniform distribution is imposed in a rectangular region extending up to $0.5D$ outside of the body in each direction (dashed line in Fig. \ref{fig:domain2D}). In this flow region, the cell aspect ratio is maintained below $4$ everywhere. Non-uniform cell distributions are used around the symmetry axis $y=z=0$ and near the body symmetry plane $x=0$. \color{black} In the low-to-moderate Reynolds number regime $\text{Re}\leq5$, we select a grid with $20$ cells per body diameter (earlier computations were performed with only $12$ cells per body diameter and minimal changes were noticed between the two resolutions). For larger Reynolds numbers, \color{black} we assume that the boundary layer thickness scales as $D\text{Re}^{-1/2}$ and make sure that at least $8$ cells stand within it. The characteristic grid size in this region is thus $\Delta \approx D/(8\text{Re}^{1/2})$. Whatever the Reynolds number, the grid is non-uniform in the outer region (\textit{i.e.} beyond the dashed rectangle in Fig. \ref{fig:domain2D}), with cell sizes following a geometric law. For $\text{Re}>5$, the growth of the cells along the domain axis is controlled to guarantee that the wake is adequately resolved. In the azimuthal direction, $32$ to $128$ planes are used, depending on the Reynolds number. \color{black} No discernible difference in the solutions returned by the $32$ and the $64$ azimuthal resolutions was noticed up to $\text{Re}=5$. \color{black} The highest azimuthal resolution ensures that the cells closest to the body have approximatively the same size in all three directions. Although the code has been extensively validated in the past, an additional validation involving a grid convergence analysis is reported in appendix \ref{app:num}. Present results are found to agree well with those of \cite{pierson2019} obtained with a distinct numerical methodology.
\section{Cylinder aligned with the upstream flow}
\label{aligned}
In this section we investigate the flow past a finite-length cylinder aligned with the incoming velocity. Our main purpose is to provide drag laws for cylinders of arbitrary aspect ratio beyond $2$ over a wide range of Reynolds number ($0.05\leq \text{Re} \leq 400$). We consider aspect ratios up to $20$ and discuss the results in ascending order of Reynolds number.
\subsection{From creeping-flow conditions to $\mathcal{O}(10)$-Reynolds numbers}
\label{creep}
No exact results for the stress distribution on a finite-length cylinder exists in the Stokes regime. \citet[]{clift1978} reported a detailed comparison of numerical and experimental results with empirical laws from the literature aimed at estimating the drag force on a cylinder aligned with the upstream flow. Approximating the cylinder geometry with a prolate spheroid of same volume yields a relative error up to $15 \%$ on the drag for small aspect ratios ($\cchi \approx 2$).
To provide a drag law valid for $\cchi \gtrsim 2$, we start from the slender-body theory. This theory provides a convenient framework to compute the forces on a slender body under creeping-flow conditions through an expansion with respect to the small parameter $ 1/\ln(2\cchi)$ for $\cchi \gg 1$ \citep{batchelor1970,cox1970}. The solution corresponding to a a straight circular fiber was obtained to third order by \citet{batchelor1970} and \citet{keller1976}. In appendix \ref{app:slender}, we improve over these predictions by computing explicitly the fourth-order term. In what follows, we compare this improved prediction with several sources, namely the numerical results of Youngren and Acrivos \cite{youngren1975}, present numerical results obtained for $\text{Re}=0.05$, and experimental results by \citet{heiss1952}.
\begin{figure}
\centering
\includegraphics[scale=0.5]{force_re0.pdf}
\caption{Drag on a finite-length cylinder aligned with the flow direction, normalized by the drag $F_{ds}$ of a sphere of same volume. Dotted, dash-dotted and dashed lines: predictions of $2^{nd}$-, $3^{rd}$- and $4^{th}$-order slender-body approximations, respectively; solid line: semiempirical formula (\ref{eq:slender_modified}); $\bullet$: numerical results of \cite{youngren1975}; $\blacksquare$: experimental results of \cite{heiss1952}; $\blacktriangledown$: present numerical results for $\text{Re}=0.05$.}
\label{fig:forcere0}
\end{figure}
Figure \ref{fig:forcere0} displays the drag force $F_d\equiv \bf{F}\cdot\bf{e_x}$ on the body, normalized by the force $F_{ds}$ on a sphere of same volume (\textit{i.e.} with diameter $\mathfrak{D}/D=\left(\frac{3}{2}\cchi)^{1/3}\right)$), as a function of $\cchi$. Clearly, the second-order slender-body approximation (see appendix \ref{app:slender}) is quite inaccurate, even for large aspect ratios. The third-order approximation provides a better agreement, but significant deviations ($>5\%$) still exist for $\cchi=20$.
The fourth-order approximation computed in appendix \ref{app:slender} approaches the experimental and numerical results down to $\cchi \approx10$ significantly better. However, all slender-body approximations inherently diverge when $\cchi\rightarrow1/2$. Actually, Fig. \ref{fig:forcere0} indicates that they are all inaccurate for $\cchi\lesssim 5$; the higher the order of the expansion the larger the aspect ratio below which the slender-body approximation becomes inaccurate. To extend the domain of validity of the theoretical approximation towards short cylinders, we empirically correct the fourth-order approximation by introducing an \textit{ad hoc} additional term.
\color{black} This term must attenuate the divergence of the slender-body approximation for $\cchi\rightarrow1/2$ and become negligible for $\cchi=\mathcal{O}(10)$. Therefore we sought it in the form $\cchi^{1/3}(\cchi-1/2)^{-p}$, in such a way that only a $(\cchi-1/2)^{-p}$ correction is introduced in the normalized force $F_d/F_{ds}$. The best fit with experimental and numerical data in the range $2\leq\cchi\leq10$ is obtained with $p=1.75$, a $\pm3\%$-difference on $p$ leading to a significantly poorer agreement. \color{black}
The full expression for the drag force incorporating this empirical term then reads
\begin{equation}
F_{d}^{Re=0} \approx 2 \pi \mu L U \left( \frac{a^{(1)}}{\ln(2\cchi)} + \frac{a^{(2)}}{(\ln(2\cchi))^2} + \frac{a^{(3)}}{(\ln(2\cchi))^3} + \frac{a^{(4)}}{(\ln(2\cchi))^4} - \frac{2.4}{\cchi ^{2/3}(\cchi -\frac{1}{2})^{1.75}} \right)\,,
\label{eq:slender_modified}
\end{equation}
\noindent with $U=\bf{U}\cdot\bf{e_x}$ and the expressions for the $a ^{(i)}$ as provided in appendix \ref{app:slender}. As Fig. \ref{fig:forcere0} shows, the modified drag law fits the numerical and experimental results well for $\cchi \gtrsim2$. It still diverges for lower $\cchi$, but the drag is then close to that experienced by the equivalent sphere ($F_d/F_{ds}\leq1.05$).
Present numerical predictions in the range $3\leq\cchi\leq10$ are also seen to agree well with results available in the literature.
The slight deviation observed for $\cchi=20$ is not unexpected. Indeed, the Reynolds number based on the body length is of $\mathcal{O}(1)$ in this case, making inertial corrections to the drag significant (see below).
\begin{figure}[h]
\centering
\includegraphics[scale=0.4]{amont_p_s_xR0v05}
\caption{Radial pressure distribution on the upstream end of the cylinder for $\text{Re}=0.05$. $\bullet$: $\cchi=3$; $\star$: $\cchi=5$; $+$: $\cchi=7$; $\blacktriangledown$: $\cchi=10$; $\blacktriangle$: $\cchi=20$. The pressure is assumed to be zero in the far field.}
\label{fig:pressure_re0}
\end{figure}
The pressure component to the drag, $F_{dp}$, results from the difference between the pressure distributions on the upstream and downstream ends of the cylinder. Owing to the fore-aft symmetry of the flow in the Stokes regime, these two distributions only differ by the sign of the corresponding pressure, since $p(r, x=L/2)=-p(r,x=-L/2)$. Figure \ref{fig:pressure_re0} displays the radial pressure distribution on the upstream end. For each $\cchi$, the pressure is seen to increase with the distance $r$ to the symmetry axis. At a given radial position, pressure variations are only mildly influenced by the aspect ratio, but this influence becomes larger as the lateral surface is approached.
\begin{figure}[h]
\centering
\includegraphics[scale=0.4]{forcep_re0}
\includegraphics[scale=0.4]{forcep_norm_re0.pdf}\\
\vspace{-1mm}
\hspace{-1.5cm}{$(a)$}\hspace{6cm}{$(b)$}
\caption{Contributions to the drag force. $(a)$: pressure contribution; $(b)$: ratio of the pressure to the shear stress contributions. $\blacktriangledown$: numerical results for $\text{Re}=0.05$; solid line: low-$\text{Re}$ prediction (\ref{eq:prolate_spheroid_Re0}) for a prolate spheroid.}
\label{fig:forces_mu_p_re0}
\end{figure}
Figure \ref{fig:forces_mu_p_re0}$(a)$ indicates that $F_{dp}$ only varies weakly with $\cchi$. It consistently decreases with the aspect ratio but its variation is less than 20\,\% from $\cchi=3$ to $\cchi=20$. Figure \ref{fig:forces_mu_p_re0}$(b)$ shows the ratio of the pressure contribution to the viscous stress contribution to the drag as a function of $\cchi$. This ratio is close to $1/3$ for $\cchi=3$. Then it decreases gradually as $\cchi$ increases, and becomes less than $1/10$ for $\cchi=20$. Hence, as expected from purely geometrical considerations, $F_{dp}$ becomes negligible with respect to $F_{d\mu}$ for large $\chi$. It is of some interest to compare this variation with that found for prolate spheroids, namely \citep{masliyah1970}
\begin{equation}
\displaystyle
\frac{F_{dp}}{F_{d\mu}} = \frac{\cchi \ln (\cchi + (\cchi ^2 -1)^{1/2})-(\cchi ^2 -1)^{1/2}}{\cchi ^2(\cchi ^2 -1)^{1/2}-\cchi \ln(\cchi + (\cchi ^2 -1)^{1/2})}\,.
\label{eq:prolate_spheroid_Re0}
\end{equation}
\noindent As Fig. \ref{fig:forces_mu_p_re0}$(b)$ indicates, the pressure drag decreases much more slowly with the aspect ratio in the case of a cylindrical body with flat ends. This difference underlines the importance of end effects even for long cylindrical bodies.\vspace{2mm}\\
\indent Next we consider the influence of inertial effects at low-to-moderate Reynolds number.
The slender-body theory initially derived under Stokes flow conditions was extended to small-but-finite Reynolds numbers in \cite{khayat1989}, still assuming $\text{Re} \ll 1$ but considering that the Reynolds number based on the body length, ${\cchi\text{Re}}$, may be arbitrarily large. \color{black} \citet{lopez2017} and \citet{roy2019} provided experimental confirmations of the relevance of the finite-$\text{Re}$ corrections derived in \cite{khayat1989} by examining the settling of long fibers ($10\lesssim\cchi \lesssim35$) in a Taylor-Green type vortical flow and the sedimentation of isolated fibers with $\cchi=20$ and $100$ in a fluid at rest, respectively. \color{black} Nevertheless, we are not aware of any detailed validation of the finite-$\text{Re}$ theory of \cite{khayat1989} in the case of a body aligned with the upstream flow. According to this theory (see also \cite{lopez2017} and \cite{roy2019}), the drag force on a long cylindrical body aligned with the flow, disregarding terms of $\mathcal{O}((1/\ln\cchi)^3)$ and higher, reads
\begin{equation}
F_d^{\cchi\text{Re}=\mathcal{O}(1)}\approx 2 \pi \mu L U \left( \frac{a^{(1)}}{\ln\cchi} + \frac{a^{(2)}-\ln2+f_\parallel}{(\ln\cchi)^2} \right)\,,
\label{eq:slender_smallinertia}
\end{equation}
where
\begin{equation}
f_\parallel =\frac{1}{2}\left( \frac{E_1(\cchi \text{Re} )+\ln(\cchi \text{Re} )-e^{-\cchi \text{Re} }+\gamma +1}{\cchi \text{Re} } + E_1(\cchi \text{Re} ) +\ln(\cchi \text{Re}) +\gamma -2\right) \,,
\label{fpKC}
\end{equation}
\noindent with $E_1(X)=\int_X^{+\infty}{\frac{e^{-t}}{t}dt}$ (related to the exponential integral $\text{Ei}(X)$ through $E_1(X)=-\text{Ei}(-X)$), and $\gamma$ the Euler constant. In the limit of small $\cchi\text{Re}$, the inertial correction factor reduces to $f_\parallel\approx\frac{1}{8}\cchi\text{Re}$. Note that in (\ref{eq:slender_modified}) and throughout the rest of the paper, the slender-body solution is expanded with respect to $1/\ln(2\cchi)$, similar to \cite{batchelor1970}, \cite{tillett1970} and \cite{,keller1976}. In contrast, an expansion with respect to $1/\ln(\cchi)$ was used in \cite{khayat1989}. Once truncated at the same order, the two formulations are of course equivalent for $\cchi\gg2$, but significant differences exist for moderate aspect ratios. This is why we keep the original formulation in (\ref{eq:slender_smallinertia}) and in similar expressions of Sec. \ref{incl1} involving the inertial corrections derived in \cite{khayat1989}. \\%We observe that inertia effects do not change the leading order term but appear in the expansion at order 2.
\indent The main shortcoming of (\ref{eq:slender_smallinertia}) is obviously the truncation at second order with respect to $1/\ln\cchi$. This limitation is confirmed in Fig. \ref{fig:small_inertia}, where the influence of inertial effects on the drag force is shown for cylindrical bodies with different aspect ratios. Significant deviations between numerical results and predictions of (\ref{eq:slender_smallinertia}) are observed whatever $\text{Re}$ and $\cchi$.
\begin{figure}[h]
\centering
{\includegraphics[scale=0.35]{force_smallinertia_chi3.pdf}\label{fig:a}}
{\includegraphics[scale=0.35]{force_smallinertia_chi5.pdf}\label{fig:b}}\\
\vspace{-4mm}
\hspace{-4cm}$(a)$\hspace{5cm}$(b)$\\
\vspace{2mm}
{\includegraphics[scale=0.35]{force_smallinertia_chi10.pdf}\label{fig:c}}
{\includegraphics[scale=0.35]{force_smallinertia_chi20.pdf}\label{fig:d}}\\
\vspace{-4mm}
\hspace{-4cm}$(c)$\hspace{5cm}$(d)$\\
\vspace{-72mm}
\hspace{-14mm}$\cchi=3$\hspace{5cm}$\cchi=5$\\
\vspace{37mm}
\hspace{-12mm}$\cchi=10$\hspace{4.8cm}$\cchi=20$
\vspace{33mm}
\caption{Influence of inertial effects on the drag of finite-length cylinders aligned with the upstream flow. The drag is normalized by that of a sphere of same volume.
$\blacktriangledown$: numerical results; solid line: semiempirical prediction (\ref{eq:slender_smallinertia2}); dash-dotted line: theoretical prediction (\ref{eq:slender_smallinertia}); dotted line: predictions (\ref{eq:slender_modified}) (upper line) and (\ref{eq:slender_smallinertia}) with $\text{Re}=0$ (lower line).}
\label{fig:small_inertia}
\end{figure}
It is thus desirable to include higher-order corrections with respect to $1/\ln(2\cchi)$ to improve the validity of (\ref{eq:slender_smallinertia}). The analysis of \cite{khayat1989} was recently extended to third order in \cite{khair2018} using the reciprocal theorem. However, the pre-factor of the third-order term is found in the form of a volume integral to be evaluated in Fourier space. To obtain a more straightforward formula accounting for small-but-finite inertial effects, we use present numerical results to empirically modify (\ref{eq:slender_smallinertia}) by taking advantage of the higher-order corrections present in (\ref{eq:slender_modified}).
The best agreement with the numerical results is obtained with the expression
\begin{equation}
F_d^{\cchi\text{Re}=\mathcal{O}(1)} \approx 2 \pi \mu L U \left[ \frac{a^{(1)}}{\ln(2\cchi)} + \frac{a^{(2)}+ f_\parallel}{(\ln(2\cchi))^2} + \frac{a^{(3)}+ f_3f_\parallel}{(\ln(2\cchi))^3} + \frac{a^{(4)}+ f_4f_\parallel }{(\ln(2\cchi))^4} - \frac{2.4}{\cchi ^{2/3}(\cchi -\frac{1}{2})^{1.75}} \right]\,,
\label{eq:slender_smallinertia2}
\end{equation}
\noindent with $f_3=(\cchi\text{Re})^{0.07\chi^{0.5}}$ and $f_4=(\cchi\text{Re})^{0.03\chi^{0.9}}$. Predictions of (\ref{eq:slender_smallinertia2}) are compared with numerical results in Fig. \ref{fig:small_inertia}. The agreement is found to be good up to $\text{Re}\approx 1$ whatever the aspect ratio. This is quite remarkable since the derivation of (\ref{eq:slender_smallinertia2}) assumes $\text{Re}\ll1$ (but possibly $\cchi\text{Re}\gtrsim1$ since $\cchi\gg1$). \citet{khair2018} also found that their analytical prediction agrees well with the numerically computed drag on a long prolate spheroid up to $\text{Re} \approx 2$ for $\cchi=10$.
\begin{figure}
\centering
\includegraphics[scale=0.213]{recirculation_x2R140}\hspace{15mm}
\includegraphics[scale=0.275]{ldc_t0x2R300.png}\\
\vspace{1mm}\hspace{-55mm}$(a)$\hspace{75mm}$(b)$
\caption{Streamlines pattern for $\cchi=2$ colored with the magnitude of the axial velocity (from $-0.25$ to $1$). $(a)$ $\text{Re}=140$; $(b)$ $\text{Re}=300$.
\label{fig:streamlines}}
\end{figure}
\subsection{From $\mathcal{O}(10)$ to $\mathcal{O}(4\times10^2)$ Reynolds numbers}
\label{moder}
\color{black}Increasing the Reynolds number, we considered six aspect ratios ranging from $2$ to $10$ and twenty Reynolds numbers from $\text{Re}=20$ to $\text{Re}=400$. \color{black} The flow past the cylinder was found to reach a steady state in all cases. Nevertheless, the flow structure revealed new features as the Reynolds number increases.\\
Figure \ref{fig:streamlines}$(a)$ displays the streamlines around a cylinder with $\cchi=2$ for $\text{Re}=140$. In this regime, the flow is attached to the body all along the lateral surface but the separation of the boundary layer at the downstream edge results in the generation of a toroidal eddy in the near wake. Simulations indicate that this standing eddy sets in for $\text{Re} \approx 10$ for $\cchi=2$ and $\text{Re} \approx 20$ for $\cchi=7$.
\begin{figure}[h]
\centering
\includegraphics[scale=0.4]{lr_t0x.pdf}
\caption{Variation of the length $l_r$ of the toroidal eddy vs. the Reynolds number past cylinders with different aspect ratios; $l_r$ is measured along the symmetry axis, from the downstream end of the cylinder to the point at which the axial velocity changes sign. $\blacktriangle$: $\cchi=1$ from \cite{auguste2010}; $\blacktriangledown$: $\cchi=2$; $\bullet$: $\cchi=3$; $\times$: $\cchi=4$; $\star$: $\cchi=5$; $+$: $\cchi=7$; solid line: $l_r(\text{Re})=0.083\text{Re}^{1/2}$.
\label{fig:toroidal}}
\vspace{-3mm}
\end{figure}
\noindent The length of the standing eddy is plotted in Fig. \ref{fig:toroidal} as a function of the Reynolds number. Remarkably, it is found to be almost independent of $\cchi$ and grows approximately as the square root of the Reynolds number.
As the Reynolds number increases, a thin secondary annular eddy sets in along the upstream part of the lateral surface of the body (Fig. \ref{fig:streamlines}$(b)$), owing to the detachment of the boundary layer along the upstream edge. The critical Reynolds number $\text{Re}_{c0}$ beyond which this flow pattern is detected is a slowly increasing function of the aspect ratio: from $\text{Re}_{c0}\approx180$ for $\cchi=2$, to $\text{Re}_{c0}\approx200$ for $\cchi=5$, $\text{Re}_{c0}\approx220$ for $\cchi=7$, until $\text{Re}_{c0}\approx240$ for $\cchi=10$. It will be seen later that this secondary eddy has a strong influence on the viscous friction experienced by the cylinder
\begin{figure}[h]
\centering
\hspace{-10mm}\includegraphics[height=4cm]{Cp_west}
\includegraphics[height=4cm]{Cp_east}
\includegraphics[height=4cm]{Cp.pdf}
\\\vspace{-3mm}
\hspace{-3cm}$(a)$\hspace{55mm}$(b)$\hspace{55mm}$(c)$
\caption{Pressure drag coefficient $C_{p}$ as a function of $\text{Re}$ for different aspect ratios. $(a)$: upstream end; $(b)$: downstream end; $(c)$: sum of the two contributions. $\blacktriangledown$: $\cchi=2$; $\times$: $\cchi=4$; $+$: $\cchi=7$; $\blacktriangle$: $\cchi=10$; solid line in $(a)$: empirical fit (\ref{eq:cp_theta0_up}); solid line in $(b)$: empirical fit (\ref{eq:cp_theta0_down}); solid line in $(c)$: sum (\ref{eq:cp_theta0_up})+(\ref{eq:cp_theta0_down}) for $\cchi=2$ (upper line) and $\cchi=10$ (lower line).
\label{fig:Cp_theta0_relarge}}
\end{figure}
\noindent Figures \ref{fig:Cp_theta0_relarge}$(a)-(b)$ display the variation with $\text{Re}$ of the pressure drag coefficient $C_p=8F_{dp}/(\pi\rho D^2U^2)$ on the upstream and downstream ends of the body for various $\cchi$. For each aspect ratio, both contributions decrease monotonically with $\text{Re}$. The pressure coefficient on the downstream end ($C_{p_{down}}$) is $3-4$ times smaller than that on the upstream end ($C_{p_{up}}$) for $\text{Re}\approx100$ and becomes only a small fraction of the latter for $\text{Re}\approx400$.
Moreover, $C_{p_{up}}$ is almost independent of $\cchi$ (except for the shortest cylinder), while $C_{p_{down}}$ gradually decreases as the aspect ratio increases. Last, for $\cchi>2$ and $\text{Re}\gtrsim300$, $C_{p_{up}}$ is seen to tend toward an almost constant value slightly larger than $0.6$. Based on these remarks, approximate expressions for the two contributions take the form
\begin{eqnarray}
\label{eq:cp_theta0_up}
C_{p_{up}}(\cchi,\text{Re}) &\approx& 0.62 + \frac{11.7}{\text{Re}^{0.9}(1+0.004\text{Re}^{0.9})}\\
\label{eq:cp_theta0_down}
C_{p_{down}}(\cchi,\text{Re}) &\approx& 2.05\cchi^{-1/4}\text{Re}^{-1/2}
\end{eqnarray}
\noindent As Fig. \ref{fig:Cp_theta0_relarge}$(c)$ confirms, numerical data for the total pressure drag coefficient $C_p=C_{p_{up}}+C_{p_{down}}$ are accurately fitted by the sum of (\ref{eq:cp_theta0_up}) and (\ref{eq:cp_theta0_down}).
\begin{figure}[h]
\centering
\includegraphics[width=6cm]{Cmu.pdf}\hspace{15mm}\includegraphics[width=6cm]{Cd.pdf}
\vspace{-3mm}\\
\hspace{-5cm}$(a)$\hspace{7cm}$(b)$\\
\caption{Drag coefficient as a function of $\text{Re}$ for different aspect ratios. $(a)$: viscous contribution $C_{\mu}$; $(b)$: total drag coefficient $C_d$. $\blacktriangledown$: $\cchi=2$; $\times$: $\cchi=4$; $+$: $\cchi=7$; $\blacktriangle$: $\cchi=10$; for each aspect ratio, the solid lines in $(a)$ and $(b)$ represent the fits (\ref{eq:cmu_theta0}) and (\ref{eq:cd_theta0}), respectively.
\label{fig:Cmu_theta0_relarge}}
\end{figure}
Figure \ref{fig:Cmu_theta0_relarge}$(a)$ displays the variation of the viscous drag coefficient, $C_{\mu}(\cchi, \text{Re})$. This contribution is seen to be an increasing function of $\cchi$. Indeed, the area of the lateral surface of the body may be expressed in the form $\pi D^2\cchi$. Hence the normalized viscous force $C_{\mu}= 8F_{d\mu}/(\pi\rho D^2U^2)$ is expected to increase almost linearly with the aspect ratio.
As usual, $C_\mu$ is also a decreasing function of the Reynolds number. However, the observed decrease is much steeper than the classical $\text{Re}^{-1/2}$ behavior expected on the basis of the boundary layer theory. The reason for this stands in the presence of the secondary annular eddy along the lateral surface. The corresponding backflow generates a negative local shear stress which lowers the overall viscous drag. For sufficiently short cylinders and large enough Reynolds numbers, this negative contribution may exceed the positive contribution of the shear stress on the rest of the lateral surface, yielding an overall negative viscous drag. This change of sign takes place at $\text{Re}\approx 420$ for $\cchi=2$. Based on the previous findings, a simple fit for $C_\mu$ is found to be
\begin{equation}
C _\mu(\cchi,\text{Re}) \approx16.45 \cchi^{0.7}\text{Re}^{-0.8}+(a_1\cchi-a_2)\text{Re}\,,\quad\mbox{with}\quad a_1=4.1\times10^{-5}\quad\mbox{and}\quad\,a_2=6\times10^{-4}\,.
\label{eq:cmu_theta0}
\end{equation}
\noindent As Fig. \ref{fig:Cmu_theta0_relarge}$(a)$ shows, the above fit describes the variations of $C_\mu$ well throughout the entire range of aspect ratios and Reynolds numbers explored numerically. The last term in the right-hand side of (\ref{eq:cmu_theta0}) accounts for the influence of the annular eddy. Note that according to (\ref{eq:cmu_theta0}), the dependence of $C _\mu$ with respect to $\cchi$ is slightly weaker than expected on the basis of the above simple geometrical argument.\\%ormula matches reasonably with the numerical results up to an upper Reynolds number in the range $[350,\,400]$, depending on $\chi$. Beyond this limit, the effect of the lateral toroidal vortex on the shear friction becomes significant and (\ref{eq:cmu_theta0}) starts to deviate significantly from the numerical results.
As Figs. \ref{fig:Cp_theta0_relarge} and \ref{fig:Cmu_theta0_relarge} evidence, pressure effects contribute less to the drag than viscous friction for $\text{Re} \lesssim 100$. For Reynolds numbers in the range $100\lesssim \text{Re}\lesssim300$ and short cylinders ($\cchi \leq 4$), both contributions have a comparable magnitude. However, due to the sharp decrease of the viscous contribution for $\text{Re}\gtrsim400$ for such short cylinders, the latter eventually becomes smaller than the pressure contribution in the upper part of the $\text{Re}$-range covered by present computations. Adding the approximate expressions (\ref{eq:cp_theta0_up}), (\ref{eq:cp_theta0_down}) and (\ref{eq:cmu_theta0}), the total drag coefficient is approached as
\begin{equation}
C_d(\cchi,\text{Re})\approx C_{p_{up}}(\cchi,\text{Re})+C_{p_{down}}(\cchi,\text{Re})+C_\mu(\cchi,\text{Re})\,.
\label{eq:cd_theta0}
\end{equation}
As Fig. \ref{fig:Cmu_theta0_relarge}$(b)$ indicates, this fit is in good agreement with the numerically predicted drag throughout the range of $\cchi$ and $\text{Re}$ covered by the simulations. Note that the upper limit of validity of this fit is presumably $\text{Re}_{max}\approx400$. In particular, the linear variation of $C_\mu$ with $\text{Re}$ predicted by (\ref{eq:cmu_theta0}) cannot continue at very large Reynolds number, as the drag coefficient is expected to decrease with $\text{Re}$ for $\text{Re}>\text{Re}_{max}$ and become $\text{Re}$-independent in the limit $\text{Re}\rightarrow\infty$. Conversely, it may be checked that the above fit properly matches the modified low-but-finite $Re$ prediction (\ref{eq:slender_smallinertia2}). Consider for instance a cylinder with $\cchi=7$ and $Re=5$. On the one hand, (\ref{eq:cp_theta0_up})-(\ref{eq:cd_theta0}) predict $C_d\approx21.6$ for this set of parameters. On the other hand, according to Fig. \ref{fig:small_inertia}, the ratio $F_d/F_{ds}$ predicted by (\ref{eq:slender_smallinertia2}) is approximately $1.85$. Keeping in mind that the diameter $\mathcal{D}$ of the equivalent sphere is related to $D$ through $\mathcal{D}/D=(\frac{3}{2}\cchi)^{1/3}$, one has $C_d=24\text{Re}^{-1}(\frac{3}{2}\cchi)^{1/3}F_d/F_{ds}\approx21.9$. Similarly, for the same Reynolds number but $\cchi=20$, one has $C_d\approx40.7$ from (\ref{eq:cp_theta0_up})-(\ref{eq:cd_theta0}) and $C_d\approx40.3$ from Fig. \ref{fig:small_inertia} where $F_d/F_{ds}\approx2.70$.
This agreement, which is confirmed with other sets of parameters, allows us to conclude that combining (\ref{eq:slender_smallinertia2}) for Reynolds numbers less than a few units with (\ref{eq:cp_theta0_up})-(\ref{eq:cd_theta0}) for larger $\text{Re}$ provides an accurate description of drag variations from $\text{Re}=0$ up to $\text{Re}=400$
\color{black}
\section{Forces and torque on a moderately inclined cylinder at low-to-moderate Reynolds number}
\label{incl1}
\begin{figure}[h]
\centering
\begin{tabular}
{>{\centering\arraybackslash} m{2cm} >{\centering\arraybackslash} m{4cm}}
\input{cylindre_incline} &
\end{tabular}
\caption{ Schematic of the force components on an inclined cylinder. In this configuration, the inertial torque is negative, tending to rotate the cylinder clockwise.
\label{fig:cylindre_incline}}
\end{figure}
We now move to the more general configuration in which the cylinder is inclined with respect to the incoming flow by a nonzero angle. For reasons discussed in Sec. \ref{intro}, we limit ourselves to maximum inclinations of $30^\circ$. The present section focuses on the low-to-moderate Reynolds number range $[ 0.1-5]$. Higher Reynolds numbers are considered in the next section.\\\color{black}
\indent The nonzero inclination breaks the flow axial symmetry. Therefore, in addition to the force component $F_{\parallel}$ parallel to the cylinder axis, a perpendicular force component ($F_\perp$) takes place, together with a spanwise torque ($T$) (Fig. \ref{fig:cylindre_incline}).
These components are linearly related to the drag and lift forces $F_d$ and $F_l$, respectively parallel and perpendicular to the incoming velocity $\textbf{U}$, via the geometric relations
\begin{eqnarray}
\label{eq:geom1}
F_\parallel(\cchi,\theta, \text{Re})&=&F_d(\cchi,\theta, \text{Re})\cos\theta-F_l(\cchi,\theta, \text{Re})\sin\theta\,,\\
\label{eq:geom2}
F_\perp(\cchi,\theta, \text{Re})&=&F_l(\cchi,\theta, \text{Re})\cos\theta+F_d(\cchi,\theta, \text{Re})\sin\theta\,.
\end{eqnarray}
In the Stokes regime, the linearity of the loads with respect to the boundary conditions implies that the force acting on the inclined cylinder is linearly related to the drag acting on the same body in the two extreme configurations $\theta=0^\circ$ and $\theta=90^\circ$ through
\begin{eqnarray}
\label{eq:Cpara_stokes}
F_{\parallel}(\cchi,\theta)&=&F_\parallel^{ \theta=0^\circ}(\cchi)\cos\theta\,, \\
\label{eq:Cperp_stokes}
F_{\perp}(\cchi,\theta)&=&F_\perp^{ \theta=90^\circ}(\cchi)\sin\theta\,.
\end{eqnarray}
\color{black}
These simple `Stokes laws' are not expected to remain valid when inertial effects become significant. To assess and possibly extend their validity, an approximate expression for $F_\perp^{ \theta=90^\circ}$, similar to (\ref{eq:slender_smallinertia2}) for $F_\parallel$, is required. In appendix \ref{perps}, we establish the fourth-order slender-body approximation of $F_\perp^{ \theta=90^\circ}(\cchi,\text{Re}=0)$, and modify it empirically to extend its validity toward aspect ratios and Reynolds numbers of $\mathcal{O}(1)$. To get some insight into the way inertial effects alter (\ref{eq:Cpara_stokes})-(\ref{eq:Cperp_stokes}), it is informative to consider the finite-Reynolds-number approximate expressions for $F_d$ and $F_l$ established for arbitrary inclinations by \citet{khayat1989}.
Evaluating these expressions in the limit $\cchi\text{Re}\ll1$ and making use of (\ref{eq:geom1})-(\ref{eq:geom2}) yields
\begin{eqnarray}
\label{eq:Cpara_stokes1}
\frac{F_{\parallel}(\cchi,\theta,\cchi\text{Re}\ll1)}{2\pi\mu UL}&\approx&\left(F_\parallel^{* \theta=0^\circ}(\cchi,\cchi\text{Re}\ll1)-\frac{1}{16}\sin^2\theta\frac{\cchi Re}{(\ln\cchi)^2}\right)\cos\theta\,, \\
\label{eq:Cperp_stokes1}
\frac{F_{\perp}(\cchi,\theta,\cchi\text{Re}\ll1)}{4\pi\mu UL}&\approx&\left(F_\perp^{* \theta=90^\circ}(\cchi,\cchi\text{Re}\ll1)+\frac{1}{16}\cos^2\theta\frac{\cchi Re}{(\ln\cchi)^2}\right)\sin\theta\,,
\end{eqnarray}
where $F_\parallel^{*}$ and $F_\perp^{*}$ stand for the dimensionless second-order expansion of the corresponding force with respect to $1/\ln\cchi$ in the limit of low-but-finite $\cchi\text{Re}$, as provided by (\ref{eq:slender_smallinertia}) and (\ref{eq:slender_perp_smallinertia}), respectively. According to (\ref{eq:slender_smallinertia}) and the asymptotic form of the inertial correction $f_\parallel$ in the limit $\cchi\text{Re}\rightarrow0$, one has $F_\parallel^{* \theta=0^\circ}(\cchi,\cchi\text{Re}\ll1)=[\text{ln}\cchi]^{-1}+(\frac{3}{2}-2\text{ln}2+\frac{1}{8}\cchi\text{Re})[\text{ln}\cchi]^{-2}$. Similarly, (\ref{eq:slender_perp_smallinertia}) and the asymptotic form of $f_\perp$ yield $F_\perp^{* \theta=0^\circ}(\cchi,\cchi\text{Re}\ll1)=[\text{ln}\cchi]^{-1}+(\frac{1}{2}-2\text{ln}2+\frac{1}{4}\cchi\text{Re})[\text{ln}\cchi]^{-2}$. Expressions (\ref{eq:Cpara_stokes1})-(\ref{eq:Cperp_stokes1}) indicate that the angular dependence of $F_\parallel$ and $F_\perp$ becomes more complex in the presence of inertial effects, involving higher-order harmonics of $\theta$. Moreover they suggest that the $\theta$-dependent inertial corrections tend to decrease $F_\parallel$ and increase $F_\perp$, compared to the prediction of the extrapolated `Stokes law' based on the finite-$\text{Re}$ drag forces $F_\parallel^{* \theta=0^\circ}$ and $F_\perp^{* \theta=90^\circ}$.
\indent Still in the Stokes regime, the spanwise torque is zero whatever $\theta$, owing to the geometrical symmetries of the cylinder and the reversibility of Stokes equations. However, nonlinearities inherent to inertial effects result in a finite torque. In the limit $Re \cchi\ll1$, the finite-Reynolds-number expression for this torque obtained in \cite{khayat1989} reduces to
\begin{equation}
\frac{T(\cchi,\theta,\cchi\text{Re}\ll1)}{\mu U(L/2)^2}\approx-\frac{5\pi}{12}\frac{\cchi Re}{(\ln\cchi)^2}\sin2\theta\,.
\label{KCt0}
\end{equation}
This negative torque tends to rotate the cylinder perpendicular to the flow direction. \vspace{1mm}\\
\indent In what follows we make use of the numerical results to examine the validity of `Stokes laws' (\ref{eq:Cpara_stokes})-(\ref{eq:Cperp_stokes}) and of the predictions of \cite{khayat1989} for low-to-moderate Reynolds numbers.
\\
\begin{figure}[h]
\centering
\includegraphics[width=4.9cm]{cx_teta_R0v1.pdf}\includegraphics[width=5cm]{cx_teta_R0v5.pdf}\includegraphics[width=5cm]{cx_teta_R5.pdf}
\vspace{-3mm}\\
\hspace{-35mm}$(a)$\hspace{47mm}$(b)$\hspace{47mm}$(c)$\\
\vspace{-13mm}
\hspace{-13mm}$\text{Re}=0.1$\hspace{38mm}$\text{Re}=0.5$\hspace{38mm}$\text{Re}=5$
\vspace{11mm}
\caption{Parallel force component $F_\parallel$ vs. the inclination angle $\theta$ for $\bullet$: $\cchi=3$, $\star$: $\cchi=5$, $\blacktriangledown$: $\cchi=10$.
Dashed line: prediction (\ref{eq:Cpara_stokes}) based on the numerical value of $F_\parallel^{ \theta=0^\circ}$; solid line: prediction (\ref{eq:Cpara_stokes}) based on the semiempirical approximation (\ref{eq:slender_smallinertia2}) for $F_\parallel^{ \theta=0^\circ}$.
\label{fig:c_par_moderate}
\vspace{-6mm}
\end{figure}\indent
Figure \ref{fig:c_par_moderate} displays the parallel force component $F_\parallel$ for three Reynolds numbers and aspect ratios $\cchi=3,\,5$ and $10$, respectively. At the lower two $\text{Re}$, variations of $F_\parallel$ closely follow the $\cos\theta$-behavior predicted by (\ref{eq:Cpara_stokes}) whatever the aspect ratio. This implies that the $\sin^2\theta$-inertial correction in (\ref{eq:Cpara_stokes1}) has only a marginal influence at moderate inclinations for $\cchi\text{Re}\lesssim5$. Indeed, for $\theta=30^\circ$, this correction is less that $2\%$ of $F_\parallel^{* \theta=0^\circ}$ whatever the aspect ratio. However, the inertial correction included in $F_\parallel^{*}$ reaches $16\%$ for $\cchi=10$, indicating that inertial effects are already significant. In other words, in the range of moderate inclinations considered here, the `Stokes law' still accurately predicts $F_\parallel(\cchi,\theta,\text{Re})$ for $\text{Re}\lesssim1$, provided the prediction makes use of the inertia-corrected drag $F_\parallel^{\theta=0^\circ}(\cchi,\text{Re})$.
\\
\indent At $\text{Re}=5$, influence of inertial effects has become dominant.
Since the magnitude of $\theta$-dependent inertial correction also increases with $\text{Re}\cchi$, the larger the aspect ratio the stronger these effects at a given $\text{Re}$. This may be appreciated in Fig. \ref{fig:c_par_moderate}$(c)$, where the difference between the computed force and the prediction of the `Stokes law' is seen to increase significantly with $\cchi$ for $\theta>10^\circ$, from less than $2\%$ for $\cchi=3$ at $\theta=30^\circ$ to $9\%$ for $\cchi=10$ at the same inclination. Interestingly, the under-prediction of $F_\parallel$ by the Stokes law is at odds with the low-$\cchi\text{Re}$ prediction (\ref{eq:Cpara_stokes1}) which suggests that this `law' should overestimate $F_\parallel$, since it ignores the negative $\sin^2\theta$-inertial contribution. This contradiction emphasizes the fact that the $\theta$-dependences of inertial effects in the low-$\cchi\text{Re}$ range and in the range $10\lesssim\cchi\text{Re}\lesssim10^2$ are drastically different.
\begin{figure}[h]
\centering
\includegraphics[width=5cm]{cy_teta_xR0v1.pdf}\includegraphics[width=5cm]{cy_teta_xR0v5.pdf}\includegraphics[width=5cm]{cy_teta_xR5.pdf}
\vspace{-3mm}\\
\hspace{-35mm}$(a)$\hspace{47mm}$(b)$\hspace{47mm}$(c)$\\
\vspace{-32mm}
\hspace{-13mm}$\text{Re}=0.1$\hspace{38mm}$\text{Re}=0.5$\hspace{38mm}$\text{Re}=5$
\vspace{30mm}\\
\caption{Perpendicular force component $F_\perp$ as a function of $\theta$ for $\bullet$: $\cchi=3$, $\star$: $\cchi=5$, $\blacktriangledown$: $\cchi=10$.
Solid line: `Stokes law' (\ref{eq:Cperp_stokes}) based on the semiempirical creeping-flow estimate (\ref{eq:fperp4m}) for $F_\perp^{ \theta=90^\circ}(\cchi)$; dashed line: same with $F_\perp^{ \theta=90^\circ}$ based on the semiempirical finite-$\text{Re}$ estimate (\ref{eq:slender_perp_smallinertia_m}). In $(a)$, the two estimates overlap for $\cchi=3$; in $(b)$, the prediction for $\cchi=10$ based on (\ref{eq:fperp4m}) overlaps with that for $\cchi=5$ based on (\ref{eq:slender_perp_smallinertia_m})
}
\label{fig:c_perp_moderate}
\vspace{-2mm}
\end{figure}
Figure \ref{fig:c_perp_moderate} shows the variations of the perpendicular force component in the same low-to-moderate $\text{Re}$-range. At $\text{Re}=0.1$ (Fig. \ref{fig:c_perp_moderate}$(a)$), $F_\perp$ is accurately approximated for the two cylinders with $\cchi=3$ and $\cchi=5$ by the `Stokes law' based on the creeping-flow prediction (\ref{eq:fperp4m}) for $F_\perp^{ \theta=90^\circ}(\cchi)$.
This is no longer the case for $\cchi=10$, where a significant underestimate may be noticed. In this case, the length-based Reynolds number $\cchi\text{Re}$ is unity, implying that inertial effects are already significant. This is why the Stokes law based on the composite expression (\ref{eq:slender_perp_smallinertia_m}) for $F_\perp^{ \theta=90^\circ}$, which incorporates the inertial correction derived in \cite{khayat1989}, closely approaches the numerical data set. The validity of the Stokes law based on (\ref{eq:slender_perp_smallinertia_m}) is maintained for $\text{Re}=0.5$ whatever the aspect ratio.
For $\text{Re}=5$ (Fig. \ref{fig:c_perp_moderate}$(c)$), predictions of the same law are found to deviate significantly from numerical results, overestimating $F_\perp$ by more than $10\%$ for $\cchi=3$ and underestimating it by a similar percentage for $\cchi=10$. This is no surprise, since the inertial corrections involved in (\ref{eq:slender_perp_smallinertia_m}) are based on the finite-$\text{Re}$ theory of \cite{khayat1989} which assumes $\text{Re}\ll1$. In this respect, the deviations observed in Fig. \ref{fig:c_perp_moderate}$(c)$ may even be considered as surprisingly small.\vspace{1mm}\\
\indent Variations of the spanwise torque with $\cchi,\,\theta$ and $\text{Re}$ are displayed in Fig. \ref{fig:T_moderate}. As expected, the torque is negative, tending to orient the cylinder broadside on. Variations with the cylinder inclination closely follow the $\sin2\theta$-dependence predicted in the low-$\cchi\text{Re}$ limit by (\ref{KCt0}). The magnitude of the torque increases with the Reynolds number, in line with its inertial nature. For a given $\text{Re}$, the shorter the cylinder the stronger the normalized torque. Numerical results are compared with the full theoretical prediction of \citet{khayat1989}, namely
\begin{eqnarray}
\nonumber
\frac{T(\cchi,\theta,\text{Re}\ll1)}{\mu U(L/2)^2}&=&\frac{2\pi}{(\ln{\cchi})^2}\left\{\cos\theta\left[\mathcal{Z}(X)-\frac{E_1(X)+\ln X+\gamma}{X}+\mathcal{Z}(Y)-\frac{E_1(Y)+\ln Y+\gamma}{Y}\right]\right.\\
&& \bigg.\frac{}{}+\mathcal{Z}(Y)-\mathcal{Z}(X)\bigg\}\sin\theta+\mathcal{O}\left(\frac{1}{(\ln{\cchi})^3}\right)\,,
\label{KCcomplet}
\end{eqnarray}
where $\mathcal{Z}(X)=\frac{2}{X}\left(1+\frac{e^{-X}-1}{X}\right)$, $X=\frac{1}{2}\cchi\text{Re}(1-\cos\theta)$, $Y=\frac{1}{2}\cchi\text{Re}(1+\cos\theta)$ and $E_1$ and $\gamma$ as defined in (\ref{fpKC}). According to Fig. \ref{fig:T_moderate}, this prediction closely approaches the numerical results for the longest cylinder up to $\text{Re=0.5}$, together with those for the intermediate cylinder at $\text{Re}=0.1$. It is no surprise that the low-but-finite $\text{Re}$ theory is unable to provide a reasonable prediction for any of the three cylinders at $\text{Re}=5$. In their experiments, \citet{roy2019} considered cylindrical fibers with $\text{Re}\approx0.15$. With $\cchi=20$ they found the theory of \cite{khayat1989} to over-predict the torque by more than $20\%$ for $\cchi=20$ at $\theta=30^\circ$, and to slightly under-predict it (by $7-8\%$) for $\cchi=100$. Present results provide a stronger support to the theory, since the difference observed in Fig. \ref{fig:T_moderate}$(a)$ is only of the order of $5\%$ for $\cchi=10$ and clearly decreases with increasing $\cchi$. The fact that the asymptotic prediction, in which terms of $\mathcal{O}\left((\ln{\cchi})^{-3}\right)$ are neglected, correctly estimates the torque on a $\cchi=5$ cylinder at $\text{Re}=0.1$ but not at $\text{Re}=0.5$, while it still provides an accurate prediction at the same Reynolds number for $\cchi=10$ is noticeable. It suggests that the conditions $\text{Re}\ll1$ and $\cchi\gg1$ on which the asymptotic prediction is grounded must rather been understood as $\text{Re}\ll1$ and $\text{Re}/\cchi^2\ll1$. Indeed, the ratio $\text{Re}/\cchi^2$ stands below $0.005$ in all three configurations correctly predicted by (\ref{KCcomplet}), while it is beyond $0.01$ in all other cases.
For practical purposes, we sought an empirical fit of the torque valid for long enough cylinders and reducing to (\ref{KCt0}) (the limit form of (\ref{KCcomplet})) for $\cchi\text{Re}\rightarrow0$. As all three panels in Fig. \ref{fig:T_moderate} indicate, numerical data corresponding to $\cchi=10$ and $\cchi=5$ are accurately approached by the formula
\begin{figure}[h]
\centering
\includegraphics[width=5.5cm]{Tz_teta_xR0v1.pdf}\includegraphics[width=5.5cm]{Tz_teta_xR0v5.pdf}\includegraphics[width=5.5cm]{Tz_teta_xR5.pdf}
\vspace{-3mm}\\
\hspace{-35mm}$(a)$\hspace{47mm}$(b)$\hspace{47mm}$(c)$\\
\vspace{-15mm}
\hspace{-14mm}$\text{Re}=0.1$\hspace{42mm}$\text{Re}=0.5$\hspace{42mm}$\text{Re}=5$
\vspace{12mm}\\
\caption{Torque as a function of $\theta$ for $\bullet$: $\cchi=3$, $\star$: $\cchi=5$, $\blacktriangledown$: $\cchi=10$.
Solid line: asymptotic prediction (\ref{KCcomplet}) for low-but-finite $\text{Re}$; dashed line: approximate fit (\ref{KCJM}) for $\cchi=5$ and $10$
}
\label{fig:T_moderate}
\vspace{-2mm}
\end{figure}
\begin{equation}
\frac{T(\cchi\gg1,\theta,\text{Re})}{\mu U(L/2)^2}\approx-\frac{5\pi}{12}\frac{\text{Re}}{(1+\cchi\text{Re}^{1.1})^{0.5}}\left\{\frac{\cchi}{(\ln\cchi)^2}+\frac{13.5-30\,\text{Re}^{1/2}}{\cchi (\ln\cchi)^3}e^{-0.7\text{Re}}\right\}\sin2\theta\,.
\label{KCJM}
\end{equation}
The $\mathcal{O}\left(\cchi^{-1}(\ln{\cchi})^{-3}\right)$-term only provides a marginal contribution for $\cchi=10$. Consequently the leading-order $\mathcal{O}\left(\cchi(\ln{\cchi})^{-2}\right)$-truncation of (\ref{KCJM}) is sufficient to correctly estimate the torque on cylindrical fibers with $\cchi\gtrsim10$ up to $\text{Re}=5$. A correction proportional to $(\ln{\cchi})^{-4}$ may certainly be incorporated in (\ref{KCJM}) to properly estimate the torque on very short cylinders.
\section{Fully inertial stationary regime}
\label{incl2}
\color{black}
As pointed out in Sec. \ref{moder}, the flow past a cylinder aligned with the incoming velocity is stationary and axisymmetric within the full range of $\text{Re}$ investigated here.
Although the axial symmetry breaks down when $\theta$ is nonzero, the flow remains stationary up to \color{black} a critical Reynolds number, $\text{Re}_c(\cchi,\,\theta)$, larger than $300$. Details on the first unsteady regimes that take place beyond this threshold are provided as Supplemental Material. Here we concentrate on the stationary non-axisymmetric regime extending from Reynolds numbers of $\mathcal{O}(10)$ up to $\text{Re}_c$. \color{black}
To limit the computational cost, simulations in this regime were only carried out for cylinders with aspect ratios $\cchi=3,5$ and $7$. However, the results to be discussed hereinafter suggest that the flow structure and loads are only weakly affected by finite-length effects beyond $\cchi\approx5$ in this range of Reynolds number, although
the inertial load coefficients defined below may well continue to depend on $\cchi$, even for $\cchi\gg1$. Consequently, the physical phenomena involved and the simulation-based approximate expressions for the loads provided below are expected to apply without significant changes to cylinders with $\cchi\gg1$, which makes them relevant to analyze the motion of long cylindrical fibers in inertia-dominated regimes.\\
\indent \color{black} In Sec. \ref{moder}, we showed that, for Reynolds numbers of $\mathcal{O}(10^2)$, the flow structure past a cylinder aligned with the incoming velocity exhibits the presence of a thin annular standing eddy along the upstream part of the lateral surface. This feature was found to significantly influence the drag force, being able to change the sign of the viscous drag for sufficiently short cylinders and large Reynolds numbers. The situation is qualitatively similar in the $(\theta,\,\text{Re})$-range considered hereinafter since, for large enough inclinations and Reynolds numbers, the flow field exhibits specific features which directly affect the loads on the body. Consequently, it is relevant to examine first how the flow close to the body varies with the control parameters, which is the purpose of Sec. \ref{separ}. Then, the possibility to use numerical data to derive simple laws for the force components is considered in Sec. \ref{stokeslaw}, before empirical fits for the axial force and spanwise torque are built in Sec. \ref{approxpt}.\color{black}
\begin{figure}[H]
\centering
\raisebox{0.4in}{}\includegraphics[scale=0.05]{ldc_x5t5R140}\quad
\raisebox{0.4in}{}\includegraphics[scale=0.05]{ldcxz_x5t5R140}\\
\vspace{-18mm}\hspace{-3.6cm}{\small{$S^-$}}\\
\vspace{-6mm} \hspace{-1.2cm}{\small{$fs^-$}}\\
\vspace{15mm} \hspace{-4cm}{$(a)$}\hspace{7cm}{$(b)$}\\
\vspace{2mm}
\raisebox{0.4in}{}\includegraphics[scale=0.05]{ldc_x5t15R140}\quad
\raisebox{0.4in}{}\includegraphics[scale=0.05]{ldcxz_x5t15R140}\\
\vspace{-27.2mm}\hspace{-1.2cm}{\small{$fs^+$}}\\
\vspace{-3.7mm}\hspace{-9.8cm}{\small{$E$}}\\
\vspace{-1.5mm}\hspace{-10cm}{\small{$S^+$}}\hspace{0.8cm}{\small{A}}\\
\vspace{1mm}\hspace{-1.2cm}{\small{$fs^-$}}\\
\vspace{-9.2mm}\hspace{5cm}{$\mathcal{R}_b$}\\
\vspace{3.2mm}\hspace{5.2cm}{\small{$S^+$}}\hspace{0.9cm}{\small{A}}\\
\vspace{2.5mm}\hspace{5cm}{$\mathcal{R}_f$}\\
\vspace{8.5mm}
\hspace{-4cm}{$(c)$}\hspace{7cm}{$(d)$}\\
\vspace{2mm}
\raisebox{0.4in}{}\includegraphics[scale=0.05]{ldc_x5t30R140}\quad
\raisebox{0.4in}{}\includegraphics[scale=0.065]{ldcxz_x5t30R140}\\
\vspace{-36mm}\hspace{-5.5cm}{\small{$fs^+$}}\\
\vspace{3mm}\hspace{-9.8cm}{\small{$E$}}\\
\vspace{0.25mm}\hspace{-10cm}{\small{$S^+$}}\hspace{0.9cm}{\small{A}}\\
\vspace{-2.5mm}\hspace{-1.2cm}{\small{$fs^-$}}\\
\vspace{-7.5mm}\hspace{5cm}{$\mathcal{R}_b$}\\
\vspace{3.8mm}\hspace{5.5cm}{\small{$S^+$}}\hspace{1.2cm}{\small{A}}\\
\vspace{2.9mm}\hspace{5cm}{$\mathcal{R}_f$}\\
\vspace{9.4mm}
\hspace{-4cm}{$(e)$}\hspace{7cm}{$(f)$}\\
\vspace{2mm}
\vspace{0mm}
\caption{Streamlines in the symmetry plane $z=0$ (left) and in the plane $y=D/2$ tangent to the lateral surface along the top generatrix (right) for a cylinder with $\cchi=5$ at $\text{Re}=140$. Streamlines are colored according to the magnitude and sign of the axial velocity from $-0.25$ (deep blue) to $+1$ (deep red).
$(a)-(b)$: $\theta=5^\circ$; $(c)-(d)$: $\theta=15^\circ$; $(e)-(f)$: $\theta=30^\circ$.
\label{fig:ldc140}}
\end{figure}
\subsection{Flow structure}
\label{separ}
Throughout the regime under consideration, the flow field exhibits a mirror symmetry with respect to the $(x,y)$-plane which contains both the body axis and the incoming velocity.
Figure \ref{fig:ldc140} helps to understand how the flow structure past the cylinder varies with the inclination angle.
Several generic features emerge. First, the front stagnation point standing in the symmetry plane $z=0$ is seen to move downward (\textit{i.e.} toward negative $y$) as $\theta$ increases, almost reaching the bottom generatrix ($y=-D/2)$ for $\theta=30^\circ$ whatever $\text{Re}$. At the back of the body, the fluid that has passed over the upper part of the lateral surface is entrained downwards. It recirculates toward the downstream end within a region whose length in the streamwise $(x)$ direction is typically of the order of the cylinder radius for $\text{Re}=140$. Examination of this recirculating region at other Reynolds numbers (not shown) indicates that its length increases gradually with $\text{Re}$, becoming of the order of $D$ for $\text{Re}=300$. Unlike the standing eddy existing in the axisymmetric configuration (Sec. \ref{moder}), this recirculating zone is no longer a closed toroid. Indeed, once the fluid entrained downward gets very close to the lowest generatrix (point $S^-$ in Fig. \ref{fig:ldc140}$(a)$), it is expelled downstream in the main flow, just above the open streamline $fs^-$ emanating from $S^-$. Out of the symmetry plane, the three-dimensional streamlines displayed in Fig. \ref{fig:ldc3d} (for $\text{Re}=300$, to magnify the regions in which the fluid recirculates) reveal that fluid particles entrapped in the recirculating region describe successive loops before being sucked into the wake and advected downstream. This scenario is similar to that observed in \cite{johnson1999} at the back of a sphere in the first (steady) non-axisymmetric wake regime.
For $\text{Re}=140$ and $\theta=5^\circ$ (Fig. \ref{fig:ldc140}$(a)$), the flow along the body remains attached everywhere to the lateral surface. The lower (resp. higher) the Reynolds number, the larger (resp. smaller) the critical inclination at which separation occurs along the upper part of this surface. For instance, the flow is still unseparated at $\theta=15^\circ$ for $\text{Re}=40$ but is already separated at $\theta=5^\circ$ for $\text{Re}=300$. Separation starts at the intersection of the upstream end and the upper generatrix (point $S^+$ in Figs. \ref{fig:ldc140}$(c)-(f)$). Beyond the corresponding critical $\theta$ and/or $\text{Re}$, separation takes place over an open surface of finite extent, the most downstream point of which (point $A$ in Figs. \ref{fig:ldc140}$(c)-(f)$) stands on the symmetry plane. Still in the plane $y=D/2$, the separating line starting at $A$ develops upstream on the sides of the cylinder (regions $\mathcal{R}_b$ and $\mathcal{R}_f$ in Figs. \ref{fig:ldc140}$(d)$ and $(f)$) before joining the incoming flow. Not visible in Fig. \ref{fig:ldc140}, the lower part of the separation surface ($y<D/2$) on the sides of the cylinder exhibits a structure qualitatively similar to that of regions $\mathcal{R}_b$ and $\mathcal{R}_f$ and ends in some intermediate plane $0<y<D/2$, the position of which depends on $\theta$ and $\text{Re}$. Moving away from the lateral surface above the upper generatrix ($y>D/2$), the extent of the separation surface in the spanwise direction decreases gradually until its trace reduces to a single point some distance above the cylinder. This apex (point $E$ in Figs. \ref{fig:ldc140}$(c)$ and $(e)$) looks like the `eye of the storm' of the separated region. The larger $\theta$ and $\text{Re}$, the larger the distance between points $S^+$ and $E$ in both the $x$- and $y$-directions. Fluid is brought in the neighborhood of $E$ from both sides of the cylinder in a way similar to that observed in Figs. \ref{fig:ldc140}$(d)$ and $(f)$ in the plane $y=D/2$ (see Fig. \ref{fig:ldc3d}). Then it is sent back toward the upstream end in between the `eye' and the cylinder. The dividing streamline joining the region of the `eye' to point $A$ gets very close to the uppermost point $S^+$ of the upstream end. There, fluid particles are deviated by the `fresh' fluid flowing along the free streamline $fs^+$ and advected downstream, just below this free streamline. Streamlines that pass closer to to the `eye' stay further away from the cylinder surface within the recirculating region. Consequently they also stay further away from $fs^+$ once they escape this region, the corresponding fluid filling the intermediate region in between $fs^+$ and the cylinder at the back of the separation surface. Overall, the flow past the cylinder looks massively separated in between the free streamlines $fs^-$ and $fs^+$, the position of the `eye' governing the flow structure in the intermediate region. A similar open separation configuration was recently observed over inclined prolate spheroids in the range $5\leq \text{Re}\leq100$ in \cite{frohlich2020}.
\begin{figure}[h]
\centering
\raisebox{0.4in}{}\includegraphics[scale=0.07]{ldc3d_x5t15R300}
\caption{Three-dimensional streamlines past an inclined cylinder with $\cchi=5$, $\theta=15^\circ$ and $\text{Re}=300$.
\label{fig:ldc3d}}
\vspace{-3mm}
\end{figure}
\subsection{Are `Stokes laws' valid for inclined cylinders at moderate-to-large Reynolds number?}
\label{stokeslaw}
\color{black} We now make use of the numerical results to build approximate force and torque laws valid for $|\theta|\lesssim30^\circ$ and $\text{Re}\gtrsim10$ in the stationary non-axisymmetric flow regime. For this purpose, we characterize the force components through the parallel and perpendicular force coefficients defined \textit{via} the identify $(F_{\perp},F_{\parallel})\equiv(C_{\perp},C_{\parallel})\rho U^2LD/2$. Similarly, we introduce the torque coefficient $C_{t}$ related to the spanwise torque $T$ through $T\equiv C_{t}\rho U^2L^2D$. \color{black}
To assess the validity of the `Stokes laws' in this regime, we inject the numerical results in (\ref{eq:geom1})-(\ref{eq:geom2}) and compare the resulting $C_\parallel(\cchi,\theta,\text{Re})$ and $C_\perp(\cchi,\theta,\text{Re})$ with the predictions of (\ref{eq:Cpara_stokes})-(\ref{eq:Cperp_stokes}) at the relevant Reynolds number and aspect ratio. To achieve this comparison, the two coefficients $C_\parallel^{ \theta=0^\circ}$ and $C_\perp^{ \theta=90^\circ}$ are required for every value of $\cchi$ and $\text{Re}$. $C_\parallel^{ \theta=0^\circ}(\cchi,\text{Re})$ is directly related to the drag determined in Sec. \ref{aligned} up to a normalization factor. More specifically, in the $\text{Re}$-range considered in Sec. \ref{moder}, $C_\parallel$ and the drag coefficient $C_d$ resulting from (\ref{eq:cp_theta0_up})-(\ref{eq:cd_theta0}) are linked through the relation
$C_\parallel^{ \theta=0^\circ}(\cchi,\text{Re})=\frac{\pi}{4}\cchi^{-1}C_d(\cchi,\text{Re})$. At lower Reynolds number, $C_\parallel^{ \theta=0^\circ}$ is readily obtained through the approximate expression (\ref{eq:slender_smallinertia2}) as $C_\parallel^{ \theta=0^\circ}(\cchi,\text{Re})=4\pi \text{Re}^{-1}\mathcal{F}(\cchi,\text{Re})$, where $\mathcal{F}(\cchi,\text{Re})$ denotes the quantity within brackets in (\ref{eq:slender_smallinertia2}). A direct graphical estimate of $\mathcal{F}(\cchi,\text{Re})$ is provided for several aspects ratios in Fig. \ref{fig:small_inertia}, noting that the relation between $\mathcal{F}(\cchi,\text{Re})$ and the normalized drag $F_d/F_{ds}$ is $\mathcal{F}(\cchi,\text{Re})=\left(\frac{81}{16}\cchi^{-2}\right)^{1/3}F_d/F_{ds}$.\\
\indent Since we did not compute the loads on a cylinder held perpendicular to the flow, we use data from the literature to estimate $C_\perp^{ \theta=90^\circ}(\cchi,\text{Re})$.
A large number of experimental data obtained with long cylinders was compiled in \cite{pruppacher1970}. The corresponding curves for the drag per unit length were fitted in \cite{clift1978} in the form
\begin{equation}
C_\perp^{ \theta=90^\circ}(\cchi\gg1,\text{Re})\approx 9.69\text{Re}^{-0.78}(1+a\text{Re}^{n})\,,
\label{eq:clift}
\end{equation}
with $a=0.227$ and $n=0.55$ for $5<\text{Re}\leq40$, and $a=0.084$ and $n=0.82$ for $40<\text{Re}\leq400$, respectively.
Comparisons between predictions of (\ref{eq:clift}) and experimental results from \cite{jayaweera1965} for finite-length cylinders falling perpendicular to their axis indicate that the drag is only marginally affected by end effects as soon as $\cchi\gtrsim2$ and $\text{Re}\gtrsim10$ \citep{clift1978}. This is why we consider that (\ref{eq:clift}) provides a relevant estimate of $C_\perp^{ \theta=90^\circ}(\cchi,\text{Re})$ throughout the range of aspect ratios and Reynolds numbers of interest here. \vspace{2mm}\\
\indent The resulting comparison between numerical results and predictions of (\ref{eq:Cpara_stokes})-(\ref{eq:Cperp_stokes}) is presented in Fig. \ref{fig18}. The parallel force coefficients are seen to decrease significantly as $\cchi$ increases. In Sec. \ref{aligned}, where the drag was normalized using the frontal area $\pi D^2/4$, we found that the pressure contribution to the drag in the configuration $\theta=0^\circ$ is almost independent of $\cchi$. For this reason, the corresponding contribution to $C_\parallel^{ \theta=0^\circ}(\cchi,\text{Re})$, which involves a normalization by $LD$, behaves as $\cchi^{-1}$ and is responsible for the most part of the large variations of $C_{\parallel}(\cchi, \theta,\text{Re})$ with $\cchi$ observed in Figs. \ref{fig18}$(a), (c),(e)$. Variations of $C_{\parallel}$ with $\theta$ are remarkable in that they clearly contradict the `Stokes law'. Indeed, it is seen that $C_{\parallel}$ is almost independent of $\theta$ for $\text{Re}=20$ (apart from a modest decrease at $\theta=30^\circ$ for $\cchi=3$), while it increases with the cylinder inclination for larger Reynolds numbers. Still modest for $\text{Re}=80$, this increase makes $C_{\parallel}$ $40-50\%$ larger for $\theta=30^\circ$ than for $\theta=0^\circ$ at $\text{Re}=300$.
Clearly the `Stokes law' (\ref{eq:Cpara_stokes}) is unable to reproduce the observed trends.\\
\begin{figure}
\centering
\begin{tabular}{cccc}
\includegraphics[width=5cm]{cx_teta_xR20}&
\includegraphics[width=5cm]{cy_teta_xR20}&
&
\vspace{-3mm}
\\
\hspace{-30mm}$(a)$ & \hspace{-30mm}$(b)$ \\
\vspace{-18mm}\\
\hspace{28mm}$\text{Re}=20$&
\hspace{33mm}$\text{Re}=20$
\vspace{12mm}\\
\end{tabular}
\centering
\begin{tabular}{cccc}
\includegraphics[width=5cm]{cx_teta_xR80}&
\includegraphics[width=5cm]{cy_teta_xR80}&
&
\vspace{-3mm}
\\
\hspace{-30mm}$(c)$ & \hspace{-30mm}$(d)$ \\
\vspace{-18mm}\\
\hspace{28mm}$\text{Re}=80$&
\hspace{33mm}$\text{Re}=80$
\vspace{12mm}\\
\end{tabular}
\centering
\begin{tabular}{cccc}
\includegraphics[width=5cm]{cx_teta_xR300}&
\includegraphics[width=5cm]{cy_teta_xR300}&
&
\vspace{-3mm}
\\
\hspace{-30mm}$(e)$ & \hspace{-30mm}$(f)$ \\
\vspace{-18mm}\\
\hspace{28mm}$\text{Re}=300$&
\hspace{33mm}$\text{Re}=300$
\vspace{12mm}\\
\end{tabular}
\caption{ Parallel and perpendicular force coefficients ($C_{\parallel},C_{\perp}$ ) vs. $\theta$ for $\bullet$: $\cchi=3$, $\star$: $\cchi=5$, +: $\cchi=7$.
Dashed lines: `Stokes laws' (\ref{eq:Cpara_stokes})-(\ref{eq:Cperp_stokes}) based on the value of $C_\parallel^{ \theta=0^\circ}(\cchi=3)$ and $C_\perp^{ \theta=90^\circ}(\cchi\gg1)$ at the relevant Reynolds number; solid line: approximate fit (\ref{corrected}).
The yellow bullets in panels $(c)-(d)$ and $(e)-(f)$ refer to the results of \cite{pierson2019} for $\cchi=3$ at $\text{Re}=75$ and $250$, respectively.
\label{fig18}}
\end{figure}\indent
The situation is markedly different with the perpendicular force coefficient, which is found to follow closely the $\sin\theta$-dependence predicted by (\ref{eq:Cperp_stokes}) throughout the entire range of $\theta$ and $\text{Re}$. Moreover, $C_\perp$ only mildly varies with the aspect ratio, with less than $20\%$ differences between the shortest and longest cylinders whatever $\text{Re}$ (Figs. \ref{fig18}$(b), (d),(f)$). However, it is also clear from these panels that the predictions of (\ref{eq:Cperp_stokes}) based on the expression (\ref{eq:clift}) for $C_\perp^{ \theta=90^\circ}(\cchi\gg1,\text{Re})$ over-predict $C_\perp(\cchi,\theta,\text{Re})$ for the highest two Reynolds numbers, the overestimate increasing with $\text{Re}$. This is actually no surprise, since in the $\text{Re}$-range considered here, the flow structure past an inclined cylinder with $0^\circ\leq\theta\leq30^\circ$ has little in common with that past a cylinder held perpendicular to the incoming velocity. While the wake of the latter becomes unsteady for $\text{Re}\approx47$ and three-dimensional for $\text{Re}\approx180$ \citep{williamson1996}, all inclined configurations considered here are stationary and inherently three-dimensional. Therefore, the connection between the two configurations becomes loose beyond Reynolds numbers of a few tens. Nevertheless, to keep the advantage of (\ref{eq:Cperp_stokes}) which is asymptotically correct in the creeping-flow limit, we sought an empirical $\text{Re}$-dependent correction capable of properly approaching the numerical results while vanishing for both $\text{Re}\rightarrow0$ and $\theta=90^\circ$. \color{black} We sought another correction to account for the dependence of $C_\perp$ with respect to the aspect ratio, requesting that this correction also vanishes for $\theta=90^\circ$ for the aforementioned reasons. Ideally, one would like this correction to recover the proper behavior $C_\perp\propto\text{Re}^{-1}(\ln\cchi)^{-1}$ for $\text{Re}\rightarrow0$. However, due to the singular nature of the problem in the limit $\text{Re}\rightarrow0,\,\cchi\rightarrow\infty$ (Stokes paradox), it is known that $C_\perp^{ \theta=90^\circ}(\cchi\gg1,\text{Re}\rightarrow0)\propto-(\text{Re}\log\text{Re})^{-1}$ \cite{Lamb}, which makes the ratio $C_{\perp}/C_\perp^{ \theta\approx90^\circ}$ ill-defined in this limit. Consequently, we merely built the finite-length correction on the basis of present numerical data. We found this correction to be almost $\text{Re}$-independent, and eventually obtained the approximate expression for $C_{\perp}(\cchi,\theta,\text{Re})$ in the form
\begin{equation}
C_{\perp}(\cchi,\theta,\text{Re})=C_\perp^{ \theta\approx90^\circ}(\cchi\gg1,\text{Re})\frac{1+1.15e^{-0.45\cchi}\cos\theta}{1+0.04\text{Re}^{1/2}\cos\theta}\sin\theta\,.
\vspace{0mm}
\label{corrected}
\end{equation}
\color{black} Figures \ref{fig18}$(b)$, $(d)$ and $(f)$ show that the main trends of the numerical data are properly captured by the above fit. \color{black} Additional improvements could easily be introduced, such as a $\sin6\theta$-correction to compensate for the slight overestimates noticed at intermediate inclinations ($\theta\approx15^\circ$) as $\cchi$ and $\text{Re}$ increase. The finite-length correction suggests that the transverse force is virtually proportional to the cylinder length beyond $\cchi\approx10$. For shorter cylinders, the increase of $C_\perp$ as $\cchi$ decreases is qualitatively reminiscent of the $(\ln\cchi)^{-1}$ low-$\text{Re}$ behavior. However finite-length effects are much weaker in the inertial regime. For instance, Fig. \ref{fig:f_perp_stokes} indicates that $C_\perp^{ \theta=90^\circ}$ is $60\%$ larger for $\cchi=3$ than for $\cchi=10$ in the creeping-flow regime, a difference reduced to $25\%$ in the fully inertial regime according to (\ref{corrected}). \color{black} From (\ref{eq:clift}) (see also figure 3 in \cite{jayaweera1965}), it may be inferred that $C_\perp^{ \theta=90^\circ}(\cchi\gg1,\text{Re})$ only weakly decreases with the Reynolds number for $\text{Re}\gtrsim100$. Therefore, the fit (\ref{corrected}) indicates that $C_{\perp}(\cchi,\theta,\text{Re})$ is almost proportional to $\text{Re}^{-1/2}$ in this range of $\text{Re}$, suggesting that the dominant contribution to the perpendicular force arises from boundary layer effects. Although (\ref{corrected}) correctly reduces to $C_{\perp}(\cchi,\theta,\text{Re})=C_\perp^{ \theta=90^\circ}(\cchi\gg1,\text{Re})$ when $\theta=90^\circ$, it is not clear up to which maximum inclination this expression provides a reliable approximation of the actual transverse force. The numerical results of \cite{pierson2019} for $\cchi=3$ suggest that this maximum is close to $50^\circ$. Computations with higher inclinations are required to clarify this issue.\\
\indent Interestingly, \citet{sanjeevi2017} recently concluded that `Stokes laws', especially the sine-squared \color{black} drag law $C_{d}(\cchi,\theta)=C_\parallel^{ \theta=0^\circ}(\cchi) + (C_\perp^{ \theta=90^\circ}(\cchi)-C_\parallel^{ \theta=0^\circ}(\cchi))\sin^2\theta$ which results from the combination of (\ref{eq:geom1})-(\ref{eq:geom2}) and (\ref{eq:Cpara_stokes})-(\ref{eq:Cperp_stokes}), \color{black} hold for prolate spheroids (and moderately oblate spheroids) up to Reynolds numbers (based on the diameter of the equivalent sphere) of $\mathcal{O}(10^3)$. They argued that the reason for this surprising agreement is due to a partial compensation between contributions to the pressure drag arising from the regions close to the two stagnation points, in such a way that the overall pressure drag follows a sine-squared law, while the viscous contribution to the drag is almost insensitive to the inclination. Clearly this scenario does not hold for cylinders with flat ends. In the present case, wake effects are strong, with massive separation at the back of the cylinder, even for $\theta=0^\circ$, as soon as $\text{Re}$ exceeds some tens (see Fig. \ref{fig:streamlines}). These effects are deeply influenced by the body inclination (see Fig. \ref{fig:ldc140}) and, as Fig. \ref{fig18} reveals, result in non-monotonic variations of $C_{\parallel}(\cchi,\theta,\text{Re})$ with $\cchi$ and $\text{Re}$ which cannot be reduced to a simple geometric law. Therefore, it must be concluded that the scenario suggested in \cite{sanjeevi2017} to explain the validity of the sine-squared drag law applies only to streamlined bodies for which wake effects weakly affect the surface stress distribution.\\
\indent To better understand why the perpendicular force follows the approximate law (\ref{corrected}) throughout the parameter range explored here, it is useful to isolate the contributions to $C_\perp$ provided by the various parts of the body surface, and split each of them into a pressure and a viscous stress term. Since the body ends are flat and we are focusing here on the perpendicular force, no pressure contribution arises from the ends. Figure \ref{fig:details_Cperp} displays the variations of the remaining four nonzero terms with $\theta$ for two markedly different Reynolds numbers. The viscous contribution arising from the downstream end ($C_{\perp\mu_{down}}$) is seen to be negligibly small in all cases \color{black}(note the $10^3$ magnification factor in Fig. \ref{fig:details_Cperp}$(d)$). \color{black} Hence, virtually no contribution to $C_\perp$ is provided by this part of the body surface, on which wake effects concentrate in the range of $\theta$ and $\text{Re}$ relevant here. Examining panels $(a)-(c)$ in Fig. \ref{fig:details_Cperp} makes it clear that the various contributions to $C_\perp$ exhibit little dependence with respect to $\chi$, apart from the viscous stress on the upstream end ($C_{\perp\mu_{up}}$) at $\text{Re}=20$. Nevertheless, this term is one order of magnitude smaller than the total (pressure+viscous stress) contribution from the lateral surface. Consequently, the behavior of $C_\perp$ is essentially dictated by the latter. Among the corresponding two terms, the viscous contribution ($C_{\perp\mu_{lat}}$) is virtually independent of $\cchi$ at large $\text{Re}$, while some finite-length influence subsists in $C_{\perp p_{lat}}$. This weak dependence with respect to the aspect ratio implies that the perpendicular force increases almost linearly with the body length, given the chosen normalization factor $\rho U^2LD/2$.
The quasi-linearity of the two dominant contributions to $C_\perp$ with respect to the inclination angle and their weak $\cchi$-dependence for $\text{Re}\gg1$ imply that, for $\text{Re}\gtrsim100$, the dimensional transverse force behaves approximately as
\begin{equation}
F_\perp(\cchi,\theta,\text{Re})\approx\frac{C_{\perp \theta\approx90^\circ}(\cchi\gg1,\text{Re})}{1+0.04\text{Re}^{1/2}\cos\theta}\rho LD\frac{||{\bf{U}}||U_\perp}{2}\,,
\label{perp}
\end{equation}
\begin{figure}%
\centering
\includegraphics[width=0.32\textwidth]{cpl_teta_x.pdf}\label{fig:a}\hspace{8mm}
\includegraphics[width=0.32\textwidth]{ctl_teta_x.pdf}\label{fig:c}\\
\vspace{-3mm}
\hspace{-35mm} $(a)$\hspace{60mm}$(b)$\\
\vspace{3mm}
\hspace{-3mm}\includegraphics[width=0.335\textwidth]{cytw_teta_x.pdf}\label{fig:d}\hspace{7mm}%
\includegraphics[width=0.335\textwidth]{cyte_teta_x.pdf}\label{fig:e
\\
\vspace{-3mm}
\hspace{-35mm} $(c)$\hspace{60mm}$(d)$\\
\vspace{1mm}
\caption{ Contributions to $C_{\perp}(\cchi,\theta,\text{Re})$ at $\text{Re}=20$ (blue) and $\text{Re}=300$ (red) for $\bullet$: $\cchi=3$, $\star$: $\cchi=5$, +: $\cchi=7$. $(a)$ pressure on the lateral surface; $(b)$ viscous stress on the lateral surface; $(c),\,(d)$ viscous stress on the upstream and downstream ends, respectively.}%
\label{fig:details_Cperp}
\end{figure}\noindent
where $U_\perp=||\bf{U}||\sin\theta$ is the component of the upstream velocity normal to the lateral surface and the Reynolds number $\text{Re}$ is based on the norm of the upstream velocity.
In this Reynolds number range, $C_\perp^{ \theta=90^\circ}$ is almost constant according to (\ref{eq:clift}), so that (\ref{perp}) indicates that the perpendicular force is approximately proportional to the power three-half of the incoming velocity, while it still varies almost linearly with $\theta$. In contrast, the Independence Principle frequently invoked in the area of vortex-induced vibrations \citep{ramberg1983,zdravkovich1997b} suggests that the perpendicular force on a long cylinder should only depend on the normal component $U_\perp$ of the incoming flow, which would imply $F_\perp=\frac{1}{2}C_\perp^{ \theta=90^\circ}(\cchi\gg1,Re_\perp)\rho LD||{\bf{U}}_\perp||U_\perp$, with $\text{Re}_\perp=\rho||{\bf{U}}_\perp||D/\mu$. If this were true in the present situation, a $\sin^2\theta$-dependence of $C_\perp$ would be observed for $\text{Re}_\perp=\mathcal{O}(10^2)$ (\textit{i.e.} $\text{Re}=300$ here in practice), and the force would vary as the square of the incoming velocity. As Fig. \ref{fig18}$(f)$ indicates, no quadratic dependence with respect to the inclination angle is noticed, which implies that the Independence Principle does not apply to the flow configurations under consideration. This is in line with the conclusions of \cite{vakil2009} where it was observed at somewhat lower Reynolds numbers that this `principle' only holds for inclinations larger than $70^\circ$ but overestimates the force by more than $50\%$ for $\theta\leq30^\circ$, even for long cylinders with $\cchi=15$. In other words, this `principle' is approximately valid when the upstream flow is almost normal to the cylinder lateral surface but can by no means be used to approximate the transverse force when the cylinder inclination is moderate.
\subsection{Approximate laws for the parallel force and spanwise torque}
\label{approxpt}
Figure \ref{fig:details_Cparal} shows the main contributions to $C_\parallel(\cchi,\theta,\text{Re})$ arising from pressure and viscous stress distributions over the various parts of the cylinder surface. \color{black} On both ends, the latter (not shown) are found to be more than one order of magnitude smaller than the former. Therefore, $C_\parallel$ is essentially controlled by the viscous stress acting on the lateral surface and the pressure distribution on both ends. \color{black} All contributions are seen to decrease for increasing aspect ratios, similar to what happens when the body is aligned with the incoming flow. This influence weakens significantly as $\text{Re}$ increases, again in line with the observations reported for $\theta=0^\circ$. In particular, the viscous contribution arising from the lateral surface (Fig. \ref{fig:details_Cparal}$(c)$) is found to be virtually independent of $\cchi$ for $\text{Re}=300$. That these features subsist for all inclinations considered here suggests that seeking an empirical expression relating $C_\parallel(\cchi,\theta,\text{Re})$ to $C_{\parallel\theta=0^\circ}(\cchi,\text{Re})$ is reasonable. Variations with $\theta$ follow different and sometimes opposite trends on the various surfaces. For instance, Fig. \ref{fig:details_Cparal}$(a)$ indicates that $C_{\parallel p_{up}}$ decreases as the inclination increases, an effect weakening at large Reynolds number.
\begin{figure}%
\centering
\hspace{-10mm}\includegraphics[width=0.33\textwidth]{cpw_teta_x.pdf}\label{fig:aa}\hspace{-1mm}
\includegraphics[width=0.335\textwidth]{cpe_teta_x.pdf}\label{fig:bb} \hspace{-2mm}
\includegraphics[width=0.33\textwidth]{cmul_teta_x.pdf}\label{fig:ee}
\vspace{-2mm}\\
\hspace{-45mm} $(a)$\hspace{48mm}$(b)$\hspace{48mm} $(c)$\\
\caption{ Contributions to $C_{\parallel}(\cchi,\theta, \text{Re})$ at $\text{Re}=20$ (blue), $\text{Re}=80$ (green) and $\text{Re}=300$ (red) for $\bullet$: $\cchi=3$, $\star$: $\cchi=5$, +: $\cchi=7$. $(a)-(b)$ pressure on the upstream and downstream ends, respectively; $(c)$ viscous stress on the lateral surface.}%
\label{fig:details_Cparal}
\end{figure}
\noindent
In contrast, $C_{\parallel p_{down}}$ (Fig. \ref{fig:details_Cparal}$(b)$) increases gradually with $\theta$. \color{black} This is because the recirculating region at the back of the cylinder tends to shrink when $\theta$ increases, as noticed in Sec. \ref{separ}. \color{black} A significant increase of $C_{\parallel\mu_{lat}}$ is also noticed at $\text{Re}=300$ (Fig. \ref{fig:details_Cparal}$(c)$), but this trend weakens as $\text{Re}$ decreases and is no longer present at $\text{Re}=20$. \\%\color{black} Some of these trends directly stem from the influence of the inclination on the size of the separated regions described in Sec. \ref{separ}; \textit{e.g.} the increase of $C_{\parallel p_{down}}$ with $\theta$, which results from the reduction of the recirculating region at the back of the cylinder as $\theta$ increases. \color{black}
\indent To approximate the variations of $C_{\parallel}$ with $\cchi,\,\theta$ and $\text{Re}$, we started from the fits (\ref{eq:cp_theta0_up})-(\ref{eq:cmu_theta0}) established for $\theta=0^\circ$ in Sec. \ref{moder}. \color{black} We took into account the constraint that $C_\parallel$ cannot depend on the sign of $\theta$ and must change sign for $\theta=90^\circ$, although this configuration is well beyond the maximum inclination considered in the simulations. This led us to assume that the leading-order angular dependence of the correction to the `Stokes law' is proportional to $\sin^2\theta\cos\theta$. Then, we first considered the case $\cchi=7$ for which finite-length effects are the weakest, and started to fit the dependence with respect to $\text{Re}$ for $\theta=30^\circ$, the maximum inclination. The behavior of $C_{\parallel}$ at median inclinations suggests that the angular dependence also involves a secondary contribution that may be approached by a term proportional to $\sin^26\theta\cos\theta$.
Last, we considered finite-length effects, starting with $\cchi=3$ for which they are most severe. All empirical pre-factors were constrained to vanish for $\text{Re}\rightarrow0$, so that $C_\parallel$ reduces to its low-$\text{Re}$ form in this limit.
The whole process was carried out iteratively, to optimize the pre-factors and exponents for the three aspect ratios over the whole range of $\text{Re}$ and $\theta$. \color{black}
\color{black} Keeping in mind that the drag coefficient $C_d(\cchi,Re)$ determined in Sec. \ref{moder} by summing (\ref{eq:cp_theta0_up})-(\ref{eq:cmu_theta0}) has to be multiplied by a factor $\frac{\pi}{4}\cchi^{-1}$ to be used in the prediction of $C_\parallel$, the final fit takes the form
\begin{figure}
\centering
\hspace{-9mm}\includegraphics[width=0.32\textwidth]{law_cx_teta_xR20.pdf}\quad
\hspace{-3.5mm}\includegraphics[width=0.32\textwidth]{law_cx_teta_xR80.pdf}\quad
\hspace{-3.5mm}\includegraphics[width=0.32\textwidth]{law_cx_teta_xR300.pdf}\\
\vspace{-4mm}
\hspace{-20mm}$(a)$\hspace{50mm}$(b)$\hspace{50mm}$(c)$ \\
\vspace{-13mm}
\hspace{-18mm}$\text{Re}=20$\hspace{43mm}$\text{Re}=80$\hspace{43mm}$\text{Re}=300$
\vspace{10mm}\\
\caption{Parallel force coefficient $C_\parallel(\cchi,\theta,\text{Re})$ vs. $\theta$ for $\bullet$: $\cchi=3$, $\star$: $\cchi=5$, +: $\cchi=7$.
Solid line: empirical fit (\ref{eq:cparfin}).}
\label{fig:cparfit}
\end{figure}
\begin{equation}
\label{eq:cparfin}
C_{\parallel}(\cchi,\theta, \text{Re}) \approx \frac{\pi}{4}\cchi^{-1}C_d(\cchi,Re)\biggl\{1+\left[(0.7-6.3\cchi^{-2})(1-e^{-0.15\text{Re}})+0.01 \text{Re}^{0.95}\right]\sin^2\theta+2\times10^{-3} \text{Re}^{0.8}\sin^26\theta\biggr\}\cos\theta\,.
\end{equation} \color{black}
As Fig. \ref{fig:cparfit} shows, this fit provides a correct estimate of $C_\parallel$ throughout the range of parameters explored in the present investigation. \color{black} Expression (\ref{eq:cparfin}) highlights the fact that inertial effects act to increase $C_\parallel$ and counteract the $\cos\theta$-decrease associated with viscous effects, and even overtake them for high enough $\text{Re}$ (for large $\cchi$ and $\text{Re}\gtrsim10$, the dominant contribution to the term within curly brackets is $1+\left(0.7+0.01 \text{Re}^{0.95}\right)\sin^2\theta$). \color{black}
Obviously, the above fit is not expected to be valid for Reynolds numbers significantly larger than the upper bound considered in the simulations, as it predicts a diverging drag in the limit $\text{Re}\rightarrow\infty$. \color{black} Similarly, (\ref{eq:cparfin}) is not expected to hold for larger inclinations: in \cite{pierson2019} it was observed that, for $\cchi=3$ and $\text{Re}=250$, $C_\parallel$ sharply decreases in the range $30^\circ<\theta<45^\circ$, a trend that the above fit is clearly unable to reproduce. \color{black}
\vspace{2mm}\\
\begin{figure}%
\centering
\hspace{-8mm}\includegraphics[width=0.35\textwidth]{Ct_X_R20}\label{fig:a}%
\hspace{-1mm}\includegraphics[width=0.35\textwidth]{Ct_X_R80}\label{fig:b}
\hspace{-2mm}\includegraphics[width=0.35\textwidth]{Ct_X_R300}\label{fig:c}\\%\\
\vspace{-4mm}
\hspace{-40mm}$(a)$\hspace{54mm}$(b)$\hspace{54mm}$(c)$ \\
\vspace{-13mm}
\hspace{-16mm}$\text{Re}=20$\hspace{47mm}$\text{Re}=80$\hspace{47mm}$\text{Re}=300$
\vspace{11mm}\\
\caption{Torque coefficient vs. $\theta$ for $\bullet$: $\cchi=3$, $\star$: $\cchi=5$, + $\cchi=7$. The yellow bullets in panels $(b)$ and $(c)$ refer to the results of \cite{pierson2019} for $\cchi=3$ at $\text{Re}=75$ and $250$, respectively
}
\label{fig:torque_theta}
\end{figure}\indent
Figure \ref{fig:torque_theta} shows the variations of the torque coefficient as a function of $\theta$. \color{black} Similar to the low-to-moderate $\text{Re}$ regime, the torque is always negative, tending to orient the cylinder axis perpendicular to the upstream flow. \color{black}
The torque coefficient exhibits a quasilinear increase with the inclination angle for $\cchi=5$ and $7$. \color{black} This may be seen as a natural extension of the $\sin2\theta$-variation characterizing the $C_t$-variations at low $\text{Re}$.
$|C_t|$ is also found to decrease for increasing $\cchi$ at $\text{Re}=20$ and $\text{Re}=80$. The behavior observed at $\text{Re}=300$ is more complex, especially in the case of the shortest cylinder for which the angular dependence is strongly nonlinear for $\theta\lesssim15^\circ$. Moreover, while the corresponding $|C_t|$ is larger than those of the other two cylinders at $\text{Re}=20$ and $\text{Re}=80$, the situation is reversed at $\text{Re}=300$. \color{black}
Some additional insight into these variations may again be obtained by splitting the torque into pressure and viscous contributions provided by each part of the cylinder surface. \color{black} The main contributions resulting from this decomposition are displayed in Fig. \ref{fig:details_Ct}. The two dominant terms are seen to result from the pressure distribution on the lateral surface ($C_{{tp}_{lat}}$) and the viscous stress on the upstream end ($C_{{t\mu}_{up}}$). Both terms decrease in magnitude with increasing $\cchi$ and $\text{Re}$, keeping a negative sign in all cases.
The third and fourth contributors, $C_{{tp}_{up}}$ and $C_{{t\mu}_{lat}}$, result from the pressure and viscous stress distributions on the same surfaces. The contributions of the downstream end (not shown) are one order of magnitude smaller than the dominant terms in all cases.
In contrast to its viscous counterpart ($C_{{t\mu}_{up}}$), $C_{{tp}_{up}}$ is seen to keep a positive sign in all cases. The three contributions $C_{{tp}_{lat}}$, $C_{{tp}_{up}}$ and $C_{{t\mu}_{up}}$ vary almost linearly with $\theta$ whatever $\cchi$ and $\text{Re}$.
The viscous contribution associated with the lateral surface, $C_{{t\mu}_{lat}}$ (Fig. \ref{fig:details_Ct}$(d)$), reveals a more complex behavior. First, its sign changes with $\cchi$ and $\text{Re}$. It stays positive whatever $\theta$ for the shortest cylinder, increasing in magnitude as $\text{Re}$ increases. Conversely, it is negative for the longest two cylinders at $\text{Re}=20$, gradually decreasing until changing sign at all inclinations for $\text{Re}=300$. Second, variations of $C_{{t\mu}_{lat}}$ with $\theta$ become increasingly nonlinear as $\cchi$ decreases and/or $\text{Re}$ increases. \color{black}
This complex behavior is responsible for the markedly nonlinear variations of $C_t$ with $\theta$ noticed above for the shortest cylinder at $\text{Re}=300$.
The open separation process discussed in Sec. \ref{separ} is responsible for these features. Indeed, the dominant local viscous contribution to the spanwise torque provided by the the lateral surface is, in dimensional form, $-\mu\frac{D}{2}\frac{{\bf{r}}\cdot{\bf{e_y}}}{||{\bf{r}}||}({\bf{n}}\cdot\nabla) u_\parallel$, where $u_\parallel$ is the fluid velocity component parallel to this surface (along the $x$-direction), ${\bf{n}}$ is the unit normal directed into the fluid and ${\bf{r}}$ is the local position with respect to the cylinder geometrical center. As far as the fluid does not recirculate, this term is positive on the lower part of the surface (${\bf{r}}\cdot{\bf{e_y}}<0$) and negative on the upper part (${\bf{r}}\cdot{\bf{e_y}}>0$). However, when separation takes place, $u_\parallel$ is negative in the corresponding region of the upper part, which then provides a positive viscous contribution, making $C_{{t\mu}_{lat}}$ positive if the recirculation is strong enough. The larger the area percentage of the lateral surface corresponding to the separated region, the larger the positive value of $C_{{t\mu}_{lat}}$. \\
\indent
\begin{figure}%
\centering
\hspace{-10mm}\includegraphics[width=0.33\textwidth]{up_Ctp_X.pdf}\label{fig:aat}\hspace{0mm}
\includegraphics[width=0.33\textwidth]{lat_Ctp_X.pdf}\label{fig:cct}\\
\vspace{5mm}
\hspace{-6.5mm}\includegraphics[width=0.33\textwidth]{up_Ctmu_X.pdf}\label{fig:ddt}\hspace{-2mm}
\includegraphics[width=0.33\textwidth]{lat_Ctmu_X.pdf}\label{fig:fft}
\vspace{-46mm}\\
\hspace{-45mm} $(a)$\hspace{55mm}$(b)$
\vspace{38.3mm}\\
\hspace{-45mm} $(c)$\hspace{55mm}$(d)$\\
\vspace{1mm}
\caption{ Contributions to $C_t(\cchi,\theta,\text{Re})$ at $\text{Re}=20$ (blue), $\text{Re}=80$ (green) and $\text{Re}=300$ (red) for $\bullet$: $\cchi=3$, $\star$: $\cchi=5$, +: $\cchi=7$. $(a)-(b)$ pressure on the upstream end and the lateral surface, respectively; $(c)-(d)$ viscous stress on the upstream end and the lateral surface, respectively.}%
\label{fig:details_Ct}
\end{figure}
\color{black} Following a fitting procedure similar to that described for $C_\parallel$, we approached the behaviors observed in Fig. \ref{fig:torque_theta} with the empirical expression
\begin{equation}
C_{t}(\cchi,\theta, \text{Re})\approx\cchi^{-0.47}\left\{-0.69\text{Re}^{-0.35-b_1(\text{Re}/\cchi)^{3.1}}\sin2\theta+1\times10^{-4}\text{Re}^{0.8+b_2(\text{Re}/\cchi)^{3.1}}\sin6\theta\right\}\,,
\label{Ctt}
\end{equation}
with $b_1=7\times10^{-8}$ and $b_2=5\times10^{-8}$.\color{black}
\begin{figure}
\centering
\hspace{-9mm}\includegraphics[width=0.33\textwidth]{law_Ct_X_R20.pdf}\quad
\hspace{-3.5mm}\includegraphics[width=0.33\textwidth]{law_Ct_X_R80.pdf}\quad
\hspace{-3.5mm}\includegraphics[width=0.33\textwidth]{law_Ct_X_R300.pdf}\\
\vspace{-4mm}
\hspace{-40mm}$(a)$\hspace{52mm}$(b)$\hspace{52mm}$(c)$ \\
\vspace{-13mm}
\hspace{-16mm}$\text{Re}=20$\hspace{45mm}$\text{Re}=80$\hspace{44mm}$\text{Re}=300$
\vspace{10mm}\\
\caption{Torque coefficient $C_t(\cchi,\theta,\text{Re})$ vs. $\theta$ for $\bullet$: $\cchi=3$, $\star$: $\cchi=5$, +: $\cchi=7$.
Solid line: empirical fit (\ref{Ctt}).}
\label{fig:ctfit}
\end{figure}
As Fig. \ref{fig:ctfit} shows, the above fit provides a correct estimate of $C_t$ throughout the range of parameters explored in the present investigation. \color{black} For large enough aspect ratios, the dominant contribution to $C_t$ is still proportional to $\sin2\theta$, similar to the low-$\text{Re}$-regime, and its magnitude varies as $\cchi^{-0.47}Re^{-0.35}$.
The torque coefficient is seen to be approximately proportional to $\cchi^{-1/2}$ whatever $\text{Re}$, which suggests that the dimensional torque behaves roughly like $L^{3/2}$ for long enough cylinders. Finite-length high-$\text{Re}$ effects associated with the open separation are translated into the above fit through the slight but sharp $(\text{Re}/\cchi)^{3.1}$-increase of the Reynolds number exponent. \color{black}
Similar to the case of $C_\parallel$, the above fit is not expected to be valid for Reynolds numbers significantly larger than the upper bound considered in the simulations. \color{black} Moreover, although (\ref{Ctt}) respects the constraint that the torque changes sign for $\theta=90^\circ$, the results of \cite{pierson2019} for a cylinder with $\cchi=3$ indicate that the behavior of $C_t$ changes significantly for $\theta\gtrsim45^\circ$, suggesting that the present fit is appropriate only below this critical inclination. \color{black}
\section{Summary and concluding remarks}
\label{Summ}
With the practical objective of providing approximate laws for predicting the translation-induced drag, lift and torque acting on long cylindrical rods and fibers, we employed fully-resolved simulations to investigate the flow around a finite-length circular cylinder held fixed in a uniform stream making some angle with the body axis. \\
\indent We first focused on the specific case where the cylinder is aligned with the incoming flow. Considering the Stokes regime and the weakly inertial regime corresponding to $\text{Re}\lesssim1$, we combined numerical results with available predictions from the slender-body theory (which we slightly improved by computing the next-order term in the expansion with respect to the small parameter $1/\ln(2\cchi)$) to build the approximate drag laws (\ref{eq:slender_modified}) and (\ref{eq:slender_smallinertia2}). The former is valid down to $\cchi\approx1.5$, while the inertial corrections contained in the latter allow an accurate estimate of the drag up to $\text{Re}=\mathcal{O}(1)$, and even up to $\text{Re}=\mathcal{O}(10)$ for $\cchi\gtrsim20$.
For larger Reynolds numbers (up to $\text{Re}=400$), the flow structure becomes more complex, although it remains stationary and axisymmetric.
Beyond a $\cchi$-dependent critical Reynolds number of the order of $200$, a second recirculating region emerges along the upstream part of the lateral surface.
Being associated with local negative shear stresses, this lateral eddy acts to reduce the friction drag, which may even become negative if $\text{Re}$ is large enough and the cylinder is short enough. We used the numerical data to build approximate fits for this friction drag and for the pressure drag contribution of the upstream and downstream ends.
With this procedure, we obtained the empirical drag law (\ref{eq:cd_theta0}) which approximates the drag well for $\cchi\gtrsim2$ in the range $20\leq \text{Re}\leq400$ and properly matches (\ref{eq:slender_smallinertia2}) for $\text{Re}=\mathcal{O}(1)$. The friction drag still represents a substantial part of the total drag at Reynolds numbers of several hundreds if the body aspect ratio is large enough ($45\%$ at $\text{Re}=400$ for a cylinder with $\cchi=10$).\\
\indent \color{black}
In the next step, we examined the case of moderately inclined cylinders ($\theta\leq30^\circ$) in the low-to-moderate Reynolds number regime ($\text{Re}\leq5$). For Reynolds numbers less than unity and aspect ratios up to $10$, we observed that the force component parallel to the cylinder axis closely follows the $\cos\theta$-variation predicted under creeping-flow conditions. The agreement deteriorates as the length-based Reynolds number $\cchi\text{Re}$ exceeds values of $\mathcal{O}(10)$. Under more inertial conditions, the $\cos\theta$-law under-predicts the actual parallel force, the difference increasing with both $\cchi\text{Re}$ and $\theta$. The force component perpendicular to the cylinder axis was found to closely follow the $\sin\theta$-variation typical of creeping-flow conditions up to $\text{Re}=5$, irrespective of $\cchi$. However, the corresponding pre-factor deviates from the creeping-flow prediction as soon as $\cchi\text{Re}\gtrsim0.5$, beyond which inertial effects become significant. Accurate predictions are obtained up to $\cchi\text{Re}\approx10$ by estimating the pre-factor of the $\sin\theta$-law through the semi-empirical formula (\ref{eq:slender_perp_smallinertia_m}) which provides a finite-Reynolds-number approximation of $F_\perp^{\theta=90^\circ}$. Throughout the low-to-moderate Reynolds number range, the inertial torque follows the $\sin2\theta$-variation predicted by the asymptotic theory of \citet{khayat1989} in the limit $\cchi\text{Re}\ll1$. However the magnitude of the torque is correctly predicted by this theory only up to $\text{Re}=\mathcal{O}(1)$ and provided $\text{Re}/\cchi^2\lesssim0.01$. To obtain a correct estimate of the torque over a broader range of conditions, we derived the semiempirical law (\ref{KCJM}) which correctly reduces to the theoretical prediction in the limit $\cchi\text{Re}\ll1, \cchi\gg1$ and closely approaches numerical data for cylinders with $\cchi\geq5$ up to $\text{Re}=5$.\color{black}\\
\indent Last we considered the three-dimensional flow past moderately inclined cylinders with aspect ratios in the range $3-7$ for $20\leq\text{Re}\leq300$.
For $\theta\leq30^\circ$, the flow remains stationary irrespective of the inclination and preserves a symmetry with respect to the plane containing the body axis and the incoming velocity.
For sufficiently low inclinations and Reynolds numbers, the flow separates only at the back of the body, the recirculating region looking like an open toroid.
In contrast, beyond a critical inclination decreasing as $\text{Re}$ increases, an open separated region emerges on the `extrados' of the lateral surface, near its `leading edge'.
In such configurations, the flow past the cylinder looks massively separated in between the two free streamlines emanating from its `trailing' and `leading' edges. \\%, \textit{i.e.} the intersection of its downstream end and its lower generatrix, and that of its upstream end and its upper generatrix, respectively. \\
\indent We used numerical data collected in this fully inertial regime to obtain approximate laws for the loads acting on the cylinder. Similar to the the low-$\text{Re}$ behavior, the perpendicular force obeys essentially a $\sin\theta$-variation with a mild dependence with respect to the aspect ratio. We found that the corresponding force coefficient $C_\perp$ may be related to the drag coefficient of a cylinder held perpendicular to the incoming flow through the emprical law (\ref{corrected}) \color{black} which involves two simple independent corrections, one proportional to $\text{Re}^{-1/2}$ accounting for inertial effects, the other for finite-length effects.\color{black
\begin{figure}
\centering
\hspace{-9mm}\includegraphics[width=0.33\textwidth]{cx_theta_R5.pdf}\quad
\hspace{-3.5mm}\includegraphics[width=0.33\textwidth]{cy_theta_R5.pdf}\quad
\hspace{-3.5mm}\includegraphics[width=0.33\textwidth]{tz_theta_R5.pdf}\\
\vspace{-4mm}
\hspace{-40mm}$(a)$\hspace{52mm}$(b)$\hspace{54mm}$(c)$ \\
\caption{Extrapolated predictions at $\text{Re}=5$ of fits derived in the fully inertial regime for $\bullet$: $\cchi=3$, $\star$: $\cchi=5$, +: $\cchi=7$. $(a)$ $C_\parallel$ and fit (\ref{eq:cparfin}); $(b)$ $C_\perp$ and fit (\ref{corrected}); $(c)$ $C_t$ and fit (\ref{Ctt}).}
\label{fig:fitsRe5}
\end{figure}
In contrast, variations of the parallel force do not follow the $\cos\theta$-law prevailing in the low-$\text{Re}$ regime. Instead, the force coefficient $C_\parallel$ barely varies with $\theta$ in the moderate-$\text{Re}$ regime and even increases with the inclination at large Reynolds number. A fitting procedure with respect to the three control parameters allowed us to mimic the influence of inertial and finite-length effects on $C_\parallel$ through the empirical law (\ref{eq:cparfin}) which reproduces all observed trends well. Variations of the spanwise torque at moderate Reynolds number are qualitatively similar to those observed for $\text{Re}\leq5$. However strong finite-length effects manifest themselves at high Reynolds number, in connection with the separation process affecting the upstream part of the lateral surface in this regime. A fitting approach similar to that employed for $C_\parallel$ yielded the empirical law (\ref{Ctt}) which correctly approximates $C_t$ throughout the explored range of $\cchi$ and $\theta$ in the fully inertial regime. \color{black} For application purposes, it is of course of interest to know how robust these fits determined from data in the range $20\leq\text{Re}\leq300$ are when the Reynolds number is decreased to the upper limit of the low-to-moderate $\text{Re}$ regime considered in Sec. \ref{incl1}. The results or this test are summarized in Fig. \ref{fig:fitsRe5}. It turns out that the empirical formula established in the fully inertial regime still perform quite well for $\text{Re}=5$. Predictions depart from numerical data by less than $5\%$ for $C_\parallel$, $10\%$ for $C_\perp$ and $15\%$ for $C_t$. Therefore, the semi-empirical predictions derived in Sec. \ref{incl1} combined with the fits established in Sec. \ref{incl2} offer a complete and almost smooth description of load variations from the creeping-flow regime up to $\text{Re}=300$. \color{black}\\
\indent Most computations only considered aspect ratios below $10$ or even $7$ for $\text{Re}\geq20$, owing to the rapid increase of computational costs with $\cchi$. However, finite-length effects were found to decrease monotonically and sharply as $\cchi$ increases, making us confident that the various empirical laws derived in the course of this study remain valid for cylindrical particles with larger aspect ratios, and therefore apply to long fibers. Obviously this does not mean that the load coefficients become independent of $\cchi$ for $\cchi\gg1$, but simply that their asymptotic dependence with respect to $\cchi$ in the limit of large aspect ratios is already captured by considering $\mathcal{O}(10)$-aspect ratios as we did. The situation is less clear regarding their range of validity with respect to the inclination angle. All of them were calibrated in the range $|\theta|\leq30^\circ$ and satisfy the required geometrical constraints for $|\theta|=90^\circ$ and the associated symmetry conditions. Nevertheless, as soon as the Reynolds number exceeds a few tens, the dynamics of the flow past a cylinder in the configuration $\theta=90^\circ$ drastically differs from that past the same cylinder for $\theta=0^\circ$. Hence, physical features that are not present in the low-to-moderate inclination range considered here take place in the near-body flow when the inclination exceeds $45^\circ$ or so, which is likely to make the empirical laws proposed here invalid for such large inclinations.\vspace{2mm}\\
\indent The present investigation leaves several important configurations and parameter ranges unexplored. First, for Reynolds numbers similar to those considered in Secs. \ref{incl1} and \ref{incl2}, the above discussion calls for a specific study focused on large inclinations, say $45^\circ\lesssim|\theta|\leq90 ^\circ$, for which the flow past the cylinder is expected to be massively separated and most of the time unsteady.
Another series of questions arises when the cylinder is allowed to rotate about an axis perpendicular to its symmetry axis and passing through its geometrical center, as rodlike particles and fibers customarily do. The torque on a slender rotating cylinder was predicted in the creeping flow limit in \cite{batchelor1970,cox1970} but no theoretical attempt to derive inertial corrections in this configuration has been reported so far. This is even more true for the general situation in which the cylinder undergoes both a translation and a rotation. In such a case, inertial effects couple the two types of motion, yielding specific contributions to the loads, which are for instance responsible for the well-known Magnus effect on a spinning sphere. To the best of our knowledge, such couplings have not been considered for slender cylinders. We are currently investigating numerically the configuration in which the cylinder undergoes an imposed rotation, and plan to apply the same methodology to the combined translation+rotation case in the near future.
\section*{Acknowledgments}
M. K.'s fellowship was provided by IFP Energies Nouvelles whose financial support is greatly appreciated. The authors thank Annaïg Pedrono for her continuous support with the use of the JADIM code. Part of the computations were carried out on the national supercomputers operated by the GENCI organization under allocation A0072B10978.
|
\section{Introduction}
\begin{quote}
First children are born,\\
later they perform miracles.
unknown Author
\end{quote}
A couple of years ago the relationsip between the Beta function
and the Euler gamma function
\begin{equation*}
\mathcal{B}(x,y)=\frac{\Gamma(x) \Gamma(y)}{\Gamma(x+y)}
\end{equation*}
gave rise to introduce a class called beta-type functions,
which are functions of the form
\begin{equation*}
{B}_f(x,y)=\frac{f(x) f(y)}{f(x+y)},
\end{equation*}
where $f$ is a positive function defined on a suitable real interval.
In this connection it was natural to ask for means in the class of beta-type functions,
which were determined in \cite{MatHim2} and \cite{MatHim4}.
The form of beta-type functions suggested very strongly to introduce three more classes of functions,
each of them related to one of the four Cauchy functional equations (cf. \cite{Kuczma2}).\\
It turned out that in three of these classes, namely in beta-type functions (also called exponential Cauchy quotients), in logarithmic and in multiplicative Cauchy quotients there exists a unique mean of $k$ variables where $k$ is an arbitrary positive integer (cf. \cite{MatHim5}),
but in the class of additive Cauchy quotients there is no mean for any number of variables greater or equal than two.
The classes of means which appeared naturally in this area of research even led to unexpected applications in stochastic analysis (cf. \cite{Burai}).
In this paper we take a closer look to the integral representation of the Beta function.
In a nutshell, we observe
that the integrand of the Beta function consists basically of two smooth exponential functions
where the bases satisfy a certain duality relation.
To avoid any misunderstanding, we go into medias res.
The Euler Beta function $\mathcal{B}:\left(
0,+\infty \right) ^{2}\rightarrow \left( 0,+\infty \right)$ is defined by
\begin{equation}
\mathcal{B}\left( x,y\right) =
\int_0^1 {t^{x-1} (1-t)^{y-1}\, dt},\RIfM@\expandafter\text@\else\expandafter\mbox\fi{ \ \ \ \ \ }x,y \in (0, +\infty).
\label{eq:Beta}
\end{equation}%
Intuitively, the ingredients of the Beta function are two smooth families of exponential functions of base $t$ and $1-t$, respectively, which are 'averaged together' in the sense that they are multiplied and then integrated over the unit interval.
More precisely, for fixed $t \in [0,1]$,
the exponential function ${\mathbb{R}} \ni x \mapsto t^{x-1} \in (0,+\infty)$ is multiplied by its dual exponential function
${\mathbb{R}} \ni y \mapsto (1-t)^{y-1} \in (0,+\infty)$ (\textit{dual} in the sense that their respective bases add to one)
and integrated over the interval $[0,1]$.
In a similar manner we introduce in this paper about three more Beta functions, each of them related to one of the other Cauchy functional equations.
\section{Multiplicative Beta function}
Since the ingredients of the Beta function, i.e. the functions under the integral in \eqref{eq:Beta}, are smooth exponential functions,
it seems natural to consider the functions
$(1, +\infty) \ni x \mapsto (x-1)^t \in (0, +\infty)$ and $(1, +\infty) \ni y \mapsto (y-1)^{1-t} \in (0, +\infty)$,
two families of smooth multiplicative functions dual to each other in the sense that the sum of their exponents is one,
as the ingredients of the multiplicative Beta function
$\mathcal{M}:\left(1,+\infty \right) ^{2}\rightarrow {\mathbb{R}}$
defined by
\begin{equation}
\mathcal{M}\left( x,y\right)=
\int_0^1 {(x-1)^t (y-1)^{1-t}\, dt}, \RIfM@\expandafter\text@\else\expandafter\mbox\fi{ \ \ \ \ \ }x,y \in (1, +\infty).
\label{eq:Beta_m:def}
\end{equation}
Unlike to the case of the Euler Beta function, which we should call exponential Beta function from a more systematic point of view,
the multiplicative Beta function $\mathcal{M}$ can be expressed by elementary functions, since, for all $x,y \in (1, +\infty), x \neq y$, we have
\begin{eqnarray}
\label{eq:B_m}
\mathcal{M}\left( x,y\right) & = & (y-1) \int_0^1 {\left(\frac{x-1}{y-1}\right)^t\, dt}\\ \nonumber
& = & (y-1) \left[ \frac{\left(\frac{x-1}{y-1}\right)^t}{\log{\left(\frac{x-1}{y-1}\right)}} \right]_0^1 \\ \nonumber
& = &\frac{x-y}{ \log{\left(\frac{x-1}{y-1}\right)}},
\end{eqnarray}
and $\mathcal{M}\left( x,x\right) = x-1$ for $x>1$ implying that $\mathcal{M}$ is the logarithmic mean where the arguments are both reduced by $1$.
\section{Additive Beta function}
The building blocks of the additive Beta function are
the two families
${\mathbb{R}} \ni x \mapsto t (x-1) \in {\mathbb{R}}$ and ${\mathbb{R}} \ni y \mapsto ({1-t})(y-1) \in {\mathbb{R}}$.
As before, we could argue that these $t$-families of continuous additive functions should be multiplied and then integrated to obtain
the additive Beta function
$\mathcal{A}_{1}: {\mathbb{R}}^2 \to {\mathbb{R}}$ defined by
\begin{equation*}
\mathcal{A}_{1} \left( x,y\right):=\int_0^1 {t (x-1) \cdot ({1-t})(y-1) \, dt},
\end{equation*}
which, for all $x,y \in{\mathbb{R}}$, simplifies to
\begin{eqnarray*}
\mathcal{A}_{1} \left( x,y\right)
&=&(x-1)(y-1) \int_0^1 {t(1-t)\, dt}\\
&=&(x-1)(y-1) {\left[ \frac{1}{2} t^2-\frac{1}{3} t^3\right]_0^1}\\
&=&\frac{1}{6} (x-1)(y-1).
\end{eqnarray*}
But one could object that in the construction of the preceding Beta functions the motivation to multiply the building blocks,
i.e. the two families of exponential functions (multiplicative functions) in the case of
the Euler Beta function (in the case of multiplicative Beta function, respectively) stems from the fact that
in the exponential Cauchy equation
\begin{equation*}
f(x+y)=f(x) \cdot f(y), \qquad x,y \in {\mathbb{R}},
\end{equation*}
(in the case of the multiplicative Cauchy equation $f(xy)=f(x)f(y)$ for all $x,y \in (0, +\infty)$, respectively)
multiplication of the function $f$ occurs on the right hand side.
If this is the right explanation
why the respective ingredients of exponential and multiplicative Beta functions are multiplied under the integral sign,
we should rather add the building blocks in the case of the additive Beta function,
since for an additive function we have
\begin{equation*}
g(x+y)=g(x) + g(y), \quad x,y \in {\mathbb{R}},
\end{equation*}
and addition occurs on the right hand side.
Thus, we may rather consider $\mathcal{A}_{2}: {\mathbb{R}}^2 \to {\mathbb{R}}$ defined by
\begin{equation}
\mathcal{A}_{2} \left( x,y\right)=\int_0^1 {\left[t (x-1) + ({1-t})(y-1)\right] \, dt},
\end{equation}
as the natural additive counterpart of the Euler Beta function.
Surprisingly, the function $\mathcal{A}_{2}$ coincides with the arithmetic mean,
since, for all $x,y \in {\mathbb{R}}$, it holds
\begin{eqnarray*}
\mathcal{A}_{2}\left( x,y\right)&=& \left[{\frac{t^2}{2} (x-1) + \left({t-\frac{t^2}{2}}\right)(y-1)}\right]_0^1\\
&=&{\frac{1^2}{2} (x-1) + \left({1-\frac{1^2}{2}}\right)(y-1)}\\
&=&\frac{1}{2} (x-1) + \frac{1}{2} (y-1)\\
&=&A(x-1,y-1),
\end{eqnarray*}
where $A:{\mathbb{R}}^2 \to {\mathbb{R}}$, defined by $A(x,y)=\frac{x+y}{2}$ for all $x,y \in {\mathbb{R}}$, is the arithmetic mean.
\section{Logarithmic Beta function}
Similarly as before, we consider the $t$-families $x \mapsto t \log{x}$ and $y \mapsto (1-t)\log{y}$ as the building blocks
of the logarithmic Beta function $\mathcal{L}_{1}: (1,+\infty) \to {\mathbb{R}}$ defined by
\begin{equation}
\mathcal{L}_{1} \left( x,y\right):=\int_0^1 { t \log{(x-1)} \cdot (1-t)\log{(y-1)} \, dt}.
\label{eq:B_l1}
\end{equation}
But since a logarithmic function satisfies the functional equation
\begin{equation*}
l(x y)= l(x) + l(y),\qquad x,y \in (0,+\infty),
\end{equation*}
and addition is the operation between $l(x)$ and $l(y)$ on the right hand side, arguing similarly when introducing the second additive pendant of the Beta function $\mathcal{A}_{2}$,
we should rather consider $\mathcal{L}_{2}: (1,+\infty) \to {\mathbb{R}}$ defined by
\begin{equation}
\mathcal{L}_{2} \left( x,y\right):=\int_0^1 { \left[t \log{(x-1)} + (1-t)\log{(y-1)} \right] \, dt}.
\label{eq:B_l2}
\end{equation}
as the logarithmic counterpart of the Euler Beta function.
It is not completely clear why in all these Beta functions the arguments are reduced by one, i.e. why $x-1$ and $y-1$, respectively, appear in the definiens of the Beta functions rather than $x$ and $y$ itself.
A historic explanation could be that $\Gamma(n)=(n-1)!$ for a natural number $n$.
Here as well we do not see any obvious reason
why the function $\Pi:(-1,+\infty) \to (0,+\infty)$ with $\Pi(x):=\Gamma{(x+1)}$ for $x>-1$, hence $\Pi(n)=n!$ for $n\in{\mathbb{N}}$,
due to Gau{\ss} was not preferred as the natural extension of the sequence $(n!)_{n\in{\mathbb{N}}}$ by the majority of mathematicians%
\footnote{The domain of the function $\Pi$ is not symmetric with respect to zero, which one may consider as not so aesthetic. Moreover, nowadays the symbol $\Pi$ is usually avoided to denote functions since it is mostly used as a symbol for products. }.
The two logarithmic Beta functions can be expressed by elementary functions since, for all $x,y \in (1, +\infty)$,
\begin{eqnarray*}
\mathcal{L}_{1} \left( x,y\right) &=& \log{(x-1) \log{(y-1) \int_0^1 {t } (1-t)} \, dt}\\
&=& \log{(x-1)} \log{(y-1)} \int_0^1 {{(t-t^2)} \, dt}\\
&=& \log{(x-1)} \log{(y-1)} {\left[\frac{t^2}{2}-\frac{t^3}{3}\right]}_0^1\\
&=& \frac{1}{6}\log{(x-1)} \log{(y-1)}.
\end{eqnarray*}
Unsurprisingly when noticing the case of the function $\mathcal{A}_{2}$, we get
\begin{eqnarray*}
\mathcal{L}_{2}\left( x,y\right)&=& \left[{\frac{t^2}{2} \log{(x-1)} + \left({t-\frac{t^2}{2}}\right)\log{(y-1)}}\right]_0^1\\
&=&{\frac{1^2}{2} \log{(x-1)} + \left({1-\frac{1^2}{2}}\right)\log{(y-1)}}\\
&=&\frac{1}{2} \log{(x-1)} + \frac{1}{2} \log{(y-1)}\\
&=&A(\log{(x-1)},\log{(y-1)}).
\end{eqnarray*}
\section{Beta functions and their respective class of Cauchy quotients}
It is natural to ask whether these newly introduced Beta functions belong to the respective class of Cauchy quotients.
\subsection{The Euler Beta function - a beta-type function}
It is known that in the case of the Euler Beta function the answer is positive due to the relation between the Euler Beta and the Gamma function
\begin{equation}
\mathcal{B}\left( x,y\right) =
\frac{\Gamma(x) \Gamma(y)}{\Gamma(x+y)},\RIfM@\expandafter\text@\else\expandafter\mbox\fi{ \ \ \ \ \ }x,y \in (0, +\infty),
\label{eq:Beta_Gamma}
\end{equation}%
which means that Euler Beta function $\mathcal{B}$ is of the form
\begin{equation*}
\frac{f(x) f(y)}{f(x+y)}
\end{equation*}
for $f=\Gamma$ and thus belongs to the class of beta-type functions (cf. \cite{MatHim2}).
\subsection{The multiplicative Beta function - a multiplicative Cauchy quotient?}
The question whether the multiplicative Beta function $\mathcal{M}$ belongs to the class of multiplicative Cauchy quotients
can be reformulated as: is there a function $f:(1,+\infty) \to (0,+\infty)$ such that, for all $x,y \in (1, +\infty)$,
\begin{equation*}
\mathcal{M} \left( x,y\right)=\frac{f(x) f(y)}{f(xy)}, \qquad x \neq y,
\end{equation*}
and \begin{equation*}
x-1=\frac{(f(x))^2}{f(x^2)}
\end{equation*}
holds true?
\iffalse
From the last equation, since $f$ is a positive function, it follows that
\begin{equation*}
f(x)=\sqrt{(x-1)f(x^2)}, \qquad x \in (1, +\infty).
\end{equation*}
\fi
By \eqref{eq:B_m}, the multiplicative Beta function $\mathcal{M}$ coincides with the logarithmic mean (cf. \cite{Stolarsky}, \cite{Stolarsky2}), which is a Cauchy mean (cf. \cite{Matkowski}).
Hence, for all $x,y \in (1, +\infty)$,
\begin{equation*}
\frac{x-y}{\log{\left(\frac{x-1}{y-1}\right)}}=\frac{f(x) f(y)}{f(xy)}, \qquad x \neq y,
\end{equation*}
and \begin{equation*}
x-1=\frac{(f(x))^2}{f(x^2)}.
\end{equation*}
\subsection{The additive Beta functions - additive Cauchy quotients?}
The additive Beta function $\mathcal{A}_1$ (respectively $\mathcal{A}_2$) is an additive Cauchy quotient
if there is an interval $I \subset {\mathbb{R}}$ which is closed under addition and a function $f: I \to (0, +\infty)$ such that, for all $x,y \in I$,
\begin{equation*}
\frac{1}{6}(x-1)(y-1)=\frac{f(x) + f(y)}{f(x+y)},
\end{equation*}
(respectively, in case of $\mathcal{A}_2$,
\begin{equation*}
\frac{(x-1)+(y-1)}{2}=\frac{f(x) + f(y)}{f(x+y)},
\end{equation*}
)
holds true.
\iffalse
Setting $x=y$, it follows
\begin{equation*}
f(x)=\frac{1}{12}(x-1) f(2x), \qquad x \in (0, +\infty)
\end{equation*}
(respectively
\begin{equation*}
f(x)=\frac{x-1}{2} f(2x), \qquad x \in (0, +\infty)
\end{equation*}).
\fi
\subsection{The logarithmic Beta function - a logarithmic Cauchy quotient?}
The logarithmic Beta function $\mathcal{L}_1$ (respectively $\mathcal{L}_2$) is a logarithmic Cauchy quotient
if there is an interval $I \subset (1,+\infty)$ which is closed under multiplication and a function $f: I \to (0, +\infty)$ such that, for all $x,y \in I$,
\begin{equation*}
\frac{1}{6}\log{(x-1)}\log{(y-1)}=\frac{f(x) + f(y)}{f(xy)},
\end{equation*}
(respectively
\begin{equation*}
\frac{\log{(x-1)}+\log{(y-1)}}{2}=\frac{f(x) + f(y)}{f(xy)},
\end{equation*}
)
holds true.
\iffalse
Setting $x=y$, it follows
\begin{equation*}
f(x)=\frac{1}{12}\log{(x-1)} f(x^2), \qquad x \in I
\end{equation*}
(respectively
\begin{equation*}
f(x)=\frac{\log(x-1)}{2} f(x^2), \qquad x \in I
\end{equation*})
\fi
\section{Cauchy Beta functions: The case $k >2$}
In this section we introduce the Cauchy Beta functions of three or more variables.
Just as in the case of $\mathcal{M}$, the multiplicative Beta function of two variables,
power functions having sum of exponents equal to one were multiplied with each other (since on the right hand side of the respective Cauchy equation $f(xy)=f(x) \cdot f(y)$ the numbers $f(x)$ and $f(y)$ are multiplied) and then integrated over the unit interval,
the multiplicative Beta function of three variables $\mathcal{M}_3:\left(1,+\infty \right)^{3}\rightarrow {\mathbb{R}}$ is
defined by
\begin{equation}
\mathcal{M}_3\left( x,y,z\right)=
\int_0^1 \int_0^1 {(x-1)^s (y-1)^{t} (z-1)^{1-(s+t)}\, ds dt}, \RIfM@\expandafter\text@\else\expandafter\mbox\fi{ \ \ \ \ \ }x,y,z \in (1, +\infty).
\label{eq:Beta_m3:def}
\end{equation}
It clearly simplifies since, for all $x,y,z \in (1,+\infty)$, to
\begin{equation}
\mathcal{M}_3\left( x,y,z\right)=
\frac{(x-z)(y-z)}{(z-1) \log{\left({\frac{x-1}{z-1}}\right)} \log{\left({\frac{y-1}{z-1}}\right)}}.
\end{equation}
Inductively, we obtain also the $k$-variable multiplicative Beta function
$\mathcal{M}_{k}:\left(1,+\infty \right)^{k}\rightarrow {\mathbb{R}}$, which for all $x_1,\ldots,x_k \in (1,+\infty)$ is
defined by
\begin{equation}
\mathcal{M}_k\left( x_1, \ldots ,x_k\right)=
\int_0^1 \int_0^1 \cdots \int_0^1 {(x_1-1)^{t_1} (x_2-1)^{t_2} \cdots (x_k-1)^{1-(t_1+ \cdots + t_{k-1})}
\, d{x_1} d{x_2} \ldots d{x_{k-1}}}.
\label{eq:Beta_mk:def}
\end{equation}
It can be written as
\begin{equation}
\mathcal{M}_k\left( x_1, \ldots ,x_k\right)=
\frac{(x_1-x_k)(x_2-x_k) \cdots (x_{k-1}-x_k) }{(x_k-1)^{k-1} \log{{\left(\frac{x_1-1}{x_k-1}\right)}} \cdot \cdots \cdot
\log{\left({\frac{x_{k-1}-1}{x_k-1}}\right)}}
\end{equation}
Analogously, the first kind three variable additive Beta function $\mathcal{A}_{1, 3}: {\mathbb{R}}^3 \to {\mathbb{R}}$ is defined by
\begin{equation}
\mathcal{A}_{1, 3} \left( x, y ,z\right)=
\int_0^1 \int_0^1 {t(x-1)s(y-1)(1-(s+t))(z-1)}{\, dt ds},
\end{equation}
simplifying to
\begin{equation}
\mathcal{A}_{1, 3} \left( x, y ,z\right)=
-\frac{1}{12} (x-1)(y-1)(z-1).
\end{equation}
The first kind four variable additive Beta function $\mathcal{A}_{1, 4}: {\mathbb{R}}^4 \to {\mathbb{R}}$
simplifies to
\begin{equation}
\mathcal{A}_{1, 4} \left( x, y ,z,w\right)=
-\frac{1}{8} (x-1)(y-1)(z-1)(w-1).
\end{equation}
The first kind five variable additive Beta function $\mathcal{A}_{1, 5}: {\mathbb{R}}^5 \to {\mathbb{R}}$
simplifies to
\begin{equation}
\mathcal{A}_{1, 5} \left( x, y ,z, w, a \right)=
-\frac{5}{48} (x-1)(y-1)(z-1)(w-1)(a-1).
\end{equation}
\iffalse
Can you find a closed formula for
the first kind $k$-variable additive Beta function $\mathcal{A}_{1, k}: {\mathbb{R}}^k \to {\mathbb{R}}$,
i.e. what is the number $c$ in front of the term $(x_1-1)(x_2-1) \cdots (x_k-1)$?
\fi
\iffalse
The second (i.e. the more natural) additive Beta function of three variables $\mathcal{A}_{2, 3}: {\mathbb{R}}^3 \to {\mathbb{R}}$ defined by
\begin{equation}
\mathcal{A}_{2, 3} \left( x, y ,z\right)=
\int_0^1 {\int_0^1 {\left[t(x-1)+s(y-1)+(1-(s+t))(z-1)\right]}}{\, dt ds},
\end{equation}
simplifying to
\begin{equation}
\mathcal{A}_{2, 3} \left( x, y ,z\right)=
...
\end{equation}
What is general form of the second kind $k$-variable additive Beta form?
\fi
\iffalse
The two kinds of logarithmic Beta function of three variables $\mathcal{L}_{1, 3}, \mathcal{L}_{2, 3}: (1,+\infty)^k \to {\mathbb{R}}$ defined by
\begin{equation}
\mathcal{L}_{1, 3} \left( x, y ,z\right)=
\int_0^1 \int_0^1 {{\left[s\log(x-1) t\log(y-1)(1-(s+t))\log(z-1)\right]}}{\, dt ds},
\end{equation}
and
\begin{equation}
\mathcal{L}_{2, 3} \left( x, y ,z\right)=
\int_0^1 \int_0^1 {{\left[s\log(x-1) + t\log(y-1)+(1-(s+t))\log(z-1)\right]}}{\, dt ds},
\end{equation}
\fi
\section{Closing Remarks}
The questions whether the newly introduced pendants of Beta functions also
belong to their respective class of beta-type functions,
namely whether $\mathcal{A}_{1,k} \in A_{f,k}$ (or, more importantly, $\mathcal{A}_{2,k} \in A_{f,k}$)),
$\mathcal{M}_{k} \in P_{f,k}$ and $\mathcal{L}_{1,k} \in L_{f,k}$ (or $\mathcal{L}_{2,k} \in L_{f,k}$)
will be answered in a future paper.
The Euler Beta function is the continuous analog of (the reciprocal of) a binomial coefficient
and consequently it appeas quite often in combinatorics.
By \eqref{eq:Beta_Gamma}, most properties of the Gamma function,
as, for instance, the duplication formula, Stirling's formula or its relation to the sine function, can be formulated
in terms of the Beta function.
Moreover, there are numerous applications of the Euler Beta function in applied sciences as String theory, Particle physics or Astrophysics.
As a result this note opens a can of new research questions, namely which properties of the Euler Beta function are shared
by its here newly introduced siblings.
Also some problems of minor importance are left today without an answer (see sections $5.2$ to $5.4$ and $6$).
From the perspective of functional equations also some probing issues strongly intrude themself, since one can introduce similar Beta functions not only for the Cauchy equations, but also related to some other functional equations.
For example, let us consider the functional equation related to the sine-addition law
\begin{equation*}
f(x+y)=f(x) g(y)+f(y)g(x).
\end{equation*}
Obviously, $f=t \sin$, for $t \in {\mathbb{R}}$, and $g=\cos$ are continuous solutions on ${\mathbb{R}}$.
The reader with an at least minor understanding of this paper will agree that
it is more than natural to call $S: {\mathbb{R}}^2 \to {\mathbb{R}}$ defined by
\begin{equation*}
S(x,y)=\int_0^1 \left({t \sin{x} \cos{y}+(1-t) \sin{y} \cos{x}}\right), \qquad x,y \in {\mathbb{R}}
\end{equation*}
(simplifying, for all $x,y \in {\mathbb{R}}$, to
\begin{equation*}
S(x,y)=\frac{1}{2} \sin{(x+y)},
\end{equation*})
the Beta function of the sine-addition functional equation (related to this given smooth solution pair).
We may summarize that the Euler Beta function indeed gives rise to introducing some kind of auxiliary Beta function
in case of a functional euqation with solution involving some degree of freedom.
\section*{Abstract (Not appropriate in this style!)}%
\else \small
\begin{center}{\bf Abstract\vspace{-.5em}\vspace{\z@}}\end{center}%
\quotation
\fi
}%
}{%
}%
\@ifundefined{endabstract}{\def\endabstract
{\if@twocolumn\else\endquotation\fi}}{}%
\@ifundefined{maketitle}{\def\maketitle#1{}}{}%
\@ifundefined{affiliation}{\def\affiliation#1{}}{}%
\@ifundefined{proof}{\def\proof{\noindent{\bfseries Proof. }}}{}%
\@ifundefined{endproof}{\def\endproof{\mbox{\ \rule{.1in}{.1in}}}}{}%
\@ifundefined{newfield}{\def\newfield#1#2{}}{}%
\@ifundefined{chapter}{\def\chapter#1{\par(Chapter head:)#1\par }%
\newcount\c@chapter}{}%
\@ifundefined{part}{\def\part#1{\par(Part head:)#1\par }}{}%
\@ifundefined{section}{\def\section#1{\par(Section head:)#1\par }}{}%
\@ifundefined{subsection}{\def\subsection#1%
{\par(Subsection head:)#1\par }}{}%
\@ifundefined{subsubsection}{\def\subsubsection#1%
{\par(Subsubsection head:)#1\par }}{}%
\@ifundefined{paragraph}{\def\paragraph#1%
{\par(Subsubsubsection head:)#1\par }}{}%
\@ifundefined{subparagraph}{\def\subparagraph#1%
{\par(Subsubsubsubsection head:)#1\par }}{}%
\@ifundefined{therefore}{\def\therefore{}}{}%
\@ifundefined{backepsilon}{\def\backepsilon{}}{}%
\@ifundefined{yen}{\def\yen{\hbox{\rm\rlap=Y}}}{}%
\@ifundefined{registered}{%
\def\registered{\relax\ifmmode{}\r@gistered
\else$\m@th\r@gistered$\fi}%
\def\r@gistered{^{\ooalign
{\hfil\raise.07ex\hbox{$\scriptstyle\rm\RIfM@\expandafter\text@\else\expandafter\mbox\fi{R}$}\hfil\crcr
\mathhexbox20D}}}}{}%
\@ifundefined{Eth}{\def\Eth{}}{}%
\@ifundefined{eth}{\def\eth{}}{}%
\@ifundefined{Thorn}{\def\Thorn{}}{}%
\@ifundefined{thorn}{\def\thorn{}}{}%
\def\TEXTsymbol#1{\mbox{$#1$}}%
\@ifundefined{degree}{\def\degree{{}^{\circ}}}{}%
\newdimen\theight
\def\Column{%
\vadjust{\setbox\z@=\hbox{\scriptsize\quad\quad tcol}%
\theight=\ht\z@\advance\theight by \dp\z@\advance\theight by \lineskip
\kern -\theight \vbox to \theight{%
\rightline{\rlap{\box\z@}}%
\vss
}%
}%
}%
\def\qed{%
\ifhmode\unskip\nobreak\fi\ifmmode\ifinner\else\hskip5\p@\fi\fi
\hbox{\hskip5\p@\vrule width4\p@ height6\p@ depth1.5\p@\hskip\p@}%
}%
\def\cents{\hbox{\rm\rlap/c}}%
\def\miss{\hbox{\vrule height2\p@ width 2\p@ depth\z@}}%
\def\vvert{\Vert
\def\tcol#1{{\baselineskip=6\p@ \vcenter{#1}} \Column} %
\def\dB{\hbox{{}}
\def\mB#1{\hbox{$#1$}
\def\nB#1{\hbox{#1}
\def\note{$^{\dag}}%
\defLaTeX2e{LaTeX2e}
\def\chkcompat{%
\if@compatibility
\else
\usepackage{latexsym}
\fi
}
\ifx\fmtnameLaTeX2e
\DeclareOldFontCommand{\rm}{\normalfont\rmfamily}{\mathrm}
\DeclareOldFontCommand{\sf}{\normalfont\sffamily}{\mathsf}
\DeclareOldFontCommand{\tt}{\normalfont\ttfamily}{\mathtt}
\DeclareOldFontCommand{\bf}{\normalfont\bfseries}{\mathbf}
\DeclareOldFontCommand{\it}{\normalfont\itshape}{\mathit}
\DeclareOldFontCommand{\sl}{\normalfont\slshape}{\@nomath\sl}
\DeclareOldFontCommand{\sc}{\normalfont\scshape}{\@nomath\sc}
\chkcompat
\fi
\def\alpha{\Greekmath 010B }%
\def\beta{\Greekmath 010C }%
\def\gamma{\Greekmath 010D }%
\def\delta{\Greekmath 010E }%
\def\epsilon{\Greekmath 010F }%
\def\zeta{\Greekmath 0110 }%
\def\eta{\Greekmath 0111 }%
\def\theta{\Greekmath 0112 }%
\def\iota{\Greekmath 0113 }%
\def\kappa{\Greekmath 0114 }%
\def\lambda{\Greekmath 0115 }%
\def\mu{\Greekmath 0116 }%
\def\nu{\Greekmath 0117 }%
\def\xi{\Greekmath 0118 }%
\def\pi{\Greekmath 0119 }%
\def\rho{\Greekmath 011A }%
\def\sigma{\Greekmath 011B }%
\def\tau{\Greekmath 011C }%
\def\upsilon{\Greekmath 011D }%
\def\phi{\Greekmath 011E }%
\def\chi{\Greekmath 011F }%
\def\psi{\Greekmath 0120 }%
\def\omega{\Greekmath 0121 }%
\def\varepsilon{\Greekmath 0122 }%
\def\vartheta{\Greekmath 0123 }%
\def\varpi{\Greekmath 0124 }%
\def\varrho{\Greekmath 0125 }%
\def\varsigma{\Greekmath 0126 }%
\def\varphi{\Greekmath 0127 }%
\def\Greekmath 0272 {\Greekmath 0272 }
\def\FindBoldGroup{%
{\setbox0=\hbox{$\mathbf{x\global\edef\theboldgroup{\the\mathgroup}}$}}%
}
\def\Greekmath#1#2#3#4{%
\if@compatibility
\ifnum\mathgroup=\symbold
\mathchoice{\mbox{\boldmath$\displaystyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\textstyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\scriptstyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\scriptscriptstyle\mathchar"#1#2#3#4$}}%
\else
\mathchar"#1#2#3#
\fi
\else
\FindBoldGroup
\ifnum\mathgroup=\theboldgroup
\mathchoice{\mbox{\boldmath$\displaystyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\textstyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\scriptstyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\scriptscriptstyle\mathchar"#1#2#3#4$}}%
\else
\mathchar"#1#2#3#
\fi
\fi}
\newif\ifGreekBold \GreekBoldfalse
\let\SAVEPBF=\pbf
\def\pbf{\GreekBoldtrue\SAVEPBF}%
\@ifundefined{theorem}{\newtheorem{theorem}{Theorem}}{}
\@ifundefined{lemma}{\newtheorem{lemma}[theorem]{Lemma}}{}
\@ifundefined{corollary}{\newtheorem{corollary}[theorem]{Corollary}}{}
\@ifundefined{conjecture}{\newtheorem{conjecture}[theorem]{Conjecture}}{}
\@ifundefined{proposition}{\newtheorem{proposition}[theorem]{Proposition}}{}
\@ifundefined{axiom}{\newtheorem{axiom}{Axiom}}{}
\@ifundefined{remark}{\newtheorem{remark}{Remark}}{}
\@ifundefined{example}{\newtheorem{example}{Example}}{}
\@ifundefined{exercise}{\newtheorem{exercise}{Exercise}}{}
\@ifundefined{definition}{\newtheorem{definition}{Definition}}{}
\@ifundefined{mathletters}{%
\newcounter{equationnumber}
\def\mathletters{%
\addtocounter{equation}{1}
\edef\@currentlabel{\arabic{equation}}%
\setcounter{equationnumber}{\c@equation}
\setcounter{equation}{0}%
\edef\arabic{equation}{\@currentlabel\noexpand\alph{equation}}%
}
\def\endmathletters{%
\setcounter{equation}{\value{equationnumber}}%
}
}{}
\@ifundefined{BibTeX}{%
\def\BibTeX{{\rm B\kern-.05em{\sc i\kern-.025em b}\kern-.08em
T\kern-.1667em\lower.7ex\hbox{E}\kern-.125emX}}}{}%
\@ifundefined{AmS}%
{\def\AmS{{\protect\usefont{OMS}{cmsy}{m}{n}%
A\kern-.1667em\lower.5ex\hbox{M}\kern-.125emS}}}{}%
\@ifundefined{AmSTeX}{\def\AmSTeX{\protect\AmS-\protect\TeX\@}}{}%
\ifx\ds@amstex\relax
\message{amstex already loaded}\makeatother\endinpu
\else
\@ifpackageloaded{amstex}%
{\message{amstex already loaded}\makeatother\endinput}
{}
\@ifpackageloaded{amsgen}%
{\message{amsgen already loaded}\makeatother\endinput}
{}
\fi
\let\DOTSI\relax
\def\RIfM@{\relax\ifmmode}%
\def\FN@{\futurelet\next}%
\newcount\intno@
\def\iint{\DOTSI\intno@\tw@\FN@\ints@}%
\def\iiint{\DOTSI\intno@\thr@@\FN@\ints@}%
\def\iiiint{\DOTSI\intno@4 \FN@\ints@}%
\def\idotsint{\DOTSI\intno@\z@\FN@\ints@}%
\def\ints@{\findlimits@\ints@@}%
\newif\iflimtoken@
\newif\iflimits@
\def\findlimits@{\limtoken@true\ifx\next\limits\limits@true
\else\ifx\next\nolimits\limits@false\else
\limtoken@false\ifx\ilimits@\nolimits\limits@false\else
\ifinner\limits@false\else\limits@true\fi\fi\fi\fi}%
\def\multint@{\int\ifnum\intno@=\z@\intdots@
\else\intkern@\fi
\ifnum\intno@>\tw@\int\intkern@\fi
\ifnum\intno@>\thr@@\int\intkern@\fi
\int
\def\multintlimits@{\intop\ifnum\intno@=\z@\intdots@\else\intkern@\fi
\ifnum\intno@>\tw@\intop\intkern@\fi
\ifnum\intno@>\thr@@\intop\intkern@\fi\intop}%
\def\intic@{%
\mathchoice{\hskip.5em}{\hskip.4em}{\hskip.4em}{\hskip.4em}}%
\def\negintic@{\mathchoice
{\hskip-.5em}{\hskip-.4em}{\hskip-.4em}{\hskip-.4em}}%
\def\ints@@{\iflimtoken@
\def\ints@@@{\iflimits@\negintic@
\mathop{\intic@\multintlimits@}\limits
\else\multint@\nolimits\fi
\eat@
\else
\def\ints@@@{\iflimits@\negintic@
\mathop{\intic@\multintlimits@}\limits\else
\multint@\nolimits\fi}\fi\ints@@@}%
\def\intkern@{\mathchoice{\!\!\!}{\!\!}{\!\!}{\!\!}}%
\def\plaincdots@{\mathinner{\cdotp\cdotp\cdotp}}%
\def\intdots@{\mathchoice{\plaincdots@}%
{{\cdotp}\mkern1.5mu{\cdotp}\mkern1.5mu{\cdotp}}%
{{\cdotp}\mkern1mu{\cdotp}\mkern1mu{\cdotp}}%
{{\cdotp}\mkern1mu{\cdotp}\mkern1mu{\cdotp}}}%
\def\RIfM@{\relax\protect\ifmmode}
\def\RIfM@\expandafter\text@\else\expandafter\mbox\fi{\RIfM@\expandafter\RIfM@\expandafter\text@\else\expandafter\mbox\fi@\else\expandafter\mbox\fi}
\let\nfss@text\RIfM@\expandafter\text@\else\expandafter\mbox\fi
\def\RIfM@\expandafter\text@\else\expandafter\mbox\fi@#1{\mathchoice
{\textdef@\displaystyle\f@size{#1}}%
{\textdef@\textstyle\tf@size{\firstchoice@false #1}}%
{\textdef@\textstyle\sf@size{\firstchoice@false #1}}%
{\textdef@\textstyle \ssf@size{\firstchoice@false #1}}%
\glb@settings}
\def\textdef@#1#2#3{\hbox{{%
\everymath{#1}%
\let\f@size#2\selectfont
#3}}}
\newif\iffirstchoice@
\firstchoice@true
\def\Let@{\relax\iffalse{\fi\let\\=\cr\iffalse}\fi}%
\def\vspace@{\def\vspace##1{\crcr\noalign{\vskip##1\relax}}}%
\def\multilimits@{\bgroup\vspace@\Let@
\baselineskip\fontdimen10 \scriptfont\tw@
\advance\baselineskip\fontdimen12 \scriptfont\tw@
\lineskip\thr@@\fontdimen8 \scriptfont\thr@@
\lineskiplimit\lineskip
\vbox\bgroup\ialign\bgroup\hfil$\m@th\scriptstyle{##}$\hfil\crcr}%
\def\Sb{_\multilimits@}%
\def\endSb{\crcr\egroup\egroup\egroup}%
\def\Sp{^\multilimits@}%
\let\endSp\endSb
\newdimen\ex@
\[email protected]
\def\rightarrowfill@#1{$#1\m@th\mathord-\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\mathord-\mkern-2mu$}\hfill
\mkern-6mu\mathord\rightarrow$}%
\def\leftarrowfill@#1{$#1\m@th\mathord\leftarrow\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\mathord-\mkern-2mu$}\hfill\mkern-6mu\mathord-$}%
\def\leftrightarrowfill@#1{$#1\m@th\mathord\leftarrow
\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\mathord-\mkern-2mu$}\hfill
\mkern-6mu\mathord\rightarrow$}%
\def\overrightarrow{\mathpalette\overrightarrow@}%
\def\overrightarrow@#1#2{\vbox{\ialign{##\crcr\rightarrowfill@#1\crcr
\noalign{\kern-\ex@\nointerlineskip}$\m@th\hfil#1#2\hfil$\crcr}}}%
\let\overarrow\overrightarrow
\def\overleftarrow{\mathpalette\overleftarrow@}%
\def\overleftarrow@#1#2{\vbox{\ialign{##\crcr\leftarrowfill@#1\crcr
\noalign{\kern-\ex@\nointerlineskip}$\m@th\hfil#1#2\hfil$\crcr}}}%
\def\overleftrightarrow{\mathpalette\overleftrightarrow@}%
\def\overleftrightarrow@#1#2{\vbox{\ialign{##\crcr
\leftrightarrowfill@#1\crcr
\noalign{\kern-\ex@\nointerlineskip}$\m@th\hfil#1#2\hfil$\crcr}}}%
\def\underrightarrow{\mathpalette\underrightarrow@}%
\def\underrightarrow@#1#2{\vtop{\ialign{##\crcr$\m@th\hfil#1#2\hfil
$\crcr\noalign{\nointerlineskip}\rightarrowfill@#1\crcr}}}%
\let\underarrow\underrightarrow
\def\underleftarrow{\mathpalette\underleftarrow@}%
\def\underleftarrow@#1#2{\vtop{\ialign{##\crcr$\m@th\hfil#1#2\hfil
$\crcr\noalign{\nointerlineskip}\leftarrowfill@#1\crcr}}}%
\def\underleftrightarrow{\mathpalette\underleftrightarrow@}%
\def\underleftrightarrow@#1#2{\vtop{\ialign{##\crcr$\m@th
\hfil#1#2\hfil$\crcr
\noalign{\nointerlineskip}\leftrightarrowfill@#1\crcr}}}%
\def\qopnamewl@#1{\mathop{\operator@font#1}\nlimits@}
\let\nlimits@\displaylimits
\def\setboxz@h{\setbox\z@\hbox}
\def\varlim@#1#2{\mathop{\vtop{\ialign{##\crcr
\hfil$#1\m@th\operator@font lim$\hfil\crcr
\noalign{\nointerlineskip}#2#1\crcr
\noalign{\nointerlineskip\kern-\ex@}\crcr}}}}
\def\rightarrowfill@#1{\m@th\setboxz@h{$#1-$}\ht\z@\z@
$#1\copy\z@\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\box\z@\mkern-2mu$}\hfill
\mkern-6mu\mathord\rightarrow$}
\def\leftarrowfill@#1{\m@th\setboxz@h{$#1-$}\ht\z@\z@
$#1\mathord\leftarrow\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\copy\z@\mkern-2mu$}\hfill
\mkern-6mu\box\z@$}
\def\qopnamewl@{proj\,lim}{\qopnamewl@{proj\,lim}}
\def\qopnamewl@{inj\,lim}{\qopnamewl@{inj\,lim}}
\def\mathpalette\varlim@\rightarrowfill@{\mathpalette\varlim@\rightarrowfill@}
\def\mathpalette\varlim@\leftarrowfill@{\mathpalette\varlim@\leftarrowfill@}
\def\mathpalette\varliminf@{}{\mathpalette\mathpalette\varliminf@{}@{}}
\def\mathpalette\varliminf@{}@#1{\mathop{\underline{\vrule\@depth.2\ex@\@width\z@
\hbox{$#1\m@th\operator@font lim$}}}}
\def\mathpalette\varlimsup@{}{\mathpalette\mathpalette\varlimsup@{}@{}}
\def\mathpalette\varlimsup@{}@#1{\mathop{\overline
{\hbox{$#1\m@th\operator@font lim$}}}}
\def\tfrac#1#2{{\textstyle {#1 \over #2}}}%
\def\dfrac#1#2{{\displaystyle {#1 \over #2}}}%
\def\binom#1#2{{#1 \choose #2}}%
\def\tbinom#1#2{{\textstyle {#1 \choose #2}}}%
\def\dbinom#1#2{{\displaystyle {#1 \choose #2}}}%
\def\QATOP#1#2{{#1 \atop #2}}%
\def\QTATOP#1#2{{\textstyle {#1 \atop #2}}}%
\def\QDATOP#1#2{{\displaystyle {#1 \atop #2}}}%
\def\QABOVE#1#2#3{{#2 \above#1 #3}}%
\def\QTABOVE#1#2#3{{\textstyle {#2 \above#1 #3}}}%
\def\QDABOVE#1#2#3{{\displaystyle {#2 \above#1 #3}}}%
\def\QOVERD#1#2#3#4{{#3 \overwithdelims#1#2 #4}}%
\def\QTOVERD#1#2#3#4{{\textstyle {#3 \overwithdelims#1#2 #4}}}%
\def\QDOVERD#1#2#3#4{{\displaystyle {#3 \overwithdelims#1#2 #4}}}%
\def\QATOPD#1#2#3#4{{#3 \atopwithdelims#1#2 #4}}%
\def\QTATOPD#1#2#3#4{{\textstyle {#3 \atopwithdelims#1#2 #4}}}%
\def\QDATOPD#1#2#3#4{{\displaystyle {#3 \atopwithdelims#1#2 #4}}}%
\def\QABOVED#1#2#3#4#5{{#4 \abovewithdelims#1#2#3 #5}}%
\def\QTABOVED#1#2#3#4#5{{\textstyle
{#4 \abovewithdelims#1#2#3 #5}}}%
\def\QDABOVED#1#2#3#4#5{{\displaystyle
{#4 \abovewithdelims#1#2#3 #5}}}%
\def\tint{\mathop{\textstyle \int}}%
\def\tiint{\mathop{\textstyle \iint }}%
\def\tiiint{\mathop{\textstyle \iiint }}%
\def\tiiiint{\mathop{\textstyle \iiiint }}%
\def\tidotsint{\mathop{\textstyle \idotsint }}%
\def\toint{\mathop{\textstyle \oint}}%
\def\tsum{\mathop{\textstyle \sum }}%
\def\tprod{\mathop{\textstyle \prod }}%
\def\tbigcap{\mathop{\textstyle \bigcap }}%
\def\tbigwedge{\mathop{\textstyle \bigwedge }}%
\def\tbigoplus{\mathop{\textstyle \bigoplus }}%
\def\tbigodot{\mathop{\textstyle \bigodot }}%
\def\tbigsqcup{\mathop{\textstyle \bigsqcup }}%
\def\tcoprod{\mathop{\textstyle \coprod }}%
\def\tbigcup{\mathop{\textstyle \bigcup }}%
\def\tbigvee{\mathop{\textstyle \bigvee }}%
\def\tbigotimes{\mathop{\textstyle \bigotimes }}%
\def\tbiguplus{\mathop{\textstyle \biguplus }}%
\def\dint{\mathop{\displaystyle \int}}%
\def\diint{\mathop{\displaystyle \iint }}%
\def\diiint{\mathop{\displaystyle \iiint }}%
\def\diiiint{\mathop{\displaystyle \iiiint }}%
\def\didotsint{\mathop{\displaystyle \idotsint }}%
\def\doint{\mathop{\displaystyle \oint}}%
\def\dsum{\mathop{\displaystyle \sum }}%
\def\dprod{\mathop{\displaystyle \prod }}%
\def\dbigcap{\mathop{\displaystyle \bigcap }}%
\def\dbigwedge{\mathop{\displaystyle \bigwedge }}%
\def\dbigoplus{\mathop{\displaystyle \bigoplus }}%
\def\dbigodot{\mathop{\displaystyle \bigodot }}%
\def\dbigsqcup{\mathop{\displaystyle \bigsqcup }}%
\def\dcoprod{\mathop{\displaystyle \coprod }}%
\def\dbigcup{\mathop{\displaystyle \bigcup }}%
\def\dbigvee{\mathop{\displaystyle \bigvee }}%
\def\dbigotimes{\mathop{\displaystyle \bigotimes }}%
\def\dbiguplus{\mathop{\displaystyle \biguplus }}%
\def\stackunder#1#2{\mathrel{\mathop{#2}\limits_{#1}}}%
\begingroup \catcode `|=0 \catcode `[= 1
\catcode`]=2 \catcode `\{=12 \catcode `\}=12
\catcode`\\=12
|gdef|@alignverbatim#1\end{align}[#1|end[align]]
|gdef|@salignverbatim#1\end{align*}[#1|end[align*]]
|gdef|@alignatverbatim#1\end{alignat}[#1|end[alignat]]
|gdef|@salignatverbatim#1\end{alignat*}[#1|end[alignat*]]
|gdef|@xalignatverbatim#1\end{xalignat}[#1|end[xalignat]]
|gdef|@sxalignatverbatim#1\end{xalignat*}[#1|end[xalignat*]]
|gdef|@gatherverbatim#1\end{gather}[#1|end[gather]]
|gdef|@sgatherverbatim#1\end{gather*}[#1|end[gather*]]
|gdef|@gatherverbatim#1\end{gather}[#1|end[gather]]
|gdef|@sgatherverbatim#1\end{gather*}[#1|end[gather*]]
|gdef|@multilineverbatim#1\end{multiline}[#1|end[multiline]]
|gdef|@smultilineverbatim#1\end{multiline*}[#1|end[multiline*]]
|gdef|@arraxverbatim#1\end{arrax}[#1|end[arrax]]
|gdef|@sarraxverbatim#1\end{arrax*}[#1|end[arrax*]]
|gdef|@tabulaxverbatim#1\end{tabulax}[#1|end[tabulax]]
|gdef|@stabulaxverbatim#1\end{tabulax*}[#1|end[tabulax*]]
|endgroup
\def\align{\@verbatim \frenchspacing\@vobeyspaces \@alignverbatim
You are using the "align" environment in a style in which it is not defined.}
\let\endalign=\endtrivlist
\@namedef{align*}{\@verbatim\@salignverbatim
You are using the "align*" environment in a style in which it is not defined.}
\expandafter\let\csname endalign*\endcsname =\endtrivlist
\def\alignat{\@verbatim \frenchspacing\@vobeyspaces \@alignatverbatim
You are using the "alignat" environment in a style in which it is not defined.}
\let\endalignat=\endtrivlist
\@namedef{alignat*}{\@verbatim\@salignatverbatim
You are using the "alignat*" environment in a style in which it is not defined.}
\expandafter\let\csname endalignat*\endcsname =\endtrivlist
\def\xalignat{\@verbatim \frenchspacing\@vobeyspaces \@xalignatverbatim
You are using the "xalignat" environment in a style in which it is not defined.}
\let\endxalignat=\endtrivlist
\@namedef{xalignat*}{\@verbatim\@sxalignatverbatim
You are using the "xalignat*" environment in a style in which it is not defined.}
\expandafter\let\csname endxalignat*\endcsname =\endtrivlist
\def\gather{\@verbatim \frenchspacing\@vobeyspaces \@gatherverbatim
You are using the "gather" environment in a style in which it is not defined.}
\let\endgather=\endtrivlist
\@namedef{gather*}{\@verbatim\@sgatherverbatim
You are using the "gather*" environment in a style in which it is not defined.}
\expandafter\let\csname endgather*\endcsname =\endtrivlist
\def\multiline{\@verbatim \frenchspacing\@vobeyspaces \@multilineverbatim
You are using the "multiline" environment in a style in which it is not defined.}
\let\endmultiline=\endtrivlist
\@namedef{multiline*}{\@verbatim\@smultilineverbatim
You are using the "multiline*" environment in a style in which it is not defined.}
\expandafter\let\csname endmultiline*\endcsname =\endtrivlist
\def\arrax{\@verbatim \frenchspacing\@vobeyspaces \@arraxverbatim
You are using a type of "array" construct that is only allowed in AmS-LaTeX.}
\let\endarrax=\endtrivlist
\def\tabulax{\@verbatim \frenchspacing\@vobeyspaces \@tabulaxverbatim
You are using a type of "tabular" construct that is only allowed in AmS-LaTeX.}
\let\endtabulax=\endtrivlist
\@namedef{arrax*}{\@verbatim\@sarraxverbatim
You are using a type of "array*" construct that is only allowed in AmS-LaTeX.}
\expandafter\let\csname endarrax*\endcsname =\endtrivlist
\@namedef{tabulax*}{\@verbatim\@stabulaxverbatim
You are using a type of "tabular*" construct that is only allowed in AmS-LaTeX.}
\expandafter\let\csname endtabulax*\endcsname =\endtrivlist
\def\@@eqncr{\let\@tempa\relax
\ifcase\@eqcnt \def\@tempa{& & &}\or \def\@tempa{& &}%
\else \def\@tempa{&}\fi
\@tempa
\if@eqnsw
\iftag@
\@taggnum
\else
\@eqnnum\stepcounter{equation}%
\fi
\fi
\global\@ifnextchar*{\@tagstar}{\@tag}@false
\global\@eqnswtrue
\global\@eqcnt\z@\cr}
\def\endequation{%
\ifmmode\ifinner
\iftag@
\addtocounter{equation}{-1}
$\hfil
\displaywidth\linewidth\@taggnum\egroup \endtrivlist
\global\@ifnextchar*{\@tagstar}{\@tag}@false
\global\@ignoretrue
\else
$\hfil
\displaywidth\linewidth\@eqnnum\egroup \endtrivlist
\global\@ifnextchar*{\@tagstar}{\@tag}@false
\global\@ignoretrue
\fi
\else
\iftag@
\addtocounter{equation}{-1}
\eqno \hbox{\@taggnum}
\global\@ifnextchar*{\@tagstar}{\@tag}@false%
$$\global\@ignoretrue
\else
\eqno \hbox{\@eqnnum
$$\global\@ignoretrue
\fi
\fi\fi
}
\newif\iftag@ \@ifnextchar*{\@tagstar}{\@tag}@false
\def\@ifnextchar*{\@tagstar}{\@tag}{\@ifnextchar*{\@tagstar}{\@tag}}
\def\@tag#1{%
\global\@ifnextchar*{\@tagstar}{\@tag}@true
\global\def\@taggnum{(#1)}}
\def\@tagstar*#1{%
\global\@ifnextchar*{\@tagstar}{\@tag}@true
\global\def\@taggnum{#1
}
\makeatother
\endinput
|
\section{Proof of the main statement}
\label{sec:MainProof}
We are finally in the position to derive the key ingredient, an iterated recursive estimate for critical points $\gamma$ of the generalized integral Menger curvature $\Mpq[p,2]$ with respect to a Sobolev norm. Based on that, we prove the main result of this paper.
\subsection{Recursive estimate and iteration}
\label{subsec:RecursiveEstimate}
In order to show analyticity of a curve, we need to bound its higher derivatives. One possible approach is to bound derivatives by the lower order ones and iterate this procedure inductively. In the following we show how such recursive estimates can be obtained from the Euler-Lagrange equations.
Let $\gamma\in C^\infty(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ be a critical point of $\Mpq[p,2]$ with respect to fixed length and $p\in(\tfrac 73,\tfrac 83)$.
By the Lagrange multiplier theorem (see e.g.\ \cite[§9.3~Theorem 1]{Luenberger:1998:OptimizationVectorSpace}), we have the existence of $\lambda \in \mathbb{R}$ such that
\[
0 = \delta\Mpq[p,2](\gamma,h) + \lambda \delta \mathcal{L}(\gamma,h).
\]
Seeing as by integration by parts and periodicity, we have
\[
\delta\mathcal{L}(\gamma,h) = \int_{\mathbb{R}/\mathbb{Z}} \left<\gamma'(u),h'(u)\right> \d u = - \int_{\mathbb{R}/\mathbb{Z}} \left<\gamma''(u),h(u)\right> \d u
\]
for $\gamma$ parametrized by arc-length.
Combining the decomposition of the first variation of $\Mpq[p,2]$ mentioned in \eqref{eq:firstVariation} with the $L^2$-forms of the main term $Q^{(p)}$ in \thref{lemma:L2MainTermViaFourier} and the remainder term $R^{(p)}$ in \thref{lemma:L2formRemainder}, we may write
\[
0 = \int_{\mathbb{R}/\mathbb{Z}} \left<12 \bigl(\ensuremath{\widetilde Q^{(p)}}(\gamma)(u) + \ensuremath{\widetilde R^{(p)}}(\gamma)(u)\bigr) - \lambda \gamma''(u), h(u)\right> \d u.
\]
Thus, we have
\[
\label{eq:RelationMainAndRemainderTerms}
\refstepcounter{equation}\tag{\theequation}
\ensuremath{\widetilde Q^{(p)}}(\gamma) = \frac {\lambda} {12} \gamma'' - \ensuremath{\widetilde R^{(p)}}(\gamma).
\]
From this, we may easily deduce the following.
\begin{corollary}
\label{cor:estimateMainTermViaRemainder}
Let $\gamma \in C^\infty(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ be a critical point of $\Mpq[p,2]$ with respect to fixed length and parametrized by arc-length.
For $m>\frac 1 2$ there exist constants
$C=C(\gamma,m,n,p)>0$, $r=r(\gamma,m,n,p)>0$, and $\mu=\mu(\gamma, n,p)$ such that
\[
\begin{split}
&\norm[H^{m+3p-7}]{\gamma^{(l+3)}}\\
&\le C \sum_{j_1 + \ldots + j_{K-1} = l+1} \binom{l+1}{j_1, \ldots, j_{K-1}} \ p^{(3n)}_{j_1} \left( \left\{\frac{(|\alpha|+1)!}{r^{|\alpha|+1}} \right\}_{|\alpha|\leq j_1}, \{ \| \gamma^{(j+1)} \|_{H^{m+ \frac 32 p -3}}\}_{\substack{j =1, \ldots, j_1\\ i = 1, \ldots, 3n}} \right) \\
& \phantom{\le} \cdot \prod_{i=2}^{K-1} {\|\gamma^{(j_i+1)}\|_{H^{m+ \frac 32 p -3}}} + \mu \norm[H^m]{\gamma^{(l+2)}}.
\end{split}
\]
\end{corollary}
\begin{proof}
By \thref{cor:Ql}, we may estimate $\norm[H^{m+3p-7}]{\gamma^{(l+3)}} \le \widetilde C \norm[H^m]{\partial^l\ensuremath{\widetilde Q^{(p)}}(\gamma)}$.
Noting that $\widetilde C$ only depends on $p$ and $n$, we only have to combine this estimate with \eqref{eq:RelationMainAndRemainderTerms} and \thref{lemma:estimateRemainder}.
\end{proof}
The obtained estimate is already recursive, however, the difference in the order of differentiability is less than 1 since
\[
3p-7 < \tfrac 32 p -3
\]
for any $p\in (\tfrac 73, \tfrac 83)$. This can lead to substantial problems in the iteration of the recursive estimate to bound higher derivatives. Therefore, we iterate the recursive estimate beforehand to get a difference of order at least 1 in the recursion.
To make the structure of the estimate more visible, set
\[
\refstepcounter{equation}\tag{\theequation}
\label{eq:definitionPhi}
\begin{split}
&\Phi(l,n,r,K;x_0,\ldots,x_l) :=\\
&\sum_{j_1 + \ldots + j_{K-1} = l} \binom{l}{j_1,\ldots,j_{K-1}} p_{j_1}^{(3n)}\left( \left\{ \frac {(\abs{\alpha} +1)!} {r^{\abs{\alpha} +1} } \right\}_{\abs{\alpha} \leq j_1}, \left\{ x_j \right\}_{\substack{j=1,\ldots,j_1\\ i=1,\ldots, 3n}} \right) \cdot \prod_{i=2}^{K-1} x_{j_i}
\end{split}
\]
and note that $\Phi$ is non-decreasing in the $x_i$.
The upper bound from \thref{cor:estimateMainTermViaRemainder} may be written as
\[
\refstepcounter{equation}\tag{\theequation}
\label{eq:upperBoundMainTermPhi}
C \Phi\left(l+1,n,r,K;\norm[H^{m+\frac 3 2 p -3}]{\gamma'}, \ldots, \norm[H^{m+\frac 3 2 p -3}]{\gamma^{(l+2)}}\right) + \mu \norm[H^m]{\gamma^{(l+2)}}.
\]
\begin{lemma}
\label{lemma:iteratedEstimateMainTerm}
Let $\gamma \in C^\infty(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ parametrized by arc-length be a critical point of $\Mpq[p,2]$ with respect to fixed length and $m\geq \frac 5 2$.
Then, there exist constants $\widehat C = \widehat C>0$, $\mu=\mu$ and $r>0$ all depending on $\gamma,m,n$ and $p$
such that
\[
\begin{split}
&\norm[H^{m+3p-7}]{\gamma^{(l+3)}}\\
& \le C \Phi\Bigl( l+1, n, r,K;\widehat{C}\Phi\left[1,n,\frac {r} {2\pi},K; \norm[H^m]{\gamma}, \norm[H^m]{\gamma'}\right] + \mu \norm[H^m]{\gamma'}, \ldots,\\
&\adjustedalignment{C \Phi\Bigl( l+1, n, r,K;\widehat{C}}{\widehat{C}}\Phi\left[l+2,n,\frac {r} {2\pi},K; \norm[H^m]{\gamma}, \ldots, \norm[H^m]{\gamma^{(l+2)}}\right] + \mu \norm[H^m]{\gamma^{(l+2)}}\Bigr)\\
&+\mu \norm[H^m]{\gamma^{(l+2)}}.
\end{split}
\]
\end{lemma}
\begin{proof}
Using \thref{cor:estimateMainTermViaRemainder} and \eqref{eq:upperBoundMainTermPhi}, it suffices to bound $\norm[H^{m + \frac 3 2 p -3}]{\gamma^{(\tilde l +1)}}$ for $\tilde l \in \{0, \ldots, l+1\}$.
Since $m\ge \frac 52$, we obtain $m + \frac 3 2 p -5 >\frac 1 2$ and may use \thref{lemma:equivalentSobolevNormsDerivative}, \thref{cor:Ql} and \thref{cor:estimateMainTermViaRemainder} to estimate
\[
\begin{split}
\norm[H^{m + \frac 3 2 p -3}]{\gamma^{(\tilde l +1)}}
&\le \norm[H^{m + \frac 3 2 p -5}]{\gamma^{(\tilde l +3)}} \\
&\le \widetilde C(p) C \Phi\left({\tilde l}+1,n,r,K;\norm[H^{m-1}]{\gamma'}, \ldots, \norm[H^{m-1}]{\gamma^{({\tilde l}+2)}}\right) + \mu \norm[H^{m + 2 - \frac 3 2 p}]{\gamma^{({\tilde l}+2)}}.
\end{split}
\]
It holds that $m+2-\frac 3 2 p < m- \frac 3 2 < m-1$ and by \thref{lemma:equivalentSobolevNormsDerivative}, $\norm[H^{m-1}]{\gamma^{(j+1)}} \le 2 \pi \norm[H^m]{\gamma^{(j)}}$.
A look at the definition of $\Phi$ together with \thref{lemma:estimateAnalyticFaaDiBruno} enables us to estimate the above by
\[
\widetilde C(p) C (2\pi)^{K-2} \Phi\left({\tilde l}+1,n,\frac {r} {2\pi},K;\norm[H^{m}]{\gamma}, \ldots, \norm[H^m]{\gamma^{({\tilde l}+1)}}\right) + \tilde \mu \norm[H^{m}]{\gamma^{({\tilde l}+1)}}
\]
for $\tilde \mu := \tilde\mu(\gamma,m,n,p)>0$.
Relabeling $\mu := \max\set{\mu,\tilde \mu}$ gives the desired statement.
\end{proof}
\subsection{Proof of Theorem \ref{thm:main}}
Now we are ready to proof the main statement of this paper by the method of majorants.
\begin{proof}[Proof of \thref{thm:main}]
Recall we have given a simple closed curve
\[
\gamma = (\gamma_1,\ldots, \gamma_n) \in C^\infty(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)
\]
parametrized by arc-length which is a critical point of integral Menger curvature $\Mpq[p,2] + \lambda \mathcal{L}$, subject to a fixed length constraint, where $p\in (\tfrac 73, \tfrac 83)$ and $\lambda \in \mathbb{R}$.
We define
\[
a_l := \norm[H^{\frac 5 2}]{\gamma^{(l)}}
\]
for all integers $l\geq 0$ by using the smoothness of the curve $\gamma$.
Our aim is show that there exist positive constants $C_\gamma$ and $r_\gamma$ such that we have for any integer $l\geq 0$
\begin{align*}
a_l \leq C_\gamma \frac{l!}{r_\gamma^l};
\end{align*}
from which the analyticity of the critical point $\gamma$ on $\mathbb{R}/\mathbb{Z}$ immediately follows by \thref{cor:TayloranalyticSob}.
First we establish a recursive estimate for the terms $a_l$. We begin with applying the criticality of the curve $\gamma$ and
\thref{lemma:iteratedEstimateMainTerm} to obtain
\[
\label{eq:recursiveEstimateA}
\refstepcounter{equation}\tag{\theequation}
\begin{split}
a_{l+3} \leq & \norm[H^{{\frac 5 2} + 3p - 7}]{ \gamma^{(l+3)} }\\
\leq & C \Phi\Bigl( l+1, n, r,K;\widehat{C}\Phi\left[1,n,\frac {r} {2\pi},K; \norm[H^{\frac 5 2}]{\gamma}, \norm[H^{\frac 5 2}]{\gamma'}\right] + \mu \norm[H^{\frac 5 2}]{\gamma'}, \ldots,\\
&\adjustedalignment{C \Phi\Bigl( l+1, n, r,K;\widehat{C}}{\widehat{C}}\Phi\left[l+2,n,\frac {r} {2\pi},K; \norm[H^{\frac 5 2}]{\gamma}, \ldots, \norm[H^{\frac 5 2}]{\gamma^{(l+2)}}\right] + \mu \norm[H^{\frac 5 2}]{\gamma^{(l+2)}}\Bigr)\\
&+ \mu \norm[H^{\frac 5 2}]{\gamma^{(l+2)}}\\
= & C \Phi\Bigl( l+1, n, r,K;\widehat{C}\Phi\left[1,n,\frac {r} {2\pi},K; a_0, a_1\right] + \mu a_1, \ldots,\\
&\adjustedalignment{C \Phi\Bigl( l+1, n, r,K;\widehat{C}}{\widehat{C}}\Phi\left[l+2,n,\frac {r} {2\pi},K; a_0, \ldots, a_{l+2}\right] + \mu a_{l+2}\Bigr) + \mu a_{l+2}.
\end{split}
\]
for positive constants $C, \widehat C, r, \mu$, all $l\ge 0$ and some $K \in \mathbb{N}$.
Next, we try to find $\tilde a_l$ that satisfy this recursive formula with equality and such that the first three coefficients $\tilde a_0, \tilde a_1$ and $\tilde a_2$ are greater than or equal to $a_0,a_1$ and $a_2$, respectively.
Define
\begin{alignat*}{2}
G: \mathbb{R}^{3n} \supseteq B_{\tilde \epsilon}(0) &\to \mathbb{R}^{3n},\\
G_i(z) &:= \widehat C \left( 1 + \frac {3n a_0 - \sum_{k=1}^{3n} z_k} {\frac {r} {2\pi}} \right)^{-2} z_i^{K-2} + \mu z_i,\\
F: \mathbb{R}^{3n} \supseteq B_{\tilde \epsilon}(0) &\to \mathbb{R}^{3n},\\
F_i(y) &:= C \left( 1 + \frac {3n DG_1((a_0,\ldots,a_0)^\intercal)(a_1, \ldots,a_1)^\intercal - \sum_{k=1}^{3n} y_k} {r} \right)^{-2} y_i^{K-2},\\
\end{alignat*}
for $\tilde \epsilon := \min\{3n a_0 + \frac {r} {2\pi}, 3n DG_1((a_0,\ldots,a_0)^\intercal)(a_1, \ldots,a_1)^\intercal + r\}$.
These functions are clearly analytic and by the Cauchy-Kovalevsky \thref{thm:CK}, there is an analytic $c: (-\epsilon,\epsilon) \to \mathbb{R}^{3n}$ satisfying
\[
\refstepcounter{equation}\tag{\theequation}
\label{eq:majorantODE}
\left\{
\begin{split}
c''(t) &= \overline{C} \cdot \bigl(F(DG(c(t)) c'(t)) + \mu c'(t)\bigr),\\
c(0) &= (a_0,\ldots,a_0)^\intercal,\\
c'(0) &= (a_1,\ldots,a_1)^\intercal,
\end{split}
\right.
\]
where
\[
\overline{C} := \max\left\{1, \frac{a_2} {F_1(DG((a_0,\ldots,a_0)^\intercal)(a_1,\ldots,a_1)^\intercal)} \right\}.
\]
A little further along, we give a formula for all derivatives of $G(c(t))$ which implies that $\overline C$ is well-defined.
Since all entries of the right hand side in \eqref{eq:majorantODE} are identical given the initial values, we have that $c_i(t) = c_1(t)$ for all valid $i$ and $t$.
Setting $\tilde a_l := c_1^{(l)}(0)$, we have by definition that
\[
\label{eq:tildeAFirstThreeTerms}
\refstepcounter{equation}\tag{\theequation}
\tilde a_l \ge a_l \text{ for } l=0,1,2.
\]
It remains to show the recursion.
An application of Leibniz' rule and \eqref{universalpoly} yields
\[
\label{eq:derviativeGc}
\refstepcounter{equation}\tag{\theequation}
\begin{split}
&\restrict{\frac {\d^l} {\d t^l} G_i(c(t))}{t=0}\\
= &\widehat C \sum_{j_1+ \ldots +j_{K-1}=l} \binom{l}{j_1, \ldots, j_{K-1}} p_l^{(3n)} \Biggl( \Biggl\{ \Biggl( 1+
\underbrace{\frac {3n a_0 - \sum_{j=1}^{3n} c_j(0)} {\frac {r} {2\pi}}}_{=0}
\Biggr)^{-\abs{\alpha} - 2} \frac {(\abs{\alpha} +1)!} {\left(\frac {r} {2\pi}\right)^{\abs{\alpha}+1}} \Biggr\}_{\abs{\alpha} \le j_1},\\
& \phantom{\widehat C \sum_{j_1+ \ldots +j_{K-1}=l} \binom{l}{j_1, \ldots, j_{K-1}} p_l^{(3n)} \Biggl(} \left\{ c_j^{(k)}(0) \right\}_{\substack{k=1,\ldots,j_1\\j=1,\ldots,3n}} \Biggr)
\cdot \prod_{k=2}^{K-1} c_i^{(j_k)}(0) + \mu c_i^{(l)}(0)\\
= & \widehat C \Phi\left(l,n,\frac {r} {2\pi},K;c_1(0),\ldots,c_1^{(l)}(0)\right) + \mu c_1^{(l)}(0)
=\widehat C \Phi\left(l,n,\frac {r} {2\pi},K;\tilde a_0,\ldots, \tilde a_l\right) + \mu \tilde a_l.
\end{split}
\]
Note that this corresponds to the inner $\Phi$ terms in \eqref{eq:recursiveEstimateA}.
For $l=1$ this is a polynomial in $\tilde a_0,\tilde a_1$ with positive coefficients, so as long as both $\tilde a_0 = a_0$ and $\tilde a_1 = a_1$ are positive, so is each component of $DG((a_0,\ldots,a_0)^\intercal)(a_1,\ldots,a_1)^\intercal$ and consequently $F_1$ applied to that.
Since $\gamma$ is not constant, neither $\gamma$ nor $\gamma'$ are $0 \in H^{{\frac 5 2}}(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ and so $a_0$ and $a_1$ have to be positive, making $\overline C$ well-defined.
Analogously to \eqref{eq:derviativeGc}, taking advantage of the identity for all components in $G$,
\[
\label{eq:derivativeRHS}
\begin{split}
&\tilde a_{l+3}
= c_1^{(l+3)}(0)
= \overline C \restrict{\frac {\d^{l+1}} {\d t^{l+1}} F_1\left(\frac {\d}{\d t} G(c(t))\right)}{t=0} + \overline C\mu c^{(l+2)}_1(0)\\
=& \overline C C \sum_{j_1 + \ldots + j_{K-1} = l+1} \binom{l+1}{j_1, \ldots, j_{K-1}}\\
&\quad p_l^{(3n)} \Biggl( \Biggl\{ \Biggl( 1+
\underbrace{\frac {3n DG_1(c(0))\dot c(0) - \sum_{j=1}^{3n} \restrict{\frac {\d} {\d t} G_j(c(t))}{t=0}} {r}}_{=0}
\Biggr)^{-\abs{\alpha} - 2} \frac {(\abs{\alpha} +1)!} {r^{\abs{\alpha}+1}} \Biggr\}_{\abs{\alpha} \le j_1},\\
&\phantom{\quad p_l^{(3n)} \Biggl(}\left\{ \restrict{\frac{\d^{k+1}} {\d t^{k+1}} G_j(c(t))}{t=0} \right\}_{\substack{k=1,\ldots,j_1\\j=1,\ldots,3n}} \Biggr) \cdot \prod_{k=2}^{K-1} \restrict{\frac {\d^{j_k+1}}{\d t^{j_k+1}} G_1(c(t))}{t=0} + \overline{C} \mu \tilde a_{l+2}\\
=&\overline{C} C \Phi\left(l+1,n,r,K; \restrict{\frac {\d}{\d t} G_1(c(t))}{t=0},\ldots,\restrict{\frac {\d^{l+2}}{\d t^{l+2}} G_1(c(t))}{t=0}\right) + \overline{C} \mu \tilde a_{l+2}\\
\ge& C \Phi\left(l+1,n,r,K; \restrict{\frac {\d}{\d t} G_1(c(t))}{t=0},\ldots,\restrict{\frac {\d^{l+2}}{\d t^{l+2}} G_1(c(t))}{t=0}\right) + \mu \tilde a_{l+2},
\end{split}
\]
hence we conclude together with \eqref{eq:derviativeGc} that
\[
\refstepcounter{equation}\tag{\theequation}
\label{eq:recursiveequation}
\begin{split}
& \tilde a_{l+3} \geq C \Phi\Bigl( l+1, n, r,K;\widehat{C}\Phi\left[1,n,\frac {r} {2\pi},K; \tilde a_0, \tilde a_1\right] + \mu \tilde a_1, \ldots,\\
&\adjustedalignment{C \Phi\Bigl( l+1, n, r,K;\widehat{C}}{\widehat{C}}\Phi\left[l+2,n,\frac {r} {2\pi},K; \tilde a_0, \ldots, \tilde a_{l+2}\right] + \mu \tilde a_{l+2}\Bigr) + \mu \tilde a_{l+2}.
\end{split}
\]
Using the fact that $\Phi$ is nondecreasing in $x_0,\ldots,x_l$ (see \eqref{eq:definitionPhi})
, we obtain
by comparing \eqref{eq:recursiveEstimateA} and \eqref{eq:recursiveequation} inductively with initial values \eqref {eq:tildeAFirstThreeTerms}, that
\[
\tilde a_{l} \geq a_{l}
\]
for all $l \in \mathbb{N}_0$.
By the analyticity of $c_1$ in zero and \thref{cor:TayloranalyticSob},
we have that there exist positive constants $C_\gamma$ and $r_\gamma$ such that
\[
\norm[H^{\frac 5 2}]{ \gamma^{(l)}} = a_l \le \tilde a_l = c_1^{(l)} (0) \le C_\gamma \frac {l!} {r_\gamma^l}
\]
for all $l\in\mathbb{N}_0$ and thus, $\gamma$ is analytic.
\end{proof}
\label{rem:smallerRegularityGain}
Note that the method presented here ought to work with even smaller differences between the orders of differentiability of the main and the remainder terms. If the regularity gain in one step is $\frac 1 m$ we would expect an analogue of \eqref{eq:recursiveEstimateA} to hold with roughly $m$ layers of nested $\Phi$-terms attainable via an analogue of \thref{lemma:iteratedEstimateMainTerm}.
This approach would possibly lead to an $m$-th order ODE in place of \eqref{eq:majorantODE}.
\section*{Acknowledgements}
Daniel Steenebrügge acknowledges support by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – project number 320021702/GRK2326 – \emph{Energy, Entropy, and Dissipative Dynamics (EDDy)} and Nicole Vorderobermeier by the Austrian Science Fund (FWF), Grant P29487. Funding from RWTH Aachen University for Nicole Vorderobermeier's research visit is also gratefully acknowledged.
In addition, the authors would like to sincerely thank Simon Blatt for his constant support and indispensable discussions throughout the development of this research paper, Heiko von der Mosel for his generous support and hospitality as well as fruitful discussions, from which Section \thref{sec:Consequences} emerged, and Philipp Reiter for having a sympathetic ear for the authors and giving important hints to deal with some filthy estimates in Section \thref{subsec:FractionalLeibnizRule}.
\section{A few simple consequences concerning critical knots}
\label{sec:Consequences}
The main result of the paper states that critical points of the generalized integral Menger curvature $\Mpq[p,2]$ are not only smooth, which was known before, cf. \cite[Theorem\,4]{BlattReiter:2015:Menger}, but also analytic. Recalling and employing special properties for analytic functions, we hence decude a few simple corollaries for critical knots of $\Mpq[p,2]$ and some other knot energies.
The following statements hold for any functional $\mathcal{E}: C^0(\mathbb{R}/\mathbb{Z},\mathbb{R}^3) \rightarrow (0,\infty]$ whose (arc-length parametrized) critical points with respect to fixed length are analytic. As a consequence, the results also apply to the generalized integral Menger curvature $\Mpq[p,2]$ and some of O'Hara's knot energies, cf. \cite{BlattVorderobermeier:2019:CriticalMoebius,Vorderobermeier:2019:CriticalOHara}.
Central to these consequences is a real version of the identity theorem. It is a well-known result for analytic functions which easily carries over to the periodic case:
\begin{proposition}
\label{propositionRealIdentityTheorem}
If two analytic functions on $\mathbb{R}/\mathbb{Z}$ are identical at infinitely many distinct points, they are identical everywhere.
\end{proposition}
\begin{proof}
Let $\gamma,\eta : \mathbb{R}/\mathbb{Z} \to \mathbb{R}^3$ be analytic and without loss of generality $t_1<t_2< \ldots \in [0,1)$ such that $\gamma(t_i)=\eta(t_i)$ for all $i \in \mathbb{N}$.
Define $g,h: (-1,2) \to \mathbb{R}^3$ as the restrictions of $\gamma$ and $\eta$ to $(-1,2)$ which are clearly also analytic.
As $t_i \in [0,1]$ for all $i \in \mathbb{N}$, we have a subsequence converging to $t_0 \in [0,1]\subseteq(-1,2)$, so by \cite[Corollary~1.2.7]{KrantzParks:2002:Aprimerofrealanalyticfunctions}, $g=h$ and thus $\gamma = \eta$.
\end{proof}
\begin{lemma}
\label{lemma:zerosAnalyticFunctions}
Let $\gamma \colon \mathbb{R}/\mathbb{Z} \to \mathbb{R}^3$ and $F: \mathbb{R}^3 \to \mathbb{R}$ both be analytic. Then, $F\circ \gamma$ either has finitely many roots or is uniformly zero.
\end{lemma}
\begin{proof}
$F \circ \gamma \colon \mathbb{R}/\mathbb{Z} \to \mathbb{R}$ is analytic as a composition of analytic functions (see e.g. \cite[Proposition~2.2.8]{KrantzParks:2002:Aprimerofrealanalyticfunctions}).
Using the zero function in \thref{propositionRealIdentityTheorem}, $F\circ \gamma$ either is uniformly zero or has only finitely many roots.
\end{proof}
\begin{proposition}
\label{corollary:UnknotEllipsoidHyperplane}
Let $\gamma \in C^1(\mathbb{R}/\mathbb{Z},\mathbb{R}^3)$ be a simple curve parametrized by arc-length which is a critical point of $\mathcal{E}$.
Then, it is the unknot or it has only finitely many intersections with any hyperplane or sphere in $\mathbb{R}^3$.
In particular, $\gamma$ cannot have straight segments.
\end{proposition}
\begin{proof}
First of all, let us consider any hyperplane with normal vector $n\in\mathbb{R}^3$ and position vector $x_0\in\mathbb{R}^3$, described by the equation $F(x)=0$, where $F:\mathbb{R}^3\rightarrow\mathbb{R}$, $F(x)= \langle n,x-x_0\rangle$. Furthermore, let $\gamma:\mathbb{R}/\mathbb{Z}\rightarrow \mathbb{R}^3$ be a critical point of $\mathcal{E}$ that is simple and parametrized by arc-length. We then observe that both $F$ and $\gamma$ are analytic, which is why we can apply \thref{lemma:zerosAnalyticFunctions} and observe that the curve $\gamma$ either intersects the hyperplane finitely many times or is entirely contained therein. In the latter case we obtain the unknot, which is the only simple planar knot, cf. \cite[p.~5]{CrowellFox:1977:Introductiontoknottheory} or \cite[Chapter~VI,~18.]{Newman:1951:ElementsOfTopology}.
If $\gamma$ contains a straight segment, this segment is in particular contained in two distinct hyperplanes.
Thus, the whole curve has to be contained in both hyperplanes and therefore has to be a straight line which is impossible for a closed $C^1$-curve of finite length.
We then turn our focus to any sphere with center $x_0\in\mathbb{R}^3$ and radius $r>0$, described by the equation $F(x) =0$, where $F:\mathbb{R}^3 \rightarrow \mathbb{R}$, $F(x) = |x-x_0|^2 -r^2$.
Therefore, similarly to the hyperplane, \thref{lemma:zerosAnalyticFunctions} gives us that any critical analytic knot $\gamma$ of $\mathcal{E}$ either intersects the sphere finitely many times or is entirely contained therein.
In the second case, we would intuitively expect to obtain the unknot by applying a stereographic projection from the sphere to the plane and then argue as above. To make this rigorous, we need to construct an ambient isotopy, cf. \cite[Chapter~8.1]{Hirsch:1976:DifferentialTopology}, to avoid any kind of self-intersection or pull-tight phenomena during the projection process.
To simplify the upcoming computations, we assume that the sphere is the unit sphere centered in the origin and that the knot $\gamma$ on the sphere does not intersect the north pole of the sphere. We omit the proof that translations and dilations of the sphere do not change the knot type of the spherical knot and remark that stereographic projections can be done from any other point on the sphere as well. The existence of such point on the sphere, which the knot $\gamma$ does not intersect, is ensured due to the fact that $\gamma$ is rectifiable. Furthermore, note that since the curve $\gamma$ is continuous, there exists an $r>0$ such that $\gamma$ does not intersect a ball around the north pole with radius $r$.
We begin with constructing a $C^1$-isotopy from $\gamma$ to the stereographic projection of $\gamma$ from the north pole onto the plane $z=0$ denoted by $\tilde \gamma$. Note that $\tilde \gamma$ is indeed a $C^1$-embedding: As $\gamma\in C^\omega(\mathbb{R}/\mathbb{Z},\mathbb{R}^3)$ is a diffeomorphism onto its image and the stereographic projection $P:\mathbb S^2\setminus{(0,0,1)} \rightarrow \mathbb{R}^2$ given by $P(x,y,z)= (\frac{x}{1-z},\frac{y}{1-z})$ with inverse function $P^{-1}(u,v) = \tfrac{1}{u^2+v^2 +1}(2u,2v,u^2+v^2 -1)$ is a diffeomorphism from the sphere without the north pole to the plane $z=0$, also $\tilde \gamma \in C^1(\mathbb{R}/\mathbb{Z},\mathbb{R}^3)$ is diffeomorphic onto its image. We then define the map $h:\mathbb{R}/\mathbb{Z} \times [0,1]\rightarrow \mathbb{R}^3$ by
\[
h_t(w):=(1-t)\gamma(w) + t \tilde \gamma (w),
\]
which is continuously differentiable in time and space.
To see that $h$ is a $C^1$-isotopy (cf.\ \cite[Section~8.1]{Hirsch:1976:DifferentialTopology}), we construct a continuously differentiable left-inverse of each $h_t$ and thus show that $h_t$ is a $C^1$-embedding.
First note that by the nature of stereographic projection, $h_t(w)$ lies on the unique straight line passing through both the north pole $N$ and $\gamma(w)$.
Since $\gamma$ lies on the sphere and does not contain $N$, this line does not contain any other points of $\gamma$ or $h(\tilde w,[0,1])$ for any $\tilde w \ne w$.
This means that for any $x \in h_t(\mathbb{R}/\mathbb{Z})$, we may obtain $w=h_t^{-1}(x)$ as $\gamma^{-1}(y)$, where $y=\lambda N + (1-\lambda)x$ with $\abs{y}=1$.
Solving for $\lambda \in \mathbb{R}$ yields either $\lambda=0$, which we may safely ignore since it corresponds to $y$ being the north pole, or $\lambda=\frac {\abs{x}^2-1} {\abs{x-N}^2}$.
As $x$ has to stay away from the north pole, we found a $C^1$ right-inverse of $h_t$ and so, $h$ is a $C^1$-embedding.
Obviously, also
\[
h(\cdot,0) = \gamma \quad \textnormal{and} \quad h(\cdot,1) = \tilde \gamma
\]
hold. Now by \cite[Section~8.1,~Exercise~4]{Hirsch:1976:DifferentialTopology} or \cite[Theorem~1.2]{Blatt:2009:NoteOnIsotopies}, we can extend the $C^1$-isotopy between $\gamma$ and $\tilde \gamma$ to an ambient isotopy of $\mathbb{R}^3$ and therefore obtain that the knot type is preserved throughout the stereographic projection.
\end{proof}
As a direct consequence of \thref{lemma:zerosAnalyticFunctions} and \thref{corollary:UnknotEllipsoidHyperplane}, we obtain the following.
\begin{corollary}
Let $\gamma \in C^1(\mathbb{R}/\mathbb{Z},\mathbb{R}^3)$ be a simple curve parametrized by arc-length which is a critical point of $\mathcal{E}$ and let $F:\mathbb{R}^3 \rightarrow \mathbb{R}$ be an analytic map that describes an implicit surface $F(x,y,z)=0$.
Then, the curve $\gamma$ either is entirely contained in the implicit surface $F(x,y,z)=0$ or has only finitely many intersections with it. In particular, if the given surface is ambient isotopic to a sphere or a hyperplane, the curve $\gamma$ is the unknot or has finitely many intersections with the surface.
\end{corollary}
\begin{corollary}
Let $\gamma,\eta$ be critical points of $\mathcal{E}$ which are parametrized by arc-length.
As soon as
\[
\gamma(t)=A\eta(t) + x_0
\]
for fixed $A\in \mathbb{R}^{3\times 3}$, $x_0 \in \mathbb{R}^3$, and infinitely many distinct $t \in [0,1)$, we have $\gamma = A \eta + x_0$.
\end{corollary}
\begin{proof}
Note that by assumption, $\gamma$ and $\eta$ are analytic.
Then, the statement follows from the fact that $A\eta + x_0$ is analytic as well and applying \thref{propositionRealIdentityTheorem}.
\end{proof}
\section{Estimate of the main term $Q^{(p)}$}
\label{sec:EstimateMainTerm}
Next we investigate the order of the main term $Q^{(p)}$, for which we need to examine its $L^2$-representation first.
\begin{lemma}
\label{lemma:L2MainTermViaFourier}
Let $p \in (\tfrac 7 3,\tfrac 8 3)$, $s \ge 0$ and $\gamma \in H^{3p-4+s}(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ be parametrized by arc-length.
Then, there is $\ensuremath{\widetilde Q^{(p)}}(\gamma) \in H^s(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ such that for all $h \in H^{\frac 3 2 p -2}(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$,
\[
\ensuremath{Q^{(p)}}(\gamma,h) = \int_{\mathbb{R}/\mathbb{Z}} \left<\ensuremath{\widetilde Q^{(p)}}(\gamma)(u), h(u)\right> \d u.
\]
In particular, $\ensuremath{\widetilde Q^{(p)}}$ maps $C^\infty(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ into itself.
The Fourier coefficients of $\ensuremath{\widetilde Q^{(p)}}$, for all $k\in \mathbb{Z}$, are given by
\begin{align*}\label{eq:fouriercoeffQ}
\widehat{\ensuremath{\widetilde Q^{(p)}}(\gamma)}(k) = q_k^{(p)} \abs{k}^{3p-4} \hat{\gamma} (k).
\end{align*}
The constants $q_k^{(p)}$ are bounded independently of $n$, and satisfy
\[
0 \leq q_k^{(p)} = c + o(1)
\]
as $\abs{k} \rightarrow \infty$ for $c>0$.
\end{lemma}
\begin{proof}
By \cite[Proposition 4.1]{BlattReiter:2015:Menger}, we have
\[
\ensuremath{Q^{(p)}}(\gamma,h) = \sum_{k \in \mathbb{Z}}\rho_k \left< \hat \gamma(k),\hat h(k)\right>_{\mathbb{C}^n}, \text{ where } \rho_k = c\abs{k}^{3p-4} + o(\abs{k}^{3p-4}) \text{ as } \abs{k} \to \infty,
\]
for any $h \in H^{\frac 3 2 p -2}(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$. In particular, the $\rho_k$ are independent of $n$, cf.\ their definition in the proof of \cite[Proposition 4.1]{BlattReiter:2015:Menger}.
Setting $q_k^{(p)} := \abs{k}^{4-3p} \rho_k$, which implies that the $q_k^{(p)}$ are also independent of $n$, we obtain
\[
\begin{split}
(1+k^2)^{\frac s 2}\abs[\big]{q_k^{(p)}\abs{k}^{3p-4}\hat \gamma(k)}
&= (1+k^2)^{\frac s 2}\abs{c+o(1)}\abs{k}^{3p-4} \abs{\hat \gamma(k)}\\
&\le (1+k^2)^{\frac s 2} \tilde c \abs{k}^{3p-4} \abs{\hat \gamma(k)}
\le \tilde c (1+k^2)^{\frac {3p-4+s} {2}} \abs{\hat \gamma(k)} \text{ as } \abs{k} \to \infty
\end{split}
\]
for some $\tilde c>0$.
The right hand side, taken as a sequence in $k \in \mathbb{Z}$, is in $\ell^2(\mathbb{Z})$ because $\gamma \in H^{3p-4+s}(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$, see \cite[Chapter~4, (3.8)]{Taylor:1996:PartialDifferentialEquationsa},
and so $\ensuremath{\widetilde Q^{(p)}}(\gamma)$, the inverse Fourier transform of $q_k^{(p)}\abs{k}^{3p-4}\hat \gamma(k)$, is well-defined and in $H^s(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$.
All $q_k^{(p)}\abs{k}^{3p-4}$ are real as $\ensuremath{Q^{(p)}}$ maps real functions to real numbers and $\overline{\hat \gamma(k)} = \hat \gamma (-k)$ due to the real image of $\gamma$.
Thus, $\ensuremath{\widetilde Q^{(p)}}(\gamma)(u) = \sum_{k\in \mathbb{Z}} q_k^{(p)}\abs{k}^{3p-4} \hat{\gamma}(k) e^{-2\pi i k u} = \hat \gamma(0) + \sum_{k \in \mathbb{N}} q_k^{(p)}\abs{k}^{3p-4} \Re(\hat \gamma(k) e^{-2 \pi i k u})$ for almost all $u \in \mathbb{R}/\mathbb{Z}$ meaning that $\ensuremath{\widetilde Q^{(p)}}(\gamma)$ maps into the real numbers.
By Parseval's relation \cite[Proposition~3.2.7~(3)]{Grafakos:2014:ClassicalFourierAnalysis}, we have
\[
\int_{\mathbb{R}/\mathbb{Z}} \left<\ensuremath{\widetilde Q^{(p)}}(\gamma)(u),h(u)\right> \d u
= \sum_{k\in \mathbb{Z}} \left< q_k^{(p)}\abs{k}^{3p-4}\hat \gamma(k), \hat h(k)\right>_{\mathbb{C}^n}
= \sum_{k \in \mathbb{Z}}\rho_k \left< \hat \gamma(k),\hat h(k)\right>_{\mathbb{C}^n}
= \ensuremath{Q^{(p)}}(\gamma,h).
\]
\end{proof}
\begin{remark}
Assuming $C^5$-regularity of $\gamma$, one may explicitly calculate
\begin{align*}
\ensuremath{\widetilde Q^{(p)}}(\gamma)(u) :=& \iint_D \abs{v-w}^{-p} \abs{v}^{2-p} \abs{w}^{2-p} \cdot \Biggl(
\frac {\frac{\difference{-w,0} \gamma} {-w} - \frac {\difference{w,0} \gamma} {w}} {w}
+ \frac {\frac {\difference{-v,0} \gamma} {-v} - \frac {\difference{v,0} \gamma} {v}} {v}\\
&+ \frac {\frac {\difference{v,0} \gamma} {v} - \frac {\difference{v-w,-w} \gamma} {v}} {w}
+ \frac {\frac {\difference{w,0} \gamma} {w} - \frac {\difference{w-v,-v} \gamma} {w}} {v}
\Biggr)
\d w \d v.
\end{align*}
The proof works via first restricting to the case $\abs{v},\abs{w}>\epsilon$ to obtain the formula and then showing existence for $\epsilon = 0$ by means of several Taylor approximations.
Since this method is quite technical and long, we choose to omit it here.
\end{remark}
From the previous \thref{lemma:L2MainTermViaFourier}, we deduce the following indispensable corollary regarding the order of the main term $Q^{(p)}$.
\begin{corollary} \label{cor:Ql}
Given real numbers $p\in (\tfrac 73, \tfrac 83)$ and $m\geq 0$, there exists a constant $ \widetilde{C}=\widetilde{C}(p)>0$ such that for all curves $\gamma \in C^\infty(\mathbb{R}/\mathbb{Z}, \mathbb{R}^n)$ and integers $l\geq 0$
\begin{align*}
\|\gamma^{(l+3)} \|_{H^{m+3p-7}} \leq \widetilde{C} \|\partial^{l} \ensuremath{\widetilde Q^{(p)}}(\gamma)\|_{H^m}
\end{align*}
holds.
\end{corollary}
\begin{proof}
By applying \thref{lemma:L2MainTermViaFourier}, well-known properties of Fourier coefficients and the elementary estimate $(1+k^2)^{3p-7} \leq (2|k|)^{2(3p-7)}$ for any $k\in \mathbb{Z} \setminus 0$, we get
\begin{align*}
\|\partial^{l}\ensuremath{\widetilde Q^{(p)}} (\gamma) \|_{H^m}^2 &= \sum_{k\in \mathbb{Z}} (1+|k|^2)^m|\widehat{\partial^{l}\ensuremath{\widetilde Q^{(p)}} (\gamma)}(k)|^2 \\
&= \sum_{k\in \mathbb{Z}} (1+|k|^2)^m (2\pi| k|)^{2l} |\widehat{\ensuremath{\widetilde Q^{(p)}} (\gamma)}(k)|^2 \\
&= \sum_{k\in \mathbb{Z}} (1+|k|^2)^m (2\pi| k|)^{2l} (q_k^{(p)})^2 |k|^{2(3p-4)} |\hat{\gamma} (k)|^2\\
&\geq \inf_{k\in\mathbb{Z}\setminus 0} \{(q_k^{(p)})^2 (2\pi)^{-6} 2^{2(7-3p)} \} \sum_{k \in \mathbb{Z}} (1+|k|^2)^{m+3p-7} (2\pi| k|)^{2(l+3)} |\hat{\gamma} (k)|^2 \\
&= \widetilde{C}^{-2} \sum_{k \in \mathbb{Z}} (1+|k|^2)^{m+3p-7} |\widehat{\gamma^{(l+3)} } (k)|^2 \\
&= \widetilde{C}^{-2} \|\gamma^{(l+3)} \|_{H^{m+3p-7}}^2,
\end{align*}
where $\widetilde{C}:=\inf_{k\in\mathbb{Z}\setminus{0}} \{\abs{q_k^{(p)}} (2\pi)^{-3} 2^{7-3p}\}^{-1}$ is a positive constant only depending on $p$.
In particular, independence of $n$ follows from \thref{lemma:L2MainTermViaFourier}.
\end{proof}
\section{Decomposition of the first variation}
\label{sec:Decomposition}
Hermes was the first to derive Gateaux differentiability and a formula for the first variation of the integral Menger curvature $\mathcal{M}_p = 2^p \textnormal{intM}^{(p,p)} $ for $p\geq 2$ in his PhD thesis \cite[Theorem\,2.33, Remark\,2.35]{Hermes:2012:Menger}. In \cite[Theorem\,3]{BlattReiter:2015:Menger}, Blatt and Reiter extended this result to the generalized integral Menger curvature $\textnormal{intM}^{(p,q)}$ for the sub-critical range $p\in (\tfrac 23 q+1, q+\tfrac 23)$ and $q>1$. However, they approached the problem differently to Hermes by taking advantage of studying a certain subdomain of integration as well as using the newly discovered characterization of energy spaces developed in \cite[Theorem\,1.1]{Blatt:2013:NoteMenger} and \cite[Theorem\,1]{BlattReiter:2015:Menger}. The latter states that injective curves in $C^1(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ parametrized by arc-length have finite generalized integral Menger curvature iff they belong to the fractional Sobolev space $W^{(3p-2)/q-1,q} (\mathbb{R}/\mathbb{Z}, \mathbb{R}^n)$ \cite[Theorem 1]{BlattReiter:2015:Menger}.
To summarize Blatt and Reiter's findings regarding the first variation of the generalized integral Menger curvature $\textnormal{intM}^{(p,q)}$, cf. \cite[Theorem 3]{BlattReiter:2015:Menger}, they showed that it is continuously differentiable on the subspace of all regular embedded $W^{(3p-2)/q-1,q}$-curves. Furthermore, they gave an explicit formula for the first variation of $\textnormal{intM}^{(p,q)}$ at any arc-length parametrized embedded $\gamma \in W^{(3p-2)/q-1,q}(\mathbb{R}/\mathbb{Z}, \mathbb{R}^n)$ in direction $h \in W^{(3p-2)/q-1,q}(\mathbb{R}/\mathbb{Z}, \mathbb{R}^n)$.
In the following we will restrict ourselves to the non-degenerate, sub-critical range of generalized integral Menger curvature, i.e. $q=2$ and $p\in(\tfrac 73, \tfrac 83)$. In \cite{BlattReiter:2015:Menger} it has turned out to be helpful in that case to decompose its first variation into a term of highest order and lower order terms in order to study the regularity of critical points.
Hence we start with recapitulating a decomposition of the first variation introduced in \cite[(4.2) and Lemma 4.2]{BlattReiter:2015:Menger}.
Note that we abbreviate
\[
\difference{x,y} f := f(u+x) - f(u+y)
\]
and
\[
D:= \set*[w \leq 1 + 2v, v \geq -1 + 2w]{(v,w) \in \left(-\tfrac 1 2,0\right) \times \left(0,\tfrac 1 2\right)}.
\]
\sloppy
We observe for any arc-length parametrized embedded $\gamma \in W^{\frac 32 p-2,2}(\mathbb{R}/\mathbb{Z}, \mathbb{R}^n)$ and $h \in W^{\frac 32 p-2,2}(\mathbb{R}/\mathbb{Z}, \mathbb{R}^n)$ that
\[
\label{eq:firstVariation}
\refstepcounter{equation}\tag{\theequation}
\delta \Mpq[p,2] (\gamma,h) = 12(\ensuremath{Q^{(p)}}(\gamma,h) + R^{(p)} (\gamma,h)),
\]
\fussy
where the main term is given by
\[
\ensuremath{Q^{(p)}}(\gamma,h) := \iiint_{\mathbb{R}/\mathbb{Z} \times D} \frac {\left< \frac {\difference{w,0} \gamma} {w} - \frac {\difference{v,0} \gamma} {v}, \frac {\difference{w,0} h} {w} - \frac {\difference{v,0} h} {v} \right>} {\abs{v-w}^p \abs{v}^{p-2} \abs{w}^{p-2}} \d w \d v \d u.
\]
The remainder term can be expressed as following.
\begin{lemma}[\protecting{\cite[Lemma 4.2]{BlattReiter:2015:Menger}}]
\label{lemma:formRemainder}
The term $R^{(p)} (\gamma,h)$ is a finite sum of terms of the form
\[
\iiint\limits_{\mathbb R / \mathbb Z \times D} \idotsint\limits_{[0,1]^K} g^p(u,v,w;s_{1},\dots,s_{K-2}) \otimes h'(u+s_{K-1}v + s_K w)
\d\theta_{1}\cdots\d\theta_{K}\d v \d w \d u
\refstepcounter{equation}\tag{\theequation}
\label{lemma:formRemainder:formula}
\]
where $\G : (0,\infty)^3 \rightarrow \mathbb R$ is an analytic function, $\otimes$ may denote any kind of product
structure, such as cross product, dot product, scalar or matrix multiplication, $s_{j}\in\set{0,\theta_{j}}$ for $j=1,\dots,K$,
\begin{multline*}
g^p(u,v,w;s_{1},\dots,s_{K-2})=
\G \left(\frac{|\T 0w|}{|w|}, \frac{|\T 0v|}{|v|}, \frac{|\T vw|}{|v-w|}\right)
\Gamma(u,v,w,s_{1},s_{2})\cdot{}\\
{}\cdot\left( \bigotimes_{i=3}^{K_1} {\gamma}'(u+ s_i v) \right) \otimes \left(\bigotimes_{j=K_1}^{K_2} {\gamma}'(u + s_i w) \right) \otimes \left(\bigotimes_{j=K_2}^{K-2} {\gamma}'(u+v + s_i (w-v)) \right),
\end{multline*}
and
$\Gamma(u,v,w,s_{1},s_{2})$ is a term of one of the four types
\begin{align*}
&\frac{\br{\gamma'(u+s_1 w) - \gamma'(u+s_1 v)}\otimes\br{\gamma'(u+s_2 w) - \gamma'(u+s_2 v)}}{|v|^{p-2} |w|^{p-2} |v-w|^{p}}, \\
&\frac{|\gamma'(u+s_1 w) - \gamma'(u+s_2 w)|^2}{|v|^{p-2} |w|^{p-2} |v-w|^{p}}, \\
& \frac{|\gamma'(u+s_1 v) - \gamma'(u+s_2 v)|^2}{|v|^{p-2} |w|^{p-2} |v-w|^{p}}, \\
& \frac{|\gamma'(u+v + s_1 (w-v)) - \gamma'(u+v+s_2 (w-v))|^2}{|v|^{p-2} |w|^{p-2} |v-w|^{p}}.
\end{align*}
\end{lemma}
\section{Properties of analytic functions}
In this section we aim to recall some statements related to analytic functions, which are used in the paper. For further information on analytic functions we refer the reader for instance to \cite{KrantzParks:2002:Aprimerofrealanalyticfunctions}.
First we note that analytic functions can be characterized as follows, cf.\ for example \cite[Corollary~3.2]{Vorderobermeier:2019:CriticalOHara}.
\begin{proposition} \label{cor:TayloranalyticSob}
Let $f\in C^\infty (\mathbb R / \mathbb Z,\mathbb{R}^n)$ and $s>0$. Then the function $f$ is analytic on $\mathbb{R}/\mathbb{Z}$ if there are positive constants $r$ and $C$ such that
\begin{align*}
\|f^{(k)}\|_{H^{1+s}} \leq C \frac{k!}{r^{k}}
\end{align*}
holds for all integers $k \geq 0$.
\end{proposition}
This can be seen by using \cite[Proposition~1.2.12]{KrantzParks:2002:Aprimerofrealanalyticfunctions} together with a standard covering argument, Morrey's inequality (see e.g. \cite[Paragraph~4.16]{AdamsFournier:2003:SobolevSpaces}), the equivalence of the $W^{1,2}$- and the $H^1$-norm on $\mathbb R / \mathbb Z$ (cf.~\cite[Lem.~1.2]{Reiter:2012:RepulsiveKnotEnergies} or \cite[7.62]{AdamsFournier:2003:SobolevSpaces}) as well as the embedding $H^s\subseteq H^t$ for any $t < s$ (cf.~\cite[Chapter~4,~Proposition~3.4]{Taylor:1996:PartialDifferentialEquationsa}).
In addition, we need a special case of the Cauchy-Kovalevsky theorem, which can be derived for example from \cite[(1.25)]{Folland:1995}.
\begin{theorem}[Cauchy-Kovalevsky -- ODE case]\label{thm:CK}
Suppose the function $G: \mathbb{R}^{2n} \rightarrow \mathbb{R}^n$ is real analytic around $(c_0,c_1)$ for some $c_0,c_1 \in\mathbb{R}^n$, and the function $f \in C^\infty((-\varepsilon, \varepsilon),\mathbb{R}^n)$ of the form $f(x)= (f_1(x),\ldots,f_n(x))$, for any $ \varepsilon > 0$, is a solution of the initial value problem
\begin{align*}
f''(x) & = G(f(x), f'(x)) \textnormal{ for } x \in (-\varepsilon,\varepsilon) \textnormal{ with}\\
f(0)& =c_0, \textnormal{ and}\\
f'(0) & = c_1.
\end{align*}
Then the function $f$ is real analytic around $0$.
\end{theorem}
The statement can be proven by the \emph{method of majorants}, e.g.\ as explained in the proof of \cite[2.4.1]{KrantzParks:2002:Aprimerofrealanalyticfunctions}. At this point we remark that the strategy of proof for the main statement, \thref{thm:main}, is motivated from this method.
\section{Faà di Bruno's formula}
Faà di Bruno's formula generalizes the chain rule to higher derivatives. In particular, the $k$-th derivative of the composition of two functions $f, g\in C^k (\mathbb{R},\mathbb{R})$ can be written as
\begin{align*}
&\left(\tfrac{d}{dt}\right)^k g(f(t)) = \\
&\sum_{\substack{m_1+2m_2+ \cdots km_k=k, \\ m_1,\ldots,m_k\in\mathbb{N}_0}} \frac{k!}{m_1!1!^{m_1}m_2!2!^{m_2}\ldots m_k!k!^{m_k}} g^{(m_1+\cdots +m_k)}(f(t))\prod_{j=1}^{k} (f^{(j)} (t))^{m_j}.
\end{align*}
R. Mishkov generalized Faà di Bruno's formula to the multivariate case \cite{Mishkov:2000:GeneralizationFormulaFaa}.
The following lemma and notation via the universal polynomial are taken directly from \cite[Section~2.3]{BlattVorderobermeier:2019:CriticalMoebius}.
\begin{lemma}\label{lem:FaadiBruno}
Let $U\subseteq\mathbb{R}$ be a neighborhood of $0$ and $f\in C^k(U, \mathbb{R}^n)$, $n\in\mathbb{N}_0$. Moreover, let $V\subseteq\mathbb{R}^n$ be a neighborhood of $f(0)$ such that $f(U)\subseteq V$ and $g\in C^k(V, \mathbb{R})$. Then we have for any $k\in\mathbb{N}$ and $x\in U$
\begin{align*}
\left(\!\tfrac{d}{dx}\!\right)^k g(f(x)) \!=\! \sum_0 \sum_1 \cdots \sum_k \frac{k!}{\prod_{i=1}^k (i!)^{r_i} \prod_{i=1}^k\prod_{j=1}^nq_{ij}! } (\partial^\alpha g) (f(x)) \prod_{i=1}^k ( f_1^{(i)}(x))^{q_{i1}} \cdots ( f_n^{(i)}(x))^{q_{in}}
\end{align*}
where the sums are over all non-negative integer solutions of the following equations
\begin{align*}
& \sum_0 : r_1 + 2r_2 + \cdots + k r_k = k, \\
& \sum_1 : q_{11} + q_{12} + \cdots + q_{1n} = r_1 \\
& \ \ \vdots \\
& \sum_k : q_{k1} + q_{k2} + \cdots + q_{kn} = r_k
\end{align*}
and $\alpha := (\alpha_1,\ldots,\alpha_n)$ with $\alpha_j := q_{1j} + q_{2j} + \cdots + q_{kj}$ for $1\leq j \leq n$.
\end{lemma}
In the following we seldom need the precise form of Faà di Bruno's formula in the multivariate case but only the fact that there is a universal polynomial $p_k^{(n)}$ with non-negative coefficients independent of $f,g$ such that
\begin{align}\label{universalpoly}
\left(\tfrac{d}{dx}\right)^k g(f(x)) = p_k^{(n)} (\{\partial_\alpha g\}_{|\alpha|\leq k}, \{ f^{(j)}_i\}_{i=1, \ldots, n, j=1, \ldots, k} )
\end{align}
for all $f \in C ^k (\mathbb R, \mathbb R ^n)$ and $g \in C^k(\mathbb R ^n, \mathbb R).$ Furthermore, $p_k^{(n)}$ is one-homogeneous in the first entries.
\begin{lemma}
\label{lemma:estimateAnalyticFaaDiBruno}
Let $k,n \in \mathbb{N}$, $C_x,C_y>0$, $x_j^{(i)}>0$ and
\begin{align}\label{eq:lemmaUpperBound}
y_\alpha \in \left(0,C_y \frac {(\abs{\alpha} +1)!} {r^{\abs{\alpha} +1}}\right)
\end{align}
for all $\alpha \in \mathbb{N}_0^n$ with $\abs{\alpha} \le k$, $i \in \{1,\ldots,k\}$ and $j \in \{1,\ldots,n\}$.
Then, for $C:=\max\{1,C_x\}$
\[
p_k^{(n)}(\{y_\alpha\}_{\abs{\alpha}\le k}, \{C_x x_j^{(i)}\}_{\substack{i = 1, \ldots, k\\j =1, \ldots, n}})
\le C_y p_k^{(n)}\left(\left\{\frac {(\abs{\alpha} +1)!} {\left(\frac r C\right)^{\abs{\alpha} +1}}\right\}_{\abs{\alpha}\le k}, \{x_j^{(i)}\}_{\substack{i = 1, \ldots, k\\j =1, \ldots, n}}\right).
\]
\end{lemma}
Note that as soon as $y_\alpha = \abs{\partial^\alpha g}$ for some function $g$ which is analytic away from the origin, the upper bound for $y_\alpha$ in \eqref{eq:lemmaUpperBound} automatically holds, see the proof of \cite[Theorem 7.2]{BlattVorderobermeier:2019:CriticalMoebius}.
\begin{proof}
Using the explicit form, we may compute
\[
\begin{split}
&p_k^{(n)}(\{y_\alpha\}_{\abs{\alpha}\le k}, \{C_x x_i^{(j)}\}_{\substack{i = 1, \ldots, k\\j =1, \ldots, n}})\\
=&\sum_0 \sum_1 \cdots \sum_k \frac{k!}{\prod_{i=1}^k (i!)^{r_i} \prod_{i=1}^k\prod_{j=1}^nq_{ij}! } y_\alpha \prod_{i=1}^k \prod_{j=1}^n (C_x x_j^{(i)})^{q_{ij}}\\
\le& C_y \sum_0 \sum_1 \cdots \sum_k \frac{k!}{\prod_{i=1}^k (i!)^{r_i} \prod_{i=1}^k\prod_{j=1}^nq_{ij}!} \frac {(\abs{\alpha} +1)!} {r^{\abs{\alpha} +1}} C_x^{\abs{\alpha}} \prod_{i=1}^k \prod_{j=1}^n (x_j^{(i)})^{q_{ij}},
\end{split}
\]
as the sum of all $q_{ij}$ is exactly $\abs{\alpha}$.
Estimating $C_x^{\abs{\alpha}} \le C^{\abs{\alpha} +1}$ yields the desired result.
\end{proof}
The explicit formula also allows us to estimate the norm of a Faà di Bruno polynomial in a Banach algebra.
\begin{lemma}
\label{lemma:estimateFaaDiBrunoBanachAlgebra}
Let $\mathcal{X}$ be a Banach algebra with norm $\norm{\cdot}$ in the sense that there is a universal constant $b$ such that for $x,y \in \mathcal{X}$, $\norm{xy} \le b \norm{x}\norm{y}$.
Furthermore, let $k \in \mathbb{N}$ and $x_j^{(i)},y_\alpha \in \mathcal{X}$ for all $\alpha \in \mathbb{N}_0^n$ with $\abs{\alpha} \le k$, $i \in \{1,\ldots,k\}$ and $j \in \{1,\ldots,n\}$.
Then,
\[
\norm*{p_k^{(n)}(\{y_\alpha\}_{\abs{\alpha}\le k}, \{x_j^{(i)}\}_{\substack{i = 1, \ldots, k\\j =1, \ldots, n}})}
\le p_k^{(n)}(\{b^{\abs{\alpha}}\norm{y_\alpha}\}_{\abs{\alpha}\le k}, \{\norm{x_j^{(i)}}\}_{\substack{i = 1, \ldots, k\\j =1, \ldots, n}}.
\]
\end{lemma}
Note that in our application scenario, the product is just pointwise multiplication between functions mapping to $\mathbb{R}$.
\begin{proof}
The statement follows from the estimate $\norm{\prod_{l=1}^m z_l} \le b^{m-1} \prod_{l=1}^m \norm{z_l}$ for all $z_l \in \mathcal{X}$ and the fact that the $q_{ij}$ sum up to $\abs{\alpha}$ together with the explicit formula
of Faa di Bruno's polynomial in Lemma \ref{lem:FaadiBruno}, in particular the one-homogeneity of its first entry.
\end{proof}
\section{An inequality for Sobolev Norms}
\begin{lemma}
\label{lemma:equivalentSobolevNormsDerivative}
Let $m>1$ and $f \in H^m(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ with $\int_0^1 f(x) \d x =0$.
Then,
\[
\frac 1 {2\pi} \norm[H^{m-1}]{f'}
\le \norm[H^m]{f}
\le \frac 1 {\sqrt{2} \pi} \norm[H^{m-1}]{f'}
\le \norm[H^{m-1}]{f'}.
\]
\end{lemma}
\begin{proof}
By \cite[Proposition~3.1.2~(10)]{Grafakos:2014:ClassicalFourierAnalysis}, we have $\widehat{f'}(k) = 2\pi i k \hat f(k)$ for all $k \in \mathbb{Z}$ and thus for $k \ne 0$,
\[
\abs{\hat f(k)}^2 = \frac 1 {(2 \pi k)^2} \abs{\widehat{f'}(k)}^2.
\]
Note that
\[
\hat f(0) = \int_0^1 f(x)e^{-2\pi \cdot 0 \cdot x} \d x =\int_0^1 f(x) \d x =0
\]
with the same holding for $\widehat{f'}(0)$ as $f$ is periodic and continuous.
Keeping the estimates $k^2 \le 1+k^2 \le 2 k^2$ in mind, we obtain
\begin{align*}
\frac 1 {(2\pi)^2} \norm[H^{m-1}]{f'}^2
&= \sum_{k\in \mathbb{Z} \setminus \{0\}} \frac 1 {(2\pi)^2} (1+k^2)^{m-1} \abs{\widehat{f'}(k)}^2
= \sum_{k\in \mathbb{Z} \setminus \{0\}} (1+k^2)^{m-1} k^2 \abs{\hat f(k)}^2\\
&\le \sum_{k\in \mathbb{Z} \setminus \{0\}} (1+k^2)^m \abs{\hat f(k)}^2
= \norm[H^m]{f}^2\\
&= \sum_{k\in \mathbb{Z} \setminus \{0\}} (1+k^2)^m \frac 1 {(2 \pi k)^2} \abs{\widehat{f'}(k)}^2
\le \frac 2 {(2\pi)^2} \sum_{k\in \mathbb{Z} \setminus \{0\}} (1+k^2)^{m-1} \abs{\widehat{f'}(k)}^2\\
&= \frac 1 {2\pi^2} \norm[H^{m-1}]{f'}^2.
\end{align*}
\end{proof}
\section{Introduction}
An easy way to produce a knot is to tie a piece of string and glue the ends together. This motivates the mathematical definition of a knot as a continuous embedding of a circle into three-dimensional Euclidean space. Depending on the initial knotting, different knot types arise. We say that two knots are of the same knot class if there exists an ambient isotopy \cite[p.~4f.]{Cromwell:2004} that deforms one knot to the other avoiding any sort of self-intersections or ``cutting and glueing'' phenomena.
In the following we aim to investigate the regularity of optimal shapes for knots within their knot classes.
Lead by the question of determining optimal representatives for given polygonal knot classes, Fukuhara \cite{Fukuhara:1988:EnergyKnot} introduced the idea to apply some sort of self-repelling potentials to given knots and follow their gradient flow towards the local minimizer. Based on that concept, O'Hara later established the notion of knot energies by defining them as real-valued functionals on the space of knots that are bounded from below and self-repulsive, i.e.~the energy blows up on a sequence of knots that converges to a curve with self-intersections \cite{OHara:2003:EnergyKnotsConformGeom}.
O'Hara was also the first to suggest such a potential knot energy for actual knots \cite{OHara:1991:EnergyKnot}, motivated from the electrostatic potential energy, in particular the self-avoidance effects described by Coulomb's law. Later, this energy was named Möbius energy due to its invariance with respect to Möbius transformations shown by Freedman, He and Wang \cite{Freedmanetal:1994:MoebiusEnergy}.
Motivated by applications to molecular biology, Gonzalez and Maddocks elaborated a completely different approach to define knot energies in search of a new notion for thickness of knots \cite{GonzalesMaddocks:1999:GlobalCurvature}.
For any continuous, closed, and rectifiable curve $\gamma:\mathbb{R}/\mathbb{Z} \rightarrow \mathbb{R}^3$, where $\mathbb{R} /\mathbb{Z} \cong S^1$ denotes the circle of length 1, they characterized the thickness of $\gamma$ by
\begin{align*}
\Delta (\gamma) = \inf_{x,y,z\in \gamma } R(x,y,z) ,
\end{align*}
where $R(x,y,z)$ stands for unique circumcircle radius of three points $x,y,z\in \mathbb{R}^n$ given by
\begin{align*}
R(x,y,z) = \frac{|y-z|\, |y-x| \, |z-x|}{2 |(y-x)\wedge (z-x)|} = \frac{|y-z|}{2 \sin \measuredangle (y-x,z-x)}.
\end{align*}
Observe that $\Delta(\gamma) = 0$ if $\gamma$ intersects itself. Therefore, the quotient $\mathcal{L}(\gamma) / \Delta(\gamma)$ of length $\mathcal{L}(\gamma)$ over thickness $\Delta(\gamma)$ called \emph{ropelength} serves as a starting point for new self-repelling potential energies.
We remark that thickness is related to the
local radius of curvature since the circumcircle of three points $x,y,z$ on a sufficiently regular curve $\gamma$ converges to the osculating circle at $x$ for $y$ and $z$ converging to $ x$ on $\gamma$. The radius of the osculating circle at $x$ on $\gamma$ is the inverse of the local curvature.
The ropelength imposes technical challenges as it consists of operators that are hard to differentiate. These drawbacks can be softened as suggested in \cite{GonzalesMaddocks:1999:GlobalCurvature} by exchanging point-wise maximization with integration, which leads to various \emph{integral Menger\footnote{Named after the Austrian-American mathematician Karl Menger (1902-1985) who used the radius of circumscribed circles on a curve to generalize the concept of curvature to metric spaces \cite{Menger:1930:UntersuchungenUeberAllgemeineMetrik}} curvature energies}
such as
\begin{align}\label{def:intMengcurv}
\mathcal{M}^{p}(\gamma) = \iiint_{(\mathbb{R}/\mathbb{Z})^3} \frac{|\gamma'(r)| \, |\gamma'(s)| \, |\gamma'(t)|}{R(\gamma(r),\gamma(s), \gamma(t))^p} \d r \d s\d t.
\end{align}
for any $p>0$. Note that for curves parametrized by arc-length $\lim_{p\rightarrow \infty} \mathcal{M}_p^{1/p} (\gamma) = \frac{1}{\Delta(\gamma)}$. Other variants between ropelength and $\mathcal M^p$ can be found in \cite{GonzalesMaddocks:1999:GlobalCurvature}.
For $p>3$, these functionals are indeed knot energies according to O'Hara's definition, as Strzelecki, Szumańska and von der Mosel showed in \cite{StrzeleckiSzumanskavonderMosel:2013:KnotEnergiesInvolvMengCurv}. Furthermore, they proved a geometric Morrey-Sobolev imbedding in three dimensions, in particular that finite energy of an arc-length parametrized curve implies $C^{1,1-\frac 3p}$-regularity and its image is $C^1$-diffeomorphic to the circle \cite{StrzeleckiSzumanskavonderMosel:2010:RegularizingandselfavoidanceeffectsofintegralMengercurvature}. Blatt completed the characterization of energy spaces in \cite{Blatt:2013:NoteMenger}. Generalizations of the integral Menger curvature to higher-dimensional objects can be found for example in \cite{StrzeleckivonderMosel:2009:IntegralMengercurvatureforsurfaces,BlattKolasinski:2012:SharpboundednessandregularizingeffectsoftheintegralMengercurvatureforsubmanifolds, Kolasinski:2014:GeometricSobolevlikeembeddingusinghighdimensionalMengerlikecurvature}.
For further information on the integral Menger curvature energies we refer the reader for example to the survey paper \cite{StrzekeckivonderMosel:2013:MengerSurvey}.
In order to tackle the main question of the present paper, the regularity of optimal shapes for knots, we first need to choose a suitable knot energy and then conduct regularity studies on its stationary points.
Now if we chose for this purpose the integral Menger curvature energies $\mathcal{M}^p$, $p>3$, we would arrive at nonlinear, non-local and degenerate Euler-Lagrange operators. A formula for the first variation can be found in \cite{Hermes:2012:Menger} and for its second and third variation in \cite{Knappmann:2020:Menger}. Hence, to produce non-degenerate energies from the integral Menger curvature energies given in \eqref{def:intMengcurv}, Blatt and Reiter \cite{BlattReiter:2015:Menger} suggested to generalize $\mathcal{M}^p$ to
\begin{align}\label{def:GenIntMengCurv}
\textnormal{intM}^{(p,q)}(\gamma) = \iiint_{(\mathbb{R}/\mathbb{Z})^3} \frac{|\gamma'(u_1)| \, |\gamma'(u_2)| \, |\gamma'(u_3)|}{R^{(p,q)}(\gamma(u_1),\gamma(u_2), \gamma(u_3))} \d u_1 \d u_2 \d u_3,
\end{align}
where they decoupled the circumcircle radius to
\begin{align*}
R^{(p,q)}(x,y,z) = \frac{(|y-z|\, |y-x| \, |z-x|)^p}{|(y-x)\wedge (z-x)|^q} = \frac{|y-z|^p |y-x|^{p-q} |z-x|^{p-q}}{ \sin \measuredangle (y-x,z-x)^q}
\end{align*}
for any $p,q>0$. Observe that $\mathcal{M}^{p}(\gamma) = 2^p \textnormal{intM}^{(p,p)}(\gamma) $. According to \cite{BlattReiter:2015:Menger}, these energies are well-defined knot energies for $p\geq \tfrac 23 q + 1$ and give finite energy values on closed curves for $p< q + \tfrac 23$. Moreover, minimizers exist within every knot class for $p\in (\tfrac 23 q +1,q+\tfrac 23)$ and $q>1$.
For the particular case $q=2$ and therefore $p\in (\tfrac 73, \tfrac 83)$, the generalized integral Menger curvature energies in \eqref{def:GenIntMengCurv} indeed lead to non-degenerate Euler-Lagrange equations, for which Blatt and Reiter proved the following regularity result by a combination of potential estimates and Sobolev-embeddings.
\begin{theorem}\label{thm:smooth}\cite[Theorem~4]{BlattReiter:2015:Menger}
For $p\in (\tfrac 73, \tfrac 83)$, let $\gamma :\mathbb{R}/\mathbb{Z}\rightarrow \mathbb{R}^n$ be a simple curve in $W^{\frac 32 p -2,2}(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ parametrized by arc-length.
If $\gamma$ is a stationary point of the generalized integral Menger curvature $\Mpq[p,2]$
with respect to fixed length,
then $\gamma\in C^\infty(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$.
\end{theorem}
Furthermore, Blatt and Reiter characterized $C^1$-curves of finite generalized integral Menger curvature energy by curves in the fractional Sobolev space $W^{\frac {3p-2}{q} -1,q}(\mathbb{R} /\mathbb{Z},\mathbb{R}^n)$ \cite[Theorem~1]{BlattReiter:2015:Menger}.
This result, combined with \thref{thm:smooth}, raises the question whether stationary points of finite generalized integral Menger curvature energies are not only smooth but also analytic.
For the Möbius energy the statement was elaborated by \cite{BlattVorderobermeier:2019:CriticalMoebius} based on the corresponding smoothness result established by \cite{BlattReiterSchikorra:2016:CriticalPointsOfMoebiusEnergyAreSmooth}. The same holds for a non-degenerate range of O'Hara's energies as shown in \cite{Vorderobermeier:2019:CriticalOHara} based on \cite{BlattReiter:2013:StationaryPointsOfOHarasKnotEnergies}. Hence it seems reasonable that it is possible to transfer the techniques to the non-degenerate range of energies $\Mpq[p,2]$ for $p\in(\tfrac 73,\tfrac 83)$.
The main method used in these papers to show analyticity is based on the method of majorants, motivated by the proof of Cauchy-Kovalevsky's theorem \cite{Kovalevskaya:1875:ZurTheoriederpartiellenDifferentialgleichungen}, see also \cite[pp.~17--33]{Cauchy:1882:OEuvrescompletesdAugustinCauchy}. For applying this technique, it is necessary to establish a recursive estimate for derivatives of critical curves.
It turns out that in such estimates, there appears
a fractional derivative of the product of two functions, which subsequently leads to a fractional Leibniz rule. Unfortunately, in contrast to the bilinear Hilbert transform in \cite{BlattVorderobermeier:2019:CriticalMoebius} and fractional Leibniz rule in \cite{Vorderobermeier:2019:CriticalOHara}, in our case this recursive estimate does not produce a sufficient difference in the order of differentiability, which is why we repeat it to obtain a difference of order $1$.
This repetition seems to lead to Gevrey regularity of order $2$ for critical curves, which is weaker than analyticity, cf. \cite[Definiton 4.1.19]{KrantzParks:2002:Aprimerofrealanalyticfunctions} or \cite[Chapter~1.4]{Rodino:1993:LinearPartialDiffOperatorInGevreySpaces},
but we are able to confirm the analyticity in the affirmative:
\begin{theorem}\label{thm:main}
For $p\in (\tfrac 73, \tfrac 83)$, let $\gamma :\mathbb{R}/\mathbb{Z}\rightarrow \mathbb{R}^n$ be a simple curve in $C^\infty(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ parametrized by arc-length.
If $\gamma$ is a critical point of the generalized integral Menger curvature $\Mpq[p,2]$
with respect to fixed length,
then the curve $\gamma$ is analytic.
\end{theorem}
A closer look at the proof of the main theorem reveals that smoothness of $\gamma$ is actually not needed. It suffices to assume that $\norm[H^{\frac 5 2}]{\gamma^{(i)}}$ is finite for $i \in \set{0,1,2}$, i.e. $\gamma \in H^{4 + \frac 1 2}(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$. All additional regularity is obtained by an inductive argument.
However, by the characterization of the energy spaces in \cite[Theorem~1]{BlattReiter:2015:Menger} and the regularity result \thref{thm:smooth}, finite energy $\Mpq[p,2]$ of a critical point implies its smoothness already. We conclude:
\begin{corollary}\label{cor:main}
Let $\gamma :\mathbb{R}/\mathbb{Z}\rightarrow \mathbb{R}^n$ be a simple curve in $C^1(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ parametrized by arc-length with $\Mpq[p,2] (\gamma) < \infty$ for $p\in(\tfrac 73, \tfrac 83)$.
If $\gamma$ is a critical point of $\Mpq[p,2]$ with respect to fixed length, then the curve $\gamma$ is analytic.
\end{corollary}
One should be able to drop the $C^1$-assumption in the previous result by obtaining sufficient initial regularity via techniques that proved to be successful in showing a geometric Morrey-Sobolev imbedding for $\mathcal{M}^p$ with $p>3$, cf.~\cite[Theorem~1.2]{StrzeleckiSzumanskavonderMosel:2010:RegularizingandselfavoidanceeffectsofintegralMengercurvature}.
Further analyticity and Gevrey regularity results in the context of non-local differential equations can be found for example in \cite{DallAcquaEtAl:2012,DallAcquaEtAl:2014,AlbaneseEtAl:2015:GevreyRegularityForIntDiffOp,Blatt:2020:AnalyticityIntegroDiffOp,Blatt:2020:AnalyticityNonlinerEllipticPDE}.
\subsection*{Exposé of the present work}
As stated above, the central idea of our proof is the same as in the one for the Cauchy-Kovalevsky theorem and goes as follows.
A curve $\gamma \in C^\infty(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ is analytic if and only if we have constants $C>0$ and $r>0$ such that for all $l \in \mathbb{N}_0$
\[
\norm[L^\infty]{\gamma^{(l)}} \le C \frac {l!} {r^l}.
\]
Suppose we had a recursive estimate
\[
\norm[L^\infty]{\gamma^{(l+1)}} \le \Phi(l;\norm[L^\infty]{\gamma},\ldots,\norm[L^\infty]{\gamma^{(l)}})
\]
and an analytic function $c:(-\epsilon, \epsilon) \to \mathbb{R}$ with
\[
c^{(l+1)}(0) = \Phi(l;c(0),\ldots,c^{(l)}(0)) \text{ and } c(0) \ge \norm[L^\infty]{\gamma}.
\]
Then, by induction,
\[
\begin{split}
0 \le \norm[L^\infty]{\gamma^{(l+1)}} \le \Phi(l+1;c(0),\ldots,c^{(l)}(0)) = c^{(l+1)}(0) \le C \frac {(l+1)!} {r^{l+1}}
\end{split}
\]
for all $l \in \mathbb{N}$ and so, $\gamma$ is analytic.
The largest part is devoted to obtaining the required recursive estimate.
In Section \ref{sec:Decomposition}, we review a decomposition of the first variation of $\Mpq[p,2]$ given in \cite[Section~4]{BlattReiter:2015:Menger} into a highest- and lower-order terms called $\ensuremath{\widetilde Q^{(p)}}$ and $\ensuremath{\widetilde R^{(p)}}$.
In a critical point, $\ensuremath{\widetilde Q^{(p)}}$ is equal to $-\ensuremath{\widetilde R^{(p)}}$ and we may use the regularity gap between those to establish the recursive estimate.
To that end, Section \ref{sec:EstimateMainTerm} yields an estimate of $\partial^{l+3} \gamma$ by $\partial^l\ensuremath{\widetilde Q^{(p)}}(\gamma)$.
In Section \ref{sec:RemainderTermNonCritical} we bound $\partial^l \ensuremath{\widetilde R^{(p)}}(\gamma)$ by $\partial^{l+2} \gamma$ via a fractional Leibniz rule.
Looking at the norms involved, it turns out that these estimates combined only lead to a regularity gain strictly between $\frac 1 2$ and $1$ which is not enough for the recursive estimate, so in Subsection \ref{subsec:RecursiveEstimate}, we iterate the above once more using the fact that $\gamma$ is a critical point.
The remainder of Section \ref{sec:MainProof} is dedicated to finding the analytic majorant $c$ which fulfils the same recursive estimate with equality.
Finally, Section \ref{sec:Consequences} discusses a few simple consequences of the analyticity of critical points, not only for $\Mpq[p,2]$, in geometric knot theory.
We expect that our approach extends to functionals whose first variation admits a decomposition with properties similar to those of $\ensuremath{\widetilde Q^{(p)}}$ and $\ensuremath{\widetilde R^{(p)}}$.
A probably valid generalization would be to assume only an arbitrarily small gain in regularity, see page \pageref{rem:smallerRegularityGain}.
\subsection*{Notation}\label{subsec:Notation}
In the following we give a short introduction to the fractional Sobolev spaces we are using in this paper. First, we present a continuation of the usual periodic Sobolev spaces of integer order $W^{k,p}(\mathbb{R}/\mathbb{Z}, \mathbb{R}^n)$ to Sobolev–Slobodeckij spaces $W^{k+s,p}(\mathbb{R}/\mathbb{Z}, \mathbb{R}^n)$ for $0<s <1$. The latter are defined by
\[
W^{k+s,p}(\mathbb{R}/\mathbb{Z}, \mathbb{R}^n) := \{ f\in W^{k,p}(\mathbb{R}/\mathbb{Z}, \mathbb{R}^n) \ | \ [f^{(k)}]_{W^{s,p}(\mathbb{R}/\mathbb{Z}, \mathbb{R}^n)} < \infty \}
\]
equipped with the norm
\[
\|f\|_{W^{k+s,p}(\mathbb{R}/\mathbb{Z}, \mathbb{R}^n)} := \|f\|_W^{k,p}(\mathbb{R}/\mathbb{Z}, \mathbb{R}^n) + [f^{(k)}]_{W^{s,p}(\mathbb{R}/\mathbb{Z}, \mathbb{R}^n)},
\]
where
\[
[f^{(k)}]_{W^{s,p}(\mathbb{R}/\mathbb{Z}, \mathbb{R}^n)} := \Big(\int_{\mathbb{R}/\mathbb{Z}} \int_{-1/2}^{1/2} \frac{|f^{(k)}(x) - f^{(k)}(y)|^p}{|x-y|^{1+ps}} \d x \d y \Big)^{\frac 1p}
\]
is the so-called Gagliardo seminorm of $k$-th derivative of $f$.
Second, we consider Bessel potential spaces of order $s \geq 0$ on the circle $\mathbb{R}/\mathbb{Z}$ that are given by
\begin{align*}
H^s(\mathbb{R}/\mathbb{Z}, \mathbb{C}^n) := \{f\in L^2 (\mathbb{R}/\mathbb{Z}, \mathbb{C}^n) \ | \ \| f \|_{H^s}:= \| f \|_{H^s(\mathbb{R}/\mathbb{Z},\mathbb{C}^n)} := \sqrt{(f,f)_{H^s}} < \infty \},
\end{align*}
equipped with the inner product
\begin{align*}
(f,g)_{H^s} := (f,g)_{H^s(\mathbb{R}/\mathbb{Z},\mathbb{C}^n)} := \sum_{k\in\mathbb{Z}} (1+k^2)^s \langle \widehat{f}(k), \widehat{g}(k) \rangle_{\mathbb{C}^n}.
\end{align*}
Here, $\hat f(k) = \int_0^1 e^{-2 \pi i k x} f(x) \d x$ is the $k$-th Fourier coefficient of $f$.
For any integer $k\geq 0$, the introduced space $H^k$ coincides with the classical Sobolev space $W^{k,2}$ and their norms $\|\cdot\|_{H^k}$ and $\|f\|_{W^{k,2}}:= (\sum_{\nu=0}^{k} \|f^{(\nu)}\|^2_{L^2})^{\frac{1}{2}}$ are equivalent (cf.~\cite[Lem.~1.2]{Reiter:2012:RepulsiveKnotEnergies} or \cite[7.62]{AdamsFournier:2003:SobolevSpaces}).
Moreover, Bessel potential spaces coincide with Sobolev–Slobodeckij spaces in the case of $p=2$, in particular, they have equivalent norms
as stated in \cite[Proposition~3.4]{DiNezzaetal:2012:HitchhikersGuide} for spaces on the domain $\mathbb{R}^d$. The proof can be modified to fit our situation.
It can be shown that $H^s(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)\subseteq H^t(\mathbb{R}/ \mathbb{Z},\mathbb{R}^n)$ for any $0 \leq t<s$ (cf.~\cite[Chapter~4,~Proposition~3.4]{Taylor:1996:PartialDifferentialEquationsa}) as well as $H^s(\mathbb{R}/\mathbb{Z},\mathbb{R}^n) \subseteq C(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ for any $s> \frac{1}{2}$ (cf.~\cite[Chapter~4,~Proposition 3.3]{Taylor:1996:PartialDifferentialEquationsa}).
Furthermore, $H^m$ satisfies the Banach algebra property for any $m\geq \frac 12$, i.e. there exists a positive constant $C = C(m)$ such that
\begin{align}\label{prop:Banachalgebra}
\|fg\|_{H^m} \leq C \|f\|_{H^m} \|g\|_{H^m}
\end{align}
for all $f,g \in H^m(\mathbb{R}/\mathbb{Z},\mathbb{R})$ (cf.~\cite[Theor.~4.39]{AdamsFournier:2003:SobolevSpaces}).
For further information on fractional Sobolev spaces, we refer the reader for instance to \cite{AdamsFournier:2003:SobolevSpaces,DiNezzaetal:2012:HitchhikersGuide,RunstSickel:1996:SobolevSpacesOfFractionalOrder, Triebel:2010:TheoryOfFunctionSpaces}.
\section{The remainder term $R^{(p)}$}
\label{sec:RemainderTermNonCritical}
Now we aim to estimate the remainder term $R^{(p)}$, which appears in the first variation of the generalized integral Menger curvature for $q=2$ and $p\in (\tfrac 73, \tfrac 83)$, and its derivatives. We observe that it is indeed of lower order in comparison to the main term $Q^{(p)}$, which has been studied in the previous section.
\subsection{Representation of the remainder term $R^{(p)}$}
In this subsection we are concerned with an alternative representation of the remainder term $R^{(p)}$ and its derivatives.
Let us first recall the following result.
\begin{lemma}[Regularity of the remainder integrand, \protecting{\cite[Lemma 4.4]{BlattReiter:2015:Menger}}]
\label{lemma:regularityIntegrandRemainder}\ \\
Let $\gamma\in W^{(3p-4)/2+\ensuremath{\sigma},2}(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ be a simple curve parametrized by arc-length.
\begin{itemize}
\item If $\ensuremath{\sigma}=0$, then $g^{p}\in L^{1}(\mathbb{R}/\mathbb{Z}\times D,\mathbb{R}^{n})$ and
\item if $\ensuremath{\sigma}>0$, then $((v,w) \mapsto g^{p}(\cdot,v,w;\dots))\in L^{1}(D,\W[\tilde\ensuremath{\sigma},1](\mathbb{R}/\mathbb{Z}, \mathbb{R}^{n}))$ for any $\tilde\ensuremath{\sigma}<\ensuremath{\sigma}$.
\end{itemize}
The respective norms are bounded independently of $s_{1},\dots,s_{K}$.
\end{lemma}
We then derive a representation of the remainder term $R^{(p)}$ corresponding to the $L^2$-re\-pre\-sen\-ta\-tion of the main term $Q^{(p)}$ in the first variation of $\Mpq[p,2]$. Recall that $\otimes$ denotes any kind of bilinear product structure, such as cross product, dot product, scalar or matrix multiplication.
\begin{lemma}
\label{lemma:L2formRemainder}
Let $\gamma \in C^\infty(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ be parametrized by arc-length and let $h \in W^{\frac 3 2 p -2,2}(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$.
Then,
\[
R^{(p)}(\gamma,h)=\int_{\mathbb{R}/\mathbb{Z}} \left<\ensuremath{\widetilde R^{(p)}}(\gamma)(u), h(u)\right> \d u
\]
with
\[
\ensuremath{\widetilde R^{(p)}}(\gamma)_o(u) := \sum_{k=1}^{k_{\max}} \iint_{D} \idotsint\limits_{[0,1]^K} \frac \d {\d u} g^p_k(u - s_{K-1}v - s_K w,v,w;s_{1},\dots,s_{K-2}) \d\theta_{1}\cdots\d\theta_{K}\d v \d w \otimes e_o.
\]
Here, $e_o \in \mathbb{R}^n$ is the $o$-th canonical unit vector, each $g^p_k$ is of the same form as $g^p$ in \thref{lemma:formRemainder} and neither $K$ nor $k_{\max}$ depend on $\gamma$.
Furthermore, $u \mapsto g^p_k(u,v,w;s_1,\ldots,s_K) \in C^\infty(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ for all $k \in \set{1,\ldots,k_{\max}}$.
\end{lemma}
\begin{proof}
Since $\gamma \in C^\infty(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$, \thref{lemma:regularityIntegrandRemainder} yields that the integrand is in fact integrable on the integration domain, allowing us to substitute $u$ by $\tilde u := u +s_{K-1}v + s_K w$.
We also obtain that $u \mapsto g^p(u,v,w;s_1,\ldots,s_K) \in C^\infty(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ and note that $\otimes$ is a bilinear operation. Thus we may use integration by parts component-wise to obtain that \eqref{lemma:formRemainder:formula} is equal to
\[
\iiint\limits_{\mathbb R / \mathbb Z \times D} \idotsint\limits_{[0,1]^K} (\tfrac \d {\d \tilde u} g^p(\tilde u - s_{K-1}v - s_K w,v,w;s_{1},\dots,s_{K-2}) ) \otimes h(\tilde u)
\d\theta_{1}\cdots\d\theta_{K}\d v \d w \d \tilde u.
\]
Note that we have no boundary terms because both $h$ and $g^p$ are periodic.
Using the bilinearity of $\otimes$, we may write \ensuremath{\widetilde R^{(p)}}{} as
\[
\int_{\mathbb{R}/\mathbb{Z}} \overline{R}^{(p)}(\gamma)(u) \otimes h(u) \d u
\]
with
\[
\overline{R}^{(p)}(\gamma)(u) := \sum_{k=1}^{k_{\max}} \iint_{D} \idotsint\limits_{[0,1]^K} \frac \d {\d u} g^p_k(u - s_{K-1}v - s_K w,v,w;s_{1},\dots,s_{K-2}) \d\theta_{1}\cdots\d\theta_{K}\d v \d w.
\]
Seeing as $x \mapsto \overline{R}^{(p)}(\gamma)(u) \otimes x \in \mathbb{R}$ is linear, we may set
\[
\ensuremath{\widetilde R^{(p)}}(\gamma)_o(u) := \overline{R}^{(p)}(\gamma)(u) \otimes e_o
\]
and obtain for $x = \sum_{o=1}^n x_oe_o \in \mathbb{R}^n$:
\[
\overline{R}^{(p)}(\gamma)(u) \otimes x = \sum_{o=1}^n x_o \cdot \overline{R}^{(p)}(\gamma)(u) \otimes e_o = \sum_{o=1}^n x_o \cdot \ensuremath{\widetilde R^{(p)}}(\gamma)_o(u) = \left<\ensuremath{\widetilde R^{(p)}}(\gamma)(u),x\right>.
\]
\end{proof}
The $l$-th derivative of the remainder representation \ensuremath{\widetilde R^{(p)}}{} can be computed as follows.
\begin{lemma}
Let $\gamma \in C^\infty(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ be parametrized by arc-length.
Then,
\[
\begin{split}
&\frac {\d^l} {\d u^l} \ensuremath{\widetilde R^{(p)}}(\gamma)_o(u)\\
=& \sum_{k=1}^{k_{\max}} \iint_{D} \idotsint\limits_{[0,1]^K} \frac {\d^{l+1}} {\d u^{l+1}} g^p_k(u - s_{K-1}v - s_K w,v,w;s_{1},\dots,s_{K-2}) \d\theta_{1}\cdots\d\theta_{K}\d v \d w\otimes e_o
\end{split}
\]
for all $l \in \mathbb{N}$.
\end{lemma}
Note that we do not need the assumption of arc-length parametrization here, as $\ensuremath{\widetilde R^{(p)}}$ is well-defined without it.
However, for other parametrizations, we do not expect $\ensuremath{\widetilde R^{(p)}}(\gamma)$ to be related to $R^{(p)}(\gamma,\cdot)$.
\begin{proof}
Since $X \mapsto X \otimes e_o$ is a fixed linear map, $\frac {\d^l} {\d u^l} (X(u) \otimes e_o) = ( \frac {\d^l} {\d u^l} \partial^l X(u)) \otimes e_o$, so we only need to worry about the integral term.
By \thref{lemma:regularityIntegrandRemainder}, $((v,w) \mapsto g^{p}_k(\cdot,v,w;\dots))\in L^{1}(D,W^{\tilde \sigma,1}(\mathbb{R}/\mathbb{Z}, \mathbb{R}^{n}))$ for all $\tilde\sigma>0$ and its norm is independent of $s_1, \ldots, s_K$.
This means that $g^p_k$ is differentiable with respect to $u$ arbitrarily many times and the derivatives are all integrable with the majorant
\[
\norm[L^\infty]{\partial_1^{l+1} g^p_k(\cdot,v,w;s_1,\ldots,s_K)}.
\]
By the Morrey- (see e.g.\ \cite[Paragraph~4.16]{AdamsFournier:2003:SobolevSpaces}) and Sobolev-embedding theorems, we can bound this above by
\[
C\norm[W^{1,2}]{\partial_1^{l+1} g^p_k(\cdot,v,w;s_1,\ldots,s_K)}
\le \widetilde C \norm[W^{l + \frac 5 2,1}]{g^p_k(\cdot,v,w;s_1,\ldots,s_K)}.
\]
As we have a uniform $L^1$-bound on the latter, we may exchange integration and differentiation.
\end{proof}
\subsection{A fractional Leibniz rule} \label{subsec:FractionalLeibnizRule}
The remainder term $R^{(p)}$ contains amongst others the factor $\Gamma$, which is basically given as a product of two functions multiplied with a singular weight. By applying a Bessel potential space norm, we obtain a sort of fractional Leibniz rule which will subsequently be discussed for $\Gamma$ and its derivatives.
We observe that the $l$-th derivative of $R^{(p)}$ leads to $k$-th derivatives of $\Gamma$, where $k\leq l+1$. Applying the generalized Leibniz rule to the $ k$-th derivative of $\Gamma$ in the next step, we obtain
\[
\begin{split}
& \frac {\d^{k}} {\d u^{k}} \Gamma(u,v,w,s_1,s_2) = \frac{1}{|v|^{p-2}|w|^{p-2}|v-w|^{p}} \\
& \cdot \sum_{o=0}^{k} \binom{k}{o} \left(\gamma^{(k-o+1)}(u+x_1) - \gamma^{(k-o+1)}(u+x_2) \right) \otimes \left( \gamma^{(o+1)}(u+x_3) - \gamma^{(o+1)}(u+x_4) \right),
\end{split}
\]
where we have the four cases
\[
\begin{split}
\textbf{Case 1:} & \quad (x_1,x_2,x_3,x_4) =(s_1w,s_1v, s_2w, s_2v), \\
\textbf{Case 2:} & \quad (x_1,x_2)=(x_3,x_4) = (s_1w,s_2w), \\
\textbf{Case 3:} & \quad (x_1,x_2)=(x_3,x_4) = (s_1v, s_2v), \\
\textbf{Case 4:} & \quad (x_1,x_2)=(x_3,x_4) = (v+s_1(w-v), v+s_2(w-v)).
\end{split}
\]
The key observation here is that $\Gamma$ as well as its derivatives can be broken down into sums of products of the form
\[
\left(\gamma_i^{(k-o+1)}(u+x_1) - \gamma_i^{(k-o+1)}(u+x_2) \right) \cdot \left( \gamma_j^{(o+1)}(u+x_3) - \gamma_j^{(o+1)}(u+x_4) \right),
\]
where $i,j=1,\ldots,n$ and $o=0,\ldots, k$. The difference in the component functions and the additional factor $|v|^{2-p}|w|^{2-p}|v-w|^{-p}$ indicate a fractional derivative of the arising products. This motivates to deduce the following general statement, which can therefore be interpreted as fractional Leibniz rule.
\begin{theorem}\label{thm:fracLeibnizrule}
Let $m> \frac{1}{2}$, $s_1,s_2 \in [0,1]$ and $p \in (\frac 73,\frac 83)$. Then there exists a positive constant $C_L = C_L(m, p) < \infty$ such that for all $f,g \in C^\infty(\mathbb{R}/\mathbb{Z},\mathbb{R})$
\begin{align*}\label{eq:fracLeibnizrule}
& \iint_D \frac{ \left\| \left(f(\cdot+x_1) - f(\cdot+x_2) \right) \left(g(\cdot+x_3) - g(\cdot+x_4) \right) \right\|_{H^m}}{|v|^{p-2}|w|^{p-2}|v-w|^{p}} \d v \d w \nonumber \\
& \leq C_L ( \|f\|_{H^{m+ \frac 32 p -3}} \|g\|_{H^{m+ \frac 32 p -3 }} ),
\end{align*}
where
\[
\begin{split}
\textbf{Case 1:} & \quad (x_1,x_2,x_3,x_4) =(s_1w,s_1v, s_2w, s_2v), \\
\textbf{Case 2:} & \quad (x_1,x_2)=(x_3,x_4) = (s_1w,s_2w), \\
\textbf{Case 3:} & \quad (x_1,x_2)=(x_3,x_4) = (s_1v, s_2v), \\
\textbf{Case 4:} & \quad (x_1,x_2)=(x_3,x_4) = (v+s_1(w-v), v+s_2(w-v)).
\end{split}
\]
\end{theorem}
\begin{proof}
We first of all note that $H^m$ is a Banach algebra due to $m>\frac 12$ (see \eqref{prop:Banachalgebra}), which implies
\begin{align*}
& \iint_D \frac{ \left\| \left(f(\cdot+x_1) - f(\cdot+x_2) \right) \left(g(\cdot+x_3) - g(\cdot+x_4) \right) \right\|_{H^m}}{|v|^{p-2}|w|^{p-2}|v-w|^{p}} \d v \d w \nonumber \\
& \leq C(m) \iint_D \frac{ \left\| f(\cdot+x_1) - f(\cdot+x_2) \right\|_{H^m} \left\|g(\cdot+x_3) - g(\cdot+x_4) \right\|_{H^m}}{|v|^{p-2}|w|^{p-2}|v-w|^{p}} \d v \d w \nonumber \\
& \leq C(m) \left(\iint_D \frac{ \left\| f(\cdot+x_1) - f(\cdot+x_2) \right\|^2_{H^m}}{|v|^{p-2}|w|^{p-2}|v-w|^{p}} \d v \d w \right)^{\frac 12} \left(\iint_D \frac{\left\|g(\cdot+x_3) - g(\cdot+x_4) \right\|^2_{H^m}}{|v|^{p-2}|w|^{p-2}|v-w|^{p}} \d v \d w \right)^{\frac 12}
\end{align*}
for some positive constant $C(m)$ only depending on $m$.
We start with \textbf{Case 1}. We first see by the definition of the fractional Sobolev norm that
\begin{align*}
& \iint_D \frac{ \left\| f(\cdot+s_1w) - f(\cdot+s_1 v) \right\|^2_{H^m}}{|v|^{p-2}|w|^{p-2}|v-w|^{p}} \d v \d w \\
& = \iint_D \sum_{k\in\mathbb{Z}} (1+k^2)^m |(\widehat{f(\cdot+s_1w) - f(\cdot+s_1v)})(k)|^2 \frac{\d v \d w }{|v|^{p-2}|w|^{p-2}|v-w|^{p}} .
\end{align*}
By substitution and the periodicity of $f$, we can rewrite the Fourier coefficients to
\begin{align}\label{eq:FouriercoeffCOV}
|\widehat{\left(f(\cdot+s_1w) - f(\cdot+s_1v)\right)}(k)| & = \left|\int_0^1 (f(u+s_1w) - f(u+s_1v)) e^{-2\pi i k u}\d u\right| \nonumber \\
& = \left|e^{2\pi i k s_1 w}\right| \, \left|\int_0^1 (f(u) - f(u+s_1(v-w))) e^{-2\pi i k u}\d u\right| \nonumber \\
& = |(\widehat{f(\cdot+s_1(v-w))-f(\cdot)})(k)|,
\end{align}
which implies
\[
\left\| f(\cdot+s_1w) - f(\cdot+s_1 v) \right\|_{H^m} = \left\| f(\cdot+s_1 (v-w)) - f(\cdot)\right\|_{H^m}.
\]
Then, as in the proof of \cite[Lemma 4.4]{BlattReiter:2015:Menger}, respectively \cite[Lemma 1]{BlattReiter:2015:Menger}, we substitute with $\Phi:(v,w) \mapsto (t,\tilde{w}):= (\frac{v}{v-w}, s_1(v-w))$, $|\det D\Phi(v,w)| = \frac{s_1}{|v-w|}$, $\Phi(D)\subset [0,1]\times [-1,0]$, such that
\begin{align*}
&\iint_D \frac{ \left\| f(\cdot+s_1w) - f(\cdot+s_1 v) \right\|^2_{H^m}}{|v|^{p-2}|w|^{p-2}|v-w|^{p}} \d v \d w \\
& = \iint_D \frac{ \left\| f(\cdot+s_1 (v-w)) - f(\cdot)\right\|^2_{H^m}}{|v|^{p-2}|w|^{p-2}|v-w|^{p}} \d v \d w \\
& \leq s_1^{-1} \int_0^1 \int_{-1}^0 \frac{ \left\| f(\cdot+\tilde{w}) - f(\cdot)\right\|^2_{H^m}} {|t \frac{\tilde{w}}{s_1}|^{p-2}|(t-1)\frac{\tilde{w}}{s_1}|^{p-2}|\frac{\tilde{w}}{s_1}|^{p-1}} \d \tilde{w} \d t \\
& = s_1^{3p-6} \left( \int_{-1}^{0} \frac{ \left\| f(\cdot+\tilde{w}) - f(\cdot)\right\|^2_{H^m}}{|\tilde{w}|^{3p-5}} \d \tilde{w} \right) \smash{\underbrace{\left(\int_0^1 \frac{dt}{|t(1-t)|^{p-2}} \right)}_{< \infty}}\\
& \leq C(p) \int_{-1}^{0} \frac{ \left\| f(\cdot+\tilde{w}) - f(\cdot)\right\|^2_{H^m}}{|\tilde{w}|^{3p-5}} \d \tilde{w}.
\end{align*}
Note that the positive constant $C(p)$ in the previous estimate does not depend on $s_1$ since $3p-6 >1$ and by definition $0\leq s_1 \leq 1$.
Now we find again by the definition of the fractional Sobolev norm
\begin{align*}
\int_{-1}^{0} \frac{ \left\| f(\cdot+\tilde{w}) - f(\cdot)\right\|^2_{H^m}}{|\tilde{w}|^{3p-5}} \d \tilde{w} = \int_{-1}^{0} \sum_{k\in\mathbb{Z}} (1+k^2)^m |(\widehat{f(\cdot+\tilde{w}) - f(\cdot)})(k)|^2 \frac{d \tilde{w}}{|\tilde{w}|^{3p-5}},
\end{align*}
whose Fourier coefficients have the form
\begin{align*}
\left|(\widehat{f(\cdot+\tilde{w}) - f(\cdot)})(k)\right|^2 & = \left|\int_0^1 f(u+\tilde{w}) e^{-2\pi i k u} \d u - \int_0^1 f(u) e^{-2\pi i k u} \d u\right|^2 \\
& = |e^{2\pi i k \tilde{w}} - 1|^2 | \hat{f}(k)|^2.
\end{align*}
Since we get by Euler's formula
\begin{align} \label{eq:orderk}
\begin{split}
\int_{-1}^{0} |e^{2\pi i k \tilde{w}} - 1|^2 \frac{d \tilde{w}}{|\tilde{w}|^{3p-5}} & = 2 \int_0^1 (1-\cos(2\pi k\tilde{w}) ) \, \frac{d \tilde{w}}{\tilde{w}^{3p-5}} \\
& \leq 2 (2\pi |k|)^{3p-6} \underbrace{\int_0^\infty \frac{1-\cos(\bar{w})}{\bar{w}^{3p-5}} \d \bar{w}}_{<\infty} \\
& \leq \widetilde C(p)\, |k|^{3p-6},
\end{split}
\end{align}
for some positive constant $\widetilde C(p)$ only depending on $p$,
as well as the elementary inequality
\begin{align}\label{eq:elementariest}
|k|^{3p-6} \leq (1+k^2)^{\frac{3p-6}{2}},
\end{align}
we obtain together with Beppo Levi's monotone convergence theorem
\begin{align*}
\int_{-1}^{0} \frac{ \left\| f(\cdot+\tilde{w}) - f(\cdot)\right\|^2_{H^m}}{|\tilde{w}|^{3p-5}} \d \tilde{w} & = \int_{-1}^{0} \sum_{k\in\mathbb{Z}} (1+k^2)^m | \hat{f}(k)|^2 |e^{2\pi i k w} - 1|^2\frac{d \tilde{w}}{|w|^{3p-5}} \\
& \leq \widetilde C(p) \sum_{k\in\mathbb{Z}} (1+k^2)^{m+\frac 32 p -3} | \hat{f}(k)|^2 \\
& = \widetilde C(p) \|f\|_{H^{m+\frac 32 p -3}}^2.
\end{align*}
After applying the same reasoning for the second factor, we finally achieve for \textbf{Case 1}
\begin{align*}
& \iint_D \frac{ \left\| \left(f(\cdot+s_1w) - f(\cdot+s_1v) \right) \left(g(\cdot+s_2w) - g(\cdot+s_2v) \right) \right\|_{H^m}}{|v|^{p-2}|w|^{p-2}|v-w|^{p}} \d v \d w \\
& \leq C(m,p) \|f\|_{H^{m+\frac 32 p -3}} \|g\|_{H^{m+\frac 32 p -3}},
\end{align*}
for some new positive constant $C(m,p)$ depending on $m$ and $p$.
In \textbf{Case 2}, we start by change of variables, cf.\ the proof of \cite[Lemma 4.4]{BlattReiter:2015:Menger}, such that
\begin{align}\label{eq:startcase2}
& \iint_D \frac{ \left\| f(\cdot+s_1w) - f(\cdot+s_2w) \right\|^2_{H^m}}{|v|^{p-2}|w|^{p-2}|v-w|^{p}} \d v \d w \nonumber \\
& \leq \int_{0}^{\frac 23} \frac{ \left\| f(\cdot+s_1w) - f(\cdot+s_2w) \right\|^2_{H^m}}{w^{p-2}} \left( \int_0^{\frac 23 - w} \frac{1}{v^{p-2}(v+w)^{p}} \d v \right) \d w \nonumber \\
& \leq \int_{0}^{\frac 23} \frac{ \left\| f(\cdot+s_1w) - f(\cdot+s_2w) \right\|^2_{H^m}}{w^{p-2}} \left( \int_0^{\infty} \frac{1}{(tw)^{p-2}(w(t+1))^{p}} w \d t \right) \d w\nonumber\\
& \leq \int_{0}^{\frac 23} \frac{ \left\| f(\cdot+s_1w) - f(\cdot+s_2w) \right\|^2_{H^m}}{w^{3p-5}} \underbrace{ \left( \int_0^{\infty} \frac{1}{t^{p-2}(1+t)^{p}} \d t \right) }_{<\infty} \d w \nonumber \\
& \leq C (p) \int_{0}^{\frac 23} \frac{ \left\| f(\cdot+s_1w) - f(\cdot+s_2w) \right\|^2_{H^m}}{w^{3p-5}} \d w,
\end{align}
where $C(p)$ is a positive constant depending on $p$.
The definition of the fractional Sobolev norm gives us
\begin{align*}
\int_{0}^{\frac 23} \frac{ \left\| f(\cdot+s_1w) - f(\cdot+s_2w) \right\|^2_{H^m}}{w^{3p-5}} \d w = \int_{0}^{\frac 23} \sum_{k\in\mathbb{Z}} (1+k^2)^m |(\widehat{f(\cdot+s_1w) - f(\cdot+s_2w)})(k)|^2 \frac{dw}{w^{3p-5}}.
\end{align*}
We then observe for the Fourier coefficients of the series, recalling of the periodicity of $f$,
\begin{align*}
|(\widehat{f(\cdot+s_1w) - f(\cdot+s_2w)})(k)|^2 & = \left|\int_0^1 f(u+s_1w) e^{-2\pi i k u}\d u - \int_0^1 f(u+s_2w)e^{-2\pi i k u}\d u\right|^2 \\
& = |e^{2\pi i k s_1 w} \hat{f}(k) - e^{2\pi i k s_2 w} \hat{f}(k)|^2 \\
& = |e^{2\pi i k (s_1 -s_2)w}-1|^2 |e^{2\pi i k s_2 w}|^2 |\hat{f}(k)|^2,
\end{align*}
and hence
\begin{align*}
\int_{0}^{\frac 23} \frac{ \left\| f(\cdot+s_1w) - f(\cdot+s_2w) \right\|^2_{H^m}}{w^{3p-5}} \d w = \int_{0}^{\frac 23} \sum_{k\in\mathbb{Z}} (1+k^2)^m |\hat{f}(k)|^2 |e^{2\pi i k (s_1 -s_2)w}-1|^2 \frac{dw}{w^{3p-5}}.
\end{align*}
Now again by Euler's formula and change of variables, we get similarly to \eqref{eq:orderk}
\begin{align*}
\int_{0}^{\frac 23} |e^{2\pi i k (s_1 -s_2)w}-1|^2 \frac{dw}{w^{3p-5}} & = 2 \int_{0}^{\frac 23} 1-\cos (2\pi k (s_1 -s_2)w) \frac{dw}{w^{3p-5}} \leq \widetilde C(p) |k|^{3p-6}
\end{align*}
where $\widetilde C(p)$ is a positive constant depending on $p$ and we used that $\abs{s_1-s_2} \le 2$.
Therefore, we obtain by \eqref{eq:elementariest} and Beppo Levi's monotone convergence theorem
\begin{align}\label{eq:case2}
\int_{0}^{\frac 23} \frac{ \left\| f(\cdot+s_1w) - f(\cdot+s_2w) \right\|^2_{H^m}}{w^{3p-5}} \d w & \leq \widetilde C(p) \sum_{k\in\mathbb{Z}} (1+k^2)^{m+\frac 32 p -3} |\hat{f}(k)|^2 \nonumber\\
& = \widetilde C (p) \|f\|^2_{H^{m+\frac 32 p -3}}.
\end{align}
After applying the same arguments for the factor of $g$, we conclude in \textbf{Case 2}
\begin{align*}
& \iint_D \frac{ \left\| \left(f(\cdot+s_1w) - f(\cdot+s_2w) \right) \left(g(\cdot+s_1w) - g(\cdot+s_2w) \right) \right\|_{H^m}}{|v|^{p-2}|w|^{p-2}|v-w|^{p}} \d v \d w \\
& \leq C(m,p) \|f\|_{H^{m+\frac 32 p -3}} \|g\|_{H^{m+\frac 32 p -3}}
\end{align*}
for some positive constant $C(m,p)$ depending on $m$ and $p$.
\textbf{Case 3} directly follows from \textbf{Case 2}. \\
Now only \textbf{Case 4} remains. Along the lines of \textbf{Case 1} and \textbf{Case 2}, in particular \eqref{eq:FouriercoeffCOV} and \eqref{eq:startcase2}, we transform with respect of the Fourier coefficients
\begin{align*}
& \iint_D \frac{ \left\| f(\cdot+v+s_1(w-v)) - f(\cdot+v+s_2(w-v)) \right\|^2_{H^m}}{|v|^{p-2}|w|^{p-2}|v-w|^{p}} \d v \d w \\
& \le \int_0^{2/3} \, \int_0^{2/3-w} \frac{ \left\| f(\cdot+s_1(w+v)) - f(\cdot+s_2(w+v)) \right\|^2_{H^m}}{|w|^{p-2}} \left( \frac{1}{v^{p-2}(v+w)^p} \right) \d v \d w \\
& \leq \int_{0}^{2/3} \frac{ \left\| f(\cdot+s_1\tilde{w}) - f(\cdot+s_2\tilde{w}) \right\|^2_{H^m}}{\tilde{w}^{3p-5}} \underbrace{\left( \int_0^{\infty} \frac{1}{t^{p-2}(1+t)^{p}} \d t \right)}_{<\infty} \d \tilde{w}.
\end{align*}
We are now in the position to apply \eqref{eq:case2} and the statement for \textbf{Case 4}
\begin{align*}
& \iint_D \left\| \left(f(\cdot+v + s_1(w-v)) - f(\cdot+v + s_1(w-v)) \right) \cdot\right.\\
& \qquad \left.\cdot \left(g(\cdot+v + s_1(w-v)) - g(\cdot+v + s_1(w-v)) \right) \right\|_{H^m} \frac {\d v \d w} {|v|^{p-2}|w|^{p-2}|v-w|^{p}} \\
& \leq C(m,p) \|f\|_{H^{m+\frac 32 p -3}} \|g\|_{H^{m+\frac 32 p -3}}
\end{align*}
for some positive constant $C(m,p)$ only dependent on $m$ and $p$
follows.
\end{proof}
\subsection{Estimate of the remainder term \texorpdfstring{$\ensuremath{\widetilde R^{(p)}}$}{R(p)}}
We may find an estimate for of $\partial^l \ensuremath{\widetilde R^{(p)}}(\gamma)$ with respect to a fractional Sobolev norm.
\begin{lemma}
\label{lemma:estimateRemainder}
Let $\gamma \in C^\infty(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$ be parametrized by arc-length and $m > \tfrac 12$.
Then there exist constants $C=C(\gamma,m,n,p)>0$ and
$r= r(\gamma,m,n,p)>0$ such that
\begin{align*}
& \norm[H^m]{\partial^l \ensuremath{\widetilde R^{(p)}}(\gamma)} \\
& \leq C \sum_{j_1 + \ldots + j_{K-1} = l+1} \binom{l+1}{j_1, \ldots, j_{K-1}} \ p^{(3n)}_{j_1} \left( \left\{\frac{(|\alpha|+1)!}{r^{|\alpha|+1}} \right\}_{|\alpha|\leq j_1}, \{ \| \gamma^{(j+1)} \|_{H^{m+ \frac 32 p -3}}\}_{\substack{j =1, \ldots, j_1\\ i = 1, \ldots, 3n}} \right) \\
& \phantom{\le} \cdot \prod_{i=2}^{K-1} {\|\gamma^{(j_i+1)}\|_{H^{m+ \frac 32 p -3}}}.
\end{align*}
Note that the second set of coefficients of $p^{(3n)}_{j_1}$ is independent of $i$, we need that parameter only for dimensional reasons.
\end{lemma}
\begin{proof}
First note that $\ensuremath{\GL^{(p)}} \circ |.|$ is an analytic function away from zero and by the fundamental theorem of calculus
\[
\begin{split}
&\ensuremath{\GL^{(p)}}\left(\frac {|\difference{v,w} \gamma|} {|v-w|}, \frac {|\difference{v,0} \gamma|} {|v|}, \frac {|\difference{w,0} \gamma|} {|w|} \right)\\
&= (\ensuremath{\GL^{(p)}} \circ |.|) \left(\int_0^1 \gamma'(u+tw)\d t, \int_0^1 \gamma'(u+tv)\d t, \int_0^1 \gamma'(u+t(w-v))\d t \right).
\end{split}
\]
Now by the generalized Leibniz formula and Faà di Bruno's formula as in \thref{lem:FaadiBruno}, we observe
\[
\begin{split}
& \frac {\d^{l+1}} {\d u^{l+1}} g_k^p(u,v,w;s_1, \ldots, s_{K-2})
\\
& = \sum_{k_1 + \ldots + k_{K-2} = l+1} \binom{l+1}{k_1, \ldots, k_{K-2}}
\\
&\cdot p^{(3n)}_{k_1}\Biggl(
\set*{ \partial^\alpha (\ensuremath{\GL^{(p)}} \circ |. |) \left(\int_0^1 \gamma'(u+tw)\d t, \int_0^1 \gamma'(u+tv)\d t, \int_0^1 \gamma'(u+t(w-v))\d t \right)}_{\abs{\alpha}\leq k_1},
\\
& \qquad
\left\{ \int_0^1 \gamma^{(j+1)}_i (u+t(x-y)) \d t \right\}_{\substack{i=1,\ldots,n, \, j =1, \ldots, k_1 ,\\ (x,y) = (v,w), (v,0), (w,0)}} \Biggr)
\\
& \cdot \frac {\d^{k_2}} {\d u^{k_2}} \Gamma(u,v,w,s_1,s_2) \\
& \cdot \left( \bigotimes_{i=3}^{K_1} \gamma^{(k_i + 1)}(u+ s_i v) \right) \otimes \left(\bigotimes_{i=K_1 +1}^{K_2} \gamma^{(k_i + 1)}(u + s_i w) \right) \otimes \left(\bigotimes_{i=K_2 +1}^{K-2} \gamma^{(k_i + 1)}(u+v + s_i (w-v)) \right).
\end{split}
\]
We then apply the $H^m$-norm on $\partial^l \ensuremath{\widetilde R^{(p)}}(\gamma)$ for some $m> \tfrac 12$ and obtain by the Banach algebra property of $H^m (\mathbb{R}/\mathbb{Z}, \mathbb{R}^n)$, cf. \eqref{prop:Banachalgebra}, together with \thref{lemma:estimateFaaDiBrunoBanachAlgebra} that
\[
\refstepcounter{equation}\tag{\theequation}
\label{lemma:estimateRemainder:eq:firstStep}
\begin{split}
& \left\| \partial^l \ensuremath{\widetilde R^{(p)}}(\gamma) \right\|_{H^m}
\\
& \leq c(m,n,p) \sum_{k=1}^{k_{\max}} \iint_{D} \idotsint\limits_{[0,1]^K} \left\| \partial^{l+1} g_k^p(\cdot,v,w;s_1, \ldots, s_{K-2})\right\|_{H^m} \d\theta_{1}\cdots\d\theta_{K}\d v \d w
\\
& \leq \tilde c(m,n,p,K) \sum_{k=1}^{k_{\max}} \iint_{D} \idotsint\limits_{[0,1]^K} \sum_{k_1 + \ldots + k_{K-2} = l+1} \binom{l+1}{k_1, \ldots, k_{K-2}}
\\
&\cdot p^{(3 n)}_{k_1}\Biggl(\!
\set*{b^{\abs{\alpha}} \norm*[H^m]{\partial^\alpha (\ensuremath{\GL^{(p)}} \circ |.|) \left(\int_0^1 \gamma'(\cdot+tw)\d t, \int_0^1 \gamma'(\cdot+tv)\d t, \int_0^1 \gamma'(\cdot+t(w-v))\d t \right)}}_{\hspace{-0.4em}\abs{\alpha}\leq k_1}\hspace{-2em},
\\
& \qquad
\left\{ \left\| \int_0^1 \gamma^{(j+1)}_i (\cdot+t(x-y)) \d t \right\|_{H^m}\right\}_{\substack{i=1,\ldots,n, \, j =1, \ldots, k_1 ,\\ (x,y) = (v,w), (v,0), (w,0)}} \Biggr)
\\
& \cdot \left\|\partial^{k_2}\Gamma(\cdot,v,w,s_1,s_2) \right\|_{H^m}
\cdot \prod_{i=3}^{K-2} \norm*[H^m]{\gamma^{(k_i+1)}} \d\theta_{1}\cdots\d\theta_{K}\d v \d w
\end{split}
\]
where $b = b(m) := \sup_{\eta_1,\eta_2 \in H^m(\mathbb{R}/\mathbb{Z},\mathbb{R})} \frac {\norm[H^m]{\eta_1}\norm[H^m]{\eta_2}} {\norm[H^m]{\eta_1\eta_2}}$ is a finite constant stemming from the Banach algebra property and $c(n,m,p)$ is an upper bound for the operator norms of $\cdot \otimes e_o : H^m(\mathbb{R}/\mathbb{Z},\mathbb{R}^n) \to H^m(\mathbb{R}/\mathbb{Z},\mathbb{R}^n)$.
Since $G^{(p)}\circ |.|$ is an analytic function away from the origin and $\gamma$ is smooth with bounded bilipschitz constant \cite[Proposition~2.1]{BlattReiter:2015:Menger},
we get from the proof of \cite[Lemma~7.2]{BlattVorderobermeier:2019:CriticalMoebius} that
$$
b^{\abs{\alpha}}\norm*[H^m]{\partial^\alpha (\ensuremath{\GL^{(p)}} \circ |.|) \left(\int_0^1 \gamma'(\cdot+tw)\d t, \int_0^1 \gamma'(\cdot+tv)\d t, \int_0^1 \gamma'(\cdot+t(w-v))\d t \right)}
\leq C_{G} \frac{(|\alpha|+1)!}{r^{|\alpha|+1}}
$$
for some positive constants $C_{G} = C_{G} (\gamma, m, n, p)$ and $r= r(\gamma,m,n,p)$ which also absorb $b$. Furthermore, we have by the triangle inequality
\[
\left\| \int_0^1 \gamma^{(j+1)}_i (\cdot+t(x-y)) \d t \right\|_{H^m} \leq \| \gamma^{(j+1)}_i \|_{H^m} \leq \| \gamma^{(j+1)} \|_{H^m}.
\]
Note that these bounds are independent of $s_3, \ldots, s_{K-2}$.
To estimate the term $\iint_D \iint_{[0,1]^2} \left\| \partial^{k_2} \Gamma(\cdot,v,w,s_1,s_2) \right\|_{H^m} \d \theta_1 \d \theta_2 \d w \d v$, we first observe by the generalized Leibniz rule
\[
\begin{split}
& \frac {\d^{k_2}} {\d u^{k_2}} \Gamma(u,v,w,s_1,s_2) = \frac{1}{|v|^{p-2}|w|^{p-2}|v-w|^{p}} \\
& \cdot \sum_{o=0}^{k_2} \binom{k_2}{o} \left(\gamma^{(k_2-o+1)}(u+x_1) - \gamma^{(k_2-o+1)}(u+x_2) \right) \otimes \left( \gamma^{(o+1)}(u+x_3) - \gamma^{(o+1)}(u+x_4) \right),
\end{split}
\]
where we have the four cases
\[
\begin{split}
\textbf{Case 1:} & \quad (x_1,x_2,x_3,x_4) =(s_1w,s_1v, s_2w, s_2v), \\
\textbf{Case 2:} & \quad (x_1,x_2)=(x_3,x_4) = (s_1w,s_2w), \\
\textbf{Case 3:} & \quad (x_1,x_2)=(x_3,x_4) = (s_1v, s_2v), \\
\textbf{Case 4:} & \quad (x_1,x_2)=(x_3,x_4) = (v+s_1(w-v), v+s_2(w-v)).
\end{split}
\]
Hence we get by \thref{thm:fracLeibnizrule}
\[
\begin{split}
&\iint_{[0,1]^2} \iint_D \left\| \partial^{k_2} \Gamma(\cdot,v,w,s_1,s_2) \right\|_{H^m} \d w \d v \d \theta_1 \d \theta_2 \\
& \leq C(m,n,p) \sum_{i=1}^n \sum_{j=1}^n \sum_{o=0}^{k_2} \binom{k_2}{o} \iint_{[0,1]^2} \iint_D \frac{1}{|v|^{p-2}|w|^{p-2}|v-w|^{p}} \\
& \cdot \norm[H^m]{\gamma_i^{(k_2-o+1)}(\cdot +x_1) - \gamma_i^{(k_2-o+1)}(\cdot +x_2)} \norm[H^m]{\gamma_j^{(o+1)}(\cdot+x_3) - \gamma_j^{(o+1)}(\cdot +x_4)} \d w \d v \d \theta_1 \d \theta_2\\
& \leq \widetilde C(m,n,p) \sum_{i=1}^n \sum_{j=1}^n \sum_{o=0}^{k_2} \binom{k_2}{o} \iint_{[0,1]^2} \| \gamma_i^{(k_2-o+1)}\|_{H^{m+ \frac 32 p -3}} \|\gamma_j^{(o+1)}\|_{H^{m+ \frac 32 p -3 }} \d \theta_1 \d \theta_2 \\
& \leq \widetilde C(m,n,p) \sum_{o=0}^{k_2} \binom{k_2}{o} \| \gamma^{(k_2-o+1)}\|_{H^{m+ \frac 32 p -3}} \|\gamma^{(o+1)}\|_{H^{m+ \frac 32 p -3 }}
\end{split}
\]
where $C(m,n,p)$ is an upper bound for the operator norms of $e_i\otimes e_j$ on $(H^m(\mathbb{R}/\mathbb{Z},\mathbb{R}))^2$.
We may combine the sum over $o$ and the binomial coefficient $\binom{k_2}{o}$ with the sum over $k_1, \ldots, k_{K-2}$ and the multinomial coefficient $\binom{l+1}{k_1,\ldots,k_{K-2}}$ from \eqref{lemma:estimateRemainder:eq:firstStep} into the sum over $j_1 + \ldots + j_{K-1}$ and the multinomial coefficient $\binom{l+1}{j_1, \ldots, j_{K-1}}$ by setting $j_2 := o$, $j_{K-1} := k_2-o$ and $j_i := k_i$ for all other $i$.
This allows us to integrate the two derivatives of $\gamma$ into the bigger product.
Then, the
statement follows from the embedding $H^s(\mathbb{R}/\mathbb{Z},\mathbb{C}^n) \subseteq H^t(\mathbb{R}/\mathbb{Z},\mathbb{C}^n)$ for any $s>t$, $s,t \geq 0$.
Since these estimates are uniform over all $g_k$, we may use an additional factor of $k_{\max}$ in order to leave out the sum.
\end{proof}
Note that $\Gamma$ is the only term for whose upper bound we need the higher-order $H^{m+\frac 3 2 p -3}$-norms.
We estimated all $H^m$-norms by these in order to get a simpler structure which will serve us in the proof of the main theorem, especially in \thref{lemma:iteratedEstimateMainTerm}.
|
\section{Introduction}
\label{sec:intro}
Recent developments in deep learning have achieved great success on semantic segmentation tasks with the help of deep convolutional neural networks (CNNs) and rich pixel-level annotations.
However, collecting a large-scale pixel-level annotated dataset requires intensive human labor, which is both expensive and time-consuming.
To end this limitation, weakly-supervised semantic segmentation (WSSS) using only image-level labels has recently attracted much attention.
One problem with using image-level annotations is that we have no information about the location of the target object; we only know whether the object is present in the image or not.
This makes semantic segmentation learning challenging.
To learn pixel-level semantic knowledge from image-level labels, it is common practice to use localization maps obtained from the classification network using class activation maps (CAMs)~\cite{zhou2016learning}.
Specifically, the discriminative region for each target class provided by CAMs is used as pixel-level supervision for segmentation network training.
However, this discriminative region is usually very sparse and only covers a small part of the object, which is not enough for semantic segmentation learning as shown in the second column in Figure \ref{figure:concept}.
Therefore, most studies in the weakly-supervised semantic segmentation field focus on expanding the object region to produce dense localization maps.
One of the recent approaches is image-level and feature-level erasure of discriminative parts~\cite{wei2017object, li2018tell, hou2018self}.
This approach strictly erases discriminative parts, letting the network focus on other non-discriminative parts.
However, they not only tend to produce undesired true negative regions when most of the discriminative parts are erased but also require a lot of additional parameters for multiple classifiers or multiple branches.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{images/concept_horizontal.pdf}
\caption{Visual comparisons of localization maps produced by original CAM, DRS, and DRS$\dagger$. DRS$\dagger$ denotes that the refinement learning is applied.}
\label{figure:concept}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{images/DRS_main.pdf}
\caption{
Diagram of our discriminative region suppression (DRS) module. DRS suppresses intermediate feature maps, diffusing the attention into adjacent non-discriminative parts.
The max-element extractor extracts $K$ maximum elements from intermediate feature maps.
These $K$ maximum elements are the maximum points of each discriminative region and are considered as starting points to be suppressed.
For convenience, $K$ maximum elements are illustrated in 5 purple points.
The controller predicts control values, which determine how much to suppress feature maps from these $K$ maximum elements.
These $K$ control values are illustrated in 5 blue arrows and the length of the arrow means how much suppress feature maps from the corresponding maximum element.
Using these $K$ maximum elements and $K$ control values, the suppressor suppresses discriminative regions and spreads the attention into adjacent non-discriminative parts.
}
\label{figure:main}
\end{figure*}
In this paper, we propose discriminative region suppression (DRS) module, which is a simple and efficient yet effective and novel approach for generating dense localization maps.
The goal of DRS is to suppress discriminative regions, not to erase them, so that attention spreads to adjacent non-discriminative regions; this mild approach helps the classifier effectively expand discriminative object regions.
DRS module consists of three components as depicted in Figure \ref{figure:main}: max-element extractor, suppression controller, and suppressor.
These components work together to produce dense localization maps by reducing the attention gap between discriminative regions and adjacent non-discriminative regions.
DRS not only effectively expands the object regions without generating much noise, but also can be plugged into any network with few or no additional parameters.
Although we can obtain dense segmentation labels from the classification network equipped with DRS, it does not recover missing parts or weak attention by itself because the objective of the classification network is classification, not localization.
To address this issue, we introduce an additional training strategy, named localization map refinement learning, inspired by \cite{jiang2019integral}
Localization map refinement learning induces self-enhancement of localization maps by recovering missing or weak attention region.
In Figure \ref{figure:concept}, we compare some results of DRS and DRS$\dagger$, where DRS$\dagger$ denotes that refinement learning is applied.
Following the convention, we generate pseudo segmentation labels from our dense localization maps and evaluate on weakly-supervised semantic segmentation task.
On the PASCAL VOC 2012 segmentation benchmark, we achieve mIoU 71.4\% on the testset using only image-level labels.
In addition, extensive experiments demonstrate the effectiveness of our approach.
In summary, the contributions of our work are as follows:
\begin{itemize}
\item We introduce a simple, effective, and novel approach for weakly-supervised semantic segmentation named discriminative region suppression (DRS) module, which requires few or no additional parameters and can be easily plugged into any network.
\item DRS effectively and efficiently suppresses discriminative regions to generate dense localization maps, bridging the gap between discriminative regions and adjacent non-discriminative regions.
\item For self-enhancement of localization maps, we introduce an additional training strategy, named localization map refinement learning.
\item Extensive experiments and analyses demonstrate the effectiveness of our DRS module and we achieve competitive performance on Pascal VOC 2012 segmentation benchmark using only image-level labels.
\end{itemize}
\section{Related Work}
Most recent studies on semantic segmentation using image-level labels as weak supervision utilize CAMs \cite{zhou2016learning} to localize object regions and focus on expanding them to non-discriminative parts of the objects.
To this end, AE-PSL \cite{wei2017object}, GAIN \cite{li2018tell}, and SeeNet \cite{hou2018self} propose erasing techniques to generate dense localization maps. However, these erasure-based approaches usually require multiple classifiers and complicated training procedures.
Moreover, erasing most of the discriminative regions may introduce true negative regions and confuse the classifier.
To avoid the repetitive training procedures of AE-PSL \cite{wei2017object}, MDC \cite{wei2018revisiting} propose a multi-dilated convolution block in which the receptive fields of various sizes capture different patterns.
As a more generalized approach, FickleNet \cite{lee2019ficklenet} aggregate diverse localization maps produced by stochastic feature selection.
Although they effectively expand the activated regions, some falsely labeled regions outside the object tend to be identified because the receptive fields of these methods are not adaptive to object size.
The recently proposed OAA \cite{jiang2019integral} accumulates attention maps at different training epochs and introduces integral attention learning to enhance attention maps.
However, it may produce undesired attention regions due to training instability in the early stage.
Some other works \cite{ahn2018learning, huang2018weakly, shimoda2019self} adopt a region-growing technique to expand initial regions.
More recently, RRM~\cite{zhang2020reliability} proposed a fully end-to-end network for joint training of classification and segmentation, and SGAN~\cite{yao2020saliency} proposed a self-attention network guided by saliency priors that can produce dense and accurate localization maps from rich contextual information.
BES~\cite{chen2020weakly} explores object boundaries to refine the semantic segmentation output.
ICD~\cite{fan2020learning} proposes an intra-class discriminator approach to separate foreground objects and the background within the same image-level class.
\begin{figure*}[t]
\centering
\includegraphics[width=0.78\linewidth]{images/network.pdf}
\caption{
Overview of the proposed method. (a) Classification network with DRS for obtaining localization maps, (b) localization map refinement learning, and (c) weakly-supervised semantic segmentation.
(a), (b), and (c) are executed sequentially, not simultaneously. Note that GAP in (a) means the global average pooling layer.
}
\label{figure:network}
\end{figure*}
\section{Method}
The overview of our method is illustrated in Figure \ref{figure:network}.
We sequentially train three different networks for classification, refinement, and segmentation.
After training the classification network with the discriminative region suppression (DRS) module, we produce dense localization maps.
Using these dense localization maps as ground truth labels for refinement learning, we train the refinement network to produce refined localization maps.
Then, pseudo segmentation labels are generated from the refined localization maps and used for training the semantic segmentation network.
We measure the segmentation performance to evaluate the quality of our localization maps.
\subsection{Observation}
\label{method:1}
We produce localization maps from the class-specific feature maps of the last convolutional layer, which have been proven by \cite{zhang2018adversarial} to be mathematically equivalent to CAMs.
We employ the VGG-16~\cite{simonyan2014very} as our classification network to produce localization maps.
To be specific, we employ modified VGG-16, where all fully connected layers are removed. On top of it, three convolutional layers with 512 channels and kernel size 3, and a convolutional layer with $C$ channels and kernel size 1 are added. Here $C$ is the number of categories.
This network produces output feature maps $F \in \mathbb{R}^{H_{out} \times W_{out} \times C}$ and classification score $P=\sigma(GAP(F))$ from input image.
$H_{out}$ and $W_{out}$ are the height and width of output feature maps, respectively; $GAP(\cdot)$ is the global average pooling; and $\sigma(\cdot)$ is the sigmoid function.
For each target category $c$, $c$-th localization map $M^{c}$ is defined as the normalized $c$-th feature map $F^{c}$:
\begin{equation}
M^c=\frac{ReLU(F^c)}{max(F^c)}.
\end{equation}
From the definition of $M$, we observe that discriminative object regions are identified with relatively high values on the feature maps $F$.
Based on this observation, we regard the high-value areas on feature maps as discriminative regions.
\begin{algorithm}[t]
\SetKwInOut{Input}{Input}
\Input{
Intermediate feature maps $X \in R^{H \times W \times K}$
}
{
\SetKwInOut{Output}{Output}
\Output{
Suppressed feature maps : $\hat{X} \in R^{H \times W \times K}$
}
\vspace{3mm}
$X_{max}$ $\gets$ $extractor(X)$ \hspace{3mm}//$X_{max} \in \mathbb{R}^{1 \times 1 \times K}$ \\
\vspace{2mm}
$G$ $\gets$ $controller(X)$ \hspace{8mm}// $G \in [0, 1]^{1 \times 1 \times K}$ \\
\vspace{1mm}
$\tau$ $\gets$ $X_{max} \cdot G$ \hspace{12mm}// upper bound, $\tau \in \mathbb{R}^{1 \times 1 \times K}$ \\
\vspace{2mm}
$\tau$ $\gets$ \textit{expand to the same shape of $X$} \\
\vspace{2mm}
$\hat{X}$ $\gets$ $min(X, \tau)$ \hspace{13mm}// suppressor \\
}
\caption{Discriminative Region Suppression}
\label{DRS_algorithm}
\end{algorithm}
\subsection{Discriminative Region Suppression}
\label{method:2}
To produce dense localization maps, we propose discriminative region suppression (DRS) module.
The main problem of segmentation label generation using CAMs is that discriminative regions only appear partially and sparsely, as shown in the second column of Figure~\ref{figure:concept}.
To address this issue, DRS aims to spread the attention on discriminative regions to adjacent non-discriminative regions.
Specifically, DRS suppresses the attention on discriminative regions, allowing the network to focus on non-discriminative regions.
Let $X \in \mathbb{R}^{H \times W \times K}$ be an intermediate feature map, where $H$, $W$, and $K$ are the height, width, and the number of channels of $X$.
DRS module consists of three parts: max-element extractor, suppression controller, and suppressor.
The max-element extractor extracts $K$ maximum elements from the intermediate feature map $X$ using global max pooling.
The output of the extractor is denoted as $X_{max} \in \mathbb{R}^{1 \times 1 \times K}$.
Based on the observation, these $K$ maximum elements are regarded as the criteria of discriminative regions and considered as starting points to be suppressed.
The suppression controller determines how much to suppress discriminative regions.
In detail, it generates $G \in [0, 1]^{1 \times 1 \times K}$ and each $k$-th control value in $G$ determines the amount of suppression in $X$ with respect to the corresponding $k$-th maximum element.
Using the $K$ maximum elements and $K$ control values, the suppressor suppresses discriminative regions. Specifically, element-wise multiplication of $X_{max}$ and $G$ is regarded as the upper bound of $X$, denoted as $\tau = X_{max} \cdot G$, $\tau \in \mathbb{R}^{1 \times 1 \times K}$.
The regions in $X$ above this upper bound are regarded as discriminative regions to be suppressed. After the upper bound $\tau$ is expanded to the same shape of $X$, the element-wise minimum operation is applied on $X$ and $\tau$ to suppress discriminative regions.
For example, if the $k$-th control value is 0.7, $X^k$ is suppressed until no element exceeds 70\% of the $k$-th maximum value.
In this way, the suppressor bridges the gap between discriminative regions and adjacent non-discriminative regions.
The whole process of DRS is described in Algorithm~\ref{DRS_algorithm} and illustrated in Figure~\ref{figure:main}.
For the suppression controller, there are two types of controller: learnable controller and non-learnable controller.
If the suppression power is too strong, the discriminative feature extraction power is weakened.
The learnable controller adaptively balances between discriminative feature extraction power and suppression power of the classification network.
Formally, the output of the learnable controller is
\begin{equation}
G = \sigma(f(GAP(X);\theta)),
\end{equation}
where $f$ is a fully connected layer, $\theta$ is a learnable parameter of the controller, and $G \in [0, 1]^{1 \times 1 \times K}$.
Since $\theta$ is trained with the classification objective, DRS with a learnable controller adaptively suppresses
discriminative regions so as not to damage the discriminative feature extraction power much.
To produce even more dense localization maps at the expense of discriminative feature extraction power, we forcibly suppress discriminative regions; this is the goal of a non-learnable controller.
For the non-learnable controller, each element of $G$ is set to a constant value $\delta$.
We set the hyperparameter $\delta$ to a value between 0 and 1, and a lower $\delta$ means more intense suppression resulting in more dense localization maps.
Compared to the learnable controller, the non-learnable controller does not require additional training parameters but requires a hyperparameter $\delta$.
In the experiment section, we analyze both learnable and non-learnable controller with quantitative and qualitative results.
Figure \ref{figure:network} (a) illustrates the process of obtaining dense localization maps from the classification network with DRS.
As shown in the third column of Figure~\ref{figure:concept}, DRS reduces the gap between the activation of discriminative regions and adjacent non-discriminative regions to obtain dense localization maps.
Note that DRS can be plugged into any layer of a network.
\subsection{Localization Map Refinement Learning}
\label{method:3}
Although DRS helps produce dense localization maps, the DRS itself lacks the ability to recover missing parts of the target objects or enhance weak attention in adjacent non-discriminative regions because the goal of the classification network is essentially classification ability, not localization map generation.
Motivated by \cite{jiang2019integral}, we introduce an additional learning strategy for localization map refinement to solve the above limitations.
This learning strategy for self-enhancement of localization maps is called localization map refinement learning, denoted as DRS$\dagger$.
After training the classification network with DRS, we exploit the output localization maps $M \in [0,1]^{H_{out} \times W_{out} \times C}$ as the ground truth localization maps for refinement learning.
The network for refinement learning, called refinement network, is based on the VGG-16; all fully-connected layers are removed and three convolutional layers with 512 channels and kernel size 3, and a convolutional layer with C channels and kernel size 1 are appended.
The refinement network directly produces refined localization maps $N \in \mathbb{R}^{H_{out} \times W_{out} \times C}$, which have the same shape as $M$.
We adopt the mean squared error (MSE) loss function as the refinement loss for the refinement network.
Refinement learning is depicted in Figure \ref{figure:network} (b).
Benefiting from refinement learning, we can obtain more dense and high-quality localization maps through self-enhancement, as shown in Figure~\ref{figure:concept} (DRS$\dagger$ $v.s.$ DRS)
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{images/delta_v2.pdf}
\caption{
Visualization of localization maps of two types of controllers and the non-learnable controller with different $\delta$.
}
\label{figure:controller}
\end{figure*}
\subsection{Weakly-Supervised Semantic Segmentation}
\label{method:4}
Using our dense localization maps obtained from the refinement network, we generate pseudo segmentation labels and use them as weak-supervision for the semantic segmentation network.
We generate pseudo segmentation labels using object cues and background cues.
We extract object cues from the localization maps by taking the pixels whose values are higher than $\alpha$ and extract background cues using salient object detection method~\cite{liu2019simple}, motivated by \cite{wei2017object, wei2018revisiting}; the pixels with saliency values lower than $\beta$ are taken as background.
Those who belong to neither of the cues are ignored.
Following the convention, we train the segmentation network such as \cite{chen2014semantic, chen2017deeplab} using the generated pseudo segmentation labels, as illustrated in Figure \ref{figure:network} (c).
The segmentation performance is compared with other methods using the same segmentation network to evaluate the quality of pseudo segmentation labels.
\section{Experiments}
\subsection{Dataset and Evaluation Metrics}
We demonstrate the effectiveness of the proposed approach on the PASCAL VOC 2012 segmentation benchmark dataset \cite{Everingham2014ThePV}, which contains 20 object categories and one background category.
Following the common practice in previous works, the training set is augmented to 10,582 images.
We evaluate the performance of our model using the mean intersection-over-union (mIoU) metric and compare it with other state-of-the-art methods on the validation (1,449 images) and test set (1,456 images).
For the test results, we submit the prediction outputs to the official PASCAL VOC evaluation server.
\subsection{Implementation details}
For the classification network, we adopt the modified VGG-16 with DRS plugged into every layer, as mentioned in the method section.
Its parameters are initialized by the VGG-16~\cite{simonyan2014very} pre-trained on ImageNet~\cite{deng2009imagenet} except for the additional convolutional layers.
We train the classification network with binary cross-entropy loss using the SGD optimizer with a weight decay of 5e-4 and a momentum of 0.9.
The initial learning rate is set to 1e-3 and is decreased by a factor of 10 at epoch 5 and 10.
For data augmentation, we apply a random crop with 321$\times$321 size, random horizontal flipping, and random color jittering.
We use a batch size of 5 and train the classification network for 15 epochs.
We optimize the refinement network for the refinement learning with MSE loss using Adam~\cite{kingma2014adam} optimizer with a learning rate of 1e-4. The batch size is 5, the total training epoch is 15, and the learning rate is dropped by a factor of 10 at epoch 5 and 10.
We apply the data same augmentation strategy as in the classification network.
For the segmentation network, we experiment with three architectures: DeepLab-Large-FOV~\cite{chen2014semantic} with VGG-16 and DeepLab-Large-FOV with ResNet-101~\cite{he2016deep} backbones, and DeepLab-ASPP~\cite{chen2017deeplab} with ResNet-101 backbone.
When generating pseudo segmentation labels, we empirically choose $\alpha = 0.2$ for object cues and $\beta = 0.06$ for background cues.
Our method is implemented on Pytorch \cite{paszke2017automatic}.
We use the DeepLab-Large-FOV code\footnote{\url{https://github.com/wangleihitcs/DeepLab-V1-PyTorch}} and DeepLab-ASPP code\footnote{\url{https://github.com/kazuto1011/deeplab-pytorch}} implemented based on the Pytorch framework, following the same hyperparameter settings for training and the conditional random field (CRF)~\cite{krahenbuhl2011efficient} as the original publications.
All experiments are performed on NVIDIA TITAN XP.
\begin{comment}
\begin{table}[t]
\centering
\caption{
Effect of the two types of controllers and the hyperparameter $\delta$ for the non-learnable controller.
}
\label{tab:controller}
\begin{adjustbox}{max width=\linewidth}
\begin{tabular}{c|ccccccc}
\hline
\multicolumn{8}{c}{suppression controller} \\
\hline
learnable & \multicolumn{7}{c}{non-learnable} \\
& $\delta$=0.90 & $\delta$=0.80 & $\delta$=0.70 & $\delta$=0.60 & $\delta$=0.55 & $\delta$=0.50 & $\delta$=0.40 \\
\hline
\textbf{62.9\%} & 51.9\% & 56.0\% & 58.7\% & 62.3\% & \textbf{62.8\%} & 62.3\% & 59.6\% \\
\hline
\end{tabular}
\end{adjustbox}
\end{table}
\end{comment}
\begin{table}[t]
\centering
\begin{adjustbox}{max width=\linewidth}
\begin{tabular}{c|c}
\hline
suppression controller & mIoU \\
\hline
learnable & \textbf{62.9\%} \\
non-learnable with $\delta$=0.90 & 51.9\% \\
non-learnable with $\delta$=0.80 & 56.0\% \\
non-learnable with $\delta$=0.70 & 58.7\% \\
non-learnable with $\delta$=0.60 & 62.3\% \\
non-learnable with $\delta$=0.55 & \textbf{62.8\%} \\
non-learnable with $\delta$=0.50 & 62.3\% \\
non-learnable with $\delta$=0.40 & 59.6\% \\
\hline
\end{tabular}
\end{adjustbox}
\caption{
Effect of the two types of controllers and the hyperparameter $\delta$ for the non-learnable controller.
}
\label{tab:controller}
\end{table}
\subsection{Analysis}
To analyze the effectiveness of the proposed method, we conduct several experiments.
Following the convention of weakly-supervised semantic segmentation, we measure the mIoU score of the segmentation network outputs to evaluate the quality of our localization maps.
For all experiments in this section, we adopt the DeepLab-Large-FOV with VGG-16 as the segmentation network and measure the mIoU score on the VOC 2012 validation set.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{images/activation.pdf}
\caption{
Visualization of feature maps on each layer. Note that element-wise averaging and normalization are applied to feature maps of each layer for visualization.
}
\label{figure:activation}
\end{figure}
\noindent\textbf{Suppression controller.}
In the method section, we introduced two types of suppression controller: learnable and non-learnable controller.
We investigate both controllers and the effect of the parameter $\delta$ for the non-learnable controller using visualization and quantitative analysis. For this analysis, we plugin the DRS to all layers of the classification network and skip the refinement learning procedure for precise effect analysis.
In the case of a non-learnable controller, we set the same $\delta$ for all layers.
Firstly, we analyze the effect of $\delta$ for the non-learnable controller and compare the output localization maps of each $\delta$ in Figure~\ref{figure:controller}.
When $\delta$ is 0.90, the localization map is mostly activated in the head of the cat.
Consequently, the gap between discriminative regions and adjacent non-discriminative regions is large, resulting in sparse localization maps.
As the $\delta$ gets smaller, activation at the body of the cat becomes higher, and the activation gap between the head and the body of the cat is smaller.
This indicates as the $\delta$ value decreases, the discriminative regions are further suppressed and the gap between discriminative regions and non-discriminative regions becomes smaller, resulting in dense localization maps.
However, if the $\delta$ value is too low ($i.e.,$ too much suppression), the gap between the background and the foreground becomes very small, resulting in a noisy localization map as shown in the rightmost result of Figure~\ref{figure:controller}.
Therefore, it is important to set an appropriate $\delta$ value for the non-learnable controller.
The quantitative results in Table \ref{tab:controller} support our arguments.
The non-learnable controller with $\delta=0.55$ achieves better performance than that of $\delta=0.90$ (62.8\% $v.s.$ 51.9\%), but in the case of over-suppression, $e.g.$ the non-learnable controller with $\delta=0.40$, the performance is rather worse than that of $\delta=0.55$ (59.6\% $v.s.$ 62.8\%).
Through this experiment, we found that $\delta=0.55$ yields the best mIoU performance.
In the case of a learnable controller, it suppresses without generating much noise, creating moderately dense localization maps as in the leftmost of Figure~\ref{figure:controller}.
Compared to the non-learnable controller with $\delta=0.55$, the learnable controller produces similar mIoU performance (62.9\% $v.s.$ 62.8\%) and localization maps.
However, the classification accuracy of the learnable controller is much higher (72.6\% $v.s.$ 68.7\%).
From these results, we can notice that the learnable controller adaptively balances between the discriminative feature extraction power and the suppression power, whereas the non-learnable controller forcibly increases the suppression power at the expense of the feature extraction power.
Note that the learnable controller is free from hyperparameter ($i.e.$, $\delta$) tuning, but requires additional training parameters (from 21.8M to 24.4M training parameters).
\begin{table}[t]
\centering
\begin{adjustbox}{max width=\linewidth}
\begin{tabular}{cccccc|c}
\hline
layer1 & layer2 & layer3 & layer4 & layer5 & layer6 & mIoU \\
\hline
- & - & - & - & - & - & 50.1\% \\
- & - & - & - & - & \checkmark & 60.8\% \\
- & - & - & - & \checkmark & \checkmark & 62.2\% \\
- & - & - & \checkmark & \checkmark & \checkmark & 62.8\% \\
- & - & \checkmark & \checkmark & \checkmark & \checkmark & \textbf{62.9\%} \\
- & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & 62.7\% \\
\checkmark & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & \textbf{62.9\%} \\
\checkmark & \checkmark & \checkmark & \checkmark & \checkmark & - & 58.2\% \\
\checkmark & \checkmark & \checkmark & \checkmark & - & - & 53.6\% \\
\hline
\end{tabular}
\end{adjustbox}
\caption{
Effect of DRS in each layer. \checkmark means DRS is applied.
}
\label{tab:each_layer}
\end{table}
\begin{table}[t]
\centering
\begin{adjustbox}{max width=\linewidth}
\begin{tabular}{c|c|c}
\hline
suppression controller & without refine & with refine \\
\hline
learnable & 62.9\% & 63.5\%\\
non-learnable ($\delta$=0.55) & 62.8\% & 63.6\% \\
\hline
\end{tabular}
\end{adjustbox}
\caption{
Effect of localization map refinement learning.
}
\label{tab:refine}
\end{table}
\noindent\textbf{Effect of DRS on each layer.}
To observe the effect of DRS on each layer, we employ two analytical methods: visualization and quantitative analysis.
For this analysis, we use the DRS module with the learnable controller and skip refinement learning for precise effect analysis.
For visualization, we apply element-wise averaging and normalization from 0 to 1 on feature maps of every layer. Figure \ref{figure:activation} shows the visualization results of the original CAM and our DRS-plugged classification network.
In lower-level layers ($i.e.$, from layer1 to layer3), we notice that the effect of DRS is minor because a network mainly focuses on the local features ($e.g.$, edge) where the gap between the discriminative and adjacent non-discriminative regions tends to be extremely large.
Meanwhile, in higher-level layers ($i.e.$, from layer4 to layer6), a network mostly focuses on the global features ($e.g.$, head of a bird) where the gap between the discriminative and adjacent non-discriminative regions is relatively small.
In this case, the effect of DRS becomes significant because it suppresses the activation of discriminative regions and expands the attention to non-discriminative regions.
For quantitative analysis, we plug in and out DRS at each layer and evaluate the performance of each case.
The results in Table \ref{tab:each_layer} show that the more we plugin the DRS at higher-level layers, the higher performance (from 50.1\% to 62.9\%).
On the other hand, applying DRS in lower-level layers has little effect (62.9\% $v.s.$ 62.7\%).
In addition, when we plug-out the DRS at higher-level layers, the performance significantly decreases (from 62.9\% to 58.2\% and 53.6\%).
From these results, we can conclude that DRS is more effective to produce dense localization maps when applied in higher-level layers.
\noindent\textbf{Improvement through refinement learning.}
As mentioned in the method section, localization map refinement learning gives a self-enhancement effect to produce high-quality dense localization maps as in Figure~\ref{figure:concept}.
The improved mIoU performance is reported in Table~\ref{tab:refine} (+0.6\% and +0.8\%).
In addition, Figure \ref{figure:result} shows some segmentation results on the PASCAL VOC 2012, where both DRS and DRS$\dagger$ show satisfactory results, but DRS$\dagger$ leads to better segmentation results.
Note that the learnable controller is used for Figure \ref{figure:concept} and \ref{figure:result}.
\begin{table}[t]
\centering
\begin{adjustbox}{max width=\linewidth}
\begin{tabular}{ccccc}
\hline
Method & $\mathcal{S}$ & val & test \\
\hline
\multicolumn{4}{c}{\textbf{Segmentation Network : DeepLab-Large-FOV (VGG-16)}} \\
\hline
AE-PSL~\cite{wei2017object} & \checkmark & 55.0\% & 55.7\% \\
GAIN~\cite{li2018tell} & \checkmark & 55.3\% & 56.8\% \\
MCOF~\cite{wang2018weakly} & \checkmark & 56.2\% & 57.6\% \\
AffinityNet~\cite{ahn2018learning} & - & 58.4\% & 60.5\% \\
SeeNet~\cite{hou2018self} & \checkmark & 61.1\% & 60.7\% \\
MDC~\cite{wei2018revisiting} & \checkmark & 60.4\% & 60.8\% \\
RRM~\cite{zhang2020reliability} & - & 60.7\% & 61.0\% \\
FickleNet~\cite{lee2019ficklenet} & \checkmark & 61.2\% & 61.8\% \\
OAA~\cite{jiang2019integral} & \checkmark & 63.1\% & 62.8\% \\
ICD~\cite{fan2020learning} & \checkmark & \textbf{64.0\%} & 63.9\% \\
BES~\cite{chen2020weakly} & - & 60.1\% & 61.1\% \\
\rowcolor{Gray} Ours (learnable) & \checkmark & 63.5\% & \textbf{64.5\%} \\
\rowcolor{Gray} Ours (non-learnable) & \checkmark & 63.6\% & 64.4\% \\
\hline
\multicolumn{4}{c}{\textbf{Segmentation Network : DeepLab-Large-FOV (ResNet-101)}} \\
\hline
MCOF~\cite{wang2018weakly} & \checkmark & 60.3\% & 61.2\% \\
SeeNet~\cite{hou2018self} & \checkmark & 63.1\% & 62.8\% \\
AffinityNet~\cite{ahn2018learning} & - & 61.7\% & 63.7\% \\
FickleNet~\cite{lee2019ficklenet} & \checkmark & 64.9\% & 65.3\% \\
RRM~\cite{zhang2020reliability} & - & 66.3\% & 65.5\% \\
OAA~\cite{jiang2019integral} & \checkmark & 65.2\% & 66.4\% \\
ICD~\cite{fan2020learning} & \checkmark & \textbf{67.8\%} & \textbf{68.0\%} \\
\rowcolor{Gray} Ours (learnable) & \checkmark & 66.5\% & 67.5\% \\
\rowcolor{Gray} Ours (non-learnable) & \checkmark & 66.8\% & 67.4\% \\
\hline
\multicolumn{4}{c}{\textbf{Segmentation Network : DeepLab-ASPP (ResNet-101)}} \\
\hline
DSRG~\cite{huang2018weakly} & \checkmark & 61.4\% & 63.2\% \\
BES~\cite{chen2020weakly} & - & 65.7\% & 66.6\% \\
SGAN~\cite{yao2020saliency} & \checkmark & 67.1\% & 67.2\% \\
\rowcolor{Gray} Ours (learnable) & \checkmark & 70.4\% & 70.7\%\\
\rowcolor{Gray} Ours (non-learnable) & \checkmark & \textbf{71.2\%} & \textbf{71.4\%} \\
\hline
\end{tabular}
\end{adjustbox}
\caption{
Comparison of state-of-the-art weakly-supervised semantic segmentation methods on the Pascal VOC 2012 dataset. $\mathcal{S}$ means the saliency map is used for extra guidance.
}
\label{tab:sota}
\vspace{-1mm}
\end{table}
\subsection{State-of-the-arts Comparison}
We compare our approach (DRS) with other state-of-the-art weakly-supervised semantic segmentation methods that use only image-level labels as supervision.
For comparison, we apply the DRS module to all layers of the classification network and perform refinement learning.
We report the performances of both learnable controller and non-learnable controller with $\delta=0.55$.
Table \ref{tab:sota} shows the mIoU performance comparison on the PASCAL VOC 2012 validation set and test set.
We fairly compare the performance of each of the three architectures of the semantic segmentation network with other works using the same network.
Note that $\mathcal{S}$ in Table \ref{tab:sota} indicates whether the saliency map is used as extra guidance.
As shown in Table \ref{tab:sota}, DRS outperforms erasing-based methods ($e.g.,$ AE-PSL~\cite{wei2017object}, GAIN~\cite{li2018tell}, SeeNet~\cite{hou2018self}), showing that suppression is more effective than erasing.
Compared to the recent state-of-the-art methods, we achieve competitive performance despite our simplicity.
In contrast to some works ($e.g.,$ DSRG~\cite{huang2018weakly}, FickleNet~\cite{lee2019ficklenet}, AffinityNet~\cite{ahn2018learning}, BES~\cite{chen2020weakly}) where CRF in the training stage slows down the training process, our method does not apply CRF during learning, thereby achieving high performance with short training time.
Although ICD~\cite{fan2020learning} achieves higher mIoU scores using an intra-class discriminator approach for separating foreground and background within the same image-level class, it requires a careful training strategy for stable optimization.
Unlike these methods, our approach enables fast and stable training procedure and is the simplest and the most effective way to achieve high segmentation performance.
The highlighted rows in Table \ref{tab:sota} show that the learnable and non-learnable controllers are both effective, with only a marginal difference in performance.
As mentioned in the analysis section, there is a trade-off between the two controller types, so we can choose depending on the situation.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{images/qualitative_results_1.pdf}
\caption{
Qualitative results on the PASCAL VOC 2012 validation set.
}
\label{figure:result}
\vspace{-1mm}
\end{figure}
\section{Conclusion}
In this paper, we propose a novel approach called DRS for enlarging the object regions highlighted by localization maps.
DRS propagates the initial attention to non-discriminative regions, generating dense localization maps.
The main advantage of our approach is that it is intuitive, efficient, and easily applicable to any classification network.
Together with refinement learning, our proposed method successfully generates dense segmentation labels that cover the entire target objects.
When applied to a weakly-supervised segmentation task, it achieves 71.4\% mIoU on pascal VOC segmentation benchmark using only image-level labels as weak supervision.
|
\section{Introduction}
After receiving paper reviews, authors may optionally submit a rebuttal to address the reviewers' comments, which will be limited to a {\bf one page} PDF file. Please follow the steps and style guidelines outlined below for submitting your author response.
Note that the author rebuttal is optional and, following similar guidelines to previous CVPR conferences, it is meant to provide you with an opportunity to rebut factual errors or to supply additional information requested by the reviewers. It is NOT intended to add new contributions (theorems, algorithms, experiments) that were not included in the original submission. You may optionally add a figure, graph or proof to your rebuttal to better illustrate your answer to the reviewers' comments.
Per a passed 2018 PAMI-TC motion, reviewers should not request additional experiments for the rebuttal, or penalize authors for lack of additional experiments. This includes any experiments that involve running code, e.g., to create tables or figures with new results. \textbf{Authors should not include new experimental results in the rebuttal}, and reviewers should discount any such results when making their final recommendation. Authors may include figures with illustrations or comparison tables of results reported in the submission/supplemental material or in other papers.
The rebuttal must adhere to the same blind-submission as the original submission and must comply with this rebuttal-formatted template.
\subsection{Response length}
Author responses must be no longer than 1 page in length including any references and figures. Overlength responses will simply not be reviewed. This includes responses where the margins and formatting are deemed to have been significantly altered from those laid down by this style guide. Note that this \LaTeX\ guide already sets figure captions and references in a smaller font.
\section{Formatting your Response}
{\bf Make sure to update the paper title and paper ID in the appropriate place in the tex file.}
All text must be in a two-column format. The total allowable width of the text
area is $6\frac78$ inches (17.5 cm) wide by $8\frac78$ inches (22.54 cm) high.
Columns are to be $3\frac14$ inches (8.25 cm) wide, with a $\frac{5}{16}$ inch
(0.8 cm) space between them. The top margin should begin
1.0 inch (2.54 cm) from the top edge of the page. The bottom margin should be
1-1/8 inches (2.86 cm) from the bottom edge of the page for $8.5 \times
11$-inch paper; for A4 paper, approximately 1-5/8 inches (4.13 cm) from the
bottom edge of the page.
Please number all of your sections and any displayed equations. It is important
for readers to be able to refer to any particular equation.
Wherever Times is specified, Times Roman may also be used. Main text should be
in 10-point Times, single-spaced. Section headings should be in 10 or 12 point
Times. All paragraphs should be indented 1 pica (approx. 1/6 inch or 0.422
cm). Figure and table captions should be 9-point Roman type as in
Figure~\ref{fig:onecol}.
List and number all bibliographical references in 9-point Times, single-spaced,
at the end of your response. When referenced in the text, enclose the citation
number in square brackets, for example~\cite{Authors14}. Where appropriate,
include the name(s) of editors of referenced books.
\begin{figure}[t]
\begin{center}
\fbox{\rule{0pt}{1in} \rule{0.9\linewidth}{0pt}}
\end{center}
\caption{Example of caption. It is set in Roman so that mathematics
(always set in Roman: $B \sin A = A \sin B$) may be included without an
ugly clash.}
\label{fig:long}
\label{fig:onecol}
\end{figure}
\subsection{Illustrations, graphs, and photographs}
All graphics should be centered. Please ensure that any point you wish to make is resolvable in a printed copy of the response. Resize fonts in figures to match the font in the body text, and choose line widths which render effectively in print. Many readers (and reviewers), even of an electronic copy, will choose to print your response in order to read it. You cannot insist that they do otherwise, and therefore must not assume that they can zoom in to see tiny details on a graphic.
When placing figures in \LaTeX, it's almost always best to use \verb+\includegraphics+, and to specify the figure width as a multiple of the line width as in the example below
{\small\begin{verbatim}
\usepackage[dvips]{graphicx} ...
\includegraphics[width=0.8\linewidth]
{myfile.eps}
\end{verbatim}
}
{\small
\bibliographystyle{ieee_fullname}
\section{Introduction}
\blfootnote{This work is financially supported by Science Foundation Ireland (SFI) under the Grant Number 15/RP/2776. We gratefully acknowledge the support of NVIDIA Corporation with the donation of the NVIDIA DGX station used for this research.}
\begin{figure*}[!htbp]
\centering
\includegraphics[width=1.\textwidth]{latex/figures/heatmap_10_compressed.pdf}
\caption{Visualization for significant features extracted by TSN, TSM and our ACTION-Net for the action \textit{`Rotate fists counterclockwise'}. Features extracted by each method are visualized by using CAM~\cite{zhou2016learning}. Compared to TSN and TSM, it can be noticed that ACTION-Net is able to extract features that are related to movements in an action especially for highlighted frames i.e., 4th and 5th columns. More examples can be referred to \textit{Supplementary Materials}.}
\label{fig:heatmap}
\end{figure*}
Video understanding has attracted an increasing amount of interest, since it is a crucial step towards real-world applications, such as Virtual Reality/Augmented Reality (VR/AR) and video-sharing social networking services. For instance, millions of videos are uploaded to TikTok, Douyin, and Xigua Video to be processed everyday, wherein video understanding acts a pivotal part. However, the explosive growth in this video streaming gives rise to challenges on performing video understanding at high accuracy and low computation cost.
Action recognition, a fundamental problem in video understanding, has been a growing demand in video-related applications, such as content moderation (i.e., recognize content in videos that break terms of service) and content recommendations (i.e., videos are ranked by most liked and recommended to similar customers). The complex actions in videos are normally temporal-dependent, which do not only contain spatial information within each frame but also include temporal information over a duration. For example, symmetric action pairs \textit{(`opening a box', `closing a box')}, \textit{(`rotate fists clockwise', `rotate fists counterclockwise')} contain similar features in spatial domains, but the temporal information is completely reversed. Traditional human action recognition is more \textit{scene-related}~\cite{soomro2012ucf101,kuehne2011hmdb,kay2017kinetics}, wherein actions are not as temporal-dependent e.g., \textit{`apply eye makeup', `walking', `running'}. With how rapid technology is developing, like VR, which requires employing gestures to interact with environments, \textit{temporal-related} action recognition has recently become a focus for research.
The mainstreams of existing methods are 3D CNN-based frameworks and 2D CNN-based frameworks. 3D CNNs have been shown to be effective in terms of spatio-temporal modeling~\cite{tran2015learning,chen20182,stroud2020d3d}, but spatio-temporal modeling is unable to capture adequate information contained in videos. The two-stream architecture was proposed to take spatio-temporal information and optical flow into account~\cite{simonyan2014two,carreira2017quo,shi2019two}, which boosted performance significantly compared to the one-stream architecture. However, computation on optical flow is very expensive, which poses challenges on real-world applications. 3D CNNs suffer from problems including overfitting and slow convergence~\cite{hara2018can}. With more large-scale datasets being released, such as Kinetics~\cite{carreira2017quo}, Moments in Time~\cite{monfort2019moments} and ActivityNet~\cite{caba2015activitynet}, optimizing 3D CNNs becomes much easier and more popular. However, heavy computations inherent in 3D CNN-based frameworks contribute to slow inferences, which would limit their deployment on real-world applications, such as VR that relies on online video recognition. Current 2D CNN-based frameworks ~\cite{karpathy2014large,simonyan2014two,wang2016temporal,zhou2018temporal,lin2019tsm} enjoy lightweight and fast inferences. These approaches operated on a sequence of short snippets (known as segments) sparsely sampled from the entire video and were initially introduced in TSN~\cite{wang2016temporal}. Original 2D CNNs lack the ability of temporal modeling, which causes losing essential sequential information in some actions e.g., \textit{`opening a box'} vs \textit{`closing a box'}. TSM~\cite{lin2019tsm} introduced temporal information to 2D CNN-based frameworks by shifting a part of channels on the temporal axis, which significantly improved the baseline for 2D CNN-based frameworks. However, TSM still lacks explicit temporal modeling for an action, such as motion information. Recent works~\cite{lee2018motion,jiang2019stm,li2020tea,liu2020teinet,liu2020tam} introduced embedded modules into 2D CNNs in terms of ResNet architecture~\cite{he2016deep}, which possessed the capability for motion modeling. In order to capture multi-type information contained by videos, previous works normally operated on input-level frames. For instance, SlowFast networks sampled raw videos at multiple rates to characterize slow and fast actions; two-stream networks utilized pre-computed optical flow for reasoning motion information. This kind of approaches commonly require multi-branch networks, which need expensive computations.
Inspired by the aforementioned observation, we propose a new \textit{plug-and-play} and \textit{lightweight} sp\textbf{A}tio-temporal, \textbf{C}hannel and mo\textbf{T}ion excitat\textbf{ION} (ACTION) module to effectively process the multi-type information on the feature level inside a single network by adopting multipath excitation. The combination of spatio-temporal features and motion features can be understood similarly as the two-stream architecture~\cite{simonyan2014two}, but we model the motion inside the network based on the feature level rather than generating another type of input (e.g., optical flow~\cite{ilg2017flownet}) for training the network, which significantly reduces computations. Inspired by SENet, the channel-wise features are extracted based on the temporal domain to characterize the channel interdependencies for the network. Correspondingly, a neural architecture equipped with such a module is dubbed ACTION-Net. ACTION comprises three components for extracting aforementioned features (1) Spatio-Temporal Excitation (STE), (2) Channel Excitation (CE) and (3) Motion Excitation (ME). Figure~\ref{fig:heatmap} visualizes features characterized by TSN, TSM, and ACTION-Net for the action \textit{`rotate fists counterclockwise'}. It can be observed that both TSN and TSM focus on recognizing objects (two fists) independently instead of reasoning an action. Compared to TSN and TSM, our proposed ACTION-Net better characterizes an action by representing feature maps that cover the two fists, especially for the highlighted 4th and 5th columns. In a nutshell, our contributions are three-fold:
\begin{itemize}
\item We propose an ACTION module that works in a \textit{plug-and-play} manner, which is able to extract appropriate spatio-temporal patterns, channel-wise features, and motion information to recognize actions.
\item A simple yet effective neural architecture referred to ACTION-Net, which we demonstrate on three backbones i.e., ResNet-50~\cite{he2016deep}, BNIception~\cite{ioffe2015batch} and MobileNet V2~\cite{sandler2018mobilenetv2}.
\item We have conducted extensive experiments and shown our superior performances on three datasets Something-Something V2~\cite{goyal2017something}, Jester~\cite{materzynska2019jester} and EgoGesture~\cite{zhang2018egogesture}.
\end{itemize}
\section{Related Works}
In this section, we discuss related works by taking 2D and 3D CNN-based frameworks into account, wherein SENet~\cite{hu2018squeeze} and TEA~\cite{li2020tea} inspired us to propose the ACTION-Net.
\subsection{3D CNN-based Framework}
The 3D CNN-based framework has a spatio-temporal modeling capability, which enhances model performance for video action recognition~\cite{tran2015learning,hara2018can,tran2015learning}. I3D~\cite{carreira2017quo} inflated the ImageNet pre-trained 2D kernels to 3D kernels for capturing spatio-temporal information. To better represent motion patterns, I3D utilized pre-computed optical flow together with RGB (also known as the two-stream architecture). SlowFast networks~\cite{feichtenhofer2019slowfast} were proposed to handle inconsistent speeds of actions in videos e.g., running and walking, which involved a slow branch and a fast branch to model slow actions and fast actions, respectively. Although 3D CNN-based approaches have achieved exciting results on several benchmark datasets, they contain massive parameters. In this case, various problems are caused, such as easily overfitting~\cite{hara2018can} and difficulty in converging~\cite{tran2018closer}, which pose challenges including ineffective inferences for online streaming videos in real-world applications. Although recent works~\cite{qiu2017learning,tran2018closer} have demonstrated that 3D convolution can be factorized to lessen computations to some extent, the computation is still much more of a burden when compared to 2D CNN-based frameworks.
\subsection{2D CNN-based Framework}
TSN~\cite{wang2016temporal} was the first proposed framework that applied 2D CNNs for video action recognition, which introduced the concept of \textit{`segment'} to process videos i.e., extract short snippets over a long video sequence with a uniform sparse sampling scheme. However, the direct use of 2D CNNs lacks temporal modeling for video sequences. TSM~\cite{lin2019tsm} firstly introduced temporal modeling to 2D CNN-based frameworks, in which a shift operation for a part of channels was embedded into 2D CNNs. However, TSM lacks explicit temporal modeling for actions such as differences among neighbouring frames. Recently, several works have proposed modules to be embedded into 2D CNNs. These modules are able to model the motion and temporal information. For instance, MFNet~\cite{lee2018motion}, TEINet~\cite{liu2020teinet} and TEA~\cite{li2020tea}, which introduced this type of modules, were demonstrated to be effective on the ResNet architecture. STM~\cite{jiang2019stm} proposed a block for modeling the spatio-temporal and motion information instead of the ordinary residual block. GSM~\cite{sudhakaran2020gate} leverages group spatial gating to control interactions in spatial-temporal decomposition.
\subsection{SENet and Beyond}
Hu~\etal~\cite{hu2018squeeze} introduced a SENet architecture. A squeeze-and-excitation (SE) block was proposed to be embedded into 2D CNNs. In this case, the learning of channel-wise features regarding image recognition tasks was enhanced by explicitly modeling channel interdependencies. To tackle this, the SE block utilized two fully connected (FC) layers in a squeeze-and-unsqueeze manner then applied a Sigmoid activation for exciting essential channel-wise features. However, it processes each image independently without considering critical information such as temporal properties for videos. To tackle this issue, TEA~\cite{li2020tea} introduces motion excitation (ME) and multiple temporal aggregation (MTA) in tandem to capture short- and long-range temporal evolution. It should be noted that MTA is specifically designed for Res2Net~\cite{gao2019res2net}, which means TEA can only be embedded into Res2Net. Inspired by these two previous works, we first propose STE and CE modules beyond the SE module by addressing the spatio-temporal perspective and channel interdependencies in temporal dimension. The ACTION module is then constructed by assembling STE, CE and ME in a parallel manner, in which case, multi-type information in videos can be activated.
\section{Design of ACTION}
In this section, we are going to introduce technical details for our proposed ACTION-Net together with ACTION module. As the ACTION module consists of three sub-modules i.e., Spatio-Temporal Excitation (STE), Channel Excitation (CE) and Motion Excitation (ME), we firstly introduce these three sub-modules respectively and then give an overview on how to integrate them to an ACTION module. Notations used in this section are: $N$ (batch size), $T$ (number of segments), $C$ (channels), $H$ (height), $W$ (width) and $r$ (channel reduce ratio). It should be noticed that all tensors outside the ACTION module are 4D i.e., (N$\times$T, C, H, W). We first reshape the input 4D tensor to 5D tensor (N, T, C, H, W) before feeding to the ACTION in order to enable the operation on specific dimension inside the ACTION. The 5D output tensor is then reshaped to 4D before being fed to the next 2D convolutional block.
\begin{figure}[!htbp]
\centering
\subfigure[]{
\includegraphics[width=.18\textwidth]{latex/figures/STE.pdf}
\label{fig:three_paths_a}
}
\hspace{20pt}
\subfigure[]{
\includegraphics[width=.15\textwidth]{latex/figures/CE.pdf}
\label{fig:three_paths_b}
}
\subfigure[]{
\includegraphics[width=.35\textwidth]{latex/figures/ME.pdf}
\label{fig:three_paths_c}
}
\caption{ACTION module consists of three sub-modules (a) Spatio-Temporal Excitation (STE) module, (b) Channel Excitation (CE) module and (c) Motion Excitation (ME) module.}
\label{fig:three_paths}
\end{figure}
\begin{figure*}[!htbp]
\centering
\includegraphics[width=.9\textwidth]{latex/figures/backbone2.pdf}
\caption{ACTION-Net architecture for ResNet-50~\cite{he2016deep}. The size of output feature map is given for each layer ($CLS$ refers to number of classes and $T$ refers to number of segments). The input video is firstly split into T segments equally and then one frame from each segment is randomly sampled~\cite{wang2016temporal}. The ACTION module is inserted at the start in each residual block. Performance of different embedded locations can be referred to \textit{Supplementary Materials}~\ref{sup-sec:location}.}
\label{fig:backbone}
\end{figure*}
\subsection{Spatio-Temporal Excitation (STE)}
STE is efficiently designed for exciting spatio-temporal information by utilizing 3D convolution. To achieve this, STE generates a spatio-temporal mask $\mathbf{M}\in\mathbb{R}^{N\times T\times 1 \times H \times W}$ that is used for element-wise multiplying the input $\mathbf{X}\in\mathbb{R}^{N\times T\times C \times H \times W}$ across all channels. As illustrated in Fig.~\ref{fig:three_paths_a}, given an input $\mathbf{X}\in\mathbb{R}^{N\times T\times C \times H \times W}$, we first average the input tensor across channels in order to get a global spatio-temporal tensor $\mathbf{F}\in \mathbb{R}^{N\times T\times 1 \times H \times W}$ with respect to the channel axis. Then we reshape $\mathbf{F}$ to $\mathbf{F}^*\in \mathbb{R}^{N\times 1\times T \times H \times W}$ to be fed to a 3D convolutional layer $\mathbf{K}$ with kernel size $3\times3\times3$, which can be formulated as
\begin{equation}
\mathbf{F}_o^* = \mathbf{K} * \mathbf{F}^*
\end{equation}
We finally reshape $\mathbf{F}_o^*$ back to $\mathbf{F}_o\in \mathbb{R}^{N\times T\times 1 \times H \times W}$ and feed it to a Sigmoid activation in order to get the mask $\mathbf{M}\in \mathbb{R}^{N\times T\times 1 \times H \times W}$, which can be represented as
\begin{equation}
\mathbf{M} = \delta(\mathbf{F}_o)
\end{equation}
The final output can be interpreted as
\begin{equation}\label{eq:sub_module_output}
\mathbf{Y} = \mathbf{X} + \mathbf{X} \odot \mathbf{M}
\end{equation}
Compared to the conventional 3D convolutional operation, STE is much more computationally efficient as the input feature $\mathbf{F}^*$ is averaged across channels. Each channel of the input tensor $\mathbf{X}$ can perceive the importance of spatio-temporal information from a refined feature excitation $\delta(\mathbf{F}_o)$.
\subsection{Channel Excitation (CE)}
CE is designed similarly to SE block~\cite{hu2018squeeze} as shown in Fig.~\ref{fig:three_paths_b}. The main difference between CE and SE is that we insert a 1D convolutional layer between two FC layers to characterize temporal information for channel-wise features. Concretely, given an input $\mathbf{X}\in\mathbb{R}^{N\times T\times C \times H \times W}$, we firstly access the global spatial information of the input feature by spatial average pooling the input, which can be represented as
\begin{equation}
\mathbf{F} = \frac{1}{H\times W}\sum^{H}_{i=1}\sum^{W}_{j=1}\mathbf{X[:,:,:,i,j]}
\label{eq:spatial_pool}
\end{equation}
where $\mathbf{F}\in \mathbb{R}^{N\times T\times C \times 1 \times 1}$. We squeeze the number of channels for $\mathbf{F}$ by a scale ratio $r$ ($r = 16$ in this work), which can be interpreted as
\begin{equation}
\mathbf{F}_r = \mathbf{K}_1 * \mathbf{F}
\label{eq:squeeze}
\end{equation}
where $\mathbf{K}_1$ is a $1\times1$ 2D convolutional layer and $\mathbf{F}_r\in \mathbb{R}^{N\times T\times \frac{C}{r} \times 1 \times 1}$. We then reshape $\mathbf{F}_r$ to $\mathbf{F}_r^*\in \mathbb{R}^{N\times \frac{C}{r} \times T \times 1 \times 1}$ to enable the temporal reasoning. A 1D convolutional layer $\mathbf{K}_2$ with kernel size 3 is utilized to process $\mathbf{F}_r^*$ as
\begin{equation}
\mathbf{F}^*_{temp} = \mathbf{K}_2 * \mathbf{F}_r^*
\end{equation}
where $\mathbf{F}^*_{temp}\in \mathbb{R}^{N\times \frac{C}{r} \times T \times 1 \times 1}$. $\mathbf{F}^*_{temp}$ is then reshaped to $\mathbf{F}_{temp}\in \mathbb{R}^{N\times T \times \frac{C}{r} \times 1 \times 1}$, which is then unsqueezed by using a $1\times 1$ 2D convolutional layer $\mathbf{K}_3$ and fed to a Sigmoid activation. These are the last two steps to obtain the channel mask $\mathbf{M}$, which can be formulated respectively
\begin{equation}
\mathbf{F}_o = \mathbf{K}_3 * \mathbf{F}_{temp}
\label{eq:unsqueeze}
\end{equation}
\begin{equation}
\mathbf{M} = \delta(\mathbf{F}_o)
\label{eq:mask}
\end{equation}
where $\mathbf{F}_o\in \mathbb{R}^{N\times T \times {C} \times 1 \times 1}$ and $\mathbf{M}\in \mathbb{R}^{N\times T \times {C} \times 1 \times 1}$. Finally, the output of CE is formulated as the same as in equation~\eqref{eq:sub_module_output} using the new generated mask.
\subsection{Motion Excitation (ME)}
ME has been explored by~\cite{jiang2019stm,li2020tea} previously, which aims to model motion information based on the feature level instead of the pixel level. Different from previous work~\cite{jiang2019stm,li2020tea} that proposed a whole block for extracting motion, we use the ME in parallel with two modules mentioned in previous two sections. As illustrated in Fig.~\ref{fig:three_paths_c}, the motion information is modeled by adjacent frames. We adopt the same squeeze and unsqueeze strategy as used in the CE sub-module by using two $1\times 1$ 2D convolutional layers, which can be referred to equation~\eqref{eq:squeeze} and equation~\eqref{eq:unsqueeze} respectively. Given the feature $\mathbf{F}_r\in \mathbb{R}^{N\times T \times \frac{C}{r} \times H \times W}$ processed by the squeeze operation, motion feature is modeled following the similar operation presented in~\cite{jiang2019stm,li2020tea}, which can be represented as
\begin{equation}
\mathbf{F}_m = \mathbf{K} * \mathbf{F}_r[:,t+1,:,:,:] - \mathbf{F}_r[:,t,:,:,:]
\end{equation}
where $\mathbf{K}$ is a $3\times 3$ 2D convolutional layer and $\mathbf{F}_m\in \mathbb{R}^{N\times 1 \times \frac{C}{r} \times H \times W}$. The motion feature is then concatenated to each other according to the temporal dimension and 0 is padded to the last element i.e., $\mathbf{F}_M = [\mathbf{F}_m(1), \dots, \mathbf{F}_m(t-1), 0], \mathbf{F}_M\in \mathbb{R}^{N\times T \times \frac{C}{r} \times H \times W}$. The $\mathbf{F}_M$ is then processed by spatial average pooling same as in equation~\eqref{eq:spatial_pool}. The feature output $\mathbf{F}_o\in \mathbb{R}^{N\times T \times C\times 1 \times 1}$ and the mask $\mathbf{M}\in \mathbb{R}^{N\times T \times {C} \times 1 \times 1}$ can then be achieved similarly as in CE through equation~\eqref{eq:unsqueeze} and equation~\eqref{eq:mask} respectively.
\subsection{ACTION-Net}
The overall ACTION module takes the element-wise addition of three excited features generated by STE, CE and ME respectively (see ACTION module block in Fig.~\ref{fig:backbone}). By doing this, the output of the ACTION module can perceive information from a spatio-temporal perspective, channel interdependencies and motion. Figure~\ref{fig:backbone} shows the ACTION-Net architecture for ResNet-50, wherein the ACTION module is inserted at the beginning in each residual block. It does not require any modification for original components in the block. Details of ACTION-Net architectures for MobileNet V2 and BNInception can be referred to \textit{Supplementary Materials}.
\section{Experiments}
We first show that ACTION-Net is able to consistently improve the performance for 2D CNNs compared to previous two fundamental works TSN~\cite{wang2016temporal} and TSM~\cite{lin2019tsm} on three datasets i.e., EgoGesture~\cite{zhang2018egogesture}, Something-Something V2~\cite{goyal2017something} and Jester~\cite{materzynska2019jester}. We then perform extensive experiments for comparing ACTION-Net with state-of-the-arts on these three datasets. Abundant ablation studies are conducted to analyze the efficacy for each excitation path in ACTION-Net. Finally, we further compare ACTION-Net with TSM on three backbones i.e., ResNet-50, MobileNet V2 and BNInception by considering performance and efficiency ($\eta$).
\subsection{Datasets}
We evaluated the performance for the proposed ACTION-Net on three large-scale and widely used action recognition datasets i.e., Something-Something V2~\cite{goyal2017something}, Jester~\cite{materzynska2019jester} and EgoGesture~\cite{zhang2018egogesture}. The Something-Something V2 dataset~\cite{lee2018motion,liu2020teinet,yang2020temporal,zolfaghari2018eco} is a large collection of humans performing actions with everyday objects. It includes 174 categories with 168,913 training videos, 24,777 validation videos and 27,157 testing videos. Jester~\cite{jiang2019stm,kopuklu2019real,zhang2020temporal,materzynska2020something,kopuklu2020drivermhg} is a third-person view gesture dataset, which has a potential usage for human computer interaction. It has 27 categories with 118,562 training videos, 14,787 validation videos and 14,743 testing videos. EgoGesture~\cite{kopuklu2019real,wang2020catnet,shi2019two,shi2019skeleton,abavisani2019improving,shi2020skeleton} is a large-scale dataset for egocentric hand gesture recognition recorded by a head-mounted camera, which is designed for VR/AR use cases. It involves 83 classes of gestures with 14,416 training samples, 4,768 validation samples and 4,977 testing samples.
\subsection{Implementation Details}
\noindent \textbf{Training.}\hspace{2pt} We conducted our experiments on video action recognition tasks by following the same strategy mentioned in TSN~\cite{wang2016temporal}. Given an input video, we firstly divided it into $T$ segments of equal duration. Then we randomly selected one frame from each segment to obtain a clip with $T$ frames. The size of the shorter side of these frames is fixed to 256 and corner cropping and scale-jittering were ultilized for data augmentation. Each cropped frame was finally resized to $224\times 224$, which was used for training the model. The input fed to the model is of the size $N\times T \times 3 \times 224 \times 224$, in which $N$ is the batch size, $T$ is the number of segments.
The models were trained on a NVIDIA DGX station with four Tesla V100 GPUs. We adopted SGD as optimizer with a momentum of 0.9 and a weight decay of $5\times 10^{-4}$. Batch size was set as $N=64$ when $T=8$ and $N=48$ when $T=16$. Network weights were initialized using ImageNet pretrained weights. For Something-Something V2, we started with a learning rate of 0.01 and reduced it by a factor of 10 at 30, 40, 45 epochs and stopped at 50 epochs. For Jester dataset, we started with a learning rate of 0.01 and reduced it by a factor of 10 at 10, 20, 25 epochs and stopped at 30 epochs. For EgoGesture dataset, we started with a learning rate of 0.01 and reduced it by a factor of 10 at 5, 10, 15 epochs and stopped at 25 epochs.
\noindent \textbf{Inference.}\hspace{2pt} We utilized the three-crop strategy following \cite{jiang2019stm,wang2018non,feichtenhofer2019slowfast} for inference. We firstly scaled the shorter side to 256 for each frame and took three crops of $256\times 256$ from scaled frames. We randomly sampled from the full-length video for 10 times. The final prediction was the averaged Softmax score for all clips.
\subsection{Improving Performance of 2D CNNs}
We compare ACTION-Net to two fundamental 2D CNN counterparts TSN and TSM. As illustrated in Table~\ref{tab:2D_baseline_compare},
\begin{table}[!htbp]
\centering
\caption{ACTION-Net consistently outperforms 2D counterparts on three representative datasets. All methods use ResNet-50 as backbone and 8 frames input for the fair comparison.}
\setlength{\tabcolsep}{0.45mm}{
\begin{threeparttable}
\begin{tabular}{cccccc}
\toprule
Dataset& Model& Top-1 & $\triangle$Top-1& Top-5& $\triangle$Top-5\\
\midrule
\multirow{3}{*}{EgoGesture\tnote{*}}& TSN & 83.1& -& 97.3&- \\
& TSM & 92.1& + 9.0&98.3 &+ 1.0\\
& \cellcolor{columbiablue}ACTION-Net & \cellcolor{columbiablue}94.2&\cellcolor{columbiablue} + 11.1&\cellcolor{columbiablue}98.7 &\cellcolor{columbiablue}+ 1.4 \\\hline
\multirow{3}{*}{SomethingV2}& TSN & 27.8& -& 57.6&-\\
& TSM &58.7& + 30.9& 84.8 &+ 27.2\\
& \cellcolor{columbiablue}ACTION-Net &\cellcolor{columbiablue}62.5&\cellcolor{columbiablue} + 34.7& \cellcolor{columbiablue}87.3&\cellcolor{columbiablue}+ 29.7\\ \hline
\multirow{3}{*}{Jester}& TSN & 81.0& -& 99.0&-\\
& TSM& 94.4& + 13.4& 99.7&+ 0.7\\
& \cellcolor{columbiablue}ACTION-Net& \cellcolor{columbiablue}97.1& \cellcolor{columbiablue}+ 16.1& \cellcolor{columbiablue}99.8&\cellcolor{columbiablue}+ 0.8\\
\bottomrule
\end{tabular}
\begin{tablenotes}
\item[*] We re-implement TSN and TSM using the official public code in~\cite{lin2019tsm}.
\end{tablenotes}
\end{threeparttable}
\label{tab:2D_baseline_compare}
}
\end{table}
ACTION-Net consistently outperforms these two 2D CNN counterparts on all three datasets. It is worth nothing that TSN does not contain any component that is able to model the temporal information. By employing a temporal shift operation to a part of channels, TSM introduces some temporal information to the network, which significantly improves the 2D CNN baseline compared to TSN. However, TSM still lacks explicit temporal modeling. By adding the ACTION module to TSN, ACTION-Net takes spatio-temporal modeling, channel interdependencies modeling and motion modeling into account. It can be noticed that the Top-1 accuracy of ACTION-Net is improved by 2\%, 3.8\% and 2.7\% compared to TSM with respect to EgoGesture, Something-Something V2 and Jester datasets.
\subsection{Comparisons with the State-of-the-Art}
\begin{table*}[!htbp]
\centering
\caption{Comparisons with the state-of-the-arts on Something-Something V2, Jester and EgoGesture datasets.}
\label{tab:SOTAs_compare}
\setlength{\tabcolsep}{.9mm}{
\begin{threeparttable}
\begin{tabular}{c|c|ccccccccc}
\toprule
\multicolumn{1}{c}{\multirow{2}{*}{Method}}&\multicolumn{1}{c}{\multirow{2}{*}{Backbone}}& \multirow{2}{*}{Plug-and-play}& \multirow{2}{*}{Pretrain}& \multirow{2}{*}{Frame}&
\multicolumn{2}{c}{Something V2}&
\multicolumn{2}{c}{Jester}&
\multicolumn{2}{c}{EgoGesture}
\\ \cline{6-7} \cline{8-9} \cline{9-11}
\multicolumn{1}{c}{}&\multicolumn{1}{c}{}&\multicolumn{1}{c}{}&&&Top-1 & Top-5& Top-1 & Top-5& Top-1 & Top-5\\
\midrule
C3D + RSTTM~\cite{zhang2018egogesture}& {-}& & {Scratch}
& 16& -& - & -& -& 89.3& -\\ \cline{0-0}\cline{2-11}
C3D~\cite{kopuklu2019real}& {ResNext-101}& & {Kinetics}
& 16& -& - & 95.9& -& 90.9& -\\ \hline
TRN Multiscale~\cite{lin2019tsm}& BNInception& & ImageNet
& 8& 48.8& 77.6 & 95.3& - & - & -\\ \cline{0-0}\cline{2-11}
MFNet-C50~\cite{lee2018motion}&\multirow{16}{*}{ResNet-50}& & {ImageNet}
& 7& -& - & 96.1& 99.7& - & -\\ \cline{0-0}\cline{3-11}
\multirow{2}{*}{TSN~\cite{lin2019tsm}}& \multirow{2}{*}{}& \multirow{2}{*}\checkmark& \multirow{2}{*}{Kinetics}
& 8& 27.8& 57.6 & 81.0 & 99.0 & 83.1 & 98.3\\
&&&&16& 30.0&60.5 & 82.3& 99.2& -&-\\ \cline{0-0}\cline{3-11}
\multirow{2}{*}{TSM~\cite{jiang2019stm}}& \multirow{2}{*}{}& \multirow{2}{*}{\checkmark}& \multirow{2}{*}{Kinetics}
& 8& 58.7& 84.8& 94.4& 99.7& 92.1& 98.3\\
&&&& 16& 61.2& 86.9& 95.3& 99.8& -&-\\
\cline{0-0}\cline{3-11}
\multirow{2}{*}{GST~\cite{luo2019grouped}}& \multirow{2}{*}{}& & \multirow{2}{*}{ImageNet}
& 8& 61.6& 87.2 & - & - & - & -\\
&&&&16& 62.6&87.9 & -& -& -&-\\ \cline{0-0}\cline{3-11}
bLVNet-TAM~\cite{liu2020teinet}& & & ImageNet
& 32& 61.7& 88.1 & - & - & - & -\\\cline{0-0}\cline{3-11}
CPNet~\cite{liu2019learning}& & & ImageNet
& 24& 57.7& 84.0 & - & - & - & -\\\cline{0-0}\cline{3-11}
\multirow{2}{*}{TEINet~\cite{liu2020teinet}}& \multirow{2}{*}{}& \multirow{2}{*}\checkmark& \multirow{2}{*}{ImageNet}
& 8& 62.7& - & - & - & - & -\\
&&&&16& 63.0&- & -& -& -&-\\ \cline{0-0}\cline{3-11}
\multirow{2}{*}{STM~\cite{jiang2019stm}}& \multirow{2}{*}{}& & \multirow{2}{*}{ImageNet}
& 8& 62.3& 88.8 & 96.6 & 99.9 & - & -\\
&&&&16& 64.2&89.8 & 96.7& 99.9& -&-\\ \cline{0-0}\cline{3-11}
\multirow{2}{*}{TEA~\cite{li2020tea}}& \multirow{2}{*}{}& & \multirow{2}{*}{ImageNet}
& 8& -& - & 96.5& 99.8& 92.3 & 98.3\\
&&&&16& \textbf{64.5}&\textbf{89.8} & 96.7& 99.8& 92.5&98.5\\
\rowcolor{columbiablue}&& &
& 8& 62.5& 87.3& 97.1& 99.8& 94.2&98.7\\
\rowcolor{columbiablue}\multirow{-2}{*}{ACTION-Net\tnote{1}}&&\multirow{-2}{*}\checkmark&\multirow{-2}{*}{ImageNet}&16& 64.0&89.3& \textbf{97.1}& \textbf{99.9}& \textbf{94.4}&\textbf{98.8}\\
\bottomrule
\end{tabular}
\begin{tablenotes}
\item[1] ACTION is inserted into each residual block in this experiment.
\end{tablenotes}
\end{threeparttable}
}
\end{table*}
We compare our approach with the state-of-the-art on Something-Something V2, Jester and EgoGesture, which is summarized in Table~\ref{tab:SOTAs_compare}. We mainly compare our approach with 2D CNN counterparts as 3D CNN-based frameworks are more favored in \textit{scene-related} datasets e.g., Kinetics~\cite{qiu2017learning,feichtenhofer2019slowfast}. The superiority of ACTION-Net on Jester and EgoGesture is quite impressive. It is clear that even ACTION-Net with 8 RGB frames as input achieves the state-of-the-art performance compared to other methods, which confirms the remarkable ability of ACTION-Net for integrating useful information from three excitation paths. In terms of Something-Something V2, ACTION-Net also achieves competitive results compared to STM and TEA. It should be noted that both STM and TEA are specifically designed for ResNet and Res2Net respectively, while our ACTION enjoys being easily equipped by other architectures i.e., MobileNet V2 and BNInception investigated in this work.
\subsection{Ablation Study}
\begin{table*}[!htbp]
\centering
\hspace{-0.5cm}
\begin{minipage}{0.6\textwidth}
\caption{Accuracy and model complexity on EgoGesture dataset. Three excitation types are compared to TSM and TSN. All methods use ResNet-50 as backbone and 8 frames input for fair comparison. The least FLOPs/$\triangle$FLOPs/Param. and the best performance for ACTION-Net and sub-modules are highlighted as bold.}
\begin{tabular}{cccccc}
\toprule
\multicolumn{2}{c}{Method}& FLOPs& $\triangle$FLOPs& Param.& Top-1 \\
\midrule
\multicolumn{2}{c}{TSN} & 33G& -& 23.68M& 83.1\\
\multicolumn{2}{c}{TSM} & 33G& -& 23.68M& 92.1\\ \hline \hline
\multirow{4}{*}{Ours} &STE & \textbf{33.1G}& \textbf{+0.1G (+0.3\%)}& \textbf{23.9M}& 93.8\\
&CE & 33.16G& +0.16G (+0.5\%)& 26.08M& 93.8\\
&ME & 34.69G& +1.69G (+5.1\%)& 25.9M& 93.9\\
&ACTION-Net & 34.75G& +1.75G (+5.3\%)& 28.08M& \textbf{94.2}\\
\bottomrule
\end{tabular}
\label{tab:three_paths}
\end{minipage}
\hspace{20pt}
\centering
\begin{minipage}{0.3\textwidth}
\caption{Ablation study of having ACTION included or not in each residual block stage regarding ResNet-50 backbone on EgoGesture dataset by using 8 frames input. More ACTION engaged yields better performance.}
\begin{tabular}{cccc}
\toprule
Stage & Top-1& Top-5 \\ \hline
\multicolumn{1}{l}{$\text{res}_2$}& 92.3& 98.2\\
\multicolumn{1}{l}{$\text{res}_{2,3}$}& 92.9& 98.5\\
\multicolumn{1}{l}{$\text{res}_{2,3,4}$}& 93.1& 98.5\\
\multicolumn{1}{l}{$\text{res}_{2,3,4,5}$}& \textbf{94.2}& \textbf{98.7}\\
\bottomrule
\end{tabular}
\label{tab:block_num}
\end{minipage}
\end{table*}
\begin{table*}[!htbp]
\centering
\caption{ACTION-Net generalizes well across backbones and datasets (TSM is used as a baseline). Accuracy and model complexity on three backbones using Something-Something V2, Jester and EgoGesture with 8 frames as input. The most significant improvements on accuracy and the least extra FLOPs are highlighted as bold.}
\begin{tabular}{ccccccc}
\toprule
Backbone& Method & FLOPs& Param.& Something V2& Jester& EgoGesture \\
\midrule
\multirow{2}{*}{ResNet-50} & TSM & 33G& 23.68M& 58.7& 94.4& 92.1\\
& \cellcolor{columbiablue}ACTION-Net & \cellcolor{columbiablue}34.75G (+5.3\%)& \cellcolor{columbiablue}28.08M& \cellcolor{columbiablue}\textbf{62.5 (+3.8)}& \cellcolor{columbiablue}\textbf{97.1 (+2.7)}&\cellcolor{columbiablue}\textbf{94.2 (+2.1)}\\\hline
\multirow{2}{*}{BNIception}& TSM & 16.39G& 10.36M& 53.5& 94.6&92.3\\
& \cellcolor{columbiablue}ACTION-Net &\cellcolor{columbiablue}16.94G (+3.4\%)& \cellcolor{columbiablue}11.59M& \cellcolor{columbiablue}56.7 (+3.2)& \cellcolor{columbiablue}96.6 (+2.0)&\cellcolor{columbiablue}93.2 (+0.9)\\ \hline
\multirow{2}{*}{MobileNet V2}& TSM& 2.55G& 2.33M& 54.9& 95.0&92.4\\
& \cellcolor{columbiablue}ACTION-Net& \cellcolor{columbiablue}\textbf{2.57G (+0.8\%)}& \cellcolor{columbiablue}2.36M& \cellcolor{columbiablue}58.5 (+3.6)& \cellcolor{columbiablue}96.7 (+1.7)&\cellcolor{columbiablue}93.5 (+1.1)\\
\bottomrule
\end{tabular}
\label{tab:TSM_STCME}
\end{table*}
\begin{figure*}
\subfigure[EgoGesture]{
\centering
\includegraphics[width=0.32\textwidth]{latex/figures/runtime_ego.pdf}
\label{fig:runtime_egogesture}
}
\subfigure[Jester]{
\centering
\includegraphics[width=0.32\textwidth]{latex/figures/runtime_jester.pdf}
\label{fig:runtime_jester}
}
\subfigure[Something-something V2]{
\centering
\includegraphics[width=0.32\textwidth]{latex/figures/runtime_sthv2.pdf}
\label{fig:runtime_sthv2}
}
\caption{Top-1 accuracy and FLOPs for ACTION-Net and TSM on three backbones i.e., ResNet-50, BNInception and MobileNet V2 using three datasets (from left to right: EgoGesture, Jester and Something-something V2). $\eta$ is calculated using equation~\eqref{eq:efficience} for each backbone on three datasets. EgoGesture: ResNet-50 $\eta=2.65\%$, BNInception $\eta=3.73\%$, MobileNet V2 $\eta=\textbf{0.71\%}$. Jester: ResNet-50 $\eta=1.96\%$, BNInception $\eta=1.68\%$, MobileNet V2 $\eta=\textbf{0.46\%}$. Something V2: ResNet-50 $\eta=1.40\%$, BNInception $\eta=1.05\%$, MobileNet V2 $\eta=\textbf{0.22\%}$.}
\label{fig:runtime}
\end{figure*}
In this section, we investigate the design of our ACTION-Net with respect to (1) efficacy of each excitation and (2) impact of the number of ACTION modules in ACTION-Net regarding the ResNet-50 architecture. We carry out ablation experiments using 8 frames as the input on the EgoGesture dataset for inspecting these two aspects.
\vspace{5pt}
\noindent
\textbf{Efficacy of Three Excitations.}\hspace{2pt} To validate the contribution of each excitation sub-module, we compare the performance for each individual module and the combination of all sub-modules (ACTION-Net) in Table~\ref{tab:three_paths}. We also provide visualization results in \textit{Supplementary Materials}. Results show that each excitation module is able to improve the performance for 2D CNN baselines provided by TSN and TSM with limited added computational cost. Concretely, STE and CE both add negligible extra computation compared to TSM by averaging channels globally and averaging spatial information globally yet they both provide useful information to the network. ME adds more computation and parameters to the network than the previous two yet it is acceptable. It captures temporal differences on the spatial domain among adjacent frames over the time and achieves better performance compared to STE and CE. When integrating all these three sub-modules to constitute the ACTION, it can be seen that the ACTION-Net achieves the highest accuracy and increases 2.1\% Top-1 accuracy together with increasing 5.3\% FLOPs. To better capture the relation between boosted performance and add-on computation, we define the efficiency formulated as
\begin{equation}
\eta = \frac{\triangle\text{FLOPs}}{\triangle\text{Top-1}}
\label{eq:efficience}
\end{equation}
where both $\triangle$FLOPs and $\triangle$Top-1 are in percent, $\eta$ is the efficiency that represents how many extra FLOPs \textit{in percent} are introduced with respect to increasing 1\% Top-1 accuracy (\textit{smaller indicates more efficient} apparently). Efficiency $\eta$ for STE, CE, ME and ACTION-Net is 0.18\%, 0.29\%, 2.83\% and 2.52\% respectively. It can be noticed that STE is the most efficient when taking $\eta$ into account.
\vspace{5pt}
\noindent
\textbf{Impact of the Number of ACTION Blocks.}\hspace{2pt} The architecture of ResNet-50 can be divided into 6 stages i.e., $\text{conv}_{1}$, $\text{res}_{2}$, $\text{res}_{3}$, $\text{res}_{4}$, $\text{res}_{5}$ and FC. The ACTION module can be inserted into any residual stage from $\text{res}_{2}$ to $\text{res}_{5}$. We investigate the impact of the number of residual stages that contain the ACTION module as shown in Table~\ref{tab:block_num}. Results show that more stages including the ACTION module results in better performance, which indicates the efficacy for our proposed approach.
\subsection{Analysis of Efficiency and Flexibility}\label{sec:Runtime}
Compared to recent 2D CNN approaches e.g., STM~\cite{jiang2019stm} and TEA~\cite{li2020tea}, our ACTION module enjoys a plug-and-play manner like TSM, which can be embedded to any 2D CNN. To validate the efficacy for our proposed approach, we compare ACTION-Net with TSM on three backbones i.e., ResNet-50, BNInception and MobileNet V2. We report FLOPs and Top-1 accuracy for ACTION-Net and TSM respectively. We also report $\eta$ calculated using equation~\ref{eq:efficience} when replacing TSM with ACTION, which indicates the penalization of computation when improving the accuracy. Table~\ref{tab:TSM_STCME} demonstrates recognition performance and computation for ACTION-Net employing three backbones. TSM is used as a baseline since TSM benefits good performance and zero extra introduced computational cost compared to TSN. It can be noticed that ACTION-Net outperforms TSM consistently regarding the accuracy for all three backbones yet with limited add-on computational cost. ResNet-50 is boosted most significantly regarding the performance and MobileNet V2 holds the least added FLOPs. Figure~\ref{fig:runtime} demonstrates the efficiency $\eta$ for ACTION-Net based on TSM on different backbones using three datasets. It can be noticed that MobileNet V2 achieves the lowest $\eta$ (the most efficient) while ResNet-50 achieves the highest $\eta$ (the least efficient) for all three datasets, which indicates that MobileNet V2 benefits mostly from the ACTION module regarding the efficiency.
\section{Conclusion}
We target at designing a module to be inserted to 2D CNN models for video action recognition and introduce a novel ACTION module that utilizes multipath excitation for spatio-temporal features, channel-wise features and motion features. The proposed module could be leveraged by any 2D CNN to build a new architecture ACTION-Net for video action recognition. We demonstrate efficacy and efficiency for ACTION-Net on three large-scale datasets. We show that ACTION-Net achieves consistently improvements compared to 2D CNN counterparts with limited extra computations introduced.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
In self-assembly of molecular systems, a disordered solution spontaneously organizes to form ordered structures or assembly patterns from its components. Examples of natural and synthetic materials where self-assembly, the resulting complex structures, and their advanced functional properties are central~\cite{whitesides2002beyond} include molecular crystals \cite{wang2020,desiraju1989crystal}, colloids \cite{evans1999thecolloidal}, lipid membranes \cite{jones1995micelles}, polymer assemblies \cite{lu2020,hu2014,stuart2010emerging}, self-assembled monolayers \cite{kumar1995patterned,smith2004} and quasicrystalline structures \cite{lifshitz2007soft}. Basic biological cell function relies on self-assembly phenomena such as spontaneous condensation \cite{banani2017biomolecular} or lipid raft formation \cite{Sezgin2017}. Also technologically, self-assembling soft materials are receiving a great deal of attention in a variety of applications including drug delivery \cite{yadav2020,branco2009self,kataoka2012block} and anticancer therapy \cite{huang2020,oerlemans2010polymeric}, nanotechnology photonics \cite{ma2019,wang2020,yang2018,shaikh2021}, polymer \cite{shaikh2021,yang2018} and other electronics \cite{yuan2009facile}, as well as biointerfaces and bioengineering \cite{lyons2020,zhang2002design}.
At the level of thermodynamics the properties of molecular self-assemblies in surfactant, biomolecular, and polymer solutions, such as micellization, emulsion formation, or self-assembling monolayers and bilayers can be explained at the level of interplay between water entropy, the relevant surface tensions and corresponding free energies \cite{nagarajan1991theory,birdi2015handbook,tadros2016emulsions}. However, understanding self-assembly systems involving multiple components, especially the mechanisms giving rise to the complex assembly characteristics in them, requires more advanced theoretical efforts.
To this end, in this manuscript, we present a general theoretical framework based on classical density functional theory (DFT) \cite{hansen_mcdonald, evans_79, lutsko2010recent} with competing molecular interactions. We demonstrate how such cases can lead to molecular solutions transitioning from uniform mixtures or simple particle formation to spontaneous ordered, structurally more complex assemblies. Qualitatively similar molecular-level aggregates range from core-shell molecular assemblies (micelles or polymer particles) to assemblies with internal composition gradients or layered structures (onion-like). Such assemblies are widely present in surfactant and biosystems (micelles, vesicles, emulsion type formulations, biomolecular droplets)~\cite{sorrenti2013,banani2017biomolecular,sheth2020multiple} but also readily formed by polymer assemblies \cite{lu2020, hu2014, stuart2010emerging,shaikh2021}. The DFT work here is based on effective potentials typically used to describe soft, intermixing species. Understanding the spontaneous assembly transitions and controlling the composition gradients, for example, hydrophobic-hydrophilic gradients, is particularly important for designing solubilization and partitioning systems e.g. for separation and extraction \cite{mortada2021} or drug delivery systems~\cite{de2012polymeric, ghosh2012core, haag2004supramolecular, kakizawa2002block}. The self-assembling micro-environments also form synthetic biology artificial cells \cite{trantidou2017,che2016} and confined catalysis systems \cite{York2017,Gaitzsch2016}.
In DFT, approximate free energy functionals are devised for particular model Hamiltonian. This enables extracting an inhomogeneous average density of the different components in the thermodynamic equilibrium based on energy minimization principles
\cite{hansen_mcdonald, evans_79, lutsko2010recent}.
In the last few years, DFT has been successfully applied to multi-component systems to describe various other forms of self-assembly, such as superlattice structures \cite{somerville2018density}, quasicrystals~\cite{ratliff2019wave, scacchi2020quasicrystal, subramanian2020density} and general multiple length scale self-assembly~\cite{scacchi2021self}. In this approach, the linear dispersion relation~\cite{archer2012solidification, archer2014solidification, archer2016generation, dispersion_relation, walters2018structural, ratliff2019wave}, characterizing the growth or decay of density modulations in the liquid state, provides the key quantity used to establish guidelines to generate appropriate interparticle potentials. The importance of the linear dispersion relation in constructing the effective interactions is also emphasized by its recent application to accurately capture non-equilibrium phenomena such as laning instability~\cite{scacchi2017dynamical} and to control the macroscopic structures of systems under external shear~\cite{scacchi2020sensitive}.
\section{Theory}
\label{sec:DFT}
Using DFT, we can obtain the equilibrium average density distribution of a binary mixture suspended in a (implicit) background fluid. In other words, our model captures the interactions of regions of two different chemical characteristics in a solution to predict how molecular species, corresponding to those chemical characteristics, would assemble and organize. Although these interaction characteristics emerge from the molecular level, we emphasize that the ``particles" in our model implicitly include interactions rising due to their solution environment. These include e.g. contributions arising from solvent quality for each of the components.
The total grand potential of a system composed by two distinct types of particles is described by
\begin{equation}
\Omega[\rho_1,\rho_2]=\mathcal{F}[\rho_1,\rho_2]+\sum_{i=1,2}\int d\textbf{r} \left[V_i^{\rm ext}(\textbf{r})-\mu_i\right]\rho_i(\textbf{r}),\label{grand_canonical}
\end{equation}
where $\mathcal{F}$ is the intrinsic Helmholtz free energy functional, $V_i^{\rm ext}(\textbf{r})$ is the one-body external potential acting on species $i$, $\mu_i$ the corresponding chemical potential and $\rho_i(\textbf{r})$ the average (inhomogeneous) one body number density. The intrinsic Helmholtz free energy can be written as
\begin{equation}
\mathcal{F}[\rho_1,\rho_2]=\mathcal{F}^{\rm id}[\rho_1,\rho_2]+\mathcal{F}^{\rm exc}[\rho_1,\rho_2],\label{free_energy}
\end{equation}
where the first term is the ideal gas contribution. The latter functional has the known form
\begin{equation}
\mathcal{F}^{id}[\rho_1,\rho_2]=k_{\rm B} T \sum_{i=1,2}\int d\textbf{r} \rho_i(\textbf{r})\left[\ln\left(\Lambda_i^d \rho_i(\textbf{r})\right)-1\right],\label{ideal}
\end{equation}
where $\Lambda_i$ is the thermal de Broglie wavelength and $d$ is the dimensionality of the system (here we focus on $d=2$ for simplicity, but our findings are expected to be valid in $d=3$ as well, where additional configurations are to be expected). The second term in Eq. (\ref{free_energy}) is the excess Helmholtz free energy functional, which describes the interactions between the particles. Following Ramakrishnan and Yussouff \cite{ramakrishnan1979first}, we opt for an expansion of the latter functional around the homogeneous fluid states $\rho_{i}^0$, $i=1,2$, in a functional Taylor expansion and truncate at second order, leaving us with
\begin{equation}
\begin{split}
\mathcal{F}^{\rm exc}&[\rho_1,\rho_2]=\mathcal{F}^{\rm exc}[\rho_{1}^0,\rho_{2}^0]+\sum_{i=1,2}\int d\textbf{r}\mu_i^{ex}\delta\rho_i(\textbf{r})\\&-\frac{1}{2\beta}\sum_{\substack{i=1,2 \\ j=1,2}}\int d\textbf{r}\int d\textbf{r}'\delta\rho_i(\textbf{r})c_{ij}(\mid \textbf{r} -\textbf{r}'\mid)\delta\rho_j(\textbf{r}'),\label{excess}
\end{split}
\end{equation}
where $\delta\rho_i(\textbf{r}) = \rho_i(\textbf{r}) - \rho_{i}^0$ and $\mu_i^{\rm ex} = \mu_i - k_{\rm B} T \ln(\rho_{i}^0\Lambda_i)$ are the excess chemical potential, $c_{ij}(\textbf{r})$ are the direct correlation functions, and $\beta=1/k_BT$. There are different ways to approximate the latter functions. The choice of the approximation depends on the type of interactions that one considers for the model. For this study, we choose the random phase approximation (RPA), which corresponds to $c_{ij}(\textbf{r})=-\beta \phi_{ij}(\textbf{r})$, where $\phi_{ij}(\textbf{r})$ are effective pair potentials~\cite{likos2001effective}. This approximation has proven to be very efficient when dealing with soft interactions in polymer, surfactant, and other macromolecular systems including dendrimers or star polymers \cite{louis2000mean, likos2001criterion, archer2002binary, archer2015soft,scacchi2021self}, and even for describing structure and dynamics of cells in living tissues and tumors~\cite{al2018dynamical}. The accuracy and the limitations of this approximation for systems relevant to this work are discussed in detail in e.g., Refs.~\cite{archer2014solidification, archer2015soft, archer2002binary}. However, if more accurate approximations were needed, one could for example use the so-called hypernetted chain (HNC) Ornstein-Zernike integral equation theory \cite{hansen_mcdonald} to obtain the direct correlation functions $c_{ij}(\textbf{r})$ in Eq.~(\ref{excess}). Such approach has recently been applied to e.g. describe nanoparticles forming quasicrystals \cite{scacchi2020quasicrystal}.
The equilibrium average density profiles $\rho_i(\textbf{r})$ are those which minimise the functional of the grand potential $\Omega[\rho_1,\rho_2]$. Therefore, they also satisfy the coupled Euler-Lagrange equations
\begin{equation}
\frac{\delta \Omega[\rho_1,\rho_2]}{\delta\rho_i}=0,
\end{equation}
for $i=1,2$. For bulk systems, i.e. for $V_i^{\rm ext}(\textbf{r}) \equiv 0$ for $i = 1, 2$ in Eq. (\ref{grand_canonical}), this is equivalent to solving the following coupled equations for $\rho_1(\textbf{r})$ and $\rho_2(\textbf{r})$
\begin{equation}\label{solver}
\ln\left(\frac{\rho_i(\textbf{r})}{\rho_i^0}\right)-\beta\sum_{j=1,2}\int d\textbf{r}'c_{ij}(\mid \textbf{r} -\textbf{r}'\mid)\delta\rho_j(\textbf{r}')=0.
\end{equation}
The results presented hereafter are obtained solving the latter equation under periodic boundary conditions. Next, we discuss the linear dispersion relation and the partial liquid structure factors. These quantities provide important guidelines with respect to the length scales and structures that are expected to be found in the inhomogeneous state~\cite{scacchi2020quasicrystal}. They can also be used to find appropriate model parameters. The Ornstein-Zernike equation for binary mixtures and different closures are also briefly presented.
\subsection{The dispersion relation}
The two branches of the dispersion relation for a binary (isotropic) mixture are \cite{dispersion_relation}
\begin{equation}\label{dispersion_relation_eq}
\omega_{\pm}(k)=\frac{1}{2}\textrm{Tr}(\textbf{M}\textbf{E})\pm \sqrt{\frac{1}{4}\textrm{Tr}(\textbf{M}\textbf{E})^2-\textrm{det}(\textbf{M}\textbf{E})},
\end{equation}
where the mobility matrix $\textbf{M}$ and the energy matrix $\textbf{E}$ are defined as
\begin{equation}
\textbf{M}=-k^2\begin{pmatrix}
D_1\rho_{1}^0 & 0\\
0 & D_2\rho_{2}^0
\end{pmatrix},
\end{equation}
and
\begin{equation}
\textbf{E}=\begin{pmatrix}
\left[\frac{1}{\rho_{1}^0}-\hat{c}_{11}(k)\right] & -\hat{c}_{12}(k)\\
-\hat{c}_{21}(k) & \left[\frac{1}{\rho_{2}^0}-\hat{c}_{22}(k)\right]
\end{pmatrix},
\end{equation}
respectively. In these, $\hat{c}_{ij}(k)$ are the Fourier transform of the direct correlation functions $c_{ij}(r)$ and $D_{i}$ is the diffusion coefficient of species $i$. If $\omega_{\pm} < 0$, for all wave numbers $k=|\textbf{k}|$, $k \in {\rm I\!R}$, then all $k$ modes decay and the uniform state is linearly stable. It is easy to see that $\omega_{-}(k) \leq \omega_{+}(k)$. Thus, if $\omega_{+} > 0$ for some values of $k$, then such modes grow over time and will dominate the density modulation of the system. The onset of linear instability, i.e. when both $\omega(k)=0$ and $d\omega(k)/{dk}=0$ corresponds to solving $\textrm{det}(\textbf{E})=0$~\cite{dispersion_relation}, which is equivalent to the condition
\begin{equation}\label{D_of_k}
\Delta(k)\equiv \left[1-\rho_{1}^0\hat{c}_{11}(k)\right]\left[1-\rho_{2}^0\hat{c}_{22}(k)\right]-\rho_{1}^0\rho_{2}^0\hat{c}_{12}^2(k)=0.
\end{equation}
It is worth mentioning that the values of the diffusion coefficients $D_1$ and $D_2$ do not influence the thermodynamic equilibrium state, i.e. the minimum of the free energy and therefore are not further discussed.
\begin{figure*}[t!]
\centering
\includegraphics[width=1\linewidth]{fig1.jpg}
\caption{Phase diagram of the different states in the ($\rho^0_1, \rho^0_2$) plane corresponding to the model of Eq. (\ref{pairpot}) and parameters in Table \ref{table1}. The letters A, B, C, D inside the diagram identify the side panel visualizations for the two densities $\rho_1 R_{11}^2$ and $\rho_2 R_{11}^2$ at the labeled state point.
The linear instability line is shown by the dash-dotted line.}
\label{pahse_diagram}
\end{figure*}
\subsection{The Ornstein-Zernike equation for binary mixtures}
The total correlation functions $h_{ij}(r)$ which are related to the radial distribution functions $g_{ij}(r)$ via $h_{ij}(r)=g_{ij}(r)-1$ can be obtained using the Ornstein-Zernike (OZ) equations. For a binary (isotropic) fluid mixture, the OZ equations are \cite{hansen_mcdonald}:
\begin{equation}
h_{ij}(r)=c_{ij}(r)+\sum_{k=1,2}\rho_{k}^0\int d\textbf{r}' c_{ik}(|\textbf{r}-\textbf{r}'|)h_{kj}(\textbf{r}').
\label{eq:OZ}
\end{equation}
These coupled equations must be solved in conjunction with the (exact) closure relations
\begin{equation}
c_{ij}(r)=-\beta\phi_{ij}(r)+h_{ij}(r)-\ln[1+h_{ij}(r)]+B_{ij}(r),
\label{eq:closure}
\end{equation}
where $B_{ij}(r)$ are the so-called bridge functions \cite{hansen_mcdonald}. For example, the HNC approximation consists of setting $B_{ij}(r)=0$ $\forall r$. The RPA approximation can be obtained by linearising the logarithm in Eq.~(\ref{eq:closure}) together with the HNC approximation. It is also worth mentioning that when modelling strongly repulsive and short-range interactions, the direct correlation functions can be obtained employing the Percus-Yevick approximation \cite{hansen_mcdonald}.
\subsection{The partial static structure factors}
The partial static structure factors, which are accessible via scattering experiments, are defined as \cite{hansen_mcdonald, dispersion_relation}:
\begin{equation}
\begin{split}
S_{11}(k)&=1+\rho_{1}^0\hat{h}_{11}(k);\\
S_{22}(k)&=1+\rho_{2}^0\hat{h}_{22}(k);\\
S_{12}(k)&=\sqrt{\rho_{1}^0\rho_{2}^0}\hat{h}_{12}(k),
\end{split}
\end{equation}
where the Fourier transform of the total correlation functions $\hat{h}_{ij}$ are
\begin{equation}
\hat{h}_{ij}(k)=\frac{N_{ij}(k)}{\Delta(k)},
\end{equation}
with the numerators $N_{ij}(k)$ given by
\begin{equation}
\begin{split}
N_{11}(k)&=\hat{c}_{11}(k)+\rho_{2}^0\left[\hat{c}_{12}^2(k)-\hat{c}_{11}(k)\hat{c}_{22}(k)\right];\\
N_{22}(k)&=\hat{c}_{22}(k)+\rho_{1}^0\left[\hat{c}_{12}^2(k)-\hat{c}_{11}(k)\hat{c}_{22}(k)\right];\\
N_{12}(k)&=\hat{c}_{12}(k).
\end{split}
\end{equation}
Note that in the case of a stable liquid, $\Delta(k)>0$ $\forall k$.
\section{Model for core-shell assemblies}\label{Sec:capsules}
\begin{figure}[b!]
\centering
\includegraphics[width=1\linewidth]{fig1_b.jpg}
\caption{Different interaction potentials $\phi_{ij}$ with the parameters reported in Table \ref{table1}.}
\label{potentials_1}
\end{figure}
For computational efficiency, we focus here on a two-dimensional (2D) model. We concentrate on a binary model that potentially self-assembles such that particle-like aggregates composed of a core and a corona, shielding the latter from bulk solution, can be expected. Practical examples of such assemblies are, e.g., surfactant micelles or block-copolymer particles. In aqueous environments, the core region is hydrophobic and the corona hydrophilic. To capture such compositional organization, we model the effective interactions of the different components as
\begin{align}\label{pairpot}
\phi_{\rm 11}(r)&=\varepsilon_{\rm 11}e^{-\left(r/R_{\rm 11}\right)^4}; \nonumber \\
\phi_{\rm 12}(r)&=\varepsilon_{\rm 12}^{-}e^{-\left(r/R_{\rm 12}^{-} \right)^2}+\varepsilon_{\rm 12}^{+}e^{-\left(r/R_{\rm 12}^{+} \right)^2};\\
\phi_{\rm 22}(r)&=\varepsilon_{\rm 22}e^{-\left(r/R_{\rm 22} \right)^2}, \nonumber
\end{align}
where $\beta\varepsilon_{\rm 12}^{-} < 0$ and $\beta\varepsilon_{\rm 12}^{+} > 0$, with the constraint $\vert \beta\varepsilon_{\rm 12}^{-}\vert < \vert \beta\varepsilon_{\rm 12}^{+}\vert$. Additionally, $R_{\rm 12}^{-} > R_{\rm 12}^{+}$. Together, these yield a short ranged repulsion and a long ranged attraction between the two species. Generally, $R_{11}$ and $R_{22}$ are a measure of the sizes of the two components (for polymeric systems these are comparable to the radius of gyration of the two species). The different variables chosen in this work for the model are reported in Table \ref{table1}, and the corresponding potentials $\phi_{ij}$ are shown in Fig.~\ref{potentials_1}.
\begin{table}[h!]
\centering
\begin{tabular}{| c | c | c | c | c | c | c | c | c |}\hline
$\beta\varepsilon_{\rm 11}$ & $\beta\varepsilon_{\rm 12}^{-}$ & $\beta\varepsilon_{\rm 12}^{+}$ & $\beta\varepsilon_{\rm 22}$ & $R_{\rm 11}$ & $R_{\rm 12}^{-}$ & $R_{\rm 12}^{+}$ & $R_{\rm 22}$\\ \hline\hline
1 & -1 & 25 & 1 & 1 & 0.5 & 0.1 & 1 \\ \hline
\end{tabular}\caption{Set of dimensionless parameters used in Eq.~(\ref{pairpot}) to model core-shell particles.}\label{table1}
\end{table}
Note that for the cross-interaction, the integrated strength $2\pi\int_0^{\infty} r \phi_{12}(r)dr=0$. The different potentials in Eq.~(\ref{pairpot}) have been chosen following the fact that in order to see the formation of clusters of purely repulsive particles, one needs in the model an intrinsic length scale that can eventually be unstable. Such property implies a symmetry breaking, and thus the formation of periodic (or aperiodic) patterns. In the current work, the interaction between particles of species 1 bear such a feature and is modelled using the so-called generalised exponential model of order $n$ (GEM-$n$) \cite{coslovich2012cluster, archer2014solidification, archer2016generation, archer2016crystallization, scacchi2018flow, caprini2019comparative, gotze2006structure, overduin2009phase, overduin2009clustering, camargo2010dynamics, camargo2011interfacial}. In this work $n=4$. In order to keep the model as simple as possible, species 2 is modelled using the Gaussian core model (GCM), which does not belong to the class of $Q\pm$-interactions \cite{likos2007ultrasoft}, i.e. it does not lead to cluster formation in single-component mixtures (except for $\varepsilon\beta > 100$~\cite{louis2000mean}, irrelevant here). The GCM has been extensively used to model polymeric interactions \cite{bolhuis2001accurate, likos2001effective, likos1998star,louis2000can, gotze2004tunable, likos2006soft, lenz2012microscopically, mladek2006formation, gotze2006structure}. The second term in the cross-interaction $\phi_{12}$ models a short range repulsion, which is necessary to avoid overlap of particles of different species. However, in order to favor the formation of a shell surrounding the core of the particle, attraction between different species is also necessary. The latter is achieved via the first term in $\phi_{12}$. The range of the attractive and repulsive part of $\phi_{12}$ are intimately related to molecular details of the components. They can be controlled, e.g., by molecular weight, solvent composition, or changing salt concentration or pH of the solution, depending on the nature of the species. It is worth mentioning that the results presented here are expected to be valid for a variety of interaction potentials which bear properties similar to those in the current model, Eq.~(\ref{pairpot}). The results can also be expected to hold for a large collection of parameters different from those reported in Table~\ref{table1}.
\subsection{Results}
\begin{figure}[h!]
\centering
\includegraphics[width=1\linewidth]{fig2.pdf}
\caption{Different onsets of the linear instability, i.e. the marginal condition for density modulation in the system, as a function of $\rho_1^0$ (dash-dotted line in the central panel of Fig.~\ref{pahse_diagram}). The critical wave numbers $k_{\rm c}$, marked by circles, shift to higher values as $\rho_1^0$ decreases, which corresponds to the lattice constant of the expected periodic structure decreasing. Additionally, the peaks in the dispersion relations broaden as the bulk density of species 1 decreases. This indicates the selectivity of the wave number $k_c$ softens.}
\label{dispersion_plot}
\end{figure}
We explore the density-density plane ($\rho^0_1, \rho^0_2$) to resolve the equilibrium self-assembly phases adopted by a system following the interaction parameters reported in Table \ref{table1}. Here $\rho^0_1$ and $ \rho^0_2$ are the bulk number densities of species 1 and species 2, respectively. The resulting assembly phase diagram and the four different assembly configurations adopted by the system in different composition regions are shown in Fig.~\ref{pahse_diagram}. For each state point, we report the density of both species 1 and species 2, labelled accordingly. The phase diagram indicates that the system adopts a homogenous mixture corresponding to uniform solution of species 1 and 2 at low concentrations of $\rho_1$ while an increase of $\rho_1$ leads to an onset of species 1 condensation into islands that exhibit hexagonal symmetry. This occurs at a concentration that is dependent on $\rho_2^0$. The data of Fig.~\ref{pahse_diagram} indicate that the system has only one phase transition. The transition happens when passing from the homogeneous state ({\it uniform}) to any state with hexagonal symmetry in species 1 assembly pattern. The presence of a single phase transition is expected, since the system has only one intrinsic length scale.
While species 1 assembles into condensation islands with hexagonal symmetry in spatial locations throughout the structure-forming part of the phase diagram, species 2 shows interesting configurational changes when exploring different concentrations. In the first two of these, species 2 adopts a distribution where it avoids the locations where species 1 is enriched as clusters. A region with very low density, i.e. a depletion region of $\rho_2$ forms around the $\rho_1$ islands. In the {\it $\rho_1$ islands in $\rho_2$} states of Fig. \ref{pahse_diagram}, most of $\rho_2$ forms a uniform background solution with density very close to the bulk value $\rho_2^0$ in the regions where $\rho_1$ is absent. An example of this configuration is shown in Fig.~\ref{pahse_diagram}, panels ${\rm A}_{\rho_1}$ and ${\rm A}_{\rho_2}$. Such assembly corresponds to, for example, multicomponent polymer solutions in which one species forms polymer particles and another remains soluble, yet immiscible with the former species. One can consider this also as two immiscible liquids, with $\rho_1$ forming droplets surrounded by the phase in which $\rho_2$ spreads uniformly to form a continuous phase. Such phase separation occurs, for example in intracellular liquid-liquid phase separation or coacervation of macromolecule solutions.
In states labeled as {\it $\rho_1$ islands in structured $\rho_2$} in Fig.~\ref{pahse_diagram}, similar holes in $\rho_2$ around the $\rho_1$ condensation sites characterize the density distribution. However, $\rho_2$ distribution shows an underlying periodic structure in the regions where particles 2 are in the liquid state (see panels ${\rm B}_{\rho_1}$ and ${\rm B}_{\rho_2}$ of Fig.~\ref{pahse_diagram}). This represents species 1 driving correlations into the structure formation of species 2, despite species 2 retaining partial liquid character. Transition from the uniform background species 2 to this correlated species 2 distribution is an indication of increasing effective attraction between the assembled species 1 islands via the species 2 condensation. This would show in practical systems as increasing solution viscosity.
The {\it $\rho_1$ islands with $\rho_2$ condensation rims} state points correspond to configurations in which species 2 gather around the peaks of $\rho_1$ defining the primary hexagonal structure. Away from these condensation islands however, the density of both $\rho_1$ and $\rho_2$ is depleted. These state points correspond to a partial (or soft) encapsulation of species 1 by a coating of species 2. An example of the latter configuration is shown in Fig.~\ref{pahse_diagram}, panels ${\rm C}_{\rho_1}$ and ${\rm C}_{\rho_2}$. Such assembly occurs, for example, in amphiphile mediated solubilization or dispersion under conditions where part of the coating species remains soluble. This assembly also corresponds to part of the species 2 assembling into micellar or core-shell particles, part remaining soluble.
Finally, the region labeled {\it $\rho_1$ islands encapsulated by $\rho_2$} correspond to structures where species 2 surrounds (almost) completely the clusters of species 1. This leads to a strong encapsulation configuration corresponding to micellar or core-shell type well-defined layered self-assembly structures in e.g. surfactant and polymer systems. This configuration is shown in panels ${\rm D}_{\rho_1}$ and ${\rm D}_{\rho_2}$ of Fig.~\ref{pahse_diagram}. The asymmetry in the shells is a result of the relatively strong and long-ranged repulsion between particles of species 2 via $\phi_{22}(r)$. The asymmetry can be reduced by diminishing the strength and range of this repulsion (see Appendix). The instability line obtained from the dispersion relation $\omega_{\pm}(k)$ is reported in the central panel of Fig.~\ref{pahse_diagram}. Left of this line the system is expected to be linearly stable which leads to the equilibrium density distribution being homogeneous. At right of this line, however, the system is expected to form patterns with lattice spacing corresponding to $\approx 2\pi/k_{\rm c}$. Here $k_{\rm c}$ is the wave number at which the system is marginally unstable. This corresponds to $\omega_{+}(k_{\rm c})=0$ and ${d\omega_{+}(k_{\rm c})}/{dk}=0$. Careful examination of the critical wave number $k_{\rm c}$ value for different pairs of bulk densities $\rho_1^0$ and $\rho_2^0$ reveals a monotonic shift such that moving along the instability line from high to low values of $\rho_1^0$, $k_{\rm c}$ increases. This means also that the lattice constant of the underlying hexagonal structure decreases. Additionally, the peak in the dispersion relation broadens, which translates into a less selective instability. The latter result is reported in Fig.~\ref{dispersion_plot}.
It is worth mentioning that at temperatures relevant here, two dimensional monodisperse systems with only hexagonal structures of lattice spacing $\approx R_{11}$, i.e. approximately constant lattice spacing for all densities, are expected~\cite{archer2015soft}. However, this is not true in three dimensions, where a series of isostructural phase transitions are expected at low temperatures~\cite{zhang2010reentrant,wilding2014demixing}.
\section{Model for mixed particles}\label{Sec:core_shell}
Exchanging the short-ranged repulsion and the long-ranged attraction with a short-ranged attraction and a long-ranged repulsion enables obtaining density distributions which are commensurate with mixed assembled particles. This means $\beta\varepsilon_{\rm 12}^{-} < 0$, $\beta\varepsilon_{\rm 12}^{+} > 0$, with the constraint $\vert \beta\varepsilon_{\rm 12}^{-}\vert > \vert \beta\varepsilon_{\rm 12}^{+}\vert$, and also $R_{\rm 12}^{-} < R_{\rm 12}^{+}$. Monodisperse suspensions interacting via a different short range attraction and long ranged repulsion (SALR) potential in two dimensions exhibit fascinating microphase separation~\cite{archer2008two}. Noteworthy is also the interesting study of phase behavior of a SARL fluid in three dimensions~\cite{archer2007phase}. Following Section \ref{Sec:capsules}, the effective pair potentials used in this model are described by Eq.~(\ref{pairpot}), however, the coefficients differ from Sec.~\ref{Sec:capsules} and are reported in Table \ref{table2}. To control the size and composition of the particles we vary the range of the attractive and repulsive contributions in $\phi_{12}$, i.e. $R_{12}^{-}$ and $R_{12}^{+}$, respectively, keeping $R_{\rm 12}^{-} < R_{\rm 12}^{+}$. Note that the integrated strength of the cross-interaction is positive for all five cases discussed here, i.e. $2\pi\int_0^{\infty} r \phi_{12}(r)dr>0$.
\begin{table}[h!]
\centering
\begin{tabular}{| c | c | c | c | c | c | c |}\hline
$\beta\varepsilon_{\rm 11}$ & $\beta\varepsilon_{\rm 12}^{-}$ & $\beta\varepsilon_{\rm 12}^{+}$ & $\beta\varepsilon_{\rm 22}$ & $R_{\rm 11}$ & $R_{\rm 22}$\\ \hline\hline
1 & -1 & 0.5 & 1 & 1 & 1 \\ \hline
\end{tabular}\caption{Set of dimensionless parameters used in Eq.~(\ref{pairpot}) to model mixed particles.}\label{table2}
\end{table}
\subsection{Results}
\begin{figure}[h!]
\centering
\includegraphics[width=1\linewidth]{fig3.pdf}
\caption{Cross-interaction potentials $\phi_{12}$ following the parameters reported in Table \ref{table2} for varying values of $R_{12}^{-}$ and $R_{12}^{+}$ in Table \ref{table3}.}
\label{potentials}
\end{figure}
\begin{figure*}[t!]
\centering
\includegraphics[width=1\linewidth]{fig4.jpg}
\caption{Average density distribution $\rho_i R_{11}^2$ for $i=1$ (top row) and for $i=2$ (bottom row) for the set of parameters corresponding to configurations E, F, G, H and I. The parameters used for modelling $\phi_{12}$ are described in Table \ref{table2} and the varying values of $R_{12}^{-}$ and $R_{12}^{+}$ are reported in Table \ref{table3}.}
\label{core_corona}
\end{figure*}
We set the average densities of the two components of the system to $\rho_1^0R_{11}^2=6.5$ and $\rho_2^0R_{11}^2=3$, respectively, and study the equilibrium configurations obtained using the DFT described in Section \ref{sec:DFT} as a function of different cross interaction potentials.
\begin{table}[h!]
\centering
\begin{tabular}{| c || c | c | c | c | c | }\hline
Model & E & F & G & H & I \\ \hline\hline
$R_{12}^{-}$ & 0.15 & 0.2 & 0.3 & 0.5 & 0.75 \\ \hline
$R_{12}^{+}$ & 0.4 & 0.5 & 0.6 & 1 & 1.5 \\ \hline
\end{tabular}\caption{Range of the attractive and repulsive contribution in the cross particles interaction potential $\phi_{12}(r)$ leading to mixed particles with different density gradients.}\label{table3}
\end{table}
The form of the inter-species interaction potentials described by the parameters in Table \ref{table3} are shown in Fig.~\ref{potentials}.
Figure \ref{core_corona} shows the resulting equilibrium density distributions of both species with species 1 as the top row and species 2 as the bottom row. As before, species 1 assembles into condensation islands with hexagonal spacing throughout the examined parameter range. Only minor changes in condensation peak distribution height and width can be observed. However, rather interesting assembly changes occur to species 2. In case E, the very short attraction causes a relatively small increment of $\rho_2$ around the locations occupied by species 1. The repulsive part of $\phi_{12}$ however causes a depleted region around the latter peaks. Here $\rho_2$ forms a thin, skin-like, layer around the species 1 condensation island and is separated from bulk by a depletion region. Beyond the depletion region, species 2 adopt an almost liquid-like assembly but showing preferred density locations as triangles formed between the $\rho_1$ neighboring peaks. This assembly pattern means that also for the species 2, hexagonal order with lattice parameter differing from that of species 1 ordering emerges.
In case F, the total attraction strength between the two components is increased. The attraction is strong enough to pull almost all species 2 particles to the sites enriched in species 1. The clusters of species 2 are very narrow for this set of parameters, and actually now species 1 forms the coating layer. However, no clear core phase exists and the two species remain mixed in the island center regions, however with strong concentration gradients. A low density honeycomb-like region of $\rho_2$ is visible far from these condensation peaks indicating some of species 2 remaining soluble outside species 1 islands.
Similar to case F, also in setup G, $\rho_2$ exhibits sharp and narrow peaks at the locations of particles 1. The width of the $\rho_2$ peaks is practically the same as those of $\rho_1$. This means that both species occupy the same region and readily mix. Moving to cases H and I, the widths of both $\rho_1$ and $\rho_2$ peaks increase. Eventually, the peaks of $\rho_2$ become larger than those of species 1 leading to a case where species 1 can be found mostly in the centre of the particles. Even in this case I, also $\rho_2$ is enriched at the same location meaning that both density and condensation have a gradient in the island.
To summarize the assembly findings, Figure \ref{cut} shows the 1D density profiles of the islands for states F, H and I. The data are normalised by the maximal values of the distributions and centered at the maximum of the peak. The figure visualizes how to control the size of the mixed particles, as well as guidelines to engineering the density gradient of the two species mixing inside the particle, but also the overall particle density profile. At a practical level, such concentration gradients mean variations in the local chemical environment. This has a major significance for applications. For example, hydrophilicity-hydrophobicity gradients control drug species partitioning in the carrier and in-out diffusion, which control the binding and release of carried species.
\begin{figure}[h!]
\centering
\includegraphics[width=1\linewidth]{fig5.pdf}
\caption{One-dimensional density distributions $\rho_1(y)$ and $\rho_2(y)$ of a single condensation peak for cases F, H and I. The density is normalised by its maximal value. Minor peak size changes are found in $\rho_1$ for the different cases. However, $\rho_2$ peaks change significantly. In case F, species 2 occupies mainly the center of the particle. In case H, the two species are very well mixed. In case I, species 1 occupies the center of the aggregate.}
\label{cut}
\end{figure}
\section{Summary and Conclusions}
Using classical density functional theory, we have presented a general framework for particle systems with competing molecular interactions. We have shown how this can lead to the spontaneous formation of structurally complex self-assembly matter such as core-shell structures and well-defined assemblies where the component species mix, yet granting control on mixing properties. We have identified the key intermolecular interactions features responsible for these complex self-assembling structures and demonstrated how the assembly and species mixing can be controlled. To this end, linear stability analysis was used as a guiding principle.
This insight can be employed to better understand multi-component self-assembly systems that emerge in, e.g., cellular function and complex biological structure formation. It also paves the way for engineering advanced complex materials solutions for, e.g., drug delivery including anticancer therapy approaches that often rely on sequential delivery of multiple drugs, see e.g. \cite{Aw2013,Wu2018}.
In this work, we applied DFT to two different cases. First, we mapped out a phase diagram where one species occupies the core of the self-assembled particles and the second species surrounds the latter either as a freely solvated species or a corona layer depending on the concentrations of the two components. This assembly response, decribed in Sec.~\ref{Sec:capsules}, can be considered to correspond to spontaneous formation of droplets (for example, biomolecular condensation), polymer core-shell particles or surfactant micelles. Such systems are ubiquitous in biological systems but also chemical engineering such as paints, cosmetics, and pharmaceutical formulations.
We also demonstrated that via the effective interactions between the species, it is possible to carefully control not only the size of the self-assembling particles but also their internal composition gradient and mixing of the species inside the particles. Furthermore, the findings give access to controlling overall density distribution. These findings are summarized in Sec.~\ref{Sec:core_shell}. A handle to self-assembling particle internal composition has applications in molecular partitioning (hydrophobicity gradients), diffusion control, chemical reactions engineering but also in controlling solubilization and dispersion.
We note that the present theory is not restricted to (ultra) soft interaction potentials that were the main focus in this work. In fact, using different approximations for the free energy functionals, the theory developed here can be applied to a wide range of systems, including, e.g., colloids and nanoparticles.
Additionally, one should note that the theoretical approach presented here can also be applied to systems exhibiting multiple length scales in their self-assembled configurations~\cite{scacchi2021self}. Examples of such systems that constitute coexisting structures with different length scales include, e.g., multi-core micelles \cite{wang2011discovering, chen2013formation, duxin2005cadmium, iatridi2011ph, ueda2011unicore, chen2012formation}, coacervate droplets in biocondensates (liquid-liquid phase-separation in biological systems) \cite{lu2020multiphase} and multiple emulsions \cite{sheth2020multiple}.
\section{Acknowledgment}
This work was supported by Academy of Finland grants No. 309324 (M.S.) and Nos. 307806 and 312298 (T.A-N.). T.A-N. has also been supported by a Technology Industries of Finland Centennial Foundation TT2020 grant. We are grateful for the support by FinnCERES Materials Bioeconomy Ecosystem. Computational resources by CSC IT Centre for Finland and RAMI -- RawMatters Finland Infrastructure are also gratefully acknowledged.
\section{Author contributions}
A.S. conceived the work, executed the theoretical calculations and wrote the first draft of the manuscript. M.S. and T.A-N. supervised the research. All authors contributed to writing.
\section{Data availability statement}
The data that support the findings of this study are available within the article. Additional data are available upon request.
|
\section{Introduction}
\label{sec:introduction}
Evolution of science and technology is episodically redirected as our understanding improves. For example, the late 19$\mathrm{^{th}}$ to mid-20$\mathrm{^{th}}$ Century recognition of nonlinear dynamical systems as not purely stochastic, or, purely deterministic was both troubling and opportune. The subsequent acknowledgment of emergent behaviors as a {\em property} of such systems can be considered such a redirection\cite{Anderson1972}. Progress along this front established dynamical systems theory as a new branch of science \cite{Journals2020}. Accordingly, it is now well-accepted that canonical statistical and parametric equation-based models applied to nonlinear systems, which turn out to be nearly ubiquitous in complex technologies and nature, are often problematic \cite{DeAngelis2015}.
Currently, machine learning is advancing new directions in data and system analysis. Techniques such as manifold learning \cite{Lin2015} and diffusion maps \cite{Coifman2006} rely on a {\em data-driven}, inductive approach, wherein the data-itself reveals underlying dynamics and behavioral complexities, rather than a presumptive, model-based deductive analysis. As effective as these methods are at finding patterns and forecasting, their scientific utility is less clear as the patterns uncovered do not typically translate into mechanistic insight. For example, identifying intervariable interactions and causal relationships between system variables.
A possible exception is a set of dynamically-based, data-driven analytics termed empirical dynamic modeling (EDM), which have developed into a useful toolset for system analysis and forecasting\cite{Sugihara1990, Sugihara1994, Dixon1999, Sugihara2012, Ye2016}. EDM operates in multidimensional state-space, either from an explicit multidimensional data set, or, from a diffeomorphic reconstruction of the state-space, for example, by application of Takens embedding theorem \cite{Takens1981, Deyle2011}. EDM is not based on parametric presumptions, fitting statistics, or, specifying equations; representing a data-driven approach amenable to nonlinear dynamics as noted by DeAngelis \cite{DeAngelis2015}. The unfamiliar reader is referred to the lucid EDM introduction in reference \cite{Chang2017}.
Since these methods are data-driven, it is possible that observational noise can interfere with accurate state-space representations, degrading inference and forecast skill. We therefore seek methods to improve state-space representations derived from noisy observations, and here, turn to empirical mode decomposition (EMD).
EMD decomposes oscillatory signals into scale-dependent modes termed intrinsic mode functions (IMF) without constraints of linearity or stationarity as presumed by Fourier, wavelet or Eigen decomposition. Application of the Hilbert transform to IMFs provides time-dependent instantaneous frequency estimates, with the combination of EMD and IMF Hilbert spectra constituting the Hilbert-Huang transform (HHT) \cite{Huang2008}. IMFs are particularly astute at isolating physically-meaningful dynamics \cite{Huang2008}, and, their data-driven isolation of noise is a cornerstone of EMD utility \cite{Dai2019}. An introduction and review are found in references \cite{Huang2008} and \cite{Looney2015}.
Since dynamics of interest are often low-dimensional and contained within a bounded state-space, univariate projections of state-space trajectories are often expressed as oscillatory signals. EMD provides a model-free and natural way to filter noise and isolate physically relevant modes of oscillatory time series, while EDM enables data-driven state-space metrics for forecasting and quantifying cross-variable interactions.
Here, we combine the model-free modal decomposition of EMD with the multivariate state-space representation inherent in EDM to improve forecasting and discovery in the presence of noisy or confounded observations. We term this empirical mode modeling (EMM).
We demonstrate the utility of this synthesis in two distinct applications. First a nonlinear, chaotic mathematical system, the Rössler attractor, and second, a nonlinear geophysical system with ecological importance, salinities in Florida Bay within Everglades National Park.
\subsection{Model fitness and application}
\label{sec:modelfitness}
Following Granger \cite{Granger1969}, we adopt a position that model predictability reflects the amount of information a model contains regarding a dynamical system. We therefore use model fidelity expressed as the Pearson correlation $\rho$ between observations and model predictions as a fitness metric. Note that the models themselves can be equation-free and nonlinear, and, that other metrics such as mutual information are equally valid.
When models are used in an operational forecast system it is typical to assess model fitness on {\it out-of-sample} data, that is, the model is trained on a subset of available data with fitness assessed on a prediction (validation) set disjoint from the training data. Out-of-sample prediction characterizes model stability and the ability to generalize to untrained data.
{\it In-sample} models, where the training and prediction sets overlap, are informative when discovering/assessing the ability of the model to represent system dynamics. For example, as a function of state-space variables and their interactions, or, in response to noise/confounded states. We use both formats to assess model fitness and dynamical information content.
\section{Empirical Mode Modeling}
\label{sec:methods}
The basis of EMM is to create a multidimensional state-space from empirical mode decomposition of observational time series. The resultant IMFs, or subsets of IMFs representing the system state-space, are then analyzed within the empirical dynamic modeling framework. The reader is referred to reference \cite{Looney2015} for details of empirical mode decomposition, and reference \cite{Chang2017} for empirical dynamic modeling.
Generically, EMM can be implemented with the following steps:
\begin{enumerate}
\item Determine the signal-to-noise ratio (SNR) of the time series. If the SNR is less than 3 dB, EMM may provide improvements in state-space representation and forecasting.
\item Apply empirical mode decomposition to the time series creating a set of intrinsic mode functions (IMFs). This can be done for multiple time series.
\item Optionally select IMFs that best represent the dynamics. This is an ongoing area of research. Here, we use two methods:
\begin{enumerate}
\item Multiview embedding as described in reference \cite{Ye2016}, selecting the set of IMFs that maximize model predictability as exemplified in section \ref{IMFSelection}.
\item Select IMFs based on spectral filtering, or, manual inspection to remove high frequency noise or low frequency oscillations external to the dynamics. This is shown in section \ref{stateSpaceModels}.
\end{enumerate}
\item Apply empirical dynamic modeling using the selected IMFs as a multivariable state-space. Application of EDM to multivariable state-space is described in reference \cite{Dixon1999}. EDM analysis can be performed using simplex\cite{Sugihara1990}, sequential locally weighted global linear maps (S-maps)\cite{Sugihara1994}, or convergent cross mapping\cite{Sugihara2012}.
\end{enumerate}
\section{Rössler Dynamics}
The Rössler attractor is a 3-dimensional coupled dynamical system known to exhibit chaotic dynamics\cite{Rossler1976}. Our implementation is defined as:
\begin{align}
\label{equ:Rosslerx}
\frac{dx}{dt} = -y-z\\
\label{equ:Rosslery}
\frac{dy}{dt} = x + ay\\
\label{equ:Rosslerz}
\frac{dz}{dt} = b + z(x-c)
\end{align}
\noindent with constants $a$ = 0.4, $b$ = 0.4, $c$ = 4, initial state values $x_0$ = 1, $y_0$ = 0, $z_0$ = 1, time increment = 0.01, and time span $T$ = [0-500]. We then subsample the integrated solutions of equations \ref{equ:Rosslerx} - \ref{equ:Rosslerz} at a time increment of 0.1, and ignore the time span from 0 to 200 resulting in a 3-D time series of 3000 points from time 200 to 500. Multispectral noise ($N$) is generated following reference \cite{Timmer1995} as a combination of brown and pink noise:
\begin{equation}
N = A~( B~ p_n + C~ b_n )
\label{equ:Noise}
\end{equation}
\noindent where $p_n$ is "pink" noise with power spectral density $1/f$, $b_n$ is low frequency "Brownian" noise with power spectrum $1/f^2$, where $f$ is frequency, and, noise amplitude $A$ ranges from 0 to 64; $B$ = 0.5, $C$ = 1. Corresponding signal-to-noise ratios are listed in Table \ref{table:SNR}.
\begin{table}[h!]
\centering
\caption{Rössler signal-to-noise ratios}
\label{table:SNR}
\begin{tabular}{r|cccccccccc}
\hline\noalign{\smallskip}
Noise A & 1 & 2 & 4 & 8 & 12 & 16 & 24 & 32 & 48 & 64\\
\noalign{\smallskip}\hline\noalign{\smallskip}
SNR dB & 10.08&7.00&4.12&1.07&-0.73&-1.89&-3.71&-4.94&-6.69&-7.95\\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
Fig. \ref{fig:RosslerNoise} shows four examples of the Rössler state-space with different levels of additive noise. It is clear that as noise increases, trajectories in the state-space become corrupted and entangled, and, we expect that methods such as manifold learning and EDM that rely on state-space representations of system dynamics can perform poorly as noise increases.
\begin{figure}
\centering
\includegraphics[width=1\textwidth]{graphics/Rossler-3D.png}
\caption{State-space time series and trajectories of Rössler dynamics with multispectral noise.}
\label{fig:RosslerNoise}
\end{figure}
\subsection{Empirical mode decomposition}
The first step in EMM is to obtain IMFs of the observation time series. Fig. \ref{fig:RosslerIMF} shows the Rössler variable $x$ and it's empirical mode decomposition IMFs at four values of additive multispectral noise amplitude $A$ from equation \ref{equ:Noise}. We can see that IMFs partition timescale variance into modes with decreasing ranges of instantaneous frequency, high frequencies in the low-order modes and lower frequencies in the higher-order modes. We will leverage this partition as a data-driven filter to select IMFs that better represent the underlying dynamics.
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{graphics/R1_IMF_Noise.png}
\caption{Rössler state-space variable $x$ time series with 3 instances of additive noise, and the resultant IMFs.}
\label{fig:RosslerIMF}
\end{figure}
\subsection{Empirical dynamic modeling}
Empirical dynamic modeling provides state-space forecasting and cross-variable analysis, here, we use the simplex algorithm to project univariate state-space dynamics to assess the fitness of state-space representations derived from traditional time-delay embeddings, and, from IMFs of observed variables. IMFs are computed over the entire time series, and, partitioned into training library and prediction sets in the same manner as any EDM candidate time series. The forecast variable is $z$ from Eq. \ref{equ:Rosslerz}, with forecasts made from three different state-space representations consisting of vectors $[x,y]$, $[x,y,z]$, or their IMFs.
The simplex algorithm is a state-space projection from a simplex of nearest neighbors in the training data ({\it library}) closest to the state-space of the prediction point. The simplex consists of $E+1$ points, where $E$ is the embedding dimension in the case the state-space is a time-delay embedding, or, simply the dimension of the multivariate state-space. Reference \cite{Sugihara1990} describes the simplex algorithm.
\subsubsection{Model comparison}
We examine four different EDM models (Table \ref{table:RosslerEDM}). The first is a full variable state-space of $[x,y,z]$ ($E\!=\!3$) with complete information. This serves as a reference. Second, a multivariable $[x,y]$ ($E\!=\!2$) state-space representing a naive predictor without information of the forecast variable $z$. Third, a Takens time-delay state-space of variables $[x,y]$ with embedding dimension $E_e\!=\!3$ resulting in a state-space dimension of $E\!=\!6$, and fourth, multivariate state-space of all IMFs of $[x,y]$. Here, the dimension $E$ is equal to the number of IMFs.
Since we are interested in the ability of these different representations to model the underlying dynamics, all models use the target time series of noiseless variable $z$, and perform simplex projection at a forecast interval of $T_{p}\!=0$. All models are evaluated in an out-of-sample prediction with the training library consisting of points [1-2000], and prediction set [2001-3000].
\begin{table}
\centering
\caption{Rössler EDM State-space Variables}
\label{table:RosslerEDM}
\begin{tabular}{lll}
\hline\noalign{\smallskip}
State-space & Variables & E \\
\noalign{\smallskip}\hline\noalign{\smallskip}
Multivariable & $x,y,z$ & 3 \\
Multivariable & $x,y$ & 2 \\
Takens time-delay& $x,y$ & 6 \\
Multivariable & $IMF_{x},IMF_{y}$ & $N_{IMF}$ (12-18) \\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
Fig. \ref{fig:RosslerEDMNoise} shows ensemble simplex projection results as function of signal-to-noise ratio for the four state-space representations over 1000 noise realisations at each noise amplitude. Here, we find the naive predictor $x,y\to z$ incapable of meaningful prediction, while the Takens time-delay embedded model (Takens $x,y\to z$) approaches that of the reference model ($x,y,z\to z$) at high signal-to-noise ratios. At high SNR, above approximately 3 dB, the Takens model outperforms EMM based on IMFs, however, as SNR falls below 3 dB EMM outperforms the time-delay representation providing information recovery closer to the full-information reference model. A SNR of 3 dB translates to signal power a factor of 2 greater than the noise power.
As noted, at SNR greater than 3 dB, EMM performs slightly worse than the time-delay embedding. It may be that the lower dimensional time-delay embedding provides a more compact state-space simplex that better represents state evolution when noise is low. Here, the time-delay state-space is 6 dimensional, while the EMM multivariate space is 12 dimensional.
An alternative EDM algorithm: Sequential Locally Weighted Global Linear Maps (S-map) \cite{Sugihara1994} applies a localisation kernel to the library of states, rather than using a fixed number of k-nearest-neighbors to define a simplex. If dimensional inflation of the simplex is the cause of degraded performance at high SNR, S-map based EMM may address this issue. However, the goal of EMM is to improve state-space models in the presence of noise.
These results suggest that when noise power is more than 1/2 the signal power, even an indiscriminate use of IMFs as EDM state-space variables instead of time-delay embedding can improve cross-variable projection $x,y\to z$, evincing a more inforamtic representation of the underlying dynamics.
\begin{figure}
\centering
\includegraphics[width=2.8in]{graphics/Rossler-V3-Noise-dB.png}
\caption{Pearson correlation and RMSE of simplex projected Rössler variable $z$ as a function of noise over 1000 noise realisations at each noise amplitude. Lines are mean values, error bars $\pm$ standard error.}
\label{fig:RosslerEDMNoise}
\end{figure}
\subsubsection{IMF selection}
\label{IMFSelection}
It was observed that low order IMFs capture high frequency noise not present in the underlying Rössler dynamics, and, that use of all IMFs as EDM state-space variables improves predictability in the presence of noise. It is then natural to ask whether subsets of IMFs can provide additional gains in state-space representation and model fidelity, or, provide equitable predictability from a lower-dimensional representation.
One method to identify noise-dominated IMFs is to examine the variance of IMF instantaneous frequencies (IF) since high variance in instantaneous frequency is indicative of noise, while low variance suggests a relatively stable oscillatory mode reflective of low-dimensional dynamics. The idea being to discard IMFs with instantaneous frequency variance above some threshold. However, this does not set a useful bound on low frequency IMFs, and requires a data-specific threshold selection.
Following the model agnostic, data driven approach wherein model predictability serves as a metric of information content, Ye \& Sugihara suggested multiview embedding as an EDM algorithm to select maximally informative state-space variables \cite{Ye2016}. Multiview embedding evaluates model fidelity across all combinations of a set of candidate variables, selecting the $D$-dimensional subset with the highest predictive skill. In standard multiview embedding, the selection process is performed with in-sample simplex projection, the top ranked variables used in an out-of-sample prediction for the final model output. However, in-sample ranking poses a problem with IMFs since low-frequency IMFs approach, and eventually express monotonic functions.
Simplex projection is a nonlinear state-space mapping from state-space variables to a target without constraint on the variables. If the variables support a unique mapping across their domain, then good {\it in-sample} predictability can be achieved. This means that arbitrary, non-constant, non-oscillatory functions that provide a unique multivariable mapping can be mapped in-sample with good fidelity to the target. Since low frequency IMFs have very little or no oscillatory content, they are likely to be selected as in-sample variables with high predictive skill. Normally, candidate state-space variables are oscillatory or physically relevant, and this is not a concern. Here, we use all available IMFs, and address IMF selection by modifying standard multiview embedding to use {\it out-of-sample} predictability as the metric for ranking top predictors.
The modified multiview embedding is performed using the complete set of IMFs from variables $x$ and $y$ with the training library of points [1-2000], prediction set [2001-3000], and forecast interval $T_P = 0$. The embedding dimension is set to $E = 1$ so that the IMFs are not time-delay embedded, but used explicitly as state-space variables. The most predictive combination of $D$ IMFs are then selected to create a multivariable state-space.
Fig. \ref{fig:MultiviewNoiseIMF} plots simplex projection skill of the most predictive IMFs determined by multiview embedding for state-space dimensions of $D$ from 3 to 8, comparing them to the $\pm$ standard error envelopes of the 1000 noise ensembles of the reference model and state-spaces using all IMFs. At SNR immediately below 3 dB the predictive skill of EMM with state-space dimension of $D = 6$ has converged to match the upper envelope of the full IMF state-space ensembles. At SNR below 0 (noise power exceeds signal power) the $D = 6$ projection exceeds that of the full IMF state-space.
To summarise, an $E = 3$ dimensional time-delay embedding on variables $x$ and $y$ to simplex project $z$ from a state-space with dimension $D$ = 6 matches that of the fully informed reference model ($[x,y,z]$ projecting $z$) at zero noise. As noise increases the Takens state-space model degrades in predictive skill, and, both an indiscriminate use of all IMFs from noisy $x$ and $y$, as well as a selected subset of $D = 6$ IMFs outperforms the Takens model at SNR less than 3 dB. We infer that EMM can provide improved state-space representations in the presence of noise.
\begin{figure}
\centering
\includegraphics[width=3in]{graphics/MultiviewNoiseIMF_dB.png}
\caption{Simplex projection skill of the most predictive IMFs of one noise realisation as a function of $D$, the number of IMFs. Dashed lines are the $\pm$ standard error envelopes of the 1000 noise ensembles of the reference model and state-spaces using all IMFs.}
\label{fig:MultiviewNoiseIMF}
\end{figure}
\section{Salinity in Florida Bay}
Coastal South Florida is fringed by national parks including Biscayne and Everglades National Parks. Florida Bay is situated between the southern edge of the Florida peninsula and Florida Keys, is part of Everglades National Park, and, is an ecologically important and proliferent multispecies marine and estuarine nursery. Hydrologically, Florida Bay is interesting in that annual evaporation exceeds rainfall, resulting in widely varying coastal salinities as shown in Fig. \ref{fig:FLBay}. Hypersalinity is a seasonal occurrence, with episodic extremes surpassing 50 psu. These extreme hypersalinity events are associated with ecological collapse \cite{Hall2016, Johnson2018}. Clarification of influence variables and the ability to forecast salinities will directly inform ecosystem resource management.
\begin{figure}
\centering
\includegraphics[width=2.8in]{graphics/FLBay_Station_Stage_Salinity.png}
\caption{Top: Aerial photograph of the southern Everglades and Florida Bay with hydrographic monitoring stations. Bottom: Station water levels (stage), salinity, and station-averaged water level and salinity over 16 years.}
\label{fig:FLBay}
\end{figure}
Our objective is to forecast salinity in Garfield Bight (GB), assessing predictive performance of EDM simplex using three different state-space representations described below. The Florida Bay data consist of 6332 days of mean water level (stage), salinity, evaporation, and rainfall. The lower panel of Fig. \ref{fig:FLBay} illustrates a nonlinear inverse relationship between salinities and coastal marsh water levels signifying that water levels in the southern Everglades are a prime determinant of coastal salinities. Other causal variables are known to include evaporation, which is related to temperature and solar radiation, rainfall, water levels in the bay, and, previous salinity.
An initial step in EDM analysis is to examine the dimensionality and nonlinearity of the data \cite{Chang2017}, here, indicating that the data are nonlinear, and, a Takens time-delay embedding dimension of 5 provides good model predictability. Results of this data discovery are presented in appendix 1 (section \ref{appendix:FLBay}).
\subsection{State-space models}
\label{stateSpaceModels}
The first state-space model is a Takens $E\!=\!5$ time-delay embedding of GB salinity. Second, a multivariate state-space is constructed from the 5 variables GB salinity, stage, evaporation, rainfall, and, coastal land water level at station S16. These variables are known influencers of salinity as determined by the Florida Bay Assessment model \cite{Park2016}. The third state-space consists of selected IMFs of GB salinity, water level, evaporation, and coastal land water level at S16. IMFs are shown in appendix 1 (section \ref{appendix:FLBay}).
\subsubsection{Model comparison}
The comparative state-space analysis defines an in-sample data set as days [1-6330] for the training library, and days [5967-6275] as the prediction set. This prediction set corresponds to the period 2016-01-01 to 2016-11-04. The EMM model uses IMFs 5,6,7 of the salinity and stage variables, and IMFs 6,7,8 for S16 water level and evaporation (figure \ref{fig:GB_IMF}). These IMFs ignore high frequency, noise dominated components, as well interannual variations not deemed important on daily timescales.
Fig. \ref{fig:Predict_IMF_MV_Tp} plots a series of in-sample GB salinity projections for the three models over a sequence of prediction horizons $T_p$. It is worth reinforcing that these are not predictions \textit{per-se}, since the model is purely in-sample, rather, we seek to evaluate internal model representations of the dynamics. Here we see that while the $T_p\!=\!1$ projections are good for all three models, the EMM model misses high frequency variations evident in the January to March and September to October time frames. As prediction interval increases, these high frequency contributions seem to dominate the multivariate and time-delay embedding projections, significantly degrading accuracy, while the EMM model maintains a high fidelity internal representation of the dynamics.
\begin{figure}
\centering
\includegraphics[width=3.3in]{graphics/Simplex_Predict_IMF_MV_Tp.png}
\caption{Simplex in-sample projections of GB salinity for the year 2016 at different prediction intervals $T_p$ for the EMM, multivariable, and Takens time-delay state-space representations.}
\label{fig:Predict_IMF_MV_Tp}
\end{figure}
\subsection{Forecasting}
We now turn to the question of whether EMM can be useful in an out-of-sample regime which is requisite for an applied forecast system. Simplex forecasts are made with the three state-space models from a training library consisting of points [1-5475] expanded in 30 day increments starting at day 5475 (2014-08-27) until the end of the record. Predictions are made at forecast intervals of 1 to 56 days in 7 day increments, starting at one day past the end of the training library. This 30-day moving window provides 28 samples at each prediction interval. The mean RMSE over each prediction interval is shown in Fig. \ref{fig:Stat_MovingPred} indicating that the EMM IMF state-space provides generally more accurate predictions of salinity than the multivariate and time-delay state-space models.
\begin{figure}
\centering
\includegraphics[width=2.5in]{graphics/Stat_MovingPred.png}
\caption{Root mean square error (RMSE) of GB salinity out-of-sample simplex forecasts for the year 2016 at different prediction intervals $T_p$ for the EMM, multivariable, and time-delay state-space representations.}
\label{fig:Stat_MovingPred}
\end{figure}
\subsubsection{Comparison to numerical model}
As noted earlier, salinity in coastal Florida Bay is an ecologically important factor, linked to the onset of widespread ecological collapse \cite{Hall2016, Johnson2018}. In 2015, environmental conditions aligned to produce historic record maximum salinities in excess of 70 psu (seawater is nominally 35 psu) followed by widespread sea-grass mortality and food-web disruption. The ability to forecast such events, or to attribute causal variables, provides actionable information for natural resource managers. Accordingly, equation-based physical models have been developed to explore these issues, and, the Bay Assessment Model (BAM) is one of the latest and best performing salinity models \cite{Park2016}. Here, we compare BAM and EMM forecasts for the 2015 hypersalinity event.
EMM forecasts are out-of-sample based on progressive training library and prediction horizon times. The training library starts with days [1-5721]. Predictions are made for 121 days starting at day 5722 (2015-5-1) over a range of prediction horizons $T_p$ from 1 to 21 days. After each set of predictions, the training library end day and prediction start day are incremented by 1 day.
Fig. \ref{fig:GB_Forecast_2015} presents comparisons of the EMM forecasts with BAM results, where we note that BAM fails to capture the extreme hypersalinity event of 2015. EMM forecasts provide accurate salinity predictions at $T_p\!=\!1$ with good forecasting of extreme salinity. Forecasts remain reasonably good at $T_p\!=\!3$, having lost good fidelity at $T_p\!=\!7$, but still outperforming the numerical model. This suggests that the method of progressive EMM projections at $T_p\!=\!1$ are good candidates for an operational forecast system.
\begin{figure}[H]
\centering
\includegraphics[width=3in]{graphics/GB_Forecast_2015.png}
\caption{Salinity forecasts in Garfield Bight during the 2015 hypersalinity event. \:a) - d) EMM forecasts (red) with observed data (blue) and results from the Bay Assessment Model (green). \:e) EMM forecast error for different prediction horizons $T_p$. \:f) EMM RMS error over the prediction period May 1 - August 31, 2015 as a function of forecast interval $T_p$.}
\label{fig:GB_Forecast_2015}
\end{figure}
\section{Conclusion}
The 21$\mathrm{^{st}}$ Century revolution in data collection and analysis has fundamentally changed our ability to infer relationships and dynamical behavior of complex systems. A powerful corollary is the emergence of data-driven, model-free analytical techniques where presumptions and models are not fit to observations, but where the data itself is represented in a state-space facilitating the discovery of dynamical relationships enabling state forecasting.
Inevitably, observational data are intertwined with noise, with potential to disrupt accurate state-space representations. Such disruptions motivate us to suggest empirical mode modeling (EMM), a synthesis of empirical mode decomposition (EMD) and empirical dynamic modeling (EDM). As the names suggest, these are data-driven, model-free techniques ideally suited for nonlinear dynamical analysis. EMM uses EMD intrinsic mode functions (IMF) as state-space vectors in the EDM framework. Since IMFs naturally decompose time series into physically-relevant modes, they are well-suited for isolation and removal of noise.
EMM was assessed by application to a synthetic data set with controlled SNR, and, it was found that for high SNR state-spaces EMM as implemented here does not improve projection skill. However, as SNR falls below 3 dB EMM is found to improve model predictability and information content of the state-space, leading us to anticipate that noise-dominated signals can benefit from EMM.
Application to a geophysical data set composed of inherently noisy data found that EMM provided superior forecasting ability in comparison to time-delay or multivariate state-space representations. Further, the EMM representation outperformed the best available numerical model in predicting the 2015 hypersalinity event in Garfield Bight of Florida Bay, an extreme occurrence without precedent in the observational record.
Collectively, this work indicates that EMM can be a useful technique to improve state-space representations in the presence of noise with signal-to-noise ratios less than 3 dB.
Although this initial implementation and assessment of EMM seems promising, there are inherent limitations in the method, and many avenues for improvement. The primary limitation is that EMD is not applicable to non-oscillatory signals. Time series such as neuronal spike trains, rainfall, or other discrete, Poisson process dynamics will not be amenable to EMM based on EMD IMFs. This might be addressed by replacing the EMD IMF with a modal decomposition that captures non-oscillatory time-dependence, for example, a discrete wavelet transform.
Regarding improvements, investigating methods to optimally select IMFs to avoid noise, or, to maximize cross-variable information assessments are warranted. Here, we have relied on EDM multiview embedding. Second, EMD has been extended beyond the original decomposition algorithm used here. For example, ensemble empirical mode decomposition (EEMD) has been found to mitigate mode-mixing and improve physical significance of modes \cite{Wu2009}. EMD has also been extended into hybrid machine learning paradigms \cite{Chen2019, Jiang2019} and multivariate implementations \cite{Rehman2009}.
Another obvious exploration is to assess complementary mixed embeddings as a state-space representation. For example, combinations of Takens time-delay embedded vectors with IMFs, and, with multivariate observations. Yet another is to use other nonlinear modal decompositions such as the discrete wavelet transform either in-place of, or, as a complement to IMFs. Finally, the use of sequential locally weighted global linear maps (S-maps)\cite{Sugihara1994} may provide improved state-space representation and forecast skill.
\section*{Conflict of interest}
The authors declare that they have no conflict of interest.
\clearpage
\section{Appendix 1: Florida Bay salinity data discovery}
\label{appendix:FLBay}
Application of EDM to a dataset is predicated on assessing whether the underlying dynamics are state-dependent (nonlinear), and if so, estimating an embedding dimension to use if univariate data are to be time-delay embedded. A primer on EDM analysis can be found in \cite{Chang2017}.
The Florida Bay data consist of 6332 daily values (September 1 1999 though December 31 2016). We use days 1-5000 as the training (library) set, and days 5001-6300 as the prediction.
Fig. \ref{fig:GB_salt_E} shows simplex prediction skill as a function of Takens embedding dimension for Garfield Bight (GB) salinity. Predictions are made at a forecast interval of $T_p\!=\!1$ timestep ahead. This suggests that embedding dimensions in the range of 2-6 provide the best model fidelity, and that the underlying dynamics of the salinity are low-dimensional.
\begin{figure}[H]
\centering
\includegraphics[width=2.5in]{graphics/GB_salt_E.png}
\caption{Out-of-sample simplex model correlation for Florida Bay station GB salinity as a function of time-delay embedding dimension. Different curves are based on different library sizes, expressed as a percentage of the total data size. The maximum library index is 5000, predictions are over days 5001-6300.}
\label{fig:GB_salt_E}
\end{figure}
Fig. \ref{fig:GB_salt_Tp} presents simplex prediction skill as a function of forecast interval $T_p$ at different embedding dimensions $E$. The saturation of predictability as $E$ approaches 5 indicates that a dimension of 5 can be a reasonable choice for EDM analysis of this data. The decline in predictability as $T_p$ increases is consistent with dynamics of a nonlinear system.
\begin{figure
\centering
\includegraphics[width=2.5in]{graphics/GB_salt_Tp.png}
\caption{Out-of-sample simplex model correlation for Florida Bay station GB salinity as a function of forecast interval of $T_p$.}
\label{fig:GB_salt_Tp}
\end{figure}
Interpreting nonlinearity as state-dependence \cite{Sugihara1994}, the EDM s-map algorithm allows one to assess nonlinearity with examination of predictive skill as a function of state-space linear localization ($\theta$). Fig. \ref{fig:GB_salt_Theta} shows this assessment for the GB salinity data, indicating a weak, but robust state-dependence, again indicative that the salinity dynamics are nonlinear.
\begin{figure
\centering
\includegraphics[width=2.5in]{graphics/GB_salt_Theta.png}
\caption{Out-of-sample s-map model correlation for Florida Bay station GB salinity as a function of nonlinear localization parameter $\theta$.}
\label{fig:GB_salt_Theta}
\end{figure}
Fig. \ref{fig:GB_IMF} presents EMD IMFs of physical variables influencing salinity in Garfield Bight, Florida Bay. Low order, high frequency IMFs are considered noise dominated and not used as EDM state-space vectors. High order, low frequency IMFs are also not used as state-space variables.
\begin{figure
\centering
\includegraphics[width=0.9\textwidth]{graphics/GB_IMF.png}
\caption{EMD IMFs of variables influencing Garfield Bight salinity. \:a) Salinity. \:b) Water level. \:c) Coastal land water level. \:d) Evaporation.}
\label{fig:GB_IMF}
\end{figure}
\clearpage
|
\section{Introduction}
Shortly after the discovery of Bose-Einstein condensation in ultracold atoms of rubidium \cite{an95}, sodium \cite{ket95} and lithium \cite{hul95,hul97}, models and theories to calculate the time-dependent condensate fraction were developed, see Refs.\,\cite{gz97,gardiner_quantum_1997,bzs00} and related works. These approaches usually consider a quantum treatment such as a nonlinear Schr\"odinger equation for the condensate and an equilibrium- or nonequilibrium-statistical description for the thermal component.
Discrepancies between early theoretical results \cite{gz97} and data for the time-dependent condensate fraction \cite{miesner_bosonic_1998} were partly due to the fact that the noncondensed particles were represented by an equilibrium particle reservoir with a fixed chemical potential -- an assumption that does not correspond to the nonequilibrium initial state that is generated through evaporative cooling.
The relaxational dynamics of the nonthermal component that occurs in the presence of the mean field of the condensate was later taken into account through kinetic equations. In particular, a semiclassical Boltzmann equation \cite{eckern84,kd85,sto97,stoof_coherent_1999,zng99,bzs00} was solved numerically in the ergodic approximation, which has been used by many authors \cite{snowo89,kss92,setk95,lrw96,hwc97,jgz97}. In this approximation, the population of a given state depends only on its energy. Such an approach provides a detailed description of the growth of the condensate based on numerical solutions of the equations for the dynamics of the thermal cloud and its mean-field interaction with the condensate \cite{bzs00}, yielding satisfactory -- though not perfect -- agreement with the available data on the time-dependent condensate fraction.
These numerical approaches to thermalization and condensate formation usually rely on an artificial seed condensate \cite{bzs00}.
In view of such involved numerical approaches it is interesting to have a simple, exactly solvable
model for the physics of condensate formation in cold atoms. For bosons, such a model has been proposed in Ref.\,\cite{gw18} and adapted to cold quantum
gases in Refs.\,\cite{gw18a,gw20}. The model is based on a nonlinear boson diffusion equation (NBDE) which has been derived from the quantum Boltzmann collision term. For energy-independent
transport coefficients, it is structurally simple, but still complicated to solve exactly due to the nonlinearity in the drift term, which corresponds to Bose stimulation and causes the system to reach the Bose-Einstein equilibrium distribution for sufficiently large times, with a singularity in the infrared. The solutions represent, in particular, evaporative cooling from an initial temperature $T_\text{i}$ to a final temperature $T_\text{f}$ that may be below the critical value $T_\text{c}$. We explore the solutions in this work to calculate the time-dependent condensate fraction based on the analytic solutions and on particle-number conservation in the cloud and the condensate, together with the time evolution of the entropy in successive and single cooling steps.
In the next section, the NBDE and its exact solutions \cite{gw18a,gw20} with the consideration of boundary conditions at the singularity are briefly reviewed. In Section\,3, the analytic time-dependent solutions of the nonlinear equation with constant transport coefficients are evaluated and compared to numerical results. In Section\,4, the time evolution of the entropy in evaporative cooling is considered for several sequential cooling steps, as well as for a single step that leads to condensate formation. In Section\,5, the time-dependent chemical potential is evaluated from the condition of particle-number conservation and the condensate fraction is discussed based on the analytic NBDE solutions. Due to the quantum-statistical properties of the bosonic system that are encoded in the NBDE, a condensate forms and grows if the final temperature is below the critical value. With parameters adapted to the evaporative cooling of $^{23}$Na, a comparison of the time-dependent condensate fraction with MIT data is presented. The conclusions are drawn in Section\,6.
\section{Nonlinear boson diffusion equation and analytic solutions}
The system of ultracold atoms is viewed as a time-dependent mean-field supplemented by a collision term. The atomic density is usually so low that interactions involving more than two particles at a time can be ignored. Only $s$-wave scattering contributes for bosons at ultracold temperatures, such that the interaction is a contact interaction with a coupling strength that is proportional to the $s$-wave scattering length.
The $N$-body density operator $\hat{\rho}_N(t)$ can be written based on $N$ single-particle wave functions of the atoms which are solutions of the time-dependent Hartree-Fock equations supplemented by a time-irreversible collision term $K_N(t)$ that accounts for the thermalization through random two-body collisions ($\hbar=c=1$)
\EQ{}{
i\,\frac{ \partial\hat{\rho}_N(t)}{\partial t} = \big[\hat{H}_\text{HF}(t),\hat{\rho}_N(t)\big] + i \hat{K}_N(t)
\label{eq1}}{}
where $\hat{H}_\text{HF}(t)$ is the self-consistent Hartree-Fock mean field of the atoms, with an external potential for ultracold atoms in a trap.
In this work we do not consider the full many-body problem and its reduction to the one-body level, but rather an approximate version that starts from the reduced ensemble-averaged single-particle density operator $\bar{\rho}_1(t)$. Its diagonal elements can be interpreted as the
probability for a particle to be in a state $|\alpha\rangle$ with energy $\epsilon_\alpha$
\EQ{}
{\big(\bar{\rho}_1(t)\big)_{\alpha,\alpha} = \langle n(\epsilon_\alpha, t)\rangle \equiv n_\alpha(\epsilon,t)}
{.}
The total number of particles is $N=\sum_\alpha n_\alpha$, and we neglect here
the off-diagonal terms of the density matrix. The occupation-number distribution $n_\alpha(t)$ in a finite Bose system obeys a Boltzmann-like collision term \cite{gw18}
\begin{align}
\frac{\partial n_\alpha}{\partial t }= \sum_{\beta,\gamma,\delta} \overline{V^2_{\alpha\beta\gamma\delta}}\,\mathcal{G}_{\alpha\beta\gamma\delta} \big[ (1+n_\alpha)(1+n_\beta)n_\gamma n_\delta\\ \nonumber
-(1+n_\gamma)(1+n_\delta)n_\alpha n_\beta \big]
\end{align}
where $\overline{V^2_{\alpha\beta\gamma\delta}}$ is the second moment of
the pairwise atom-atom interaction. $\mathcal{G}_{\alpha\beta\gamma\delta}$ is an energy-conserving function which has a finite width, because it is the total energy of mean field plus collision term that must be conserved. It becomes a $\delta$-function as in the Boltzmann-Nordheim equation \cite{no28} only if the mean-field energy content stays constant in time -- which is, however, usually not the case in cold-atom systems. Whereas in an infinite homogeneous system that is described using a quantum Boltzmann
equation with an energy-conserving $\delta$-function an infinite amount of time is required to create a condensate, this is not expected to be the case for an energy-conserving function that has a finite width. Hence, the theoretical framework is consistent with a finite condensate formation time as required from experiment.
This equation has been transformed into a nonlinear partial differential equation in Refs.\,\cite{gw18,gw18a}. One starts by defining transition probabilities from state $\gamma$ to $\alpha$ as
\EQ{}{
W_{\gamma \rightarrow \alpha} = \sum_{\beta,\delta} \overline{V^2_{\alpha\beta\gamma\delta}} \,\mathcal{G}_{\alpha\beta\gamma\delta} (1+n_\beta)\,n_\delta
}{.}
The collision term can be written as a
master equation where the transitions into and out of the state with energy $\epsilon_\alpha$ are more explicit
\EQ{}
{
\frac{\partial n_\alpha}{\partial t }= (1+n_\alpha) \sum_\gamma W_{\gamma \rightarrow \alpha} n_\gamma
- n_\alpha \sum_\gamma W_{\alpha \rightarrow \gamma} (1+n_\gamma)
}{.}
Introducing the density of states $g_\alpha = g(\epsilon_\alpha)$, such that $W_{\gamma \rightarrow \alpha} = W_{\gamma,\alpha} g_\alpha$
and using the fact that quantum particles are indistinguishable
$W_{\alpha,\gamma} = W_{\gamma,\alpha} = W(\epsilon_\alpha, \epsilon_\gamma)$ one arrives at
\EQ{}
{
\frac{\partial n_\alpha}{\partial t }= \int_0^\infty W_{\alpha,\gamma} \left[ g_\alpha (1+n_\alpha) n_\gamma - g_\gamma (1+n_\gamma) n_\alpha \right] \text{d} \epsilon_\gamma
}{.}
Defining the transport coefficients $v\equiv v\,(\epsilon_\alpha,t)$ and $D\equiv D\,(\epsilon_\alpha,t)$ as first and second moments of the transition probability \cite{gw18},
we have obtained the nonlinear boson diffusion equation (NBDE) for the single-particle expectation-value occupation-number distribution of the energy eigenstates $\epsilon_\alpha$, $n\equiv n_\alpha \equiv \langle n(\epsilon_\alpha,t)\rangle$, in Refs.\,\cite{gw18,gw18a} as \footnote{The derivative-term of the diffusion coefficient has been modified as compared to Refs.\,\cite{gw18,gw18a,gw20} in order to secure the correct stationary solution}
\begin{equation}
\frac{\partial n}{\partial t}=-\frac{\partial}{\partial\epsilon}\Bigl[v\,n\,(1+n)+n\frac{\partial D}{\partial \epsilon}\Bigr]+\frac{\partial^2}{\partial\epsilon^2}\bigl[D\,n\bigr]\,.
\label{boseq}
\end{equation}
The drift term $v(\epsilon,t)$ accounts for dissipative effects, the term $D(\epsilon,t)$ for diffusion of particles in the energy space.
The many-body physics is contained in these
transport coefficients, which depend on energy, time, and the second moment of the interaction.
A prerequisite for the reduction to $1+1$ dimensions in the above formulation is spatial and momentum isotropy. This corresponds to the assumption of sufficient ergodicity, which has been widely discussed in the literature \cite{snowo89,kss92,setk95,lrw96,hwc97,jgz97}. For the thermal cloud of cold atoms around a Bose-Einstein condensate (BEC), it is expected to be a reasonable assumption, even though the condensate in a trap is spatially anisotropic. Regarding the role of different spatial dimensions in view of BEC formation, this enters our present formulation only through the density of states, which differs according to the number of spatial dimensions, and the type of confinement. The model calculations in this work are for a 3d system, and we shall investigate results for the density of states of a free Bose gas, and bosonic atoms confined in a harmonic trap. One-dimensional systems where no BEC should be formed have not yet been studied.
To derive the stationary solution $n_\infty(\epsilon)$ for the equation with
variable transport coefficients, we rewrite Eq.\,(\ref{eq1}) and set the time derivative
to zero
\begin{equation}
0= \frac{\partial}{\partial \epsilon}{\left[ v\,n_\infty \,(1+n_\infty)- D\,\frac{\partial n_\infty}{\partial \epsilon} \right]}\,,
\end{equation}
such that
\EQ{}
{v\,n_\infty(1+n_\infty) - D\,\frac{\partial n_\infty}{\partial \epsilon} = c_1}{.}
Dividing by $n_\infty(1+n_\infty)D$ and integrating over $\epsilon$ yields
\EQ{}
{
\int \frac{\text{d} n_\infty/\text{d} \epsilon}{n_\infty(1+n_\infty)}\text{d}\epsilon =
\int \left( \frac{v}{D} - \frac{c_1}{n_\infty(1+n_\infty)D} \right)\text{d} \epsilon
}{.}
We integrate the l.h.s, resulting
in $[\ln(n_\infty) - \ln(1+n_\infty)] = \ln(1-\frac{1}{1+n_\infty})$ plus an integration constant $c_2$.
Solving for $n_\infty$ one obtains
\EQ{}
{n_\infty = \left[
\exp\left(
\int \left(- \frac{v}{D} + \frac{c_1}{n_\infty(1+n_\infty)D}\right)\text{d} \epsilon +c_2 \right) -1
\right]^{-1}.
}{}
In order to reduce to a Bose-Einstein distribution, $c_1 = 0$ is required, and the ratio $v/D$ must have no energy dependence for ${t\rightarrow\infty}$ so that it can
be pulled out of the integral.
It follows that $\lim \limits_{t \to \infty}[-v(\epsilon,t)/D(\epsilon,t)] \equiv 1/T$ and $c_2\equiv -\mu/T$ such that the stationary distribution equals the Bose-Einstein equilibrium distribution ($k_\text{B}=1$)
\begin{equation}
n_\infty(\epsilon)=n_\text{eq}(\epsilon)=\frac{1}{e^{(\epsilon-\mu)/T}-1}
\label{Bose-Einstein}
\end{equation}
with the chemical potential $\mu<0$ in a finite Bose system.
In the limit of energy-independent transport coefficients the nonlinear boson diffusion equation
for the occupation-number expectation-value distribution $n(\epsilon,t)$
becomes
\begin{equation}
\frac{\partial n}{\partial t}=-v\,\frac{\partial}{\partial\epsilon}\Bigl[n\,(1+n)\Bigr]+D\,\frac{\partial^2n}{\partial\epsilon^2}\,.
\label{bose}
\end{equation}
As in case of Eq.\,(\ref{boseq}), the thermal equilibrium distribution $n_\text{eq}$ is a stationary solution
with $\mu<0$ and $T=-D/v$.
In spite of its simple structure, the NBDE with constant transport coefficients thus
preserves the essential features of Bose-Einstein
statistics which are contained in the bosonic Boltzmann equation. For a given initial condition $n_\text{i}(\epsilon)$, it can be solved exactly using the nonlinear transformation outlined
in Ref.\,\cite{gw18a}. The resulting solution is
\begin{align}
n(\epsilon,t) = -\frac{D}{v} \frac{\partial}{\partial\epsilon}\ln{\mathcal{Z}(\epsilon,t)} -\frac{1}{2}= -\frac{D}{v}\frac{1}{\mathcal{Z}} \frac{\partial\mathcal{Z}}{\partial\epsilon} -\frac{1}{2}
\label{eq:Nformula}
\end{align}
where the time-dependent partition function ${\mathcal{Z}(\epsilon,t)}$ obeys a linear diffusion equation
\begin{align}
\frac{\partial}{\partial t}{\mathcal{Z}}(\epsilon,t) = D \frac{\partial^2}{\partial\epsilon^2}{\mathcal{Z}}(\epsilon,t)\,.
\label{eq:diffusionequation}
\end{align}
The partition function $\mathcal{Z}(\epsilon,t)$
\begin{align}
\mathcal{Z}(\epsilon,t)= a(t)\int_{-\infty}^{+\infty} G(\epsilon,x,t)\,F(x)\,\text{d}x
\label{eq:partitionfunctionZ}
\end{align}
is an integral over Green's function $G(\epsilon,x,t)/\sqrt{4\pi\,Dt}$ of Eq.\,(\ref{eq:diffusionequation}), and an exponential function $F(x)$ that depends on
the initial occupation-number distribution $n_\text{i}$
\begin{align}
F(x) = \exp\Bigl[ -\frac{1}{2D}\bigl( v x+2v \int_0^x n_{\mathrm{i}}(y)\,\text{d}y \bigr) \Bigr]\,.
\label{ini}
\end{align}
The time-dependent prefactor $a(t)=1/\sqrt{4\pi\,Dt}$ in Eq.\,(\ref{eq:partitionfunctionZ}) cancels out when taking the logarithmic derivative in Eq.\,(\ref{eq:Nformula}), and for the same reason the
definite integral over the initial conditions taken at the lower limit in Eq.\,(\ref{ini}) drops out in the calculation of $n(\epsilon,t)$. Hence, the definite integral in Eq.\,(\ref{ini}) can be replaced \cite{rgw20} by the indefinite integral $A_{\mathrm{i}}(x)$ over the initial distribution with $\partial_x A_{\mathrm{i}} (x) = n_{\mathrm{i}}(y)$.
With these modifications that do not affect the accuracy, it becomes possible to compute the partition function and the overall solution for the occupation-number distribution function Eq.\,(\ref{eq:Nformula}) analytically.
When solving the problem with the free Green's function of Eq.\,(\ref{eq:diffusionequation}), the physically correct solution with the Bose-Einstein equilibrium limit is attained in the UV region, but not in the IR \cite{gw18}. To solve this problem, one has to consider the boundary conditions at the singularity $\epsilon = \mu < 0$ \cite{gw20}.
For fixed chemical potential $\mu$, an initial temperature $T_\text{i}$, and a final temperature $T_\text{f} = - D/v$, the combined initial- and boundary value problem has been solved exactly in Refs.\,\cite{rgw20,gw20} using the above nonlinear transformation from Eq.\,(\ref{eq:Nformula}) \cite{gw18,gw20}, and an infinite series expansion. The analytic solutions are briefly reconsidered before we use them for further calculations and in particular, to obtain the time-dependent condensate fraction.
With the boundary condition \(\lim_{\epsilon \downarrow \mu} n(\epsilon,t) = \infty\) \,$\forall$ \(t\) at the singularity $\epsilon = \mu$ one obtains \( \mathcal{Z} (\mu,t) = 0\), and the energy range is restricted to $\epsilon \ge \mu$. This requires a new Green's function \cite{gw20}
that equals zero at \(\epsilon = \mu\) $\forall \,t$. It can be written as
\begin{align}
{G} (\epsilon,x,t) = G_\text{free}(\epsilon - \mu,x,t) - G_\text{free}(\epsilon - \mu,-x,t)\,,
\label{eq:newGreens}
\end{align}
and the partition function with this boundary condition becomes
\begin{align}
{\mathcal{Z}} (\epsilon,t) = \int_0^\infty {G} (\epsilon, x, t)\,F(x+\mu)\, \text{d}x\,.
\label{eq:newformulaforZ}
\end{align}
The function $F$ remains unaltered with respect to Eq.\,(\ref{ini}), but its argument is shifted by the chemical potential.
\begin{figure}[b!]
\centering
\includegraphics[width = 0.47\textwidth]{fig1.eps}
\caption{The argument $\ln{F(x)}$ of the exponential function $F(x)$ in Eq.\,(\ref{ini}) (solid curve) with a singularity at \(x = \mu<0\) in the definite integral of an initial distribution $n_\text{i}$ given by a Bose-Einstein distribution with $T=240$\,nK and $\mu=-8$\,nK that is truncated at $30$ nK (dotted vertical line).
}
\label{fig1}
\end{figure}
As an initial condition that is appropriate for a schematic description of evaporative cooling {\cite{an95,dmk95,lrw96}} to demonstrate our method, a truncated thermal equilibrium distribution that is cut off at a maximum energy $\epsilon_\text{i}$ beyond which high-velocity atoms are removed has been chosen in our previous works
\begin{equation}
n_\text{i}(\epsilon)=\frac{1}{e^{(\epsilon-\mu)/T}-1}\, \theta (1-\epsilon/\epsilon_\text{i})\,.
\label{inibec}
\end{equation}
The distribution is truncated at a maximum energy $\epsilon_\text{i}$ beyond which high-velocity atoms are removed. In a trapped ultracold gas, these atoms leave the trap: The system of atoms in the trap is thus not isolated and its entropy decreases in the course of cooling. In the subsequent re-thermalization (and for $T_\text{f}<T_\text{c}$, condensate formation), the entropy of the atoms in the trap rises again, but remains below the initial thermal value. This will be discussed in more detail in Section\,5.
For such a truncated initial thermal distribution, the
logarithm of the function $F(x)$ with a singularity at {$x=\mu$}\\ (before shifting it by $\mu$) is shown in Fig.\,\ref{fig1} with the following set of parameters: $T_\text{i} = 240$ nK,
$\epsilon_\text{i} = 30$ nK, $D = 100\,$(nK)$^2$\,ms$^{-1}$, $v = -1\,$nK\,ms$^{-1}$, $T_\text{f} = - D/v = 100$ nK.
Due to the singularity, $F(x)$ vanishes at $x = \mu=-8$\,nK. The logarithm of $F(x)$ is continuous, but not differentiable at $x = \epsilon_\text{i}$, which holds the key to the equilibration in the UV region.
Using Eq.\,(\ref{inibec}) as initial distribution, the occupation-number distribution
can be evaluated exactly for $T_\text{i} \ne T_\text{f}$ with the time-dependent partition function \cite{rgw20}
\begin{align}
{\mathcal{Z}}(\epsilon,t) = \sqrt{4 D t} \, \exp\Bigl(-\frac{\mu}{2 T_{\mathrm{f}}}\Bigr) \sum_{k=0}^{\infty} \binom{\frac{T_{\mathrm{i}}}{T_{\mathrm{f}}}}{k} \left( -1 \right)^k \times \notag \\
\Bigg( \text{e}^{\alpha_k^2 D t} \left[ \text{e}^{\alpha_k (\epsilon - \mu)} \Lambda_1^k (\epsilon,t) - \text{e}^{\alpha_k (\mu - \epsilon)} \Lambda_2^k (\epsilon, t) \right] \notag \\
+ \exp\Bigl( \frac{(\mu - \epsilon_i)k}{T_{\mathrm{i}}}\Bigr) \exp\Bigl({\frac{D t}{4 T_{\mathrm{f}}^2}}\Bigr) \times \notag \\
\Big[ \exp\Bigl(\frac{\epsilon-\mu}{2 T_{\mathrm{f}}}\Bigr) \Lambda_3 (\epsilon,t) - \exp\Bigl(\frac{\mu - \epsilon}{2 T_{\mathrm{f}}}\Bigr) \Lambda_4 (\epsilon,t) \Big] \Bigg)\, \label{eq:Zarbtemp}
\end{align}
where $\alpha_k=1/T_\text{f}-k/T_\text{i}$ and the auxiliary functions are
\begin{align}
\Lambda_1^{k} (\epsilon,t) =& \,\text{erf}\,\Bigl(\frac{\epsilon - \mu +2 D t \alpha_k}{\sqrt{4 D t}}\Bigr) \notag \\
&\qquad - \,\text{erf}\,\Bigl(\frac{\epsilon - \epsilon_i + 2 D t \alpha_k}{\sqrt{4 D t}}\Bigr)\,,\\
\Lambda_2^{k} (\epsilon,t) =& \,\text{erf}\,\Bigl(\frac{\mu-\epsilon+ 2 D t \alpha_k}{\sqrt{4 D t}}\Bigr) \notag \\
&\qquad - \,\text{erf}\,\Bigl(\frac{2 \mu - \epsilon - \epsilon_i + 2 D t \alpha_k}{\sqrt{4 D t}}\Bigr)\,,\\
\Lambda_3 ( \epsilon,t) =& \,\text{erfc}\,\Bigl(\frac{\epsilon_i - \epsilon + t v }{\sqrt{4 D t}}\Bigr)\,, \\
\Lambda_4 ( \epsilon,t) =& \,\text{erfc}\,\Bigl(\frac{\epsilon - 2 \mu + \epsilon_i + t v}{\sqrt{4 D t}}\Bigr)\,.
\end{align}
\begin{figure}[t!]
\centering
\includegraphics[width = 0.47\textwidth]{fig2}
\caption{\label{fig2} Schematic representation of evaporative cooling in a a finite Bose system based on the nonlinear evolution according to Eq.\,(\ref{bose})
starting from a truncated Bose-Einstein distribution Eq.\,(\ref{inibec}), upper solid curve with cutoff at $\epsilon_\text{i}$\,=\,30\,nK and $\mu=-8$\,nK.
The transport coefficients are $D = 100$\,(nK)$^2$\,ms$^{-1},~\,v = -1$\,nK$\,$ms$^{-1}$. The initial temperature is $T = 240$\,nK, the final temperature
$ T_\text{f} = T = -D/v = 100$\,nK.
The time evolution of the single-particle occupation-number distributions is shown at $t = 0.001, 0.01, 0.1, 0.4$, and $0.8$\, ms (top to bottom in the IR). For comparison, the numerical results using Matlab are shown as open circles.
}
\end{figure}
To obtain the time-dependent occupation-number distribution function from Eq.\,(\ref{eq:Nformula}), the derivative $\partial {\mathcal{Z}}/\partial\epsilon$ is also required. It can be calculated analytically \cite{rgw20}, such that an exact expression for the time-dependent occupation-number distribution function for evaporative cooling results that can be directly compared with numerical solutions of the NBDE, see the following section
\section{Occupation-number distributions}
The time-dependent analytic distribution functions that solve the NBDE exactly with boundary conditions at the singularity are displayed in Fig.\;\ref{fig2} for the same initial conditions as in Fig.\;\ref{fig1}. The analytic results agree precisely with numerical solutions (open circles) of the basic equation, {and predict the time-dependent cooling from a thermal distribution with temperature $T_\text{i}$ that is truncated at $\epsilon_\text{i}$ to a Bose-Einstein distribution with $T_\text{f} < T_\text{i}$, which is the thermal distribution for $t\rightarrow \infty$. The approach is similar to the kinetic theory of evaporative cooling in works such as Ref.\,\cite{lrw96}, but now an analytically solvable model is formulated.}
The corresponding numerical solutions (circles in Fig.\;\ref{fig2}) have been obtained using Matlab's routine \textit{pdepe}\rm\, for the solution of partial differential equations with given initial and boundary values \cite{skeel1990method}. Basic finite-difference algorithms like the Crank-Nicolson method did not provide sufficient accuracy for this nonlinear problem.
The implementation of \textit{pdepe} is based on the algorithm described in Ref.~\cite{skeel1990method} and is suited for (nonlinear) parabolic
partial differential equations.
Due to the singularity at $\epsilon=\mu$, the integration was started at $\mu+\delta<0$. Regarding the associate boundary condition, we use the known stationary solution $n_\infty(\epsilon)$ to determine the boundary values in the IR as $n_\infty(\mu+\delta)$, and accordingly, in the UV. This minimizes the numerical errors at the boundaries. The value of $\delta$ is chosen as small as possible, but large enough to prevent numerical inaccuracies. In this particular case $\delta=0.2$ and an upper boundary of $\epsilon_\text{u}=3 \times T_\text{i}$ yields satisfactory agreement with the exact solutions.
{The values of the transport coefficients $v, D$ in this specific model calculation have been derived from their relations to the equilibrium temperature $T=-D/v$ and the equilibration time $\tau_\text{eq}=4D/(9v^2)$ \cite{gw18}, with $T=100$ nK and $\tau_\text{eq}=44$ ms. They do not yet correspond to a specific experimental situation, and the cut-off temperature
$\epsilon_\text{i}=30$\,nK is articifially low to better demonstrate the behaviour of the analytic solutions. For direct comparisons with data, these values shall be adapted to the corresponding experiment, see Section\,5.}
In the infrared, thermalization occurs faster \cite{gw20} than in case of a linear relaxation ansatz: With the parameter set of Fig.\;\ref{fig2}, the thermal distribution is reached within $t\simeq 1$ ms in the IR. The buildup of the thermal slope in the UV is, however, slower in the nonlinear model as compared to the linear relaxation ansatz, because the latter enforces {an exponential approach to {the} Boltzmann-like tail even for $\epsilon\rightarrow\infty$ \cite{gw20}.
\begin{figure}[t!]
\centering
\includegraphics[width = 0.47\textwidth]{fig3}
\caption{\label{fig3} Convergence properties of the analytic NBDE solution with constant transport coefficients according to Eqs.\,(\ref{eq:Nformula}) and (\ref{eq:Zarbtemp}).
The initial condition is as in Fig.\,\ref{fig2}, upper solid curve with cutoff at $\epsilon_\text{i}$\,=\,30\,nK.
The time-dependent nonequilibrium solution is shown for $t = 0.4$\,ms, lower solid curve. The convergence is displayed for maximum expansion parameters
$k_\text{max}=10, 12, 14, 20$, and $200$ (from bottom to top) in the series expansion of the partition function.
}
\end{figure}
The convergence of our analytical method is displayed in Fig.\;\ref{fig3} using the above NBDE solution from Fig.\;\ref{fig2} for $t=0.4$\,ms as an example. Results according to Eq.\,(\ref{eq:Nformula}) are shown for $k_\text{max}=10, 12, 14, 20$, and $200$ (from bottom to top) in the series expansion of the partition function
Eq.\,(\ref{eq:Zarbtemp}). Here, $k_\text{max}$ is the maximum expansion coefficient that is used in the calculation. A solution for $k_\text{max}=40$ is already indistinguishable from the plotted one for $k_\text{max}=200$. It is thus sufficiently close to the exact solution, which would require $k_\text{max}=\infty$. We shall later check the convergence properties also for derived quantities such as the time-dependent chemical potential $\mu(t)$ that is needed to maintain particle-number conservation in the course of the time evolution, and for the condensate fraction, in Section\, 5.
\section{Evaporative cooling and time-dependent entropy}
To achieve the phase transition to the condensate, successive evaporative cooling is used, thereby removing high-velocity atoms. In the course of the subsequent equilibration, the number of condensed particles rises due to a transfer from the nonlinear kinetic region into the coherent region \cite{svi91,kss92,kas97} and an isotropic tail \cite{an95,BookPitaevskii} develops in the ultraviolet, smearing out the sharp cut that corresponds to evaporative cooling.
Solutions of the nonlinear boson diffusion equation fulfil the physically reasonable condition \cite{gw20} that the entropy drops in the course of an evaporative cooling step and then increases with time towards the equilibrium value that is determined by the final Bose-Einstein distribution, such that there is an interplay of cooling and re-thermalization.
To compute the entropy, we first consider a bosonic system in statistical equilibrium that will be the limit of the time-dependent case for $t\rightarrow\infty$,
\begin{align}
\label{eq:gen_ent}
S_\text{eq} = \int g(\epsilon)\Bigl [ \bigl (1+n_\text{eq}(\epsilon)\bigr) \ln \bigl(1+n_\text{eq}(\epsilon)\bigr)\\ \nonumber
- n_\text{eq}(\epsilon) \ln n_\text{eq}(\epsilon)\Bigr]\,\text{d}\epsilon\,.
\end{align}
Here the density of states $g(\epsilon)$ for a three-dimensional isotropic Bose gas without external potential is given by
\begin{align}
g(\epsilon) = g_0 \, \sqrt{\epsilon}
\end{align}
with
\begin{align}
g_0=(2m)^{3/2}\,V/(4\pi^2)\,,
\end{align}
as obtained from the substitution of a summation over the quantum numbers of the associated states with an energy integration \cite{BookPitaevskii}. For a harmonic oscillator potential the dependence is $g(\epsilon) = g^\text{HO}_0 \epsilon^2$.
In order to check the numerical integration in Eq.\,(\ref{eq:gen_ent}) against an exact result, we derive $S_\text{eq}^{\mu\rightarrow 0}$ for a free three-dimensional Bose gas as
\begin{align}
\label{eq:exact_s}
S_{\text{eq}}^{\mu\rightarrow 0}/g_0 = T^{3/2}\Biggl( \zeta\Bigl(\frac{5}{2}\Bigr) \Gamma\Bigl(\frac{5}{2}\Bigr)\qquad\qquad\qquad\\ \nonumber
+ \frac{\sqrt{\pi}}{2} \sum_{k=1}^{\infty}
\frac{1}{k} \Bigl[ \zeta(3/2,k) - \zeta(3/2,k+1)\Bigr]\Biggr)
\end{align}
with the Hurwitz zeta function $\zeta(s,q)$, yielding $S_{\text{eq}}^{\mu\rightarrow 0}/(g_0\,T^{3/2}) = 2.9721553$ as compared to the result $2.97216$ of a numerical integration from zero to infinity.
If the final temperature $T_\text{f}$ is below the initial one as is the case for cooling, the entropy drops below the initial equilibrium value and then rises again towards the final value in the course of the equilibration process. For $T_\text{f}$ below critical value for condensate formation $T_\text{c}$, particles also occupy the condensed state, but the total entropy still equals the entropy of the atoms in the thermal cloud. As has been emphasized in Ref.\,\cite{scul18}, this is the case even though the entropy of the particles in the ground state is nonzero, because the latter is cancelled by the so-called correlation entropy due to the fixed number of particles distributed among the quantum states.
It is therefore sufficient for a calculation of the total time-dependent entropy $S(t)$ to consider only the thermal cloud. The analytical solutions based on Eqs.\,(\ref{eq:Nformula}) and (\ref{eq:Zarbtemp}) for constant chemical potential $\mu$ and boundary conditions at the singularity $\epsilon=\mu$ are used. The entropy in a bosonic system
for an average number of particles $n(\epsilon,t)$ per single-particle state can be reformulated \cite{yam86,gw20} as
\begin{eqnarray}
S(t)=\int_0^\infty g(\epsilon)\Bigl[\ln\bigl(1+n(\epsilon,t)\bigr)\qquad\qquad\notag\\
+n(\epsilon,t)\ln\bigl(1+1/n(\epsilon,t)\bigr)\Bigr]\text{d}\epsilon\,.
\label{entropy}
\end{eqnarray}
Again, the spatial dimensionality and the external confinement enter the present formulation only through the density of states. The properties of the trapping potential and its effect on the density of states in evaporative cooling have been previously discussed in Ref.\,\cite{dmk95}}.
\begin{figure}
\centering
\includegraphics[scale=0.86]{fig4
\caption{\label{fig4} Time evolution of the entropy $S(t)/g_0$ in an equilibrating Bose gas for $k=0-6$ evaporative cooling steps starting with $T_\text{i}=1000$ nK and $T_k = 1000 \times (0.8)^k$ nK as calculated
from the analytical solution of the NBDE Eq.\,(\ref{bose}) with constant transport coefficients and constant chemical potential $\mu = - 8$ nK, solid curves. In each step, $S(t)$ drops instantaneously due to cooling and then rises in the course of re-thermalization. } \end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.56]{fig5
\caption{\label{fig5} Time evolution of the entropy $S(t)/g_0$ in an equilibrating Bose gas in the course of single-step evaporative cooling from $T_\text{i}=240$ nK with a cut at $\epsilon_\text{i}=80$\,nK to $T_\text{f}=100$ nK for $\mu=-8$\,nK as calculated
from the analytical solution of the NBDE Eq.\,(\ref{bose}), solid curve. {$S(t=0)$ is the entropy following evaporative cooling}. The dashed curve is the wave entropy, the dotted curve the particle entropy, and the dot-dashed curve the result from the linear relaxation ansatz \cite{gw20}. The dotted horizontal line indicates the equilibrium value {at $T_\text{f}=100$ nK} calculated from Eq.\,(\ref{eq:gen_ent}).}
\end{figure}
Results for successive cooling as displayed in Fig.\;\ref{fig4} show that the entropy decreases in each cooling step and subsequently rises due to re-thermalization. Correspondingly, for $T_\text{f}<T_\text{c}$ the number of condensed particles rises due to a transfer from the nonlinear kinetic region into the condensate \cite{svi91,kas97}, as will be discussed in the next section. In the example shown here
with $T_\text{i}\simeq 1000$ nK and $T_k = 1000 \times (0.8)^k$ nK, $\epsilon_{\text{cut},k} = 1000 \times (0.7)^{k+1}$ nK, $k = 0\ldots6$.
The entropy of the initial thermal distribution -- without the cut -- is about four times larger than the one of the final distribution.
The first term {in the entropy Eq.\,(\ref{entropy})} is referred to as wave entropy, it yields the largest contribution when the single-particle state is occupied by many particles, as in the IR. The second term is the particle entropy, which is more relevant in case of low occupation $n(\epsilon,t)<1$, as in the UV \cite{yam86,gw20}.
Results for the time-dependent contributions to the entropy are shown in Fig.\,\ref{fig5}, calculated with the analytical solutions of the NBDE. Here, the equilibrium value of the entropy at the initial temperature $T_\text{i}\simeq 240$ nK and for $\mu=-8$ \,nK is $S_\text{eq}/g_0\,(T_\text{i})\simeq10666$ (nK)$^{3/2}$. With a sharp cutoff at $\epsilon_\text{i}=80$\,nK
to account for evaporative cooling, the entropy of the initial (cooled) nonequilibrium distribution is reduced to $S_\text{i}/g_0\,(T_\text{i})\simeq 1218$\, (nK)$^{3/2}$.
Thermalization during the time evolution then occurs through the analytical solutions of the NBDE, and the result for the rising entropy $S(t)/g_0$ is shown in the solid curve
in Fig.\,\ref{fig5}.
The new equilibrium value of the entropy at the final temperature $T_\text{f}\simeq 100$ nK following evaporative cooling and thermalization is $S_\text{eq}/g_0\,(T_\text{f})\simeq 2740$ \,(nK)$^{3/2}$, dotted horizontal line.
The dashed curve is the wave entropy. It is most relevant for large occupation numbers, which are present at all times in the IR region $\epsilon<\epsilon_\text{i}$, and therefore, this contribution shows a rather weak time dependence. The dotted curve is the particle entropy, which is initially smaller than the wave entropy, because the occupation in the UV beyond the cut
is negligible at small times. In the course of thermalization, however, it exceeds the wave entropy at $t\simeq 10$ ms for the parameter set used in this calculation, and rises subsequently. Hence, the entropy at large times is mostly determined by the slowly rising contribution of the thermal tail.
In contrast, results for the corresponding linear relaxation-time approximation \cite{gw20} (dot-dashed curve in Fig.\,\ref{fig5}) reach the thermal limit that is given by
Eq.\,(\ref{eq:gen_ent}) much faster because an exponential buildup of the thermal tail towards $n_\text{eq}$ is enforced.
The entropy approaches the same equilibrium value, but at shorter times than the nonlinear solution.
It is noted that the final state does not change if both transport coefficients are scaled by the same amount because $T_\text{f}=-D/v$, but the time scale varies since $\tau_\text{eq}=4D/(9v^2)$. By measuring the time it takes for a given system to thermalize we can thus determine the absolute magnitude of the transport coefficients. To obtain the time scale from first principles, a microscopic calculation of the transport coefficients from a many-body theory is required, which is beyond the scope of our present approach.
Up to now, the results for the time-dependent occupation-number distributions and the entropy have been discussed for constant chemical potential. When computing the condensate fraction, however, particle-number conservation must be considered and $\mu$ becomes time-dependent, see the following section.
\begin{figure}
\centering
\includegraphics[scale=0.9]{fig6
\caption{\label{fig6} Chemical potential $\mu(t)$ in the course of evaporative cooling from $T_\text{i}=240$ nK to $T_\text{f}=100$ nK as calculated
based on particle-number conservation from the analytical solution of the NBDE Eq.\,(\ref{bose}) with $k_\text{max}\equiv K= 5, 10, 20, 200$ in the series expansion of the exact solution with $\epsilon_\text{i}=100$\,nK and $\mu_\text{i}=-8$\,nK.}
\end{figure}
\section{Time-dependent condensate fraction}
{We proceed to calculate the rate of increase of atoms in the condensate $N_\text{c}(t)$ for conserved total particle number $N = N_\text{c}(t)+N_\text{th}(t)$ and $T_\text{f}<T_\text{c}$
based on the exact analytic NBDE-solutions Eq.\,(\ref{eq:Nformula}) for any given $\mu$. Particle-number conservation -- which is a necessary condition for condensate formation to occur -- requires a time-dependent chemical potential, with $\mu(t)\rightarrow0$ corresponding to the condensed state.
To compute $\mu(t)$ in accordance with overall particle-number conservation, we
note that the particle number of the thermal cloud is given by
\EQ{eq:n_thermal}
{N_\text{th}(t) = \int_0^{\infty} g(\epsilon)\, \tilde{n}(\epsilon,t)\, \text{d}\epsilon
}{,}
where $\tilde{n}(\epsilon,t)$ is the distribution function with the time-dependent value of $\mu\equiv \mu(t)$. For each time step,
$\mu(t)$ is calculated such that the particle number is conserved, $N = N_\text{th}$, until
$\mu$ reaches the value of zero and the largest possible value of $N_\text{th}$ is attained. This instant marks
the moment of condensation and any future difference between the initial particle number and the thermal cloud particle number
is accounted for by the condensate
\EQ{}
{N - N_\text{th}(t) \equiv N_c(t)\,, \quad \text{for } \mu=0
}{.}
A typical time evolution of $\mu(t)$ is displayed in Fig.\,\ref{fig6}, for different values of $k_\text{max}\equiv K$ to show the convergence. Here the initial chemical potential is
$\mu_\text{i}=-8$ nK, the final one $\mu_\text{f}=0$. With increasing $k_\text{max}$ the accuracy increases rapidly, precise results are obtained already for $k_\text{max}=20$. For the parameters used in Fig.\,\ref{fig6}, condensation sets in at around $400$ ms.
As is obvious from Eq.\,\eqref{eq:n_thermal}, $\mu(t)$ and hence, the onset of condensation, also depends on the density of states.
Due to the statistical properties of the bosonic system that are encoded in the NBDE, no condensate forms if the final temperature remains above the critical value. We have confirmed this in calculations for different final temperatures $T_\text{f}$, with the condensate fraction $N_\text{c}(t)/N=0\, \forall \, t \land T_\text{f}>T_\text{c}$.
It is also self-evident because $T_\text{c}$ is derived from the Bose-Einstein equilibrium solution, which is the stationary limit of the NBDE for $t\rightarrow\infty$.
\begin{figure}
\centering
\includegraphics[scale=0.88]{fig7
\caption{\label{fig7} Condensate fraction $N_\text{c}(t)/N$ in an equilibrating Bose gas of $^{23}$Na subsequent to fast evaporative cooling in a single step from $T_\text{i}=876$ nK to $T_\text{f}=750$ nK as calculated
from the analytical solution of the NBDE Eq.\,(\ref{bose}) with $k_\text{max}\equiv K=5, 10, 20, 40$ in the series expansion of the exact solution, cutoff energy
$\epsilon_\text{i}=2190$\,nK, $\mu_\text{i}=-8$\,nK, and the density of states for a free Bose gas.
The transport coefficients are $D = 3750$\,(nK)$^2$\,ms$^{-1},~\,v = -5$\,nK$\,$ms$^{-1}$.
The MIT data for the condensate fraction (crosses, no error bars) are from Ref.\,\cite{miesner_bosonic_1998}.}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.88]{fig8
\caption{\label{fig8} Condensate fraction $N_\text{c}(t)/N$ as in Fig.\,\ref{fig7}, following evaporative cooling (RF-sweep) in a single step from $T_\text{i}=876$ nK to $T_\text{f}=660$ nK as calculated
from the analytical solution of the NBDE Eq.\,(\ref{bose}) with $k_\text{max}\equiv K=5, 10, 20, 40$, $\epsilon_\text{i}=1810$\,nK, $\mu_\text{i}=-18$\,nK, and the density of states in a harmonic trap.
Due to the constant transport coefficients, the diffusion into the condensate starts instantaneously.
The transport coefficients are $D = 3300$\,(nK)$^2$\,ms$^{-1},~\,v = -5$\,nK$\,$ms$^{-1}$.
The MIT data for $^{23}$Na are from Ref.\,\cite{miesner_bosonic_1998}, with an enlarged time scale. }
\end{figure}
In order to compare our results for $T_\text{f}<T_\text{c}$ to data we use the MIT measurements from Ref.\,\cite{miesner_bosonic_1998} with parameters from
Ref.\,\cite{bzs00}, where a nonlinear Schr\"odinger equation coupled to a
Boltzmann-like quantum collision term has been solved numerically.
In particular, the final number of
condensate atoms
is taken from Ref.\,\cite{miesner_bosonic_1998}. The initial
number of atoms in the trap is unknown but in Ref.\,\cite{bzs00}, $N_0$ = \num{60e6} has been used
in order to achieve agreement of a numerical simulation and the data. We adopt their values of $T_\text{i} = 876$\,{nK} and $\epsilon_\text{i}/T_\text{i}$ = \num{2.5}. In contrast to Ref.\,\cite{bzs00} our treatment also requires the specification of the final
temperature which we infer from energy conservation in a numerical calculation, yielding $T_\text{f} = 750$\,nK for the free gas, and $T_\text{f} = 660$\,nK with the density of states in a harmonic oscillator potential as in the MIT trap.
After solving the NBDE for $n(\epsilon,t)$ at fixed initial chemical potential we calculate $\mu(t)$ at all times and finally
compute $N_c(t)/N_\text{tot}$ which is shown in Fig.\,\ref{fig7} together with the convergence of the analytic solution for different values of $k_\text{max}$ in the
series expansion of the exact analytic solution. In this result we have used the density of states for a free gas, to be compared with the one for a harmonic potential
in Fig.\,\ref{fig8} (here the timescale is enlarged to emphasize the small-time region).
In both cases, the time axis has been shifted to give the best possible fit. The transport coefficients were adapted
such that the observed time scale was reproduced. With these adjustments, data and model result are found to agree.
The solutions are steeper for the free gas at small times compared to the harmonic-oscillator density of states.
One feature both solutions fail to account for, however,
is the measured initial slow increase of the condensate fraction in the first {30}\,{ms}, which cannot be reproduced in our model with constant
transport coefficients. Should the small-time gradual rise of the condensate persist in future measurements, we can adapt the model using
time-dependent transport coefficients, such that the diffusion into the condensate builds up slowly with time: For constant drift and diffusion coefficients,
the diffusion into the condensate sets in instantaneously with full strength and without a seed condensate.
It is recognized that the experiments deal with an inhomogeneous trapped Bose gas, whereas the theory in the present form is homogeneous. In particular, we consider the difference between a finite system with and without a trapping potential only through the difference in the densities of states. This is certainly a limitation of the model, it is not meant to present a full description of the dynamics of condensate formation. But it yields physically reasonable results on a purely nonequilibrium-statistical basis.
\section{Conclusion}
Exact solutions of the nonlinear boson diffusion equation which take account of the singularity in the initial conditions at $\epsilon=\mu<0$ and the necessary boundary conditions at the singularity have been explored in this work. The analytical solutions for constant transport coefficients are compared in detail with numerical solutions of the NBDE obtained with Matlab and agreement is found. The exact solutions are used
to calculate the time-dependent entropy in a cooling Bose gas. Whereas cooling drastically reduces the entropy, the subsequent re-thermalization causes a gradual increase towards the equilibrium value, which coincides with the Bose-Einstein result and is significantly below the initial value before cooling. Wave and particle entropy are discussed. Together with particle-number conservation and a time-dependent chemical potential, the condensate fraction in a trap has been calculated as a function of time.
The solutions of the NBDE are thus shown to properly describe equilibration processes that occur in quantum gases in the course of evaporative cooling, and the associated condensate formation for $T_\text{f}<T_\text{c}$.
They are specifically applied to evaporative cooling and condensate formation in $^{23}$Na. Agreement with MIT data for the time-dependent condensate fraction is found.
The convergence properties of the infinite series expansion in the analytic solution of the NBDE have been tested.
In this work, we have not aimed at a detailed treatment of the quantum-mechanical properties of the condensate, which is coupled to the time-dependent evolution of the thermal cloud. Instead, the characteristic nonequilibrium-statistical features of the bosonic system and in particular, the boson stimulation that is encoded in the NBDE, have been used together with particle-number conservation to infer the time-dependent evolution of the condensate fraction.
Further refinements of the model such as variable transport coefficients to describe a gradual buildup of the condensate, and the coupling to the quantum fluctuations or excitations of the condensate are conceivable, but are unlikely to allow for analytic solutions. Extensions of the NBDE itself to higher dimensions in order to better account for anisotropic systems should also be tested.
More extensive comparisons of the results to new data from cold-atom experiments with an improved signal-to-noise ratio and systematic as well as statistical error bars would be most welcome. In particular, time-dependent measurements for other bosonic alkali atoms such as rubidium or lithium with substantially different timescales for condensate formation and equilibration are very desirable. This could, in particular, offer the opportunity to learn more about the initial BEC formation phase in the first 50 ms. Here, the quantum Boltzmann-type approaches often use an artificial seed condensate. Although this is needed to trigger condensate growth and achieve agreement with the data, it seems difficult to develop a sound physical justification for the initial condensate occupation. In contrast, in the NBDE approach no seed is needed. \\
\acknowledgments
We thank both referees for questions and comments, and Joseph Indekeu of Leuven University for the editorial work.
|
\section{Introduction}
Any complex Enriques surface is doubly covered by a K3 surface.
On the other hand, a K3 surface \(X\) can cover infinitely many Enriques surfaces. The set \(\mathrm{Enr}(X)\) of isomorphism classes of Enriques surfaces doubly covered by \(X\), though, is always finite by a result of Ohashi \cite{Ohashi:number.Enriques}.
We call its cardinality \(|{\mathrm{Enr}(X)}|\) the \emph{Enriques number} of the K3 surface \(X\).
The Enriques number \(|{\mathrm{Enr}(X)}|\) only depends on the transcendental lattice of~\(X\). Shimada and the second author~\cite{Shimada.Veniani:Enriques.involutions.singular.K3} described a procedure to determine \(|{\mathrm{Enr}(X)}|\) and applied it to K3 surfaces of maximal Picard rank~\(20\).
A K3 surface \(X\) of Picard rank~\(19\) can be seen as the generic element of a pencil of K3 surfaces. Its transcendental lattice \(T_X\) is an even lattice of signature \((2,1)\).
By a result of Brandhorst, Sonel and the second author~\cite{Brandhorst.Sonel.Veniani:idoneal.genera}, the surface \(X\) covers an Enriques surface only if \(4\) divides \(\det(T_X)\), but this condition is not sufficient. In this paper we analyze in detail what happens when \(|{\det(T_X)}|\) is small, more precisely
\begin{equation} \label{eq:det(T_X)<16}
|{\det(T_X)}| < 16.
\end{equation}
Henceforth, let \(X\) be a K3 surface of Picard rank \(19\) with transcendental lattice \(T_X\). In the case \(T_X \cong \bU \oplus [2n]\), \(n \geq 1\), it was already noted by Hulek and Schütt~\cite{Hulek.Schuett:Enriques.surf.jacobian.ell} that \(\mathrm{Enr}(X) \neq \emptyset\) if and only if \(n\) is even.
Indeed, we prove in \autoref{lem:|det(T_X)|<16} under assumption~\eqref{eq:det(T_X)<16} that \(\mathrm{Enr}(X) \neq \emptyset\) if and only if
\[
T_X \cong \bU \oplus [4],\,\bU \oplus [8] \text{ or } \bU \oplus [12].
\]
The main reason for bound \eqref{eq:det(T_X)<16} is to keep computations feasible.
In particular the enumeration of jacobian elliptic fibrations on K3 surfaces with \(T_X \cong \bU \oplus [16]\) already becomes quite hard. Moreover, the pencil of K3 surfaces with \(T_X \cong \bU(2) \oplus [4]\) is not of the form \(\bU \oplus [2n]\), but it still holds \(\mathrm{Enr}(X) \neq \emptyset\), as its generic element is a Kummer surface~\cite{Keum:Kummer.Enriques}.
Quite interestingly, the first two pencils already feature prominently in Kond\={o}'s classification of Enriques surfaces with finite automorphism group \cite{Kondo:Enriques.finite.aut}, which we briefly recall.
There are seven families of such Enriques surfaces, numbered \(\I\) to \(\VII\).
Families \(\I\) and \(\II\) are \(1\)-dimensional, while families \(\III\) to \(\VII\) are \(0\)-dimensional.
The K3 surfaces covering the generic Enriques surface of type~\(\I\) and \(\II\) have transcendental lattice \(T_X \cong \bU \oplus [4]\) and \(T_X \cong \bU \oplus [8]\), respectively.
The third pencil with \(T_X \cong \bU \oplus [12]\) has also been extensively studied, because of its arithmetical properties and its appearance in several seemingly unrelated physical contexts (see \cite{Festi.vanStraten:Bhabha.scattering, Peters.Stienstra:pencil.Apery}).
Following Bertin and Lecacheux~\cite{Bertin.Lecacheux:apery-fermi.2-isogenies}, who classified the elliptic fibrations on its generic element (\autoref{tab:genus_Apery_Fermi}), we call it the \emph{Apéry--Fermi pencil}.
The aim of this paper is to enumerate and investigate the Enriques surfaces covered by these three pencils.
More precisely, for each \(m \in \set{1,2,3}\) we consider a K3 surface \(X\) with \(T_X \cong \bU \oplus [4m]\) and do the following:
\begin{itemize}
\item we compute the Enriques number \(|{\mathrm{Enr}(X)}|\);
\item we classify all jacobian elliptic fibrations on \(X\) using the extension of the Kneser--Nishiyama method explained in \cite{Festi.Veniani:counting.ell.fibr.K3};
\item we relate the special elliptic pencils on the Enriques quotients to the elliptic fibrations on~\(X\).
\end{itemize}
We summarize here our findings.
Fix \(m \in \IZ\), \(m \geq 1\), and let \(\omega\) be the number of prime divisors of \(2m\) and \(X\) a K3 surface with \(T_X \cong \bU \oplus [4m]\), \(m \geq 1\).
Among the Enriques quotients of \(X\) there are \(2^{\omega-1}\) which we call \emph{of Barth--Peters type} (\autoref{lem:BP.quotients}).
Such quotients admit a cohomologically trivial involution (see~\cite{Mukai:numerically.trivial.inv.Kummer.type,Mukai.Namikawa:aut.Enriques.which.act.trivially}) and their presence is explained by the fact that our pencils are subfamilies of the \(2\)-dimensional Barth--Peters family, a fact already noted by Hulek and Schütt \cite{Hulek.Schuett:Enriques.surf.jacobian.ell,Hulek.Schuett:arithmetic.singular.Enriques}.
It turns out that if \(m = 1\), then \(X\) covers only one Enriques surface \(Y\) (\autoref{thm:Enr(X)_Kondo_I}). Therefore, the Enriques surface \(Y\) is of Barth--Peters type and, moreover, coincides with Kond\={o}'s quotient, so it has finite automorphism group.
The list of the \(9\) elliptic fibrations on \(X\) appears in other papers by Scattone~\cite{Scattone:on.the.compactification}, Dolgachev~\cite{Dolgachev:Mirror.symm.latt.pol.K3} and Elkies and Schütt~\cite{Elkies.Schuett:K3.families.high.Picard}, and we confirm it here (\autoref{tab:genus_Kondo_I}).
If \(m = 2\), then \(X\) covers two Enriques surfaces \(Y',Y''\), of which only one, say \(Y'\), is of Barth--Peters type. We show that the other surface \(Y''\) is Kond\={o}'s quotient with finite automorphism group.
We include the classification of elliptic fibrations on \(X\) up to automorphisms (\autoref{tab:genus_Kondo_II}).
One subtlety arises in this case: two of the \(17\) elliptic fibrations on \(X\) (No. 12 and 13 in \autoref{tab:genus_Kondo_II}) have the same Mordell-Weil group and two singular fibers of type~\(\I_4^*\).
Nonetheless, the two fibrations are not equivalent under the action of \(\Aut(X)\), as they have different frames.
We determine which one is the pullback of a special elliptic pencil on \(Y'\) and which one is the pullback of a special elliptic pencil on \(Y''\) (\autoref{rmk:W11,W12}).
Finally, if \(m = 3\), then \(X\) covers three Enriques surfaces \(Y',Y'',Y'''\), of which two, say \(Y'\) and \(Y''\) are of Barth--Peters type (\autoref{thm:Apery_Fermi}).
Applying a construction by Hulek and Schütt \cite[§3]{Hulek.Schuett:Enriques.surf.jacobian.ell} and using a particular configuration of curves on \(X\) found by Peters and Stienstra, we determine a simple description of an explicit Enriques involution for \(Y'''\).
In this way we find a configuration of smooth rational curves on \(Y'''\) whose dual graph is the union of a tetrahedron and a complete graph of degree~\(6\) (\autoref{rmk:Apery-Fermi.ell.fibr}).
\subsection*{Acknowledgments}
We warmly thank Simon Brandhorst, Klaus Hulek, Matthias Schütt and Ichiro Shimada for their valuable comments. We are also grateful to the anonymous referee for carefully reading the manuscript and for their useful remarks.
\section{Preliminary results}
In this section, after explaining our conventions on lattices in \autoref{sec:lattices}, we collect results regarding K3 surfaces with transcendental lattice \(T_X \cong \bU \oplus [2m]\), \(m \in \IZ\), especially regarding their jacobian elliptic fibrations in \autoref{sec:fibrations}.
In~\autoref{sec:Enriques.numbers} we recall the enumeration formula for Enriques quotients contained in \cite{Shimada.Veniani:Enriques.involutions.singular.K3} and we prove the lemma that motivates the whole paper.
Finally, in \autoref{sec:Barth-Peters.type} we introduce the notion of Enriques quotient of Barth--Peters type.
\subsection{Lattices} \label{sec:lattices}
In this paper, a \emph{lattice of rank \(r\)} is a finitely generated free \(\IZ\)-module \(L \cong \IZ^r\) endowed with an integral symmetric bilinear form \(L \times L \rightarrow \IZ\) denoted \((v,w) \mapsto v \cdot w\).
The \emph{signature} of \(L\) is the signature of the induced real symmetric form on \(L \otimes \IR\).
We say that \(L\) is \emph{even} if \(v^2 \coloneqq v \cdot v \in 2 \IZ\) for every \(v \in L\).
The \emph{dual} \(L^\vee \coloneqq \hom(L,\IZ)\) of \(L\) can be identified with \(\Set{w \in L \otimes \IQ}{w \cdot v \in \IZ \text{ for all } v \in L}\). The \emph{discriminant group} of \(L\) is defined as
\[
L^\sharp \coloneqq L^\vee/L,
\]
which is a finite abelian group.
We denote by \(\ell(L^\sharp)\) its \emph{length}, i.e. the minimal number of generators. For a prime number \(p\) we denote by \(\ell_p(L^\sharp)\) its \emph{\(p\)-length}, i.e. the minimal number of generators of its \(p\)-part.
If \(L\) is an even lattice, then \(L^\sharp\) acquires naturally the structure of a finite quadratic form \(L^\sharp \rightarrow \IQ/2\IZ\).
There is a natural homomorphism \(\OO(L) \rightarrow \OO(L^\sharp)\) denoted \(\gamma \mapsto \gamma^\sharp\).
We write \(\bU\) for the indefinite unimodular even lattice of rank~\(2\), and \(\bA_n,\bD_n,\bE_n\) for the negative definite ADE lattices. The notation~\([m]\), with \(m \in \IZ\), denotes the lattice of rank~\(1\) generated by a vector of square~\(m\).
We adopt Miranda--Morrison's notation~\cite{Miranda.Morrison} for the elementary finite quadratic forms \(\bu_k,\bv_k,\bw^\varepsilon_{p,k}\).
We recall that \(\bu_k\) (resp. \(\bv_k\)) is generated by two elements of order \(2^k\), both of square \(0 \in \IQ/2\IZ\) (resp. \(1 \in \IQ/2\IZ\)), such that their product is equal to \(1/2^k \in \IQ/\IZ\). The forms \(\bw_{2,k}^\varepsilon\), with \(\varepsilon \in \set{1,3,5,7}\), are generated by one element of order \(2^k\) and square \(\varepsilon/2^k \in \IQ/2\IZ\).
For an odd prime \(p\) the forms \(\bw_{p,k}^\varepsilon\), with \(\varepsilon \in \set{\pm 1}\), are generated by one element of order \(p^k\) and square \(a/p^k \in \IQ/2\IZ\), where \(a\) is a square modulo \(p\) if and only if \(\varepsilon = 1\).
The \emph{genus} of a lattice \(L\) is defined as the set of isomorphism classes of lattices \(M\) with \(\sign(L) = \sign(M)\) and \(L^\sharp \cong M^\sharp\).
A genus is always a finite set (see \cite[Satz~21.3]{Kneser:quadratische.Formen}).
An embedding of lattices \(\iota\colon M \hookrightarrow L\) is called \emph{primitive} if \(L/\iota(M)\) is a free group. We denote by \(\iota(M)^\perp \subset L\) the \emph{orthogonal complement} of \( M \) inside \( L \).
We quickly summarize Nikulin's theory of primitive embeddings \cite{Nikulin:int.sym.bilinear.forms}.
By \cite[Prop.~1.5.1]{Nikulin:int.sym.bilinear.forms} a primitive embedding of even lattices \(M \hookrightarrow L\) is given by a subgroup \(H \subset M^\sharp\) and an isometry \[
\gamma \colon H \rightarrow H' \coloneqq \gamma(H) \subset (\iota(M)^\perp(-1))^\sharp.
\]
If \(\Gamma\) denotes the graph of \(\gamma\) in \(M^\sharp \oplus (\iota(M)^\perp(-1))^\sharp\), the following identification between finite quadratic forms holds (the finite quadratic form on the right side being induced by the one on \(M^\sharp \oplus (\iota(M)^\perp(-1))^\sharp\)):
\begin{equation} \label{eq:L^sharp=Gamma^perp/Gamma}
L^\sharp \cong \Gamma^\perp/\Gamma.
\end{equation}
In this paper we call \(H\), \(\gamma\) resp. \(\Gamma\) the \emph{gluing subgroup}, \emph{gluing isometry} resp. \emph{gluing graph} of \(M \hookrightarrow L\).
Equivalently by \cite[Prop.~1.15.1]{Nikulin:int.sym.bilinear.forms}, assuming that \(L\) is unique in its genus, a primitive embedding \(M \hookrightarrow L\) is given by a subgroup \(K \subset L^\sharp\) and an isometry
\[
\xi\colon K \rightarrow K' \coloneqq \xi(K) \subset M(-1)^\sharp.
\]
If \(\Xi\) denotes the graph of \(\xi\) in \(L^\sharp \oplus M(-1)^\sharp\), the following identification between finite quadratic forms holds (the finite quadratic form on the right side being induced by the one on \(L^\sharp \oplus M(-1)^\sharp\)):
\begin{equation} \label{eq:(M^perp)^sharp=Xi^perp/Xi}
(\iota(M)^\perp)^\sharp \cong \Xi^\perp/\Xi.
\end{equation}
In this paper we call \(K\), \(\xi\) resp. \(\Xi\) the \emph{embedding subgroup}, \emph{embedding isometry} resp. \emph{embedding graph} of \(M \hookrightarrow L\).
\subsection{Elliptic fibrations} \label{sec:fibrations}
Given a K3 surface \(X\), we denote \(T_X\) its transcendental lattice, \(S_X\) its Néron--Severi lattice, and \(\cJ_X\) the set of jacobian elliptic fibrations on \(X\).
The \emph{frame genus} of \(X\) is defined as the genus \(\cW_X\) of negative definite lattices~\(W\) with \(\rank(W) = \rank(S_X) - 2\) and \(W^\sharp \cong S_X^\sharp\).
The lattices in \(\cW_X\) are called \emph{frames}.
The classes of a fiber and a section of a jacobian elliptic fibration induces a primitive embedding \(\iota \colon \bU \hookrightarrow S_X\).
As explained in \cite{Festi.Veniani:counting.ell.fibr.K3}, there is a well-defined function
\[
\fr_X \colon \cJ_X/{\Aut(X)} \rightarrow \cW_X
\]
which sends each jacobian fibration to the isomorphism class of \(\iota(\bU)^\perp \subset S_X\).
\begin{lemma} \label{lem:BP.fibrations}
If \(X\) is a K3 surface with transcendental lattice \(T_X \cong \bU \oplus [2n]\), \(n \geq 1\),
then on \(X\) there are exactly \(2^{\omega-1}\) jacobian elliptic fibrations with frame \(W \coloneqq \bE_8^2 \oplus [-2n]\) up to automorphisms, where \(\omega\) is the number of prime divisors of~\(2n\).
\end{lemma}
\proof
Essentially by \cite[Thm.~2.8]{Festi.Veniani:counting.ell.fibr.K3} we want to prove that
\[
|{\OO_\hdg^\sharp(T_X)}\backslash{\OO(T_X^\sharp)}/{\OO(W)}| = 2^{\omega-1},
\]
If \(n = 1\), then \(\OO(T_X^\sharp) = \set{\id}\) and we conclude immediately.
Suppose that \(n \geq 2\). Since \(\ell(T_X^\sharp) = 1\), the discriminant form \(T_X^\sharp\) is the direct sum of forms \(\bw_{p,k}^\varepsilon\).
It holds \(\OO(q \oplus q') \cong \OO(q) \times \OO(q')\) if \(q\) and \(q'\) are finite quadratic forms with \(|q|\) and \(|q'|\) coprime, and \(|{\OO(\bw^{\varepsilon}_{p,k})}| = 2\) if \(p\) is odd or \(p = 2\) and \(k \geq 2\).
Hence, \(\OO(T_X^\sharp)\) is a \(2\)-elementary group of length~\(\omega\). In particular,
\[
|{\OO(T_X^\sharp)}| = 2^{\omega},
\]
As \(\rank(T_X)\) is odd, it holds \(\OO_\hodge^\sharp(T_X) = \set{\pm \id}\) (see for instance \cite[Cor.~3.3.5]{Huybrechts:lectures.K3}).
Note, moreover, that \(\id \neq -\id\) in \(T_X^\sharp\).
The orthogonal group of \(W\) is the direct sum of \(\OO(\bE_8^2)\), which has trivial action on the discriminant group because \(\bE_8\) is unimodular, and \(\OO([-2n]) = \set{\pm \id}\).
Therefore, it also holds \(\OO^\sharp(W) = \set{\pm \id}\), so we have
\[
|{\OO_\hodge^\sharp(T_X)}\backslash{\OO(T_X^\sharp)}/{\OO^\sharp(W)}| = |{\OO(T_X^\sharp)}/{\set{\pm \id}}| = |{\OO(T_X^\sharp)}| / |{\set{\pm \id}}| = 2^{\omega-1}. \qedhere
\]
\endproof
The \emph{Mordell--Weil group}, i.e. the group of sections of a jacobian elliptic fibration, is naturally endowed with a rational symmetric bilinear form denoted by \((P,Q) \mapsto \langle P, Q \rangle \in \IQ\), called the \emph{Mordell--Weil lattice}. The \emph{height} of a section is defined as \(\height(P) \coloneqq \langle P, P \rangle\). For a clear exposition of this topic we refer to Shioda's original paper~\cite{Shioda:on.MW.lattices}.
\begin{remark} \label{rmk:Hulek.Schutt.II^*}
Consider one of the elliptic fibrations \(\pi \colon X \rightarrow \IP^1\) as in \autoref{lem:BP.fibrations} and for simplicity assume that \(n \geq 2\).
Since \(W_\rootlattice \cong \bE_8^2\), the fibration \(\pi\) has two singular fibers of Kodaira type~\(\II^*\).
As already remarked by Hulek and Schütt~\cite[§4.2.2]{Hulek.Schuett:Enriques.surf.jacobian.ell}, starting from the fibration \(\pi\) we can construct an involution on \(X\) which turns out to be an Enriques involution if \(n\) is even.
We repeat here their construction directly on the lattice \(S_X \cong \bU \oplus \bE_8^2 \oplus [-2n]\).
In the following computations we let \(\OO(S_X)\) act on \(S_X\) \emph{from the right},
so the composition of two isometries in \(\OO(S_X)\) corresponds to the product of their associated matrices in reversed order.
Let \(s_1,\ldots,s_{19}\) be a system of generators of \(S_X\) such that the corresponding Gram matrix is the standard one.
Then, \(S_X^\sharp\) is generated by \(s_{19}/(2n)\).
In these coordinates, consider the vectors
\begin{align*}
F & \coloneqq ( 1, 0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), \\
O & \coloneqq (-1, 1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), \\
P & \coloneqq (n-1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1).
\end{align*}
Note that \(F^2 = 0, O^2 = P^2 = -2, F \cdot O = F \cdot P = 1, P \cdot O = n-2\).
We can suppose that \(s_3,\ldots,s_{18}\), generating the two copies of \(\bE_8\), correspond to the components of the singular fibers which do not intersect \(O\), \(F\) to the class of a fiber, \(O\) to a section which we take as origin and \(P\) to a section of height (cf. \cite[eq. (8.19)]{Shioda:on.MW.lattices})
\[
\height(P) = \langle P,P \rangle = 2 \chi(\cO_X) + 2 P \cdot O = 2 \cdot 2 + 2 \cdot (n-2) = 2n.
\]
Let \(Q \coloneqq \boxminus P\) be the inverse section of \(P\). Then \(\height(Q) = \height(P) = 2n\), hence \(Q\cdot O = n-2\).
It follows from \(\langle P,Q \rangle = - \langle P,P \rangle\) that (cf. \cite[eq. (8.18)]{Shioda:on.MW.lattices})
\[
P \cdot Q = \chi(\cO_X) + P \cdot O + Q \cdot O - \langle P, Q \rangle = 2 + (n-2) + (n-2) + 2n = 4n- 2.
\]
Recalling moreover that \(F \cdot Q = 1\) and \(Q\cdot Q = -2\), we see that in our basis we can write
\[
Q = (n-1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,-1).
\]
Let \(t_{P} \in \OO(S_X)\) be the pullback of the automorphism induced by the translation by \(P\). We have \(t_P(Q) = O, t_P(O) = P, t_P(F) = F\) and \(t_P\) acts trivially on the other components of the fibers of type~\(\II^*\).
Therefore, \(\tau_P\) is given by the following matrix:
\[
t_P = \begin{pmatrix}
1 & & & \\
n & 1 & & 1 \\
& & I_{16} & \\
2n & & & 1 \\
\end{pmatrix}
\]
Let now \(\imath \in \OO(S_X)\) be the isometry given by
\[
\imath = \begin{pmatrix}
I_2 & \\
& & I_{8} & \\
& I_8 & \\
& & & -1 \\
\end{pmatrix}.
\]
Note that \(\imath\) swaps the two fibers of type \(\II^*\) and that \(\imath(F) = F\), \(\imath(P) = Q\) and \(\imath(Q) = P\).
Constructing an ample divisor as in \cite[Prop.~2.7]{Festi.Veniani:counting.ell.fibr.K3}, it is easy to see that \(\imath\) preserves the ample cone. Moreover, \(\imath^\sharp = -\id \in S_X^\sharp\). Therefore, by the Torelli theorem \(\imath\) is the pullback of a non-symplectic involution (whose quotient is a rational surface).
Consider \(\varepsilon \coloneqq \imath \circ t_P\), whose matrix is then given by
\[
\varepsilon = \begin{pmatrix}
1 \\
n & 1 & & & -1\\
& & & I_{8} & \\
& & I_8 & \\
2n & & & & -1 \\
\end{pmatrix}.
\]
A computation shows that for even \(n\), the invariant lattice of \(\varepsilon\) is isometric to \(\bE_8(2) \oplus \bU(2)\), hence it corresponds to an Enriques involution by Nikulin's classification~\cite[Thm.~4.2.2]{Nikulin:quotient.groups.hyperbolic.forms.2-reflections}.
The coinvariant lattice is isometric to \(\bE_8(2) \oplus [-2n]\).
\end{remark}
\subsection{Enriques numbers} \label{sec:Enriques.numbers}
Let \(X\) be a K3 surface with Néron--Severi lattice \(S_X\) and transcendental lattice \(T_X\).
We recall briefly the formula for \(|{\mathrm{Enr}(X)}|\) proved in \cite{Shimada.Veniani:Enriques.involutions.singular.K3}.
We define
\[
\bM \coloneqq \bU(2) \oplus \bE_8(2).
\]
By Nikulin's classification~\cite{Nikulin:quotient.groups.hyperbolic.forms.2-reflections}, if \(\varepsilon \in \OO(S_X)\) is the pullback of an Enriques involution, then the invariant sublattice
\(S_X^\varepsilon \coloneqq \Set{x \in S_X}{\varepsilon(x) = x}\)
is isomorphic to \(\bM\). We denote by \((S_X)_\varepsilon \coloneqq (S_X^\varepsilon)^\perp\) the coinvariant lattice, whose isometry class depends on the involution \(\varepsilon\).
Given a primitive embedding \(\iota \colon \bM \hookrightarrow S_X\), we put
\[
\OO(S_X,\iota) \coloneqq \{\, \varphi \in \OO(S_X) \,\vert\, \varphi(\iota(\bM)) = \iota(\bM)\,\}.
\]
The Hodge structure on \(\HH^2(X,\IZ)\) induces a Hodge structure on \(T_X\). We write \(\OO_\hodge(T_X)\) for the group of Hodge isometries of \(T_X\).
We fix an anti-isometry \(T_X^\sharp \cong S_X^\sharp\) (cf. \cite[Prop.~1.6.1]{Nikulin:int.sym.bilinear.forms}), so that we can identify \(\OO(T_X^\sharp) \cong \OO(S_X^\sharp)\).
We denote the images of \(\OO_\hodge(T_X)\) and \(\OO(S_X,\iota)\) under the natural morphisms \(\OO(T_X) \rightarrow \OO(T_X^\sharp)\) and \(\OO(S_X) \rightarrow \OO(S_X^\sharp)\) by \(\OO_\hodge^\sharp(T_X)\) and \(\OO^\sharp(S_X,\iota)\), respectively.
\begin{theorem}[{\cite[Thm.~3.1.9]{Shimada.Veniani:Enriques.involutions.singular.K3}}] \label{thm:Shimada.Veniani}
For any K3 surface \(X\) it holds
\[
|{\mathrm{Enr}(X)}| = \sum
|{\OO_\hodge^\sharp(T_X)}\backslash{\OO(T_X^\sharp)}/{\OO^\sharp(S_X,\iota)}|,
\]
where the sum runs over all primitive embeddings \(\iota\colon \bM \hookrightarrow S_X\) up to the action of \(\OO(S_X)\) such that there exists no \(v \in \iota(\bM)^\perp\) with \(v^2 = -2\).
\end{theorem}
The main topic of the present paper are K3 surfaces of Picard rank~\(19\) covering an Enriques surface.
For computational reasons, we restrict ourselves to K3 surfaces whose transcendental lattice has discriminant \(|{\det(T_X)}| < 16\). By the next lemma, we have three cases to consider.
\begin{lemma} \label{lem:|det(T_X)|<16}
Let \(X\) be a K3 surface of Picard rank~\(19\) and transcendental lattice \(T_X\) and suppose that \(|{\det(T_X)}| < 16\). Then, \(\mathrm{Enr}(X) \neq \emptyset\) if and only if \(T_X \cong \bU \oplus [4m]\) with \(m \in \{1,2,3\}\).
\end{lemma}
\proof
If \(T_X \cong \bU \oplus [4m]\), \(m \geq 1\), then \(X\) covers an Enriques surface by~\cite[Proposition 4.2]{Hulek.Schuett:Enriques.surf.jacobian.ell}.
Conversely, suppose that \(\mathrm{Enr}(X) \neq \emptyset\). By~\cite[Thm.~1.1]{Brandhorst.Sonel.Veniani:idoneal.genera} the lattice \(T_X\) has a Gram matrix of the form
\[
\begin{pmatrix}
2a_{11} & a_{12} & a_{13} \\
a_{12} & 4a_{22} & 2a_{23} \\
a_{13} & 2a_{23} & 4a_{33},
\end{pmatrix}, \quad a_{ij} \in \IZ.
\]
Therefore, \(\det(T_X)\) is divisible by \(4\).
Now, the discriminant group \(T_X^\sharp\) is a finite quadratic form on an abelian group of order \(|{\det(T_X)}|\) and of signature \(2-1 = 1\), because \(T_X\) has signature \((2,1)\).
We classify such finite quadratic forms \(q\) using Miranda and Morrison's normal form \cite{Miranda.Morrison}.
For each \(q\) in the list we find a lattice \(T\) such that \(T^\sharp \cong q\), obtaining the following table.
\begin{table}
\centering
\begin{tabular}{c|llllll}
\toprule
\({\det(T)}\) & \(-4\) & \(-8\) & \(-8\) & \(-8\) & \(-12\) & \(-12\) \\[2mm]
\(T^\sharp\) & \(\bw^1_{2,2}\) & \(\bu_1 \oplus \bw^1_{2,1}\) & \(\bw^1_{2,3}\) & \(\bw^5_{2,3}\)& \(\bw^{3}_{2,2} \oplus \bw^{1}_{3,1}\) & \(\bw^{7}_{2,2} \oplus \bw^{-1}_{3,1}\) \\[2mm]
\(T\) & \(\bU \oplus [4]\) & \(\bU(2) \oplus [2]\) & \(\bU \oplus [8]\) & {\small \(\begin{pmatrix} 2 & -1 & 0 \\ -1 & 2 & -1 \\ 0 & -1 & -2 \end{pmatrix}\)} & \(\bU \oplus [12]\) & {\small \(\begin{pmatrix} 2 & -1 & 0 \\ -1 & -2 & -1 \\ 0 & -1 & 2 \end{pmatrix}\)} \\
\bottomrule
\end{tabular}
\end{table}
In all cases except the second one, \(T\) is unique by \cite[Thm.~1.14.2]{Nikulin:int.sym.bilinear.forms}. In the case \(T = \bU(2) \oplus [2]\), \(T\) is unique because \(T = T'(2)\), with \(T'\) a unimodular indefinite lattice.
If \(T_X \cong T\) is such that \(T^\sharp \cong \bw^5_{2,3}\) or \(T^\sharp \cong \bw^7_{2,2} \oplus \bw^{-1}_{3,1}\), then \(\mathrm{Enr}(X) = \emptyset\) because of \cite[Prop.~3.9]{Brandhorst.Sonel.Veniani:idoneal.genera} (in the notation of \cite{Brandhorst.Sonel.Veniani:idoneal.genera}, the two forms do not satisfy condition C(1)). The case \(T_X \cong \bU(2) \oplus [2]\) is excluded because of \cite[Thm.~1.1]{Brandhorst.Sonel.Veniani:idoneal.genera} (the lattice is an `exceptional lattice').
Therefore, the only cases left are \(T_X \cong T \in \set{\bU \oplus [4],\bU \oplus [8],\bU \oplus [12]}\).
\endproof
\subsection{Enriques quotients of Barth--Peters type} \label{sec:Barth-Peters.type}
Let now \(X\) be a K3 surface with \(T_X \cong \bU \oplus [4m]\), \(m \geq 1\).
A primitive embedding \(\iota \colon \bM \hookrightarrow S_X\) depends in general on several data (cf. \autoref{sec:lattices} and \cite[Prop.~1.15.1]{Nikulin:int.sym.bilinear.forms}),
but in this case one only has to consider the orthogonal complement of the image, thanks to the next lemma.
\begin{lemma} \label{lem:emb.M-->S_X.for.T_X=U+4m}
Let \(X\) be a K3 surface with transcendental lattice \(T_X \cong \bU \oplus [4m]\), \(m \geq 1\).
If \(\iota \colon \bM \hookrightarrow S_X\) is a primitive embedding, then \(\iota(\bM)^\perp \cong N(2)\), where \(N\) is a lattice in the genus of \(\bE_8 \oplus [-2m]\). Conversely, for each such lattice \(N\) there exists exactly one primitive embedding \(\iota \colon \bM \hookrightarrow S_X\) with \(\iota(\bM)^\perp \cong N(2)\) up to the action of \(\OO(S_X)\).
\end{lemma}
\proof
The Neron--Séveri lattice \(S_X\) is isomorphic to \(\bU \oplus \bE_8^2 \oplus [-4m]\).
Consider a primitive embedding \(\iota\colon \bM \hookrightarrow S_X\) with embedding subgroup \(K \subset S_X^\sharp\) and embedding graph \(\Xi\) (see~\autoref{sec:lattices}).
Since \(\bM^\sharp \cong 5\bu_1\) is \(2\)-elementary and \(S_X^\sharp\) has length \(1\), it holds either \(|K| = 1\) or \(|K| = 2\).
The first case, though, is impossible, as otherwise \((\iota(\bM)^\perp)^\sharp\) would have length \(11 > \rank(\iota(\bM)^\perp)\).
Therefore, it must be \(|K| = 2\), so there is only one choice for the subgroup \(K \subset S_X^\sharp\), which is generated by an element of order \(2\) and square \(0 \in \IQ/2\IZ\).
Moreover, when taking \(\Xi^\perp/\Xi\) in the identification~\eqref{eq:(M^perp)^sharp=Xi^perp/Xi} one copy of \(\bu_1\) gets killed. Hence, it holds
\[
(\iota(\bM)^\perp)^\sharp \cong 4\bu_1 \oplus [-4m]^\sharp,
\]
and in particular \(\ell_2((\iota(\bM)^\perp)^\sharp) = 9 = \rank(\iota(\bM)^\perp)\).
Therefore (see for instance \cite[Lemma~3.10]{Brandhorst.Sonel.Veniani:idoneal.genera}), it holds \(\iota(\bM)^\perp \cong N(2)\), with \(N\) an even lattice. The genus of \(N(2)\) determines the genus of \(N\), so we see that \(N\) is in the genus of \(\bE_8 \oplus [-2m]\).
The converse holds by \cite[Prop.~1.15.1]{Nikulin:int.sym.bilinear.forms}, because
\(S_X\) is unique in its genus, \(K\) is uniquely determined and \(\OO(\bM) \rightarrow \OO(\bM^\sharp)\) is surjective (see for instance \cite[p. 388]{Barth.Peters:aut.enriques}).
\endproof
Note that a lattice \(N' \cong N(2)\), with \(N\) an even lattice, does not contain vectors of square \(-2\).
Therefore, \autoref{lem:emb.M-->S_X.for.T_X=U+4m} essentially says that the terms in the sum of \autoref{thm:Shimada.Veniani} are in one-to-one correspondence with the lattices in the genus of \(\bE_8 \oplus [-2m]\).
In particular, one of them corresponds to \(\bE_8 \oplus [-2m]\) itself, which we presently consider more in detail.
Barth--Peters introduced a \(2\)-dimensional family of K3 surfaces, whose general element \(\fX\) has transcendental lattice \(T_\fX \cong \bU \oplus \bU(2)\) and Néron--Severi lattice \(S_\fX \cong \bU(2) \oplus \bE_8^2\).
Ohashi~\cite[Remark~4.9(2)]{Ohashi:number.Enriques} proved that \(|{\mathrm{Enr}(\fX)}| = 1\).
The coinvariant lattice of an Enriques involution on \(\fX\) is isomorphic to \(\bE_8(2)\).
In the situation of \autoref{lem:emb.M-->S_X.for.T_X=U+4m}, if \(\bE_8(2)\) embeds into \((S_X)_\varepsilon \cong \iota(\bM)^\perp\), then \(\iota(\bM)^\perp \cong \bE_8(2) \oplus [-4m]\).
This motivates the following definition.
\begin{definition}
We say that an Enriques involution \(\varepsilon \in \Aut(X)\) on \(X\) is \emph{of Barth--Peters type} if \((S_X)_\varepsilon \cong \bE_8(2) \oplus [-4m]\).
The corresponding Enriques quotient is also called \emph{of Barth--Peters type}.
\end{definition}
The following lemma provides the number of Enriques quotients of Barth--Peters type up to isomorphisms.
\begin{lemma} \label{lem:BP.quotients}
If \(X\) is a K3 surface with transcendental lattice \(T_X \cong \bU \oplus [4m]\), \(m \geq 1\),
and \(\iota\colon \bM \hookrightarrow S_X\) is a primitive embedding with \(\iota(\bM)^\perp \cong \bE_8(2) \oplus [-4m]\), then it holds
\[
|{\OO_\hodge^\sharp(T_X)}\backslash{\OO(T_X^\sharp)}/{\OO^\sharp(S_X,\iota)}| = 2^{\omega-1},
\]
where \(\omega\) is the number of prime divisors of~\(2m\).
\end{lemma}
\proof
As in the proof of \autoref{lem:BP.fibrations}, it holds \(|{\OO(T_X^\sharp)}| = 2^{\omega}\).
As \(\rank(T_X)\) is odd, it holds \(\OO_\hodge^\sharp(T_X) = \set{\pm \id}\) (see for instance \cite[Cor.~3.3.5]{Huybrechts:lectures.K3}).
Note, moreover, that \(\id \neq -\id\) in~\(T_X^\sharp\).
We now want to determine \(\OO_\hodge^\sharp(S_X,\iota)\) using the identification \eqref{eq:L^sharp=Gamma^perp/Gamma}.
Let \(s \in \bE_8(2) \oplus [-4m]\) be the generator of the copy of \([-4m]\), \(H\) be the gluing subgroup, \(\gamma \colon H \rightarrow H'\) be the gluing isometry, and \(\Gamma\) the gluing graph of \(\iota\) (see \autoref{sec:lattices}). By the identification \eqref{eq:L^sharp=Gamma^perp/Gamma}, it holds \(|H| = |H'| = 2^9\).
Therefore, \(S_X^\sharp \cong \Gamma^\perp/\Gamma\) is generated by an element of the form \((\alpha,s/4m)\), with \(\alpha \in \bM^\sharp\).
Recall now that \(\OO(\bM) \rightarrow \OO(\bM^\sharp)\) is surjective (see \cite[p. 388]{Barth.Peters:aut.enriques}) and that an isometry of a definite lattice preserves its decomposition in irreducible lattices up to order (see for instance \cite[Satz~27.2]{Kneser:quadratische.Formen}), so \(\OO(\bE_8(2) \oplus [-4m]) \cong \OO(\bE_8(2)) \times \OO([-4m])\).
These facts imply that the group \(\OO_\hodge(S_X,\iota)\) can only act as \(\pm \id\) on \((\alpha,s/4m)\), i.e. \(\OO_\hodge^\sharp(S_X,\iota) = \set{\pm \id}\).
Therefore, we have
\[
|{\OO_\hodge^\sharp(T_X)}\backslash{\OO(T_X^\sharp)}/{\OO^\sharp(S_X,\iota)}| = |{\OO(T_X^\sharp)}/{\set{\pm \id}}| = |{\OO(T_X^\sharp)}| / |{\set{\pm \id}}| = 2^{\omega-1}. \qedhere
\]
\endproof
\begin{remark} \label{rmk:pullbacks}
Recall that any elliptic pencil on an Enriques surface has exactly two multiple fibers \(2F\), \(2F'\). The divisors \(F\) and \(F'\) are called \emph{half-pencils} (necessarily of type \(\I_m\) for some \(m \geq 0\)).
An elliptic pencil on an Enriques surface is said to be \emph{special} if it has a \(2\)-section which is a smooth rational curve.
As noted by Kond\={o} \cite[Lem.~2.6]{Kondo:Enriques.finite.aut}, the pullback of a special elliptic pencil induces a jacobian elliptic fibration \(\pi\) on the K3 surface \(X\).
Such pullbacks satisfy the following condition:
if the fibration \(\pi\) has exactly \(n_i\) fibres type of type \(\J_i\) (for \(i = 1,\ldots,r\)), where \(\J_i, \J_j\) are pairwise distinct Kodaira types if \(i \neq j\),
then at most two coefficients \(n_i\) can be odd;
moreover, if \(n_i\) is odd, then \(\J_i = \I_{2m}\) for some \(m\geq 0\).
The last sentence comes from the fact that one of the fibers of type \(\J_i\) is necessarily the pullback of a half-pencil.
\end{remark}
\begin{remark} \label{rmk:Barth-Peters.ell.fibr}
Let \(n = 2m\) be an even integer and consider one of the \(2^{\omega-1}\) elliptic fibrations with two fibers of type~\(\II^*\) given in \autoref{lem:BP.fibrations}. By the construction of \autoref{rmk:Hulek.Schutt.II^*}, we obtain one of the \(2^{\omega-1}\) Enriques quotients \(Y\) of Barth--Peters type of \autoref{lem:BP.quotients}.
In the notation of \autoref{rmk:Hulek.Schutt.II^*}, the vector
\[
R \coloneqq (m + 1, 2, -4, -5, -7, -10, -8, -6, -4, -2, -2, -3, -4, -6, -5, -4, -3, -2, 1)
\]
has square \(-2\), satisfies \(R \cdot F = 2\) and has intersection number \(1\) with \(e_3\) and \(e_{18}\). Therefore, it represents a smooth rational curve. Moreover, \(R \cdot \varepsilon(R) = 0\).
Thus, the surface \(Y\) contains ten smooth rational curves which are the images of \(R\) and of the components of the fibers of type \(\II^*\). They form the following dual graph, where the white vertex represents the image of \(R\), which is a \(4\)-section of the highlighted elliptic pencil.
\begin{center}
\begin{tikzpicture}[scale=0.5]
\draw[very thick] (-2, 0) -- (-1, 0);
\draw[very thick] ( 1, 0) -- ( 2, 0);
\draw[very thick] (-2, 0) -- (-1, 1);
\draw[very thick] (-1, 1) -- ( 0, 2);
\draw[very thick] (-2, 0) -- (-1,-1);
\draw[very thick] (-1,-1) -- ( 0,-2);
\draw ( 2, 0) -- ( 1, 1);
\draw ( 1, 1) -- ( 0, 2);
\draw[very thick] ( 2, 0) -- ( 1,-1);
\draw[very thick] ( 1,-1) -- ( 0,-2);
\draw[fill=black] (-2, 0) circle (4pt) node {};
\draw[fill=black] (-1, 0) circle (4pt) node {};
\draw[fill=black] (-1, 1) circle (4pt) node {};
\draw[fill=black] ( 0, 2) circle (4pt) node {};
\draw[fill=white] ( 1, 1) circle (4pt) node {};
\draw[fill=black] ( 2, 0) circle (4pt) node {};
\draw[fill=black] ( 1, 0) circle (4pt) node {};
\draw[fill=black] (-1,-1) circle (4pt) node {};
\draw[fill=black] ( 1,-1) circle (4pt) node {};
\draw[fill=black] ( 0,-2) circle (4pt) node {};
\end{tikzpicture}
\end{center}
This graph appears in \cite[Thm.~1.7(i)]{Kondo:Enriques.finite.aut} and is related to the fact that \(Y\) has a cohomologically trivial automorphism (such automorphisms were studied by Mukai and Namikawa~\cite{Mukai:numerically.trivial.inv.Kummer.type,Mukai.Namikawa:aut.Enriques.which.act.trivially}).
On the above graph we can recognize three more special elliptic pencils up to symmetries (dashed lines indicates half-pencils):
\begin{center}
\begin{tikzpicture}[scale=0.5]
\draw[very thick] (-2, 0) -- (-1, 0);
\draw[very thick] ( 1, 0) -- ( 2, 0);
\draw[very thick] (-2, 0) -- (-1, 1);
\draw (-1, 1) -- ( 0, 2);
\draw[very thick] (-2, 0) -- (-1,-1);
\draw[very thick] (-1,-1) -- ( 0,-2);
\draw[very thick] ( 2, 0) -- ( 1, 1);
\draw ( 1, 1) -- ( 0, 2);
\draw[very thick] ( 2, 0) -- ( 1,-1);
\draw[very thick] ( 1,-1) -- ( 0,-2);
\draw[fill=black] (-2, 0) circle (4pt) node {};
\draw[fill=black] (-1, 0) circle (4pt) node {};
\draw[fill=black] (-1, 1) circle (4pt) node {};
\draw[fill=white] ( 0, 2) circle (4pt) node {};
\draw[fill=black] ( 1, 1) circle (4pt) node {};
\draw[fill=black] ( 2, 0) circle (4pt) node {};
\draw[fill=black] ( 1, 0) circle (4pt) node {};
\draw[fill=black] (-1,-1) circle (4pt) node {};
\draw[fill=black] ( 1,-1) circle (4pt) node {};
\draw[fill=black] ( 0,-2) circle (4pt) node {};
\end{tikzpicture}
\qquad
\begin{tikzpicture}[scale=0.5]
\draw[very thick] (-2, 0) -- (-1, 0);
\draw ( 1, 0) -- ( 2, 0);
\draw[very thick] (-2, 0) -- (-1, 1);
\draw[very thick] (-1, 1) -- ( 0, 2);
\draw[very thick] (-2, 0) -- (-1,-1);
\draw[very thick] (-1,-1) -- ( 0,-2);
\draw ( 2, 0) -- ( 1, 1);
\draw[very thick] ( 1, 1) -- ( 0, 2);
\draw ( 2, 0) -- ( 1,-1);
\draw[very thick] ( 1,-1) -- ( 0,-2);
\draw[fill=black] (-2, 0) circle (4pt) node {};
\draw[fill=black] (-1, 0) circle (4pt) node {};
\draw[fill=black] (-1, 1) circle (4pt) node {};
\draw[fill=black] ( 0, 2) circle (4pt) node {};
\draw[fill=black] ( 1, 1) circle (4pt) node {};
\draw[fill=white] ( 2, 0) circle (4pt) node {};
\draw[fill=black] ( 1, 0) circle (4pt) node {};
\draw[fill=black] (-1,-1) circle (4pt) node {};
\draw[fill=black] ( 1,-1) circle (4pt) node {};
\draw[fill=black] ( 0,-2) circle (4pt) node {};
\end{tikzpicture}
\qquad
\begin{tikzpicture}[scale=0.5]
\draw (-2, 0) -- (-1, 0);
\draw ( 1, 0) -- ( 2, 0);
\draw[very thick,densely dashed] (-2, 0) -- (-1, 1);
\draw[very thick,densely dashed] (-1, 1) -- ( 0, 2);
\draw[very thick,densely dashed] (-2, 0) -- (-1,-1);
\draw[very thick,densely dashed] (-1,-1) -- ( 0,-2);
\draw[very thick,densely dashed] ( 2, 0) -- ( 1, 1);
\draw[very thick,densely dashed] ( 1, 1) -- ( 0, 2);
\draw[very thick,densely dashed] ( 2, 0) -- ( 1,-1);
\draw[very thick,densely dashed] ( 1,-1) -- ( 0,-2);
\draw[fill=black] (-2, 0) circle (4pt) node {};
\draw[fill=white] (-1, 0) circle (4pt) node {};
\draw[fill=black] (-1, 1) circle (4pt) node {};
\draw[fill=black] ( 0, 2) circle (4pt) node {};
\draw[fill=black] ( 1, 1) circle (4pt) node {};
\draw[fill=black] ( 2, 0) circle (4pt) node {};
\draw[fill=white] ( 1, 0) circle (4pt) node {};
\draw[fill=black] (-1,-1) circle (4pt) node {};
\draw[fill=black] ( 1,-1) circle (4pt) node {};
\draw[fill=black] ( 0,-2) circle (4pt) node {};
\end{tikzpicture}
\end{center}
In our case we retrieve the jacobian elliptic fibrations on \(X\) with respectively two fibres of type~\(\I_4^*\), two fibres of type~\(\III^*\), and one fibre of type \(\I_{16}\) (hence the corresponding root lattices \(W_\rootlattice\) of the frame contain the sublattices \(\bD_8^2\), \(\bE_7^2\) and \(\bA_{15}\), respectively).
\end{remark}
\section{The three pencils}
This section is divided into three subsections, in which we study K3 surface \(X\) with transcendental lattice \(T_X \cong \bU \oplus [4]\) (\autoref{sec:Kondo-I}), \(T_X \cong \bU \oplus [8]\) (\autoref{sec:Kondo-II}), and \(T_X \cong \bU \oplus [12]\) (\autoref{sec:Apery.Fermi}).
In each case we determine \(|{\mathrm{Enr}(X)}|\) and \(|\cJ_X/{\Aut(X)}|\), then we focus on their Enriques quotients, especially those \emph{not} of Barth--Peters type (because those of Barth--Peters type were already considered in \autoref{sec:Barth-Peters.type}).
Moreover, we show that all jacobian elliptic fibrations satisfying the condition in~\autoref{rmk:pullbacks} are indeed pullbacks of elliptic pencils on some Enriques quotient.
\subsection{\texorpdfstring{Kond\={o}}{Kondo}'s pencil I} \label{sec:Kondo-I}
Let \(X\) be a K3 surface with transcendental lattice
\[
T_X \cong \bU \oplus [4].
\]
\begin{theorem} \label{thm:Enr(X)_Kondo_I}
It holds \({|\mathrm{Enr}(X)|} = 1\).
\end{theorem}
\proof
The lattice \(\bE_8 \oplus [-2]\) is unique in its genus by the mass formula. By \autoref{lem:emb.M-->S_X.for.T_X=U+4m}, the sum in \autoref{thm:Shimada.Veniani} has only one term, which is equal to \(1\) by \autoref{lem:BP.quotients}.
\endproof
Therefore, the surface \(X\) admits only one Enriques quotient \(X \rightarrow Y\). Necessarily, the Barth--Peters quotient of \autoref{lem:BP.quotients} coincides with Kond\={o}'s quotient~\cite{Kondo:Enriques.finite.aut} (in particular, \(Y\) has a finite automorphism group).
Indeed, the graph of nodal curves contained in \(Y\), which is pictured in \cite[Fig.~1.4]{Kondo:Enriques.finite.aut}, contains the Barth--Peters graph as a subgraph.
This Enriques quotient was also studied by Hulek and Schütt~\cite[§4.6]{Hulek.Schuett:Enriques.surf.jacobian.ell}.
For the sake of completeness, we enumerate all jacobian elliptic fibrations on \(X\) up to automorphisms (the same list is contained in an unpublished paper by Elkies and Schütt~\cite{Elkies.Schuett:K3.families.high.Picard}).
\begin{proposition}
The frame genus \(\cW_X\) contains exactly \(9\) isomorphism classes, listed in \autoref{tab:genus_Kondo_I}, whose Gram matrices are contained in the arXiv ancillary file \verb+genus_Kondo_I.sage+.
Moreover, it holds
\[
|\cJ_X/{\Aut(X)}| = 9.
\]
\end{proposition}
\proof
It holds \(|\cJ_X/{\Aut(X)}| = |\cW_X|\) by \cite[Cor.~2.10]{Festi.Veniani:counting.ell.fibr.K3}.
In order to determine \(\cW_X\), we apply the Kneser--Nishiyama method with \(T_0 = \bD_7\).
The list is complete because the mass formula holds:
\[
\sum_{i = 1}^{9} \frac{1}{|{\OO(W_i)}|} = \frac{642332179}{18881368343036559360000} = \mass(\cW_X). \qedhere
\]
\endproof
\input{tables/genus_Kondo_I}
\subsection{\texorpdfstring{Kond\={o}}{Kondo}'s pencil II} \label{sec:Kondo-II}
Let \(X\) be a K3 surface with transcendental lattice
\[
T_X \cong \bU \oplus [8].
\]
\begin{theorem} \label{thm:Enr(X)_Kondo_II}
It holds \({|\mathrm{Enr}(X)|} = 2\).
\end{theorem}
\proof
There is only one more lattice in the genus of \(\bE_8 \oplus [-4]\), namely \(\bD_9\), as the mass formula shows.
By \autoref{lem:emb.M-->S_X.for.T_X=U+4m} the sum in \autoref{thm:Shimada.Veniani} has two terms, both equal to \(1\) as it holds \(\OO_\hodge^\sharp(T_X) = \OO(T_X^\sharp)\).
\endproof
By \autoref{lem:BP.quotients}, one of the two Enriques quotient, say \(X \rightarrow Y'\), is of Barth--Peters type. The corresponding coinvariant lattice is isometric to \(\bE_8(2) \oplus [-8]\) and contains \(240\) vectors of square \(-4\).
By Kond\={o}'s classification, the surface \(X\) admits an Enriques quotient \(X \rightarrow Y''\) with finite automorphism group.
Kond\={o}'s quotient \(Y''\) was also studied by Hulek and Schütt \cite[§4.7 and §4.8]{Hulek.Schuett:Enriques.surf.jacobian.ell}.
We argue that \(Y'\) is not isomorphic to \(Y''\).
Geometrically, this follows from the fact that \(Y''\) contains exactly \(12\) smooth rational curves whose dual graph is pictured on \cite[p.~207, Fig.~2.4]{Kondo:Enriques.finite.aut}. This dual graph does not contain the graph pictured in \autoref{rmk:Barth-Peters.ell.fibr} as a subgraph.
Algebraically, we can distinguish the two quotients in the following way.
\begin{enumerate}
\item The surface \(X\) contains \(24\) smooth rational curves \(F_1^+,F_1^-,\ldots,F_{12}^+,F_{12}^-\), which intersect as in \cite[p.~207, Fig.~2.3]{Kondo:Enriques.finite.aut} and generate the Néron--Severi lattice \(S_X\).
\item Kond\={o}'s Enriques involution exchanges \(F_i^+\) with \(F_i^-\), \(i = 1,\ldots,12\).
\item Computing explicitly the coinvariant lattice of Kond\={o}'s Enriques involution in \(S_X\), we see that it contains \(144\) vectors of square \(-4\), so it must be isomorphic to \(\bD_9(2)\). In particular, Kond\={o}'s quotient is not of Barth--Peters type.
\end{enumerate}
We now enumerate all jacobian elliptic fibrations on~\(X\) up to automorphisms.
\input{tables/genus_Kondo_II}
\begin{proposition}
The frame genus \(\cW_X\) contains exactly \(17\) isomorphism classes, listed in \autoref{tab:genus_Kondo_I}, whose Gram matrices are contained in the arXiv ancillary file \verb+genus_Kondo_II.sage+.
Moreover, it holds
\[
|\cJ_X/{\Aut(X)}| = 17.
\]
\end{proposition}
\proof
It holds \(|\cJ_X/{\Aut(X)}| = |\cW_X|\) by \cite[Cor.~2.10]{Festi.Veniani:counting.ell.fibr.K3}.
In order to determine \(\cW_X\), we apply the Kneser--Nishiyama method with \(T_0 = \bA_7\).
Note that there are two different primitive embeddings \(\bA_7 \hookrightarrow \bD_8\) (cf. \cite[Lem.~4.2]{Nishiyama:Jacobian.fibrations.K3.MW}), leading to two distinct frames \(W\) with \(W_\rootlattice \hookrightarrow \bD_8^3\), namely \(W_{11}\) and~\(W_{12}\) (cf. \autoref{rmk:W11,W12}).
The list is complete because the mass formula holds:
\[
\sum_{i = 1}^{17} \frac{1}{|{\OO(W_i)}|} = \frac{642332179}{73755345089986560000} = \mass(\cW_X). \qedhere
\]
\endproof
\begin{remark} \label{rmk:Kondo-II.ell.fibr}
The surface \(Y''\) contains \(12\) curves on whose dual graph one can recognize the following elliptic pencils (dashed lines indicate half-pencils):
\begin{center}
\begin{tikzpicture}[scale=0.5]
\draw (-2, 0) -- (-1, 2);
\draw[very thick,densely dashed] (-1, 2) -- ( 0, 3);
\draw[very thick,densely dashed] ( 1, 2) -- ( 0, 3);
\draw ( 1, 2) -- ( 2, 0);
\draw ( 2, 0) -- ( 2,-1);
\draw[very thick] ( 2,-1) -- ( 1,-1);
\draw[very thick] ( 1,-1) -- (-1,-1);
\draw[very thick] (-1,-1) -- (-2,-1);
\draw (-2,-1) -- (-2, 0);
\draw (-2, 0) -- (-1, 0);
\draw[very thick] (-1,-1) -- (-1, 0);
\draw[very thick,densely dashed] (-1, 2) -- ( 0, 1);
\draw[very thick,densely dashed] ( 1, 2) -- ( 0, 1);
\draw ( 2, 0) -- ( 1, 0);
\draw[very thick] ( 1,-1) -- ( 1, 0);
\draw[fill=white] (-2, 0) circle (4pt) node {};
\draw[fill=black] (-1, 0) circle (4pt) node {};
\draw[fill=black] ( 1, 0) circle (4pt) node {};
\draw[fill=white] ( 2, 0) circle (4pt) node {};
\draw[fill=black] (-2,-1) circle (4pt) node {};
\draw[fill=black] (-1,-1) circle (4pt) node {};
\draw[fill=black] ( 1,-1) circle (4pt) node {};
\draw[fill=black] ( 2,-1) circle (4pt) node {};
\draw[fill=black] (-1, 2) circle (4pt) node {};
\draw[fill=black] ( 0, 1) circle (4pt) node {};
\draw[fill=black] ( 0, 3) circle (4pt) node {};
\draw[fill=black] ( 1, 2) circle (4pt) node {};
\end{tikzpicture}
\qquad
\begin{tikzpicture}[scale=0.5]
\draw[very thick] (-2, 0) -- (-1, 2);
\draw[very thick] (-1, 2) -- ( 0, 3);
\draw[very thick] ( 0, 3) -- ( 1, 2);
\draw[very thick] ( 1, 2) -- ( 2, 0);
\draw[very thick] ( 2, 0) -- ( 2,-1);
\draw ( 2,-1) -- ( 1,-1);
\draw ( 1,-1) -- (-1,-1);
\draw (-1,-1) -- (-2,-1);
\draw[very thick] (-2,-1) -- (-2, 0);
\draw[very thick] (-2, 0) -- (-1, 0);
\draw (-1,-1) -- (-1, 0);
\draw (-1, 2) -- ( 0, 1);
\draw ( 1, 2) -- ( 0, 1);
\draw[very thick] ( 2, 0) -- ( 1, 0);
\draw ( 1,-1) -- ( 1, 0);
\draw[fill=black] (-2, 0) circle (4pt) node {};
\draw[fill=black] (-1, 0) circle (4pt) node {};
\draw[fill=black] ( 1, 0) circle (4pt) node {};
\draw[fill=black] ( 2, 0) circle (4pt) node {};
\draw[fill=black] (-2,-1) circle (4pt) node {};
\draw[fill=white] (-1,-1) circle (4pt) node {};
\draw[fill=white] ( 1,-1) circle (4pt) node {};
\draw[fill=black] ( 2,-1) circle (4pt) node {};
\draw[fill=black] (-1, 2) circle (4pt) node {};
\draw[fill=white] ( 0, 1) circle (4pt) node {};
\draw[fill=black] ( 0, 3) circle (4pt) node {};
\draw[fill=black] ( 1, 2) circle (4pt) node {};
\end{tikzpicture}
\qquad
\begin{tikzpicture}[scale=0.5]
\draw[very thick] (-2, 0) -- (-1, 2);
\draw[very thick] (-1, 2) -- ( 0, 3);
\draw[very thick] ( 0, 3) -- ( 1, 2);
\draw[very thick] ( 1, 2) -- ( 2, 0);
\draw[very thick] ( 2, 0) -- ( 2,-1);
\draw[very thick] ( 2,-1) -- ( 1,-1);
\draw[very thick] ( 1,-1) -- (-1,-1);
\draw[very thick] (-1,-1) -- (-2,-1);
\draw[very thick] (-2,-1) -- (-2, 0);
\draw (-2, 0) -- (-1, 0);
\draw (-1,-1) -- (-1, 0);
\draw (-1, 2) -- ( 0, 1);
\draw ( 1, 2) -- ( 0, 1);
\draw ( 2, 0) -- ( 1, 0);
\draw ( 1,-1) -- ( 1, 0);
\draw[fill=black] (-2, 0) circle (4pt) node {};
\draw[fill=white] (-1, 0) circle (4pt) node {};
\draw[fill=white] ( 1, 0) circle (4pt) node {};
\draw[fill=black] ( 2, 0) circle (4pt) node {};
\draw[fill=black] (-2,-1) circle (4pt) node {};
\draw[fill=black] (-1,-1) circle (4pt) node {};
\draw[fill=black] ( 1,-1) circle (4pt) node {};
\draw[fill=black] ( 2,-1) circle (4pt) node {};
\draw[fill=black] (-1, 2) circle (4pt) node {};
\draw[fill=white] ( 0, 1) circle (4pt) node {};
\draw[fill=black] ( 0, 3) circle (4pt) node {};
\draw[fill=black] ( 1, 2) circle (4pt) node {};
\end{tikzpicture}
\qquad
\begin{tikzpicture}[scale=0.5]
\draw[very thick,densely dashed] (-2, 0) -- (-1, 2);
\draw (-1, 2) -- ( 0, 3);
\draw ( 0, 3) -- ( 1, 2);
\draw[very thick,densely dashed] ( 1, 2) -- ( 2, 0);
\draw[very thick,densely dashed] ( 2, 0) -- ( 2,-1);
\draw[very thick,densely dashed] ( 2,-1) -- ( 1,-1);
\draw[very thick,densely dashed] ( 1,-1) -- (-1,-1);
\draw[very thick,densely dashed] (-1,-1) -- (-2,-1);
\draw[very thick,densely dashed] (-2, 0) -- (-2,-1);
\draw (-2, 0) -- (-1, 0);
\draw (-1,-1) -- (-1, 0);
\draw[very thick,densely dashed] (-1, 2) -- ( 0, 1);
\draw[very thick,densely dashed] ( 1, 2) -- ( 0, 1);
\draw ( 2, 0) -- ( 1, 0);
\draw ( 1,-1) -- ( 1, 0);
\draw[fill=black] (-2, 0) circle (4pt) node {};
\draw[fill=white] (-1, 0) circle (4pt) node {};
\draw[fill=white] ( 1, 0) circle (4pt) node {};
\draw[fill=black] ( 2, 0) circle (4pt) node {};
\draw[fill=black] (-2,-1) circle (4pt) node {};
\draw[fill=black] (-1,-1) circle (4pt) node {};
\draw[fill=black] ( 1,-1) circle (4pt) node {};
\draw[fill=black] ( 2,-1) circle (4pt) node {};
\draw[fill=black] (-1, 2) circle (4pt) node {};
\draw[fill=black] ( 0, 1) circle (4pt) node {};
\draw[fill=white] ( 0, 3) circle (4pt) node {};
\draw[fill=black] ( 1, 2) circle (4pt) node {};
\end{tikzpicture}
\end{center}
The first three pencils are special pencils and correspond to the elliptic fibrations on \(X\) with frames \(W_1\), \(W_{11}\) (see \autoref{rmk:W11,W12}) and \(W_{16}\), respectively.
The fourth pencil is not special: the highlighted curves on \(Y''\) form a half-pencil and the white vertices represent \(4\)-sections.
Indeed, the pullback on \(X\) correspond to an elliptic fibration with a fiber of type \(\I_{18}\), namely
\[
F_1^+ + F_4^- + F_3^- + F_5^- + F_6^- + F_7^- + F_9^- + F_{10}^- + F_{11}^- + F_1^- + F_4^+ + F_3^+ + F_5^+ + F_6^+ + F_7^+ + F_9^+ + F_{10}^+ + F_{11}^+.
\]
This fibration is not jacobian, as it does not appear in \autoref{tab:genus_Kondo_II}.
\end{remark}
\begin{remark} \label{rmk:W11,W12}
The two frames \(W_{11}\) and \(W_{12}\) are not isometric, but they can be distinguished neither by the pair \((W_\rootlattice,W/W_\rootlattice)\) nor by their number of automorphisms \(|{\OO(W)}|\).
Using the command \verb+is_globally_equivalent_to+ of the Sage class \verb+QuadraticForm+,
we can check that the frame \(W_{11}\) corresponds to the fibration with fiber
\begin{align*}
F & = F_6^+ + F_8^- + 2F_5^+ + 2F_3^+ + 2F_2^+ + 2F_1^+ + 2F_{11}^+ + F_{10}^+ + F_{12}^+ \\
& = F_6^- + F_8^+ + 2F_5^- + 2F_3^- + 2F_2^- + 2F_1^- + 2F_{11}^- + F_{10}^- + F_{12}^-.
\end{align*}
which is the pullback of the second special pencil on \(Y''\) listed in \autoref{rmk:Kondo-II.ell.fibr}.
On the other hand, with the same command we can check that the frame \(W_{12}\) corresponds to the fibration with fiber
\[
F_2^+ + F_5^+ + 2F_3^+ + 2F_4^+ + 2F_1^- + 2F_2^- + 2F_3^- + F_4^- + F_5^-,
\]
which is then the pullback of a special pencil on \(Y'\) (cf.~\autoref{rmk:Barth-Peters.ell.fibr}).
\end{remark}
\subsection{Apéry--Fermi pencil} \label{sec:Apery.Fermi}
Let \(X\) be a K3 surface with transcendental lattice
\[
T_X \cong \bU \oplus [12].
\]
The classification of the jacobian elliptic fibrations on \(X\) was carried out by Bertin and Lecacheux~\cite{Bertin.Lecacheux:apery-fermi.2-isogenies} and then refined in~\cite{Festi.Veniani:counting.ell.fibr.K3}.
For the reader's convenience we reproduce in \autoref{tab:genus_Apery_Fermi} the same table as \cite[Table~7]{Festi.Veniani:counting.ell.fibr.K3}.
\input{tables/genus_Apery_Fermi}
\begin{theorem} \label{thm:Enr(X)_Apery_Fermi}
It holds \({|\mathrm{Enr}(X)|} = 3\).
\end{theorem}
\proof
The genus of \(\bE_8 \oplus [-6]\) contains two lattices, namely \(\bA_2 \oplus \bE_7\) and \(\bE_8 \oplus [-6]\) itself, as the mass formula shows.
Thus, by \autoref{lem:emb.M-->S_X.for.T_X=U+4m}, the sum in \autoref{thm:Shimada.Veniani} consists of two terms, one of which is equal to~\(2\) by \autoref{lem:BP.quotients}.
Fix a primitive embedding \(\iota \colon \bM \hookrightarrow S_X\) with \(\iota(\bM)^\perp \cong \bA_2(2) \oplus \bE_7(2)\). Note that it holds
\[
S_X^\sharp \cong \bw^5_{2,2}\oplus \bw^{-1}_{3,1}, \quad \bA_2(2)^\sharp \cong \bv_1 \oplus \bw^{-1}_{3,1}, \quad \bE_7(2)^\sharp = 3\bu_1 \oplus \bw^1_{2,2}.
\]
Let \(H \subset \bM^\sharp\) be the gluing subgroup (see \autoref{sec:lattices}).
By the identification~\eqref{eq:L^sharp=Gamma^perp/Gamma} we have \(|H| = 2^9\).
Thus, the image \(H' \coloneqq \gamma(H) \subset (\iota(\bM)^\perp(-1))^\sharp\) of the gluing isometry is the sum of the copy of \(\bv_1\) in \(\bA_2(2)\) and the whole group \(\bE_7(2)^\sharp\) (with inverted sign).
Consider the isometry \(\alpha \in \OO(\iota(\bM)^\perp)\) defined as \(-\id\) on the copy of \(\bA_2(2)\) and as \(\id\) on the copy of \(\bE_7(2)\).
Since the natural homomorphism \(\OO(\bM) \rightarrow \OO(\bM^\sharp)\) is surjective, \(\alpha\) extends to an isometry \(\tilde \alpha \in \OO(S_X,\iota)\) by \cite[Cor.~1.5.2]{Nikulin:int.sym.bilinear.forms}.
By construction of \(\alpha\) and by the above description of the gluing isometry \(\gamma\), the element \(\tilde \alpha^\sharp\) acts as \(-\id\) on the \(3\)-part of \(S_X^\sharp\) and as \(\id\) on the \(2\)-part of \(S_X^\sharp\). In particular, \(\OO^\sharp(S_X,\iota)\) contains at least three different elements, namely \(\id\), \(-\id\) and \(\tilde \alpha^\sharp\).
On the other hand, \(\OO(T_X^\sharp)\) contains exactly four elements, as it is generated by multiplication by \(-1\) and by \(5\).
Therefore, we have \(\OO^\sharp(S_X,\iota) = \OO(T_X^\sharp)\), which implies \[|{\OO_\hodge^\sharp(T_X)} \backslash{\OO(T_X^\sharp)}/ {\OO^\sharp(S_X,\iota)}| = 1.\]
In total we get \(|{\mathrm{Enr}(X)}| = 3\).
\endproof
Let \(Y',Y'',Y'''\) be the three Enriques quotients of \(X\) up to automorphisms.
We can suppose that \(Y', Y''\) are of Barth--Peters type (see \autoref{sec:Barth-Peters.type}).
Here we are interested in studying \(Y \coloneqq Y'''\).
Peters and Stienstra \cite{Peters.Stienstra:pencil.Apery} showed that \(X\) contains \(20 + 12\) smooth rational curves, called \emph{\(L\)-lines} and \emph{\(M\)-lines}, forming a particular configuration, which we call the \emph{Peters--Stientra cube}.
The dual graph of the \(L\)-lines is pictured in \cite[Fig.~1]{Peters.Stienstra:pencil.Apery}.
We do not reproduce it here, but we follow the same notation.
The intersection numbers of the \(M\)-lines are described in \cite[Lem.~1]{Peters.Stienstra:pencil.Apery}.
In order to make a connection with the construction of \autoref{rmk:Hulek.Schutt.II^*}, we first look for a fibration with two fibers of type \(\II^*\) or, equivalently, with frame \(W_{2}\) in \autoref{tab:genus_Apery_Fermi}.
We can suppose
\begin{align*}
F & = 2 L_{++0} + 3 M_{2-+} + 4 L_{+++} + 6 L_{+0+} + 5 L_{+-+} + 4 L_{0-+} + 3 M_{1--} + 2 L_{0+-} + L_{-+-} \\
& = 2 L_{--0} + 3 M_{1++} + 4 L_{---} + 6 L_{0--} + 5 L_{+--} + 4 L_{+0-} + 3 M_{2+-} + 2 L_{-0+} + L_{-++},
\end{align*}
as pictured below in the Peters--Stienstra cube (note that \(M_{1++}\) and \(M_{2+-}\) and the other \(M\)-lines are not displayed):
\[
\begin{tikzpicture}[x={(1cm,0cm)},y={(3.85mm, 3.85mm)},z={(0cm,1cm)},baseline=0]
\coordinate (M2mp) at (-2, 0, 2);
\draw (-1,-1,-1) -- (-1,-1, 1);
\draw (-1,-1,-1) -- (-1, 1,-1);
\draw[very thick] (-1,-1,-1) -- ( 1,-1,-1);
\draw (-1,-1, 1) -- (-1, 1, 1);
\draw (-1,-1, 1) -- ( 1,-1, 1);
\draw (-1, 1,-1) -- (-1, 1, 1);
\draw (-1, 1,-1) -- ( 1, 1,-1);
\draw ( 1,-1,-1) -- ( 1,-1, 1);
\draw ( 1,-1,-1) -- ( 1, 1,-1);
\draw (-1, 1, 1) -- ( 1, 1, 1);
\draw[very thick] ( 1,-1, 1) -- ( 1, 1, 1);
\draw ( 1, 1,-1) -- ( 1, 1, 1);
\draw[very thick] (-1, 1,-1) -- ( 0, 1,-1);
\draw[very thick] ( 1, 1, 0) -- ( 1, 1, 1);
\draw[very thick] ( 1,-1,-1) -- ( 1, 0,-1);
\draw[very thick] (-1,-1,-1) -- (-1,-1, 0);
\draw[very thick] ( 0,-1, 1) -- ( 1,-1, 1);
\draw[very thick] (-1, 1, 1) -- (-1, 0, 1);
\draw[very thick] ( 1, 0, 1) -- (M2mp);
\draw (-1, 0,-1) -- (M2mp);
\draw[very thick] ( 0,-1, 1) -- ( 0,-1.5,-1.5);
\draw[very thick] ( 0, 1,-1) -- ( 0,-1.5,-1.5);
\draw[fill=black] (-1,-1,-1) circle (2pt) node[above right] {};
\draw[fill=white] (-1,-1, 1) circle (2pt) node[below] {};
\draw[fill=black] (-1, 1,-1) circle (2pt) node[above] {};
\draw[fill=black] (-1, 1, 1) circle (2pt) node[left] {};
\draw[fill=black] ( 1,-1,-1) circle (2pt) node[right] {};
\draw[fill=black] ( 1,-1, 1) circle (2pt) node[below] {};
\draw[fill=white] ( 1, 1,-1) circle (2pt) node[above] {};
\draw[fill=black] ( 1, 1, 1) circle (2pt) node[below right] {};
\draw[fill=black] ( 0,-1,-1) circle (2pt) node[below] {};
\draw[fill=black] ( 0,-1, 1) circle (2pt) node[below] {};
\draw[fill=black] ( 0, 1,-1) circle (2pt) node[above] {};
\draw[fill=white] ( 0, 1, 1) circle (2pt) node[above] {};
\draw[fill=white] (-1, 0,-1) circle (2pt) node[left] {};
\draw[fill=black] (-1, 0, 1) circle (2pt) node[left] {};
\draw[fill=black] ( 1, 0,-1) circle (2pt) node[right] {};
\draw[fill=black] ( 1, 0, 1) circle (2pt) node[right] {};
\draw[fill=black] (-1,-1, 0) circle (2pt) node[below right] {};
\draw[fill=white] (-1, 1, 0) circle (2pt) node[above left] {};
\draw[fill=white] ( 1,-1, 0) circle (2pt) node[below right] {};
\draw[fill=black] ( 1, 1, 0) circle (2pt) node[above left] {};
\draw[fill=black] (M2mp) circle (2pt) node[left] {\(M_{2-+}\)};
\draw[fill=black] ( 0,-1.5,-1.5) circle (2pt) node[left] {\(M_{1--}\)};
\end{tikzpicture}
\]
In the coordinate system of \autoref{rmk:Hulek.Schutt.II^*}, up to substituting \(P\) with \(Q\), we can suppose that
\begin{align*}
L_{-+0} & = O, \\
L_{++0} & = (0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), \\
L_{--0} & = (0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0), \\
L_{++-} & = (4, 4, -6, -8, -11, -16, -13, -10, -7, -4, -6, -9, -12, -18, -15, -12, -8, -4, 1).
\end{align*}
The coordinates of all other \(L\)-lines and \(M\)-lines are determined by these choices.
In order to construct the Enriques involution corresponding to \(Y\), we now consider a fibration with frame \(W_{19}\) in \autoref{tab:genus_Apery_Fermi}.
As \((W_{19})_\rootlattice \cong \bA_2^2 \bE_6^2\), the fibration has two fibers of type \(\I_3\) (or \(\IV\)) and two fibers of type \(\IV^*\).
As pictured below in the Peters--Stienstra cube (omitting the \(M\)-lines) we choose
\begin{align*}
F_{19} & \coloneqq L_{+-+} + L_{++-} + 2L_{+0+} + 2L_{++0} + 3L_{+++} + 2L_{0++} + L_{-++} = M_{3+-} + M_{1+-} + M_{2+-} \\
& = L_{+--} + L_{-+-} + 2L_{0--} + 2L_{-0-} + 3L_{---} + 2L_{--0} + L_{--+} = M_{3--} + M_{1--} + M_{2--}.
\end{align*}
\[
\begin{tikzpicture}[x={(1cm,0cm)},y={(3.85mm, 3.85mm)},z={(0cm,1cm)},baseline=0]
\draw[very thick] (-1,-1,-1) -- (-1,-1, 1);
\draw[very thick] (-1,-1,-1) -- (-1, 1,-1);
\draw[very thick] (-1,-1,-1) -- ( 1,-1,-1);
\draw (-1,-1, 1) -- (-1, 1, 1);
\draw (-1,-1, 1) -- ( 1,-1, 1);
\draw (-1, 1,-1) -- (-1, 1, 1);
\draw (-1, 1,-1) -- ( 1, 1,-1);
\draw ( 1,-1,-1) -- ( 1,-1, 1);
\draw ( 1,-1,-1) -- ( 1, 1,-1);
\draw[very thick] (-1, 1, 1) -- ( 1, 1, 1);
\draw[very thick] ( 1,-1, 1) -- ( 1, 1, 1);
\draw[very thick] ( 1, 1,-1) -- ( 1, 1, 1);
\draw[fill=black] (-1,-1,-1) circle (2pt) node[above right] {};
\draw[fill=black] (-1,-1, 1) circle (2pt) node[below] {};
\draw[fill=black] (-1, 1,-1) circle (2pt) node[above] {};
\draw[fill=black] (-1, 1, 1) circle (2pt) node[left] {};
\draw[fill=black] ( 1,-1,-1) circle (2pt) node[right] {};
\draw[fill=black] ( 1,-1, 1) circle (2pt) node[below] {};
\draw[fill=black] ( 1, 1,-1) circle (2pt) node[above] {};
\draw[fill=black] ( 1, 1, 1) circle (2pt) node[below right] {};
\draw[fill=black] ( 0,-1,-1) circle (2pt) node[below] {};
\draw[fill=white] ( 0,-1, 1) circle (2pt) node[below] {};
\draw[fill=white] ( 0, 1,-1) circle (2pt) node[above] {};
\draw[fill=black] ( 0, 1, 1) circle (2pt) node[above] {};
\draw[fill=black] (-1, 0,-1) circle (2pt) node[left] {};
\draw[fill=white] (-1, 0, 1) circle (2pt) node[left] {};
\draw[fill=white] ( 1, 0,-1) circle (2pt) node[right] {};
\draw[fill=black] ( 1, 0, 1) circle (2pt) node[right] {};
\draw[fill=black] (-1,-1, 0) circle (2pt) node[below right] {};
\draw[fill=white] (-1, 1, 0) circle (2pt) node[above left] {};
\draw[fill=white] ( 1,-1, 0) circle (2pt) node[below right] {};
\draw[fill=black] ( 1, 1, 0) circle (2pt) node[above left] {};
\end{tikzpicture}
\]
Moreover, we choose \(O_{19} \coloneqq L_{+-0}\) as origin. Then, \(P_{19} \coloneqq L_{0+-}\) and \(Q_{19} \coloneqq L_{-0+}\) become the two \(3\)-torsion sections, because it holds \(\langle P_{19}, P_{19} \rangle = \langle Q_{19}, Q_{19} \rangle = 0\), whereas
\(R_{19} \coloneqq L_{-+0}\) becomes a section of infinite order. From~\cite[eq. (8.12) and Table (8.16)]{Shioda:on.MW.lattices} it follows
\[
\langle R_{19}, R_{19} \rangle = 2\chi + 2 L_{-+0}\cdot L_{+-0} -\sum \textit{contr}_v (R_{19})= 2\cdot 2 + 2 \cdot 0 - 2\cdot \frac{4}{3}= \frac{4}{3}.
\]
\begin{theorem} \label{thm:Apery_Fermi}
There is an Enriques involution \(\varepsilon \in \Aut(X)\) which acts on the \(L\)-lines by exchanging all subscripts `\(+\)' with `\(-\)' and on the \(M\)-lines by exchanging \(M_{k+\beta}\) with \(M_{k-\beta}\), for all \(k \in \set{1,2,3}\), \(\beta \in \set{+,-}\). Moreover, \(\varepsilon\) is not of Barth--Peters type.
\end{theorem}
\proof
Let \(S_{19} \coloneqq \boxminus R_{19}\) be the section given by the inverse of \(R_{19}\) in the Mordell--Weil group.
Then we clearly have \(\langle S_{19}, S_{19}\rangle = \langle R_{19}, R_{19}\rangle\) and \(\langle S_{19}, R_{19}\rangle = -\langle R_{19}, R_{19}\rangle\).
From these equalities, using~\cite[Theorem 8.6]{Shioda:on.MW.lattices} we obtain \(O_{19} \cdot S_{19}=0\) and \(R_{19} \cdot S_{19}=2\).
These intersection numbers explicitly determine \(S_{19}\) in the coordinate system of \autoref{rmk:Hulek.Schutt.II^*}:
\[
S_{19}= (19, 17, -27, -42, -54, -81, -66, -51, -34, -17, -27, -42, -54, -81,
-66, -51, -34, -17, 4)\, .
\]
We are then able to compute the translation by \(R_{19}\), denoted by \(t\), and involution \(\imath\) as in Hulek and Schütt's construction \cite[§3]{Hulek.Schuett:Enriques.surf.jacobian.ell}.
Explicit computations show that the invariant lattice of \(\varepsilon \coloneqq t \circ \imath\) is isomorphic to \(\bM\), so that \(\varepsilon\) is the pullback of an Enriques involution.
We can verify directly that \(\varepsilon\) acts on the \(L\)-lines and \(M\)-lines as described in the statement of the theorem.
By~\autoref{lem:emb.M-->S_X.for.T_X=U+4m} and the proof of \autoref{thm:Enr(X)_Apery_Fermi},
we know that the coinvariant lattice \((S_X)_\varepsilon\) is isomorphic to either \(\bA_2(2) \oplus \bE_7(2)\) or \( \bE_8(2) \oplus [-12] \).
An explicit computation shows that \((S_X)_\varepsilon\) contains \(132\) vectors of square \(-4\), so it is necessarily isomorphic to \(\bA_2(2) \oplus \bE_7(2)\), i.e. \(\varepsilon\) is not of Barth--Peters type.
We refer to the ancillary file \texttt{calc.Apery\char`_Fermi.sage} for the actual computations in Sage.
\endproof
\begin{remark} \label{rmk:Apery-Fermi.ell.fibr}
Thanks to the description of the Enriques involution in \autoref{thm:Apery_Fermi}, it is immediate to see that the images of the \(L\)-lines in \(Y\) form a tetrahedron, while the images of the \(M\)-lines form a complete graph with \(6\) vertices in which three pairs of curves intersect doubly.
The tetrahedron and the complete graph are connected in the following way, where double intersections are marked with a double edge.
\[
\begin{tikzpicture}
\coordinate (A) at (0,0);
\coordinate (B) at (2,-0.5);
\coordinate (C) at (3,0.5);
\coordinate (D) at (1.4,2);
\coordinate (Z) at (8,0.75);
\draw (A) -- (D) coordinate [midway] (E);
\draw (B) -- (C) coordinate [midway] (F);
\draw (A) -- (B) coordinate [midway] (G);
\draw (C) -- (D) coordinate [midway] (H);
\draw (A) -- (C) coordinate [midway] (I);
\draw (B) -- (D) coordinate [midway] (L);
\path (Z) arc (0:0:1.25) coordinate (M)
(Z) arc (0:60:1.25) coordinate (N)
(Z) arc (0:120:1.25) coordinate (O)
(Z) arc (0:180:1.25) coordinate (P)
(Z) arc (0:240:1.25) coordinate (Q)
(Z) arc (0:300:1.25) coordinate (R);
\draw[double] (M) -- (N);
\draw (M) -- (O) (M) -- (P) (M) -- (Q) (M) -- (R) (N) -- (O) (N) -- (P) (N) -- (Q) (N) -- (R);
\draw[double] (O) -- (P);
\draw (O) -- (Q) (O) -- (R) (P) -- (Q) (P) -- (R);
\draw[double] (Q) -- (R);
\draw[double] (H) -- (O);
\draw[double] (L) .. controls (4,-0.4) .. (Q);
\draw[double] (F) .. controls (4,-2) and (10,-2) .. (M);
\draw[double] (E) .. controls (-1, 3.5) and (8, 3.5) .. (N);
\draw[double] (I) .. controls (0,-2) and (8,-2) .. (R);
\draw[double] (G) .. controls (-2,-2) and (5,-2) .. (P);
\foreach \node in {A,B,C,D,E,F,G,H,I,L,M,N,O,P,Q,R}
\draw[fill=white] (\node) circle (2pt);
\end{tikzpicture}
\]
The following pencils (we omit here the images of the \(M\)-lines) are special pencils on \(Y\) whose pullbacks correspond to the elliptic fibrations on \(X\) with frames \(W_{19}\), \(W_{20}\), \(W_{24}\) and \(W_{26}\), respectively.
\[
\begin{tikzpicture}
\coordinate (A) at (0,0);
\coordinate (B) at (2,-0.5);
\coordinate (C) at (3,0.5);
\coordinate (D) at (1.4,2);
\draw[very thick] (A) -- (D) coordinate [midway] (E);
\draw (B) -- (C) coordinate [midway] (F);
\draw (A) -- (B) coordinate [midway] (G);
\draw[very thick] (C) -- (D) coordinate [midway] (H);
\draw (A) -- (C) coordinate [midway] (I);
\draw[very thick] (B) -- (D) coordinate [midway] (L);
\foreach \node in {A,B,C,D,E,L,H}
\draw[fill=black] (\node) circle (2pt);
\foreach \node in {I,G,F}
\draw[fill=white] (\node) circle (2pt);
\end{tikzpicture}
\quad
\begin{tikzpicture}
\coordinate (A) at (0,0);
\coordinate (B) at (2,-0.5);
\coordinate (C) at (3,0.5);
\coordinate (D) at (1.4,2);
\draw (A) -- (D) coordinate [midway] (E);
\path (B) -- (C) coordinate [midway] (F);
\path (A) -- (B) coordinate [midway] (G);
\draw (C) -- (D) coordinate [midway] (H);
\path (A) -- (C) coordinate [midway] (I);
\draw (B) -- (D) coordinate [midway] (L);
\draw[very thick,densely dashed] (A) -- (B);
\draw[very thick,densely dashed] (B) -- (C);
\draw[very thick,densely dashed] (C) -- (A);
\foreach \node in {A,G,B,F,C,I,D}
\draw[fill=black] (\node) circle (2pt);
\foreach \node in {E,L,H}
\draw[fill=white] (\node) circle (2pt);
\end{tikzpicture}
\quad
\begin{tikzpicture}
\coordinate (A) at (0,0);
\coordinate (B) at (2,-0.5);
\coordinate (C) at (3,0.5);
\coordinate (D) at (1.4,2);
\draw (A) -- (D) coordinate [midway] (E);
\draw (B) -- (C) coordinate [midway] (F);
\draw (A) -- (B) coordinate [midway] (G);
\draw (C) -- (D) coordinate [midway] (H);
\draw[very thick] (A) -- (C) coordinate [midway] (I);
\draw (B) -- (D) coordinate [midway] (L);
\draw[very thick] (E) -- (A) -- (G);
\draw[very thick] (H) -- (C) -- (F);
\foreach \node in {A,C,E,F,G,H,I,L}
\draw[fill=black] (\node) circle (2pt);
\foreach \node in {B,D}
\draw[fill=white] (\node) circle (2pt);
\end{tikzpicture}
\quad
\begin{tikzpicture}
\coordinate (A) at (0,0);
\coordinate (B) at (2,-0.5);
\coordinate (C) at (3,0.5);
\coordinate (D) at (1.4,2);
\draw[very thick] (A) -- (D) coordinate [midway] (E);
\draw[very thick] (B) -- (C) coordinate [midway] (F);
\draw[very thick] (A) -- (B) coordinate [midway] (G);
\draw[very thick] (C) -- (D) coordinate [midway] (H);
\draw (A) -- (C) coordinate [midway] (I);
\draw (B) -- (D) coordinate [midway] (L);
\foreach \node in {A,B,C,D,E,F,G,H}
\draw[fill=black] (\node) circle (2pt);
\foreach \node in {I,L}
\draw[fill=white] (\node) circle (2pt);
\end{tikzpicture}
\]
\end{remark}
\bibliographystyle{amsplain}
|
\section{Introduction}
\label{intro}
Modelling match outcomes in association football (referred to simply as "football" hereafter) undoubtedly represents an element of primary interest in the field of sports analysis. In order to do this, two different but interconnected strategies can be considered: the results-based (or direct) approach and the goals-based (or indirect) approach. Given a specific match between two competing teams, the former focuses on modelling the categorical ordinal variable taking the three possible result values (home win-draw-away win) typically through a regression model in which the probabilities of the three final outcomes are estimated on the basis of some external variables \citep[see, e.g.,][]{koning2000balance, goddard2004forecasting, schauberger2018predicting, carpita2019exploring, groll2019hybrid}. Instead, the goals-based approach considers a broader framework in which the purpose is to model the number of goals scored by the two teams during that specific match.
Since estimating the probability of each possible combination of home goal and away goal allows the estimate of the probability of home win-draw-away win to be consequently obtained, the two approaches are nested. The goals-based approach, being more general, has also some intriguing practical consequences, such as allowing types of bet different from traditional 1-X-2 (e.g. Under/Over). In addition, the difference in terms of performance between the two strategies is investigated by \citet{goddard2005regression} and \citet{koopman2019forecasting}: whereas the former highlights no relevant differences between them, the latter finds evidence that the goals-based approach provides more precise forecasts. In light of all these reasons, in the rest of the paper we will focus on the goals-based strategy.
The first articles in this field are \citet{moroneyj} and \citet{reep1971skill}, in which the Poisson distribution and the negative binomial distribution are proposed to model the aggregated number of goals scored per game. To the best of our knowledge, \citet{maher1982modelling} represents the first work aimed at modelling the number of goals scored by individual teams: specifically, the number of goals scored by the home team and the away team defines two independent Poisson random variables whose parameters depend on the attack and defence skills of the two teams. In the same paper, the author also carries out a bivariate Poisson model which keeps the two marginal distributions unchanged, driven by the fact that the initial proposal tends to underestimate the proportion of draws. In the nineteen-nineties two different models were proposed in the same year: \citet{lee1997modeling} keeps the general structure of \citet{maher1982modelling} assuming independence between the two Poisson random variables, whereas the article of \citet{dixon1997modelling} moves from the pioneering work of Maher by introducing some crucial innovations. First of all, the authors introduce a specific dependence structure by specifying a parameter $\rho$ which allows the joint probability to be different from the product of the marginal probabilities. Secondly, they include a weighting function $\phi$ which down-weights old matches in the likelihood in order to obtain estimates of the parameters that are mainly based on recent performances of the teams. In order to improve the aspect just mentioned, \citet{rue2000prediction} propose a Bayesian dynamic generalized linear model leaving the attack and defence parameters free to change randomly over time, while the aforementioned bivariate Poisson model is reconsidered and extended by \citet{karlis2003analysis}. A completely different approach is the one developed by \citet{baio2010bayesian}: in this case, their Bayesian hierarchical model considers two conditionally independent Poisson random variables for the numbers of goal scored, but the dependence is introduced through a careful choice of the hyper-parameters. A recent development of this approach is given by \citet{egidi2018combining}, where the betting odds are included in the model specification together with other modifications. \citet{owen2011dynamic} implements a dynamic generalized linear model whose evolution component is specified as a random walk for the attack and defence parameters, whereas \citet{mchale2011modelling} use copulas to allow dependence between the two Poisson random variables under the assumption that the dependence parameter can be expressed as a linear function of the rank difference of the two teams. As argued by the two authors, the copula model represents a more flexible solution than the bivariate Poisson since it also permits negative correlation. Another interesting proposal able to capture the main features mentioned so far is the one provided by \citet{koopman2015dynamic}, where a non-Gaussian state space model assumes a bivariate Poisson distribution whose attack and defence parameters are allowed to vary stochastically over time.
Taking into account the variety of approaches presented in the literature, it is evident that the definition of a proper dependence structure between the goals scored by the two teams represents an essential issue in modelling the final outcome of a football game. The aim of this work is to enrich the current literature
presenting a discussion of the common dependence assumptions of the models for football match outcomes and introducing an innovative method able to balance flexibility and conceptual simplicity.
Specifically, moving from the Dixon and Coles approach, which still represents a point of reference in the betting industry \citep[see, e.g., the methodology used at][]{mercurius} due to its conceptual simplicity, we introduce a more comprehensive dependence structure to improve the overall forecasting performance, while maintaining the easiness in the parameter interpretation and data generator process definition. In order to do that, we define the joint probability mass function describing the number of goals scored by the two teams by means of carefully chosen marginal probability mass functions (Mar-) and conditional probability mass functions (-Co). In so doing, the resulting model (Mar-Co) still manages the dependence between the home and the away goals by using an univariate parameter as in the Dixon and Coles approach.
The article is organised as follows: in Section \ref{sec::dependence} a broad discussion about the aforementioned dependence structure (focusing in particular on the one proposed by Dixon and Coles) is proposed; we present our adjustments to the Dixon and Coles model in Section \ref{sec::mar-co_mod}, whereas in Section \ref{sec::results} an application compares our proposal with the Dixon and Coles approach. Finally, Section \ref{sec::conclusion} provides some concluding remarks.
\section{The dependence assumption}
\label{sec::dependence}
\subsection{On the existence of dependence}
\label{sec::exis_dep}
Differently from the results-based strategy, the goals-based one requires a careful management of the dependence between the number of goals scored by the two competing teams, a notoriously tough task \citep{karlis2009bayesian}. Indeed, different studies draw very different conclusions depending on the championship, period of time and statistical tool considered. For example, \citet{karlis2000modelling} highlight a small positive correlation combining the evidence from 24 championships of different European countries by means of the method proposed by \citet{hasselblad199421}. In \citet{mchale2007modelling}, the correlations computed on the English Premier League data from August 2003 to March 2006 suggest that, in contrast to shots, the goals scored by the two teams show only slight positive or no correlation, whereas the work of \citet{mchale2011modelling} shows statistically significant negative correlation considering matches between national teams.
In this regard, a fundamental aspect needs to be clarified. Let $x$ be the number of goals scored by the home team in a generic match of a given championship and let $y$ be the number of goals scored by the opposing team (i.e. the away team). Similarly, let us define $x_{ij}$ and $y_{ij}$ the number of goals scored by teams $i$ and $j$ respectively in the specific match in which $i$ plays at home and $j$ plays away. Variables $x$ and $y$ are different random objects from $x_{ij}$ and $y_{ij}$ (although they are interconnected somehow), and so the development of an empirical study assessing the dependence between $x$ and $y$ is of limited practical use since the purpose is to model the team-specific variables $x_{ij}$ and $y_{ij}$. For instance, one is justified in expecting a team scoring many goals during a game to concede a low number of goals (i.e. negative correlation between $x$ and $y$) since in that specific occasion it totally outclasses the opposing team. On the other hand, a different dynamic can occur when two specific teams - characterised by specific attack and defence skills - play against each other: for example, a strong team playing against a weak one may score less goals than those expected since it mainly focuses on conserving energy for future more challenging matches. A discussion on the topic is also provided by \citet{dixon1998birth} and \citet{rue2000prediction}.
However, the crucial empirical study of the dependence between $x_{ij}$ and $y_{ij}$ is practically unfeasible since each couple of teams plays against just twice during a season (and only once team $i$ plays at home and team $j$ away). Considering more than one season represents an intuitive solution, but it seems to be unsuitable since the strength of a team can vary widely from season to season (due to newly signed players, the sacking of a manager, etc\dots). In view of this, in the rest of Section \ref{sec::exis_dep} we will focus on the dependence between the aggregated variables $x$ and $y$ in order to have an indication about the general behaviour, but all the results provided should be approached with particular caution in light of the considerations just expressed.
In order to investigate the dependence, the data we consider---obtained from \url{http://www.football-data.co.uk/}---refers to the 9130 matches of the English Premier League, the French Ligue 1, the German Bundesliga, the Italian Serie A and the Spanish La Liga played between the 2014-2015 season and the 2018-2019 season. A first, fundamental study concerns the existence of the most familiar kind of dependence---i.e. correlation---between $x$ and $y$. To do so, we consider a bootstrap test aimed at evaluating the Pearson correlation coefficient with the samples drawn under the hypothesis of independence.
One may argue that Pearson correlation is specific to investigate linear correlation and it could be not suitable with count data with possible zero inflation presence. However, testing the Pearson coefficient value is sufficient to have indication against independence if linear correlation is observed in the data.
The observed value of the Pearson correlation coefficient (computed considering all the 9130 matches) is $-0.085$ and the p-value is $<0.001$. The same strong evidence against the hypothesis of independence can be found considering each championship separately, with the only small exception of the Italian Serie A, whose p-value is equal to $0.026$. This result confirms the evidence provided by \citet{mchale2011modelling} and consequently the importance of a proper modelling of the dependence between the number of goals scored during a match.
\subsection{The Dixon and Coles dependence structure}
\label{sec:dix_col_dep_str}
Among the various goals-based approaches proposed in this field, the model introduced by \citet{dixon1997modelling} represents one of the most famous, innovative and performing. Since the model proposed in this paper moves from it, an overview of the Dixon and Coles approach is provided as follows.
Let $x_{ij}^k$ be the number of goals scored by the home team $i$ against the away team $j$ in match $k$, $y_{ij}^k$ be the number of goals scored by the away team $j$ on the same occasion and let $m$ be the total number of teams considered. In order not to overcomplicate the notation, hereafter $x_k$ ($y_k$ respectively) will be used instead of $x_{ij}^k$ ($y_{ij}^k$ respectively) with the home team $i$ and the away team $j$ implicitly assigned to every match $k$.
First of all, the two authors define $x_k$ and $y_k$ as follows:
\begin{eqnarray*}
x_k &\sim& \text{Pois}(\lambda_k), \label{eq:mar-pois} \\
y_k &\sim& \text{Pois}(\mu_k), \nonumber \\
\log(\lambda_k) &=& \gamma + \alpha_{i(k)} + \beta_{j(k)}, \nonumber\\
\log(\mu_k) &=&\alpha_{j(k)} + \beta_{i(k)}, \nonumber
\end{eqnarray*}
with $\text{Pois}(\lambda)$ denoting a Poisson distribution with mean $\lambda$, $i(k)$ and $j(k)$ indices which identify the home and away teams playing match $k$, and where $\alpha$, $\beta$, $\gamma$ are the attack, defence and home effect parameter, respectively.
By considering the matches chronologically, the two authors divide the seasons into a series of half-weekly time points and construct the following function for each time point $t$:
\begin{align}
\label{eq::pseudoveros_DC}
\mathcal{L}_t(\alpha_i,\beta_i,\rho,\gamma; i=1,\dots,m)&=\prod_{k \in A_t} \biggr\{ \tau_{\lambda_k, \mu_k}(x_k, y_k) \, e^{-\lambda_k} \, \lambda_k^{x_k} \, e^{-\mu_k} \, \mu_k^{y_k} \biggr\} ^{e^{-\xi \, (t-t_k)}}, \\
A_t &= \{ k: t_k < t \}, \nonumber
\end{align}
with $t_k$ the time that game $k$ is played, $\tau_{\lambda_k,\mu_k}(x,y)$ the function depending on parameter $\rho$ which manages the dependence between $x_k$ and $y_k$, and $\xi \geq 0$ the parameter that regulates the down-weighting of old matches. Consistently with the notation of the original paper, the constraint $\sum_{i=1}^m \alpha_i=m$ is included for identifiability
Hence, the two authors obtain the estimates of the parameters by numerically maximising the function in Equation (\ref{eq::pseudoveros_DC}) at each time point $t$ after choosing $\xi$.
The choice of $\xi$ is particularly tough because Equation (\ref{eq::pseudoveros_DC}) defines a sequence of non-independent functions that makes difficult to obtain the value of $\xi$ that maximises the overall predictive capability of the model. In order to overcome this problem, the two authors focus on the prediction of match outcomes rather than match scores and define the value of $\xi$ as the value maximising
\begin{equation*}
S(\xi)=\sum_{k} \left(\delta_{k}^{H} \log p_{k}^{H}+\delta_{k}^{D} \log p_{k}^{D}+\delta_{k}^{A} \log p_{k}^{A}\right)
\end{equation*}
with
\begin{equation*}
p_{k}^{r}=\sum_{h, a \in B_{r}} \operatorname{pr}\left(x_{k}=h, y_{k}=a\right),
\end{equation*}
which is implicitly a function of $\xi$ since the score probabilities $\operatorname{pr}\left(x_{k}=h, y_{k}=a\right)$ are estimated from the maximisation of the function in (\ref{eq::pseudoveros_DC}) at $t_k$ with weighting parameter set at $\xi$ and with $r=\{H,D,A\}$, $B_{H}=\{(h, a): h>a\}, B_{D}=\{(h, a): h=a\}, B_{A}=\{(h, a): h<a\}$, and $\delta_{k}^{H}$ ($\delta_{k}^{D}$, $\delta_{k}^{A}$ respectively) the delta function equal to 1 if the final result of game $k$ is a home win (draw, away win respectively).
The last, fundamental aspect introduced by Dixon and Coles concerns the description of the dependence structure, that is defined by means of the function
\begin{equation}
\tau_{\lambda_k,\mu_k}(x,y)=\begin{cases}
1-\lambda_k \mu_k \rho & \text{if $x=0$, $y=0$} \\
1+\lambda_k \rho & \text{if $x=0$, $y=1$} \\
1+\mu_k \rho & \text{if $x=1$, $y=0$} \\
1- \rho & \text{if $x=1$, $y=1$} \\
1 & \text{otherwise}
\end{cases}
\label{eq::function_tau}
\end{equation}
subject to
\begin{equation*}
\max \biggr(-\frac{1}{\lambda_k},-\frac{1}{\mu_k}\biggr) \leq \rho \leq \min\biggr( \frac{1}{\lambda_k \mu_k},1\biggr).
\end{equation*}
By using function (\ref{eq::function_tau}), the marginal distributions of $x_k$ and $y_k$ are Poisson with means $\lambda_k$ and $\mu_k$ respectively and the independence between $x_k$ and $y_k$ is obtained when $\rho=0$.
\begin{figure}
\centering
\makebox{\includegraphics[width=0.99\textwidth]{Fig1.eps}}
\caption{\label{fig:dc}Each panel shows, under the Dixon and Coles model, the difference between the probability of the exact outcomes in match $k$, given a certain value of $\rho$ (equal to -0.1 on the left, 0 in the middle and 0.1 on the right), and the probability of the exact outcomes in match $k$ in case of independence (i.e. $\rho=0$). The marginal means $\lambda_k$ and $\mu_k$ are set equal to 1 and 1.5, respectively. }
\end{figure}
Figure \ref{fig:dc} provides an intuitive representation of the dependence structure, highlighting the difference between the joint probability mass functions of the goals scored assuming independence and dependence.
The only outcomes affected by a change in probability, when $\rho$ varies, are those in the black boxes of the picture. This means that the probability that one team scores at least two goals does not depend on the number of goals scored by the opposing team in the same match, since
$$
\sum_{h=0}^1 \sum_{a=0}^1 \text{pr}(x_k=h, y_k=a) = \sum_{h=0}^1 \sum_{a=0}^1 \text{pr}(x_k=h) \text{pr}(y_k=a)
$$
holds.
In other terms, the Dixon and Coles model is based on the strong assumption that the number of goals scored by the home team is independent of the number of goals scored by the away team, conditionally on observing an outcome not included in the set $\{(0,0), (1,0), (0,1), (1,1)\}$.
Looking at the 9130 matches played during the period from August 2014 to May 2019 in the top five European championships, this assumption seems no longer reasonable, as suggested by testing it as null hypothesis in the following bootstrap procedure.
Let $\tilde{x}_k$ and $\tilde{y}_k$ denote the home and away number of goals considering scoring either zero or one goal as unique event---i.e. $\tilde{x}_k=\max(1,x_k)$, $\tilde{y}_k=\max(1,y_k)$. Hence, under the null hypothesis, the variables $\tilde{x}_k$ and $\tilde{x}_k$ are independent. To perform the test, we generated 1000 data sets with $\tilde{x}_k$ and $\tilde{y}_k$ independently re-sampled and we computed the sample correlation between the two variables in each data set. The negative correlation computed on the observed data set lies in the left tail of the bootstrap correlation distribution, encouraging the rejection of the null hypothesis with a p-value $<0.001$.
\begin{table}
\caption{\label{tab:rep_dc} Ratios between the observed frequencies of 25 possible match outcomes and the joint probability function obtained by multiplying the observed marginal probability functions of home and away scores. Bootstrap standard errors are reported in parentheses, while grey cells indicate ratios that are significantly different from 100 at level $0.05$. To facilitate the reading, the numbers are multiplied by 100.}
\fbox{%
\begin{tabular}{crrrrr}
Home& \multicolumn{5}{c}{Away goals}\\
goals & 0 & 1 & 2 & 3 & 4 \\
\hline
0 & 99.47 (2.25) & \cellcolor{light-gray} 92.95 (2.02) & 100.68 (3.10) & \cellcolor{light-gray} 114.96 (5.13) & \cellcolor{light-gray} 139.57 (9.53) \\
1 & \cellcolor{light-gray} 94.46 (1.77) & 102.11 (1.65) & 102.75 (2.35) & 103.21 (3.91) & 107.72 (6.86)\\
2 & 99.01 (2.10) & 103.40 (1.98) & 102.69 (2.92) & \cellcolor{light-gray} 89.23 (4.55) & \cellcolor{light-gray} 84.39 (7.75) \\
3 & \cellcolor{light-gray} 108.13 (3.18) & 101.23 (2.95) & \cellcolor{light-gray} 91.24 (4.12) & 99.11 (7.13) & \cellcolor{light-gray} 59.27 (10.11)\\
4 & \cellcolor{light-gray} 121.20 (5.29) & 98.78 (4.73) & 88.63 (6.65) & \cellcolor{light-gray} 68.19 (9.49) & \cellcolor{light-gray} 52.58 (14.97)
\end{tabular}}
\end{table}
As a further evidence against the Dixon and Coles dependence structure, we replicated on the above mentioned recent data the study reported in Section 3 of \citet{dixon1997modelling}. Table \ref{tab:rep_dc} reports the ratio $f_{x,y}(x_k,y_k)/\{f_x(x_k)f_y(y_k)\}$, where $f_{x,y}(x_k,y_k)$, $f_x(x_k)$, and $f_y(y_k)$ are the joint and marginal observed probability mass functions for the home and away number of goals, respectively. We notice that nowadays, unlike what observed by Dixon and Coles at the time, the frequency $f_{x,y}(x_k,y_k)$ of several match outcomes is significantly different from the product of the marginal frequencies of goals scored by home and away teams $f_x(x_k)f_y(y_k)$.
These empirical results, together with the evidence that a kind of dependence between the home and away goals exists, suggest that a new, more complex dependence structure should be considered.
Furthermore, the Dixon and Coles model aims to properly estimate the probabilities of home win-draw-away win, which are directly influenced by the dependence structure proposed, since each of these three probabilities changes according to $\rho$. Vice versa, such dependence structure cannot adjust the probability of events defined by other types of bet, as we will see in details in Section \ref{sec::results}.
Motivated by these considerations, in the next section, starting from the Dixon and Coles model, we propose a new model characterised by a more general dependence structure, such that the induced joint probability mass function is different from the product of the marginal probability mass functions for each possible match outcome.
\section{The Mar-Co model}
\label{sec::mar-co_mod}
Inspired by \citet{berkhout2004bivariate} and moving from the considerations discussed in the previous section, we propose a new model which differs from the Dixon and Coles one by the choice of the dependence structure between $x_k$ and $y_k$.
Given a certain marginal distribution of $x_k$, an intuitive approach to allow the specification of a sufficiently general dependence structure consists in modelling the conditional distribution $y_k \mid x_k$.
By indicating, with a slight abuse of notation, $F_{\mu_k}(x)$
as the cumulative distribution function evaluated at $x$ of a Poisson random variable with mean $\mu_k$---i.e. $F_{\mu_k}(x)=\exp\{-\mu_k\} \sum_{i=0}^x( \mu_k^i/i!) \; \forall \, x \in \mathbb{N}_{0}$--
we specify
\begin{eqnarray*}
(y_{k} \mid x_{k} = h) &\sim& \text{Pois}\{\psi_{y}(\mu_{k},h)\}, \qquad \text{with}\\
\psi_{Y}(\mu_{k},h) &=& \exp[\theta_{1} + \theta_{2}\log{\mu_{k}} + \theta_{3} \text{logit}\{F_{\mu_k}(h)\}],
\end{eqnarray*}
where logit$(p)=\log\{p/(1-p)\}$ and $\boldsymbol{\theta}=(\theta_1, \theta_2, \theta_3)^\top \in \mathbb{R}^{3}$.
From an interpretative point of view, the mean of $y_k \mid x_k$ depends on the ability of the away team, characterized by $\mu_{k}$, to score no more goals than those scored in that occasion by the home team.
As in the Dixon and Coles model, an univariate parameter ($\theta_3$ in this case) regulates the dependence between $x_k$ and $y_k$, where $\theta_{3} \neq 0$ implies dependence existence.
In addition, it is possible to obtain the above mentioned independence case corresponding to the Dixon and Coles model with $\rho=0$ by assuming $\boldsymbol{\theta}=(0,1,0)^\top$ and by modelling $x_k$ as the Poisson random variable defined in (\ref{eq:mar-pois}).
Hence, assuming $x_k \sim \text{Pois}(\lambda_k)$, we can exploit the definition of joint probability mass function as product of conditional and marginal probability mass functions to define
\begin{equation}
\label{eq:P_A}
\text{pr}_A(x_k=h, y_k=a)=\frac{e^{-\psi_y(\mu_{k},h)} \psi_y(\mu_{k},h)^a}{a!} \frac{e^{-\lambda_{k}} \lambda_{k}^h}{h!}.
\end{equation}
On the other hand, if we follow symmetrical steps specifying
\begin{eqnarray*}
y_k &\sim& \text{Pois}(\mu_k),\\
(x_k\mid y_k =a) &\sim& \text{Pois}\{\psi_{x}(\lambda_{k},a)\},\\
\psi_{x}(\lambda_{k},a) &=& \exp[\theta_{1} + \theta_{2}\log{\lambda_{k}} + \theta_{3} \text{logit}\{F_{\lambda_k}(a)\}],
\end{eqnarray*}
we can define the symmetrical joint probability mass function
\begin{equation}
\label{eq:P_B}
\text{pr}_B(x_k=h, y_k=a)=\frac{e^{-\psi_x(\lambda_{k},a)} \psi_x(\lambda_{k},a)^h}{h!} \frac{e^{-\mu_{k}} \mu_{k}^a}{a!}.
\end{equation}
As we do not have formal reasons to favour one of the two specifications, we define the distribution of the joint outcome $(x_k, y_k)$ as an equally weighted mixture of the two distributions defined by $\text{pr}_A$ and $\text{pr}_B$. Therefore, under the Mar-Co model, the likelihood related to the $k$-th match can be expressed as
\begin{eqnarray*}
\mathcal{L}_k^{M}(\alpha_i, \beta_i, \alpha_j, \beta_j, \gamma, \boldsymbol{\theta}) &\propto& \text{pr}_A(x_k=h, y_k=a; \alpha_i, \beta_i, \alpha_j, \beta_j, \boldsymbol{\theta}, \gamma) \\
&\quad& + \text{pr}_B(x_k=h, y_k=a,\alpha_i, \beta_i, \alpha_j, \beta_j, \boldsymbol{\theta}, \gamma ). \nonumber
\label{eq:Lt}
\end{eqnarray*}
As a merely football-based argument in favour of this model construction,
the presence of two symmetrical data generation processes appears to be plausible, since one is justified in expecting sometimes the away team to react to the performance of the home team, sometimes vice versa.
\begin{figure}
\centering
\makebox{\includegraphics[width=0.99\textwidth]{Fig2.eps}}
\caption{\label{fig:marco} Each panel shows, under the Mar-Co model, the difference between the probability of the exact outcomes in match $k$, given a certain value of $\theta_3$ (equal to -0.05 on the left, 0 in the middle and 0.05 on the right), and the probability of the exact outcomes in match $k$ in case of independence (i.e. $\theta_3=0$). The marginal means $\lambda_k$ and $\mu_k$ are set equal to 1 and 1.5, respectively, while $\theta_1=0$ and $\theta_2=1$. }
\end{figure}
The key difference between the Mar-Co model and the Dixon and Coles one lies in the specification of a joint distribution that permits $x_k$ and $y_k$ to be dependent also conditionally on observing an outcome not included in the set $\{(0,0),(1,0),(0,1),(1,1)\}$. This aspect is clearly displayed in Figure \ref{fig:marco}, where the grey scale identifies the difference between the Mar-Co joint probability mass function considering three possible values of $\theta_3$ and the probability mass function obtained when $\theta_3=0$. A comparison of the left and right panels of Figure \ref{fig:marco} with the corresponding panels in Figure \ref{fig:dc} immediately highlights the discrepancy between the two models in terms of outcomes affected by the introduction of dependence.
Formally,
\begin{align*}
\sum_{h =0}^1 \sum_{a=0}^1 \text{pr}(x_k=h, y_k=a) &=
\sum_{h =0}^1 \sum_{a=0}^1 \frac{\text{pr}_A(x_k=h, y_k=a) + \text{pr}_B(x_k=h, y_k=a)}{2} \\
&\neq \sum_{h =0}^1 \sum_{a=0}^1\text{pr}(x_k=h) \text{pr}(y_k=a),
\end{align*}
when $\theta_3 \neq 0$.
The parameter $\theta_3$ presents also an immediate interpretation, which can be easily described through well known team behaviours unlike the parameter $\rho$ in the Dixon-Coles models or other complicate dependence structures. Indeed, $\theta_3>0$ means that teams tend to adapt their attacking performance to the opponent performance such that several tight matches are generated. In other words, teams tend to increase their attacking level against well-performing teams and to lower it when big attacking efforts are not necessary to win. Vice versa, if $\theta_3<0$, teams tend to have better (worse) performance than their \textit{marginal} behaviour when they face teams playing poorly (very well). In fact, both these behaviours are likely and they generally depend on the football culture of a team or, more often, of a country, as will be discussed in Section \ref{sec::results}.
Relying on distinct parameter estimation processes for the different leagues, one may interpret the estimates of $\theta_3$ as an average league measure of the competitive balance within the matches, with $\theta_3$ positive indicating a league where matches are often tighter than what expected by looking at the marginal strength of the teams involved.
A careful analysis of the properties and a proper definition of $\theta_3$ as an index of competitiveness may lead to the inclusion of a further dimension in the framework of competitive balance measurements, a field that is gaining large attention from researchers in recent years \citep{manasis2014, manasis2022}.
Although the latter objective is beyond the aim of this work, the potentialities of our proposed dependence structure in terms of interpretation represent a further distinct and interesting trait with respect to other more complex models.
On the face of a more comprehensive dependence structure, the Mar-Co model does not present the desired properties in terms of marginal distributions that characterise the Dixon and Coles model.
Indeed, in case of dependence under the Mar-Co model, both the conditional and the marginal distributions of $x_k$ and $y_k$ are not Poisson, since each of these distributions can be described as the mixture between a Poisson and a non-Poisson distribution.
When $\theta_3 \neq 0$, the marginal distributions $x_k$ and $y_k$ are characterised by non finite moments, since they linearly depends on the non finite expected values $E([F_{\lambda_k}(y)/\{1-F_{\lambda_k}(y)\}]^{\theta_3})$ or $E([F_{\mu_k}(y)/\{1-F_{\mu_k}(x)\}]^{\theta_3})$, where $F_{\lambda_k}(y)$ and $F_{\mu_k}(x)$ are uniformly distributed in $(0,1)$. Although this fact is not appealing in terms of property demonstrations, it suggests a robust behaviour of the marginal distributions with respect to large results, due to the fat tails.
The non finite first moments of the marginal distributions does not allow one to analytically recover the correlation between home and away goals. Then, the induced correlation between $x_k$ and $y_k$ at varying of $\theta_3$ is reported in Figure \ref{fig:correlation} via Monte Carlo estimates of the Spearman correlation coefficient, which is robust to non finite moments.
\begin{figure}
\centering
\includegraphics[width=0.99\textwidth]{Fig7.eps}
\caption{Monte Carlo estimates of the Spearman correlation coefficient at varying of $\theta_3$, with $\theta_{1}=0$, $\theta_{2}=1$, $\mu_k=1.75$, and $\lambda_{k}=1.25$. The estimates are based on $250$ computations of sample Spearman coefficient of $1000$-large samples of possible match outcomes.}
\label{fig:correlation}
\end{figure}
The lack of Poisson marginal distributions appears in contrast with one of the most widely accepted assumptions in the literature. As mentioned in the Introduction, several authors assume Poisson marginal distributions on the basis of the empirical behaviour of the marginal home and away number of goals in a championship \citep[see, e.g., ][]{dixon1997modelling}.
Nevertheless, a parametric bootstrap test conducted on the same data presented in Section \ref{sec::exis_dep} questions the validity of this assumption, suggesting to reject the null hypothesis of Poisson-distributed variables $x$ and $y$.
In details, we compared the ratio between the sample variance and the sample mean---i.e. an estimate of the dispersion parameter---computed on the observed data with those computed on 1000 data sets drawn from the Poisson distribution with mean equal to the observed sample mean. In so doing, we observed a p-value $<0.001$ both for the home and the away number of goals.
By replicating the test on each championship separately, we obtained similar strong indications against the null hypothesis for each marginal distribution considered. In addition, the same evidence, although slighter, was obtained by performing an alternative parametric bootstrap test.
This later test is based again on 1000 data sets drawn from the Poisson distribution with mean equal to the observed sample mean and it considers as test statistic the Kullback-Leibler divergence between the empirical distribution of each sample and the theoretical Poisson distribution used to generate the samples.
Following the considerations discussed in Section \ref{sec::exis_dep}, a further crucial element should be taken into account.
The studies of the marginal behaviour of the number of goals, such as that reported above, generally investigate the properties of the distributions of $x$ and $y$, i.e. the number of goals observed in a generic match of a championship, while the goals-based models usually make assumptions on the distributions of $x_{ij}$ and $y_{ij}$, i.e. the number of goals scored in a match between two specific teams $i$ and $j$. The article of \citet{karlis2000modelling} represents a progress in this discussion, since the authors analyse the dispersion parameter of the distribution of the number of goals in matches played by a given team $i$, i.e. $x_{i}$ and $y_i$,
providing an evidence that we can expect to be more similar to that we could obtain by analysing the distribution of $x_{ij}$ and $y_{ij}$. The study, conducted on 456 teams in 24 different European leagues, suggests the presence of overdispersion, providing a further element against the hypothesis of Poisson-distributed variables $x_{ij}$ and $y_{ij}$.
Therefore, although we cannot conclude that $x_{ij}$ and $y_{ij}$ are undoubtedly non-Poisson random variables, it is hard to argue that the assumption of marginal Poisson distributions is usually a convenient choice in terms of modelling, rather than a consequence of well established behaviours.
In view of this perspective, it could be worth having non-Poisson marginal distributions, if it allows to include a more flexible dependence structure able to improve the overall predictive capability of the model, as we try to demonstrate via an application study reported in the next section.
As regards the estimation process, let $\boldsymbol{\alpha}$ and $\boldsymbol{\beta}$ denote the vectors including the parameters $\alpha_i$ and $\beta_i$ for $i=1,\ldots,m$.
The estimates of the parameters at time point $t$ are obtained by maximising the function
\begin{equation}
\mathcal{L}_t^{M}(\boldsymbol{\alpha},\boldsymbol{\beta},\gamma, \boldsymbol{\theta})=\prod_{k \in A_t} \biggr\{ \text{pr}_A(x_k, y_k; \boldsymbol{\alpha}, \boldsymbol{\beta}, \gamma, \boldsymbol{\theta}) + \text{pr}_B(x_k, y_k;\boldsymbol{\alpha}, \boldsymbol{\beta}, \gamma, \boldsymbol{\theta} )\biggr\}^{e^{-\xi \, (t-t_k)}}, \nonumber
\label{eq::pseudoveros_M}
\end{equation}
where $\text{pr}_A$ and $\text{pr}_B$ refer to the functions in (\ref{eq:P_A}) and (\ref{eq:P_B}) respectively, while the set $A_t$ and the parameter $\xi$ are defined consistently with the notation of the Dixon and Coles model presented in Section \ref{sec:dix_col_dep_str}.
Given the shape of the mixture distribution, the presence of local modes cannot be excluded and, at the same time, the maximisation algorithms could encounter computational troubles in evaluating $\mathcal{L}_t^M$ at points that are far from the mode.
\begin{comment}
In view of these issues, we propose the following maximisation procedure.
Firstly, we obtain a first estimate of the parameters through a plug-in estimation process \citep{gongsamaniego}.
Specifically, after computing $\hat{\alpha}^{(0)}, \hat{\beta}^{(0)}, \hat{\gamma}^{(0)}$ by maximising $\mathcal{L}_t^{M}(\alpha,\beta,\gamma, \theta=(0,1,0)^T)$ with $\theta$ fixed---which is equivalent to maximise the Dixon and Coles function $\mathcal{L}_t$ with $\rho$ set to 0---we find $\hat{\theta}^{(0)}$ as the argument that maximises $\mathcal{L}_t^{M}(\hat{\alpha}^{(0)},\hat{\beta}^{(0)},\hat{\gamma}^{(0)}, \theta)$ over $\mathbb{R}^3$.
Secondly, we obtain the final estimates $\hat{\alpha}$, $\hat{\beta}$, $\hat{\gamma}$, and $\hat{\theta}$ by optimising the function $\mathcal{L}_t^{M}$ in the neighbourhood of $(\hat{\alpha}^{(0)},\hat{\beta}^{(0)},\hat{\gamma}^{(0)}, \hat{\theta}^{(0)})$, which is used as the starting point of the maximisation step.
Such algorithm, although unconventional, presents advantages both from a computational and an interpretative point of view. Indeed, it avoids computational troubles related to the evaluation of $\mathcal{L}_t^{M}$ where such function assumes low values, and, at the same time, it allows an easy interpretation of the estimates of the team fixed effects, since they are close to those obtained under the independence assumption.
\end{comment}
However, empirical results suggest that the maximisation procedure followed in this work generally leads to stable and satisfying results.
\section{An application to European leagues}
\label{sec::results}
In this section, a detailed comparison between the model presented in Section \ref{sec::mar-co_mod} and the Dixon and Coles one is provided. The two models are evaluated on a real data application which considers the five most important European leagues (the English Premier League, the French Ligue 1, the German Bundesliga, the Italian Serie A, the Spanish La Liga) and different betting types. Although sports betting is not the primary tool to assess the performance of a model, in this specific case it is particularly useful since it allows to compare the two models in different frameworks and to highlight strengths and weaknesses
of each approach.
First of all, we focus on the most famous betting type, i.e. Home Win-Draw-Away Win (known simply as 1-X-2).
In order to compare the two models, the first step is the choice of the parameter $\xi$, that is made for each combination of league and model separately by means of the procedure proposed by Dixon and Coles and described in Section \ref{sec:dix_col_dep_str}. In fact, consistently with the work of \citet{diquigiovanni2019analysis}, hereafter the time point $t$ will refer to the specific day of the year on which a given match takes place. This is due to the fact that nowadays, unlike when Dixon and Coles carried out the study, the teams play almost every day and so a more precise subdivision of the season is required. Specifically, the matches of three consecutive seasons (2012-2013, 2013-2014, 2014-2015) are used to choose $\xi$ and the probabilities of home win-draw-away win are not estimated for matches played between May and September in order to include promoted teams and to avoid misleading results due to lack of effort of some teams in the final part of the season. In so doing, the estimates of the probabilities are obtained starting from the first game played in October 2013 on the basis of all the information available at that time and $S(\xi)$ is computed considering that game together with all the subsequent matches.
Table \ref{tab:opt_xi} displays the values of $\xi$ maximising $S(\xi)$ based on a grid search, with $\xi>0$ indicating exponentially decreasing importance of a match over the time.
\begin{table}
\caption{\label{tab:opt_xi}Values of $\xi$ maximising $S(\xi)$ for each combination of league (columns) and model (rows).}
\centering
\fbox{%
\begin{tabular}{lccccc}
arg$\max_{\xi}$ $S(\xi)$ &England & France & Germany & Italy & Spain \\
\hline
Dixon and Coles Model & 0 & 0 &0.0046&0.0053&0.0021 \\
Mar-Co Model & 0 & 0 &0.0025&0.0045&0.0021 \\
\end{tabular}}
\end{table}
The results seem to show on the one hand no relevant dissimilarities between the two models, and on the other different dynamics in the leagues considered.
Although it is not an aspect of primary interest, this evidence suggests a fascinating insight into European leagues: as the greater the value of $\xi$, the less importance is given to the oldest matches, then the current physical and psychological condition of the two competing teams seems to particularly affect the outcome of a match in the German Bundesliga and the Italian Serie A.
Once the value of $\xi$ is set, we use the Ranked Probability Score \citep[or RPS;][]{epstein1969scoring} to compare the performance of the two models.
Despite the well known limitations of it \citep{wheatcroft2019evaluating}, RPS still represents one of the most famous and used scoring rules in this field due to its conceptual simplicity and easiness of implementation.
Formally, the RPS related to a specific match and betting type is defined as
\begin{equation*}
\frac{1}{r-1} \sum_{i=1}^{r-1}\left(\sum_{j=1}^{i}\left(\hat{p}_{j}-e_{j}\right)\right)^{2}
\end{equation*}
with $\hat{p}_j$ the estimate of the probability of the $j$-th possible betting outcome and $e_j=1$ if the outcome $j$ is observed and 0 otherwise. In the 1-X-2 framework, we therefore obtain: $r=3$, $\hat{p}_1$ ($\hat{p}_2$, respectively) equal to the estimate of the probability of home win (draw, respectively) and $e_1=1$ ($e_2=1$, respectively) if the final result of the match is a home win (draw, respectively) and 0 otherwise.
A detailed description of the RPS and its properties can be found in \citet{constantinou2012solving} and references therein.
\begin{figure}
\includegraphics[width=0.99\textwidth]{Fig3.eps}
\caption{Cumulative difference between the RPSs computed for the two competing models. Betting type: 1-X-2.}
\label{fig:1}
\end{figure}
We compute the RPSs for the matches of the 2015-2016, 2016-2017, 2017-2018, 2018-2019, 2019-2020 (until interruption due to the COVID-19 outbreak) seasons.
As for the choice of $\xi$, the first RPSs are obtained starting from the second season and they are computed only for matches played between October and April.
Figure \ref{fig:1} shows the cumulative difference over time between the RPSs computed for the Dixon and Coles model and the RPSs computed for the Mar-Co model considering the matches of all the leagues
(solid blue line) and considering each league separately (dashed coloured lines). Since a smaller RPS indicates better predictive performance, values
above the horizontal white line suggest that the Mar-Co model outperforms the Dixon and Coles one, and vice versa. The trend over time seems to be quite promising: after an initial period characterised by some fluctuations, the Mar-Co model provides more accurate predictions than those outputted by the Dixon and Coles one considering both all the leagues together and each league separately, with the sole exception of the French Ligue 1. Although the overall evidence seems to suggest that the modification introduced by the Mar-Co model can be used profitably in the long run, the analysis of Figure \ref{fig:1} should be accompanied by a study that verifies whether the two models are statistically different or not. To do that, for every match considered, the RPS
computed for the Mar-Co model and the RPS
computed for the Dixon and Coles one are switched with probability $0.5$ in order to obtain a new sample. By replicating this procedure $n_{b}$ times, it is possible to compute the $n_{b}$ mean differences of the reshuffled samples: in so doing, if the evidence in favour of the Mar-Co model is frequently stronger than the observed one, we conclude that the observed difference between the two models is mainly due to chance. By applying this procedure with $n_{b}=10000$, we obtain that 14\% of the time a stronger evidence is obtained: we conclude that, although it cannot be excluded that the two models are identical in terms of performance, the Mar-Co model seems to represent a promising modification of the starting model.
\begin{table}
\caption{\label{tab:theta} Estimates $\hat{\theta}_3$ for each league and the related bootstrap $95\%$ confidence intervals (CI) based on $250$ replications.}
\centering
\fbox{%
\begin{tabular}{lrl}
Ligue &$\hat{\theta}_3$ & $95\%$ CI \\
\hline
England & $-0.013$ & $(-0.038, 0.015)$ \\
France & $-0.027$ & $(-0.051, 0.000)$ \\
Germany & $-0.078$ & $(-0.119, -0.040 )$ \\
Italy & $0.003$ & $(-0.044, 0.043 )$ \\
Spain & $0.015$ & $(-0.011, 0.046)$ \\
\end{tabular}}
\end{table}
Since the two models differ with regard to their dependence structure, parameter $\theta$ represents the key element in determining the satisfactory results obtained by the Mar-Co model. As a consequence, the estimate $\hat{\theta}_3$ is a quantity of deep interest that allows to discover fundamental league-specific dynamics. In Table \ref{tab:theta}, we report for every league the estimates obtained at the end of the period considered and the related bootstrap $95\%$ confidence intervals.
The estimate for the Italian Serie A seems to indicate independent behaviours of home and away team performances. On the other hand, the Spanish La Liga is characterised by a positive value of $\hat{\theta}_3$. This evidence seems to confirm the hypothesis made in Section \ref{sec::exis_dep}, namely that the home (away) team tends to underperform when it plays against a team able to score less goals $y$ than a \textit{median} performance of the home (away) team (i.e., $F_{\lambda_k}(y)<0.5$).
Correspondingly, a home (away) team tends to overperform when it plays against a team able to score many goals (e.g., $F_{\lambda_k}(y)>0.5$), thus creating a more balanced match. Conversely, the opposite dynamics characterise the French Ligue 1, the German Bundesliga, and partially the English Premier League.
Such dynamics seem to be mirrored in the common opinion according to which is easy to observe in the French Ligue 1 and the German Bundesliga more unbalanced matches than what expected by looking at the \textit{marginal} value of the teams.
In view of this, the ability to properly manage these different behaviours according to the specific league considered seems to represent the key feature of the Mar-Co model.
Another betting type able to highlight the differences in terms of dependence modelling between the two approaches is the so-called Under/Over bet. In this case, the aim is to predict whether the overall number of goals scored during a game will be less (Under) or greater than (Over) a certain threshold. For the sake of clarity, Figure \ref{fig:bett}
\begin{figure}
\includegraphics[width=0.99\textwidth]{Fig4.eps}
\caption{Combinations of home and away goals characterising each betting outcome. The black boxes highlight the four match outcomes whose probabilities depend on $\rho$ in the Dixon and Coles model.}
\label{fig:bett}
\end{figure}
shows the combinations of home and away goals characterising the two possible betting outcomes when the threshold is set equal to 1.5 (in the middle) and 2.5 (on the right). This type of bet is particularly interesting since, as noted in Section \ref{sec:dix_col_dep_str}, the Dixon and Coles parameter $\rho$ allows only the probabilities of the results (0,0), (1,0), (0,1), (1,1) to be reshuffled. Conversely, the newly introduced parameter $\theta_3$ allows also the probabilities of the other results to vary.
As a result, assuming the other parameters as known, one is justified in expecting both dependence structures to have an impact on the prediction of 1-X-2 (see the left of Figure \ref{fig:bett}) and Under/Over 1.5, while only the dependence structure of the Mar-Co model is supposed to affect the prediction of Under/Over 2.5 since in that specific case the four cells contained in the black boxes in Figure \ref{fig:bett} belong to just one of the two possible events, i.e. Under 2.5.
\begin{figure}
\includegraphics[width=0.99\textwidth]{Fig5.eps}
\caption{Cumulative difference between the RPSs computed for the two competing models. Betting type: Under/Over 1.5.}
\label{fig:2prova}
\end{figure}
\begin{figure}
\includegraphics[width=0.99\textwidth]{Fig6.eps}
\caption{Cumulative difference between the RPSs computed for the two competing models. Betting type: Under/Over 2.5.}
\label{fig:3}
\end{figure}
Figure \ref{fig:2prova} and Figure \ref{fig:3} show the cumulative difference between the RPSs for Under/Over 1.5 and Under/Over 2.5, respectively. Focusing on Figure \ref{fig:2prova}, the Dixon and Coles model appears to outperform the Mar-Co one, with the sole exception of the Italian Serie A. However, the difference seems to be mainly due to chance: indeed, more than 40\% of the time the reshuffled samples obtained by replicating the aforementioned procedure provide a stronger evidence in favour of the Dixon and Coles model than the observed one. In view of this, the observed difference is not sufficient to draw conclusions about the Under/Over 1.5 framework.
As regards the Under/Over 2.5 betting type, Figure \ref{fig:3} highlights a totally different behaviour as the Mar-Co model largely outperforms the Dixon and Coles one, again with the sole exception of the French Ligue 1.
The scale fixed scale of the y-axis in graphs \ref{fig:1}--\ref{fig:3} helps in showing the magnitude of the difference between the two models in this betting type, with respect to the others.
In this case, only $\sim$ 2\% of the time the reshuffled samples provide a stronger evidence in favour of the Mar-Co model than the observed one, and so the Mar-Co model seems to represent an interesting improvement of the Dixon and Coles model in the Under/Over 2.5 framework. This evidence is not entirely surprising given the specific betting type taken into account as any value of $\rho$ does not modify the probability of the event \emph{Under 2.5} (and consequently also of the event \emph{Over 2.5}) obtained with $\rho=0$. As a consequence, the fact that, given the estimates of the fixed effects and of $\gamma$, the estimates of the probabilities of Under 2.5 and Over 2.5 obtained by the Dixon and Coles model are equal to the estimates obtained assuming independence regardless the value of $\hat{\rho}$ represents ad undeniable limit in the Dixon and Coles procedure that may explain the improved performance obtained by the newly introduced Mar-Co model.
\section{Final remarks}
\label{sec::conclusion}
This work aimed to enrich the literature debate about goals-based models, by providing in-depth analyses on the distribution of the number of goals and by proposing an innovative adjustment of a classic and widely used model.
According to our contribution, nowadays the structure of the dependence between the number of goals scored by two competing teams in a match cannot be simplified to an adjustment of the probabilities of the outcomes 0-0, 1-0, 0-1, 1-1. This finding, jointly with the necessity of keeping a simple interpretation of the marginal distribution and their relation, motivates the alternative model proposed in this paper, which, moving from the Dixon and Coles approach, presents a more flexible and comprehensive representation of the dependence structure.
The particular specification of the term regulating the dependence, which is related to the conditional strength of the two competing teams, also plays a key role in providing interesting insights and interpretations of league dynamics.
Despite the lack of a full formalized protocol of comparison, the first encouraging results in terms of absolute predictive capability and further highlighted by the comparison between the Mar-Co model and the Dixon and Coles one confirm the validity of our proposal. This is particularly true in case of the Under/Over 2.5 bet, where the probability of the two betting outcomes, under the Dixon and Coles model, is not influenced by the dependence parameter.
However, some aspects must be pointed out. First of all, alternative types of bet (e.g. Asian Handicap bet) are worth exploring in order to obtain further evidences about the relation between dependence structure and prediction effectiveness.
Secondly, the inclusion of external covariates, such as information on the players conditions or on the teams motivations, should be considered to make the model even more attractive for an effective use in the world of betting.
Finally, the lack of a recursive estimation process represents an undeniable limit in terms of computational effort required, since new crucial information is provided after each match.
However, although an extended analysis on a longer period and on more leagues would be desirable, the promising results achieved shed new light on the importance of properly choosing a suitable dependence structure to model the number of goals in a football match.
\section*{Declarations}
\subsection*{Funding} Not applicable.
\subsection*{Conflicts of interest} The authors declare that they have no conflict of interest.
\subsection*{Availability of data and material} Data are publicly available at \url{http://www.football-data.co.uk/}.
\subsection*{Code availability} The {\sf R} code used in this study is available from the corresponding author upon reasonable request.
\bibliographystyle{apa}
|
\section{Introduction}
In general, humans are able to reason about the identity of an object and the object's motion independently, thereby implying that identity and motion are considered as disentangled generative attributes \cite{Higgins2018, bengio1}. In other words, a change to an object's motion does not affect the object's identity. For example, in sign language translation, the canonical form of a gesture exists independently of the identity or appearance of the signer. In order to reason independently about the latent factors underlying identity and motion, it is therefore desirable to seek disentanglement.
Achieving disentangled representations is a long standing goal for machine learning, and supports causal reasoning \cite{Parascandolo2018,Besserve2019, suter2018, Vowels2020b}, fair machine learning \cite{locatello2019fairness, louizos2017, creager2019, Vowels2020}, generalizability \cite{hosoya2020, MLVAE}, structured/controllable inference and prediction \cite{Yan2019}, attribute transfer \cite{bousmalis2016,press1}, and improved performance on downstream tasks \cite{bengio1, Lake2017, Ridgeway2016, Raffin2019, gatedvae}. Unfortunately, being able to consistently learn a disentangled representation such that the factors correspond with meaningful attributes has been shown to be both theoretically and empirically impossible without the use of some form of supervision or inductive bias \cite{dai3, locatello}. However, acquiring high-quality supervision is expensive and time-consuming. Whilst many methods rely on such supervision, we consider how the implicit structure embedded in video data can be leveraged and reflected in the structure of the model, in order to achieve unsupervised disentanglement.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{ sprites_action_swap.pdf}
\caption{VDSM action transfer for the Sprites dataset \cite{Li2018b}. The target action embedding is transferred swapped with one from a different identity and the sequence is generated.}
\label{fig:sprites_action_swap}
\end{figure}
We propose Video Disentanglement via State-Space-Modeling (VDSM). VDSM is motivated by a careful consideration of the generative structure of a video sequence, which is assumed to be composed of identity (i.e. the canonical appearance of an object or person), an action (i.e. dictating the dynamics governing change over time), and pose (i.e. the time varying aspects of appearance).
In summary, VDSM:
1) Is a completely unsupervised approach that avoids the need for adversarial training.
2) Incorporates a novel structure designed to factorise appearance and motion, using a strong mixture of decoders to separate identity.
3) Produces embeddings that achieve state-of-the-art classification performance.
4) Far exceeds GAN based approaches in sequence generation (evidenced by FID scores).
5) Produces superior disentanglement compared to approaches which use forms of supervision.
6) Exceeds accuracy consistency by over 30\% compared to the nearest competing approach.
7) Produces qualitative results that reflect the quantitative performance in terms of disentanglement and image quality.
The rest of this paper is structured as follows. First we discuss related work in Section 2 before describing the structure and training of VDSM in Section 3. We provide qualitative and quantitative experiments in Section 4 and conclude in Section 5.
\section{Related Work}
Disentanglement is an ongoing challenge for machine learning. However, achieving consistent, interpretable disentanglement without some form of supervision or inductive bias is impossible \cite{locatello}. A large body of recent work seeks disentanglement via the incorporation of various levels of weak- \cite{gatedvae, locatello2020, denton2017, MLVAE, shu2019, chen2019} or semi-supervision \cite{louizos2017, moyer1, DIVA, siddharth2017}. Many of these methods seek disentanglement between static and time-varying factors in sequences (such as content and pose). Impressive results have been achieved in extending this goal to multi-object tracking \cite{Akhundov2019,Kosiorek2018, Eslami2016, Kossen2019,Stelzner2019, Hsieh2018}. Such an approach has the advantage of trading the need for explicit supervision with structural inductive bias. For example, Grathwohl \& Wilson (2016) \cite{Grathwohl2016} utilize variational inference with hierarchical latent factors and a slow-feature regularizing penalty to disentangle content and pose. Denton \& Birodkar (2017) \cite{denton2017} disentangle content from pose by combining adversarial training with two deterministic, recurrent encoder-decoders. They incorporate an adversarial component which discriminates between content encodings from within the same or between different sequences. Other methods which seek disentanglement between static and time-varying factors include: S3VAE \cite{Zhu2020}, which uses recurrent variational autoencoders; VideoVAE \cite{He2018}, a semi-supervised technique that enables attribute control for video generation; Factorizing VAE (FAVAE) \cite{Yamada2020}, which uses multi-scale time-based convolution with variational autoencoders; Disentangled Sequental Autoencoders (DSA) \cite{Li2018b} and Disentangled State Space Model (DSSM) \cite{Miladinovic2019} which undertake structured variational inference; and G3AN \cite{Wang2020b} and MoCoGAN \cite{Tulyakov2017} both Generative Adversarial Networks (GANs).
Other networks which leverage structure to model sequence dynamics include a wide range of latent variable and state-space variational methods. For instance, Deep Kalman Filters \cite{Krishnan2015} extend the structure of the traditional Kalman filter to incorporate recent deep learning techniques for video prediction. Structured inference networks \cite{Krishnan2016} leverage similar structural considerations for sequential data, facilitating time-series causal inference from the disentangled representations. These networks are similar to other stochastic latent variable networks for sequence modeling, which tend to vary according to the structure imposed in the generative and inference models (see \textit{e.g.} \cite{Becker2019, YinND, Watter2015, Hafner2019b, Hafner2019, Rybkin2019, Schmeckpeper2019, Bayer2015, Chung2016, Marino2019, Aksan2019, Watter2015, Gregor2019, Fraccaro2016, Goyal2017}).
Not all methods are concerned with achieving factor disentanglement, and many are primarily designed to generate or predict future video frames. These networks may still include structured variational and/or adversarial techniques. Recent attempts at video generation include SinGAN \cite{Shaham2019} which produces a video sequence from a single image; various multi-scale generative adversarial networks \cite{Mathieu2016b, Clark2019, Acharya2018}; the Latent Video Transformer \cite{Rakhimov2020} which applies transformers \cite{Vaswani2017} to discrete latent representations learned in an autoencoder \cite{VQVAEs}; and pixel-level autoregressive methods \cite{Weissenborn2019, Kalchbrenner2017}.
Evidently, the tasks of video generation and disentanglement overlap, particularly when structured networks are used. The more inductive bias is used, the more domain-specific the network tends to become. For example, the network may incorporate inductive bias corresponding to the physical laws of interaction and motion \cite{Fraccaro2017, Battaglia2016, Ye2018, Chen2019c, Toth2019, Greydanus2019}. We prefer to keep our method general, such that it may be applied to non-vision related tasks for which the benefits of such inductive bias may be inappropriate.
In terms of structure, our method `VDSM', is probably closest to DSA \cite{Li2018b}, Structured Inference Networks \cite{Krishnan2016}, DSSM \cite{Miladinovic2019} and Factorized Hierarchical VAE (FSVAE) \cite{Hsu2017}. In terms of the computer vision applications, VDSM is most similar to the recently released S3VAE \cite{Zhu2020} and G3AN \cite{Wang2020b}. The latter two methods seek to disentangle motion and appearance for video data, and incorporate various forms of weak- or semi-supervision, such as optical flow \cite{Zhu2020}, or labels for conditional motion generation \cite{Wang2020b}. Importantly, and in contrast with these two methods, our network does not use any form of supervision.
Whilst adversarial methods such as G3AN and MoCoGAN are popular and have been shown to work well for density estimation and sequence generation, they are also notoriously difficult and unreliable to train \cite{moyer1, lezama, gabbay2019}. Furthermore, recent work has highlighted that adversarial training is potentially unnecessary, and that non-adversarial methods can achieve comparable or better results across a wide range of tasks \cite{moyer1, VQVAEs, Ho2020, Vahdat2020NVAE, Xiao2020VAEBM, gabbay2019}. Finally, supervision may not always be available, and unsupervised methods are more generally applicable in scenarios where labels (even partial labels) are not available. VDSM intends to address the considerable challenge of disentangling key generative factors in sequences (pose, identity, dynamics) without supervision, and without adversarial training.
\begin{figure}
\centering
\includegraphics[width=.8\linewidth]{ SHVAE_arch.pdf}
\caption{Block diagram of VDSM architecture. Gray lines indicate mixture of expert selection, dashed lines indicate inference, solid lines indicate generation, and looped arrows indicate autoregressive dependency.}
\label{fig:block}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=.85\linewidth]{ SHVAE_gen.pdf}
\caption{Generative model for VDSM with images $\mathbf{x}$, latent pose $\mathbf{z}$, motion dynamics $\mathbf{d}$, and static factors / content $\mathbf{s}$. Grey shading indicates pre-trained components.}
\label{fig:genmodel}
\end{figure}
\section{Video Disentanglement with State-Space Modeling (VDSM)}
Consider $N$ sequences of natural images $\{\mathbf{x}_{1:T_n}^n \}_{n=1}^N$ for which we may assume each image in a sequence lies on some significantly lower-dimensional manifold. We wish to encode the images via a stochastic mapping $p_\theta(\mathbf{z}_t^n, \mathbf{s}^n, \mathbf{d}^n | \mathbf{x}_{t=1:T_n}^n)$ where $\mathbf{x}_{t=1:T_n}^n$ is the $n^{th}$ sequence of $T_n$ images, and $\mathbf{z} \in \mathbb{R}^{\kappa_z}$, $\mathbf{s} \in \mathbb{R}^{\kappa_s}$ and $\mathbf{d} \in \mathbb{R}^{\kappa_d}$ provide compact representations for the time varying latent factors (\textit{e.g.} pose), static latent factors (\textit{e.g.} identity), and the action dynamics (\textit{e.g.} waving), respectively.
A block diagram for VDSM is shown in Figure \ref{fig:block}. Inspecting this figure, it can be seen that images from a sequence are first encoded individually into two latent partitions $\mathbf{h}_t^n$ (which vary over the sequence), and $\mathbf{s}^n$ (which remain constant throughout a sequence. The time varying features are summarized by $\mathbf{d}^n$ which represents the dynamics or action being performed, and then decoded following a sequence-to-sequence (seq2seq) structure \cite{sutskever2}. A combination network (in yellow) takes the output of the seq2seq network and parameterizes a posterior distribution $\mathbf{z}_t^n$ which, when combined with the dynamics $\mathbf{d}^n$ and the static factors $\mathbf{s}^n$, are decoded to reconstruct the original images. In the decoding stage, the static factors $\mathbf{s}^n$ are also used to blend weights from a bank of decoders that form the Mixture of Experts (MoE) decoder which specializes in reconstructing information relevant to identity.
According to the task of image reconstruction through a structured, probabilistic bottleneck, VDSM resembles a variational autoencoder \cite{kingma, rezende2}. We incorporate inductive bias in the form of hierarchical latent structure, dynamic priors, Markov factorization, and architectural constraints (\textit{e.g.} MoE and seq2seq). The \textbf{generative model} for VDSM is shown in Figure \ref{fig:genmodel} and can be factorized as follows:
\vspace{-2mm}
\begin{equation}
\begin{split}
p_{\theta}(\mathbf{x}^n_{t=1:T_n}, \mathbf{s}^n, \mathbf{d}^n, \mathbf{z}^n_{t=1:T_n}) = \\ p_{\theta}(\mathbf{s}^n)p_{\theta}(\mathbf{d}^n)p_{\theta}(\mathbf{z}_{t=1}^n)\prod_{t=2}^{T_n} p_{\theta}(\mathbf{x}^n_t|\mathbf{s}^n, \mathbf{z}_t^n) p_{\theta}(\mathbf{z}_{t}^n|\mathbf{z}_{t-1}^n, \mathbf{d}^n)\\
p_{\theta}(\mathbf{s}^n) = \mathcal{N}(0, \frac{1}{\mathrm{N}_s})\\
\hat{\mathbf{s}}^n = \mbox{softmax}\left(\mathbf{s}^n / \tau_{s} \right) \\
p_{\theta}(\mathbf{d}^n) = \mathcal{N} \left( \mathbf{0}, \mathbf{1} \right) \\
p_{\theta}(\mathbf{x}^n_t|\mathbf{s}^n, \mathbf{z}_t^n) = \mbox{Bern}\left(\mbox{DEC}_{MoE}(\hat{\mathbf{s}}^n, \{\mathbf{s}^n, \mathbf{z}_t^n\} \right) )\\
p_{\theta}(\mathbf{z}^n_{t=1}) = \mathcal{N} \left( \mathbf{0}, \mathbf{1} \right)\\
p_{\theta}(\mathbf{z}^n_{t>1}) = \mathcal{N} \left( \mu_{tr}(\mathbf{z}_{t-1}^n, \mathbf{d}^n), \sigma_{tr}(\mathbf{z}_{t-1}^n, \mathbf{d}^n) \right)
\end{split}
\label{eq:generativeparams}
\end{equation}
Considering this factorization and Figure \ref{fig:genmodel}, for sequence $n$ we sample an action from the dynamics/action factor $\mathbf{d}^n$ and an initial pose $\mathbf{z}^n_{t=1}$. These are fed into a transition network (green) which produces the next pose factor $\mathbf{z}^n_{t=2}$ and we repeat this for $T_n$ timepoints. The transition network outputs the location $\mu_{tr}(\mathbf{z}_{t-1}^n, \mathbf{d}^n)$ and scale $\sigma_{tr}(\mathbf{z}_{t-1}^n, \mathbf{d}^n)$ for a diagonal Gaussian parameterizing the next timepoint $\mathbf{z}_t^n$. We then sample an identity/static factor $\mathbf{s}^n$. For each timepoint, we concatenate this factor with the pose $\mathbf{z}_t$ for timepoints $t=1:T_n$ and pass them into a Mixture of Experts (MoE) Decoder, resulting in the conditional likelihood $p_{\theta}(\mathbf{x}^n_t|\mathbf{s}^n, \mathbf{z}_t^n)$. The MoE weights are determined by $\hat{\mathbf{s}}^n$ which is derived by dividing factors $\mathbf{s}^n$ by temperature $\tau_s$ and applying a softmax operation. The conditional likelihood (i.e., the distribution of generated images) is Bernoulli distributed according to Eq. \ref{eq:generativeparams} but can be chosen according to the data. $\mathrm{N}_s$ (line 3 Eq. \ref{eq:generativeparams}) represents a prior `guess' as to the number of distinct individuals/identities. $\hat{\mathbf{s}}^n$ is simplectic and similar to the Dirichlet distribution; it sums to 1, and the parameter $\tau_{s}$ controls the temperature \cite{Gelman1996}.\footnote{Similar to the Gumbel Softmax, or Concrete distributions \cite{Maddison2017, jang2017}.} A Dirichlet distribution was found to be less stable to train.
\begin{figure}
\centering
\includegraphics[width=.95\linewidth]{ SHVAE_inf.pdf}
\caption{Inference model for VDSM with images $\mathbf{x}$, latent pose $\mathbf{z}$, motion dynamics $\mathbf{d}$, encoder bi-LSTM and decoder LSTM hidden states $\mathbf{h}$ and $\bar{\mathbf{h}}$ resp., and static factors / content $\mathbf{s}$. \cite{Kingma2016}. Together the encoder and decoder LSTMs form a seq2seq model with dynamics $\mathbf{d}^n$ as the bottleneck/summary vector. Grey background shading indicates pre-trained components.}
\label{fig:infmodel}
\end{figure}
In order to undertake this inference, we leverage stochastic variational inference (SVI) \cite{Jordan1999, rezende2, kingma} to learn an approximate posterior distribution $q_{\phi}(.)$ according to the VDSM \textbf{inference model} shown in Equation \ref{eq:inf2}. In words, the static factor representing the identity of an individual $\mathbf{s}^n$ is represented as a diagonal-Gaussian distribution. The location and scale for this Gaussian are parameterized as functions $f_s$ of the average of the identity encodings of the images in a sequence using encoder $\mbox{ENC}_s(\mathbf{x}^n_{t=1:T_n})$. The function $f_s$ is a linear/fully-connected layer with a non-linear activation. By taking the average, we force the network to learn factors which remain constant over the course of a sequence (i.e., the identity). Before the softmax operation, $\tau_s$ is used to anneal the temperature of this distribution during training. The time-varying pose embeddings from the encoder $\mbox{ENC}_z$ (not shown in Eq. \ref{eq:inf2}) are fed into a recurrent, bidirectional Long Short Term Memory (bi-LSTM) \cite{hochreiterLSTM} network, the final hidden state of which $\mathbf{h}^n$ is fed through functions $f_{d\mu}(.)$ and $f_{d\sigma}(.)$ (also linear layers with non-linear activations) to parameterize a Gaussian distribution for dynamics factors $\mathbf{d}^n$. The dynamics embedding $\mathbf{d}^n$ is duplicated and used as the initial hidden and cell states for a decoding (uni-directional) LSTM. This decoder LSTM produces hidden states at each timestep $\bar{\mathbf{h}}_t$ which are fed to the combiner functions $\mu_{comb}$ and $\sigma_{comb}$, with $\mathbf{z}_{t-1}^n$ from the previous timestep.
\vspace{-3mm}
\begin{equation}
\begin{split}
q_{\phi}(\mathbf{s}^n, \mathbf{d}^n, \mathbf{z}_{t=1:T_n} | \mathbf{x}^n_{t=1:T_n}) =\\ q_{\phi}(\mathbf{s}^n|\mathbf{x}^n_{t=1:T_n})q_{\phi}( \mathbf{d}^n|\mathbf{x}^n_{t=1:T_n})q_{\phi}(\mathbf{z}^n_{t=1},|\mathbf{x}^n_{t=1:T_n}, \mathbf{d}^n)\\ \prod_{t=2}^{T_n}q_{\phi}(\mathbf{z}^n_t| \mathbf{z}^n_{t-1},\mathbf{x}^n_{t=1:T_n}, \mathbf{d}^n)\\
q_{\phi}(\mathbf{s}^n|\mathbf{x}^n_{t=1:T_n}) = \mathcal{N} \left( s_{loc}, s_{scale} \right)\\
s_{loc} = f_{s,loc}\left(\frac{1}{T_n}\sum_{t=1}^{T_n} (\mbox{ENC}_s(\mathbf{x}^n_{t}))\right) \\
s_{scale} =f_{s,scale}\left(\frac{1}{T_n}\sum_{t=1}^{T_n} (\mbox{ENC}_s(\mathbf{x}^n_{t}))\right)\\
q_{\phi}(\mathbf{d}^n | \mathbf{x}^n_{t=1:T_n}) = \left(\mathcal{N} f_{d\mu}(\mathbf{h}^n), f_{d\sigma}(\mathbf{n}^n) \right) \\
q_{\phi}(\mathbf{z}^n_{t} | \mathbf{x}^n_{t=1:T_n}, \mathbf{d}^n) =\\ \mathcal{N}(\mu_{comb}(\mathbf{z}_{t-1}^n,\bar{\mathbf{h}}_t^n, \mathbf{d}^n), \sigma_{comb}(\mathbf{z}_{t-1}^n,\bar{\mathbf{h}}_t, \mathbf{d}^n)))
\end{split}
\label{eq:inf2}
\end{equation}
\vspace{-2mm}
\begin{equation}
\mbox{max}_{\phi, \theta} \Braket{ \Braket{ \frac{p_{\theta}(\mathbf{x}^n_{t=1:T_n}, \mathbf{s}^n, \mathbf{d}^n, \mathbf{z}^n_{t=1:T_n})}{q_{\phi}(\mathbf{s}^n, \mathbf{d}^n, \mathbf{z}_{t=1:T_n}^n\mid \mathbf{x}^n_{i=1:T_n})) }}_{q_{\phi}}}_{p_{\mathcal{D}(\mathbf{x}^n)}}
\label{eq:elbo}
\end{equation}
The goal of the inference model is to make the problem of latent inference possible through the use of parametric approximating posteriors. Together, the generation and inference models can be optimized as part of a stochastic, amortized, variational inference objective, known as the Evidence Lower Bound (ELBO), given in Equation \ref{eq:elbo}. The ELBO objective may be derived similarly to \cite{Krishnan2016} (also see supplementary material) and is shown in Equation \ref{eq:fullobj}. In Eq. \ref{eq:fullobj}, each $\lambda$ term represents a hyperparameter used for annealing the corresponding KL (Kullback-Liebler divergence) objective term during training.
\vspace{-2mm}
\begin{equation}
\begin{split}
\mathcal{L}(\mathbf{x}^n_{1:T_n};(\theta,\phi)) =\\ \sum_{t=1}^{T_n} \mathbb{E}_{q_{\phi}(\mathbf{z}_t^n, \mathbf{s}^n | \mathbf{x}^n_{1:T_n})}[\log p_{\theta}(\mathbf{x}^n_t|\mathbf{z}_t^n, \mathbf{s}^n)]\\
-\lambda_{d}(\mbox{KL}(q_{\phi}(\mathbf{d}^n|\mathbf{x}^n_{1:T_n})||p_{\theta}(\mathbf{d}^n))) \\
-\lambda_{s}(\mbox{KL}(q_{\phi}(\mathbf{s}^n|\mathbf{x}^n_{1:T_n})||p_{\theta}(\mathbf{s}^n))) \\
-\lambda_{z}(\mbox{KL}(q_{\phi}(\mathbf{z}^n_1|\mathbf{x}^n_{1:T_n}, \mathbf{d}^n)||p_{\theta}(\mathbf{z}^n_1)) \\
- \lambda_{z} \sum_{t=2}^{T_n}\mathbb{E}_{q_{\phi}(\mathbf{z}_{t-1}^n|\mathbf{x}^n_{1:T_n},\mathbf{d}^n)}\\\mbox{KL}(q_{\phi}(\mathbf{z}_t^n|\mathbf{z}_{t-1}^n,\mathbf{d}^n, \mathbf{x}^n_{1:T_n})||p_{\theta}(\mathbf{z}_t^n|\mathbf{z}_{t-1}^n, \mathbf{d}^n)))
\label{eq:fullobj}
\end{split}
\end{equation}
\subsection{Functional Form}
As depicted in Figures \ref{fig:genmodel} and \ref{fig:infmodel}, there are a number of functions with learnable parameters in the generative and inference models of VDSM.
\textbf{Encoder and Static Factors}: The encoder functions $\mbox{ENC}_{\{s, z\}}(.)$ (pink block in Figure \ref{fig:infmodel}) are used to infer the static and time varying embeddings of the images at each timestep. It comprises 5 layers of anti-aliased convolution \cite{Zhang2019conv} downsampling, as well as two separate fully-connected embedding layers with non-linear activations. The identity embedding is averaged over a sequence and fed through a fully-connected layer $f_s(.)$ (light gray block in Figure \ref{fig:infmodel}) which is used to infer the identity/static factor for a sequence $\mathbf{s}^n$.
\textbf{Dynamics Layer and LSTMs: } The Dynamics layer (red block Figure \ref{fig:infmodel}) comprises $f_{d\mu}(.)$ and $f_{d\sigma}(.)$ which are fully-connected neural network layers with non-linear activations used to infer the location and scale of the dynamics/action factor $\mathbf{d}^n$ and are fed with the last hidden state from the encoder bi-LSTM. It therefore represents a bottleneck inside a seq2seq network. The seq2seq network's uni-directional LSTM decoder uses the inferred $\mathbf{d}^n$ as the initial hidden and initial cell states, and produces per-timestep hidden states $\bar{\mathbf{h}}_t^n$ which are fed to the combiner.
\textbf{The Combiner}: The Combiner (yellow block in Figure~\ref{fig:infmodel}) is used to infer the current latent pose factor at time $t$ given the current hidden state from the decoder LSTM, the previous latent pose factor, and the dynamics: $\mathbf{z}_t^n|\mathbf{z}_{t-1}^n, \bar{\mathbf{h}^n_t}$. The function is parameterized as follows:
\begin{equation}
\begin{split}
\hat{\mathbf{h}}_t^n = \{\bar{\mathbf{h}}_t^n, \mathbf{d}^n\}\\
\mathbf{c}_t^n = 0.5(\mbox{tanh}(f_7(\mathbf{z}^n_{t-1})) + \hat{\mathbf{h}}_t^n)\\
\mbox{loc-comb}_t^n = f_8(\mathbf{c}_t^n)\\
\mbox{scale-comb}_t^n = \log(1 + \mbox{exp}(f_{9}(\mathbf{c}_t^n)))\\
\mathbf{z}_t^n \sim \mathcal{N}(\mbox{loc-comb}_t^n, \mbox{scale-comb}_t^n)
\end{split}
\end{equation}
Where $\{\bar{\mathbf{h}}_t^n, \mathbf{d}^n\}$ is the concatenation of the LSTM decoder hidden state and the dynamics, and functions $f_{7-9}(.)$ are fully-connected neural network layers with non-linear activations. During development we explored the use of Inverse Autogressive Flows (IAFs) to increase the expressivity of the approximating posterior. IAFs are an adapted form of Normalizing Flow \cite{Kobyzev2020, rezende} designed for efficient transformation of simple posterior approximating distributions. IAFs leverage an invertible transformation function with a particularly tractable Jacobian determinant \cite{rezende, Kobyzev2020, Kingma2016}. However, at least in our particular experiments, we found that the concomitant increase in computational complexity was not justified by the negligible performance increase.
\textbf{Transition Network: } Similar to the state-space transition model in \cite{Krishnan2015}, the transition network (green block Figure \ref{fig:genmodel}) parameterizes the latent time-varying factors $\mathbf{z}^n_t$ with functions $\mu_{tr}(.), \sigma_{tr}(.)$, each with the following parameterization:
\vspace{-2mm}
\begin{equation}
\begin{split}
\mathbf{g}_t^n = \mbox{sigmoid}(f_2(\mbox{ReLU}(f_1(\{\mathbf{z}_{t-1}^n, \mathbf{d}^n\}))))\\
\mathbf{h}_t^n = f_4(\mbox{ReLU}(f_3(\{\mathbf{z}_{t-1}^n, \mathbf{d}^n\})))\\
\mbox{loc-trans}_t^n = \mathbf{g}_t\circ \mathbf{h}_t^n + (1-\mathbf{g}_t^n)\circ f_5(\{\mathbf{z}_{t-1}^n, \mathbf{d}^n\}) \\
\mbox{scale-trans}_t^n = \log(1 + \mbox{exp}(f_6(\mbox{ReLU}(\mathbf{h}_t^n))))
\end{split}
\end{equation}
where $\circ$ indicates the Hadamard/elementwise product, $\{\mathbf{z}_{t-1}^n, \mathbf{d}^n\}$ is the concatenation of the previous pose vector with the dynamics vector for sequence $n$, and the functions $f_{1-6}(.)$ are fully-connected neural network layers with non-linear activations. This network is used as part of the generation process to generate successive latent pose factors $\mathbf{z}_t^n|\mathbf{z}^n_{t-1}$.
\begin{figure*}
\centering
\includegraphics[width=0.8\linewidth]{ mixed_results_1.pdf}
\caption{\textbf{Left}: Sprites dataset \cite{Li2018b} action transfer examples for S3VAE \cite{Zhu2020}, MN (MonkeyNet) \cite{Siarohin2019MonkeyNet} and DSA \cite{Li2018b} (results adapted from \cite{Zhu2020}). \textbf{Middle} and \textbf{right}: VDSM action transfer on Sprites and moving MNIST datasets \cite{Srivastava2016} respectively. `Target Action' indicates the desired action, whilst `Source ID' indicates the identity to which the action is transferred during sequence generation.}
\label{fig:mixed_results}
\end{figure*}
\textbf{Mixture of Experts Decoder: } The Mixture of Experts (MoE) \cite{Zhang2018, Shazeer2017, Jordan1993, Eigen2013} decoder $\mbox{DEC}_{MoE}(.)$ (blue block in Figure \ref{fig:genmodel}) constitutes a bank of $\mathrm{N}_s$ upsampling convolution image decoders. Each decoder comprise a series of 2D transpose convolution and Leaky ReLU activation layers. The weights and biases for the transpose convolution operations of the bank of MoE decoders are \textit{blended} according to a weighted average, where the weighting is determined by the identity mixture parameter $\hat{\mathbf{s}}^n$. In addition, $\mathbf{s}^n$ (which is $\hat{\mathbf{s}}^n$ before temperature annealing and the softmax operation) is concatenated with $\mathbf{z}_t^n$ and fed to the input of the decoder. As the temperature parameter $\tau_s$ is annealed during training, the mixture parameter becomes more `peaked' resulting in a tendency to select individual decoders from the bank, rather than mix between them. Each decoder in the bank is therefore given the opportunity to specialize in reconstructing a particular identity, whilst also leveraging information encoded in the weights in the entire bank. The MoE decoder is used as part of the generation process to generate the images $\mathbf{x}_t^n | \mathbf{z}_t^n, \mathbf{s}^n $.
\subsection{Training and Testing VDSM}
\textbf{Training:} Training is split into two stages. Without two stages, the network tends to push both pose and identity information into $\mathbf{d}$ to avoid the averaging operation associated with $\mathbf{s}$. We start by \textbf{pretraining} the encoder $\mbox{ENC}_{\{s,z\}}$, MoE decoder $\mbox{DEC}_{MoE}$ and the Static Factors Layer $f_s$ without modeling the autoregressive structure of the data (this is illustrated in the gray shaded regions of Figures \ref{fig:genmodel} and \ref{fig:infmodel}. In other words, we begin by treating the data as non-sequential, and group images from the same individuals/identities into batches in order to facilitate the inference of the identity factor $\mathbf{s}^n$ and i.i.d. $\mathbf{z}_t^n$. During this first stage, the objective function is reduced to the following:
\vspace{-2.5mm}
\begin{equation}
\begin{split}
\mathcal{L}(\mathbf{x}^n_{1:T_n}; (\theta', \phi')) =
\frac{1}{N} \sum_{i=1}^{T}(\mathbb{E}_{q_{\phi'}(\mathbf{z}_i, \mathbf{s}| \mathbf{x}_i)}[\log p_{\theta'}(\mathbf{x}_i | \mathbf{z}_i, \mathbf{s})]\\
- \lambda_z \mbox{KL}[q_{\phi'}(\mathbf{z}_i | \mathbf{x}_i) \| p_{\theta'}(\mathbf{z})])
- \lambda_s \mbox{KL}[q_{\phi'}(\mathbf{s} | \mathbf{x}_i) \| p_{\theta'}(\mathbf{s})])
\end{split}
\label{eq:ELBOpretrain}
\end{equation}
Here we deliberately use subscript $i$ rather than $t$ to emphasize that the images need not be in any sequential order, they just need to be grouped according to the same identity. Furthermore, we indicate that we are only training a subset of the inference and generation parameters with $(\theta', \phi')$. During pretraining, the KL weight on the static factors is inversely annealed $\lambda_s = \{ 0.1, ..., 1.0 \}$ (i.e. low to high) with respect to the KL weight on the pose factor $\lambda_z = \{ 30, ..., 1.0 \}$ (i.e. high to low), with the number of steps equal to the number of training epochs. This forces information to flow through the static factor $\mathbf{s}^n$ because the high weight $\lambda_z$ pushes the time varying factors $\mathbf{z}_i$ to zero. The static factor is computed as an average over embeddings and as such, it is forced to represent the information which is consistent over the group of images (i.e. the identity). Simultaneously, the temperature parameter $\tau_{s}$ is gradually increased, resulting in a shift from uniform MoE decoder blending weights, to increasing specialization. Then, as the weight $\lambda_z$ falls, the pose factors $\mathbf{z}_i$ start to become useful in encoding information that varies across the groups of images of the same individual. As such, the pretraining enables us to learn the two most highly-parameterized deterministic functions used for mapping images to latent factors and vice versa: the encoder and the MoE decoder. It is worth noting that the pretraining stage yields a valuable model in its own right, that disentangles identity from pose. For instance, in Figure~\ref{fig:pretrained_MUG_swap}, the expression can be swapped with any ID and vice versa.
The \textbf{second stage of training} is concerned with disentangling the sequence dynamics, and introduces the seq2seq encoder and decoder LSTMs, the Combiner, and the transition network. The training objective is given by the full ELBO in Eq. \ref{eq:fullobj}. The weights and biases for the encoder and MoE decoder are all frozen at the end of pretraining, apart from the final layers preceding the parameterizations of $\mathbf{s}^n$ and $\mathbf{z}_t^n$ which are allowed to vary during this second stage. Allowing these layers to be fine-tuned is important in ensuring that the model learns a valid ELBO during the sequential modeling stage (i.e. by allowing the parameters to vary we facilitate variational inference). Annealing is used with $\lambda_z = \{0.1, ..., 1.0\}$ to improve inference of $\mathbf{z}_t$ and to help prevent posterior collapse during training \cite{disentanglement}.
To infer $\mathbf{s}^n$, which is primarily learned during the first stage of training (but fine-tuned in the second stage), we only need to be able to group images of the same individual (and these may or may not be from the same sequence). For $\mathbf{d}^n$ the images need to be from sequences, but no supervision is required. This allows us to sample multiple images of the same individual, without needing labels for which individual is being sampled. So long as there is, on average, some variation across identity between groups of images (or between sequences), the network can infer the identity of the individual. VDSM is trained using the Adam \cite{adam} optimizer with the Stochastic Variational Inference algorithm in the Pyro probabilistic programming language \cite{Bingham2019pyro}. Minimal hyperparameter tuning was undertaken.\footnote{Network details can be found in supplementary material.}
\textbf{Testing:} At test time, the generative model may either be used unconditionally (where the factors $\mathbf{s}^n, \mathbf{d}^n$, and $\mathbf{z}^n_t$ are sampled from their priors), or be used conditioned on some initial state. In the latter case, the inference network may be used to derive $\mathbf{s}^n, \mathbf{d}^n, \mathbf{h}^n_{t=1}$, and $\mathbf{z}^n_{t=1}$ from a sample sequence, or even a single image. The generative model is then conditioned on these initial factors, and further sampling is undertaken according to Figure \ref{fig:genmodel}. The benefit of this flexible structure is that it allows different factors to be swapped or controlled independently in order to mix any identity with any action dynamics.
\begin{figure}
\centering
\includegraphics[width=0.85\linewidth]{ MUG_results_1.pdf}
\caption{Downsampled sequence generation samples from G3AN \cite{Wang2020b}, MoCoGAN \cite{Tulyakov2017}, S3VAE \cite{Zhu2020}, and VDSM (ours). Figure adapted from \cite{Zhu2020, Wang2020b}.}
\label{fig:MUG_seq_gen}
\end{figure}
\section{Experiments}
VDSM is evaluated on four datasets: MUG \cite{Aifanti2010MUG}; colored Sprites \cite{Li2018b}; synthetic swinging pendula; moving MNIST (similar to \cite{Srivastava2016}). MUG \cite{Aifanti2010MUG} comprises 3528 videos of 52 individuals performing different facial expressions for anger, disgust, fear, happiness, sadness, and surprise. The performances vary in length, and were downsampled by a factor of two (to approximately 8 fps) to improve data efficiency. The images were aligned using OpenFace \cite{openface2}, centre cropped, and resized to 64x64 (similar to \cite{Tulyakov2017, Zhu2020}). Random segments of length 16 frames were sampled from the sequences for training, and a 15\% holdout set was used for testing.
The Sprites dataset comprises 64x64 sequences of cartoon characters from the Universal LPC SpriteSheet Character Generator performing 3, 8-frame long action sequences (spellcast, slashing, and walking), from 3 viewing angles. Following \cite{Li2018b} we create characters with 7 body types, 4 shirts, 8 hairstyles, and 5 pants, resulting in 1120 identities. A 10\% holdout set was used for testing. The moving MNIST dataset \cite{Srivastava2016} comprises 55,000 16-frame long sequences of randomly sampled MNIST digits \cite{MNIST} moving in random trajectories. A 10\% holdout set was used for testing. Finally, the pendulum dataset represents a synthetic dataset which comprises seven different colored pendula swinging at two speeds (fast and slow).
We evaluate quantitatively in terms of (1) Fr\'echet Inception Distance (FID) \cite{Heusel2017FID, Borji2018GAN} using a 3D ResNeXt 101 \cite{Xie2017ResNeXt, Hara2017} pre-trained on Kinetics \cite{Kay2017Kinetics} as per \cite{Wang2020b}, (2) consistency at identity and action classification between real and generated sequences \cite{Li2018, Zhu2020}, and (3) in terms of identity and action classification score using the separate identity and dynamics embeddings.\footnote{Additional results may be found in supplementary material.} For (2), the consistency is measured by comparing the predictions of a classifier trained to predict ground truth factors from real images with predictions from sampled images. If the quality of the images is high and the network is encoding information about the identity and action, then the predictions for real and generated images should be close. For (3) we expect that disentangled identity and action embeddings to be informative for predicting their respective factors, but not informative for predicting each other's factors.
Finally, we evaluate quantitatively, in terms of identity swapping, dynamics swapping, and sequence generation quality.
\begin{figure}
\centering
\includegraphics[width=.9\linewidth]{ face_swap.pdf}
\caption{Action transfer on the MUG dataset \cite{Aifanti2010MUG} for S3VAE \cite{Zhu2020}, DSA \cite{Li2018b}, MonkeyNet \cite{Siarohin2019MonkeyNet}, and VDSM (ours). Figure adapted from \cite{Zhu2020}.}
\label{fig:MUG_action_transfer}
\end{figure}
\begin{table}[]
\small
\begin{tabular}{llll} \hline
\textbf{Attr.} & \textbf{Acc. Cons}. $\uparrow$ & \textbf{ID} \textbf{Acc}. & \textbf{Dyn. Acc}. \\ \hline
Body (Sprites) & 99.67 & 98.67 $\uparrow$& 52.67 $\downarrow$\\
Shirt (Sprites)& 99.67 & 99.67 $\uparrow$& 47.33 $\downarrow$ \\
Pant (Sprites)& 99.00 & 97.67 $\uparrow$& 31.67 $\downarrow$\\
Hair (Sprites)& 100.00 & 99.33 $\uparrow$& 21.00 $\downarrow$\\
Action (Sprites) & 99.33 & 45.00 $\downarrow$ & 99.33 $\uparrow$\\
\hline
ID (MUG) & 98.67 & 99.00 $\uparrow$ & 65.67 $\downarrow$\\
Action (MUG) & 88.00 & 57.33 $\downarrow$ & 83.33 $\uparrow$\\
\hline \hline
\end{tabular}
\caption{For each ground-truth attribute of Sprites and MUG, this table shows a breakdown of VDSM's consistency (Acc. Cons.) and disentanglement accuracy scores in percent using the separate identity (ID) and dynamics/action (Dyn.) embeddings. Arrows indicate whether a higher $\uparrow$ or lower $\downarrow$ score is preferred.}
\label{tab:disent}
\end{table}
\begin{table}[]
\small
\begin{tabular}{lllll} \hline
\textbf{Method} & Acc. \textbf{Sprites} & Acc. \textbf{MUG} & $H(y)$ & $H(y|x)$ \\ \hline
VDSM & \textbf{99.53} & \textbf{93.33} & \textbf{2.21} & 0.203 \\
S3VAE \cite{Zhu2020}& \textbf{99.49} & 70.51& 1.760& \textbf{0.135}\\
DSA \cite{Li2018b} & 90.73 & 54.29 & 1.657 & 0.374\\
MoCo \cite{Tulyakov2017} & 92.89 & 63.12 &1.721 & 0.183\\
\hline \hline
\end{tabular}
\caption{Results for accuracy consistency (averaged across attributes) between different methods for the Sprites and MUG datasets, and results for Inter-Entropy $H(y)$ (higher is better) and Intra-Entropy $H(y|x)$ (lower is better) \cite{He2018} on the MUG dataset.}
\label{tab:consall}
\end{table}
\subsection{Quantitative Evaluation}
Quantitative results were obtained for 1000 samples from the test sets. The results in Table \ref{tab:disent} show that the ID embedding is predictive of ID or ID-related attributes (acc. $>97\%$ for Sprites), but not of action (acc.~$=45\%$ for Sprites). Similarly, the dynamics embedding was highly predictive of action (acc. $>99\%$ for Sprites) but not of ID-related attributes (acc. $<53\%$ for Sprites). Interestingly, the ID could be classified using the dynamics embedding for MUG (acc. 66\%). We believe this may be because individuals exhibit unique action dynamics that make them identifiable. This behavior may have relevant application elsewhere (\textit{e.g.} gait recognition). Future work should explore whether this behavior is exhibited by other methods.
The breakdown of VDSM's accuracy consistency scores in Table \ref{tab:disent} demonstrates that there is high corroboration between action and ID-related attribute classifier predictions for real and generated images. Indeed, the results for VDSM exceed those from competing methods. In particular, see the results in Table \ref{tab:consall} where MUG performance was~$>30\%$ higher than the best competing method (S3VAE) despite this method using additional supervision (\textit{e.g.} optical flow). Table \ref{tab:consall} also shows the results for Inter- and Intra-Entropy \cite{He2018} for the same classifier used for the disentanglement evaluation. These are computed as $H(y) = -\sum_y p(y) \log p(y)$ and $H(y|x) = -\sum_y p(y|x) \log p(y|x)$ respectively, where $y$ is the predicted attribute label, and $x$ is the given sequence. The inter- and intra-entropies are averaged across identity and action prediction results. The results for inter-entropy provide a proxy for diversity, and indicate that, in generation mode, VDSM produces the most diverse samples. The results for intra-entropy demonstrate competitive performance with another unsupervised method, DSA. However, together with the disentanglement and accuracy results in Table \ref{tab:consall} suggest that VDSM's primary strength is superior, unsupervised, disentangled representation learning.
\begin{figure}[!h]
\centering
\includegraphics[width=0.6\linewidth]{ expression_swap.pdf}
\caption{Identity/pose transfer using the pre-trained (non-sequential) components of VDSM on the MUG dataset \cite{Aifanti2010MUG}. }
\label{fig:pretrained_MUG_swap}
\end{figure}
FID scores for MUG are shown in Table \ref{tab:FID}. For this evaluation 1000 samples were generated firstly by using $\mathbf{d}^n$ and $\mathbf{s}^n$ embeddings derived from real sequences in the test set (denoted `reconstruction' in the table), and secondly by generating images unconditionally from the priors (denoted `generation' in the table). In both case, the FID substantially improves upon competing adversarial methods.
\begin{table}[]
\centering
\small
\begin{tabular}{lll} \hline
\textbf{Method} &\textbf{FID} \\ \hline
VDSM (ours, generation) & \textbf{44.12} \\
VDSM (ours, reconstruction) & \textbf{17.64} \\
G3AN \cite{Wang2020b} & 67.12 \\
MoCo \cite{Tulyakov2017} & 87.11 \\
VGAN \cite{Vondrick2016VGAN} & 160.76 \\
TGAN \cite{Saito2017TGAN} & 97.07 \\ \hline \hline
\end{tabular}
\caption{FIDs for MUG. We include VDSM in generation and reconstruction modes, where samples are drawn unconditionally from priors or are drawn from approximate posteriors derived from test set sequences, respectively.}
\label{tab:FID}
\end{table}
\subsection{Qualitative Evaluation}
Figures \ref{fig:sprites_action_swap} and \ref{fig:mixed_results} illustrate VDSM's action transfer performance on the Sprites dataset, and some comparisons are included in the leftmost part of Figure \ref{fig:mixed_results}. It can be seen that VDSM facilitates disentangled transfer, with clean separation of time varying and time static information. In contrast, DSA and MonkeyNet struggle to transfer the action, and S3VAE requires auxiliary information to achieve comparably (\textit{e.g.} optical flow). The rightmost part of Fig. \ref{fig:mixed_results} shows action transfer of VDSM on the moving MNIST dataset.
Figure \ref{fig:MUG_seq_gen} compares VDSM's sequence generation performance against G3AN, MocoGAN, and S3VAE, where it can be seen that VDSM outperforms the adversarial methods, and performs comparably to S3VAE, despite not requiring additional information (\textit{e.g.} optical flow).\footnote{Sequences are downsampled - for full sequences see supplementary material.} Figure \ref{fig:MUG_action_transfer} compares the action transfer performance of VDSM against DSA, S3VAE and MonkeyNet, again highlighting that VDSM is comparable to S3VAE despite not requiring additional information during training (\textit{e.g.} optical flow). Figure \ref{fig:pretrained_MUG_swap} demonstrates the \textit{pre-trained} components of VDSM performing identity and expression transfer, illustrating how the model can disentangle pose and identity without needing to model sequences.
Results on the synthetic pendulum dataset are shown in Figure 11 in the supplementary material. These results demonstrate that, even though VDSM was only trained on sequence segments which were 16 frames long, the network can generate sequences following the dynamics of a pendulum with an arbitrary number of frames (40 frames are shown in the figure). The figure also demonstrates action swapping, where the dynamics of one pendulum is transferred to another.
\section{Summary}
We presented VDSM, an unsupervised state-space model for video disentanglement. VDSM incorporates a range of inductive biases including a hierarchical latent structure, a dynamic prior, a seq2seq network, and a mixture of experts decoder. The evaluation demonstrated that the network informatively embeds and disentangles static and time-varying factors, as well as generating quality video many frames into the future. VDSM's performance matches or exceeds that of methods requiring additional supervision (such as optical flow) or adversarial training.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}\label{introduction}
Ongoing advances in gravitational wave astronomy offer unprecedented
opportunities to study the complex and rich physics of neutron stars. The
milestone detection GW170817 of two merging neutron stars
\cite{2017PhRvL.119p1101A,2017ApJ...848L..12A,2017ApJ...848L..13A} (and
subsequent events such as GW190425 \cite{2020ApJ...892L...3A}) triggered
an enormous amount of studies and insights to a wide range of questions on
neutron stars, nuclear physics, cosmology, and fundamental physics~\cite{2017ApJ...850L..19M,2017ApJ...850L..34B,2017PhRvL.119y1301B,2017PhRvL.119y1303S,2017PhRvL.119y1304E,2018ApJ...852L..29R,2018PhRvL.121p1101A}. Among the
key observables that should be detectable in the future by advanced detectors
are the $f$-mode frequencies of massive post-merger objects or rotating
neutron stars (both isolated as well as part of a binary system). Those are a
promising laboratory to study extreme nuclear physics and strong field gravity
in more detail and will help constrain the nuclear equation of state
(henceforth EOS)~\cite{2014PhRvD..90b3002B, 2014PhRvL.113i1104T}.
While binary mergers involving at least one neutron star will provide a
plethora of observations that can be utilized in order to put constraints on
current unknowns in nuclear physics, we will in this study focus on the
sub-field of \emph{gravitational wave
asteroseismology}~\cite{1996PhRvL..77.4134A, 1998MNRAS.299.1059A,
2005ApJ...629..979L}. Starting in the 1970s, helioseismology has proven
extremely successful in gaining a highly detailed knowledge on the composition
as well as processes operating in the interior of the Sun, based on the
observation of acoustic modes visible on the solar
surface~\cite{1976Natur.259...89C, 2002RvMP...74.1073C}. Asteroseismology of
neutron stars will never be able to reproduce that level of detail known from
solar studies, however, it will play a crucial role in the long-lasting
effort to constrain the nuclear EOS; the inverse problem has to be tackled.
Even though neutron stars exhibit a rich spectrum of different oscillation
modes~\cite{1941MNRAS.101..367C, 1985ApJ...297L..37M, 1992MNRAS.255..119K}, we
will focus on the fundamental acoustic mode, or more precisely, the $l=|m|=2$
$f$-mode, as it is the fluid mode strongest coupled to gravitational radiation
and hence most likely to be detected via future gravitational wave
observations. The computation of the $f$-modes of arbitrarily fast rotating
neutron stars is numerically very complex and has a long history
\cite{1967ApJ...149..591T, 1983ApJS...53...73L, 2002MNRAS.334..933J,
2008PhRvD..78f4063G, PhysRevD.102.064026}. Asteroseismology often relies on
the availability of certain \emph{universal relations} that are independent of
the nuclear EOS, i.e., they allow deductions from mode frequencies to bulk
properties of the observed star by solving the inverse problem. Several such
universal relations have been proposed for various oscillation modes of
neutron stars; they often utilise ``basic'' neutron star properties such as mass,
radius and rotation rate to parameterize the mode frequency
\cite{1998MNRAS.299.1059A, 2008PhRvD..78f4063G, 2013PhRvD..88d4052D} but more
recently also more complex quantities such as the effective compactness or tidal
deformability have been employed in the universal relations
\cite{2005ApJ...629..979L, 2010ApJ...714.1234L, 2014PhRvD..90l4023C,
2015PhRvD..92l4004D} in order to improve the accuracy of the estimate. In our
study, we will mostly focus on the most recently proposed universal relation
in Ref.~\cite{Kruger:2019zuz}, as it is the first fitting formula for $f$-mode
frequencies that does not rely on simplifications in their determination.
In this work we adopt a Bayesian framework that allows us to quantify the connection between future measurements of the co- and counterrotating $l=|m|=2$ $f$-modes with key neutron star properties, such as their stellar structure and underlying EOS. This particular choice for the modes is motivated from earlier studies that predict them to be more relevant for typical observations than other modes of the spectrum. The Bayesian framework is related to the one presented recently in Ref.~\cite{Volkel:2020daa}, where it was applied to reconstructing parametrized black hole space-times from their quasi-normal mode spectrum.
Gravitational wave observations of different types of neutron star systems may also come along with complementary information on some of the system properties: e.g., the remnant mass after a neutron star merger may be estimated from the inspiral signal; or the neutron star spin known from radio observation may be accompanied by glitch induced GW signal \cite{2020PhRvD.101j3009H}. Hence, we incorporate mass and spin into our framework. More specifically, we study scenarios in which different levels of prior knowledge on the neutron star mass or independent measurements of the rotation rate are available.
Because the extraction of $f$-modes from complex numerical simulations are computationally time consuming and difficult to automatize, we calculate the $f$-mode frequencies for a sufficiently dense set of equilibrium configurations across the possible parameter space of each considered EOS (they are the same as those used in Ref.~\cite{Kruger:2019zuz}) and interpolate the frequencies of intermediate neutron star models. We call this framework, which always requires an initial choice for the underlying EOS, in the following the {\it EOS method}. The second, complementary approach to the inverse problem is based on using universal relations (UR), which have been reported in the same work \cite{Kruger:2019zuz}. URs allow to compute $f$-mode frequencies from analytic functions of key neutron star parameters up to percent level. The URs have been constructed by fitting the extracted $f$-mode frequencies to a simple analytical function. We call this approach in the following the {\it UR method}.
We find major advantages and disadvantages between the two methods, which are
strongly related to the specific context and what additional information on
neutron star parameters is known. One difference is that the EOS method has
two parameters to uniquely compute the $f$-modes for a given EOS, while the UR
method requires three parameters to do so. If informed prior knowledge on the
remnant mass or additional constraints on the rotation rate are known, meaning
one already knows them to some extent, the UR method provides reliable
constraints on neutron star bulk properties when both $f$-modes have been
observed. Since the currently available universal relation we are using is not
directly incorporating the neutron star radius, it can not directly be used to
recover all neutron star parameters, for which some knowledge of the EOS is
still needed, but most of them. The EOS method has the advantage of being, in
principle, independent of approximations and directly provides all neutron
star parameters that have been computed for the equilibrium solutions
previously. The disadvantage is that the true EOS is not known and assuming a
wrong EOS, i.e., one that does not reflect the physical reality, will in
general yield biased parameters; in our case, our method will yield a neutron
star with different radius and rotation rate as the underlying one which we
used to generate the $f$-mode frequencies (see Sec.~\ref{app_EOS}). The case,
in which two neutron stars constructed using different EOS and stellar structure yield similar $f$-mode frequencies, simply shows that the unique reconstruction requires in general additional information.
This work is structured as follows. Sec.~\ref{Theory_Methods} provides an overview of the theory and methods being used, which are then applied in Sec.~\ref{applications}. We discuss our findings in Sec.\ref{Discussion} and conclude in Sec.~\ref{Conclusions}.
\section{Theory and Methods}\label{Theory_Methods}
In the following we summarize our methods being used in this work. We start with a review of the underlying theoretical framework to study neutron star oscillations in Sec.~\ref{mathematical_formulation}, before we outline the actual computation of $f$-modes in the present work in Sec.~\ref{interp_f}. We then describe the adopted universal relations in Sec.~\ref{subsec_ur} and the Bayesian analysis in Sec.~\ref{Bayesian_framework}.
\subsection{Mathematical formulation}\label{mathematical_formulation}
The framework in which we calculate the $f$-mode frequencies is laid out in detail in a previous article (see Ref. \cite{PhysRevD.102.064026}), however, for completeness, we will repeat the fundamentals here. In this work we assume standard general relativity (with units in which $G=c=1$), whose field equations, along with the law for the conservation of energy-momentum, are given in terms of the Einstein tensor $G_{\mu \nu}$ and the energy-momentum tensor $T_{\mu \nu}$ as
\begin{align}
G_{\mu \nu} = 8 \pi T_{\mu \nu}
\qquad\text{and}\qquad
\nabla_\mu T^{\mu \nu} = 0.
\end{align}
As is common practice in mode studies, we restrict ourselves to linear perturbations around equilibrium.
To model the neutron star, we consider a perfect fluid with energy density $\epsilon$ and pressure $p$; the energy momentum tensor then takes the form
\begin{align}
T^{\mu \nu} = \left( \epsilon + p\right) u^{\mu} u^{\nu} + p g^{\mu \nu},
\end{align}
where $u^\mu$ is the 4-velocity and $g^{\mu \nu}$ the underlying metric. Its line element, which describes an axisymmetric neutron star, can be written in quasi-isotropic coordinates as
\begin{align}
\text{d}s^2
= &-e^{2\nu} \text{d}t^2 + e^{2\psi} r^2 \sin^2 \theta \left(\text{d}\varphi - \omega \text{d} t\right)^2 \nonumber \\
& + e^{2\mu} \left(\text{d}r^2 + r^2 \text{d}\theta^2 \right).
\end{align}
In order to construct neutron star equilibrium solutions (for which we employ
the \texttt{rns}-code \cite{1995ApJ...444..306S, 1998A&AS..132..431N, rns-v1.1}), it is necessary to close the system of equations by providing an equation of state that relates $p$ and $\epsilon$. Since the true EOS is still part of current research and subject to large uncertainties, we resort to commonly used proposals for realistic EOSs for our simulations, which are based on various approaches such as detailed microscopic calculations, relativistic mean-field theory, or Skyrme models. In particular, we will use the proposed piecewise-polytropic approximations \cite{2009PhRvD..79l4032R} of the EOSs APR4, H4, MPA1, MS1, SLy4, and WFF1. While different observational constraints, such as a lower bound for the maximum mass \cite{2010Natur.467.1081D} or the radius of a neutron star \cite{2020NatAs...4..625C}, may rule out certain EOSs, our choice of EOSs is intended to cover a broad range of the parameter space.
The field equations then are expanded up to linear order around an equilibrium solution and then evolved in time. While the full oscillation problem has been studied for many decades, and many different family of modes and relations are known from theoretical computations, we here only focus on some part of the so-called $f$-mode spectrum.
\subsection{Computation of $f$-mode frequencies}\label{interp_f}
While our linear perturbation code allows the determination of $f$-mode frequencies at comparatively low computational expense (when compared to non-linear simulations in full general relativity), that task is still tedious and requires manual tweaking of parameters. Hence, we determine the mode frequencies for a sufficiently large number of equilibrium configurations with rotation rates up to the Kepler limit and use linear interpolation to estimate the frequencies of an arbitrary model (based on the same EOS) in the following way: For each of the considered EOS, we work with a tabulated grid which contains the $f$-mode frequencies as function of the star's gravitational mass $M$ and its equatorial radius $R$. Note that in the rotating case, for a given one-parameter EOS, the star's rotation rate $\Omega$ and other parameters are uniquely related to $M$ and $R$. Thus, one is left with a two-dimensional interpolation to relate a specific star with its $f$-mode frequencies. For the technical reasons described above, our grids have a resolution that can in practice not easily be increased arbitrarily. However, we performed several MCMC analysis, beyond the precision reported later in this work, and find that our grids would need to be refined if $f$-modes with below percent precision are studied.
\subsection{Universal relations}\label{subsec_ur}
The estimation of $f$-mode frequencies, based on interpolation as described in the previous Sec.~\ref{interp_f}, are EOS dependent and rely on the availability of sufficiently dense tabulated input for the interpolation. While this can in principle be done for specific choices for EOS, we also study EOS independent properties, commonly known as universal relations. These are obtained by scaling key stellar properties with oscillation modes or damping times and allow to constrain, some, but not all stellar parameters.
In the following, we provide such an alternative approach by utilizing numerically fitted universal relations in order to compute the frequencies of the co- and counterrotating $f$-modes. Such a relation has been proposed in Ref.~\cite{Kruger:2019zuz} and is given by
\begin{align}\label{universal_relation}
\hat{\sigma}^{i} = \left(c^{i}_1 + c^{i}_2 \hat{\Omega} + c^{i}_3\hat{\Omega}^2 \right) + \left(d^{i}_1 + d^{i}_3 \hat{\Omega}^2 \right) \eta.
\end{align}
Here $i$ denotes the branch of the $f$-mode (co- or counterrotating) and $\hat{\sigma}^{i} = \bar{M} \sigma^{i}/\mathrm{kHz}$ and $\hat{\Omega} = \bar{M} \Omega/\mathrm{kHz}$, where $\bar{M} = M/M_{\astrosun}$. Furthermore, $\eta$ is the effective compactness, which is related to the mass and moment of inertia $I$ of the star via $\eta = \sqrt{\bar{M}^3/I_{45}}$, with $I_{45} = I/10^{45} \mathrm{g \, cm^2}$. The numerical values of the coefficients have been reported in the same work and were obtained by fitting a range of different realistic EOS with stellar sequences \cite{Kruger:2019zuz}.
The simplicity of universal relations does not only enhance the qualitative understanding of parameter dependencies, but it also allows for back-of-the-envelope estimates. The trade-off is that universal relations are not exact, but come with an intrinsic uncertainty due to their simplicity; in this case the error is at percent level as long as $\Omega$ is not too close to the Kepler limit (see Ref.~\cite{Kruger:2019zuz}). Also note that the universal relation comes at the price of requiring three neutron star parameters, whereas the EOS dependent method is satisfied with only two of them; in a certain way, the EOS is encoded as the third parameter in the universal relation.
\subsection{Bayesian framework}\label{Bayesian_framework}
In the following we briefly summarize the basics of the Bayesian analysis as used in this work. It is similar to the analysis for black hole QNMs, which one of the authors reported recently in Ref.~\cite{Volkel:2020daa}.
Bayes' theorem connects the parameters $\theta$ of a model with the observed data $D$ via
\begin{align}
P\left(\theta | D \right) = \frac{ P \left(D | \theta \right) P\left(\theta \right)}{P\left( D \right)}.
\end{align}
Here $P\left(\theta | D \right)$ is the posterior, which describes the probability distribution of the parameters given the data.
It is equal to the likelihood $P \left(D | \theta \right)$, the probability distribution of the data given the parameters, times the prior $P\left(\theta \right)$, the probability distribution of the parameters before looking at the data. The normalization is given by the evidence $P\left( D \right)$, which is the probability of the data itself. In our work we will utilize different levels of informed/uninformed priors for the parameters, which for us here means the distributions are described by narrow/wide Gaussians.
The $f$-mode doublet is the quantity which we consider as data.
Without real analyzed data, one has to make some assumptions for the likelihood.
In the following we assume that a future experiment would provide the modes and that the likelihood can be described by two Normal distributions $\mathcal{N}(\mu_i, \sigma_i)$ with $\mu_i = \nu$ and $\sigma_i = \nu \delta_\nu$. Here $\delta_\nu$ is the dimensionless relative error with which an experiment has determined the $f$-mode frequency $\nu$ (in order to avoid confusion with the standard deviation, we have denoted the $f$-mode frequency here with $\nu$). It can be seen as a free parameter that can be used to qualitatively study the prospects of future gravitational wave detectors.
The actual computation of the posteriors is done using a Markov chain Monte Carlo (MCMC) analysis based on the Python probabilistic programming framework \textsc{PyMC3} \cite{pymc3}.
The MCMC requires to compute $f$-mode frequencies in each step of the chain; since this is, as mentioned above, not a trivial task, we interpolate it from our data set and couple it via a custom theano (Python library) function to the workflow of \textsc{PyMC3}.
\section{Applications}\label{applications}
In this section, we apply the EOS and UR methods to hypothetically observed $f$-mode frequencies coming from a sample of representative neutron stars with the EOS described previously. In all cases we assume that the co- and counterrotating $f$-mode frequencies are Normal distributed, whereas the one sigma relative error and the prior knowledge on $M$ are being varied depending on the application. Since detailed computations on how precise $f$-mode frequencies can be reconstructed from future observations are highly non-trivial, we assume 3\%-5\%, which seems to be in reach, at least for future detectors, e.g., the Einstein Telescope (see Ref.~\cite{Pratten:2019sed} for a recent study). First, we start with the EOS method in Sec.~\ref{app_EOS} and then apply the UR method in Sec.~\ref{app_UR}. We provide results for their joint application in Sec.~\ref{app_EOS_UR}.
\subsection{EOS Method}\label{app_EOS}
The EOS method requires the assumption of a particular EOS which we deem to be the physically ``correct'' one to infer the stellar parameters with the MCMC. We have studied this for multiple EOS and varied the priors of the parameters. In the following, we discuss two representative cases.
In the first application, which can also be seen as proof of principle, we
chose the H4 EOS and take the $f$-mode frequencies from our data set for a
typical neutron star with $M = 1.8 M_\odot$ and $R = 15\,$km;\footnote{We could also have used two typical values for the $f$-mode frequencies rather than using our data set to simulate a proper neutron star model, however, in that case, we would not be able to test our method by comparing how well it reconstructs the chosen neutron star.} this fully specifies the neutron star and hence all its bulk parameters. These are the true neutron star parameters that we want to infer from the ``observed'' $f$-modes by {\it using the same EOS.} The result, which is represented by blue data, is shown in Fig.~\ref{FIG_eos_methods} and demonstrates that the framework gives reasonable results. The red cross and lines indicate our chosen values for $M$ and $R$.
In the second application, we use the same $f$-modes of the first application, but now assume a different EOS to infer the neutron star parameters. This reflects a less optimistic situation, since there is no hint in the observation that would point towards only one unique EOS. In a ``lucky'' case, it might be such that many of the hitherto proposed EOSs do not support an equilibrium configuration that possesses the specifically observed $f$-modes and would thus be ruled out. However, in general, we do not expect such constraints from observed $f$-mode frequencies as the possible range is rather large; instead, we would expect to recover biased parameters describing a different neutron star with coinciding $f$-mode frequencies.
We find and report a case in which an alternative EOS yields biased results, which is represented by the orange data in
Fig.~\ref{FIG_eos_methods}. Here we have used the MPA1 EOS for the parameter
estimation. Looking only at the posteriors, there is no indication as to which
of the two EOS is the ``correct'' one (i.e. the one we have selected to
generate the frequencies), or if maybe yet another EOS would explain the
observation better. In both applications we have assumed that $M$ and $R$ have uninformed priors, and that both $f$-modes have a relative error of $3\%$.
Furthermore we verified that the expected distributions of $f$-modes from both EOS are in agreement with the initially provided ones. This has been done by computing their distributions from draws of both of the here shown posteriors. We find no significant deviation that could be used to easily rule out the ``false'' EOS.
Finally, we also studied how prior knowledge of $M$ changes the above findings. Since precise knowledge of $M$ reduces the given inverse problem to finding the unknown equatorial radius $R$ from two mode frequencies, one would expect that two different EOS become distinguishable. For the above example we report that knowing $M$ within $10\%$ causes strong tensions between posterior and prior of $M$ for the MPA1 EOS, as well as discrepancies for the simulated versus observed $f$-modes. We have verified this by sampling $f$-modes using the posterior distributions of both EOS and compared these with the ones describing the observed $f$-modes. Thus, as expected, one can distinguish the correct from the false EOS.
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{eos_methods.png}
\caption{We compare the EOS method assuming the H4 EOS (blue) and the MPA1 EOS (orange). The diagonal panels show the sampled posterior distribution of $M$ and $R$, while the main panel combines a scatter plot with logarithmic contour lines. The red cross and red line indicate the true H4 parameters that belong to the assumed $f$-mode data being used for both EOS for the parameter estimation. The $f$-mode relative error is assumed to be $3\%$.}
\label{FIG_eos_methods}
\end{figure}
\subsection{Universal Relation Method}\label{app_UR}
In the following, we demonstrate the application of the UR method to a range of different pairs of $f$-mode frequencies, representing neutron star models employing different EOSs. Since the UR method is (intentionally) blind to the underlying EOS, but only depends on a subset of the neutron star parameters, the application is straightforward.
We start with the reconstruction of the rotation rate $\Omega$ and its
relative error. Both are shown as a function of the prior
knowledge of $M$ in Fig.~\ref{FIG_UR}. First, it seems that the relative errors of $\Omega$ are almost independent of the prior knowledge of the mass $M$. Second, the absolute values of the relative errors depend more strongly on the underlying EOS than on the knowledge of $M$. However, in all cases it seems to be possible to constrain the rotation rate $\Omega$, for the given $5\%$ precision for the $f$-modes, almost independently from the prior knowledge of $M$ to within $10\%$ to $20\%$.
In the bottom panel of Fig.~\ref{FIG_UR}, we show the corresponding analysis
for the effective compactness $\eta$, again as function of the standard deviation of the prior Normal
distribution for the mass $M$.
Since we assume only two $f$-mode frequencies as observation, but the UR is a function of three parameters, one can only expect to constrain some part of the parameter space. However, if the mass is known to within a few or tens of percent, the parameter space is already strongly confined. We find bounds on $\eta$ which scale for a wide range roughly linear with the uncertainty of $M$. This scaling is comparable for all of the considered EOSs, which suggests that this scaling is also universal.
In order to compare how the two very different scalings depend on the precision of the provided $f$-modes, we repeated the above analysis with smaller relative errors of $3\%$. As expected, the qualitative scaling remains unchanged, but now provides a bit more stringent bounds.
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{ur_rel_omega.pdf}
\includegraphics[width=1.0\linewidth]{ur_rel_eta.pdf}
\caption{Here we show the relative error of the reconstructed rotation
rate $\Omega$ (top panel) and effective compactness $\eta$ (bottom panel)
as function of the relative error of the mass $M$ for different EOS
(different colors). The dashed lines are the linear interpolation of the
individual points. The central (solid) lines are the mean value of the reconstruction, while the upper and lower lines show the $68\%$ highest credible interval. The black dashed lines indicate a uniform linear scaling. The $f$-mode relative error is assumed to be $5\%$.}
\label{FIG_UR}
\end{figure}
\subsection{Joint EOS and UR Method}\label{app_EOS_UR}
While the EOS method can in principle recover all neutron star parameters, we already argued that the result will, in general, be biased and resemble a different neutron star with similar $f$-modes. In order to quantify how robustly the EOS method is able to constrain the same parameters that can be inferred from the UR method, we have injected a particular $f$-mode pair and applied both methods. The observed $f$-modes have been produced using the H4 EOS, and the EOS method reconstruction is then done via the ``correct'' H4 EOS, the ``wrong'' MPA1 EOS, as well as the UR method.
We report our findings in Fig.~\ref{FIG_eos_vs_ur}, which shows the posterior distribution of the normalized rotation rate $\Omega$ and effective compactness $\eta$ for different prior knowledge on $M$. The normalization is with respect to the true H4 EOS values for the provided $f$-modes.
It is evident that the H4 EOS method (blue lines) and UR method (green lines) yield very similar results for the rotation rate $\Omega$,
while assuming the MPA1 EOS (orange lines) indicates a value that is larger than the correct one. Note that both observations hold independent
of the specific prior knowledge of $M$ assumed here ($30\%$ and $10\%$).
The situation for the effective compactness $\eta$ is qualitatively different. First, the prior knowledge of $M$ plays a big role for the UR method, but is less important for the EOS methods. For those we find that the correctly assumed H4 EOS is almost independent of uncertainties in $M$, while the posterior distribution obtained by the MPA1 EOS is shifted. Note that the rather different scaling behavior of the UR method is in agreement with the findings of Sec.~\ref{app_UR}.
Finally, while the posteriors of $\Omega$ are very smooth, one observes small
``bumps'' for the H4 EOS, e.g. at $\eta/\eta_0 \approx 1.02$. We have verified that this does not originate from a too small sample size of the MCMC sampling, but most likely is an artifact from the finite resolution and particular range of the used H4 $f$-mode data, as described in Sec.~\ref{interp_f}. This directly sets the scale of how precise our currently implemented EOS data can be used to resolve the underlying parameters, which is of order percent level.
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{eos_vs_ur_omega_int.png}
\includegraphics[width=1.0\linewidth]{eos_vs_ur_eta_int.png}
\caption{Here we show the posterior distributions of the rotation rate $\Omega$ (top panel) and effective compactness $\eta$ (bottom panel) normalized to the injected H4 values ($\Omega_0, \eta_0$). Posteriors are obtained by using the EOS method with H4 EOS (blue), the MPA1 EOS (orange), as well as the UR method (green). Solid lines correspond to $30\%$ relative error on the prior mass $M$ and dashed lines to $10\%$. We indicate each mean of the posteriors as vertical lines. The $f$-mode relative error is assumed to be $3\%$.}
\label{FIG_eos_vs_ur}
\end{figure}
\section{Discussion}\label{Discussion}
In the following we discuss our EOS and UR method based findings in Sec.~\ref{discussion_comparison_EOS_ur}, comment on the uniqueness of the inverse problem in Sec.~\ref{discussion_uniquness}, provide some computational details in Sec.~\ref{discussion_computational}, and provide a brief outlook in Sec.~\ref{discussion_outlook}.
\subsection{EOS and UR methods}\label{discussion_comparison_EOS_ur}
Knowing the correct underlying EOS for hypothetically observed $f$-modes is a very optimistic assumption, since the current variety of proposed realistic EOS still produces quite different neutron stars and $f$-mode spectra. Consequently, by assuming we know the exact EOS one would naively expect to find overly optimistic and biased bounds on the reconstructed stellar parameters. By applying the EOS and UR methods in different contexts in Sec.~\ref{applications}, we have quantified several related aspects and discuss them in the following.
In order to quantify the expectation of biased neutron star parameters, we
applied the EOS method twice to the same pair of $f$-mode frequencies in
Sec.~\ref{app_EOS}. In the first case the same EOS was assumed, in the second
case a different realistic EOS was chosen. The correlations and posteriors for the neutron star mass $M$ and the equatorial radius $R$, both presented in Fig.~\ref{FIG_eos_methods}, clearly show bias. Both reconstructed parameters differ by order $10\%$, and most importantly, the posterior distributions do not overlap significantly. This particular example demonstrates that EOS based reconstruction using interpolation schemes and MCMC is in principle very powerful, but only reliable if the correct EOS is known. Results from any realistic application where the underlying EOS is not known have thus to be interpreted with great care, even if the shape of the posteriors do not show any obvious flaws.
Since the UR method relies on a simple analytic function with previously fitted coefficients, it can be applied very efficiently. We therefore used it to infer the effective compactness and rotation rate using $f$-modes provided from various realistic EOS and furthermore assumed a wide range of different prior knowledge of the mass in Sec.~\ref{app_UR}. For the effective compactness we find a roughly EOS independent scaling reaching up to few percent level reconstruction of $\eta$, assuming the mass is known with similar precision. However, the posterior distribution of the rotation rate $\Omega$ is almost independent of the prior on $M$ and slightly more EOS dependent. Because the UR method can only be used to infer a combination of mass $M$, effective compactness $\eta$ and rotation rate $\Omega$, it can---by construction---not be used to directly infer the equatorial radius $R$. The most important finding is the robust and thus mostly EOS independent reconstruction of $\eta$ and $\Omega$. This implies that those two parameters can be reliably extracted and their accuracy is mainly limited by the precision with which the $f$-modes can be measured, as well as the prior of $M$.
In the third application, provided in Sec.~\ref{app_EOS_UR}, we directly compared the reconstruction of effective compactness and rotation rate using the same $f$-modes with the two EOS methods from Sec.~\ref{app_EOS} and the UR method. As is evident from Fig.~\ref{FIG_eos_vs_ur} the latter one yields results comparable to those obtained by assuming the correct EOS. For the rotation rate we confirm what can partially already be expected from the results in Sec.~\ref{app_UR}. The UR method includes the correct value, but with larger uncertainties than those coming from the correct EOS method and are quasi independent of the prior of $M$. Since the wrong EOS method seems to converge towards a value being larger than the correct one, but without obvious flaws, one finds biased results, similar to the ones for $M$ and $R$ reported in Sec.~\ref{app_EOS}.
\subsection{Uniqueness of the Inverse Problem}\label{discussion_uniquness}
Unlike in the non-rotating case, the relation between mass and radius is in general not unique. It will crucially depend on the rotation rate that deforms the neutron star, which implies that there are now two unknowns that need to be determined from observations. Measuring two $f$-mode frequencies allows, in principle for any given one-parameter EOS, to uniquely determine the stellar structure. However, this system of equations is not over determined and can thus not be used to rule out certain EOS, unless the observed $f$-modes naturally can not be explained for any neutron star model of that EOS. The consequence is that for the given information, the inverse problem is not uniquely solvable, and in some cases, the biased parameters simply describe viable neutron stars with different structure properties, but very similar $f$-mode spectrum. Only if additional observations, e.g. the mass or rotation rate, can be provided with high accuracy, it is possible to distinguish among the remaining EOS. Especially the robust reconstruction of the rotation rate seems to indicate that glitch induced GW signals could be particularly valuable.
\subsection{Computational aspects}\label{discussion_computational}
Since realistic $f$-mode computations, as well as Bayesian analysis in terms of a MCMC analysis are both computationally expensive we want to make a few comments.
Reliably and robustly extracting $f$-mode frequencies from a single time
evolution, as done in Ref.~\cite{Kruger:2019zuz}, requires at least a few
dozens of hours on a regular workstation, as well as human interaction to guarantee the robustness of the extracted modes. Having a sufficiently dense parameter space for a specific EOS is therefore cumbersome. However, once it has been obtained, applying the interpolation scheme described in Sec.~\ref{interp_f} allows for almost instant mode computation. Even faster than this is the simple analytic formula describing the universal relation. Depending on the exact size of this parameter space, we find that the UR method speeds up the MCMC analysis by at least one to two orders of magnitude compared to a given EOS method. A typical analysis with sufficient sample size will take a few minutes for the UR method, and at maximum up to a few hours for the EOS method, both on a regular workstation. Especially the application to numerically involved parameter estimation problems demonstrates the enormous advantages of UR based approaches.
\subsection{Outlook}\label{discussion_outlook}
Our analysis only focused on the observation of the co- and counterrotating $f$-mode frequencies, as well as on various choices for informed priors being motivated from different observational scenarios. Other, in principle, available and related quantities are the associated damping times as well as overtones or higher order modes of the acoustic mode spectrum. It can be expected that the reconstruction of stellar parameters would improve if additional modes, e.g. $r$-modes or $g$-modes, were observed, and also the potential to exclude certain EOS. This is particularly interesting for the UR method. As demonstrated in Sec.~\ref{app_UR}, one requires an informed prior on $M$ to obtain reconstructed properties. We argue that this is mainly due to the fact that the universal relations Eq.~\eqref{universal_relation} involve three unknowns $M, \Omega$ and $\eta$. Thus, any additional mode or damping time would lead to at least as many observables as unknowns and therefore, in principle, to a unique reconstruction, modulo intrinsic uncertainties. Another interesting extension of our work could be to perform a Bayesian model comparison to further quantify our observations from Sec.~\ref{app_EOS}, namely how much prior knowledge on neutron star parameters is necessary to distinguish different EOS in cases where they include similar $f$-modes.
\section{Conclusions}\label{Conclusions}
In this work we have studied the inverse problem of rotating neutron stars with realistic equations of state by assuming that measurements of the co- and counterrotating $f$-mode frequencies become available with next generation gravitational wave detectors. These frequencies, along with universal relations, have been reported recently in Ref.~\cite{Kruger:2019zuz,PhysRevD.102.064026}, which defines the theoretical framework of this work. To solve the inverse problem we have conducted a Bayesian analysis by performing MCMC simulations. We provide results using two complementary methods, each of them coming with their own strong and weak aspects.
The EOS method assumes that the underlying EOS is known and is used to recover the stellar parameters. To perform the computationally expensive analysis, we compute $f$-mode frequencies by interpolating from previously obtained multidimensional tables. The UR method is EOS independent and purely based on universal relations. This allows a fully analytic computation of $f$-modes and thus major computational advantages, which are beneficial for a Bayesian analysis. In order to account for different astrophysical scenarios, in which prior knowledge of the neutron star mass $M$ could be obtained from complementary observations, we study different cases of informed priors for $M$. Furthermore we have selected various representative neutron star models of different EOS, computed their $f$-mode spectrum, and applied the UR method.
The main findings of this work are the following. The UR method is powerful in the reconstruction of the effective compactness, depending on the prior knowledge of $M$, and yields EOS independent results up to a few percent. For all of the here studied cases the reconstructed rotation rate is only mildly depending on the prior of $M$ and its posterior distribution is including the correct value. By construction, the UR method can not be used to directly recover the equatorial radius. The latter one can only be reconstructed using the EOS method by assuming a specific EOS or by performing further calculations. Here we report that if the correct EOS is assumed, the one used to produce the observed $f$-modes, the reconstruction is well behaved and converges towards the injected stellar parameters. However, using a different EOS can in general point towards biased parameters, whose posterior distributions do not necessarily show any flaws, but can admit some shifts once more precise data is included. The falsification of realistic EOS, unless some extreme values of neutron star $f$-modes are considered, will in general require additional and very informed knowledge on other stellar parameters, e.g., the rotation rate or the mass. Finally, we expect that further inclusion of additional universal relations will provide a quantitative and computationally feasible framework to solve the inverse problem of rotating neutron stars.
~\\
\begin{acknowledgments}
SV acknowledges financial support provided under the European Union's H2020 ERC Consolidator Grant ``GRavity from Astrophysical to Microscopic Scales'' grant agreement no. GRAMS-815673. CK acknowledges financial support by DFG research Grant No. 413873357.
\end{acknowledgments}
|
\section{Introduction}
Evaporation and condensation are ubiquitous processes which play crucial roles in different areas, from physics and chemistry to biology.
The fundamental understanding of the heat and mass transfer across the liquid-vapor interface is not only of theoretical interest but also of considerable relevance to a wide range of applications.
Examples include the distillation process in high vacuum conditions~\cite{ST13,LX14}, the heating and evaporation of fuel droplets~\cite{S17}, the thermal management of modern electronic devices~\cite{JESGEL19}, and the flash boiling in gasoline direct-injection sprays~\cite{CLPP20}.
The modelling of fluid flows with phase change is challenging since multiple characteristic time and length scales are involved.
The liquid and vapor bulk phases are well described at the macroscopic scale by hydrodynamic equations, while the separating region has a more complex structure, with sharp gradients in flow variables which manifest as jumps at the macroscopic scale.
In non-equilibrium conditions, the region between the two bulk phases comprises two sub-regions, namely the liquid-vapor interface, whose characteristic length is of the order of the molecular diameter~\cite{RW13}, and the tipically larger Knudsen layer that extends a few mean free paths into the vapor phase~\cite{S00}.
The standard kinetic theory studies on evaporation and condensation processes focus on the vapor dynamics in the Knudsen layer next to the liquid phase.
The liquid-vapor interface is simplified to a structureless surface bounding the vapor and the molecular exchanges with the liquid phase are dealt with using a phenomenological boundary condition. More specifically, the molecular flux coming out of the liquid-vapor surface is assumed to have two contributions: atoms spontaneously leaving the liquid bulk (evaporated atoms) and atoms which are backscattered into the vapor after impinging on the liquid-vapor interface (reflected atoms).
The distribution function of evaporating atoms is usually described by a half-Maxwellian while the Maxwell's scattering kernel is the most widely used choice to describe the molecular reflection from the liquid phase.
Two phenomenological coefficients are also introduced to weight the relative importance of the evaporation and condensation fluxes.
In spite of the increasing number of experimental, theoretical and numerical studies, the full understanding of evaporation processes is still lacking and numerous issues remain unresolved. As an example, while the evaporation and condensation coefficients are generally accepted to be lower than unity, experimental measurements for water have provided values that span three orders of magnitude~\cite{PW16}. Furthermore, a temperature discontinuity has been experimentally found across the water liquid-vapor interface that is much larger in magnitude and in the opposite direction to that predicted by classical kinetic theory or nonequilibrium thermodynamics~\cite{fang1999temperature}.
A lot of effort has been thus expended in assessing the physical appropriateness of kinetic boundary conditions at the liquid-vapor interface mostly through molecular dynamics (MD) simulations and mean-field kinetic approximations of simple liquids.
These studies can be roughly grouped into two categories depending on whether the liquid evaporation occurs into vapor~\cite{C36,BL19,TTM99,MFYH04,XSC11,XCS12,KKW14,KSKFW17} or into near vacuum conditions~\cite{ZA97,IYF04, FGL05, FGLS18, HV19}. In this latter process the backscattered vapor component is virtually absent and, therefore, the distribution function of evaporated atoms is evaluated without any ambiguity.
Accurate MD simulations of liquid argon evaporating into vapor have shown that atoms leaving the liquid-vapor interface are distributed with excellent approximation according to a half-Maxwellian while, separately, evaporated and reflected atoms are not~\cite{TTM99,MFYH04}.
More specifically, the velocity distribution function of evaporated atoms shows a fatter tail compared to a Maxwellian while the one of reflected atoms has a reduced tail and larger density in the low velocity region. However, deviations are described differently, namely through the dependency of the condensation/evaporation coefficient on the normal-velocity component of atoms impinging on the liquid surface~\cite{TTM99} or by a drifted half-Maxwellian~\cite{MFYH04}.
Using MD simulations, deviations from a half-Maxwellian have also been revealed in the case of n-dodecane, i.e. more complex molecules, but the distribution function of evaporated atoms is fitted by an anisotropic half-Maxwellian (also referred to as bi-Maxwellian) with the temperature normal to the liquid-vapor interface larger than the parallel one~\cite{XSC11,XCS12}.
The hypothesis that an anisotropic half-Maxwellian provides a better approximation of the distribution function close to the evaporating surface was conjectured for the first time in Ref.~\cite{C36} and the theoretical analysis has been more recently extended to include the case of an evaporation coefficient less than one~\cite{BL19}.
According to review above, most of the MD studies of liquids evaporating into vapor agree that the distribution function of evaporated atoms deviates from the half-Maxwellian, albeit there is no consensus on the fitting function. The simulation results of liquids evaporating into near vacuum conditions are even less conclusive in that deviations have not always been found.
In some MD studies, it has been shown that, in the limit of a low-density vapor, evaporated argon atoms are distributed according to a half-Maxwellian~\cite{ZA97, ZKLSA19}. This conclusion has been explained by the fact that atoms leave the condensed phase due to collisions from the edge of the liquid-vapor interface where the binding energy is negligible, whence the half-Maxwellian shape.
In other MD studies, the velocity distribution function of spontaneously evaporating atoms has been determined to be nearly a half-Maxwellian at low temperatures, but a tail fatter than the one of a half-Maxwellian has been found normal to the liquid-vapor interface at high temperatures~\cite{IYF04,HV19}.
The validity of standard kinetic boundary conditions at the liquid-vapor interface has also been assessed based on the Enskog-Vlasov (EV) equation.
This kinetic equation has the capability of describing both the liquid and vapor phases, including the interface region and has many attractive features. Compare to MD simulations, its numerical solution requires less computational effort and, unlike macroscopic approaches, such as diffuse interface models and sharp interface methods, it can capture the nonequilibrium effects in the liquid-vapor interface and in the Knudsen layer.
The EV equation has been used to study the evaporation into near vacuum of a thin liquid film~\cite{FGL05}, the steady evaporation/condensation flow between two planar liquid films kept at different temperatures~\cite{KKW14}, and, more recently, the evaporation of multi-component substances into vapor and vacuum~\cite{KSKFW17,FGLS18}.
Remarkably, in spite of the simplified treatment of pair correlations in the dense phase, the results obtained by the EV equation closely resemble the ones provide by MD simulations. More specifically, it is generally concluded that evaporated atoms are distributed according to a half-Maxwellian with liquid bulk temperature even though a slight anisotropy between the velocity components normal and parallel to the liquid-vapor interface has been found at high evaporation temperatures.
The foregoing literature review shows that uncertainties remain on the statistical features of atoms spontaneously emitted from the liquid phase. Most of the studies observed that evaporated atoms are not distributed according to a half-Maxwellian but none has systematically evaluated the deviations as a function of the evaporation temperature. Furthermore, the physical mechanism which leads the velocity distribution function of evaporated atoms to differ from a half-Maxwellian has not been convincingly explained.
The aim of the present work is thus twofold.
First, to accurately determine the distribution function of atoms evaporating into near-vacuum conditions by numerically solving the EV equation.
Compared to MD simulations, the numerical solution of this kinetic equation requires less computational effort and, therefore, permits one to get results with the required level of accuracy for the analysis undertaken in this study~\cite{FB17}.
Second, to provide a convincing explanation of the deviations from the half-Maxwellian.
The rest of the paper is organised as follows.
Section~\ref{sec:formulation} outlines the mathematical formulation of the evaporation of a monatomic single-component liquid into near-vacuum condition based on the EV equation and outlines the particle method of solution.
Section~\ref{sec:results} contains a brief description of the computational setup and presents the main results of the paper. In particular, it is shown that the reduced distribution function of spontaneously evaporated atoms are well approximated by a drifted anisotropic half-Maxwellian. Furthermore, it is shown, by numerical evidence and a simple mathematical model, that deviations from an isotropic half-Maxwellian are due to the atoms' collisions in the liquid-vapor interface.
Section~\ref{sec:conclusions} summarises and comments on the main findings of the paper.
\section{Mean-field kinetic theory approach}
\label{sec:formulation}
\subsection{The Enskog-Vlasov equation}
Let us consider a fluid composed of spherical and identical atoms of mass $m$ and diameter $a$ interacting through the Sutherland potential given by a superposition of a hard sphere potential and an attractive soft potential tail:
\begin{equation}
\phi(\rho)=
\begin{cases}
+\infty \,,\quad &\rho < a, \\
\displaystyle{-\phi_a\left( \frac{\rho}{a}\right)^{-\gamma}}\, ,\quad &\rho\geq a,
\end{cases}
\end{equation}
where $\rho=||\bm{r}_1-\bm{r}||$ is the distance between the atoms at position $\bm{r}_1$ and $\bm{r}$ while the two positive constants $\phi_a$ and $\gamma$
define the depth of the potential well and the range of the soft interaction, respectively.
Karkheck et al \cite{KS81} derived the {\it exact} evolution equation for the one-particle distribution function of this system of atoms, $f(\bm{r},\bm{v},t)$, but this equation is of little use since it involves the two-particle distribution function. In order to obtain a closed equation, two simplifying assumptions are introduced in the particle dynamics, namely long-range particle correlations are neglected while short-range particle correlations are approximated by following the Enskog theory originally derived for dense gases.
By adopting these assumptions, the following closed equation for the evolution of $f(\bm{r},\bm{v},t)$ is obtained:
\begin{subequations}
\label{eq:ev}
\begin{equation}
\label{eq:ev_general}
\frac{\partial f}{\partial t}
+\bm{v}\cdot\nabla_{\bm{r}}f+\frac{\mathcal{F}[n]}{m} \cdot \nabla_{\bm{v}}f=\mathcal{C}_E[f],
\end{equation}
where square brackets denote functional dependence.
In Eq.~\eqref{eq:ev_general}, $\mathcal{F}(\bm{r},t)$ is the self-consistent force field generated by the soft attractive tail which reads:
\begin{equation}
\mathcal{F}[n]=\int_{||\bm{r}_1-\bm{r}||>a} \frac{d\phi_a(\rho
)}{d\rho}\frac{\bm{r}_1-\bm{r}}{||\bm{r}_1-\bm{r}||}n(\bm{r_1})d\bm{r_1},
\end{equation}
and the hard-sphere collision integral $\mathcal{C}_E(f,f)$ is given by:
\begin{multline}
\mathcal{C}_E[f]=
a^2\int(\bm{v}_r\cdot\bm{\hat{k}})^+d\bm{v}_1d^2\bm{\hat{k}}
\left\{\chi\left[n\left(\bm{r}+\frac{a}{2}\bm{\hat{k}},t\right)\right] f(\bm{r}+a\bm{\hat{k}},\bm{v}_1^*,t) f(\bm{r},\bm{v}^*,t)-\right.\\
\left.\chi\left[n\left(\bm{r}-\frac{a}{2}\bm{\hat{k}},t\right)\right] f(\bm{r}-a\bm{\hat{k}},\bm{v}_1,t) f(\bm{r},\bm{v},t) \right\},\label{eq:enskog}
\end{multline}
where $(\cdot)^{+}$ indicates that the surface integral is restricted to the half-sphere for which $\bm{v}_r\cdot\bm{\hat{k}}>0$ and $\chi[n]$ is the contact value of the pair correlation function in a hard-sphere fluid in equilibrium with number density $n$.
\end{subequations}
In the Standard Enskog Theory (SET), $\chi$ is approximated by using the value of the pair correlation function in a fluid in uniform equilibrium with the value of the density at the contact point of the two colliding atoms. An approximate, but accurate expression for $\chi_{\mbox{\tiny SET}}$ can be obtained from the equation of state of the hard-sphere fluid proposed by Carnahan and Starling~\cite{CS69}, as:
\begin{equation}
\chi_{\mbox{\tiny SET}}(n)=\frac{1}{nb}\left(\frac{p^{hs}}{n k_B T}-1\right)=\frac{1}{2}\frac{2-\eta}{(1-\eta)^3},
\qquad b=\frac{2\pi a^3}{3},\qquad \eta=\frac{\pi a^3 n}{6}
\label{eq:carnstar}.
\end{equation}
where $p^{hs}$ is the pressure of a system of hard spheres and $k_B$ is the Boltzmann constant.
In the present work, following the Fischer-Methfessel approach~\cite{FM80}, the actual value of the density at the contact point is replaced with the value of the density field averaged over a spherical volume of radius $a$, namely:
\begin{subequations}
\begin{equation}
\chi[n](\bm{r},\bm{r}-a\bm{\hat{k}})=\chi_{\mbox{\tiny SET}}\left(\overline{n}\left(\bm{r}-a \frac{\bm{\hat{k}}}{2}\right)\right),
\end{equation}
where
\begin{equation}
\overline{n}(\bm{r},t)=\frac{3}{4\pi a^3}\int_{\mathbb{R}^3} n(\bm{r}_*,t)w(\bm{r},\bm{r}_*)\,d\bm{r}_*, \hspace{1cm}
w(\bm{r},\bm{r}_*)=\left\{
\begin{array}{cc}
1, &\qquad \|\bm{r}_*-\bm{r}\|<a, \\
0, & \qquad \|\bm{r}_*-\bm{r}\|>a.
\end{array}
\right.
\end{equation}
\end{subequations}
The kinetic equation~\eqref{eq:ev} is usually referred to as the Enskog-Vlasov (EV) equation~\cite{S67,G71,KS81,FGL05,BB19}.
This kinetic equation has been applied to investigate a wide range of two-phase flows, including condensation/evaporation processes~\cite{FGL05,KKW14,FBG19}, liquid menisci in nano-channels~\cite{BFG15}, and its mathematical properties have been extensively studied in connection with the liquid-vapor transition~\cite{TMHH18,BB18}.
\subsection{Particle method of solution}
In this work, the EV equation is solved numerically by an extension of the original Direct Simulation Monte-Carlo (DSMC) scheme to dense fluids \cite{F97b}. A thorough description of the numerical scheme and the analysis of its computational complexity is given in Ref.~\cite{FBG19}.
For EV simulations, the main framework of DSMC scheme used to solve the Boltzmann equation is preserved, with modifications occurring in the collision algorithm due to the nonlocal structure of the Enskog collision operator.
The distribution function is represented by $N$ computational particles:
\begin{equation}
f(\bm{r},\bm{v},t)= \sum_{i=1}^{N} \delta{\left(\bm{r}-\bm{r}_i(t)\right)} \delta(\bm{v}-\bm{v}_i(t)),
\end{equation}
where $\bm{r}_i$ and $\bm{v}_i$ are the position and the velocity of the $i$th particle at time $t$, respectively.
The distribution function is updated by a fractional-step method based on the time-splitting of the evolution operator in two sub-steps, namely free streaming and collision.
In the first stage, the collisions between particles are neglected and the distribution function is advanced from $t$ to $t+\Delta t$ by solving the equation:
\begin{equation}
\label{eq:stage_I}
\frac{\partial f}{\partial t}
+\bm{\bm{v}}\cdot\nabla_{\bm{r}}f+\frac{\mathcal{F}[n]}{m}\cdot\nabla_{\bm{\bm{v}}}f=0,
\end{equation}
which translates into updating the positions and velocities of the computational particles according to:
\begin{subequations}
\begin{align}
\bm{r}_i(t+\Delta t)&=\bm{r}_i(t)+\bm{v}_i\Delta t+\frac{\mathcal{F}[n(t)]}{m}\frac{(\Delta t)^2}{2},\\
\bm{v}_i(t+\Delta t)&=\bm{v}_i(t)+\frac{\mathcal{F}[n(t)]}{m}\Delta t.
\end{align}
\end{subequations}
In the second stage, the short range hard-sphere interactions are considered and the updating rule is given by:
\begin{equation}
f(\bm{r},\bm{v},t+\Delta t)= \tilde{f}(\bm{r},\bm{v},t+\Delta t)+\mathcal{C}_E[\tilde{f}]\Delta t.
\end{equation}
During this stage, particles' positions $\bm{x}_i$ are unchanged while their velocities $\bm{v}_i$ are modified according to stochastic rules which essentially correspond to the Monte Carlo evaluation of the collision integral given by Eq.~\eqref{eq:enskog}.
The macroscopic quantities are obtained by time averaging the particles' microscopic states. Note that for steady flows simulations, as the ones considered below, the averaging time can be long enough to obtain accurate results without the need of using a large number of computational particles.
\section{Evaporation of monatomic liquid into near vacuum}
\label{sec:results}
\begin{figure}
\includegraphics[width=0.65\linewidth]{setup.pdf}
\caption{Schematic of the computational setup. The simulation domain is a finite symmetric interval with perfectly absorbing boundary conditions at $z=\pm 60a$. The initial condition consists of a liquid slab in equilibrium with its vapor placed in the middle of the simulation domain.}
\label{fig:schematic}
\end{figure}
\subsection{Computational setup}
\label{sec:computational_setup}
The schematic of the computational setup is shown in Fig.~\ref{fig:schematic}.
The simulation domain is a finite symmetric interval $z\in[-L,L]$ with $L=60a$ and homogeneous along the $x$ and $y$ axis. Perfectly absorbing boundary conditions are assumed at $z=\pm L$, namely computational particles reaching the boundaries are removed from the simulation domain. Initially, a liquid slab in equilibrium with its vapor is considered in the middle of the simulation domain. This initial state has been obtained by placing a liquid slab in $z\in[-L_s,L_s]$, with $L_s=50a$, gas elsewhere, and using periodic boundary conditions at $z=\pm L$. Afterwards, the system has been let to evolve until the equilibrium steady state is reached.
The distance between the liquid-vapor interfaces and the absorbing surfaces is about $10a$, namely large enough not to influence the interface but, at the same time, sufficiently small to minimise the backscattered flow.
The system is clearly symmetric with respect to the $z$ axis and, therefore, in principle, the computational setup might be simplified by considering only half of the domain. However, this would lead to the need of imposing a specular boundary condition at the center of the liquid slab which, among other things, would require a tricky treatment of the mean force field. Furthermore, results in the two halves of the domain can be superimposed and, therefore, the larger computational effort of simulating all the system is used to reduce the statistical noise.
The simulation domain is divided into 2400 cells with size $\Delta z=a/20$ and the time step is $\Delta t=2\times 10^{-4}\,a/(R T_0)^{1/2}$, where $R=k_B/m$ is the specific gas constant and $T_0$ is the reference temperature.
The number of computational particles is set to $1.2\times10^6$ and
made equal to the number of real atoms by a proper choice of the cross section normal to the non-homogeneous direction $z$.
The interaction parameters were chosen to be $\phi_a/(k_B T_0)=1$ and $\gamma=6$ so as to match the same far field behavior as the 12-6 Lennard-Jones potential~\cite{HBC64}.
The study of evaporation of a liquid slab is carried out at the following temperature values $T_\ell/T_c=\{0.53,0.596,0.663,0.729\}$, where $T_c$ is the critical temperature $T_c/T_0=0.754632$ as a consequence of the interaction parameters chosen~\cite{FGL05}.
In this temperature range, the vapor phase contains a number of particles sufficient to limit the statistical noise of the results (which are obtained by averaging particles' properties) but, at the same time, is dilute enough to behave as an ideal gas. The Andersen thermostat~\cite{A80} is applied in the central part of the liquid slab, $10a$ wide, to prevent the progressive cooling of the system and keep its temperature to the constant value of $T_\ell$. This thermostat is chosen for its simplicity and computational efficiency.
Note that, in principle, during the evaporation into vacuum, the liquid slab is consumed and the liquid-vapor interfaces slowly recede from the absorbing surfaces. The flow is thus an unsteady process and the possibility to evaluate the macroscopic quantities by time-averaging is jeopardized.
In order to circumvent this difficulty, in the present work the evaporation process is studied in a frame of reference fixed relative to the liquid-vapor interface using the following procedure.
The simulation advances until the number of particles reaching the absorbing surfaces equals or exceeds $2N_\ell$, where $N_\ell$ is the average number of particles per cell in the liquid bulk at temperature $T_\ell$. When this happens, the simulation stops. All particles are then moved by $\Delta z$ towards the closest absorbing surface and the empty gap, which forms in the centre of the simulation domain, is filled with $2N_\ell$ particles sampled from a Maxwellian with temperature $T_\ell$. Afterwards, the simulation is restarted. Note that the interface movement is negligibly small during two successive applications of this procedure. In this steady evaporation framework, simulations are run for the time duration $2000\,a/(RT_0)^{1/2}$ (equivalent to $10^7$ iterations).
\subsection{Simulation results}
\subsubsection{Macroscopic quantities}
\label{sec:macro}
\begin{figure}
\begin{subfigure}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{allmacro-040.pdf}
\subcaption{$T_\ell/T_c=0.530$}
\end{center}
\end{subfigure} \hfill
\begin{subfigure}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{allmacro-055.pdf}
\subcaption{$T_\ell/T_c=0.729$}
\end{center}
\end{subfigure}
\caption{Dimensionless density, $na^3$, mean velocity in the $z$-direction, $V_z/(RT_0)^{1/2}$, normal, $T_\perp/T_c$, parallel, $T_\parallel/T_c$, and total, $T/T_c$, temperatures for (a) the lowest and (b) the highest liquid bulk temperatures considered in the simulation campaign.
\label{fig:macro_all}}
\end{figure}
A broad picture of a liquid slab evaporating into near-vacuum is provided by observing the behaviour of the macroscopic quantities across the the liquid and vapor phases shown in Fig.~\ref{fig:macro_all} for the lowest ($T_\ell/T_c=0.53$) and the highest ($T_\ell/T_c=0.729$) temperatures considered. Due to the symmetry of the system, only half of the simulation domain is shown.
The minimum density in the liquid phase is found in the central part of the domain where the liquid is thermostated, while the maximum is reached close to the liquid-vapor interface, with an almost linear profile in between them. This behavior is due to the evaporation cooling of the liquid slab, which causes a temperature decrease in the liquid region which is not thermostated.
After the drop in the liquid-vapor interface, the temperature in the vapor phase levels off but still exhibits a small gradient. This clearly indicates that some collisions between atoms occur in the gap region between the edge of the interface at the vapor side and the absorbing surface.
The normal and parallel temperatures, which are defined based on the velocity components normal and parallel to the liquid-vapor interface, superimpose up to a point located at about half of the interface width. Their separation indicates that the local thermodynamic quasi-equilibrium condition breaks down, marking the beginning of a transition layer which extends a few molecular diameters into the low density region. In the following, we refer to the point at which the separation of temperatures occurs as the `separation point', denoted $z_s$, and the corresponding value of the temperature as the `separation temperature', denoted $T_s$.
After the separation point the parallel temperature profile follows a similar slope as the one in the liquid bulk, while the normal temperature exhibits a sharp drop coupled with an increase in the mean velocity component normal to the liquid-vapor interface.
Note that the normal temperature and the mean velocity in the $z$-component show gradients in a thin layer close to the boundary of the computational domain, especially for the highest temperature case shown in Fig.~\ref{fig:macro_all}(b). This behaviour is not a numerical artifact but it is due to the the discontinuity of the density field at the absorbing surface, which results in a mean force field directed towards the liquid acting on the atoms in this boundary region.
\subsubsection{Velocity distribution function of evaporated atoms}
\label{sec:vdf}
\begin{table}[t!]
\begin{center}
\begin{tabular*}{\columnwidth}{@{\extracolsep{\stretch{1}}}*{5}{lcccc}@{}}
\toprule
& Parameters & {Undrifted anisotropic} & {Drifted isotropic}& {Drifted anisotropic} \\ \hline
$T_\ell/T_c=0.530$ & $\xi/(RT_0)^{1/2}$ & --- & $0.026366\pm0.00388$ & $0.03427\pm0.01364$ \\
& $\theta_\perp/T_c$ & $0.5278\pm0.00254$ & $0.51238\pm0.00176$ & $0.5072\pm0.00852$ \\
& $\theta_\parallel/T_c$ & $0.51238\pm0.00176$ & $0.51238\pm0.00176$ & $0.51238\pm0.00176$ \\
& $\chi^2$ &$0.03016$ &$0.02866$ &$0.02856$\\
\hline
$T_\ell/T_c=0.729$ & $\xi/(RT_0)^{1/2}$ & --- & $0.10903\pm0.00246$ & $0.1628\pm0.0056$\\
& $\theta_\perp/T_c$ &$0.66987\pm0.00348$&$0.59937\pm0.00087$ & $0.55924\pm0.00397$ \\
& $\theta_\parallel/T_c$ & $0.59937\pm0.00087$ & $0.59937\pm0.00087$ & $0.59937\pm0.00087$ \\
& $\chi^2$ &$0.03911$&$0.01163$&$0.00688$ \\
\hline \hline
\end{tabular*}
\end{center}
\caption{Estimates of the fitting parameters, $\xi,\, \theta_\perp$, $\theta_\parallel$ alongside the asymptotic standard error, and the residual sum of squares, $\chi^2$, of the three considered fitting velocity distribution functions for the lowest and the highest liquid bulk temperatures considered in the simulation campaign.}
\label{tab:fits}
\end{table}
The statistical features of spontaneously evaporated atoms have not been systematically assessed until now. According to some studies their velocity distribution function can be approximated by an anisotropic half-Maxwellian~\cite{CLPG11,IYF04,FGLS18} while others pointed out the presence of a velocity drift~\cite{MFYH04}.
Here, the velocity distribution function of evaporated atoms is evaluated based on the particles collected at the absorbing surfaces. As specified in Sec.~\ref{sec:computational_setup}, these surfaces are placed at the edge of the liquid-vapor interface where kinetic boundary conditions can be formulated. Three different Maxwellian-like functional forms are tested for fitting:
\begin{subequations}
\begin{align}
\label{eq:I}
\hspace{-0.36cm} \mbox{Undrifted Anisotropic:}\hspace{0.1cm} & f(v_\parallel,v_\perp) = \frac{C_1}{(2\pi R)^{3/2}\theta_\perp\theta_\parallel^{1/2}}
\exp{ \left[ -\frac{v_\perp^2}{2R \theta_\perp}
-\frac{v_\parallel^2}{2R \theta_\parallel} \right]}, \hspace{0.2cm} v_\perp >0,\\
\label{eq:II}
\hspace{-0.36cm} \mbox{Drifted Isotropic:}\hspace{0.1cm} & f(v_\parallel,v_\perp) = \frac{C_2}{(2\pi R\theta_\parallel)^{3/2}}
\exp{ \left[ -\frac{\left(v_\perp-\xi\right)^2}{2R \theta_\parallel}
-\frac{v_\parallel^2}{2R \theta_\parallel} \right]}, \hspace{0.2cm} v_\perp >0,\\
\label{eq:III}
\hspace{-0.36cm} \mbox{Drifted Anisotropic:}\hspace{0.1cm} & f(v_\parallel,v_\perp) = \frac{C_3}{(2\pi R)^{3/2}\theta_\perp\theta_\parallel^{1/2}}
\exp{ \left[ -\frac{\left(v_\perp-\xi\right)^2}{2R \theta_\perp}
-\frac{v_\parallel^2}{2R \theta_\parallel} \right]}, \hspace{0.2cm} v_\perp >0,
\end{align}
\label{eq:guess}
\end{subequations}
where $C_i$ are constants which make the velocity distribution functions normalised to unity, $v_\perp$, $v_\parallel$ are the velocity components normal and parallel to the liquid-vapor interface, and $\xi$, $\theta_\perp$, $\theta_\parallel$ are the fitting free parameters.
Note that $\xi$ cannot be identified with the mean velocity in the normal $z$-direction since the velocity distribution function is defined only for $v_\perp>0$. Likewise, $\theta_\perp$ and $\theta_\parallel$ cannot be identified with the normal and parallel temperatures.
The values of the fitting parameters are listed in Table~\ref{tab:fits} for the lowest and highest values of the liquid bulk temperatures considered in the simulation campaign. The sum of squares of residuals, $\chi^2$, which is an indicator of the goodness of the fit, is also reported. According to $\chi^2$, the drifted anisotropic half-Maxwellian provides the best approximation, especially for the highest temperature $T_\ell/T_c=0.729$.
\begin{figure}[p!]
\begin{subfigure}[b]{0.76\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{dist_drift_fit_TL040.pdf}
\subcaption{Normal molecular flux}
\end{center}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{dist_drift_fit_TL040_rez.pdf}
\subcaption{Residuals}
\end{center}
\end{subfigure} \hfill
\begin{subfigure}[b]{0.395\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{rezqqnorm040.pdf}
\subcaption{Q-Q plot}
\end{center}
\end{subfigure}
\caption{(a) Reduced normal molecular fluxes normalized to unity of evaporated molecules, (b) residual distribution, and (c) corresponding Q-Q plots of the residuals for the lowest liquid bulk temperature considered in the simulation campaign ($T_\ell/T_c=0.53$). Dashed-dotted line: Fitted undrifted anisotropic half-Maxwellian; Dashed: Fitted isotropic drifted half-Maxwellian; Solid line: Fitted drifted anisotropic half-Maxwellian. Inset: The tail of the distributions in logarithmic scale, with x axis tick values matched to the ticks of the plot.}
\label{fig:log_plot040}
\end{figure}
\begin{figure}[p!]
\begin{subfigure}[b]{0.76\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{dist_drift_fit_TL055.pdf}
\subcaption{Normal molecular flux}
\end{center}
\end{subfigure} \hfill
\begin{subfigure}[b]{0.5\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{dist_drift_fit_TL055_rez.pdf}
\subcaption{Residuals}
\end{center}
\end{subfigure} \hfill
\begin{subfigure}[b]{0.4\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{rezqqnorm055.pdf}
\subcaption{Q-Q plot}
\end{center}
\end{subfigure}
\caption{(a) Reduced normal molecular fluxes normalized to unity of evaporated molecules, (b) residual distribution, and (c) corresponding Q-Q plots of the residuals for the highest liquid bulk temperature considered in the simulation campaign ($T_\ell/T_c=0.729$). Dashed-dotted line: Fitted undrifted anisotropic half-Maxwellian; Dashed: Fitted isotropic drifted half-Maxwellian; Solid line: Fitted drifted anisotropic half-Maxwellian. Inset: The tail of the distributions in logarithmic scale, with x axis tick values matched to the ticks of the plot.}
\label{fig:log_plot055}
\end{figure}
The direct comparison between the fitting curves and the data is shown in Figs.~\ref{fig:log_plot040} and \ref{fig:log_plot055} alongside their residuals and residuals Q-Q scatter plots which provide a more quantitative assessment of the goodness of fit. More specifically, the Q-Q scatter plot compares the quantiles of the residuals distribution functions against one another. The alignment along the bisector of first quadrant indicates that residuals are normally distributed as it should be for the ideal fitting.
In the lowest temperature case shown in Fig.~\ref{fig:log_plot040}, all the proposed distributions superimpose almost perfectly and their Q-Q scatter plots indicate a near normal distribution of residuals. By contrast, in the highest temperature case shown in Figs.~\ref{fig:log_plot055}, deviations can be observed between the different distributions, with the drifted anisotropic half-Maxwellian clearly providing the best fit.
The direct inspection of residuals distributions, Fig.~\ref{fig:log_plot040}b and Fig.~\ref{fig:log_plot055}b, shows that the undrifted anisotropic half-Maxwellian and the drifted isotropic half-Maxwellian do not provide a good fit in the peak and tail regions, while residuals of the drifted anisotropic half-Maxwellian are randomly distributed except for the tail region where a weak pattern is visible.
None of the fitting functions perfectly matches the data but the analysis above indicates that the drifted anisotropic half-Maxwellian provides the best approximation, albeit at the cost of an extra fitting parameter. In the considered range of temperatures, using a distribution function with only one fitting parameter leads to errors in the mean velocity and temperature within one percent but errors rapidly increases if one considers higher-order moments which are more sensitive to the accuracy of the fitting in the tail region. Accordingly, for the remainder of the paper, we assume that the velocity distribution function of evaporated atoms is an anisotropic drifted half-Maxwellian.
\begin{figure}[t!]
\begin{subfigure}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{dist_j0-T040.pdf}
\subcaption{$T_\ell/T_c=0.530$}
\end{center}
\end{subfigure} \hfill
\begin{subfigure}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{dist_j0-T045.pdf}
\subcaption{$T_\ell/T_c=0.596$}
\end{center}
\end{subfigure}
\vspace{0.5cm}
\begin{subfigure}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{dist_j0-T050.pdf}
\subcaption{$T_\ell/T_c=0.663$}
\end{center}
\end{subfigure} \hfill
\begin{subfigure}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{dist_j0-T055.pdf}
\subcaption{$T_\ell/T_c=0.729$}
\end{center}
\end{subfigure}
\caption{Reduced velocity distribution function and molecular flux of evaporated molecules, parallel and normal to the liquid-vapor interface, respectively, at different liquid bulk temperatures. Coloured histograms are the numerical results of the EV equation normalized to unity; solid and dashed lines are their best fits based on a drifted anisotropic half-Maxwellian with parameters $\xi$, $\theta_\perp$, and $\theta_\parallel$.}
\label{fig:vdf}
\end{figure}
The reduced distribution function of the parallel velocity components $v_x$ and $v_y$, denoted $f_{x,y}$, and the reduced flux of the normal-velocity component $v_z$, denoted $v_zf_z$, are shown in Fig.~\ref{fig:vdf} while the values of $\xi$, $\theta_\perp$, and $\theta_\parallel$ are summarised in Table~\ref{table:temp_anis} and plotted in Fig.~\ref{fig:temp_anis}. It can be observed from Fig.~\ref{fig:temp_anis}(b) that as the liquid bulk temperature increases, the difference between the normal and parallel fitting parameters becomes larger. The drift velocity also grows monotonically with the bulk temperature, as it can be seen in Fig.~\ref{fig:temp_anis}(a).
It is worth stressing that, the compressibility factor of the vapor is only slightly less than one in the entire range of evaporation temperatures considered in the simulation campaign (see Table~\ref{table:temp_anis}). Accordingly, the velocity drift and temperature anisotropy pointed out in this study are expected to show up even in vapors whose behaviour is only slightly non-ideal.
Note that, in the previous molecular dynamics studies~\cite{CLPG11,IYF04,MFYH04} the deviations from the undrifted isotropic half-Maxwellian were generically attributed to the collisions of atoms in the liquid-vapor interface. We will examine this argument in more depth in the next section.
\begin{table}
\begin{center}
\begin{tabular}{ccccccc} \hline \hline
$T_\ell/T_c$&\qquad $T_s/T_c$ &\qquad$\xi/(RT_0)^{1/2}$ &\qquad$\theta_\perp/T_c$ &\qquad$\theta_\parallel/T_c$ &\qquad$n_v a^3$ &\qquad$Z$\\ \hline
0.530 &\qquad $0.518$ &\qquad 0.034 &\qquad$0.507$ &\qquad$0.512$ &\qquad 0.0008 &\qquad 0.990 \\
0.596 &\qquad$0.561$ &\qquad 0.081 &\qquad$0.526$ &\qquad$0.547$ &\qquad 0.0018 &\qquad 0.981\\
0.663 &\qquad$0.597$ &\qquad 0.133 &\qquad$0.541$ &\qquad$0.579$ &\qquad 0.0028 &\qquad 0.971\\
0.729 &\qquad$0.628$ &\qquad 0.163 &\qquad$0.559$ &\qquad$0.599$ &\qquad 0.0040 &\qquad 0.961 \\ \hline \hline
\end{tabular}
\end{center}
\caption{Separation temperature, $T_s/T_c$, velocity drift, $\xi/(RT_0)^{1/2}$, normal and parallel temperature, $\theta_\perp/T_c$ and $\theta_\parallel/T_c$, of the velocity distribution function of spontaneously evaporated atoms, number density in the vapor region, $n_v a^3$, and vapor compressibility index, $Z=p_v/(n_v k_B T_\ell)$, being $p_v$ the pressure in the vapor phase and $k_B$ the Boltzmann constant, as a function of the liquid bulk temperatures, $T_\ell/T_c$.}
\label{table:temp_anis}
\end{table}
\begin{figure}[t!]
\begin{subfigure}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{u_drift.pdf}
\subcaption{Velocity drift.}
\end{center}
\end{subfigure} \hfill
\begin{subfigure}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{temp_anis.pdf}
\subcaption{Parallel and normal temperatures.}
\end{center}
\end{subfigure}
\caption{Parameters of the velocity distribution function of spontaneously evaporating atoms as a function of the liquid bulk temperature.}
\label{fig:temp_anis}
\end{figure}
\subsection{Velocity drift and temperature anisotropy}
\label{sec:anis}
\begin{figure}[t!]
\begin{subfigure}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{last1coll-T040-with-macro.pdf}
\subcaption{$T_\ell/T_c=0.530$}
\end{center}
\end{subfigure} \hfill
\begin{subfigure}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{last1coll-T055-with-macro.pdf}
\subcaption{$T_\ell/T_c=0.729$}
\end{center}
\end{subfigure}
\caption{Histogram of the last collision cell of evaporated atoms for the lowest and highest liquid bulk temperatures considered in the simulation campaign. The peak of the distribution is close to the separation point $z_s$ which is marked by the vertical black line.}
\label{fig:lastcoll}
\end{figure}
\subsubsection{Numerical study}
In order to track the origin of the statistical properties of evaporated atoms, their trajectories are traced backwards, to the points where they collide for the last time. The spontaneously evaporating atoms can thus be identified as atoms that, coming from different regions of the domain, reach the absorbing surfaces in free-molecular motion under the action of the conservative mean force field.
These atoms are then divided into two groups, i.e. those whose last collision takes place before and after the separation point. Finally, the distribution functions of these two groups of atoms and the corresponding fitting parameters are evaluated.
Figure~\ref{fig:lastcoll} shows the histograms of the last collision cell of evaporated atoms for $T_\ell/T_c=0.53$ and $T_\ell/T_c=0.729$.
As reported in Table~\ref{tab:before_after}, most of the atoms collected at the absorbing surfaces had their last collision after the separation point, i.e. they represent more than $80\%$ of the total evaporated atoms. Herein, the total number of evaporated atoms is denoted $N_e$.
Remarkably, the velocity drift and temperature anisotropy is much larger for atoms whose last collision was before $z_s$, where the local velocity distribution functions are undrifted and isotropic Maxwellians. It is thus reasonable to focus on this group of atoms to shed light on the mechanism that leads to the velocity drift and temperature anisotropy.
\begin{table}[t!]
\begin{center}
\begin{tabular*}{\columnwidth}{@{\extracolsep{\stretch{1}}}*{9}{c}@{}}
\toprule
& \multicolumn{4}{c}{Before $z_s$} & \multicolumn{4}{c}{After $z_s$} \\ \hline
$T_\ell/T_c$ & $\xi/(RT_0)^{1/2}$ & $\theta_\perp/T_c$ & $\theta_\parallel/T_c$ & $N_{e,z<z_s}/N_e$
& $\xi/(RT_0)^{1/2}$ & $\theta_\perp/T_c$ & $\theta_\parallel/T_c$ & $N_{e,z>z_s}/N_e$ \\ \hline
0.530
& 0.154 & 0.455 & 0.495 & 17.8\%
& 0.036 & 0.512 & 0.518 & 82.2\% \\
0.729
& 0.344 & 0.484 & 0.567 & 11.2\%
& 0.161 & 0.559 & 0.602 & 88.8\% \\ \hline \hline
\end{tabular*}
\end{center}
\caption{Parameters of the velocity distribution function of spontaneously evaporating atoms before and after the separation temperature for the lowest and largest liquid bulk temperatures considered.}
\label{tab:before_after}
\end{table}
\begin{figure}[t!]
\begin{subfigure}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{evap-vs-local-040-contours.pdf}
\subcaption{$T_\ell/T_c=0.53$}
\end{center}
\end{subfigure} \hfill
\begin{subfigure}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{evap-vs-local-055-contours.pdf}
\subcaption{$T_\ell/T_c=0.729$}
\end{center}
\end{subfigure}
\caption{Isocontours of the distribution functions of evaporated atoms (dashed lines) and `potentially' evaporating atoms (solid lines). Atoms originate from locations before the separation point.}
\label{fig:fz_contours}
\end{figure}
In Fig.~\ref{fig:fz_contours} we plot the isocontours of the velocity distribution functions of the evaporated atoms originating before the separation point (group A) and of the `potentially' evaporating atoms (group B). This latter group comprises all the atoms coming from the same locations of the ones of group A under the assumption that they can reach the absorbing surface without suffering any backscattering due to collisions in the interface region.
Both velocity distribution functions are normalized to unity.
The velocity distribution function of the group B is an undrifted isotropic half-Maxwellian with the weighted average temperature of the region where they originate.
This is not unexpected, since, as also proved in Section~\ref{sec:toy_model}, an undrifted Maxwellian velocity distribution function of atoms moving in a conservative force field remains Maxwellian with the same temperature.
The comparison clearly shows that the evaporated atoms have predominantly higher normal velocities and lower parallel velocities. This result can be understood as the result of the interplay between normal and parallel velocities due to collisions in the liquid-vapor interface. Indeed, the higher is the normal-velocity component, the less likely an atom suffers a collision in the liquid-vapor interface because the lower is the time spent in that region. However, at the same time, the larger is the speed, the greater becomes the probability of collisions (see also Eq.~\eqref{eq:colf}). This interplay also explains why atoms coming from the region before the separation point mostly contribute to the velocity drift and temperature anisotropy. Indeed, these atoms must travel a larger distance before reaching the absorbing surface and are thus more likely to be back-scattered by collisions. Likewise, the liquid-vapor interface gets wider as the liquid bulk temperature increases and, therefore, the effect of preferential evaporation of atoms with large normal and lower parallel velocity components is enhanced.
\subsubsection{An instructive model}
\label{sec:toy_model}
\begin{table}[t!]
\begin{center}
\begin{tabular*}{0.6\columnwidth}{@{\extracolsep{\stretch{1}}}*{4}{cccc}@{}}
\toprule
$T_\ell/T_c$ & $\bar{d}/a$ & $\bar{n}a^3$ & $\bar{T}/T_c$ & $\Delta\mathcal{U}/(k_B T_s)$ \\ \hline
0.530 & 9.8 & 0.00157 & 0.44 & 0.01686 \\
0.729 & 12.4 & 0.00554 & 0.497 & 0.04936 \\ \hline \hline
\end{tabular*}
\end{center}
\caption{Numerical values of the parameters which enter in the model for the lowest and highest liquid bulk temperatures considered in the simulation campaign
}
\label{tab:toy}
\end{table}
\begin{figure}[t!]
\begin{subfigure}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{evap-vs-local-040-detailed-contours.pdf}
\subcaption{EV solution, $T_\ell/T_c=0.53$}
\end{center}
\end{subfigure} \hfill
\begin{subfigure}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{evap-toy-040k.pdf}
\subcaption{Model prediction, $T_\ell/T_c=0.53$}
\end{center}
\end{subfigure}
\vspace{0.5cm}
\begin{subfigure}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{evap-vs-local-055-detailed-contours.pdf}
\subcaption{EV solution, $T_\ell/T_c=0.729$}
\end{center}
\end{subfigure} \hfill
\begin{subfigure}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{evap-toy-055k.pdf}
\subcaption{Model prediction, $T_\ell/T_c=0.729$}
\end{center}
\end{subfigure}
\caption{Isocontours of distribution functions of evaporated particles for two evaporation temperatures.}
\label{fig:toy_contours}
\end{figure}
The results presented in the previous section suggest that the velocity drift and the temperature anisotropy of spontaneously evaporating atoms is a consequence of the collisions in the liquid-vapor interface which filter out atoms with a lower normal-velocity component.
In order to support this conclusion, here we approximately evaluate the distribution function of atoms which, being initially at the separation point, cross the liquid-vapor interface, and are collected at the absorbing surface placed at a distance $\bar{d}$. For simplicity, it is assumed that atoms move in a region of constant density, $\bar{n}$, and temperature, $\bar{T}$.
At the separation point, atoms are supposed to be distributed according to an isotropic Maxwellian with zero drift and temperature $T_s$:
\begin{equation}
\label{eq:f_s}
f_s\left(v_{\parallel},v_\perp \right) = \frac{n_s}{\left(2\pi R T_s\right)^{3/2}} \exp\left( -\frac{v_\parallel^2+v_\perp^2}{2RT_s} \right).
\end{equation}
Only a fraction of these atoms can reach the absorbing surface, namely the ones which (i) have a speed sufficiently large to overcome the potential barrier of the mean force field which pushes atoms towards the liquid phase and (ii) are not backscattered due to the collisions with other atoms. Note that:
\begin{itemize}
\item[(i)] The conservation of mechanical energy implies:
\begin{equation}
\label{eq:energy}
\frac{1}{2} m v_s^2 + {\mathcal U}_s = \frac{1}{2} m v_e^2 + {\mathcal U}_e \, \Longrightarrow \, v_{\perp,s}^2 = v_{\perp,e}^2 +v_{\perp,min}^2, \hspace{0.5cm} v_{\perp,min}=\left(2 \frac{\Delta \mathcal{U}}{m}\right)^{1/2},
\end{equation}
where $(v_s,{\mathcal U}_s)$ and $(v_e, {\mathcal U}_e)$ are the atom's velocity and the potential energy of the mean force field at the separation point and at the absorbing surface, respectively, and $\Delta \mathcal{U}=\mathcal{U}_e-\mathcal{U}_s$.
In Eq.~\eqref{eq:energy}, it has been used that $v_{\parallel,s}=v_{\parallel,e}$ since the mean force field acts along the $z$-direction. It is plain that only atoms whose $z$-component of the velocity is larger than $v_{\perp,min}$ can reach the absorbing surface, all the others turn back, under the action of the force field.
\item[(ii)] The number of collisions experienced by an atom in the time interval $\Delta t$ can be assumed to follow a Poisson distribution:
\begin{subequations}
\begin{equation}
Pr(\mbox{$N$ Collisions in the time interval $\Delta t$})=\frac{(\bar{\nu} \Delta t)^Ne^{-\bar{\nu} \Delta t}}{N!},
\end{equation}
where the collision rate $\bar{\nu}$ can be estimated as:
\begin{equation}
\bar{\nu}(v)= \chi(\bar{n}) \frac{\pi a^2 \bar{n}}{\sqrt{\pi}\beta}
\left[e^{-\beta^2 v^2}+\left(2 \beta v+\frac{1}{\beta v}\right)\frac{\pi}{2}\erf(\beta v)\right],
\label{eq:colf}
\end{equation}
where $\beta^2=1/(2R\bar{T})$~\cite{K38}.
Accordingly, the probability that an atom reaches the absorbing surface without suffering any collision is:
\begin{equation}
Pr(\mbox{$0$ Collisions in the time interval $\Delta t$})=e^{-\bar{\nu}(v) \Delta t}=e^{-\bar{\nu}(v)\frac{\bar{d}}{v_\perp}},
\end{equation}
\end{subequations}
where it has been used that the time needed to cross the liquid-vapor interface is $\Delta t=\bar{d}/v_\perp$.
\end{itemize}
By adopting cylindrical coordinates and using (i)-(ii), the net mass flux of atoms across the liquid-vapor interface can be written as:
\begin{equation}
\label{eq:balance_interface}
\int_{0}^{+\infty} dv_{\parallel} \, v_\parallel \int_{v_{\perp,min}}^{+\infty} dv_\perp \, f_s\left(v_{\parallel},v_\perp \right) v_\perp =
\int_{0}^{+\infty} d\tilde{v}_{\parallel} \, \tilde{v}_\parallel \int_{0}^{+\infty} d\tilde{v}_\perp f_e \left(\tilde{v}_{\parallel},\tilde{v}_\perp \right) \tilde{v}_\perp,
\end{equation}
where $f_e$ is given by Eq.~\eqref{eq:f_s} and $f_e$ is the unknown distribution function of evaporated atoms, i.e. atoms collected at the absorbing surface.
Note that in Eq.~\eqref{eq:balance_interface}, the integral over the azimuthal angle cancels out due to the symmetry of the system.
The first integral on the right hand side of Eq.~\eqref{eq:balance_interface} can be simplified by making the change of variables
$\left( v_\perp, v_\parallel\right) \rightarrow \left( (\tilde{v}_\perp^2+\tilde{v}_{\perp,min}^2)^{1/2}, \tilde{v}_\parallel\right)$ suggested by the conservation of mechanical energy, Eq.~\eqref{eq:energy}. The distribution function of atoms collected at the absorbing surface can thus be readily obtained:
\begin{multline}
\label{eq:f_p}
f_e\left(\tilde{v}_{\parallel},\tilde{v}_\perp \right)=
\frac{n_s}{\left(2\pi R T_s\right)^{3/2}}
\exp\left( -\frac{\tilde{v}_{\perp,min}^2}{2RT_s} \right) \\
\exp\left( -\frac{\tilde{v}_\parallel^2+\tilde{v}_\perp^2}{2RT_s} \right)
\exp\left[ -\frac{\bar{\nu}\left( (\tilde{v}^2+v^2_{\perp,min})^{1/2}\right)
\bar{d}}{\left( \tilde{v}_{\perp}^2+v_{\perp,min}^2 \right)^{1/2}} \right].
\end{multline}
Note that if one disregards collisions, $\bar{\nu}=0$, Eq.~\eqref{eq:f_p} simplifies to a half-Maxwellian with the parameter density given by $n_s$ reduced by the Boltzmann factor $\exp{[-\Delta \mathcal{U}/(k_B T_s)]}$.
The dimensionless values of the width of the liquid-vapor interface, $\bar{d}/a$, the mean density, $\bar{n}a^3$, and temperature, $\bar{T}/T_c$, of the liquid-vapor interface are estimated from the simulation results presented in the previous sections. The potential jump $\Delta\mathcal{U}/(k_B T_s)$ is the one that occurs at the edge of a slab having length $\bar{d}/a$ and density $\bar{n}a^3$ placed next to vacuum. The numerical values of all these parameters are listed in Table~\ref{tab:toy}.
Figure~\ref{fig:toy_contours} shows the comparison between the normalised velocity distribution function of evaporated atoms obtained by numerically solving the Enskog-Vlasov equation, and the predictions given by Eq.~\eqref{eq:f_p}. The very good qualitative agreement strongly suggests that the collisions in the liquid-vapor interface may be responsible for deviations from the isotropic half-Maxwellian.
\section{Conclusions}
\label{sec:conclusions}
The Enskog-Vlasov (EV) equation has been used to study the one-dimensional steady evaporation of a monatomic liquid into near vacuum conditions.
The main aim has been to elucidate the statistical features of atoms spontaneously emitted by the liquid bulk. This is a key step in formulating kinetic boundary conditions at the liquid-vapor interface.
The mean-field kinetic theory approach is used in this study since it is by far less computationally demanding than molecular dynamics simulations and permits one to get results with the required high level of accuracy.
The velocity distribution function of spontaneously evaporating atoms are commonly assumed to be an half-Maxwellian at the temperature of the liquid-vapor interface.
By contrast, the main results of this work show that:
\begin{itemize}
\item Evaporated atoms are distributed according to a drifted anisotropic half-Maxwellian. Deviations from the isotropic half-Maxwellian become more pronounced as the liquid bulk temperature increases.
\item The velocity drift and the temperature anisotropy are the results of collisions in the liquid-vapor interface region which preferentially backscatter atoms with a lower normal-velocity component.
\end{itemize}
Note that previous studies based on molecular dynamics simulations have already pointed out that the distribution function of evaporating atoms deviate from an isotropic half-Maxwellian but there hasn't been unanimous agreement on the functional form, i.e. an anisotropic half-Maxwellian~\cite{TTM99,IYF04} or a drifted half-Maxwellian~\cite{MFYH04}. Furthermore, deviations were noticed only for high evaporation temperatures and, accordingly, attributed to the non-ideal vapor behaviour.
The novelty of the present work is thus threefold.
First, it establishes that the drifted anisotropic half-Maxwellian provides the best fitting out of the considered distributions.
Second, it shows that deviations from the half-Maxwellian occur even when the vapor is only slightly non-ideal, namely in conditions where, in principle, the usual kinetic-theory treatment of the evaporation process can still be used~\cite{FBG19}.
Third, it brings evidence, through accurate numerical results and a simple model, that collisions in the interface may explain the velocity drift and the temperature anisotropy.
It is worth stressing that the results above are expected to be valid even in presence of net condensation since it is commonly assumed that the evaporation flux only depends on the state of the fluid in the liquid phase. However, a more detailed study would be needed to assess to what extent the vapor dynamics affects the structure of the liquid-vapor interface and, in turn, the evaporation mass flux.
This work paves the way to a reformulation of boundary conditions usually adopted at the liquid-vapor interface in kinetic theory studies of evaporation/condensation processes. In this respect, an interesting research perspective consists in determining the dependence of the drift-velocity and temperature anisotropy on the properties of the liquid bulk by means of the balance equations of mass, momentum, and energy derived from the EV equation.
\begin{acknowledgments}
This work has been financially supported in the UK by EPSRC grants (EP/N016602/1, EP/R007438/1, EP/S029966/1 \& EP/P031684/1) and the Leverhulme Trust.
\end{acknowledgments}
|
\section{Introduction}
The time evolution of unsteady, electrically conducting fluids in the presence of magnetic field is described by the magnetohydrodynamic (MHD) system. Mathematical theory of MHD is widely applied in astrophysics and thermonuclear reactions, among many others. A simplified and well-accepted model for compressible isentropic MHD system admits the form (see for instance \cite{HW10}):
\begin{equation}\label{pde}
\left\{\begin{aligned}
& \pd _t \varrho+{\rm div} (\varrho \mathbf{u})=0,\\
& \pd _t (\varrho\mathbf{u})+{\rm div} (\varrho \mathbf{u} \otimes \mathbf{u} )+\nabla p(\varrho)={\rm div} \mathbb{S}(\nabla \mathbf{u})+{\rm Curl}\mathbf{B}\times \mathbf{B},\\
& \pd _t \mathbf{B}={\rm Curl} (\mathbf{u} \times \mathbf{B})- \alpha {\rm Curl} ({\rm Curl} \mathbf{B}),\\
& {\rm div} \mathbf{B}=0, \\
\end{aligned}\right.
\end{equation}
in the time-space domain $(0,T)\times \Omega, \; \Omega \subset \mathbb{R}^d,\,d=2,3$. Here, $t\in (0,T)$ and $x\in \Omega$ represent the time and space variables, respectively. We denote by $\varrho=\varrho(t,x)$ the density of the fluids, $\mathbf{u}=\mathbf{u}(t,x)\in \mathbb{R}^d$ the velocity field, $\mathbf{B}=\mathbf{B}(t,x)\in \mathbb{R}^d$ the magnetic field and $p=p(\varrho)$ the scalar pressure. $\alpha>0$ is the resistivity coefficient acting as the magnetic diffusion. $\mathbb{S}=\mathbb{S}(\nabla \mathbf{u})$ stands for the Newtonian viscous stress tensor given by
\
\mathbb{S}(\nabla \mathbf{u})=\mu \left( \nabla \mathbf{u}+\nabla^T \mathbf{u}-\frac{2}{d}{\rm div} \mathbf{u} \mathbb{I} \right)+\lambda {\rm div} \mathbf{u} \mathbb{I},
\]
where $\mu>0$ and $\lambda \geq 0$ are the shear and bulk viscosity coefficients, respectively. The pressure $p$ is assumed to satisfy the isentropic law
\begin{equation}\label{plaw}
p(\varrho)=a \varrho^{\gamma},\,\,a >0,
\end{equation}
where $\gamma>1$ is the adiabatic exponent. System \eqref{pde} is supplemented with the boundary conditions ($\vc{n}$ denotes the unit outward normal on the boundary $\partial \Omega$):
\begin{equation}\label{BCs}
\left\{\begin{aligned}
& \text{either non-slip boundary conditions: }\mathbf{u}|_{\partial \Omega}=\mathbf{0},\quad
\,\,
\mathbf{B}\times \vc{n}|_{\partial \Omega}=\mathbf{0},\\
& \text{ or periodic boundary conditions: }\Omega=\mathcal{T}^d=([0,1]_{\{0,1\}})^d,\\
\end{aligned}\right.
\end{equation}
together with the initial conditions:
\begin{equation}\label{ini_c}
(\varrho,\varrho\mathbf{u},\mathbf{B})|_{t=0}=(\varrho_0,\mathbf{m}_0,\mathbf{B}_0).
\end{equation}
The well-posedness of the compressible MHD system~\eqref{pde} has been studied in several occasions. We refer to Vol'pert and Hudjaev \cite{VoHu} for local well-posedness with positive initial density, Fan and Yu \cite{FY} for local well-posedness with initial vacuum. The existence of global weak solutions with finite energy initial data was studied by Hu and Wang \cite{HW10} for $\gamma>\frac{3}{2}$ in three dimensions. Moreover, we refer to \cite{Ka84,LiZh13} for global well-posedness theory with smallness of initial data, either close to equilibrium state or smallness of initial energy but possibly large oscillations.
The convergence analysis of numerical solutions for compressible viscous fluids was first reported by Karper~\cite{Karper} and further studied by Feireisl et al.~\cite{FL18,FLMS_FVNS,FeLMMiSh}. When a magnetic field is coupled to compressible viscous fluids,
as far as we know, the only result on numerical convergence was done by Ding and Mao~\cite{Ding}. They studied the convergence of a mixed finite volume (FV)-finite element (FE) approximation towards (a suitable subsequence of) weak solutions requiring the technical assumption on the adiabatic exponent $\gamma>3$. Unfortunately it excludes physically relevant parameters, e.g. $\gamma =7/5$ for the diatomic gas. Therefore, it is significant to study the case of ``small" $\gamma$ that covers the physical parameters.
The aim of this paper is to establish a general framework for the convergence analysis of numerical approximations for the compressible MHD system \eqref{pde} for the full range of $\gamma\in(1,\infty)$. As a byproduct, we also prove global solvability to the compressible MHD system for any $\gamma>1$ and large initial data. The strategy is built on the concepts of \emph{dissipative weak solutions} and \emph{consistent approximation}, see respectively Definition~\ref{def_dw} and Definition~\ref{def_ca}.
A dissipative weak (DW) solution allows concentration and oscillation defects that can be controlled by the dissipation defects of the energy stability. It should be stressed that constructing a weak solution for ``small" $\gamma<d/2$ is a challenging task for compressible viscous fluids while the dissipative character of the DW solution allows us to work with ``small" $\gamma$.
Though the DW solution is weaker than the standard finite energy weak solution, it satisfies the weak--strong uniqueness principle, meaning that a DW solution coincides with a classical solution as long as the latter exists. By this argument, the convergence of a numerical solution reduces to the convergence towards a DW solution. Upon realizing a DW solution can be obtained by the limit (discretization parameter $h \to 0$) of a consistent approximation interpreting the stability and consistency properties of the numerical solution, we find a generalized Lax equivalence theorem:
\begin{center}
convergence (to a DW or classical solution) $\Longleftrightarrow$ stability + consistency =:\; consistent approximation.
\end{center}
More specifically, our strategy reads:
\begin{itemize}
\item Proving the weak--strong uniqueness principle in the class of DW solution, see Theorem~\ref{Th1}.
\item Passing to the limit ($h \to 0$) from the consistent approximation to construct a DW solution in the sense of Definition~\ref{def_dw}.
\item Showing that a numerical solution is a consistent approximation in the sense of Definition~\ref{def_ca} that reflects the stability and consistency of the numerical solution.
\end{itemize}
In this paper, the convergence theory is aimed for the class of numerical schemes that preserves the divergence free of the magnetic field weakly. Other properties such as conservation of the mass, positivity of the density, and stability of the total energy are also expected for the numerical solutions.
The rest of the paper is organized as follows.
In Sect.~\ref{sec_main} we first introduce the concepts of DW solutions and consistent approximation. Then we state the main theorems; these are the weak--strong uniqueness property and convergence of a consistent approximation for the compressible MHD system. In Sect.~\ref{dmv_strong} we establish the relative energy inequality in the framework of DW solutions and prove the weak--strong uniqueness principle through the Gronwall-type argument.
In Sect.~\ref{sec_con}, we prove the convergence of a consistent approximation and apply it to the convergence analysis of numerical solutions of two mixed finite volume--finite element methods. The paper ends up with the conclusion.
\section{Main results}\label{sec_main}
\subsection{Preliminaries}
Let $\mathcal{M}\left( \overline{\Omega}\right)$ signify the space of signed Borel measures over $ \overline{\Omega}$ and let $\mathcal{M}^{+}\left( \overline{\Omega}\right)$ be the non-negative ones. Moreover, we recall that $\psi \in L^2_0(\Omega)$ means $\varphi \in L^2(\Omega)$ with zero mean. We now introduce the concept of DW solutions.
\begin{Definition}[DW solution]
\label{def_dw}
A triple $(\varrho,\mathbf{u},\mathbf{B})$ is said to be a dissipative weak solution to the MHD system \eqref{pde}--\eqref{ini_c} provided that
\begin{itemize}
\item{ Regularity of solution
\[
\varrho \geq 0,\,\, \varrho \in L^{\infty}(0,T;L^{\gamma}(\Omega)),\,\,\sqrt{\varrho}\mathbf{u} \in L^{\infty}(0,T;L^{2}(\Omega;\mathbb{R}^d)),
\]
\[
\mathbf{B} \in L^{\infty}(0,T;L^{2}(\Omega;\mathbb{R}^d)),\,\,\nabla \mathbf{u}, {\rm Curl} \mathbf{B} \in L^2(0,T;L^{2}(\Omega;\mathbb{R}^{d\times d}));
\]
}
\item{ The continuity equation
\begin{equation}\label{lbb3}
\int_0^{\tau}\int_{\Omega} \Big ( \varrho \partial_t \varphi + \varrho\mathbf{u}\cdot \nabla \varphi \Big )\dx \dt
=\left[ \int_{\Omega} \varrho \varphi \,{\rm d} {x}\right]_{t=0}^{t=\tau}
\end{equation}
for a.e. $\tau \in (0,T)$, any $\varphi \in C_c^1([0,T]\times \overline{\Omega})$;
}
\item { The momentum equation
\[
\int_0^{\tau}\int_{\Omega} \Big ( \varrho\mathbf{u}\cdot \p_t \bfv + \varrho\mathbf{u}\otimes \mathbf{u}: \nabla \bfv+p(\varrho){\rm div} \bfv-\mathbb{S}(\nabla \mathbf{u}):\nabla \bfv
+\left({\rm Curl} \mathbf{B} \times \mathbf{B}\right)\cdot \bfv
\Big ) \dx \dt
\]
\begin{equation}\label{lbb4}
+
\int_0^{\tau}\int_{\overline{\Omega}} \nabla \bfv :{\rm d} \mu_c(t)\,{\rm d} t
+
\int_0^{\tau}\int_{\overline{\Omega}} \bfv \cdot {\rm d} \mu_B(t)\,{\rm d} t
=\left[ \int_{\Omega} \varrho\mathbf{u} \cdot\bfv \,{\rm d} {x} \right]_{t=0}^{t=\tau}
\end{equation}
for a.e. $\tau \in (0,T)$, any $\bfv \in C_c^1([0,T]\times \Omega;\mathbb{R}^d)$ and some $\mu_c\in L^{\infty}(0,T;\mathcal{M}(\overline{\Omega};\mathbb{R}^{d \times d}_{sym}))$, $\mu_B\in L^{2}(0,T;\mathcal{M}(\overline{\Omega};\mathbb{R}^{d \times d}))$;
}
\item { The Maxwell's equation
\begin{equation}\label{lbb5}
\int_0^{\tau}\int_{\Omega} \Big (\mathbf{B}\cdot \p_t \bfv+ (\mathbf{u} \times \mathbf{B})\cdot {\rm Curl} \bfv
-\alpha {\rm Curl} \mathbf{B} \cdot {\rm Curl} \bfv \Big ) \dx \dt
=\left[ \int_{\Omega} \mathbf{B}\cdot\bfv \,{\rm d} {x} \right]_{t=0}^{t=\tau}
\end{equation}
for a.e. $\tau \in (0,T)$, any $\bfv \in C_c^1 ([0,T]\times \overline{\Omega};\mathbb{R}^d),\,\bfv\times \vc{n}|_{\p \Omega}=\mathbf{0}$};
\item { Divergence free of magnetic field
\begin{equation}\label{lbb6}
\int_{\Omega} \mathbf{B} \cdot \nabla \varphi\,{\rm d} {x}
=0
\end{equation}
for any
${\varphi} \in C^1( \overline{\Omega}) \cap L_0^2(\Omega) $};
\item { Balance of total energy
\[
\int_{\Omega} \left[\frac{1}{2}\varrho |\mathbf{u}|^2+\frac{1}{2}|\mathbf{B}|^2+\mathcal{H}(\varrho)\right](\tau,x)\,{\rm d} {x}
+\int_0^{\tau}\int_{\Omega} \left ( \mathbb{S}(\nabla\mathbf{u}):\nabla \mathbf{u}
+\alpha \left| {\rm Curl} \mathbf{B} \right|^2
\right )\dx \dt
\]
\begin{equation}\label{lbb7}
+\int_{\overline{\Omega} }{\rm d}\mathfrak{D}(\tau)
+ \int_0^{\tau}
\int_{\overline{\Omega}} {\rm d}\mathfrak{C}
\leq
\int_{\Omega}\left[ \frac{1}{2}\frac{|\mathbf{m}_0|^2}{\varrho_0}+\frac{1}{2}|\mathbf{B}_0|^2+\mathcal{H}(\varrho_0)\right]\,{\rm d} {x}
\end{equation}
for a.e. $\tau \in (0,T)$ and some $\mathfrak{D}\in L^{\infty}(0,T; \mathcal{M}^{+}(\overline{\Omega}))$, $\mathfrak{C} \in \mathcal{M}^{+}([0,T]\times \overline{\Omega})$
}
\item { Compatibility conditions
\begin{equation}\label{lbb8}
\int_0^T \psi(t)\int_{\overline{\Omega}}{\rm d}|\mu_c(t)|\,{\rm d} t \lesssim \int_0^T \psi(t)\int_{\overline{\Omega}}{\rm d}\mathfrak{D}(t)\,{\rm d} t ,
\end{equation}
\begin{equation}\label{lbb8-2}
\int_0^T \psi(t)\int_{\overline{\Omega}}{\rm d}|\mu_B(t)|\,{\rm d} t \lesssim \frac{1}{\epsilon}\int_0^T \psi(t)\int_{\overline{\Omega}}{\rm d}\mathfrak{D}(t)\,{\rm d} t +\epsilon \int_0^T
\int_{\overline{\Omega}} \psi(t) {\rm d}\mathfrak{C}
\end{equation}
for any $\epsilon>0,\psi \in C([0,T]),\psi \geq 0$.
}
\end{itemize}
\end{Definition}
\medskip
\begin{Remark}
In (\ref{lbb4}), the measure $\mu_c$ denotes the oscillation and concentration defects due to the nonlinear terms $\varrho \mathbf{u}\otimes \mathbf{u}$ and $p(\varrho)$, while $ \mu_B$ reflects that of ${\rm Curl} \mathbf{B} \times \mathbf{B}$. In (\ref{lbb7}), the non-negative measure $\mathfrak{D}$ represents the defects from the total energy $\frac{1}{2}\varrho |\mathbf{u}|^2+\frac{1}{2}|\mathbf{B}|^2+\mathcal{H}(\varrho)$, while $\mathfrak{C}$ means the defects from the dissipative terms $
\mathbb{S}(\nabla\mathbf{u}):\nabla \mathbf{u}+\alpha \left| {\rm Curl} \mathbf{B} \right|^2$. Furthermore, these measures are interrelated through the compatibility conditions (\ref{lbb8}) and (\ref{lbb8-2}), which play a crucial role in proving the weak--strong uniqueness principle.
\end{Remark}
Next, let us define the concept of consistent approximation in terms of the stability and consistency of a numerical solution.
\begin{Definition}[Consistent approximation]\label{def_ca}
Let the differential operator $\nabla_h$ (resp. ${\rm div}_h$ and ${\rm Curl}_h$) be compatible\footnote{A simple example of such compatibility is that $\nabla_h = \nabla$ element-wisely, see~\cite[Section 11.4]{FeLMMiSh}.} with $\nabla$ (resp. ${\rm div}$ and ${\rm Curl}$) in the sense of \cite[Definition 5.8]{FeLMMiSh}.
We say a numerical approximation $(\varrho_h,\mathbf{u}_h,\vB_h)$ of the MHD system~\eqref{pde}--\eqref{ini_c} is a consistent approximation if the following stability and consistency conditions hold:
\begin{enumerate}
\item {\bf Stability.}
The numerical approximation is stable in the sense that
\begin{equation}\label{es}
E_h(\tau) + \int_0^\tau \intOB{ \S(\nabla_h \mathbf{u}_h) : \nabla_h \mathbf{u}_h
+ \alpha | {\rm Curl} \vB_h |^2 } \leq E_h(0), \quad \forall \tau \in [0,T],
\end{equation}
where $ E_h = \intO{ \left(\frac{1}{2} \varrho_h \abs{\Pi_h \mathbf{u}_h }^2 + \mathcal{H}(\varrho_h )+\frac{1}{2} \abs{\vB_h }^2 \right) } $ the total energy, $\mathcal{H}(\varrho_h) = \frac{a}{\gamma-1}\varrho_h^\gamma$ the pressure potential, and $\Pi_h$ is either identity or a piecewise constant projection operator satisfying $\norm{\Pi_h \mathbf{u}_h - \mathbf{u}_h }_{L^2(\Omega)} \leq h \norm{\nabla_h \mathbf{u}_h}_{L^2(\Omega)}$.
\item {\bf Consistency.}
The numerical approximation is consistent if:
\begin{subequations}\label{cP}
$\bullet$ {\bf Continuity equation.} It holds for any $\phi \in C_c^1([0,T) \times \Ov{\Omega})$ that
\begin{equation} \label{cP1}
\intTOB{ \varrho_h \partial_t \phi + \varrho_h \mathbf{u}_h \cdot \nabla \phi }
=- \intO{ \varrho_h^0 \phi(0,\cdot) }
+ e_{1,h}[\phi],
\end{equation}
where $e_{1,h}[\phi] \to 0 \mbox{ as } h \to 0 \mbox{ for any } \phi \in C_c^M([0,T) \times \Ov{\Omega}) \mbox{ for some integer } M\geq 1;$
$\bullet$ {\bf Balance of momentum.} It holds for any $\bfv \in C_c^1([0,T) \times \Omega; \mathbb{R}^d)$ that
\begin{equation} \label{cP2}
\begin{aligned}
\intTOB{ \varrho_h \Pi_h \mathbf{u}_h \cdot \partial_t \bfv + \varrho_h \Pi_h \mathbf{u}_h \otimes \mathbf{u}_h : \nabla \bfv + p_h {\rm div} \bfv
- \S( \nabla_h \mathbf{u}_h) : \nabla \bfv }
\\ +\intTO{ ( {\rm Curl}_h \vB_h \times \vB_h ) \cdot \bfv}
= - \intO{ \varrho_h^0 {\Pi_h \mathbf{u}_h^0} \cdot \bfv(0,\cdot) }
+e_{2,h}[\bfv]
\end{aligned}
\end{equation}
where $e_{2,h}[\bfv] \to 0$ as $ h \to 0$ for any $ \bfv \in C_c^M([0,T) \times \Omega; \mathbb{R}^d)$ for some integer $M\geq 1$;
$\bullet$ {\bf The Maxwell equation} It holds for any $\mathbf{C} \in C_c^1([0,T) \times \Ov{\Omega}; \mathbb{R}^d)$, $\mathbf{C} \times \vc{n}|_{\p \Omega}=\mathbf{0}$ that
\begin{equation} \label{cP3}
\intTOB{ \vB_h \cdot \pd _t \mathbf{C} + (\mathbf{u}_h \times \vB_h - \alpha {\rm Curl}_h \vB_h ) \cdot {\rm Curl} \mathbf{C}}
=- \intO{ \vB_h^0 \cdot \mathbf{C}(0,\cdot)} +e_{3,h}[\mathbf{C}]
\end{equation}
where $e_{3,h}[\mathbf{C}] \to 0$ as $ h \to 0$ for any $ \mathbf{C} \in C_c^M([0,T) \times \Omega; \mathbb{R}^d)$ for some integer $M\geq 1$;
$\bullet$ {\bf Weakly divergence free of magnetic field.} It holds for any $\psi \in C^1(\Ov{\Omega})\cap L^2_0(\Omega)$ that
\begin{equation}\label{cP4}
\intO{ \vB_h \cdot \nabla \psi} = e_{4,h}[\psi]
\end{equation}
where $e_{4,h}[\psi] \to 0$ as $ h \to 0$ for any $ \psi \in C^M(\Ov{\Omega})\cap L^2_0(\Omega)$ for some integer $M\geq 1$.
\end{subequations}
\end{enumerate}
\end{Definition}
\medskip
\subsection{Main theorems}
Our main results in this paper are summarized in the following two theorems. The first one is concerned with the stability of classical solutions within DW solutions.
\begin{Theorem}[weak--strong uniqueness]\label{Th1}
Let $(\varrho,\mathbf{u},\mathbf{B})$ be a DW solution to \eqref{pde}--\eqref{ini_c} with the initial data $(\varrho_0,\varrho_0\mathbf{u}_0,\mathbf{B}_0)$.
Suppose that $(\widetilde{\varrho},\widetilde{\mathbf{u}},\widetilde{\mathbf{B}})$ is a classical solution to \eqref{pde}--\eqref{ini_c} starting from the same initial data with $\varrho_0>0,{\rm div} \mathbf{B}_0=0$ and belonging to the class
\begin{equation}\label{STC}
\varrho \in C^1([0,T]\times \overline{\Omega}),\,\,\,\, \mathbf{u},\mathbf{B} \in
C^2([0,T]\times \overline{\Omega};\mathbb{R}^d).
\end{equation}
Then $\mu_c=\mathbf{0},\,\,\,\, \mu_B=\mathbf{0},\,\,\,\, \mathfrak{D}=0,\,\,\,\, \mathfrak{C}=0$ and
\[
\varrho=\tilde{\varrho},\,\,\,\, \mathbf{u}=\tilde{\mathbf{u}},\,\,\,\,\mathbf{B}=\tilde{\mathbf{B}},
\text{ in } (0,T)\times \Omega.
\]
\end{Theorem}
\medskip
The second one gives the convergences of numerical solutions.
\begin{Theorem}[Convergence]\label{Th2}
Let $(\varrho_h,\mathbf{u}_h,\vB_h)$ be a consistent approximation of the MHD system in the sense of Definition~\ref{def_ca}. Then the following convergences hold:
\begin{enumerate}
\item {\bf Convergence to DW solution.}
There exists a subsequence of $(\varrho_h,\mathbf{u}_h,\vB_h)$ not relabelled such that
\begin{align*}
\varrho_h \rightarrow & \; \varrho \mbox{ weakly-(*) in } L^\infty(0,T;L^\gamma(\Omega)),\\
\mathbf{u}_h \rightarrow & \; \mathbf{u} \mbox{ weakly in } L^2 ((0,T)\times \Omega; \mathbb{R}^d)\\
\vB_h \rightarrow & \; \mathbf{B} \text{ weakly-(*) in } L^\infty(0,T; L^2( \Omega; \mathbb{R}^d)),
\end{align*}
where the triple $(\varrho, \mathbf{u},\mathbf{B})$ represents a DW solution to the MHD system in the sense of Definition~\ref{def_dw}.
\item {\bf Convergence to classical solution.}
In addition, let the MHD system \eqref{pde}--\eqref{ini_c} admit a classical solution in the class \eqref{STC}. Then the above weak limit is unconditional (no need of subsequence but the whole sequence) and the limit quantity $(\varrho,\mathbf{u},\mathbf{B})$ coincides with the classical solution.
\end{enumerate}
\end{Theorem}
\medskip
\section{Weak--strong uniqueness}\label{dmv_strong}
In this section, we aim to prove the weak--strong uniqueness theory (Theorem~\ref{Th1}) for the DW solutions given in Definition~\ref{def_dw}. To this end, we invoke the relative entropy functional to measure the distance between a DW solution and a classical solution. For definiteness, we shall proceed in case of Dirichlet boundary conditions, while the periodic case can be carried out exactly in the same way.
\subsection{Relative energy inequality}\label{rela}
The goal of this part is to establish the relative energy inequality in the context of DW solutions.
Let $(\varrho,\mathbf{u},\mathbf{B})$ be a DW solution to (\ref{pde})--(\ref{ini_c}) and $(r,\mathbf{U},\mathbf{b})$ be subject to
\begin{equation*}\label{lbn1}
\left\{\begin{aligned}
& r \in C^1([0,T]\times \overline{\Omega}),\,\,\,\,r>0,\\
& \mathbf{U}\in C^1([0,T]\times \overline{\Omega};\mathbb{R}^d),\,\,\,\,\mathbf{U}|_{\partial \Omega}=\vc{0},\\
& \mathbf{b}\in C^1([0,T]\times \overline{\Omega};\mathbb{R}^d),\,\,\,\,\mathbf{b} \times \vc{n}|_{\partial \Omega}=\mathbf{0},\,\,\,\,{\rm div} \mathbf{b}=0.\\
\end{aligned}\right.
\end{equation*}
Inspired by \cite{FNS} in the context of finite energy weak solutions, we introduce the \emph{relative entropy} in the framework of DW solutions
\begin{equation*}\label{lbn2}
\mathcal{E}\Big((\varrho,\mathbf{u},\mathbf{B})\,\Big|\,(r,\mathbf{U},\mathbf{b})\Big)(\tau)
=\int_{\Omega}\left[\frac{1}{2}\varrho |\mathbf{u}-\mathbf{U}|^2+\frac{1}{2} |\mathbf{B}-\mathbf{b}|^2
+\mathcal{H}(\varrho)-\mathcal{H}(r)-\mathcal{H}'(r)(\varrho-r)
\right](\tau,\cdot) \,{\rm d} {x}.
\end{equation*}
Notice that we may rewrite the relative entropy in an equivalent form as follows
\begin{equation}\label{lbn3}
\begin{split}
& \mathcal{E}\Big((\varrho,\mathbf{u},\mathbf{B})\,\Big|\,(r,\mathbf{U},\mathbf{b})\Big)(\tau)
=\int_{\Omega}\left( \frac{1}{2}\varrho |\mathbf{u}|^2+\frac{1}{2}|\mathbf{B}|^2+\mathcal{H}(\varrho) \right) \,{\rm d} {x}
+\int_{\Omega}\frac{1}{2} \varrho |\mathbf{U}|^2\,{\rm d} {x}
\\& -\int_{\Omega}\varrho \mathbf{u} \cdot \mathbf{U} \,{\rm d} {x}
-\int_{\Omega} \mathbf{B}\cdot \mathbf{b} \,{\rm d} {x}
-\int_{\Omega} \varrho \mathcal{H}'(r) \,{\rm d} {x}
+\int_{\Omega} p(r) \,{\rm d} {x}+\frac{1}{2}\int_{\Omega}|\mathbf{b}|^2 \,{\rm d} {x}.
\end{split}
\end{equation}
The crucial observation is that the integrals on the right-hand side of (\ref{lbn3}) can be expressed through (\ref{lbb3})--(\ref{lbb7}) with suitable choices of test functions. To handle the density-dependent terms, we first test the continuity equation (\ref{lbb3}) with $\frac{1}{2}|\mathbf{U}|^2$ to derive
\begin{equation}\label{lbn4}
\left[ \int_{\Omega} \frac{1}{2} \varrho |\mathbf{U}|^2 \,{\rm d} {x} \right]_{t=0}^{t=\tau}
=\int_0^{\tau}\int_{\Omega} \Big ( \varrho \mathbf{U}\cdot \pd _t \mathbf{U}+ \varrho\mathbf{u} \cdot \nabla \mathbf{U} \cdot \mathbf{U} \Big )\dx \dt.
\end{equation}
Moreover, we take $\mathcal{H}'(r)$ as a test function in (\ref{lbb3}) to find
\begin{equation}\label{lbn5}
\left[ \int_{\Omega} \varrho \mathcal{H}'(r) \,{\rm d} {x} \right]_{t=0}^{t=\tau}
=\int_0^{\tau}\int_{\Omega} \Big ( \varrho\pd _t \mathcal{H}'(r)+ \varrho\mathbf{u} \cdot \nabla \mathcal{H}'(r) \Big )\dx \dt.
\end{equation}
Upon choosing $\mathbf{U}$ as a test function of the momentum equation (\ref{lbb4}), we observe that
\[
\left[ \int_{\Omega} \varrho\mathbf{u} \cdot \mathbf{U} \,{\rm d} {x} \right]_{t=0}^{t=\tau}
=\int_0^{\tau}\int_{\Omega} \Big ( \varrho\mathbf{u} \cdot \pd _t \mathbf{U} + \varrho\mathbf{u}\otimes \mathbf{u} : \nabla \mathbf{U}+p(\varrho){\rm div} \mathbf{U}
\]
\begin{equation}\label{lbn6}
-\mathbb{S}(\nabla\mathbf{u} ):\nabla \mathbf{U}
+\left({\rm Curl} \mathbf{B} \times \mathbf{B} \right)\cdot \mathbf{U}
\Big ) \dx \dt
+
\int_0^{\tau}\int_{\overline{\Omega}} \nabla \mathbf{U} :{\rm d} \mu_c(t)\,{\rm d} t
+
\int_0^{\tau}\int_{\overline{\Omega}} \mathbf{U} \cdot {\rm d} \mu_B(t)\,{\rm d} t .
\end{equation}
Next, to calculate the term involved with the magnetic field, we choose $\mathbf{b}$ as a test function in (\ref{lbb5}) to deduce that
\begin{equation}\label{lbn7}
\left[ \int_{\Omega} \mathbf{B} \cdot\mathbf{b} \,{\rm d} {x} \right]_{t=0}^{t=\tau}
=\int_0^{\tau}\int_{\Omega} \Big (\mathbf{B} \cdot \pd _t \mathbf{b}+ (\mathbf{u} \times \mathbf{B}) \cdot {\rm Curl} \mathbf{b}
-\alpha{\rm Curl} \mathbf{B} \cdot {\rm Curl} \mathbf{b} \Big ) \dx \dt.
\end{equation}
Finally, combining (\ref{lbn4})--(\ref{lbn7}) with the balance of total energy (\ref{lbb7}), we obtain the \emph{relative energy inequality} as follows
\[
\left[\mathcal{E}\Big((\varrho,\mathbf{u},\mathbf{B})\,\Big|\,(r,\mathbf{U},\mathbf{b})\Big)\right]_{t=0}^{t=\tau}
+\int_0^{\tau}\int_{\Omega} \mathbb{S}(\nabla\mathbf{u}-\nabla \mathbf{U}):(\nabla \mathbf{u}-\nabla \mathbf{U})
\dx \dt
\]
\[
+\alpha \int_0^{\tau}\int_{\Omega} \left| {\rm Curl} (\mathbf{B} -\mathbf{b})\right|^2\dx \dt
+\int_{\overline{\Omega} }{\rm d}\mathfrak{D}(\tau)
+ \int_0^{\tau}
\int_{\overline{\Omega}} {\rm d}\mathfrak{C}
\]
\[
\leq -\int_0^{\tau}\int_{\Omega} \Big ( \varrho\mathbf{u} \cdot \pd _t \mathbf{U} + \varrho\mathbf{u}\otimes \mathbf{u}: \nabla \mathbf{U}+p(\varrho){\rm div} \mathbf{U}\Big ) \dx \dt
\]
\[
+\int_0^{\tau}\int_{\Omega} \Big ( \varrho \mathbf{U}\cdot \pd _t \mathbf{U}+ \varrho\mathbf{u} \cdot \nabla \mathbf{U} \cdot \mathbf{U} \Big )\dx \dt
+\int_0^{\tau}\int_{\Omega} \mathbb{S} (\nabla \mathbf{U}):
\left(
\nabla \mathbf{U}-\nabla \mathbf{u}
\right) \dx \dt
\]
\[
+\int_0^{\tau}\int_{\Omega}
\left[
\left( 1-\frac{\varrho}{r} \right)p'(r)\pd _t r
-\varrho\mathbf{u}\cdot \frac{p'(r)}{r} \nabla r
\right]
\dx \dt
+\alpha \int_0^{\tau}\int_{\Omega}
{\rm Curl} \mathbf{b} \cdot
\left(
{\rm Curl} \mathbf{b}-{\rm Curl} \mathbf{B}
\right)
\dx \dt
\]
\[
+\int_0^{\tau}\int_{\Omega}
\Big(
\pd _t \mathbf{b}\cdot
\left(
\mathbf{b}-\mathbf{B}
\right)
-(\mathbf{u} \times \mathbf{B}) \cdot {\rm Curl} \mathbf{b}
\Big)
\dx \dt
-\int_0^{\tau}\int_{\Omega}
\left({\rm Curl} \mathbf{B} \times \mathbf{B} \right)\cdot \mathbf{U}
\dx \dt
\]
\begin{equation}\label{lbn8}
-\int_0^{\tau}\int_{\overline{\Omega}} \nabla \mathbf{U} :{\rm d} \mu_c(t)\,{\rm d} t
-\int_0^{\tau}\int_{\overline{\Omega}} \mathbf{U} \cdot {\rm d} \mu_B(t)\,{\rm d} t .
\end{equation}
\subsection{Weak--strong uniqueness principle}\label{west}
The aim of this part is to estimate the right hand side of \eqref{lbn8} towards the proof of weak--strong uniqueness principle. The strategy consists of the following steps:
\begin{itemize}
\item{
Setting the classical solution $(\widetilde{\varrho},\widetilde{\mathbf{u}},\widetilde{\mathbf{B}})$ as the test function $(r,\mathbf{U},\mathbf{b})$ in the relative energy inequality (\ref{lbn8});
}
\item{
Estimating each term on the right-hand side of the relative energy inequality \eqref{lbn8} in a suitable manner;
}
\item{
Applying Gronwall-type argument to derive the expected results.
}
\end{itemize}
Let $(\widetilde{\varrho},\widetilde{\mathbf{u}},\widetilde{\mathbf{B}})$ be a classical solution to (\ref{pde})--(\ref{ini_c}) starting from the smooth initial data $(\varrho_0,\mathbf{u}_0,\mathbf{B}_0)$ with strictly positive $\varrho_0$ and ${\rm div} \mathbf{B}=0$. Let $(\varrho,\mathbf{u},\mathbf{B})$ be a DW solution to (\ref{pde})--(\ref{ini_c}) emanating from $(\varrho_0,\varrho_0\mathbf{u}_0,\mathbf{B}_0)$. It follows from (\ref{lbn8}) that
\[
\mathcal{E}\Big((\varrho,\mathbf{u},\mathbf{B})\,\Big|\,(\widetilde{\varrho},\widetilde{\mathbf{u}},\widetilde{\mathbf{B}})\Big)(\tau)
+\int_0^{\tau}\int_{\Omega} \mathbb{S}(\nabla\mathbf{u}-\nabla \widetilde{\mathbf{u}}):(\nabla \mathbf{u}-\nabla \widetilde{\mathbf{u}})
\dx \dt
\]
\[
+\alpha \int_0^{\tau}\int_{\Omega} | {\rm Curl} (\mathbf{B} -\widetilde{\mathbf{B}})|^2\dx \dt
+\int_{\overline{\Omega} }{\rm d}\mathfrak{D}(\tau)
+ \int_0^{\tau}
\int_{\overline{\Omega}} {\rm d}\mathfrak{C}
\]
\[
\leq -\int_0^{\tau}\int_{\Omega} \Big ( \varrho\mathbf{u}\cdot \pd _t \widetilde{\mathbf{u}} + \varrho\mathbf{u}\otimes \mathbf{u}: \nabla \widetilde{\mathbf{u}}+p(\varrho){\rm div} \widetilde{\mathbf{u}}\Big ) \dx \dt
\]
\[
+\int_0^{\tau}\int_{\Omega} \Big ( \varrho\widetilde{\mathbf{u}}\cdot \pd _t \widetilde{\mathbf{u}}+ \varrho\mathbf{u} \cdot \nabla \widetilde{\mathbf{u}} \cdot \widetilde{\mathbf{u}} \Big )\dx \dt
+\int_0^{\tau}\int_{\Omega} \mathbb{S} (\nabla \widetilde{\mathbf{u}}):
\left(
\nabla \widetilde{\mathbf{u}}-\nabla \mathbf{u}
\right) \dx \dt
\]
\[
+\int_0^{\tau}\int_{\Omega}
\left[
\left( 1-\frac{\varrho}{\widetilde{\varrho}} \right)p'(\widetilde{\varrho})\pd _t \widetilde{\varrho}
-\varrho\mathbf{u}\cdot \frac{p'(\widetilde{\varrho})}{\widetilde{\varrho}} \nabla \widetilde{\varrho}
\right]
\dx \dt
+\alpha \int_0^{\tau}\int_{\Omega}
{\rm Curl} \widetilde{\mathbf{B}} \cdot
\left(
{\rm Curl} \widetilde{\mathbf{B}}-{\rm Curl} \mathbf{B}
\right)
\dx \dt
\]
\[
+\int_0^{\tau}\int_{\Omega}
\left[
\pd _t \widetilde{\mathbf{B}}\cdot
\left(
\widetilde{\mathbf{B}}-\mathbf{B}
\right)
-(\mathbf{u} \times \mathbf{B}) \cdot {\rm Curl} \widetilde{\mathbf{B}}
\right]
\dx \dt
-\int_0^{\tau}\int_{\Omega}
({\rm Curl} \mathbf{B} \times \mathbf{B} )\cdot \widetilde{\mathbf{u}}
\dx \dt
\]
\begin{equation}\label{lbn9}
-\int_0^{\tau}\int_{\overline{\Omega}} \nabla \widetilde{\mathbf{u}} :{\rm d} \mu_c(t)\,{\rm d} t
-\int_0^{\tau}\int_{\overline{\Omega}} \widetilde{\mathbf{u}} \cdot {\rm d} \mu_B(t)\,{\rm d} t .
\end{equation}
In light of the compatibility conditions (\ref{lbb8}), it holds that
\[
\left|-\int_0^{\tau}\int_{\overline{\Omega}} \nabla \widetilde{\mathbf{u}} :{\rm d} \mu_c(t)\,{\rm d} t
-\int_0^{\tau}\int_{\overline{\Omega}} \widetilde{\mathbf{u}} \cdot {\rm d} \mu_B(t)\,{\rm d} t
\right|
\]
\begin{equation*}\label{lbn10}
\lesssim
\epsilon \int_0^{\tau}
\int_{\overline{\Omega}} {\rm d}\mathfrak{C}
+ \frac{1}{\epsilon}
\int_0^{\tau} \int_{\overline{\Omega}}{\rm d}\mathfrak{D}(t)\,{\rm d} t ,
\end{equation*}
where $\epsilon>0$ is chosen to be sufficiently small.
Using the hypothesis that $(\widetilde{\varrho},\widetilde{\mathbf{u}},\widetilde{\mathbf{B}})$ solves (\ref{pde})--(\ref{ini_c}) in the classical sense, i.e.,
\begin{equation}\label{lbn11}
\left\{\begin{aligned}
& \pd _t \widetilde{\varrho}+{\rm div} (\widetilde{\varrho} \,\widetilde{\mathbf{u}})=0,\\
& \widetilde{\varrho} \left(\pd _t \widetilde{\mathbf{u}}+\widetilde{\mathbf{u}}\cdot \nabla \widetilde{\mathbf{u}} \right)
+\nabla p(\widetilde{\varrho})={\rm div} \mathbb{S}(\nabla \widetilde{\mathbf{u}})+{\rm Curl} \widetilde{\mathbf{B}}\times \widetilde{\mathbf{B}},\\
& \pd _t \widetilde{\mathbf{B}}={\rm Curl} (\widetilde{\mathbf{u}} \times \widetilde{\mathbf{B}})- \alpha{\rm Curl} ({\rm Curl} \widetilde{\mathbf{B}}),\\
& {\rm div} \widetilde{\mathbf{B}}=0, \\
\end{aligned}\right.
\end{equation}
we furthermore simplify the right-hand side of (\ref{lbn9}) as follows. Since this process is straightforward and similar to the compressible Navier-Stokes system (see \cite{FGSGW}), the details are omitted.
\[
\mathcal{E}\Big((\varrho,\mathbf{u},\mathbf{B})\,\Big|\,(\widetilde{\varrho},\widetilde{\mathbf{u}},\widetilde{\mathbf{B}})\Big)(\tau)
+\int_0^{\tau}\int_{\Omega} \Big( \mathbb{S}(\nabla\mathbf{u}-\nabla \widetilde{\mathbf{u}}):(\nabla \mathbf{u}-\nabla \widetilde{\mathbf{u}})
\Big)\dx \dt
\]
\[
+\alpha \int_0^{\tau}\int_{\Omega} | {\rm Curl} (\mathbf{B} -\widetilde{\mathbf{B}})|^2\dx \dt
+\int_{\overline{\Omega} }{\rm d}\mathfrak{D}(\tau)
+ \int_0^{\tau}
\int_{\overline{\Omega}} {\rm d}\mathfrak{C}
\]
\[
\stackrel{<}{\sim} \int_0^{\tau}\int_{\Omega}
\varrho(\mathbf{u}-\widetilde{\mathbf{u}}) \cdot \nabla \widetilde{\mathbf{u}}\cdot (\widetilde{\mathbf{u}}-\mathbf{u})
\dx \dt
+\int_0^{\tau}\int_{\Omega} \mathbb{S} (\nabla \widetilde{\mathbf{u}}):
\left(
\nabla \widetilde{\mathbf{u}}-\nabla \mathbf{u}
\right) \dx \dt
\]
\[
+\int_0^{\tau}\int_{\Omega}
\varrho(\widetilde{\mathbf{u}}-\mathbf{u}) \cdot
\frac{1}{\widetilde{\varrho}} {\rm div} \mathbb{S}(\nabla \widetilde{\mathbf{u}})
\dx \dt
-\int_0^{\tau}\int_{\Omega}
\Big( p(\varrho)-p(\widetilde{\varrho}) -p'(\widetilde{\varrho}) (\varrho-\widetilde{\varrho})\Big) {\rm div} \widetilde{\mathbf{u}}
\dx \dt
\]
\[
+\alpha \int_0^{\tau}\int_{\Omega}
{\rm Curl} \widetilde{\mathbf{B}} \cdot
\left(
{\rm Curl} \widetilde{\mathbf{B}}-{\rm Curl} \mathbf{B}
\right)
\dx \dt
+\int_0^{\tau}\int_{\Omega}
\varrho(\widetilde{\mathbf{u}}-\mathbf{u}) \cdot
\frac{1}{\widetilde{\varrho}} \left({\rm Curl} \widetilde{\mathbf{B}}\times \widetilde{\mathbf{B}}\right)
\dx \dt
\]
\[
+\int_0^{\tau}\int_{\Omega}
\Big(
\pd _t \widetilde{\mathbf{B}}\cdot
(
\widetilde{\mathbf{B}}-\mathbf{B}
)
-(\mathbf{u} \times \mathbf{B}) \cdot {\rm Curl} \widetilde{\mathbf{B}}
\Big)
\dx \dt
\]
\begin{equation}\label{lbn12}
-\int_0^{\tau}\int_{\Omega}
({\rm Curl} \mathbf{B} \times \mathbf{B} )\cdot \widetilde{\mathbf{u}}
\dx \dt
+ \int_0^{\tau} \int_{\overline{\Omega}}{\rm d}\mathfrak{D}(t)\,{\rm d} t .
\end{equation}
Notice that the integrals involved with the magnetic field may be rewritten as, using (\ref{lbn11})$_3$,
\[
\alpha \int_0^{\tau}\int_{\Omega}
{\rm Curl} \widetilde{\mathbf{B}} \cdot
\left(
{\rm Curl} \widetilde{\mathbf{B}}-{\rm Curl} \mathbf{B}
\right)
\dx \dt
+\int_0^{\tau}\int_{\Omega}
\varrho(\widetilde{\mathbf{u}}-\mathbf{u}) \cdot
\frac{1}{\widetilde{\varrho}}\left( {\rm Curl} \widetilde{\mathbf{B}}\times \widetilde{\mathbf{B}}\right)
\dx \dt
\]
\[
+\int_0^{\tau}\int_{\Omega}
\Big(
\pd _t \widetilde{\mathbf{B}}\cdot
\left(
\widetilde{\mathbf{B}}-\mathbf{B}
\right)
-(\mathbf{u} \times \mathbf{B} )\cdot {\rm Curl} \widetilde{\mathbf{B}}
\Big)
\dx \dt
-\int_0^{\tau}\int_{\Omega}
({\rm Curl} \mathbf{B} \times \mathbf{B} )\cdot \widetilde{\mathbf{u}}
\dx \dt
\]
\[
=\int_0^{\tau}\int_{\Omega}
(\varrho-\widetilde{\varrho})(\widetilde{\mathbf{u}}-\mathbf{u}) \cdot
\frac{1}{\widetilde{\varrho}}\left( {\rm Curl} \widetilde{\mathbf{B}}\times \widetilde{\mathbf{B}} \right)
\dx \dt
+\int_0^{\tau}\int_{\Omega}
(\widetilde{\mathbf{u}}-\mathbf{u}) \cdot
\left( {\rm Curl} \widetilde{\mathbf{B}}\times \widetilde{\mathbf{B}} \right)
\dx \dt
\]
\[
+\int_0^{\tau}\int_{\Omega}
\left[
{\rm Curl} (\widetilde{\mathbf{u}} \times \widetilde{\mathbf{B}}) \cdot
\left(
\widetilde{\mathbf{B}}-\mathbf{B}
\right)
-(\mathbf{u} \times \mathbf{B})\cdot {\rm Curl} \widetilde{\mathbf{B}}
\right]
\dx \dt
+\int_0^{\tau}\int_{\Omega}
{\rm Curl} \mathbf{B} \cdot
(\widetilde{\mathbf{u}}\times \mathbf{B} )
\dx \dt
\]
\[
=\int_0^{\tau}\int_{\Omega}
(\varrho-\widetilde{\varrho})(\widetilde{\mathbf{u}}-\mathbf{u}) \cdot
\frac{1}{\widetilde{\varrho}}\left( {\rm Curl} \widetilde{\mathbf{B}}\times \widetilde{\mathbf{B}} \right)
\dx \dt
+\int_0^{\tau}\int_{\Omega}
{\rm Curl} (\mathbf{B} -\widetilde{\mathbf{B}})
\cdot
\Big( \widetilde{\mathbf{u}} \times (\mathbf{B}-\widetilde{\mathbf{B}})
\Big )
\dx \dt
\]
\begin{equation}\label{lbn13}
+\int_0^{\tau}\int_{\Omega}
{\rm Curl} \widetilde{\mathbf{B}} \cdot
\Big( (\mathbf{u}-\widetilde{\mathbf{u}})\times(\widetilde{\mathbf{B}}-\mathbf{B})\Big)
\dx \dt.
\end{equation}
Moreover, it holds that
\[
\left|
\int_0^{\tau}\int_{\Omega}
{\rm Curl} (\mathbf{B} -\widetilde{\mathbf{B}})
\cdot
\Big( \widetilde{\mathbf{u}} \times (\mathbf{B}-\widetilde{\mathbf{B}})
\Big )
\dx \dt
\right|
\]
\begin{equation}\label{lbn14}
\stackrel{<}{\sim} \epsilon \int_0^{\tau}\int_{\Omega} \left| {\rm Curl} (\mathbf{B} -\widetilde{\mathbf{B}})\right|^2\dx \dt
+c(\epsilon) \int_0^{\tau}\int_{\Omega}
|\mathbf{B}-\widetilde{\mathbf{B}}|^2
\dx \dt;
\end{equation}
\[
\left|
\int_0^{\tau}\int_{\Omega}
{\rm Curl} \widetilde{\mathbf{B}} \cdot
\Big( (\mathbf{u}-\widetilde{\mathbf{u}})\times(\widetilde{\mathbf{B}}-\mathbf{B})\Big)
\dx \dt
\right|
\]
\begin{equation}\label{lbn15}
\stackrel{<}{\sim} \epsilon \int_0^{\tau}\int_{\Omega} |\mathbf{u}-\widetilde{\mathbf{u}}|^2 \dx \dt
+c(\epsilon) \int_0^{\tau}\int_{\Omega}
|\mathbf{B}-\widetilde{\mathbf{B}}|^2
\dx \dt.
\end{equation}
Due to the generalized Korn-type inequality,
\begin{equation}\label{lbn16}
\int_0^{\tau}\int_{\Omega} |\mathbf{u}-\widetilde{\mathbf{u}}|^2\dx \dt
\lesssim
\int_0^{\tau}\int_{\Omega} \mathbb{S}(\nabla\mathbf{u}-\nabla \widetilde{\mathbf{u}}):(\nabla \mathbf{u}-\nabla \widetilde{\mathbf{u}})
\dx \dt.
\end{equation}
In addition, the isentropic law of pressure function yields
\[
\left|
\int_0^{\tau}\int_{\Omega}
\Big( p(\varrho)-p(\widetilde{\varrho}) -p'(\widetilde{\varrho}) (\varrho-\widetilde{\varrho})\Big) {\rm div} \widetilde{\mathbf{u}}
\dx \dt
\right|
\]
\begin{equation}\label{lbn17}
\stackrel{<}{\sim}
\int_0^{\tau}\int_{\Omega}
\Big(\mathcal{H}(\varrho)-\mathcal{H}(\widetilde{\varrho}) -\mathcal{H}'(\widetilde{\varrho}) (\varrho-\widetilde{\varrho})\Big)
\dx \dt.
\end{equation}
Consequently, combining (\ref{lbn12})--(\ref{lbn17}) and choosing $\epsilon>0$ suitably small give rise to
\[
\mathcal{E}\Big((\varrho,\mathbf{u},\mathbf{B})\,\Big|\,(\widetilde{\varrho},\widetilde{\mathbf{u}},\widetilde{\mathbf{B}})\Big)(\tau)
+\int_0^{\tau}\int_{\Omega} \mathbb{S}(\nabla\mathbf{u}-\nabla \widetilde{\mathbf{u}}):(\nabla \mathbf{u}-\nabla \widetilde{\mathbf{u}})
\dx \dt
\]
\[
+\alpha \int_0^{\tau}\int_{\Omega} | {\rm Curl} (\mathbf{B} -\widetilde{\mathbf{B}})|^2\dx \dt
+\int_{\overline{\Omega} }{\rm d}\mathfrak{D}(\tau)
+ \int_0^{\tau}
\int_{\overline{\Omega}} {\rm d}\mathfrak{C}
\]
\[
\stackrel{<}{\sim}
\int_0^{\tau}\int_{\Omega} \mathbb{S} (\nabla \widetilde{\mathbf{u}}):
\left(
\nabla \widetilde{\mathbf{u}}-\nabla \mathbf{u}
\right) \dx \dt
+\int_0^{\tau}\int_{\Omega}
\varrho(\widetilde{\mathbf{u}}-\mathbf{u}) \cdot
\frac{1}{\widetilde{\varrho}} \,{\rm div} \mathbb{S}(\nabla \widetilde{\mathbf{u}})
\dx \dt
\]
\[
+\int_0^{\tau}\int_{\Omega}
(\varrho-\widetilde{\varrho})(\widetilde{\mathbf{u}}-\mathbf{u}) \cdot
\frac{1}{\widetilde{\varrho}}\left( {\rm Curl} \widetilde{\mathbf{B}}\times \widetilde{\mathbf{B}} \right)
\dx \dt
\]
\begin{equation}\label{lbn18}
+
\int_0^{\tau}
\mathcal{E}\Big((\varrho,\mathbf{u},\mathbf{B})\,\Big|\,(\widetilde{\varrho},\widetilde{\mathbf{u}},\widetilde{\mathbf{B}})\Big)(t)
\,{\rm d} t
+\int_0^{\tau} \int_{\overline{\Omega}}{\rm d}\mathfrak{D}(t)\,{\rm d} t .
\end{equation}
Following \cite{FGSGW,FNS}, we estimate the remaining integrals as follows. Let $\chi$ be a cut-off function such that
\begin{equation*}\label{lbn19}
\left\{\begin{aligned}
& \chi \in C _c^{\infty}((0,\infty)),\\
& 0\leq \chi \leq 1,\\
& \chi(\varrho)=1 \text{ if }\varrho \in [\inf{\widetilde{\varrho}},\sup{\widetilde{\varrho}}].\\
\end{aligned}\right.
\end{equation*}
Then we may write
\[
\left|
\int_0^{\tau}\int_{\Omega}
(\varrho-\widetilde{\varrho})(\widetilde{\mathbf{u}}-\mathbf{u}) \cdot
\frac{1}{\widetilde{\varrho}}\left( {\rm Curl} \widetilde{\mathbf{B}}\times \widetilde{\mathbf{B}} \right)
\dx \dt
\right|
\]
\begin{equation}\label{lbn20}
\stackrel{<}{\sim}
\int_0^{\tau}\int_{\Omega}
\chi(\varrho)|\varrho-\widetilde{\varrho}||\widetilde{\mathbf{u}}-\mathbf{u}|
\dx \dt
+
\int_0^{\tau}\int_{\Omega}
(1-\chi(\varrho))|\varrho-\widetilde{\varrho}||\widetilde{\mathbf{u}}-\mathbf{u}|
\dx \dt.
\end{equation}
The first integral on the right-hand side of (\ref{lbn20}) is bounded through
\[
\int_0^{\tau}\int_{\Omega}
\chi(\varrho)|\varrho-\widetilde{\varrho}||\widetilde{\mathbf{u}}-\mathbf{u}|
\dx \dt
\]
\[
\stackrel{<}{\sim}
\int_0^{\tau}\int_{\Omega}
\frac{1}{2}\varrho|\widetilde{\mathbf{u}}-\mathbf{u}|^2
\dx \dt
+
\int_0^{\tau}\int_{\Omega}
\frac{1}{2}\frac{\chi^2(\varrho)}{\varrho} |\varrho-\widetilde{\varrho}|^2
\dx \dt
\]
\begin{equation}\label{lbn21}
\stackrel{<}{\sim}
\int_0^{\tau}
\mathcal{E}\Big((\varrho,\mathbf{u},\mathbf{B})\,\Big|\,(\widetilde{\varrho},\widetilde{\mathbf{u}},\widetilde{\mathbf{B}})\Big)(t)
\,{\rm d} t .
\end{equation}
To estimate the second integral on the right-hand side of (\ref{lbn20}), we make a further decomposition, i.e.,
\[
1-\chi(\varrho)=\chi_1(\varrho)+\chi_2(\varrho)
\]
such that
\[
\text{supp}\chi_1 \subset [0,\inf{\widetilde{\varrho}}],\,\,
\text{supp}\chi_2 \subset [\sup{\widetilde{\varrho}},\infty].
\]
It follows from (\ref{lbn16}) that
\[
\int_0^{\tau}\int_{\Omega}
\chi_1(\varrho)|\varrho-\widetilde{\varrho}||\widetilde{\mathbf{u}}-\mathbf{u}|
\dx \dt
\]
\[
\stackrel{<}{\sim}
\epsilon
\int_0^{\tau}\int_{\Omega} |\mathbf{u}-\widetilde{\mathbf{u}}|^2 \dx \dt
+c(\epsilon)
\int_0^{\tau}\int_{\Omega}
\chi_1^2(\varrho)|\varrho-\widetilde{\varrho}|^2
\dx \dt
\]
\begin{equation}\label{lbn22}
\stackrel{<}{\sim}
\epsilon\int_0^{\tau}\int_{\Omega} \Big( \mathbb{S}(\nabla\mathbf{u}-\nabla \widetilde{\mathbf{u}}):(\nabla \mathbf{u}-\nabla \widetilde{\mathbf{u}})
\Big)\dx \dt
+c(\epsilon)
\int_0^{\tau}
\mathcal{E}\Big((\varrho,\mathbf{u},\mathbf{B})\,\Big|\,(\widetilde{\varrho},\widetilde{\mathbf{u}},\widetilde{\mathbf{B}})\Big)(t)
\,{\rm d} t .
\end{equation}
Clearly,
\[
\int_0^{\tau}\int_{\Omega}
\chi_2(\varrho)|\varrho-\widetilde{\varrho}||\widetilde{\mathbf{u}}-\mathbf{u}|
\dx \dt
\]
\[
\stackrel{<}{\sim}
\int_0^{\tau}\int_{\Omega}
\chi_2(\varrho)\varrho|\widetilde{\mathbf{u}}-\mathbf{u}| ^2
\dx \dt
+\int_0^{\tau}\int_{\Omega}
\chi_2(\varrho)\varrho
\dx \dt
\]
\begin{equation}\label{lbn23}
\stackrel{<}{\sim}
\int_0^{\tau}
\mathcal{E}\Big((\varrho,\mathbf{u},\mathbf{B})\,\Big|\,(\widetilde{\varrho},\widetilde{\mathbf{u}},\widetilde{\mathbf{B}})\Big)(t)
\,{\rm d} t .
\end{equation}
Taking (\ref{lbn20})--(\ref{lbn23}) into account, we see
\[
\left|
\int_0^{\tau}\int_{\Omega}
(\varrho-\widetilde{\varrho})(\widetilde{\mathbf{u}}-\mathbf{u}) \cdot
\frac{1}{\widetilde{\varrho}}\left( {\rm Curl} \widetilde{\mathbf{B}}\times \widetilde{\mathbf{B}} \right)
\dx \dt
\right|
\]
\begin{equation}\label{lbn24}
\stackrel{<}{\sim}
\epsilon\int_0^{\tau}\int_{\Omega} \Big( \mathbb{S}(\nabla\mathbf{u}-\nabla \widetilde{\mathbf{u}}):(\nabla \mathbf{u}-\nabla \widetilde{\mathbf{u}})
\Big)\dx \dt
+
c(\epsilon)
\int_0^{\tau}
\mathcal{E}\Big((\varrho,\mathbf{u},\mathbf{B})\,\Big|\,(\widetilde{\varrho},\widetilde{\mathbf{u}},\widetilde{\mathbf{B}})\Big)(t)
\,{\rm d} t .
\end{equation}
Finally, notice also that the first two integrals on the right-hand side of (\ref{lbn18}) are estimated as above upon observing that
\[
\int_0^{\tau}\int_{\Omega} \mathbb{S} (\nabla \widetilde{\mathbf{u}}):
\left(
\nabla \widetilde{\mathbf{u}}-\nabla \mathbf{u}
\right) \dx \dt
+\int_0^{\tau}\int_{\Omega}
\varrho(\widetilde{\mathbf{u}}-\mathbf{u}) \cdot
\frac{1}{\widetilde{\varrho}} {\rm div} \mathbb{S}(\nabla \widetilde{\mathbf{u}})
\dx \dt
\]
\begin{equation}\label{lbn25}
=
\int_0^{\tau}\int_{\Omega}
(\varrho-\widetilde{\varrho})(\widetilde{\mathbf{u}}-\mathbf{u}) \cdot
\frac{1}{\widetilde{\varrho}}
{\rm div} \mathbb{S}(\nabla \widetilde{\mathbf{u}})
\dx \dt.
\end{equation}
Combining (\ref{lbn18}), (\ref{lbn24})--(\ref{lbn25}) and fixing $\epsilon>0$ sufficiently small shows that
\[
\mathcal{E}\Big((\varrho,\mathbf{u},\mathbf{B})\,\Big|\,(\widetilde{\varrho},\widetilde{\mathbf{u}},\widetilde{\mathbf{B}})\Big)(\tau)
+\int_0^{\tau}\int_{\Omega} \Big( \mathbb{S}(\nabla\mathbf{u}-\nabla \widetilde{\mathbf{u}}):(\nabla \mathbf{u}-\nabla \widetilde{\mathbf{u}})
\Big)\dx \dt
\]
\[
+\alpha \int_0^{\tau}\int_{\Omega} | {\rm Curl} (\mathbf{B} -\widetilde{\mathbf{B}})|^2\dx \dt
+\int_{\overline{\Omega} }{\rm d}\mathfrak{D}(\tau)
+ \int_0^{\tau}
\int_{\overline{\Omega}} {\rm d}\mathfrak{C}
\]
\begin{equation*}\label{lbn26}
\stackrel{<}{\sim}
\int_0^{\tau}
\mathcal{E}\Big((\varrho,\mathbf{u},\mathbf{B})\,\Big|\,(\widetilde{\varrho},\widetilde{\mathbf{u}},\widetilde{\mathbf{B}})\Big)(t)
\,{\rm d} t + \int_0^{\tau} \int_{\overline{\Omega}}{\rm d}\mathfrak{D}(t)\,{\rm d} t .
\end{equation*}
As a direct application of Gronwall's inequality, we conclude that
\[
\varrho=\tilde{\varrho},\,\,\,\, \mathbf{u}=\tilde{\mathbf{u}},\,\,\,\,\mathbf{B}=\tilde{\mathbf{B}},
\text{ in } (0,T)\times \Omega,
\]
\[
\mu_c=\mathbf{0},\,\,\,\, \mu_B=\mathbf{0},\,\,\,\, \mathfrak{D}=0,\,\,\,\, \mathfrak{C}=0,
\]
thus completely finishing the proof of Theorem \ref{Th1}.
\qed
\section{Convergence}\label{sec_con}
In this section we prove another main result, that is the convergence of a consistent approximation stated in Theorem~\ref{Th2}. As an application of this theorem, we will also show the convergence analysis of two mixed finite volume-finite element methods.
\subsection{Convergence of a consistent approximation}
In this subsection we prove Theorem~\ref{Th2} for the convergence of a consistent approximation $(\varrho_h, \mathbf{u}_h, \vB_h)$ in two steps, that are the convergences towards a DW solution and towards a classical solution.
\subsubsection{Convergence to a DW solution}
\begin{proof}[Proof of Item 1 of Theorem~\ref{Th2}]
As $(\varrho_h, \mathbf{u}_h, \vB_h)$ is a consistent approximation in the sense of Definition~\ref{def_ca}, it satisfies the stability property~\eqref{es}. Consequently, we derive for suitable subsequences, not relabelled, that
\[
\begin{split}
\varrho_h &\to \varrho \ \mbox{weakly-(*) in}\ L^\infty(0,T; L^\gamma(\Omega)),\ \varrho \geq 0,
\\
\mathbf{u}_h, \Pi_h \mathbf{u}_h &\to \mathbf{u} \ \mbox{weakly in}\ L^2((0,T)\times \Omega; \mathbb{R}^d), \mbox{ where } \mathbf{u} \in L^2(0,T; W^{1,2}(\Omega;\mathbb{R}^d)),\\
& \mbox{ and in the case of Dirichlet boundary conditions } \mathbf{u} \in L^2(0,T; W_0^{1,2}(\Omega;\mathbb{R}^d)), \\
\varrho_h \Pi_h \mathbf{u}_h &\to \overline{\varrho \mathbf{u}} \ \ \mbox{weakly-(*) in}\ L^\infty(0,T; L^{\frac{2\gamma}{\gamma + 1}}(\Omega; \mathbb{R}^d)),
\\
\vB_h &\to \mathbf{B} \ \mbox{weakly-(*) in}\ L^\infty(0,T; L^2(\Omega;\mathbb{R}^d)),
\\
{\rm Curl}_h \vB_h &\to {\rm Curl} \mathbf{B} \ \mbox{weakly in}\ L^2((0,T)\times \Omega; \mathbb{R}^d),
\\
\mathbf{u}_h\times \vB_h &\to \overline{\mathbf{u} \times \mathbf{B}} \ \mbox{weakly in}\ L^2(0,T; L^{\f32}(\Omega;\mathbb{R}^d)),
\\
\varrho_h \Pi_h \mathbf{u}_h \otimes \mathbf{u}_h + p(\varrho_h) \mathbb{I} &
\to \overline{ 1_{\varrho>0}\frac{\mathbf{m}\otimes \mathbf{m}}{\varrho} +p(\varrho)\mathbb{I} }
\ \mbox{weakly-(*) in} \ L^\infty(0,T; \mathcal{M}(\overline{\Omega};\mathbb{R}^{d\times d}_{sym})) ,
\\
{\rm Curl}_h \vB_h \times \vB_h & \to \overline{ {\rm Curl} \mathbf{B}\times \mathbf{B}}
\ \mbox{weakly-(*) in} \ L^2(0,T; \mathcal{M}(\overline{\Omega};\mathbb{R}^d)) ,
\\
\S(\nabla_h \mathbf{u}_h) : \nabla_h \mathbf{u}_h &\to \overline{\mathbb{S}(\nabla \mathbf{u}):\nabla \mathbf{u}} \mbox{ in }\mathcal{M}^{+}([0,T]\times \overline{\Omega}),
\\
|{\rm Curl}_h \vB_h|^2 &\to \overline{ |{\rm Curl} \mathbf{B}|^2 }\mbox{ in }\mathcal{M}^{+}([0,T]\times \overline{\Omega}),
\\
\frac{1}{2} \varrho_h \abs{\Pi_h \mathbf{u}_h}^2 + \mathcal{H}(\varrho_h) +\frac{1}{2} \abs{\vB_h}^2 &\to \overline{ \frac{1}{2}\varrho |\mathbf{u}|^2+\mathcal{H}(\varrho)+\frac{1}{2}|\mathbf{B}|^2 }\mbox{ weakly-(*) in }L^{\infty}(0,T; \mathcal{M}^{+}(\overline{\Omega})).
\end{split}
\]
With the uniform bounds at hand, we may invoke Lemma 3.7 in Abbatiello et al. \cite{AnFeNo} to deduce that
\[
\overline{\varrho \mathbf{u}}=\varrho\mathbf{u},\,\,\,\, \overline{\mathbf{u} \times \mathbf{B}}=\mathbf{u}\times \mathbf{B}.
\]
We then set
\[
\mu_c:=\overline{ 1_{\varrho>0}\frac{\mathbf{m}\otimes \mathbf{m}}{\varrho} +p(\varrho)\mathbb{I} }
-\left(
1_{\varrho>0}\frac{\mathbf{m}\otimes \mathbf{m}}{\varrho} +p(\varrho)\mathbb{I}
\right),
\]
\[
\mu_B:=\overline{ {\rm Curl} \mathbf{B}\times \mathbf{B}}
-( {\rm Curl} \mathbf{B}\times \mathbf{B} ),
\]
\[
\mathfrak{D}:=\overline{ \frac{1}{2}\varrho |\mathbf{u}|^2+\mathcal{H}(\varrho)+\frac{1}{2}|\mathbf{B}|^2 }
-\left(\frac{1}{2}\varrho |\mathbf{u}|^2+\mathcal{H}(\varrho)+\frac{1}{2}|\mathbf{B}|^2\right),
\]
\[
\mathfrak{C}=\Big(\overline{\mathbb{S}(\nabla \mathbf{u}):\nabla \mathbf{u}}
-\mathbb{S}(\nabla \mathbf{u}):\nabla \mathbf{u} \Big)+\alpha \Big(\overline{ |{\rm Curl} \mathbf{B}|^2 }
-|{\rm Curl} \mathbf{B}|^2 \Big),
\]
Knowing the above limit, we are ready to pass to the limit $h \to 0$ in the consistency formulation \eqref{cP1}--\eqref{cP4} and the energy stability \eqref{es}. We get the following formulae for the limit functions
\begin{equation*} \label{M1}
\left[ \intO{ \varrho \phi } \right]_{t=0}^{t=\tau} =
\int_0^\tau \intO{ \Big( \varrho \pd _t \phi + \varrho \mathbf{u} \cdot \nabla \phi \Big)} \,{\rm d} t ,
\end{equation*}
for any $\phi \in C_c^M([0,T] \times \Ov{\Omega})$ and for some $M \geq 1$;
\begin{equation*} \label{M2}
\begin{split}
& \left[ \intO{ \varrho \mathbf{u} \cdot \bfv } \right]_{t=0}^{t=\tau} =
\int_0^{\tau} \intO{ \Big( \varrho \mathbf{u} \cdot \partial_t \bfv + \varrho \mathbf{u} \otimes \mathbf{u} +p(\varrho) \mathbb{I} : \nabla \bfv \Big)} \,{\rm d} t
\\&
- \int_0^{\tau} \intO{ \S( \nabla \mathbf{u}) : \nabla \bfv} \,{\rm d} t
+ \int_0^{\tau} \intO{ ( {\rm Curl} \mathbf{B} \times \mathbf{B}) \cdot \bfv}\,{\rm d} t
\\ &
+
\int_0^{\tau}\int_{\overline{\Omega}} \nabla \bfv :{\rm d} \mu_c(t)\,{\rm d} t
+
\int_0^{\tau}\int_{\overline{\Omega}} \bfv \cdot {\rm d} \mu_B(t)\,{\rm d} t
\end{split}
\end{equation*}
for any $\bfv \in C_c^M([0,T] \times \Omega; \mathbb{R}^d)$ and for some $M \geq 1$;
\begin{equation*} \label{M3}
\left[ \intO{ \mathbf{B} \cdot \mathbf{C} }\right]_{t=0}^{t=\tau}
= \int_0^{\tau} \int_{\Omega} \Big(\mathbf{B} \cdot \pd _t \mathbf{C}
- \alpha {\rm Curl} \mathbf{B} \cdot {\rm Curl} \mathbf{C}
+ (\mathbf{u}\times \mathbf{B}) \cdot {\rm Curl}\mathbf{C}\Big)
\dx \dt
\end{equation*}
for any $\mathbf{C} \in C_c^M([0,T]\times \overline{\Omega};\mathbb{R}^d),\,\bfv\times \vc{n}|_{\p \Omega}=\mathbf{0}$ and for some $M \geq 1$;
\begin{equation*}\label{M4}
\intO{ \mathbf{B} \cdot \nabla \psi} = 0
\end{equation*}
for any $\psi \in C^M( \overline{\Omega}) \cap L_0^2(\Omega) $ and for some $M \geq 1$;
\[
\int_{\Omega} \left[\frac{1}{2}\varrho |\mathbf{u}|^2+\frac{1}{2}|\mathbf{B}|^2+\mathcal{H}(\varrho)\right](\tau,x)\,{\rm d} {x}
+\int_0^{\tau}\int_{\Omega} \left ( \mathbb{S}(\nabla\mathbf{u}):\nabla \mathbf{u}
+\alpha \left| {\rm Curl} \mathbf{B} \right|^2
\right )\dx \dt
\]
\begin{equation*} \label{M5}
+\int_{\overline{\Omega} }{\rm d}\mathfrak{D}(\tau)
+ \int_0^{\tau}
\int_{\overline{\Omega}} {\rm d}\mathfrak{C}
\leq
\int_{\Omega}\left[ \frac{1}{2}\frac{|\mathbf{m}_0|^2}{\varrho_0}+\frac{1}{2}|\mathbf{B}_0|^2+\mathcal{H}(\varrho_0)\right]\,{\rm d} {x}.
\end{equation*}
for a.e. $\tau \in (0,T)$;
Finally, the compatibility conditions (\ref{lbb8}) and (\ref{lbb8-2}) hold owing to \cite[Lemma 2.1]{FGSGW} and the crucial observations:
\[
\Big|\varrho \mathbf{u} \otimes \mathbf{u} +p(\varrho)\mathbb{I}\Big|
\lesssim \frac{1}{2}\varrho |\mathbf{u}|^2+\mathcal{H}(\varrho)
\lesssim \frac{1}{2}\varrho |\mathbf{u}|^2+\mathcal{H}(\varrho)+\frac{1}{2}|\mathbf{B}|^2,
\]
\[
\big| {\rm Curl} \mathbf{B} \times \mathbf{B} \big|
\lesssim
\epsilon |{\rm Curl} \mathbf{B}|^2 +\frac{1}{\epsilon}|\mathbf{B}|^2
\lesssim
\epsilon \big( |{\rm Curl} \mathbf{B}|^2 +\mathbb{S}(\nabla\mathbf{u}):\nabla \mathbf{u} \big)
\]
\[
+\frac{1}{\epsilon}\left[\frac{1}{2}\varrho |\mathbf{u}|^2+\frac{1}{2}|\mathbf{B}|^2+\mathcal{H}(\varrho)\right], \,\,\text{ for any }\epsilon>0.
\]
Consequently, collecting the above identities and relations, we conclude that the weak limit $(\varrho,\mathbf{u},\mathbf{B})$, with the associated $\mu_c,\mathfrak{D},\mathfrak{C}$, generated by the consistent approximation $(\varrho_h, \mathbf{u}_h, \vB_h)$ for $h \to 0$, represents a DW solution of the MHD system \eqref{pde} in the sense of Definition \ref{def_dw}. This proves Item 1 of Theorem~\ref{Th2}.
\end{proof}
\subsubsection{Convergence to the classical solution}
\begin{proof}[Proof of Item 2 of Theorem~\ref{Th2}]
Combining Item 1 of Theorem~\ref{Th2} and Theorem~\ref{Th1} we immediately obtain Item 2 of Theorem~\ref{Th2}; that is the convergence of the consistent approximation $(\varrho_h, \mathbf{u}_h, \vB_h)$ towards the classical solution.
\end{proof}
\medskip
We have built a general theory in Theorem~\ref{Th2} that a consistent approximation converges to the DW solution as well as the classical solution (on its lifespan). Next, we show the application of this theory in the convergence analysis of numerical solutions by two examples. The only gap here is whether the numerical solution is a consistent approximation in the sense of Definition~\ref{def_ca}.
\subsection{Example-I}
In the first example, we propose a mixed FV--FE approximation adapted from the Navier-Stokes solver of Karper~\cite{Karper} and the magnetic solver of Ding and Mao~\cite{Ding} with no-slip boundary conditions.
To begin, we introduce the necessary notations.
\paragraph{Mesh.} Let $\Omega_h$ be a regular and quasi-uniform triangulation of the bounded domain $\Omega$, $\mcE$ be the set of all $(d-1)$-dimensional faces, $\mcEe=\mcE \cap \partial \Omega$ be the exterior faces, $\mcEi=\mcE \setminus \mcEe$ be the interior faces, and $\faces(K)$ be the set of all faces of an arbitrary element $K$.
We denote $\sigma=K|L\in \mcEi$ as the common face of two neighbouring elements $K$ and $L$.
Further, we denote $\vc{n}_\sigma$ as the outer normal of a face $\sigma \in \mathcal{E}$ and $\vc{n}_{\sigma,K}$ as the unit normal vector pointing outwards $K$ if $\sigma \in \faces(K)$.
The size of the mesh (maximal diameter of all elements) is supposed to be a positive parameter $h<1$.
\paragraph{Function spaces.}\label{sec:spaces}
We define on $\Omega_h$ discrete function spaces $Q_h, V_h, \Nh$, that are spaces of piecewise constants, piecewise linear Crouzeix--Ravairt elements, and lowest order $ H({\rm curl}) $-N\'{e}d\'{e}lec edge elements, respectively.
\begin{equation*}
Q_h \equiv \left\{ v \in L^2(\Omega) \middle| \; v \vert_K \in \mcP^1_0(K) \ K \in \Omega_h \right\},
\end{equation*}
\begin{equation*}
V_h \equiv \left\{ \bfv \in L^2(\Omega) \middle| \; \bfv \vert_K \in \mcP^d_1(K) \forall\, K \in \Omega_h;\int_{\sigma} { \jump{\bfv}} =0 \; \forall\, \sigma \in \mcEi \right\},
\end{equation*}
\begin{equation*
\Nh \equiv \left\{ \bfv, {\rm Curl} \bfv \in L^2(\Omega) \middle| \; \bfv \vert_K \in \mcP_0^d\oplus\mcP_0^1 x \; \forall\; K \in \Omega_h;
\int_{\sigma} { \jump{\bfv \times \vc{n}}} =0 \; \forall\; \sigma \in \mcEi
\right\},
\end{equation*}
where $\mcP^d_n(K)$ (resp. $\mcP^1_n(K)$) denotes the space of polynomials of degree not greater than $n$ on element $K$ for $d$-dimensional vector valued functions (resp. for scalar functions).
The spaces $Q_h, V_h$ and $\Nh$ shall be used for the approximation of the discrete density, velocity and magnetic field, respectively.
In addition, for the purpose of designing weakly divergence-free magnetic field (see Lemma~\ref{lem_b}), we introduce the following space
\[W_h = \left\{ v \in W^{1,2} \cap L^2_0 \middle| \; v\vert_K \in \mcP_1^1(K) \; \forall \; K \in \Omega_h \right\}.
\]
It should be mentioned that for any $\psi_h \in W_h$ we have $\nabla \phi_h \in \Nh$.
The interpolation operators associated to the function spaces $V_h$, $\Nh$ and $W_h$ are given by
\begin{equation*}\label{proj1}
\Pi_V :\ W^{1,2}(\Omega) \rightarrow V_h, \quad
\Pi_N :\ W^{1,2}(\Omega) \rightarrow \Nh, \quad
\Pi_W :\ L^2_0(\Omega) \rightarrow W_h
\end{equation*}
satisfy the following interpolation estimates
\begin{equation}\label{IE}
\norm{\bfv - \PiV \vv }_{L^p} \stackrel{<}{\sim} h \norm{\bfv}_{C^1},
\;
\norm{\bfv - \Pi_W \bfv }_{L^p} \stackrel{<}{\sim} h \norm{\bfv}_{C^1},
\;
\norm{\mathbf{u} - \Pi_N \mathbf{u} }_{L^p} +\norm{{\rm Curl} (\mathbf{u} - \Pi_N \mathbf{u}) }_{L^p} \stackrel{<}{\sim} h \norm{\mathbf{u}}_{C^2}.
\end{equation}
for any $\bfv \in C^1(\Omega)$, $\mathbf{u} \in C^2(\Omega)$, $p\in[1,\infty]$, see e.g. the monograph of Brezzi et al.~\cite{Brezzi}.
For simplicity of notation, we denote $\FS=Q_h \times \CR \times \ND$, where
\[ \CR = \left\{ \bfv \in V_h \middle| \int_{\sigma} {\bfv} =0\ \forall\, \sigma \in \mcEe \right\} \mbox{ and }
\ND=\left\{ \bfv \in \Nh \middle| \int_{\sigma} {\bfv\times \vc{n}} =0 \; \forall\; \sigma \in \mcEe \right\}.
\]
Next, we suppose the time step size $\Delta t \approx h$ and denote $t^k= k\Delta t$ for $k=1,\ldots,N_T(=T/\Delta t)$.
For a generic discrete function $v_h$ at time $t^k= k\Delta t$ we denote it by $v_h^k$ and write $v_h \in L_{\Delta t}(0,T;Y)$ if $v_h^k \in Y$ for all $k=1,\dots, N_T$ with $Y \in \{Q_h, \CR, \ND, \FS \}$ and
\[ v_h(t,\cdot) =v_h^0 \mbox{ for } t \leq 0,\ v_h(t,\cdot)=v_h^k \mbox{ for } t\in ((k-1)\Delta t,k\Delta t],\ k=1,2,\ldots,N_T.
\]
\paragraph{Some discrete operators.}
We define the discrete time derivative by the backward Euler method
\[
D_t v_h = \frac{v_h (t) - v_h^{\triangleleft}(t)}{\Delta t} \mbox{ for } t\in(0,T) \quad \mbox{ with } \quad v_h^{\triangleleft}(t) = v_h(t - \Delta t) .
\]
For any piecewise continuous function $f$, we define its trace on a generic edge
as
\begin{equation*}
\begin{aligned}
f^{\rm in}|_{\sigma} = \lim_{\delta \rightarrow 0^+} f(\bfx -\delta \bfn_{\sigma} ), \ \forall\ \sigma \in \mcE,
\qquad
f^{\rm out}|_{\sigma} = \lim_{\delta \rightarrow 0^+} f(\bfx +\delta \bfn_{\sigma} ), \ \forall\ \sigma \in \mcEi.
\end{aligned}
\end{equation*}
Note that $f^{\rm out}|_{\mcEe}$ is determined by the boundary condition.
Further, we define the jump and average operators at an edge $\sigma \in \mcE$ as
\begin{equation}\label{op_diff}
\jump{f}_{\sigma} = f^{\rm out} - f^{\rm in} \mbox{ and }
\avg{f} = \frac{f^{\rm out} + f^{\rm in}}{2} ,
\end{equation}
respectively, and denote $\Pi_h \equiv \Pi_Q$ as the element-wise constant projection, where
\begin{equation*}\label{op_avgk}
\Pi_Q: L^1(\Omega) \to Q_h. \quad \left. \Pi_Q f\right|_K = \left. \avc{f}\right|_K \equiv \frac{1}{\abs{K}} \int_{K} f , \ \forall\ K \in \Omega_h.
\end{equation*}
Next, we introduce the upwind flux for any function $r_h \in Q_h$ at a generic face $\sigma \in \faces^I$
\begin{align*}\label{Up}
{\rm Up} [r_h, \mathbf{u}_h] =r_h^{\rm up} u_\sigma
=r_h^{\rm in} [ u_\sigma ]^+ + r_h^{\rm out} [ u_\sigma ]^-,
\end{align*}
where $\mathbf{u}_h \in V_h$ is the velocity field and
\begin{equation*}
u_\sigma = \frac{1}{|\sigma|} \int_{\sigma} {\mathbf{u}_h} \cdot \vc{n}_\sigma, \quad
[f]^{\pm} = \frac{f \pm |f| }{2} \quad \mbox{and} \quad
r_h^{\rm up} =
\begin{cases}
r_h^{\rm in} & \mbox{if } \ u_\sigma \geq 0, \\
r_h^{\rm out} & \mbox{if } \ u_\sigma < 0.
\end{cases}
\end{equation*}
Furthermore, we consider a diffusive numerical flux function of the following form for $\varepsilon>0$
\begin{equation}\label{num_flux}
\begin{aligned}
{\rm F}_h^{\varepsilon}(r_h,\mathbf{u}_h)
={\rm Up}[r_h, \mathbf{u}_h] - h^\varepsilon \jump{ r_h }
\end{aligned}
\end{equation}
It is easy to check for any $\varrho_h \in Q_h$ and $\mathbf{u}_h \in V_h$ that
\begin{equation}\label{fluxes}
\begin{aligned}
& \sum_{ \sigma \in \faces^I } \int_{\sigma} { \left({\rm F}_h^{\varepsilon}(\varrho_h \avc{\mathbf{u}_h},\mathbf{u}_h) \jump{\avc{\mathbf{u}_h}} - {\rm F}_h^{\varepsilon}(\varrho_h ,\mathbf{u}_h) \jump{ \frac{|\avc{\mathbf{u}_h}|^2}{2}} \right)}
\\ & = - \sum_{ \sigma \in \faces^I } \int_{\sigma} { \left(\frac{1}2 \varrho_h^{{\rm up}} \abs{ u_\sigma }
+h^\varepsilon \avg{\varrho_h} \right)\abs{\jump{\avc{\mathbf{u}_h}}}^2 } .
\end{aligned}
\end{equation}
For simplicity, we denote
$\co{a}{b} =\left[\min(a,b), \max(a,b) \right]$
and write
$ a \stackrel{<}{\sim} b \mbox{ if } a \le cb$ if $c$ is a positive constant that is independent of the mesh size and time step used in the scheme.
We shall frequently use the abbreviation $\norm{\cdot}_{L^p}$ and $\norm{\cdot}_{L^pL^q}$ for $\norm{\cdot}_{L^p(\Omega)}$ and $\norm{\cdot}_{L^p\left(0,T;L^q(\Omega)\right)}$, respectively.
\paragraph{The numerical method.}
Using the above notations we propose a mixed FV--FE method for the approximation of the MHD system \eqref{pde} named as {\bf Scheme-I}.
\begin{tcolorbox}
{\bf Scheme-I.}
Given the initial values \eqref{ini_c} we set $(\varrho_h^0,\mathbf{u}_h^0, \vB_h^0) =(\Pi_Q \varrho_0, \Pi_V \mathbf{u}_0, \Pi_N \mathbf{B}_0)$ and seek $(\varrho_h,\mathbf{u}_h, \vB_h) \in L_{\Delta t}(0, T;\FS) $ such that
\begin{subequations}\label{scheme}
\begin{equation}\label{schemeD}
\intO{ D_t \varrho_h \phi_h } - \sum_{ \sigma \in \faces^I } \int_{\sigma} { {\rm F}_h^{\varepsilon}(\varrho_h ,\mathbf{u}_h )
\jump{\phi_h} } = 0 \quad \mbox{for all } \phi_h \in Q_h;
\end{equation}
\begin{equation}\label{schemeM}
\begin{aligned}
\intO{ D_t (\varrho_h \widehat{\vuh} ) \cdot \vv_h } -
\sum_{ \sigma \in \faces^I } \int_{\sigma} { {\rm F}_h^{\varepsilon}(\varrho_h \widehat{\vuh} ,\mathbf{u}_h ) \cdot
\jump{ \widehat{\vv_h} } }
+ \mu \intO{ \nabla_h \mathbf{u}_h : \nabla_h \vv_h } \\ + \intO{( \nu {\rm div}_h \mathbf{u}_h -p_h ) {\rm div}_h \vv_h }
- \intO{ ({\rm Curl}_h \vB_h \times \vB_h^\triangleleft ) \cdot \vv_h}
=0 \quad \mbox{for all } \vv_h \in \CR ;
\end{aligned}
\end{equation}
\begin{equation}\label{schemeB}
\intOB{ D_t \vB_h \cdot \vC_h
+ \alpha {\rm Curl}_h \vB_h \cdot {\rm Curl}_h \vC_h
- (\mathbf{u}_h \times \vB_h^\triangleleft) \cdot {\rm Curl}_h \vC_h
} \quad \mbox{for all } \vC_h \in \ND .
\end{equation}
\end{subequations}
where $\nu = \frac{d-2}{d}\mu+\lambda$,
the discrete operators ${\rm div}_h, \nabla_h$ and ${\rm Curl}_h$ are the same as the continuous case on each element. Moreover, the artificial diffusion parameter $\varepsilon$ follows
\begin{equation*
\varepsilon >0 \mbox{ if } \gamma \geq 2 \quad \mbox{ and }\quad \varepsilon \in(0, 2 \gamma-1 -d/3) \mbox{ if } \gamma \in(4d /(1+3d),2).
\end{equation*}
\end{tcolorbox}
{\bf Scheme-I} enjoys the following properties:
\begin{Lemma}[Existence, mass conservation, renormalized continuity, positivity, divergence free]\label{lem_b}
\hfill
\begin{enumerate}
\item {\bf Existence of a numerical solution.}
There exists at least one solution to {\bf Scheme-I}.
\item {\bf Mass conservation.}
The numerical method~\eqref{scheme} preserves the total mass.
\[
\intO{ \varrho_h (t) } = \intO{ \varrho_h (0) } = \intO{ \varrho_0 } , \quad \forall\; t \in[0,T].
\]
\item {\bf Renormalized continuity equation.}
Let $(\varrho_h,\mathbf{u}_h)\in Q_h \times \CR$ satisfy the discrete continuity equation \eqref{schemeD} and $b=b(\varrho)\in C^2(0,\infty)$. Then the discrete continuity equation \eqref{schemeD} can be renormalized in the sense that
\begin{equation}\label{r1}
\begin{aligned}
&\intO{\left(D_t b(\varrho_h ) -\big(\varrho_h b'(\varrho_h )-b(\varrho_h )\big) {\rm div}_h \mathbf{u}_h \right) }
\\&=
- \frac{\Delta t}2 \intO{ b''(\xi)|D_t \varrho_h |^2 }
- \sum_{ \sigma \in \faces^I } \int_{\sigma} { b''(\zeta) \jump{ \varrho_h } ^2 \left(h^\varepsilon + \frac12 | u_\sigma | \right) }.
\end{aligned}
\end{equation}
where $\xi \in \co{\varrho_h^{\triangleleft}}{\varrho_h }$ and $\zeta \in \co{\varrho_h^{\rm in}}{\varrho_h^{\rm out}}$.
\item {\bf Positivity of the density.} Let $\varrho_0>0$. Then any solution to the discrete problem \eqref{scheme} satisfies $\varrho_h(t)>0$ for $t\in(0,T)$.
\item {\bf Weakly divergence free of magnetic field.}
Let ${\rm div} \mathbf{B}_0 =0$. Then {\bf Scheme-I} preserves divergence free of magnetic field weakly, meaning that
$\intO{ \vB_h \cdot \nabla_h \psi_h} =0$ for any $\psi_h \in W_h$
\end{enumerate}
\end{Lemma}
\begin{proof}
\begin{itemize}
\item
The existence of a numerical solution to \eqref{scheme} can be proven exactly in the same way as in \cite{Ding} via the theorem of topological degree, see also similar result in \cite[Lemma 11.3]{FeLMMiSh}
\item
Taking $\phi_h \equiv 1$ in the equation of continuity \eqref{schemeD} immediately yields the mass conservation.
\item
We refer to~\cite[Lemma 4.1]{Karper} for the proof of renormalized continuity equation.
\item
Concerning the positivity of density, we refer to \cite[Lemma 8.3]{FeLMMiSh} for the proof.
\item We refer to \cite[Remark 6]{Ding} for the proof of weakly divergence-free of magnetic field.
\end{itemize}
\end{proof}
\subsubsection{Stability}
The solution of {\bf Scheme-I} (see \eqref{scheme}) satisfies the stability criteria~\eqref{es} of the consistent approximation. More precisely, we have the following energy estimates.
\begin{Theorem}[Stability of {\bf Scheme-I}]\label{Tm1}\hspace{1em}\newline
Let $(\varrho_h,\mathbf{u}_h, \vB_h)$ be a solution of {\bf Scheme-I}. Then there exist $\xi \in \co{\varrho_h^{\triangleleft}}{\varrho_h }$ and $\zeta \in \co{\varrho_h^{\rm in}}{\varrho_h^{\rm out}}$ for any $\sigma \in \faces^I$ such that
\begin{equation} \label{ke}
\begin{split}
D_t & \intO{ \left(\frac{1}{2} \varrho_h \abs{ \widehat{\vuh} }^2 + \mathcal{H}(\varrho_h )
+\frac{1}{2} \abs{\vB_h }^2 \right) }
+ \mu \norm{\nabla_h \mathbf{u}_h }_{L^2} ^2 + \nu \norm{ {\rm div}_h \mathbf{u}_h }_{L^2}^2
+ \norm{{\rm Curl}_h \vB_h }_{L^2}^2
\\ =&
- \frac{\Delta t}{2} \intO{ \varrho_h^{\triangleleft}\abs{D_t \mathbf{u}_h }^2 }
- \frac{ \Delta t}{2} \intO{ \abs{D_t \vB_h }^2 }
- \frac{\Delta t}2 \intO{ \mathcal{H}''(\xi)\abs{D_t \varrho_h }^2 }
\\&
- \sum_{ \sigma \in \faces^I } \int_{\sigma} { \left( \varrho_h^{{\rm up}} \frac{\abs{ u_\sigma } }{2}
+h^\varepsilon \avg{\varrho_h} \right)\abs{\jump{\avc{\mathbf{u}_h}}}^2 }
- \sum_{ \sigma \in \faces^I } \int_{\sigma} { \mathcal{H}''(\zeta) \jump{ \varrho_h } ^2 \left(h^\varepsilon +\frac{\abs{ u_\sigma } }{2} \right)}
\leq 0.
\end{split}
\end{equation}
\end{Theorem}
\begin{proof}
First, summing up \eqref{schemeD} and \eqref{schemeM} with the test functions $\phi_h = -\frac{\abs{ \widehat{\vuh} }^2}{2}$ and $\vv_h = \mathbf{u}_h $ implies the discrete kinetic energy balance
\begin{equation}\label{ke1}
\begin{aligned}
&\intO{D_t \left(\frac12 \varrho_h \abs{ \widehat{\vuh} }^2 \right) } + \mu \norm{\nabla_h \mathbf{u}_h }_{L^2}^2 + \nu \norm{ {\rm div}_h \mathbf{u}_h }_{L^2}^2 - \intO{ p_h {\rm div}_h \mathbf{u}_h }
\\& \quad
+ \frac{\Delta t}{2} \intO{ \varrho_h^{\triangleleft} \abs{D_t \mathbf{u}_h }^2}
+ \sum_{ \sigma \in \faces^I } \int_{\sigma} { \left(\frac12 \varrho_h^{{\rm up}} \abs{ u_\sigma }
+h^\varepsilon \avg{\varrho_h} \right)\abs{\jump{\avc{\mathbf{u}_h}}}^2 }
\\ &
= \intO{ ( {\rm Curl}_h \vB_h \times \vB_h^{\triangleleft} ) \cdot \mathbf{u}_h }
=- \intO{ {\rm Curl}_h \vB_h \cdot (\mathbf{u}_h \times \vB_h^{\triangleleft} ) }
.
\end{aligned}
\end{equation}
where we have used \eqref{fluxes} and the following equality
\[
D_t (\varrho_h \widehat{\vuh} ) \cdot \mathbf{u}_h- D_t \varrho_h \frac{| \widehat{\vuh} |^2}{2}
= D_t \Big(\frac12 \varrho_h |\mathbf{u}_h |^2 \Big) + \frac{\Delta t}2 \varrho_h^{\triangleleft} |D_t\mathbf{u}_h |^2 .
\]
Next, by setting $\vC_h =\vB_h $ in \eqref{schemeB}, we derive
\begin{equation}\label{ke2}
\intOB{ ( \mathbf{u}_h \times \vB_h^{\triangleleft} ) \cdot {\rm Curl}_h \vB_h - \alpha | {\rm Curl}_h \vB_h |^2 }
=\intO{ D_t \vB_h \cdot \vB_h }
=\intOB{ D_t \frac{|\vB_h|^2}{2} + \frac{\Delta t}2 |D_t \vB_h|^2 }
\end{equation}
Upon setting $b=\mathcal{H}(\varrho)$ in the renormalized continuity equation~\eqref{r1} and noticing the equality $\varrho \mathcal{H}'(\varrho) - \mathcal{H} (\varrho)=p(\varrho)$, we obtain the balance of internal energy
\begin{equation}\label{ke3}
\intO{ \left( D_t \mathcal{H}(\varrho_h ) - p_h {\rm div}_h \mathbf{u}_h \right) }
=
- \frac{\Delta t}2 \intO{\mathcal{H}''(\xi)|D_t \varrho_h |^2 }
- \sum_{ \sigma \in \faces^I } \int_{\sigma} { \mathcal{H}''(\zeta) \jump{ \varrho_h } ^2 \left(h^\varepsilon + \frac12 | u_\sigma | \right) }.
\end{equation}
where $\xi \in \co{\varrho_h^{\triangleleft}}{\varrho_h }$ and $\zeta \in \co{\varrho_h^{\rm in} }{\varrho_h^{\rm out} }$ are the same as in the renormalized continuity equation~\eqref{r1}.
Finally, we finish the proof by summing up the identities \eqref{ke1}--\eqref{ke3}.
\end{proof}
\medskip
\paragraph{Uniform bounds.} As a consequence of the energy estimates~\eqref{ke}
and Sobolev's inequality, we deduce the following bounds.
\begin{align}
\label{est_B}
\norm{\mathbf{u}_h}_{L^{2}L^6} \stackrel{<}{\sim} \norm{\nabla_h \mathbf{u}_h}_{L^2 L^2} \stackrel{<}{\sim} 1, \;
\Delta t^{1/2} \norm{D_t \vB_h}_{L^{\infty}L^2} \stackrel{<}{\sim} 1, \;
\norm{\vB_h}_{L^{\infty}L^2} \stackrel{<}{\sim} 1, \;
\norm{{\rm Curl}_h \vB_h}_{L^{2}L^2} \stackrel{<}{\sim} 1.
\end{align}
\subsubsection{Consistency}\label{sec_Consistency}
Another step towards the consistent approximation is the consistency.
The numerical solution of {\bf Scheme-I} satisfies the consistency criteria (\ref{cP}) of a consistent approximation. More precisely, we have the following consistency formulation.
\begin{Theorem}[Consistency of the {\bf Scheme-I}]\label{Tm2} \hspace{1em}\newline
Let $(\varrho_h, \mathbf{u}_h,\vB_h)$ be a solution of the discrete problem \eqref{scheme} on the time interval $[0,T]$ with $\Delta t\approx h$ and $\gamma > \frac{4d}{1+3d}$.
Then there exists some positive constant $\beta$ such that
\begin{subequations}\label{AcP}
\begin{equation} \label{AcP1}
\int_0^\tau \intO{ \left[ \varrho_h \partial_t \phi + \varrho_h \mathbf{u}_h \cdot \nabla_h \phi \right]} \,{\rm d} t
= - \intO{ \varrho_h^0 \phi(0,\cdot) } +
{\mathcal O}(h^{\beta})
\end{equation}
for any $\phi \in C_c^2([0,T) \times \Ov{\Omega})$;
\begin{equation} \label{AcP2}
\begin{split}
&
\intTO{ \left[ \varrho_h \widehat{\vuh} \cdot \partial_t \bfv + \varrho_h \widehat{\vuh} \otimes \mathbf{u}_h : \nabla \bfv + p_h {\rm div} \bfv \right]}
- \intTO{ \S( \nabla_h \mathbf{u}_h) : \nabla \bfv}
\\& + \intTO{ ({\rm Curl}_h \vB_h \times \vB_h ) \cdot \bfv}
=-\intO{ \varrho_h^0 \avc{\mathbf{u}_h^0} \cdot \bfv(0,\cdot) } + {\mathcal O}(h^{\beta})
\end{split}
\end{equation}
for any $\bfv \in C^2_c([0,T) \times {\Omega}; \mathbb{R}^d)$;
\begin{equation} \label{AcP3}
\intO{ \vB_h^0 \cdot \mathbf{C}(0,\cdot)}
+ \intTOB{ \vB_h \cdot \pd _t \mathbf{C}
- \alpha {\rm Curl}_h \vB_h \cdot {\rm Curl}\mathbf{C} +( \mathbf{u}_h \times \vB_h )\cdot {\rm Curl}\mathbf{C}}
= {\mathcal O}(h^{\beta})
\end{equation}
for any $\mathbf{C} \in C_c^2([0,T)\times \overline{\Omega};\mathbb{R}^d ),\,\mathbf{C} \times \vc{n}|_{\p \Omega}=0$;
\begin{equation}\label{AcP4}
\intO{\vB_h \cdot \nabla \psi} ={\mathcal O}(h)
\end{equation}
for any $\psi \in C^2( \overline{\Omega}) \cap L^2_0(\Omega)$.
\end{subequations}
\end{Theorem}
\begin{proof}
First, recalling \cite[Theorem 13.2]{FeLMMiSh} we know \eqref{AcP1} holds and there exists a $\beta>0$ such that
\begin{equation}\label{Ac1}
\begin{aligned}
& \intTO{ D_t (\varrho_h \widehat{\vuh} ) \cdot \vv_h } -
\int_0^T \sum_{ \sigma \in \faces^I } \int_{\sigma} { {\rm F}_h^{\varepsilon}(\varrho_h \widehat{\vuh} ,\mathbf{u}_h ) \cdot
\jump{ \widehat{\vv_h} } } \,{\rm d} t
\\&\quad + \mu \intTO{ \nabla_h \mathbf{u}_h : \nabla_h \vv_h }
+ \intTO{( \nu {\rm div}_h \mathbf{u}_h -p_h ) {\rm div}_h \vv_h }
\\ &= - \intO{ \varrho_h^0 \avc{\mathbf{u}_h^0} \cdot \bfv(0) } -
\int_0^T \intO{ \left[ \varrho_h \widehat{\vuh} \cdot \partial_t \bfv + \varrho_h \widehat{\vuh} \otimes \mathbf{u}_h : \nabla \bfv + p_h {\rm div} \bfv \right]} \,{\rm d} t
\\& \quad
+ \int_0^T \intO{ \S( \nabla_h \mathbf{u}_h) : \nabla \bfv} \,{\rm d} t
+ h^{\beta}
.
\end{aligned}
\end{equation}
Then we derive \eqref{AcP2} by combining \eqref{Ac1} with the following estimates
\begin{align*}
&\abs{ \intTOB{ ( {\rm Curl}_h \vB_h \times \vB_h^{\triangleleft} ) \cdot \Pi_V\bfv - ( {\rm Curl}_h \vB_h \times \vB_h ) \cdot \bfv} }
\\&= \abs{
\intTOB{ ( {\rm Curl}_h \vB_h \times \vB_h ) \cdot (\Pi_V\bfv -\bfv) +
({\rm Curl}_h \vB_h \times (\vB_h-\vB_h^{\triangleleft})) \cdot\bfv} }
\\&
\stackrel{<}{\sim} \norm{{\rm Curl}_h \vB_h}_{L^2L^2}
\left( \norm{\vB_h}_{L^2L^2} h \norm{\bfv}_{C^1}
+ \Delta t \norm{D_t \vB_h}_{L^2L^2} \norm{\bfv}_{C^0}
\right)
\stackrel{<}{\sim} h + \Delta t^{1/2} \stackrel{<}{\sim} h^{1/2}.
\end{align*}
where we have used H\"older's inequality, the uniform bounds \eqref{est_B} as well as the interpolation estimate~\eqref{IE}.
We are left with the proof of \eqref{AcP3} and \eqref{AcP4}. To proceed, we set $\vC_h =\Pi_N \mathbf{C}$ as the test function in \eqref{schemeB} and analyze each term in the following. First, for the time derivative term we have
\[ \begin{aligned}
&\int_0^T \intO{D_t \vB_h \cdot \Pi_N \mathbf{C}}
=\frac{1}{\Delta t}\int_0^T \intO{ \vB_h(t)\cdot \Pi_N \mathbf{C}(t)}
- \frac{1}{\Delta t}\int_{-\Delta t}^{T-\Delta t} \intO{ \vB_h(t)\cdot \Pi_N \mathbf{C}(t+\Delta t)}
\\& =
-\int_0^T \intO{ \vB_h(t)\cdot D_t \Pi_N \mathbf{C} }
- \frac{1}{\Delta t}\int_{-\Delta t}^{0} \intO{ \vB_h(t)\cdot \Pi_N \mathbf{C}(t+\Delta t)}
\\ & \quad + \frac{1}{\Delta t}\int_{T-\Delta t}^{T} \intO{ \vB_h(t)\cdot \Pi_N \underbrace{\mathbf{C}(t+\Delta t)}_{=0}}
\\& =
-\int_0^T \intO{ \vB_h(t)\cdot D_t \Pi_N \mathbf{C} }
- \intO{ \vB_h^0\cdot \int_0^{\Delta t}\Pi_N \mathbf{C}(t)\,{\rm d} t }
\\& =
-\int_0^T \intO{ \vB_h(t)\cdot \pd _t \mathbf{C} }
- \intO{ \vB_h^0\cdot \mathbf{C}(0)}
+I_1+I_2,
\end{aligned}
\]
where
\[I_1 = \int_0^T \intO{ \vB_h(t)\cdot (\pd _t \mathbf{C} - D_t \Pi_N \mathbf{C}) }\,{\rm d} t , \quad
I_2 = \intO{ \vB_h(0)\cdot \left(\mathbf{C}(0) - \int_0^{\Delta t}\Pi_N \mathbf{C}(t)\,{\rm d} t \right)}.
\]
By H\"older's inequality and the estimates \eqref{est_B} we have
\[
\abs{ I_1
\stackrel{<}{\sim} \norm{\vB_h}_{L^2L^2} \Delta t \norm{\mathbf{C}}_{C^2} \stackrel{<}{\sim} h, \quad
\abs{ I_2
\stackrel{<}{\sim} \norm{\vB_h^0}_{L^1} \Delta t \norm{\mathbf{C}}_{C^2} \stackrel{<}{\sim} h.
\]
Next, using H\"older's inequality again with the uniform bounds \eqref{est_B} and interpolation estimate~\eqref{IE} we derive
\[
\int_0^T \intO{ {\rm Curl}_h \vB_h \cdot ( {\rm Curl}_h \Pi_N \mathbf{C} - {\rm Curl} \mathbf{C})}\,{\rm d} t
\stackrel{<}{\sim} h\norm{\mathbf{C}}_{C^2} \norm{{\rm Curl}_h \vB_h}_{L^2L^2}
\stackrel{<}{\sim} h,
\]
and
\[
\begin{aligned}
& \intTOB{ ( \mathbf{u}_h\times \vB_h^{\triangleleft} ) \cdot {\rm Curl}_h \Pi_N \mathbf{C} - ( \mathbf{u}_h\times \vB_h ) \cdot {\rm Curl} \mathbf{C}}
\\& =
\intTO{ ( \mathbf{u}_h\times \vB_h^{\triangleleft} ) \cdot ({\rm Curl}_h \Pi_N \mathbf{C} - {\rm Curl} \mathbf{C})}
+ \intTO{ (\mathbf{u}_h\times (\vB_h^{\triangleleft} -\vB_h) ) \cdot {\rm Curl} \mathbf{C} }
\\& \stackrel{<}{\sim} \norm{\mathbf{u}_h}_{L^2L^6} \left( h \norm{\vB_h}_{L^\infty L^2} \norm{\mathbf{C}}_{C^2} + \Delta t \norm{D_t \vB_h}_{L^\infty L^2} \norm{\mathbf{C}}_{C^1} \right)
\stackrel{<}{\sim} h + \Delta t^{1/2} \stackrel{<}{\sim} h^{1/2}.
\end{aligned}
\]
Consequently, summing up the above terms finishes the proof of \eqref{AcP3}.
Finally, concerning the proof of \eqref{AcP4}, we recall Item 5 of Lemma~\ref{lem_b} to deduce
\[
\intO{\vB_h\cdot \nabla \psi} = \intO{\vB_h\cdot \nabla(\psi-\Pi_W\psi)} \stackrel{<}{\sim} \norm{\vB_h}_{L^\infty L^2} h \norm{\psi}_{C^2} \stackrel{<}{\sim} h,
\]
which completes the proof.
\end{proof}
\subsubsection{Convergence}
Now we are ready to prove the convergence of {\bf Scheme-I}.
\begin{Theorem}[Convergence of {\bf Scheme-I}]\label{Th_SA}\hspace{1em}\newline
Let $(\varrho_h, \mathbf{u}_h, \vB_h)$ be a solution to {\bf Scheme-I} with $\Delta t \approx h$ and $\gamma > \frac{4d}{1+3d}$. Then it converges in the sense of Theorem~\ref{Th2}.
\end{Theorem}
\begin{proof}
Note that the compatibility of the discrete differential operators has been presented in \cite[Section 11.4 and Section 13.4]{FeLMMiSh}.
Combing Theorem~\ref{Tm1} and Theorem~\ref{Tm2} we conclude that the numerical solution of {\bf Scheme-I} is a consistent approximation of the MHD system in the sense of Definition~\ref{def_ca}.
Applying Theorem~\ref{Th2} we derive the convergence for {\bf Scheme-I}.
\end{proof}
\subsection{Example-II}\label{oth_num}
In this example we introduce {\bf Scheme-II} on a periodic domain identified with the flat torus.
On one hand, we use the same discretization as {\bf Scheme-I} for the magnetic field. On the other hand, we follow Feireisl et al.~\cite{FLMS_FVNS} with piecewise constant discretizations for the approximation of the density, velocity, and pressure for the Navier-Stokes part.
\begin{tcolorbox}
{\bf Scheme-II.}\label{def_fv}
Let $\Omega= \mathbb{T}^d =\left( [0,1]|_{\{0,1\}}\right)^d$ and $p$ satisfy \eqref{plaw}. Given the initial data \eqref{ini_c} we set $(\varrho_h^0,\mathbf{u}_h^0, \vB_h^0) =(\Pi_Q \varrho_0, \Pi_V \mathbf{u}_0, \Pi_N \mathbf{B}_0)$ and seek $(\varrho_h , \mathbf{u}_h , \vB_h ) \in L_{\Delta t}(0,T; Q_h \times(Q_h)^d \times \Nh)$ such that
\eqref{schemeB} holds for any $\vC_h \in \Nh$ and
\begin{equation*}
D_t \varrho _K + \sum_{\sigma \in \faces(K)} \frac{|\sigma|}{|K|} {\rm F}_h^{\varepsilon}(\varrho_h ,\mathbf{u}_h ) =0,\quad \mbox{ for all } K\in \Omega_h;
\end{equation*}
\begin{equation*}
\begin{aligned}
& D_t (\varrho_h \mathbf{u}_h )_K + \sum_{\sigma \in \faces(K)} \frac{|\sigma|}{|K|}
\left( {\rm F}_h^{\varepsilon}(\varrho_h \mathbf{u}_h ,\mathbf{u}_h ) + \avg{p_h } \vc{n}
- \mu \frac{\jump{\mathbf{u}_h }}{d_\sigma}
-(\mu + \lambda) \avg{{\rm div}_h \mathbf{u}_h } \vc{n}\right)
\\& =- \frac{1}{|K|}\int_{K} {{\rm Curl}_h \vB_h \times \vB_h^{\triangleleft}},
\quad \mbox{ for all } K\in \Omega_h;
\end{aligned}
\end{equation*}
where the artificial diffusion parameter $\varepsilon$ satisfies
\begin{equation*
\varepsilon >0 \mbox{ if } \gamma \geq 2 \quad \mbox{ and }\quad \varepsilon \in(0, 2 \gamma-1 -d/3) \mbox{ if } \gamma \in(1,2).
\end{equation*}
\end{tcolorbox}
Here, the discrete operators $\jump{\cdot}$, $\avg{\cdot}$ and the numerical flux ${\rm F}_h^{\varepsilon}$ are defined in \eqref{op_diff} and \eqref{num_flux}, $\Omega_h$ is a uniform structured mesh discretization of $\Omega$ consisting of rectangles in 2D or cuboids in 3D, $d_\sigma = h$ denotes the distance between the centers of neighboring elements. Moreover, the discrete divergence operator for the piecewise constant velocity $\mathbf{u}_h \in (Q_h)^d$ is given by
\begin{equation*}
({\rm div}_h \mathbf{u}_h)_K =
\frac{1}{|K|}\sum_{\sigma\in \faces(K)}|\sigma| \avg{\mathbf{u}_h} \cdot \vc{n} \quad \forall \; K \in \Omega_h.
\end{equation*}
\begin{Remark}
The difference between {\bf Scheme-I} and {\bf Scheme-II} mainly relies in the discretization of the Navier-Stokes part.
It is analogous to check that {\bf Scheme-II} also satisfies all the properties stated in Lemma~\ref{lem_b}, e.g. conservation of mass, positivity of density, weakly divergence free of magnetic field.
Further, noticing that the stability and consistency of the Navier-Stokes part of {\bf Scheme-II} have been analyzed in \cite[Chapter 11]{FeLMMiSh} with $\gamma>1$ and $\Delta t \approx h$, we may analogously show that the numerical solution of {\bf Scheme-II} is a consistent approximation of the MHD system in the sense of Definition~\ref{def_ca}. Systematically, we have the following convergence result.
\end{Remark}
\begin{Proposition}[Convergence of {\bf Scheme-II}]\label{thm_cons}
Let $(\varrho_h, \mathbf{u}_h, \vB_h)$ be a solution to {\bf Scheme-II} with $\Delta t \approx h$ and $\gamma >1$. Then it converges in the sense of Theorem~\ref{Th2}.
\end{Proposition}
\section{Conclusion}\label{sec_end}
We introduced the concept of DW solution and consistent approximation for multi-dimensional compressible MHD system~\eqref{pde}--\eqref{ini_c}.
We derived the weak--strong uniqueness property for the DW solution, meaning that the DW solution coincides with the classical solution (emanating from the same initial data) as long as the latter exists.
Further, we proved the convergence of the consistent approximation towards the DW solution as well as the classical solution on the lifespan of the latter. Interpreting the consistent approximation as the stability and consistency of a numerical solution, we established a generalized Lax equivalence theory. Finally, we applied this theory for the convergence analysis of two mixed finite volume--finite element methods. These two methods preserve the conservation of mass, positivity of density, stability of total energy, and weakly divergence free of the magnetic field.
\bigskip
\clearpage \newpage
\centerline{\bf \large{Acknowledgements}}
The research of Y.~Li is supported by National Natural Science Foundation of China under grant No. 12001003. The research of B.~She is supported by Czech Science Foundation, Grant Agreement 21-02411S. The institute of Mathematics of the Czech Academy of Sciences is supported by RVO:67985840.
\def\cprime{$'$} \def\ocirc#1{\ifmmode\setbox0=\hbox{$#1$}\dimen0=\ht0
\advance\dimen0 by1pt\rlap{\hbox to\wd0{\hss\raise\dimen0
\hbox{\hskip.2em$\scriptscriptstyle\circ$}\hss}}#1\else {\accent"17 #1}\fi}
|
\section{Introduction}
The controlled generation and manipulation of massively entangled quantum states is one of the central tasks of modern quantum technology platforms \cite{Preskill2012Arxiv}. In the context of $S=1/2$ (or qubit) ensembles, a fundamental class of entangled quantum states is represented by spin-squeezed states \cite{Kitagawa1993PRA, Ma2011PR, Pezze2018RMP}, namely states which feature a net polarization of the collective spin, along with suppressed fluctuations of a spin component transverse to the polarization axis. Introducing the collective spin of $N$ qubits, $\hat{\bm J} = \sum_{i=1}^N \hat{\bm S}_i$ -- where $\hat{\bm S}_i$ is a $S=1/2$ spin operator -- and assuming that $\langle \hat J^{y,z} \rangle = 0$ while $\langle \hat J^x \rangle \neq 0$, a state is spin-squeezed if \cite{Wineland1994PRA}
\begin{equation}
\xi_R^2 =
\frac{N \min_{\perp}[{\rm Var}(\hat J^{\perp})]}{\langle \hat J^x \rangle^2}
< 1,
\end{equation}
where $\min_{\perp}$ indicates the minimization over the spin components transverse to the polarization axis $x$. A state exhibiting squeezing is entangled \cite{Sorensen2001}; specifically, if $\xi_R^2 < 1/k$ with integer $k$, the state cannot be represented as a separable state among clusters of less than $k+1$ spins \cite{Pezze2009PRL, Hyllus2012PRA, Toth2012PRA}; moreover the achievement of strong spin squeezing allows a quantum state to violate many-body Bell inequalities, as recently shown experimentally \cite{Schmied2016, Engelsen2017PRL}. Most importantly, spin squeezing offers a fundamental metrological advantage over separable states when used as the input for Ramsey interferometry \cite{Wineland1994PRA} allowing one to beat the standard quantum limit, as already demonstrated in seminal experiments (see e.g. Refs.~\cite{LouchetChauvet2010NJP,Leroux2010PRL,Muessel2014PRL,PedrozoPenafiel2020}). Identifying different many-body mechanisms leading to spin squeezing is therefore of paramount importance: the entanglement of the resulting states can be certified, and their metrological potential exploited by accessing their most basic physical property, namely the collective spin.
The paradigmatic spin-squeezing dynamics is the one generated by the so-called one-axis-twisting (OAT) Hamiltonian \cite{Kitagawa1993PRA}
\begin{equation}
\hat{\cal H}_{\rm OAT} = \frac{({\hat J}^z)^2}{2I} ~,
\label{e.OAT}
\end{equation}
namely, the Hamiltonian of a planar rotor with moment of inertia $I$.
For the energy to be extensive, one must assume that $I \sim N$. Under this assumption, starting from the initial state $|{\rm CSS}_x\rangle = \otimes_{i=1}^{N}|\rightarrow_x\rangle_i$ (the coherent spin state - CSS - polarized along the $x$ axis) the strongest squeezing is achieved at a time $t_{\rm min} \sim N^{-2/3} I \sim N^{1/3}$; and it scales as $\xi^2_{R,\min} \sim N^{-2/3}$ \cite{Kitagawa1993PRA}. This Hamiltonian (as well as related ones) and the corresponding squeezing dynamics have been realized in seminal experiments exploiting interactions in Bose-Einstein condensates \cite{Esteve2008, Riedel2010, Hosten2016} and in trapped ions \cite{Bohnet2016}, to cite a few relevant examples.
Interestingly, the OAT Hamiltonian of Eq.~\eqref{e.OAT} plays also a special role in condensed matter, to explain the mechanism of spontaneous breaking of a continuous symmetry in quantum spin models \cite{Beekman2019SPPLN, Lauchli2016, Tasaki2018JSP}.
Hereafter, we consider models with an axial rotational ($U(1)$) symmetry, and we shall choose the $z$ axis as the symmetry axis.
Such models have a ground state which is also $U(1)$ symmetric for any finite size. As suggested by Anderson in pioneering works \cite{Anderson1952PR,Anderson1997}, a quantum spin model in the thermodynamic limit can break a continuous symmetry (such as $U(1)$) by developing a finite order parameter due to the existence of a set of low-energy Hamiltonian eigenstates -- the so-called Anderson's tower of states (ToS). These states are approximately eigenstates of a planar rotor Hamiltonian of the kind of Eq.~\eqref{e.OAT}; and their energy decreases as ${\cal O}(1/N)$ in the thermodynamic limit (due to the scaling of the moment of inertia $I$), making them nearly degenerate with the $U(1)$ symmetric ground state. Spontaneous symmetry breaking (SSB) is therefore the result of the collapse of the ToS.
The central insight of the present work is that generic quantum spin Hamiltonians can produce a persistent spin-squeezing dynamics thanks to the emergent ToS structure of their low-energy eigenstates -- in other words, several models with $U(1)$ symmetry generate the same dynamics as that of the OAT Hamiltonian Eq.~\eqref{e.OAT} at sufficiently short times, irrespective of the nature (long-range vs. short-range) of their interactions.
This is strictly true when the dynamics is initialized in a CSS polarized in the symmetry plane. Such a state, maximizing the total spin $\langle \hat {\bm J}^2 \rangle = J_{\rm max}(J_{\rm max}+1)$ with $J_{\rm max} = N/2$ (for an ensemble of qubits), has a very strong overlap with the ToS, because the latter states have in turn an anomalously large average value of $\hat {\bm J}^2$ among all the states in the same energy range -- namely they behave as ``quantum many-body scars" \cite{Turner2018NP, Serbyn2021NP}, whose presence alters profoundly the dynamics of the system with respect to a generic initial state with the same initial energy.
We underpin this insight with a combination of exact diagonalization (ED) and time-dependent variational calculations of the dynamics of various $S=1/2$ $XX$ models, fully corroborating the universal picture of spin-squeezing dynamics.
Our results imply that a large variety of current quantum simulation setups implementing $U(1)$ symmetric models of quantum magnetism \cite{Brydges2019, Browaeys2020NP, Jepsen2020, Chiaro2022PRR, Guo2021} can be viewed as generators of spin-squeezed states, of potential immediate interest for quantum metrology tasks.
{The structure of the paper is as follows. In Sec.~\ref{section:model_method} we briefly present the $\alpha-XX$ model and our numerical approach to simulate quantum dynamics; in Sec.~\ref{section:tos} we identify the Tower of States for the model under study; Sec.~\ref{section:squeezing} discusses our variational results for spin squeezing and Sec.~\ref{section:squeezing_and_tos} describes the connection between the robustness of squeezing dynamics and the Tower of States.}
\section{Models and method}
\label{section:model_method}
We specialize our attention to the case of $XX$ models with power-law decaying interactions ($\alpha-XX$ models), whose Hamiltonian reads
\begin{equation}
\hat{\cal H} = - \sum_{i<j} {\cal J}^{(\alpha)}_{ij} \left ( \hat S_i^x \hat S_j^x + \hat S_i^y \hat S_j^y \right ) ~,
\end{equation}
where ${\cal J}^{(\alpha)}_{ij} = {\cal J} \, |\bm r_i - \bm r_j|^{-\alpha}$ with $\alpha \geq 0$. The limit $\alpha = 0$ corresponds to infinite-range interactions (reproducing the OAT Hamiltonian), while the opposite limit $\alpha \to \infty$ corresponds to nearest-neighbor interactions~\footnote{The $\alpha-XX$ Hamiltonian features extensive eigenvalues only for $\alpha>d$, $d$ being the number of dimensions; later, when appropriate, we will adopt a Kac normalization of the coupling constant in order to reinstate the extensive nature of the energy.}.
The above Hamiltonian occupies a prominent role in quantum simulation, as it is currently implemented on various lattice geometries by a rich variety of very different quantum simulation platforms, including trapped ions ($0 < \alpha < 3$) \cite{Brydges2019}; Rydberg atoms ($\alpha = 3$) \cite{Browaeys2020NP}; spinor gases ($\alpha = \infty$) \cite{Jepsen2020}; and superconducting circuits ($\alpha = \infty$) \cite{Chiaro2022PRR, Guo2021}. The generation and metrological exploitation of spin squeezed states in these platforms has been already discussed in the recent past (see e.g. Refs.~\cite{Gil2014PRL, Bohnet2016, Qu2019PRA, VanDamme2021PRA, Groszkowski2022PRX}), although not in connection with the operation mode implementing the $\alpha-XX$ model.
Throughout the paper, we will consider the quench dynamics generated by $\alpha-XX$ Hamiltonians starting from the $|{\rm CSS}_x\rangle$ state, $|\Psi(t)\rangle = \exp(-i\hat{\cal H} t) |{\rm CSS}_x\rangle$. Under this assumption, the sign of the ${\cal J}$ coupling is irrelevant as long as one follows the expectation value of operators which are real matrices (e.g. on the computational basis of the eigenstates of the $\hat{S}_i^z$ operators) \cite{Frerot2018PRL}. Hereafter we will assume ${\cal J} >0$ for definiteness. We shall study this model on lattices with $N=L^d$ sites (in $d=1$ and 2) with periodic boundary conditions.
Studying the real-time dynamics generated by the Hamiltonian with $\alpha > 0$ is generically a challenging problem.
Recent results based on a semi-classical approach \cite{Perlin2020PRL} already indicate the robustness of squeezing when moving away from the $\alpha=0$ limit. Here we adopt a different strategy, going beyond any semi-classical framework.
Specifically, we use exact diagonalization for small systems \cite{Weinberg2017SP, Weinberg2019SP}, and a time-dependent variational Monte Carlo (tVMC) approach to tackle large system sizes. The latter are both relevant to current experimental realizations and instrumental to our scaling analysis of the spin-squeezing dynamics discussed below.
Our tVMC calculations are based on the pair-product Ansatz (or two-spin long-range entangled-plaquette state, 2LR-EPS \cite{Thibaut2019PRB})
\begin{equation}
|\Psi(t)\rangle =:
\sum_{\bm \sigma} \prod_{i<j} \psi_{ij}(\sigma_i,\sigma_j;t) |\bm \sigma\rangle,
\label{e.ansatz}
\end{equation}
where $|\bm \sigma \rangle = |\{ \sigma_i \}\rangle$ is the joint eigenstate of all $\hat{S}_i^z$ operators. The evolution of the variational parameters $\psi_{ij}(\sigma_i,\sigma_j;t)$ is dictated by the time-dependent variational principle, and it requires Monte Carlo sampling of the probability distribution $|\langle {\bm \sigma} | \Psi(t) \rangle|^2$ \cite{Carleo2012SR, Becca2017}.
{As discussed in the following,} the chosen Ansatz has the advantage of reproducing the \emph{exact} dynamics in the $\alpha=0$ limit; and, as we shall see, it remains very accurate for $\alpha > 0$.
{\subsection{Pair-product Ansatz is exact for the one-axis twisting Hamiltonian}
The exact evolution of the CSS under the OAT Hamiltonian is restricted to the $J = J_{\rm max}$ sector, and it reads
\begin{equation}
\begin{aligned}
&e^{-i \hat{\cal{H}}_{\rm OAT} t} |{\rm CSS}_x \rangle
=\\
&\sum_{J^z=-J_{\rm max}}^{J_{\rm max}}
\langle J_{\rm max}, J^z | {\rm CSS}_x \rangle
\exp\left[-it \frac{(J^z)^2}{2 I} \right] |J_\mathrm{max}, J^z \rangle,
\end{aligned}
\end{equation}
where we set $\hbar = 1$ and where
\begin{equation}
\langle J_{\rm max}, J^z | {\rm CSS}_x \rangle =
2^{-J_{\rm max}}
\sqrt{\begin{pmatrix} 2 J_{\rm max} \\ J_{\rm max}-J^z \end{pmatrix} } ~.
\end{equation}
Provided that a certain variational Ansatz correctly represents the initial state, then it can also exactly reproduce the time-evolved state if the coefficients $\langle {\bm \sigma} | \Psi(t) \rangle$ can take values proportional to $\exp[-i t (J^z)^2 / (2 I)]$, with $J^z = \sum_{i=1}^N \sigma_i$ and where $\sigma_i = \pm 1/2$ is the eigenvalue of $\hat{S}_i^z$.
The pair-product Ansatz in Eq.~\eqref{e.ansatz} has this property, as we show by construction.
If we set its coefficients to be
\begin{equation}
\psi_{jk}(\sigma_j, \sigma_k; t) =
\exp\left[w_{jk} \sigma_j \sigma_k \right] ~,
\qquad
w_{jk} = -\frac{i t}{I} ~,
\label{se.ansatz_for_OAT}
\end{equation}
then the value of $|\Psi(t)\rangle$ on a given basis state $|{\bm \sigma}\rangle$ reads
\begin{equation}
\begin{aligned}
\langle {\bm \sigma} | \Psi(t) \rangle &=
\prod_{j<k} \psi_{jk}(\sigma_j, \sigma_k; t) =
\exp
\left[ -\frac{i t}{I} \sum_{j<k} \sigma_j \sigma_k \right]
=\\
&=
\exp\left[-i t \, \frac{(J^z)^2}{2I} \right]
\exp\left[ i t \, \frac{N}{8I} \right].
\end{aligned}
\end{equation}
This corresponds to the required form of $\langle {\bm \sigma} | \Psi(t) \rangle$, up to an irrelevant global phase factor.
Therefore the pair-product Ansatz can describe the exact time evolution of $|{\rm CSS}_x\rangle$. Note that the initial state is trivially represented with this Ansatz by setting all coefficients $\psi_{jk}(\sigma_j, \sigma_k ; t = 0)$ equal to each other.
The Hamiltonian of the $\alpha-XX$ model with $\alpha=0$,
\begin{equation}
\hat{\cal{H}} =
\frac{\mathcal{J}}{2} \left[(\hat{J}^z)^2 - \hat{\bm J}^2\right] + \mathrm{const}~,
\end{equation}
corresponds to the OAT Hamiltonian with moment of inertia $I = 1 / {\cal J}$, up to a constant shift and to an additional $\hat{\bm J}^2$ term (which only adds an irrelevant global phase factor to the time-evolved state).
Thus the pair-product Ansatz is also exact for the $\alpha-XX$ model with $\alpha = 0$.
We explicitly verify this fact by comparing the tVMC dynamics of spin squeezing to the exact expression for the OAT squeezing dynamics \cite{Kitagawa1993PRA}. As shown in Fig. \ref{f.jastrow_OAT}, the tVMC calculation for the $\alpha-XX$ model with $\alpha=0$ is exact.
Note that the dynamics of variational parameters is obtained through the tVMC scheme \cite{Becca2017} {based on the time-dependent variational principle}, that is, without postulating the expected expression in Eq. \eqref{se.ansatz_for_OAT}.
\begin{figure}[hbt]
\centering
\includegraphics[width=0.8\linewidth]{figsupp_1D_alpha_0.pdf}
\caption{Spin squeezing generated by the $\alpha-XX$ Hamiltonian, with $d=1$ and $\alpha=0$. For each system size $N$, the result of tVMC with the pair-product Ansatz (solid lines) perfectly matches with the exact solution for the OAT Hamiltonian \cite{Kitagawa1993PRA}.}
\label{f.jastrow_OAT}
\end{figure}
}
\section{Tower of states as quantum many-body scars}
\label{section:tos}
The existence of an Anderson's ToS in the low-energy spectrum of the $\alpha-XX$ model can be directly verified using ED on small system sizes \cite{Lauchli2016}. Fig.~\ref{f.ToS} shows the low-energy spectrum of a $N=16$ chain with $\alpha = 1, 3$ and $\infty$; the ToS is clearly visible when plotting the energies as a function of the quantum number $J^z$, since the ground states in each sector have an energy which grows almost exactly as $(J^z)^2$, as expected for the eigenstates of $\hat{\cal H}_{\rm OAT}$. Remarkably, a ToS is also observed for $\alpha \geq 3$, although in this case it does not collapse onto the ground state faster than the spin-wave excitations do \cite{Frerot2017PRB}, so that SSB is not realized in the ground state \cite{Maghrebi2017PRL}.
\begin{figure*}[ht!]
\includegraphics[width=0.95\textwidth]{fig1_1D_L16_spectra_and_scars.pdf}
\caption{Low-energy spectrum of the $d=1$ $\alpha-XX$ model for (a)-(d) $\alpha = 1$; (b)-(e) $\alpha = 3$; (c)-(f) $\alpha = \infty$. The spectrum is shown as a function of the two quantum numbers $E = \langle \hat{\cal H} \rangle$ and $(J^z)^2$ in panels (a-c), and as a function of $E$ and the average total spin $\langle \hat {\bm J}^2 \rangle$ in panels (d-f). Black dots correspond to states of the ToS, while orange (gray) dots are outside the ToS; straight lines are linear fits of the ToS states.}
\label{f.ToS}
\end{figure*}
The states of the ToS, being ground states of different $J^z$ sectors, can be naturally expected to exhibit non-typical features among all the states in the corresponding energy range. In particular, when mapping the $\alpha-XX$ model onto hardcore bosons \cite{Matsubara1956PTP}, $J^z$ parametrizes the particle number, such that the ToS is readily understood as the set of the (quasi-)condensate ground states for different particle numbers~\footnote{The states of the ToS, being ground states of hardcore bosons with long-range hopping, are expected to exhibit area-law scaling of entanglement entropy, at least for $\alpha > d$, while logarithmic scaling of the entropy with subsystem size can be exactly proven for $\alpha=0$. This is in stark contrast with typical states at finite energy density.}. In particular, the ToS contains the states featuring the most slowly decaying correlations $C^{x(y)}_{ij} = \langle \hat S_i^{x(y)} \hat S_j^{x(y)} \rangle$ (corresponding to the one-body density matrix for bosons) among all the states at the same $J^z$: therefore they are the states maximizing the total angular momentum $\langle \hat {\bm J}^2 \rangle = (J^z)^2 + \sum_{ij} C^{x(y)}_{ij}$. Fig.~\ref{f.ToS} shows that they are also the states with the largest $\langle \hat{\bm J}^2 \rangle$ throughout their energy range. As such, they represent a paradigmatic example of quantum many-body scars \cite{Turner2018NP}.
Due to their anomalously large value of $\langle \hat{\bm J}^2 \rangle$, the states of the ToS are naturally expected to play a significant role in the dynamics of the system when starting from the $|{\rm CSS}_x\rangle$ state, given that for this state $\langle \hat{\bm J}^2 \rangle$ is maximal. Indeed, as shown in Appendix~\ref{app:overlap}, the CSS has maximal overlap with the ToS.
\section{Squeezing dynamics and its scaling}
\label{section:squeezing}
The existence of a set of quantum-scar states with an energy dependence on $J^z$ akin to that of the OAT Hamiltonian in Eq.~\eqref{e.OAT}, and which are strongly overlapping with the CSS, suggests that, at sufficiently short times, the Hamiltonian evolution starting from the CSS will strongly resemble that of the OAT model -- namely it will feature squeezing. Probing the persistence of this squeezing dynamics and its quantitative relationship with the Anderson ToS will be the goal of the rest of this work.
Fig.~\ref{f.squeezing} clearly exhibits the existence of squeezing dynamics for different $\alpha$ values in $d=1$ and for variable system sizes up to $N=128$ spins. Our tVMC results are indistinguishable from the exact ones (which we obtain for $N=16$) up to $\alpha \approx 1$, remaining accurate for all values of $\alpha$ -- see also additional data in Appendix~\ref{app:squeezing_data}.
\begin{center}
\begin{figure*}[htb]
\includegraphics[width=0.75 \textwidth]{fig2_1D_squeezing.pdf}
\caption{Dynamics of the squeezing parameter for the 1$d$ $\alpha-XX$ Hamiltonian with $\alpha = 1$ (a), $\alpha = 3$ (b) and $\alpha = \infty$ (c), for system size from $N=16$ (bottom line) to $N=128$ (top line).
Time is rescaled by the Kac prefactor $K_N^{(\alpha)}$ -- see text. The dashed line is the exact result for $N=16$, while the other data are obtained via tVMC (solid lines).
}
\label{f.squeezing}
\end{figure*}
\end{center}
{ For each $\alpha$, we identify the minimal value of the squeezing parameter $\xi^2_{R,\min}$ and the corresponding optimal time $t_{\rm min}$;
Figs. \ref{f.squeezing_scaling}(a) and \ref{f.squeezing_scaling}(b) show the scaling of these two quantities with the system size.}
\begin{figure}[ht!]
\centering
\includegraphics[width=\linewidth]{fig2_1D_exponents}
\caption{Finite-size scaling of the optimal time $t_\mathrm{min}$ (a) and of the minimal squeezing $\xi^2_{R,\mathrm{min}}$ (b), for several values of $\alpha$. Dots are tVMC results; straight lines are power-law fits with $t_\mathrm{min}\mathcal{J} K_N^{(\alpha)} \sim N^{\mu}$ and $\xi^2_{R,\mathrm{min}} \sim N^{-\nu}$. The fit range is $32 \leq N \leq 128$.
(c) Exponent $\mu$ for the scaling of the optimal time. (d) Exponent $\nu$ for the scaling of the minimum squeezing parameter. In panels (c-d), the horizontal black line shows the OAT exponents, and the vertical dotted line marks $\alpha=5d/3$ (see text).
}
\label{f.squeezing_scaling}
\end{figure}
The results of the scaling analysis in the $1d$ $\alpha-XX$ model are summarized in Fig.~\ref{f.squeezing_scaling}(c-d). Following the example of the OAT model, we postulate \cite{FossFeig2016Arxiv, Perlin2020PRL} the power-law scalings
\begin{equation}
\xi^2_{R,\min} \sim N^{-\nu}, ~~~~~~~ t_{\rm min} K^{(\alpha)}_N \sim N^{\mu}~.
\end{equation}
Here we have introduced the Kac normalization
\begin{equation}
K^{(\alpha)}_N = N^{-1} \sum_i \sum_{j \neq i} |\bm r_i - \bm r_j|^{-\alpha}
\end{equation}
of the power-law couplings in order to properly scale time, by considering evolutions with Hamiltonians with extensive energies~\footnote{In particular, in the limit $N\to\infty$, we have that $K^{(\alpha)}_N \sim N^0$ for $\alpha > 1$, $K^{(\alpha)}_N \sim \ln N$ for $\alpha = 1$ and $K^{(\alpha)}_N \sim N^{1-\alpha}$ for $\alpha < 1$.} -- this is appropriate for all the quantum-simulation platforms cited above. The system sizes we have considered (up to $N=128$) do not necessarily capture the asymptotic scaling limit, but they are comparable with the typical sizes achieved by state-of-the-art quantum simulators for the $\alpha-XX$ model.
In Fig.~\ref{f.squeezing}(d-e) we observe that the scaling properties of the OAT limit $\alpha = 0$ (namely $\mu = 1/3$ and $\nu = 2/3$) are essentially maintained throughout the range $0 < \alpha \lesssim 1.5$. This observation reveals (far beyond what is accessible to ED) the dominant role that the ToS has on the dynamics of the system in this range of $\alpha$, exceeding the regime of long-range interactions $\alpha \leq 1$. On the other hand for $\alpha \gtrsim 2$ both $\mu$ and $\nu$ are found to sharply drop to zero; this observation signals that the dynamics leaks significantly out of ToS manifold, and that further Hamiltonian eigenstates entering in the dynamics have the effect of curbing the growth of squeezing and suppressing its scaling behavior. {In the following section, using spin-wave (SW) theory \cite{Frerot2017PRB,Frerot2018PRL}, we can estimate that $\alpha < 5d/3$ (for $d\leq 3$) is a necessary condition for the squeezing dynamics of the OAT to be reproduced at $\alpha > 0$. This would translate to $\alpha < 1.6666...$ in $d=1$, in apparent agreement with the observation of Fig.~\ref{f.squeezing_scaling}(d).}
{ \subsection{Necessary condition on the persistence of $\alpha=0$ scaling of squeezing from spin-wave theory }
In a (Kac-normalized) OAT Hamiltonian, with all-to-all interactions, the optimal squeezing time scales as
\begin{equation}
t_{\rm min} \sim N^{1/3} = L^{d/3}~.
\end{equation}
Taking the OAT Hamiltonian literally, its dynamics is characterized by the total absence of retardation effects: any signal propagates instantaneously from end to end of the system. As a consequence, the dynamics of the spin system is completely captured by that of the collective spin variables, and there is no relative dynamics between spins.
Even in the picture of the OAT model, there is still a finite time required for the establishment of maximal squeezing, scaling with system size. This scaling also contains the Kac normalization, namely the interactions have no retardation, but their strength is decreasing as $N^{-1}$ to keep the energy extensive. This has the fundamental effect of making the optimal time increase with system size.
When considering instead $\alpha > 0$, the relevant excitations involved in the dynamics starting from the CSS$_{x}$ state are not only the states from Anderson's ToS, but also spin-wave (SW) excitations. These excitations may have in general a finite group velocity, or a group velocity diverging with system size sufficiently slowly, so that they could lead to retardation effects in the dynamics of the collective spin, thereby altering substantially the picture of the OAT model.
SW theory for the $\alpha-XX$ model \cite{Frerot2017PRB,Frerot2018PRL} predicts that SW excitations have a dispersion relation $\omega \sim k^z$ at small $k$, with dynamical exponents $z$ taking values
\begin{equation}
z = \begin{cases} 1 ~~~~~~~~~~~ \alpha \geq d+2 \\ \frac{\alpha - d}{2} ~~~~~~~ d \leq \alpha \leq d+2 \\ 0 ~~~~~~~~~~~ \alpha \leq d \end{cases}~.
\end{equation}
The associated group velocity is therefore $v_g \sim k^{z-1}$. On a finite system of linear size $L$ the maximum group velocity is therefore scaling as $v_{g, {\rm max}} \sim k_{\rm min}^{z-1} \sim L^{1-z}$ where $k_{\rm min} = 2\pi/L$. Associated with this maximal group velocity there is an intrinsic minimal retardation time
\begin{equation}
t_{\rm SW} = \frac{L}{v_{g,\rm max}} \sim L^z
\end{equation}
which is the time needed for the fastest spin-wave excitations to traverse the entire system of linear size $L$.
When $\alpha \leq d$, the vanishing of the $z$ exponent implies that the retardation time does not scale with system size, so that the spatial decay of interactions is not expected to affect the collective spin dynamics at the time scale of $t_{\rm min}$ (which instead grows with $L$). On the other hand, when $\alpha > d$ one has that $z>0$, so that retardation effects can play a role. Given that $t_{\rm SW}$ is the minimal retardation time in the collective spin dynamics, one can argue that a \emph{necessary} condition for retardation effects not to affect the collective spin dynamics up to the optimal squeezing time $t_{\rm min}$ is that
\begin{equation}
t_{\rm SW} < t_{\rm min}~.
\end{equation}
For this condition to remain valid for large system sizes, one should then require that $L^z < L^{d/3}$, namely
\begin{equation}
z < \frac{d}{3} ~~ \Longrightarrow ~~ \alpha < \frac{5}{3} d ~.
\end{equation}
The above necessary condition is valid if $\frac{5}{3} d \leq d+2$, namely if $d \leq 3$, covering all situations of interest.
For $d=1$, the necessary condition reads $\alpha<5/3=1.666..$, and our numerical results suggest that this necessary condition may also be sufficient -- see Fig.~\ref{f.squeezing_scaling}(c-d).
{If the same condition is sufficient as well for $d>1$, this would imply that the OAT scaling of spin squeezing dynamics is expected e.g. in dipolar systems ($\alpha = 3$) in $d=2$, a situation which is highly relevant for experiments on arrays of Rydberg atoms \cite{Browaeys2020NP} or for dipolar molecules in optical lattices \cite{Yan2013}, to cite two important examples.}
{We also remark that, in the case of the $\alpha-XX$ model in $d=1$ discussed here, the persistence of the OAT squeezing dynamics for $\alpha < 5/3$ is accompanied by the persistence of a large value of the total spin length, $\langle {\bm J}^2 \rangle$ for the largest system sizes we considered, as shown in Appendix~\ref{app:Jsquare}; while for larger $\alpha$ values $\langle {\bm J}^2 \rangle$ decays to values which decrease rather fast with system size. This aspect must also be related to the $\alpha$-dependence of the low-temperature thermodynamics of the system, which in principle dictates the long-time behavior of the non-equilibrium evolution within a picture of thermalizing dynamics, expected for the $\alpha-XX$ chain which is non-integrable for all $0 < \alpha < \infty$. Establishing a quantitative relationship is nonetheless beyond the scope of the current work, and it will be subject of a future work.}
\begin{center}
\begin{figure}[ht!]
\includegraphics[width=\columnwidth]{fig3_squeezing_and_Irot.pdf}
\caption{\emph{Universal short-time squeezing dynamics.} Squeezing dynamics for the $\alpha-XX$ model in (a) $d=1$ ($N=16$) and (b) $d=2$ ($N=4\times 4$), for various values of $\alpha$ from $\alpha=0$ (darkest line) to $\alpha=\infty$ (lightest line). Collapse of the curves at short times is obtained by rescaling time with the $I^{(\alpha)}$ parameter; (c-d) Comparison between the three estimates of the moment of inertia: from the Hamiltonian projection onto Dicke states ($I^{(\alpha)}$); from the Fourier transform of the coupling constant ($I_{\gamma}^{(\alpha)}$); and from the ToS spectrum ($I_{\rm ToS}^{(\alpha)}$). System Hamiltonian and sizes as in panels (a-b).}
\label{f.universal}
\end{figure}
\end{center}
}
\section{Relationship of the squeezing dynamics to the ToS spectrum}
\label{section:squeezing_and_tos}
The connection between the OAT dynamics and the short-time dynamics of generic $\alpha-XX$ models can be made fully quantitative by identifying the moment of inertia $I^{(\alpha)}$ of the effective OAT Hamiltonian emerging from the $\alpha-XX$ model. The general way to do this is to assume that the short-time dynamics of the $\alpha-XX$ model prepared in the CSS remains confined in the sector with maximum total angular momentum $J_{\rm max}$ -- this assumption will be justified below. If this is true, then the effective Hamiltonian governing the dynamics is given by the restriction of the $\alpha-XX$ model onto the subspace of Dicke states $|J_{\rm max},J^z\rangle$ which are eigenstates of $\hat{\bm J}^2$ with $J = N/2$, as well as of $\hat J^z$; given that $\hat {\cal H}$ commutes with $\hat J^z$ and it is invariant under inversion of the spins along $z$, the resulting projected Hamiltonian is necessarily an even function of $J^z$
\begin{eqnarray}
\langle J_{\rm max}, J^z | \hat {\cal H} | J_{\rm max}, (J^z)' \rangle = ~~~~~~~~~~~~~~~~~~~~~&&\nonumber \\
\delta_{J^z, (J^z)'} [ {\rm const.} + (J^z)^2/(2I^{(\alpha)}) + {\cal O}(J^z)^4]~. &&
\end{eqnarray}
The eigenvalues of the Hamiltonian projected onto the Dicke states turn out to be almost perfect quadratic functions of $J^z$ (Appendix \ref{app:diagonal}), which allow us to systematically extract the corresponding moment of inertia $I^{(\alpha)}$. The picture of the effective dynamics projected onto the Dicke-state manifold would then predict that all $\alpha-XX$ models squeeze the fluctuations of the CSS in the same way as the OAT Hamiltonian does -- so that universal dynamics should be manifested when properly rescaling time by $t_{\alpha} = I^{(\alpha)}$. This is indeed observed in Fig.~\ref{f.universal}(a-b), exhibiting a perfect collapse for the short-time evolution of the squeezing parameter. Yet the time over which the $\alpha-XX$ model reproduces the spin squeezing dynamics of the OAT is strongly dependent on $\alpha$ and on $d$, as a result of the properties of the ToS in the Hamiltonian spectrum.
To quantitatively connect the universal spin-squeezing dynamics observed above with the existence of the ToS in the spectrum one can adopt a Fourier decomposition of the $\alpha-XX$ model \cite{Beekman2019SPPLN} as $\hat {\cal H} = \sum_{\bm q} \hat{\cal H}_{\bm q} + {\rm const.}$, where
\begin{equation}
\hat{\cal H}_{\bm q} =
-\frac{\cal J}{2N}
(\gamma_{\bm q, N}^{(\alpha)}+1)
\left ( \hat J_{\bm q}^x \hat J_{-\bm q}^x + \hat J_{\bm q}^y \hat J_{-\bm q}^y \right)~.
\end{equation}
We have introduced the Fourier transform of the collective spin
\begin{equation}
\hat J_{\bm q}^{x(y)} = \frac{1}{\sqrt{N}} \sum_i e^{i\bm q \cdot {\bm r}_i} \hat S^{x(y)}_i~,
\end{equation}
while
\begin{equation}
\gamma^{(\alpha)}_{\bm q, N} = \sum_{\bm r \neq 0} e^{i\bm q \cdot {\bm r}} |\bm r|^{-\alpha}
\end{equation}
is the Fourier transform of the couplings calculated on an $N$-site system. The $\hat{\cal H}_{{\bm q}=0}$ Hamiltonian corresponds to a OAT model with $[I_\gamma^{(\alpha)}]^{-1} = {\cal J}(\gamma^{(\alpha)}_{0,N}+1)/N$.
The emergence of a ToS in the spectrum of the $\alpha-XX$ model can be understood as the fact that the $\hat{\cal H}_{\bm q\neq 0}$ terms in the Hamiltonian perturb only weakly the eigenstates of the ${\cal H}_{{\bm q}=0}$ part with maximal $\langle \hat{\bm J}^2 \rangle$; as a consequence there exist special Hamiltonian eigenstates which are parametrically closer to Dicke states with $J=J_{\rm max}$, the smaller $\alpha$ is.
These states have energies $E \approx (J^z)^2/(2 I^{(\alpha)}_{\rm ToS}) + {\rm const.}$, with a moment of inertia $I^{(\alpha)}_{\rm ToS}$ that can be extracted by fitting the energy spectrum as in Fig.~\ref{f.ToS}. The $I^{(\alpha)}_{\rm ToS}$ parameter differs from the $I_\gamma^{(\alpha)}$ parameter in that it is renormalized by the $\hat{\cal H}_{\bm q\neq 0}$ perturbations; and it differs from the $I^{(\alpha)}$ parameter, because it includes the mixing of the $J_{\rm max}$ sector of Hilbert space with all the other $J$ sectors. Yet these three definitions of the effective moment of inertia -- a dynamical one ($I^{(\alpha)}$), an Hamiltonian one ($I_{\gamma}^{(\alpha)}$) and a spectral one ($I^{(\alpha)}_{\rm ToS}$) -- coincide for $\alpha = 0$, and they turn out to be very close to each other for all $\alpha$ values, as shown in Fig.~\ref{f.universal}(c-d) for both $d=1$ (for $\alpha \lesssim 2$) and $d=2$. This result corroborates the picture in which the $\alpha-XX$ Hamiltonians with high connectivity couple weakly the $J_{\rm max}$ sector with all the other sectors of Hilbert space, resulting simultaneously in 1) the existence of quantum scars (the ToS) in the spectrum, predominantly overlapping with the $J_{\rm max}$ sector; and 2) the existence of robust spin-squeezing dynamics at short times, which remains temporarily confined in the $J_{\rm max}$ sector (see Appendix~\ref{app:Jsquare} for the dynamics of $\langle \hat{{\bm J}}^2 \rangle$), reproducing the behavior of the OAT model with an effective moment of inertia related to the energy spectrum of the ToS.
{ This picture is to be contrasted with the one for models lacking a ToS.
As we showed recently, short-time spin-squeezing dynamics starting from a coherent spin state is a very general phenomenon common to a large class of Hamiltonians featuring bilinear spin-spin interactions with parity conservation~\cite{Roscilde2021PRA}. The ability of an Hamiltonian to induce spin squeezing at short times has therefore \emph{a priori} no specific relationship to the existence of an Anderson ToS in the spectrum.
{In Appendix~\ref{app:dimerized} we describe in detail an example (the dimerized $XX$ chain) where the ToS is absent. While short-time squeezing dynamics is also observed for that Hamiltonian (in compliance with the theorem of Ref.~\cite{Roscilde2021PRA}), we show that its characteristic time scale cannot be directly related to features of the spectrum, because the states participating in the dynamics (namely overlapping significantly with the CSS) lie in the ``bulk" of the spectrum and do not form a well-defined tower of quantum scars.} }
\section{Conclusions}
Focusing on $U(1)$ symmetric models, in this work we have unveiled a dynamical manifestation of the Anderson's tower of states -- the fundamental spectral mechanism of spontaneous breaking of continuous symmetries in quantum mechanics -- in the form of an effective squeezing dynamics of coherent spin states, fully equivalent at short times to that generated by a planar-rotor Hamiltonian.
In particular we find that the one-dimensional $XX$ model with power-law interactions generates squeezing that scales with system size when the decay exponent takes values $\alpha \lesssim 5/3$.
These results establish an important link between the simulation of non-equilibrium dynamics of $U(1)$ quantum spin Hamiltonians -- a common task of nearly all quantum simulation platforms -- and quantum metrology.
\begin{acknowledgments} We acknowledge useful discussions with Johannes Schachenmayer. Exact results for small systems are obtained via the QuSpin package \cite{Weinberg2017SP,Weinberg2019SP}.
This work is supported by ANR (``EELS" project) and by QuantERA (``MAQS" project). All numerical simulations have been performed on the PSMN cluster of the ENS of Lyon. Data and additional details about the tVMC numerical simulations are made publicly available \cite{Comparin2022PRA_suppdata}.
\end{acknowledgments}
|
\section{Introduction}
\label{SectIntro}
\IEEEPARstart{M}{atrices} with low-rank structure are omnipresent in signal processing, data analysis, scientific computing and machine learning applications including system identification \cite{FazelPST13}, subspace clustering \cite{CaiLi15}, matrix completion \cite{EftekhariYW18}, background subtraction \cite{AybatI15, MFKDeJSTSP18}, least-squares regression \cite{ClarWood2017}, hyperspectral imaging \cite{Rasti18, ErichsonMWK18}, anomaly detection \cite{FowlerDu12,KaDeICASSP17}, subspace estimation over networks \cite{ChenRS7859344}, genomics \cite{DarnellGME17, UbaruSaad19}, tensor decompositions \cite{7038247}, and sparse matrix problems \cite{DavisRS2016}.
\stepcounter{footnote}\footnotetext {This paper has been accepted for publication in the IEEE Transactions on Signal Processing (submitted December 2019, accepted March 2021).}
\stepcounter{footnote}\footnotetext {Maboud F. Kaloorazi is with Xi'an Shiyou University, Xi'an, China. This work was done when he was a postdoctoral researcher with School of Marine Science and Technology, Northwestern Polytechnical University, Xi'an, China. J. Chen is with School of Marine Science and Technology, Northwestern Polytechnical University, Xi'an, China.}
\stepcounter{footnote}\footnotetext {Matlab code for this paper is available at \url{https://maboudfkaloorazi.com/codes/}}
The SVD (singular value decomposition) \cite{GolubVanLoan96}, CPQR (column-pivoted QR) \cite{Chan87}, \cite{GolubVanLoan96} and p-QLP (pivoted QLP) \cite{StewartQLP} provide a full factorization of an input matrix. The p-QLP is closely associated with CPQR, as it uses CPQR in its computation procedure. These methods are costly in arithmetic operations (though the last two are more efficient) as well as the communication cost, i.e., moving data between the slow and fast memory \cite{DemGGX15,DuerschGu2017, Dongarraetal18, MartinssonQH2019}. Traditional methods to compute a truncated version of the SVD, CPQR and p-QLP, though efficient in arithmetic cost, need to repeatedly access the data. This is a serious bottleneck, as the data access on advanced computational platforms is considerably more expensive compared with arithmetic operations.
Randomized methods \cite{Rokhlin09}, \cite{HMT2011}, \cite{Gu2015}, \cite{ClarWood2017}, \cite{MFKDeTSP18}, \cite{MFKDeJSTSP18}, \cite{Martinsson18}, \cite{Saibaba19}, \cite{MFKJC19}, \cite{MartinssonQH2019} provide an approximation to the foregoing deterministic decompositions. The general strategy of this type of methods is as follows: first, the input matrix is transformed to a lower-dimensional space by means of random sampling. Second, an SVD or CPQR is utilized to process the reduced-size matrix. Finally, the processed matrix is projected back to the original space. Randomized methods are more efficient in arithmetic cost, highly effective, particularly if a low-rank approximation is desired, and harness parallel architectures better in comparison to their classical counterparts.
A bottleneck associated with existing randomized methods, however, is the utilization of the SVD or CPQR; as mentioned, standard methods to compute these factorizations encounter difficulties for parallel implementation, particularly for very large matrices. The goal of this paper is, thus, to devise a randomized matrix factorization method that averts this bottleneck by using only the QR factorization (without pivoting), that, compared with methods that utilize the SVD and CPQR, lends itself much readily to parallel processing.
\subsection{Our Contributions}
\label{subsec_I_Contrib}
In this paper, we devise a matrix factorization algorithm called PbP-QLP (projection-based partial QLP). PbP-QLP makes use of random sampling and the unpivoted QR factorization, and establishes an approximation to the p-QLP, primarily for matrices with low numerical rank. Given ${\bf A} \in \mathbb R^{n_1 \times n_2}$, a large and dense matrix with numerical rank $k$ and $n_1 \ge n_2$, and an integer $k \le d \le n_2$, PbP-QLP generates an approximation
$\hat{\bf A}_\text{PbP-QLP}$ such as:
\begin{equation}
\hat{\bf A}_\text{PbP-QLP}= {\bf QLP}^T,
\label{eq_PbPQLP}
\end{equation}
where ${\bf Q}\in \mathbb R^{n_1 \times d}$ and ${\bf P} \in \mathbb R^{n_2 \times d}$ are orthonormal matrices that constitute approximations to the numerical range of $\bf A$ and ${\bf A}^T$, respectively. ${\bf L}\in \mathbb R^{d \times d}$ is a lower triangular matrix, and its diagonals constitute approximations to the first $d$ singular values of $\bf A$. With theoretical analysis and numerical examples, we show that, analogous to the p-QLP, PbP-QLP is rank-revealing, that is, the rank of $\bf A$ is revealed in the $k \times k$ principal submatrix of $\bf L$. Due to incorporated randomization, as well as using only the unpivoted QR factorization as the decompositional tool, PbP-QLP (i) is efficient in arithmetic operations as its computation incurs the cost $\mathcal{O}(n_1n_2d)$, (ii) makes only a constant number of passes over $\bf A$, and (iii) can harness modern computing environments. We apply PbP-QLP to real data and data matrices from various applications.
\subsection{Notation}
We denote matrices/vectors by bold-face upper-case/lower-case
letters. For any matrix $\bf M$, ${\|{\bf M}\|_2}$ and ${\|{\bf M}\|_F}$ denote the spectral and the Frobenius norm, respectively. The $i$-th largest and the smallest singular value of $\bf M$ are indicated by $\sigma_i(\bf M)$ and $\sigma_\text{min}({\bf M})$, respectively. The notations $\mathcal{R}({\bf M})$ and $\mathcal{N}({\bf M})$ indicate the numerical range and null space of $\bf M$, respectively. $\texttt{orth}({\bf M})$ constructs an orthonormal basis for the range of $\bf M$. $\texttt{qr}({\bf M})$, $\texttt{cpqr}({\bf M})$ and $\texttt{svd}({\bf M})$ give the unpivoted QR, column-pivoted QR and SVD of $\bf M$, respectively. $\texttt{randn}(\cdot)$ generates a matrix whose entries are independent, identically distributed Gaussian random variables of zero mean and variance one. ${\bf I}_r$ indicates an identity matrix of order $r$, and the dagger $\dagger$ indicates the Moore-Penrose inverse.
The remainder of this paper is organized as follows. Section
\ref{section_PriorWork} briefly surveys the related works. Section \ref{section_PbPQLP} describes our
proposed method (PbP-QLP). In Section \ref{section_PbPQLPAnalysis}, we establish a
thorough analysis for PbP-QLP. In Section \ref{sec_NumerExp}, we present and
discuss the results of our numerical experiments, and Section
\ref{section_concl} gives our conclusions.
\section{Related Works}
\label{section_PriorWork}
This section gives a review of related prior works: Section \ref{subsec_II_DeterMeth} presents traditional deterministic matrix factorization methods, while Section \ref{subsec_II_RandMeth} describes more recent methods
based on randomization for low-rank matrix approximations. In this work, we consider a matrix $\bf A$ defined in Section \ref{subsec_I_Contrib}.
\subsection{Deterministic Methods}
\label{subsec_II_DeterMeth}
Let matrix $\bf A$ have a decomposition as follows:
\begin{equation}
{\bf A} = {\bf UTV}^T.
\label{eq_UTV_RelWork}
\end{equation}
\begin{itemize}
\item The Singular Value Decomposition (SVD). If ${\bf U}\in \mathbb R^{n_1 \times n_2}$ and ${\bf V} \in \mathbb R^{n_2 \times n_2}$ have orthonormal columns, and ${\bf T}\in \mathbb R^{n_2 \times n_2}$ is diagonal (with non-negative entries), the factorization in \eqref{eq_UTV_RelWork} is called the SVD \cite{GolubVanLoan96}. To be more
precise, an SVD of ${\bf A}$ is defined as:
\begin{equation}
\begin{aligned}
{\bf A} = {\bf U}{\bf \Sigma V}^T =
\begin{bmatrix} {{\bf U}_k \quad {\bf U}_\perp} \end{bmatrix}
\begin{bmatrix}
{\bf \Sigma}_k & {\bf 0} \\
{\bf 0} & {\bf \Sigma}_\perp
\end{bmatrix}\begin{bmatrix}
{\bf V}_k^T \\
{\bf V}_\perp^T
\end{bmatrix},
\end{aligned}
\label{eq_SVD}
\end{equation}
where the columns $\{{\bf u}_i\}_{i=1}^k$ of ${\bf U}_k \in \mathbb R^{n_1 \times k}$ and $\{{\bf u}_i\}_{i=k+1}^{n_2}$ of ${\bf U}_\perp \in \mathbb R^{n_1 \times (n_2-k)}$ span $\mathcal{R}({\bf A})$ and $\mathcal{N}({\bf A}^T)$, respectively. The diagonal matrix ${\bf \Sigma}$ contains the singular values $\sigma_i$'s in a nonincreasing order: ${\bf \Sigma}_k$ comprises the first $k$ and ${\bf \Sigma}_\perp$ the remaining $n_2-k$ singular values. The columns $\{{\bf v}_i\}_{i=1}^k$ of ${\bf V}_k \in \mathbb R^{n_2 \times k}$ and $\{{\bf v}_i\}_{i=k+1}^{n_2}$ of ${\bf V}_\perp \in \mathbb R^{n_2 \times (n_2-k)}$ span $\mathcal{R}({\bf A}^T)$ and $\mathcal{N}({\bf A})$, respectively. For any integer $1 \le r < n_2$, the SVD establishes the best rank-$r$ approximation ${\bf B}$ to ${\bf A}$:
\begin{equation}\notag
\begin{aligned}
& \underset{\text{rank}({\bf B})\le r}{\text{argmin}}
&& \|{\bf A} - {\bf B}\|_\xi = \|{\bf \Sigma}_{r+1}\|_\xi,
\end{aligned}
\end{equation}
where $\xi \in \{2, F\}$, and $\|{\bf \Sigma}_{r+1}\|_2=\sigma_{r+1}$, and $\|{\bf \Sigma}_{r+1}\|_F=\Big(\sum_{i=r+1}^{n_2}{\sigma_i^2}\Big)^{1/2}$.
\item The Column-Pivoted QR (CPQR). If ${\bf U} \in \mathbb R^{n_1 \times n_2}$ is orthonormal, ${\bf T} \in \mathbb R^{n_2 \times n_2}$ is upper triangular, and ${\bf V} \in \mathbb R^{n_2 \times n_2}$ is a permutation matrix, it is called the
CPQR or rank-revealing QR decomposition \cite{GolubVanLoan96, Chan87, Stewart98}. Columns of ${\bf U}$ span $\mathcal{R}({\bf A})$ and $\mathcal{N}({\bf A}^T)$, and diagonals of ${\bf T}$ constitute approximations to the singular values. Let ${\bf T}$ be partitioned as:
\begin{equation}\notag
{\bf T} =
\begin{bmatrix}
{\bf T}_{11} & {\bf T}_{12} \\
{\bf 0} & {\bf T}_{22}
\end{bmatrix},
\end{equation}
where ${\bf T}_{11} \in \mathbb R^{k \times k}$, ${\bf T}_{12} \in \mathbb R^{k \times (n_2-k)}$, and ${\bf T}_{22} \in \mathbb R^{(n_2-k) \times (n_2-k)}$. The rank-revealing property of CPQR implies that $\sigma_\text{min}({\bf T}_{11})= \mathcal{O}(\sigma_k)$, and $\|{\bf T}_{22}\|_2 = \mathcal{O}(\sigma_{k+1})$.
\item The pivoted QLP (p-QLP). If ${\bf U} \in \mathbb R^{n_1 \times n_2}$ and ${\bf V} \in \mathbb R^{n_2 \times n_2}$ are orthonormal, and ${\bf T} \in \mathbb R^{n_2 \times n_2}$ is lower triangular, it is called the p-QLP \cite{StewartQLP}. The p-QLP is built on the CPQR, and addresses its drawbacks: the CPQR (i) establishes fuzzy approximations to the singular values, and (ii) does not explicitly establish orthogonal bases for $\mathcal{R}({\bf A}^T)$ and $\mathcal{N}({\bf A})$. Let $\bf A$ have a CPQR factorization as follows:
\begin{equation}
{\bf A\Pi}_A = {\bf Q}_A{\bf R}_A.
\label{eq_CPQR_1pQLP}
\end{equation}
The p-QLP is obtained by performing a CPQR on ${\bf R}_A^T$:
\begin{equation}
{\bf R}_A^T\acute{\bf \Pi} = \acute{\bf P}\acute{\bf L}^T.
\label{eq_CPQR_2pQLP}
\end{equation}
Thus, ${\bf A} = {\bf Q}_A\acute{\bf \Pi}\acute{\bf L}\acute{\bf P}^T{\bf \Pi}_A^T$. Orthogonal ${\bf Q}_A\acute{\bf \Pi}$ and ${\bf \Pi}_A\acute{\bf P}$ provide bases for the column space and row space of $\bf A$, respectively. Diagonals of $\acute{\bf L}$ approximate the singular values of $\bf A$. The p-QLP is rank-revealing \cite{StewartQLP, MathiasStewart93, HuckabyChan05}.
\item Complete Orthogonal Decompositions (CODs). If ${\bf U} \in \mathbb R^{n_1 \times n_2}$ and ${\bf V} \in \mathbb R^{n_2 \times n_2}$ are orthonormal, and ${\bf T} \in \mathbb R^{n_2 \times n_2}$ is upper or lower triangular, it is called COD \cite{GolubVanLoan96}. CODs
are also called UTV decompositions \cite{ Stewart98, FierroHan97}. They are rank-revealing in the same sense as CPQR.
\end{itemize}
Computing a full factorization of $\bf A$ by these deterministic methods needs $\mathcal{O}(n_1n_2^2)$ arithmetic operations, which is demanding. However, computing a truncated version of them, say a rank $k$ factorization, which is done by terminating the computation procedure after $k$ steps, incurs the cost $\mathcal{O}(n_1n_2k)$. Though this is regarded as efficient, the significant drawback encountered in computing traditional methods is the data movement, as they require to repeatedly access the date \cite{DemGGX15, DuerschGu2017, Dongarraetal18}. To be more specific, the SVD of a matrix is usually computed using the classical bidiagonalization method in two stages: (i) the matrix is reduced to a bidiagonal form, and (ii) the bidiagonal form is reduced to a diagonal form. In the first stage, a large portion of the operations are performed in terms of level-1 and level-2 BLAS (the Basic Linear Algebra Subprograms) routines. While the method only utilizes level-1 BLAS routines in the second stage \cite{Dongarraetal18}, \cite{MartinssonQH2019}. On the other hand, roughly half of the operations of CPQR and most operations of the unpivoted QR factorization are in level-3 BLAS \cite{DuerschGu2017}, \cite{MartinssonQH2019}. Level-1 and level-2 BLAS routines are memory-bound, and as such can not attain high performance on modern computing platforms. Level-3 BLAS routines, however, are CPU-bound; they can take advantage of the data locality in advanced architectures, and thus attain higher performance, very close to the peak performance rendered by the processors.
\subsection{Randomized Methods}
\label{subsec_II_RandMeth}
Recently, matrix approximation methods based on randomized sampling have gained momentum. These methods
combine the ideas from random matrix theory and classical
methods to approximate a matrix, especially matrices with low numerical rank. The advantages offered by randomized methods are (i) computational efficiency, and (ii) parallelization on modern computers. Using randomized methods to factor a matrix may lead to the loss of accuracy. However, the optimality is not necessary in many practical applications.
The earliest works were based on subset selection or
pseudo-skeleton approximation \cite{GoreinovTZ97, FriezeKVS04}. These methods are referred to as CUR decompositions \cite{MahDri09, DeshpandeV2006, RudelsonV07, Friedland11, BoutsidisWood17}, where matrices $\bf C$ and $\bf R$ are formed by the $k$ (actual) columns and rows (selected according to a probability distribution) of an input matrix, respectively, and matrix $\bf U$ is subsequently obtained via various procedures. The works in \cite{ClarWood2017, Sarlos06, PapaRTV00, NelsonNg2013} were based on the concept of \emph{random projections} \cite{AilonCh09} in which the authors argued that projecting a low-rank matrix onto a random subspace produces a good approximation. This is due to the linear dependency of the matrix's rows. The work presented in \cite{Rokhlin09} made use of random Gaussian matrix to firstly compress the input matrix. The rank-$k$ approximation is subsequently obtained by means of an SVD on the small matrix. In \cite{HMT2011}, Halko, Martinsson, and Tropp devised the R-SVD (randomized SVD) algorithm based on random column sampling that approximates the SVD of a given matrix. R-SVD constructs a factorization to matrix $\bf A$ as described in Algorithm \ref{Alg_RSVD}.
\begin{algorithm}
\caption{Randomized SVD (R-SVD)}
\begin{algorithmic}[1]
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\Require
\Statex $n_1 \times n_2$ matrix ${\bf A}$, and integers $1\le k\le d \le n_2$.
\renewcommand{\algorithmicrequire}{\textbf{Output:}}
\Require
\Statex Orthonormal $\widetilde{\bf U} \in \mathbb R^{n_1 \times d}$ that approximates $\mathcal{R}({\bf A})$, diagonal $\widetilde{\bf \Sigma}\in \mathbb R^{d \times d}$ that approximates the $d$ leading singular values of $\bf A$, orthonormal $\widetilde{\bf V} \in \mathbb R^{n_2 \times d}$ that approximates $\mathcal{R}({\bf A}^T)$ and, accordingly, an approximation constructed as $\hat{\bf A} = \widetilde{\bf U}\widetilde{\bf \Sigma}\widetilde{\bf V}^T$.
\Function{[$\widetilde{\bf U}, \widetilde{\bf \Sigma}, \widetilde{\bf V}$]}{}$=$\texttt{{R\_SVD}}(${\bf A}, d$)
\State ${\bf \Omega}=\texttt{randn}(n_2, d)$
\State ${\bf F} = {\bf A}{\bf \Omega}$
\State $\bar{\bf U} = \texttt{orth}({\bf F})$
\State ${\bf G} = \bar{\bf U}^T{\bf A}$
\State $[\Bar{\Bar{\bf U}}, \widetilde{\bf \Sigma}, \widetilde{\bf V}] = \texttt{svd}({\bf G}) \rightarrow \widetilde{\bf U}\coloneqq \bar{\bf U}\Bar{\Bar{\bf U}}$
\State\hskip-\ALG@thistlm \textbf{end function}
\EndFunction
\end{algorithmic}
\label{Alg_RSVD}
\end{algorithm}
Gu \cite{Gu2015} modified R-SVD, by using a truncated SVD, and presented a new error analysis. He applied the modified R-SVD to improve the subspace iteration methods. The work in \cite{MFKDeTSP18} proposed a \textit{fixed-rank} matrix approximation method,
that is, a rank-$k$ factorization of an input matrix, termed SOR-SVD (subspace-orbit randomized SVD). Through the random column and row sampling, SOR-SVD first reduces the dimension of the input matrix. Then, the matrix is transformed into a lower dimensional space (compared with matrix ambient dimensions). Lastly, a truncated SVD follows to construct an approximate SVD of the given matrix. The work in \cite{MFKDeJSTSP18} presented a rank-revealing algorithm using the randomized scheme termed compressed randomized UTV (CoR-UTV) decomposition. This algorithm is described in Algorithm \ref{Alg_CoRUTV}.
\begin{algorithm}
\caption{Compressed Randomized UTV (CoR-UTV)}\begin{algorithmic}[1]
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\Require
\Statex $n_1 \times n_2$ matrix ${\bf A}$, and integers $1\le k\le d \le n_2$.
\renewcommand{\algorithmicrequire}{\textbf{Output:}}
\Require
\Statex $\widetilde{\bf U} \in \mathbb R^{n_1 \times d}$ that approximates $\mathcal{R}({\bf A})$, upper triangular ${\bf T}\in \mathbb R^{d \times d}$ whose diagonals approximate the $d$ leading singular values of $\bf A$, orthonormal $\widetilde{\bf V} \in \mathbb R^{n_2 \times d}$ that approximates $\mathcal{R}({\bf A}^T)$ and, accordingly, an approximation constructed as $\hat{\bf A} = \widetilde{\bf U}\widetilde{\bf T}\widetilde{\bf V}^T$.
\Function{[$\widetilde{\bf U}, \widetilde{\bf T}, \widetilde{\bf V}$]}{}$=$\texttt{{CoR\_UTV}}(${\bf A}, d$)
\State ${\bf \Omega}=\texttt{randn}(n_2, d)$
\State ${\bf F}_1 = {\bf A}{\bf \Omega}$
\State ${\bf F}_2 = {\bf A}^T{\bf F}_1$
\State $\bar{\bf U} = \texttt{orth}({\bf F}_1)$
\State $\bar{\bf V} = \texttt{orth}({\bf F}_2)$
\State ${\bf G} = \bar{\bf U}^T{\bf A}\bar{\bf V}$ $\rightarrow$ \big(An approximation to $\bf G$, which does\\ \hspace{.4cm} not need a pass over $\bf A$, can be computed as follows: \\
\hspace{.4cm} $\hat{\bf G}= \bar{\bf U}^T{\bf F}_1(\bar{\bf V}^T{\bf F}_2)^\dagger)$.\big)
\State $[\Bar{\Bar{\bf U}}, \widetilde{\bf T}, \Bar{\Bar{\bf \Pi}}] = \texttt{cpqr}({\bf G}) \rightarrow \widetilde{\bf U}\coloneqq \bar{\bf U}\Bar{\Bar{\bf U}}$, and $ \widetilde{\bf V} \coloneqq \bar{\bf V}\Bar{\Bar{\bf \Pi}}$
\State\hskip-\ALG@thistlm \textbf{end function}
\EndFunction
\end{algorithmic}
\label{Alg_CoRUTV}
\end{algorithm}
The work in \cite{XiaoGL17} presented the spectrum-revealing QR factorization (SRQR) algorithm for low-rank matrix approximation. The workhorse of the algorithm is a randomized version of CPQR \cite{DuerschGu2017}. The accuracy of SRQR can match that of CPQR. (For a comparison of CPQR, the SVD and randomized methods in terms of approximation accuracy, see, e.g., \cite{DuerschGu2017}, \cite{MFKDeJSTSP18}.) The authors in \cite{FengXG19} presented the Flip-Flop SRQR factorization algorithm. The algorithm first applies SSQR to compute a partial CPQR of the given matrix. Next, an LQ factorization is performed on the R-factor obtained. The final approximation is then given through an SVD of the transformed matrix. The authors made use of Flip-Flop SRQR for robust principal component analysis \cite{MFKJC19} and tensor decomposition applications. The work in \cite{WuX20} presented randomized QLP decomposition algorithm; the authors have replaced the SVD in R-SVD \cite{HMT2011} by p-QLP \cite{StewartQLP} and presented error bounds (in expected value) for the approximate leading singular values of the matrix.
The current randomized algorithms make use of the SVD or CPQR algorithms in order to factor the transformed (or reduced) input matrix. When dealing with large matrices, the drawback of such deterministic factorization algorithms, as mentioned earlier, is that standard methods to compute them can not efficiently exploit modern parallel architectures, owning to the utilization of level-1 and level-2 BLAS routines \cite{DemGGX15, DuerschGu2017, Dongarraetal18}. In this paper, we develop a randomized matrix factorization algorithm that uses only the unpivoted QR factorization. As most operations of this factorization are in level-3 BLAS, the proposed algorithm thereby weeds out the drawback associated with existing randomized algorithms and in turn lends itself much easily to parallel implementation.
\section{Projection-based Partial QLP (PbP-QLP)}
\label{section_PbPQLP}
We describe in this section our proposed PbP-QLP algorithm. It approximates the p-QLP \cite{StewartQLP}, providing a factorization in the form of \eqref{eq_PbPQLP}. PbP-QLP is rank-revealing, and its development has been motivated by the p-QLP’s procedure. However, in contrast to the p-QLP, PbP-QLP (i) uses randomization, and (ii) does not use CPQR; it only employs the unpivoted QR as the decompositional apparatus. In addition to the basic form of PbP-QLP described in Section \ref{subsec_SimplePbP}, we present in Section \ref{subsec_PIPbPQLP} a variant of the algorithm that utilizes the power iteration scheme that enhances its accuracy and robustness.
\subsection{Computation of PbP-QLP}
\label{subsec_SimplePbP}
Given matrix ${\bf A}$ and an integer $1\le k\le d\le n_2$, we generate in the first step of computing PbP-QLP an $n_1 \times d$ standard Gaussian matrix ${\bf \Phi}=\texttt{randn}(n_1, d)$. We then transform ${\bf A}$ to a low dimensional space using ${\bf \Phi}$:
\begin{equation}
{\bf C} = {\bf A}^T{\bf \Phi}.
\label{eq_c}
\end{equation}
We next construct an orthonormal basis $\bar{\bf P}\in \mathbb R^{n_2 \times d}$ for $\mathcal{R}({\bf C})$:
\begin{equation}
\bar{\bf P} = \texttt{orth}({\bf C}).
\label{eq_Pbar}
\end{equation}
This step can efficiently be performed by a call to a packaged
QR decomposition. Then, we form a matrix ${\bf D}\in \mathbb R^{n_1 \times d}$ by right-multiplying ${\bf A}$ with $\bar{\bf P}$:
\begin{equation}
{\bf D} = {\bf A}\bar{\bf P}.
\label{eq_D}
\end{equation}
Afterwards, we use the unpivoted QR to factor $\bf D$:
\begin{equation}
{\bf D}= {\bf Q}{\bf R}.
\label{eq_DQR}
\end{equation}
We now carry out another QR factorization on ${\bf R}^T\in \mathbb R^{d \times d}$:
\begin{equation}
{\bf R}^T = \widetilde{\bf P}\widetilde{\bf R}.
\label{eq_RTPtRt}
\end{equation}
Lastly, we construct an approximation $\hat{\bf A}_\text{PbP-QLP}$ to $\bf A$:
\begin{equation}
\hat{\bf A}_\text{PbP-QLP} = {\bf Q}\widetilde{\bf R}^T \widetilde{\bf P}^T \bar{\bf P}^T \coloneqq {\bf QLP}^T,
\label{eq_Ahat_simplePbP}
\end{equation}
where ${\bf L}\coloneqq \widetilde{\bf R}^T$, and ${\bf P} \coloneqq \bar{\bf P}\widetilde{\bf P}$. Orthonormal matrices ${\bf Q}\in \mathbb R^{n_1 \times d}$ and ${\bf P}\in \mathbb R^{n_2 \times d}$ approximate $\mathcal{R}({\bf A})$ and $\mathcal{R}({\bf A}^T)$, respectively. ${\bf L}\in \mathbb R^{d \times d}$ is lower triangular, where its diagonals approximate the first $d$ singular values of $\bf A$, and its $k\times k$ leading block reveals the numerical rank $k$ of $\bf A$.
\noindent \textbf{Justification and Relation to p-QLP.} As expounded in Section \ref{subsec_II_DeterMeth}, the p-QLP for $\bf A$ involves two steps, each performing one CPQR factorization (equations \eqref{eq_CPQR_1pQLP} and \eqref{eq_CPQR_2pQLP}). Stewart \cite{StewartQLP} argues that in the second step of the computation, column pivoting is not necessary, due to pivoting done in the first step. To relate it to PbP-QLP, equations \eqref{eq_c}-\eqref{eq_DQR} emulate (up to $d$ columns) the first step of p-QLP. Since $\bar{\bf P}$ \eqref{eq_Pbar} approximates $\mathcal{R}({\bf A}^T)$ and $\bf D$ is formed as in \eqref{eq_D}, matrix $\bf R$ \eqref{eq_DQR} reveals the numerical rank of $\bf A$. As such, we proceed with an unpivoted QR on ${\bf R}^T$ for the second step in PbP-QLP’s computation. We will show matrix $\bf R$ reveals the gap in $\bf A$’s spectrum through theoretical analysis and simulations.
\subsection{Power Iteration-coupled PbP-QLP}
\label{subsec_PIPbPQLP}
PbP-QLP provides highly accurate approximations to the
SVD of matrices with rapidly spectral decay. However, it
may produce less accurate approximations for matrices whose singular values decay slowly. To enhance the accuracy of approximations, we thus utilize the power iteration (PI) scheme \cite{Rokhlin09, HMT2011, Gu2015, VoroninMartin16, YuLi18}. The reasoning behind using the PI scheme is to replace the input matrix with an intimately related matrix whose singular values decay more rapidly. In particular, ${\bf A}^T$ in \eqref{eq_c} is substituted by ${\bf A}_\text{PI}$ defined as:
\begin{equation}\notag
{\bf A}_\text{PI} = ({\bf A}^T{\bf A})^q{\bf A}^T,
\end{equation}
where $q$ is the PI factor. It is seen that ${\bf A}^T$ and ${\bf A}_\text{PI}$ have the same singular vectors, whereas the latter has singular values $\sigma_i^{2q+1}$. It should be noted that the performance improvement brought by the PI technique incurs an extra cost, as PbP-QLP needs more arithmetic operations and also more passes through $\bf A$, if $\bf A$ is stored externally.
The computation of ${\bf A}_\text{PI}$ in floating point arithmetic is prone to round-off errors. To put it precisely, let $\epsilon_\text{machine}$ be the machine precision. Then, any singular component less than $\sigma_1 \epsilon_\text{machine}^{1/(2q+1)}$ will be lost. To compensate this loss of accuracy, an orthonormalization of the sample matrix between each application of ${\bf A}^T$ and ${\bf A}$ is needed \cite{HMT2011,Gu2015}. The resulting PbP-QLP algorithm is presented in Algorithm \ref{Alg_PbPQLP_PI}.
\begin{algorithm}
\caption{Power Iteration-coupled PbP-QLP}
\begin{algorithmic}[1]
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\Require
\Statex $n_1 \times n_2$ matrix ${\bf A}$, and integers $1\le k\le d \le n_2$, and $q\ge 1$.
\renewcommand{\algorithmicrequire}{\textbf{Output:}}
\Require
\Statex Orthonormal ${\bf Q} \in \mathbb R^{n_1 \times d}$ that approximates $\mathcal{R}({\bf A})$, lower triangular ${\bf L}\in \mathbb R^{d \times d}$ whose diagonals approximate the $d$ leading singular values of $\bf A$, orthonormal ${\bf P} \in \mathbb R^{n_2 \times d}$ that approximates $\mathcal{R}({\bf A}^T)$ and, accordingly, an approximation constructed as $\hat{\bf A} = {\bf Q}{\bf L}{\bf P}^T$.
\Function{[${\bf Q}, {\bf L}, {\bf P}$]}{}$=$\texttt{{PbP\_QLP}}(${\bf A}, d$)
\State ${\bf \Phi}=\texttt{randn}(n_1, d)$
\State ${\bf C} = {\bf A}^T{\bf \Phi}$
\State $\bar{\bf P} = \texttt{orth}({\bf C})$
\State \textbf{for} $i=1:q$
\State \quad ${\bf C} = {\bf A}\bar{\bf P}$; $\bar{\bf P} = \texttt{orth}({\bf C})$
\State \quad ${\bf C} = {\bf A}^T\bar{\bf P}$; $\bar{\bf P} = \texttt{orth}({\bf C})$
\State \textbf{end for}
\State ${\bf D} = {\bf A}\bar{\bf P}$
\State $[{\bf Q}, {\bf R}] = \texttt{qr}({\bf D})$
\State $[\widetilde{\bf P}, \widetilde{\bf R}] = \texttt{qr}({\bf R}^T) \rightarrow {\bf P}\coloneqq \bar{\bf P}\widetilde{\bf P}, {\bf L}\coloneqq \widetilde{\bf R}^T$
\State\hskip-\ALG@thistlm \textbf{end function}
\EndFunction
\end{algorithmic}
\label{Alg_PbPQLP_PI}
\end{algorithm}
\section{Theoretical Analysis of PbP-QLP}
\label{section_PbPQLPAnalysis}
This section provides a detailed account of theoretical
analysis for the PbP-QLP algorithm.
\subsection{Rank of $\bf A$ is revealed in $\bf R$}
Here, we show that matrix $\bf R$ generated by the PbP-QLP
algorithm (equation \eqref{eq_DQR}, or Step 10 of Algorithm \ref{Alg_PbPQLP_PI}) reveals the numerical rank of $\bf A$. To be precise, let $\bf A$ have an SVD as defined in \eqref{eq_SVD}, and the matrix $\bf R$ be partitioned as:
\begin{equation}
{\bf R} =
\begin{bmatrix}
{\bf R}_{11} & {\bf R}_{12} \\
{\bf 0} & {\bf R}_{22}
\end{bmatrix},
\label{eq_Rparti}
\end{equation}
where ${\bf R}_{11} \in \mathbb R^{k \times k}$, ${\bf R}_{12} \in \mathbb R^{k \times (d-k)}$, and ${\bf R}_{22} \in \mathbb R^{(d-k) \times (d-k)}$. We call the diagonals of $\bf R$, R-values. The definition of rank-revealing factorizations in the literature \cite{Chan87, Stewart98, ChandrasekaranIpsen94} implies that
\begin{equation} \label{eq_sigmak_R11}
\sigma_\text{min}({\bf R}_{11})= \mathcal{O}(\sigma_k).
\end{equation}
\begin{equation} \label{eq_2normR22}
\|{\bf R}_{22}\|_2 = \mathcal{O}(\sigma_{k+1}).
\end{equation}
We prove \eqref{eq_2normR22} first, and then \eqref{eq_sigmak_R11}.
\begin{theorem}
Let ${\bf A} \in \mathbb R^{n_1\times n_2}$ be an input matrix with an SVD defined in \eqref{eq_SVD}, $\bf R$ be generated by PbP-QLP partitioned as in \eqref{eq_Rparti}, integer $p\ge 0$, and $k+p\le d$. Defining
\begin{equation}\notag
\begin{bmatrix}
{\bf \Phi}_1 \\
{\bf \Phi}_2
\end{bmatrix} \coloneqq {\bf U}^T{\bf \Phi},
\label{eqG12_ThrLRA}
\end{equation}
where ${\bf \Phi}_1 \in \mathbb R^{(d-p)\times d}$ and ${\bf \Phi}_2 \in \mathbb R^{(n_2-d+p)\times d}$, we then have
\begin{equation}\label{eq_2nR22_bound}
\begin{aligned}
\|{\bf R}_{22}\|_2 & \le \sigma_{k+1} + \frac{\delta_k^{2q+2}\sigma_1\|{\bf \Phi}_2\|_2\|{\bf \Phi}_1^\dagger\|_2}{\sqrt{1 + \delta_k^{4q+4}\|{\bf \Phi}_2\|_2^2\|{\bf \Phi}_1^\dagger\|_2^2}},
\end{aligned}
\end{equation}
where $\delta_k = \frac{\sigma_{d-p+1}}{\sigma_k}$.
\label{Th_R22bound}
\end{theorem}
\textit{Proof.} The proof is given in Appendix \ref{Proo_R22bound}.
\begin{remark}
The relation \eqref{eq_2nR22_bound} implies that one of the conditions through which the rank of $\bf A$ is revealed by matrix $\bf R$ depends on the ratio $\frac{\sigma_{d-p+1}}{\sigma_k}$. Provided that (i) there exists a large gap
in the singular values of $\bf A$ ($\sigma_k \gg \sigma_{k+1}$), and (ii) $\sigma_k$ is not substantially smaller than $\sigma_1$, for any $p \ge 0$ the trailing block of $\bf R$ is sufficiently
small in magnitude. In addition, utilizing the PI technique
exponentially derives down the quotient in \eqref{eq_2nR22_bound} to zero. It is therefore expected that PI considerably improves the accuracy of PbP-QLP in forming a rank-revealing factorization.
\end{remark}
The following theorem bounds the minimum singular value of the leading block of $\bf R$.
\begin{theorem}\label{Th_boundssigmak}
Under the notation and hypotheses of Theorem \ref{Th_R22bound}, we have
\begin{equation}\label{eq_Th_boundssigmak}
\sigma_{k+1} \le \sigma_k({\bf R}_{11}) \le \sigma_k + \sigma_{k+1}
\end{equation}
\end{theorem}
\textit{Proof.} The proof is given in Appendix \ref{Proo_boundssigmak}.
\begin{remark}
The bounds presented in this theorem together with that of Theorem 1 assert that if there is a substantial gap in the singular values of $\bf A$, and $\sigma_1$ is not substantially larger than $\sigma_k$, the rank of $\bf A$ is revealed in $\bf R$.
\end{remark}
\subsection{Approximate Subspaces and Low-Rank Approximation}
The closeness of any two subspaces, say ${\bf S}_1$ and ${\bf S}_2$, is measured by the \textit{distance} between them \cite{GolubVanLoan96}. Denoted by $\text{dist}({\bf S}_1, {\bf S}_2)$, it corresponds to the largest canonical angle between the two subspaces. The following theorem establishes the distances between ${\bf U}_k$ and ${\bf V}_k$ (the principal left and right singular vectors of $\bf A$, respectively) and their corresponding approximations ${\bf Q}_1$ and ${\bf P}_1$ by PbP-QLP, formed by the first $k$ columns of ${\bf Q}$ and ${\bf P}$.
\begin{theorem} \label{Th_distUkVkQ1P1}
Let ${\bf A} \in \mathbb R^{n_1\times n_2}$ be an input matrix with an SVD defined in \eqref{eq_SVD}, and ${\bf Q}$ and ${\bf P}$ be generated by PbP-QLP. Let further
\begin{equation}\notag
\begin{aligned}
\begin{bmatrix}
{\bf \Phi}_{11} & {\bf \Phi}_{12} \\
{\bf \Phi}_{21} & {\bf \Phi}_{22}
\end{bmatrix} \coloneqq {\bf U}^T{\bf \Phi},
\end{aligned}
\end{equation}
where ${\bf \Phi}_{11} \in \mathbb R^{k \times k}$, ${\bf \Phi}_{21} \in \mathbb R^{(n_2-k) \times k}$, ${\bf \Phi}_{12} \in \mathbb R^{k \times (d-k)}$ and ${\bf \Phi}_{22} \in \mathbb R^{(n_2-k) \times (d-k)}$. Let ${\bf \Phi}_{11}$ be full rank, and ${\bf Q}_1$ and ${\bf P}_1$ be formed by the first $k$ columns of ${\bf Q}$ and ${\bf P}$. Then
\begin{equation}\label{eq_Th_distance1}
\text{dist}(\mathcal{R}({\bf U}_k), \mathcal{R}({\bf Q}_1)) \le \Big(\frac{\sigma_{k+1}}{\sigma_k}\Big)^{2q+2}\|{\bf \Phi}_{21}{\bf \Phi}_{11}^{-1}\|_2.
\end{equation}
\begin{equation}\label{eq_Th_distance2}
\text{dist}(\mathcal{R}({\bf V}_k), \mathcal{R}({\bf P}_1)) \le \Big(\frac{\sigma_{k+1}}{\sigma_k}\Big)^{2q+1}\|{\bf \Phi}_{21}{\bf \Phi}_{11}^{-1}\|_2.
\end{equation}
\end{theorem}
\textit{Proof.} The proof is provided in Appendix \ref{Proo_distance}.
\begin{remark} \label{Remark_Basis}
Provided ${\sigma_k}>{\sigma_{k+1}}$, Theorem \ref{Th_distUkVkQ1P1} shows that the subspaces $\mathcal{R}({\bf Q}_1)$ and $\mathcal{R}({\bf P}_1)$ converge respectively to $\mathcal{R}({\bf U}_k)$ and $\mathcal{R}({\bf V}_k)$ at a rate proportional to $\Big(\frac{\sigma_{k+1}}{\sigma_k}\Big)^q$. The convergence rate also depends on the distribution of the random matrix employed. The virtue of powering the input matrix laid out by Theorem \ref{Th_distUkVkQ1P1} can be seen in the accuracy of PI-coupled PbP-QLP: for large enough $q$, ${\bf Q}_1^T{\bf U}_\perp \approx 0$, which vanishes the first term in \eqref{eq_2n2ndTermR22}, as well as the second term in the right-hand side of the last relation in \eqref{eq_R11_exp}. This, simply, implies that with an appropriately chosen $q$, the rank of the input matrix is revealed in submatrix $\bf R$ (Step 6 of Alg. \ref{Alg_PbPQLP_PI}), regardless of the gap in the spectrum being substantial or not.
\end{remark}
\begin{theorem}\label{Th_LRapprox}
Under the notation of Theorem \ref{Th_R22bound}, let further $\bf Q$ be constructed by PbP-QLP. Then
\begin{equation}\label{eq_2nLRAErrbound}
\begin{aligned}
\|({\bf I} - {\bf QQ}^T){\bf A}\|_2 & \le \|{\bf \Sigma}_\perp\|_2 + \frac{\delta_k^{2q+2}\sigma_1\|{\bf \Phi}_2\|_2\|{\bf \Phi}_1^\dagger\|_2}{\sqrt{1 + \delta_k^{4q+4}\|{\bf \Phi}_2\|_2^2\|{\bf \Phi}_1^\dagger\|_2^2}}.
\end{aligned}
\end{equation}
\end{theorem}
\textit{Proof.} The proof is provided in Appendix \ref{Proo_LRapprox}.
\begin{remark}
This theorem shows the PI scheme considerably enhances
the approximation quality of PbP-QLP: the accuracy of
a low-rank approximation generated by PbP-QLP coupled with PI depends on the ratio $\frac{\sigma_{d-p+1}}{\sigma_k}$. Provided that $\sigma_k > \sigma_{k+1}$, for any $p \ge 0$, the extra quotient in the bound \eqref{eq_2nLRAErrbound} is driven to zero exponentially fast.
\end{remark}
\subsection{Estimated Singular Values and Rank-Revealing Property of PbP-QLP}
Here, we establish the relation between the first $k$ estimated singular values of $\bf A$ computed by PbP-QLP, i.e., diagonals of $\bf L$, and the first $k$ singular values given by the SVD.
\begin{theorem}
Under the notation of Theorem \ref{Th_R22bound}, let further $\bf D$ be generated by PbP-QLP. Then, for $i = 1,..., k$, we have
\begin{equation}\label{eq_Th_SinValD_bound}
\sigma_i \ge \sigma_i({\bf D}) \ge \frac{\sigma_i}{\sqrt{1 + \delta_i^{4q+4}\|{\bf \Phi}_2\|_2^2\|{\bf \Phi}_1^\dagger\|_2^2}},
\end{equation}
where $\delta_i= \frac{\sigma_{d-p+1}}{\sigma_i}$.
\label{Theorem_SinValueApprox}
\end{theorem}
\textit{Proof.} The proof is given in Appendix \ref{Proo_SinValueApprox1}.
\begin{theorem} \label{Th_SinValueApprox2}
Under the notation of Theorem \ref{Th_R22bound}, let further $\bf D$ and ${\bf L}$ be generated by PbP-QLP, and ${\bf L}$ be partitioned as:
\begin{equation} \label{eq_Lparti}
\begin{aligned}
{\bf L} =\begin{bmatrix}
{\bf L}_{11} & {\bf 0} \\
{\bf L}_{21} & {\bf L}_{22}
\end{bmatrix},
\end{aligned}
\end{equation}
where ${\bf L}_{11} \in \mathbb R^{k \times k}$, ${\bf L}_{21} \in \mathbb R^{(d-k) \times k}$, and ${\bf L}_{22} \in \mathbb R^{(d-k) \times (d-k)}$. Then, for $i=1,..., k$, we have
\begin{equation}\label{eq_sigmaL11sigmai}
\begin{aligned}
\frac{{\sigma_i({\bf L}_{11})}}{{\sigma_i}} \ge
\frac{\Bigg[1 - \mathcal{O}\Bigg(\frac{\|{\bf L}_{12}\|_2^2}{(1-\rho^2)\sigma_k^2({\bf L}_{11})}\Bigg)\Bigg]}{\sqrt{1 + \delta_i^{4q+4}\|{\bf \Phi}_2\|_2^2\|{\bf \Phi}_1^\dagger\|_2^2}}.
\end{aligned}
\end{equation}
where $\rho = \frac{\|{\bf L}_{22}\|_2}{\sigma_k({\bf L}_{11})}$.
\end{theorem}
\textit{Proof.} The proof is given in Appendix \ref{Proo_SinValueApprox2}.
We have thus far shown matrix $\bf R$ generated by PbP-QLP reveals the numerical rank of $\bf A$. The rank-revealing property of PbP-QLP is guaranteed by Theorem 2.1 of \cite{MathiasStewart93}: let $\bf R$ and $\bf L$ be constructed by PbP-QLP and partitioned as in \eqref{eq_Rparti} and \eqref{eq_Lparti}, respectively. Then, we have
\begin{equation}\notag
\sigma_\text{min}({\bf L}_{11}) \ge \sigma_\text{min}({\bf R}_{11}),
\end{equation}
\begin{equation}\notag
\sigma_1({\bf L}_{22}) \le \sigma_1({\bf R}_{22}).
\end{equation}
\begin{theorem}\label{Th_boundsA'ssigmak}
Under the notation of Theorem \ref{Th_distUkVkQ1P1}, let further $\bf L$ be constructed by PbP-QLP and partitioned as in \eqref{eq_Lparti}. Then
\begin{equation}\notag
\sigma_{\text{min}}({\bf L}_{11}) \le \sigma_k \le \sigma_{\text{min}}({\bf L}_{11}) + \sigma_1\Big(\frac{\sigma_{k+1}}{\sigma_k}\Big)^{2q+2}\|{\bf \Phi}_{21}{\bf \Phi}_{11}^{-1}\|_2.
\end{equation}
\end{theorem}
\textit{Proof.} The proof is provided in Appendix \ref{Proo_boundsA'ssigmak}.
\subsection{High-Probability Error Bounds}
The following theorem provides high-probability bounds
which give further insights into the accuracy of PbP-QLP.
\begin{theorem}\label{Th_SinVal_PbP}
Let ${\bf A} \in \mathbb R^{n_1\times n_2}$ be an input matrix with an SVD defined in \eqref{eq_SVD}, $0\le p\le d-k$, and $\bf R$ and $\hat{\bf A}_\text{PbP-QLP}$ be generated by PbP-QLP. Let further $0<\Upsilon \ll 1$, and define
\begin{equation}\label{key}
C_\Upsilon = \frac{e\sqrt{d}}{p+1}\Big(\frac{2}{\Upsilon}\Big)^\frac{1}{p+1}\Big(\sqrt{n-d+p}+\sqrt{d} + \sqrt{2\text{log}\frac{2}{\Upsilon}}\Big).
\end{equation}
Then with probability not less than $1 - \Upsilon$, we have
\begin{equation}\label{eq_HighProB_R22}
\sigma_1({\bf R}_{22}) \le \sigma_k +\delta_k^{2q+2}\sigma_1C_\Upsilon,
\end{equation}
\begin{equation}\label{eq_HighProB_SVs}
\begin{aligned}
\frac{{\sigma_i(\hat{\bf A}_\text{PbP-QLP})}}{{\sigma_i}} \ge
\frac{\Bigg[1 - \mathcal{O}\Bigg(\frac{\|{\bf L}_{12}\|_2^2}{(1-\rho^2)\sigma_k^2({\bf L}_{11})}\Bigg)\Bigg]}{\sqrt{1 + \delta_i^{4q+4}C_\Upsilon^2}},
\end{aligned}
\end{equation}
and
\begin{equation}\label{eq_HighProB_LRAE}
\begin{aligned}
\|({\bf I} - {\bf QQ}^T){\bf A}\|_2 & \le \|{\bf \Sigma}_\perp\|_2 + \delta_k^{2q+2}\sigma_1C_\Upsilon.
\end{aligned}
\end{equation}
\end{theorem}
\textit{Proof.} The proof is given in Appendix \ref{Proo_SinVal_PbP}.
For the distance between approximate subspaces $\mathcal{R}({\bf Q}_1)$ and $\mathcal{R}({\bf P}_1)$ and the corresponding singular subspaces, we have with high probability
\begin{equation}\notag
\text{dist}(\mathcal{R}({\bf U}_k), \mathcal{R}({\bf Q}_1)) \le c\sqrt{k(n_1-k)}\Big(\frac{\sigma_{k+1}}{\sigma_k}\Big)^{2q+2}.
\end{equation}
\begin{equation}\notag
\text{dist}(\mathcal{R}({\bf V}_k), \mathcal{R}({\bf P}_1)) \le c\sqrt{k(n_1-k)}\Big(\frac{\sigma_{k+1}}{\sigma_k}\Big)^{2q+1}.
\end{equation}
Here $c = c_1c_2$, where $c_1$ and $c_2$ are positive absolute constants. These results follow by applying the bounds for standard Gaussian matrices ${\bf \Phi}_{11}$ and ${\bf \Phi}_{21}$ devised in \cite{Edelman88, Szarek91}:
\begin{equation}\notag
\|{\bf \Phi}_{11}^{-1}\|_2 \le c_1\sqrt{k}, \text{and} \quad \|{\bf \Phi}_{21}\|_2 \le c_2\sqrt{n_1 - k}. \end{equation}
\subsection{Computational Cost}
\label{subsec_computcost}
Computation of the simple form of PbP-QLP for matrix $\bf A$ needs the following arithmetic operations:
\begin{itemize}
\item Generating a matrix $\bf \Phi$ with Gaussian random variables costs $\mathcal{O}(n_1d)$.
\item Forming matrix $\bf C$ in \eqref{eq_c} costs $\mathcal{O}(n_1n_2d)$.
\item Constructing the basis $\bar{\bf P}$ in \eqref{eq_Pbar} costs $\mathcal{O}(n_2d^2)$.
\item Forming matrix $\bf D$ in \eqref{eq_D} costs $\mathcal{O}(n_1n_2d)$.
\item Constructing matrices ${\bf Q}$ and ${\bf R}$ in \eqref{eq_DQR} costs $\mathcal{O}(n_1d^2)$.
\item Constructing matrices $\widetilde{\bf P}$ and $\widetilde{\bf R}$ in \eqref{eq_RTPtRt} costs $\mathcal{O}(d^3)$.
\item Forming matrix $\bf P$ costs $\mathcal{O}(n_2d^2)$.
\end{itemize}
The above cost is dominated by multiplications of $\bf A$ and ${\bf A}^T$ by associated matrices. Thus,
\begin{equation}\notag
\mathcal{C}_\text{PbP-QLP} = \mathcal{O}(n_1n_2d).
\label{eq_CostBasicRPTSOD}
\end{equation}
The computation of PbP-QLP is more efficient provided that the matrix $\bf A$ is sparse. In this case, the flop count is proportional to the number of non-zero entries $z$ of $\bf A$ and satisfies $\mathcal{O}(zd)$. Considering $\bf A$ to be \emph{large}, that is, $\bf A$ is stored out-of-memory, the simple form of PbP-QLP makes two passes over $\bf A$ to construct an approximation. However, if PbP-QLP is coupled with the PI scheme, it needs $2q+2$ passes over $\bf A$ and, as a result, its flop count satisfies $(q+1)\mathcal{C}_\text{PbP-QLP}$.
In addition to the arithmetic cost, another cost imposed on any algorithm is the communication cost \cite{DemGGX15, DuerschGu2017}. This cost, which constitutes the primary factor for traditional methods to be unsuitable on high performance computing devices, is associated with the exploitation of level-1, 2 and 3 BLAS routines: it is determined by moving data between processors working in parallel, and also data movement between different levels of the memory hierarchy. On modern devices, the communication cost considerably dominates the cost of an algorithm. This has motivated researchers to incorporate randomized sampling paradigm into matrix factorization algorithms \cite{DemGGX15, DuerschGu2017, MartinssonQH2019}, where the share of level-3 BLAS operations is higher. In addition, among classical methods applied to the reduced matrices by randomized methods, the unpivoted QR factorization algorithm needs the least data access (the least communication cost), as the large majority of its operations are in terms of level-3 BLAS. As PbP-QLP utilizes only the QR factorization to construct the approximation, its operations can be cast almost completely in terms of level-3 BLAS operations. Consequently, it can be executed more efficiently on highly parallel machines.
\section{Nunerical Simulations}
\label{sec_NumerExp}
This section reports our numerical results obtained by conducting a set of experiments with randomly generated data as well as real-world data. They show the performance behavior of the PbP-QLP algorithm in comparison with several existing algorithms. The experiments are run in MATLAB on a 4-core Intel Core i7 CPU running at 1.8 GHz and 8GB RAM.
\subsection{Runtime Comparison}
We compare the speed of proposed PbP-QLP against two most representative randomized algorithms, namely R-SVD \cite{HMT2011} (Algorithm \ref{Alg_RSVD}) and CoR-UTV \cite{MFKDeJSTSP18} (Algorithm \ref{Alg_CoRUTV}), in factoring input matrices with various dimensions. We generate square, dense matrices of order $n$, and consider three cases for the sampling size parameter $d$, specifically $d=0.04n$, $d=0.2n$ and $d=0.3n$. Note that for the runtime comparison, the distribution of singular values of matrices is immaterial. In this experiment, we have excluded deterministic algorithms for two reasons: firstly, they are considerably slower than their randomized counterparts due to computational cost. Secondly, there are no optimized MATLAB implementations for a truncated version of these algorithms. This, accordingly, enables us to clearly display the behavior of considered algorithms. The results for the three algorithms with and without the power iteration technique are reported in Tables \ref{Table1RuntimePbPQLP}-\ref{Table3RuntimePbPQLP}. For each runtime measurement, the result is averaged over 10 independent runs. It is seen that for the first scenario, where $d=0.04n$ (Table \ref{Table1RuntimePbPQLP}), PbP-QLP shows similar performance as R-SVD and CoR-UTV in most cases. However, as the dimension of the input matrix grows, PbP-QLP begins to edge out. For the second and third scenarios (Tables \ref{Table2RuntimePbPQLP} and \ref{Table3RuntimePbPQLP}), we observe that PbP-QLP outperforms the other two algorithms in all cases. Moreover, for larger matrices we observe larger discrepancies in execution time. This is because, as mentioned earlier, PbP-QLP only makes use of the unpivoted QR factorization (to factor the reduced-size matrix) whose vast majority of operations (contrary to the SVD and CPQR) are in level-3 BLAS operations. These results simply show that the ``quality" of operations matters, and factorization algorithms that incorporate more BLAS-3 routines consume less time. As such, if implemented on a highly parallel machine, PbP-QLP would show yet better results compared to R-SVD and CoR-UTV. (Compared to R-SVD and PbP-QLP, CoR-UTV needs more arithemtic operations, but almost all of them are in level-3 BLAS. Thus, we expect it shows better performance if implemented in parallel.)
\begin{table}[!htb]
\caption{Computational time (in seconds) for different randomized algorithms. $d$ is the sampling size parameter, and $q$ is the power iteration factor.}
\begin{tabular}{p{1.2cm} p{0.5cm} p{2.5cm} p{2.5cm}}
\noindent\rule{8.1cm}{0.4pt}\\
Algorithms & &
\begin{tabular}{p{0.65cm} p{0.65cm} p{0.65cm} p{0.65cm} p{0.65cm}}
\multicolumn{5}{c}{$n$ $(d=0.04n)$} \\
\hline
5000 & 10000 & 15000 & 20000 & 25000 \\
\end{tabular} \\
\noindent\rule{8.1cm}{0.4pt}\\
R-SVD &
\begin{tabular}{|p{0.5cm} p{.8cm} p{0.65cm} p{0.6cm} p{0.65cm} p{0.5cm}}
$q$=0 & 0.29 & 2.3 & 7.2 & 16.9 & 35.1 \\
$q$=1 & 0.51 & 4.1 & 13.0 & 29.7 & 59.5 \\
$q$=2 & 0.74 & 5.9 & 18.6 & 42.8 & 84.1 \\
\end{tabular} \\ \\
CoR-UTV &
\begin{tabular}{|p{0.5cm} p{.8cm} p{0.65cm} p{0.6cm} p{0.65cm} p{0.5cm}}
$q$=0 & 0.32 & 2.3 & 7.5 & 17.0 & 34.4 \\
$q$=1 & 0.58 & 4.2 & 13.2 & 29.6 & 59.3 \\
$q$=2 & 0.80 & 6.0 & 18.8 & 42.7 & 83.7 \\
\end{tabular} \\\\
PbP-QLP &
\begin{tabular}{|p{0.5cm} p{.8cm} p{0.65cm} p{0.6cm} p{0.65cm} p{0.5cm}}
$q$=0 & 0.29 & 2.2 & 7.1 & 16.0 & 31.6 \\
$q$=1 & 0.53 & 4.1 & 12.7 & 28.7 & 56.6 \\
$q$=2 & 0.81 & 5.9 & 18.6 & 41.4 & 81.7 \\
\end{tabular} \\
\noindent\rule{8.1cm}{0.4pt}\\
\end{tabular}
\label{Table1RuntimePbPQLP}
\end{table}
\begin{table}[!htb]
\caption{Computational time (in seconds) for different randomized algorithms. $d$ is the sampling size parameter, and $q$ is the power iteration factor.}
\begin{tabular}{p{1.2cm} p{0.5cm} p{2.5cm} p{2.5cm}}
\noindent\rule{8.1cm}{0.4pt}\\
Algorithms & &
\begin{tabular}{p{0.65cm} p{0.65cm} p{0.65cm} p{0.65cm} p{0.65cm}}
\multicolumn{5}{c}{$n$ $(d=0.2n)$} \\
\hline
5000 & 10000 & 15000 & 20000 & 25000 \\
\end{tabular} \\
\noindent\rule{8.1cm}{0.4pt}\\
R-SVD &
\begin{tabular}{|p{0.5cm} p{.75cm} p{0.65cm} p{0.6cm} p{0.65cm} p{0.5cm}}
$q$=0 & 2.4 & 20.2 & 65.7 & 195 & 429 \\
$q$=1 & 3.6 & 29.1 & 94.8 & 280 & 613 \\
$q$=2 & 4.7 & 37.9 & 124 & 364 & 782 \\
\end{tabular} \\ \\
CoR-UTV &
\begin{tabular}{|p{0.5cm} p{.75cm} p{0.65cm} p{0.6cm} p{0.65cm} p{0.5cm}}
$q$=0 & 2.7 & 22.7 & 74.4 & 236 & 469 \\
$q$=1 & 3.8 & 31.7 & 103 & 316 & 767 \\
$q$=2 & 4.9 & 39.7 & 133 & 408 & 888 \\
\end{tabular} \\\\
PbP-QLP &
\begin{tabular}{|p{0.5cm} p{.75cm} p{0.65cm} p{0.6cm} p{0.65cm} p{0.5cm}}
$q$=0 & 2.1 & 15.4 & 48.7 & 160 & 319 \\
$q$=1 & 3.3 & 24.1 & 77.5 & 248 & 496 \\
$q$=2 & 4.3 & 33.4 & 107 & 332 & 654 \\
\end{tabular} \\
\noindent\rule{8.1cm}{0.4pt}\\
\end{tabular}
\label{Table2RuntimePbPQLP}
\end{table}
\begin{table}[!htb]
\caption{Computational time (in seconds) for different randomized algorithms. $d$ is the sampling size parameter, and $q$ is the power iteration factor.}
\begin{tabular}{p{1.2cm} p{0.5cm} p{2.5cm} p{2.5cm}}
\noindent\rule{8.1cm}{0.4pt}\\
Algorithms & &
\begin{tabular}{p{0.65cm} p{0.65cm} p{0.65cm} p{0.65cm} p{0.65cm}}
\multicolumn{5}{c}{$n$ $(d=0.3n)$} \\
\hline
5000 & 10000 & 15000 & 20000 & 25000 \\
\end{tabular} \\
\noindent\rule{8.1cm}{0.4pt}\\
R-SVD &
\begin{tabular}{|p{0.5cm} p{.75cm} p{0.65cm} p{0.6cm} p{0.65cm} p{0.5cm}}
$q$=0 & 4.7 & 40.5 & 142 & 387 & 856 \\
$q$=1 & 6.3 & 53.2 & 193 & 504 & 1148 \\
$q$=2 & 8.0 & 66.6 & 244 & 622 & 1423 \\
\end{tabular} \\ \\
CoR-UTV &
\begin{tabular}{|p{0.5cm} p{.75cm} p{0.65cm} p{0.6cm} p{0.65cm} p{0.5cm}}
$q$=0 & 5.9 & 48.8 & 179 & 481 & 993 \\
$q$=1 & 7.7 & 61.7 & 230 & 618 & 1325 \\
$q$=2 & 9.5 & 74.3 & 280 & 732 & 1626 \\
\end{tabular} \\\\
PbP-QLP &
\begin{tabular}{|p{0.5cm} p{.75cm} p{0.65cm} p{0.6cm} p{0.65cm} p{0.5cm}}
$q$=0 & 3.5 & 25.9 & 95.0 & 285 & 601 \\
$q$=1 & 5.2 & 38.7 & 146 & 402 & 896 \\
$q$=2 & 7.0 & 51.5 & 197 & 530 & 1201 \\
\end{tabular} \\
\noindent\rule{8.1cm}{0.4pt}\\
\end{tabular}
\label{Table3RuntimePbPQLP}
\end{table}
\begin{figure}[t]
\begin{center}
\input{Plots/ASV_Mat1_LG}
\captionsetup{justification=centering,font=scriptsize}
\caption{Singular value approximations for \texttt{LowRankLargeGap}. Left: basic PbP-QLP. Right: PI-coupled PbP-QLP.}
\label{fig_SVM1_LG}
\end{center}
\end{figure}
\subsection{Test Matrices with Randomly Generated Variables}
\label{subsec_LRA_SynData}
We construct four classes of input matrices. The first two
classes contain one or multiple gaps in the spectrum and are particularly designed to investigate the rank-revealing property of PbP-QLP. The second two classes have fast and slow decay singular values. For Matrices 1, 3 and 4, we further investigate the low-rank approximation accuracy of PbP-QLP in relation to those of the optimal SVD and some existing methods. We generate square matrices of order $n = 1000$.
\begin{itemize}
\item Matrix 1 (low-rank plus noise). This rank-$k$ matrix, with $k = 20$, is formed as follows:
\begin{equation}\label{eq_A1PlusA2}
{\bf A} = {\bf A}_1 + {\bf A}_2,
\end{equation}
where ${\bf A}_1 = {\bf U\Sigma V}^T$. Matrices $\bf U$ and $\bf V$ are random orthogonal, and $\bf \Sigma$ is diagonal whose entries ($\sigma_i$s) decrease linearly from 1 to $10^{-25}$, and $\sigma_{k+1} =...=\sigma_n= 0$. ${\bf A}_2 = \mu \sigma_k {\bf N}$, where $\bf N$ is a normalized Gaussian matrix. We consider two cases for $\mu$:
\begin{itemize}
\item [i)] $\mu = 0.005$ in which the matrix has a gap $\approx 200$. This matrix is denoted by \texttt{LowRankLargeGap}.
\item [ii)] $\mu = 0.02$ in which the matrix has a gap $\approx 50$. This matrix is denoted by \texttt{LowRankSmallGap}.
\end{itemize}
\item Matrix 2 (the devil’s stairs \cite{StewartQLP}). This challenging matrix has multiple gaps in its spectrum. The singular values are arranged analogues to a descending staircase with each step consisting of $\ell = 15$ equal singular values.
\item Matrix 3 (fast decay). This matrix is generated as ${\bf A}_1$ in \eqref{eq_A1PlusA2}, however the diagonal elements of $\bf \Sigma$ have the form $\sigma_i = e^{-i/6}$, for $i=1,..., n$.
\item Matrix 4 (slow decay). This matrix is also formed as ${\bf A}_1$, but the diagonal entries of $\bf \Sigma$ take the form $\sigma_i = i^{-2}$, for $i=1,..., n$.
\end{itemize}
\begin{figure}[t]
\begin{center}
\input{Plots/ASV_Mat1_SG}
\captionsetup{justification=centering,font=scriptsize}
\caption{Singular value approximations of \texttt{LowRankSmallGap}. Left: basic PbP-QLP. Right: PI-coupled PbP-QLP with $q=2$.}
\label{fig_SVM1_SG}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\input{Plots/ASV_Mat2}
\captionsetup{justification=centering,font=scriptsize}
\caption{Singular value approximations for Matrix 2. Left: basic PbP-QLP. Right: PI-coupled PbP-QLP with $q=2$.}
\label{fig_SVM2}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\input{Plots/ASV_Mat3}
\captionsetup{justification=centering,font=scriptsize}
\caption{Singular value approximations for Matrix 3. Left: basic PbP-QLP. Right: PI-coupled PbP-QLP with $q=2$.}
\label{fig_SVM3}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\input{Plots/ASV_Mat4}
\captionsetup{justification=centering,font=scriptsize}
\caption{Singular value approximations for Matrix 4. Left: basic PbP-QLP. Right: PI-coupled PbP-QLP with $q=2$.}
\label{fig_SVM4}
\end{center}
\end{figure}
The results for singular values estimation are plotted in Figs. \ref{fig_SVM1_LG}-\ref{fig_SVM4}. We make several observations:
\begin{enumerate}
\item The numerical rank of both matrices \texttt{LowRankLargeGap} and \texttt{LowRankSmallGap} is strongly revealed in $\bf R$ generated by PbP-QLP with no PI technique. This is due to the fact that the gaps in the spectrums of these matrices are well-defined. R-values of PI-coupled PbP-QLP are as accurate as those of the optimal SVD. Though CPQR reveals the gaps, it substantially underestimates the principal singular values of the two matrices. Figs. \ref{fig_SVM1_LG} and \ref{fig_SVM1_SG} show that PbP-QLP is a rank-revealer.
\item For Matrix 2, R-values of basic PbP-QLP do not clearly disclose the gaps in matix's spectrum. This is because the gaps are not substantial. However, basic PbP-QLP strongly reveals the gaps, which shows the procedure that leads to the formation $\bf R$ provides a good first step for PbP-QLP. R-values of PI-coupled PbP-QLP clearly disclose that gaps, due to the effect of PI (see Remark \ref{Remark_Basis}). CPQR reveals the gaps, however it underestimates the singular values.
\item For Matrices 3 and 4, PbP-QLP, though it only uses QR factorization, provides highly accurate singular values, showing similar performance as R-SVD and CoR-UTV.
\end{enumerate}
The results presented in Figs. \ref{fig_SVM1_LG}-\ref{fig_SVM4} demonstrate the applicability of PbP-QLP in accurately estimating the singular values of matrices from different classes.
For Matrices 1 (\texttt{LowRankSmallGap}), 3 and 4, we now compare the accuracy of PbP-QLP, together with other considered algorithms, for rank-$d$ approximations, where $d\ge 1$. We calculate the error as follow:
\begin{equation}\notag
\text{Error} = \|{\bf A} - {\bf A}_\text{Approx}\|_2.
\end{equation}
Here, ${\bf A}_\text{Approx}$ is an approximation constructed by each algorithm. The results are shown in Figs. \ref{fig_LRAE_M1}-\ref{fig_LRAE_M4}. We make two observations:
\begin{enumerate}
\item For Matrix 1, approximations by basic PbP-QLP for $1\le d \le 20$ are highly accurate, while for $d>20$, the approximations, similar to those of R-SVD and CoR-UTV, are poorer than the CPQR, p-QLP, and SVD. However, PI-coupled PbP-QLP with $q=2$ produces approximations as accurate as p-QLP and the SVD for any rank parameter $d$.
\item For Matrices 3 and 4 CPQR shows a better performance compared with R-SVD, CoR-UTV and PbP-QLP methods with no PI. While PI-coupled PbP-QLP with $q=2$ constructs approximations as good as the optimal SVD.
\end{enumerate}
\begin{figure}[t]
\begin{center}
\input{Plots/BLRE_Mat1}
\captionsetup{justification=centering,font=scriptsize}
\caption{Low-rank approximation errors for Matrix 1 (\texttt{LowRankSmallGap}). Left: basic PbP-QLP. Right: PI-coupled PbP-QLP with $q=2$.}
\label{fig_LRAE_M1}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\input{Plots/BLRE_Mat3}
\captionsetup{justification=centering,font=scriptsize}
\caption{Low-rank approximation errors for Matrix 3. Left: basic PbP-QLP. Right: PI-coupled PbP-QLP with $q=2$.}
\label{fig_LRAE_M3}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\input{Plots/BLRE_Mat4}
\captionsetup{justification=centering,font=scriptsize}
\caption{Low-rank approximation errors for Matrix 4. Left: basic PbP-QLP. Right: PI-coupled PbP-QLP with $q=2$.}
\label{fig_LRAE_M4}
\end{center}
\end{figure}
\subsection{Low-Rank Image Reconstruction}
This experiment investigates the performance of PbP-QLP on real-world data, where we reconstruct a gray-scale low-rank image of a butterfly of size $969\times 812$. We compare the results of PbP-QLP against those produced by the (optimal) truncated SVD, truncated CPQR and p-QLP, R-SVD, and CoR-UTV. The reconstructions with $rank = 80$ are displayed in Fig. \ref{fig:Butterfly}. We observe, with close scrutiny, that (i) the reconstructions by CPQR and the three randomized methods with no PI contains more noise than those of other methods, (ii) the reconstruction by CPQR, in addition to adding noise to the image, slightly distorts
butterfly’s antennas and upper parts of forewings (close to the head), and (iii) the approximation by PI-coupled PbP-QLP with $q = 1$ is as good as those of the p-QLP and SVD.
Fig. \ref{fig_Butterfly_FroErr} displays the reconstruction errors in terms of the Frobenius norm for considered methods against the
approximation rank. We observe that (i) randomized methods with no PI demonstrate similar performance and their reconstructions generate more errors, and (ii) PbP-QLP-coupled with $q = 2$ outperforms p-QLP and constructs approximations with almost no loss of accuracy compared to the optimal SVD.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.42\textwidth,height=15cm]{Plots/Butterfly10_PbP}
\caption{Low-rank image reconstruction. This figure shows the results of reconstructing a butterfly image with dimension $969\times 812$ using $rank=80$ by different methods.}
\label{fig:Butterfly}
\end{figure}
\begin{figure}[t]
\begin{center}
\input{Plots/Cbfly_FErr_PbPQLP}
\caption{Image reconstruction Frobenius norm approximation error. This figure displays the errors incurred by different methods in reconstructing a butterfly image. Here, ${\bf A}_{\text{approx}}$ is an approximation generated by either method. The approximation by PI-coupled PbP-QLP with $q=2$ shows almost no loss of accuracy compared to the optimal truncated SVD.}
\label{fig_Butterfly_FroErr}
\end{center}
\end{figure}
\subsection{Test Matrices from Applications}
In this experiment, we investigate the effectiveness of the PbP-QLP algorithm on five matrices of size $256 \times 256$ described in Table \ref{Table_PracMat} from different applications \cite{Hansen_RegTools}. These matrices have been used in other works, e.g., \cite{DemGGX15, Ayala19}.
\begin{table}[!htb]
\caption{Test matrices from applications.}
\begin{tabular}
{p{0.08cm} p{1cm} p{6.5cm} }
\noindent\rule{8.4cm}{0.4pt}\\
No. & Matrix & Description \\
\noindent\rule{8.4cm}{0.4pt}\\
1 & \texttt{Baart} & Discretization of a first-kind Fredholm integral equation. \\
2 & \texttt{Deriv2} & Computation of the second derivative. \\
3 & \texttt{Foxgood} & Severely ill-posed problem. \\
4 & \texttt{Gravity} & 1D gravity surveying problem. \\
5 & \texttt{Heat} & Inverse heat equation. \\
\noindent\rule{8.4cm}{0.4pt}
\end{tabular}
\label{Table_PracMat}
\end{table}
\subsubsection{Low-Rank Approximation}
With matrices from Table \ref{Table_PracMat} as our inputs, similar to the experiment in Section \ref{subsec_LRA_SynData}, we construct rank-$d$ approximations using different algorithms and compute the approximation errors. The results are shown in Figs. \ref{fig_Baart}-\ref{fig_Heat}. We observe that while PbP-QLP with no PI scheme provides fairly accurate low-rank approximations for \texttt{Baart}, \texttt{Foxgood} and \texttt{Gravity}, the approximations for \texttt{Deriv2} and \texttt{Heat}, similar to those of R-SVD, are rather poor. This is because these two matrices have slowly decaying singular values. However, the errors incurred by PI-coupled PbP-QLP overlap those produced by the optimal SVD for all five matrices, showing the high accuracy of PbP-QLP. We further observe in Figs. \ref{fig_Baart} and \ref{fig_Foxgood} that for \texttt{Baart} and \texttt{Foxgood}, basic CoR-UTV produces less accurate results as the rank parameter increases. This is due to the application of input matrix to the sample matrix (Step 4 of Algorithm \ref{Alg_CoRUTV}) without any orthonormalization being applied (see Section \ref{subsec_PIPbPQLP}). To be precise, in this case, the best accuracy attainable by the algorithm is $\sigma_1 \epsilon_\text{machine}^{1/2}$. Let $\epsilon_\text{machine}=10^{-16}$, and we have the largest singular values of \texttt{Baart} and \texttt{Foxgood} equal to $3.2$ and $0.8$, respectively. Thus, the best accuracy attinable by basic CoR-UTV for \texttt{Baart} is $3.2 \times 10^{-8}$, and for \texttt{Foxgood} is $8\times 10^{-9}$. This issue, however, is resolved by applying the orthonormalization scheme, as the plots show.
\subsubsection{Matrix $\ell_2$-norm Estimation}
For any matrix $\bf A$, $\|{\bf A}\|_2$ is equal to its largest singular value. We compute an estimation to the largest singular value of each matrix from Table \ref{Table_PracMat} using PbP-QLP, and compare the results with those of CPQR and p-QLP. Fig. \ref{fig_Ratio} displays the ratios of the estimated singular values to the exact norms. It is seen that (i) basic PbP-QLP substantially outperforms CPQR in estimating the matrix $\ell_2$-norm, while it produces comparable results with p-QLP, and (ii) PI-coupled PbP-QLP provides excellent estimations to the first singular values and its performance exceeds that of p-QLP for all matrices.
\begin{figure}[t]
\begin{center}
\input{Plots/ELRE_Baart}
\captionsetup{justification=centering,font=scriptsize}
\caption{Low-rank approximation errors for \texttt{Baart}. Left: basic PbP-QLP. Right: PI-coupled PbP-QLP with $q=2$.}
\label{fig_Baart}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\input{Plots/ELRE_Derive2}
\captionsetup{justification=centering,font=scriptsize}
\caption{Low-rank approximation errors for \texttt{Derive2}. Left: basic PbP-QLP. Right: PI-coupled PbP-QLP with $q=2$.}
\label{fig_Derive2}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\input{Plots/ELRE_Foxgood}
\captionsetup{justification=centering,font=scriptsize}
\caption{Low-rank approximation errors for \texttt{Foxgood}. Left: basic PbP-QLP. Right: PI-coupled PbP-QLP with $q=2$.}
\label{fig_Foxgood}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\input{Plots/ELRE_Gravity}
\captionsetup{justification=centering,font=scriptsize}
\caption{Low-rank approximation errors for \texttt{Gravity}. Left: basic PbP-QLP. Right: PI-coupled PbP-QLP with $q=2$.}
\label{fig_Gravity}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\input{Plots/ELRE_Heat}
\captionsetup{justification=centering,font=scriptsize}
\caption{Low-rank approximation errors for \texttt{Heat}. Left: basic PbP-QLP. Right: PI-coupled PbP-QLP with $q=2$.}
\label{fig_Heat}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\input{Plots/DRatio1_5}
\captionsetup{justification=centering,font=scriptsize}
\caption{Ratios of the estimated $\ell_2$-norm to the exact norm for matrices of Table \ref{Table_PracMat}.}
\label{fig_Ratio}
\end{center}
\end{figure}
\section{Conclusion}
\label{section_concl}
We presented in this paper the rank-revealing PbP-QLP algorithm, which, by utilizing randomization, constructs an approximation to the pivoted QLP and truncated SVD. PbP-QLP is primarily designed to approximate low-rank matrices. It consists of two stages, each performing only the unpivoted QR factorization to factor the associated small matrices. With theoretical analysis, we showed that the numerical rank of a given matrix is revealed in the first stage. We further furnished a detailed theoretical analysis for PbP-QLP, which brings an insight into the rank-revealing property as well as the accuracy of the algorithm. Through numerical tests conducted on several classes of matrices, we showed our proposed PbP-QLP (i) outperforms R-SVD and CoR-UTV in runtime, and (ii) establishes highly accurate approximations, as accurate as those of the optimal SVD, to the matrices.
\appendices
\section{Proof of Theorem \ref{Th_R22bound}}
\label{Proo_R22bound}
We first prove \eqref{eq_2nR22_bound} for the case $q = 0$ (the basic PbP-QLP). To do so, we write $\bf D$ and its $\bf Q$ and $\bf R$ factors, equations \eqref{eq_D} and \eqref{eq_DQR}, with partitioned matrices:
\begin{equation}\label{key}
\begin{aligned}
{\bf D} = {\bf A}\bar{\bf P} = {\bf QR} =[{\bf D}_1\quad {\bf D}_2]=[{\bf Q}_1\quad {\bf Q}_2]
\begin{bmatrix}
{\bf R}_{11} & {\bf R}_{12} \\
{\bf 0} & {\bf R}_{22}
\end{bmatrix},
\end{aligned}
\notag
\end{equation}
where ${\bf D}_1$ and ${\bf Q}_1$ contain the first $k$ columns, and ${\bf D}_2$ and ${\bf Q}_2$ contain the remaining $d-k$ columns of ${\bf D}$ and ${\bf Q}$, respectively.
We now define a matrix $\bf X$ constructed by the interaction of $\bf V$ (right singular vectors of $\bf A$) and $\bar{\bf P}$:
\begin{equation}
{\bf X} \coloneqq {\bf V}^T\bar{\bf P} = \begin{bmatrix}
{\bf V}_k^T \\
{\bf V}_\perp^T
\end{bmatrix}[\bar{\bf P}_1\quad \bar{\bf P}_2]
= \begin{bmatrix}
{\bf X}_{11} & {\bf X}_{12} \\
{\bf X}_{21} & {\bf X}_{22}
\end{bmatrix},
\label{eq_X}
\end{equation}
where $\bar{\bf P}_1$ contains the first $k$ columns, and $\bar{\bf P}_2$ contains the remaining $d-k$ columns of $\bar{\bf P}$. With a simple computation, we will obtain the following four equalities:
\begin{equation}\label{eq_D1}
{\bf D}_1 = {\bf U}_k{\bf \Sigma}_k{\bf X}_{11} + {\bf U}_{\perp}{\bf \Sigma}_{\perp}{\bf X}_{21}.
\end{equation}
\begin{equation}\label{eq_D2}
\begin{aligned}
{\bf D}_2 = {\bf U}_k{\bf \Sigma}_k{\bf X}_{12} + {\bf U}_{\perp}{\bf \Sigma}_{\perp}{\bf X}_{22}.
\end{aligned}
\end{equation}
\begin{equation}\label{key}
{\bf D}_1 = {\bf Q}_1{\bf R}_{11}.
\notag
\end{equation}
\begin{equation}\label{key}
\begin{aligned}
{\bf P}_{{\bf D}_1^\perp} {\bf D}_2 = {\bf Q}_2{\bf R}_{22} = ({\bf I} - {\bf D}_1{\bf D}_1^\dagger){\bf D}_2 =
({\bf I} - {\bf Q}_1{\bf Q}_1^T){\bf D}_2.
\end{aligned}
\notag
\end{equation}
Hence, ${\bf R}_{22}$ is obtained as
\begin{equation}
{\bf R}_{22} = {\bf Q}_2^T ({\bf I} - {\bf Q}_1{\bf Q}_1^T){\bf D}_2.
\notag
\end{equation}
We take the $\ell_2$-norm of the above identity:
\begin{equation}\label{key}
\begin{aligned}
\|{\bf R}_{22} \|_2 & \le \|{\bf Q}_2^T\|_2\|({\bf I} - {\bf Q}_1{\bf Q}_1^T){\bf D}_2\|_2\\&
\le \|({\bf I} - {\bf Q}_1{\bf Q}_1^T){\bf D}_2\|_2.
\end{aligned}
\notag
\end{equation}
The last relation follows since for any orthonormal matrix $\bf M$, $\|{\bf M}\|_2\le 1$. By replacing ${\bf D}_2$ \eqref{eq_D2} and applying the triangle inequality, we obtain
\begin{equation}\label{key}
\begin{aligned}
\|{\bf R}_{22}\|_2 &\le \|({\bf I} - {\bf Q}_1{\bf Q}_1^T){\bf U}_k{\bf \Sigma}_k{\bf X}_{12}\|_2 \\ &+\|({\bf I} - {\bf Q}_1{\bf Q}_1^T){\bf U}_{\perp}{\bf \Sigma}_{\perp}{\bf X}_{22}\|_2.
\end{aligned}
\notag
\end{equation}
By substituting ${\bf X}_{12}$ and ${\bf X}_{22}$ \eqref{eq_X} into the above equation, and from the orthonormality of $\bar{\bf P}_2$, we will have
\begin{equation}\label{eq_2nR22_2terms}
\|{\bf R}_{22}\|_2 \le \|({\bf I} - {\bf Q}_1{\bf Q}_1^T){\bf A}_k\|_2 +\|({\bf I} - {\bf Q}_1{\bf Q}_1^T){\bf U}_{\perp}{\bf \Sigma}_\perp \|_2,
\end{equation}
where ${\bf A}_k \coloneqq {\bf U}_k{\bf \Sigma}_k{\bf V}_k^T$. We now bound the two terms on the right-hand side of \eqref{eq_2nR22_2terms}. To bound
the first term, firstly
\begin{equation}
\mathcal{R}({\bf D}) = \mathcal{R}({\bf A}\bar{\bf P})= \mathcal{R}({\bf AA}^T{\bf \Phi}).
\notag
\end{equation}
The second equality holds as $\mathcal{R}(\bar{\bf P})=\mathcal{R}({\bf A}^T{\bf \Phi})$. Hence,
\begin{equation}
\begin{aligned}
{\bf AA}^T{\bf \Phi} = {\bf U \Sigma}^2{\bf U}^T {\bf \Phi} ={\bf U}\begin{bmatrix}
{\bf \Sigma}_1^{2} & {\bf 0} & {\bf 0} \\
{\bf 0} & {\bf \Sigma}_2^{2} & {\bf 0}\\
{\bf 0} & {\bf 0} & {\bf \Sigma}_3^{2}
\end{bmatrix}
\begin{bmatrix}
{\bf \Phi}_1\\
\\
{\bf \Phi}_2
\end{bmatrix}= {\bf Q}{\bf R},
\end{aligned}
\notag
\end{equation}
where ${\bf \Sigma}_1 \in \mathbb R^{k \times k}$, ${\bf \Sigma}_2 \in \mathbb R^{(d-p-k) \times (d-p-k)}$, and ${\bf \Sigma}_3 \in \mathbb R^{(n_2-d+p) \times (n_2-d+p)}$. Assuming that ${\bf \Phi}_1$ is full row rank and its Moore-Penrose inverse satisfies
\begin{equation}
{\bf \Phi}_1{\bf \Phi}_1^\dagger ={\bf I},
\notag
\end{equation}
we define a non-singular matrix $\bf Y$ as follows:
\begin{equation}
{\bf Y} = \bigg[
{\bf \Phi}_1^\dagger
\begin{pmatrix}
{\bf \Sigma}_1^2 & {\bf 0}\\
{\bf 0} & {\bf \Sigma}_2^2
\end{pmatrix}^{-1}, \bar{\bf Y}\bigg],
\notag
\end{equation}
where $\bar{\bf Y}\in \mathbb R^{d \times p}$ is chosen so that ${\bf Y}\in \mathbb R^{d \times d}$ is non-singular and ${\bf \Phi}_1\bar{\bf Y} = \bf 0$. We then compute the matrix product:
\begin{equation}
\begin{aligned}
{\bf A}{\bf A}^T {\bf \Phi Y} =
& {\bf U}\begin{bmatrix}
\begin{pmatrix}
{\bf \Sigma}_1^2 & {\bf 0}\\
{\bf 0} & {\bf \Sigma}_2^2
\end{pmatrix}{\bf \Phi}_1\\
{\bf \Sigma}_3^2 {\bf \Phi}_2
\end{bmatrix}{\bf Y}= {\bf U} \begin{bmatrix}
{\bf I} & {\bf 0} & {\bf 0} \\
{\bf 0} & {\bf I} & {\bf 0}\\
{\bf Z}_1 & {\bf Z}_2 & {\bf Z}_3
\end{bmatrix},
\end{aligned}
\label{eq_AAtPhiY_part}
\end{equation}
where ${\bf Z}_1 = {\bf \Sigma}_3^2{\bf \Phi}_2{\bf \Phi}_1^\dagger {\bf \Sigma}_1^{-2} \in \mathbb R^{(n_2-d+p) \times k}$, ${\bf Z}_2 = {\bf \Sigma}_3^2{\bf \Phi}_2{\bf \Phi}_1^\dagger {\bf \Sigma}_2^{-2} \in \mathbb R^{(n_2-d+p) \times (d-p-k)}$, and
${\bf Z}_3 = {\bf \Sigma}_3^2{\bf \Phi}_2 \bar{\bf Y} \in \mathbb R^{(n_2-d+p) \times p}$. Let the matrix in \eqref{eq_AAtPhiY_part} have a QR factorization:
\begin{equation}
\begin{aligned}
{\bf U} \begin{bmatrix}
{\bf I} & {\bf 0} & {\bf 0} \\
{\bf 0} & {\bf I} & {\bf 0}\\
{\bf Z}_1 & {\bf Z}_2 & {\bf Z}_3
\end{bmatrix} =
{\bf Q}^\prime {\bf R}^\prime = \begin{bmatrix}
{{\bf Q}_1^\prime}^T\\
{{\bf Q}_2^\prime}^T\\
{{\bf Q}_3^\prime}^T
\end{bmatrix}^T
\begin{bmatrix}
{\bf R}_{11}^\prime & {\bf R}_{12}^\prime & {\bf R}_{13}^\prime \\
{\bf 0} & {\bf R}_{22}^\prime & {\bf R}_{23}^\prime\\
{\bf 0} & {\bf 0} & {\bf R}_{33}^\prime
\end{bmatrix}
\end{aligned}
\notag
\end{equation}
which gives
\begin{equation}\label{eq_Q1prime}
{\bf U} \begin{bmatrix}
{\bf I} \\
{\bf 0} \\
{\bf Z}_1
\end{bmatrix} = {\bf Q}_1^\prime{\bf R}_{11}^\prime.
\end{equation}
Since matrix $\bf Y$ is non-singular, by \cite[Lemma 4.1]{Gu2015}, we have ${\bf Q}{\bf Q}^T = {\bf Q}^\prime {{\bf Q}^\prime}^T$. Exploiting \eqref{eq_Q1prime}, it follows that
\begin{equation}
\begin{aligned}
&{\bf I} - {\bf Q}_1 {{\bf Q}_1}^T = {\bf U}
\begin{bmatrix}
{\bf I} - \bar{\bf Z}^{-1} & {\bf 0} & -\bar{\bf Z}^{-1}{\bf Z}_1^T \\
{\bf 0} & {\bf I} & {\bf 0}\\
- {\bf Z}_1 \bar{\bf Z}^{-1} & {\bf 0} &
{\bf I} - {\bf Z}_1 \bar{\bf Z}^{-1}{\bf Z}_1^T
\end{bmatrix} {\bf U}^T,
\end{aligned}
\notag
\end{equation}
where ${\bf Z}_1$ is defined in \eqref{eq_AAtPhiY_part}, and $\bar{\bf Z}^{-1}$ is defined as follows:
\begin{equation}\label{eq_ZbarInv}
\begin{aligned}
\bar{\bf Z}^{-1} \coloneqq {{\bf R}_{11}^\prime}^{-1} {{\bf R}_{11}^\prime}^{-T} =
({{\bf R}_{11}^\prime}^T {\bf R}_{11}^\prime)^{-1} = ({\bf I}+{\bf Z}_1^T{\bf Z}_1) ^{-1}.
\end{aligned}
\notag
\end{equation}
Writing ${\bf A}_k = {\bf U}[{\bf \Sigma}_1 \quad {\bf 0}\quad {\bf 0}]^T{\bf V}^T$, we obtain
\begin{equation}
\begin{aligned}
({\bf I} - {\bf Q}_1{\bf Q}_1^T){\bf A}_k = {\bf U}
\begin{bmatrix}
({\bf I} - \bar{\bf Z}^{-1}){\bf \Sigma}_1 \\
{\bf 0} \\
- {\bf Z}_1 \bar{\bf Z}^{-1}{\bf \Sigma}_1
\end{bmatrix} {\bf V}^T.
\end{aligned}
\notag
\end{equation}
It follows that
\begin{equation}
\begin{aligned}\label{eq_2nIminQ1Q1tAk2}
\|({\bf I} - {\bf Q}_1{\bf Q}_1^T){\bf A}_k \|_2^2 & = \|{\bf \Sigma}_1 ({\bf I} - \bar{\bf Z}^{-1}){\bf \Sigma}_1\|_2 \\ & \le \|{\bf \Sigma}_1\|_2^2\|{\bf I} - \bar{\bf Z}^{-1}\|_2,
\end{aligned}
\end{equation}
where we have used the following relation that holds for any
matrix $\bf M$ with ${\bf I} +{\bf M}$ being non-singular \cite{HendersonS81}:
\begin{equation}
({\bf I} +{\bf M})^{-1} = {\bf I} - {\bf M}({\bf I} +{\bf M})^{-1} = {\bf I} - ({\bf I} +{\bf M})^{-1}{\bf M}.
\notag
\end{equation}
Let ${\bf K}\coloneqq {\bf I}-\bar{\bf Z} ^{-1}$. Matrix $\bf K$ is positive semidefinite, and its eigenvalues satisfy \cite[p.148]{StewartSun90}:
\begin{equation}
\begin{aligned}
\lambda_i({\bf K}) = \frac{{\sigma}_i^2({\bf Z}_1)}{1+{\sigma}_i^2({\bf Z}_1)}, \quad i=1,...k.
\end{aligned}
\notag
\end{equation}
The largest singular value of ${\bf Z}_1$ satisfies:
\begin{equation}\notag
\sigma_1 ({\bf Z}_1) \le \delta_k^2\|{\bf \Phi}_2\|_2\|{\bf \Phi}_1^\dagger\|_2.
\end{equation}
Accordingly,
\begin{equation}
\begin{aligned}
\lambda_1({\bf K}) \le \frac{\delta_k^4\|{\bf \Phi}_2\|_2^2\|{\bf \Phi}_1^\dagger\|_2^2}{1+\delta_k^4\|{\bf \Phi}_2\|_2^2\|{\bf \Phi}_1^\dagger\|_2^2},
\end{aligned}
\notag
\end{equation}
Plugging this result into \eqref{eq_2nIminQ1Q1tAk2} and taking the square root, it follows
\begin{equation}\label{eq_2n1ndTermR22}
\begin{aligned}
\|({\bf I} - {\bf Q}_1{\bf Q}_1^T){\bf A}_k \|_2\le \frac{\delta_k^{2}\sigma_1\|{\bf \Phi}_2\|_2\|{\bf \Phi}_1^\dagger\|_2}{\sqrt{1 + \delta_k^{4}\|{\bf \Phi}_2\|_2^2\|{\bf \Phi}_1^\dagger\|_2^2}},
\end{aligned}
\end{equation}
For the second term on the right-hand side of \eqref{eq_2nR22_2terms}, we have
\begin{equation}
\begin{aligned}
\|({\bf I} - {\bf Q}_1{\bf Q}_1^T){\bf U}_{\perp}{\bf \Sigma}_\perp\|_2 & \le \|{\bf I} - {\bf Q}_1{\bf Q}_1^T\|_2\|{\bf U}_\perp\|_2\|{\bf \Sigma}_\perp\|_2\\ & \le \sigma_{k+1}.
\end{aligned}
\label{eq_2n2ndTermR22}
\end{equation}
By substituting the results in \eqref{eq_2n1ndTermR22} and \eqref{eq_2n2ndTermR22} into \eqref{eq_2nR22_2terms}, the
theorem for the basic version of PbP-QLP follows.
When the PI scheme with power parameter $q$ is used, $\bf A$ is supplanted by $({\bf A}^T{\bf A})^q{\bf A}^T$, and as a result
\begin{equation}
\mathcal{R}({\bf D}) = \mathcal{R}({\bf A}\bar{\bf P})= \mathcal{R}({\bf A}({\bf A}^T{\bf A})^q{\bf A}^T{\bf \Phi}).
\notag
\end{equation}
Therefore, considering
\begin{equation}\notag
{\bf A}({\bf A}^T{\bf A})^q{\bf A}^T{\bf \Phi} = {\bf U \Sigma}^{2q+2}{\bf U}^T {\bf \Phi},
\end{equation}
we now define a non-singular matrix $\bf Y$ as:
\begin{equation}\notag
{\bf Y} = \bigg[
{\bf \Phi}_1^\dagger
\begin{pmatrix}
{\bf \Sigma}_1^{2q+2} & {\bf 0}\\
{\bf 0} & {\bf \Sigma}_2^{2q+2}
\end{pmatrix}^{-1}, \bar{\bf Y}\bigg],
\end{equation}
and compute the product
\begin{equation}\notag
\begin{aligned}
{\bf A}({\bf A}^T{\bf A})^q{\bf A}^T{\bf \Phi Y} =
{\bf U} \begin{bmatrix}
{\bf I} & {\bf 0} & {\bf 0} \\
{\bf 0} & {\bf I} & {\bf 0}\\
{\bf Z}_1 & {\bf Z}_2 & {\bf Z}_3
\end{bmatrix},
\end{aligned}
\label{}
\end{equation}
where ${\bf Z}_1 = {\bf \Sigma}_3^{2q+2}{\bf \Phi}_2{\bf \Phi}_1^\dagger {\bf \Sigma}_1^{-({2q+2})}$, ${\bf Z}_2 = {\bf \Sigma}_3^{2q+2}{\bf \Phi}_2{\bf \Phi}_1^\dagger {\bf \Sigma}_2^{-({2q+2})}$, and ${\bf Z}_3 = {\bf \Sigma}_3^{2q+2}{\bf \Phi}_2 \bar{\bf Y}$. Proceeding further with the procedure as described for $q=0$, the result for PI-incorporated PbP-QLP follows. \QEDB
\section{Proof of Theorem \ref{Th_boundssigmak}}
\label{Proo_boundssigmak}
The left inequality in \eqref{eq_Th_boundssigmak} is easily obtained by the Cauchy’s interlacing theorem \cite{StewartSun90}. To prove the right inequality, we exploit the following theorem from \cite{HornJohnson94}.
\begin{theorem} (Horn and Johnson \cite{HornJohnson94}). Let ${\bf M}, {\bf N} \in \mathbb R^{n_1 \times n_2}$ and $\ell=min\{n_1, n_2\}$. Then for $1\le i, j \le \ell$ and $i+j \le \ell+1$,
\begin{equation}\notag
\sigma_i({\bf MN}) \le \sigma_1({\bf M})\sigma_i({\bf N}).
\end{equation}
\begin{equation}\notag
\sigma_{i+j-1}({\bf M}+{\bf N}) \le \sigma_i({\bf M})+ \sigma_j({\bf N}).
\end{equation}
\begin{equation}
\sigma_{i+j-1}({\bf MN}) \le \sigma_i({\bf M})\sigma_j({\bf N}).
\label{eq1_HornJohn94}
\end{equation}
\label{Thr_HornJohn94}
\end{theorem}
By the definition of ${\bf D}_1$ \eqref{eq_D1}, we have
\begin{equation}\label{eq_R11_exp}
\begin{aligned}
{\bf R}_{11} & = {\bf Q}_1^T{\bf D}_1 = {\bf Q}_1^T({\bf U}_k{\bf \Sigma}_k{\bf X}_{11} + {\bf U}_{\perp}{\bf \Sigma}_{\perp}{\bf X}_{21})\\ &
= {\bf Q}_1^T{\bf U}_k{\bf \Sigma}_k{\bf X}_{11} + {\bf Q}_1^T{\bf U}_{\perp}{\bf \Sigma}_{\perp}{\bf X}_{21}.
\end{aligned}
\end{equation}
Exploiting Theorem \ref{Thr_HornJohn94} results in
\begin{equation}\label{eq_sigmak_R11_proo}
\begin{aligned}
\sigma_k({\bf R}_{11}) & \le \sigma_k({\bf Q}_1^T{\bf U}_k{\bf \Sigma}_k{\bf X}_{11}) + \sigma_1({\bf Q}_1^T{\bf U}_{\perp}{\bf \Sigma}_{\perp}{\bf X}_{21})\\
& \le \sigma_k({\bf Q}_1^T{\bf U}_k{\bf \Sigma}_k{\bf X}_{11}) + \sigma_{k+1},
\end{aligned}
\end{equation}
and $\sigma_k$ of ${\bf Q}_1^T{\bf U}_k{\bf \Sigma}_k{\bf X}_{11}$ is bounded above by:
\begin{equation}\notag
\begin{aligned}
\sigma_k({\bf Q}_1^T{\bf U}_k{\bf \Sigma}_k{\bf X}_{11}) & \le \sigma_1({\bf Q}_1^T)\sigma_k({\bf U}_k{\bf \Sigma}_k{\bf X}_{11})\\ & \le \sigma_1({\bf U}_k)\sigma_k({\bf \Sigma}_k{\bf X}_{11}) \le
\sigma_k.
\end{aligned}
\end{equation}
The last relation follows by applying \eqref{eq1_HornJohn94} with $i=k$ and $j=1$. Plugging the result into \eqref{eq_sigmak_R11_proo} gives
\begin{equation}\notag
\sigma_k({\bf R}_{11}) \le \sigma_k + \sigma_{k+1},
\end{equation}
which completes the proof. \QEDB
\section{Proof of Theorem \ref{Th_distUkVkQ1P1}}
\label{Proo_distance}
We provide proofs for the case the PI is used (for the simple form of PbP-QLP, $q = 0$). According to Theorem 2.6.1 of \cite{GolubVanLoan96}, we have for the distance between $\mathcal{R}({\bf U}_k)$ and $\mathcal{R}({\bf Q}_1)$
\begin{equation}
\begin{aligned}
\text{dist}(\mathcal{R}({\bf U}_k), \mathcal{R}({\bf Q}_1)) = \|{\bf U}_k{\bf U}_k^T - {\bf Q}_1{\bf Q}_1^T\|_2 = \|{\bf U}_\perp^T{\bf Q}_1\|_2.
\label{eq_dist_def}
\end{aligned}
\end{equation}
For the range of $\bf Q$, we have
\begin{equation}\notag
\mathcal{R}({\bf Q}) = \mathcal{R}({\bf D}) = \mathcal{R}({\bf A}\bar{\bf P})= \mathcal{R}({\bf A}({\bf A}^T{\bf A})^q{\bf A}^T{\bf \Phi}).
\label{}
\end{equation}
We therefore have
\begin{equation}\notag
\begin{aligned}
{\bf A}({\bf A}^T{\bf A})^q{\bf A}^T{\bf \Phi} & = {\bf U}\begin{bmatrix}
{\bf \Sigma}_k^{2q+2} & {\bf 0} \\
{\bf 0} & {\bf \Sigma}_\perp^{2q+2}
\end{bmatrix} \begin{bmatrix}
{\bf \Phi}_{11} & {\bf \Phi}_{12} \\
{\bf \Phi}_{21} & {\bf \Phi}_{22}
\end{bmatrix}\\
& = [{\bf Q}_1\quad {\bf Q}_2]
\begin{bmatrix}
{\bf R}_{11} & {\bf R}_{12} \\
{\bf 0} & {\bf R}_{22}
\end{bmatrix}.
\end{aligned}
\end{equation}
Let
\begin{equation}\notag
\begin{aligned}
{\bf H}\coloneqq {\bf U}\begin{bmatrix}
{\bf \Sigma}_k^{2q+2} & {\bf 0} \\
{\bf 0} & {\bf \Sigma}_\perp^{2q+2}
\end{bmatrix} \begin{bmatrix}
{\bf \Phi}_{11} \\
{\bf \Phi}_{21}
\end{bmatrix} = {\bf Q}_1{\bf R}_{11}.
\end{aligned}
\end{equation}
We now define a matrix ${\bf J} \in \mathbb R^{k \times k}$ as:
\begin{equation}\notag
{\bf J} \coloneqq {\bf \Phi}_{11}^{-1}{\bf \Sigma}_k^{-(2q+2)},
\end{equation}
and compute a QR factorization of the product $\bf HJ$:
\begin{equation}\notag
\begin{aligned}
{\bf HJ}= {\bf U}\begin{bmatrix}
{\bf I}_k \\
{\bf \Sigma}_\perp^{2q+2}{\bf \Phi}_{21}{\bf \Phi}_{11}^{-1} {\bf \Sigma}_k^{-(2q+2)}
\end{bmatrix} = \ddot{\bf Q}\ddot{\bf R},
\end{aligned}
\end{equation}
which gives
\begin{equation}\notag
\begin{aligned}
\begin{bmatrix}
{\bf I}_k \\
{\bf \Sigma}_\perp^{2q+2}{\bf \Phi}_{21}{\bf \Phi}_{11}^{-1} {\bf \Sigma}_k^{-(2q+2)}
\end{bmatrix} = \begin{bmatrix}
{\bf U}_k^T \\
{\bf U}_\perp^T
\end{bmatrix} \ddot{\bf Q}\ddot{\bf R}.
\end{aligned}
\end{equation}
From ${\bf U}_k^T\ddot{\bf Q}\ddot{\bf R} = {\bf I}_k$, it is derived that $\ddot{\bf R}^{-1} = {\bf U}_k^T\ddot{\bf Q}$. Accordingly,
\begin{equation}\notag
{\bf U}_\perp^T \ddot{\bf Q} = {\bf \Sigma}_\perp^{2q+2}{\bf \Phi}_{21}{\bf \Phi}_{11}^{-1} {\bf \Sigma}_k^{-(2q+2)}{\bf U}_k^T\ddot{\bf Q}.
\end{equation}
From \eqref{eq_dist_def}, it follows that
\begin{equation}\notag
\|{\bf U}_\perp^T \ddot{\bf Q}\|_2 \coloneqq \|{\bf U}_k{\bf U}_k^T - \ddot{\bf Q}\ddot{\bf Q}^T\|_2
\end{equation}
By \cite[Lemma 4.1]{Gu2015}, ${\bf Q}_1{\bf Q}_1^T=\ddot{\bf Q}\ddot{\bf Q}^T$. As a result,
\begin{equation}\notag
\begin{aligned}
\text{dist}(\mathcal{R}({\bf U}_k), \mathcal{R}({\bf Q}_1)) & =
\|{\bf U}_\perp^T \ddot{\bf Q}\|_2 \\ & \le \|{\bf \Sigma}_\perp^{2q+2}{\bf \Phi}_{21}{\bf \Phi}_{11}^{-1} {\bf \Sigma}_k^{-(2q+2)}\|_2\|{\bf U}_k^T\ddot{\bf Q}\|_2,
\end{aligned}
\end{equation}
from which the theorem for \eqref{eq_Th_distance1} follows.
To prove \eqref{eq_Th_distance2}, since ${\bf P} = \bar{\bf P}\widetilde{\bf P}$, we have
\begin{equation}\notag
\mathcal{R}({\bf P}) = \mathcal{R}(\bar{\bf P})= \mathcal{R}(({\bf A}^T{\bf A})^q{\bf A}^T{\bf \Phi}).
\end{equation}
The first relation follows because $\widetilde{\bf P}$ is an orthogonal matrix of order $d$. Thus, it suffices to bound from above $\text{dist}(\mathcal{R}({\bf V}_k), \mathcal{R}(\bar{\bf P}_1))$, where $\bar{\bf P}_1$ is the first $k$ columns of $\bar{\bf P}$. The rest of the proof is similat to that of \eqref{eq_Th_distance1}, and we therefore omit it.
\QEDB
\section{Proof of Theorem \ref{Th_LRapprox}}
\label{Proo_LRapprox}
For the left-hand side term in \eqref{eq_2nLRAErrbound}, we write
\begin{equation}\label{eq_LRAErr_proof}
\begin{aligned}
\|({\bf I} - {\bf QQ}^T){\bf A}\|_2 & \le \|({\bf I} - {\bf QQ}^T){\bf A}_k\|_2 + \|({\bf I} - {\bf QQ}^T){\bf A}_\perp\|_2\\ & \le \|({\bf I} - {\bf Q}_1{\bf Q}_1^T){\bf A}_k\|_2 + \|{\bf \Sigma}_\perp\|_2.
\end{aligned}
\end{equation}
In the last relation, the first term on the right-hand side results from the fact that $\mathcal{R}({\bf Q}_1) \subset \mathcal{R}({\bf Q})$, which for any matrix $\bf A$ yields \cite[Proposition 8.5]{HMT2011}:
\begin{equation}\notag
\|({\bf I} - {\bf QQ}^T){\bf A}\|_2 \le \|({\bf I} - {\bf Q}_1{\bf Q}_1^T){\bf A}\|_2.
\end{equation}
By substituting the bound in \eqref{eq_2n1ndTermR22} into \eqref{eq_LRAErr_proof}, the theorem follows for the case $q = 0$. The theorem for the case when the PI technique is used follows similarly. \QEDB
\section{Proof of Theorem \ref{Theorem_SinValueApprox}}
\label{Proo_SinValueApprox1}
First, we write ${\bf D}= {\bf Q}{\bf Q}^T{\bf D}$. It follows that
\begin{equation}\notag
\begin{aligned}
{\bf D}^T{\bf D} & = {\bf D}^T{\bf Q}{\bf Q}^T{\bf D}\\
{\bf D}^T{\bf D} & \succeq {\bf D}^T{\bf Q}_1{\bf Q}_1^T{\bf D}.
\end{aligned}
\end{equation}
By the Cauchy’s interlacing theorem, we therefore have
\begin{equation}\notag
\lambda_i({\bf A}^T{\bf A}) \ge \lambda_i({\bf D}^T{\bf D}) \ge \lambda_i({\bf D}^T{\bf Q}_1{\bf Q}_1^T{\bf D}).
\end{equation}
Furthermore, we have ${\bf Q}_1{\bf Q}_1^T = {\bf Q}_1^\prime {{\bf Q}_1^\prime}^T$. Thus, by replacing $\bf D$ and ${\bf Q}_1^\prime$, for the last term of the above equation we obtain
\begin{equation}\notag
\begin{aligned}
& {\bf D}^T{\bf Q}_1^\prime {{\bf Q}_1^\prime}^T{\bf D} = \bar{\bf P}{\bf A}^T{\bf Q}_1^\prime {{\bf Q}_1^\prime}^T{\bf A}\bar{\bf P}\\ &
= \bar{\bf P}{\bf V}^T
\begin{bmatrix}
{\bf \Sigma}_1\bar{\bf Z}^{-1}{\bf \Sigma}_1 & {\bf 0} & {\bf \Sigma}_1\bar{\bf Z}^{-1}{\bf Z}_1^T{\bf \Sigma}_3 \\
{\bf 0} & {\bf 0} & {\bf 0}\\
{\bf \Sigma}_3{\bf Z}_1 \bar{\bf Z}^{-1}{\bf \Sigma}_1 & {\bf 0} &
{\bf \Sigma}_3{\bf Z}_1 \bar{\bf Z}^{-1}{\bf Z}_1^T{\bf \Sigma}_3
\end{bmatrix}{\bf V}^T\bar{\bf P}.
\end{aligned}
\end{equation}
It is seen that ${\bf \Sigma}_1\bar{\bf Z}^{-1}{\bf \Sigma}_1 $ is a submatrix of ${\bf D}^T{\bf Q}_1^\prime {{\bf Q}_1^\prime}^T{\bf D}$. By replacing $\bar{\bf Z}^{-1}$, we thus obtain the following relation:
\begin{equation}\notag
\begin{aligned}
\lambda_i({\bf A}^T{\bf A}) \ge \lambda_i({\bf D}^T{\bf D}) & \ge \lambda_i({\bf D}^T{\bf Q}_1{\bf Q}_1^T{\bf D})\\
& \ge \lambda_i({\bf \Sigma}_1({\bf I} + {\bf Z}_1^T{\bf Z}_1)^{-1}{\bf \Sigma}_1).
\end{aligned}
\end{equation}
By applying the properties of partial ordering, it follows that
\begin{equation}\notag
\begin{aligned}
{\bf Z}_1^T{\bf Z}_1 & \preceq \sigma_{d-p+1}^{4q+4}\|{\bf \Phi}_2\|_2^2\|{\bf \Phi}_1^\dagger\|_2^2 {\bf \Sigma}_1^{-(4q+4)} \\
& = {\bf \Delta}^{4q+4}\|{\bf \Phi}_2\|_2^2\|{\bf \Phi}_1^\dagger\|_2^2,
\end{aligned}
\end{equation}
where ${\bf \Delta} = \text{diag}(\delta_1, ..., \delta_k)$ is a $k \times k$ matrix with entries $\delta_i= \frac{\sigma_{d-p+1}}{\sigma_i}$. In addition, the following relation holds
\begin{equation}\notag
{\bf \Sigma}_1({\bf I} + {\bf Z}_1^T{\bf Z}_1)^{-1}{\bf \Sigma}_1 \succeq {\bf \Sigma}_1({\bf I} + {\bf \Delta}^{4q+4}\|{\bf \Phi}_2\|_2^2\|{\bf \Phi}_1^\dagger\|_2^2)^{-1}{\bf \Sigma}_1,
\end{equation}
which results in
\begin{equation}\notag
\begin{aligned}
\lambda_i({\bf A}^T{\bf A}) \ge \lambda_i({\bf D}^T{\bf D}) & \ge \lambda_i({\bf \Sigma}_1({\bf I} + {\bf Z}_1^T{\bf Z}_1)^{-1}{\bf \Sigma}_1) \\ & \ge \frac{\sigma_i^2}{1 + \delta_i^{4q+4}\|{\bf \Phi}_2\|_2^2\|{\bf \Phi}_1^\dagger\|_2^2}.
\end{aligned}
\end{equation}
By taking the square root of the last identity, the theorem
follows. \QEDB
\section{Proof of Theorem \ref{Th_SinValueApprox2}}
\label{Proo_SinValueApprox2}
According to Theorem 2.1 of \cite{HuckabyChan05}, the singular values of ${\bf L}_{11}$ and $\bf D$ are related as follows:
\begin{equation} \label{eq_sL11si}
\begin{aligned}
\frac{\sigma_i({\bf L}_{11})}{\sigma_i({\bf D})} \ge 1 - \mathcal{O}\Bigg(\frac{\|{\bf L}_{12}\|_2^2}{(1-\rho^2)\sigma_k^2({\bf L}_{11})}\Bigg), \quad i=1,..., k.
\end{aligned}
\end{equation}
We also have the following relation:
\begin{equation}\label{eq_quotientMult}
\begin{aligned}
\frac{\sigma_i({\bf L}_{11})}{\sigma_i}= \frac{\sigma_i({\bf L}_{11})}{\sigma_i({\bf D})} \times \frac{\sigma_i({\bf D})}{\sigma_i}.
\end{aligned}
\end{equation}
By substituting bounds \eqref{eq_Th_SinValD_bound} and \eqref{eq_sL11si} into \eqref{eq_quotientMult}, the theorem follows.\QEDB
\section{Proof of Theorem \ref{Th_boundsA'ssigmak}}
\label{Proo_boundsA'ssigmak}
Let ${\bf A}_k$ and $\hat{\bf A}_k$ be rank-$k$ approximations of $\bf A$ constructed by the SVD and PbP-QLP, respectively. From the perturbation theory for singular values
\cite[Corollary 7.3.5]{HornJohnson852012}, for $i=1,...,k$, we have
\begin{equation}\notag
\begin{aligned}
|\sigma_i - \sigma_i({\bf L}_{11})| \le \|{\bf A}_k - \hat{\bf A}_k\|_2.
\end{aligned}
\end{equation}
Writing ${\bf A}_k={\bf U}_k{\bf U}_k^T{\bf A}$, $\hat{\bf A}_k={\bf Q}_1{\bf Q}_1^T{\bf A}$, and applying Theorem \ref{Th_distUkVkQ1P1}, we obtain
\begin{equation}\notag
|\sigma_i - \sigma_i({\bf L}_{11})| \le \|{\bf A}\|_2\text{dist}(\mathcal{R}({\bf U}_k), \mathcal{R}({\bf Q}_1)),
\end{equation}
from which the theorem follows. \QEDB
\section{Proof of Theorem \ref{Th_SinVal_PbP}}
\label{Proo_SinVal_PbP}
To prove this theorem, we use a key result from \cite[Theorem 5.8]{Gu2015}, stated with our notation as follows:
\begin{equation}\label{eq_HighProGu}
\mathbb P\{\|{\bf \Phi}_2\|_2\|{\bf \Phi}_1^\dagger\|_2\ge C_\Upsilon\}\le \Upsilon.
\end{equation}
To prove \eqref{eq_HighProB_R22}, we observe that \eqref{eq_2nR22_bound} is simplified to
\begin{equation}\notag
\begin{aligned}
\|{\bf R}_{22}\|_2 \le \sigma_{k+1} + {\delta_k^{2q+2}\sigma_1\|{\bf \Phi}_2\|_2\|{\bf \Phi}_1^\dagger\|_2}.
\end{aligned}
\end{equation}
Substituting the bound in \eqref{eq_HighProGu} into the above equation gives the desired result. The bounds \eqref{eq_HighProB_SVs} and \eqref{eq_HighProB_LRAE} are likewise obtained, by plugging \eqref{eq_HighProGu} into \eqref{eq_sigmaL11sigmai} and \eqref{eq_2nLRAErrbound}, respectively. \QEDB
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Hybrid materials composed of organic molecules deposited on inorganic substrates have been intensively investigated in the last decade due to their potential for optoelectronic applications.~\cite{Kubatkin2003,Moth-Poulsen2009,nagata2013apl, feng2014jmca, schulz2014afm} Depending on the nature of the constituents, different types of chemical bonds can be formed at the interface.~\cite{brau+09am} Covalent interactions, corresponding to molecular chemisorption, which are typically obtained in the presence of anchor groups, give rise to large degrees of interfacial charge transfer and to pronounced electronic hybridization.~\cite{calz+12jpcc,hofmann2013jcp,lang+14afm,timp+2014cm,turk+19ats,matt+20jpcc} The resulting hybrid system exhibits new characteristics compared to those of its individual building blocks. On the other hand, when molecules are \textit{physisorbed} on a surface, they preserve their intrinsic features to a large extent. Nonetheless, their electronic and optical properties undergo significant variations, including band-gap renormalization~\cite{garc+09prb,della+11prl,pusc+12prb,desp+13prb,nabo+19cm} and reduction of exciton binding energies.~\cite{desp+13prb,fu+17pccp,kerf+18jcp}
Capturing these effects with the required accuracy represents one of the main challenges for modern electronic structure theory.~\cite{draxl2014acr} Many-body perturbation theory (MBPT), including the $GW$ approximation and the solution of the Bethe-Salpeter equation, is the state-of-the-art first-principles method to compute quasi-particle band structures and optical absorption spectra including correlated electron-hole pairs (\textit{i.e.}, excitons).~\cite{onid+02rmp} Unfortunately, the computational costs of these calculations remain prohibitive for realistic materials described by hundreds of atoms, in spite of the remarkable efforts devoted lately to scale up the performance of MBPT calculations.~\cite{intr1,intr2,intr3,intr4,intr5} Therefore, the \textit{ab initio} treatment of complex systems including nanostructured substrates and/or mesoscopic elements, such as nanoparticles and nanotips, remains unfeasible with purely atomistic schemes.
A viable workaround for the aforementioned scenarios is provided by implicit schemes that capture \textit{effectively} the electrostatic interactions between photo-active systems and their environment.
Such methods have been applied for decades in computational chemistry to describe molecules in solution. The polarizable-continuum model (PCM)~\cite{tomasi2005cr} and the conductor-like screening model~\cite{cosmo} are successful examples of approaches that have been specifically tailored to deal with isotropic (liquid) media surrounding an active solute. More recently, embedding schemes coupled to MBPT calculations have been proposed and successfully applied~\cite{intr10,duch+18cs,wehn+18jctc,tiri+20jcp,kshi+20jctc} even in the context of periodic systems.~\cite{intr11,intr12,intr13}
However, these advances were devised to deal with molecules in solution or on bulk substrates, but not with substrates consisting of multiple stacked confined layers.
The advent of two-dimensional (2D) materials~\cite{novo+04sci,sple+10nl} and their heterostructures,~\cite{geim-grig13nat} which are known to feature also complex patterns,~\cite{zhan+17sciadv,pan+18nl} have made this scenario very realistic, paving the way for hybrid materials with nano-engineered inorganic substrates.~\cite{intr6,intr7,intr8,intr9}
Suitable theoretical methods that are able to treat this complexity at affordable computational costs are therefore needed for substantial advances in the field.
In this work, we present an extension of the PCM coupled to real-time time-dependent density-functional theory (RT-TDDFT)\cite{liang+2012jpca, nguyen+2012jpcl, pipolo+2014ctc} for an efficient and yet accurate description of the electrostatic interactions between quantum-mechanically modelled physisorbed molecules and implicit dielectric substrates. Envisioning the application of this method to stacked 2D materials, we allow the substrate to feature a layered structure exhibiting in-plane/out-of-plane anisotropy. We demonstrate that the developed formalism can be employed to calculate the polarization-induced renormalization of frontier energy levels of the adsorbate. For simulations of laser-induced ultrafast dynamics with RT-TDDFT, we additionally supply a way of calculating electric fields that arise from Fresnel-reflection at the substrate. We show that, depending on the complexity of the layered substrate, these reflected fields can assume non-trivial shapes and profoundly affect the electron dynamics induced in the molecule. As an example, we consider a substrate containing a deep-lying dielectric mirror, which gives rise to a delayed field after irradiation with a laser pulse. This leads to interference effects in the coherent dynamics of the molecule, similar to those of multi-pulse-based spectroscopy.
We finally compare our numerical results with those obtained from an exactly solvable two-level model, in order to disclose the dynamics of the excited electrons and the role of polarization effects therein.
This paper is organized as follows. In Section~\ref{pcm.sec}, we provide a brief review of the non-equilibrium flavor of the PCM in the integral-equation formalism, laying out the essential equations. On this basis, we present our extension of the model in Section~\ref{deriv.sec}, together with the proposed method of calculating the Fresnel-reflected fields. In Section~\ref{impl.sec}, we review how the PCM is interfaced with TDDFT, and provide computational details. Finally, in Section~\ref{results.sec}, we report numerical results that verify our implementation, and showcase different types of substrate-related effects that can be described with the presented method. Conclusions are given in Section~\ref{conclu.sec}.
\section{Background: The polarizable-continuum model}\label{pcm.sec}
The PCM poses a well-established way of coupling quantum-chemically modelled molecules to macroscopic polarizable environments.~\cite{tomasi2005cr} It was originally conceived to study the influence of electrostatic solute-solvent interactions on the properties of solvated systems.~\cite{onsager1936jacs, miertus+1981} In this framework, the solute molecule resides in a molecularly shaped cavity cut into an infinitely extended dielectric medium, which is typically characterized by its dielectric constant (see Fig.~\ref{fig.sketch}). The nuclear and electronic charge densities of the molecule polarize the surrounding medium, giving rise to polarization surface charge density $\sigma_0$ spread on the cavity boundary~$\Gamma$. These polarization charges cause an electrostatic field, the so-called \textit{reaction field}, that, in turn, polarizes the solute inside the cavity. The Schr{\"o}dinger equation for the solute within the reaction field and the Poisson equation for the molecular charge densities coupled to the continuum dielectric are solved self-consistently.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{sketch.pdf}
\caption{Sketch of the cavity, indicating the quantities that enter in the calculation of the PCM response function. $q_0$ stands for a free charge within the cavity that gives rise to the electrostatic potential $V_0$.}
\label{fig.sketch}
\end{figure}
For a given molecular charge density that generates the electrostatic potential $V_0$, the polarization surface charge density is calculated as
\begin{align}\sigma_0(\textbf{s}) = \int_{\Gamma}\text ds'\,{\cal Q}_0(\textbf{s},\textbf{s}' )V_0(\textbf{s}'),
\label{asc_static.eq}\end{align}
where ${\cal Q}_0$ is the PCM response function, connecting the potential at the cavity point $\textbf{s}'$ with a polarization-induced apparent surface charge $\sigma_0(\textbf{s})\text ds$ at the cavity point $\textbf{s}$ (see Fig.~\ref{fig.sketch}). The combined effect of all these charges gives rise to the electrostatic reaction potential
\begin{align}\label{reaction_pot.eq}
V_{\text{R}}(\textbf{r}) = \int_\Gamma \frac{\sigma_0(\textbf{s})\text ds}{|\textbf{r}-\textbf{s}|}.
\end{align}
Applying Eqs.~\eqref{asc_static.eq} and \eqref{reaction_pot.eq} corresponds to solving Poisson's equation within the cavity with a "free" charge density given by the molecular electronic and nuclear charges, correctly taking into account all boundary conditions at the dielectric interfaces. As the next step, the charge density of the molecule subject to the field of Eq.~\eqref{reaction_pot.eq} is recalculated, and the new electrostatic potential, $V_0$, is inserted back into Eq.~\eqref{asc_static.eq}. This procedure is repeated until self-consistency is achieved.
Within the integral-equation formalism,~\cite{cances1997jcp, cances1998jmc} the response function ${\cal Q}_0$, which is the crucial quantity in the PCM, is calculated from the static dielectric constant of the solvent, $\varepsilon$, via
\begin{align}
{\cal Q}_0 = &{\cal Q}(\varepsilon) = \left\lbrace\left[2\pi I-D_e(\varepsilon)\right]S_i+S_e(\varepsilon)\left(2\pi I+D_i^*\right)\right\rbrace^{-1}\nonumber\\&\times\left\lbrace S_e(\varepsilon)S_i^{-1}\left(2\pi I+D_i\right)+\left[2\pi I+D_e(\varepsilon)\right]\right\rbrace\label{pcmmat.eq}
\end{align}
where $I$ is the identity operator, and $S_{i}$, $D_{i}$ and $D^*_i$, as well as $S_{e}$ and $D_{e}$ are integral operators, defined by their action on an function $u$ supported on the cavity surface $\Gamma$ as\cite{cances1998jmc}
\begin{subequations}\label{eq.internal_matrices}
\begin{align}
\left(S_{i}u\right)(\textbf{s}) &= \int_\Gamma\text ds'\, G_{i}(\textbf{s},\textbf{s}')u(\textbf{s}') \label{simat.eq}\\
\left(D_{i}u\right)(\textbf{s}) &= \int_\Gamma\text ds'\left[\hat{\textbf{n}}(\textbf{s}')\cdot \nabla' G_{i}(\textbf{s},\textbf{s}')\right]u(\textbf{s}') \label{dimat.eq}\\
\left(D^*_iu\right)(\textbf{s}) &= \int_\Gamma\text ds'\left[\hat{\textbf{n}}(\textbf{s})\cdot \nabla G_{i}(\textbf{s},\textbf{s}')\right]u(\textbf{s}')\label{dimatstar.eq},
\end{align}
with the electrostatic Green's function for vacuum, $G_i(\textbf{r}, \textbf{r}')=1/|\textbf{r}-\textbf{r}'|$, and
\end{subequations}
\begin{subequations}\label{eq.external_matrices}
\begin{align}
\left[S_{e}(\varepsilon)u\right](\textbf{s}) &= \int_\Gamma\text ds'\, G_{e}(\textbf{s},\textbf{s}';\varepsilon)u(\textbf{s}') \label{semat.eq}\\
\left[D_{e}(\varepsilon)u\right](\textbf{s}) &= \varepsilon_{\mathrm{solv}}\int_\Gamma\text ds'\left[\hat{\textbf{n}}(\textbf{s}')\cdot \nabla' G_{e}(\textbf{s},\textbf{s}';\varepsilon)\right]u(\textbf{s}').\label{demat.eq}
\end{align}
\end{subequations}
Here, $\hat{\textbf{n}}(\textbf{s})$ is the normal vector at the surface point $\textbf{s}$, and $\varepsilon_{\mathrm{solv}}$ is the dielectric constant of the medium encompassing the cavity (see Fig.~\ref{fig.sketch}). $G_e$, appearing in Eqs.~\eqref{eq.external_matrices}, solves Poisson's equation for a free point charge in the environment, disregarding the response of the cavity surface itself. The application of $S_x$ and $D_x$ operators to a function $u$ thus yields the electrostatic potential of a polarization surface charge density $u$ and a surface dipole density $u\hat{\textbf{n}}$, respectively, \textit{in vacuo} ($x=i$) or in the environment ($x=e$). For an isotropic solvent with dielectric constant $\varepsilon_{\mathrm{solv}}$, $G_e(\textbf{r}, \textbf{r}';\varepsilon)=G_e^0(\textbf{r}, \textbf{r}';\varepsilon_{\mathrm{solv}})=1/\varepsilon_{\mathrm{solv}} |\textbf{r}-\textbf{r}'|$. In the generalized case, $\varepsilon$ represents all dielectric constants characterizing the environment.
While the formalism is based on electrostatics, it can be adopted for time-dependent calculations as well. The short distances between solute and surrounding solvent molecules justify a quasi-static treatment of the dynamical electromagnetic interactions, \textit{i.e.}, the neglect of retardation. In this case, however, it is not very accurate to characterize the solvent by its dielectric constant alone. Often, a non-equilibrium version~\cite{marcus1956jcp, aguilar+1993jcp, cammi+tomasi1995ijqc, menucci+jcp1998} of time-dependent PCM is employed, which additionally includes the optical dielectric constant of the solvent, $\varepsilon_d$. Time scales are separated: The inert rotational degrees of freedom of the solvent are associated with the static dielectric constant, whereas the fast polarization response of the electrons in the solvent is described by $\varepsilon_d$. In ultrafast dynamical processes such as vertical ionizations or optical excitations, this non-equilibrium approach assumes that the electronic degrees of freedom are able to instantaneously follow the rapidly varying electrostatic forces generated by the charge densities of the solute, while the rotational (and in general the nuclear) solvent degrees of freedom remain fully frozen. The time-dependent apparent surface charge density is then given by~\cite{aguilar+1993jcp, cammi+tomasi1995ijqc, menucci+jcp1998, ingrosso+2003jml, caricato+2005jcp}
\begin{align}\sigma(\textbf{s},t) &= \sigma_0(\textbf{s})+\int_{\Gamma}\text ds'\,{\cal Q}_d(\textbf{s},\textbf{s}') [V(\textbf{s}',t)-V_0(\textbf{s}')],
\label{asc.eq}\end{align}
where $V(t)$ is the instantaneous electrostatic potential caused by the molecule in its dynamical perturbed state, and ${\cal Q}_d$ is the dynamical response function, calculated as ${\cal Q}_d$~=~${\cal Q}(\varepsilon_d)$ with the same prescription as in Eq.~\eqref{pcmmat.eq}. Eq.~\eqref{asc.eq} is valid as long as the solvent has no optical excitations in the considered frequency range. For a more general treatment, one should account for the frequency-dependent and complex-valued nature of the dielectric function of the solvent,~\cite{ding2015jcp, corni2015jpca, pipoleCorni2016jpcc} which, however, is not straightforward and goes beyond the scope of the present work.
While originally devised for molecules immersed in isotropic solvents, the reformulation of PCM within the integral-equation formalism\cite{cances1997jcp, cances1998jmc} allows for the overhead-free featuring of more general dielectric environments, such as anisotropic\cite{cances1997jcp} and ionic\cite{mennucci+1997jpcb, cances1998jmc} solvents, metal surfaces,~\cite{corni2003jcp} close-by nanoparticles,~\cite{corni2001jcp, delgado2013jcp} and interfaces between different fluid phases.~\cite{frediani2004jcp,frediani2004jpcb}
The Green's function for the interior of the cavity, $G_i$, remains unchanged, whereas the one for the environment, $G_e$, assumes a different form. In Section~\ref{deriv.sec}, we determine $G_e$ for a stack of anisotropic dielectric layers.
\section{Electromagnetic response of the substrate}\label{deriv.sec}
In this section, we derive the equations for the description of molecule-substrate-solvent interactions (Section~\ref{green.sec}) and of the reflection behavior of an electromagnetic plane wave impinging on the substrate (Section~\ref{fresnel.sec}). The whole environment of the thiophene ring considered in this example is built up as sketched in Fig.~\ref{env.fig}. The molecule is immersed in a solvent with dielectric constant $\varepsilon_{\mathrm{solv}}$, residing in a vacuum cavity with boundary $\Gamma$ (region I). The vacuum cavity represents the solvent-inaccessible surrounding of the thiophene ring.~\cite{tomasi2005cr} The layered substrate is characterized by the origin and normal vector of its surface, by the perpendicular and parallel components of the dielectric constant, $\varepsilon_{n,\perp}$ and $\varepsilon_{n,\parallel}$, respectively, as well as by the thicknesses $d_n$ of the $N$ intermediate layers (region II). The semi-infinite bottom layer is characterized by dielectric constants $\varepsilon_{N+1,\parallel}$ and $\varepsilon_{N+1,\perp}$ (region III). In the following, we refer to the parallel direction ($\parallel$) as ``in-plane", and to the perpendicular one ($\perp$) as ``out-of-plane".
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{environment_bw.pdf}
\caption{Sketch of the dielectric environment of a thiophene molecule residing in a cavity with surface $\Gamma$. The molecule is surrounded by a solvent with dielectric constant $\varepsilon_\mathrm{solv}$. The substrate is formed by $N$ layers of thickness $d_n$ between the coordinates $z=z_{n-1}$ and $z=z_{n}$, and has in-plane and out-of-plane dielectric constants $\varepsilon_{n,\parallel}$ and $\varepsilon_{n,\perp}$, respectively. The final layer ($N+1$) extends down to $z=-\infty$.}
\label{env.fig}
\end{figure}
\subsection{The Green's function of the environment}\label{green.sec}
We make use of the integral-equation formalism of the PCM to calculate the reaction field stemming from the polarization of the substrate and the solvent. The missing ingredient is the electrostatic Green's function of the environment, $G_e$, to be inserted into Eqs.~\eqref{eq.external_matrices}, which depends parametrically on the sets of dielectric constants $\{\varepsilon\}$, on the layer thicknesses $\{d\}$, and on the shape and position of the cavity. To keep the notation simple, we omit these dependencies in the following. The dynamical PCM response function, ${\cal Q}_d$, to be plugged into Eq.~\eqref{asc.eq} for non-equilibrium time-dependent calculations, is obtained by replacing $\{\varepsilon\}$ by the set of optical dielectric constants, $\{\varepsilon_d\}$.
The environmental Green function, $G_e$~=~$G_e(\textbf{r},\textbf{r}')$, is the total electrostatic potential at a point~$\textbf{r}$ generated by a unit point charge situated at~$\textbf{r}'$, ignoring the electrostatic response of the vacuum cavity. From Eqs.~\eqref{reaction_pot.eq}, \eqref{pcmmat.eq}, and \eqref{eq.external_matrices}, it is evident that this function is ultimately evaluated for $\textbf{r}$ and $\textbf{r}'$ lying on the cavity surface $\Gamma$; thus, both these vectors can be assumed to be within region I (see Fig.~\ref{env.fig}). The electrostatic potential in this region is the sum of the direct solvent-screened Coulomb potential and of a second term stemming from the polarization of the substrate, $G_{\text{img}}$, such that:
\begin{align}
G_e(\textbf{r}, \textbf{r}') = G_e^0(\textbf{r}, \textbf{r}')+G_{\text{img}}(\textbf{r},\textbf{r}').\label{solv.eq}
\end{align}
By definition, this function solves Poisson's equation, which, restricting ourselves to the solvent region, reads
\begin{align}
-\varepsilon_{\mathrm{solv}}\nabla^2G_e(\textbf{r}, \textbf{r}')= \delta(\textbf{r}-\textbf{r}').\label{poisson.eq}
\end{align}
From Eqs. \eqref{solv.eq} and \eqref{poisson.eq}, it is clear that the polarization term satisfies Laplace's equation,
\begin{align}
-\nabla^2G_{\mathrm{img}}(\textbf{r}, \textbf{r}')= 0.\label{laplaceSolv.eq}
\end{align}
Extending previous work within the PCM,~\cite{corni2003jcp, frediani2004jcp,frediani2004jpcb} the following derivations exploit the cylindrical symmetry of the dielectric environment, which implies that $G_e(\textbf{r}, \textbf{r}')$ depends on the in-plane variables $x$, $x'$, $y$, and $y'$ only via $\Delta r_\parallel$ = $|\Delta\textbf{r}_\parallel|$, where $\Delta\textbf{r}_\parallel= (x-x',y-y')$.
Applying a partial spatial Fourier transform $\textbf{r}_\parallel\rightarrow\textbf{q}_\parallel$, we can write\cite{corni2002jcp}
\begin{align}
G_e(\textbf{r},\textbf{r}') = \frac{1}{2\pi}\int_0^\infty\text dq_\parallel\,q_\parallel J_0(q_\parallel\Delta r_\parallel)G(q_\parallel, z, z'),\label{greenPartial.eq}
\end{align}
where $J_0$ is the zeroth-order Bessel function of first kind.~\cite{temme1996} Due to in-plane isotropy, the transformed Green's function, $G(q_\parallel, z, z')$, depends only on the absolute value~$q_\parallel$ of the in-plane wave vector~$\textbf{q}_\parallel$.
By transformation of Eq.~\eqref{solv.eq}, we obtain\cite{fuchs1981prb}
\begin{align}
G_e(q_\parallel, z, z') = \frac{2\pi}{\varepsilon_{\mathrm{solv}} q_\parallel}e^{-q_\parallel|z-z'|}+G_{\text{img}}(q_\parallel, z, z'),
\end{align}
where $G_{\text{img}}(q_\parallel, z, z')$ satisfies the transformed version of Eq.~\eqref{laplaceSolv.eq},
\begin{align}
\left(-q_\parallel^2+\frac{\partial^2}{\partial z^2}\right)G_{\text{img}}(q_\parallel, z, z') = 0.\label{greenSolvent.eq}
\end{align}
Eq.~\eqref{greenSolvent.eq} is evidently solved by
\begin{align}
G_{\text{img}}(q_\parallel, z, z') = B_0(q_\parallel,z')\,e^{-q_\parallel(z-z_0)},\label{imageFourier.eq}
\end{align}
where the exponentially increasing solution has been discarded, as it diverges for $z\rightarrow\infty$. To simplify later expressions, the zero of $z$ is set to the surface at $z_0$. All electrostatic effects due to the polarization of a specific substrate are contained in the constant of integration, $B_0(q_\parallel,z')$. We obtain the latter by enforcing appropriate boundary conditions at the surface and at all underlying interfaces in the substrate. To this end, we need to know the electrostatic potential inside the layers. With the assumed in-plane/out-of-plane anisotropy, the dielectic tensor in layer $n$ in region II is given by $\varepsilon_n = \text{diag}\{\varepsilon_{n,\parallel},\varepsilon_{n,\parallel},\varepsilon_{n,\perp}\}$, and the electrostatic potential satisfies a generalized Laplace's equation,
\begin{align}
\left(-\varepsilon_{n,\parallel} q_\parallel^2+\varepsilon_{n,\perp}\frac{\partial^2}{\partial z^2}\right)G_{\text{e},n}(q_\parallel,z,z') &= 0,\label{anisoLaplace.eq}
\end{align}
which is a direct consequence of Gauss' law, $\nabla\cdot\textbf{D} = \nabla\cdot(\varepsilon\textbf{E})= 0$. Eq. \eqref{anisoLaplace.eq} is solved by
\begin{align}
G_{\text{e},n}(q_\parallel,z,z') &= A_ne^{\xi_n q_\parallel (z-z_{n-1})}+ B_ne^{-\xi_n q_\parallel (z-z_{n-1})}, \label{greenSubstrate.eq}
\end{align}
where $\xi_n$~=~$(\varepsilon_{n,\parallel}/\varepsilon_{n,\perp})^{1/2}$ is a measure of the in-plane/out-of-plane anisotropy, and $z_{n-1}$ marks the interface in the direction of the surface.
The next step towards calculating $B_0(q_\parallel,z')$ [see Eq.~\eqref{imageFourier.eq}] consists of connecting the potentials in Eq.~\eqref{greenSubstrate.eq} at the interfaces inside the substrate according to the boundary conditions $\delta\textbf{E}_\parallel$~=~$0$ and $\delta\textbf{D}_\perp$~=~$0$, which are derived from the basic electrostatic equations $\nabla\times\textbf{E}$~=~$0$ and $\nabla\cdot\textbf{D}$~=~$0$, respectively. Since its curl vanishes, the electric field $\textbf{E}$ is the gradient of the potential~$G_\text{e}$, and the boundary conditions are equivalent to $\delta G_\text{e}(q_\parallel, z, z')$~=~$0$ and $\delta [\varepsilon_\perp\partial G_\text{e}(q_\parallel, z, z')/\partial z]$~=~$0$, where the $\delta$-variation is related to the $z$~coordinate. We consider the interface between the layers $n$ and $n+1$, which is located at $z_n<z_0$. Thus, the continuity condition, $\delta G(q_\parallel, z, z') = 0$, corresponds to
\begin{subequations}\label{eq.cont_and_deriv}
\begin{align}
A_ne^{-\xi_nq_\parallel d_n}+ B_ne^{\xi_n q_\parallel d_n} = A_{n+1}+ B_{n+1},\label{cont}
\end{align}
where $d_n = z_{n-1}-z_n$ is the thickness of the layer $n$. The second condition,
$\delta [\varepsilon_\perp\partial G(q_\parallel, z, z')/\partial z] = 0$, reads
\begin{align}
\xi_n\varepsilon_{n,\perp}\left(A_ne^{-\xi_nq_\parallel d_n}- B_ne^{\xi_n q_\parallel d_n}\right) = \xi_{n+1}\varepsilon_{n+1,\perp}\left(A_{n+1}- B_{n+1}\right).\label{deriv}
\end{align}
\end{subequations}
Eqs. \eqref{eq.cont_and_deriv} constitute a linear system that can be rewritten in matrix form as
\begin{align}
\begin{pmatrix}
A_n\\B_n
\end{pmatrix}
=
{\cal D}_n(q_\parallel)
\begin{pmatrix}
A_{n+1}\\B_{n+1}
\end{pmatrix},
\end{align}
where
\begin{widetext}
\begin{align}
{\cal D}_n(q_\parallel) =
\frac{1}{2\xi_n\varepsilon_{n,\perp}}\begin{pmatrix}
(\varepsilon_{n,\perp}\xi_n+\varepsilon_{n+1,\perp}\xi_{n+1})e^{q_\parallel\xi_n d_n} &
(\varepsilon_{n,\perp}\xi_n-\varepsilon_{n+1,\perp}\xi_{n+1})\,e^{q_\parallel\xi_n d_n}
\\
(\varepsilon_{n,\perp}\xi_n-\varepsilon_{n+1,\perp}\xi_{n+1})\,e^{-q_\parallel\xi_n d_n} &
(\varepsilon_{n,\perp}\xi_n+\varepsilon_{n+1,\perp}\xi_{n+1})e^{-q_\parallel\xi_n d_n}
\end{pmatrix}.
\end{align}
\end{widetext}
The matrix ${\cal D}_n$ connects the potentials in the $n$'th and $(n+1)$'th layers. The coefficients in the first layer, $n$~=~1, and in the final semi-infinite layer, $n$~=~$N+1$, are thus connected via
\begin{align}
\begin{pmatrix}
A_1\\B_1
\end{pmatrix}
=
{\cal T}(q_\parallel)
\begin{pmatrix}
A_{N+1}\\0
\end{pmatrix}, \label{transferMatrixProp.eq}
\end{align}
where the \textit{transfer matrix} is defined as
\begin{align}
{\cal T}(q_\parallel) = \prod_{n=1}^{N}{\cal D}_n(q_\parallel).
\end{align}
In Eq.~\eqref{transferMatrixProp.eq}, we have made explicit that the coefficient $B_{N+1}$ vanishes in region III, since it corresponds to an unphysical exponentially rising solution for $z\rightarrow-\infty$.
Finally, we consider the boundary conditions at the surface, \textit{i.e.} at the interface between regions I and II, where $z=z_0$. From Eqs. \eqref{imageFourier.eq} and \eqref{greenSubstrate.eq} for $n=1$, we have
\begin{subequations}\label{eq.surface_boundary}
\begin{align}
\frac{2\pi}{\varepsilon_{\mathrm{solv}}q_\parallel}e^{q_\parallel(z_0-z')}+B_0 &= A_1 + B_1\label{surfaceBoundary1.eq}\\
2\pi e^{q_\parallel(z_0-z')}-{\varepsilon_0}q_\parallel B_0 &= q_\parallel\varepsilon_1^\perp\xi_1(A_1 - B_1).\label{surfaceBoundary2.eq}
\end{align}
\end{subequations}
One of the constants, $A_1$ or $B_1$, can be eliminated by using Eq.~\eqref{transferMatrixProp.eq}, which implies $A_1/B_1 = {\cal T}_{11}/{\cal T}_{21}$. Consequently, Eqs. \eqref{eq.surface_boundary} form a linear system with two equations for two unknowns ($B_0$ and $A_1$ or $B_1$), which can be solved to yield the desired constant $B_0$, and thus an expression for the polarization term of the Green's function:
\begin{align}
G_{\text{img}}(q_\parallel,z,z') = \frac{2\pi}{\varepsilon_{\mathrm{solv}}q_\parallel}\frac{\varepsilon_{\mathrm{solv}}-\varepsilon_{\mathrm{subs}}(q_\parallel)}{\varepsilon_{\mathrm{solv}}+\varepsilon_{\mathrm{subs}}(q_\parallel)}e^{-q_\parallel(z'+z-2z_0)},\label{bZero.eq}
\end{align}
where $\varepsilon_{\mathrm{subs}}(q_\parallel) = \varepsilon_{1,\perp}\xi_1[\,{\cal T}_{11}(q_\parallel)-{\cal T}_{21}(q_\parallel)\,]/[\,{\cal T}_{11}(q_\parallel)+{\cal T}_{21}(q_\parallel)\,]$.
In order to build $D_e$ according to Eq.~\eqref{demat.eq}, the $\textbf{r}'$-gradient of the Green's function is required. With the polarization term in the form of Eq.~\eqref{greenPartial.eq} and the relations
\begin{subequations}
\begin{align}
\frac{\partial G_{\text{img}}(q_\parallel,z,z')}{\partial z'} &= -q_\parallel G_{\text{img}}(q_\parallel,z,z'),\\
\frac{\text dJ_0}{\text dx} &= -J_1(x),\\
\text{and}\hspace{0.5cm}\nabla'(\Delta r_\parallel) &= -\frac{\Delta\textbf{r}_\parallel}{\Delta r_\parallel} = -\Delta\hat{\textbf{r}}_\parallel,
\end{align}
\end{subequations}
where $J_1$ is the first-order Bessel function of first kind,~\cite{temme1996} we obtain\cite{corni2002jcp}
\begin{align}
\nabla' G_{\text{img}}(\textbf{r},\textbf{r}')
&=-\frac{1}{2\pi}\int_0^\infty\text dq_\parallel\,q_\parallel^2 G_{\text{img}}(q_\parallel, z, z')\nonumber\\
&\times\left[ J_0(q_\parallel\Delta r_\parallel)\hat{\mathbf{z}}-J_1(q_\parallel\Delta r_\parallel){\Delta\hat{\textbf{r}}_\parallel}\right].\label{nablaGreenPartial.eq}
\end{align}
Eqs.~\eqref{bZero.eq} and \eqref{nablaGreenPartial.eq} are used to calculate the electrostatic response functions ${\cal Q}_0$ and ${\cal Q}_d$.
\subsection{Fresnel reflection}\label{fresnel.sec}
\subsubsection{Reflectivity}
The electrons and nuclei polarize the substrate, causing an electric field that acts back on the molecule. If an external time-dependent electric field, \textit{e.g.} associated with electromagnetic radiation, acts on the system, the corresponding charge densities change in time, leading to a time-dependent reaction field. However, the external field interacts also directly with the layered substrate. At each interface, it is partially reflected, and the reflected beams interfere with the incident one. To determine the field acting locally on the molecule, these effects should be accounted for. In order to calculate the reflected fields, we make use once more of the transfer matrix formalism to calculate the frequency-dependent reflectivity of the substrate. This is common practice for anisotropic dielectric materials with arbitrary numbers of layers.~\cite{yeh1979josa, yeh1980ss, puschnig2006aem, vorwerk2016cpc, passler2017josab, passler2020prb} Full anisotropy requires dealing with 4$\times$4 matrices, as it mixes $s$- and $p$-modes at the interfaces between different dielectrics. In the present case of mere in-plane/out-of-plane anisotropy (uniaxial birefringence), the symmetry of the problem is still high enough to inhibit this type of mixing. Consequently, the 4$\times$4 matrix equations collapse into separate 2$\times$2 systems for $s$- and $p$-polarizations. Thus, we can avoid cumbersome numerical procedures and solve the problem analytically. Finally, with the knowledge of the frequency-dependent reflectivity, we can compute the total field as a superposition of the incident wave and of all partially reflected waves in real time.
The reaction field is calculated in the quasistatic limiting case, which amounts to the neglect of retardation. This approximation is known to be accurate at short distances, but cannot be applied to the external field: unlike the short-ranged field emitted by the molecule, the external, incident field is a plane wave, which is not attenuated over distance. Hence, it interacts also with deep substrate layers, whereas the substrate-molecule coupling is restricted to sub-nanometric depths within the surface. Hence, the full set of Maxwell's equations has to be considered. As the application of the transfer-matrix formalism to calculate reflectivities is well-established,~\cite{yeh1979josa, yeh1980ss, puschnig2006aem, vorwerk2016cpc, passler2017josab, passler2020prb} we present here only a reduced derivation of the formulas and refer the reader to the Appendix for further details.
The differential Maxwell's equations within a layer with in-plane/out-of-plane anisotropy can be cast into a Fourier-space wave equation
\begin{align}\label{wave.eq}
k^2\textbf{E}+\frac{\varepsilon_\perp-\varepsilon_\parallel}{\varepsilon_\parallel}\textbf{k}k_zE_z + \left(\frac{\omega}{c}\right)^2\varepsilon\textbf{E}= 0.
\end{align}
The condition for nontrivial solutions of the linear system in Eq.~\eqref{wave.eq} gives rise to a number of allowed modes, characterized by their wave vector $\textbf{k}$; the corresponding solutions $\textbf{E}$ determine the associated polarization directions of the field. We assume that $\textbf{k}=\textbf{k}_\mathrm{in}$ is known outside of the substrate. Its in-plane component, $\textbf{k}_\parallel$, remains unchanged throughout all layers. The determinant is a polynomial of 6th order in $k_z$ and, thus, has six independent roots. However, Gauss' law, $\nabla\cdot\textbf{D} = \nabla\cdot(\varepsilon\textbf{E}) = 0$, imposes additional constraints, resulting in only four possible wave vectors:
\begin{subequations}
\begin{align}
\textbf{k}^{(s)}_\pm &= \left(k_\parallel,0, \pm\sqrt{\left(\frac{\omega}{c}\right)^2\varepsilon_\parallel-k_\parallel^2}\right) = \left(k_\parallel,0, \pm k_z^{(s)}\right),\label{kS} \\
\textbf{k}^{(p)}_\pm &= \left(k_\parallel, 0, \pm\sqrt{\left(\frac{\omega}{c}\right)^2\varepsilon_\parallel-\frac{\varepsilon_\parallel}{\varepsilon_\perp}k_\parallel^2}\right) = \left(k_\parallel,0, \pm k_z^{(p)}\right)\label{kP},
\end{align}
\end{subequations}
where the coordinate system is rotated in the $x$-$y$ plane such that $\textbf{k}_\parallel$ points to the $x$ direction. We also introduce for convenience the $z$-direction wavenumbers, $k_z^{(s)}$ and $k_z^{(p)}$. The corresponding polarization vectors are
\begin{subequations}
\begin{align}
{\boldsymbol{\gamma}}^{(s)}_\pm &= \left(0, 1, 0\right)\label{pS}\\
{\boldsymbol{\gamma}}^{(p)}_\pm &\propto \left(1, 0, \mp\frac{\varepsilon_\parallel}{\varepsilon_\perp}\frac{k_\parallel}{k_z^{(p)}}\right).\label{pP}
\end{align}
\end{subequations}
The modes characterized by the wave vectors in Eq.~\eqref{kS} and by the polarizations in Eq.~\eqref{pS} are forward (+) and backward (-) propagating $s$-polarized waves, with polarization perpendicular to the plane of incidence. As the corresponding field is exclusively in-plane, these modes are not affected by the anisotropy. Consequently, they are purely transversal modes, \textit{i.e.}, $\nabla\cdot\textbf{E}\sim i\textbf{k}^{(s)}\cdot{\boldsymbol{\gamma}}^{(s)} = 0$. The polarization vectors of the $p$-polarized counterparts, Eq. \eqref{kP} and Eq. \eqref{pP}, lie in the plane of incidence. Thus, they have components in both in-plane and out-of-plane directions and are affected by the corresponding anisotropy. As a consequence, $\textbf{k}^{(p)}$ and ${\boldsymbol{\gamma}}^{(p)}$ are not mutually perpendicular. In other words, $p$-polarized waves induce space charge densities $\nabla\cdot\textbf{E}\neq 0$ in anisotropic layers and therefore gain a longitudinal component.
Since $s$- and $p$-polarizations are decoupled, they can be treated individually. For the $s$-polarization, we start from the electric field. In the Appendix, we show how the reflectivity of the whole layered structure, $r^{(s)}(\omega)$, can be calculated by matching the field amplitudes of adjacent layers at the interfaces, in accordance with the boundary conditions $\delta\textbf{E}_\parallel = 0$ and $\delta\textbf{B}_\parallel = 0$. The final result is
\begin{align}
r^{(s)}(\omega) = {\cal T}^{(s)}_{21}/{\cal T}^{(s)}_{11},\label{rS}
\end{align}
where the transfer matrix related to the electric field is defined as
\begin{align}\label{ts}
\left.
{\cal T}^{(s)} =
\begin{pmatrix}
k_{\mathrm{solv},z}^{(s)} & -1 \\
k_{\mathrm{solv},z}^{(s)} & 1
\end{pmatrix}
\right(\prod_{n=1}^{N}
{\cal D}_n^{(s)}
\left)
\begin{pmatrix}
1 & 1 \\
-k_{N+1,z}^{(s)} & k_{N+1,z}^{(s)}
\end{pmatrix},\right.
\end{align}
with
\begin{align}\label{ds}
{\cal D}_n^{(s)} =
\begin{pmatrix}
\cos(k_{n,z}^{(s)}d_n) & -i\sin(k_{n,z}^{(s)}d_n)/k_{n,z}^{(s)} \\
-ik_{n,z}^{(s)}\sin(k_{n,z}^{(s)}d_n) & \cos(k_{n,z}^{(s)}d_n)
\end{pmatrix}.
\end{align}
The case of a $p$-polarized field is more conveniently treated starting from the magnetic field. In this case, we end up with the electric-field reflectivity
\begin{align}
r^{(p)}(\omega) = -{\cal T}^{(p)}_{21}/{\cal T}^{(p)}_{11},\label{rP}
\end{align}
including the magnetic-field-related transfer matrix
\begin{widetext}
\begin{align}\label{tp}
\left.
{\cal T}^{(p)} =
\begin{pmatrix}
k_{\mathrm{solv},z}^{(p)}/\varepsilon_{\mathrm{solv}} & -1 \\
k_{\mathrm{solv},z}^{(p)}/\varepsilon_{\mathrm{solv}} & 1
\end{pmatrix}
\right(\prod_{n=1}^{N}
{\cal D}_n^{(p)}
\left)
\begin{pmatrix}
1 & 1 \\
-k_{N+1,z}^{(p)}/\varepsilon_{N+1,\parallel} & k_{N+1,z}^{(p)}/\varepsilon_{N+1,\parallel}
\end{pmatrix},
\right.
\end{align}
\end{widetext}
where
\begin{align}\label{dp}
{\cal D}_n^{(p)} =
\begin{pmatrix}
\cos(k_{n,z}^{(p)}d_n) & -i(k_{n,z}^{(p)}/\varepsilon_{n,\parallel})^{-1}\sin(k_{n,z}^{(p)}d_n) \\
-i(k_{n,z}^{(p)}/\varepsilon_{n,\parallel})\sin(k_{n,z}^{(p)}d_n) & \cos(k_{n,z}^{(p)}d_n)
\end{pmatrix}.
\end{align}
We continue by using the determined reflectivity to calculate from an arbitrary incident field the reflected field in time domain.
\subsubsection{Total field close to the surface}
The interaction between the molecule and the external electric field is usually treated in the dipole approximation, since the molecules are typically much smaller ($\sim$1~nm) than the wavelengths of interest (100-1000~nm), and higher multipole components couple to spatial variations of the field rather than to its absolute values. Within this approximation, the molecule is subject to a spatially uniform electric field. We write the incident field as
\begin{align}
\textbf{E}_{\text{in}}(t) = \hat{\boldsymbol{\gamma}}E_{\text{in}}(t),\label{incField.eq}
\end{align}
with the complex-valued polarization vector $\hat{\boldsymbol{\gamma}}$, and the real-valued temporal profile $E_{\text{in}}(t)$. From this, the time-dependence of the two components of the reflected field is calculated through causality-respecting convolutions
\begin{subequations}\label{eq.response_both}
\begin{align}
E_{\text r}^{(s)}(t) &= \int_{-\infty}^t\text dt'\,r^{(s)}(t-t')E_{\text{in}}(t')\label{response_s.eq}\\
E_{\text r}^{(p)}(t) &= \int_{-\infty}^t\text dt'\,r^{(p)}(t-t')E_{\text{in}}(t'),\label{response_p.eq}
\end{align}
\end{subequations}
where the response functions, $r^{(s)}(t-t')$ and $r^{(p)}(t-t')$, are obtained through inverse Fourier transformations of the reflectivities $r^{(s)}(\omega)$ and $r^{(p)}(\omega)$ [Eqs. \eqref{rS} and \eqref{rP}], respectively. The reflected fields of Eqs.~\eqref{eq.response_both} are weighted by the projection of the polarization vector of the incident field, $\hat{\boldsymbol{\gamma}}$, onto the $s$- and $p$-polarization vectors. The calculation of these polarization vectors requires the knowledge of the plane of incidence. Even though the spatial dependence of the field is neglected, the wavevector ${\textbf{k}}_{\text{in}}$ of the incident field is needed as an input, since the plane of incidence is spanned by ${\textbf{k}}_{\text{in}}$ and the surface normal vector $\hat{\textbf{n}}$. Since the direction of ${\textbf{k}}_{\text{in}}$ alone is sufficient, and this vector is perpendicular to the polarization $\hat{\boldsymbol{\gamma}}$, this corresponds to specifying a single additional parameter, which is the angle of incidence. We obtain
\begin{subequations}
\begin{align}
\hat{\boldsymbol{\gamma}}^{(s)}_{\text{in}} &= {\textbf{k}}_{\text{in}}\times\hat{\textbf{n}}/|{\textbf{k}}_{\text{in}}\times\hat{\textbf{n}}|,\label{incidenta.eq}\\
\hat{\boldsymbol{\gamma}}^{(p)}_{\text{in}} &= {\textbf{k}}_{\text{in}}/|\textbf{k}_{\text{in}}|\times\hat{\boldsymbol{\gamma}}^{(s)}_{\text{in}}.\label{incidentb.eq}
\end{align}
\end{subequations}
The reflected field propagates with the surface-reflected wavevector $\textbf{k}_\text{in}-2\hat{\textbf{n}}\left(\hat{\textbf{n}}\cdot{\textbf{k}}_{\text{in}}\right)$. The associated polarization vectors are
\begin{subequations}
\begin{align}
\hat{\boldsymbol{\gamma}}^{(s)}_r &= \hat{\boldsymbol{\gamma}}^{(s)}_{\text{in}}\label{reflecteda.eq}\\
\hat{\boldsymbol{\gamma}}^{(p)}_\text{r} &= -\left[{\textbf{k}}_{\text{in}}-2\hat{\textbf{n}}\left(\hat{\textbf{n}}\cdot{\textbf{k}}_{\text{in}}\right)\right]/|\textbf{k}_{\text{in}}|\times\hat{\boldsymbol{\gamma}}^{(s)}_{\text{in}}.\label{reflectedb.eq}
\end{align}
\end{subequations}
Having defined these quantities, we can write the total field acting on the molecule as
\begin{align}
\textbf{E}_{\text{tot}}(t) = \textbf{E}_{\text{in}}(t) + \hat{\boldsymbol{\gamma}}^{(s)}_\text{r}\left(\hat{\boldsymbol{\gamma}}^{(s)}_\text{in}\cdot\hat{\boldsymbol{\gamma}}\right)E_\text{r}^{(s)}(t)+\hat{\boldsymbol{\gamma}}^{(p)}_\text{r}\left(\hat{\boldsymbol{\gamma}}^{(p)}_\text{in}\cdot\hat{\boldsymbol{\gamma}}\right)E_\text{r}^{(p)}(t).
\end{align}
With this expression, it is implicitly assumed that $\textbf{r}$~=~0 for all fields, \textit{i.e.} the field is evaluated directly at the surface. Further away, the reflected field would have an additional delay related to the time it takes to travel from the molecule to the surface and back. This retardation effect, noticeable only for relatively large distances (on the order of visible light wavelengths), can be accounted for by including a corresponding vacuum layer on top of the substrate.
\section{Implementation}\label{impl.sec}
\subsection{Interface with density-functional theory}
The presented formalism can be interfaced easily with density functional theory\cite{hohenberg1964} (DFT) and its time-dependent extension\cite{rungegross} (TDDFT). DFT is a quantum-mechanical framework to determine ground-state electronic structure of a many-electron system. This is achieved by solving the Kohn-Sham equation,\cite{kohnsham1965}
\begin{align}\label{ks.eq}
\hat{\cal H}_{\text{KS}}[\rho_0]\psi_k(\textbf{r}) = E_k\psi_k(\textbf{r}),
\end{align}
for the Kohn-Sham orbitals $\psi_k(\textbf{r})$ and eigenvalues $E_k$. Eq.~\eqref{ks.eq} must be solved self-consistently, since the Kohn-Sham Hamiltonian depends on the solutions $\psi_k(\textbf{r})$ of the equation via the ground-state electron density
\begin{align}
\rho_0(\textbf{r}) = \sum_k^{\text{occ}}|\psi_k(\textbf{r})|^2.
\end{align}
The Kohn-Sham Hamiltonian \textit{in vacuo} is
\begin{align}\label{ks_ham.eq}
\hat{\cal H}_{\text{KS},0}[\rho_0] = -\frac{\nabla^2}{2}+ V_{\text{nuc}}(\textbf{r})+V_{\text{H}}[\rho_0](\textbf{r})+V_{\text{xc}}[\rho_0](\textbf{r}),
\end{align}
where $V_{\text{xc}}[\rho_0](\textbf{r})$ is the exchange-correlation potential. This term has to be approximated, since its exact form is unknown. $V_{\text{nuc}}(\textbf{r})$ and
\begin{align}
V_{\text{H}}[\rho_0](\textbf{r}) = \int\text d^3r'\,\frac{\rho_0(\textbf{r}')}{|\textbf{r}-\textbf{r}'|}
\end{align}
are the electrostatic potentials due to the nuclei and the electron cloud, respectively, up to a sign. They can thus be directly be inserted into Eq.~\eqref{asc_static.eq} via
\begin{align}
V_0(\textbf{s}') = -\lbrace V_{\mathrm{nuc}}(\textbf{s}')+V_H[\rho_0](\textbf{s}')\rbrace
\end{align}
to obtain $\sigma_0$, \textit{i.e.} the static polarization surface charge density. The polarizing effect of the reaction field on the solute is included by adding a corresponding electrostatic energy to Eq.~\eqref{ks_ham.eq},
\begin{align}\label{total_static_ham.eq}
\hat{\cal H}_{\text{KS}} =
\hat{\cal H}_{\text{KS},0}
-\int_{\Gamma}\text ds\,\frac{\sigma_0(\textbf{s})}{|\textbf{r}-\textbf{s}|}.
\end{align}
Including the polarization of the environment adds another layer of self-consistency to the calculation.
In TDDFT, the time-dependent Kohn-Sham equation
\begin{align}
i\frac{\partial}{\partial t}\psi_k(\textbf{r},t) = \hat{\cal H}_{\text{KS}}[\rho](t)\psi_k(\textbf{r},t),
\end{align}
is used to propagate in time the orbitals $\psi_k(\textbf{r},t)$ and the dynamical electron density
\begin{align}
\rho(\textbf{r},t) &= \sum_{k}^{\mathrm{occ}}|\psi_k(\textbf{r},t)|^2,
\end{align}
where the initial condition assumes that the system is in its ground state, \textit{i.e.}, $\psi_k(\textbf{r},t=0)$~=~$\psi_k(\textbf{r})$. In the adiabatic approximation, the full time-dependent Kohn-Sham Hamiltonian including the dynamical reaction field\cite{pipolo+2014ctc,corni2015jpca} is given by
\begin{align}
\hat{\cal H}_{\text{KS}}[\rho](t) = \hat{\cal H}_{\mathrm{KS},0}[\rho(t)]-\int_{\Gamma}\text ds\,\frac{\sigma(\textbf{s},t)}{|\textbf{r}-\textbf{s}|}+\textbf{r}\cdot\textbf{E}(t),
\end{align}
where $\sigma(t)$ is calculated from Eq.~\eqref{asc.eq} with
\begin{align}
V(\textbf{s}',t) &=
V_H[\rho(t)](\textbf{s}')+V_{\mathrm{nuc}}(\textbf{s}').
\end{align}
The coupling to the external electric field $\textbf{E}(t)$ is included within the dipole approximation. At this stage, we are not including the so-called cavity field effects.~\cite{pipolo+2014ctc, gil2019jctc}
For the analysis of the electron dynamics, we monitor the number of excited electrons,~\cite{krumland2020jcp}
\begin{align}\label{population.eq}
N_e(t) = N - \sum\limits_{i,j}^{\text{occ}}\left|\langle\psi_i(0)|\psi_j(t)\rangle\right|^2,
\end{align}
as well as the time-dependent dipole moment,
\begin{align}\label{dipole.eq}
\boldsymbol\mu(t) = - \int\text d^3r\, \textbf{r} \, \rho(\textbf{r}, t).
\end{align}
$N_e(t)$ and $\boldsymbol\mu(t)$ give indications about the excited-state populations and coherences of the many-electron system, respectively.~\cite{krumland2020jcp}
\subsection{Computational Details}\label{comp_details.sec}
The presented formalism has been implemented in a module named \textit{LayerPCM} in a locally modified version of the \textsc{Octopus} code (v9.0),~\cite{marques2003cpc, castro2006pssb, andrade2015pccp} incorporating the previous installment of PCM.~\cite{tancogne2020jctc, gil2019jctc, delgado2015jcp} Orbitals are represented on a real-space grid inside a simulation box formed by interlocking spheres centered at the atomic positions. For the first considered system, the He$^+$ ion (Section~\ref{benchmark.sec}), we set the radius to 9.45~bohr; for the second one, the thiophene molecule (1T, Section~\ref{thiophene.sec}), we use 11.34~bohr. We employ grid spacings of 0.28~bohr and 0.36~bohr for He$^+$ and 1T, respectively. For the exchange-correlation potential, we choose the (adiabatic) local-density approximation in the Perdew-Zunger parametrization.~\cite{ekar1984prb, perd1981prb} For $\Delta$self-consistent field ($\Delta$SCF) calculations, we use the corresponding spin-polarized formalism for the 1T$^+$ ion. In time-dependent runs, we employ the approximated enforced time-reversal symmetry propagator and a time step of 0.045~a.u.\cite{cast2004jcp}
The PCM cavities are also build from interlocking spheres.~\cite{pascual1990jcp} In the case of He$^+$, we center one sphere at (0,0,0), which is the position of the ion, and another one at (1~bohr,\,1~bohr,\,1~bohr), with radii of 2.27~bohr each. For 1T, we use a cavity formed by the union of spheres that are centered on the atomic positions; we use radii of 3.78~bohr for C, and 4.54~bohr for S. We do not put spheres on H based on the united-atom approach,~\cite{barone1997jcp} giving rise to a smoother cavity. All these values correspond to the respective van-der-Waals radii of the atoms scaled by a factor 1.2. The cavity surface $\Gamma$ is discretized into $N_t$ surface elements (\textit{tesserae}), characterized by their position $\textbf{s}_k$ and total charge $q_k$. The tessellation transforms the operators in Eqs.~\eqref{eq.internal_matrices} and \eqref{eq.external_matrices}, and thus the PCM response function [Eq.~\eqref{pcmmat.eq}], into square matrices. The reaction potentials become
\begin{align}
-\int_{\Gamma}\text ds\,\frac{\sigma(\textbf{s})}{|\textbf{r}-\textbf{s}|}\approx -\sum_{k=1}^{N_t}\frac{q_k}{|\textbf{r}-\textbf{s}_k|}.
\end{align}
We employ tessellation densities of 240/sphere and 60/sphere for He$^+$ and 1T, respectively. The total number of tesserae is reduced by discarding those that reside inside the union of the spheres.
The unbounded integrals of Eqs.~\eqref{greenPartial.eq} and \eqref{nablaGreenPartial.eq} are evaluated numerically with a 20-point Gauss-Laguerre quadrature.~\cite{abramowith1974}
\section{Results}\label{results.sec}
In this section, we showcase results that validate the developed approach and the corresponding numerical implementation, including benchmarks against analytical results (Section~\ref{benchmark.sec}). Subsequently, we apply the method to describe a realistic setup in the framework of TDDFT (Section~\ref{thiophene.sec}).
\subsection{He$^+$ ion}\label{benchmark.sec}
To put the developed method and its numerical implementation to the test, we choose the He$^+$ ion as a benchmark system. Due to its spherical symmetry, this cation can be assumed to be reasonably similar to an ideal point charge $+e$. It thus lends itself for comparisons to (semi-)analytical results from elementary electrostatics. As described in Section~\ref{comp_details.sec}, we employ as a cavity the union of two diagonally displaced spheres. Conventionally, a spherical cavity would be used for a single spherical ion; we employ an irregular one, instead, to check the robustness of the implementation also beyond high-symmetry cases. As long as the entire ionic charge density lies within the cavity and the environment is completely on the outside, the shape of the cavity should not matter.
As a first, textbook example, we consider the $+e$ charge to be within a medium of dielectric constant $\varepsilon_{\mathrm{solv}}$ (which we call \textit{solvent}), placed at a distance $d$ away from the interface to a second semi-infinite medium of dielectric constant $\varepsilon_{\mathrm{subs}}$ (\textit{substrate}). The charge polarizes the solvent and the substrate, giving rise to a reaction field that, in turn, stabilizes the charge. The interaction free energy, defined as the difference of the free energies obtained with and without substrate, is given as
\begin{align}\label{interactionDistance.eq}
{\cal G}_{\text{int}}(d) = -\frac{\varepsilon_{\mathrm{subs}}-\varepsilon_{\mathrm{solv}}}{\varepsilon_{\mathrm{subs}}+\varepsilon_{\mathrm{solv}}}\frac{1}{4\varepsilon_{\mathrm{solv}}d}.
\end{align}
To obtain a comparable quantity from DFT/PCM, we explicitly calculate the free energies in the solvent with and without the substrate, and subtract the resulting values. Looking at the interaction energy as a function of the distance $d$ from the substrate, we find perfect agreement with the analytical expression Eq.~\eqref{interactionDistance.eq} for different combinations of $\varepsilon_{\mathrm{solv}}$ and $\varepsilon_{\mathrm{subs}}$ [see Fig.~\ref{interactionEnergy.fig}a)]. This ideal match has a subtle implication. In the DFT/PCM calculation, the charge is spatially extended and must be hosted by a vacuum cavity. The point charge, on the other hand, does not have such a cavity around it. Thus, there is an additional dielectric interface and also an additional surface charge density in the DFT/PCM scenario. Note that we are referring here to the actual interfacial charge densities, not to the abstract apparent surface charges comprising the reaction of the entire dielectric environment. Interactions between the charges at the cavity boundary and the other polarization-induced charge densities in the substrate should, in principle, give rise to contributions to the interaction energy that are captured by the DFT/PCM calculation, but not by the analytical expression Eq.~\eqref{interactionDistance.eq}. However, the perfect agreement between the two approaches proves that such coupled contributions do not play a role here.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{interactionEnergy.pdf}
\caption{Interaction free energy between a) a charged particle within a solvent with dielectric constant $\varepsilon_{\mathrm{solv}}$ and its induced polarization charges on the surface towards a substrate with dielectric constant $\varepsilon_{\mathrm{subs}}$ at a distance $d$ away [Eq. \eqref{interactionDistance.eq}]. For the turquoise curve, $\varepsilon_{\mathrm{solv}}$~=~$\varepsilon_{\mathrm{subs}}$~=~5; for the blue one, $\varepsilon_{\mathrm{solv}}$~=~2 and $\varepsilon_{\mathrm{subs}}$~=~5; for the red one, $\varepsilon_{\mathrm{solv}}$~=~1 and $\varepsilon_{\mathrm{subs}}$~=~5. b) Interaction free energy between a charged particle (no solvent) and the charges induced in an anisotropic layer of finite thickness $T$ with perpendicular and parallel dielectric constants $\varepsilon_\perp$ and $\varepsilon_\parallel$, respectively [Eq. \eqref{interactionThickness.eq}]. The horizontal dashed line in panel (b) marks the $T\rightarrow\infty$ limit. In both panels, the solid lines are not guides to the eye, but correspond to the analytic electrostatic formulas, whereas the discrete points result from the PCM/DFT formalism, applied to a helium cation (He$^+$).}
\label{interactionEnergy.fig}
\end{figure}
As a second example, we test the description of spatially confined, anisotropic layers. The charge is situated \textit{in vacuo}, at a distance $d$~=~6~bohr away from a single dielectric layer of varying thickness $T$, with perpendicular and parallel dielectric constants $\varepsilon_\perp$ and $\varepsilon_\parallel$, respectively. Since no solvent is included here, it is evident that in this case the cavity and the charges supported on its boundary $\Gamma$ do not bear any physical meaning: $\Gamma$ does not mark an interface between dielectrics, but is situated in free space. Unfortunately, to the best of our knowledge, an analytical benchmark for this dielectric setup does not exist. We therefore use Eqs.~\eqref{greenPartial.eq} and \eqref{bZero.eq} instead. These equations are significantly simplified, since $\Delta r_\parallel$~=~0 and $z'+z-2z_0$~=~$2d$ for a point charge interacting with its own reaction field. We end up with
\begin{align}\label{interactionThickness.eq}
{\cal G}_{\text{int}}(T) &= -\frac{[(\varepsilon_\perp\varepsilon_\parallel)^{1/2}+1][(\varepsilon_\perp\varepsilon_\parallel)^{1/2}-1]}{2}\times\nonumber\\&\times\int_0^\infty\text dq_\parallel\,\frac{e^{-2q_\parallel d}}{1+\varepsilon_\perp\varepsilon_\parallel+2(\varepsilon_\perp\varepsilon_\parallel)^{1/2}\coth(q_\parallel(\varepsilon_\parallel/\varepsilon_\perp)^{1/2} T)}.
\end{align}
Comparing the interaction energies obtained from the DFT/PCM calculation and from Eq.~\eqref{interactionThickness.eq} as a function of the layer thickness $T$ and for different anisotropies, we get again essentially identical results [Fig.~\ref{interactionEnergy.fig}b)]. On a side note, Fig. \ref{interactionEnergy.fig}b) highlights the effect on the anisotropy on the reaction field. The product $\varepsilon_\perp\varepsilon_\parallel$ is the same for all three displayed cases, which, according to Eq.~
\eqref{interactionThickness.eq}, means that the curves converge to the same value for $T\rightarrow\infty$. However, the higher the value of the in-plane dielectric constant $\varepsilon_\parallel$ with respect to the out-of-plane one $\varepsilon_\perp$, the more rapidly this limit is approached. Thus, large in-plane dielectric constants give rise to stronger reaction fields than large out-of-plane ones.
The reflection-related part of our implementation [Eqs.~\eqref{rS} and \eqref{rP}] can be compared to analytical formulas available for limiting cases. Since the transfer-matrix formalism does not entail any additional approximations, it has to reproduce these results identically. The first special case is a single interface between two dielectric, semi-infinite layers. The first layer has an optical dielectric constant $\varepsilon_\mathrm{1}$, the second one $\varepsilon_{\mathrm{2},\parallel}$ and $\varepsilon_{\mathrm{2},\perp}$. As shown in Fig.~\ref{fresnel.fig}, the transfer-matrix method yields the correct angle dependence of the reflectivity; the results are identical to those obtained from the generalized Fresnel expression, ${\cal R}^{(s/p)}(\theta) = |\zeta_-^{(s/p)}(\theta)/\zeta_+^{(s/p)}(\theta)|^2$, with
\begin{subequations}
\begin{align}
\zeta_\pm^{(s)}(\theta) &= \varepsilon_\mathrm{1}\cos\theta\pm\varepsilon_\mathrm{1}^{1/2}\left(\varepsilon_{\mathrm{2},\parallel}-\varepsilon_\mathrm{1}\sin^2\theta\right)^{1/2}
\end{align}
and
\begin{align}
\zeta_\pm^{(p)}(\theta) &= \varepsilon_{\mathrm{2},\parallel}\cos\theta\pm\xi_\mathrm{2}\varepsilon_\mathrm{1}^{1/2}\left(\varepsilon_{\mathrm{2},\perp}-\varepsilon_\mathrm{1}\sin^2\theta\right)^{1/2}.
\end{align}
\end{subequations}
\begin{figure}
\centering
\includegraphics[width=.47\textwidth]{fresnel.pdf}
\caption{Angle-dependent reflectivity of an interface between media with optical dielectric constants $\varepsilon_1$ and $\varepsilon_2$. The red curves corresponds to optical dielectric constants of $\varepsilon_1$~=~1 and $\varepsilon_2$~=~2, the blue one to $\varepsilon_1$~=~3 and $\varepsilon_2$~=~40, and the turquoise one to $\varepsilon_1$~=~1, $\varepsilon_{2,\parallel}$~=~40, and $\varepsilon_{2,\perp}$~=~10.}
\label{fresnel.fig}
\end{figure}
As a second test case, we consider a dielectric mirror that consists of $m$ double layers of materials with optical dielectric constants $\varepsilon_1$ and $\varepsilon_2$. For a given wavelength $\lambda$, the reflectivity of this mirror is maximized for quarter-wave layers, \textit{i.e.} for thicknesses $d_1=\lambda/4\varepsilon_1^2$ and $d_2=\lambda/4\varepsilon_2^2$. In this case, the reflection coefficient for normal incidence is given as ${\cal R} = |\zeta_-/\zeta_+|^2$, with
\begin{align}
\zeta_\pm = \varepsilon_{\mathrm{solv}}^2\varepsilon_2^{4m}\pm\varepsilon_N^2\varepsilon_1^{4m},
\end{align}
where $\varepsilon_{\mathrm{solv}}$ and $\varepsilon_N$ are the dielectric constants of the solvent and substrate above and below the stack, respectively.~\cite{dbr} Also in this case, analytic and matrix results yield equal values (not shown). These test cases demonstrate that the transfer-matrix formalism captures the dependence of the reflectivity on the angle of incidence, the anisotropy, and the thickness of the layers.
\subsection{Thiophene@MoS$_2$ on top of a dielectric mirror}\label{thiophene.sec}
In this section, we test the whole formalism for a realistic system formed by a thiophene molecule adsorbed on a MoS$_2$ monolayer on top of a Si-based dielectric mirror. For convenience, we switch from atomic to conventional eV/\AA\,\,units. We consider the 1T at a distance 3~\AA\,away from the MoS$_2$ single-layer substrate atop the distributed Si/SiO$_2$ Bragg mirror separated by a large SiO$_2$ spacer layer (see Fig. \ref{parameters.fig}). The spacer layer has a depth on the order of microns, giving rise to notable delays of internally reflected fields. The thicknesses of the layers constituting the mirror are dictated by the wavelength that it is supposed to reflect, and thus on the nanometer scale. We do not use different values for the static and optical dielectric constants, since their values are fairly similar for the chosen materials close to the surface.\cite{laturia2018npj,malitson1965josa} This would not hold for (polar) solvents, which possess additional orientational degrees of freedom.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{parameters_new.pdf}
\caption{Schematic representation of the layered substrate used in the example. For the MoS$_2$ monolayer, we set\cite{laturia2018npj} $\varepsilon_\parallel$~=~15, $\varepsilon_\perp$~=~6 and $d$~=~4~\AA. It is separated by the next layer by 1~\AA\,\,of vacuum ($\varepsilon$~=~1). The spacer layer is made of SiO$_2$ ($\varepsilon$~=~2, $d$~=~1~$\mathrm{\mu}$m).~\cite{malitson1965josa} The two dual Si ($\varepsilon$~=~16, $d$~=~19.3~nm)\cite{aspnes1983prb} / SiO$_2$ ($d$~=~55.3~nm) layers underneath constitute a wavelength-selective mirror. The bulk substrate is of Si.}
\label{parameters.fig}
\end{figure}
We begin by examining the first ionization potential (IP) of 1T. In exact KS-DFT, this value is given as $-E_{\text{HOMO}}$.~\cite{perd1981prb} Within the adopted local-density approximation, we obtain values of $-E_{\text{HOMO}}$~=~6.10~eV and 6.13~eV \textit{in vacuo} and on top of the substrate, respectively. In other words, no relevant renormalization is obtained. However, if we estimate the IP by a $\Delta$SCF calculation, a different picture is observed: the IP decreases from 9.23~eV \textit{in vacuo} to 8.65~eV on the substrate. The absolute values are substantially higher than those of the KS levels, as expected, since the latter are always underestimated by the local-density approximation, as a consequence of self-interaction errors.~\cite{perdewZunger1981prb} More importantly, however, there is a sizeable reduction of the IP once the molecule is adsorbed on the substrate, which is a universal phenomenon. This shift is absent in the KS levels that are computed implicitly assuming a frozen substrate polarization, as no relaxation of the environment is taken into account. In the $\Delta$SCF calculation, on the other hand, the substrate is completely equilibrated with the ionized species. We note that the assumption of complete equilibrium is justified by the very similar values of the static and optical dielectric constants of the substrate; if remarkable differences were to exist, a non-equilibrium approach should be employed to allow only the fast environmental degrees of freedom to adapt to the ionization-induced charge-density variations.~\cite{pavel2008jpca,pavel2009jacs} Still, a significant renormalization of the energy levels would occur. This effect is not captured by $E_{\text{HOMO}}$ when simulating atomistically the combined system at the mere DFT level of theory with conventional exchange-correlation functionals.~\cite{liu2017jcp} A more sophisticated treatment of electron correlations, such as in the \textit{GW} approach,~\cite{garciaLastra2009prb, neaton2006prl, thygesen2009prl} is required for this purpose.
Proceeding with time-dependent simulations, we recall that an accurate model should take into account energy transfer between the molecule and the substrate. This can be expected to play a significant role in this case, since the excitation frequency of 1T overlaps with optical excitations of MoS$_2$ and SiO$_2$. A corresponding extension of the model, already developed for the case of pure solvent environments~\cite{ding2015jcp, corni2015jpca} and plasmonic nanoparticles,~\cite{pipoleCorni2016jpcc} will be the subject of future work; the results shown in this section represent an initial survey of the problem. We apply to the system a normally incident, Gaussian-enveloped laser pulse
\begin{align}
E_{\text{in}}(t) = E_0e^{-(t-t_\mu)^2/2t_\sigma^2}\cos(\omega_pt),
\end{align}
where $t_\mu=8$~fs and $t_\sigma=2$~fs; the amplitude $E_0$ corresponds to a peak intensity of 1$\times$10$^{10}$~W/cm$^2$. The carrier frequency, $\omega_p$~=~5.6~eV, and the polarization vector, $\hat{\boldsymbol{\gamma}}$, are chosen to ensure resonance with the first bright excitation, corresponding to the transition from the highest occupied to the lowest unoccupied molecular orbital of 1T.~\cite{krumland2020jcp, cocchidraxl2015prb} The resonance frequency is not strongly affected by the presence of the substrate, implying that the band-gap renormalization does not translate into a reduced optical gap. This is quite generally the case; the electronic properties are more strongly influenced by polarizable environments than the optical properties. Energy levels are associated with charged systems, which evoke a strong polarization responses of the surrounding medium, as they create an attractive image charge that stabilizes the ion. The transition densities related to optical excitations, on the other hand, are overall charge-neutral and interact with the environment only via dipole-dipole and higher multipolar interactions. In the framework of MBPT, this corresponds to a reduced quasi-particle gap being compensated by the decrease of the exciton binding energy.
Before analyzing the induced electron dynamics, we consider the characteristics of the reflected field [see Fig.~\ref{dbr.fjg}a)]. Some of the radiation is immediately reflected at the interfaces close to the surface and interferes with the incident pulse. Due to the reflection at the deep-lying Bragg reflector, a strong delayed pulse leaves the substrate at $t=19$~fs. The shape of the pulse is altered with respect to the incident one, as clearly visible in its prolonged tail. This is a consequence of the strong frequency dependence of the reflectivity in the spectral region of the incident laser [Fig. \ref{dbr.fjg}b)]; the power spectrum is situated on the right flank of the main stopband of the Bragg mirror. A second field recurrence can be spotted around 30~fs. The corresponding wave does not exit the substrate after the reflection at the mirror, but is reflected at the surface, and reflected once more at the mirror before being transmitted through the surface. The combination of the incident field and the strong delayed pulses effectively turns probing the adsorbed molecule into a multi-pulse scenario. The delay between the pulses as well as their relative phase is related to the angle of incidence and the thickness of the spacer layer.
Modulating the substrate through the application of pressure to the substrate can be anticipated as an effective way to tune the phase relationship of the pulses.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{fields.pdf}
\caption{a) Incident and reflected fields in time domain. The sketches in the insets show the light pathways corresponding to the different reflected pulses. b) Reflectivity $\cal R$ (black) of the layered substrate characterized by the parameters given in Fig. \ref{parameters.fig} as well as the spectral energy density of the incident laser field (red).}
\label{dbr.fjg}
\end{figure}
We now turn to examining the electron dynamics induced by these fields. If the contribution of the reflected field is not included, the incident pulse leads to a level of excited-state population that remains constant after the pulse ends (solid red line in Fig.~\ref{populations.fig}). Taking into account the polarization response of the substrate, the achieved excited-state population increases (dashed red line in Fig. \ref{populations.fig}). Thus, the molecule/substrate coupling enhances the absorption strength. When the reflected field is included, the time evolution of the number of excited electrons becomes non-trivial. The initial increase is weaker, since the surface-reflected field destructively interferes with the incident one. The second pulse decreases the excited-state population to a minimum before raising it back to slightly above the initial level. This is an interference effect between the oscillating dipoles due to the incident and reflected fields, which have some arbitrary phase relationship.~\cite{coccia2019jcp} At the minimum around 18~fs, the induced dipole moment of the system is reduced to almost zero, while part of the excited-state population is retained. Afterwards, the dipole moment is built back up by absorption. The third pulse ultimately causes a small decrease of the excited-state population. This final decrease is equivalent to the one at the beginning of the second pulse, since the phase difference between the first and second pulse is the same as that between the second and the third. Notably, the final number of excited electrons is lower when the polarization response of the substrate is taken into account. This is a consequence of different phase relationships between the fields and the dipole moment, which can lead to either additional absorption or to deexcitation. The phase difference stems from the subtle polarization-induced shift of the transition frequency, which dictates the oscillation frequency of the dipole moment.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{populations.pdf}
\caption{Excited-state population in laser-excited thiophene. The upper panel shows the incident (red) and reflected laser pulses (blue). The lower panel shows the number of excited electrons [Eq. \eqref{population.eq}] for four different cases; in the first two (red), the reflected field is not taken into, whereas in the other two (blue), it is included in the simulation. For the solid curves, the polarization response of the substrate is neglected, for the dashed ones, it is taken into account.}
\label{populations.fig}
\end{figure}
The phase relationship between the incident and reflected pulses can be tuned by changing the thickness of the spacer layer separating the surface and the Bragg mirror (see Fig. \ref{parameters.fig}). This can give rise to different interference effects that are reflected in the population dynamics.~\cite{giulia2020pccp} By deliberately choosing the layer thickness to have coherent incident and reflected fields, the reflected pulse further excites the molecule [see Fig.~\ref{phases.fig}a) and b)]. In this case, the reflected field is phase-shifted by $-\pi/2$ with respect to the dipole moment induced by the incident pulse. In the second case, the phase difference is $+\pi/2$ [Fig.~\ref{phases.fig}a) and c)]: This corresponds to the previously discussed scenario (Fig.~\ref{populations.fig}). The dipole moment is reduced to approximately zero, at which point it starts increasing again, as the reflected field has enforced the dipole-field phase relationship proper for absorption. The vanishing dipole moment at the minimum does not imply a stationary electronic system; the electron dynamics induced by the phase-shifted incident and reflected pulses overlap such that the dipole moments induced by such fields compensate each other at this point. A layer thickness bringing the dipole moment and the reflected field in phase with respect to each other leads to a depopulation of the excited state [Fig.~\ref{phases.fig}a) and d)]. At the end, the molecule is essentially back in its ground state.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{bloch.pdf}
\caption{a) Number of excited electrons calculated from PCM/TDDFT (solid), TDDFT without reaction field (dotted, only black), and the two-level system (dashed). The different shades of grey correspond to different relative phases of the induced dipole moment and the reflected laser [panels b)-d)].
These phase relationships correspond to spacer layer thicknesses of
b) 1564.93~nm, c) 1623.61~nm, and
d) 1604.05~nm
respectively. The time window displayed in panels b)-d) is indicated in a) by the vertical dashed lines. e) Trajectories of the Bloch vector generated by the pulses with in panels b)-d) and calculated from the two-level system driven by scaled-up fields (see text). The same shades of grey are used for reference. The dashed lines in the sphere mark its intersections with the three Cartesian planes.}
\label{phases.fig}
\end{figure*}
In order to spot possible artifacts arising from adiabatic TDDFT,~\cite{maitra2016jcp} we compare the results of such calculations with those obtained from a two-level system, ruled by the exactly solvable Hamiltonian
\begin{align}\label{two_level.eq}
\hat{\cal H}(t) = \Omega_{ge}|e\rangle\langle e|-\boldsymbol{\mu}_{ge}\cdot\textbf{E}(t)\big(| g\rangle\langle e|+|e\rangle\langle g|\big).
\end{align}
In Eq.~\eqref{two_level.eq}, the ground-state energy is shifted to zero, such that $\Omega_{ge}$ corresponds to the excitation energy from the ground state $|g\rangle$ to the excited state $|e\rangle$, with transition dipole moment $\boldsymbol{\mu}_{ge}$. The small static dipole moments of 1T in $|g\rangle$ and in $|e\rangle$ are neglected. Writing the total wavefunction as
\begin{align}
|\Psi(t)\rangle = c_g(t)|g\rangle + c_e(t)|e\rangle,
\end{align}
the time-dependent Schr\"odinger equation $i\partial/\partial t|\Psi(t)\rangle$ = $\hat{\cal H}(t)|\Psi(t)\rangle$ can be cast into matrix form and solved numerically by propagating in time the coefficients $c_g(t)$ and $c_e(t)$. The adopted system-related parameters, $\Omega_{ge}$~=~5.6~eV and $\boldsymbol{\mu}_{ge}$~=~1.86~D, result from a linear-response TDDFT calculation \textit{in vacuo} of the absorption spectrum of 1T. For consistency, the same electric field $\textbf{E}(t)$ is used in both RT-TDDFT and model calculations. The propagation of $c_g(t)$ and $c_e(t)$ is done with a fourth-order Runge-Kutta scheme.
We find qualitative agreement between the populations $|c_e(t)|^2$ of the excited state $|e\rangle$ and the number of excited electrons $N_e(t)$ calculated from TDDFT using Eq.~\eqref{population.eq},
see Fig. \ref{phases.fig}d): in all three cases, the trend is the same. However, the final population predicted by TDDFT is notably lower compared to the model in cases of constructive interference [Fig. \ref{phases.fig}a)] and intermediate phase relationship [Fig. \ref{phases.fig}c)]. This, however, does not necessarily expose failures of adiabatic TDDFT;~\cite{krumland2020jcp} differences may as well be due to additional states playing a role in the dynamics. If this is the case, the reason for the quantitative mismatch lies in the insufficiency of the two-level system as a model for the molecule. Regardless, the qualitatively identical results obtained with respect to the model prove that adiabatic TDDFT can be a useful tool to explore the coherent dynamics of adsorbed molecules.~\cite{krumland2020jcp}
The achieved qualitative agreement between the results of the exact two-level model and TDDFT allows us to exploit this comparison also in the interpretation of the dynamics. Let us consider the Bloch vector, $\textbf{R}(t)$, in a frame rotating with the transition frequency $\Omega_{ge}$. This vector has components
\begin{subequations}
\begin{align}
R_x(t) &= 2\Re{\left\lbrace c_g^*(t)c_e(t)e^{-i\Omega_{ge}t}\right\rbrace}\\
R_y(t) &= 2\Im{\left\lbrace c_g^*(t)c_e(t)e^{-i\Omega_{ge}t}\right\rbrace}\\
R_z(t) &= |c_e(t)|^2-|c_g(t)|^2.
\end{align}
\end{subequations}
The absence of decoherence mechanisms in the adopted formalism implies that $\textbf{R}(t)$ always has unit length, \textit{i.e.} lies on the Bloch sphere depicted in Fig.~\ref{phases.fig}e). As the system is initially in its ground state, the trajectories start at the south pole of the sphere. During the first pulse, the three $\textbf{R}(t)$, corresponding to the different phases of the reflected pulse, congruently move up the side of the sphere towards the branching point $B$. This first path segment from the ground state to $B$ is within the $R_x$~=~0 plane, indicating resonance. The differences in the delayed reflected fields, however, lead to a splitting; $\textbf{R}(t)$ moves away from $B$ in different directions. The angles with which $\textbf{R}(t)$ emerges from $B$ are directly related to the phase difference between the first and second pulse. Trajectories are curved, indicating a time-dependent detuning, which is a consequence of the altered spectrum of the reflected pulse. The shape of the second segment starting at $B$ is essentially the same in all three cases, and they are just rotated versions of each other. This suggests that increasing the thickness of the spacer layer leads to a continuous rotation of the second segment around the branching point $B$ within the Bloch sphere, giving rise to the variable observed population and coherence dynamics. We notice that the presented line of reasoning holds exactly in the two level system. A quantitative analysis of this sort on RT-TDDFT calculations is not straightforward for the lack of the many-body wave function of the system, and thus goes beyond the scope of this work. However, we acknowledge that it may be potentially useful to reveal departure of the results from the two-state model.
All of the above-described phenomena take place within the linear regime. The trajectories in Fig.~\ref{phases.fig}e) are magnified: They have been obtained by multiplying all field strengths by a factor of 13. This scaling does not noticeably affect the shape of the curves, as the maximum excited-state population is still well below 50\% (marked by the equator of the sphere), where stimulated emission, \textit{i.e.} a nonlinear effect, starts to outweigh. Avoiding fields that are strong and/or long enough to cause population inversion and Rabi oscillations is imperative for obtaining reliable results with adiabatic RT-TDDFT.~\cite{krumland2020jcp,fuks2011pbr,raghunathan2011jctc,raghunathan2012jcp}
\section{Summary, conclusions, and outlook}\label{conclu.sec}
We have presented LayerPCM, an effective approach, based on the well-established polarizable continuum model combined with RT-TDDFT, to account for the electrostatic interactions between quantum-mechanically modelled molecules physisorbed on dielectric substrates.
The key feature of the proposed method, which has been implemented in the \textsc{Octopus} code (v9.0), is to effectively model the dielectric properties of anisotropic layered substrates.
It is therefore ideally applicable to low-dimensional materials and heterostructures, including complex patterns hosting adsorbed molecules.
After having tested the robustness of the implementation, we have shown that the LayerPCM can be successfully adopted to compute the polarization-induced renormalization of the frontier energy levels of the adsorbate molecules.
Furthermore, we have demonstrated its build-in suitability for laser-induced ultrafast dynamics simulations within RT-TDDFT.
In these calculations, we have explicitly accounted for the electric fields emerging from Fresnel-reflection at the substrate, which, according to the complexity of the layered surface, can assume non-trivial shapes and thus profoundly affect the electron dynamics in the molecule. Interference effects in the coherent dynamics of the adsorbate can emerge upon pronounced retardation effects of the field after laser irradiation.
The comparison of our numerical results with those obtained from an exactly solvable two-level model enables the quantitative analysis of the excited-state population, including the polarization effects induced by the substrate.
The numerical efficiency of the proposed approach offers new opportunities to simulate complex hybrid structures with nano-engineered substrates from first principles at affordable computational efforts. Coupling LayerPCM with DFT enables to capture the key effects of the substrate on the electronic structure of the adsorbed molecules, such as the energy level renormalization, without the need for atomistic simulation of the entire hybrid system. Also, it allows for a computationally efficient modulation of the substrate characteristics, such as the layer thickness, the dielectric function, and the anisotropy, in order to explore different scenarios.
The advantages of LayerPCM in the time-domain are even more striking. The inclusion of reflected fields allows reproducing non-trivial physical effects that resemble those induced by multi-pulse spectroscopies.
The presented implementation is the first step towards a comprehensive description of electromagnetic interactions between molecules and layered substrates, treated implicitly.
Future steps entail the extension of LayerPCM to include frequency-dependent dielectric functions and the coupling with near fields.
\section*{Data Availability Statement}
The data that support the findings of this study are available from the corresponding author upon reasonable request.
\section*{Acknowledgements}
We are thankful to Prof. Roberto Cammi for useful discussions on some aspects of the theory. This work was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - project number 182087777 - SFB 951, by the German Federal Ministry of Education and Research (Professorinnenprogramm III), and by the State of Lower Saxony (Professorinnen für Niedersachsen). Computational resources were provided by the North-German Supercomputing Alliance (HLRN), project bep00076. GG and SC acknowledge funding from the ERC under the grant ERC-CoG-681285 TAME-Plasmons.
|
\section{Introduction}
The currently observed accelerated expansion of the Universe is widely supported by experimental evidence \citep{1999ApJ...517..565P,1998Natur.391...51P,1998AJ....116.1009R,2003ApJ...594....1T,2003Sci...299.1532B,2003ApJS..148....1B,2003ApJS..148..135H,2003ApJS..148..161K,2003ApJS..148..175S,2005ApJ...633..560E}.
The concordance paradigm assumes that a fluid whose corresponding density, $\rho_{\rm de}$, under the form of a cosmological constant, $\Lambda$, with equation of state (EoS), say $\omega_{\rm de}=P_{\rm de}/\rho_{\rm de}\equiv\omega_{\Lambda}=-1$, sufficiently negative to counterbalance the action of gravity and to speed up the Universe today \citep{2000IJMPD...9..373S,2006IJMPD..15.1753C}. Any departure in terms of barotropic fluids stands for dark energy (DE) \citep{1998AJ....116.1009R,1999ApJ...517..565P,2003RvMP...75..559P,2013PhR...530...87W} with the purpose of overcoming the main caveats of the standard $\Lambda$CDM paradigm \citep{2003PhR...380..235P}, i.e., constructed using $\Lambda$. Explaining the DE nature passes through the use of first principles \citep{2012Ap&SS.342..155B,2003PhR...380..235P,2006IJMPD..15.2105S,2006IJMPD..15.1753C, 2000IJMPD...9..373S}, and/or in terms of extended/modified theories of gravity \citep{2017PhR...692....1N, 2007IJGMM..04..209S} and so on. All these approaches, although profoundly different, rely on the hypothesis that DE is an additional fluid different from baryons and dark matter (DM).\footnote{For a different perspective see, e.g., \cite{Luongo2018}.} Among the different possibilities of studying competing DE models \citep{2006IJMPD..15.1753C,2012Ap&SS.342..155B,2009IJMPA..24.1545M}, it could be possible to formulate some sort of thermodynamic acceleration, i.e., treating the Universe as a thermodynamic system, where thermodynamic considerations over the whole ensemble of fluids permit acceleration of the Universe today, adopting a single fluid that unifies DE and DM.
Relevant typologies of thermodynamic models that satisfy the above requirement are the so-called logotropic models, which have been introduced to overcome the cusp-core problem \citep{2015EPJP..130..130C}. These models attempt to unify the dark sector since the logotropic fluid recovers DE and/or DM in limiting regimes, in analogy to the Chaplygin gas \citep{Bertacca2007,Bertacca2008,Piattella2010,Bertacca2011}. A particular class of logotropic models has been recently introduced within the framework of Anton-Schmidt EoS \citep{AntonSchmidt1997, Mayeretal2003}. This class of models is similar to genuine logotropic paradigms \citep{2015EPJP..130..130C,2017PhLB..770..213F,2019EPJC...79..332B} and can be matched with modified versions of the Chaplygin gas. The advantage of this class of models is that the Anton-Schmidt EoS is physically interpreted as the deformation of the Universe under the action of cosmic expansion \citep{Capozziello} and describes the transition from a pressureless status (state) to one with negative pressure while satisfying the Debye approximation \citep{Debye1912}.
In this paper, we propose a proper treatment employing linear perturbations for a set of four classes of logotropic models, including Anton-Schmidt gases. Doing so, we analyze the dynamical and experimental features of two logotropic models and two Anton-Schmidt gases. In particular, we first consider the original version of the logotropic fluid and then we introduce a new paradigm in which the logotropic counterpart explicitly distinguishes the role of DM from baryons. Afterwards, we investigate the most accredited versions of the Anton-Schmidt fluids, where the Gr\"uneisen parameter $\gamma_{\rm G}$ is constant throughout the Universe evolution. At first we allow it to be free and then we fix it to the special value, namely $\gamma_{\rm G}=\frac{5}{6}$, or alternatively $n=-\frac{1}{6}-\gamma_{\rm G}=-1$. We then demonstrate that Anton-Schmidt gases can be seen as generalized logotropic fluids. We, thus, investigate how structures evolve, generalizing the growth factor equation by taking into account the effects of the EoS and sound speed of each model.
To describe the evolution of the inhomogeneous energy shift we parameterize the growth function $f=\frac{\mathrm{d}\ln{\delta}}{\mathrm{d}\ln{a}}$ in terms of the growth index $\gamma$. Expanding in Taylor series, we get the corresponding approximate normalized growth function. Afterwards, numerical results are viewed in terms of Monte Carlo Markov Chain (MCMC) analyses based on the type Ia Supernova (SNe Ia) Pantheon data catalog, Hubble rates at different redshifts and redshift-space distortions, and through $\sigma_8$ data points based on the linear growth function for the large scale structures.
Inconsistencies among models are discussed with respect to the standard $\Lambda$CDM model. We show that the Anton-Schmidt gases are disfavored by intermediate redshift observations of the redshift space distortions with respect to pure logotropic models. Even though we demonstrate that logotropic models work better, we show statistical inconsistencies even for such scenarios with respect to both the $\Lambda$CDM, $\omega$CDM paradigms and Chevallier-Polarski-Linder (CPL) parametrization \cite{Chevallier2001,Linder2003}.
We conclude that a possible solution to the above-raised issues of our underlying models could be to take a varying Gr\"uneisen index that depends upon the temperature, namely $\gamma_{\rm G}=\gamma_{\rm G}(T)$. This would enable the sound speed to be always positive definite throughout the Universe evolution, cancelling out any perturbation instabilities.
The paper is thus structured as follows. In Sec.~\ref{sezione2} we exploit the concept of thermodynamic acceleration in the context of logotropic models and Anton-Schmidt gases. To do so, we highlight the basic properties of our four classes of thermodynamic models, confronting the genuine logotropic paradigm with Anton-Schmidt fluids. In Sec.~\ref{sezione3}, we work out linear perturbations for each model. We underline the basic differences and we evaluate the growth factor and the growth index $\gamma$. In Sec.~\ref{sezione4}, we present our fitting procedures, whose main results are analyzed and interpreted in Secs.~\ref{sezione4bis} and~\ref{sezione5}. Finally, in Sec.~\ref{sezione6}, we report our conclusions and perspectives.
\section{Thermodynamic dark energy: logotropic fluids}\label{sezione2}
Logotropic corrections to the Universe EoS are an attractive feature worth investigating. The original formulation of the Anton-Schmidt fluid can be clearly matched with logotropic DE models \citep{2015EPJP..130..130C} and Chaplygin gas \citep{2001PhLB..511..265K}. The simplest approach to determine a barotropic Anton-Schmidt EoS leads to \footnote{Please note that throughout the work we conventionally adopt natural units where $c=1$.}
\begin{equation}\label{eq:sample1}
\omega(\rho) = A\left(\frac{\rho^{-n(T)-1}}{\rho_{\ast}^{-n(T)}}\right)\ln\left(\frac{\rho}{\rho_{\ast}}\right)\,,
\end{equation}
where $\omega$ is the background EoS $\omega \equiv P/\rho$, whereas $\rho_{\ast}$ and $\rho$ are the reference and matter densities, respectively. The constant $A$ is a normalization factor. The index $n$ depends on the absolute temperature $T$ of the environment, i.e., the Universe, and can easily be approximated to a constant in epochs where $T$ does not significantly evolve.
Arguably, from solid state physics, we can write $n =-{\frac{1}{6}}-\gamma_{\rm G}$, where $\gamma_{\rm G}$ is the Gr\"uneisen parameter, closely associated to the physical properties of the fluid itself (for details see, e.g., \citep{Gruneisen1912}). For $n = 0$, Eq.~(\ref{eq:sample1}) reduces to the genuine logotropic cosmological models \citep{2015EPJP..130..130C}.
Considering the continuity equation in a Friedmann-Robertson-Walker spacetime
\begin{equation*}
\frac{\mathrm{d}\rho}{\mathrm{d}t}+3H(\rho+P) = 0\,,
\end{equation*}
one gets the Hubble parameter
\begin{equation}\label{eq:sample4}
H^2 \equiv \left(\frac{\dot{a}}{a}\right)^2 = \frac{8 \pi G \epsilon}{3}\,,
\end{equation}
where one can split the total energy, $\epsilon$, into matter and DE counterparts \citep{Capozziello,Capozziello2}.
We distinguish four relevant cases. In the first, $n$ is a free parameter of the model, assumed constant throughout the cosmic evolution. This case corresponds to the simplest Anton-Schmidt gas, i.e., the one in which the Universe temperature has (small) influence over the evolution of the fluid itself. In the second case we set $n=-1$. This leads to a negligible effect of the temperature on $n$ that is no longer a free parameter of the model itself. Statistically speaking, \emph{a priori}, this case may be favored with respect to the first case and has been investigated in \citep{Capozziello2}. These two approaches correspond to limiting cases of the most general logotropic models, discussed as a third case. The latter represents a genuine logotropic paradigm, which simply requires $n\rightarrow0$. The last case is the formulation of a further logotropic model that we introduce, assuming a specific case of the pressure $P$, as we will detail below. For each model we will present below the most relevant expressions, needful to study the evolution of linear perturbations.
The total energy density can be split into two components, i.e., matter and DE as $\epsilon = \epsilon_{\rm m}+\epsilon_{\rm de}$ and the energy density parameters are defined as $\Omega_{\rm m,0}\equiv\epsilon_{\rm m,0}/\epsilon_{\rm c}$, $\Omega_{\rm de,0}\equiv\epsilon_{\rm de,0}/\epsilon_{\rm c} = 1-\Omega_{\rm m,0}$, where we used the critical energy density $\epsilon_{\rm c}\equiv 3H_0^2/(8\pi G)$.
Knowing the expression for the pressure $P$ and the density $\rho$, the adiabatic sound speed for a barotropic fluid is defined by
\begin{equation}\label{eq:cs2}
c_{\rm s,a}^2 = \left(\frac{\partial P}{\partial a}\right)\left(\frac{\partial\rho}{\partial a}\right)^{-1}\,.
\end{equation}
This represents a key quantity, entering the linear perturbation equations, that determines the stability of perturbations.
Thus, to investigate the cosmological features of all our models, we employ the original formulation of the Anton-Schmidt pressure \cite{Capozziello2} and adopt the first law of thermodynamics, $\mathrm{d}\epsilon=\left(\frac{\epsilon+P}{\rho}\right)\mathrm{d}\rho$, which can be integrated as
\begin{equation}\label{lointegrale}
\epsilon = \rho + \rho \int^{\rho}\mathrm{d}\rho^{\prime}\frac{P(\rho^{\prime})}{{\rho^{\prime}}^2}\,.
\end{equation}
Considering the pressure of a logotropic model, we can express $\epsilon$ in terms of $\rho$
\begin{equation}\label{eq:total rho}
\epsilon = \rho + \frac{A}{2}\left(\frac{\rho}{\rho_{\ast}}\right)\ln^2{\left(\frac{\rho}{\rho_{\ast}}\right)}\,.
\end{equation}
Clearly, the limits for $a\ll 1$ and $a\gg 1$ lead to a matter and DE-dominated Universe, respectively.
Below we highlight the different models, concentrating on the total EoS, say $\omega=P/\epsilon$, the DE EoS, namely $\omega_{\rm de}=P_{\rm de}/\epsilon_{\rm de}$ and the adiabatic sound speed, as above defined by Eq.~(\ref{eq:cs2}).
\emph{Case I: ($n$ and $\gamma_{\rm G}$ as free coefficients)} For clarity, $n$ is a function of the temperature and so, in principle, it is free to vary throughout the cosmic evolution. However, given a particular cosmic era, it is plausible that $n$ only slightly evolves and, therefore, it could be considered roughly constant \cite{Capozziello}.
Assuming that deviations from the case of constant $n$ are negligible, we have
\begin{equation}\label{eq:sample72}
\epsilon_{\rm de} = \epsilon_{\rm de,0}a^{3n} + \frac{3A}{n+1}\left(\frac{\rho_{\rm m,0}}{\rho_{\ast}}\right)^{-n}a^{3n}\ln{a}\,,
\end{equation}
where $\epsilon_{\rm m,0}$ and $\epsilon_{\rm de,0}$ are the matter and DE densities at current time, respectively. We further have
\begin{equation}\label{eq:sampleT1}
\epsilon_{\rm de,0} = -\frac{A}{n+1}\left(\frac{\rho_{\rm m,0}}{\rho_{\ast}}\right)^{-n} \left[\ln{\left(\frac{\rho_{\rm m,0}}{\rho_{\ast}}\right)}+ \frac{1}{n+1}\right]\,.
\end{equation}
\noindent Eq.~(\ref{eq:sample4}) can be written as
\begin{equation}\label{eq:sample74}
H^{(I)} = H_0\left[\Omega_{\rm m,0}a^{-3} + \Omega_{\rm de,0}\left(1+3B\ln{a}\right) a^{3n}\right]^{\frac{1}{2}}\,,
\end{equation}
where the superscript $(I)$ denotes that the underlying quantity refers to model I. The characteristic parameter $B$ is
\begin{equation}\label{eq:B}
B \equiv -\left[\ln{\left(\frac{\rho_{\ast}}{\rho_{\rm m,0}}\right)} +\frac{1}{n+1}\right]^{-1}\,.
\end{equation}
The parameter $B$ is related to a dimensionless logotropic temperature that is assumed to be constant for simplicity of computation. Further, we get
\begin{eqnarray}
\omega^{(I)} & = & -\frac{\Omega_{\rm de,0}\left[B+(n+1)\left(1+3B\ln{a}\right)\right] a^{3n}}{\Omega_{\rm m,0}a^{-3}+\Omega_{\rm de,0}\left(1+3B\ln{a}\right)a^{3n}}\,, \label{eq:sample79}\\
\omega_{\rm de}^{(I)} & = & -\left(n+1\right)-\frac{B}{1+3B\ln{a}}\,, \label{eq:sample79b}\\
c_{\rm s, a}^{2\,\,(I)} & = & \left(\frac{\Omega_{\rm de,0}}{\Omega_{\rm m,0}}\right)a^{3(n+1)} \times \nonumber \\
&& \left[(1+2n)B + n(n+1)(1+3B\ln{a})\right]\,,\label{eq:sample84}
\end{eqnarray}
respectively, the total EoS, the Anton-Schmidt EoS and the adiabatic sound speed. The latter is positive, leading to stable perturbations, only if the argument of the second parenthesis is positive. This requirement is essential for structure formation theory.
\emph{Case II: (fixed $n=-1$ and $\gamma_{\rm G}=\tfrac{5}{6}$)} Here, the effect of the temperature does not influence the overall evolution. For this reason, $n$ is fixed to a precise value and the corresponding EoS aims at describing both the deceleration and acceleration epochs.
Thus, handling the total density, splitting it into two contributions ($\epsilon = \epsilon_{\rm m} + \epsilon_{\rm de}$), with $\epsilon_{\rm m} = \epsilon_{\rm m,0}a^{-3}$, we have
\begin{equation}\label{eq:rho_de}
\epsilon_{\rm de} = \frac{\epsilon_{\rm de,0} }{a^3} - \frac{3A}{a^3}\left(\frac{\rho_{\rm m,0}}{\rho_{\ast}}\right)\ln{a}\ln{ \left(\frac{\rho_{\rm m,0}}{\rho_{\ast}}a^{-3/2}\right)}\,,
\end{equation}
for the DE density. We defined
\begin{equation}
\epsilon_{\rm de,0} = \frac{A}{2}\left(\frac{\rho_{\rm m,0}}{\rho_{\ast}}\right)\ln^2{\left(\frac{\rho_{m,0}}{\rho_\ast}\right)}\,.
\end{equation}
As for model I, we can define the characteristic parameter $B$
\begin{equation}
B\equiv\ln^{-1}\left(\frac{\rho_{\rm m,0}}{\rho_{\ast}}\right)\,,
\end{equation}
which is quite different from the one given for the logotropic models
\begin{equation}
B_{\rm log}\equiv \left[\ln{\left(\frac{\rho_{\ast}}{\rho_{\rm m,0}}\right)}-1\right]^{-1}\,.
\end{equation}
The latter, as stated above, depends on the logotropic temperature and holds a precise physical meaning. Contrary to its particular interpretation, in our present case, we expect to have $B<0$ as $\rho_{\ast}\gg 1$. Differently from previous results in the literature \cite{estuba}, we demonstrate here that the characteristic density is not necessarily the Planck density. Moreover, our experimental fits will show that, even considering the Planck density as extreme case for $\rho_{\ast}$, the model fails to predict high redshift evolution of the Universe.
Using the expression for $\epsilon_{\rm de,0}$, we can write the Hubble parameter as
\begin{equation}
H^{(II)} = H_0\left[\Omega_{\rm m,0}a^{-3}+\Omega_{\rm de,0}(1-6B\ln{a}+9B^2\ln^2{a})a^{-3}\right]^{\frac{1}{2}}\,.
\end{equation}
As done for model I, we now present the expressions for the EoS of the total fluid and of the DE component and for the adiabatic sound speed, respectively:
\begin{eqnarray}
\omega^{(II)} & = & \frac{2B - 6B^2\ln{a}}{\Omega_{\rm de,0}^{-1}-6B\ln{a} + 9B^2\ln^2{a}}\,,\\
\omega_{\rm de}^{(II)} & = & \frac{2B}{1-3B\ln{a}}\,,\\
c_{\rm s, a}^{2\,\,(II)} & = & \frac{A\left[1+\ln{\left(\frac{\rho}{\rho_{\ast}}\right)}\right]}{\rho_{\ast}+\frac{A}{2}\left[2+\ln{\left(\frac{\rho}{\rho_{\ast}}\right)}\right]\ln{\left(\frac{\rho}{\rho_{\ast}}\right)}}\,.
\end{eqnarray}
For model II, we added the superscripts $(II)$, in analogy to model I. At small $z$, one gets $\omega_{\rm de}\approx 2B+6B^2(a-1)$, and the $\Lambda$CDM paradigm, for which $\omega_{de}\equiv-1$, is recovered when $B\rightarrow-1/2$ at $a=1$.
The expression for the sound speed can be written in terms of the characteristic parameter $B$. After simple manipulations we get
\begin{equation}\label{eq:cs nuovo}
c_{\rm s, a}^{2\,\,(II)} = \frac{2B\Omega_{\rm de,0}(1+B-3B\ln{a})}{\Omega_{\rm m,0}}\,.
\end{equation}
Again, its sign depends upon the choice of the free constants here involved. This limitation of the model will reflect to our experimental fits.
\emph{Case III: (fixed $n=0$ and $\gamma_{\rm G}=-\frac{1}{6}$)} This case deals with a pure logotropic framework. In particular, we recover the third model as limit case of model I. Once again, the Universe temperature is here negligible with the peculiar choice $n=0$. This assumption unifies DE and DM as two byproducts of the same single fluid.
As a possible physical justification of $n=0$, let us assume the DE EoS can be used for galactic dynamics. Hence, taking it within hydrodynamic equilibrium equations and noticing that DE and DM can be considered as single dark fluid, the pressure may describe both cosmic evolution and complicated galactic DM structures.
In particular, if DM halos are subject to hydrostatic equilibrium, in Newtonian regime we have
\begin{equation}\label{logojust1}
\nabla P + \rho\nabla\Phi = {\bf 0}\,,
\end{equation}
and considering a polytropic relation $P=K\rho^{\gamma}$, one finds
\begin{equation}\label{logojust2}
K\gamma \rho^{\gamma-1}\nabla \rho + \rho\nabla\Phi = {\bf 0}\,.
\end{equation}
To avoid a central cusp, the pressure should be constant. For the pressure gradient to counterbalance gravity, we require, however, $K\gamma>0$ and we can get a unified EoS by assuming $\gamma\rightarrow 0$ and $K\rightarrow \infty$. This leads to $A=K\gamma$, which is finite, and in such a limit we write
\begin{equation}\label{logojust3}
A \frac{\nabla\rho}{\rho} + \rho\nabla\Phi = {\bf 0}\,.
\end{equation}
Comparing Eqs.~(\ref{logojust1}) and~(\ref{logojust3}), it is evident that
\begin{equation}\label{logojust4}
P = A\ln{\rho} + C\,,
\end{equation}
where $A$ and $C$ are two integration constants. The logotropic EoS can be simplified fixing $C$, i.e., the cosmological constant contribution and so an easier form of the above equation for the pressure becomes
\begin{equation}\label{eq:sampleT3}
P = A\ln{\left(\frac{\rho}{\rho_{\ast}}\right)}\,,
\end{equation}
where the cosmological constant has been suitably removed.
The main purpose is now to determine the energy density $\epsilon$, knowing $\rho$. Assuming an adiabatic evolution, one immediately gets
\begin{equation}\label{eq:sampleT0}
\epsilon = \epsilon_{\rm m,0}a^{-3} + \epsilon_{\rm de,0} \left(1+3B\ln{a} \right)\,,
\end{equation}
and the Hubble rate becomes
\begin{equation}\label{eq:sampleT4}
H^{(III)} = H_0 \left[\Omega_{\rm m,0}a^{-3} + \Omega_{\rm de,0}\left(1+3B\ln{a}\right)\right]^{\frac{1}{2}}\,,
\end{equation}
that, clearly, can be recovered from Eq.~(\ref{eq:sample74}) when $n=0$ and with the superscript $(III)$ that hereafter is used to distinguish model III from models I and II. Here, the parameter $B$ is defined as
\begin{equation}\label{eq:sampleT5}
B = \frac{A}{\epsilon_{\rm c}\Omega_{\rm de,0}}\,.
\end{equation}
Analogously with the previous two paradigms, the relevant thermodynamic quantities for this model read
\begin{align}
\omega^{(III)} = &\, -\frac{\Omega_{\rm de,0}\left(B+1+3B\ln{a}\right)} {\Omega_{\rm m,0}a^{-3} + \Omega_{\rm de,0}\left(1+3B\ln{a} \right)}\,, \label{eq:sampleT1a} \\
\omega_{\rm de}^{(III)} = &\, -1-\frac{B}{1+3B\ln{a}}\,, \label{eq:sampleT1b}\\
c_{\rm s, a}^{2\,\,(III)} = &\, B a^{3} \left(\frac{\Omega_{\rm de,0}}{\Omega_{\rm m,0}}\right)\,, \label{eq:sampleT2}
\end{align}
which represent the total EoS, that of the DE and the adiabatic sound speed, respectively. The $\Lambda$CDM model is recovered for $B=0$. Taking into account our previous calculations, we can easily relate the adiabatic sound speed in pure Anton-Schmidt's paradigm, model I, with Eq.~\eqref{eq:sampleT2} by
\begin{equation}\label{comparison}
c_{\rm s,a}^{2\,\,(I)} = \frac{2(1+B)(3B\ln{a}-B-1)}{a^3}\, c_{\rm s,a}^{2\,\,(III)}\,,
\end{equation}
with $B$ given as in (\ref{eq:B}).
\emph{Case IV: (modified $n=0$ logotropic model)} The logotropic version of DE, considered in case III, takes into account the basic assumption that the DE EoS can be used for galactic dynamics.
However, the main disadvantage of models I, II and III is that their sound speeds are negative definite in broad regions of the Universe's evolution. This could be a direct consequence of how the model unifies DE with DM. A general solution of this issue, as we will stress below, is offered by removing the constant temperature approximation on $n$. This speculation deserves, however, accurate investigations since we do not know \emph{a priori} how the Gr\"uneisen index depends on the temperature. The prerogative of understanding which is the most suitable $\gamma_{\rm G}=\gamma_{\rm G}(T)$ function will be object of future investigations and would help to unify inflation with dark energy epochs.
Hence, we propose the simplest generalization of model III, where DM and baryons are unified with DE. This would guarantee the sound speed to be positive definite in the wider domain of the Universe's expansion unlike models I, II and III. Thus, assuming again the hydrostatic equilibrium, i.e., Eq.~\eqref{logojust1}, one can extend the polytropic EoS through a double polytropic of the form
\begin{equation}\label{doublelog}
P = K_1\rho^{\gamma_1} + K_2\rho^{\gamma_2}\,,
\end{equation}
where only two constants, namely $(K_1,\gamma_1)$, behave as above, i.e., $K_1\gamma_1=const$, with $(K_1,\gamma_1)\rightarrow(\infty,0)$. The other two, namely $(K_2,\gamma_2)$, vary freely. In particular, we have
\begin{equation}\label{logojustother}
K_1\gamma_1 \rho^{\gamma_1-1}\nabla \rho + K_2\gamma_2 \rho^{\gamma_2-1}\nabla \rho+\rho\nabla\Phi = {\bf 0}\,.
\end{equation}
\noindent It is interesting to work out the case $(K_2,\gamma_2)\rightarrow(C, 1)$. In this respect, we get
\begin{equation}\label{extendedlogotropicmodel13}
P = -A\ln{\left(\frac{\rho}{\rho_{\ast}}\right)} + C\left(\frac{\rho}{\rho_\ast}\right)\,,
\end{equation}
where $A\equiv -K_1\gamma_1$. For the sake of clarity, the above procedure can be extended up to an arbitrary order of polytropic equations of state. However, we limit our attention to the simplest case provided by the choice \eqref{doublelog} that has the intriguing advantage to reduce to $\sim \rho$ and $\sim\ln{\rho}$ for very large and small density $\rho$, respectively. Plugging Eq.~\eqref{extendedlogotropicmodel13} into \eqref{lointegrale}, we obtain
\begin{equation}\label{epvsrhocase4}
\epsilon = \rho + \left[A + C\left(\frac{\rho}{\rho_{\ast}}\right)\right]\ln\left(\frac{\rho}{\rho_{\ast}}\right) + A\,.
\end{equation}
Note that $A$, although under the form of a cosmological constant, does not have the meaning of a cosmological constant contribution, since it is formally given as the product by $K_1$ and $\gamma_1$. Following the same strategy of the three previous models, we get $\epsilon \equiv \epsilon_{\rm b} + \epsilon_{\rm cdm} + \epsilon_{\rm de}$, explicitly showing the contributions of baryons ($\epsilon_{\rm b}$) and cold DM ($\epsilon_{\rm cdm}$) that arise from the term $\sim C\rho$, in Eq.~\eqref{extendedlogotropicmodel13}. We can then define the total matter contribution as $\rho_{\rm m} = \rho_{\rm cdm}+\rho_{\rm b}$. By comparison with case III, we impose
\begin{subequations}
\begin{align}
\rho_{\rm b} & = \frac{\rho_{\rm b,0}}{a^3}\,,\\
\rho_{\rm cdm} & = \frac{\rho_{\rm cdm,0}}{a^3} = \frac{C}{a^3}\left(\frac{\rho_{\rm b,0}}{\rho_\ast}\right)\ln\left(\frac{\rho_{\rm b,0}}{\rho_\ast}\right)\,,\\
\rho & = A\left[1+\ln\left(\frac{\rho_{\rm b,0}}{\rho_\ast}\right)\right]-3\left[A+\frac{C}{a^3}\left(\frac{\rho_{\rm b,0}}{\rho_\ast}\right)\right]\ln{a}\,,
\end{align}
\end{subequations}
where we defined
\begin{align}
B = &\, -\left[1+\ln\left(\frac{\rho_{\rm b,0}}{\rho_{\ast}}\right)\right]^{-1}\,,\\
A = &\, -\left(1-\Omega_{\rm m,0}\right)B \rho_{\rm c}\,.
\end{align}
\noindent We are now in the position to get the Hubble rate
\begin{eqnarray}
H^{(IV)} = H_0\Big[\Omega_{\rm m,0}a^{-3} & + & (1-\Omega_{\rm m,0})\left(1+3B\ln{a}\right) + \nonumber\\
& + & \Omega_{\rm cdm,0}\left(\frac{3B}{B+1}\right)a^{-3}\ln{a}\Big]^{\frac{1}{2}}\,,
\end{eqnarray}
so that the total EoS, the DE EoS and the adiabatic sound speed are, respectively,
\begin{eqnarray}
\omega^{(IV)} & = & -1 - \frac{B\left(1-\Omega_{\rm m,0}\right)-\Omega_{\rm m,0}\,a^{-3}}{E^2} + \nonumber \\
& & \qquad - \frac{\Omega_{\rm cdm,0}}{E^2}
\left(\frac{B}{B+1}\right)a^{-3}\left(1-3\ln{a}\right)\,,\\
\omega_{\rm de}^{(IV)} & = & -1 - \frac{B\left(1-\Omega_{\rm m,0}\right)}{E^2-\Omega_{\rm m,0}a^{-3}}+\nonumber\\
& &-\frac{\Omega_{\rm cdm,0}}{a^3 E^2-\Omega_{\rm m,0}}\left(\frac{B}{B+1}\right)\left(1-3\ln{a} \right),\,\,\,\,\,\label{DEeos4}\\
c_{\rm s,a}^{2\,\,(IV)} & = & \frac{B}{\Omega_{\rm b,0}} \left[\left(1-\Omega_{\rm m,0}\right)a^3 - \frac{\Omega_{\rm cdm,0}}{B+1}\right]\,. \label{soundspeedmod}
\end{eqnarray}
In the above expressions, the superscript $(IV)$ refers to model IV, $E=H/H_0$ and $\Omega_{\rm m,0}=\Omega_{\rm b,0}+\Omega_{\rm cdm,0}$.
\section{Evolution of the growth factor}\label{sezione3}
The models described in the previous section are all specific cases of a wider formalism where DM and DE are unified into a single dark fluid \citep{Bertacca2007,Bertacca2008,Piattella2010,Bertacca2011}. Both components represent limiting cases of the more general fluid at early and late times, respectively. This has two beneficial consequences: the first is that we only need a single component to explain both the structure formation and the observed accelerated expansion; the second is that we can treat DM and DE at the perturbation level in exactly the same way.
In this section, we concentrate on the analysis of linear perturbations, by deriving the appropriate differential equation for the growth factor. Note that, while formally perturbations are linear only at early times and/or on large scales, the growth factor equation is valid only on small scales, and a fundamental assumption is that it is valid on subhorizon scales. We also remind the reader that the growth factor is one of the main ingredients of the halo mass function, making it, therefore, an important quantity to study.
In the literature, the following expression is widely used
\begin{equation}
\ddot{\delta} + 2H\dot{\delta} - 4\pi G \rho_{\rm m} \delta = 0\,,
\end{equation}
where $\rho_{\rm m}$ is the (total) matter density. This equation, valid on small scales and for linear perturbations, implicitly assumes that matter is the clustering component. In this case, $c_{\rm s}^2=\omega=0$. However, our setup is more general than that and the correct way of proceeding needs to take into account the additional degrees of freedom of our model. In this respect, our physical setup is very similar to what has been recently done for generalized DM \citep{Kopp2016,Thomas2016} by \cite{Pace2020}. One of the differences with respect to that work is that in our case both the background EoS and the sound speed are, in general, time-dependent and both need to be properly evaluated. We have no freedom to set one of them to zero to simplify our expressions.
Before starting, we need to define correctly which perturbations are we talking about. From our previous discussion, we said that our fluid can be decomposed into two components, one representing matter and the other one resembling a smooth DE component driving the accelerated expansion of the Universe. Of these two fluids, only DM is clustering and, therefore, modifications to the standard growth factor equation will be related to the DM component.
To derive our equations, we start from the equations for a fluid with pressure and density perturbations, so that $c_{\rm s}^2=\delta P/\delta\rho$ and follow the derivation in \cite{Pace2017}. However, since the models considered here are adiabatic, the sound speed entering into perturbations is the adiabatic one, i.e., $c_{\rm s}^2=c_{\rm s,a}^2$. This will enormously simplify our analysis, in that we do not need to use any scalar field descriptions for the models.
The continuity and Euler equations read, respectively,
\begin{align}
\delta^{\prime} + 3(s-w)\delta + (1+w)\tilde{\theta} = &\, 0\,, \label{eq:continuity}\\
\tilde{\theta}^{\prime} + \left(2+\frac{H^{\prime}}{H}\right)\tilde{\theta} + \frac{3}{2}(1+3s)\Omega(a)\delta = &\, 0\,, \label{eq:euler}
\end{align}
where the prime represents the derivative with respect to $\ln{a}$, $\delta=\delta\rho/\rho$ is the dimensionless density perturbation of the fluid, $s=c_{\rm s,a}^2$ is the dimensionless adiabatic sound speed for perturbations and $\tilde{\theta}=\theta/H$, where $\theta$ represents the divergence of the peculiar velocity $\boldsymbol{u}$, $\theta=\vec{\nabla}\cdot\boldsymbol{u}$. In the Euler equation, $\Omega(a)$ represents the energy density parameter of the perturbed fluid, i.e., the matter component, so that $\Omega(a) = \Omega_{\rm m}(a)$, which is defined as
\begin{equation}
\Omega_{\rm m}(a) = \frac{\rho_{\rm m}}{\rho_{\rm m}+\rho_{\rm de}}\,.
\end{equation}
We are now in a position to derive a second order equation for $\delta$. To do so, we take the derivative of Eq.~(\ref{eq:continuity}) and we substitute in it Eq.~(\ref{eq:euler}). For simplicity of notation, we define the following variables:
\begin{equation*}
A = 3(s-w)\,, \quad B = 1+\omega\,, \quad f = \frac{3}{2}(1+3s)\Omega(a)\,.
\end{equation*}
\noindent The final equation then reads
\begin{equation}\label{eq:gf}
\delta^{\prime\prime} + \left(A\delta\right)^{\prime} +
\left[\left(2+\frac{H^{\prime}}{H}\right)-\frac{B^{\prime}}{B}\right]
(\delta^{\prime}+A\delta) - Bf\delta = 0\,.
\end{equation}
This expression is similar in its form, and fully equivalent, to what obtained in \cite{Abramo2007}.
We can now specify the two free functions in Eq.~\eqref{eq:gf}. Since the clustering component is DM, we set $\omega=0$, whereas the adiabatic sound speed is applied to the whole model. Note the specularity with clustering DE models, where $c_{\rm s}^2=0$ to allow a clustering similar to that of DM. We can then simplify Eq.~(\ref{eq:gf}) to
\begin{equation}\label{eq:gf1}
\delta^{\prime\prime} + 3\left(s\delta\right)^{\prime} +
\left(2+\frac{H^{\prime}}{H}\right)(\delta^{\prime} + 3s\delta) - \frac{3}{2}(1+3s)\Omega_{\rm m}(a)\delta = 0\,.
\end{equation}
It is often interesting to consider the logarithmic derivative of the growth factor $f=\mathrm{d}\ln{\delta}/\mathrm{d}\ln{a}$. Its equation, in light of the modified Eq.~(\ref{eq:gf1}), reads
\begin{equation}\label{eqn:exact_f}
f^{\prime} + f^2 + 3(s^{\prime}+sf) + \left(2+\frac{H^{\prime}}{H}\right)(f+3s) - \frac{3}{2}(1+3s)\Omega_{\rm m} = 0\,.
\end{equation}
For many models, it is possible to give a phenomenological solution for $f$ \citep{Paul}
\begin{equation}\label{eq:f}
f \approx \Omega_{\rm m}^{\gamma}(a)\,,
\end{equation}
where $\gamma$ is the so-called growth index.
To study the evolution of $\gamma$, the simplest thing to do is to plug the approximate solution for $f$ into (\ref{eq:f}). This leads to a first-order differential equation for $\gamma$ which reads:
\begin{align}
\label{totalgamma}
\gamma^{\prime} + & \frac{3\Omega_{\rm de}\omega_{\rm de}}{\ln{\Omega_{\rm m}}}\gamma + \frac{\Omega_{\rm m}^{\gamma}}{\ln{\Omega_{\rm m}}} + 3\frac{s+s^{\prime}\Omega_{\rm m}^{-\gamma}}{\ln{\Omega_{\rm m}}} \\
+ & \frac{1-3\Omega_{\rm de}\omega_{\rm de}}{2\ln{\Omega_{\rm m}}}\left(1+3s\Omega_{\rm m}^{-\gamma}\right) - \frac{3(1+3s)}{2\ln{\Omega_{\rm m}}}\Omega_{\rm m}^{1-\gamma} = 0\,. \nonumber
\end{align}
To make progress in solving this equation, we will make some assumptions which allow us to linearize it. When $\Omega_{\rm m}\approx\mathcal{O}(1)$, we can write $\ln{\Omega_{\rm m}}\approx-\Omega_{\rm de}$ and $\Omega_{\rm m}^{\gamma}\approx 1-\gamma\Omega_{\rm de}$. Under these approximations, the evolution of $\gamma$ is described by
\begin{align}\label{eq:gamma_linear}
\gamma^{\prime} + & \left[1-3\omega_{\rm de}+\frac{3}{2}+\frac{3}{2}s\left(2+3\Omega_{\rm de}\omega_{\rm de}\right)-3s^{\prime}\right]\gamma = \nonumber \\
& \frac{3}{2}(1+6s)(1-\omega_{\rm de}) + \frac{3s^{\prime}}{\Omega_{\rm de}}\,,
\end{align}
and we will assume that this approximation is roughly valid also at later times.
For $s=0$, assuming constant $\omega_{\rm de}$, $\gamma$ recovers the solution of the $\omega$CDM model
\begin{equation}\label{eq:gamma_wCDM}
\gamma = \frac{3(\omega_{\rm de}-1)}{6\omega_{\rm de}-5}\,,
\end{equation}
which reduces to $\gamma=6/11$ for the standard $\Lambda$CDM paradigm \citep{1998ApJ...508..483W}.
The equation above has a formal solution which can be expressed via the integral of the coefficient of $\gamma$ and of the source term. However, this expression will provide a very limited insight into the physics of the model.
Whether an analytical solution is possible or not, Eq.~(\ref{eq:gamma_linear}) is showing that the adiabatic sound speed acts as a correction to the standard picture. To better see why this is the case, we will also assume that both $\omega_{\rm de}$ and $s$ are constant. We will further consider, consistently for the derivation of the equation for $\gamma$, that $2+3\Omega_{\rm de}\omega_{\rm de}\approx 2$. Under these assumptions, Eq.~(\ref{eq:gamma_linear}) becomes
\begin{equation}
\gamma^{\prime} + \left(1-3\omega_{\rm de}+\frac{3}{2}+3s\right)\gamma = \frac{3}{2}(1+6s)(1-\omega_{\rm de})\,,
\end{equation}
whose solution is
\begin{equation}\label{eq:gamma_s_w}
\gamma = \frac{3(1+6s)(1-\omega_{\rm de})}{5+6(s-\omega_{\rm de})}\,.
\end{equation}
The evolution of $\gamma$ for our models will be discussed in the next sections, in comparison with the predictions of the standard model, i.e. $\gamma=\frac{6}{11}$, of the $\omega$CDM, through the use of Eq.~\eqref{eq:gamma_s_w}, and evolving DE. Finally, we stress that Eq.~\eqref{eq:gamma_s_w} provides a prominent role in approximating $\gamma$ for models I, II and III as we will outline below.
\section{Experimental limits}\label{sezione4}
One of the main purposes of this paper is to understand which model better approximates the Universe dynamics among the four paradigms described above.
Understanding which model is effectively the most suitable one to pass through higher redshift data domains in which the degeneracy problem is somehow healed is needed. In particular, to fix cosmological bounds over the different paradigms, we employed the standard low-redshift data surveys based on: observational Hubble data set (OHD) \citep{2002ApJ...573...37J}, SNe Ia with the {\it Pantheon} catalog \citep{2018ApJ...859..101S} and higher-redshift points coming from the data based on the use of the so-called growth function $f$ for large scale structure, together with the normalization of matter power spectrum, $\sigma_8$.
\subsection{Likelihood analysis}
Here, we perform a set of MCMC analyses involving all the above cases. The best set of parameters is hereafter dubbed {\bf x}, entering the total log-likelihood function, $\ln{\mathcal{L}}$
\begin{equation}
\ln{\mathcal{L}} = \ln{\mathcal{L}_{\rm OHD}} + \ln{\mathcal{L}_{\rm SN}} + \ln{\mathcal{L}_{\rm f}} + \ln{\mathcal{L}_{\rm \sigma_8}}\,.
\end{equation}
Below, we introduce the log-likelihood for each of the probes.
\begin{table}[!t]
\begin{center}
\renewcommand*{\arraystretch}{0.8}
\caption{\label{tab:HOD} $H(z)$ measurements for the OHD data used in the text, reported in units km\,s$^{-1}$\,Mpc$^{-1}$, with uncertainties (second column). In the first column we report the observed redshift $z$, whereas in the third column we report the reference paper in which the corresponding measurement has been first presented.}
\begin{ruledtabular}
\begin{tabular}{|ccc|}
$z$ &$ H \pm \sigma_H$ & Reference \\
\hline
0.0708 & $69.0 \pm 19.68$ & \cite{Zhang2014} \\
0.09 & $69.0 \pm 12.0$ & \cite{Jimenez2002} \\
0.12 & $68.6 \pm 26.2$ & \cite{Zhang2014} \\
0.17 & $83.0 \pm 8.0$ & \cite{Simon2005} \\
0.179 & $75.0 \pm 4.0$ & \cite{Moresco2012} \\
0.199 & $75.0 \pm 5.0$ & \cite{Moresco2012} \\
0.20 & $72.9 \pm 29.6$ & \cite{Zhang2014} \\
0.27 & $77.0 \pm 14.0$ & \cite{Simon2005} \\
0.28 & $88.8 \pm 36.6$ & \cite{Zhang2014} \\
0.35 & $82.1 \pm 4.85$ & \cite{Chuang2012} \\
0.352 & $83.0 \pm 14.0$ & \cite{Moresco2016} \\
0.3802 & $83.0 \pm 13.5$ & \cite{Moresco2016} \\
0.4 & $95.0 \pm 17.0$ & \cite{Simon2005} \\
0.4004 & $77.0 \pm 10.2$ & \cite{Moresco2016} \\
0.4247 & $87.1 \pm 11.2$ & \cite{Moresco2016} \\
0.4497 & $92.8 \pm 12.9$ & \cite{Moresco2016} \\
0.4783 & $80.9 \pm 9.0$ & \cite{Moresco2016} \\
0.48 & $97.0 \pm 62.0$ & \cite{Stern2010} \\
0.593 & $104.0 \pm 13.0$ & \cite{Moresco2012} \\
0.68 & $92.0 \pm 8.0$ & \cite{Moresco2012} \\
0.781 & $105.0 \pm 12.0$ & \cite{Moresco2012} \\
0.875 & $125.0 \pm 17.0$ & \cite{Moresco2012} \\
0.88 & $90.0 \pm 40.0$ & \cite{Stern2010} \\
0.9 & $117.0 \pm 23.0$ & \cite{Simon2005} \\
1.037 & $154.0 \pm 20.0$ & \cite{Moresco2012} \\
1.3 & $168.0 \pm 17.0$ & \cite{Simon2005} \\
1.363 & $160.0 \pm 33.6$ & \cite{Moresco2015} \\
1.43 & $177.0 \pm 18.0$ & \cite{Simon2005} \\
1.53 & $140.0 \pm 14.0$ & \cite{Simon2005} \\
1.75 & $202.0 \pm 40.0$ & \cite{Simon2005} \\
1.965 & $186.5 \pm 50.4$ & \cite{Moresco2015} \\
\end{tabular}
\end{ruledtabular}
\end{center}
\end{table}
\begin{itemize}
\item[(a)]\emph{Hubble rate likelihood}: To evaluate the Hubble rate likelihood, we notice that OHD points are cosmology-independent measurements of the Hubble rate at various $z$ through the differential age method \cite{2002ApJ...573...37J,ohd1}.
The Hubble rate is written by the identity $H(z)=-(1+z)^{-1}\Delta z/\Delta t$. Thus, from spectroscopic measurements of the age difference $\Delta t$, and redshift difference $\Delta z$, of couples of passively evolving galaxies that formed at the same time one infers the set of Hubble points \cite{ohd2}. The corresponding log-likelihood function is then given by
\begin{eqnarray}
\ln{\mathcal{L}_{\rm OHD}} & = & -\frac{1}{2}\sum_{i=1}^{N_{\rm OHD}}\ln{\left(2\pi\sigma_{\rm H_i}^2\right)} \nonumber\\
&& -\frac{1}{2}\sum_{i=1}^{N_{\rm OHD}}\left[\frac{H_{\rm i}-H\left({\bf x},z_i\right)}{\sigma_{\rm H_i}}\right]^2\,, \label{chisquared1}
\end{eqnarray}
where $N_{\rm OHD}$ corresponds to the OHD data points, as reported in Table~\ref{tab:HOD}.
\item[(b)] \emph{Pantheon likelihood}: The Pantheon data set is the most updated SN Ia sample composed of $1048$ sources \citep{2018ApJ...859..101S}. The standardization of their light curves involves the following corrections: 1) the luminosity-stretch coefficient $\alpha$ and factor $\mathcal{X}_1$ and 2) the luminosity-colour coefficient $\beta$ and factor $\mathcal{C}$, and the distance corrections $\Delta_{\rm M}$ and $\Delta_{\rm B}$, based on SN host galaxy mass and predicted biases, respectively. Once these corrections are applied, all SN Ia light curves become standard and the associated distance moduli are defined as
\begin{equation}
\mu = m_{\rm B} - \left(\mathcal{M} -\alpha \mathcal{X}_1 + \beta \mathcal{C} - \Delta_{\rm M} - \Delta_{\rm B}\right)\,,
\end{equation}
where $\mathcal{M}$ is the $B$-band absolute magnitude and $m_{\rm B}$ the $B$-band apparent magnitude \citep{2011ApJS..192....1C}. The log-likelihood function is given by \citep{2001A&A...380....6G}
\begin{equation}\label{eqn:chimarg}
\ln{\mathcal{L}_{\rm SN}} = -\frac{1}{2}\left(a + \ln{\frac{e}{2 \pi}} - \frac{b^2}{e}\right)\,,
\end{equation}
where $a\equiv\Delta\vec{\mathbf{\mu} }^{T}\mathbf{C}^{-1}\Delta\vec{\mathbf{\mu} }$, $b\equiv\Delta\vec{\mathbf{\mu} }^{T}\mathbf{C}^{-1}\vec{\mathbf{1}}$, and $e \equiv
\vec{\mathbf{1}}^T\mathbf{C}^{-1} \vec{\mathbf{1}}$, in which $\Delta\mu\equiv \mu - \mu_{\rm th}\left({\bf x},z\right)$ is the vector of residuals between the observed distance moduli, $\mu$, and the theoretical ones, $\mu_{\rm th}$. Finally, $\mathbf{C}$ is the covariance matrix which is related to statistical and intrinsic systematic uncertainties of SNe \citep{2011ApJS..192....1C}. For the sake of clarity, with SNe Ia only the $H_0$ value cannot be constrained. This is another reason to combine such a data set with the others, presented here.
\item[(c)] \emph{Matter growth likelihood}: In linear theory, the growth of matter fluctuations is described by the growth function $f$, defined by Eqs.~(\ref{eqn:exact_f}) and (\ref{eq:f}). The log-likelihood is therefore given by \cite{mat1,mat2}
\begin{equation}\label{flikelihood}
\ln{\mathcal{L}_{\rm f}} = -\frac{1}{2}\sum_{i=1}^{N_{\rm f}}\left\{\ln{\left(2\pi\sigma_{\rm f_i}^2\right)} + \left[\frac{f_{\rm i}-f\left({\bf x},z_i\right)}{\sigma_{\rm f_i}}\right]^2\right\}\,.
\end{equation}
\item[(c)] \emph{$\sigma_8$ likelihood}: An alternative observational probe of $\delta(z)$ is the rms mass fluctuation $\sigma_8(z)$. It is linked to $\delta(z)$ via
\begin{equation}
\sigma_8(z) = \sigma_8(0)\frac{\delta(z)}{\delta(0)} = \sigma_8(0) e^{\int^{\frac{1}{1+z}}_{1}\Omega_{\bf m}^\gamma(a)\mathrm{d}a}\,,
\end{equation}
where $\sigma_8(0)$ is the value at $z=0$. Most of the currently available data points originate from the observed redshift evolution of the flux power spectrum of the Ly--$\alpha$ forest \cite{mat1,mat2}. To avoid the use of the additional parameter, $\sigma_8(0)$, in the fitting procedure, an alternative parameter can be used, i.e.,
\begin{equation}
s_8(z_1,z_2) = \frac{e^{\int^{\frac{1}{1+z_1}}_{1}\Omega_{\bf m}^\gamma(a)\mathrm{d}a}}{e^{\int^{\frac{1}{1+z_2}}_{1}\Omega_{\bf m}^\gamma(a)\mathrm{d}a}}\,.
\end{equation}
The corresponding log-likelihood becomes
\begin{eqnarray}
\ln{\mathcal{L}_{\rm s8}} & = & -\frac{1}{2}\sum_{i=1}^{N_{\rm s_8}}\ln{\left(2\pi\sigma_{\rm s_{8,i}}^2\right)} \nonumber\\
&& -\frac{1}{2}\sum_{i=1}^{N_{\rm s_8}}\left[\frac{s_{\rm 8,i}-s_8\left({\bf x},z_i,z_{i+1}\right)}{\sigma_{\rm s_{8,i}}}\right]^2\,,\label{chisquared2}
\end{eqnarray}
where $\sigma_{\rm s_{8,i}}$ is derived by error propagation from the errors of $s_8(z_i)$ and $s_8(z_i+1)$.
\end{itemize}
\begin{table}[!t]
\caption{Statistical comparison among the $\Lambda$CDM, $\omega$CDM paradigms, the CPL parametrization, and models I--IV. The reference scenario is $\Lambda$CDM model, so that $\Delta{\rm AIC(BIC)}={\rm AIC(BIC)}_{\rm i}-{\rm AIC(BIC)}_{\rm \Lambda CDM}$.}
\centering
\begin{tabular}{|l|ccccc|}
\hline
\hline
Model & $\ln \mathcal{L}_{max}$ & AIC & BIC & $\Delta$AIC & $\Delta$BIC\\
\hline
\hline
$\Lambda$CDM & $-640.24$ & $1284.48$ & $1294.50$ & $0$ & $0$\\
$\omega$CDM & $-640.48$ & $1286.97$ & $1301.99$ & $2.48$ & $7.49$\\
CPL & $-640.22$ & $1288.44$ & $1308.48$ & $3.96$ & $13.98$\\
\hline
I & $-639.35$ & $1284.69$ & $1299.72$ & $0.21$ & $5.22$\\
II & $-895.91$ & $1795.83$ & $1805.84$ & $511.32$ & $511.32$\\
III & $-640.30$ & $1284.64$ & $1294.66$ & $0.16$ & $0.16$\\
IV & $-640.29$ & $1284.58$ & $1294.59$ & $0.09$ & $0.09$\\
\hline
\hline
\end{tabular}
\label{tab:params1}
\end{table}
\section{Statistical analyses}\label{sezione4bis}
We now present the statistical performances of the various models through the use of the Bayesian selection criteria to measure the evidence of a given model against a reference scenario \citep{Kunz2006}, conventionally chosen as the $\Lambda$CDM paradigm. Specifically, we consider the Akaike information criterion (AIC) \citep{Akaike1974} and the Bayesian information criterion (BIC) \citep{Schwarz1978}, defined, respectively, as
\begin{subequations}
\begin{align}
\text{AIC} \equiv & -2\ln{\mathcal{L}}_{\rm max} + 2p \,, \\
\text{BIC} \equiv & -2\ln{\mathcal{L}}_{\rm max} + p\ln{N}\,.
\end{align}
\end{subequations}
Here, $\mathcal{L}_{\rm max}$ is the maximum likelihood estimate, $p$ is the number of free parameters of the model, and $N$ the total number of data points. We note that, for high $N$, the BIC criterion penalizes the model with a large number of free parameters more severely than AIC.
Using these definitions, we calculated the differences $\Delta$AIC and $\Delta$BIC with respect to the reference scenario to measure the amount of information lost by adding extra parameters in the statistical fitting. Negative values of $\Delta$AIC and $\Delta$BIC would indicate that the model under investigation performs better than the reference model, while for positive values, one needs to know that
\begin{itemize}\label{AICintervals}
\item[(a)] $\Delta\text{AIC(BIC)}\in [0,2]$ indicates a weak evidence in favour of the reference model, leaving open the question on which model is the most suitable one;
\item[(b)] $\Delta\text{AIC(BIC)}\in (2,6]$ indicates a mild evidence against the given model with respect to the reference paradigm;
\item[(c)] $\Delta\text{AIC(BIC)}> 6$ indicates a strong evidence against the given model, which should be rejected.
\end{itemize}
We report the $\Delta$AIC and $\Delta$BIC values for the different cosmological models in Table \ref{tab:params1}. They assume positive values for all the models indicating that data are either weakly or strongly in favor of the $\Lambda$CDM model. In more detail, according to the AIC criterion, the DE models (CPL and $\omega$CDM) are quite disfavored with respect to the reference model, while model II is strongly disfavored. Models I, III and IV are only weakly disfavored, in that $\Delta$AIC$<1$, so one can not really establish which model is favored. According to the BIC criterion, all the models are rather strongly disfavored (once again, model II in particular), but III and IV are slightly favored for the $\Lambda$CDM model, as $\Delta$BIC$<1$.
\begin{table*}[!t]
\caption{Best-fit parameters for flat $\Lambda$CDM, $\omega$CDM, CPL, and models I--IV with $1$--$\sigma$ (and $2$--$\sigma$) errors. Hereafter, $h_0\equiv H_0/100s^{-1}$Mpc/km.}
\centering
\begin{tabular}{|l|cccccc|}
\hline
\hline
Model & $h_0$ & $\Omega_{\rm m,0}$ & $\Omega_{\rm cdm,0}$ & $n$ & $w_0$& $w_1$\\
\hline
$\Lambda$CDM & $0.695^{+0.027 (+0.045)}_{-0.027 (-0.043)}$ & $0.290^{+0.028 (+0.047)}_{-0.026 (-0.042)}$ & - & - & - & - \\
$\omega$CDM & $0.693^{+0.033 (+0.047)}_{-0.033 (-0.049)}$ & $0.270^{+0.047 (+0.070)}_{-0.040 (-0.056)}$ & - & - & $-0.92^{+0.10 (+0.15)}_{-0.13 (-0.21)}$ & - \\
CPL & $0.690^{+0.039 (+0.057)}_{-0.036 (-0.049)}$ & $0.269^{+0.052 (+0.076)}_{-0.051 (-0.071)}$ & - & - & $-1.02^{+0.27 (+0.37)}_{-0.19 (-0.22)}$ & $0.63^{+0.72 (+0.74)}_{-1.79 (-2.54)}$\\
\hline
I & $0.689^{+0.034 (+0.053)}_{-0.033 (-0.046)}$ & $0.308^{+0.040 (+0.062)}_{-0.039 (-0.057)}$ & - & $0.004^{+0.015 (+0.027)}_{-0.009 (-0.022)}$ & - & - \\
II & $0.492^{+0.017 (+0.028)}_{-0.017 (-0.028)}$ & $0.510^{+0.070 (+0.118)}_{-0.067 (-0.109)}$ & - & -1 & - & - \\
III & $0.695^{+0.027 (+0.044)}_{-0.028 (-0.044)}$ & $0.291^{+0.028 (+0.046)}_{-0.026 (-0.042)}$ & - & 0 & - & - \\
IV & $0.693^{+0.028 (+0.045)}_{-0.028 (-0.045)}$ & $0.291^{+0.029 (+0.049)}_{-0.028 (-0.045)}$ & $0.269^{+0.029 (+0.049)}_{-0.028 (-0.045)}$ & - & - & - \\
\hline
\hline
\end{tabular}
\label{tab:params2}
\end{table*}
\begin{figure*}[!ht]
\centering
\includegraphics[width=1.2\columnwidth,clip]{MCMC_n.eps}\hfill
\includegraphics[width=0.8\columnwidth,clip]{MCMC_nm1.eps}
\includegraphics[width=0.8\columnwidth,clip]{MCMC_n0.eps}\hfill
\includegraphics[width=0.8\columnwidth,clip]{MCMC_4.eps}
\caption{$1$--$\sigma$ (dark gray) and $2$--$\sigma$ (light gray) contour plots and best-fit parameters (black dot). In the top left (right) panel we show model I (II), while in the bottom left (right) panel, we present model III (IV).}
\label{fig:MCMC}
\end{figure*}
\section{Discussion on theoretical results}\label{sezione5}
In this section, we describe all our numerical findings. They have been reported in Table~\ref{tab:params2}, where we included best fit parameters, and in Table~\ref{tab:params1}, for the AIC and BIC values. We also presented the results of our MCMC runs in Fig.~{\ref{fig:MCMC}}, where we show models I and II in the top panels and models III and IV in the bottom panels.
In all our fits, $\rho_{\star}$ corresponds to the Planck density, $\rho_{\rm Pl}$, in agreement with theoretical predictions. This assumption works well even for high redshift data, except for model II. In fact, the choice $n=-1$ is clearly ruled out by using the intermediate redshift data that we employed.
Moreover, for model II even if $\rho_{\star}\neq\rho_{\rm Pl}$ (letting $\rho_{\star}$ be free to vary) the model does not work well with our data catalogs at small and intermediate redshifts.
Hubble constant constraints, for all our models, except again for model II, appear in tension with what found by \cite{Riess2019}. The error bars up to $2$--$\sigma$ confidence levels are not large enough to avoid the tension with Riess measurements at $\geq 1$--$\sigma$. In principle such a result may be affected by systematics and may be reconsidered in view of future developments. In all our cases, the Hubble parameters are compatible with Planck results \cite{Planck2016}.
Focusing on each model, we can compare our results with the concordance $\Lambda$CDM paradigm, and with two evolving DE scenarios, i.e., here the $\omega$CDM model and the CPL parametrization. In model I, $n$ severely differs from the outcomes obtained in \cite{Capozziello}. This is closely related to the higher redshift data sets that we adopted here. This fact, if confirmed by future analyses, suggests the first model is non-unequivocally constrained by cosmic data, leading to a clear limitation of the model itself. In particular, the value of $n$ is positive but very small. We show that if one adopts only small redshift data, the model is quite unbounded and the predictions over $n$ are unconstrained, i.e., severely different from our findings. In addition, here $n$ is compatible with zero at both $1$-- and $2$--$\sigma$ confidence levels. The values for both matter density $\Omega_{\rm m,0}$ and Hubble parameter are compatible with the $\Lambda$CDM scenario at $1-\sigma$ confidence level, making it, according to the AIC criterion, only slightly disfavored.
Our numerical results remark that negative $n$ are excluded for Anton-Schmidt frameworks, in disagreement with previous studies, e.g., \cite{Capozziello}. In view of this, it is easy to stress again that even statistical criteria establish that model II is highly disfavored by data.
The genuine logotropic paradigm, model III, works quite well in describing the Universe dynamics. Both matter density and the Hubble function are in good agreement with the values found for the reference $\Lambda$CDM model. As a consequence, both the AIC and the BIC are small, $(\sim 0.16)$, showing that the $\Lambda$CDM model is only weakly favored. The genuine logotropic model is very similar to model IV, in which one includes an additional parameter, the baryon density $\Omega_{\rm b}$. As it is not possible to constrain $\Omega_{\rm b}$ within our redshift range, we take it fixed to the Planck constraint, $\Omega_{\rm b}\simeq 0.02242$. The pure logotropic model is largely more predictive than Anton-Schmidt paradigms, certifying that:
\begin{itemize}
\item[(a)] in general, Anton-Schmidt paradigms (models I and II) seem to be less predictive than pure logotropic models (models III and IV),
\item[(b)] generalized versions of the pure logotropic model (model III) do not significantly change the experimental expectations over the free coefficients.
\end{itemize}
An immediate interpretation of the above two points is the following: the Anton-Schmidt models are disfavored than pure logotropic scenarios because of the approximations made on $n$. In fact, as already stated above, in the original Anton-Schmidt picture, $n$ is a general function of the temperature. This fact could severely influence the goodness of Anton-Schmidt models, for example enabling the sound speed to be always positive definite.
However, comparing the AIC and BIC criteria as reported in Table~\ref{tab:params1}, we notice only a weak evidence for the statistical goodness of logotropic models with respect to Anton-Schmidt paradigms. The statistical significance of logotropic models, for both models III and IV, is lower than the concordance paradigm, as a consequence of the additional free parameters. In other words, the standard cosmological model is still favored in describing the cosmic expansion history even with respect to genuine logotropic models. Foe the sake of completeness, it is interesting to note that for models I (top left panel), III and IV (bottom panels) there is a small anticorrelation between $h_0$ and $\Omega_{\rm m}$, while this does not occur for model II. This is one of the reasons why this model performs so badly. Note that there is also a positive correlation between the free index $n$ and the matter density parameter $\Omega_{\rm m,0}$: this means that it is difficult to modify $n$ without finding unreasonable values of $\Omega_{\rm m,0}$.
Analogous comparisons can be performed among our models and evolving DE frameworks. So, regarding the CPL parametrization, both parameters ($w_0$ and $w_1$) are compatible with the cosmological constant scenario, i.e., $w_0=-1$ and $w_a=0$. Whilst the value of $h_0$ is very similar to what we found for our reference model, the matter density parameter is about $7\%$ smaller, which corresponds to less than $1$--$\sigma$ difference. Nevertheless, this model has two additional parameters with respect to the $\Lambda$CDM model and is, therefore, heavily penalized by the selection criteria. This explains the large $\Delta$AIC and $\Delta$BIC values, compared with the concordance paradigm. Similar considerations can be done for the $\omega$CDM model, for which we found values very similar to the CPL parametrization. It is interesting to notice that with $w_0=-0.92$, the model is in the ``quintessence regime", but still compatible to a $\Lambda$CDM model within $1$--$\sigma$. So, concerning the $\omega$CDM model, the $\Delta$AIC and $\Delta$BIC are lower than for the CPL one, making it statistically preferred with respect to the latter. This is clearly due to the smaller number of parameters. Comparing our paradigms with these results show that models I, III and IV seem to be more statistically favored than CPL and $\omega$CDM frameworks. However, we note that the statistical significance is weak and there is no reason \emph{a priori} to imagine that these models behave better than $\omega$CDM and CPL throughout the Universe evolution.
\subsection{Properties of the best fit models}
We now study the background properties, the adiabatic sound speed for linear perturbations and the evolution of the growth factor and index for our models.
\begin{figure}[!ht]
\centering
\includegraphics[width=\columnwidth,clip]{Total.eps}
\caption{Plots of the thermodynamic functions described for our four models in terms of $a$. Hereafter, for brevity we removed any superscripts. From top to bottom: the Hubble rate $H(a)$, normalized with the $\Lambda$CDM model, $H_{\Lambda{\rm CDM}}$, the total EoS parameter, $\omega$, the DE EoS, $\omega_{\rm de}$, and the adiabatic sound speed, $c_{\rm s,a}^2$. The values of the chosen parameters are summarized in Table~\ref{tab:params2}.}
\label{ris:image01}
\end{figure}
\begin{figure}[!ht]
\centering
\includegraphics[width=\columnwidth,clip]{Total2.eps}
\caption{The linear density perturbations over the scale factor $\delta/a$ (top panel) and the growth $f$ (bottom panel) as a function of $a$ for the four models. The values of the parameters chosen to plot each model are summarized in Table~\ref{tab:params2}.}
\label{ris:image02}
\end{figure}
In Fig.~\ref{ris:image01}, we present the evolution of the Hubble parameter normalized with the reference $\Lambda$CDM model (upper panel), the total EoS $\omega$ (second panel), the EoS for the DE component only, namely $\omega_{\rm de}$ (third panel), and the adiabatic sound speed (bottom panel) for the four models underlined in this work. For comparison, we also plot the expectations obtained from the $\Lambda$CDM, $\omega$CDM and CPL scenarios. From a qualitative point of view, model II with $n=-1$ is the one showing the largest differences with respect to the other logotropic models and to the two DE models considered in this work. Analyzing the background evolution from a more quantitative point of view, we note that model II deviates from the others at all redshifts, with differences much higher than 10\%. All the other models differ by at most a few percent, with a maximum of 5\% at early times. It is worth noticing that at late time the CPL model expansion rate is higher than for the $\Lambda$CDM model, while for $a<0.2$ it is a few percent smaller.
In the third panel from top, we present the evolution of the DE EoS $\omega_{\rm de}$. As expected, for the $\Lambda$CDM ($\omega$CDM) model, $\omega_{\rm de}=-1$ ($\omega_{\rm de}=-0.92$) constant throughout the cosmic history and for the CPL model $\omega_{\rm de}$ grows steadily as $w_1=0.63$. The evolution of the logotropic models is instead more interesting. For models I--III, the EoS of the DE component is rather constant and very close to $-1$, therefore having a behavior similar to that of the cosmological constant, $\Lambda$. Model IV, instead, is close to the cosmological constant up to $a=0.6$ and at earlier times it grows rapidly such that at $a=0.2$, we have $\omega_{\rm de}=-0.5$. This discussion helps us to understand the behavior of the total EoS of our models, presented in the second panel from top. In particular, for model II, the total EoS is practically null over the whole cosmic history, explaining why the model represents a very poor fit to the data.
The quantity of more interest for the evolution of perturbations is the adiabatic sound speed of each models. For adiabatic models, this quantity determines the stability of a given model against perturbations and ultimately its validity.
This happens because, in general, we can write the evolution of perturbations as
\begin{equation}
\ddot{\delta} + A(t)\dot{\delta} + c_{\rm s}^2\delta = 0\,,
\end{equation}
where $A(t)$ represents a generic damping term associated with the cosmic expansion. We also generically denoted the sound speed as $c_{\rm s}^2$, as this discussion has general validity. When $c_{\rm s}^2>0$, the solution is a damped harmonic oscillator, hence, perturbations are stable and bound. On the contrary, when $c_{\rm s}^2<0$, its solution is an exponential function and perturbations grow unbounded. Thus, the model is unstable and we can see the developing instability in the top panel. The instability is evident at very early times, where the quantity $\delta/a$ appears to be diverging. This is also reflected in the growth where $f$ is approximately constant. In fact, if $f\approx c_f$, with $c_f$ constant, then $\delta\propto a^{c_f}$ at all times (at odds with the physical intuition that an accelerated expansion of the Universe would lead structures to grow less).
All the models, except IV, have a roughly constant sound speed, which allows us to determine an approximate solution for the growth index $\gamma$, as shown in Sec.~\ref{sezione3}. Models I and III have a small (close to zero) sound speed. At early time the sound speed is small and clustering properties are due to DM, while at late time the sound speed slightly grows, to reach the value of $\approx 2\times 10^{-2}$. Model II has a similar behavior, with an adiabatic sound speed squared constant but negative ($c_{\rm s}^2\approx -10^{-2}$). This implies that the model is unstable and explains once more why the model performs badly against the observational data. Finally, for model IV, at early times, the adiabatic sound speed is positive and of the order of $5\times 10^{-2}$, but at late times it becomes negative, reaching in absolute value approximately the same value at early times. Since the sound speed is in general also small when positive, the effects of instabilities at late times can be substantial.
The discussion about the sound speed is a useful introduction to the study of the evolution of the density perturbations and the growth. Models I and III, as expected from the previous analysis (especially in terms of Bayesian evidence) show a behavior very close to that of the $\Lambda$CDM model in terms of $\delta$, but differ more in the evolution of $f$ (see Fig.~\ref{ris:image02}). This is easily understood remembering that the growth is directly proportional to the time derivative of $\delta$ and that differential quantities highlight differences. It is, therefore, crucial, to also consider this quantity when comparing models. Note that model II is fully nonpredictive as stressed several times in this work and $f_{\rm CPL}\simeq f_{\rm \Lambda CDM}$. We now consider model IV. To do so we notice that throughout the cosmic history, perturbations grow similarly to the $\Lambda$CDM model. It is, however, more interesting to consider its logarithmic derivative, $f$. Even if its evolution is qualitatively different from that of the reference model, we notice a peculiar behavior at very early and late times. Thus, at early times, $f$ first flattens and is of the order unity as expected, but then it tends to decrease. This is symptomatic of a departure from a limiting Einstein-de Sitter regime in that the sound speed at early times is small but non-negligible. At late times, we see the effects of a negative sound speed: since perturbations now grow unbounded, the late time cosmic acceleration does not suppress structure formation and $f$ first flattens and then starts to grow again. This makes the value of the growth more similar to the $\Lambda$CDM model, thus explaining why, despite the unwanted feature of negative sound speed, the model represents a good fit to the data. In other words, since the instabilities arising at low redshift ($z\approx 0.33$), they did not have the time to considerably affect the evolution of perturbations. Finally, the time evolution of the growth index $\gamma$ (Fig.~\ref{ris:image03}) for models where the adiabatic sound speed and the DE EoS are approximately constant, is approximately constant as well, justifying our simplifying assumption. In general, $c_{\rm s,a}^2$ and $\omega_{\rm de}$ can be well approximated to a constant for the $\Lambda$CDM and $\omega$CDM models and for the logotropic models I--III, but this is not the case for the CPL parametrization and for the model IV introduced in this work. It is interesting to notice that for models I and III, $\gamma$ evolves like the corresponding adiabatic sound speeds. This is easy to understand as this is the only relevant quantity evolving in our analytical approximation. For model IV our approximation is no longer valid since both $\omega_{\rm de}$ and $\gamma$ evolve rather strongly with time. The effects of the negative adiabatic sound speed emerge in the behavior of the growth index: since its value reduces and it is about 30\% lower than for the $\Lambda$CDM model, the growth would be enhanced, despite the recent accelerated expansion.
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth,clip]{Total3.eps}
\caption{The growth index $\gamma$ computed numerically from Eq.~\eqref{totalgamma} (top panel) and from the analytical solution in Eq.~\eqref{eq:gamma_s_w} (bottom panel) as a function of $a$ for the four models. The parameters chosen to plot each model are summarized in Table~\ref{tab:params2}. Each model is displayed, as indicated in the legend. Line styles will be the same for all the following figures.}
\label{ris:image03}
\end{figure}
Analyzing both panels in Fig.~\ref{ris:image03} more quantitatively, we see that our analytical prediction reproduces the numerical solution at the percent level, with the exception, as already discussed, of model IV. For this model, only its qualitative behavior is reproduced.
\section{Final remarks}\label{sezione6}
In this paper, we investigated four thermodynamic models that fuel current acceleration through a single fluid unifying DE with DM. The four models under examination are part of two typologies: Anton-Schmidt and logotropic fluids, intertwined among them by fixing the Gr\"uneisen index to precise values. These scenarios have recently been introduced as alternatives to the standard cosmological model with the purpose of extending modified Chaplygin gas models. In particular, we investigated the Anton-Schmidt paradigms for two relevant cases of the Gr\"uneisen index. Afterwards, we studied the logotropic framework, corresponding to $n=0$, and then we proposed a novel approach, namely model IV, in which DM and baryons are disentangled.
We first discussed the basic features of each model and then evaluated the general picture of small perturbations, computing the growth $f=\mathrm{d}\ln{\delta}/\mathrm{d}\ln{a}$ (with $\delta$ the linearly perturbed density) and growth index $\gamma$ for each model. We portrayed and confronted the functional behaviors of the adiabatic sound speed, total and DE EoS and Hubble parameters for each model.
Further, we compared our models with data at intermediate redshifts, through a MCMC procedure based on the Metropolis algorithm. To do so, we took into account SNeIa, OHD and linear growth observations for the large scale structures. We fixed tight constraints over the thermodynamic free parameters of our four models and we inferred the statistical significance of them through the AIC and BIC statistical criteria.
We showed that the logotropic models behave better than Anton-Schmidt ones and the case $n=-1$ is fully ruled out.
Additionally, we demonstrated that the logotropic models are less stable than the standard cosmological paradigm, stressing that there is no statistical advantage to handle a logotropic and/or Anton-Schmidt universe, if compared with the $\Lambda$CDM, $\omega$CDM and CPL scenarios. In support of this, we outlined where perturbations become unstable due to negative sound speed for all the underlying models.
In the future, we will investigate a possible time dependence of the Gr\"uneisen index $n$ for Anton-Schmidt models, as predicted by solid state physics, and check if it would enable the sound speed to be positive definite at all times. We will also attempt unifying DE with inflationary epochs through a single description based on Anton-Schmidt and/or logotropic models.
\begin{acknowledgments}
K. B., O. L. and M. M. acknowledge the Ministry of Education and Science of the Republic of Kazakhstan, Grant No. IRN AP08052311. MM is supported by INFN as part of the MoonLIGHT-2 experiment in the framework of the research activities of CSN2. F. P. acknowledges the support from the Grant No. ASI n.2018-23-HH.0.
\end{acknowledgments}
|
\section{Introduction}
The advection of scalar fields by external flows enters in a wide
range of phenomena, such as the fluctuations of temperature or salinity
in the ocean, or the dispersion of pollutants in the atmosphere.
When the carrier
flow is turbulent, understanding
the statistical properties and the mixing of the scalar
become a difficult and fundamental issue.
A simplifying assumption is to consider as passive
the advected scalar field, which means that it has a negligible backreaction onto the
carrier flow, such that the dynamics of the latter is unaffected.
This assumption is realistic in particle-ladden flows only at sufficiently small concentration.
However, even in this simpler case, the complete
characterization of the properties of the scalar is still challenging, and we focus
on this case in this work.
For a high-Reynolds number carrier flow,
one can distinguish several regimes for the scalar
depending on the Schmidt number, which is the ratio
of the fluid viscosity to the scalar diffusivity.
We refer to Ref.~\onlinecite{Sreenivasan18175} for an overview.
In this work, we focus on the inertial-convective range, which corresponds
to Schmidt numbers of order one, and which spans scales between
the energy injection scale of the scalar and the scale at which energy is dissipated,
for typical scales of the velocity of the turbulent fluid in the inertial range.
The spectrum of the scalar field in the inertial-convective range
was established by Obukhov and Corrsin \cite{Obukhov49,Corrsin51}.
Other pioneering studies are due to Yaglom, Batchelor, and Kraichnan
\cite{Yaglom49,batchelor_1959,Kraichnan_PoF_68}. Barring possible
corrections due to intermittency, the scalar spectrum in this
range is determined via the energy cascade picture and shows a power-law decay with the
same exponent $-5/3$ in three dimensions as the spectrum of the turbulent fluid carrying it.
The spectrum is related to the two-point correlation function of the scalar field
at equal times. Characterizing its behavior at different times
is also fundamental to understand the mixing properties of the scalar.
In particular, determining the temporal dependence of the Eulerian correlations of the scalar
is a difficult task and few is known so far on their properties.
In this work, we use the framework of the
functional renormalization group (FRG) (see Ref.~\onlinecite{Delamotte:2007pf} for an introduction and Ref.~\onlinecite{Dupuis:2020fhh} for a recent review) to make progress on this issue.
The FRG approach was recently employed to study Navier-Stokes (NS) turbulence, where it proved
to be fruitful. By exploiting the symmetries of the field
theoretical formulation of the stochastically forced Navier-Stokes equation, it
yielded predictions for the general form of the spatio-temporal dependence
of any $n$-point correlation function of the turbulent velocity field in the limit of large wavenumbers \cite{Canet:2016nmg,Tarpin:2017uzn}.
These predictions were accurately confirmed
by Direct Numerical Simulations \cite{Canet:2016nmg,Gorbunova:2020uxl,Gorbunova:2021a}.
The FRG approach was also extended to study the direct cascade in 2D turbulence \cite{Tarpin:2018yvs}.
In this work, we follow a similar strategy for the passive scalar, in order to construct
an approximation scheme which is not based on a
small-coupling expansion, and thus offers a nonperturbative approach.
Exploiting the symmetries of the field theory of the advected scalar,
we provide the general expression of the temporal dependence of the two-point Eulerian correlations
of the scalar field in the limit of large wavenumbers.
These expressions are obtained both for
a scalar advected by a turbulent NS flow, and for a scalar advected by a ``synthetic'' stochastic velocity field, as prescribed in the Kraichnan model
\cite{Kraichnan_PoF_68,Kraichnan94prl}. In the latter case,
the velocity statistics is Gaussian and the associated covariance is white-in-time. Although these
delta-correlations yield simpler explicit expressions, they have a significant impact on the
temporal behavior of the scalar field at small time. Indeed, we show that for a scalar in a turbulent NS flow,
the two-point correlation function decays at small $t$ as a Gaussian in the variable $tk$ where $t$ is the time delay
and $k$ the wavenumber, and as an exponential at large $t$.
The scalar field in the inertial-convective regime thus inherits in this case temporal correlations which are
very similar to the ones of the turbulent fluid (reported {\it eg.} in Ref.~\onlinecite{Tarpin:2018yvs}).
For the Kraichnan scalar, the two-point correlations always decay exponentially, the Gaussian small-time behavior is completely washed out by the instantaneous decorrelation of the synthetic velocity field.
The paper is organized as follows.
In Sec.~\ref{sec:Kraichnan-model}, we introduce the Kraichnan model,
the associated field theory, and we expound its symmetries and the related Ward identities.
In Sec.~\ref{sec:preliminaries-and-non-renormalization-th},
we establish some preliminary results for this model that will be used to derive a closed
flow equation for the correlation functions of the Kraichnan scalar in Sec.~\ref{sec:closed-FRGE-Kraichnan}.
In Sec.~\ref{sec:closed-FRGE-for-advected-scalars}, we generalize
the derivation and results to the case of scalar fields advected by turbulent NS flows.
We summarize our findings in Sec.~\ref{sec:summary}.
In Appendix \ref{app:Yaglom_rel}, we discuss the Yaglom relation
from a field theoretical viewpoint, and in Appendix \ref{app:dissipative},
we establish the expression of the spectrum in the near-dissipative range from
the Dyson equation.
\section{Field theory and symmetries of the Kraichnan model} \label{sec:Kraichnan-model}
The Kraichnan model was proposed as
a simplified model of scalar turbulence where the advecting velocity field
is taken as a random vector field with Gaussian statistics to replace
the NS fluctuating turbulent velocity field \cite{Kraichnan_PoF_68}.
The key simplifying feature of this model is that the velocity covariance
is white in time, $\langle v^{i}\left(t,\vec{x}\right)v^{j}\left(t^{\prime},\vec{x}\right)\rangle\propto\delta\left(t-t^{\prime}\right)$.
This model has been widely studied since many aspects can be investigated analytically.
In particular, several techniques have been employed to compute
intermittency corrections via suitable expansions \cite{ChertkovETal1995,Gawedzki:1995zz,ChertkovFalkovich1996,Bernard:1996um,BernardGK1998}.
In the perturbative regime, intermittency corrections were computed also
via various field theoretical approaches \cite{PhysRevE.58.1823,PhysRevE.63.025303,PhysRevE.58.7381,Kupiainen:2006em}, including the functional renormalization group \cite{Pagani:2015hna}.
In this section, we present the path integral formulation of the Kraichnan
model, introduce the FRG formalism, and then discuss the symmetries of the Kraichnan field theory and
the associated Ward identities. We shall focus on incompressible flows
and take the velocity field to be divergenceless.
\subsection{The Kraichnan model}
The Kraichnan model describes the advection and diffusion of a scalar
field $\theta\left(t,\vec{x}\right)$ following the dynamics
\begin{eqnarray}
\partial_{t}\theta\left(t,\vec{x}\right)+v^{i}\left(t,\vec{x}\right)\partial_{i}\theta\left(t,\vec{x}\right)-\frac{\kappa}{2}\partial^{2}\theta\left(t,\vec{x}\right) & = & f\left(t,\vec{x}\right)\,,\label{eq:Kraichnan-model_SDE}
\end{eqnarray}
where $\kappa$ is the molecular diffusivity and $f\left(t,\vec{x}\right)$
is a stochastic forcing characterized by Gaussian statistics with
covariance $\langle f\left(t,\vec{x}\right)f\left(t^{\prime},\vec{y}\right)\rangle=\delta\left(t-t^{\prime}\right)M\left(\frac{|\vec{x}-\vec{y}|}{L}\right)$. In this model,
the velocity $v^{i}\left(t,\vec{x}\right)$ is a stochastic vector field chosen with
a Gaussian distribution. It has zero average and the following covariance:
\begin{eqnarray}
\Bigr\langle v^{i}\left(t,\vec{x}\right)v^{j}\left(t^{\prime},\vec{y}\right)\Bigr\rangle & = & \delta\left(t-t^{\prime}\right)D_{0}\int_{\vec{p}} \frac{e^{i \vec{p}\cdot \left(\vec{x}-\vec{y}\right)} \,P_{ij}(\vec{p})}{\left(p^{2}+m^{2}\right)^{\frac{d}{2}+\frac{\varepsilon}{2}}}
\,,\label{eq:vel-2pt-correlation-function}
\end{eqnarray}
with $P_{ij}(\vec{p})=\delta_{ij}-\frac{p_{i}p_{j}}{p^{2}}$ the transverse
projector. The mass $m$ in Eq.~(\ref{eq:vel-2pt-correlation-function})
prevents any infrared divergence in the velocity covariance.
Note that we use for the integrals the
shorthand notation
\begin{equation*}
\int_{t\vec{x}} = \int dt\int d^{d}\vec{x}\quad,\quad \quad \int_{\omega \vec{q}} = \int\frac{d\omega}{2\pi}\int\frac{d^{d}\vec{q}}{\left(2\pi\right)^{d}}\,
\end{equation*}
and for the Fourier transforms the convention
\begin{eqnarray}
g\left(\omega,\vec{p}\right) & \equiv & \int_{t\vec{x}}e^{i\omega t-i\vec{p}\cdot\vec{x}}g\left(t,\vec{x}\right),\nonumber\\
g\left(t,\vec{x}\right) & \equiv & \int_{\omega \vec{p}}e^{-i\omega t+i\vec{p}\cdot \vec{x}}g\left(\omega,\vec{p}\right)\,.
\label{eq:Fourier}
\end{eqnarray}
In the limit $\varepsilon\rightarrow0$, the coinciding-point
covariance (\ref{eq:vel-2pt-correlation-function}) diverges logarithmically.
Formally, a finite limit can be obtained by defining $D_{0}\equiv D_{0}^{\prime}\varepsilon$.
We shall use either $D_{0}$ or $D_{0}^{\prime}$, keeping in mind that they are just fixed numbers for a
given value of $\varepsilon$.
The stochastic differential equation (SDE) Eq.~(\ref{eq:Kraichnan-model_SDE})
is a multiplicative one and is interpreted
in the Stratonovich sense. It can also be rewritten
in the Ito sense, but we shall work here with the Stratonovich convention unless otherwise stated.
The Martin-Siggia-Rose-Janssen-de Dominicis (MSRJD) formalism provides
a well established procedure to map a SDE
to an equivalent path integral formulation \cite{Martin73,Janssen76,Dominicis76,Dominicis79}.
The MSRJD approach allows one to compute an arbitrary correlation
function by introducing the appropriate generating functionals.
A key advantage of employing the MSRJD formalism is that
it allows one to compute Eulerian correlation functions at different
times in a simple way.
The generating functional for the Kraichnan model is obtained as
\begin{equation}
{Z} = \int {\cal D}\theta{\cal D}\bar{\theta} {\cal D}c{\cal D}\bar{c}{\cal D}\vec{v} e^{-S_{{\rm Kr}} + \int_{t,\vec{x}} \phi {J}}
\end{equation}
with the following action:
\begin{eqnarray}
S_{{\rm Kr}} & = & \int_{t\vec{x}}\,\bar{\theta}\left(t,\vec{x}\right)\Biggr\{\partial_{t}\theta\left(t,\vec{x}\right)+v^{i}\left(t,\vec{x}\right)\partial_{i}\theta\left(t,\vec{x}\right)-\frac{\kappa}{2}\partial^{2}\theta\left(t\vec{x}\right)\Biggr\}\nonumber \\
& & -\frac{1}{2}\int_{t\vec{x} \vec{y}} \bar{\theta}\left(t,\vec{x}\right)M\left(\frac{\vec{x}-\vec{y}}{L}\right)\bar{\theta}\left(t,\vec{y}\right)\nonumber \\
& & +\int_{t\vec{x}}\,\bar{c}\left(t,\vec{x}\right)\Biggr\{\partial_{t}c\left(t,\vec{x}\right)+v^{i}\left(t,\vec{x}\right)\partial_{i}c\left(t,\vec{x}\right)-\frac{\kappa}{2}\partial^{2}c\left(t,\vec{x}\right)\Biggr\} \nonumber \\
& & +\frac{1}{2}\int_{t\vec{x}}\,v^{i}\left(t,\vec{x}\right)\frac{\left(-\partial^{2}+m^{2}\right)^{\frac{d}{2}+\frac{\varepsilon}{2}}}{D_{0}}v^{i}\left(t,\vec{x}\right) \,.\label{eq:Kraichnan-model-action}
\end{eqnarray}
Several comments regarding this action are in order.
The field $\bar{\theta}$ is usually called the response field
since it is related to response functions. The quadratic term in
$\bar{\theta}$ originates from integrating out the stochastic forcing
$f$ of the scalar: the function $M\left(\frac{\vec{x}-\vec{y}}{L}\right)$ is essentially the
forcing covariance and $L$ represents the large length scale at which
energy is injected to stir the scalar. The third line in (\ref{eq:Kraichnan-model-action})
contains the Grassmannian odd fields $c$ and $\bar{c}$ which are
introduced to express the functional determinant involved in the procedure via an action and are referred
to as ghosts. This determinant is written in the Stratonovich
convention. The source term ${J}$ gathers the different sources,
respectively $J_{\theta},\bar{J}_\theta,\vec{j},\eta,\bar{\eta}$, for the fields $\theta,\bar{\theta},\vec{v},c,\bar{c}$
gathered in the field multiplet $\phi$.
\subsection{The functional renormalization group formalism} \label{subsec:FRG-formalism}
The renormalization group has been instrumental for understanding
the scaling properties of physical systems at criticality. Also away
from equilibrium the RG constitutes a crucial conceptual and calculational
framework to investigate several phenomena. As far as turbulence is
concerned, the renormalization group approach has been employed in
many studies \cite{Dominicis79,Fournier83,Yakhot1986,Canuto96,Adzhemyan03,Adzhemyan1999}.
However, perturbative techniques have often been
hampered by the lack of a small expansion parameter.
The functional renormalization group implements the Wilsonian renormalization
program at a functional level by adding to the microscopic action
a term $\Delta S_{k}$ which introduces a wavenumber scale $k$ and suppresses
the integration over low wavenumber modes $|\vec{p}|\lesssim k$
\cite{Wetterich:1992yh,Reuter:1993kw,Reuter:1994sg,Ellwanger:1993mw,Morris:1993qb}.
The fluctuations are hence progressively averaged thereby building up the effective
theory at scale $k$.
The modified effective action in the presence of this scale-dependent term
is called the effective average action
(EAA) and its $k$-dependence is governed by an exact equation, which
can be solved by implementing some approximation scheme. We refer
to Ref.~\onlinecite{Delamotte:2007pf} for a pedagogical introduction and to Ref.~\onlinecite{Dupuis:2020fhh}
for a recent review.
To achieve the separation of fluctuation modes, the term $\Delta S_{k}$ is chosen quadratic in the
fields:
\begin{eqnarray*}
\Delta S_{k}\left[\phi\right] & \equiv & \int_{t\vec{x}}\frac{1}{2}\phi R_{k}\left(-\partial^{2}\right)\phi\,,
\end{eqnarray*}
where $R_{k}\left(-\partial^{2}\right)$ is a suitable cutoff kernel.
It is required to be large for modes with wavenumber $|\vec{p}| \lesssim k$
and to vanish for modes with $|\vec{p}| \gtrsim k$ such that only these modes are integrated out:
\begin{equation}
R_k(p^2) \stackrel{|\vec{p}|\lesssim k}{\sim} k^2\quad\quad\quad R_k(p^2) \stackrel{|\vec{p}|\gtrsim k}{\longrightarrow}0
\label{eq:cutoff}
\end{equation}
The modified functional integral takes the form
\begin{eqnarray*}
Z_k = e^{W_{k}\left[J\right]} & \equiv & \int{\cal D}\phi\,e^{-S\left[\phi\right]-\Delta S_{k}\left[\phi\right]+\int_{t\vec{x}} {J}\phi}\,,
\end{eqnarray*}
where $W_{k}\left[J\right]$ is the (modified) generating functional
of the connected correlation functions.
The EAA is defined by the Legendre transform of $W_{k}\left[J\right]$
and is denoted by $\Gamma_{k}\left[\varphi\right]$:
\begin{eqnarray}
\Gamma_{k}\left[\varphi\right]+\Delta S_{k}\left[\varphi\right] & = & \int_{t\vec{x}} J\varphi-W_{k}\left[\varphi\right]\,,
\label{eq:Legendre}
\end{eqnarray}
where $\varphi$ is the average field. The scale dependence of the
EAA is governed by an exact RG flow equation that takes the following general
form:
\begin{eqnarray}
\partial_{s}\Gamma_{k}\left[\varphi\right] & = & \frac{1}{2}\mbox{Tr}\left[\left(\Gamma_{k}^{\left(2\right)}\left[\varphi\right]+R_{k}\right)^{-1}\partial_{s}R_{k} \right]\,,
\label{eq:FRGE-generic_EAA}
\end{eqnarray}
where $s\equiv\log (k/\Lambda)$ with $\Lambda$ a UV scale.
At this scale, one can show
that $\Gamma_\Lambda$ identifies with the bare action, since no fluctuation is yet incorporated.
When $k\to 0$, the regulator is removed and one obtains the actual properties of the model, when all fluctuations
have been integrated over. Eq.~(\ref{eq:FRGE-generic_EAA}) provides the exact interpolation between these two scales.
After this general introduction, let us discuss the application of
the FRG formalism to the Kraichnan model. In particular, we follow
the approach discussed in Ref.~\onlinecite{Canet:2014dta}, where the scale
$k$ is identified with the forcing scale of the scalar, i.e. $k\equiv L^{-1}$, which is now running,
and $M\to M_k$ in Eq.~(\ref{eq:Kraichnan-model-action}).
Indeed, since the forcing covariance is peaked at large
length scales and vanishes at small length scales, it satisfies the requirements (\ref{eq:cutoff}) and
can therefore be interpreted as a cutoff. Besides this ``effective forcing'' cutoff,
we also introduce cutoff kernels for the other fields. The generalized
cutoff action is defined by
\begin{eqnarray}
\Delta S_{k}\left[\phi\right] & \equiv & \int_{t\vec{x}}\,\Biggr\{\frac{\kappa}{2}\bar{\theta}R_{k}\left(-\partial^{2}\right)\theta-\frac{1}{2}\bar{\theta}M_{k}\left(-\partial^{2}\right)\bar{\theta}\nonumber \\
& & +\frac{1}{2D_{0}}v^{i}R_{k,vv}\left(-\partial^{2}\right)v^{i}+\frac{\kappa}{2}\bar{c}R_{k}\left(-\partial^{2}\right)c\Biggr\}\,.\label{eq:cutoff-action-Kraichnan}
\end{eqnarray}
With this addition the total action finally reads
\begin{eqnarray}
&& S = \int_{t\vec{x}}\,\bar{\theta}\left(t,\vec{x}\right)\Biggr\{\partial_{t}\theta\left(t,\vec{x}\right)+v^{i}\left(t,\vec{x}\right)\partial_{i}\theta\left(t,\vec{x}\right)-\frac{\kappa}{2}\partial^{2}\theta\left(t,\vec{x}\right)\Biggr\}\nonumber \\
& & + \int_{t\vec{x}}\,v^{i}\left(t,\vec{x}\right)\frac{\left(-\partial^{2}+m^{2}\right)^{\frac{d}{2}+\frac{\varepsilon}{2}}}{2 D_{0}}v^{i}\left(t,\vec{x}\right)\nonumber \\
& & +\int_{t\vec{x}}\,\bar{c}\left(t,\vec{x}\right)\Biggr\{\partial_{t}c\left(t,\vec{x}\right)+v^{i}\left(t,\vec{x}\right)\partial_{i}c\left(t,\vec{x}\right)-\frac{\kappa}{2}\partial^{2}c\left(t,\vec{x}\right)\Biggr\} \label{eq:S_final_total} \\
& & +\int_{t\vec{x}}\,\Biggr\{\frac{\kappa}{2}\bar{\theta}\left(t,\vec{x}\right)R_{k}\left(-\partial^{2}\right)\theta\left(t,\vec{x}\right)-\frac{1}{2}\bar{\theta}\left(t,\vec{x}\right)M_{k}\left(-\partial^{2}\right)\bar{\theta}\left(t,\vec{x}\right)\nonumber \\
& & +v^{i}\left(t,\vec{x}\right) \frac{R_{k,vv}\left(-\partial^{2}\right)}{2D_{0}} v^{i}\left(t,\vec{x}\right)+\frac{\kappa}{2}\bar{c}\left(t,\vec{x}\right)R_{k}\left(-\partial^{2}\right)c\left(t,\vec{x}\right)\Biggr\}\,, \nonumber
\end{eqnarray}
where the last two lines contain the cutoff action which incorporates the scalar forcing term.
\subsection{Vertex functions and notation}
The EEA $\Gamma_k$ introduced in section \ref{subsec:FRG-formalism} is the generating functional of the one-particle-irreducible (1-PI) correlation functions,
also called vertices, which are often introduced in field theoretical treatments.
The set of $n$-point connected correlation functions, denoted generically $W_k^{(n)}$, is equivalent
to the set of $n$-point 1-PI correlation functions, denoted $\Gamma_k^{(n)}$, in the sense that one can be constructed from
the other. Both types of correlation functions are used in the following.
Let us introduce some notation, in order to specify the fields entering these functions.
By taking $n$ functional derivatives of the EEA, one obtains the $n$-point corresponding vertex function
\begin{eqnarray*}
&& \frac{\delta^{n}\Gamma_{k}}{\delta\varphi_{1}\left(t_{1},\vec{x}_{1}\right)\cdots\delta\varphi_{2}\left(t_{n_{\varphi_{1}}+1},\vec{x}_{n_{\varphi_{1}}+1}\right)\cdots} = \\
&& \qquad\qquad\qquad
\Gamma_{k}^{\left(n_{\varphi_{1}},n_{\varphi_{2}},\cdots\right)}\left(t_{1},\vec{x}_{1},\cdots,t_{n_{\varphi_{1}}+1},\vec{x}_{n_{\varphi_{1}}+1},\cdots\right)\,,
\end{eqnarray*}
where $n=\sum_i n_{\varphi_{i}}$ and $\varphi_i$ denote any of the fields contained in the multiplet $\varphi$.
The Fourier transform of a vertex is denoted by
\begin{eqnarray*}
&& \Gamma_{k}^{\left(n_{\varphi_{1}},n_{\varphi_{2}},\cdots\right)}\left(\omega_{1},\vec{p}_{1}\cdots\right) = \\
&& \qquad\qquad\qquad
{\cal F}\left[\Gamma_{k}^{\left(n_{\varphi_{1}},n_{\varphi_{2}},\cdots\right)}\left(t_{1},\vec{x}_{1}\cdots\right)\right]
\left(\omega_{1},\vec{p}_{1}\cdots\right)\,
\end{eqnarray*}
with ${\cal F}$ the Fourier transform defined as Eq.~(\ref{eq:Fourier}) for each variables $(\omega_i,\vec{p}_i)$.
A fully analogous notation will be employed for
the connected correlation functions $W_k^{(n)}$.
In order to take into account translation invariance in space and time, we also introduce the
following generic $n$-point function:
\begin{eqnarray*}
&& \Gamma_{k}^{\left(n_{\varphi_{1}},n_{\varphi_{2}},\cdots\right)}\left(\omega_{1},\vec{p}_{1}\cdots\right)
= \left(2\pi\right)^{d+1}\delta\left(\sum_{i=1}^{n}\omega_{i}\right)\delta^d\left(\sum_{i=1}^{n}\vec{p}_{i}\right) \\
&& \qquad\qquad\qquad
\bar{\Gamma}_{k}^{\left(n_{\varphi_{1}},n_{\varphi_{2}},\cdots\right)}\left(\omega_{1},\vec{p}_{1}\cdots,\omega_{n-1},\vec{p}_{n-1}\right)\,,
\end{eqnarray*}
where $\bar{\Gamma}_{k}^{\left(n\right)}$ hence denotes a 1-PI vertex ``stripped''
of the wavenumber and frequency conserving delta functions. A fully analogous notation
is introduced for the $n$-point correlation function, $\bar{W}_{k}^{\left(n\right)}$.
In particular, a propagator stripped of the delta functions will
be written as $\bar{G}_{\varphi_{a}\varphi_{b}}\left(\omega,\vec{p}\right)$.
In order not to overburden the notation in certain equations, we shall
sometimes employ the DeWitt notation. In this case, the average field
is denoted simply by $\varphi_{a}$ and the sources by $J^{a}$. The
summation convention is intended to
mean $\varphi_{a}J^{a}\equiv\int_{t\vec{x}}\varphi_{\alpha}\left(t,\vec{x}\right)J^{\alpha}\left(t,\vec{x}\right)$.
\subsection{Symmetries and Ward identities \label{subsec:Symmetries-and-WI}}
In this section we discuss the symmetries of the action (\ref{eq:S_final_total})
and the associated Ward identities for the EAA. It turns out
that it is particularly
useful to consider not only the strict symmetries of the action but
also the field transformations which leave the action invariant
up to terms linear in the fields since in this latter case one can
actually write down Ward identities which are even more constraining.
We first explicitly expound
the case of the shift symmetry, and for the other symmetries, we only
provide the field transformation, the action variation, and the ensuing Ward
identity. We refer the interested reader to Refs.~\onlinecite{Canet:2014cta,Tarpin:2018yvs}
where a similar analysis is detailed for the case of the NS field theory.
With abuse of notation, we denote the average fields by
employing the same symbols used for the fluctuating fields,
i.e. $\varphi\equiv\left(\theta,\bar{\theta},v^{i},c,\bar{c}\right)$.
Before considering continuous symmetries, let us note that the action
(\ref{eq:S_final_total}) possesses the following $\mathbb{Z}_{2}$-symmetry:
$\theta\rightarrow-\theta$ and $\bar{\theta}\rightarrow-\bar{\theta}$.
An analogous symmetry is present in the ghost sector. Moreover, one can
show that any term in the EAA must depend on at least one
response field, the EAA cannot have terms depending
on $\theta$ (or $c$) only, see Ref.~\onlinecite{Canet11b} for a detailed discussion
on this point.\\
\emph{Scalar field shifts. }Let us consider the transformation $\theta\left(t,\vec{x}\right)\rightarrow\theta\left(t,\vec{x}\right)+\varepsilon\left(t\right)$,
which is a time gauged shift of the advected scalar field. The variation
of the action reads
\begin{eqnarray*}
\delta S & = & \int_{t\vec{x}}\,\bar{\theta}\left(t,\vec{x}\right)\Biggr\{\partial_{t}+\frac{\kappa}{2}R_{k}\left(0\right)\Biggr\} \varepsilon\left(t\right)\,.
\end{eqnarray*}
By writing that this field transformation must leave the functional integral $Z$ unchanged, one obtains
the following Ward identity
\begin{eqnarray*}
&& \int_{t\vec{x}}\,J_{\theta}\left(t,\vec{x}\right)\varepsilon\left(t\right) =
\Bigr\langle\int_{t\vec{x}}\,\bar{\theta}\left(t,\vec{x}\right)\Biggr\{\partial_{t}+\frac{\kappa}{2}R_{k}\left(0\right)\Biggr\} \varepsilon\left(t\right)\Bigr\rangle
\end{eqnarray*}
Expressing the sources as functional derivatives of $\Gamma_k$
using the definition (\ref{eq:Legendre}) of the Legendre transform then yields
\begin{eqnarray*}
&& \int_{t\vec{x}}\,\left(\frac{\delta\Gamma_{k}}{\delta\theta\left(t,\vec{x}\right)}+\bar{\theta}\left(t,\vec{x}\right)\frac{\kappa}{2}R_{k}\left(0\right)\right)\varepsilon\left(t\right) = \\
&&\quad\quad\quad\quad\quad\int_{t\vec{x}}\,{\bar{\theta}\left(t,\vec{x}\right)}\Biggr\{\partial_{t}+\frac{\kappa}{2}R_{k}\left(0\right)\Biggr\} \varepsilon\left(t\right)
\end{eqnarray*}
Since this relation holds for an arbitrary infinitesimal $\varepsilon(t)$,
this leads to the final Ward identity
\begin{equation}
\int_{\vec{x}}\,\frac{\delta\Gamma_{k}}{\delta\theta\left(t,\vec{x}\right)} = -\int_{\vec{x}}\,\partial_{t}\bar{\theta}\left(t,\vec{x}\right)\,,\label{eq:WI-t-local-shift-1}
\end{equation}
which is local in time. The Ward identity (\ref{eq:WI-t-local-shift-1})
has an immediate interpretation: the term $\int\bar{\theta}\partial_{t}\theta$
in the microscopic action $S$ is not renormalized and it appears
identically in the associated EAA. By taking further functional derivatives
of this identity with respect to the other fields, one obtains the following relations
in Fourier space
\begin{eqnarray}
&& \bar{\Gamma}^{(1,1,0,0,0)}_{k}(\omega_{1},\vec{p}_1=0) = i\omega_{1} \nonumber \\
&& \bar{\Gamma}_{k}^{\left(n_{\theta}\geq1,n_{\bar{\theta}},n_{v},n_{c},n_{\bar{c}}\right)}\left(\omega_{\theta},\vec{p}_{\theta}=0,\cdots\right) = 0\,.\label{eq:WI-t-local-shift-2}
\end{eqnarray}
Equation (\ref{eq:WI-t-local-shift-2}) entails that any vertex having at least one vanishing wavenumber carried by a $\theta$
field actually vanishes.\\
\emph{Response field shifts. }Let us consider the transformation $\bar{\theta}\left(t,\vec{x}\right)\rightarrow\bar{\theta}\left(t,\vec{x}\right)+\bar\varepsilon\left(t\right)$.
By exploiting the fact that the velocity field is divergenceless,
the variation of the action reads
\begin{eqnarray*}
\delta S & = & \int_{t\vec{x}}\,\bar\varepsilon(t)\Biggr\{\partial_{t}\theta\left(t,\vec{x}\right)+\frac{\kappa}{2} R_{k}\left(0\right)\theta\left(t,\vec{x}\right)
-M_{k}\left(0\right)\bar{\theta}\left(t,\vec{x}\right)\Biggr\}\,.
\end{eqnarray*}
One deduces that
\begin{eqnarray}
&& \Gamma_{k}^{\left(n_{\theta},n_{\bar{\theta}}\geq1,n_{v},n_{c},n_{\bar{c}}\right)}\left(\cdots,\omega_{\bar{\theta}},\vec{p}_{\bar{\theta}}=0,\cdots\right) = 0\,.\label{eq:WI-t-local-Qb-shift-1}
\end{eqnarray}
which yields that any vertex having at least one vanishing wavenumber carried by a $\bar\theta$
field also vanishes.\\
\emph{Time gauged Galilei transformation. }
We now consider an extended Galilean transformation, which corresponds to the infinitesimal field transformation
\begin{eqnarray}
v^{j}\left(t,\vec{x}\right) & \rightarrow & v^{j}\left(t,\vec{x}\right)-\varepsilon^{i}\left(t\right)\partial_{i}v^{j}\left(t,\vec{x}\right)+\dot{\varepsilon}^{j}\left(t\right)\nonumber \\
\bar{\theta}\left(t,\vec{x}\right) & \rightarrow & \bar{\theta}\left(t,\vec{x}\right)-\varepsilon^{i}\left(t\right)\partial_{i}\bar{\theta}\left(t,\vec{x}\right)\nonumber\\
\theta\left(t,\vec{x}\right) & \rightarrow & \theta\left(t,\vec{x}\right)-\varepsilon^{i}\left(t\right)\partial_{i}\theta\left(t,\vec{x}\right)\,,\label{eq:Galilei-transformation}
\end{eqnarray}
and similarly for the ghost fields.
This transformation can be interpreted as a time-gauged extension of space translations.
The standard Galilean transformation is recovered for $\vec \varepsilon(t)= \vec \varepsilon \,t$.
One finds
\begin{eqnarray*}
\delta S & = & \int_{t\vec{x}}\,v^{i}\left(t,\vec{x}\right)\left[\frac{1}{D_{0}}m^{d+\varepsilon}+\frac{1}{D_0} R_{k,vv}\left(0\right)\right]\p_t{\varepsilon}^{i}\left(t\right)\,.
\end{eqnarray*}
The associated Ward identity constrains the vertices for which one
of the velocity has a zero wavenumber. The explicit expression reads
\begin{eqnarray}
&& \Gamma_{\gamma_{1}\cdots\gamma_{n_{v}}\gamma_{n_{v}+1}}^{\left(n_{\theta},n_{\bar{\theta}},n_{v}+1,n_{c},n_{\bar{c}}\right)}\Bigr(\cdots,\underbrace{\omega_\ell,\vec{p}_\ell=0}_{\ell={\rm velocity \;index}},\cdots\Bigr) = \nonumber \\
&& \qquad
-\sum_{i=1}^{n}\frac{p_{i}^{\gamma_\ell}}{\omega}\Gamma_{\gamma_{1}\cdots\gamma_{n_{v}} }^{\left(n_{\theta},n_{\bar{\theta}},n_{v},n_{c},n_{\bar{c}}\right)}\Bigr(\cdots,\underbrace{\omega_{i}+\omega,\vec{p}_{i}}_{i^{{\rm th}}{\rm field}},\cdots\Bigr),\label{eq:WI-t-local-Galilei-1}
\end{eqnarray}
where $\gamma_{1}\cdots\gamma_{n_{v}+1}$ are the spatial
indices of the velocity fields, $\vec{p}_{i}$ is the wavevector of the $i$th
field in the vertex, and $n=n_{\theta}+n_{\bar{\theta}}+n_{v}+n_{\theta}+n_{\bar{\theta}}$.\\
\emph{Time-gauged rotations}. The time-gauged version of spatial rotations
is also an extended symmetry of the action $S_{{\rm Kr}}$. This symmetry
was first identified for the two-dimensional NS equation, see Ref.~\onlinecite{Tarpin:2018yvs}.
In the present work, we will not exploit this symmetry and leave its
study for future work.\\
\emph{BRS (Becchi-Rouet-Stora) symmetry}. We consider
the transformation $\theta\rightarrow\theta+\varepsilon c,\:\bar{c}\rightarrow\bar{c}-\varepsilon\bar{\theta}$
where $\epsilon$ is now an anti-commuting Grassmann parameter.
This transformation hence mixes the ``bosonic'' (scalar fields) and the ``fermionic'' (ghosts)
sectors of the action. One observes that this transformation leaves the action invariant $\delta S=0$.
It follows that
\begin{eqnarray*}
\int_{t\vec{x}}\,\left[\frac{\delta\Gamma_{k}}{\delta\theta\left(t,\vec{x}\right)}c\left(t,\vec{x}\right)+\bar{\theta}\left(t,\vec{x}\right)
\frac{\delta\Gamma_{k}}{\delta\bar{c}\left(t,\vec{x}\right)}\right] & = & 0\,.
\end{eqnarray*}
By taking further functional derivatives one deduces relations between
vertices involving ghosts. This is particularly useful to show that the kinetic
term of the velocity field is not renormalized\footnote{This non-renormalization of the velocity covariance
can also shown using different arguments, as in the approach of
Ref.~\onlinecite{PhysRevE.58.1823}.} and to show that the terms in the EAA which depend on the velocity
alone are at most quadratic in the fields.
Let us prove this last assertion, by explicitly
using the flow equation of the EAA. Let us assume that at the UV scale
$k=\Lambda$, the EAA is such that $\Gamma_{\Lambda}\left[0,0,v,0,0\right]$
is at most quadratic in the velocity fields, which is the case for the microscopic
action $S$ in (\ref{eq:S_final_total}).
Given this condition, we want to determine if the flow equation generates
terms which are of higher order in $v$. The associated flow equation
reads
\begin{eqnarray*}
&& \partial_{s}\Gamma_{k}\left[0,0,v,0,0\right] = \frac{1}{2}\mbox{Tr}\left[\left(\Gamma_{k,vv}^{\left(2\right)}+R_{k}\right)^{-1}\dot{R}_{k,vv}\right]\\
&& \qquad
+\mbox{Tr}\left[\left(\Gamma_{k,\theta\bar{\theta}}^{\left(2\right)}+R_{k}\right)^{-1}\dot{R}_{k}\right]-\mbox{Tr}\left[\left(\Gamma_{k,c\bar{c}}^{\left(2\right)}+R_{k}\right)^{-1}\dot{R}_{k}\right]\\
&& \qquad
= \frac{1}{2}\mbox{Tr}\left[\left(\Gamma_{k,vv}^{\left(2\right)}+R_{k}\right)^{-1}\dot{R}_{k,vv}\right]\,,
\end{eqnarray*}
where the last equality ensues from the BRS symmetry.
Since by assumption $\Gamma_{k,vv}^{\left(2\right)}$ has no dependence
on $v$, one finds that $\partial_{s}\Gamma_{k}\left[0,0,v,0,0\right]$ is
just a constant. Thus we can conclude that the quadratic-in-velocity nature of $\Gamma_{k}\left[0,0,v,0,0\right]$
is preserved by the flow, i.e. it holds true at any scale $k$.
It is instructive to sketch how the non-renormalization of the kinetic
term of the velocity arises within the FRG approach.
For this, one considers the flow equation for $\Gamma^{(2)}_{k,v v}$,
which can be obtained by taking two functional derivatives of the exact flow Eq. (\ref{eq:FRGE-generic_EAA})
with respect to velocity fields.
The vertices entering this flow equation are 3-point vertices with at least
one velocity field (explicitly $\Gamma_k^{(1,1,1,0,0)}$, $\Gamma_k^{(0,2,1,0,0)}$, $\Gamma_k^{(0,0,1,1,1)}$ and
$\Gamma_k^{(0,0,3,0,0)}$), and 4-point vertices with at least two velocity fields
(explicitly $\Gamma_k^{(1,1,2,0,0)}$, $\Gamma_k^{(0,2,2,0,0)}$, $\Gamma_k^{(0,0,2,1,1)}$ and
$\Gamma_k^{(0,0,4,0,0)}$).
The BRS Ward identities lead to the cancellation of all the diagrams with vertices
including the scalar fields and the ghosts. One is left with only pure velocity vertices.
However, since we have shown that $\Gamma_k$ is quadratic in $v$, it follows that
$\Gamma_k^{(0,0,3,0,0)}=\Gamma_k^{(0,0,4,0,0)}=0$, and this implies the non-renormalization of
the velocity kinetic term.\\
\emph{Further symmetries in the ghost sector}. The ghost sector shares
the same symmetries as the scalar sector, i.e., shift and $\mathbb{Z}_{2}$
symmetries. Moreover the ghost action is also invariant under $\bar{c}\rightarrow e^{\alpha}\bar{c}$
and $c\rightarrow e^{-\alpha}c$.\\
\emph{Synthesis}.
We have derived from the symmetries -- in an extended sense -- of the action,
a set of Ward identities which will lead to the exact closure
of the RG flow equation of any $n$-point correlation functions in the limit of large
wavenumbers. The key feature that will be exploited is that
the identities (\ref{eq:WI-t-local-shift-2}),
(\ref{eq:WI-t-local-Qb-shift-1}), and (\ref{eq:WI-t-local-Galilei-1})
imply that any 1-PI $n$-point vertex carrying one vanishing wavevector
is either zero or can be expressed in terms of lower-order $\left(n-1\right)$-point
vertices.
\section{Anomalous dimensions} \label{sec:preliminaries-and-non-renormalization-th}
In this section, we establish certain fundamental properties which are needed
in the derivation of the closed flow equation for $n$-point correlation functions.
\subsection{Propagator and bare scaling dimensions}
The exact flow equation (\ref{eq:FRGE-generic_EAA})
involves as a fundamental ingredient the regularized propagator $\left(\Gamma_{k}^{\left(2\right)}+R_{k}\right)^{-1}$.
To obtain it, one writes the matrix $\big(\Gamma_{k}^{\left(2\right)} +R_{k}\big)$ of second functional
derivatives of the EEA and regulator action. Since in Fourier space, it is diagonal in frequency and wavevector,
one can then simply invert the matrix.
Evaluated for vanishing values of the average field, $\varphi=0$, the regularized propagator
takes the following form in field space:
\begin{eqnarray}
\left(\bar{\Gamma}_{k}^{\left(2\right)}+R_{k}\right)^{-1} & = & \left(\begin{array}{ccccc}
\bar{G}_{\theta\theta} & \bar{G}_{\theta\bar{\theta}} & 0 & 0 & 0\\
\bar{G}_{\bar{\theta}\theta} & 0 & 0 & 0 & 0\\
0 & 0 & \bar{G}_{vv} & 0 & 0\\
0 & 0 & 0 & 0 & \bar{G}_{c\bar{c}}\\
0 & 0 & 0 & \bar{G}_{\bar{c}c} & 0
\end{array}\right)\,, \label{eq:propag}
\end{eqnarray}
with
\begin{eqnarray}
\bar{G}_{\theta\bar\theta}(\omega,\vec{p}) &=& \frac{1}{\Gamma_{k}^{\left(1,1,0,0,0\right)}(-\omega,\vec{p})+R_{k}(\vec{p})}\,, \nonumber\\
\bar{G}_{\theta\theta}(\omega,\vec{p}) &=& -\frac{\Gamma_{k}^{\left(0,2,0,0,0\right)}(\omega,\vec{p})-M_{k}(\vec{p})}{\left|\Gamma_{k}^{\left(1,1,0,0,0\right)}(\omega,\vec{p})+R_{k}(\vec{p})\right|^{2}},
\end{eqnarray}
and similarly for the ghost sector. $G_{vv}$ is the regularized velocity propagator, and
since the velocity kinetic term is not renormalized,
it writes
\begin{eqnarray}
\bar{G}_{v^{i}v^{j }}\left(\omega,\vec{p}\right) & = & D_{0}\frac{P_{ij }(\vec{p})}{\left(p^{2}+m^{2}\right)^{\frac{d}{2}+\frac{\varepsilon}{2}}+R_{k,vv}\left(\vec{p}\right)}\,.\label{eq:velocity-propagator}
\end{eqnarray}
Note that $m$ can be safely set to zero in this formalism thanks to the presence of the regulator.
Let us establish the bare scaling dimensions, which are defined
by the scaling properties of the fields as deduced from the microscopic
action (\ref{eq:S_final_total}). From the velocity kinetic term, it follows that the bare scaling
dimension of the velocity field should satisfy $TL^{d}L^{-d-\varepsilon}\left[v\right]^{2}=1$,
so that $\left[v\right]=L^{\varepsilon/2}/T^{1/2}$.\footnote{Note that this implies that the engineering dimension of the constant $D_{0}$
is not the same as its (trivial) scaling dimension.} For the quadratic term in $\theta$ and $\bar{\theta}$, one
obtains that
$TL^{d}T^{-1}\left[\theta\bar{\theta}\right]$=1, which yields $\left[\theta\bar{\theta}\right]=L^{-d}$.
The dimension of the molecular viscosity reads $\left[\kappa\right]=L^{2}/T$.
The scaling dimension of $\bar{\theta}$ depends on the particular
form of the forcing covariance chosen. In the present case
we consider
\begin{eqnarray}
\int_{t\vec{x}}\,\frac{1}{2}\bar{\theta} M_{k}\left(-\partial^{2}\right)\bar{\theta}
& \equiv & \int_{t\vec{x}}\,\frac{1}{2}\bar{\theta} \left(-\partial^{2}\right)e^{\frac{\partial^{2}}{k^{2}}}\bar{\theta} \,,\label{eq:def-forcing-M_k}
\end{eqnarray}
to be dimensionless. The microscopic scaling dimension of $\bar{\theta}$
is then defined by $TL^{d}L^{-2}\left[\bar{\theta}\right]^{2}=1$
implying $\left[\bar{\theta}\right]=T^{-1/2}L^{\left(2-d\right)/2}$.
We parametrize the RG flow by introducing the following quantities:
the field renormalizations $\left(Z_{k,\theta},Z_{k,\bar{\theta}},Z_{k,v},Z_{k,c},Z_{k,\bar{c}}\right)$
and the running molecular diffusivity $\kappa_k/2$.
The precise definition of the field renormalization constants
in terms of the EAA is given in Sec.~\ref{subsec:Non-renormalization-theorem-and-anomalous-dim}.
For convenience, we also introduce a running coupling $\lambda_k$ for the three-point
vertex present in the bare action $\bar\theta v_i \partial_i \theta$, although
it is not renormalized (see below). Its bare scaling dimension is determined by
$TL^{d}\left[\bar{\theta}\theta\right]\left[v\right]L^{-1}\left[\lambda_k\right]=1$,
which leads to $[\lambda_k] = L^{1-\varepsilon/2}T^{-1/2}$.
\subsection{Non-renormalization theorems and anomalous dimensions \label{subsec:Non-renormalization-theorem-and-anomalous-dim}}
We associate to each field in the EEA the corresponding field renormalization $Z_{k,\varphi}^{1/2}$.
Omitting the ghosts, the EAA then reads
\begin{eqnarray*}
&& \Gamma_{k} =
\int_{t\vec{x}}\,Z_{k,\bar{\theta}}^{1/2}\bar{\theta} \Biggr\{\partial_{t}+\lambda_k Z_{k,v}^{1/2}v^{i} \partial_{i}-\frac{\kappa_k}{2}\partial^{2}\Biggr\} Z_{k,\theta}^{1/2}\theta \\
& & \qquad\qquad\qquad
+\frac{1}{2}\int_{t\vec{x}}\,Z_{k,v}v^{i} \frac{\left(-\partial^{2}+m^{2}\right)^{\frac{d}{2}+\frac{\varepsilon}{2}}}{D_{0}}v^{i} +\cdots\,,
\end{eqnarray*}
The field renormalizations can thus be obtained through the relations
\begin{align}
Z_{k,\bar{\theta}}^{1/2}Z_{k,\theta}^{1/2}&=\frac{\p}{\p\left(i\omega\right)}\Gamma_k^{\left(1,1,0,0,0\right)}\left(\omega=0,\vec{p}=0\right)\nonumber\\
Z_{k,\bar{\theta}}^{1/2}Z_{k,\theta}^{1/2}\frac{\kappa_k}{2}&=\frac{\p^2}{\p p^2}\Gamma_k^{\left(1,1,0,0,0\right)}\left(\omega=0,\vec{p}=0\right)\nonumber\\
\frac{Z_{k,v}}{D_0}&=\frac{\p}{\p p^{d+\varepsilon}}\Gamma_k^{\left(0,0,2,0,0\right)}\left(\omega=0,\vec{p}=0\right)
\end{align}
The anomalous dimension of a field is then defined by $\eta_{\phi}\equiv-Z_{\phi}^{-1}\partial_{s}Z_{\phi}$,
and the anomalous dimension associated to the molecular diffusivity
by $\eta_{\kappa}\equiv-\kappa_k^{-1}\partial_{s}\kappa_k$.
The Ward identities derived in
Sec.~\ref{subsec:Symmetries-and-WI} provide crucial constraints on these
running constants. We now spell out the consequences
of these identities and of the non-renormalization theorems presented in Sec.~\ref{subsec:Symmetries-and-WI}.
First, we showed that the velocity kinetic term is not renormalized.
This immediately implies $Z_{k,v}=1$ and $\eta_{v}=0$. The Ward identity associated with the time-gauged
shifts Eq.~(\ref{eq:WI-t-local-shift-2})
implies $Z_{k,\bar{\theta}}^{1/2}Z_{k,\theta}^{1/2}=1$ and thus
$Z_{k,\theta}=Z_{k,\bar{\theta}}^{-1}$ and $\eta_{\theta}=-\eta_{\bar{\theta}}$.
It follows that the Ward identity associated with the time-gauged Galilean symmetry Eq.~(\ref{eq:WI-t-local-Galilei-1}) imposes that this
vertex is not renormalized, and thus $\lambda_k=1$.
Note that we redefine the cutoff action (\ref{eq:cutoff-action-Kraichnan})
such that the field renormalizations also appear in the cutoff
kernels. In the present case, this simply amounts to the subsitution
$M_{k}\left(-\partial^{2}\right)\rightarrow Z_{k,\bar{\theta}}M_{k}\left(-\partial^{2}\right)$.
To study the existence of a fixed point, one usually
switches to a dimensionless formulation, by rescaling all quantities
by suitable powers of the RG scale. In particular, dimensionless
wavevectors $\hat{\vec{p}}$ are defined by $\hat{\vec{p}}\equiv \vec{p}/k$ and dimensionless
frequencies by $\hat{\omega}\equiv\omega/\left(\kappa_k k^{2}\right)$.
At a fixed point, the RG flow of the dimensionless couplings vanishes.
In this respect, the non-renormalization of the coupling $\lambda_k$
has a striking consequence. Indeed, the associated dimensionless coupling
reads $\hat{\lambda_k}\equiv\lambda_k k^{1-\varepsilon/2}\left(\kappa_k k^{2}\right)^{-1/2}=\lambda_k k^{-\varepsilon/2}\kappa_k^{-1/2}$.
Hence, since $\p_s \lambda_k=0$, its flow equation is simply given by the linear contribution
\begin{eqnarray}
\partial_{s}\hat{\lambda_k} & = & \left(-\frac{\varepsilon}{2}+\frac{\eta_{\kappa}}{2}\right)\hat{\lambda_k}\,.\label{eq:beta-lambda-dimless-kraichnan}
\end{eqnarray}
This implies that at a non-Gaussian fixed point, satisfying $\hat{\lambda}\neq0$, then $\eta_{\kappa}=\varepsilon$.
Given the relations $\eta_{\theta}=-\eta_{\bar{\theta}}$, $\eta_{\kappa}=\varepsilon$,
and $\eta_{v}=0$, one deduces that there is a single
anomalous dimension left to be determined at a fixed point.
\subsection{Energy budget}
In this section, we outline
an argument which allows one to fix the value of the remaining anomalous dimension $\eta_{\bar{\theta}}$.
In principle, in order to determine $\eta_{\bar{\theta}}$ within
the FRG framework, one has to derive the flow equations for the two-point functions $\Gamma_k^{(2)}$,
integrate them, and read off the value of $\eta_{\bar{\theta}}$ at the fixed point.
In this work, we shall employ a shortcut first proposed in Ref.~\onlinecite{Canet:2014dta} for the case of NS equations.
Since we are interested in the stationary turbulent state,
this requires that the mean energy rate injected by the forcing is constant and matches
the mean energy rate dissipated by the scalar field.
The argument is thus based, in 3D, on imposing a
finite value to $\langle f\theta\rangle$ in the ideal limit $L\rightarrow\infty$,
i.e. $k\rightarrow0$, or equivalently that $\langle\kappa\partial_{i}\theta\partial_{i}\theta\rangle$
remains finite in the limit $\kappa\to 0$.
The averaged injected power for the scalar can be estimated as
\begin{eqnarray}
&& \langle f\left(t,\vec{x}\right)\theta\left(t,\vec{x}\right)\rangle \propto
\int_{\vec{y}}\Big\langle M_{k}\left(\vec{x}-\vec{y}\right)\bar{\theta}\left(t,\vec{y}\right)\theta\left(t,\vec{x}\right)\Big\rangle
\label{eq:estimate-energy-inj-1} \\
&& = \int_{\omega \vec{q}}M_{k}\left(\vec{q}\right)G_{\theta\bar{\theta}}\left(\omega,\vec{q}\right)\,\propto\,k^{d+2-\eta_{\bar{\theta}}}\int_{\hat{\omega}\hat{\vec{q}}}\hat{M}_{k}\left(\hat{\vec{q}}\right)\hat{G}_{\theta\bar{\theta}}\left(\hat{\omega},\hat{\vec{q}}\right)\,. \nonumber
\end{eqnarray}
For $\langle f\theta\rangle$ to have a finite
limit when $k\to 0$ imposes $\eta_{\bar{\theta}}=d+2$.
Let us remark that the estimate (\ref{eq:estimate-energy-inj-1})
is based on assuming standard scale invariance, in the sense that there is agreement
between the RG scaling (in $k$) and the actual scaling properties of 1-PI vertices (in $\omega,\vec{q}$).
However, although such correspondence is true in general in critical phenomena,
it can be violated in the case of turbulence, we refer to Ref.~\onlinecite{Canet:2014dta}
for a discussion of this point in the context of the NS equations. Such
violations may be at the source of deviations from the Kolmogorov scaling.
In this work, we limit ourselves to assuming standard
scale invariance to fix the anomalous dimension $\eta_{\bar{\theta}}$.
This in turn yields the standard Obukhov-Corrsin scaling \cite{Obukhov49,Corrsin51}.
Let us show that considering the energy rate dissipated by the
scalar leads to the same result.
In the Kraichnan model, one may define the analog of the Kolmogorov
dissipation scale by $\eta_{{\rm diss}}\sim\left(2\kappa/D_{0}^{\prime}\right)^{1/\varepsilon}$,
see eg.~Ref.~\onlinecite{Kupiainen:2006em}. The average dissipated power
can be estimated as
\begin{eqnarray*}
\kappa\langle\partial_{i}\theta\partial_{i}\theta\rangle & = & \kappa\int_{\omega}\int_{\vec{q}, \,|\vec{q}|<\eta_{{\rm diss}}^{-1}}q^{2}G_{\theta\theta}\left(\omega,\vec{q}\right)\,,
\end{eqnarray*}
where $\eta_{{\rm diss}}^{-1}$ is a UV cutoff. The integral is dominated
by the UV contribution, which we can estimate by assuming standard
scale invariance. It follows that
\begin{eqnarray}
\lim_{\kappa\rightarrow0}\kappa\langle\partial_{i}\theta\partial_{i}\theta\rangle & \sim & \lim_{\kappa\rightarrow0}\kappa\left(\frac{1}{\eta_{{\rm diss}}}\right)^{d+2-\eta_{\bar{\theta}}+\eta_{\kappa}}\,.\label{eq:dissipation-limit-kappa_to_zero-and-kappa-power}
\end{eqnarray}
By inserting $\eta_{{\rm diss}}\sim\left(2\kappa/D_{0}^{\prime}\right)^{1/\varepsilon}$
and $\eta_{\kappa}=\varepsilon$, one deduces that reaching
a finite limit requires $\eta_{\bar{\theta}}=d+2=5$.
\section{Correlation functions in the Kraichnan
model} \label{sec:closed-FRGE-Kraichnan}
The (connected) correlation functions $W_k^{(n)}$ can be expressed in terms
of the 1-PI correlation functions $\Gamma_k^{(n)}$ via a sum of tree diagrams.
In the FRG framework, the $\Gamma_k^{(n)}$ are obtained by deriving
the associated flow equations, solving them, and eventually taking the limit
$k\rightarrow0$. For instance, by taking two functional derivatives
of the FRG flow equation (\ref{eq:FRGE-generic_EAA}), one obtains
an exact flow equation for the 2-point functions $\Gamma_k^{(2)}$ (the inverse propagator).
However, the RHS
of this flow equation depends on the $3$-point and $4$-point vertices,
which are governed by their own flow equation depending on higher-order vertices.
It is then clear that an infinite hierarchy is generated
and that a suitable approximation scheme must be implemented.
In the present section, we consider a closure scheme based on the Ward
identities detailed in Sec.~\ref{subsec:Symmetries-and-WI}. The
crucial approximation underlying this scheme is the following one.
All the vertices, which enter into the flow equation of a given vertex $\Gamma_k^{(n)}(\omega_i,\vec{p}_i)$ ,
are expanded in the loop wavevector $\vec{q}\simeq 0$.
The rationale of this approximation is that the loop
wavevector is controlled by the cutoff derivative $\partial_{s}R_{k}\left(\vec{q}\right)$,
which vanishes for $|\vec{q}| \gtrsim k$.
Hence, if one considers large external wave-vectors $\vec{p}_i$, one may expand all the vertices
about zero $\vec{q}$. This type of approximation is inspired by the
Blaizot-M\'endez-Galain-Wschebor (BMW) scheme \cite{Blaizot:2005xy,Blaizot:2005wd,Blaizot:2006vr}.
It becomes exact in the limit where all $|\vec{p}_i|\to \infty$.
In the context of NS turbulence, this large wavenumber expansion was developed and analyzed
in Ref.~\onlinecite{Canet:2014dta,Canet:2016nmg,Tarpin:2017uzn,Tarpin:2018yvs}.
The crucial feature of this scheme for the NS equations, which renders it particularly powerful,
is that all the vertices with one wavector set to zero either vanish or are given exactly
in terms of lower-order vertices through the Ward identities.
This results in a closure of the flow equation of any $\Gamma_k^{(n)}$, which
is thus expressed in terms of vertices with $m \leq n$ only, without
any further approximation than the large wavenumber limit.
We show in the following that this closure can also be achieved for the Kraichnan model,
thanks to the Ward identities derived in Sec.~\ref{subsec:Symmetries-and-WI}.
In order to do so, we closely follow the derivation detailed in Ref.~\onlinecite{Tarpin:2017uzn}
to apply it to the Kraichnan model. Since the calculations which lead to
the final form of the closed flow equation are formally identical, we limit
ourselves to outlining the main steps of the derivation and refer to Ref.~\onlinecite{Tarpin:2017uzn}
for details. In this work, we derive the general closed flow equation
for a generic $n$-point correlation function, but we only focus on the solution of
the 2-point function of the scalar field.
\subsection{Closed flow equations for $n$-point correlation functions\label{subsec:Closed-flow-equations}}
For this derivation, it is more convenient to consider the exact FRG equation for
the generating functional of the connected correlation functions
$W_{k}\left[J\right]$, which reads \cite{Delamotte:2007pf}
\begin{eqnarray*}
\partial_{s}W_{k}\left[J\right] & = & -\frac{1}{2}\partial_{s}R_{k,ij}\left[\frac{\delta^{2}W_{k}\left[J\right]}{\delta J^{i}\delta J^{j}}+\frac{\delta W_{k}\left[J\right]}{\delta J^{i}}\frac{\delta W_{k}\left[J\right]}{\delta J^{j}}\right]\,,
\end{eqnarray*}
where we use deWitt notation, in particular the integrals are implicit. By taking functional derivatives
of this equation,
one obtains the exact flow equations for $n$-point correlation functions as
\begin{eqnarray}
&& \partial_{s}\frac{\delta^{n}W_{k}\left[J\right]}{\delta J^{1}\cdots\delta J^{n}} =
-\frac{1}{2}\partial_{s}R_{k,ij}\Biggr[\frac{\delta^{n+2}W_{k}\left[J\right]}{\delta J^{i}\delta J^{j}\delta J^{1}\cdots\delta J^{n}} \nonumber \\
&& \quad
+\sum_{\stackrel{\left(\left\{ a_{k}\right\} ,\left\{ a_{\ell}\right\} \right)}{k+\ell=n}}\frac{\delta^{k+1}W_{k}\left[J\right]}{\delta J^{i}\delta J^{a_{1}}\cdots\delta J^{a_{k}}}\frac{\delta^{\ell+1}W_{k}\left[J\right]}{\delta J^{j}\delta J^{a_{k+1}}\cdots\delta J^{a_{k+\ell}}}\Biggr]\,,\label{eq:FRGE-W(n)-1}
\end{eqnarray}
where $\left(\left\{ a_{k}\right\} ,\left\{ a_{\ell}\right\} \right)$
indicates all possible bipartions of the indices $1,\cdots,n$.
Let us first consider the second term in Eq.~(\ref{eq:FRGE-W(n)-1}).
It is straightforward to check that in wavevector space the cutoff derivative
term $\partial_{s}R_{k,ij}$ is a function of the total wavevector
carried by $J^{a_{1}},\cdots,J^{a_{k}}$, say $\vec{p}\equiv \vec{p}_{a_{1}}+\cdots+ \vec{p}_{a_{k}}$.
In the large wavenumber limit, the cutoff derivative $\partial_{s}R_{k,ij}\left(\vec{p}\right)$
is exponentially suppressed. Therefore, in this limit
the second term of Eq.~(\ref{eq:FRGE-W(n)-1}) is negligible and one can discard it.
Let us now turn to the first term in Eq.~(\ref{eq:FRGE-W(n)-1}).
We first rewrite it in terms of functional derivatives with respect
to the average field $\varphi$, rather than functional derivatives
with respect to the sources, since the former will generate 1-PI vertices to which
one can apply the intended approximation scheme ($\vec{q}$-expansion and Ward identities).
This leads to
\begin{eqnarray}
&& \partial_{s}R_{k,ij}\frac{\delta^{n+2}W_{k}\left[J\right]}{\delta J^{i}\delta J^{j}\delta J^{1}\cdots\delta J^{n}} = \nonumber \\
&& \qquad\quad
-\frac{1}{2}\partial_{s}R_{k,ab}\frac{\delta\varphi_{i}}{\delta J^{a}}\frac{\delta\varphi_{j}}{\delta J^{b}}\frac{\delta}{\delta\varphi_{i}}\frac{\delta}{\delta\varphi_{j}}\frac{\delta^{n}W_{k}\left[J\right]}{\delta J^{1}\cdots\delta J^{n}}
\nonumber \\
&& \qquad\quad
-\frac{1}{2}\partial_{s}R_{k,ab}\left(\frac{\delta^{3}W_{k}\left[J\right]}{\delta J^{a}\delta J^{b}\delta J^{j}}\right)\frac{\delta}{\delta\varphi_{j}}\frac{\delta^{n}W_{k}\left[J\right]}{\delta J^{1}\cdots\delta J^{n}}\,.\label{eq:FRGE-W(n)-step-2}
\end{eqnarray}
At vanishing sources and for corresponding vanishing average fields,
the term $\p_s R_k W_k^{(3)}$ in (\ref{eq:FRGE-W(n)-step-2}) is proportional to
the flow of $\frac{\delta W_{k}}{\delta J^{j}}=\varphi_{j}$, which
itself vanishes. Thus, one concludes that also the second term in
(\ref{eq:FRGE-W(n)-step-2}) vanishes. The flow equation
hence takes the form
\begin{eqnarray}
\partial_{s}\frac{\delta^{n}W_{k}\left[J\right]}{\delta J^{1}\cdots\delta J^{n}} & = &
-\frac{1}{2}\partial_{s}R_{k,ab}\frac{\delta\varphi_{i}}{\delta J^{a}}\frac{\delta\varphi_{j}}{\delta J^{b}}\frac{\delta^2}{\delta\varphi_{i}\delta\varphi_{j}}\frac{\delta^{n}W_{k}\left[J\right]}{\delta J^{1}\cdots\delta J^{n}}\,,
\qquad
\label{eq:FRGE-W(n)-3}
\end{eqnarray}
where the RHS must be though of as a functional of the average field
which is eventually evaluated at $\varphi=0$.
At this point, one can proceed and evaluate the functional derivatives
$\frac{\delta}{\delta\varphi_{i}}$ and $\frac{\delta}{\delta\varphi_{j}}$
in (\ref{eq:FRGE-W(n)-3}). In order to implement the closure strategy,
we observe that each of these functional derivative carries a wavevector $\vec{q}$
which is controlled by the cutoff derivative $\partial_{s}R_{k,ab}\left(\vec{q}\right)$.
In turn this implies that $\frac{\delta}{\delta\varphi_{i}}$ will generate
a 1-PI vertex which has its wavevector carried by $\varphi_i$ equal to $\vec{q}$.
The large wavenumber expansion then leads to setting $\vec{q}=0$ in
all such vertices. One can then apply the Ward identities studied in
Sec.~\ref{subsec:Symmetries-and-WI}. They
have an immediate stricking consequence: the only functional derivatives
$\frac{\delta}{\delta\varphi_{i}}$ which lead to a non-zero contribution
are those with respect to the velocity field only, since when $\vec{q}$ is associated
to any other fields, then the Ward identities yield that the corresponding
vertex vanishes.
Therefore, one can further simplify the flow equation, which takes
the form
\begin{eqnarray}
\partial_{s} W^{\left(n\right)}_{k}
& = &
-\frac{1}{2}\left(G_{v^{i}v^{\alpha}}\partial_{s}R_{k,v^{\alpha}v^{\beta}}G_{v^{\beta}v^{j}}\right)\frac{\delta^2}{\delta v^{i} \delta v^{j}}
W^{\left(n\right)}_{k} \,. \qquad
\label{eq:FRGE-W(n)-4}
\end{eqnarray}
As it is, Eq.~(\ref{eq:FRGE-W(n)-4}) is not closed yet. More
explicitly, this equation reads
\begin{eqnarray*}
\partial_{s} W^{\left(n\right)}_{k} & = & -\frac{1}{2}\int_{\omega \vec{q}}\left(\bar{G}_{v^{i}v^{\alpha}}\partial_{s}R_{k,v^{\alpha}v^{\beta}}\bar{G}_{v^{\beta}v^{j}}\right)\left(\omega,\vec{q}\right) \\
&& \times
\frac{\delta}{\delta v^{i}\left(-\omega,-\vec{q}\right)}\frac{\delta}{\delta v^{j}\left(\omega,\vec{q}\right)}W^{\left(n\right)}_{k} \,.
\end{eqnarray*}
However, acting on $W_{k}^{\left(n\right)}$ with $\frac{\delta}{\delta v}$
generates 1-PI vertices which are controlled by (\ref{eq:WI-t-local-Galilei-1})
after setting $\vec{q}=0$.
One can show that the action of the velocity functional derivative yields
\begin{eqnarray}
&& \left(\frac{\delta}{\delta v^{i}\left(\omega,\vec{q}\right)}W^{\left(n\right)}_{k} \left(\omega_{1},\vec{p}_{1},\cdots \right)\right)\Bigr|_{J=0, \vec{q}=0} \nonumber \\
&& = \sum_{k=1}^{n}-\frac{p_{k}^{i}}{\omega}W_{k}^{\left(n\right)}\left(\omega_{1},\vec{p}_{1},\cdots,\omega_{k}+\omega,\vec{p}_{k},\cdots\right)\nonumber \\
&& \equiv {\cal D}^{i}\left(\omega\right) W_{k}^{\left(n\right)}\left(\omega_{1},\vec{p}_{1},\cdots\right) \,.\label{eq:one-q-zero-v-derivatives}
\end{eqnarray}
and a further functional derivative can be expressed as
\begin{eqnarray}
&& \frac{\delta}{\delta v^{i}\left(-\omega,-\vec{q}\right)}\frac{\delta}{\delta v^{j}\left(\omega,\vec{q}\right)}W_{k}^{\left(n\right)}\left(\omega_{1},\vec{p}_{1},\cdots\right)\Bigr|_{\vec{q}=0} \nonumber \\
&& \qquad\qquad\quad =
{\cal D}^{i}\left(-\omega\right){\cal D}^{j}\left(\omega\right)W_{k}^{\left(n\right)}\left(\omega_{1},\vec{p}_{1},\cdots\right)\,.\label{eq:two-q-zero-v-derivatives}
\end{eqnarray}
Note that $\vec{q}$ is set to zero only at the level of vertices (not in the $W_k^{(n)}$)
and also the Ward identities are applied to these vertices.
The proof of the resulting identities (\ref{eq:one-q-zero-v-derivatives}) and (\ref{eq:two-q-zero-v-derivatives})
at the level of the correlation functions $W_k^{(n)}$
is quite lengthy, but it poses no difficulty
as it is strictly similar to Ref.~\onlinecite{Tarpin:2017uzn}, to which
we refer the reader for the detailed demonstration.
Using the identities (\ref{eq:one-q-zero-v-derivatives}) and (\ref{eq:two-q-zero-v-derivatives}) to express
the RHS of Eq.~(\ref{eq:two-q-zero-v-derivatives})
then leads to the final expression for the flow equation of $W_k^{n}$ at large wavenumbers
\begin{eqnarray}
&& \partial_{s}W_{k}^{n}\left(\cdots,\omega_{k},\vec{p}_{k},\cdots\right)
=
\frac{1}{2}\int_{\omega \vec{q}}\left(\bar{G}_{v^{i}v^{\alpha}}\partial_{s}R_{k,v^{\alpha}v^{\beta}}\bar{G}_{v^{\beta}v^{j}}\right)\left(\omega,\vec{q}\right)
\qquad
\nonumber \\
&& \quad
\sum_{k,\ell=1}^{n}\frac{p_{k}^{i}p_{\ell}^{j}}{\omega^{2}}W_{k}^{\left(n\right)}\left(\cdots,\omega_{k}+\omega,\vec{p}_{k},\cdots,\omega_{\ell}-\omega,\vec{p}_{\ell},\cdots\right)\,,\label{eq:FRGE-W(n)-5}
\end{eqnarray}
which does not involved any higer-order correlation functions $W_{k}^{\left(m>n\right)}$,
showing that it is closed.
\subsection{Two-point correlation function in the large wavenumber limit \label{subsec:2pt-function-in-large-P}}
We now specify to the flow equation for $\bar{G}_{\theta\theta}\left(\omega,\vec{p}\right)=\bar{W}_{k}^{\left(2,0,0,0,0\right)}\left(\omega,\vec{p}\right)$.
Starting from equation (\ref{eq:FRGE-W(n)-5}), a straightforward
calculation leads to
\begin{eqnarray*}
&& \partial_{s}\bar{G}_{\theta\theta}\left(\omega_{1},\vec{p}\right) = \frac{1}{2}\int_{\omega \vec{q}}\left(\bar{G}_{v^{i}v^{\alpha}}\partial_{s}R_{k,v^{\alpha}v^{\beta}}\bar{G}_{v^{\beta}v^{j}}\right)\left(\omega, \vec{q}\right) \times \\
&&
\frac{p^{i}p^{j}}{\omega^{2}}\Biggr\{2\bar{G}_{\theta\theta}\left(\omega_{1}, \vec{p}\right)-\bar{G}_{\theta\theta}\left(\omega_{1}+\omega,\vec{p}\right)-\bar{G}_{\theta\theta}\left(\omega_{1}-\omega,\vec{p}\right)\Biggr\}\,.
\end{eqnarray*}
By inverse Fourier transforming to real time, one obtains
\begin{eqnarray}
&& \partial_{s}\bar{G}_{\theta\theta}\left(t,\vec{p}\right) = \frac{1}{2}\int_{\omega \vec{q}}\left(\bar{G}_{v^{i}v^{\alpha}}\partial_{s}R_{k,v^{\alpha}v^{\beta}}\bar{G}_{v^{\beta}v^{j}}\right)\left(\omega,\vec{q}\right)p^{i}p^{j} \nonumber \\
&& \qquad\qquad\qquad\qquad\quad
\times \Biggr\{\frac{2-2\cos\left(t\omega\right)}{\omega^{2}}\Biggr\}\bar{G}_{\theta\theta}\left(t,\vec{p}\right)\,.\label{eq:FRGE-Gqq(tp)-1}
\end{eqnarray}
Equation (\ref{eq:FRGE-Gqq(tp)-1}) is written in terms of the scalar
field and the velocity propagators. At this point, one of the peculiarities
of the Kraichnan model intervenes, namely that the velocity propagator
is known exactly, Eq.~(\ref{eq:velocity-propagator}), and
it does not depend on the frequency. This allows one to perform
explicitly the integration over the internal frequency, which gives
\begin{eqnarray}
&&
\partial_{s}\bar{G}_{\theta\theta}\left(t,\vec{p}\right) = \frac{d-1}{2d}\int_{\vec{q}}\frac{D_{0}}{\left(\left(q^{2}+m^{2}\right)^{\frac{d}{2}+\frac{\varepsilon}{2}}+R_{k,vv}\right)^{2}}\partial_{s}R_{k,vv}\left(\vec{q}\right) \nonumber \\
&& \qquad\qquad\qquad\qquad\qquad\qquad\qquad\quad
\times p^{2}\left|t\right|\bar{G}_{\theta\theta}\left(t,\vec{p}\right)\,.
\label{eq:kraichnan-dimful-FRGE}
\end{eqnarray}
We are interested in the fixed-point solution of this equation.
It is thus convenient to switch
to dimensionless variables. We thus introduce $\hat{\vec{p}}=\vec{p}/k$,
$\hat{t}\equiv t\kappa_k k^{2}$,
\begin{eqnarray*}
\bar{G}_{\theta\theta}\left(t,\vec{p}\right) & = & k^{\eta_{\kappa}-\eta_{\bar{\theta}}}\hat{\bar{G}}_{\theta\theta}\left(\hat{t},\hat{\vec{p}}\right)\,,
\end{eqnarray*}
and the dimensionless regulator following $R_{k,vv}(\vec{q}) = k^{d+\epsilon}\hat{R}_{k,vv}(\hat{\vec{q}})$, and thus
$$ \p_s R_{k,vv}(\vec{q}) = k^{d+\epsilon}\Big((d+\epsilon)\hat{R}_{k,vv}(\hat{\vec{q}})-\hat{q}\p_{\hat{q}}\hat{R}_{k,vv}(\hat{\vec{q}})\Big).$$
We define
\begin{eqnarray*}
\hat {\alpha}_k & \equiv & \frac{d-1}{2d}
\int_{\hat{\vec{q}}}\frac{\partial_{s}\hat{R}_{k,vv}\left(\hat{\vec{q}}\right)}{\left(\left(\hat{q}^{2}+\hat{m}^{2}\right)^{\frac{d}{2}+\frac{\varepsilon}{2}}+\hat{R}_{k,vv}\right)^{2}}\,,
\end{eqnarray*}
which, at the fixed point, becomes a pure number that can be calculated given any explicit cutoff kernel $R_{k,vv}$, and eg. $m=0$.
The dimensionless flow equation then reads
\begin{eqnarray}
&& \left(\partial_{s}+\left(\eta_{\kappa}-\eta_{\bar{\theta}}\right)+\left(2-\eta_{\kappa}\right)\hat{t}\partial_{\hat{t}}-\hat{\vec{p}}\partial_{\hat{\vec{p}}}\right)\hat{\bar{G}}_{\theta\theta}\left(\hat{t},\hat{\vec{p}}\right)
= \nonumber\\
&&\qquad\qquad\qquad\qquad\qquad
= D_{0} \hat {\alpha}_k \hat{\lambda_k}^{2}\hat{p}^{2}\left|\hat{t}\right|\hat{\bar{G}}_{\theta\theta}\left(\hat{t},\hat{\vec{p}}\right)\,\nonumber\\
\label{eq:dimless-FRGE-Gqq-1}
\end{eqnarray}
where $\hat{\lambda}_k$ is the dimensionless coupling associated to the non-renormalized bare one
$\lambda_k=1$.
By inserting $\eta_{\kappa}=\varepsilon$ and $\eta_{\bar{\theta}}=5$, and considering the fixed-point
(i.e. $\p_s=0$ in Eq.~(\ref{eq:dimless-FRGE-Gqq-1})), one obtains
\begin{eqnarray}
&& \left(-\left(5-\varepsilon\right)+\left(2-\varepsilon\right)\hat{t}\partial_{\hat{t}}-\hat{\vec{p}}\partial_{\hat{\vec{p}}}\right)\hat{\bar{G}}_{\theta\theta}\left(\hat{t},\hat{\vec{p}}\right)\nonumber \\
&& \qquad\qquad\qquad\qquad\qquad
= D_{0}\hat{\alpha}\hat{\lambda}^{2}\hat{p}^{2}\left|\hat{t}\right|\hat{\bar{G}}_{\theta\theta}\left(\hat{t},\hat{\vec{p}}\right)\,.\label{eq:FP-FRGE-Gqq-1}
\end{eqnarray}
where $\hat{\lambda}$ and $\hat{\alpha}$ are the fixed-point values of $\hat{\lambda}_k$ and $\hat\alpha_k$.
At equal time $\hat{t}=0$, Eq.~(\ref{eq:FP-FRGE-Gqq-1})
takes a very simple form since the RHS and the logarithmic
$\hat{t}$-derivative vanish, and the solution is a power law
\begin{eqnarray*}
\hat{\bar{G}}_{\theta\theta}\left(0,\hat{\vec{p}}\right) & \sim & \hat{p}^{-\left(5-\varepsilon\right)}\,.
\end{eqnarray*}
This power law corresponds to the scaling expected in the inertial
range ${\bar{G}}_{\theta\theta}\left(0,{x}\right)-{\bar{G}}_{\theta\theta}\left(0,0\right)\sim x^{2-\varepsilon}$.
The general solution of Eq.~(\ref{eq:FP-FRGE-Gqq-1}) is given
by
\begin{eqnarray}
\hat{\bar{G}}_{\theta\theta}\left(\hat{t},\hat{\vec{p}}\right) & = & \hat{p}^{-\left(5-\varepsilon\right)}e^{-D_{0}\frac{\hat{\alpha}}{\varepsilon}\hat{\lambda}^{2}\hat{p}^{2}|\hat{t}|}\hat{f}\left(\hat{t}\hat{p}^{2-\varepsilon}\right)\,,\label{eq:dimless-FP-sol-Gqq-1}
\end{eqnarray}
where $f\left(\hat{t}\hat{p}^{2-\varepsilon}\right)$ is a universal scaling
function, which can be determined from FRG methods. To do so,
it is not sufficient to solve the fixed-point equation but one has
to evolve the flow equation from an initial condition corresponding to the bare action (\ref{eq:S_final_total}) down
to $k=0$ (in fact one can stop at the scale $k$ where the fixed point is reached).
The exponential present in the solution (\ref{eq:dimless-FP-sol-Gqq-1})
does not have a smooth limit as $\varepsilon\rightarrow0$. This is
not surprising since the velocity two-point correlation function
(\ref{eq:vel-2pt-correlation-function}) itself is not well defined
as $\varepsilon\rightarrow0$. In terms of the parameter $D_{0}^{\prime}$
the solution has a smooth behavior in $\varepsilon$ (note that
$\hat{\alpha}$ is also a function of $\varepsilon$) which reads,
switching back to the dimensionful physical quantities,
\begin{eqnarray}
\bar{G}_{\theta\theta}\left(t,\vec{p}\right)
& = & p^{-\left(5-\varepsilon\right)}e^{-D_{0}^{\prime}\hat{\alpha} k^{-\varepsilon}p^{2}|t|} {f}\left(tp^{2-\varepsilon}\right)\,.\label{eq:dimless-FP-sol-Gqq-3}
\end{eqnarray}
Note that the exponential in (\ref{eq:dimless-FP-sol-Gqq-3}) depends
explicitly on $k$. This is a sign of the breaking of standard
scale invariance. When considering the fixed point for $k\rightarrow0$,
one typically identifies the RG scale $k$ with the inverse of the energy injection
scale, $k=L^{-1}$ since the flow essentially stops when crossing this scale.
However, for the Kraichnan model, it makes sense
to identify $k$ with the mass scale present in the velocity propagators
since the coefficient $\hat{\alpha}$ is fully determined by it.
This yields
\begin{eqnarray}
\bar{G}_{\theta\theta}\left(t,\vec{p}\right) & = & p^{-\left(d+2-\varepsilon\right)}e^{-D_{0}^{\prime}\hat{\alpha} m^{-\varepsilon}p^{2}|t|}f\left(tp^{2-\varepsilon}\right)\,.\label{eq:dimless-FP-sol-Gqq-4}
\end{eqnarray}
which is one of the main results of this work. Let us emphasize that this expression
is valid at large wavenumber $p$ but for arbitrary time delays $t$ in the stationary state. The temporal decorrelation of the scalar
in the Kraichan model is always exponential at any time scale. This is due to the complete decorrelation in time of the stochastic velocity field in this model. One can indeed remark that if this feature is relaxed by introducing some time correlation in the covariance Eq.~(\ref{eq:vel-2pt-correlation-function}),
the expression (\ref{eq:dimless-FP-sol-Gqq-4}) is no longer valid at small $t$, but is replaced by a Gaussian decay as occurs for the real scalar, see Sec.~\ref{sec:closed-FRGE-for-advected-scalars}.
\subsection{Two-point correlation function from the Dyson equation \label{subsec:2pt-correlation-function-from-Dyson-eq}}
The simplifying assumptions on the velocity field in the Kraichnan model
lead to the important consequence that the equal-time
two-point correlation function $\langle\theta\left(t,x\right)\theta\left(t,y\right)\rangle$
can be computed exactly, see Ref.~\onlinecite{RevModPhys.73.913} and references
therein. Such an exact solution is obtained by deriving a closed form of
the associated Hopf equation, which allows one to write an exact
differential equation for the two-point correlation function,
see, eg. Refs.~\onlinecite{RevModPhys.73.913,Falkovich1996,Eyink2013}.
The connection between this exact differential equation for the two-point
correlation function and the field theoretical calculation was worked out in
Ref.~\onlinecite{PhysRevE.58.7381}, which showed how to retrieve
the differential equation from the Dyson equation
in a steady state.
In this section,
we follow the approach used in Ref.~\onlinecite{PhysRevE.58.7381} and extend it
to compute the two-point Eulerian correlation function of the scalar \emph{at unequal times},
thereby allowing for a comparison
with the result obtained by FRG methods, Eq.~(\ref{eq:dimless-FP-sol-Gqq-4}).
The aim is to compute the self-energies, which can be defined from the propagators as
\begin{eqnarray*}
\bar{G}_{\theta \bar{\theta}}\left(\varpi,\vec{p}\right) & \equiv & \frac{1}{-i\varpi+\frac{\kappa}{2}p^{2}-\Sigma_{\theta\bar{\theta}}\left(\varpi,\vec{p}\right)}\\
\bar{G}_{\theta\theta}\left(\varpi,\vec{p}\right) & \equiv & \frac{\left(M_{k}\left(\vec{p}\right)+\Sigma_{\bar{\theta}\bar{\theta}}\left(\varpi,\vec{p}\right)\right)}{\left|-i\varpi+\frac{\kappa}{2}p^{2}-\Sigma_{\theta\bar{\theta}}\left(\varpi,\vec{p}\right)\right|^{2}}\,.
\end{eqnarray*}
We consider the one-loop diagrams
which enter in the determination of $\Sigma_{\bar{\theta}\theta}$
and $\Sigma_{\theta\theta}$, and express them in terms of the exact
propagators. Using the expression (\ref{eq:vel-2pt-correlation-function}) for the velocity propagator,
one finds \cite{PhysRevE.58.7381}
\begin{eqnarray*}
\Sigma_{\theta\bar{\theta}}\left(\varpi,\vec{p}\right) & = &
-\int_{\omega \vec{q}}\bar{G}_{\theta\bar{\theta}}\left(\omega+\varpi,\vec{q}+\vec{p}\right)
p^{\alpha}\bar{G}_{v^{\alpha}v^{\beta}}\left(\vec{q}\right)(p^{\beta}+q^{\beta})\\
& = & -p^{2}\left[D_{0}\frac{\left(d-1\right)}{2d}\frac{\Gamma\left(\varepsilon/2\right)m^{-\varepsilon}}{\left(4\pi\right)^{d/2}\Gamma\left(d/2+\varepsilon/2\right)}\right]\,
\end{eqnarray*}
This a striking feature of the Kraichnan model:
The expression of $\Sigma_{\bar{\theta}\theta}\left(\varpi,\vec{p}\right)$
is \emph{frequency independent} because the velocity propagator is
itself frequency independent, allowing one to shift the frequency
in the integral. The frequency integration simply yields
$\int_{\omega}\bar{G}_{\theta\bar{\theta}}(\omega,\vec{p})=\frac{1}{2}$,
and the wavevector integral can then also be carried out, leading
to an extremely simple $\vec{p}$-dependence as well.
The 1-PI vertex $\Gamma^{\left(1,1,0,0,0\right)}$ then takes the following
simple form
\begin{eqnarray*}
&&
\Gamma^{\left(1,1,0,0,0\right)}\left(\varpi,\vec{p}\right) \\
&&= i\varpi + p^2\left[\frac{\kappa}{2} + D_{0}\frac{\left(d-1\right)}{2d}\frac{\Gamma\left(\varepsilon/2\right)m^{-\varepsilon}}{\left(4\pi\right)^{d/2}\Gamma\left(d/2+\varepsilon/2\right)}\right]\\
&& \equiv i\varpi + p^2\frac{\kappa_{{\rm ren}}}{2}\, ,
\end{eqnarray*}
where we introduced the renormalized molecular viscosity $\kappa_{{\rm ren}}$.
The self-energy of the response field is given by
\begin{eqnarray*}
&&
\Sigma_{\bar{\theta}\bar{\theta}}\left(\varpi,\vec{p}\right) \\
&&
= \int_{\omega \vec{q}}\bar{G}_{v^{\alpha}v^{\beta}}\left(\vec{q}\right)
\left(p^{\alpha}+q^{\alpha}\right)(p^{\beta}+q^{\beta})G_{\theta\theta}\left(\omega+\varpi,\vec{q}+\vec{p}\right)\\
&&
= \int_{\omega \vec{q}}\frac{D_{0}}{\left(q^{2}+m^{2}\right)^{\frac{d+\varepsilon}{2}}}\left(p^{2}-\frac{\left(\vec{p}\cdot \vec{q}\right)^{2}}{q^{2}}\right) \\
&& \qquad\qquad\qquad\qquad
\times \frac{M_{k}\left(\vec{q}+\vec{p}\right)+\Sigma_{\bar{\theta}\bar{\theta}}\left(\varpi+\omega,\vec{q}+\vec{p}\right)}{\left(\omega+\varpi\right)^{2}
+\left(\frac{\kappa_{{\rm ren}}}{2}\left(\vec{q}+\vec{p}\right)^{2}\right)^{2}}\,.
\end{eqnarray*}
A shift in the integration frequency shows
that this self-energy is also independent of $\varpi$. It is clear that
the frequency independence of the self-energies is a peculiarity of
the Kraichnan model.
At this point, one can express the two-point correlation function as
\begin{eqnarray*}
\bar{G}_{\theta\theta}\left(\varpi,\vec{p}\right) &=&
\frac{1}{\varpi^{2}+\left(\frac{\kappa_{{\rm ren}}}{2}p^{2}\right)^{2}}
\Biggr(M_{k}\left(\vec{p}\right)+\int_{\vec{q}}\frac{D_{0}}{\left(q^{2}+m^{2}\right)^{\frac{d+\varepsilon}{2}}} \\
& \times &
\left(p^{2}-\frac{\left(\vec{p}\cdot \vec{q}\right)^{2}}{q^{2}}\right)\frac{M_{k}\left(\vec{q}+\vec{p}\right)
+\Sigma_{\bar{\theta}\bar{\theta}}\left(\vec{q}+\vec{p}\right)}{\kappa_{{\rm ren}}\left(\vec{q}+\vec{p}\right)^{2}} \Biggr)\\
& \equiv & \frac{F\left(\vec{p}\right)}{\varpi^{2}+\left(\frac{\kappa_{{\rm ren}}}{2}p^{2}\right)^{2}}\,,
\end{eqnarray*}
where $F\left(\vec{p}\right)$ is simply a shorthand for the expression in the
parentheses. It is straightforward to transform back to real time and
obtain
\begin{eqnarray}
\bar{G}_{\theta\theta}\left(t,\vec{p}\right) & = & e^{-\frac{\kappa_{{\rm ren}}}{2}p^{2}\left|t\right|}\frac{F\left(\vec{p}\right)}{\kappa_{{\rm ren}}p^{2}}\,,\label{eq:Gqq-t-p-via-Dyson-fin}
\end{eqnarray}
which provides another expression for the two-point correlation function of the scalar, that we compare
with the FRG one in the following section.
\subsection{Comparison of the two approaches and remarks}
Let us comment on the two expressions we have derived for $\bar{G}_{\theta\theta}\left(t,\vec{p}\right)$.
First of all, the temporal dependence
of (\ref{eq:Gqq-t-p-via-Dyson-fin}) is of the same form as the one obtained
by FRG methods, namely an exponential with an exponent proportional
to $p^{2}\left|t\right|$. It is also interesting to compare the coefficients
appearing in front of this factor. In (\ref{eq:Gqq-t-p-via-Dyson-fin})
the coefficient $\frac{\kappa_{{\rm ren}}}{2}$ has two contributions:
one is just the microscopic molecular diffusivity and the other reads
\begin{eqnarray}
\frac{\kappa_{{\rm ren}}}{2}-\frac{\kappa}{2} & = & D_{0}\frac{\left(d-1\right)}{2d}\frac{\Gamma\left(\varepsilon/2\right)m^{-\varepsilon}}{\left(4\pi\right)^{d/2}\Gamma\left(d/2+\varepsilon/2\right)}\,.\label{eq:kappa_ren-minus-kappa}
\end{eqnarray}
It is clear that the FRG result should be compared against this
expression since $\kappa$ is negligible in the inertial range.
The expression (\ref{eq:kappa_ren-minus-kappa}) exhibits three main features:
it is proportional to $D_{0}$, it diverges as $1/\varepsilon$ as $\varepsilon\rightarrow0$,
and it is proportional to $m^{-\varepsilon}$ (which signals an IR
singularity as $m\rightarrow0$). All these three aspects are shared
by the coefficient appearing in the FRG formula (\ref{eq:dimless-FP-sol-Gqq-4}).
Now, even neglecting the molecular viscosity, the two coefficients share many
similarities but are not equal, which is to be expected, since the setting and assumptions
are different. For instance the coefficient (\ref{eq:kappa_ren-minus-kappa})
is IR divergent if no IR mass/regulator is added, which renders
difficult a meaningful comparison beyond the qualitative features
outlined above. However, note that if one chooses a mass cutoff
kernel defined by $S_{{\rm vel}}+\Delta S_{k}=\int\frac{1}{2}v\left(-\partial^{2}+k^{2}\right)^{\frac{d+\varepsilon}{2}}v$
and eventually sets $k=m$, one finds that the exponent in (\ref{eq:dimless-FP-sol-Gqq-4})
has the same expression as in (\ref{eq:kappa_ren-minus-kappa}).
In principle, one can reconstruct the correlations of any microscopic theory by
solving the FRG equation exactly, strictly taking the limit $k\rightarrow 0$, and keeping $k\neq m$.\\
Let us now make two additional remarks.
First, the same framework of the Dyson equations can be used
to determine the behavior of the spectrum in the mildly non-linear regime,
that is in the dissipative range, but close to the inertial range where the convection still plays a role.
We find that the energy spectrum in this range exhibits an intermediate power law regime following $p^{-(d+2+\varepsilon)}$,
which is shown in Appendix~\ref{app:dissipative}.
Finally, let us comment on the discretization scheme
of the SDE
(\ref{eq:Kraichnan-model_SDE}).
This equation
is written in the Stratonovich convention.
In the Ito convention, it reads (see, e.g., Ref.~\cite{Kupiainen:2006em})
\begin{eqnarray}
\partial_{t}\theta +v^{i} \partial_{i}\theta
-\left(\frac{\kappa}{2}\delta^{ij}+\frac{1}{2}D^{ij}\left(0\right)\right)\partial_{i}\partial_{j}\theta
& = & f \,,\label{eq:Kraichnan-model-Ito-SDE-1}
\end{eqnarray}
where $D^{ij}\left(\vec{x}-\vec{y}\right)$ denotes the spatial part of the velocity
covariance (\ref{eq:vel-2pt-correlation-function}), explicitly
$\langle v^{i}\left(t,\vec{x}\right)v^{j}\left(t^{\prime},\vec{y}\right)\rangle=\delta\left(t-t^{\prime}\right)D^{ij}\left(\vec{x}-\vec{y}\right)$.
The new term in the SDE can be expressed as
\begin{eqnarray*}
\frac{1}{2}D^{ij}\left(0\right) & = & \frac{1}{2}D_{0}\int_{\vec{q}}\frac{P_{ij}(\vec{q})}{\left(q^{2}+m^{2}\right)^{\frac{d}{2}+\frac{\varepsilon}{2}}} \nonumber \\
&=&
D_{0}\frac{\left(d-1\right)}{2d}\frac{\Gamma\left(\varepsilon/2\right)m^{-\varepsilon}}{\left(4\pi\right)^{d/2}\Gamma\left(d/2+\varepsilon/2\right)}\delta_{ij}\, .
\end{eqnarray*}
Hence, one observes that the coefficient in front of $\delta_{ij}$ is exactly
the quantity appearing in (\ref{eq:kappa_ren-minus-kappa}). It follows
that Eq.~(\ref{eq:Kraichnan-model-Ito-SDE-1}) can be written
as
\begin{eqnarray}
\partial_{t}\theta\left(t,\vec{x}\right)+v^{i}\left(t,\vec{x}\right)\partial_{i}\theta\left(t,\vec{x}\right)-\frac{\kappa_{{\rm ren}}}{2}\partial^{2}\theta\left(t,\vec{x}\right) & = & f\left(t,\vec{x}\right)\,. \quad
\label{eq:Kraichnan-model-Ito-SDE-2}
\end{eqnarray}
From the SDE (\ref{eq:Kraichnan-model-Ito-SDE-2}), one
can estimate the temporal dependence of $\bar{G}_{\theta\theta}\left(t,\vec{p}\right)$
via the following argument. Let us neglect the convection and the
forcing term. In this approximation
$\partial_{t}\langle\theta\left(t,\vec{x}\right)\theta\left(0,\vec{y}\right)\rangle=\frac{\kappa_{{\rm ren}}}{2}\partial_{x}^{2}\langle\theta\left(t,x\right)\theta\left(0,y\right)\rangle$,
which leads precisely to the temporal behavior found via field theoretical
methods, namely $\bar{G}_{\theta\theta}\left(t,\vec{p}\right)\propto e^{-\frac{\kappa_{{\rm ren}}}{2}p^{2}\left|t\right|}$.
Summarizing, the effect of computing $\Sigma_{\theta\bar{\theta}}$
(or, equivalently, to renormalize the microscopic molecular diffusivity)
simply amounts to introduce an ``Ito correction'' to the molecular
diffusivity. In this sense, the Ito convention makes the mixing between
the velocity and the scalar more transparent.
A more refined approach has been developed in Ref.~\onlinecite{MitraPandit2005},
where a differential equation for the two-point correlation function
at unequal time was found and investigated numerically confirming
the exponential decay of the two-point correlation function.
\section{Closed flow equation for the correlation functions of advected
scalars} \label{sec:closed-FRGE-for-advected-scalars}
In the Kraichnan model, the temporal dependence of the two-point correlation function
of the scalar can be obtained from various approaches, as shown in
Sec.~\ref{subsec:2pt-correlation-function-from-Dyson-eq}. This is due
to the key simplifying feature of this model, which is the white-in-time
Gaussian statistics of the velocity. However, the resulting temporal dependence
for the scalar is far from that of scalars in actual flows. However, the approximation
scheme developed in Secs.~\ref{subsec:Closed-flow-equations} and
\ref{subsec:2pt-function-in-large-P} is not based on the white-in-time
statistics of the velocity but on the symmetries of the action functional.
As we shall show in Sec.~\ref{subsec:Action-and-symmetries-for-NS_plus_scalar},
the symmetries for scalars advected by a velocity field satisfying
the NS equation are not much different from the symmetries of the Kraichnan model,
such that the same scheme is also applicable to real scalars, contrarily to the perturbative
approach. The main objective of this section is thus to derive
and solve the closed flow equation for the two-point correlation
function of the scalar field advected by NS turbulence by exploiting the strategy already followed
for the Kraichnan model.
\subsection{Action and symmetries for scalars advected by a NS velocity \label{subsec:Action-and-symmetries-for-NS_plus_scalar}}
To obtain the action functional for the scalar field advected by a NS turbulent flow,
we apply the MSRJD procedure to the system of equations
composed by the Navier-Stokes equation for an incompressible fluid and by the advection-diffusion equation (\ref{eq:Kraichnan-model_SDE}) for the scalar. In order to describe a stationary
turbulent state, we consider the NS equation equipped with a
Gaussian stochastic forcing. Following he conventions adopted in Ref.~\onlinecite{Canet:2014dta},
the action functional associated to the
NS equation is given by
\begin{eqnarray}
S_{{\rm NS}} & = & \int_{t\vec{x}}\left[\bar{v}^{i}\left(\partial_{t}+v^{j}\partial_{j}-\kappa_{\nu}\partial^{2}\right)v^{i}+\bar{v}^{i}\partial_{i}p+\bar{p}\partial_{i}v^{i}\right]\nonumber\\
&-&\int_{t\vec{x}\vx'}\frac{1}{2}\bar{v}^{i}N_{k,ij}\bar{v}^{j}\,,
\label{eq:action-NS}
\end{eqnarray}
where $\kappa_{\nu}$ denotes the kinematic viscosity (the notation $\nu$ is avoided to prevent confusion with the velocity $v$), $N_{k}$
is the covariance of the stochastic forcing of the fluid velocity
in the NS equation, for which the integral scale $L$ of the fluid
is identified with the inverse of the RG scale $k$,
$p$ is the pressure, and $\bar{p}$ serves
as a Lagrange multiplier which ensures the velocity to be divergenceless.
The pressure sector is fully non-renormalized, as can be shown by considering
gauged shifts in $p$ and $\bar{p}$, see Ref.~\onlinecite{Canet:2014dta} for the explicit Ward identities.
The covariance $N_{k,ij}$ can be chosen diagonal in components $N_{k,ij}=\delta_{ij}N_k$ because of incompressibility,
and $N_k$ is taken of the same form as the
one of the scalar forcing $M_{k}$ defined in (\ref{eq:def-forcing-M_k}).
The total action of the NS velocity plus scalar system is thus given by
\begin{eqnarray}
S & = & \int_{t\vec{x}}\left[\bar{\theta}\left(\partial_{t}+v^{i}\partial_{i}-\frac{\kappa}{2}\partial^{2}\right)\theta\right]-\int_{t\vec{x}\vx'}\frac{1}{2}\bar{\theta}M_{k}\bar{\theta}+S_{{\rm NS}}\,.\nonumber\\
\label{eq:action-NS_plus_scalar}
\end{eqnarray}
with the ghost sector implicit.
The action (\ref{eq:action-NS_plus_scalar}) shares some similarity with the Kraichnan
action (\ref{eq:Kraichnan-model-action}) with the main difference
that now the velocity does not appear quadratically but in the more
complicated form $S_{{\rm NS}}$. However, the symmetries
are essentially unaltered, as we now discuss.
In the FRG context, the symmetries and extended symmetries of the NS action (\ref{eq:action-NS})
have been discussed in Ref.~\onlinecite{Canet:2014dta}, to which we refer for
a detailed presentation. Here we limit ourselves to discussing the most
relevant points and the extension to the action (\ref{eq:action-NS_plus_scalar}).\\
\emph{Time-gauged Galilei transformation}. The Galilean transformation
has the same expression as (\ref{eq:Galilei-transformation}) with the response velocity field
and pressure fields transforming as scalars
$$ \bar{\varphi}\left(t,\vec{x}\right) \rightarrow \varphi\left(t,\vec{x}\right)-\varepsilon^{i}\left(t\right)\partial_{i}\varphi\left(t,\vec{x}\right) $$
for $\varphi=\bar{v}^{j}, p$ or $\bar{p}$.
The associated Ward identity reads
\begin{eqnarray}
&0& =
\int_{\vec{x}}\partial_{t}^{2}\bar{v}^{i}\left(t,\vec{x}\right)
+\int_{\vec{x}} \Biggr[\partial_{t}\frac{\delta\Gamma_{k}\left[\varphi\right]}{\delta v^{i}\left(t,\vec{x}\right)}+\frac{\delta\Gamma_{k}\left[\varphi\right]}{\delta v^{j}\left(t,\vec{x}\right)}\partial_{i}v^{j}\left(t,\vec{x}\right)
\nonumber \\
& & +\frac{\delta\Gamma_{k}\left[\varphi\right]}{\delta\bar{\theta}\left(t,\vec{x}\right)}
\partial_{i}\bar{\theta}\left(t,\vec{x}\right)+\frac{\delta\Gamma_{k}\left[\varphi\right]}{\delta\theta\left(t,\vec{x}\right)}\partial_{i}\theta\left(t,\vec{x}\right) +\frac{\delta\Gamma_{k}\left[\varphi\right]}{\delta\bar{p}\left(t,\vec{x}\right)}
\partial_{i}\bar{p}\left(t,\vec{x}\right)\nonumber \\
& & +\frac{\delta\Gamma_{k}\left[\varphi\right]}{\delta p\left(t,\vec{x}\right)}\partial_{i} p\left(t,\vec{x}\right) +\frac{\delta\Gamma_{k}\left[\varphi\right]}{\delta\bar{v}^{j}\left(t,\vec{x}\right)}\partial_{i}\bar{v}^{j}\left(t,\vec{x}\right)
\Biggr]\,. \qquad
\label{eq:WI-T-gauged-Galiei-for-NS_plus_scalar}
\end{eqnarray}
As in the case of the Kraichnan model, the Ward identity for time-gauged
Galilean invariance is the crucial ingredient for the closure of the
flow equation of any $n$-point correlation function in the limit of large wavenumber
since it allows one to express a $\left(n+1\right)$-vertex
with one velocity wavevector set to zero in terms of $n$-point vertices.\\
\emph{Time-gauged shifts of the response fields.} An important extended symmetry of the NS
action is related to the following time-gauged transformation
$\delta\bar{v}^{i}\left(t,\vec{x}\right)=\bar{\varepsilon}^{i}\left(t\right)$
and $\delta\bar{p}\left(t,\vec{x}\right)=\bar{\varepsilon}^{i}\left(t\right){v}^{i}\left(t,\vec{x}\right)$.
It yields the Ward identity
\begin{eqnarray*}
\int_{\vec{x}}\frac{\delta\Gamma_{k}}{\delta\bar{v}_{i}\left(t,\vec{x}\right)} & = & \int_{\vec{x}}\partial_{t}v^{i}+\int_{\vec{x}}v^{j}\left(t,\vec{x}\right)\partial_{j}v^{i}\left(t,\vec{x}\right)\,,
\end{eqnarray*}
which entails the non-renormalization of the Lagrangian time derivative
term $\int\bar{v}\left(\partial_{t}+v\cdot\partial\right)v$
in the bare action, and the vanishing of all the vertices of $n>2$ points with
one zero wavevector carried by a response velocity. \\
\emph{Shift symmetry of the scalar fields.}
The time-gauged shift symmetry $\theta\left(t,\vec{x}\right)\rightarrow\theta\left(t,\vec{x}\right)+\epsilon\left(t\right)$ is the same as in the Kraichnan model and
the Ward identity (\ref{eq:WI-t-local-shift-1}) is also valid for the NS scalar.
Thus, any $\left(n>2\right)$-point vertex with one zero wavevector carried by a
scalar field vanishes.
For the response scalar field, the time-gauged shift now involves the coupled transformation $\bar{\theta}\left(t,\vec{x}\right)\rightarrow\bar{\theta}\left(t,\vec{x}\right)+\varepsilon\left(t\right)$
and $\bar{p}\left(t,\vec{x}\right)\rightarrow\bar{p}\left(t,\vec{x}\right)+\varepsilon\left(t\right)\theta\left(t,\vec{x}\right)$, which leads to the modified Ward identity
\begin{eqnarray*}
0 & = & -\int_{\vec{x}}\left[\partial_{t}\theta\left(t,\vec{x}\right)+v^{i}\left(t,\vec{x}\right)\partial_{i}\theta\left(t,\vec{x}\right)\right]+\int_{\vec{x}}\frac{\delta\Gamma\left[\varphi\right]}{\delta\bar{\theta}\left(t,\vec{x}\right)}\,.
\end{eqnarray*}
This identity entails that the term $\int \bar\theta v_j\p_j \theta$ is not renormalized.
Furthermore, it yields that any vertex with a wavevector carried by a $\bar{\theta}$ set
to zero is actually given by its bare expression, and thus vanishes when $n>2$.\\
\emph{Synthesis}.
The key point is that the Ward identities for the scalar advected by the NS flow, although
slightly modified compared to the ones of the Kraichnan scalar, imply the same
consequence for 1-PI vertices, which is the cornerstone of the large wavenumber closure:
if at least one wavevector of the vertex is set to zero, then it vanishes
except when the wavevector is carried
by a velocity field, in which case it is controlled by (\ref{eq:WI-T-gauged-Galiei-for-NS_plus_scalar}).\footnote{Possible contributions to the flow equation from any bare-like three-point
vertex actually vanishes by incompressibility.}
\subsection{Anomalous dimensions}
In this section, we discuss the constraints on the anomalous
dimensions imposed
by the symmetries analyzed in Sec.~\ref{subsec:Action-and-symmetries-for-NS_plus_scalar}.
The analysis proceeds along the same line as for the Kraichnan model in
Sec.~\ref{subsec:Non-renormalization-theorem-and-anomalous-dim}, and the outcome
is similar in that the anomalous dimensions are essentially fixed by the symmetries
and the requirement of stationary turbulence.
We parametrize the EEA for the scalar advected by the NS flow in a similar way as the one for the Kraichnan scalar
\begin{eqnarray*}
\Gamma_{k} & = & \int_{t\vec{x}}\left[Z_{k,\bar{v}}^{1/2}\bar{v}^{i}\left(\partial_{t}+Z_{k,v}^{1/2}\lambda_{k,v}v^{j}\partial_{j}-{\kappa_{k,\nu}}\partial^{2}\right)Z_{k,v}^{1/2}v^{i}\right]\\
& & +\int_{t\vec{x}}\left[Z_{k,\bar{\theta}}^{1/2}\bar{\theta}\left(\partial_{t}+Z_{k,v}^{1/2}\lambda_{k,\theta}v^{i}\partial_{i}-\frac{\kappa_{k}}{2}\partial^{2}\right)Z_{k,\theta}^{1/2}\theta\right]+\cdots\,,
\end{eqnarray*}
where we introduced the field renormalizations $Z_{k,\varphi^{i}}$,
the couplings $\lambda_{k,v}$ and $\lambda_{k,\theta}$, and the running
viscosity and running molecular diffusivity, $\kappa_{k,\nu}$ and $\kappa_{k}$
respectively. In full analogy with the analysis of section \ref{subsec:Non-renormalization-theorem-and-anomalous-dim},
one deduces from the non-renormalization theorems the following constraints: $Z_{k,\bar{v}}^{1/2}Z_{k,v}^{1/2}=1$,
$Z_{k,v}^{1/2}\lambda_{k,v}=1$, $Z_{k,\bar{\theta}}^{1/2}Z_{k,\theta}^{1/2}=1$,
and $Z_{k,v}^{1/2}\lambda_{k,\theta}=1$. In terms of couplings and anomalous
dimensions, it follows that $\eta_{\bar{v}}=-\eta_{v}$, $\eta_{\bar{\theta}}=-\eta_{\theta}$,
and $\lambda_{k,\theta}=\lambda_{k,v}\equiv \lambda_k=1$, which plays the role of the coupling
appearing in the Galilean covariant derivative. We also associate
anomalous dimensions $\eta_{\kappa_{\nu}}$ and $\eta_{\kappa}$ to the viscosity
and the molecular diffusivity, respectively.
We assume that the scalar is passive, which means that its backreaction onto
the velocity flow is negligible. In the framework of the RG, this
translates into the statement that none of the quantities related
to the velocities, in particular $\eta_{v}$, $\eta_{\kappa_{\nu}}$, and $\lambda_{k,v}$,
are affected by the scalar field and its fluctuations. Therefore,
in order to determine such quantities, we simply neglect the
terms in the EAA which contains a scalar field. This part of
the EAA then simply corresponds to the NS equations, already studied in Ref.~\onlinecite{Canet:2014dta},
and of which we only give a concise account. In essence, a
dimensionless coupling $\hat{\lambda}_{k,v}$ is introduced as
$\lambda_{k,v}\equiv\hat{\lambda}_{k,v}{\kappa_{k,\nu}}^{3/2}k^{-d/2+1}$
and its beta function is simply given by $\partial_{s}\hat{\lambda}_{k,v}=\left(3\eta_{\kappa_{\nu}}-\eta_{\bar{v}}+d-2\right)\hat{\lambda}_{k,v}/2$
since $\lambda_{k,v}$ is not renormalized.
It follows that any non-Gaussian fixed point satisfies $3\eta_{\kappa_{\nu}}-\eta_{\bar{v}}+d-2=0$.
Finally, in analogy with the arguments of Sec.~\ref{subsec:Non-renormalization-theorem-and-anomalous-dim},
one may determine the anomalous dimensions by requiring $\lim_{k\rightarrow0}\langle f\cdot v\rangle$
and $\lim_{\kappa_\nu\rightarrow0}{\kappa_\nu}\langle\partial_{i}v_{j}\partial_{i}v_{j}\rangle$
to be non-zero constants. This leads in $d=3$ to fix $\eta_{\bar{v}}=d+2=5$,
which then yields that at a non-Gaussian fixed point $\eta_{\kappa_\nu}=4/3$.
Let us now turn to the anomalous dimensions and the couplings inherent to
the scalar dynamics. We introduce the dimensionless coupling $\hat{\kappa}_k\equiv\kappa_{k}/{\kappa_{k,\nu}}$.
The associated beta function is given by $\partial_{s}\hat{\kappa}_k=\left(-\eta_{\kappa}+\eta_{\kappa_\nu}\right)\hat{\kappa}_k$.
At any non-trivial fixed point of $\hat{\kappa}$, one thus has $\eta_{\kappa}=\eta_{\kappa_\nu}$.
This means that the dynamical critical
exponent $z$ is controlled by the viscosity and that it also
applies to the scalar sector. To determine the anomalous dimensions
of the scalars, we proceed as in Sec.~\ref{subsec:Non-renormalization-theorem-and-anomalous-dim}.
One may either require that $\lim_{k\rightarrow0}\langle f_{\theta}\theta\rangle$
approaches a constant, which implies $\eta_{\bar{\theta}}=d+2$, or
that $\lim_{\kappa\rightarrow0}\kappa\langle\partial_{i}\theta\partial_{i}\theta\rangle$
is finite, which leads to the same result. In this latter case, the difference with respect to
Eq.~(\ref{eq:dissipation-limit-kappa_to_zero-and-kappa-power}) is that
the dissipation scale is provided by $\eta_{{\rm diss}}^{-1}\equiv\left(\epsilon_{\theta}/\kappa^{3}\right)^{1/4}$,
with $\epsilon_{\theta}$ the scalar energy density conserved
through the cascade.
\subsection{Closed flow equation for the two-point correlation function}
Since the general form of the flow equation and the structure of the Ward
identities for the passive scalar in the NS flow are the same as in the Kraichnan
model, one can closely follow the derivation of Secs.~\ref{subsec:Closed-flow-equations}
and \ref{subsec:2pt-function-in-large-P} to obtain a closed flow equation for any $n$-point correlation
function of the scalar field in the limit of large wavenumber.
The inverse propagator $\big(\Gamma_k^{(2)}+{R}_k\big)$ is now a $8\times 8$ matrix,
comprising the scalar field sector ($\theta,\bar{\theta},c,\bar{c}$) and the velocity
sector ($v_i,\bar{v}_i,p,\bar{p}$), but it has a simple structure since it is block-diagonal,
and also diagonal in frequency and wavevector in Fourier space.
It follows that the renormalized propagator bares a similar structure as Eq.~(\ref{eq:propag}), with
the $G_{vv}$ now replaced by a $4\times 4$ matrix whose general form can be found in Ref.~\onlinecite{Canet:2014dta}.
One can show that the pressure sector decouples (as the ghost sector) and the only part which plays a role is the velocity
sector, whose propagator is given by
\begin{eqnarray*}
\left(\bar{\Gamma}_{k}^{\left(2\right)}+R_{k}\right)^{-1}_{v\bar{v}} & = & \left(\begin{array}{cc}
\bar{G}_{v^\alpha v^\beta} & \bar{G}_{v^\alpha\bar{v}^\beta} \\
\bar{G}_{\bar{v}^\alpha v^\beta} & 0
\end{array}\right)\,,
\end{eqnarray*}
with
\begin{eqnarray*}
\bar{G}_{v^\alpha \bar{v}^\beta}( \omega,\vec{q}) &=&
\frac{P_{\alpha\beta}(\vec{q})}{\bar{\Gamma}^{(2)}_{v^\alpha \bar{v}^\beta}( -\omega,\vec{q}) + R_{k,v\bar{v}}(\vec{q})}\,, \\
\bar{G}_{v^\alpha {v}^\beta}( \omega,\vec{q})&=& -P_{\alpha\beta}(\vec{q})
\frac{\bar{\Gamma}^{(2)}_{\bar{v}^\alpha \bar{v}^\beta}( \omega,\vec{q}) -2 N_{k,\alpha\beta}(\vec{q})}{\left| \bar{\Gamma}^{(2)}_{v^\alpha \bar{v}^\beta}+R_{k,v\bar{v}}(\vec{q})\right|^2},
\end{eqnarray*}
where the field functional derivative are indicated as indices on the $\Gamma^{(2)}$ to alleviate notation.
We now consider the generic flow equation (\ref{eq:FRGE-W(n)-1}) for a $n$-point correlation function $W_{k}^{\left(n\right)}$.
In the large wavenumber limit, since we have shown that all vertices with one zero-wavevector
vanish but the ones for which it is carried by a velocity field, the only remaining term in
this flow equation is
\begin{eqnarray*}
&&
\partial_{s}\frac{\delta^{n}W_{k}\left[J\right]}{\delta J^{1}\cdots\delta J^{n}}
= -\frac{1}{2} H_{k,i j} \frac{\delta}{\delta v^{i}}\frac{\delta}{\delta v^{j}}\frac{\delta^{n}W_{k}\left[J\right]}{\delta J^{1}\cdots\delta J^{n}}\,,
\end{eqnarray*}
where we introduced the notation
\begin{eqnarray*}
&&
H_{k,i j}\left(\omega,\vec{q}\right) \\
&& \qquad\qquad
\equiv \left(2\bar{G}_{v^{i}v^{\alpha}}\partial_{s}R_{k,v^{\alpha}\bar{v}^{\beta}} \Re (\bar{G}_{\bar{v}^{\beta}v^{j}})+\bar{G}_{v^{i}\bar{v}^{\alpha}}\partial_{s}N_{k,\alpha\beta}\bar{G}_{\bar{v}^{\beta}v^{j}}\right)\,.
\end{eqnarray*}
Hence this equation can be closed exploiting the time-gauged Galilean Ward identity
\begin{eqnarray*}
&&
\frac{\delta}{\delta v^{i}\left(-\omega,-\vec{q}\right)}\frac{\delta}{\delta v^{j}\left(\omega,\vec{q}\right)}W_{k}^{\left(n\right)}\left(\omega_{1},\vec{p}_{1},\cdots\right)\Bigr|_{\vec{q}=0} \\
&& \qquad\qquad\qquad\qquad
= {\cal D}^{i}\left(-\omega\right){\cal D}^{j}\left(\omega\right)W_{k}^{\left(n\right)}\left(\omega_{1},\vec{p}_{1},\cdots\right)\,.
\end{eqnarray*}
One thus also sobtains for the scalar in a turbulent NS flow a closed flow equation for any $n$-point correlation
function of any fields since at this stage, the $J^m$ are any of the sources.
In the following, we focus on the two-point correlation function of the scalar field. The explicit expression
of its flow equation, with an inverse Fourier transform to revert to time-wavevector coordinates, is given at large $p$ by
\begin{eqnarray}
&&
\partial_{s}\bar{G}_{\theta\theta}\left(t,\vec{p}\right) \nonumber\\
&&= \frac{1}{2}\int_{\omega \vec{q}} H_{k,ij }\left(\omega,\vec{q}\right)
p^{i}p^{j }\Biggr\{\frac{2-2\cos\left(t\omega\right)}{\omega^{2}}\Biggr\}\bar{G}_{\theta\theta}\left(t,\vec{p}\right)\,.\label{eq:FRGE-Gqq(tp)-advected-scalar-1}
\end{eqnarray}
Since $H_{k,ij }= H_k P_{ij }$, one can further simplify
Eq.~(\ref{eq:FRGE-Gqq(tp)-advected-scalar-1}) replacing
$H_{k,ij } p^i p^j \to \frac{d-1}{d} H_k p^2$ in the integral using isotropy.
Eq.~(\ref{eq:FRGE-Gqq(tp)-advected-scalar-1}) is formally very
similar to Eq.~(\ref{eq:FP-FRGE-Gqq-1}) obtained for the
Kraichnan scalar. However, there are some crucial differences.
First, in the case of the NS passive scalar, the velocity
propagators are renormalized and are not known exactly. Second,
the velocity propagators, which enter in (\ref{eq:FRGE-Gqq(tp)-advected-scalar-1}) through $H_{k}$,
have a non-trivial frequency dependence, contrarily to the
case of the Kraichnan model. Since their exact form is not known,
one cannot obtain an explicit expression for the resulting
frequency integral. However, these integrals can be simplified in both the short-time and long-time limits,
which we study below.
\subsubsection{Short-time limit of the two-point correlation function}
Let us consider the limit $t\rightarrow0$ of Eq.~(\ref{eq:FRGE-Gqq(tp)-advected-scalar-1}).
In this limit, the cosine can be expanded which yields
\begin{eqnarray*}
\partial_{s}\bar{G}_{\theta\theta}\left(t,\vec{p}\right) & = & \frac{d-1}{2d}\int_{\omega \vec{q}}H_{k}\left(\omega,\vec{q}\right)p^2 t^{2}\bar{G}_{\theta\theta}\left(t,\vec{p}\right)\,,
\end{eqnarray*}
since the frequency integration converges
thanks to the function $H_{k}$.
The dimensionless two-point function is defined by $\bar{G}_{\theta\theta}\left(t,\vec{p}\right)\equiv k^{\eta_{\kappa}-\eta_{\bar{\theta}}}\hat{\bar{G}}_{\theta\theta}\left(\hat{t},\hat{\vec{p}}\right)$,
and it satisfies the flow equation
\begin{eqnarray*}
&&
\left(\partial_{s}+\left(\eta_{\kappa}-\eta_{\bar{\theta}}\right)+\left(2-\eta_{\kappa}\right)\hat{t}\partial_{\hat{t}}-\hat{\vec{p}}\partial_{\hat{\vec{p}}}\right)\hat{\bar{G}}_{\theta\theta}
\left(\hat{t},\hat{\vec{p}}\right) \\
&& \qquad\qquad\qquad\qquad
= \hat{\alpha}_k \hat{\lambda}_k^{2}\hat{p}^{2}\hat{t}^{2}\hat{\bar{G}}_{\theta\theta}\left(\hat{t},\hat{\vec{p}}\right)\,,
\end{eqnarray*}
where we introduced the scale dependent parameter $\hat{\alpha}_k\equiv\frac{d-1}{2d}\int_{\hat{\omega}\hat{\vec{q}}}\hat{H}_k\left(\hat{\omega},\hat{\vec{q}}\right)$.
At the fixed point, this parameter is just a number $\hat{\alpha}_k\to \hat{\alpha}_{{\rm s}}$
and the fixed-point equation reads, now specifying to $d=3$
\begin{eqnarray}
&&
\left(-\frac{11}{3}+\frac{2}{3}\hat{t}\partial_{\hat{t}}-\hat{\vec{p}}\partial_{\hat{\vec{p}}}\right)\hat{\bar{G}}_{\theta\theta}\left(\hat{t},\hat{\vec{p}}\right)
\nonumber \\
&& \qquad\qquad\qquad\qquad\qquad
= \hat{\alpha}_{{\rm s}}\hat{\lambda}^{2}\hat{p}^{2}\hat{t}^{2}\hat{\bar{G}}_{\theta\theta}\left(\hat{t},\hat{\vec{p}}\right)\,.\label{eq:FP-FRGE-Gqq-short-time}
\end{eqnarray}
The solution of this equation is
\begin{eqnarray*}
\hat{\bar{G}}_{\theta\theta}\left(\hat{t},\hat{\vec{p}}\right) & = & \hat{p}^{-\frac{11}{3}}e^{-\frac{3}{2}\hat{\alpha}_{{\rm s}}\hat{\lambda}^{2}\hat{p}^{2}\hat{t}^{2}}\hat{f}\left(\hat{p}^{2/3}\hat{t}\right)\,,
\end{eqnarray*}
where $\hat{f}\left(\hat{p}^{2/3}\hat{t}\right)$ is a universal scaling function, which could be computed by explicitly
integrating the flow from the initial condition.
Reverting to dimensionful variables, and neglecting the subleading time dependence of the scaling function, one obtains
\begin{eqnarray}
G_{\theta\theta}\left(t,\vec{p}\right) = C_{\rm s} \epsilon_\theta \epsilon^{-1/3}_v
{p}^{-\frac{11}{3}} e^{-\hat{\alpha}_{\rm s} (L/\tau_0)^2 p^2 t^2}\,,
\quad
\label{eq:2pt-correlation-short-time}
\end{eqnarray}
where $\tau_0\equiv \left(L^2/\epsilon_v \right)^{-1/3}$ denotes the eddy-turnover time at the energy injection scale
and $C_{\rm s}$ and $\hat{\alpha}_{\rm s}$ are non-universal constants (the factor 3/2 was absorbed in
the latter).
In Eq.~(\ref{eq:2pt-correlation-short-time}) $\epsilon_\theta$ and $\epsilon_v$
denote the mean rate of energy injection of the scalar and of the velocity,
respectively.
\subsubsection{Large-time limit of the two-point correlation function}
We now consider the limit $t\rightarrow\infty$ of Eq.~(\ref{eq:FRGE-Gqq(tp)-advected-scalar-1}).
In this limit, it is straighforward to check that the frequency integration
is dominated by the term in the curly brackets of Eq.~(\ref{eq:FRGE-Gqq(tp)-advected-scalar-1}).
Performing the frequency integration, one obtains
\begin{eqnarray*}
\partial_{s}\bar{G}_{\theta\theta}\left(t,\vec{p}\right) & = & \frac{d-1}{2d}\int_{\vec{q}}H_{k}\left(0,\vec{q}\right)p^{2}\left|t\right|\bar{G}_{\theta\theta}\left(t,\vec{p}\right)\,.
\end{eqnarray*}
In analogy with the short-time limit, we thus obtain the
fixed point equation in $d=3$
\begin{eqnarray}
&&
\left(-\frac{11}{3}+\frac{2}{3}\hat{t}\partial_{\hat{t}}-\hat{\vec{p}}\partial_{\hat{\vec{p}}}\right)\hat{\bar{G}}_{\theta\theta}\left(\hat{t},\hat{\vec{p}}\right)
\nonumber \\
&& \qquad\qquad\qquad\qquad\qquad\quad
= \hat{\alpha}_{{\rm \ell}}\hat{\lambda}^{2}\hat{p}^{2}\left|\hat{t}\right|\hat{\bar{G}}_{\theta\theta}\left(\hat{t},\hat{p}\right)\,.\label{eq:FP-FRGE-Gqq-large-time}
\end{eqnarray}
where $\hat{\alpha}_{{\rm \ell}}$ is the fixed point value of $\frac{1}{3}\int_{\hat{\vec{q}}}\hat{H}_k\left(0,\hat{\vec{q}}\right)$.
The solution to equation (\ref{eq:FP-FRGE-Gqq-large-time}) is given by
\begin{eqnarray*}
\hat{\bar{G}}_{\theta\theta}\left(\hat{t},\hat{\vec{p}}\right) & = & \hat{p}^{-\frac{11}{3}}e^{-\frac{3}{4}\hat{\alpha}_{{\rm \ell}}\hat{\lambda}^{2}\hat{p}^{2}|\hat{t}|}\hat{f}\left(\hat{p}^{2/3}\hat{t}\right)\,,
\end{eqnarray*}
which in dimensionful variables implies
\begin{eqnarray}
G_{\theta\theta}\left(t,\vec{p}\right) = C_{\rm \ell} \epsilon_\theta \epsilon^{-1/3}
{p}^{-\frac{11}{3}} e^{-\hat{\alpha}_{\rm \ell} \frac{L^2}{\tau_0} p^2 |t|}\,,
\label{eq:2pt-correlation-long-time}
\end{eqnarray}
where $C_{\rm \ell}$ and $\hat{\alpha}_{\rm \ell}$ are non-universal constants (with the numerical factor absorbed in the latter).
Equations (\ref{eq:2pt-correlation-short-time}) and (\ref{eq:2pt-correlation-long-time}) constitute the main results of this work
and generalize the temporal dependence found for the Kraichnan model
to scalar fields advected by a NS flow. As anticipated, when the carrier turbulent velocity field
have some temporal correlations, which is always the case for any realistic flows, then the short-time behavior
of the scalar is Gaussian in $t$. The exponential decay is only observed at sufficiently large time scales.
This is the main difference with the scalar advected by the idealized Kraichnan stochastic velocity,
where the short-time regime is eliminated by the time delta-correlation of the velocity covariance.
\section{Summary and perspectives} \label{sec:summary}
In this work, we have studied the temporal dependence
of the two-point correlation function
of turbulent passive scalar fields in the inertial-convective range. We analyzed
the symmetries, and extended symmetries, of the action functional of the system,
and derived the corresponding Ward identities. These identities
are the crucial ingredients which enabled us to obtain, within the FRG framework in the limit of large wavenumbers, a closed flow equation for
generic Eulerian $n$-point correlation functions of the scalar field, advected both by the Gaussian stochastic velocity
of the Kraichnan model, or by a turbulent NS velocity field.
We have focused on the solution at the fixed point of the two-point
correlation function of the scalar. We have shown that
its temporal decay exhibits two regimes: a Gaussian decay at small time delays,
which crosses over to an exponential decay at large time delays,
with a coefficient proportional to $k^{2}$ in both regimes.
This time dependence explicitly breaks standard scale invariance.
This demonstrates that the scalar field, in the inertial-convective range considered here,
inherits the time correlations of the carrier fluid, since this exactly
corresponds to the behavior of a NS velocity
field \cite{Tarpin:2018yvs}.
The Kraichnan model replaces the carrier NS fluid velocity by a Gaussian
stochastic field with delta-correlation in time.
This leads to great simplifications, as was already recognized in
many works. As a consequence, we obtained an explicit expression for the prefactor
in the exponential, and we also derived the temporal behavior
of the two-point function in this model using a perturbative approach,
which is not possible for the NS flow. This approach yields
results compatible with the FRG approach, and provides a useful connection with former results.
The prize of this simplification is that it significantly alters
the temporal behavior of the advected scalar, since the small-time
regime is destroyed in this case.
The approximation scheme used in this work, based on a large wavenumber expansion, is rooted in the symmetries of
the system. Since the symmetries of the advected scalar field are also
present in other regimes, say the viscous-convective regime,
it is conceivable that this approach can be extended
outside the inertial-convective range. We leave this task for future
work. Let us emphasize that all the results obtained in this work
on the temporal behavior of turbulent passive scalar fields
can be tested in direct numerical simulations, which is underway \cite{Gorbunova:2021b}.
These results may also initiate experimental investigations.
Another important direction which deserves further study is the combined use
of numerical solutions to the FRG flow equations, via ans{\"a}tze, and the symmetry
arguments presented in this paper. This will allow one in particular to
study correlation functions which include
composite operators, and investigate their anomalous scalings. \footnote{We refer to
Ref.~\onlinecite{Pagani:2016pad,Pagani:2017tdr,Pagani:2020ejb,Rose:2015bma}
for details regarding the composite operators within the FRG framework.}
We hope to tackle these issues in the near future.
\begin{acknowledgments}
The authors thank B.~Delamotte and M.~Tarpin for discussions and
collaboration at the initial stages of the project,
A.~Gorbunova for discussions and collaboration on
related projects,
and G.~Eyink for useful references.
L.C.~acknowledges support from the ANR-18-CE92-0019 Grant NeqFluids and support from Institut Universitaire de France.
\end{acknowledgments}
|
\section{Introduction}
Human pose estimation is a fundamental problem in computer vision, which aims at locating anatomical keypoints (\emph{e.g.}, wrist, ankle, etc.) or body parts. It has enormous applications in diverse domains such as security, violence detection, crowd riot scene identification, human behavior understanding, and action recognition \cite{liu2016hierarchical}. Earlier methods \cite{wang2008multiple,wang2013beyond,zhang2009efficient,sapp2010cascaded} adopt the probabilistic graphical model or the pictorial structure model.
Recent methods have built upon the success of deep convolutional neural networks (CNNs) \cite{Cao_2017_CVPR, Toshev_2014_CVPR, Wei_2016_CVPR, Fang_2017_ICCV, newell2016stacked, chu2017multi, luo2018lstm, liu2021aggregated, liu2019towards}, achieving outstanding performance in this task. Unfortunately, most of the recent state-of-the-art methods are designed for static images, with greatly diminished performance when handling video input.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.95\linewidth]{Figures/Fig1-PTM.jpg}
\end{center}
\caption{An illustration of our Pose Temporal Merger (PTM) network. \textbf{(a):} Original video sequence in the datasets, and we aim to detect poses in the current frame ${F}_c$.
\textbf{(b):} Each person in the original video sequence is assembled to a cropped clip and a single-person joints detector gives preliminary estimations of keypoint heatmaps (illustrated for right wrist).
\textbf{(c)-left:} Merged keypoint heatmaps for the right wrist, generated by our PTM network through encoding the keypoint spatial contexts. The color intensity encodes spatial aggregation.
\textbf{(c)-right:} Zoomed-in view of the merged keypoint heatmaps.}
\label{fig:heatmap_fuse}
\vspace{-0.5em}
\end{figure}
In this paper, we focus on the problem of multi-person pose estimation in video sequences. Conventional image-based approaches disregard the temporal dependency and geometric consistency across video frames. Dissevering these additional cues results in failure cases when dealing with challenging situations that inherently occurs in video sequences such as motion blur, video defocus, or pose occlusions. Effectively leveraging the temporal information in video sequences is of great significance to facilitate pose estimation and often plays an indispensable role for detecting heavily occluded or blurry joints.
A direct and intuitive approach to tackle this issue is to employ recurrent neural networks (RNNs) such as Long-Short Term Memory (LSTM), Gate Recurrent Unit (GRU) or 3DCNNs to model geometric consistency as well as temporal dependency across video frames. \cite{luo2018lstm} uses convolutional LSTM to capture temporal and spatial cues, and directly predicts the keypoint heatmap sequences for videos. This RNN based approach is more effective when the human subjects are spatially sparse such as single-person scenes with minimal occlusion. However, performance is severely hindered in the case of occlusion commonly occurring in multi-person pose estimation and even self-occlusion in the single-person case. \cite{wang2020combining} proposes a 3DHRNet (extension of HRNet \cite{sun2019deep} to include a temporal dimension) for extracting spatial and temporal features across video frames to estimate pose sequences. This model has shown excellent results particularly for adequately long duration single-person sequences. Another line of work considers fine-tuning the primary prediction with high confidence keypoints from the adjacent frames. \cite{song2017thin, pfister2015flowing} propose to compute the dense optical flow between every two frames, and leverage the additional flow based representations to align the predictions. This approach is promising when the optical flow can be computed precisely. However in cases involving motion blur or defocus, the poor image qualities lead to imprecise optical flows which translates to performance drops.
To address the shortcomings of existing methods, we propose to incorporate consecutive frames from dual temporal directions to improve pose estimation in videos. Our framework, termed Dual Consecutive network for pose estimation (DCPose), first encodes the spatial-temporal keypoint context into localized search scopes, computes pose residuals, and subsequently refines the keypoint heatmap estimations. Specifically, we design three task-specific modules within the DCPose pipeline. 1) As illustrated in Fig.~\ref{fig:heatmap_fuse}, a Pose Temporal Merger (PTM) network performs keypoints aggregation over a continuous video segment (\emph{e.g.}, three consecutive frames) with group convolution, thereby localizing the search range for the keypoint. 2) A Pose Residual Fusion (PRF) network is introduced to efficiently obtain the pose residuals between the current frame and adjacent frames. PRF computes inter-frame keypoint offsets by explicitly utilizing the temporal distance. 3) Finally, a Pose Correction Network (PCN) comprising five parallel convolution layers with different dilation rates is proposed for resampling keypoint heatmaps in the localized search range.
It is worth mentioning that the architecture of our network extends the successful PoseWarper architecture \cite{bertasius2019learning} in three ways. (1) PoseWarper focuses on enabling effective label propagation between frames, while we aim to refine the pose estimation of current frame using the motion context and temporal information from unlabeled neighboring frames. (2) Information from two directions are utilized and we explicitly consider weighted residuals between frames. (3) Instead of applying the learned warping operation to a heatmap from one adjacent frame, the new network fuses together heatmaps from the adjacent frames and the current frame.
To summarize, our key contributions are: 1) A novel dual consecutive pose estimation framework is proposed. DCPose effectively incorporates bidirectional temporal cues across frames to facilitate the multi-person pose estimation task in videos. 2) We design 3 modular networks within DCPose to effectively utilise the temporal context: i) a novel Pose Temporal Merger network for effectively aggregating keypoint across frames and identifying a search scope, ii) a Pose Residual Fusion network to efficiently compute weighted pose residuals across frames, and iii) a Pose Correction Network that updates the pose estimation with the refined search scope and pose residual information. 3) Our method achieves state-of-the-art results on PoseTrack2017 and PoseTrack2018 Multi-frame Person Pose Estimation Challenge. To facilitate future research, our source code is released at \url{https://github.com/Pose-Group/DCPose}.
\begin{figure*}
\begin{center}
\includegraphics[width=0.95\linewidth]{Figures/Fig2-Pipeline.png}
\end{center}
\caption{Overall pipeline of our DCPose framework. The goal is to locate the keypoint positions for the current frame ${F}_c$. First, an individual person $i$ is assembled into an input sequence $\mathbf{Clip}_i(p,c,n)$, and a HRNet backbone predicts initial keypoint heatmaps $\mathbf{h}_i(p),\mathbf{h}_i(c),\mathbf{h}_i(n)$.
Our Pose Temporal Merger (PTM) and Pose Residual Fusion (PRF) networks work concurrently to obtain an effective search scope $\Phi_i(p,c,n)$ and pose residuals $\Psi_i(p,c,n)$, respectively. These are then fed into our Pose Correction Network (PCN) which refines the keypoint estimation for person $i$ in ${F}_c$.}
\label{fig:pipeline}
\end{figure*}
\section{Related Work}
\subsection{Imaged Based Multi Person Pose Estimation}
Earlier image-based human pose estimation works generally fall within a pictorial structure model paradigm, in which the human body is represented as a tree-structured model \cite{wang2008multiple,wang2013beyond,zhang2009efficient,sapp2010cascaded} or a forest model \cite{sun2012conditional,dantone2013human}. Despite allowing for efficient inference, these approaches tend to be insufficient in modeling complex relationships between body parts, and this weakness is accentuated when temporal information enters the picture.
Recently, neural networks based methods \cite{lin2018learning, liu2015graph, tang2019learning, artacho2020unipose, cheng2020higherhrnet, huang2020devil, zhang2020distribution, varamesh2020mixture, su2019multi, Chengzhiyong, ijcai2020} have been in the spotlight due to their superior performance in various fields. One line of work \cite{carreira2016human} outputs skeletal joints coordinates directly by regressing image features. Another approach \cite{newell2016stacked, pfister2015flowing} utilises probability heatmaps to represent joints locations. Due to the reduced difficulty for optimization, heatmap based pose estimation have since been widely adopted. In general, these methods can be classified into part-based framework (bottom-up) and two-step framework (top-down). The bottom-up approach \cite{Cao_2017_CVPR, kocabas2018multiposenet, kreiss2019pifpaf,li2019crowdpose} first detects individual body parts, then assembles these constituent parts into the entire person. \cite{Cao_2017_CVPR} builds a bottom-up pipeline and utilizes part affinity fields to capture pairwise relationships between different body parts. Conversely, the top-down approach \cite{Fang_2017_ICCV, xiao2018simple, Wei_2016_CVPR, sun2019deep, newell2016stacked, moon2019posefix} first performs person detection, then proceeds with single-person pose estimation on each individual. \cite{Wei_2016_CVPR} proposes a sequential architecture of convolutional pose machines, which follows a strategy of iteratively refining the output of each network stage. \cite{Fang_2017_ICCV} designs a symmetric spatial transformer network for extracting a high-quality single person region from an inaccurate bounding box. A recent work in \cite{sun2019deep} proposes a HRNet that performs multi-scale fusion to retain high resolution feature maps. This improves spatial precision in keypoint heatmaps and achieves the state-of-the-art on several image-based benchmarks.
\subsection{Video Based Multi Person Pose Estimation}
Directly applying the existing image-level methods to video sequences produces unsatisfactory predictions, primarily due to the failure to capture temporal dependency among video frames. Consequently, these models fail to handle motion blur, video defocus, or pose occlusions which are frequently encountered in video inputs. \cite{pfister2015flowing, song2017thin, weinzaepfel2013deepflow} compute the dense optical flow between every consecutive frames with the flow representations providing additional cues for aligning predictions. However, motion blur, defocus or occlusion occurrences hinder optical flow computation and affect performance. \cite{luo2018lstm} replaces the convolutional pose machines in \cite{Wei_2016_CVPR} with convolutional LSTMs for modeling temporal information in addition to spatial contexts. A principal shortcoming of such an approach is being severely impacted by occlusion. \cite{bertasius2019learning} proposes to learn an effective video pose detector from sparsely labeled videos through a warping mechanism and has turned out to be very successful, dominating the PoseTrack leaderboard for a long time. \cite{wang2020combining} extends HRNet \cite{sun2019deep} with temporal convolutions and proposes 3DHRNet, which is successful in handling pose estimation and tracking jointly.
\section{Our Approach}\label{section:method}
The pipeline of our proposed DCPose is illustrated in Fig.~\ref{fig:pipeline}. To improve keypoint detection for the current frame ${F}_c$, we make use of additional temporal information from a previous frame ${F}_p$ and a future frame ${F}_n$. ${F}_p$ and ${F}_n$ are selected within a frame window $[c-T,c+T]$, where $p \in [c-T, c)$ and $n \in (c, c+T]$ respectively denote the frame indices. The bounding boxes for individual persons in $F_c$ are first obtained by a human detector. Each bounding box is enlarged by 25\% and is further used to crop the same person in $F_p$ and ${F}_n$. Individual $i$ in the video will thus be composed of a cropped video segment, which we denote as $\mathbf{Clip}_i(p,c,n)$.
$\mathbf{Clip}_i(p,c,n)$ is then fed into a backbone network that serves to output preliminary keypoint heatmap estimations $\mathbf{h}_i(p,c,n)$. The pose heatmaps $\mathbf{h}_i(p,c,n)$ is then processed in parallel through two modular networks, Pose Temporal Merger (PTM) and Pose Residual Fusion (PRF). PTM outputs $\Phi_i(p,c,n)$, {which encodes the spatial aggregation}, and PRF computes $\Psi_i(p,c,n)$, {which captures pose residuals in two directions}. Both feature tensors $\Phi_i(p,c,n)$ and $\Psi_i(p,c,n)$ are then simultaneously fed into our Pose Correction Network (PCN) to refine and improve upon the initial pose estimations. In what follows, we introduce the three key components in detail.
\subsection{Pose Temporal Merger}
The motivation for our Pose Temporal Merger (PTM) comes from the following observations and heuristics. 1) Although existing pose estimation methods such as \cite{sun2019deep, Fang_2017_ICCV} suffer from performance deterioration on videos, we observe that their predictions do still provide useful information for approximating the keypoint spatial positions. 2) Another heuristic is on temporal consistency, \emph{i.e.,} the pose of an individual does not undergo dramatic and abrupt changes across a very few frame intervals (typically 1/60 to 1/25 of a second per frame). Therefore, we design PTM to encode the keypoint spatial contexts based on initial predictions (from a backbone network), providing a compressed search scope that facilitates refinement and correction of pose prediction within a confined range.
For person $i$, the backbone network returns initial keypoint heatmaps $\mathbf{h}_i(p),\mathbf{h}_i(c),\mathbf{h}_i(n)$. Naively, we could merge them through a direct summation $\mathbf{H}_i(p,c,n)=\mathbf{h}_i(p)+\mathbf{h}_i(c)+\mathbf{h}_i(n)$. However, we expect that the additional information that may be extracted from ${F}_p$ and ${F}_n$ is inversely proportional to their temporal distances from current frame ${F}_c$. We formalize this intuition as:
\begin{equation}
\begin{aligned}
\mathbf{H}(p,c,n) = \frac{n-c}{n-p}\mathbf{h}_i(p)+\mathbf{h}_i(c)+\frac{c-p}{n-p}\mathbf{h}_i(n). \label{eqn:mergedheatmap}
\end{aligned}
\end{equation}
Recall that $p,c,n$ are frame indices. We explicitly assign higher weights to the frames that are temporally nearer to the current frame.
Based on the important fact that convolution operations serve to adjust (feature) weights, we utilise convolutional neural networks to practically implement the idea
of Eq.~\ref{eqn:mergedheatmap}. However, including all joint channels in the computation for the merged keypoint heatmap of a single joint will result in redundancy. For example, when encoding the spatial context of the left wrist, involving other joints such as the head and ankle at different times will likely not to have any bearing and may even breed confusion.
Therefore, for each joint, we only include its own specific temporal information for computing its merged keypoint heatmap. This is implemented via a group convolution. We regroup the keypoint heatmaps $\mathbf{h}_i(p),\mathbf{h}_i(c),\mathbf{h}_i(n)$ according to joint, and stack them to a feature tensor $\phi_i$, which can be expressed as:
\begin{equation}
\begin{aligned}
\phi_i(p,c,n) = \bigoplus_{j=1}^{N} \frac{n-c}{n-p}\mathbf{h}_i^j(p)\oplus\mathbf{h}_i^j(c)\oplus\frac{c-p}{n-p}\mathbf{h}_i^j(n)
\end{aligned}
\end{equation}
where $\oplus$ denotes the concatenate operation and the superscript $j$ index the $j$-th joint for a total of $N$ joints. Subsequently, the feature tensor $\phi_i$ is fed into a stack of 3 $\times$ 3 residual blocks (adapted from the residual steps block in RSN \cite{cai2020learning}), producing the merged keypoint heatmaps $\Phi_i(p,c,n)$:
\begin{equation}
\begin{aligned}
\phi_i(p,c,n) \xrightarrow[\text{residual blocks}]{\text{stack of }3\times3} \Phi_i(p,c,n).
\end{aligned}
\end{equation}
This group convolution not only eliminates the disturbance of irrelevant joints, but also removes redundancy and shrinks the amount of model parameters required. It is also advantageous to directly summing keypoint heatmaps in Eq.~\ref{eqn:mergedheatmap} since the group CNN operation allow different weights
at the pixel level and benefits learning an end-to-end model.
Visual results of aggregated keypoint heatmaps following our PTM is illustrated in Fig.~\ref{fig:heatmap_fuse}.
\subsection{Pose Residual Fusion}
\label{res}
Parallel to spatial aggregation of keypoint heatmaps in PTM, our Pose Residual Fusion (PRF) branch aims to compute the pose residuals which will serve as additional favorable temporal cues. Given keypoint heatmaps $\mathbf{h}_i(p),\mathbf{h}_i(c),\mathbf{h}_i(n)$, we compute the pose residual features as follows:
\begin{equation}
\begin{aligned}
\psi_i(p,c)&=\mathbf{h}_i(c)-\mathbf{h}_i(p)\\
\psi_i(c,n)&=\mathbf{h}_i(n)-\mathbf{h}_i(c)\\
\psi_i&=\psi_i(p,c)\oplus\psi_i(c,n)\\
&\oplus\frac{n-c}{n-p}\psi_i(p,c)\oplus\frac{c-p}{n-p}\psi_i(c,n).
\end{aligned}
\end{equation}
$\psi_i$ concatenates the original pose residuals $\psi_i(p,c)$, $\psi_i(c,n)$, and their weighted versions, where the weights are obtained according to the temporal distances. Similar to PTM, $\psi_i$ is then processed via a stack of $3\times3$ residual blocks to give the final pose residual feature $\Psi_i(p,c,n)$:
\begin{equation}
\begin{aligned}
\psi_i(p,c,n) \xrightarrow[\text{residual blocks}]{\text{stack of }3\times3} \Psi_i(p,c,n).
\end{aligned}
\end{equation}
\subsection{Pose Correction Network}
Given the merged keypoint heatmaps $\Phi_i(p,c,n)$ and pose residual feature tensor $\Psi_i(p,c,n)$, our Pose Correction Network is employed to refine the initial keypoint heatmap estimation $\mathbf{h}_i(c)$, yielding adjusted final keypoint heatmaps. Primarily, the pose residual feature tensor $\Psi_i(p,c,n)$ is used as the input for five parallel $3\times3$ convolution layers with different dilation rates $d\in\{3,6,9,12,15\}$. This computation gives five groups of offsets for the five kernels of the subsequent deformable convolution layer. Formally, the offsets are computed as:
\begin{equation}
\begin{aligned}
\Phi_i(p,c,n)\oplus\Psi_i(p,c,n) \xrightarrow[\text{residual blocks}]{\text{stack of }3\times3}\xrightarrow[\text{convolution layers}]{\text{dilation rate } d} O_{i,d}.
\end{aligned}
\end{equation}
Different dilation rates correspond to varying the size of the effective receptive field whereby enlarging the dilation rate \cite{yu2015multi} increases the scope of the receptive field. A smaller dilation rate focuses on local appearance, which is more sensitive for capturing subtle motion contexts.
Conversely, using a large dilation rate allows us to encode global representations and capture relevant information of a larger spatial scope. In addition to the offset computation, we feed the merged keypoint heatmaps to similar convolution layers and obtain five sets of masks $M_d$ as:
\begin{equation}
\begin{aligned}
\Phi_i(p,c,n)\oplus\Psi_i(p,c,n) \xrightarrow[\text{residual blocks}]{\text{stack of }3\times3}\xrightarrow[\text{convolution layers}]{\text{dilation rate } d} M_{i,d}.
\end{aligned}
\end{equation}
The parameters of two dilation convolution structures for offset $O$ and mask $M$ computation are independent. A mask $M_d$ can be considered as the weight matrix for a convolution kernel.
We implement the pose correction module through the deformable convolution $V2$ network (DCN v2 \cite{zhu2019deformable}) at various dilation rates $d$. DCN v2 takes the following inputs:
1) the merged keypoint heatmaps $\Phi_i(p,c,n)$, 2) the kernel offsets $O_{i,d}$, and 3) the masks $M_{i,d}$, and outputs a pose heatmap for person $i$ at dilation rate $d$:
\begin{equation}
\begin{aligned}
\left(\Phi_i(p,c,n),O_{i,d},M_{i,d}\right)\xrightarrow[\text{DCN v2}]{\text{dilation rate } d} \mathbf{H}_{i,d}(c).
\end{aligned}
\end{equation}
The five outputs for five dilation rates are summarized and normalized to yield the final pose prediction for person $i$:
\begin{equation}
\begin{aligned}
\sum_{d\in\{3,6,9,12,15\}}\mathbf{H}_{i,d}(c)\xrightarrow[]{\text{normalization}} \mathbf{H}_{i}(c).
\end{aligned}
\end{equation}
Ultimately, the above procedure is performed for each individual $i$. By effectively utilising the additional cues from $F_p$ and ${F}_n$ in our DCPose framework, the final pose heatmaps are enhanced and improved.
\subsection{Implementation Details}
\textbf{Backbone Model}\quad
Our network is highly adaptable and we can seamlessly integrate any image based pose estimation architecture as our backbone. We employ the state-of-the-art Deep High Resolution Network (HRNet-W48 \cite{sun2019deep}) as our backbone joints detector, since its superior performance for single image pose estimation will be beneficial for our approach.
\renewcommand\arraystretch{1.2}
\begin{table}
\resizebox{0.48\textwidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c|c|c}
\hline
Method &Head &Shoulder &Elbow &Wrist &Hip &Knee &Ankle &{\bf Mean}\cr
\hline
PoseFlow\cite{xiu2018pose} &$66.7$ & $73.3$ &$68.3$ &$61.1$ &$67.5$ &$67.0$ &$61.3$ &{$\bf 66.5$}\cr
JointFlow\cite{doering2018joint} & - & - &- &- &- &- &- &{\bf $\bf 69.3$}\cr
FastPose\cite{zhang2019fastpose} &$80.0$ &$80.3$ &$69.5$ &$59.1$ &$71.4$ &$67.5$ &$59.4$ &{$\bf 70.3$}\cr
SimpleBaseline\cite{xiao2018simple} &$81.7$ &$83.4$ &$80.0$ &$72.4$ &$75.3$ &$74.8$ &$67.1$ &{$\bf 76.7$}\cr
STEmbedding\cite{jin2019multi} &$83.8$ &$81.6$ &$77.1$ &$70.0$ &$77.4$ &$74.5$ &$70.8$ &{$\bf 77.0$}\cr
HRNet\cite{sun2019deep} &$82.1$ &$83.6$ &$80.4$ &$73.3$ &$75.5$ &$75.3$ &$68.5$ &{$\bf 77.3$}\cr
MDPN\cite{guo2018multi} &$85.2$ &$88.5$ &$83.9$ &$77.5$ & $79.0$&$77.0$ &$71.4$ &{$\bf 80.7$}\cr
PoseWarper\cite{bertasius2019learning} &$81.4$ &$88.3$ &$83.9$ &$ 78.0$ &$82.4$ &$80.5$ &$73.6$ &{ $\bf 81.2$}\cr
\hline
\bf DCPose &$\bf 88.0$ &$\bf 88.7$ &$\bf 84.1$ &$\bf 78.4$&$\bf 83.0$ &$\bf 81.4$&$\bf 74.2$ &$\bf 82.8$\cr
\hline
\end{tabular}}
\vspace{0.2em}\caption{\footnotesize{Quantitative Results (\textbf{AP}) on PoseTrack2017 \bf validation set}.} \label{tab1}
\end{table}
\renewcommand\arraystretch{1.2}
\begin{table}
\resizebox{0.48\textwidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c|c|c}
\hline
Method &Head&Shoulder &Elbow &Wrist &Hip &Knee &Ankle &{\bf Total}\cr
\hline
PoseFlow\cite{xiu2018pose} &$64.9$ &$67.5$&$65.0$ &$59.0$ &$62.5$ &$62.8$ &$57.9$ &{\bf $\bf 63.0$}\cr
JointFlow\cite{doering2018joint} &- &- &- &$53.1$ &- &- &$50.4$ &{$\bf 63.4$}\cr
KeyTrack\cite{snower202015} &- &- &- &$71.9$ &- &- &$65.0$ &{$\bf 74.0$}\cr
DetTrack\cite{wang2020combining} &- &- &- &$69.8$ &- &- &$65.9$ &{$\bf 74.1$}\cr
SimpleBaseline\cite{xiao2018simple} &$80.1$ &$80.2$ &$76.9$ &$71.5$ &$72.5$ &$72.4$ &$65.7$ &{$\bf 74.6$}\cr
HRNet\cite{sun2019deep} &$80.1$ &$80.2$ &$76.9$ &$72.0$ &$73.4$ &$72.5$ &$67.0$ &{$\bf 74.9$}\cr
PoseWarper\cite{bertasius2019learning} &$79.5$ &$84.3$ &$80.1$ &$75.8$ &$77.6$ &$76.8$ &$70.8$ &$\bf 77.9$\cr
\hline
\bf DCPose&$\bf 84.3$&$\bf 84.9$&$\bf 80.5$&$\bf 76.1$&$\bf 77.9$&$\bf 77.1$&$\bf 71.2$&$\bf 79.2$\cr
\hline
\end{tabular}}
\vspace{0.2em}\caption{\footnotesize{Performance comparisons on the PoseTrack2017 \textbf{test} set.}} \label{tab2}
\end{table}
\renewcommand\arraystretch{1.2}
\begin{table}
\resizebox{0.48\textwidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c|c|c}
\hline
Method &Head &Shoulder &Elbow &Wrist &Hip &Knee &Ankle &{\bf Mean}\cr
\hline
AlphaPose\cite{Fang_2017_ICCV} &$63.9$ &$78.7$&$77.4$ &$71.0$ &$73.7$ &$73.0$ &69.7 &{\bf $\bf 71.9$}\cr
MDPN\cite{guo2018multi} &$75.4$ &$81.2$ &$79.0$ &$74.1$ &$72.4$ &$73.0$ &$69.9$ &{$\bf 75.0$}\cr
PoseWarper\cite{bertasius2019learning} &$79.9$&$86.3$&$82.4$&$77.5$&$79.8$&$78.8$&$73.2$&{ $\bf 79.7$}\cr
\hline
\bf DCPose&$\bf 84.0$&$\bf 86.6$&$\bf 82.7$&$\bf 78.0$&$\bf 80.4$&$\bf 79.3$&$\bf 73.8$&$\bf 80.9$\cr
\hline
\end{tabular}}
\vspace{0.2em}\caption{\footnotesize{Quantitative Results(\textbf{AP}) on PoseTrack2018 \bf validation set}.} \label{tab3}
\end{table}
\renewcommand\arraystretch{1.2}
\begin{table}
\resizebox{0.48\textwidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c|c|c}
\hline
Method &Head &Shoulder &Elbow &Wrist &Hip &Knee &Ankle &{\bf Total}\cr
\hline
AlphaPose++\cite{Fang_2017_ICCV,guo2018multi} &- &- &- &$66.2$ &- &- &$65.0$ &{\bf $\bf 67.6$}\cr
DetTrack \cite{wang2020combining} &- &- &- &$69.8$ &- &- &$67.1$ &{\bf $\bf 73.5$}\cr
MDPN\cite{guo2018multi} &- &- &- &74.5 &- &- &69.0 &{$\bf 76.4$}\cr
PoseWarper\cite{bertasius2019learning} &$78.9$&$\bf 84.4$&$\bf 80.9$&$76.8$&$75.6$&$77.5$&$71.8$&{ $\bf 78.0$}\cr
\hline
\bf DCPose&$\bf 82.8$&$ 84.0$&$ 80.8$&$\bf 77.2$&$\bf 76.1$&$\bf 77.6$&$\bf 72.3$&$\bf 79.0$\cr
\hline
\end{tabular}}
\vspace{0.2em}\caption{\footnotesize{Performance comparisons on the PoseTrack2018 \textbf{test} set.}} \label{tab4}
\end{table}
\textbf{Training}\quad
Our Deep Dual Consecutive Network is implemented in PyTorch. During training, we use the ground truth person bounding boxes to generate the $\mathbf{Clip}_i(p,c,n)$ for person $i$ as the input sequence to our model. For boundary cases, we apply same padding. In other words, if there are no frames to extend forward and backward from ${F}_c$,
${F}_p$ or ${F}_n$ will be replaced by ${F}_c$. We utilize the HRNet-W48 pretrained on the PoseTrack dataset as our backbone, and freeze the backbone parameters throughout training, only backpropagating through the subsequent components in DCPose.
\textbf{Loss function}\quad
We employ the standard pose estimation loss function as our cost function. Training aims to minimize the total Euclidean or $L2$ distance between prediction and ground truth heatmaps for all joints. The cost function is defined as:
\begin{equation}
\begin{aligned}
L = \frac{1}{N} * \sum_{j=1}^N v_j * ||G(j) - P(j)||^2
\end{aligned}
\end{equation}
Where $G(j)$, $P(j)$ and $v_j$ respectively denote the ground truth heatmap, prediction heatmap and visibility for joint $j$. During training, the total number of joints is set to $N=15$. The ground truth heatmaps are generated via a 2D Gaussian centered at the joint location.
\section{Experiments}
In this section, we present our experimental results on two large-scale benchmark datasets: Posetrack2017 and PoseTrack2018 Multi-frame Person Pose Estimation Challenge datasets.
\subsection{Experimental Settings}
\textbf{Datasets}\quad
PoseTrack is a large-scale public dataset for human pose estimation and articulated tracking in video and includes challenging situations with complicated movement of highly occluded people in crowded environments. The PoseTrack2017 dataset \cite{Iqbal_2017_CVPR} contains 514 video clips and 16,219 pose annotations, with 250 clips for training, 50 clips for validation, and 214 clips for testing. The PoseTrack2018 dataset greatly increased the number of video clips to a total of 1,138 clips and 153,615 pose annotations. The training, validation, and testing splits consist of 593, 170, and 375 clips respectively. The 30 frames in the center of the training video clips are densely annotated. For the validation clips, annotations are provided every four frames. Both PoseTrack2017 and PoseTrack2018 identify 15 joints, with an additional annotation label for joint visibility. We evaluate our model only for visible joints with the average precision (\textbf{AP}) metric \cite{Iqbal_2017_CVPR, Andriluka_2018_CVPR}.
\renewcommand\arraystretch{1.1}
\begin{table*}
\centering
\resizebox{1\textwidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c|c|c}
\hline
Method & Head & Shoulder & Elbow & Wrist & Hip & Knee & Ankle & {\bf Mean}\cr
\hline
\textbf{DCPose, complete, $T=1$}
& $\bf 88.0$ & $\bf 88.7$ & $\bf 84.1$ & $\bf 78.4$ & $\bf 83.0$ & $\bf 81.4$ & $\bf 74.2$ & $\bf 82.8$ \cr
\hline
r/m PTM, $\Phi(p,c,n)\leftarrow\mathbf{h}(p)+\mathbf{h}(c)+\mathbf{h}(n)$
& 87.3 & 88.4 & 83.6 & 77.6 & 82.8 & 78.4 & 73.7 & 82.0 \cr
r/m PTM, $\Phi(p,c,n)\leftarrow\mathbf{h}(c)$
& 87.5 & 88.5 & 83.7 & 78.1 & 82.9 & 80.8 & 73.4 & 82.2 \cr
\hline
r/m PRF
& 87.4 & 88.3 & 83.6 & 77.6 & 82.7 & 80.5 & 73.4 & 82.1 \cr
\hline
1 dilation convolution, $d=3$
& 87.3 & 88.5 & 83.8 & 72.6 & 82.9 & 78.4 & 73.7 & 81.7 \cr
2 dilation convolutions, $d\in\{3,6\}$
& 87.5 & 88.2 & 83.6 & 74.7 & 82.8 & 79.8 & 73.6 & 81.9 \cr
3 dilation convolutions, $d\in\{3,6,9\}$
& 87.9 & 88.6 & 83.9 & 78.3 & 82.9 & 81.4 & 74.0 & 82.7 \cr
4 dilation convolutions , $d\in\{3,6,9,12\}$
& 87.9 & 88.6 & 84.0 & 76.9 & 83.0 & 80.9 & 74.0 & 82.6 \cr
\hline
r/m previous frame
& 87.8 & 88.4 & 83.8 & 77.8 & 82.6 & 80.6 & 73.7 & 82.2 \cr
r/m next frame
& 88.0 & 88.6 & 83.7 & 77.7 & 82.8 & 80.7 & 73.2 & 82.3 \cr
\hline
$T=2$
& 87.7 & 88.4 & 83.8 & 78.0 & 83.0 & 80.8 & 73.2 & 82.6 \cr
$T=3$
& 87.6 & 88.4 & 83.6 & 77.7 & 82.6 & 80.7 & 71.7 & 82.1 \cr
$T=4$
& 81.2 & 87.4 & 83.3 & 76.7 & 81.9 & 80.2 & 72.1 & 81.7 \cr
\hline
\end{tabular}}
\vspace{0.2em}\caption{\normalsize{Ablation study of different components in our DCPose performed on PoseTrack2017 validation set. ``r/m X" refers to removing X module in our network. The complete DCPose consistently achieves the best results which are highlighted.}} \label{tab5}
\end{table*}
\textbf{Parameter Settings}\quad
During training, we incorporate data augmentation including random rotations, scaling, truncating, and horizontal flipping to increase variation. Input image size is fixed to $384\times288$. The default interval between ${F}_c$ and ${F}_p$ or ${F}_n$ is set to 1. Backbone parameters are fixed to the pretrained HRNet-W48 model weights.
All subsequent weight parameters are initialized from a Gaussian distribution with $\mu=0$ and $\sigma=0.001$, while bias parameters are initialized to $0$. We employ the Adam optimizer with a base learning rate of $0.0001$ that decays by $10\%$ every 4 epochs. We train our model for a batch size of 32 for 20 epochs with 2 Nvidia GeForce Titan X GPUs.
\subsection{Comparison with State-of-the-art Approaches}
\textbf{Results on the PoseTrack2017 Dataset}\quad
We evaluate our approach on PoseTrack2017 validation set and full test set using the widely adopted average precision (\textbf{AP}) metric \cite{xiao2018simple,xiu2018pose,doering2018joint,zhang2019fastpose}.
Table \ref{tab1} presents the quantitative results of different approaches in terms of AP on PoseTrack2017 validation set.
We benchmark our \textbf{DCPose} model against eight existing methods \cite{xiu2018pose}, \cite{doering2018joint}, \cite{zhang2019fastpose},
\cite{xiao2018simple}, \cite{jin2019multi}, \cite{sun2019deep}, \cite{guo2018multi}, \cite{bertasius2019learning}. In Table \ref{tab1}, the APs of key joints, such as Head, Shoulder, Knee, and Elbow, are reported, as well as the mAP (mean AP) for all joints.
Results on the test set are provided in Table \ref{tab2}. These results are obtained by uploading our prediction results to the PoseTrack evaluation server:\url{https://posetrack.net/leaderboard.php} because the annotations for test set are not public. Our DCPose network achieves state-of-the-art results for multi-frame person pose estimation challenge for both the validation and test sets. DCPose consistently outperforms existing methods and achieves a \textbf{mAP} of 79.2. The performance boost for relatively difficult joints is also encouraging: we obtain an \textbf{mAP} of 76.1 for the wrist and an \textbf{mAP} of 71.2 for the ankle. Some sample results are displayed in Fig.~\ref{fig:results}, which are indicative of the effectiveness of our method in complex scenes. More visualized results can be found in \url{https://github.com/Pose-Group/DCPose}.
{\textbf{Results on the PoseTrack2018 Dataset}}\quad
We also evaluate our model on the PoseTrack2018 dataset. The validation and test set \text{AP} results are tabulated in Table \ref{tab3} and Table \ref{tab4}, respectively. As shown in the tables, our approach once again delivers the state-of-the-art results. We achieve an \textbf{mAP} of 79.0 on the test set, and obtain an \textbf{mAP} of 77.2 for the difficult \emph{wrist} joint and 72.3 for the \emph{ankle}.
\begin{figure*}
\begin{center}
\includegraphics[width=1\linewidth]{Figures/Fig3-Results.jpg}
\end{center}
\caption{Visual results of our model on the PoseTrack2017 and PoseTrack2018 datasets comprising complex scenes: high-speed movement, occlusion, and multiple persons.}
\label{fig:results}
\end{figure*}
\subsection{Ablation Experiments}
Extensive ablation experiments are performed on the PoseTrack2017 dataset to study the effectiveness of various components in our DCPose framework. Through ablating the various modular networks including Pose Temporal Merger (PTM), Pose Residual Fusion (PRF) and Pose Correction Network, we evaluate their contributions towards the overall performance. We also investigate the efficacy of including information from both temporal directions as well as the impact of modifying the temporal distance, \emph{i.e.,} frame intervals between ${F}_c$ and ${F}_p$, ${F}_n$. The results are reported in Table~\ref{tab5}.
\textbf{Pose Temporal Merger}\quad
For this ablation setting, we remove the PTM and instead obtain the merged pose heatmaps $\Phi(p,c,n)$ as follows:
i) $\mathbf{h}(p)+\mathbf{h}(c)+\mathbf{h}(n)\xrightarrow[\text{convolution layer}]{3\times3}\Phi(p,c,n)$; ii) $\mathbf{h}(c)\xrightarrow[]{}\Phi(p,c,n)$. The \textbf{mAP} falls from 82.8 to 82.0 for (i) and 82.2 for (ii). This significant performance degradation upon removal of the PTM module can be attributed to the failure of obtaining an effective search range for the joints, which lead to decreased accuracy in locating the joint in the subsequent pose correction stage.
\textbf{Pose Residual Fusion}\quad
We investigate removing the PRF and compute the pose residual maps $\Psi(p,c,n)$ with the following scheme:
$\mathbf{h}(c)-\mathbf{h}(p)\oplus\mathbf{h}(n)-\mathbf{h}(c)\xrightarrow[\text{convolution layer}]{3\times3}\Psi(p,c,n)$. This results in the \textbf{mAP} dropping 0.7 to 82.1. This significant reduction in performance highlights the important role of PRF in providing accurate pose residual cues for computing offsets and guiding the keypoint localization.
\begin{figure}
\begin{center}
\includegraphics[width=1\linewidth]{Figures/Fig4-Contrast.jpg}
\end{center}
\caption{Visualization of the pose predictions of our DCPose model($\bf a$), SimpleBaseline($\bf b$), HRNet($\bf c$), and PoseWarper($\bf d$) on the challenging cases from the $ PoseTrack 2017$ and $ PoseTrack 2018$ datasets.
Each column from left to right represents the Rapid-Motion, Nearby-Person, Occlusions, and Video-Defocus scene, respectively.
Inaccurate predictions are highlighted with the red solid circles.}
\label{fig:vis}
\vspace{-0.5em}
\end{figure}
\textbf{Pose Correction Network}\quad
We study the effects of adopting different sets of dilation rates for the convolutions in PCN. This corresponds to different effective receptive fields. We experiment with four different dilation settings: $d=3$, $d\in\{3, 6\}$, $d\in\{3, 6, 9\}$ and $d\in\{3, 6, 9, 12\}$ whereas the complete DCPose framework setting has $d\in\{3, 6, 9, 12, 15\}$. From the results in Table \ref{tab5}, we observe the gradual improvement of the \textbf{mAP} with increasing levels of dilation rates, from $81.7\to81.9\to82.7\to82.6\to82.8$. This is in line with our intuitions that increasing the depth and scope of the effective receptive fields, \emph{i.e.}, by varying the range of dilation rates, the Pose Correction Network is able to model local and global contexts more efficiently, leading to more accurate estimations of the joint locations.
\textbf{Bidirectional Temporal Input}\quad
In addition, we examine the effects of removing a single temporal direction, \emph{i.e.}, removal of either ${F}_p$ or ${F}_n$ from the input to PTM and PRF. In removing the previous (respectively next) frame $F_p$ (respectively $F_n$), the \textbf{mAP} drops 0.6 (respectively 0.5). This highlights the importance of leveraging dual temporal directions as each direction allows access to useful information that are beneficial in improving pose estimation for videos.
\textbf{Time Interval $T$}\quad
The time interval $T$ described in Section \ref{section:method} is a hyper-parameter whose default value is set to $1$. In other words, DCPose looks at 3 consecutive frames as the shorter time/frame interval ensures the validity of our assumptions of temporal consistency. We experiment with enlarging the interval between frames, where $T$ is set to 2, 3, 4. Indices $p$ and $n$ are randomly selected with $p\in [c-T,c)$ and $n\in (c,c+T]$. The results reflect a performance drop with an increase in $T$, whereby the \textbf{mAP} decreases from 82.8 for $T=1$ to 82.6, 82.1, 81.7 as $T$ goes to 2, 3, 4 respectively. This is in accordance with our expectations, since frame-wise difference is small for shorter time intervals so that ${F}_p$ and ${F}_n$ can provide more abundant and effective temporal and spatial cues for our PTM and PRF. Conversely, a large time interval would invalidate our assumption of time consistency as the frame differences would be too great and the additional cues would lack efficacy. This ablation experiment explains our choice of modeling consecutive frames. The continuous frames of a video clip facilitate a precise aggregation of spatiotemporal context cues and improve the robustness of our model.
\subsection{Comparison of Visual Results}
In order to evaluate the capability of our model to adapt to sophisticated scenarios, we illustrate in Fig.~\ref{fig:vis} the side-by-side comparisons of our DCPose network against state-of-the-art approaches. Each column depicts different scenario complications including \emph{rapid motion, nearby-person, occlusions} and \emph{video defocus}, whereas each row displays the joint detection results from different methods. We compare our a) DCPose against 3 state-of-the-art methods, namely b) SimpleBaseline \cite{xiao2018simple}, c) HRNet-W48 \cite{sun2019deep} and d) PoseWarper \cite{bertasius2019learning}. It is observed that our method yields more robust detection for such challenging cases. SimpleBaseline and HRNet-W48 are trained on static images and fail to capture temporal dependency among video frames, resulting in suboptimal joint detection. On the other hand, PoseWarper leverages the temporal information between video frames to warp the initial pose heatmaps but only employing one adjacent video frame as the auxiliary frame. Our DCPose network makes full use of the temporal context by processing consecutive frames in dual directions. Through our principled design of PTM and PRF modules for better encoding of these information as well as PCN for refining pose detection, our method achieves new state-of-the-art both quantitatively and qualitatively.
\section{Conclusion}
In this paper, we propose a dual consecutive network for multi-frame person pose estimation which significantly outperforms existing state-of-the-art methods on the benchmark datasets. We design a Pose Temporal Merger and a Pose Residual Fusion module that allows abundant auxiliary information to be drawn from the adjacent frames, providing a localized and pose residual corrected search range for location keypoints. Our Pose Correction Network employs multiple effective receptive fields to refine pose estimation in this search range, achieving notable improvements and is able to handle complex scenes.
\section{Acknowledgements}
This work was supported by the Natural Science Foundation of Zhejiang Province, China No. LQ19F020001, the National Natural Science Foundation of China No. 61902348, the Humanities and Social Science Fund of the Ministry of Education of China under Grant No.17YJCZH076, and the Zhejiang Science and Technology Project under Grant No.LGF18F020001.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
Since the first introduction of deep convolutional neural networks (CNN) and the rise of GANs, more and more powerful image processing applications have been proposed. A still popular topic in the literature discusses how these image processing techniques can be efficiently extended to videos.
A natural way to transfer image processing algorithms to videos is to perform such transformation frame by frame. However, the latter scheme necessarily causes unpleasant visual artifacts (\textit{e.g.} flickering) originating from temporal inconsistency. For instance, successive frames are not necessarily transformed coherently since image trained algorithms fail to capture the temporal dimension of videos. The literature offers several types of approaches to circumvent such problems: video-to-video translation algorithms and post-processing techniques of per-frame transformed videos. The former consists in conditional video generation algorithms aiming at generating temporally coherent outputs, while the latter post-processes videos transformed using image-trained algorithms to make them temporally coherent. On the one hand, video-to-video translation algorithms exhibit interesting performances even for relatively high-resolution videos, however, they require video datasets which are very scarce especially for supervised algorithms. On the other hand, post-processing techniques offer satisfactory results and do not necessitate any retraining of image processing algorithms, but often fail to reduce localized flickering without generating new visual artifacts and tend to modify the overall color of videos.
\begin{figure}[t!]
\centering
\centering
\animategraphics[autoplay,loop,scale=0.5, poster=40]{30}{files/videos/cows_new_folder/frame-}{0}{79}
\caption{Our model successfully removes flickering from the raw processed video without altering the overall style nor the brightness. \cite{lai2018learning} on the contrary seems to remove flickering at the cost of style alteration as the video becomes darker. This video is best seen using Adobe Acrobat Reader.}
\label{fig:cows}
\end{figure}
We propose a blind post-processing algorithm in the form of a recurrent neural network. The main contributions involved in the model are:
\begin{itemize}
\item[•] Following up on recent work \cite{park2019arbitrary} in the image style transfer literature, we introduce a novel style preserving perceptual loss which reduces deviation from the overall style of the original processed video.
\item[•] We propose to enforce long term consistency by adapting the Ping Pong training procedure \cite{Chu_2020} to this task and its corresponding Ping Pong Loss.
\item[•] We empirically highlight some weaknesses of the perceptual distance as a measure of style preservation.
\item[•] Our model offers state-of-the-art results in flicker removal for any transformation and does not deteriorate the visual quality of videos when enforcing time consistency, nor generate new temporal artifacts.
\end{itemize}
\section{Background and related work}
When independently applying image-trained algorithms (\textit{e.g.} image-trained style transfer algorithms) to successive frames in a video, several temporal artifacts tend to appear. In that context, the recent literature has focused on two approaches to tackle the problem of video processing and temporal inconsistency.
The first approach, namely video-to-video translation, consists in adapting image processing algorithms to take into account the temporal dimension of the input. Most approaches involve computing optical flows during training and consider networks taking several inputs including the current frames (raw and stylized), the previous output frame, and the previous stylized frame warped using the estimated optical flow \cite{Ruder_2016}. Optical flow estimation is computationally costly, thus, later approaches also considered similar methods but circumvent the problem of computing optical flows during the inference stage such as \cite{gao2020}. Recent approaches offer satisfying results but tend to require complex training procedures such as a two-stage progressive training strategy and need to be retrained for each new application considered.
The second method, task-independent post-processing, aims at considering time consistency in a more agnostic manner not focusing on any type of image processing. Blind post-processing consists in (i) first transforming a video using any algorithm and (ii) using an algorithm, independent of the transformation previously performed, to make the video temporally consistent. This approach requires to consider a twofold problem: on the one hand, videos must be made temporally coherent, on the other hand, the style of videos must not be altered.
Recent papers have obtained satisfying results using general approaches totally blind to the filter applied to videos. For instance, \cite{bonneel:hal-01264081} and \cite{yao2017} display very satisfactory results, however, the computational complexity of their proposed algorithm makes it difficult to process long video sequences.
More recently, in \cite{lai2018learning} the authors propose a neural network to enforce time consistency in videos. The model takes as an input the raw unprocessed video frames, the processed video frames, and the previous output of the recurrent network. The model also makes use of optical flow estimation for temporal consistency loss computation and of the VGG19 \cite{simonyan2014deep} network for perceptual loss computation. Their model offers interesting results but fails to coherently learn long-term temporal consistency and generates new temporal artifacts when the motion of an object changes direction. Moreover, their model tends to alter the overall style of the video once post-processed, thus reducing the quality of the video transformation. Recent work in the style transfer literature has offered possible solutions to foster style preservation when learning temporal coherence.
Neural style transfer consists in transferring the style of an image (\textit{e.g.} a painting) to another one while preserving the content of the latter image. Style transfer in the form of deep neural networks was first introduced in \cite{Gatys}. Their seminal model makes use of the VGG19 \cite{simonyan2014deep} network to perform style transfer through both a content and style loss. Their approach opened a wide range of applications making use of deep features (\textit{e.g.} VGG19 \cite{simonyan2014deep}) as measures of similarity between images. Recent approaches such as \cite{park2019arbitrary}
have offered a better balance between the structure and the style of the output by using moments of different orders (mean and standard deviation) of the pre-trained features.
\section{Proposed method}
\label{nn}
\subsection{Task}
We argue that post-processing techniques can manage to efficiently remove any temporal artifacts on videos originating from per-frame processing. This approach is all the more appealing since it needs not any retraining for each task; this is a key aspect since it can be challenging to find a satisfying equilibrium in the case of video-to-video translation algorithms. We thus propose a blind post-processing model in the form of a recurrent neural network which improves current state-of-the-art results. Our approach does not require optical flow estimation during inference but necessitates to dispose of the original non-transformed video.
\subsection{Architecture}
The deep network is comprised of a classical encoder-decoder architecture. The encoder part is composed of two downsampling strided convolutional layers, each followed by Instance Normalization. The encoder is then followed by 5 residual blocks and a ConvLSTM module \cite{NIPS2015_5955}.
The decoder placed after the ConvLSTM module is composed of two transposed convolutional layers also followed by Instance Normalization. We also include skip-connections via concatenation from the encoder to the decoder.
ConvLSTM modules are well-suited to our problem since they can compress all the previous information into a hidden state which can be used to estimate the current state. This hidden state can capture the spatial information of the whole input sequence and allows the model to learn coherently time consistency. In tandem with temporal losses, ConvLSTM modules were shown to offer very satisfactory results in fostering temporal consistency on video style transfer networks.
Let us denote the original unprocessed frames $\{I_t\}_{t=1,\hdots,T }$, the per-frame processed videos $\{P_t\}_{t=1,\hdots,T }$ and $\{O_t\}_{t=1,\hdots,T }$ the corresponding outputs of the model. Since we consider residual blocks in our network, $O_t$ can be seen as a residual transformation of $P_t$ to make it temporally consistent. To output $O_t$, the network takes as an input $I_t, I_{t-1}, P_t$ and $O_{t-1}$. Formally
\[
O_t = P_t + \mathcal{F}(I_t, I_{t-1}, P_t, O_{t-1})
\]
where $\mathcal{F}(.)$ denotes our network. The recurrent form of the network thus originates from the fact that our network makes use of the previous output to generate the current output. Let us note here that for of each sequence $P_1 = O_1$.
\subsection{Losses}
\label{losse}
The aim of the proposed model is to reduce the temporal inconsistency while maintaining the perceptual similarity in both content and style between the output frame of the model and the processed frame. Therefore, two types of loss must be considered: perceptual losses and temporal losses.
\subsubsection{Perceptual losses}
The pre-trained VGG19 classification network is vastly used in a large number of application as it has been shown to be a good measure of human perception. To preserve the content of the video, as in \cite{lai2018learning} we set the content perceptual loss to be
\begin{equation}
\label{eq:perceptual_loss}
\mathcal{L}_p = \sum_{t=2}^T \sum_{i=1}^N \sum_{l} \bigg{|}\bigg{|}\phi_l\big{(}O_t^{(i)}\big{)} - \phi_l\big{(}P_t^{(i)}\big{)} \bigg{|}\bigg{|}_1
\end{equation}
where $O_t^{(i)} \in \mathbb{R}^3$ represent the RGB vector of pixel $i$ in the output frame $O_t$ composed of $N$ pixels, while $\phi_l(.)$ represents the feature activation of the l-th block of the VGG19 network $\phi$ (we consider a sum of all feature activation until the 4th block, \textit{i.e.} \texttt{relu1-2}, \texttt{relu2-2}, \texttt{relu3-3} and \texttt{relu4-3}).
To ensure that our model does not alter the style of the video, we also introduce a novel loss which makes use of the VGG network discussed above. For instance, we consider
\begin{equation}
\label{eq:perceptual_loss}
\begin{array}{rcl}
\mathcal{L}_{sp} &=& \displaystyle{\sum_{t=2}^T \sum_{l}} \{ \big{|}\big{|}\mu\big{(}\phi_l(O_t)\big{)} - \mu\big{(}\phi_l(P_t)\big{)} \big{|}\big{|}_2^2 \\
& & + \big{|}\big{|}\sigma\big{(}\phi_l(O_t)\big{)} - \sigma\big{(}\phi_l(P_t)\big{)} \big{|}\big{|}_2^2 \}
\end{array}
\end{equation}
where $\mu(.)$ denotes the average in the channel dimension and $\sigma(.)$ the standard deviation.
Such approach is well suited to our case since pixel-wise comparisons between $P_t$ and $O_t$ can be too conservative, while considering only moments of the first and second order tend to capture well global style of a frame.
We also consider its temporal equivalent to avoid excessive deviations from the global style of the raw processed videos. For instance, since images are generated in an online manner, minor deviation from the style of the video tend to accumulate as frames are being processed.
\begin{equation}
\label{eq:perceptual_loss}
\begin{array}{lll}
\mathcal{L}_{sp}^{temp} &=& \displaystyle{\sum_{t=2}^T \sum_{l}} \{ \big{|}\big{|}\mu\big{(} \phi_l(O_{t})\big{)} - \mu\big{(}\phi_l( O_{t-1})\big{)} \big{|}\big{|}_2^2 \\
& & + \big{|}\big{|}\sigma\big{(}\phi_l( O_{t})\big{)} - \sigma\big{(} \phi_l(O_{t-1})\big{)} \big{|}\big{|}_2^2 \}
\end{array}
\end{equation}
In both losses, the feature activation considered are \texttt{relu1-2} and \ \texttt{relu2-2}. We thus consider the following style preserving perceptual loss
\[
\mathcal{L}_{SP} = \mathcal{L}_{sp} + \mathcal{L}_{sp}^{temp}
\]
\subsubsection{Temporal losses}
During training, we randomly fetch a sequence of $k$ successive frames in a video which we transform to construct a temporal cycle centered around the last frame of the selected sequence. For instance, for each randomly selected sequence in the training dataset $\{P_t, \hdots , P_{t+k-1}, P_{t+k} \} $ the following ping pong sequence is constructed $\{P_t, \hdots , P_{t+k-1}, P_{t+k}, P_{t+k-1}, \hdots, P_t \}$. In the recurrent framework detailed in the previous section, one obtains a similar sequence of output frames
$$
\{O_t, \hdots , O_{t+k-1}, O_{t+k}, O_{t+k-1}', \hdots, O_t' \}
$$
where frames with <<'>> denote frames constructed in a backward direction. In other words, $O_t'$ is estimated by our network taking as inputs $I_t, I_{t+1}, O_{t+1}'$ and $P_t$.
\noindent Considering such procedure, \cite{Chu_2020} propose the following Ping Pong loss
\begin{equation}
\label{eq:pingpong}
\mathcal{L}_{PP} = \sum_{t=1}^{k-1} \big{|}\big{|}O_t - O_t'\big{|}\big{|}_2
\end{equation}
Such loss which combines short term consistency for frames close to $n$, \textit{e.g.} $||O_{n-1} - O_{n-1}'||_2$ and long term consistency for factors far from $n$, \textit{e.g.} $||O_1 - O_1'||_2$, should reduce the accumulation of unwanted features throughout the video.
The short term temporal error is based on the warping error of successive output frames. Notice that the Ping Pong training procedure modifies classical short term temporal loss by involving backward warping error
\begin{equation}
\label{eq:st_loss}
\begin{array}{ll}
\mathcal{L}_{st} = & \displaystyle{\sum_{t=2}^T \sum_{i=1}^N} M_{t,t-1}^{(i)} \big{|}\big{|} O_t^{(i)} - warp(O_{t-1}, F_{t,t-1})^{(i)} \big{|}\big{|}_1 \\
& + M_{t,t+1}^{(i)} \big{|}\big{|} O_t'^{(i)} - warp(O_{t+1}', F_{t,t+1})^{(i)} \big{|}\big{|}_1
\end{array}
\end{equation}
where $warp(O_{t-1}, F_{t,t-1})$ is the $O_{t-1}$ frame warped to time $t$ using the backward optical flow $F_{t,t-1}$ estimated between $I_t$ and $I_{t-1}$. $M_{t,t-1} = \exp(-\alpha ||I_t - warp(I_{t-1}, F_{t,t-1})||_2^2)$ is the visibility mask calculated from the warping error between the input frame $I_t$ and the warped input frame $warp(I_{t-1}, F_{t,t-1})$ where $\alpha$ is set to be 50 \cite{lai2018learning}. Optical flows are estimated using FlowNet2 \cite{flownet2_ilg017}.
The previously described loss is used to enforce consistency between adjacent frames, however there is no guarantee regarding long-term consistency. The proposed long-term temporal loss consists in the warping error between the first frame of the sequence and all of the output frames of the forward sequence,
\begin{equation}
\label{eq:lt_loss}
\mathcal{L}_{lt} = \sum_{t=2}^T \sum_{i=1}^N M_{t,1}^{(i)} \big{|}\big{|} O_t^{(i)} - warp(O_{1}, F_{t,1})^{(i)} \big{|}\big{|}_1
\end{equation}
Following up on \cite{chen2020optical} we formulate the hypothesis on stable stylized/processed videos: non-occluded regions should be low-rank representations.
Consider during training (i) a sequence of $k$ consecutive frames, $\{I_t, \hdots , I_{t+k-1}, I_{t+k} \} $ (ii) a reference time (which we simply set to $k/2$) to which all frames are warped (iii) and occlusion masks at each time step.
Denoting $R_t$ the Hadamard product (element-wise product) between the occlusion mask and the warped frame, consider
$
\chi = [vec(R_0), \hdots, vec(R_k)]^\top \in \mathbb{R}^{k \times N}
$
where $N = H \times W$ the number of pixels in the image and $vec(.)$ simply denotes the flattening operator of a two-dimensional image.
Based on the formulated hypothesis, the rank of $\chi$ constructed using the raw input frames $I_t$, $\chi_I$ and using the output frames of the model $O_t$, $\chi_O$ should not be too different from each other. \cite{chen2020optical} propose the low rank loss using the convex relaxation of the rank defined as
\begin{equation}
\label{eq:lowrank}
\mathcal{L}_{rank} = (||\chi_I||_* - ||\chi_O||_*)^2
\end{equation}
where $||.||_*$ is the nuclear norm defined as the sum of all singular values of a matrix.
\subsubsection{Overall loss}
The overall loss is a weighted sum of all previously defined losses
\begin{equation}
\label{eq:refined_overall_loss}
\begin{array}{rcl}
\mathcal{L} & = &\lambda_p\mathcal{L}_{p} + \lambda_{SP}\mathcal{L}_{SP} + \lambda_{st}\mathcal{L}_{st} + \\
&& \lambda_{lt} \mathcal{L}_{lt} + \lambda_{rank} \mathcal{L}_{rank} + \lambda_{PP} \mathcal{L}_{PP}
\end{array}
\end{equation}
where $\lambda_p$, $\lambda_{SP}$, $\lambda_{st}$, $\lambda_{lt}$, $\lambda_{rank}$ and $\lambda_{PP}$ are hyper-parameters to be defined, that represent the weight given to each loss in the overall training of the network.
\section{Experimental results}
\subsection{Evaluation metrics}
\label{pd}
Two features must be evaluated to assess the quality of the proposed model : (i) the output frames $O_t$ must be as close as possible to unprocessed frames $I_t$ in terms of time consistency and (ii) the output frames $O_t$ must visually look like the input processed frames $P_t$. Two types of metrics are used in the literature to evaluate the quality of a model: a time consistency metric based on the warping error and a perceptual similarity metric.
The literature dealing with time consistency has offered several metrics to measure time consistency in a video. The most vastly used is the mean warping error between successive frames of a video. The latter is defined as the difference between a warped frame to time $t$ and the true frame at time $t$.
Formally, the warping error between two frames $V_t$ and $V_{t+1}$ is defined as
\[
\omega_{t,t+1}^V = \frac{1}{M}\sum_{i=1}^NM_t^{(i)} || V_t^{(i)} - warp(V_{t+1},F_{t,t+1})^{(i)}||_2^2
\]
where $M_t \in \{0,1\}^{N}$ is the non-occlusion mask indicating non-occluded regions, $N$ is the number of pixels in each frame and $M = \sum_{i=1}^N M_t^{(i)}$. To estimate those regions, we resort to the method given in \cite{Ruder_2016} which consists in performing a backward-forward consistency check of the optical flows based on the method introduced in \cite{Bro10e}. To obtain a metric for a whole video $V$, one averages the warping error between successive frames over the sequence.
Perceptual similarity is evaluated using the LPIPS metric proposed in \cite{zhang2018unreasonable} to evaluate the perceptual similarity between the raw processed and output videos.
In \cite{lai2018learning}, the authors resort to the LPIPS metric proposed in \cite{zhang2018unreasonable} and the SqueezeNet \cite{squeeze} (noted $\mathcal{S}$) to evaluate the perceptual similarity between the raw processed videos $P$ and output $O$ videos.
\[
D_{perceptual}(P,O) = \frac{1}{T-1}\sum_{t=2}^T \mathcal{S}(O_t,P_t)
\]
We discuss this metric more deeply in the next section and argue that this metric fails to correctly capture localized style deviations or brightness alteration.
\subsection{Datasets and training}
For comparability and convenience, we use the dataset made available by \cite{lai2018learning} which is constituted of videos from DAVIS-2017 dataset \cite{Perazzi2016} and videvo.net videos. The height of each video in the training set is scaled to 480 while preserving the aspect ratio. The training set contains in total 25,735 frames.
The applications considered in the dataset are the following: artistic style transfer, Colorization, Image enhancement, Intrinsic image decomposition, and Image-to-image translation.
Results presented hereafter are those obtained after training with values of lambdas set empirically : $
\lambda_p=10, \lambda_{SP} = 10, \ \lambda_{st}=100, \ \lambda_{lt}=100, \ \lambda_{PP} = 100, \ \lambda_{rank} = 0.00001
$.
We set the number of epoch to 100, each composed of 1000 batches of size 4. The sequences considered in every batch are composed of 5 successive frames each, and thus 9 once the Ping Pong sequence is constructed. We consider Adam optimizer for parameter optimization.
\subsection{Qualitative evaluation}
\begin{figure}[h!]
\centering
\begin{subfigure}{0.165\textwidth}
\centering
\includegraphics[scale=0.5]{files/images/frame-15_ECCV18_square.png}
\caption{Lai et al.}
\label{fig:eccv18_p13}
\end{subfigure}
\begin{subfigure}{0.165\textwidth}
\centering
\includegraphics[scale=0.5]{files/images/frame-15_mine_square.png}
\caption{Ours + SP Loss}
\label{fig:retrained_p13}
\end{subfigure}
\caption{(a) displays a frame of a video processed using make-up synthesis (lipstick) and post-processed using \cite{lai2018learning} which removes the lip flickering but at the cost of a red trail following the lips as the head moves. (b) shows the same frame post-processed using our proposed model which removes flickering without generating any trail. The rectangular shape seen around the lips in both (a) and (b) originates from the processing algorithm which we used.}
\label{fig:retrained_vs_eccv18_trail}
\end{figure}
We propose to compare our model to the current state-of-the-art post-processing technique of \cite{lai2018learning}. The latter model successfully removes flickering artifacts in most cases but suffers from two drawbacks : (i) it tends to generate new temporal artifacts when flickering is localized on videos or when the motion of an object changes direction and (ii) the overall color of the video is often modified after using their post-processing method. An example of (i) can be seen in Fig. \ref{fig:eccv18_p13} where make-up synthesis was applied which generates flickering on the lips. Once post-processed using the model of \cite{lai2018learning} a trail following the lips appears. We believe this type of artifacts to originate from the form of the long-term loss considered in their model which forces the model to be coherent with the first frames of the sequence.
Our model visually outperforms the current state-of-the-art model in terms of style conservation and flicker removal. For instance Fig. \ref{fig:retrained_p13} shows that our model successfully removes flickering and does not generate trails as \cite{lai2018learning}. Similarly Fig. \ref{fig:cows} shows how our model manages to better remove flickering while not suffering from perceptual deterioration of the overall style of the video. For instance, in Fig \ref{fig:cows} one can see the videos post-processed using \cite{lai2018learning} becoming darker frame after frame, while our model coherently keeps the overall look of the video while not altering the brightness nor the style of the video. More examples showing the performance of our model are available in the supplementary material.
\subsection{Quantitative comparisons}
To evaluate the temporal coherence of videos we resort to the warping error metric previously discussed. Table \ref{tab:time_met} shows that our model outperforms the model of \cite{lai2018learning} for both DAVIS and videvo.net datasets for a large majority of applications when the style preserving loss is considered and for all test applications considered when not including the style preserving loss.
Let us note that we obtain better results than \cite{bonneel:hal-01264081} on applications where they outperformed \cite{lai2018learning}. Also note that the average warping error for the raw processed frame are $0.052$ for the DAVIS dataset and $0.053$ for the videvo.net dataset.
\begin{table}[h!]
\caption{\textbf{Warping Error} of post-processing models }
\centering
\tiny
\begin{tabular}{l|c c c | c c c}
\toprule
\multirow{3}{*}{Task} & \multicolumn{3}{c}{DAVIS} & \multicolumn{3}{c}{VIDEVO.NET} \\
& \multirow{2}{*}{Lai et al.} & \multirow{2}{*}{Ours} & Ours + & \multirow{2}{*}{Lai et al.} & \multirow{2}{*}{Ours} & Ours + \\
& & & SP loss & & & SP loss \\
\midrule
\textbf{WCT/antimono} & $0.031$ & $\mathbf{0.017}$ & $0.024$ & $0.022$ & $\mathbf{0.013}$ & $0.019$ \\
\textbf{WCT/asheville} & $0.059$ & $\mathbf{0.029}$ & $0.043$ & $0.047$ & $\mathbf{0.024}$ & $0.037$\\
\textbf{WCT/candy} & $0.045$ & $\mathbf{0.025}$ & $0.039$ & $0.034$ & $\mathbf{0.020}$ & $0.031$ \\
\textbf{WCT/feathers} & $0.039$ & $\mathbf{0.022}$ & $0.036$ & $0.034$ & $\mathbf{0.021}$ & $0.034$\\
\textbf{WCT/sketch} & $0.030$ & $\mathbf{0.015}$ & $0.024$ & $0.030$ & $\mathbf{0.014}$ & $0.023$ \\
\textbf{WCT/wave} & $0.036$ & $\mathbf{0.020}$ & $0.032$ & $0.057$ & $\mathbf{0.017}$ & $0.027$ \\
\textbf{Fast-neural-style/princess}& $0.060$ & $\mathbf{0.036}$ & $0.060$ & $0.060$ & $\mathbf{0.039}$ & $0.065$\\
\textbf{Fast-neural-style/udnie} & $0.023$ & $\mathbf{0.012}$ & $0.017$ & $0.017$ & $\mathbf{0.009}$ & $0.013$ \\
\textbf{DBL/expertA} & $0.013$ & $\mathbf{0.008}$ & $0.011$ & $0.012$ & $\mathbf{0.008}$ & $0.011$ \\
\textbf{DBL/expertB} & $0.011$ & $\mathbf{0.007}$ & $0.009$ & $0.009$ & $\mathbf{0.006}$ & $0.010$\\
\textbf{Intrinsic/reflectance} & $0.010$ & $\mathbf{0.006}$ & $0.009$ & $0.010$ & $\mathbf{0.007}$ & $0.010$ \\
\textbf{Intrinsic/shading} & $0.009$ & $\mathbf{0.005}$ & $0.007$ &$0.007$ & $\mathbf{0.005}$ & $0.007$ \\
\textbf{CycleGAN/photo2ukiyoe} & $0.020$ & $\mathbf{0.013}$ & $0.019$ & $0.017$ & $\mathbf{0.012}$ & $0.018$ \\
\textbf{CycleGAN/photo2vangogh} & $0.027$ & $\mathbf{0.020}$ & $0.028$ &$0.025$ & $\mathbf{0.017}$ & $0.028$ \\
\textbf{Colorization} & $0.012$ & $\mathbf{0.007}$ & $0.009$ & $0.009$ & $\mathbf{0.006}$ & $0.008$ \\
\textbf{Colorization} & $0.010$ & $\mathbf{0.006}$ & $0.008$ & $0.008$ & $\mathbf{0.006}$ & $0.008$ \\
\midrule
\textbf{Average} & $0.027$ & $\mathbf{0.015}$ & $0.023$ & $0.023$ & $\mathbf{0.014}$ & $0.021$\\
\bottomrule
\end{tabular}
\label{tab:time_met}
\end{table}
When improving temporal consistency in a video, one usually faces a trade-off between making a video more coherent and not altering the overall style of the video.
The metric presented in section \ref{pd} fails to fully capture style conservation. For instance, one major drawback of this metric is its inability to correctly account for new temporal artifacts appearing: in Fig. \ref{fig:retrained_vs_eccv18_trail} and the video available in the supplementary material one can see that the post-processing model of \cite{lai2018learning} generates a trail following the lips, decreasing the resemblance between the original processed video and the post-processed video. However, the perceptual distance seems to favor \cite{lai2018learning} rather than our proposed model despite such phenomenon ($0.026$ for the post-processing model of \cite{lai2018learning} and $0.052$ for our proposed model). Similarly, \cite{lai2018learning} tends to alter the overall style of videos by either changing the contrast or the brightness of videos, while our model successfully avoids such pitfalls (see Fig. \ref{fig:brightness} where both post-processing models obtain a perceptual distance of $0.038$ over the whole video) but the metric can still favor the competing model. We believe that the metric only captures blurriness and major color drift, but fails to measure true resemblance and style preservation. Our model obtains an average perceptual distance over all applications of respectively $0.058$ and $0.056$ with and without our style preserving perceptual loss for the DAVIS dataset, as well as $0.051$ and $0.053$ for the videvo.net dataset, while \cite{lai2018learning} obtained respectively $0.017$ and $0.012$.
\subsection{Discussion}
\begin{figure}[h!]
\centering
\includegraphics[scale=0.4]{files/images/udnie_comparison_2.png}
\caption{Both models are compared using the temporal warping error and the perceptual distance for the application fast-neural-style/udnie. Our model significantly outperforms \cite{lai2018learning} in terms of warping error and obtains slightly higher perceptual distance.}
\label{fig:graph}
\end{figure}
\noindent As seen in Fig. \ref{fig:graph}, we obtain the lowest temporal warping error at the cost of a slightly larger perceptual distance than \cite{lai2018learning}. Note that when making their hyperparameters vary, \cite{lai2018learning} manage to improve their temporal coherence at the cost of a much higher perceptual distance, but they never attain as low perceptual distance as our model. The blue dots in Fig. \ref{fig:graph} correspond to the model of \cite{lai2018learning} for several sets of hyperparameters.
When including the style preserving perceptual loss, we lightly deteriorate both the perceptual distance and temporal warping error.
However as seen in Fig. \ref{fig:brightness} the style preserving perceptual loss reduces deterioration of the brightness throughout videos and style deviation. This supports the idea that the perceptual distance fails to fully capture style preservation. For instance, both our model and the model of \cite{lai2018learning} can alter the brightness of videos and our proposed loss manages to reduce such issue.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.2]{files/images/couple_SPLOSS.png}
\caption{Without our style preserving perceptual loss the video tends to become darker when post-processed.}
\label{fig:brightness}
\end{figure}
\section{Conclusion}
\label{conclusion}
We discussed questions regarding temporal consistency linked to independently processing videos frame by frame using image trained algorithms. We put forward a new model to remove flickering artifacts, it includes recent techniques proposed in approaches orthogonal to our work in the field of video processing, as well as a novel loss to circumvent problems linked to style deterioration throughout videos.
Thanks to these modifications, we obtain state-of-the-art results regarding flickering removal and also improve the visual output of the post-processed videos.
We believe that future work in the field of video temporal post-processing ought to focus on algorithms that do not necessitate to have the original non-flickering video for inference.
\bibliographystyle{IEEEbib}
|
\section{Introduction} \label{sec intro}
Multiplicities of zeros of polynomials constitute a central elementary concept with a wide range of generalizations and applications. One of the most basic results states that, if $ F $ is a non-zero polynomial over a field and $ a_1, a_2, \ldots, a_n $ are all of its $ n $ distinct roots (in the same field), with multiplicities $ r_1, r_2, \ldots, r_n $, respectively, then
\begin{equation}
\sum_{i=1}^n r_i \leq \deg(F),
\label{eq degree bound intro}
\end{equation}
where equality holds for all $ F $ if, and only if, the field is algebraically closed. The bound (\ref{eq degree bound intro}) is equivalent to Hermite interpolation (in the sense of Theorems \ref{th hermite interpolation nonconj naive case} and \ref{th hermite interpolation II} using Hasse derivatives \cite{hasse} in positive characteristic) and the invertibility of confluent Vandermonde matrices \cite{kalman}.
However, multiplicities in non-commutative algebra are not understood to the same extent. The study of zeros of conventional polynomials over division rings was initiated in \cite{wedderburn} and later \cite{gordon}. Zeros of quaternionic regular functions (which include polynomials) were studied in \cite{gentili}, and their multiplicities were considered in \cite[Def. 5.5]{gentili-zero}.
In an alternative line of research, Lam and Leroy \cite{lam, lam-leroy} described exactly the structure of zeros (without counting multiplicities) of general skew polynomials \cite{ore}, which include conventional polynomials over division rings \cite{wedderburn}, linearized polynomials over finite fields \cite{orespecial}, and differential polynomials \cite{jacobson-derivation}, among others. Lam and Leroy's works characterize evaluation points where (\ref{eq degree bound intro}) holds without counting multiplicities. Such sets of points are called \textit{P-independent}. Lam and Leroy's results admit natural generalizations to (free) multivariate skew polynomial rings \cite{lin-multivariateskew, multivariateskew}.
In \cite{eric, gentili-zero, johnson-thesis}, an element $ a $ is considered to be a zero (say, on the right) of a skew polynomial $ F $ with multiplicity $ r \in \mathbb{N} $ if $ (x-a)^r $ divides $ F $ on the right. Although this notion works fine in the quaternionic case \cite{bolotnikov-conf}, we will show that it becomes quite pathological for general skew polynomials (see Examples \ref{ex power of linear term is no multiplicity} and \ref{ex full conjugacy class pathology for naive mult}).
An interesting alternative notion of multiplicity for zeros of quaternionic polynomials was recently given by Bolotnikov \cite{bolotnikov-conf, bolotnikov-zeros}. Bolotnikov defines $ a $ as a zero of $ F $ of multiplicity $ r \in \mathbb{N} $ if there exists a skew polynomial of degree $ r $ that factors completely
\begin{equation}
P = (x-a_r) \cdots (x - a_2) (x-a_1)
\label{eq P intro}
\end{equation}
such that $ P $ divides $ F $ on the right, $ a = a_1 $ and $ a_1 $ is the only zero of $ P $ (on the right). Over the quaternions, this latter condition on $ P $ holds if $ a_1 = a_2 = \ldots = a_r $ (see Corollary \ref{cor multiplicity over quaternions}), but in contrast with the commutative case, the converse does not hold. In other words, $ a_1 $ may be the only right zero of $ P $, yet the set $ \{ a_1, a_2, \ldots, a_r \} $ may have more than one element. As we will show, even the direct implication does not hold for general skew polynomials. In other words, $ (x-a)^r $ may have more than one zero without counting multiplicities (see Example \ref{ex power of linear term is no multiplicity}).
In this manuscript, we study separately both definitions of multiplicity of zeros of general skew polynomials. We extend classical results, including the degree bound (\ref{eq degree bound intro}) above (Theorems \ref{th upper bound multiplicities by degree general}, \ref{th deg bound for nonconj naive case} and \ref{th upper bound multiplicities by degree II}), Hermite interpolation based on Hasse derivatives or confluent Vandermonde matrices (Theorems \ref{th lagrange interpolation}, \ref{th hermite interpolation nonconj naive case} and \ref{th hermite interpolation II}), or multiplicity enhancements of Hilbert's Theorem 90 (Theorem \ref{th min skew pol of conjugacy class with mult} and Corollary \ref{cor hilbert 90}). We will also characterize skew polynomials that have a unique factorization in linear terms (Theorem \ref{th charact multiplicity seqs in alg conj}).
\section{High-degree evaluation points} \label{sec high-degree zeros}
In this section, we provide an abstract framework for generalizing the study of zeros of skew polynomials by Lam and Leroy \cite{lam, lam-leroy}. The final objective is to be able to handle multiplicities of classical zeros of skew polynomials (Sections \ref{sec multi from powers} and \ref{sec multi from linear terms}).
Skew polynomial rings were introduced by Ore \cite{ore} as follows. Fix a division ring $ \mathbb{F} $, let $ \mathbb{N} $ be the set of natural numbers including $ 0 $, and set $ \mathbb{Z}_+ = \mathbb{N} \setminus \{ 0 \} $. Let $ \mathcal{R} $ be the left vector space over $ \mathbb{F} $ with basis $ \{ x^i \mid i \in \mathbb{N} \} $, where we denote $ 1 = x^0 $ and $ x = x^1 $. The degree of a non-zero $ F = \sum_{i \in \mathbb{N}} F_i x^i \in \mathcal{R} $, where $ F_i \in \mathbb{F} $ for all $ i \in \mathbb{N} $, is the maximum $ i $ such that $ F_i \neq 0 $, and is denoted by $ \deg(F) $. We define $ \deg(F) = - \infty $ if $ F = 0 $.
A product in $ \mathcal{R} $ turns it into a (non-commutative) ring with identity $ 1 $, with $ \mathbb{F} \subseteq \mathcal{R} $ as a subring, where $ x^i x^j = x^{i+j} $, for all $ i,j \in \mathbb{N} $, and $ \deg(FG) = \deg(F) + \deg(G) $ for all $ F,G \in \mathcal{R} $, if, and only if, there exist $ \sigma, \delta : \mathbb{F} \longrightarrow \mathbb{F} $ such that
\begin{equation}
xa = \sigma(a) x + \delta(a),
\label{eq product over constants and variables}
\end{equation}
for all $ a \in \mathbb{F} $, where $ \sigma : \mathbb{F} \longrightarrow \mathbb{F} $ is a ring endomorphism and $ \delta : \mathbb{F} \longrightarrow \mathbb{F} $ is a \textit{$ \sigma $-derivation}: That is, $ \delta $ is additive and $ \delta(ab) = \sigma(a)\delta(b) + \delta(a)b $, for all $ a,b \in \mathbb{F} $.
For each such pair $ (\sigma,\delta) $, we denote $ \mathcal{R} = \mathbb{F}[x; \sigma, \delta] $ when the product is given by (\ref{eq product over constants and variables}), and we call $ \mathbb{F}[x; \sigma, \delta] $ the \textit{skew polynomial ring} over $ \mathbb{F} $ with morphism $ \sigma $ and derivation $ \delta $. The \textit{conventional polynomial ring} $ \mathbb{F}[x] $ is obtained by choosing the identity morphism $ \sigma = {\rm Id} $ and the zero derivation $ \delta = 0 $. Moreover, $ \mathbb{F}[x; \sigma, \delta] $ is commutative if, and only if, $ \mathbb{F} $ is a field, $ \sigma = {\rm Id} $ and $ \delta = 0 $. This will be called \textit{the commutative case}.
The ring $ \mathbb{F}[x; \sigma, \delta] $ is a right Euclidean domain \cite{ore}. Thanks to this property, Lam and Leroy \cite{lam, lam-leroy} provided a natural definition of evaluation by forcing a ``Remainder Theorem'' on the right. We now extend this notion of evaluation as follows.
\begin{definition} [\textbf{High-degree evaluation}] \label{def evaluation high degree}
Given $ F,P \in \mathbb{F}[x;\sigma,\delta] $, we define the evaluation of $ F $ at $ P $, denoted by $ F(P) $, as the remainder of $ F $ when divided by $ P $ by Euclidean division on the right. Thus $ F(P) \in \mathbb{F}[x;\sigma, \delta] $ is a skew polynomial such that $ \deg(F(P)) < \deg(P) $. We say that $ P $ is a zero of $ F $ if $ F(P) = 0 $.
\end{definition}
The evaluation of $ F \in \mathbb{F}[x;\sigma,\delta] $ in $ a \in \mathbb{F} $ given in \cite{lam, lam-leroy} corresponds to the evaluation of $ F $ in $ x - a \in \mathbb{F}[x;\sigma,\delta] $ according to Definition \ref{def evaluation high degree}, and will be denoted $ F(a) \in \mathbb{F} $.
In this manuscript, we will consider non-empty sets $ \mathcal{U} \subseteq \mathbb{F}[x;\sigma,\delta] $ such that $ \deg(F) \geq 1 $, for all $ F \in \mathcal{U} $. The set $ \mathcal{U} $ indicates where we may evaluate, thus where we may take zeros from. The case of classical zeros of skew polynomials is recovered by considering
\begin{equation}
\mathcal{U}_0 = \{ x-a \in \mathbb{F}[x;\sigma,\delta] \mid a \in \mathbb{F} \}.
\label{eq universal set 0}
\end{equation}
We will identify a set $ \Omega \subseteq \mathcal{U}_0 $ with its underlying set $ \{ a \in \mathbb{F} \mid x-a \in \Omega \} \subseteq \mathbb{F} $. In Section \ref{sec multi from powers}, we will count multiplicities by considering zeros in the universal set
\begin{equation}
\mathcal{U}_1 = \left\lbrace (x-a)^r \in \mathbb{F}[x;\sigma,\delta] \mid a \in \mathbb{F}, r \in \mathbb{Z}_+ \right\rbrace ,
\label{eq universal set I}
\end{equation}
and from Section \ref{sec multi from linear terms} on, we will consider the universal set
\begin{equation}
\mathcal{U}_2 = \left\lbrace P_\mathbf{a} \in \mathbb{F}[x;\sigma,\delta] \mid Z(P_\mathbf{a}) = \{ a_1 \}, \mathbf{a} \in \mathbb{F}^r, r \in \mathbb{Z}_+ \right\rbrace ,
\label{eq universal set II}
\end{equation}
where $ P_\mathbf{a} = (x-a_r) \cdots (x-a_2)(x-a_1) \in \mathbb{F}[x;\sigma,\delta] $ if $ \mathbf{a} = (a_1, a_2, \ldots, a_r) \in \mathbb{F}^r $. Whereas it trivially holds that $ \mathcal{U}_0 \subseteq \mathcal{U}_1 $ and $ \mathcal{U}_0 \subseteq \mathcal{U}_2 $, neither the inclusion $ \mathcal{U}_1 \subseteq \mathcal{U}_2 $ nor the inclusion $ \mathcal{U}_2 \subseteq \mathcal{U}_1 $ hold in general (see Example \ref{ex power of linear term is no multiplicity} and Section \ref{sec particular cases}).
The following definition is inspired by classical Algebraic Geometry.
\begin{definition} [\textbf{High-degree zeros}]
Given $ A \subseteq \mathbb{F}[x;\sigma,\delta] $, its zero set in $ \mathcal{U} $ is
$$ Z_\mathcal{U}(A) = \{ P \in \mathcal{U} \mid F(P) = 0, \textrm{ for all } F \in A \}. $$
Similarly, given a set $ \Omega \subseteq \mathcal{U} $, the ideal of skew polynomials vanishing at $ \Omega $ is
$$ I(\Omega) = \{ F \in \mathbb{F}[x;\sigma,\delta] \mid F(P) = 0, \textrm{ for all } P \in \Omega \}. $$
\end{definition}
Some properties of zero sets and ideals will require the following generalization of set inclusion (we also use it in Corollary \ref{cor monotonicity}, Proposition \ref{prop monotonicity multiplicity I} and Theorem \ref{th P-independence multiplicity seqs}).
\begin{definition} \label{def subseteq leq}
Given sets $ \Psi, \Omega \subseteq \mathbb{F}[x;\sigma, \delta] $, we say that $ \Psi \leq \Omega $ if, for every finite non-empty subset $ \{ Q_1, Q_2, \ldots, $ $ Q_n \} $ $ \subseteq \Psi $ of size $ n $, there exists a subset $ \{ P_1, P_2, \ldots, P_n \} \subseteq \Omega $, also of size $ n $, such that $ Q_i $ divides $ P_i $ on the right, for $ i = 1,2, \ldots, n $.
\end{definition}
\begin{example}
Clearly $ \Psi \leq \Omega $ whenever $ \Psi \subseteq \Omega $. However, $ \Psi \leq \Omega $ holds in many other cases. For example, if $ \Psi = \{ (x-a)^r \} $, for $ a \in \mathbb{F} $ and $ r \in \mathbb{Z}_+ $, then $ \Psi \leq \Omega $ for any $ \Omega = \{ (x-a)^s \} $, where $ s \geq r $. The interpretation, in this case, is that $ \Psi $ and $ \Omega $ correspond to the point $ a $ with multiplicities $ r $ and $ s $, and $ \Psi \leq \Omega $ indicates that the same holds for their multiplicities, i.e., $ r \leq s $. More generally, $ \Psi \leq \Omega $ for $ \Psi = \{ (x-a_1)^{r_1}, \ldots, (x-a_n)^{r_n} \} $ and $ \Omega = \{ (x-a_1)^{s_1}, \ldots, (x-a_n)^{s_n} \} $, where $ a_1, a_2, \ldots, a_n \in \mathbb{F} $ and $ r_i ,s_i \in \mathbb{Z}_+ $ are such that $ r_i \leq s_i $, for $ i = 1,2, \ldots, n $, and $ \Omega $ has size $ n $. Note that, in these cases, $ \Psi \subseteq \Omega $ does not hold unless $ r_i = s_i $ for all $ i = 1,2, \ldots, n $.
\end{example}
Throughout the manuscript, given $ F \in \mathbb{F}[x;\sigma,\delta] $, we denote by $ (F) = \{ GF \mid G \in \mathbb{F}[x;\sigma,\delta] \} $ the left ideal in $ \mathbb{F}[x;\sigma,\delta] $ generated by $ F $. An important feature of this abstract framework is that the following basic properties still hold as in classical Algebraic Geometry.
\begin{proposition} \label{prop properties alg geom}
Let $ \Omega, \Omega_1, \Omega_2 \subseteq \mathcal{U} $ and $ A, A_1, A_2 \subseteq \mathbb{F}[x;\sigma,\delta] $ be arbitrary sets.
\begin{enumerate}
\item
The set $ I(\Omega) $ is a left ideal of $ \mathbb{F}[x;\sigma,\delta] $.
\item
Given $ P \in \mathcal{U} $, it holds that $ I(\{ P \}) = (P) $ and $ P \in Z_\mathcal{U}(\{ P \}) $.
\item
Assume that if $ P \in \mathcal{U} $, $ Q \in \mathbb{F}[x;\sigma,\delta] $ divides $ P $ on the right and $ \deg(Q) \geq 1 $, then $ Q \in \mathcal{U} $. Then $ Z_\mathcal{U}(\{ P \}) = \{ P \} $ if, and only if, $ P $ is irreducible.
\item
$ I(\varnothing) = (1) $ and $ Z_\mathcal{U}(\{ 1 \}) = \varnothing $.
\item
If $ \Omega_1 \leq \Omega_2 $, then $ I(\Omega_2) \subseteq I(\Omega_1) $.
\item
If $ A_1 \subseteq A_2 $, then $ Z_\mathcal{U}(A_2) \subseteq Z_\mathcal{U}(A_1) $.
\item
$ I(\Omega_1 \cup \Omega_2) = I(\Omega_1) \cap I(\Omega_2) $.
\item
$ Z_\mathcal{U}(A) = Z_\mathcal{U}((A)) $ and $ Z_\mathcal{U}(A_1 \cup A_2) = Z_\mathcal{U}((A_1) + (A_2)) = Z_\mathcal{U}(A_1) \cap Z_\mathcal{U}(A_2) $.
\item
$ \Omega \subseteq Z_\mathcal{U}(I(\Omega)) $ and equality holds if, and only if, $ \Omega = Z_\mathcal{U}(B) $ for some $ B \subseteq \mathbb{F}[x;\sigma,\delta] $.
\item
$ A \subseteq (A) \subseteq I(Z_\mathcal{U}(A)) $ and equality holds if, and only if, $ A = I(\Psi) $ for some $ \Psi \subseteq \mathcal{U} $.
\end{enumerate}
\end{proposition}
\begin{proof}
We prove the following items:
\begin{enumerate}
\item
If $ F \in \mathbb{F}[x;\sigma,\delta] $, $ G,H \in I(\Omega) $ and $ P \in \Omega $, then $ P $ divides $ G $ and $ H $ on the right. Thus $ P $ also divides $ FG $ and $ G+H $ on the right.
\item
$ F \in I(\{ P \}) $ if, and only if, $ P $ divides $ F $ on the right.
\item
If $ \mathcal{U} $ satisfies such a property, then $ Z_\mathcal{U}(\{ P \}) $ is formed by all the right divisors of $ P $ of degree at least $ 1 $.
\item[5.]
Let $ F \in I(\Omega_2) $ and $ P \in \Omega_1 $. Since $ \Omega_1 \leq \Omega_2 $, then there is $ Q \in \Omega_2 $ such that $ P $ divides $ Q $ on the right. Since $ F \in I(\Omega_2) $, then $ Q $ divides $ F $ on the right, hence $ P $ divides $ F $ on the right and $ F \in I(\Omega_1) $.
\end{enumerate}
The remaining items are trivial from the definitions. Regarding equality in item 9, note that if $ \Omega = Z_\mathcal{U}(B) $, then $ B \subseteq I(Z_\mathcal{U}(B)) = I(\Omega) $ by item 10, thus $ Z_\mathcal{U}(I(\Omega)) \subseteq Z_\mathcal{U}(B) = \Omega $. Since $ \Omega \subseteq Z_\mathcal{U}(I(\Omega)) $, then equality holds. Similarly for the equality in item 10.
\end{proof}
We may now extend Lam and Leroy's concepts of algebraic sets, minimal skew polynomials, P-closed sets, P-independent sets, P-generators and P-bases \cite{lam, algebraic-conjugacy, lam-leroy}
\begin{definition} [\textbf{Algebraic sets}] \label{def algebraic sets}
We say that a set $ \Omega \subseteq \mathcal{U} $ is algebraic if $ I(\Omega) \neq \{ 0 \} $.
\end{definition}
\begin{definition} [\textbf{Minimal skew polynomial}]
If $ \Omega \subseteq \mathcal{U} $ is algebraic, we define its minimal skew polynomial, denoted $ F_\Omega $, as the unique monic non-zero skew polynomial of minimum possible degree in $ I(\Omega) $. If $ \Omega \subseteq \mathcal{U} $ is not algebraic, we define $ F_\Omega = 0 $.
\end{definition}
For any set $ \Omega \subseteq \mathcal{U} $, the skew polynomial $ F_\Omega $ exists and generates $ I(\Omega) $.
In order to bound the degree of the minimal skew polynomial in Proposition \ref{prop deg of min skew pol upper bound by sum}, we will use the following relation, see \cite{ore} or \cite[Deg. Eq. 4.1]{lam-leroy-wedI}.
\begin{lemma} [\textbf{\cite{ore}}] \label{lemma ores cross cut formula}
Given skew polynomials $ F,G \in \mathbb{F}[x;\sigma,\delta] $, let $ D, M \in \mathbb{F}[x;\sigma,\delta] $ be their greatest right common divisor and least left common multiple. Then
$$ \deg(D) + \deg(M) = \deg(F) + \deg(G). $$
\end{lemma}
The first step in extending the concept of P-independence is the following proposition, which extends \cite[Prop. 6]{lam} to high-degree zeros.
\begin{proposition} \label{prop deg of min skew pol upper bound by sum}
Given a non-empty set $ \Omega \subseteq \mathcal{U} $, it holds that
$$ \deg(F_\Omega) \leq \sum_{P \in \Omega} \deg(P). $$
\end{proposition}
\begin{proof}
First, if $ \Omega $ is not finite, then the right-hand side is infinite (recall that $ \deg(P) \geq 1 $ for all $ P \in \mathcal{U} $), hence the inequality holds. Second, if $ \Omega $ is not algebraic, then $ F_\Omega = 0 $ and $ \deg(F_\Omega) = - \infty $, hence the inequality also holds. Finally assume that $\Omega $ is finite and algebraic. By definition, $ F_\Omega $ is the least left common multiple of the skew polynomials in $ \Omega $. Therefore, the result follows from Lemma \ref{lemma ores cross cut formula} applied recursively.
\end{proof}
We now extend the notion of \textit{P-independence} \cite[Sec. 4]{lam} to high-degree points.
\begin{definition} [\textbf{P-independence}] \label{def P-independence}
$ \Omega \subseteq \mathcal{U} $ is P-independent if it is finite, algebraic and
$$ \deg(F_\Omega) = \sum_{P \in \Omega} \deg(P). $$
We also say that a subset $ \Psi = \{ a_1 , a_2, \ldots, a_n \} \subseteq \mathbb{F} $ is P-independent if the corresponding set $ \{ x-a_1, x-a_2, \ldots, x-a_n \} \subseteq \mathcal{U}_0 $ is P-independent.
\end{definition}
P-independent sets are precisely those for which we may obtain an upper bound on the number of zeros of a skew polynomial by its degree (see also Theorems \ref{th deg bound for nonconj naive case} and \ref{th upper bound multiplicities by degree II}).
\begin{theorem} [\textbf{Degree bound}] \label{th upper bound multiplicities by degree general}
An algebraic set $ \Omega \subseteq \mathcal{U} $ is P-independent if, and only if, for every non-zero $ F \in I(\Omega) $, it holds that
\begin{equation}
\sum_{P \in \Omega} \deg(P) \leq \deg(F).
\label{eq degree bound for proof}
\end{equation}
Moreover, when this is the case, then equality in (\ref{eq degree bound for proof}) is attained by $ F $ if, and only if, $ F = a F_\Omega $, where $ a \in \mathbb{F}^* $.
\end{theorem}
\begin{proof}
First assume that $ \Omega $ is P-independent. By definition, $ F_\Omega \neq 0 $ and if $ F \in I(\Omega) $, then $ F_\Omega $ divides $ F $ on the right. By definition of P-independence, we conclude that
$$ \sum_{P \in \Omega} \deg(P) = \deg(F_\Omega) \leq \deg(F). $$
Conversely, assume that $ \Omega $ is not P-independent. Then (\ref{eq degree bound for proof}) does not hold for $ F_\Omega \in I(\Omega) $ by definition (note that $ F_\Omega \neq 0 $ since $ \Omega $ is algebraic).
\end{proof}
We may now define \textit{P-closed sets}.
\begin{definition} [\textbf{P-closed sets}] \label{def P-closed sets}
For a given set $ \Omega \subseteq \mathcal{U} $, we define its P-closure (in $ \mathcal{U} $) as $ \overline{\Omega} = Z_\mathcal{U}(I(\Omega)) $. We say that $ \Omega $ is P-closed (in $ \mathcal{U} $) if $ \overline{\Omega} = \Omega $.
\end{definition}
The next concepts to generalize are those of P-generators and P-bases.
\begin{definition} [\textbf{P-bases}] \label{def P-bases}
Let $ \Omega \subseteq \mathcal{U} $ be a P-closed set. We say that $ \mathcal{G} \leq \Omega $ is a set of P-generators for $ \Omega $ if $ \Omega = \overline{\mathcal{G}} $. Finally, we say that $ \mathcal{B} \leq \Omega $ is a P-basis of $ \Omega $ if $ \mathcal{B} $ is P-independent and a set of P-generators of $ \Omega $.
\end{definition}
The following lemma ensures that all P-bases have the same ``size'' (counting degrees). It follows from the definitions and Proposition \ref{prop properties alg geom}.
\begin{lemma} \label{lemma P-generators iff min skew pols equal}
Given a P-closed set $ \Omega \subseteq \mathcal{U} $ and a set $ \mathcal{G} \leq \Omega $, then $ \mathcal{G} $ is a set of P-generators of $ \Omega $ if, and only if, $ F_\mathcal{G} = F_\Omega $. In particular, if $ \mathcal{B} \leq \Omega $ is a P-basis of $ \Omega $, then
$$ \deg(F_\Omega) = \sum_{P \in \mathcal{B}} \deg(P) . $$
The number $ {\rm Rk}(\Omega) = deg(F_\Omega) $ will be called the rank of $ \Omega $.
\end{lemma}
We next obtain a general Lagrange interpolation theorem. This result will be a prequel for a Hermite interpolation theorem (see Theorems \ref{th hermite interpolation nonconj naive case} and \ref{th hermite interpolation II}).
\begin{theorem} [\textbf{Lagrange interpolation}] \label{th lagrange interpolation}
Let $ \Omega \subseteq \mathcal{U} $ be a finite algebraic set given by $ \Omega = \{ P_1, P_2, \ldots, P_n \} $, of size $ n $, and define $ N = \deg(F_\Omega) $. The following are equivalent:
\begin{enumerate}
\item
$ \Omega $ is P-independent.
\item
The left $ \mathbb{F} $-linear map
$$ \phi : \mathbb{F}[x;\sigma, \delta]_N \longrightarrow \prod_{i=1}^n \frac{\mathbb{F}[x;\sigma,\delta]}{(P_i)} $$
given by $ \phi (F) = \left( F(P_i) \right)_{i=1}^n $, for $ F \in \mathbb{F}[x;\sigma, \delta]_N $, is a left vector space isomorphism.
\item
The left $ \mathbb{F} $-linear map
$$ \psi : \mathbb{F}[x;\sigma, \delta] \longrightarrow \prod_{i=1}^n \frac{\mathbb{F}[x;\sigma,\delta]}{(P_i)} $$
given by $ \psi (F) = \left( F(P_i) \right)_{i=1}^n $, for $ F \in \mathbb{F}[x;\sigma, \delta] $, is surjective.
\end{enumerate}
Here, $ \mathbb{F}[x;\sigma, \delta]_N = \{ F \in \mathbb{F}[x;\sigma, \delta] \mid \deg(F) < N \} $ and $ \frac{\mathbb{F}[x;\sigma,\delta]}{(P)} $ denotes the left quotient module of $ \mathbb{F}[x;\sigma,\delta] $ by the left ideal $ (P) $, for $ P \in \mathbb{F}[x;\sigma,\delta] $.
\end{theorem}
\begin{proof}
Before proving the corresponding implications, we show that the map $ \phi $ is always injective. Let $ F \in \mathbb{F}[x;\sigma,\delta]_N $ and assume that $ \phi(F) = 0 $. Then $ F \in I(\Omega) $, and therefore, $ F_\Omega $ divides $ F $ on the right. If $ F \neq 0 $, then we would have that $ \deg(F) \geq \deg(F_\Omega) = N $, which is absurd since $ F \in \mathbb{F}[x;\sigma,\delta]_N $. Thus $ F = 0 $ and $ \phi $ is injective.
We now prove the following equivalences:
$ 1. \Longleftrightarrow 2. $: By counting dimensions, it is trivial that item 2 implies item 1. Assume now that item 1 holds, that is, $ \Omega $ is P-independent. Then the domain and co-domain of $ \phi $ have the same dimension on the left over $ \mathbb{F} $. Since $ \phi $ is always injective, we conclude that $ \phi $ is a left vector space isomorphism.
$ 2. \Longleftrightarrow 3. $: It is trivial that item 2 implies item 3. Assume now that item 3 holds. Since $ \phi $ is always injective, we only need to show that, given $ F \in \mathbb{F}[x;\sigma,\delta] $, there exists $ G \in \mathbb{F}[x;\sigma,\delta]_N $ such that $ F(P_i) = G(P_i) $, for $ i = 1,2, \ldots, n $. To achieve this, let $ G $ be the remainder of $ F $ divided by $ F_\Omega $ on the right (note that $ F_\Omega \neq 0 $ since $ \Omega $ is algebraic).
\end{proof}
We now deduce the following monotonicity property of P-independence, which will be useful later (see Proposition \ref{prop monotonicity multiplicity I} and Theorem \ref{th P-independence multiplicity seqs}).
\begin{corollary} [\textbf{Monotonicity}] \label{cor monotonicity}
Let $ \Psi, \Omega \subseteq \mathcal{U} $ be finite non-empty sets such that $ \Psi \leq \Omega $. If $ \Omega $ is P-independent, then so is $ \Psi $.
\end{corollary}
\begin{proof}
First observe that, if $ \Omega $ is algebraic, then so is $ \Psi $ by Proposition \ref{prop properties alg geom}. By Definition \ref{def subseteq leq}, if we set $ \Psi = \{ Q_1, Q_2, \ldots, Q_n \} $ of size $ n $, then there exists a subset $ \{ P_1, P_2, \ldots, P_n \} \subseteq \Omega $ of size $ n $, such that $ Q_i $ divides $ P_i $ on the right, for $ i = 1,2, \ldots, n $. In particular, we have that $ (P_i) \subseteq (Q_i) $, for $ i = 1,2, \ldots, n $. Therefore, we have a natural surjective map
$$ \prod_{i=1}^n \frac{\mathbb{F}[x;\sigma, \delta]}{(P_i)} \longrightarrow \prod_{i=1}^n \frac{\mathbb{F}[x;\sigma, \delta]}{(Q_i)} . $$
Hence if $ \Omega $ is P-independent, then so is $ \Psi $ by item 3 in Theorem \ref{th lagrange interpolation}.
\end{proof}
\section{Hasse derivatives of skew polynomials} \label{sec hasse derivatives}
Throughout this manuscript, given $ r \in \mathbb{Z}_+ $ and a sequence $ \mathbf{a} = (a_1, a_2, \ldots , a_r) \in \mathbb{F}^r $, we define its associated skew polynomial as
\begin{equation}
P_\mathbf{a} = (x-a_r) \cdots (x-a_2) (x - a_1) \in \mathbb{F}[x;\sigma,\delta] .
\label{eq def multiplicity seq polynomial}
\end{equation}
In this section, we introduce a skew polynomial extension of Hasse derivatives \cite{hasse}. The definition is given so that a point is a zero of a skew polynomial with multiplicity $ r \in \mathbb{Z}_+ $ if, and only if, the corresponding first $ r $ consecutive Hasse derivatives of the skew polynomial at that point are zero. This is not the case with previous notions of derivatives of skew polynomials \cite{bolotnikov-conf, eric} if $ \mathbb{F} $ has positive characteristic.
\begin{definition} [\textbf{Hasse derivative}] \label{def derivative}
Given $ F \in \mathbb{F}[x;\sigma,\delta] $, $ r \in \mathbb{Z}_+ $, and $ \mathbf{a} = (a_1, a_2, $ $ \ldots, a_r) \in \mathbb{F}^r $, we define the corresponding Hasse derivative of order $ r $ as the unique $ D_{\mathbf{a}} (F) \in \mathbb{F} $ such that there exist $ G, H \in \mathbb{F}[x;\sigma,\delta] $, where $ H $ is monic of degree $ r-1 $ and $ F = G P_{\mathbf{a}} + D_{\mathbf{a}} (F) H $, for $ P_{\mathbf{a}} \in \mathbb{F}[x;\sigma,\delta] $ as in (\ref{eq def multiplicity seq polynomial}).
\end{definition}
The existence and uniqueness of Hasse derivatives, and the following proposition, follow directly from right Euclidean division.
\begin{proposition} [\textbf{Taylor expansion}] \label{prop taylor expansion}
Let $ r \in \mathbb{Z}_+ $ and $ \mathbf{a} = (a_1, a_2, \ldots, a_r) \in \mathbb{F}^r $. For every skew polynomial $ F \in \mathbb{F}[x;\sigma,\delta] $, there exists $ G \in \mathbb{F}[x;\sigma,\delta] $ such that
$$ F = G P_{\mathbf{a}_r} + D_{\mathbf{a}_{r}} (F) P_{\mathbf{a}_{r-1}} + \cdots + D_{\mathbf{a}_{2}} (F) P_{\mathbf{a}_1} + D_{\mathbf{a}_1} (F) , $$
where $ \mathbf{a}_i = (a_1, a_2, \ldots, a_i) \in \mathbb{F}^i $ and $ P_{\mathbf{a}_i} \in \mathbb{F}[x;\sigma,\delta] $ is as in (\ref{eq def multiplicity seq polynomial}), for $ i = 1,2, \ldots, r $. In particular, we have that $ D_{\mathbf{a}_1}(F) = F(a_1) \in \mathbb{F} $ (see Definition \ref{def evaluation high degree})
\end{proposition}
We conclude with the corresponding \textit{confluent Vandermonde matrices}.
\begin{definition} [\textbf{Confluent Vandermonde matrices}] \label{def confluent vandermonde}
Let $ n, r_i \in \mathbb{Z}_+ $ and $ \mathbf{a}_i = (a_{i,1}, a_{i,2}, $ $ \ldots , $ $ a_{i,r_i}) $ $ \in \mathbb{F}^{r_i} $, and define $ \mathbf{a}_{i,j} = (a_{i,1}, a_{i,2} , \ldots, a_{i,j}) \in \mathbb{F}^j $, for $ j = 1,2, \ldots r_i $ and $ i = 1,2, \ldots, n $. Given $ N \in \mathbb{Z}_+ $, define the confluent Vandermonde matrix of order $ N $ on $ \mathbf{a}_i $ as the $ N \times r_i $ matrix with entries in $ \mathbb{F} $ given by
$$ V_N (\mathbf{a}_i) = \left( \begin{array}{cccc}
D_{\mathbf{a}_{i,1}} (1) & D_{\mathbf{a}_{i,2}} (1) & \ldots & D_{\mathbf{a}_{i,r_i}} (1) \\
D_{\mathbf{a}_{i,1}} (x) & D_{\mathbf{a}_{i,2}} (x) & \ldots & D_{\mathbf{a}_{i,r_i}} (x) \\
\vdots & \vdots & \ddots & \vdots \\
D_{\mathbf{a}_{i,1}} (x^{N-1}) & D_{\mathbf{a}_{i,2}} (x^{N-1}) & \ldots & D_{\mathbf{a}_{i,r_i}} (x^{N-1}) \\
\end{array} \right) , $$
for $ i = 1,2, \ldots, n $. Finally, we define the confluent Vandermonde matrix of order $ N $ on $ \mathbf{a}_1, \mathbf{a}_2, \ldots, $ $ \mathbf{a}_n $ as the $ N \times (\sum_{i=1}^n r_i) $ matrix with entries in $ \mathbb{F} $ given by
$$ V_N(\mathbf{a}_1, \mathbf{a}_2, \ldots, \mathbf{a}_n) = \left( V_N (\mathbf{a}_1), V_N (\mathbf{a}_2), \ldots, V_N (\mathbf{a}_n) \right) . $$
\end{definition}
Classical confluent Vandermonde matrices \cite{kalman} are recovered from Definition \ref{def confluent vandermonde} when $ \sigma = {\rm Id} $, $ \delta = 0 $ and $ a_{i,1} = a_{i,2} = \ldots = a_{i,r_i} $, for $ i = 1,2, \ldots, n $. Skew polynomial Vandermonde matrices \cite{lam, lam-leroy} are recovered from Definition \ref{def confluent vandermonde} when $ r_1 = r_2 = \ldots = r_n = 1 $. Finally, quaternionic confluent Vandermonde matrices \cite{bolotnikov-conf} are recovered from Definition \ref{def confluent vandermonde} when $ \mathbb{F} $ denotes the quaternions, $ \sigma = {\rm Id} $, $ \delta = 0 $ and $ \mathbf{a}_i $ is a spherical chain (what we will call a multiplicity sequence in Definition \ref{def multiplicity seq}), for $ i = 1,2, \ldots, n $.
\section{Multiplicities from powers of a linear polynomial} \label{sec multi from powers}
In this section, we provide a first definition of multiplicities of zeros of skew polynomials (similar to \cite{eric, gentili-zero, johnson-thesis}). We consider a different definition in Sections \ref{sec multi from linear terms}, \ref{sec multiplicity seqs} and \ref{sec particular cases}.
\begin{definition} [\textbf{Multiplicity I}] \label{def multiplicity I}
Given $ F \in \mathbb{F}[x;\sigma,\delta] $, $ a \in \mathbb{F} $ and $ r \in \mathbb{Z}_+ $, we say that $ a $ is a zero of $ F $ of multiplicity $ r $ if $ (x-a)^r \in \mathbb{F}[x;\sigma,\delta] $ divides $ F $ on the right.
\end{definition}
We may now characterize multiplicities (as in Definition \ref{def multiplicity I}) in terms of Hasse derivatives (Definition \ref{def derivative}). This result follows from Proposition \ref{prop taylor expansion}.
\begin{proposition} [\textbf{Derivative criterion I}] \label{prop charact multiplicities derivatives I}
Given $ F \in \mathbb{F}[x;\sigma,\delta] $ and $ a \in \mathbb{F} $, denote by $ F^{(i)} (a) = D_{\mathbf{a}_{i+1}}(F) \in \mathbb{F} $ the Hasse derivative (Definition \ref{def derivative}) of $ F $ at $ \mathbf{a}_{i+1} = (a,a,\ldots, a) \in \mathbb{F}^{i+1} $, for $ i \in \mathbb{N} $. Then $ a $ is a zero of $ F $ of multiplicity $ r \in \mathbb{Z}_+ $ (according to Definition \ref{def multiplicity I}) if, and only if, $ F^{(0)}(a) = F^{(1)}(a) = \ldots = F^{(r-1)}(a) = 0 $.
\end{proposition}
We now use the general framework from Section \ref{sec high-degree zeros} for the type of multiplicities from Definition \ref{def multiplicity I}. To this end, we consider the universal set $ \mathcal{U}_1 $ from (\ref{eq universal set I}).
By Theorems \ref{th upper bound multiplicities by degree general} and \ref{th lagrange interpolation}, we see that bounding the number of zeros with multiplicity and Hermite interpolation are equivalent to considering P-independent subsets of $ \mathcal{U}_1 $. Therefore, the real challenge is to identify such P-independent subsets. A first result is the particular case of monotonicity (Corollary \ref{cor monotonicity}) applied to $ \mathcal{U}_1 $.
\begin{proposition} [\textbf{Monotonicity}] \label{prop monotonicity multiplicity I}
Let $ \Omega = \{ (x-a_1)^{r_1}, (x-a_2)^{r_2}, \ldots, (x-a_n)^{r_n} \} \subseteq \mathbb{F}[x;\sigma,\delta] $ be of size $ n \in \mathbb{Z}_+ $, where $ a_1,a_2,\ldots, a_n \in \mathbb{F} $ and $ r_1, r_2,\ldots, r_n \in \mathbb{Z}_+ $.
\begin{enumerate}
\item
If $ \Omega $ is P-independent, then so is $ \Psi = \{ (x-a_1)^{s_1}, (x-a_2)^{s_2}, \ldots, (x-a_n)^{s_n} \} $ $ \subseteq \mathbb{F}[x;\sigma,\delta] $, for all $ s_1, s_2, \ldots, s_n \in \mathbb{Z}_+ $ such that $ s_i \leq r_i $, for $ i = 1,2, \ldots, n $.
\item
If $ \Omega $ is P-independent, then the underlying set $ \Psi = \{ a_1, a_2, \ldots, a_n \} $ $ \subseteq \mathbb{F} $ has size $ n $ and is P-independent.
\end{enumerate}
\end{proposition}
However, the reversed implications do not hold, as the following example shows.
\begin{example} \label{ex power of linear term is no multiplicity}
Let $ q $ be a power of an odd prime number and let $ m \in \mathbb{N} $ be an even number with $ m \geq 2 $. Let $ \mathbb{F} = \mathbb{F}_{q^m} $ be the finite field with $ q^m $ elements, let $ \sigma : \mathbb{F}_{q^m} \longrightarrow \mathbb{F}_{q^m} $ be given by $ \sigma(a) = a^q $, for $ a \in \mathbb{F}_{q^m} $, and let $ \delta = 0 $. Let $ \gamma \in \mathbb{F}_{q^m}^* $ be a primitive element (a generator of the cyclic multiplicative group $ \mathbb{F}_{q^m}^* $ \cite[Th. 2.8]{lidl}) and consider
$$ a = \gamma^{ \frac{1}{2} \cdot \frac{q^m - 1}{q-1}} \quad \textrm{and} \quad b = - \gamma^{ \frac{1}{2} \cdot \frac{q^m - 1}{q-1}} . $$
Since $ q $ is odd and $ m $ is even, $ 2 $ divides $ (q^m-1)/(q-1) = 1 + q + q^2 + \cdots + q^{m-1} $, thus $ a $ and $ b $ are well defined, and since $ q $ is odd, it holds that $ a \neq b $. We have
$$ (x-a)^2 = (x-b)^2 \in \mathbb{F}_{q^m}[x;\sigma,\delta]. $$
Thus the doubleton set $ \Psi = \{ a,b \} \subseteq \mathbb{F}_{q^m} $ is P-independent, but the doubleton set $ \Omega = \{ (x-a)^2, (x-b)^3 \} \subseteq \mathbb{F}_{q^m}[x;\sigma,\delta] $ is not P-independent
This example also shows that the minimal skew polynomial of $ \Psi = \{ a,b \} $ is $ F_\Psi = (x-a)^2 = (x-b)^2 $, which has zeros of multiplicity $ 2 $, that is, its zeros are not all simple.
This is according to Definition \ref{def multiplicity I}. Multiplicities as in Definition \ref{def multiplicity II} in the following section do not suffer from such pathologies (see Theorem \ref{th P-independence multiplicity seqs}).
\end{example}
However, a union theorem \cite[Th. 22]{lam} still holds in general for multiplicities as in Definition \ref{def multiplicity I}. To that end, we need some preliminary definitions and lemmas. The first of these is the equivalence relation given by \textit{conjugacy} \cite[Eq. (2.5)]{lam-leroy}.
\begin{definition} [\textbf{Conjugacy \cite{lam-leroy}}] \label{def conjugate}
Given $ a,b \in \mathbb{F} $, we say that $ b $ is a $ (\sigma,\delta) $-conjugate (or simply conjugate) of $ a $ if there exists $ \beta \in \mathbb{F}^* $ such that $ b = a^\beta $, where
$$ a^\beta = \sigma(\beta) a \beta^{-1} + \delta(\beta) \beta^{-1}. $$
The conjugacy class of $ a $ is defined as $ C^{\sigma,\delta}(a) = \left\lbrace a^\beta \mid \beta \in \mathbb{F}^* \right\rbrace $.
\end{definition}
Conjugacy allows us to introduce the powerful product rule \cite[Th. 2.7]{lam-leroy}.
\begin{lemma} [\textbf{Product rule \cite{lam-leroy}}] \label{lemma product rule}
Let $ F,G \in \mathbb{F}[x;\sigma,\delta] $ and let $ a \in \mathbb{F} $. If $ G(a) = 0 $, then $ (FG)(a) = 0 $. On the other hand, if $ \beta = G(a) \neq 0 $, then
$$ (FG)(a) = F (a^\beta) G(a). $$
\end{lemma}
Finally, we need the following auxiliary lemma.
\begin{lemma} \label{lemma min skew pol expression for union}
Assume that $ \Omega = \{ (x-a_1)^{r_1}, (x-a_2)^{r_2}, \ldots, (x-a_n)^{r_n} \} \subseteq \mathbb{F}[x;\sigma,\delta] $ is P-independent of size $ n \in \mathbb{Z}_+ $, where $ a_1,a_2,\ldots, a_n \in \mathbb{F} $ and $ r_1, r_2,\ldots, r_n \in \mathbb{Z}_+ $. Then there exist elements $ \beta_{i,j} \in \mathbb{F}^* $, for $ j = 1,2, \ldots, r_i $ and $ i = 1,2, \ldots, n $, such that, if $ \Omega_{i,j} = \{ (x-a_1)^{r_1}, (x-a_2)^{r_2}, \ldots, (x-a_{i-1})^{r_{i-1}}, (x-a_i)^{j} \} $ (thus $ \Omega_{n,r_n} = \Omega $), then
\begin{equation}
F_{\Omega_{i,j}} = \left( x-a_i^{\beta_{i,j}} \right) \cdots \left(x-a_i^{\beta_{i,1}} \right) \cdots \left(x-a_1^{\beta_{1,r_1}} \right) \cdots \left(x-a_1^{\beta_{1,1}} \right) ,
\label{eq betas for min skew pol of multiplicities condition}
\end{equation}
for $ j = 1,2, \ldots, r_i $ and $ i = 1,2, \ldots, n $.
\end{lemma}
\begin{proof}
We prove the lemma by induction on $ r = r_1 + r_2 + \cdots + r_n $. First, the case $ n = r_1 = r= 1 $ is trivial. Assume now that $ r \geq 2 $. We will assume for convenience that $ r_n \geq 2 $, but the case $ r_n = 1 $ (still with $ r \geq 2 $) is proven analogously.
Define $ \Psi = \{ (x-a_1)^{r_1}, (x-a_2)^{r_2}, \ldots, (x-a_n)^{r_n - 1} \} $, which is P-independent by Corollary \ref{cor monotonicity} since $ \Psi \leq \Omega $. Furthermore, $ \Psi $ has size $ n $ since $ \Omega $ is P-independent. By induction hypothesis, there exist $ \beta_{i,j} \in \mathbb{F}[x;\sigma,\delta] $ satisfying Equation (\ref{eq betas for min skew pol of multiplicities condition}), for $ j = 1,2, \ldots, r_i $ if $ 1 \leq i \leq n-1 $, and for $ j = 1,2, \ldots, r_n-1 $ if $ i = n $. In particular,
$$ F_\Psi = \left( x-a_n^{\beta_{n,r_n-1}} \right) \cdots \left(x-a_n^{\beta_{n,1}} \right) \cdots \left(x-a_1^{\beta_{1,r_1}} \right) \cdots \left(x-a_1^{\beta_{1,1}} \right). $$
Since $ \Psi \leq \Omega $, then $ F_\Psi $ divides $ F_\Omega $ on the right. Furthermore, by counting degrees, since $ \Omega $ is P-independent, there exists $ b \in \mathbb{F} $ with $ F_\Omega = (x-b) F_\Psi $. Now, by definition of $ F_\Psi $ and since $ \Omega $ is P-independent, we have that $ (x-a_n)^{r_n-1} $ divides $ F_\Psi $ on the right, but $ (x-a_n)^{r_n} $ does not. That is, there exists $ G \in \mathbb{F}[x;\sigma,\delta] $ such that
$$ F_\Psi = G (x-a_n)^{r_n-1} \quad \textrm{but} \quad G(a_n) \neq 0. $$
Moreover, since $ (x-a_n)^{r_n} $ divides $ F_\Omega $ on the right, there exists a skew polynomial $ H \in \mathbb{F}[x;\sigma,\delta] $ such that $ F_\Omega = (x-b) G (x-a_n)^{r_n-1} = H (x-a_n)^{r_n} $. Thus, we deduce that $ (x-b) G = H (x-a_n) $. In particular, $ a_n $ is a zero of $ (x-b) G $ but not of $ G $. By the product rule (Lemma \ref{lemma product rule}), we deduce that $ G(a_n) \neq 0 $ and $ a_n^{G(a_n)} = b $. Therefore, we conclude by choosing $ \beta_{n,r_n} = G(a_n) $.
\end{proof}
The following theorem extends \cite[Th. 22]{lam} to multiplicities as in Definition \ref{def multiplicity I}.
\begin{theorem} [\textbf{Union}] \label{th union}
Let $ \Omega = \{ (x-a_1)^{r_1}, \ldots, (x-a_n)^{r_n} \} $ and $ \Psi = \{ (x-b_1)^{s_1}, \ldots, (x-b_m)^{s_m} \} $ be P-independent sets of sizes $ n,m \in \mathbb{Z}_+ $, respectively, and where $ a_1,\ldots, a_n, b_1, $ $ \ldots, $ $ b_m $ $ \in \mathbb{F} $ and $ r_1, \ldots, r_n, s_1, \ldots, s_m \in \mathbb{Z}_+ $. If no element in $ \{ a_1, a_2, \ldots, a_n \} $ is conjugate to an element in $ \{ b_1, b_2, \ldots, b_m\} $, then $ \Omega \cup \Psi $ has size $ n+m $ and is P-independent.
\end{theorem}
\begin{proof}
We proceed by induction on $ t = \sum_{i=1}^n r_i + \sum_{j=1}^m s_j $. Since $ n \geq 1 $ and $ m \geq 1 $, then $ t \geq 2 $. Observe that the case $ r_1 = \ldots = r_n = s_1 = \ldots = s_m = 1 $ is precisely Lam's union theorem \cite[Th. 22]{lam}. Hence the basis step $ t = 2 $ follows (since, in that case, $ n = m = r_1 = s_1 = 1 $), and for the inductive step, we may assume without loss of generality that $ s_m \geq 2 $. In other words, we may define
$$ \Psi^\prime = \{ (x-b_1)^{s_1}, (x-b_2)^{s_2}, \ldots, (x-b_m)^{s_m -1} \}, $$
which is P-independent by Corollary \ref{cor monotonicity} since $ \Psi^\prime \leq \Psi $. By induction hypothesis, $ \Omega \cup \Psi^\prime $ is P-independent. We proceed now by contradiction and assume that $ \Omega \cup \Psi $ is not P-independent. Since $ \Omega \cup \Psi^\prime \leq \Omega \cup \Psi $, then $ F_{\Omega \cup \Psi^\prime} $ divides $ F_{\Omega \cup \Psi} $ on the right. By counting degrees, we conclude that $ F_{\Omega \cup \Psi^\prime} = F_{\Omega \cup \Psi} $. By Lemma \ref{lemma min skew pol expression for union}, there exist elements $ \beta_{i,j} \in \mathbb{F}^* $, for $ j = 1,2, \ldots, r_i $ and $ i = 1,2, \ldots, n $, such that
$$ F_{\Omega \cup \Psi} = F_{\Omega \cup \Psi^\prime} = \left( x-a_n^{\beta_{n,r_n}} \right) \cdots \left(x-a_n^{\beta_{n,1}} \right) \cdots \left(x-a_1^{\beta_{1,r_1}} \right) \cdots \left(x-a_1^{\beta_{1,1}} \right) F_{\Psi^\prime}. $$
Since $ \Psi $ is P-independent, then $ (x-b_m)^{s_m-1} $ divides $ F_{\Psi^\prime} $ on the right, but $ (x-b_m)^{s_m} $ does not. Hence there exists a skew polynomial $ G \in \mathbb{F}[x;\sigma,\delta] $ such that $ F_{\Psi^\prime} = G (x-b_m)^{s_m-1} $ and $ G(b_m) \neq 0 $. Thus we deduce that $ b_m $ is a zero of the skew polynomial
$$ \left( x-a_n^{\beta_{n,r_n}} \right) \cdots \left(x-a_n^{\beta_{n,1}} \right) \cdots \left(x-a_1^{\beta_{1,r_1}} \right) \cdots \left(x-a_1^{\beta_{1,1}} \right) G, $$
while $ G(b_m) \neq 0 $. By the product rule (Lemma \ref{lemma product rule}), we deduce that there exists $ \gamma \in \mathbb{F}^* $ and indices $ j = 1,2, \ldots, r_i $ and $ i = 1,2, \ldots, n $, such that $ b_m^\gamma = a_i^{\beta_{i,j}} $. This contradicts the hypothesis that no element in $ \{ a_1, a_2, \ldots, a_n \} $ is conjugate to an element in $ \{ b_1, b_2, \ldots, b_m\} $, and we are done.
\end{proof}
We may now extend the degree bound and Hermite interpolation. First we combine Theorem \ref{th union} with the fact that singleton sets are always P-independent.
\begin{corollary} \label{cor non-conj multiple points are P-indep}
Let $ \Omega = \{ (x-a_1)^{r_1}, (x-a_2)^{r_2}, \ldots, (x-a_n)^{r_n} \} \subseteq \mathbb{F}[x;\sigma,\delta] $, where $ a_1,a_2,\ldots, a_n \in \mathbb{F} $ and $ r_1, r_2,\ldots, r_n \in \mathbb{Z}_+ $. If $ a_1, a_2, \ldots, a_n $ are pair-wise non-conjugate, then $ \Omega $ has size $ n $ and is P-independent.
\end{corollary}
The next result combines Corollary \ref{cor non-conj multiple points are P-indep} with Theorem \ref{th upper bound multiplicities by degree general}.
\begin{theorem} [\textbf{Degree bound I}] \label{th deg bound for nonconj naive case}
Let $ a_1, a_2, \ldots, a_n \in \mathbb{F} $ be pair-wise non-conjugate. If $ F \in \mathbb{F}[x;\sigma,\delta] $ is not zero and has $ a_i $ as a zero of multiplicity $ r_i \in \mathbb{Z}_+ $ (according to Definition \ref{def multiplicity I}), for $ i = 1,2, \ldots, n $, then it holds that
$$ \sum_{i=1}^n r_i \leq \deg(F). $$
Furthermore, equality holds if, and only if, $ F = a F_\Omega $, where $ a \in \mathbb{F}^* $ and $ F_\Omega $ is the least left common multiple of $ (x-a_1)^{r_1} , (x-a_2)^{r_2}, \ldots, (x-a_n)^{r_n} $.
\end{theorem}
The next results follows from Corollary \ref{cor non-conj multiple points are P-indep} with Theorem \ref{th lagrange interpolation}.
\begin{theorem} [\textbf{Hermite interpolation I}] \label{th hermite interpolation nonconj naive case}
Let $ a_1, a_2, \ldots, a_n \in \mathbb{F} $ be pair-wise non-conjugate. Let the notation of Hasse derivatives be as in Proposition \ref{prop charact multiplicities derivatives I}. For all $ b_{i,j} \in \mathbb{F} $, for $ j = 1,2,\ldots, r_i $ and $ i = 1,2, \ldots, n $, there exists a unique $ F \in \mathbb{F}[x;\sigma,\delta] $ with $ \deg(F) < \sum_{i=1}^n r_i $ such that $ F^{(j-1)}(a_i) = b_{i,j} $, for $ j = 1,2,\ldots, r_i $ and $ i = 1,2, \ldots, n $. In particular, if $ N = \sum_{i=1}^n r_i $ and $ \mathbf{a}_i = (a_i, a_i, \ldots, a_i) \in \mathbb{F}^{r_i} $, for $ i = 1,2, \ldots, n $, then the confluent Vandermonde matrix $ V_N(\mathbf{a}_1, \mathbf{a}_2, \ldots, \mathbf{a}_n) $ from Definition \ref{def confluent vandermonde} is invertible.
\end{theorem}
\section{Multiplicities from linear products with a single zero} \label{sec multi from linear terms}
In this section, we turn to a different notion of multiplicities, which allows a set of multiple zeros to be P-independent if, and only if, the underlying set of zeros is P-independent (see Theorem \ref{th P-independence multiplicity seqs}). The following definition is inspired by Bolotnikov's works on \textit{spherical chains} of quaternions (see \cite[Lemma 4.1]{bolotnikov-zeros} or \cite[Th. 3.1]{bolotnikov-conf}).
\begin{definition} [\textbf{Multiplicity sequences \cite{bolotnikov-zeros}}] \label{def multiplicity seq}
For $ r \in \mathbb{Z}_+ $, we say that $ \mathbf{a} = (a_1, $ $ a_2, \ldots, $ $ a_r) \in \mathbb{F}^r $ is a $ (\sigma,\delta) $-multiplicity sequence if $ a_1 $ is the only zero o
$$ P_{\mathbf{a}} = (x-a_r) \cdots (x-a_2) (x - a_1) \in \mathbb{F}[x;\sigma,\delta] . $$
\end{definition}
In contrast with the commutative case, a $ (\sigma,\delta) $-multiplicity sequence does not need to satisfy $ a_1 = a_2 = \ldots = a_r $. This was first noted by Leroy in \cite[Ex. 1.15.3]{leroy-noncommutative}. In Section \ref{sec multiplicity seqs}, we will provide sufficient and necessary conditions for a sequence to be a $ (\sigma,\delta) $-multiplicity sequence. In particular, the existence and (lack of) uniqueness of $ (\sigma,\delta) $-multiplicity sequences will be discussed in Corollary \ref{cor existence multiplicity seqs} and Section \ref{sec particular cases}.
One of the key properties of $ (\sigma,\delta) $-multiplicity sequences is their monotonicity.
\begin{proposition} [\textbf{Monotonicity}] \label{prop monotonicity multiplicity seqs}
If $ (a_1, a_2, \ldots, a_r) \in \mathbb{F}^r $ is a $ (\sigma,\delta) $-multiplicity sequence, then so are the sequences $ (a_1, a_2, \ldots, a_i) \in \mathbb{F}^i $, for $ i = 1,2, \ldots, r-1 $.
\end{proposition}
From now on, we only consider the following definition of multiplicity, which extends Bolotnikov's definition for quaternionic polynomials \cite{bolotnikov-zeros}.
\begin{definition} [\textbf{Multiplicity II \cite{bolotnikov-zeros}}] \label{def multiplicity II}
Given $ F \in \mathbb{F}[x;\sigma,\delta] $, $ r \in \mathbb{Z}_+ $, and a $ (\sigma,\delta) $-multiplicity sequence $ \mathbf{a} = (a_1, a_2, \ldots, a_r) \in \mathbb{F}^r $, we say that $ a_1 $ is a zero of $ F $ of multiplicity $ r $ via $ \mathbf{a} $ if the skew polynomial $ P_{\mathbf{a}} $ in (\ref{eq def multiplicity seq polynomial}) divides $ F $ on the right (i.e, $ F(P_\mathbf{a}) = 0 $ according to Definition \ref{def evaluation high degree}). We say that $ a \in \mathbb{F} $ is a zero of multiplicity $ r $ of $ F $ if it is a zero of multiplicity $ r $ of $ F $ via some $ (\sigma,\delta) $-multiplicity sequence starting from $ a $.
\end{definition}
Similarly to Proposition \ref{prop charact multiplicities derivatives I}, we have the following characterization of multiplicities based on Hasse derivatives (Definition \ref{def derivative}), which follows from Proposition \ref{prop taylor expansion}.
\begin{proposition} [\textbf{Derivative criterion II}] \label{prop charact multiplicities derivatives II}
Given $ F \in \mathbb{F}[x;\sigma,\delta] $, $ r \in \mathbb{Z}_+ $ and a $ (\sigma,\delta) $-multiplicity sequence $ \mathbf{a} = (a_1, a_2, \ldots, a_r) \in \mathbb{F}^r $, it holds that $ a_1 $ is a zero of $ F $ of multiplicity $ r $ via $ \mathbf{a} $ if, and only if, $ D_{\mathbf{a}_1} (F) = D_{\mathbf{a}_2} (F) = \ldots = D_{\mathbf{a}_r} (F) = 0 $, where $ \mathbf{a}_i = (a_1, a_2, \ldots, a_i) \in \mathbb{F}^i $, for $ i = 1,2, \ldots, r $.
\end{proposition}
We may also deduce the following monotonicity property.
\begin{proposition} [\textbf{Monotonicity}]
Let $ F \in \mathbb{F}[x;\sigma,\delta] $, and let $ \mathbf{a} = (a_1, a_2, \ldots, a_r) \in \mathbb{F}^r $ be a $ (\sigma,\delta) $-multiplicity sequence. If $ a_1 $ is a zero of $ F $ of multiplicity $ r $ via $ \mathbf{a} $, then it is also a zero of $ F $ of multiplicity $ i $ via $ \mathbf{a}_i = (a_1, a_2, \ldots, a_i) $, for $ i = 1,2, \ldots, r $.
\end{proposition}
We now use the general framework from Section \ref{sec high-degree zeros} for the type of multiplicities from Definition \ref{def multiplicity II}. To this end, we use in this section the universal set $ \mathcal{U}_2 $ from (\ref{eq universal set II}).
We will also use the concept of \textit{centralizers}, introduced in \cite[Eq. (3.1)]{lam-leroy}.
\begin{definition} [\textbf{Centralizer \cite{lam-leroy}}] \label{def centralizer}
We define the $ (\sigma,\delta) $-centralizer of $ a \in \mathbb{F} $ (or simply centralizer) as the division subring of $ \mathbb{F} $ given by $ K^{\sigma,\delta}_a = \{ \beta \in \mathbb{F} \mid \sigma(\beta) a + \delta(\beta) = a \beta \} $.
\end{definition}
The following characterization is the main result of this section.
\begin{theorem} \label{th P-independence multiplicity seqs}
Let $ \mathbf{a}_i = (a_{i,1}, a_{i,2}, \ldots, a_{i,r_i}) \in \mathbb{F}^{r_i} $ be a $ (\sigma,\delta) $-multiplicity sequence, for $ i = 1,2, \ldots, n $. Assume that $ a_{1,1}, a_{2,1}, \ldots, a_{n,1} $ are distinct. Let also $ a_1, a_2, \ldots , a_\ell \in \mathbb{F} $ be pair-wise non-conjugate elements, $ n_i \in \mathbb{Z}_+ $ and $ \beta_{i,1}, \beta_{i,2}, \ldots, \beta_{i,n_i} \in \mathbb{F}^* $ such that
$$ \{ a_{1,1}, a_{2,1}, \ldots, a_{n,1} \} = \bigcup_{i=1}^\ell \left\lbrace a_i^{\beta_{i,1}}, a_i^{\beta_{i,2}}, \ldots , a_i^{\beta_{i,n_i}} \right\rbrace , $$
for $ i = 1,2, \ldots, \ell $, and $ n = n_1 + n_2 + \cdots + n_\ell $. The following are equivalent:
\begin{enumerate}
\item
$ \Omega = \{ P_{\mathbf{a}_1}, P_{\mathbf{a}_2}, \ldots, P_{\mathbf{a}_n} \} \subseteq \mathcal{U}_2 $ is P-independent.
\item
$ \Psi = \{ a_{1,1}, a_{2,1}, \ldots, a_{n,1} \} \subseteq \mathbb{F} $ is P-independent.
\item
$ \beta_{i,1}, \beta_{i,2}, \ldots , \beta_{i,n_i} \in \mathbb{F}^* $ are right $ K^{\sigma,\delta}_{a_i} $-linearly independent, for $ i = 1,2, \ldots, \ell $.
\end{enumerate}
\end{theorem}
\begin{proof}
The equivalence between items 2 and 3 is due to Lam and Leroy (see \cite[Cor. 21]{lam} and \cite[Th. 4.5]{lam-leroy}). We now prove the equivalence between items 1 and 2. First, if $ \Omega $ is P-independent, then so is $ \Psi $ by Corollary \ref{cor monotonicity}. Hence, we only need to prove the reversed implication. We proceed by induction on $ \sum_{i=1}^n r_i $, being the case $ n = r_1 = 1 $ trivial. Next, $ \Omega $ and $ \Psi $ are always P-independent if $ n = 1 $. Furthermore, in the case $ r_1 = r_2 = \ldots = r_n = 1 $, it holds by definition that $ \Omega $ is P-independent if, and only if, so is $ \Psi $. Therefore, we may assume without loss of generality that $ n \geq 2 $ and $ r_n \geq 2 $.
Since $ r_n \geq 2 $, we may define $ \mathbf{a}_n^\prime = (a_{n,1}, a_{n,2}, \ldots, a_{n,r_n-1}) \in \mathbb{F}^{r_n-1} $, which is a $ (\sigma,\delta) $-multiplicity sequence by Proposition \ref{prop monotonicity multiplicity seqs}. Define $ \Omega^\prime = \{ P_{\mathbf{a}_1}, P_{\mathbf{a}_2}, \ldots, P_{\mathbf{a}_{n-1}}, P_{\mathbf{a}_n^\prime} \} $. Since $ n \geq 2 $, we may also define $ \Phi = \{ P_{\mathbf{a}_1}, P_{\mathbf{a}_2}, \ldots, P_{\mathbf{a}_{n-1}} \} $. By induction hypothesis, both $ \Omega^\prime $ and $ \Phi $ are P-independent, which means by Definition \ref{def P-independence} that
$$ \deg(F_{\Omega^\prime}) = \sum_{i=1}^n r_i - 1 \quad \textrm{and} \quad \deg(F_{\Phi}) = \sum_{i=1}^{n-1} r_i. $$
Assume now that $ \Omega $ is not P-independent. Since $ \Omega^\prime \leq \Omega $, then $ F_{\Omega^\prime} $ divides $ F_\Omega $ on the right. By Proposition \ref{prop deg of min skew pol upper bound by sum}, we deduce that $ F_{\Omega^\prime} = F_\Omega $. In other words,
$$ \deg(F_\Omega) = \sum_{i=1}^n r_i - 1 = \deg(F_\Phi) + \deg(P_{\mathbf{a}_n}) - 1. $$
Observe that, by definition, $ F_\Omega $ is the least left common multiple of $ F_\Phi $ and $ P_{\mathbf{a}_n} $. By Lemma \ref{lemma ores cross cut formula}, we deduce that the greatest right common divisor of $ F_\Phi $ and $ P_{\mathbf{a}_n} $ has degree exactly $ 1 $, that is, it is of the form $ D = x - b \in \mathbb{F}[x;\sigma,\delta] $, for some $ b \in \mathbb{F} $. Now, since $ D $ divides $ P_{\mathbf{a}_n} $ on the right and $ \mathbf{a}_n $ is a $ (\sigma,\delta) $-multiplicity sequence, we conclude that $ b = a_{n,1} $. However, this also means that $ D $ divides $ P_{\mathbf{a}_n^\prime} $ on the right. In other words, $ D $ is a right common divisor of both $ P_{\mathbf{a}_n^\prime} $ and $ F_\Phi $. However, the least left common multiple of $ P_{\mathbf{a}_n^\prime} $ and $ F_\Phi $ is $ F_{\Omega^\prime} $ by definition. Thus, using again Lemma \ref{lemma ores cross cut formula}, we deduce that
$$ \deg(F_{\Omega^\prime}) \leq \deg(P_{\mathbf{a}_n^\prime}) + \deg(F_\Phi) - 1 = \sum_{i=1}^n r_i - 2, $$
which contradicts the fact that $ \Omega^\prime $ is P-independent, and we are done.
\end{proof}
We now derive the following particular case of Theorem \ref{th upper bound multiplicities by degree general} from Theorem \ref{th P-independence multiplicity seqs}.
\begin{theorem} [\textbf{Degree bound II}] \label{th upper bound multiplicities by degree II}
Let $ \Psi = \{a_1, a_2, \ldots, a_n \} \subseteq \mathbb{F} $ be a set of size $ n \in \mathbb{Z}_+ $ and let $ r_1, r_2, \ldots, r_n \in \mathbb{Z}_+ $. Then $ \Psi $ is P-independent if, and only if, for every non-zero $ F \in \mathbb{F}[x;\sigma,\delta] $ that has $ a_i $ as a zero of multiplicity $ r_i $, for $ i = 1,2, \ldots, n $, it holds that
$$ \sum_{i=1}^n r_i \leq \deg(F). $$
Furthermore, when this is the case, then equality holds for $ F $ if, and only if, $ F = a F_\Omega $, where $ a \in \mathbb{F}^* $, $ F_\Omega $ is the least left common multiple of $ P_{\mathbf{a}_1} , P_{\mathbf{a}_2} ,\ldots, P_{\mathbf{a}_n} $, and $ \mathbf{a}_i \in \mathbb{F}^{r_i} $ is a $ (\sigma,\delta) $-multiplicity sequence starting at $ a_i $, for $ i = 1,2, \ldots, n $.
\end{theorem}
We may analogously rephrase Theorem \ref{th lagrange interpolation} as a Hermite interpolation theorem.
\begin{theorem} [\textbf{Hermite interpolation II}] \label{th hermite interpolation II}
Given a set $ \Psi = \{a_1, a_2, \ldots, a_n \} \subseteq \mathbb{F} $ of size $ n \in \mathbb{Z}_+ $ and given $ r_1, r_2, \ldots, r_n \in \mathbb{Z}_+ $, the following are equivalent:
\begin{enumerate}
\item
$ \Psi $ is P-independent.
\item
For $ b_{i,j} \in \mathbb{F} $ and $ (\sigma,\delta) $-multiplicity sequences $ \mathbf{a}_{i,j} = (a_{i,1}, a_{i,2}, $ $ \ldots, a_{i,j}) $ with $ a_{i,1} = a_i $, for $ j = 1,2, \ldots, r_i $ and $ i = 1,2, \ldots, n $, there exists a unique $ F \in \mathbb{F}[x;\sigma,\delta] $ with $ \deg(F) < \sum_{i=1}^n r_i $ such that $ D_{\mathbf{a}_{i,j}}(F) = b_{i,j} $, for $ j = 1,2,\ldots, r_i $ and $ i = 1,2, \ldots, n $.
\item
For all $ (\sigma,\delta) $-multiplicity sequences $ \mathbf{a}_i = (a_{i,1} , a_{i,2}, \ldots, a_{i,r_i}) \in \mathbb{F}^{r_i} $ such that $ a_{i,1} = a_i $, for $ i = 1,2, \ldots, n $, and setting $ N = \sum_{i=1}^n r_i $, the $ N \times N $ confluent Vandermonde matrix $ V_N(\mathbf{a}_1, \mathbf{a}_2, \ldots, \mathbf{a}_n) $ from Definition \ref{def confluent vandermonde} is invertible.
\end{enumerate}
\end{theorem}
\section{Multiplicity sequences in algebraic conjugacy classes} \label{sec multiplicity seqs}
In this section, we characterize $ (\sigma,\delta) $-multiplicity sequences (Definition \ref{def multiplicity seq}) over algebraic conjugacy classes assuming that $ \sigma $ is surjective, see Theorem \ref{th charact multiplicity seqs in alg conj}. Then, in Theorem \ref{th min skew pol of conjugacy class with mult}, we study what happens when all points in a single conjugacy class are zeros of the same multiplicity of a given skew polynomial.
We start with some auxiliary lemmas. The following result is \cite[Cor. 6.3]{lam-leroy-wedI}.
\begin{lemma} [\textbf{\cite{lam-leroy-wedI}}] \label{lemma algebraic conj class 1}
Let $ a \in \mathbb{F} $ and assume that its conjugacy class $ C^{\sigma,\delta}(a) \subseteq \mathbb{F} $ is algebraic, that is, $ I(C^{\sigma,\delta}(a)) \neq \{ 0 \} $ (Definition \ref{def algebraic sets}). If a skew polynomial $ F \in \mathbb{F}[x;\sigma,\delta] $ has $ a $ as a left zero (that is, $ x-a $ divides $ F $ on the left), then $ F $ has some element $ b \in C^{\sigma,\delta}(a) $ as a right zero (that is, $ x-b $ divides $ F $ on the right).
\end{lemma}
Another useful ingredient is the following characterization, which is \cite[Th. 5.10]{algebraic-conjugacy}.
\begin{lemma} [\textbf{\cite{algebraic-conjugacy}}] \label{lemma algebraic conj class 2}
For $ a \in \mathbb{F} $, the conjugacy class $ C^{\sigma,\delta}(a) \subseteq \mathbb{F} $ is algebraic if, and only if, $ \mathbb{F} $ has finite right dimension over $ K^{\sigma,\delta}_a $ (Definition \ref{def centralizer}). In such a case, $ C^{\sigma,\delta}(a) $ is P-closed and
$$ {\rm Rk}(C^{\sigma,\delta}(a)) = \dim^R_{K^{\sigma,\delta}_a}(\mathbb{F}) . $$
\end{lemma}
We will assume that $ \sigma $ is surjective due to the following result \cite[Eq. (19)]{ore}.
\begin{lemma} [\textbf{\cite{ore}}] \label{lemma ore left right isomorph}
If $ \sigma $ is surjective, then $ \mathbb{F}[x;\sigma,\delta] $ is a left Euclidean domain. In particular, if $ \sigma $ is surjective, then all definitions and results stated up to this point on the right hold in the same way on the left, and vice versa.
\end{lemma}
We will also use the notion of evaluation from \cite[Def. 3.1]{hilbert90} and \cite[Eq. (2.7)]{leroy-pol}
\begin{definition} [\textbf{\cite{hilbert90, leroy-pol}}] \label{def lin op pols}
Given $ a \in \mathbb{F} $, we define the right $ K^{\sigma,\delta}_a $-linear map $ \mathcal{D}^{\sigma,\delta}_a : \mathbb{F} \longrightarrow \mathbb{F} $ such that $ \mathcal{D}^{\sigma,\delta}_a(\beta) = \sigma(\beta) a + \delta(\beta) $, for $ \beta \in \mathbb{F} $. We will denote $ \mathcal{D}_a = \mathcal{D}^{\sigma,\delta}_a $ if there is no confusion about the pair $ (\sigma,\delta) $. For a skew polynomial $ F = F_d x^d + \cdots + F_1 x + F_0 \in \mathbb{F}[x;\sigma,\delta] $, we define $ F^{\mathcal{D}_a} = F_d \mathcal{D}_a^d + \cdots + F_1 \mathcal{D}_a + F_0 {\rm Id} \in \mathbb{F}[\mathcal{D}_a] $.
\end{definition}
We will also use the following formula from \cite[Th. 2.8]{leroy-pol} (see also \cite[Lemma 24]{linearizedRS}).
\begin{lemma} [\textbf{\cite{leroy-pol}}] \label{lemma connecting evaluations}
For $ F \in \mathbb{F}[x;\sigma,\delta] $, $ a \in \mathbb{F} $ and $ \beta \in \mathbb{F}^* $, we have
$$ F ( a^\beta ) = F^{\mathcal{D}_a}(\beta) \beta^{-1}. $$
\end{lemma}
The following result is our main characterization of $ (\sigma,\delta) $-multiplicity sequences. It extends \cite[Lemma 4.1]{bolotnikov-zeros} (see also \cite[Th. 3.1]{bolotnikov-conf}) from quaternionic polynomials to general skew polynomials.
\begin{theorem} \label{th charact multiplicity seqs in alg conj}
Assume that $ \sigma $ is surjective and all conjugacy classes are algebraic. Let $ \mathbf{a} = (a_1, a_2, $ $ \ldots, a_r) \in \mathbb{F}^r $, where $ r \geq 2 $. The following are equivalent.
\begin{enumerate}
\item
$ \mathbf{a} $ is a $ (\sigma,\delta) $-multiplicity sequence, that is, $ a_1 $ is the only (right) zero of $ P_{\mathbf{a}} $.
\item
If $ b_1, \ldots, b_r \in \mathbb{F} $ satisfy $ P_{\mathbf{a}} = (x-b_r) \cdots (x-b_1) $, then $ b_i = a_i $, for $ i = 1,2, \ldots, r $.
\item
The pair $ (a_i, a_{i+1}) $ is a $ (\sigma,\delta) $-multiplicity sequence, for $ i = 1,2, \ldots, r-1 $.
\item
For $ i = 1,2, \ldots, r-1 $, there exists $ \beta_i \in \mathbb{F}^* $ such that $ a_{i+1} = a_i^{\beta_i} $ and
$$ \beta_i \notin \left\lbrace \left. \left( a_i^\beta - a_i \right) \beta \right| \beta \in \mathbb{F} \right\rbrace = \left\lbrace \left. \mathcal{D}_{a_i}^{\sigma,\delta}(\beta) - a_i \beta \right| \beta \in \mathbb{F} \right\rbrace. $$
\item
$ x - a_r $ is the only linear skew polynomial that divides $ P_{\mathbf{a}} $ on the left.
\end{enumerate}
\end{theorem}
\begin{proof}
We prove the following implications separately:
$ 3. \Longrightarrow 1. $: We proceed by induction in $ r \geq 2 $, being the case $ r = 2 $ trivial. Assume that item 3 holds for some $ r \geq 3 $. By induction hypothesis, $ (a_1, a_2, \ldots, a_{r-1}) \in \mathbb{F}^{r-1} $ is a $ (\sigma,\delta) $-multiplicity sequence. Define $ P = (x-a_{r-2}) \cdots (x-a_2) (x - a_1) $ and $ Q = (x-a_r) (x-a_{r-1}) $. Hence $ a_1 $ is the only zero of $ P $ by Proposition \ref{prop monotonicity multiplicity seqs}. Assume that $ \mathbf{a} = (a_1, a_2, \ldots, a_r) \in \mathbb{F}^r $ is not a $ (\sigma,\delta) $-multiplicity sequence. In other words, there exists $ b \in \mathbb{F} $ such that $ b \neq a_1 $ and $ P_{\mathbf{a}}(b) = 0 $. By the product rule (Lemma \ref{lemma product rule}), it holds that
$$ 0 = P_{\mathbf{a}}(b) = Q \left( b^{P(b)} \right) P(b), $$
where $ P(b) \neq 0 $ since $ b \neq a_1 $. Therefore we deduce that $ Q \left( b^{P(b)} \right) = 0 $. Since $ (a_{r-1},a_r) $ is a $ (\sigma,\delta) $-multiplicity sequence, we have that $ b^{P(b)} = a_{r-1} $. However, in that case, we have that $ R(b) = 0 $ by the product rule (Lemma \ref{lemma product rule}), for
$$ R = (x-a_{r-1}) P = (x-a_{r-1}) \cdots (x-a_2) (x - a_1), $$
which is absurd since $ b \neq a_1 $ but $ (a_1, a_2, \ldots, a_{r-1}) \in \mathbb{F}^{r-1} $ is a $ (\sigma,\delta) $-multiplicity sequence. Thus it must hold that $ \mathbf{a} \in \mathbb{F}^r $ is a $ (\sigma,\delta) $-multiplicity sequence, and we are done.
$ 2. \Longleftrightarrow 3. $: It is trivial that item 2 implies item 3. Assume now that item 3 holds. Since item 3 implies item 1, we deduce that $ (a_i, a_{i+1}, \ldots, a_r) \in \mathbb{F}^{r-i+1} $ is a $ (\sigma,\delta) $-multiplicity sequence, for $ i = 1,2, \ldots, r $. Assume that $ P_{\mathbf{a}} = (x-b_r) \cdots (x-b_2)(x-b_1) $ for $ b_1, b_2, \ldots, b_r \in \mathbb{F} $. Let $ j \in \mathbb{N} $ be the maximum number such that $ 1 \leq j \leq r $ and $ b_i = a_i $, for $ i = 1,2, \ldots, j $. If item 2 does not hold, then $ j \leq r-1 $. In that case,
$$ (x - b_r) \cdots (x - b_{j+2}) (x - b_{j+1}) = (x - a_r) \cdots (x - a_{j+2}) (x - a_{j+1}). $$
However, since $ b_{j+1} \neq a_{j+1} $, then $ (a_{j+1}, a_{j+2}, \ldots, a_r) $ is not a $ (\sigma,\delta) $-multiplicity sequence. This is a contradiction, thus item 2 holds.
$ 1. \Longrightarrow 3. $: We proceed by induction on $ r \geq 2 $, being the case $ r = 2 $ trivial. Assume that item 1 holds for some $ r \geq 3 $. By induction hypothesis, $ (a_i, a_{i+1}) $ is a $ (\sigma,\delta) $-multiplicity sequence, for $ i = 1,2, \ldots, r-2 $. Assume however that $ (a_{r-1},a_r) $ is not a $ (\sigma,\delta) $-multiplicity sequence. By definition of minimal skew polynomial, we deduce that $ F_\Psi = (x-a_r) (x-a_{r-1}) $, where $ \Psi = \{ a_{r-1}, b \} \subseteq \mathbb{F} $ is P-independent and $ b \neq a_{r-1} $. Define now $ P = (x-a_{r-2}) \cdots (x-a_2) (x - a_1) $. Hence, we have that $ P_{\mathbf{a}} = F_\Psi P $.
By hypothesis and Proposition \ref{prop monotonicity multiplicity seqs}, $ (a_1, a_2, \ldots, a_{r-2}) $ is a $ (\sigma,\delta) $-multiplicity sequence, thus $ C^{\sigma,\delta}(a_1) \cap Z(P) = \{ a_1 \} $. Hence by Lemma \ref{lemma connecting evaluations}, $ {\rm Ker} \left( P^{\mathcal{D}_{a_1}} \right) = K^{\sigma,\delta}_{a_1} $. We also have that $ m = \dim^R_{K^{\sigma,\delta}_{a_1}} (\mathbb{F}) < \infty $ by Lemma \ref{lemma algebraic conj class 2}, since $ C^{\sigma,\delta}(a_1) $ is algebraic. Therefore
\begin{equation}
\dim^R_{K^{\sigma,\delta}_{a_1}} \left( {\rm Im} \left( P^{\mathcal{D}_{a_1}} \right) \right) = m - 1.
\label{eq proof equivalences multiplicity 2}
\end{equation}
Since $ C^{\sigma,\delta}(a_{r-1}) $ and $ C^{\sigma,\delta}(b) $ are algebraic and $ \mathbf{a} $ is a $ (\sigma,\delta) $-multiplicity sequence, we deduce from Proposition \ref{prop monotonicity multiplicity seqs} and Lemma \ref{lemma algebraic conj class 1} that there exist $ \alpha, \beta \in \mathbb{F}^* $ such that $ a_{r-1} = a_1^\alpha $ and $ b = a_1^\beta $. Next we prove that
\begin{equation}
{\rm Im} \left( P^{\mathcal{D}_{a_1}} \right) \cap \langle \alpha, \beta \rangle^R_{K^{\sigma,\delta}_{a_1}} = \{ 0 \} ,
\label{eq proof equivalences multiplicity 1}
\end{equation}
where $ \langle \alpha, \beta \rangle^R_{K^{\sigma,\delta}_{a_1}} $ denotes the right $ K^{\sigma,\delta}_{a_1} $-linear span of $ \alpha $ and $ \beta $. Let $ \gamma \in {\rm Im} \left( P^{\mathcal{D}_{a_1}} \right) \cap \langle \alpha , \beta \rangle^R_{K^{\sigma,\delta}_{a_1}} $. Assume that $ \gamma \neq 0 $. Since $ \gamma \in {\rm Im} \left( P^{\mathcal{D}_{a_1}} \right) $, there exists $ \eta \in \mathbb{F}^* $ such that $ \gamma = P^{\mathcal{D}_{a_1}}(\eta) $. Observe that, by Lemma \ref{lemma connecting evaluations}, we have that $ P \left( a_1^\eta \right) = P^{\mathcal{D}_{a_1}}(\eta) \eta^{-1} = \gamma \eta^{-1} \neq 0 $. Now, since $ \gamma \in \langle \alpha, \beta \rangle^R_{K^{\sigma,\delta}_{a_1}} $, we deduce by Theorem \ref{th P-independence multiplicity seqs} that $ a_1^\gamma $ is P-dependent on $ a_{r-1} $ and $ b $. In other words, we have that
$$ 0 = F_\Psi \left( a_1^\gamma \right) = F_\Psi \left( a_1^{P \left( a_1^\eta \right) \eta } \right). $$
Hence, by the product rule (Lemma \ref{lemma product rule}), it holds that
$$ P_{\mathbf{a}} \left( a_1^\eta \right) = F_\Psi \left( a_1^{P \left( a_1^\eta \right) \eta } \right) P \left( a_1^\eta \right) = 0. $$
Since $ \mathbf{a} $ is a $ (\sigma,\delta) $-multiplicity sequence, we deduce that $ a_1^\eta = a_1 $, that is, $ \eta \in K^{\sigma,\delta}_{a_1} $. However, since $ {\rm Ker} \left( P^{\mathcal{D}_{a_1}} \right) = K^{\sigma,\delta}_{a_1} $, we would have that $ \gamma = P^{\mathcal{D}_{a_1}}(\eta) = 0 $, a contradiction. Thus (\ref{eq proof equivalences multiplicity 1}) must hold. Finally, since $ \Psi = \{ a_{r-1}, b \} $ is P-independent, we conclude by Theorem \ref{th P-independence multiplicity seqs} that $ \dim^R_{K^{\sigma,\delta}_{a_1}} \left( \langle \alpha, \beta \rangle^R_{K^{\sigma,\delta}_{a_1}} \right) = 2 $. Combined with (\ref{eq proof equivalences multiplicity 2}) and (\ref{eq proof equivalences multiplicity 1}), we have
$$ \dim^R_{K^{\sigma,\delta}_{a_1}} \left( {\rm Im} \left( P^{\mathcal{D}_{a_1}} \right) + \langle \alpha, \beta \rangle^R_{K^{\sigma,\delta}_{a_1}} \right) = m+1. $$
This is absurd since $ \dim^R_{K^{\sigma,\delta}_{a_1}}(\mathbb{F}) = m < \infty $ by assumption. Therefore, $ (a_{r-1},a_r) $ must be a $ (\sigma,\delta) $-multiplicity sequence, and we are done.
$ 2. \Longleftrightarrow 5. $: It follows from Lemma \ref{lemma ore left right isomorph} and the equivalence between items 1 and 2, since item 2 can be read from left to right or from right to left.
$ 3. \Longleftrightarrow 4. $: First, if item 3 holds, then $ a_1, a_2, \ldots, a_r \in C^{\sigma,\delta}(a_1) $ by Lemma \ref{lemma algebraic conj class 1}, since all conjugacy classes are algebraic and $ (a_1, a_2, \ldots, a_i ) \in \mathbb{F}^i $ is a $ (\sigma,\delta) $-multiplicity sequence, for $ i = 1,2, \ldots, r $, by the fact that item 3 implies item 1. Thus, in both items 3 and 4 it holds that $ a_1, a_2, \ldots, a_r \in C^{\sigma,\delta}(a_1) $.
Now fix $ i = 1,2, \ldots, r-1 $ and let $ \beta_i, \beta \in \mathbb{F}^* $ be such that $ a_{i+1} = a_i^{\beta_i} $ and $ a_i^\beta \neq a_i $, which is equivalent to $ \beta \in \mathbb{F} \setminus K^{\sigma,\delta}_{a_i} $. By the product rule (Lemma \ref{lemma product rule}),
$$ P ( a_i^\beta ) = \left( a_i^{ \left( a_i^\beta - a_i \right) \beta } - a_i^{\beta_i} \right) \left( a_i^\beta - a_i \right) $$
for $ P = (x-a_{i+1})(x-a_i) $. Thus it holds that $ P ( a_i^\beta ) = 0 $ for some $ \beta \in \mathbb{F}^* $ such that $ a_i^\beta \neq a_i $ if, and only if, $ a_i^{ \left( a_i^\beta - a_i \right) \beta } = a_i^{\beta_i} $, which is equivalent to
$$ \beta_i \in \left\langle \left\lbrace \left. \left( a_i^\beta - a_i \right) \beta \right| \beta \in \mathbb{F} \setminus K^{\sigma,\delta}_{a_i} \right\rbrace \right\rangle^R_{K^{\sigma,\delta}_{a_i}}. $$
We are done by the definition of $ \mathcal{D}_{a_i}^{\sigma,\delta} $ (Definition \ref{def lin op pols}), and by noticing that
$$ \left\langle \left\lbrace \left. \left( a_i^\beta - a_i \right) \beta \right| \beta \in \mathbb{F} \setminus K^{\sigma,\delta}_{a_i} \right\rbrace \right\rangle^R_{K^{\sigma,\delta}_{a_i}} = \left\lbrace \left. \left( a_i^\beta - a_i \right) \beta \right| \beta \in \mathbb{F} \right\rbrace . $$
\end{proof}
\begin{remark} \label{remark all conj classes are algebraic}
The assumption that all conjugacy classes are algebraic may look inconvenient. However, it holds in many classical particular cases (see Section \ref{sec particular cases})
\end{remark}
A first consequence of Theorem \ref{th charact multiplicity seqs in alg conj} is that $ (\sigma,\delta) $-multiplicity sequences exist starting at any point of an algebraic conjugacy class. We will need the following lemma.
\begin{lemma} \label{lemma image of varphi is hyperplane}
If $ a \in \mathbb{F} $ and $ m = \dim^R_{K^{\sigma,\delta}_a}(\mathbb{F}) < \infty $, then $ \dim^R_{K^{\sigma,\delta}_a}(\mathcal{V}^{\sigma,\delta}_a) = m-1 $, where $ \mathcal{V}^{\sigma,\delta}_a = \{ \mathcal{D}^{\sigma,\delta}_a(\beta) - a \beta \mid \beta \in \mathbb{F} \} \subseteq \mathbb{F} $.
\end{lemma}
\begin{proof}
The map $ \varphi^{\sigma,\delta}_a : \mathbb{F} \longrightarrow \mathbb{F} $ given by $ \varphi^{\sigma,\delta}_a(\beta) = \mathcal{D}^{\sigma,\delta}_a(\beta) - a \beta $, for all $ \beta \in \mathbb{F} $, is right $ K^{\sigma,\delta}_a $-linear and satisfies that $ {\rm Ker} (\varphi^{\sigma,\delta}_a) = K^{\sigma,\delta}_a $ and $ {\rm Im}(\varphi^{\sigma,\delta}_a) = \mathcal{V}^{\sigma,\delta}_a $
\end{proof}
We now prove the existence of $ (\sigma,\delta) $-multiplicity sequences over algebraic conjugacy classes. This result follows from Lemma \ref{lemma algebraic conj class 2}, item 4 in Theorem \ref{th charact multiplicity seqs in alg conj} and Lemma \ref{lemma image of varphi is hyperplane}.
\begin{corollary} \label{cor existence multiplicity seqs}
For all $ a \in \mathbb{F} $ such that $ \dim^R_{K^{\sigma,\delta}_a}(\mathbb{F}) < \infty $, and for all $ r \in \mathbb{Z}_+ $, there exists a $ (\sigma,\delta) $-multiplicity sequence $ \mathbf{a} = (a_1, a_2, \ldots, a_r) \in \mathbb{F}^r $ such that $ a_1 = a $.
\end{corollary}
On the other hand, we do not have in general uniqueness of $ (\sigma,\delta) $-multiplicity sequences starting at a point in an algebraic conjugacy class. However, there is a unique way to consider multiplicities as in Definition \ref{def multiplicity II} over a full conjugacy class, as the following theorem shows. We will assume again that all conjugacy classes are algebraic. We also need an auxiliary lemma, obtained from \cite[Lemma 5.2]{algebraic-conjugacy} or \cite[Prop. 18]{multivariateskew}.
\begin{lemma} [\textbf{\cite{algebraic-conjugacy, multivariateskew}}] \label{lemma ideal of conjugacy twosided}
Given $ a \in \mathbb{F} $, the left ideal $ I(C^{\sigma,\delta}(a)) $ is two-sided.
\end{lemma}
\begin{theorem} \label{th min skew pol of conjugacy class with mult}
Let $ a \in \mathbb{F} $ and $ r \in \mathbb{Z}_+ $, and assume that all conjugacy classes are algebraic. Let $ a_1, a_2, \ldots, a_m \in C^{\sigma,\delta}(a) $ form a P-basis of $ C^{\sigma,\delta}(a) $. The following hold:
\begin{enumerate}
\item
If $ \mathbf{b} = (b_1, b_2, \ldots, b_r) \in \mathbb{F}^r $ is a $ (\sigma,\delta) $-multiplicity sequence such that $ b_1 \in C^{\sigma,\delta}(a) $, then $ P_\mathbf{b} $ divides $ F_{C^{\sigma,\delta}(a)}^r $ on the right.
\item
It holds that $ F_\Omega = F_{C^{\sigma,\delta}(a)}^r $, where $ \Omega = \{ P_{\mathbf{a}_1}, P_{\mathbf{a}_2}, \ldots, P_{\mathbf{a}_m} \} $ and $ \mathbf{a}_i = (a_{i,1}, a_{i,2}, \ldots, $ $ a_{i,r}) \in \mathbb{F}^r $ are $ (\sigma,\delta) $-multiplicity sequences with $ a_{i,1} = a_i $, for $ i = 1,2, \ldots, m $.
\item
The subset of $ \mathcal{U}_2 $ from Equation (\ref{eq universal set II}) given by
$$ C^{\sigma,\delta}_r(a) = \{ P_{\mathbf{b}} \mid \mathbf{b} \in \mathbb{F}^i, (\sigma,\delta) \textrm{-multiplicity seq.}, b_1 \in C^{\sigma,\delta}(a), 1 \leq i \leq r \} $$
is P-closed with minimal skew polynomial $ F_{C^{\sigma,\delta}(a)}^r $, P-basis $ \Omega $ and $ {\rm Rk}(C^{\sigma,\delta}_r(a)) = r \cdot {\rm Rk}(C^{\sigma,\delta}(a)) $. It may be seen as the conjugacy class of $ a $ with multiplicity $ r $.
\end{enumerate}
\end{theorem}
\begin{proof}
Denote $ P = F_{C^{\sigma,\delta}(a)} $ for simplicity. We start by proving item 1. We will show by induction in $ r \in \mathbb{Z}_+ $ that $ P_\mathbf{b} $ divides $ P^r $ on the right. The case $ r = 1 $ is trivial by definition. Consider now that $ r \geq 2 $. Since $ b_1 \in C^{\sigma,\delta}(a) $, there exists $ G \in \mathbb{F}[x;\sigma,\delta] $ such that $ P = G(x-b_1) $. By Lemma \ref{lemma ideal of conjugacy twosided}, there exists $ H \in \mathbb{F}[x;\sigma,\delta] $ such that
\begin{equation}
P^r = P^{r-1} G (x-b_1) = H P^{r-1} (x-b_1).
\label{eq proof multiplicities on full conjugacy}
\end{equation}
By Theorem \ref{th charact multiplicity seqs in alg conj}, $ \mathbf{b}^\prime = (b_2, b_3, \ldots, b_r) \in \mathbb{F}^{r-1} $ is also a $ (\sigma,\delta) $-multiplicity sequence such that $ b_2 \in C^{\sigma,\delta}(a) $. Thus, by induction hypothesis, $ P_{\mathbf{b}^\prime} $ divides $ P^{r-1} $ on the right. Hence $ P_\mathbf{b} = P_{\mathbf{b}^\prime} (x-b_1) $ divides $ P^r $ on the right by (\ref{eq proof multiplicities on full conjugacy}).
Now, we prove item 2. Since $ \Omega $ is P-independent by Theorem \ref{th P-independence multiplicity seqs}, we deduce that $ \deg(F_\Omega) = \deg(P^r) = mr $. Moreover, as $ P_{\mathbf{a}_i} $ divides $ P^r $ on the right, for $ i = 1,2, \ldots, m $, we conclude that $ F_\Omega $ divides $ P^r $ on the right. Since they have the same degree, then $ F_\Omega = P^r $ and we are done.
Finally, we prove item 3. With the results obtained above, combined with the product rule (Lemma \ref{lemma product rule}) and Proposition \ref{prop properties alg geom}, it is easy to see that $ I(C^{\sigma,\delta}_r(a)) = (F_{C^{\sigma,\delta}(a)}^r) $ and $ Z_{\mathcal{U}_2}(I(C^{\sigma,\delta}_r(a))) = Z_{\mathcal{U}_2}(\{ F_{C^{\sigma,\delta}(a)}^r \}) = C^{\sigma,\delta}_r(a) $, for $ \mathcal{U}_2 $ as in Equation (\ref{eq universal set II}). Therefore, $ C^{\sigma,\delta}_r(a) $ is P-closed with minimal skew polynomial $ F_{C^{\sigma,\delta}(a)}^r $. Furthermore, $ \Omega $ is P-independent by Theorem \ref{th P-independence multiplicity seqs}, and it is a set of P-generators of $ C^{\sigma,\delta}_r(a) $ by Lemma \ref{lemma P-generators iff min skew pols equal}, since $ F_\Omega = F_{C^{\sigma,\delta}(a)}^r $. Thus, $ \Omega $ is a P-basis of $ C^{\sigma,\delta}_r(a) $.
\end{proof}
It follows that, when considering a full conjugacy class, multiplicities as in Definition \ref{def multiplicity II} work as expected and do not depend on $ (\sigma,\delta) $-multiplicity sequences
\begin{corollary} \label{cor multiplicities on full conjugacy}
Let the notation and assumptions be as in Theorem \ref{th min skew pol of conjugacy class with mult}. Given $ F \in \mathbb{F}[x;\sigma,\delta] $, and with multiplicities as in Definition \ref{def multiplicity II}, the following are equivalent:
\begin{enumerate}
\item
Every element in $ C^{\sigma,\delta}(a) $ is a zero of $ F $ of multiplicity $ r $.
\item
Every element in a P-basis of $ C^{\sigma,\delta}(a) $ is a zero of $ F $ of multiplicity $ r $.
\item
$ F_{C^{\sigma,\delta}(a)}^r $ divides $ F $ (on the right, on the left or in between, by Lemma \ref{lemma ideal of conjugacy twosided}).
\end{enumerate}
\end{corollary}
It is natural to ask if we could have arrived at Corollary \ref{cor multiplicities on full conjugacy} from the notion of multiplicity in Definition \ref{def multiplicity I}. We now show that this is not the case.
\begin{example} \label{ex full conjugacy class pathology for naive mult}
Consider the setting of Example \ref{ex power of linear term is no multiplicity}, and further assume that $ m = 2 $. Using that $ a^{q-1} = -1 $, the reader may verify that $ a = \gamma^{ \frac{q+1}{2} } $ and $ b = - \gamma^{ \frac{q+1}{2} } $ are conjugate. Observe also that $ \{ a,b \} $ is P-independent, $ K^{\sigma,\delta}_a = \mathbb{F}_q $ and, by Lemma \ref{lemma algebraic conj class 2}, $ {\rm Rk}(C^{\sigma,\delta}(a)) = \dim_{\mathbb{F}_q}(\mathbb{F}_{q^m}) = m = 2 $. Therefore, $ \{ a,b \} $ forms a P-basis of $ C^{\sigma,\delta}(a) $. However,
$$ F_{C^{\sigma,\delta}(a)} = (x-a)^2 = (x-b)^2 = x^2 + \gamma^{q+1} \in \mathbb{F}_q[x^2] . $$
Hence $ F_{C^{\sigma,\delta}(a)} $ vanishes at a P-basis of $ C^{\sigma,\delta}(a) $ with multiplicity $ 2 $ according to Definition \ref{def multiplicity I}. Clearly $ F_{C^{\sigma,\delta}(a)}^2 $ does not divide $ F_{C^{\sigma,\delta}(a)} $, and $ \deg(F_{C^{\sigma,\delta}(a)}) = 2 $, hence Corollary \ref{cor multiplicities on full conjugacy} and Theorem \ref{th upper bound multiplicities by degree II} do not hold for multiplicities as in Definition \ref{def multiplicity I}.
\end{example}
\section{Particular cases and worked examples} \label{sec particular cases}
In this section, we study particular cases for which all conjugacy classes are algebraic, hence the results in the previous section hold. For worked examples, see \cite{skew-multi}.
We start with Bolotnikov's characterization of multiplicity sequences of quaternions \cite{bolotnikov-conf, bolotnikov-zeros} (called \textit{spherical chains} there). Quaternionic Lagrange interpolation has recently been motivated by the study of tight frames \cite{waldron-interpol, waldron-frames}. Denote by $ \mathbb{R} $ the real field. The division ring of quaternions is given by
$$ \mathbb{H} = \{ a + b \mathbf{i} + c \mathbf{j} + d \mathbf{k} \mid a,b,c,d \in \mathbb{R} \}, $$
where $ \mathbf{i} $, $ \mathbf{j} $ and $ \mathbf{k} $ commute with real numbers and $ \mathbf{i}^2 = \mathbf{j}^2 = \mathbf{k}^2 = \mathbf{i} \mathbf{j} \mathbf{k} = - 1 $. Given a quaternion $ q = a + b \mathbf{i} + c \mathbf{j} + d \mathbf{k} \in \mathbb{H} $, where $ a,b,c,d \in \mathbb{R} $, the real and imaginary parts of $ q $ are given, respectively, by $ {\rm Re}(q) = a $ and $ {\rm Im}(q) = b \mathbf{i} + c \mathbf{j} + d \mathbf{k} $. The quaternionic conjugate of $ q \in \mathbb{H} $ is given by $ \overline{q} = {\rm Re}(q) - {\rm Im}(q) \in \mathbb{H} $, whereas its quaternionic modulus is given by $ |q| = \sqrt{q \overline{q}} $. Two quaternions $ q, t \in \mathbb{H} $ are conjugate, in the sense of division rings, if there exists $ u \in \mathbb{H}^* $ such that $ t = u q u^{-1} $. This is actually equivalent to $ {\rm Re} (q) = {\rm Re} (t) $ and $ |q| = |t| $ (see \cite[p. 333]{brenner}). The following result is a consequence of Theorem \ref{th charact multiplicity seqs in alg conj} and was obtained in \cite[Lemma 4.1]{bolotnikov-zeros} (see also \cite[Th. 3.1]{bolotnikov-conf}).
\begin{corollary} [\textbf{Quaternions \cite{bolotnikov-zeros}}] \label{cor multiplicity over quaternions}
Let $ \mathbb{F} = \mathbb{H} $ and $ (\sigma, \delta) = ({\rm Id}, 0) $. The vector $ (a_1, a_2, $ $ \ldots, $ $ a_r) \in \mathbb{H}^r $ is a $ (\sigma,\delta) $-multiplicity sequence if, and only if, either $ a_1 = a_2 = \ldots = a_r \in \mathbb{R} $, or $ a_1, a_2, \ldots, a_r \in \mathbb{H} \setminus \mathbb{R} $ and for $ i = 1,2, \ldots, r-1 $, it holds that
$$ {\rm Re}(a_{i+1}) = {\rm Re}(a_i), \quad |a_{i+1}| = |a_i|, \quad \textrm{and} \quad \overline{a}_{i+1} \neq a_i . $$
In particular, $ (a,\ldots,a) \in \mathbb{H}^r $ is a $ (\sigma,\delta) $-multiplicity sequence, for $ a \in \mathbb{H} $ and $ r \in \mathbb{Z}_+ $.
\end{corollary}
\begin{proof}
First, since $ \dim_{\mathbb{R}}(\mathbb{H}) = 4 $, all conjugacy classes are algebraic by Lemma \ref{lemma algebraic conj class 2}. Second, if $ a \in \mathbb{R} $, then $ C^{\sigma,\delta}(a) = \{ a \} $, and if $ a \in \mathbb{H} \setminus \mathbb{R} $, then $ C^{\sigma,\delta}(a) \cap \mathbb{R} = \varnothing $. From these facts and Theorem \ref{th charact multiplicity seqs in alg conj}, if $ a,b \in \mathbb{H} $ are conjugate in the sense of division rings, then $ (a,b) \in \mathbb{H}^2 $ is not a $ (\sigma,\delta) $-multiplicity sequence if, and only if, $ a,b \in \mathbb{H} \setminus \mathbb{R} $ and $ b = \overline{a} $, and the characterization follows.
\end{proof}
We now illustrate Corollary \ref{cor multiplicities on full conjugacy} for $ \mathbb{F} = \mathbb{H} $ and $ (\sigma,\delta) = ({\rm Id},0) $. Recall that, given $ a \in \mathbb{H} $, its conjugacy class is $ C^{\sigma,\delta}(a) = \{ a \} $ if $ a \in \mathbb{R} $, and $ C^{\sigma,\delta}(a) = \{ b \in \mathbb{H} \mid {\rm Re}(b) = {\rm Re}(a), |b| = |a| \} $ if $ a \in \mathbb{H} \setminus \mathbb{R} $. In the case $ a \in \mathbb{H} \setminus \mathbb{R} $, this is due to the fact that
$$ F_{C^{\sigma,\delta}(a)} = (x-a) (x - \overline{a}) = x^2 -2 {\rm Re}(a) x + |a|^2 \in \mathbb{R}[x]. $$
Thus the following result is immediate from Corollary \ref{cor multiplicities on full conjugacy}.
\begin{corollary}
Consider $ \mathbb{F} = \mathbb{H} $, $ (\sigma,\delta) = ({\rm Id},0) $, $ a \in \mathbb{H} $, $ F \in \mathbb{H}[x] $ and $ r \in \mathbb{Z}_+ $.
\begin{enumerate}
\item
If $ a \in \mathbb{R} $, then $ a $ is a zero of $ F $ of multiplicity $ r $ (according to Definition \ref{def multiplicity II}) if, and only if, the central polynomial $ (x-a)^r \in \mathbb{R}[x] $ divides $ F $.
\item
If $ a \in \mathbb{H} \setminus \mathbb{R} $, then $ a $ and $ \overline{a} $ are zeros of $ F $, each of multiplicity $ r $ (according to Definition \ref{def multiplicity II}) if, and only if, $ F $ is divisible by the central polynomial
$$ (x-a)^r (x - \overline{a})^r = \left( x^2 -2 {\rm Re}(a) x + |a|^2 \right)^r \in \mathbb{R}[x]. $$
\end{enumerate}
\end{corollary}
We now turn to cyclic Galois extensions of fields. This includes the case of skew polynomials over finite fields, whose evaluations give projective polynomials \cite{projective-pols} (Definition \ref{def evaluation high degree}) and linearized polynomials \cite{orespecial} (Definition \ref{def lin op pols}). This case is of interest in Coding Theory. For instance, linearized Reed-Solomon codes \cite{linearizedRS} provide PMDS codes with the smallest field sizes known so far \cite{cai-field, gopi-field, pmds-conjugacy, universal-lrc}. \textit{Multiplicity enhancements} are of interest in similar areas of Computer Science \cite{dvir}.
If $ \mathbb{F} $ is a field and $ \sigma $ has finite order $ m \in \mathbb{Z}_+ $, then for each $ a \in \mathbb{F}^* $, it holds that $ K^{\sigma,\delta}_a = K = \{ \beta \in \mathbb{F} \mid \sigma(\beta) = \beta \} $, and $ K \subseteq \mathbb{F} $ is a Galois extension with cyclic Galois group $ \mathcal{G} = \{ {\rm Id}, \sigma, \sigma^2, \ldots, \sigma^{m-1} \} $ of order $ m \in \mathbb{Z}_+ $. In particular, all conjugacy classes are algebraic by Lemma \ref{lemma algebraic conj class 2}. In this setting, we may assume without loss of generality that $ \delta = 0 $ (any other derivation would be inner for some $ c \in \mathbb{F} $, and the results are simply translated by $ c $, see \cite[Sec. 8.3]{cohn} or \cite[Sec. 4.1]{linearizedRS}). We will denote by $ {\rm T}_{\mathbb{F}/K}, {\rm N}_{\mathbb{F}/K} : \mathbb{F} \longrightarrow K $ the classical trace and norm maps, given for $ a \in \mathbb{F} $ by
$$ {\rm T}_{\mathbb{F}/K}(a) = \sum_{i=0}^{m-1} \sigma^i(a) \quad \textrm{and} \quad {\rm N}_{\mathbb{F}/K}(a) = \prod_{i=0}^{m-1} \sigma^i(a). $$
\begin{corollary} [\textbf{Cyclic Galois extensions}] \label{cor multiplicity over cyclic galois}
Assume that $ \mathbb{F} $ is a field, $ \sigma $ has finite order $ m \in \mathbb{Z}_+ $, let $ \delta = 0 $ and define $ K = \{ \beta \in \mathbb{F} \mid \sigma(\beta) = \beta \} \subseteq \mathbb{F} $. The vector $ \mathbf{a} = (a_1, a_2, \ldots, a_r) \in \mathbb{F}^r $ is a $ (\sigma,\delta) $-multiplicity sequence if, and only if, either $ a_1 = a_2 = \ldots = a_r = 0 $ or $ a_1 \neq 0 $ and, for $ i = 1,2, \ldots, r-1 $, there exists $ \beta_i \in \mathbb{F}^* $ such that
$$ a_{i+1} = \sigma(\beta_i)\beta_i^{-1} a_i \quad \textrm{and} \quad {\rm T}_{\mathbb{F}/K} \left( \beta_i a_i^{-1} \right) \neq 0. $$
In particular, $ (a,\ldots, a) $ is a $ (\sigma,\delta) $-multiplicity sequence if, and only if, $ {\rm T}_{\mathbb{F}/K} \left( a^{-1} \right) \neq 0 $.
\end{corollary}
\begin{proof}
First, all conjugacy classes are algebraic by Lemma \ref{lemma algebraic conj class 2}. By Theorem \ref{th charact multiplicity seqs in alg conj}, $ \mathbf{a} $ is a $ (\sigma,\delta) $-multiplicity sequence if, and only if, there exists $ \beta_i \in \mathbb{F}^* $ such that $ a_{i+1} = \sigma(\beta_i)\beta_i^{-1} a_i $ and $ \beta_i \notin \left\lbrace \left( \sigma(\beta) - \beta \right)a_i \mid \beta \in \mathbb{F} \right\rbrace $, for $ i = 1,2, \ldots, r-1 $. For a given $ i = 1,2,\ldots, r-1 $, the latter condition on $ \beta_i \in \mathbb{F}^* $ does not hold if, and only if, $ {\rm Tr}_{\mathbb{F}/K} \left( \beta_i a_i^{-1} \right) = 0 $, by Hilbert's additive Theorem 90 \cite[p. 290, Th. 6.3]{lang}, and the first statement follows. Second, if $ (a,\ldots,a) \in \mathbb{F}^r $, then $ \beta \in \mathbb{F}^* $ satisfies that $ a = \sigma(\beta)\beta^{-1} a $ if, and only if, $ \beta \in K $. In that case, $ {\rm T}_{\mathbb{F}/K}(\beta a^{-1}) = \beta {\rm T}_{\mathbb{F}/K}(a^{-1}) $, and the last statement follows.
\end{proof}
We may now rephrase Corollary \ref{cor multiplicities on full conjugacy} for cyclic Galois extensions of fields. In the setting of Corollary \ref{cor multiplicity over cyclic galois} above, Hilbert's Theorem 90 \cite[p. 288, Th. 6.1]{lang} is equivalent to saying that the conjugacy class of $ a \in \mathbb{F}^* $ is an algebraic P-closed set with minimal skew polynomial $ F_{C^{\sigma,\delta}(a)} = x^m - N_{\mathbb{F}/K}(a) \in K[x^m] $ (see \cite[Th. 19]{lam}). Thus the following consequence of Corollary \ref{cor multiplicities on full conjugacy} can be seen as a Hilbert Theorem 90 with multiplicities
\begin{corollary} [\textbf{Multiplicity Hilbert 90}] \label{cor hilbert 90}
Assume that $ \mathbb{F} $ is a field, $ \sigma $ has finite order $ m \in \mathbb{Z}_+ $, $ \delta = 0 $ and define $ K = \{ \beta \in \mathbb{F} \mid \sigma(\beta) = \beta \} \subseteq \mathbb{F} $. Given $ F \in \mathbb{F}[x;\sigma,\delta] $ and $ a \in \mathbb{F}^* $, and with multiplicities as in Definition \ref{def multiplicity II}, the following are equivalent:
\begin{enumerate}
\item
Every element in $ C^{\sigma,\delta}(a) $ is a zero of $ F $ of multiplicity $ r $.
\item
Every element in a P-basis of $ C^{\sigma,\delta}(a) $ is a zero of $ F $ of multiplicity $ r $.
\item
The central skew polynomial $ \left( x^m - N_{\mathbb{F}/K}(a) \right)^r \in K[x^m] $ divides $ F $.
\end{enumerate}
In particular, if $ a_1, a_2, \ldots, a_\ell \in \mathbb{F}^* $ are pair-wise non-conjugate, $ F \neq 0 $, and every element in $ C^{\sigma,\delta}(a_i) $ is a zero of $ F $ of multiplicity $ r_i \in \mathbb{Z}_+ $, for $ i = 1,2, \ldots, \ell $, then
$$ m \sum_{i=1}^\ell r_i \leq \deg(F) . $$
Furthermore, equality holds if, and only if, $ F = b \prod_{i=1}^\ell \left( x^m - N_{\mathbb{F}/K}(a_i) \right)^{r_i} $, where $ b \in \mathbb{F}^* $.
\end{corollary}
A particular case of interest is that of skew polynomials over finite fields (note that Examples \ref{ex power of linear term is no multiplicity} and \ref{ex full conjugacy class pathology for naive mult} above are of this form).
\begin{example} [\textbf{Finite fields}] \label{ex multiplicity over finite fields}
Let $ q $ be a power of a prime number and let $ m \in \mathbb{Z}_+ $. Let $ \mathbb{F}_q \subseteq \mathbb{F}_{q^m} $ be an extension of finite fields of sizes $ q $ and $ q^m $, respectively. The Galois group of this extension is cyclic and generated by the $ q $-Frobenius automorphism $ \sigma : \mathbb{F}_{q^m} \longrightarrow \mathbb{F}_{q^m} $, given by $ \sigma (a) = a^q $, for all $ a \in \mathbb{F}_{q^m} $. As above, we may assume that $ \delta = 0 $.
In this setting, Corollary \ref{cor multiplicity over cyclic galois} says that $ \mathbf{a} = (a_1, a_2, \ldots, a_r) \in \mathbb{F}_{q^m}^r $ is a $ (\sigma,\delta) $-multiplicity sequence if, and only if, $ a_1 = a_2 = \ldots = a_r = 0 $ or $ a_1 \neq 0 $ and there exist $ \beta_i \in \mathbb{F}_{q^m}^* $ such that
$$ a_{i+1} = \beta_i^{q-1} a_i \quad \textrm{and} \quad {\rm T}_{\mathbb{F}_{q^m} / \mathbb{F}_q} \left( \beta_i a_i^{-1} \right) = \sum_{j=0}^{m-1} \beta_i^{q^j} a_i^{-q^j} \neq 0, $$
for $ i = 1,2, \ldots, r-1 $. In particular, if $ a_1 \neq 0 $, then we have that
$$ {\rm N}_{\mathbb{F}_{q^m} / \mathbb{F}_q} \left( a_{i+1} a_i^{-1} \right) = \prod_{j=0}^{m-1} a_{i+1}^{q^j} a_i^{-q^j} = 1. $$
Observe that we may replace $ \sigma $ by any other generator of the Galois group, that is, we may consider $ \sigma(a) = a^{q^s} $, for all $ a \in \mathbb{F}_{q^m} $, where $ s \in \mathbb{Z}_+ $ is coprime with $ m $.
Now, consider the particular setting of Examples \ref{ex power of linear term is no multiplicity} and \ref{ex full conjugacy class pathology for naive mult}. That is, consider $ q $ odd, $ m $ even with $ m \geq 2 $, and define
$$ a = \gamma^{ \frac{1}{2} \cdot \frac{q^m - 1}{q-1}} \quad \textrm{and} \quad b = - \gamma^{ \frac{1}{2} \cdot \frac{q^m - 1}{q-1}} . $$
where $ \gamma \in \mathbb{F}_{q^m}^* $ is a primitive element. According to Example \ref{ex power of linear term is no multiplicity}, $ (a,a) \in \mathbb{F}_{q^m}^* $ is not a $ (\sigma,\delta) $-multiplicity sequence since a simple computation shows that $ (x-a)^2 = (x-b)^2 $ and $ a \neq b $. We may verify this again by checking the conditions in Corollary \ref{cor multiplicity over cyclic galois}. We simply note that, since $ a^q = -a $, then
$$ {\rm T}_{\mathbb{F}_{q^m}/\mathbb{F}_q} \left( a^{-1} \right) = \frac{m}{2} \cdot a^{-1} - \frac{m}{2} \cdot a^{-1} = 0, $$
since $ m $ is even. Thus Corollary \ref{cor multiplicity over cyclic galois} also states that $ (a,a) $ is not a $ (\sigma,\delta) $-multiplicity sequence.
\end{example}
Another interesting example is that of complex conjugation.
\begin{example} [\textbf{Complex conjugation}] \label{ex multiplicity over complex conj}
Consider the cyclic Galois extension $ \mathbb{R} \subseteq \mathbb{C} $ of order $ 2 $, where $ \mathbb{C} $ denotes the complex field. As is well known, the generator of the Galois group of this extension is $ \sigma : \mathbb{C} \longrightarrow \mathbb{C} $, given by complex conjugation, that is,
$$ \sigma(a) = \overline{a} = {\rm Re}(a) - {\rm Im}(a), $$
for all $ a \in \mathbb{C} $. As above, we may assume that $ \delta = 0 $.
In this setting, Corollary \ref{cor multiplicity over cyclic galois} says that $ \mathbf{a} = (a_1, a_2, \ldots, a_r) \in \mathbb{C}^r $ is a $ (\sigma,\delta) $-multiplicity sequence if, and only if, $ a_1 = a_2 = \ldots = a_r = 0 $ or $ a_1 \neq 0 $ and there exists $ \beta_i \in \mathbb{C}^* $ such that
$$ a_{i+1} = \overline{\beta}_i \beta_i^{-1} a_i \quad \textrm{and} \quad {\rm T}_{\mathbb{C} / \mathbb{R}} \left( \beta_i a_i^{-1} \right) = \beta_i a_i^{-1} + \overline{\beta_i a_i^{-1}} \neq 0, $$
for $ i = 1,2, \ldots, r-1 $. By a straightforward calculation we deduce that, if $ a_1 \neq 0 $, then $ \mathbf{a} $ is a $ (\sigma,\delta) $-multiplicity sequence if, and only if,
$$ |a_{i+1}| = | a_i | \quad \textrm{and} \quad \overline{a}_{i+1} \neq - a_i, $$
for $ i = 1,2, \ldots, r-1 $. It is interesting to note that $ \mathbf{a} = (a,a, \ldots, a) \in \mathbb{C}^r $ is a $ (\sigma,\delta) $-multiplicity sequence if, and only if, $ a $ is not purely imaginary, that is, $ \overline{a} \neq -a $. We may verify the reversed implication directly by noting that, if $ a = ib \in \mathbb{C} $, where $ b \in \mathbb{R}^* $ and $ i $ denotes the imaginary unit, then
$$ (x-a)^2 = x^2 - b^2 \in \mathbb{R}[x^2], $$
operating inside the ring $ \mathbb{C}[x;\sigma,\delta] $. Therefore, $ (x-a)^2 $ has two distinct zeros, namely, $ b $ and $ -b $. We leave to the reader the verification that $ (x-a)^2 \in \mathbb{C}[x;\sigma,\delta] $ only has one zero (without counting multiplicities) if $ a $ is not purely imaginary.
\end{example}
Another common scenario is that of standard derivations over fields, which appears in \textit{differential Galois Theory} \cite{jacobson-derivation, singer}. In this case, the corresponding skew polynomial ring is commonly called a \textit{differential polynomial ring}. The following result is straightforward by item 4 in Theorem \ref{th charact multiplicity seqs in alg conj}. Again in this case, all conjugacy classes are algebraic.
\begin{corollary} [\textbf{Standard derivations}] \label{cor multiplicity over derivations}
Assume that $ \mathbb{F} $ is a field, $ \sigma = {\rm Id} $ is the identity morphism, and $ \delta $ is an arbitrary $ {\rm Id} $-derivation (that is, a standard derivation of the field $ \mathbb{F} $). Let $ K = \{ \beta \in \mathbb{F} \mid \delta(\beta) = 0 \} \subseteq \mathbb{F} $ denote the subfield of constants of $ \delta $, and assume that $ \dim_K(\mathbb{F}) < \infty $. The vector $ (a_1, a_2, \ldots, a_r) \in \mathbb{F}^r $ is a $ (\sigma,\delta) $-multiplicity sequence if, and only if, there exist $ \beta_i \in \mathbb{F}^* $ such that, for $ i = 1,2, \ldots, r-1 $,
$$ a_{i+1} - a_i = \delta(\beta_i) \beta_i^{-1} \quad \textrm{and} \quad \beta_i \notin \delta(\mathbb{F}). $$
\end{corollary}
In the setting of Corollary \ref{cor multiplicity over derivations} above, the dimension of $ \mathbb{F} $ over $ K $ can be determined from the minimal polynomial of the derivation $ \delta $. The following result is a particular case of \cite[Th. 5.10]{algebraic-conjugacy}.
\begin{lemma}[\textbf{\cite{algebraic-conjugacy}}] \label{lemma algebraic derivations}
Assume that $ \mathbb{F} $ is a field and let $ \delta $ be an arbitrary derivation. If $ K = \{ \beta \in \mathbb{F} \mid \delta(\beta) = 0 \} \subseteq \mathbb{F} $, then $ \mathbb{F} $ has finite dimension over $ K $ if, and only if, there exists a non-zero polynomial $ P \in \mathbb{F}[y] $ such that $ P(\delta) = 0 $, that is, $ \delta $ is algebraic. In such a case, $ \dim_K(\mathbb{F}) $ is the degree of the minimum polynomial of $ \delta $ over $ \mathbb{F} $.
\end{lemma}
In particular, in positive characteristic, the \textit{D-fields} introduced by Jacobson in \cite[Sec. III]{jacobson-derivation} fall into the setting of Corollary \ref{cor multiplicity over derivations}. Perhaps the simplest non-trivial examples of D-fields are given by rational functions in positive characteristic.
\begin{example} [\textbf{Rational function fields}] \label{ex multiplicity over rational function fields}
Let $ k $ be a field of characteristic $ p > 0 $. Define $ \mathbb{F} = k(z) $ to be the field of rational functions over $ k $ in $ z $. Let $ \delta : \mathbb{F} \longrightarrow \mathbb{F} $ be a non-zero derivation (note that they are all given by $ \delta = c d/dz $, where $ c \in \mathbb{F}^* $ and $ d/dz $ is the usual derivation). As noted by Jacobson \cite{jacobson-derivation}, since $ p > 0 $, it holds that $ \delta^p $ is again a derivation. Furthermore, it is trivial to see that $ \delta^p = 0 $ in $ k[z] $. Using the defining properties of derivations, we deduce that
$$ \delta^p (f g^{-1}) = f \delta^p(g^{-1}) + \delta^p(f) g^{-1} = - f \delta^p(g) g^{-2} = 0, $$
for all $ f, g \in k[z] $ with $ g \neq 0 $. Hence we deduce that $ \delta^p = 0 $ in $ \mathbb{F} $. In fact, $ P = y^p - 1 \in \mathbb{F}[y] $ is the minimal polynomial of $ \delta $ over $ \mathbb{F} $. We may check the validity of Lemma \ref{lemma algebraic derivations} by noting that
$$ K = \{ \beta \in \mathbb{F} \mid \delta(\beta) = 0 \} = k(z^p), $$
and $ 1,z,z^2, \ldots, z^{p-1} \in \mathbb{F} $ form a basis of $ \mathbb{F} $ as a vector space over $ K $ (the proof is straightforward and left to the reader). Now, by writing an element in $ \mathbb{F} $ as $ f(z) = \sum_{i=0}^{p-1} f_i(z^p) z^i $, where $ f_i(z^p) \in K $, for $ i = 0,1, \ldots, p-1 $, it is easy to see that
\begin{equation}
\delta (\mathbb{F}) = \langle 1,z, z^2, \ldots, z^{p-2} \rangle_K .
\label{eq example rational functions}
\end{equation}
We may also check the validity of Lemma \ref{lemma image of varphi is hyperplane} by noting that
$$ \dim_K (\delta(\mathbb{F})) = p-1 = \dim_K(\mathbb{F}) - 1. $$
Now, taking (\ref{eq example rational functions}) into account and setting $ \sigma = {\rm Id} $, Corollary \ref{cor multiplicity over derivations} says that $ \mathbf{a} = (a_1, a_2, \ldots, a_r) \in \mathbb{F}^r $ is a $ (\sigma,\delta) $-multiplicity sequence if, and only if, there exists
$$ \beta_i = \sum_{j=0}^{p-1} f_{i,j}(z^p) z^j \in k(z), $$
where $ f_{i,j}(z^p) \in k(z^p) $, for $ j =0,1, \ldots, p-1 $, $ f_{i,p-1}(z^p) \neq 0 $ (that is, $ \beta_i \notin \delta(\mathbb{F}) $), and
$$ a_{i+1} = a_i + \delta(\beta_i) \beta_i^{-1}, $$
for $ i = 1,2, \ldots, r-1 $. Fix now an index $ i = 1,2, \ldots, r-1 $. It holds that
$$ \delta(\beta_i) = \sum_{j=1}^{p-1} f_{i,j}(z^p) j \delta(z) z^{j-1} \neq 0, $$
since $ f_{i,p-1}(z^p) (p-1) \delta(z) \in K^* $ and $ 1, z, \ldots, z^{p-2} $ are linearly independent over $ K $. Therefore, we conclude that for all $ a \in \mathbb{F} $ and all $ r \in \mathbb{N} $ such that $ r \geq 2 $, it holds that $ \mathbf{a} = (a,a,\ldots, a) \in \mathbb{F}^r $ is not a $ (\sigma,\delta) $-multiplicity sequence, which is a particularly pathological case.
As in previous examples, we may directly verify that $ (x-a)^2 \in \mathbb{F}[x;\sigma,\delta] $ has a zero $ b \in \mathbb{F} $ such that $ b \neq a $, in this case for all $ a \in \mathbb{F} $. A direct computation shows that this condition holds if, and only if, there exists $ c \in \mathbb{F}^* $ such that
$$ c^2 = - \delta(c), $$
which indeed does not depend on $ a \in \mathbb{F} $. If we set $ \delta(z) = 1 $ for simplicity (that is, $ \delta = d/dz $ is the usual derivation), then we are done by taking the element $ c = z^{-1} $.
\end{example}
{\footnotesize
|
\section{Introduction}
The field of hydrogen-based superconductivity has progressed enormously since 1968, when Ashcroft first proposed that pressurized hydrogen may become a high temperature superconductor \cite{1}. While the first discovered superconductors were Th$_4$H$_{15}$ \cite{2} in 1970 and PdH \cite{2B} in 1972, with not very promising critical temperatures ($T_c$) of 7.6 K and 5 K respectively, the more recent experimental discoveries at high pressures show that superconductivity on hydrogen-based compounds can span from a few Kelvin to room temperature. Few examples are the recently synthesized PrH$_9$\cite{4F}, where $T_c=7$ K at 125 GPa, AlH$_3$ where no superconductivity was observed above 4 K despite the predictions \cite{4E}, and, on the other extreme, H$_3$S \cite{DrozdovEremets_Nature2015}, YH$_9$ \cite{4D,snider2020superconductivity}, YH$_6$ \cite{troyan2020anomalous}, and LaH$_{10}$ \cite{4,4B} all reaching critical temperatures well above 200 K at megabar pressures. In addition, the recent observation of a $T_c$ of 288 K at 267 GPa in a compound formed by sulfur, carbon, and hydrogen \cite{4C} confirms that hydrogen-based superconductors can be room-temperature superconductors. This observation finally dismisses the maximal limits for $T_c$ initially suggested by Cohen and Anderson for electron-phonon driven superconductivity \cite{9}. Hydrogen-based compounds are thus the best currently available candidates to reach ambient temperature and pressure superconductivity.
In order to discover new compounds with high $T_c$ at low pressures, a simple physical-chemical understanding of the properties enhancing the critical temperatures in hydrogen-based systems is necessary. In this regard, the hundreds of compounds predicted to be superconductors with first principles crystal structure prediction techniques constitute a rich working dataset to extract conclusions \cite{7,15,doi:10.1146/annurev-conmatphys-031218-013413}. Among these predictions, the highest $T_c$ values are 300 K for pure metallic hydrogen \cite{5,6} and 326 K for the YH$_{10}$ binary compound \cite{8}. Attempts have been made to increase $T_c$ further through ternary compounds, for instance with H$_3$S$_{1-x}$P$_x$ \cite{10} and Li$_2$MgH$_{16}$ \cite{11}. Aiming at extracting useful information from this dataset, two main routes are being explored: on the one hand, machine learning methods \cite{12,13,14} are starting to be employed to further increase the list of predicted systems, although the obtained new compounds so far do not beat the already known; on the other hand, additional efforts are being invested into classifying these superconductors using simple footprints based on structural, chemical, and electronic properties \cite{13,15,17}. These studies suggest hydrogen rich systems with highly symmetrical structures and high density of states (DOS) at the Fermi level are the best candidates for high-temperature superconductivity.
Even if these properties are able to suggest good trends, they serve necessary but not sufficient conditions. This ultimately means there are not good optimizers thus far: improving these parameters will not necessarily lead to an improvement of the superconducting critical temperatures. In other words, even if the footprints for a good superconductor are somewhat clear, we cannot yet rely on simple variables to estimate the superconducting temperatures, clarify the reason for such a broad spectrum of $T_c$ values, and, ultimately, chemically engineer better superconductors.
In this work we investigate the chemical, structural, and electronic properties through {\it ab initio} methods based on density functional theory (DFT) for a set of 178 hydrogen-based superconductors previously predicted in the literature \cite{7}, including pure hydrogen and binary compounds. Our ultimate goal is to provide a simple understanding of the origin of the high $T_c$ in these compounds. We focus mainly on the electronic and structural properties by means of chemical bonding descriptors, hydrogen-hydrogen distance, electronic charge, and density of states at the Fermi level. We review the impact on the predicted $T_c$ of many of these descriptors, which have been already somewhat studied on a case to case basis in the literature but do not reveal conclusive. We identify that the electron pairing and delocalization play instead the final role. This takes us to propose a universal descriptor based on the identification of electronic delocalization networks, identified by means of the electron localization function (ELF) \cite{18,19,20,21,22}, which quantifies the network of the electronic bonding. We define a simple magnitude, the {\it networking value}, which is easily obtained from the calculation of ELF isosurfaces. Such quantity reveals useful to have a first estimate of the superconducting critical temperature without the need of performing electron-phonon coupling calculations. To the best of our knowledge, it is the first time that such a descriptor is proposed in the literature. We believe the analysis of the {\it networking value} is the first step for a high-throughput analysis of potential new high-$T_c$ systems. It also provides clear paths for chemically engineering better hydrogen-based superconductors, guiding the quest for high-$T_c$ compounds among the vast possibilities offered by ternary compounds.
\section{Results}
\noindent
\textbf{Chemical composition and bonding categories.}
In order to provide a comprehensive understanding of the different type of hydrogen-based compounds predicted to exist in the literature, it is convenient to categorize them in families according to the nature of the chemical bonding of the hydrogen atoms in the system. Our classification is guided by the study of the ELF \cite{18,19,20,21,22} and the atomic charge distribution obtained through the Bader analysis \cite{23,24,25,26} (see Methods for more details). The ultimate goal of this investigation is to better understand what kind of chemical interaction is the most beneficial for superconductivity among these compounds.
After a thorough analysis of the ELF in our test set, we identify six different families according to the nature of the chemical bonding, namely, \textit{molecular} systems, \textit{covalent} systems, systems driven by \textit{weak covalent hydrogen-hydrogen interactions}, systems with \textit{electride} behavior, \textit{ionic} systems, and \textit{isolated} systems. In each case, the nature of the bonds is identified through the analysis of the ELF saddle points between different atoms. Since bonding properties are mainly local, each system can belong to more than just one family. However, in order to simplify the analysis, we focus on the most dominant feature for each compound. A representative for each family, together with the distribution of the families through the groups of the periodic table and the amount of hydrogen fraction in each case, is shown in Fig. \ref{fig:Group}. The hydrogen fraction ($H_f$) takes the following form:
\begin{equation}
H_f = \frac{N_H}{N_H +N_X},
\end{equation}
where $N_H$ and $N_X$ are the number of hydrogen and non-hydrogen atoms in the primitive cell, respectively.
\begin{figure*}[ht]
\includegraphics[width=\textwidth]{Fig1.pdf}
\caption{\textbf{Classification of hydrogen-based superconductors according to their bonding categories.} The top panel shows representative systems for the different categories with ELF isosurfaces at different values: TeH$_4$ for {\it molecular} systems (magenta), PH$_2$ for {\it covalent} systems (purple), ScH$_4$ for systems with \textit{weak covalent hydrogen-hydrogen interactions} (dark green), CrH for {\it elecrtides} (yellow), PdH for {\it isolated} systems (orange), and ScH$_2$ for {\it ionic} systems (light green). The lower panel shows respectively the critical temperature as a function of the host atom periodic group (left panel) and the hydrogen fraction in the compounds (right panel). In the left panel, the name of some compounds is explicitly given together with the index (in parenthesis) given to each compound in Extended Data Tables I - III. In the right panel, and also in Figs. \ref{fig:Electronic} and \ref{fig:Network}, just the index is given for these compounds for the sake of brevity.}
\label{fig:Group}
\end{figure*}
The \textit{molecular} family describes all systems having at least one pair of hydrogen atoms forming a molecule. The latter can be identified through the ELF analysis by locating an isosurface surrounding an isolated molecule at very high values of ELF (see the magenta surface around the hydrogen molecules in TeH$_4$ in Fig. \ref{fig:Group}). A system is chosen to be molecular if an isolated pair of hydrogen atoms appears connected at a value of ELF higher than 0.85, i.e., if the minimum ELF value in between the hydrogen atoms is above 0.85. Systems with {\it molecular} behavior appear between groups 1 to 4 and 13 to 17. All of them have a value of $H_f>0.\overline{6}$, meaning that in the unit cells there are at least two hydrogen atoms per host. We note that {\it molecular} systems tend to have very high values of $H_f$, which reflects that in most cases several molecules exist per host atom. The highest critical temperatures for these {\it molecular} compounds have been predicted for ScH$_{12}$(33) and TeH$_4$(150), with $T_c$ values around 150 K (the number in parenthesis after a given compound corresponds to the index given for each compound in Supplementary Data Tables I - III). The critical temperatures of the {\it molecular} compounds span from few Kelvin to these high values.
The \textit{covalent} family is composed of systems where covalent bonds between hydrogen and the host atoms are dominant. Throughout our discussion we shall label the host atom as X, with X$\neq$H. For H-X bonds, the covalent character can be identified by an elongation toward the host atom of the ELF isosurface surrounding the hydrogen (i.e. a polarized covalent bond). This is exemplified by the purple surfaces pointing from the H (small spheres) to the P atoms (big spheres) in PH$_2$ in Fig. \ref{fig:Group}. {\it Covalent} systems appear for groups 13 to 17, and are mostly related to the host atom's $p$ type orbital character. Some of the highest critical temperature for these systems have been predicted for H$_3$S(138) at 200 K, BH$_3$(77) at 125 K, and H$_3$Se(148) at 110 K.
The \textit{weak covalent hydrogen-hydrogen interaction} family is dominated by compounds with predominant weak H-H covalent interactions. The difference with the molecular case is that hydrogen molecules or clusters appear elongated or quasi dissociated. This is illustrated with the dark green surfaces in ScH$_4$ in Fig. \ref{fig:Group}, where bonds between hydrogen atoms appear at much lower values of ELF. From a quantitative point of view we assume that a group of hydrogen atoms is weakly bonded if the ELF at the bond point is within the range [0.4 - 0.85]. These systems mostly appear between groups 1 to 5. In this bonding type, interactions seem to be purely related to hydrogen atoms whilst the host atoms appear as inert or acting as a chemical pre-compressor or electron donor. Compounds with this bonding characteristic tend to contain many hydrogen atoms per host, as it happens for the molecular compounds. For compounds with the host atom in a low group of the periodic table, the host atoms valence electrons are donated to the hydrogen atoms resulting in a weakening of the H-H bonds, which translates into an increment of the H-H distance. This family shows the highest predicted critical temperatures, the highest being 326 K for YH$_{10}$(27) and 300 K for metallic hydrogen H(76).
The \textit{ionic} family is formed by those systems whose hydrogen atoms show an ionic character. This is identified by an isolated proto-spherical ELF isosurfaces surrounding the hydrogen atoms. In addition, for a system to be hereby considered as \textit{ionic} the mean extra charge per hydrogen atom must be more than 0.5 electrons. This is illustrated by the spherical light green surfaces around hydrogen in ScH$_2$ in Fig. \ref{fig:Group}, where the charge on H is 0.67 electrons. Ionic behavior was observed between groups 2 to 5 of the periodic table, in all cases with low values of $H_f$. The ionic bonding originates from the strong difference in electronegativity between host and hydrogen atoms, which is always increased under pressure \cite{31}. This family transforms into the \textit{weak covalent hydrogen-hydrogen interaction} class when $H_f$ increases, since the donated charge becomes small in respect to the number of recipient hydrogen atoms. Critical temperatures for these systems are low, with the highest being 45 K for BeH$_2$(4).
The \textit{electride} family contains systems featuring electride behavior, i.e. compounds with electrons localized in the voids. The latter can be identified by isolated pockets of localized electrons in empty space of the crystal as the ones shown in yellow for CrH in Fig. \ref{fig:Group}. From a quantitative viewpoint, electride behavior is characterized in terms of isolated isosurfaces not surrounding any nuclei with ELF maximum values in between 0.35 to 0.7. Note that metallic compounds are also included in this family. Metallic cases also show isolated bubbles of ELF occupying the voids, but their profile is flatter. Given the difficulties to set a quantitative barrier, we have merged them in a unique family. Electrides and metals only appear mainly between groups 5 and 10, and are among the systems with the lowest value of $H_f$ reaching a maximum of 3 hydrogen atoms per host ($H_f = 0.75$). Critical temperatures for these systems are low, not reaching above 50K, exception made for the Si$_2$H$_6$(94) with a $T_c$ of 153 K, where the electride-metallic behavior seems to be arising from H-Si covalent bonds (not justifying a further differentiation between electrides and metals).
The remaining family includes all the materials featuring extremely weak bonds between hydrogen and host atoms. These systems have been named as \textit{isolated} and are identified by the lack of any kind of connection of the ELF isosurfaces above 0.25. These systems have low critical temperatures not reaching above 40 K and appear mainly between groups 5 and 12 of the periodic table. They also show a weak capacity of hosting a large number of H atoms per X atom, as they have the lowest values of $H_f$.
Overall our results highlight that characterizing the bonding type of a solid thanks to these families enables to discard a great number of compounds as potential high-$T_c$ compounds. Covalent interactions, be it weak H-H or X-H are the most favorable for high-temperature superconductivity. This allows to identify the potential interesting combination of elements, especially with respect to the increasing search among ternary compounds. The lowest $T_c$ values appear for {\it electrides} and {\it isolated} compounds, mostly appearing between groups 5 and 12, which do not show lots of potential as high-$T_c$ compounds. These families also show the lowest values of $H_f$.
\newline
\begin{figure*}[ht]
\includegraphics[width=\textwidth]{Fig2.pdf}
\caption{\textbf{Structural and electronic properties.} Panel \textbf{a)} shows $T_c$ as a function of the shortest hydrogen-hydrogen distance for all the compounds. Panel \textbf{b)} shows $T_c$ as a function of the mean extra electrons per hydrogen atom. Panels \textbf{c)}, \textbf{d)}, and \textbf{e)}, respectively, show $T_c$ as a function of the total DOS at the Fermi level, the DOS at the Fermi level projected on the hydrogen $s$ orbitals per hydrogen, and the fraction of the total DOS at the Fermi level coming from the hydrogen orbitals.}
\label{fig:Electronic}
\end{figure*}
\noindent
\textbf{Hydrogen-hydrogen distance and electronic properties.}
After categorizing the different bonding families of hydrogen-based superconductors, we focus on understanding the trends of the predicted $T_c$ with structural and electronic properties. The results are summarized in Fig. \ref{fig:Electronic}. The analysis focuses on the shortest hydrogen-hydrogen distance for each compound, the charge distribution on hydrogen atoms, and the density of states at the Fermi level with the ultimate goal of finding correlations between such descriptors and the critical temperature.
Even if no general trend is observed when plotting $T_c$ as a function of the H-H shortest distance, several conclusions can be drawn. The structural analysis highlights an increment of the superconducting critical temperature with the increase of the shortest hydrogen-hydrogen distance for those systems where the bonding is driven by pure hydrogen interactions, meaning the {\it molecular} and \textit{weak covalent hydrogen-hydrogen interaction} families. For these two families the H-H distance spans from 0.74 \AA~for systems with $T_c$ below 1 K to a maximum of around 1.35 \AA~for the compounds with highest critical temperatures. In the region between 0.9 and 1.35 \AA~lie the currently predicted compounds with the highest superconducting temperatures reaching values as high as 300 K. In other words, elongated H-H interactions promote $T_c$. Interestingly, our analysis highlights that such H-H distance variation is not related to a variation of pressure, meaning that on a broad level, increasing the pressure does not necessarily yield an increment of the bonding distance, so composition rather than pressure would be a more relevant variable to tweak (see Supplementary Data Table I to see the pressure at which each compound is studied).
The shortest hydrogen-hydrogen distance for the {\it covalent} family spans between 1 and 2.5 \AA. Low symmetry {\it covalent} systems appear between 1 to 1.45 \AA\ where the short H-H distance is due to inter hydrogen bonds appearing beside the dominant hydrogen-host bonds. The highest reported superconducting temperature for these systems is about 135 K for BH$_3$(77). Interestingly, at around 1.55 \AA\ the {\it covalent} family shows a sharp spike in the predicted $T_c$ through systems sharing linear $X=H=X$ bonds originating through the host $p$ orbitals, a $\bar{3}$m point group, a value of $H_f$ equal to 0.75, and a lack of direct hydrogen-hydrogen bonds. Here lay systems as H$_3$S(138), H$_3$Se(148), GaH$_3$(80), and GeH$_3$(106). Interestingly, the Si$_2$H$_6$(94) {\it electride} also shares these features even if the bonding nature is slightly different. All these compounds are shown in Fig. \ref{fig:HHHComp} with a representative ELF isosurface. For H-H distances above approximately 1.55 \AA\ a sharp drop in $T_c$ appears, with systems not reaching above 50 K. This zone lacks direct hydrogen-hydrogen bonds and features systems with a low percentage of hydrogen, with mostly {\it isolated}, {\it electride}, and {\it ionic} behavior. Hence, covalent elongated bonds, H-H or H-X, with high $H_f$ seem to be the best candidates to increase $T_c$.
\begin{figure*}[ht]
\includegraphics[width=\textwidth]{Fig3.pdf}\\
\caption{\textbf{Highly symmetric high-$T_c$ systems .} This figure shows five compounds related to the $T_c$ spike around a hydrogen-hydrogen distance of 1.55 \AA . The pink spheres refer to hydrogen atoms, while the rest refer to the respective host atoms. An ELF isosurface is depicted in each case. Also the shortest H-H ($D_{HH}$) and H-X ($D_{HX}$) distances are noted. Most of these compounds belong to the {\it covalent} family and show a purple ELF isosurface. The Si$_2$H$_6$ compound is an {\it electride} (note the small ELF bubbles in empty sites) and its ELF isosurface is shown in orange.}
\label{fig:HHHComp}
\end{figure*}
\noindent
Panel \textbf{b} of Fig. \ref{fig:Electronic} shows the mean extra electron per hydrogen atom obtained through the analysis of the Bader charge (see Methods). The {\it covalent} family has been excluded from this panel due to the unreliability of the Bader anlysis, which arises from the difficulties in assigning shared electrons in the H-X bonds to each atom. We estimate the mean extra electron per hydrogen atom ($\bar{\rho} $) as
\begin{equation}
\bar{\rho} = \frac{\sum_{i=1}^{N_H} Q_i - N_H}{N_H},
\end{equation}
where $Q_i$ is the number of electrons assigned to the $i$th hydrogen atom by the Bader analysis, while $N_H$ is the total number of hydrogen atoms in the primitive cell.
Overall hydrogen atoms tend to gain electrons due to their higher electronegativity with respect to other atoms in the cell. We observe that the highest $T_c$ values are associated with small charge transfers, i.e. from 0 to 0.25 extra electrons per hydrogen atom. As the extra charge increases, $T_c$ drops sharply below 50 K, exception made for the BeH$_2$(5), compound belonging to the {\it ionic} family with $T_c=97$ K, and Si$_2$H$_6$(94), with $T_c=153$ K and part of the {\it elecrtide} family.
It is often mentioned in the literature that an increment of (negative) charge on the hydrogen atom leads to a weakening of the hydrogen-hydrogen bonds \cite{11,104,112,https://doi.org/10.1002/wcms.1330}. Our results show that both {\it molecular} and \textit{weak covalent hydrogen-hydrogen interaction} families show non-negligible extra electrons on the hydrogen atoms. We also observe that the extra charge on the hydrogen is responsible for a slight increment of the H-H distance for the \textit{weak covalent hydrogen-hydrogen interaction} family. In fact, compounds within this family with a shorter H-H distance tend to have less extra electrons per hydrogen atom. The maximum extra electron per hydrogen among this family is around 0.5, a value that increases up to 1 in the {\it molecular} systems.
Panels \textbf{c}, \textbf{d}, and \textbf{e} in Fig. \ref{fig:Electronic} report the results for the DOS analysis. A simplified equation for the electron-phonon coupling constant $(\lambda)$ may be given by \cite{PhysRev.167.331}
\begin{equation}
\lambda = \frac{N(E_F)\langle I^2 \rangle}{M\langle \omega^2 \rangle},
\end{equation}
where $N(E_F)$ is the DOS at the Fermi energy, $\langle I^2 \rangle$ is a term that is related to the electron-phonon matrix elements, $M$ is the mean mass of the primitive cell, and $\langle \omega^2\rangle$ is the mean quadratic phonon frequency. This suggests that a higher $N(E_F)$ should give rise to a higher electron-phonon coupling and, thus, also a higher superconducting critical temperature. Although this is a favorable condition for good superconductors, the results for the DOS prove this condition not to be univocally sufficient \cite{16}. The total DOS per electron at the Fermi level shown in panel \textbf{c} of Fig. \ref{fig:Electronic} shows a sharp increment in the highest superconducting temperatures for a DOS value of around 0.015 $eV^{-1}$. However, for such values of the DOS, compounds with very low $T_c$ can still be found, making the total DOS at the Fermi level not a good descriptor of high-$T_c$ compounds.
Considering that H atoms due to their light mass are responsible for the large values of $T_c$ in these compounds, we analyze the projection of the DOS at the Fermi level onto hydrogen atoms (see panel \textbf{d} of Fig. \ref{fig:Electronic}). However, the trends are similar to those obtained for the total DOS. In this case, the chance to have a higher superconducting critical temperature is related to an increment of DOS up to a value of 0.06 $eV^{-1}N_H^{-1}$. In particular, our results suggest that the key for high superconductivity is not strictly related to the value of the DOS. Instead, the fraction of active hydrogen atoms at the Fermi energy (see panel \textbf{e}) seems to be more relevant. In agreement with our previous findings, only {\it molecular} and \textit{weak covalent hydrogen-hydrogen interaction} families are able to reach high amounts of DOS coming from the hydrogen active atoms. This is directly related to the fact that the H-H bonds in those systems have a great contribution to the HOMO and LUMO. Nevertheless, false positives appear where the $T_c$ appears very low despite the large contribution of hydrogen states to the Fermi surface. In fact, the DOS misses the information on how electrons are coupled with the lattice vibrations. Systems with very high values of the DOS at the Fermi energy but with very low electron-phonon coupling will not exhibit high superconducting critical temperatures.
\noindent
\textbf{Networking through the ELF.}
In order to screen new superconductors and guide the quest for new superconducting compounds, we need to look for an easily computable variable that will tell us not only about the interatomic bonding properties, but also how prone the system is to electron-phonon coupling. Even if some trends can be observed as discussed above, the descriptors analyzed so far in Figs. \ref{fig:Group} and \ref{fig:Electronic} are unable to capture when electrons couple more strongly to lattice vibrations and, thus, do not correlate with $T_c$. In this section we propose a new observable based on the study of electron (de)localization that, instead, correlates well with the predicted superconducting critical temperature.
The ELF is a proper function to analyze the degree of electronic localization as high value isosurfaces of the ELF reveal regions in space where electrons localize. In fact, for isosurfaces with values close to 1 the electrons are localized generally on atomic sites, and they start to delocalize toward neighbors and form bonds as the ELF value decreases (see Methods). In order to analyze this delocalization on a crystal size scale, we define the \textit{networking value} $\phi$ as the highest value of the ELF that creates an isosurface spanning through the whole crystal in all three cartesian directions. The $\phi$ value can thus be easily extracted by calculating the ELF and determining at which value a crystal sized isosurface is created when lowering the ELF value from 1. This isosurface encloses most of the atoms in the crystal, but not necessarily all. The ELF saddle points reveal crucial for the determination of the {\it networking value}, especially for hydrogen based compounds. For these systems, where hydrogen-hydrogen bonds are dominant, the ELF saddle points identify the interatomic bonds that pave the crystal sized network. Thus, for the determination of the {\it networking value} it is sufficient to identify the ensemble of ELF saddle points at the highest value of ELF able to bridge the gap between different atoms and create the 3D network.
In Fig. \ref{fig:Network} we provide the ELF isosurface related to the $\phi$ value as well as the network created by the saddle points for this ELF value for PdH(70), YH$_4$(15), and H$_3$S(138). These three cases are related to the most common types of networks identified during the analysis: {\it isolated}, \textit{weak covalent hydrogen-hydrogen interaction}, and {\it covalent} families, respectively. In PdH(70) the network includes both Pd and H atoms. The isolated behavior of the atoms leads to the ELF bubbles around the atoms getting in contact at very low values of ELF, 0.19. For YH$_4$(15) the highest 3D connecting network appears at ELF=0.43, and is constructed only by hydrogen atoms showing a weak covalent interaction. This is one of the cases in which the 3D network subsisting at the highest value of ELF does not include all the atoms in the unit cell. For the case of H$_3$S(138), two interlaced networks appear at ELF=0.68 due to its high symmetry, which is supported by the weak $H=S$ covalent bonds. Another type of network is the one arising from the {\it electride} systems, where the connection appears through the isolated pockets of charge in the empty zones of the unit cell.
Interestingly, the {\it networking value} correlates rather well with $T_c$ as shown in Fig. \ref{fig:Network}, clearly much better than any other descriptor based on the structure or the electronic properties studied so far in the literature and in Figs. \ref{fig:Group} and \ref{fig:Electronic}. We attribute this positive correlation to the capacity of this descriptor to somehow measure how the lattice vibrations affect the electronic cloud. This is not surprising, because if atoms are connected among them with highly localized electrons, phonon vibrations are prone to affect more electrons. The fact that the network is required to span through all the crystal seems also reasonable, as all, or at least many, phonon modes are expected to affect the electronic cloud.
Few attempts were made to improve the correlation between $\phi$ and $T_c$ by introducing additional conditions for the network identification, such as having all the hydrogen atoms be part of the network, connecting the atoms through short direct paths, and creating connections between close hydrogen atoms up to a given distance. The introduction of these restrictions however produced only minor improvements at the prize of complicating the definition of $\phi$. On the opposite side, relaxing the condition about the 3D nature of the network to simply two dimensional sheets or one dimensional worms worsened the correlation. This suggests that forming a 3D bonding network is crucial, which we believe it is related to the fact that in this case the networking value better captures the overall coupling of the phonon modes with the electrons as argued above.
As seen in Fig. \ref{fig:Network}, an improvement in the correlation is obtained by multiplying $\phi$ by the hydrogen fraction $H_f$:
\begin{equation}
\Phi = \phi H_f.
\end{equation}
A reasonable explanation for such improvement is that the hydrogen fraction is a rough estimation of the multiplicity of hydrogen bonds. Systems with few H atoms will tend to form less bonds in which H atoms participate. This is in contrast with hydrogen rich systems with an incredible number of bonds in which H atoms participate. An example is HSe(148) with just 7 bonds per formula unit and $H_f = 0.5$, in contrast with LaH$_{10}$(30) with 20 bonds per formula unit and $H_f$ = 0.90.
We have also investigated the possible correlation between the {\it networking value} and the DOS which was found missing (see Fig. \ref{fig:NetworkvsDOS}). The fact the {\it networking value} does not correlate at all with the DOS at the Fermi level , not even with the DOS coming from H atoms, underlines that these two descriptors are measuring different things. While $\phi$ is able to capture the effect of the electron-phonon interaction, the DOS cannot, and consequently does not correlate well with $T_c$. On the contrary, $\phi$, or its product with the hydrogen fraction $\Phi$, determine for the first time a magnitude that shows a striking correlation with $T_c$, and constitutes a first positive and cheap identification of hydrogen-based superconductors.
\begin{figure*}[ht]
\includegraphics[width=\textwidth]{Fig4.pdf}
\caption{\textbf{Networking value.} The upper panel shows the ELF isosurface and the three dimensional network spanning through all the crystal, which is formed by the ELF saddle points and the atoms, associated with the $\phi$ value for PdH(70), YH$_4$(15), and H$_3$S(138). The {\it networking value} is given for each case. The bottom panels show the critical temperature $T_c$ as a function of the networking value $\phi$ (left) and the networking value multiplied by $H_f$, $\Phi = \phi H_f$ (right). In both cases there is a linear correlation with the $T_c$, although $\Phi$ correlates slightly better.}
\label{fig:Network}
\end{figure*}
\noindent
\begin{figure}[ht]
\includegraphics[width=\columnwidth]{hdrDOS.png}
\caption{\textbf{Networking value with respect to the DOS.} The figure shows the {\it networking value} $\phi$ as a function of the hydrogen contribution to the DOS at the Fermi energy per hydrogen atom.}
\label{fig:NetworkvsDOS}
\end{figure}
\section{Discussion}
After the large analysis presented in this work, we can conclude that the highest critical temperatures are achieved in the {\it molecular}, {\it covalent} and \textit{weak covalent hydrogen-hydrogen interaction} families. These three families are different expressions of covalent bonds, where electrons are strongly localized. The highest critical temperatures appear between groups 1 and 5 of the periodic table, where bonds are mainly driven by covalent hydrogen-hydrogen interactions, and 13 and 16, where covalent bonds are predominantly between hydrogen and host atoms. It seems that these {\it covalent} compounds reach their highest $T_c$ values for systems with large symmetry and hydrogen-host bonds at a distance of about 1.55 \AA, without direct hydrogen-hydrogen bonding.
It is important to remark that our work shows how the {\it molecular} family transitions towards the \textit{weak covalent hydrogen-hydrogen interaction} one, with an associated increase in $T_c$. The transition is smooth, starting from low $T_c$ systems with only hydrogen molecules, such as H$_4$I(164), going then through a mixed phase where molecules expand and intermolecular H-H interactions start to be present as for ScH$_9$(31) and ScH$_7$(34), to finally transition towards a full weak interacting behavior with no molecules as found for YH$_{10}$(27) and LaH$_{10}$(30), which have the largest {\it networking value} and $T_c$. This suggests that stretching hydrogen molecules is beneficial for superconductivity. The same conclusion is reached looking at the two pure hydrogen phases studied here (systems 176 and 177). The hydrogen phase ascribed to the \textit{weak covalent hydrogen-hydrogen interaction} behavior (176) shows a far higher $T_c$ of around 300 K compared to the purely molecular phase (177), with $T_c=109$ K. Therefore, stretching hydrogen molecules by chemical or mechanical means in systems containing many H$_2$ units seems a very promising path to discovering new high-$T_c$ compounds. This seems to put in context the extraordinary prediction of a critical temperature of 473 K in Li$_2$MgH$_{16}$ \cite{11}, where doping a molecular MgH$_{16}$ compound with Li brakes the molecular units, transforming the system into one with \textit{weak covalent hydrogen-hydrogen interactions}.
The {\it networking value} $\phi$ defined here is able to capture effectively how sensitive the electronic cloud is on average to lattice vibrations, and, consequently, correlates well with $T_c$. The {\it networking value} improves all other structural or electronic descriptors previously studied \cite{13,15,17}. As extracting $\phi$ simply requires the analysis of ELF isosurfaces, which can be easily obtained post-processing DFT ground state calculations, it offers a simple way of screening hydrogen-based superconductors, as well as showing the correct directions to chemically engineering better hydrogen-based superconductors. Interestingly, as the definition of $\phi$ is completely general, not limited to the presence of hydrogen in the system, it could potentially be used to estimate the $T_c$ of all phonon-mediated superconductors. Eventually, it may be also worth studying it in unconventional superconductors, as the $\phi$ is only based on the analysis of the electronic cloud and correlations could also appear.
As a final word, we would like to underline that the superconducting critical temperatures used to find correlations are extracted directly from the literature, without being recalculated (see Supplementary Data Table I to check the reference from which the $T_c$ value was taken for each case). All the $T_c$ values were obtained by first principles DFT calculations, but at different levels of theory, for instance, for the estimation of the critical temperature. The wide grey area in Fig. \ref{fig:Network} could be the result of such inconsistencies. In addition, most of these $T_c$ values have been obtained assuming that the ground state structure is the one given by the minimum of the Born-Oppenheimer energy surface (classical approximation) and that lattice vibrations can be described within the harmonic approximation around these positions. However, in hydrogen-based superconductors, recent calculations have shown that the crystal structure can be largely modified by ionic quantum effects and that anharmonicity strongly renormalizes the obtained harmonic phonon spectra, which can strongly impact the predicted $T_c$\cite{PhysRevLett.111.177002,PhysRevB.89.064302,33,34,Errea81,6,8,Bianco2018,hou2021strong}. These effects could produce a narrower correlation between $T_c$ and $\phi$ by introducing variations due to corrections on the structure and $T_c$.
\section{Methods}
\noindent
\textbf{DFT Calculations.}
We carry out our analysis on a sample of 178 compounds containing hydrogen that had been previously predicted to be superconductors in the literature. Most of the chosen compounds were those summarized in Ref. \cite{7}, which had been analyzed before in the literature \cite{scList,100,101,102,103,104,105,106,107,108,109,110,111,112,113,115,116,117,118,119,120,121,122,123,124,125,126,127,128,129,130,131,132,133,134,135,136,137,138,139,140,141,142,143,144,145,146,147,148,149,150,151,152,153,154,155,156,157,158,159,160,161,162,163,164,165,166,167,168,169,170,171,172,173,174,175,176,177,178,179,180,181,182,183,184,185,186,187,188,189,190}. We took each of these compounds, relax them classically at a given pressure with DFT at the Born-Oppenheimer minimum position, and calculated electronic properties for them. Due to the immense work required, we did not perform the $T_c$ calculations, but took the values predicted in the literature. Among the 178 compounds, 43 were discarded due to lack of information on the atomic structure, which made impossible their analysis. Supplementary Data Tables I-III summarize all the calculated data presented here, as well as from which reference the $T_c$ value was taken.
All DFT calculations were performed with the plane-wave {\sc Quantum ESPRESSO} (QE) package \cite{27,28}. The exchange correlation potential was approximated with the Perdew-Burke-Ernzerhof parametrization \cite{GGA-PBE}. At least the first few upper core orbitals were included in the pseudopotential for the host atom. The cutoff for the wave-functions and the density were respectively 70 Ry and 700 Ry. Integrations over the Brillouin zone were performed with the Methfessel-Paxton smearing technique \cite{PhysRevB.40.3616}, with a 0.02 Ry broadening. These integrations were performed with dense $\mathbf{k}$ point grids, where a volume of 0.001 {\AA}$^{-3}$ was occupied per $\mathbf{k}$ point in the Brillouin zone for the self-consistent calculation and a volume of 0.0002 {\AA}$^{-3}$ for the non self-consistent calculation. The electronic properties such as the ELF, the DOS, and the charge distribution were calculated for each system using the QE post-processing tools through the results obtained for the non self-consistent calculations.
\noindent
\textbf{Bader charges. }
There have been numerous approaches in order to determine the charge associated to an atom in a molecule. Probably, one of the most useful in solid state is that derived from the electron density, introduced by Bader and coworkers in what is called the Quantum Theory of Atoms in Molecules \cite{BA90}.
In an ordinary solid, the electron density has its maxima (cusps) at the nuclei and decays exponentially as the electron density moves away from the nuclei. The resulting topology looks like an assemblage of mountains, each of which is identified as an atom.
The zero gradient surface around these maxima are well defined surfaces that lead to atoms as non-overlapping units. This allows determining their charge by mere integration of the electron density within their associated region of space. Since these regions are non-overlapping these charges have the advantageous properties of being additive.
\noindent
\textbf{The Electron Localization Function. }
The electron localization function (ELF) was developed by Becke and Edgecombe in 1990 \cite{BE90} for the analysis in real space of electron localization, and later on reinterpreted by Savin \cite{KS96} in terms of the Pauli kinetic energy density ($t_p$) corrected by the homogeneous electron gas kinetic energy density ($t_w$):
\begin{eqnarray}
\chi=\frac{1}{1+(t_p/t_w)^2}.
\label{eta}
\end{eqnarray}
$\chi$ is then mapped to run from 0 to a maximum value of 1:
\begin{eqnarray}
ELF=\frac{1}{1+\chi^2}.
\label{elf}
\end{eqnarray}
Values close to 1 appear in those places where electrons are localized. Hence, maxima appear in the bonds (as well as in cores and lone pairs).
It should be noted that hydrogen constitutes a peculiar case. A molecule such as N$_2$ will feature a maximum in the middle of the inter-nitrogen distance associated with the N-N bond and separated from the N cores. Since a hydrogen molecule only has 2 electrons, a maximum does not appear for the H-H bond, but rather a surface with very high ELF that encapsulates the H$_2$ molecule.
It is easy to see from Eq. \ref{elf} that the value $ELF=$0.5 is associated with the distribution in a homogeneous electron gas of the same density as the point of study. Indeed, metals are characterized by very flat ELF profiles slightly deviating from 0.5.
Whereas the maxima of ELF provide a measure of how localized electrons are, its value in between these maxima characterize delocalization in between these regions, i.e. how easy it is for electrons to go from one localized unit to another \cite{deloc}.
When the ``easiest pathway'' is analyzed in the full crystalline cell, we end up with a picture of how easy it is for electrons to move across the crystal.
\section{Data availability}
All the results presented in this work are summarized in Supplementary Data Tables I-III. Further data or details are available from the corresponding author upon reasonable request.
\section{Code availability}
Quantum ESPRESSO is an open-source suite of computational tools available at https://www.quantum-espresso.org.
\def\bibsection{\section*{\refname}}
\section{Supplementary Data Tables}
\begin{longtable}{c|c|c|c|c|c|c}
\caption{Table reporting item number used in the manuscript, chemical formula, space group, superconducting critical temperature as predicted, pressure at which it has has been predicted, hydrogen fraction $H_f$, and the reference from which the $T_c$ value has been extracted. The empty cells refer to missing values from the literature or to the impossibility to perform calculations.} \\
\label{table:Sup1}
Item & Chemical formula & Space group & $T_c$ (K) & Pressure (GPa) & $H_f$ & Reference \\
\hline
0 & LiH$_{2}$ & $ P4/mbm $ & 0 & 150 & 0.666 & \cite{100} \\
1 & LiH$_{2}$ & $ R\bar{3}m $ & 38.34 & 140 & 0.857 & \cite{100} \\
2 & LiH$_{8}$ & $ I422 $ & 31.04 & 200 & 0.888 & \cite{100} \\
3 & KH$_{6}$ & $ C2/c $ & 69.8 & 166 & 0.8 & \cite{101} \\
4 & BeH$_{2}$ & $ Cmcm $ & 45.1 & 204 & 0.666 & \cite{102}\\
5 & BeH$_{2}$ & $ P4/nmm $ & 97 & 349 & 0.666 & \cite{102}\\
6 & MgH$_{6}$ & $ Im\bar{3}m $ & 271 & 300 & 0.857 & \cite{103} \\
7 & CaH$_{6}$ & $ Im\bar{3}m $ & 235 & 150 & 0.857 & \cite{104}\\
8 & SrH$_{6}$ & $ R\bar{3}m $ & 156 & 250 & 0.857 & \cite{105}\\
9 & BaH$_{2}$ & $ R\bar{3}m $ & 0 & 60 & - & \cite{106}\\
10 & BaH$_{6}$ & $ P4/mmm $ & 38 & 70 & - & \cite{106}\\
11 & ScH$_{3}$ & $ - $ & 19.3 & 18 & - & \cite{107} \\
12 & LaH$_{3}$ & $ Cmcm $ & 22.5 & 11 & - & \cite{107} \\
13 & YH$_{3}$ & $ - $ & 40 & 17 & - & \cite{107} \\
14 & ScH$_{2}$ & $ - $ & 38.11 & 80 & 0.666 & \cite{108}\\
15 & YH$_{4}$ & $ I4/mmm $ & 95 & 120 & 0.8 & \cite{109}\\
16 & YH$_{6}$ & $ Im\bar{3}m $ & 264 & 120 & 0.857 & \cite{109}\\
17 & ScH$_{4}$ & $ I4/mmm $ & 98 & 200 & 0.8 & \cite{110}\\
18 & ScH$_{6}$ & $ Im\bar{3}m $ & 169 & 350 & 0.857 & \cite{111} \\
19 & PrH$_{9}$ & $ F\bar{4}3m $ & 0 & 100 & 0.9 & \cite{112}\\
20 & CeH$_{10}$ & $ Fm\bar{3}m $ & 134 & 200 & 0.909 & \cite{113}\\
21 & CeH$_{9}$ & $ P6_3/mmc $ & 50 & 100 & 0.9 & \cite{112}\\
22 & LaH$_{9}$ & $ Cc $ & 30 & 50 & 0.9 & \cite{112}\\
23 & LaH$_{6}$ & $ R\bar{3}c $ & 170 & 100 & 0.857 & \cite{112}\\
24 & YH$_{9}$ & $ P6_3/mmc $ & 250 & 150 & 0.9 & \cite{112}\\
25 & ScH$_{9}$ & $ P6_3/mmc $ & 180 & 400 & 0.9 & \cite{112}\\
27 & YH$_{10}$ & $ Fm\bar{3}m $ & 326 & 250 & 0.909 & \cite{112,115} \\
28 & LaH$_{4}$ & $ I4/mmm $ & 10 & 300 & 0.8 & \cite{115} \\
29 & LaH$_{8}$ & $ C2/m $ & 131 & 300 & 0.888 & \cite{115} \\
30 & LaH$_{10}$ & $ Fm\bar{3}m $ & 274 & 150 & 0.909 & \cite{115} \\
31 & ScH$_{9}$ & $ I4_1md $ & 163 & 300 & 0.9 & \cite{111}\\
32 & ScH$_{10}$ & $ Cmcm $ & 120 & 250 & 0.909 & \cite{111}\\
33 & ScH$_{12}$ & $ Immm $ & 141 & 350 & 0.923 & \cite{111}\\
34 & ScH$_{7}$ & $ Cmcm $ & 169 & 300 & 0.875 & \cite{111}\\
35 & ScH$_{3}$ & $ P6_3/mmc $ & 1 & 400 & - & \cite{111}\\
36 & ScH$_{2}$ & $ P6/mmm $ & 4 & 300 & - & \cite{111}\\
37 & ScH$_{6}$ & $ P6_3/mmc $ & 119 & 130 & 0.857 & \cite{111}\\
38 & TiD$_{0.74}$ & $ - $ & 4.43 & 30 & - & \cite{116}\\
39 & TiH$_{2}$ & $ Fm\bar{3}m $ & 7 & 0 & 0.666 & \cite{117} \\
40 & TiH$_{2}$ & $ I4/mmm $ & 0.002 & 0 & 0.666 & \cite{117} \\
41 & ZrH & $ Cmcm $ & 11 & 120 & 0.5 & \cite{118} \\
42 & HfH$_{2}$ & $ I4/mmm $ & 0 & 0 & 0.666 & \cite{119} \\
43 & HfH$_{2}$ & $ Cmma $ & 6 & 180 & 0.666 & \cite{119} \\
44 & HfH$_{2}$ & $ P2_1/m $ & 12 & 260 & 0.666 & \cite{119} \\
45 & NbH$_{2}$ & $ P6_3mc $ & 0.5 & 60 & 0.666 & \cite{120} \\
46 & VH$_{2}$ & $ Fm\bar{3}m $ & 0.5 & 0 & 0.666 & \cite{120} \\
47 & NbH$_{2}$ & $ Fm\bar{3}m $ & 1.5 & 0 & 0.666 & \cite{120} \\
48 & NbH$_{2}$ & $ Pnma $ & 4 & 60 & 0.666 & \cite{120} \\
49 & NbH$_{4}$ & $ I4/mmm $ & 47 & 300 & 0.8 & \cite{121} \\
50 & TaH$_{2}$ & $ Pnma $ & 7.1 & 200 & 0.666 & \cite{122} \\
51 & TaH$_{4}$ & $ R\bar{3}m $ & 31 & 250 & 0.8 & \cite{122} \\
52 & TaH$_{6}$ & $ Fdd2 $ & 138.1 & 300 & 0.857 & \cite{122} \\
53 & CrH & $ P6_3/mmc $ & 10.6 & 0 & 0.5 & \cite{123} \\
54 & CrH$_{3}$ & $ P6_3/mmc $ & 37.1 & 81 & 0.75 & \cite{123} \\
55 & TcH$_{2}$ & $ I4/mmm $ & 10.64 & 200 & 0.666 & \cite{124} \\
56 & TcH$_{2}$ & $ Cmcm $ & 8.61 & 300 & 0.666 & \cite{124} \\
57 & TcH$_{3}$ & $ P4_2/mmc $ & 9.94 & 300 & 0.75 & \cite{124} \\
58 & FeH$_{6}$ & $ Cmmm $ & 42.9 & 150 & 0.857 & \cite{125} \\
59 & FeH$_{5}$ & $ I4/mmm $ & 42.6 & 150 & 0.833 & \cite{125} \\
60 & FeH$_{5}$ & $ I4/mmm $ & 51 & 130 & - & \cite{126} \\
61 & RuH & $ Fm\bar{3}m $ & 0.41 & 100 & 0.5 & \cite{127} \\
62 & RuH$_{3}$ & $ Pm\bar{3}m $ & 3.57 & 125 & 0.75 & \cite{127} \\
63 & RuH$_{3}$ & $ Pm\bar{3}n $ & 1.25 & 200 & 0.75 & \cite{127} \\
64 & OsH & $ Fm\bar{3}m $ & 2.1 & 100 & 0.5 & \cite{128} \\
65 & CoH & $ Fm\bar{3}m $ & 0.11 & 5 & 0.5 & \cite{129} \\
66 & R$_H$H & $ Fm\bar{3}m $ & 2.5 & 4 & - & \cite{scList} \\
67 & IrH & $ Fm\bar{3}m $ & 7 & 8 & - & \cite{scList} \\
68 & PdH & $ - $ & 9 & 0 & 0.5 & \cite{130} \\
69 & PdD & $ - $ & 11 & 0 & - & \cite{130} \\
70 & PdH & $ Fm\bar{3}m $ & 5 & 0 & 0.5 & \cite{131} \\
71 & PdD & $ Fm\bar{3}m $ & 6.5 & 0 & - & \cite{131} \\
72 & PdT & $ Fm\bar{3}m $ & 6.9 & 0 & - & \cite{131} \\
73 & PtH & $ Fm\bar{3}m $ & 15 & 100 & 0.5 & \cite{131} \\
74 & PtH & $ P6_3/mmc $ & 25 & 80 & 0.5 & \cite{131} \\
75 & AuH & $ Fm\bar{3}m $ & 21 & 200 & 0.5 & \cite{scList} \\
76 & BH & $ P6/mmm $ & 21.4 & 170 & 0.5 & \cite{132} \\
77 & BH$_{3}$ & $ Pbcn $ & 125 & 360 & 0.75 & \cite{133} \\
78 & AlH$_{3}$ & $ Pm\bar{3}n $ & 11.5 & 73 & 0.75 & \cite{134,135} \\
79 & AlH$_{3}$(H$_{2}$) & $ P2_1/m $ & 146 & 250 & 0.833 & \cite{136} \\
80 & GaH$_{3}$ & $ Pm\bar{3}n $ & 102 & 120 & 0.75 & \cite{137} \\
81 & InH$_{5}$ & $ P2_1/m $ & 27.1 & 150 & 0.833 & \cite{138} \\
82 & InH$_{3}$ & $ R\bar{3} $ & 40.5 & 200 & 0.75 & \cite{138} \\
83 & SiH$_{4}$ & $ C2/c $ & 55 & 125 & 0.8 & \cite{139} \\
84 & SiH$_{4}$ & $ P6/mmm $ & 74 & 120 & 0.8 & \cite{140} \\
85 & SiH$_{4}$ & $ Pmna $ & 166 & 202 & - & \cite{141} \\
86 & SiH$_{4}$ & $ C2/c $ & 30 & 300 & - & \cite{142} \\
87 & SiH$_{4}$ & $ P2_1/c $ & 35 & 400 & 0.8 & \cite{142} \\
88 & SiH$_{4}$ & $ C2/m $ & 110 & 610 & 0.8 & \cite{142} \\
89 & SiH$_{4}$ & $ P\bar{3} $ & 35.1 & 300 & 0.8 & \cite{142} \\
90 & SiH$_{4}$ & $ Cmca $ & 20 & 150 & - & \cite{143} \\
91 & SiH$_{4}$ & $ Pbcn $ & 16.5 & 190 & 0.8 & \cite{144} \\
92 & SiH$_{4}$H$_{2}$ & $ Cmca $ & 107 & 250 & - & \cite{145} \\
93 & Si$_2$H$_{6}$ & $ P\bar{1} $ & 80 & 200 & 0.75 & \cite{146} \\
94 & Si$_2$H$_{6}$ & $ Pm\bar{3}m $ & 153 & 275 & 0.75 & \cite{146} \\
95 & Si$_2$H$_{6}$ & $ C2/c $ & 42 & 300 & 0.75 & \cite{146} \\
96 & Si$_2$H$_{6}$ & $ Cmcm $ & 25 & 220 & - & \cite{146} \\
97 & GeH$_{4}$ & $ C2/c $ & 64 & 220 & 0.8 & \cite{147} \\
98 & GeH$_{4}$ & $ Cmmm $ & 47 & 20 & - & \cite{148} \\
99 & GeH$_{4}$ & $ Ama2 $ & 57 & 250 & 0.8 & \cite{149} \\
100 & GeH$_{4}$ & $ C2/c $ & 84 & 500 & 0.8 & \cite{149} \\
101 & GeH$_{4}$(H$_{2}$)2 & $ P2_1/c $ & 90 & 250 & 0.888 & \cite{150} \\
102 & GeH$_{3}$ & $ Cccm $ & 80 & 300 & - & \cite{151} \\
103 & GeH$_{4}$ & $ C2/m $ & 67 & 280 & 0.8 & \cite{152} \\
104 & Ge3H$_{11}$ & $ I\bar{4}m2 $ & 43 & 295 & 0.785 & \cite{152} \\
105 & GeH$_{3}$ & $ P4_2/mmc $ & 90 & 180 & 0.75 & \cite{153} \\
106 & GeH$_{3}$ & $ Pm\bar{3}m $ & 140 & 180 & 0.75 & \cite{153} \\
107 & SnH$_{8}$ & $ I\bar{4}m2 $ & 72 & 250 & 0.888 & \cite{154} \\
108 & SnH$_{4}$ & $ Ama2 $ & 22 & 120 & 0.8 & \cite{155} \\
109 & SnH$_{4}$ & $ P6_3/mmc $ & 62 & 200 & 0.8 & \cite{155} \\
110 & SnH$_{4}$ & $ P6/mmm $ & 80 & 120 & - & \cite{156} \\
111 & SnH$_{4}$ & $ C2/m $ & 95 & 600 & 0.8 & \cite{157} \\
112 & SnH$_{4}$ & $ I4/mmm $ & 92 & 220 & - & \cite{158} \\
113 & SnH$_{12}$ & $ C2/m $ & 93 & 250 & 0.923 & \cite{158} \\
114 & SnH$_{14}$ & $ C2/m $ & 97 & 300 & 0.933 & \cite{158} \\
115 & PbH$_{4}$(H$_{2}$)2 & $ C2/m $ & 107 & 130 & 0.888 & \cite{158} \\
116 & PH$_{3}$ & $ - $ & 100 & 226 & - & \cite{159} \\
117 & PH & $ I4/mmm $ & 81 & 230 & - & \cite{160} \\
118 & PH$_{2}$ & $ I4/mmm $ & 86 & 230 & - & \cite{160,161} \\
119 & PH$_{4}$ & $ C2/m $ & 1.9 & 80 & 0.8 & \cite{162} \\
120 & PH$_{2}$ & $ Cmmm $ & 29.5 & 80 & 0.666 & \cite{163} \\
121 & AsH & $ Cmcm $ & 21.2 & 300 & 0.5 & \cite{164} \\
122 & AsH$_{8}$ & $ C2/c $ & 151.4 & 450 & 0.888 & \cite{164} \\
123 & SbH & $ Pnma $ & 14.6 & 175 & 0.5 & \cite{164} \\
124 & SbH$_{3}$ & $ Pmmn $ & 25.9 & 300 & 0.875 & \cite{164} \\
125 & SbH$_{4}$ & $ P6_3/mmc $ & 102.2 & 150 & 0.8 & \cite{164} \\
126 & BiH$_{2}$ & $ Pnma $ & 39 & 125 & 0.666 & \cite{165} \\
127 & BiH$_{3}$ & $ I4_1/amd $ & 65 & 270 & 0.75 & \cite{165} \\
128 & SbH$_{3}$ & $ Pnma $ & 68 & 170 & 0.75 & \cite{165} \\
129 & BiH & $ P6_3/mmc $ & 30 & 250 & 0.5 & \cite{166} \\
130 & BiH$_{2}$ & $ P2_1/m $& 65 & 300 & 0.666 & \cite{166} \\
131 & BiH$_{4}$ & $ Pmmm $ & 93 & 150 & 0.8 & \cite{166} \\
132 & BiH$_{5}$ & $ C2/m $ & 119 & 300 & - & \cite{166} \\
133 & BiH$_{6}$ & $ P\bar{1} $ & 113 & 300 & 0.857 & \cite{166} \\
134 & H$_{2}$S & $ Cmca $ & 82 & 258 & - & \cite{167} \\
135 & H$_{2}$S & $ P\bar{1} $ & 60 & 258 & - & \cite{167} \\
136 & (H$_{2}$S)2H$_{2}$ & $ Im\bar{3}m $ & 204 & 200 & 0.75 & \cite{168} \\
137 & (H$_{2}$S)2H$_{2}$ & $ R\bar{3}m $ & 166 & 130 & - & \cite{168} \\
138 & H$_{3}$S & $ Im\bar{3}m $ & 225 & 150 & 0.75 & \cite{169} \\
139 & H$_{3}$S & $ R\bar{3}m $ & 214 & 170 & - & \cite{170} \\
140 & H$_{4}$S$_3$ & $ Pnma $ & 2.1 & 140 & 0.571 & \cite{171} \\
141 & H$_{5}$S$_2$ & $ P\bar{1} $ & 79 & 130 & 0.714 & \cite{172} \\
142 & D$_3$S & $ Im\bar{3}m $ & 188 & 200 & - & \cite{173} \\
143 & H$_{3}$S$_{0.925}$P$_{0.075}$ & $ Im\bar{3}m $ & 280 & 250 & - & \cite{174} \\
144 & H$_{3}$S$_{0.9}$P$_{0.1}$ & $ Im\bar{3}m $ & 240 & 200 & - & \cite{174} \\
145 & H$_{3}$S$_{0.96}$Si$_{0.04}$ & $ Im\bar{3}m $ & 275 & 250 & - & \cite{174} \\
146 & HSe$_2$ & $ C2/m $ & 5 & 300 & 0.333 & \cite{175} \\
147 & HSe & $ P4/nmm $ & 42 & 300 & 0.5 & \cite{175} \\
148 & H$_{3}$Se & $ Im\bar{3}m $ & 110 & 200 & 0.75 & \cite{175} \\
149 & HSe & $ P2_1/c $ & 32 & 300 & 0.5 & \cite{175} \\
150 & H$_{4}$Te & $ P6/mmm $ & 104 & 170 & 0.8 & \cite{176} \\
151 & H$_{5}$Te$_2$ & $ C2/m $ & 58 & 200 & 0.714 & \cite{176} \\
152 & HTe & $ P4/nmm $& 28 & 150 & 0.5 & \cite{176} \\
153 & H$_{4}$Te & $ R\bar{3}m $ & 76 & 270 & 0.8 & \cite{176} \\
154 & HTe & $ P6_3/mmc $ & 44.2 & 300 & 0.8 & \cite{176} \\
155 & PoH$_{4}$ & $ C2/c $ & 53.6 & 250 & 0.8 & \cite{177} \\
156 & PoH & $ P6_3/mmc $ & 0.65 & 300 & 0.5 & \cite{177} \\
157 & PoH$_{2}$ & $ Pnma $ & 0 & 200 & 0.666 & \cite{177} \\
158 & PoH$_{6}$ & $ C2/m $ & 4.68 & 200 & 0.857 & \cite{177} \\
159 & HBr & $ P2_1/m $ & 51 & 200 & 0.5 & \cite{178} \\
160 & HCl & $ P2_1/m $ & 40 & 360 & 0.5 & \cite{179} \\
161 & HBr & $ C2/m $ & 27 & 150 & 0.5 & \cite{180} \\
162 & HCl & $ C2/m $ & 20 & 250 & 0.5 & \cite{181} \\
163 & H$_{2}$I & $ Cmcm $ & 8 & 100 & 0.666 & \cite{182} \\
164 & H$_{4}$I & $ P6/mmm $ & 12.5 & 300 & 0.8 & \cite{183} \\
165 & H$_{2}$I & $ Pnma $ & 5.3 & 100 & 0.666 & \cite{183} \\
166 & H$_{2}$I & $ R\bar{3}m $ & 33 & 240 & 0.666 & \cite{183} \\
167 & XeH & $ Immm $ & 28 & 100 & - & \cite{184} \\
168 & XeH$_{2}$ & $ Cmcm $ & 26 & 400 & - & \cite{184} \\
169 & MgH$_{2}$ & $ P6_3/mmc $ & 23 & 180 & 0.666 & \cite{185} \\
170 & MgH$_{4}$ & $ Cmcm $ & 37 & 100 & 0.8 & \cite{185} \\
171 & MgH$_{12}$ & $ R\bar{3} $ & 60 & 140 & 0.923 & \cite{185} \\
172 & PH$_{3}$ & $ C2/m $ & 81 & 200 & - & \cite{186,187} \\
173 & PH$_{2}$ & $ C2/m $ & 75 & 200 & 0.666 & \cite{186,187} \\
174 & H$_{2}$Br & $ Cmcm $ & 12.1 & 240 & - & \cite{188} \\
175 & H$_{4}$Br & $ P6_3/mmc $ & 2.4 & 240 & - & \cite{188} \\
176 & H & $ I4_1/amd $ & 300 & 500 & 1 & \cite{189} \\
177 & H & $ Cmca\bar{4} $ & 109 & 450 & 1 & \cite{190}
\end{longtable}
\clearpage
\begin{longtable}{c|c|c|c|c}
\caption{Table reporting the item index, chemical formula, {\it networking value} $\phi$, dominant bonding family assigned, and shortest H-H distance for all the compounds. The empty cells refer to missing values from the literature or to the impossibility to perform calculations. } \\
\label{table:Sup2}
Item & Chemical formula & {\it Networking value} $\phi$ & Dominant bonding family & H-H Distance (\AA) \\
\hline
0 & LiH$_{2}$ & 0.24 & Molecular & 0.759 \\
1 & LiH$_{2}$ & 0.37 & Molecular & 0.803 \\
2 & LiH$_{8}$ & 0.22 & Molecular & 0.844 \\
3 & KH$_{6}$ & 0.19 & Molecular & 0.887 \\
4 & BeH$_{2}$ & 0.35 & Ionic & 1.404 \\
5 & BeH$_{2}$ & 0.4 & Ionic & 1.368 \\
6 & MgH$_{6}$ & 0.63 & Weak H Interactions & 1.093 \\
7 & CaH$_{6}$ & 0.62 & Weak H Interactions & 1.237 \\
8 & SrH$_{6}$ & 0.33 & Weak H Interactions & 1.033 \\
10 & BaH$_{6}$ & - & - & - \\
9 & BaH$_{2}$ & - & - & - \\
11 & ScH$_{3}$ & - & - & - \\
12 & LaH$_{3}$ & - & - & - \\
13 & YH$_{3}$ & - & - & - \\
14 & ScH$_{2}$ & 0.33 & Ionic & 2.091 \\
15 & YH$_{4}$ & 0.43 & Weak H Interactions & 1.352 \\
16 & YH$_{6}$ & 0.58 & Weak H Interactions & 1.301 \\
17 & ScH$_{4}$ & 0.48 & Weak H Interactions & 1.198 \\
18 & ScH$_{6}$ & 0.6 & Weak H Interactions & 1.112 \\
19 & PrH$_{9}$ & 0.49 & Weak H Interactions & 1.028 \\
20 & CeH$_{10}$ & 0.51 & Weak H Interactions & 0.943 \\
21 & CeH$_{9}$ & 0.51 & Weak H Interactions & 0.909 \\
22 & LaH$_{9}$ & 0.13 & Molecular & 0.550 \\
23 & LaH$_{6}$ & - & - & 0.944 \\
24 & YH$_{9}$ & 0.57 & Weak H Interactions & 1.045 \\
25 & ScH$_{9}$ & 0.65 & Weak H Interactions & 0.983 \\
27 & YH$_{10}$ & 0.58 & Weak H Interactions & 1.047 \\
28 & LaH$_{4}$ & - & - & 1.189 \\
29 & LaH$_{8}$ & 0.52 & Weak H Interactions & 0.890 \\
30 & LaH$_{10}$ & 0.52 & Weak H Interactions & 0.840 \\
31 & ScH$_{9}$ & 0.59 & Molecular & 0.848 \\
32 & ScH$_{10}$ & 0.33 & Molecular & 0.866 \\
33 & ScH$_{12}$ & 0.33 & Molecular & 0.909 \\
34 & ScH$_{7}$ & 0.49 & Molecular & 0.960 \\
35 & ScH$_{3}$ & - & - & - \\
36 & ScH$_{2}$ & - & - & - \\
37 & ScH$_{6}$ & 0.26 & Molecular & 1.022 \\
38 & TiD$_{0.74}$ & - & - & - \\
39 & TiH$_{2}$ & 0.29 & Ionic & 2.210 \\
40 & TiH$_{2}$ & 0.38 & Ionic & 1.628 \\
41 & ZrH & 0.32 & Ionic & 2.753 \\
42 & HfH$_{2}$ & 0.23 & Ionic & 2.149 \\
43 & HfH$_{2}$ & 0.23 & Molecular & 0.765 \\
44 & HfH$_{2}$ & 0.27 & Isolated & 1.359 \\
45 & NbH$_{2}$ & 0.29 & Electride & 2.009 \\
46 & VH$_{2}$ & 0.3 & Ionic & 2.105 \\
47 & NbH$_{2}$ & 0.27 & Ionic & 2.277 \\
48 & NbH$_{2}$ & 0.36 & Electride & 1.955 \\
49 & NbH$_{4}$ & 0.27 & Weak H Interactions & 1.209 \\
50 & TaH$_{2}$ & 0.41 & Electride & 1.728 \\
51 & TaH$_{4}$ & 0.27 & Isolated & 1.313 \\
52 & TaH$_{6}$ & 0.38 & Weak H Interactions & 1.085 \\
53 & CrH & 0.4 & Electride & 2.171 \\
54 & CrH$_{3}$ & 0.35 & Isolated & 1.615 \\
55 & TcH$_{2}$ & - & - & 1.700 \\
56 & TcH$_{2}$ & - & - & 1.241 \\
57 & TcH$_{3}$ & 0.31 & Isolated & 1.299 \\
58 & FeH$_{6}$ & 0.36 & Molecular & 0.733 \\
59 & FeH$_{5}$ & 0.17 & Weak H Interactions & 1.300 \\
60 & FeH$_{5}$ & - & - & - \\
61 & RuH & 0.28 & Isolated & 2.626 \\
62 & RuH$_{3}$ & 0.29 & Isolated & 1.807 \\
63 & RuH$_{3}$ & 0.29 & Isolated & 1.542 \\
64 & OsH & 0.33 & Ionic & 2.676 \\
65 & CoH & 0.27 & Metallic & 2.570 \\
66 & R$_H$H & - & - & - \\
67 & IrH & - & - & - \\
68 & PdH & 0.19 & Isolated & 2.911 \\
69 & PdD & - & - & - \\
70 & PdH & 0.19 & Isolated & 2.911 \\
71 & PdD & - & - & - \\
72 & PdT & - & - & - \\
73 & PtH & 0.27 & Isolated & 2.723 \\
74 & PtH & 0.27 & Ionic & 2.563 \\
75 & AuH & 0.27 & Isolated & 2.667 \\
76 & BH & 0.51 & Covalent & 10000 \\
77 & BH$_{3}$ & 0.56 & Covalent & 1.024 \\
78 & AlH$_{3}$ & 0.36 & Covalent & 1.593 \\
79 & AlH$_{3}$(H$_{2}$) & - & - & 0.842 \\
80 & GaH$_{3}$ & 0.37 & Covalent & 1.531 \\
81 & InH$_{5}$ & 0.27 & Molecular & 0.770 \\
82 & InH$_{3}$ & 0.28 & Molecular & 0.884 \\
83 & SiH$_{4}$ & 0.27 & Covalent & 1.435 \\
84 & SiH$_{4}$ & 0.51 & Covalent & 1.081 \\
85 & SiH$_{4}$ & - & - & - \\
86 & SiH$_{4}$ & - & - & - \\
87 & SiH$_{4}$ & 0.45 & Covalent & 1.043 \\
88 & SiH$_{4}$ & 0.49 & Covalent & 1.009 \\
89 & SiH$_{4}$ & 0.49 & Molecular & 0.750 \\
90 & SiH$_{4}$ & - & - & - \\
91 & SiH$_{4}$ & 0.32 & Covalent & 1.372 \\
92 & SiH$_{4}$H$_{2}$ & - & - & - \\
93 & Si$_2$H$_{6}$ & 0.48 & Covalent & 1.247 \\
94 & Si$_2$H$_{6}$ & 0.58 & Electride & 1.592 \\
95 & Si$_2$H$_{6}$ & 0.48 & Molecular & 0.592 \\
96 & Si$_2$H$_{6}$ & - & - & - \\
97 & GeH$_{4}$ & 0.39 & Molecular & 0.878 \\
98 & GeH$_{4}$ & - & - & - \\
99 & GeH$_{4}$ & 0.34 & Molecular & 0.803 \\
100 & GeH$_{4}$ & 0.4 & Covalent & 0.973 \\
101 & GeH$_{4}$(H$_{2}$)2 & 0.37 & Molecular & 0.845 \\
102 & GeH$_{3}$ & - & - & - \\
103 & GeH$_{4}$ & 0.33 & Covalent & 1.057 \\
104 & Ge3H$_{11}$ & 0.33 & Covalent & 1.281 \\
105 & GeH$_{3}$ & 0.31 & Covalent & 1.360 \\
106 & GeH$_{3}$ & 0.32 & Covalent & 1.553 \\
107 & SnH$_{8}$ & 0.24 & Molecular & 0.860 \\
108 & SnH$_{4}$ & 0.22 & Molecular & 0.792 \\
109 & SnH$_{4}$ & 0.24 & Molecular & 0.816 \\
110 & SnH$_{4}$ & - & - & - \\
111 & SnH$_{4}$ & 0.29 & Molecular & 0.807 \\
112 & SnH$_{4}$ & - & - & - \\
113 & SnH$_{12}$ & 0.34 & Molecular & 0.754 \\
114 & SnH$_{14}$ & 0.37 & Molecular & 0.783 \\
115 & PbH$_{4}$(H$_{2}$)2 & 0.28 & Molecular & 0.785 \\
116 & PH$_{3}$ & - & - & - \\
117 & PH & - & - & - \\
118 & PH$_{2}$ & - & - & - \\
119 & PH$_{4}$ & 0.19 & Molecular & 0.740 \\
120 & PH$_{2}$ & 0.15 & Molecular & 0.741 \\
121 & AsH & 0.45 & Covalent & 1.580 \\
122 & AsH$_{8}$ & 0.39 & Molecular & 0.813 \\
123 & SbH & 0.35 & Covalent & 1.905 \\
124 & SbH$_{3}$ & 0.27 & Weak H Interactions & 0.882 \\
125 & SbH$_{4}$ & 0.29 & Molecular & 0.842 \\
126 & BiH$_{2}$ & 0.32 & Molecular & 0.805 \\
127 & BiH$_{3}$ & 0.31 & Weak H Interactions & 0.924 \\
128 & SbH$_{3}$ & 0.35 & Molecular & 0.878 \\
129 & BiH & 0.29 & Isolated & 3.033 \\
130 & BiH$_{2}$ & 0.31 & Molecular & 0.812 \\
131 & BiH$_{4}$ & 0.29 & Molecular & 0.834 \\
132 & BiH$_{5}$ & - & - & - \\
133 & BiH$_{6}$ & 0.37 & Molecular & 0.905 \\
134 & H$_{2}$S & - & - & - \\
135 & H$_{2}$S & - & - & - \\
136 & (H$_{2}$S)2H$_{2}$ & 0.69 & Covalent & 1.493 \\
137 & (H$_{2}$S)2H$_{2}$ & - & - & - \\
138 & H$_{3}$S & 0.7 & Covalent & 1.532 \\
139 & H$_{3}$S & - & - & - \\
140 & H$_{4}$S$_3$ & 0.44 & Covalent & 1.554 \\
141 & H$_{5}$S$_2$ & 0.56 & Covalent & 1.543 \\
142 & D$_3$S & - & - & - \\
143 & H$_{3}$S$_{0.925}$P$_{0.075}$ & - & - & - \\
144 & H$_{3}$S$_{0.9}$P$_{0.1}$ & - & - & - \\
145 & H$_{3}$S$_{0.96}$Si$_{0.04}$ & - & - & - \\
146 & HSe$_2$ & 0.44 & Covalent & 1.762 \\
147 & HSe & 0.52 & Covalent & 2.133 \\
148 & H$_{3}$Se & 0.53 & Covalent & 1.571 \\
149 & HSe & 0.52 & Covalent & 1.277 \\
150 & H$_{4}$Te & 0.4 & Molecular & 0.844 \\
151 & H$_{5}$Te$_2$ & 0.41 & Weak H Interactions & 0.916 \\
152 & HTe & 0.41 & Covalent & 2.465 \\
153 & H$_{4}$Te & 0.35 & Molecular & 0.832 \\
154 & HTe & 0.37 & Molecular & 0.834 \\
155 & PoH$_{4}$ & 0.34 & Molecular & 0.813 \\
156 & PoH & 0.34 & Ionic & 2.335 \\
157 & PoH$_{2}$ & 0.35 & Molecular & 0.792 \\
158 & PoH$_{6}$ & 0.36 & Molecular & 0.797 \\
159 & HBr & 0.37 & Covalent & 1.411 \\
160 & HCl & 0.37 & Covalent & 1.266 \\
161 & HBr & 0.38 & Covalent & 1.492 \\
162 & HCl & 0.38 & Covalent & 1.403 \\
163 & H$_{2}$I & 0.36 & Molecular & 0.792 \\
164 & H$_{4}$I & 0.45 & Molecular & 0.799 \\
165 & H$_{2}$I & 0.36 & Molecular & 0.797 \\
166 & H$_{2}$I & 0.35 & Molecular & 0.797 \\
167 & XeH & - & - & - \\
168 & XeH$_{2}$ & - & - & - \\
169 & MgH$_{2}$ & 0.33 & Ionic & 1.763 \\
170 & MgH$_{4}$ & 0.26 & Molecular & 0.776 \\
171 & MgH$_{12}$ & 0.36 & Molecular & 0.819 \\
172 & PH$_{3}$ & - & - & - \\
173 & PH$_{2}$ & 0.32 & Covalent & 1.491 \\
174 & H$_{2}$Br & - & - & - \\
175 & H$_{4}$Br & - & - & - \\
176 & H & 0.64 & Weak H Interactions & 0.985 \\
177 & H & 0.35 & Molecular & 0.779
\end{longtable}
\clearpage
\begin{longtable}{c|c|c|c|c}
\caption{ Table reporting the item index, chemical formula, total DOS per electron, hydrogen contribution to the DOS per hydrogen atom, and hydrogen fraction of the DOS at the Fermi energy. The empty cells refer to missing values from the literature or to the impossibility to perform calculations.} \\
\label{table:Sup3}
Item & Chemical formula & Total DOS ($eV^{-1}$) & Hydrogen DOS ($eV^{-1}$) & Hydrogen DOS fraction \\
\hline
0 & LiH$_{2}$ & 0.00298 & 0.00709 & 0.94924 \\
1 & LiH$_{2}$ & 0.02220 & 0.02933 & 0.88063 \\
2 & LiH$_{8}$ & 0.03063 & 0.039 & 0.92588 \\
3 & KH$_{6}$ & 0.05004 & 0.1429 & 0.87854 \\
4 & BeH$_{2}$ & 0.01675 & 0.02111 & 0.42005 \\
5 & BeH$_{2}$ & 0.01743 & 0.02091 & 0.39974 \\
6 & MgH$_{6}$ & 0.02064 & 0.0408 & 0.74105 \\
7 & CaH$_{6}$ & 0.01872 & 0.0456 & 0.91338 \\
8 & SrH$_{6}$ & 0.01656 & 0.038 & 0.86045 \\
9 & BaH$_{2}$ & - & - & - \\
10 & BaH$_{6}$ & - & - & - \\
11 & ScH$_{3}$ & - & - & - \\
12 & LaH$_{3}$ & - & - & - \\
13 & YH$_{3}$ & - & - & - \\
14 & ScH$_{2}$ & 0.03600 & 0.00638 & 0.02726 \\
15 & YH$_{4}$ & 0.03105 & 0.04555 & 0.39112 \\
16 & YH$_{6}$ & 0.04132 & 0.053 & 0.45269 \\
17 & ScH$_{4}$ & 0.03179 & 0.0356 & 0.29856 \\
18 & ScH$_{6}$ & 0.03896 & 0.0356 & 0.32246 \\
19 & PrH$_{9}$ & 0.01967 & 0.01129 & 0.16150 \\
20 & CeH$_{10}$ & 0.01466 & 0.01309 & 0.27900 \\
21 & CeH$_{9}$ & 0.02029 & 0.02036 & 0.29135 \\
22 & LaH$_{9}$ & 0.03072 & 0.01651 & 0.16123 \\
23 & LaH$_{6}$ & - & - & - \\
24 & YH$_{9}$ & 0.02762 & 0.04033 & 0.65701 \\
25 & ScH$_{9}$ & 0.02564 & 0.03135 & 0.55013 \\
27 & YH$_{10}$ & 0.03547 & 0.04992 & 0.67017 \\
28 & LaH$_{4}$ & - & - & - \\
29 & LaH$_{8}$ & 0.01605 & 0.02265 & 0.38922 \\
30 & LaH$_{10}$ & 0.02601 & 0.04064 & 0.50395 \\
31 & ScH$_{9}$ & 0.02952 & 0.03491 & 0.53215 \\
32 & ScH$_{10}$ & 0.02176 & 0.02453 & 0.53668 \\
33 & ScH$_{12}$ & 0.02376 & 0.02673 & 0.58697 \\
34 & ScH$_{7}$ & 0.02505 & 0.03605 & 0.55962 \\
35 & ScH$_{3}$ & - & - & - \\
36 & ScH$_{2}$ & - & - & - \\
37 & ScH$_{6}$ & 0.02758 & 0.0336 & 0.42992 \\
38 & TiD$_{0.74}$ & - & - & - \\
39 & TiH$_{2}$ & 0.17927 & 0.00066 & 0.00053 \\
40 & TiH$_{2}$ & 0.05533 & 0.01472 & 0.03800 \\
41 & ZrH & 0.07636 & 0.01206 & 0.01214 \\
42 & HfH$_{2}$ & 0.01242 & 0.00186 & 0.00787 \\
43 & HfH$_{2}$ & 0.01072 & 0.02352 & 0.11542 \\
44 & HfH$_{2}$ & 0.00907 & 0.01225 & 0.07105 \\
45 & NbH$_{2}$ & 0.03424 & 0.00216 & 0.00844 \\
46 & VH$_{2}$ & 0.08699 & 0.00352 & 0.00539 \\
47 & NbH$_{2}$ & 0.06102 & 0.00306 & 0.00668 \\
48 & NbH$_{2}$ & 0.06066 & 0.00522 & 0.01147 \\
49 & NbH$_{4}$ & 0.04581 & 0.021 & 0.10784 \\
50 & TaH$_{2}$ & 0.02064 & 0.0091 & 0.03040 \\
51 & TaH$_{4}$ & 0.01307 & 0.02355 & 0.23240 \\
52 & TaH$_{6}$ & 0.01519 & 0.0403 & 0.48226 \\
53 & CrH & 0.06706 & 0.0064 & 0.00636 \\
54 & CrH$_{3}$ & 0.07060 & 0.0282 & 0.07048 \\
55 & TcH$_{2}$ & - & - & - \\
56 & TcH$_{2}$ & - & - & - \\
57 & TcH$_{3}$ & 0.03144 & 0.0138 & 0.07314 \\
58 & FeH$_{6}$ & 0.02460 & 0.01662 & 0.18419 \\
59 & FeH$_{5}$ & 0.02408 & 0.00616 & 0.06091 \\
60 & FeH$_{5}$ & - & - & - \\
61 & RuH & 0.04832 & 0.0133 & 0.01618 \\
62 & RuH$_{3}$ & 0.04596 & 0.024 & 0.08244 \\
63 & RuH$_{3}$ & 0.02571 & 0.022 & 0.13508 \\
64 & OsH & 0.02037 & 0.00498 & 0.00788 \\
65 & CoH & 0.24733 & 0.02135 & 0.00479 \\
66 & R$_H$H & - & - & - \\
67 & IrH & - & - & - \\
68 & PdH & 0.02481 & 0.0745 & 0.15803 \\
69 & PdD & - & - & - \\
70 & PdH & 0.02481 & 0.0745 & 0.15803 \\
71 & PdD & - & - & - \\
72 & PdT & - & - & - \\
73 & PtH & 0.01163 & 0.0562 & 0.14640 \\
74 & PtH & 0.01735 & 0.047 & 0.08208 \\
75 & AuH & 0.02019 & 0.01531 & 0.02230 \\
76 & BH & 0.02755 & 0.03 & 0.27223 \\
77 & BH$_{3}$ & 0.02614 & 0.028 & 0.53550 \\
78 & AlH$_{3}$ & 0.03379 & 0.0458 & 0.67764 \\
79 & AlH$_{3}$(H$_{2}$) & - & - & - \\
80 & GaH$_{3}$ & 0.01841 & 0.0208 & 0.56491 \\
81 & InH$_{5}$ & 0.02133 & 0.02644 & 0.34432 \\
82 & InH$_{3}$ & 0.02221 & 0.03366 & 0.28419 \\
83 & SiH$_{4}$ & 0.03007 & 0.0359 & 0.59684 \\
84 & SiH$_{4}$ & 0.05725 & 0.0464 & 0.40524 \\
85 & SiH$_{4}$ & - & - & - \\
86 & SiH$_{4}$ & - & - & - \\
87 & SiH$_{4}$ & 0.02126 & 0.02338 & 0.54972 \\
88 & SiH$_{4}$ & 0.03087 & 0.02732 & 0.44245 \\
89 & SiH$_{4}$ & 0.03394 & 0.01705 & 0.25117 \\
90 & SiH$_{4}$ & - & - & - \\
91 & SiH$_{4}$ & 0.0195 & 0.0243 & 0.62307 \\
92 & SiH$_{4}$H$_{2}$ & - & - & - \\
93 & Si$_2$H$_{6}$& 0.03554 & 0.0297 & 0.35811 \\
94 & Si$_2$H$_{6}$& 0.04522 & 0.0392 & 0.37144 \\
95 & Si$_2$H$_{6}$& 0.04734 & 0.03976 & 0.35998 \\
96 & Si$_2$H$_{6}$& - & - & - \\
97 & GeH$_{4}$ & 0.01901 & 0.0314 & 0.36688 \\
98 & GeH$_{4}$ & - & - & - \\
99 & GeH$_{4}$ & 0.01805 & 0.03362 & 0.41384 \\
100 & GeH$_{4}$ & 0.01439 & 0.03255 & 0.50236 \\
101 & GeH$_{4}$(H$_{2}$)2 & 0.01920 & 0.0279 & 0.52831 \\
102 & GeH$_{3}$ & - & - & - \\
103 & GeH$_{4}$ & 0.01449 & 0.02803 & 0.42974 \\
104 & Ge3H$_{11}$ & 0.01079 & 0.02473 & 0.47560 \\
105 & GeH$_{3}$ & 0.02187 & 0.03453 & 0.27853 \\
106 & GeH$_{3}$ & 0.03444 & 0.0484 & 0.24793 \\
107 & SnH$_{8}$ & 0.01736 & 0.02607 & 0.54617 \\
108 & SnH$_{4}$ & 0.01782 & 0.0241 & 0.30038 \\
109 & SnH$_{4}$ & 0.0223 & 0.0372 & 0.37070 \\
110 & SnH$_{4}$ & - & - & - \\
111 & SnH$_{4}$ & 0.01920 & 0.0331 & 0.38303 \\
112 & SnH$_{4}$ & - & - & - \\
113 & SnH$_{12}$ & 0.02066 & 0.02871 & 0.64144 \\
114 & SnH$_{14}$ & 0.02083 & 0.02999 & 0.71983 \\
115 & PbH$_{4}$(H$_{2}$)2 & 0.02688 & 0.03232 & 0.43717 \\
116 & PH$_{3}$ & - & - & - \\
117 & PH & - & - & - \\
118 & PH$_{2}$ & - & - & - \\
119 & PH$_{4}$ & 0.02710 & 0.01734 & 0.28436 \\
120 & PH$_{2}$ & 0.03010 & 0.01784 & 0.16935 \\
121 & AsH & 0.01920 & 0.052 & 0.16926 \\
122 & AsH$_{8}$ & 0.02111 & 0.03243 & 0.53430 \\
123 & SbH & 0.02155 & 0.0482 & 0.13976 \\
124 & SbH$_{3}$ & 0.01877 & 0.037 & 0.62715 \\
125 & SbH$_{4}$ & 0.03026 & 0.0465 & 0.32348 \\
126 & BiH$_{2}$ & 0.02709 & 0.0446 & 0.19364 \\
127 & BiH$_{3}$ & 0.02584 & 0.0537 & 0.34636 \\
128 & SbH$_{3}$ & 0.02447 & 0.0408 & 0.27787 \\
129 & BiH & 0.02447 & 0.09 & 0.22987 \\
130 & BiH$_{2}$ & 0.02626 & 0.0485 & 0.21725 \\
131 & BiH$_{4}$ & 0.02858 & 0.0402 & 0.29603 \\
132 & BiH$_{5}$ & - & - & - \\
133 & BiH$_{6}$ & 0.02549 & 0.04146 & 0.46471 \\
134 & H$_{2}$S & - & - & - \\
135 & H$_{2}$S & - & - & - \\
136 & (H$_{2}$S)2H$_{2}$ & 0.06064 & 0.0867 & 0.47654 \\
137 & (H$_{2}$S)2H$_{2}$ & - & - & - \\
138 & H$_{3}$S & 0.05348 & 0.0782 & 0.48732 \\
139 & H$_{3}$S & - & - & - \\
140 & H$_{4}$S$_3$ & 0.01120 & 0.01403 & 0.22762 \\
141 & H$_{5}$S$_2$ & 0.03182 & 0.0442 & 0.40847 \\
142 & D$_3$S & - & - & - \\
143 & H$_{3}$S$_{0.925}$P$_{0.075}$ & - & - & - \\
144 & H$_{3}$S$_{0.9}$P$_{0.1}$ & - & - & - \\
145 & H$_{3}$S$_{0.96}$Si$_{0.04}$ & - & - & - \\
146 & HSe$_2$ & 0.01736 & 0.0819 & 0.14294 \\
147 & HSe & 0.02240 & 0.0544 & 0.14283 \\
148 & H$_{3}$Se & 0.02796 & 0.0857 & 0.48394 \\
149 & HSe & 0.01874 & 0.0494 & 0.15501 \\
150 & H$_{4}$Te & 0.03724 & 0.0464 & 0.24914 \\
151 & H$_{5}$Te$_2$ & 0.02416 & 0.04592 & 0.25675 \\
152 & HTe & 0.02680 & 0.065 & 0.14264 \\
153 & H$_{4}$Te & 0.02674 & 0.0381 & 0.28495 \\
154 & HTe & 0.03070 & 0.0424 & 0.27614 \\
155 & PoH$_{4}$ & 0.02522 & 0.0359 & 0.28468 \\
156 & PoH & 0.02491 & 0.0864 & 0.20401 \\
157 & PoH$_{2}$ & 0.02682 & 0.0405 & 0.16772 \\
158 & PoH$_{6}$ & 0.03239 & 0.03846 & 0.32384 \\
159 & HBr & 0.01632 & 0.052 & 0.17697 \\
160 & HCl & 0.01352 & 0.0432 & 0.17748 \\
161 & HBr & 0.01836 & 0.0454 & 0.13735 \\
162 & HCl & 0.03575 & 0.0388 & 0.13565 \\
163 & H$_{2}$I & 0.02955 & 0.0269 & 0.09580 \\
164 & H$_{4}$I & 0.01606 & 0.0262 & 0.31063 \\
165 & H$_{2}$I & 0.02681 & 0.0298 & 0.11697 \\
166 & H$_{2}$I & 0.02684 & 0.042 & 0.16466 \\
167 & XeH & - & - & - \\
168 & XeH$_{2}$ & - & - & - \\
169 & MgH$_{2}$ & 0.00801 & 0.01701 & 0.35378 \\
170 & MgH$_{4}$ & 0.00576 & 0.01346 & 0.66683 \\
171 & MgH$_{12}$ & 0.01311 & 0.0221 & 0.91945 \\
172 & PH$_{3}$ & - & - & - \\
173 & PH$_{2}$ & 0.04305 & 0.061 & 0.40477 \\
174 & H$_{2}$Br & - & - & - \\
175 & H$_{4}$Br & - & - & - \\
176 & H & 0.03300 & 0.033 & 1 \\
177 & H & 0.01878 & 0.01878 & 1 \\
\end{longtable}
|
\section{Introduction} \label{sec:intro}
RV~Tauri-type variable stars are F through K supergiants with luminosity classes between Ia and II. They extend the brightest part of the Type II Cepheid (i.e., RR~Lyr, BL~Her, W~Vir, and RV~Tau stars) instability strip on the H-R diagram, as well as the brightest part of their period-luminosity relation, with radial pulsation periods longer than $20$~days \citep{Soszynski2017}. However, it has been shown that RV~Tauri variables follow a steeper period-luminosity relation compared to their shorter-period Type II Cepheid counterparts \citep[e.g.][]{Bodi2019}.
The signature characteristic of RV~Tauri variables are the alternating deep and shallow minima in their light curves: the time between a deep minimum and its successive shallower minimum is called the ``fundamental'' period; the time between two successive deep minima is the ``formal'' period. The alternation of deep and shallow minima is not strict in nature, as the light curves also show strong cycle-to-cycle variability, where the depths of the minima can vary randomly and may be attributed to low-dimensional chaotic behavior \citep[e.g.][and references therein]{Plachy2014,Plachy2018}. In addition to pulsations, a photometric subgroup (RVb-type) displays an additional, slower, periodic variation in mean brightness that ranges between 470 and 2800~days \citep{Soszynski2017}. There are $\sim$300 identified RV~Tauri variables in the Galaxy \citep{Soszynski2020}.
It has been established that most RV~Tauri variables are a subclass of post-asymptotic giant branch (post-AGB) stars in binary systems surrounded by a dusty circumbinary disk (CBD), and there is argument for the companion to be an unevolved main-sequence star \citep{Vanwinckel2009,Manick2019}. The post-AGB primary is evolved from a low-to-intermediate-mass progenitor (0.8--8~M$_{\odot}$), on their way to forming a planetary nebula, based on high luminosities ($\sim$10$^3$--$10^4$~L$_{\odot}$), mass-loss history, and infrared excess (indicative of dust) in their spectral energy distribution \citep[SED;][]{Gehrz1972,LloydEvans1985, Jura1986, Alcock1998,vanwinckel2003}. \citet{Groenewegen_Jurkovic2017} recently found mass estimates for RV~Tauri variables in the Magellanic Clouds that showed either very high ($\gtrsim$~1~M$_{\odot}$) or very low mass values ($\lesssim$0.5~M$_{\odot}$), which are in conflict with the standard single-star evolution of a post-AGB object, likely due to different evolutionary channels and perhaps revealing the effects of binarity. It may be that some of these post-AGB systems may evolve too slowly to ever become planetary nebulae \citep{vanWinckel2017}. \citet{Bodi2019} showed similar mass estimate results to those of \citet{Groenewegen_Jurkovic2017}, for a small sample of Galactic RV~Tauri variables.
It was recently discovered that some of the less luminous RV~Tauri variables are actually post-red giant branch (post-RGB) stars, a class of objects that have similar spectroscopic stellar parameters to their post-AGB counterparts but appear at lower luminosities on the H-R diagram \citep{Kamath2016, Manick2018}. These objects are thought to have evolved off the RGB instead of the AGB as a result of binary interaction.
The CBDs around post-AGB (and post-RGB) binaries are described as second-generation `scaled-up' analogs to proto-planetary disks surrounding young stellar objects \citep[YSOs;][]{Hillen2017,Vanwinckel2018}. The near-IR excess observed in the SEDs of these bright post-AGB binaries, especially the RV~Tauri variables, comes from hot dust close to the central source, where a disk-like nature of the near-IR emission has been confirmed by interferometric observations \citep{Deroo2006}. The mid- to far-IR excess then comes from cooler dust in the disk \citep{DeRuyter2006PostAGBs,Hillen2015ACHer}.
The CBDs are found to be relatively compact with sufficient angular momentum, in combination with the binary system's gravity, to settle and form the rotating disk from ejected material, especially from the huge mass loss after the AGB phase (up to $10^{-4}$~M$_\odot$~yr$^{-1}$) of the evolved primary \citep{Bujarrabal2013}. Keplerian rotation of the circumbinary material around these stars has been resolved by narrow CO line profiles, as well as the presence of large grains with a high degree of crystallinity, indicating longevity and stability, that form these disks \citep{Gielen2011} with diameters ranging between $\sim$100 and 2000~au \citep{Bujarrabal2005,Bujarrabal2013,Bujarrabal2018}.
The slower photometric phenomenon from the RVb subset of RV~Tauri systems (and in other post-AGB binaries) has been attributed to an extrinsic variable extinction. An inclined CBD can shadow the primary at certain phases of its orbit \citep[e.g.,][]{Manick2017}, blocking the light, and causing extinction and scattering along our line of sight. \citet{Vega_2017} used ultra-precise flux measurements from the \textit{Kepler} telescope and found that the decrease in pulsation amplitude of DF~Cyg (the only RV~Tauri system in \textit{Kepler}'s original field of view) perfectly tracked the decrease in flux from its RVb oscillation, showing that the long-term minima are due to the disk obscuring the pulsating primary. \citet{Kiss2017} further extended this study to all known RVbs in the Galaxy (by also adopting flux units instead of the inverse logarithmic magnitude system) to confirm that all RVbs displayed the correlation found in DF~Cyg. Furthermore, by using interferometric observations, \citet{Kluska2019} have found that the RVb sources in their sample have high inclinations, implying a very high disk scale height and allowing the disk-shadowing interpretation to be correct.
The presence of disks has also been linked to the phenomenon known as depletion, which is found in post-AGB binaries. Depletion is a systematic under-abundance of refractory elements, in the photospheres of post-AGB stars, that correlates with the condensation temperature of an element \citep{Giridhar2000ApJ}. In order to get a depleted photosphere, the stellar radiation pressure on circumstellar material separates dust grains (containing refractory elements) from the volatile-rich gas that gets reaccreted onto the stellar surface. Since dust grains experience a much larger radiation pressure, they do not get accreted. Depletion is thought to be caused by accretion of metal-poor gas from the CBD \citep{Waters1992}. Even though the presence of a disk seems to be needed, it is not a sufficient condition for depletion to occur since not all post-AGB binaries with disks are depleted \citep{Gezer2015}.
Though it is clear that binarity plays a very important role in the dynamics and evolution of post-AGB systems, the details of binary interaction processes are still not well understood \citep{Vanwinckel2018} and have been an important topic of investigation. New advances on the discovery of high-velocity outflows in post-AGB binaries have brought to light unique orbital phase-dependent variations in the H$\alpha$ profiles and have proven to be rather common in post-AGB binaries \citep{Gorlova2012_AAP_timespec_jet_bd46442,Gorlova2014,Gorlova2015,Bollen2017,Bollen2019,Bollen2020}. Studies associate this phenomenon with binary interaction and show that these high-velocity outflows (i.e. a bipolar jet) are launched by an accretion disk around the companion (i.e. circumcompanion accretion disk) that produces a P Cygni profile from the H$\alpha$ emission line, as the jet from the companion transits the bright post-AGB primary. Based on accretion models for two well-sampled post-AGB binaries, \citet{Bollen2020} concluded that the CBD is likely feeding a circumcompanion accretion disk. This agrees with the observed depletion patterns of refractory elements in post-AGB binaries \citep{Oomen2020}. \citet{Manick2019} have already shown similar H$\alpha$ variations that point to the presence of jets for two RV~Tauri systems: RV~Tau and DF~Cyg.
In the context of the recent developments in the field of post-AGB binaries, we report new observations that may provide insight into how binarity plays a role in these systems. In this paper we present the most comprehensive, multiwavelength analysis to date of a classical RVb variable: U~Monocerotis (U~Mon). The analysis includes an X-ray detection---never before reported for any RV~Tauri system---and an unprecedented photometric time series spanning 130 yr. Our results focus on consolidating several characteristics displayed in previous measurements, such as the very long time baseline photometry. We introduce recent space- and ground-based data collected for U~Mon from the XMM-Newton X-ray satellite and the Submillimeter Array (SMA), respectively.
We give an introduction to the U~Mon system in Section~\ref{sec:umon} and describe each data set in Section~\ref{sec:data}. We present our results in Section~\ref{sec:results}, including evidence for an even longer-term secular variation in the U~Mon light curve that may recur every $\sim$60 yr and that may represent a persistent feature in the U~Mon CBD; a refined determination of the disk's inner hole, which appears to coincide with the size of the binary orbit; a new characterization of the binary companion star; and evidence for its possible interaction with the CBD's inner edge. In Section~\ref{sec:discussion} we discuss the implications of our findings and present U~Mon as a template for the discovery of X-rays in the RV~Tauri subclass of post-AGB binaries. Finally, we summarize our conclusions in Section~\ref{sec:conclusions}.
\section{The U~Monocerotis System}\label{sec:umon}
In this section we provide an observational overview of the U~Mon system and review the physical properties that have been determined from previous studies. Table~\ref{Tab:Properties} summarizes the observed and derived physical properties that we use in our analysis.
\begin{table}
\caption{Observed and Derived Physical Properties for the U~Mon System Used in Our Analysis}
\begin{center}
\begin{tabular}{rrcr}
\hline
\hline
Properties & Value && Reference \\
\hline
Spectral type & K0Ibpv && 1 \\
${\rm [Fe/H]}$ & $-0.8$ && 2 \\
${\rm [C/O]}$ & $0.8$ && 2 \\
T\textsubscript{eff} & 5000~K && 2 \\
Distance & $1111_{-102}^{+137}$~pc && 3$^a$ \\
Radius & $100_{-13.2}^{+18.9}$ R$_\odot$ && 3 \\
Luminosity & $5480_{-882}^{+1764}$ L$_\odot$ && 3 \\
Mass\textsubscript{Post-AGB}&$2.07_{-0.9}^{+1.4}$~M$_\odot$ && 3$^b$ \\
Pulsation formal period & $91.48$~days && 4 \\
RVb long-term period & $2451$~days && 4 \\
\hline
\bf{Binary} & && \\
Inclination & {$75^\circ$} && 5 \\
Orbital period & {$2451$~days} && (fixed) 6 \\
$e$ & {$0.31 \pm 0.04$} && 6 \\
T\textsubscript{0} & {$2452203\pm 17$~days} && 6 \\
$\omega$ & {$95^\circ \pm 7^\circ$} && 6 \\
K\textsubscript{1} & {$13.5 \pm 0.7$~km\,s$^{-1}$} && 6 \\
Mass function & {$0.54 \pm 0.12$~M$_{\odot}$} && 6 \\
Mass\textsubscript{Companion}& $2.22^{+1.0}_{-0.75}$~M$_\odot$&& 6 \\
Semimajor axis & {$5.78_{-1.4}^{+2.7}$~au} && 6 \\
\hline
\bf{Disk} & && \\
Diameter & $\lesssim 550$~au && 6 \\
Dust mass & $\sim$~$4\times10^{-4}$~M$_{\odot}$ && 6 \\
Inner-disk edge radius & $\sim$~4.5--9.0~au && 6 \\
\hline
\end{tabular}
\begin{tablenotes}
\small
\item References. (1) \citet{He2014}; (2) \citet{Giridhar2000ApJ}; (3) \citet{Bodi2019}; (4) \citet{Kiss2017}; (5) \citet{Oomen2018}; (6) this study. \\$^a$ Based on \textit{Gaia} measurements. However, see note in text (Section~\ref{sec:binary}) regarding the \textit{Gaia} parallax accuracy and uncertainty. \\ $^b$Average of the values from the \citet{Groenewegen_Jurkovic2017} period--luminosity--mass relationships adopted in \citet{Bodi2019}.
\end{tablenotes}
\label{Tab:Properties}
\end{center}
\end{table}
\subsection{U~Mon as an RV~Tauri Variable of RVb Type}
U~Mon, a yellow supergiant variable \citep[K0Ibpv;][]{He2014}, is located at a distance of $1111_{-102}^{+137}$~pc \citep{Bodi2019}. It is the second-brightest RV~Tauri variable (after R~Sct) with a magnitude range of 5.45--7.67 in V \citep{Watson2006}, mean color index $B - V = 1.05$~mag at RVb brightness maximum, $1.11$~mag at RVb minimum \citep{Pollard1996}, and a metallicity of [Fe/H] $= -0.8$ \citep{Giridhar2000ApJ}. Its formal pulsation period (deep + shallow minima cycle) is 91.48~days and its fundamental period is 45.74~days. U~Mon also exhibits a long-term periodic modulation in mean brightness (RVb phenomenon) with a period of $\sim$2451~days \citep{Kiss2017}, where the large-amplitude difference in mean brightness is $\sim$3 mag.
\subsection{U~Mon as a Binary Star System}\label{sec:binary}
With BVRI photometry and high-resolution spectra, \citet{Pollard1995} concluded that U~Mon is an eccentric binary (e = 0.43) exhibiting a radial velocity amplitude of 30~km~s$^{-1}$ (full amplitude) and an orbital period of $\sim$2597~days, which is similar to its photometric RVb period, as found in other RVb systems \citep{Manick2017}.
\citet{Oomen2018} recently updated the orbital properties of the U~Mon binary system using data from the HERMES spectrograph on the 1.2 m Mercator telescope. They found an orbital period of $2549 \pm 143$~days. By assuming a typical post-AGB mass for the primary of 0.6~M$_\odot$ and an inclination of $75^{\circ}$, they estimated a projected semi-major axis of $3.38 \pm 0.31$~au, a mass function of $0.79 \pm 0.18$~M$_\odot$, and hence a minimum mass of 1.64~M$_\odot$ for the companion. However, most recently \citet{Bodi2019} used \textit{Gaia} \citep[DR2;][]{Gaia2018} data and two different period--luminosity--mass--temperature--metallicity relations derived by \citet{Groenewegen_Jurkovic2017} (based on hydrodynamical atmosphere modeling of fundamental-mode pulsators from \citet{Bono2000,Marconi2015}), to empirically infer the mass of the post-AGB star in U~Mon to be between 2.00 and 2.13~M$_\odot$. For our analysis we adopt the mean of these last two values (Section~\ref{ssec:orbital_properties}).
We note that the \textit{Gaia} distance is likely biased by the orbital movement of the binary. The renormalized unit weight error (RUWE) for U~Mon (RUWE = 2.4) indicates a poor single-star solution. High RUWE can be caused by a variety of factors, including resolved components in U~Mon \citep[see][]{2020MNRAS.496.1922B}. The astrometric excess noise ($\epsilon_{Gaia}$) for U~Mon in the DR2 catalog is 0.40~mas. This parameter is the excess uncertainty that must be added in quadrature to the formal uncertainties to obtain a statistically acceptable astrometric solution in the DR2 pipeline \citep{Lindegren2012}. \citet{Gandhi2020} considered using $\epsilon_{Gaia}$ as a proxy for the expected astrometric orbital wobble ($\omega$) to help identify potential X-ray binaries. Given that the \textit{Gaia} DR2 observations\footnote{Section~1.3.1 Time coverage: \url{https://gea.esac.esa.int/archive/documentation/GDR2/}} occurred over only $\sim$25$\%$ of the orbit of the U~Mon system, the maximum expected orbital wobble for our orbit solution for U~Mon at a distance of 1.1~kpc is $\sim0.3$~mas, which is comparable to the reported $\epsilon_{Gaia}$ value, suggesting that the excess uncertainty could be mostly due to orbital motion. Indeed, the newly updated \textit{Gaia} Early Data Release 3 parallax\footnote{\textit{Gaia} EDR3 was released on 2020 December 01.} of 1.28$\pm$0.12~mas differs from the \textit{Gaia} DR2 parallax by 0.36~mas, again very similar to $\epsilon_{Gaia}$. Therefore, if we adopt $\epsilon_{Gaia}$ as the uncertainty in the parallax measurement, then the distance to U~Mon ranges between $\sim$770~pc and $\sim$2~kpc. However, using the same period-luminosity-mass relations by \citet{Groenewegen_Jurkovic2017} as were used originally by \citet{Bodi2019}, we find that the stellar and orbital quantities in Table~\ref{Tab:Properties} do not change significantly relative to the already-quoted uncertainties. Therefore, we adopt the nominal distance of 1.1~kpc as in \citet{Bodi2019}.
\subsection{The U~Mon Circumbinary Disk}
As with many other RV~Tauri variables in the literature, the U~Mon binary star system is surrounded by a CBD. \citet{Bujarrabal2013} estimated the size of the U~Mon CBD to be $\sim$300~au using observations at the IRAM 30 m telescope in 2012--2013. Though they did not detect the Keplerian rotation of the CBD, they derived an upper limit for the molecular gas mass ($<$9~$\times$~$10^{-4}$ M$_\odot$) from their \mbox{$^{13}$CO} \mbox{$J$=1$-$0} data.
In addition, \citet{Kluska2019} were able to reproduce their recent Very Long Baseline Interferometer (VLTI)/PIONIER near-IR (H-band) observations for U~Mon by using their most complex model that includes a binary and inner ring. They suggested an an inner-disk diameter of $\sim$5.9~AU.
\subsection{Magnetic Activity in U~Mon}
\citet{sabin2015} were the first to find magnetic fields at the surface of U~Mon. The Stokes~Q and U profiles were observed at RVb phase~$\sim$0.63 and the Stokes~V profile at phase~$\sim$0.84 in Figure \ref{fig:lc_min_var}, respectively. They analyzed high-resolution spectropolarimetric ESPaDOnS (CFHT) data and found a clear Zeeman signature in the Stokes~V profile. They measured a longitudinal magnetic field (i.e., in the line of sight) of $10.2\pm1.7$~G in the photosphere of U~Mon. Although the sample is poor, this is to date the strongest surface field directly detected for a post-AGB star. The Stokes~Q and U profiles indicated the presence of shocks, and the authors suggested the possible amplification of the magnetic field due to the atmosphere dynamics.
\section{Data}\label{sec:data}
\subsection{Radial Velocity Observations}\label{ssec:rv}
We adopted radial velocity observations of U~Mon, corrected for the effects of surface pulsations, reported by \citet[][available on VizieR]{Oomen2018}. Most photometric RVb periods of RV~Tauri variables are similar to the orbital periods of their binaries \citep{Manick2017}. To calculate a new orbital solution for the binary in U~Mon, we fixed the orbital period at 2451~days \cite[U~Mon's photometric RVb period;][] {Bodi2019}, and we refit the data with a single-lined spectroscopic, Keplerian binary orbit model using {\tt PHOEBE} \citep{Prsa2016}. The resulting fit is shown in Figure~\ref{fig:orbit}, and the fit parameters are summarized in Table~\ref{Tab:Properties}.
We discuss our findings in Section~\ref{ssec:orbital_properties}.
\begin{figure}[!ht]
\centering
\includegraphics[width=\linewidth,trim=100 70 70 80,clip]{fig1_umon_rv_fit.pdf}
\caption{Single-lined orbit solution for U~Mon as fit to the radial velocity observations of \citet{Oomen2018}.}
\label{fig:orbit}
\end{figure}
\subsection{Light-curve Observations}
We use two sources of long-term photometric monitoring of U~Mon in order to explore secular changes in the U~Mon light curve on timescales of decades or longer. The full light-curve data set is represented in Figure~\ref{fig:umon_lc}.
\begin{figure*}[!ht]
\centering
\includegraphics[width=\linewidth]{fig2_umon_aavso-dasch_full_apastron_Halpha_Xray_Bfield.pdf}
\caption{Final combined AAVSO (gray points) and DASCH (black points) light curve of U~Mon, binned by 5 days. Both data sets overlap in time between 1945 and 1954, and during two long-term RVb cycles between 1974 and 1988. The scatter is due to the short-term pulsation variability. The times of apastron passage are represented by golden vertical dotted lines; the apastron times match with the RVb minima throughout the entire light curve (see Section~\ref{ssec:orbital_properties}). The arrows represent the observation times of the enhanced H$\alpha$ (purple); the Stokes Q, U, and V profiles (gold); and X-rays (teal). The light curve in this plot is available as Data behind the Figure (DbF).}
\label{fig:umon_lc}
\end{figure*}
\begin{figure}[!ht]
\centering
\includegraphics[width=0.5\textwidth]{fig3_umon_aavso-dasch_minima2_zoom_lc.pdf}
\caption{Two long-term RVb minima where the DASCH (teal pentagons) and AAVSO data (gray points) overlap the most in time. The grey dashed line on the AAVSO points (yellow dotted line for DASCH), though affected by the 5-day binning of the data, mainly highlights the pulsation variation of U~Mon. Pulsation amplitudes during the RVb maxima may extend to as low as the mean magnitude at RVb minima, while the pulsations during RVb minima are always smaller in amplitude, showing the effect of disk obscuration of the pulsating post-AGB.}
\label{fig:umon_zoom_lc}
\end{figure}
\subsubsection{{\normalfont AAVSO}}
The American Association of Variable Star Observers (AAVSO) is a global network of amateur and professional astronomers dedicated to monitoring variable stars. The earliest observation of U~Mon in the AAVSO archive was made by Ernest E. Markwick on 1888 December 25 (JD 2,410,997.0), and after a few observations there followed a 49 yr gap between 1896 and 1945. More regular monitoring began in the mid-1940s and continues to the present day. We downloaded all (including ``discrepant"-flagged) Mag$_{V}$ data from the AAVSO database up to 2020 May 26. We chose to keep discrepant data since we are not focused here on the pulsation variability. We also wanted to make sure we had the data between 1888 and 1896 (which are all marked discrepant) because they capture a long-term RVb cycle; these data are nonetheless useful for our purposes and fill a gap in time that is otherwise unavailable. We only excluded 36 Mag$_{V}$ data points tagged as upper limits.
\subsubsection{{\normalfont DASCH}}
The Digital Access to a Sky Century at Harvard (DASCH) survey is an ongoing effort to digitize about 0.5 million photographic plates covering the northern and southern sky from 1880 to 1985 \citep{Grindlay2009}.
Data for U~Mon were released in the DASCH Data Release 6 (DR6).
We downloaded the light curve of U~Mon from the DASCH Light Curve Access pipeline website\footnote{\url{http://dasch.rc.fas.harvard.edu/lightcurve.php}} using the default search radius of $5\arcsec$. Since the majority of the Harvard plates are close to Johnson B, we chose the data from the APASS B photometric calibration catalog, which yields the most accurate photometry \citep{Tang2013}.
The DASCH (APASS B-band catalog) light curve of U~Mon (ID: T540046991) contained 3824 magnitude data points from approximately 3436 plates at the time of download. The light curve has a mean magnitude of 8.77 in B and a baseline going back to 1888 January 25 (JD 2410661.7) with occasional gaps, the largest one being the ``Menzel'' gap in the 1950s and 1960s, when the plate-making operation was halted by the Harvard Observatory director at the time owing to financial concerns. We excluded DASCH data points that had estimated errors of the locally corrected magnitude measurement ({\verb magcal_local_rms }) values greater than 0.6 mag (which included {\verb magcal_local_rms } values set to 99.0; these are tagged magnitudes from images dimmer than the limiting value of the image). We also excluded magnitude-dependent corrected magnitude ({\verb magcal_magdep }) values brighter than 2.0 and dimmer than 11.0 (which were well away from the mean magnitude of the overall light curve).
\subsubsection{Combining {\normalfont AAVSO} and {\normalfont DASCH} Light-curve Data}
To analyze the AAVSO and DASCH data sets together, we converted the light curves to flux units using arbitrary zero-points (ZPs). For AAVSO we adopted ZP~=~25 \citep{Kiss2017}. Because the DASCH effective bandpass is different from the AAVSO Mag$_V$ bandpass, we experimented with ZP values for DASCH to empirically determine a scaling factor and ensure a proper match of the AAVSO and DASCH data during two RVb cycle amplitudes (between 1974 October and 1988 August), where the AAVSO and DASCH data sets overlap to the greatest degree in time. We adopted ZP~=~27.5 for DASCH as the best fit but found that we also had to scale the DASCH fluxes by a factor of 1.21 and then subtract $1.1\times 10^{7}$ flux units to match the two overlapping RVb cycles (see Figure~\ref{fig:umon_zoom_lc}). After we converted to arbitrary flux units, we made another cut to exclude data above $\sim$ $6 \times 10^{7}$ flux units and $\leq$0.
The final combined AAVSO+DASCH light curve is shown in Figure~\ref{fig:umon_lc}. It is an impressive data set, spanning the period from $\sim$1890 to $\sim$2020. The light curve is binned by 5~days to average out errors of individual observations and to reduce phase smearing due to the binning \citep{Kiss2017}. What seems to be leftover scatter is actually due to pulsations of different amplitudes, as well as large gaps in the data (see, e.g., Figure~\ref{fig:umon_zoom_lc}). Several features are present in the secular changes of the light curve on timescales of decades, and we discuss this in Section~\ref{ssec:secular_variations}.
\subsection{{\normalfont SMA} Observations}
The SMA is an interferometer, composed of eight 6 m dishes, that observed U~Mon on three occasions. On 2018 February 1, we observed the source in the subcompact array configuration with projected baselines of 7--49~m and covered two frequency ranges at 223.6--231.6~GHz and 239.6--247.5~GHz. On 2018 October 2, observations were made at higher frequencies, from 328.8 to 360.8~GHz, and at longer baselines of 6--70~m.
On 2019 March 15, we finally observed U~Mon in the very extended SMA configuration (VEX) with baselines from 32 to 514~m. The covered frequency ranges were 209.1--212.8~GHz, 214.9--216.7~GHz, 225.1--226.9~GHz, 229.2--238.8~GHz, 332.8--340.8~GHz, and 348.8--356.8~GHz. The spectral coverage in the VEX observations is smaller than on the earlier dates owing to a nonfunctional quadrant of the SMA correlator SWARM \citep{SWARM} at that time. All eight SMA antennas were used for our first two observing runs except for observations in VEX at low frequencies for which only six antennas had properly functioning receivers.
Because the target was observed mainly as a filler project,\footnote{\url{http://sma1.sma.hawaii.edu/call_filler.html}} the final {\it uv} coverage is suboptimal, in particular, very inhomogeneous at higher frequencies. The complex antenna gains were calibrated in all observing runs by observations of quasars J0730--116 and J0725--009. The bandpass was calibrated using long integrations of quasars 3C 279 and 3C 84. An absolute flux scale was established by observations of Uranus (on February~1), Neptune (on October~2), and Callisto (in 2019). The calibration was performed in {\tt MIR} software using standard procedures\footnote{\url{http://www.cfa.harvard.edu/~cqi/mircook.html}}. Further data processing, including imaging, was performed in {\tt CASA} \citep{CASA}.
Calibrated visibilities measured in the different array configurations were combined within the two atmospheric windows near 345~GHz ($\sim$0.8 mm) and 230~GHz (1.3~mm) and their weights rescaled to the actual noise levels. This resulted in continuum sensitivities of $\sigma_{345}=4.0$~mJy~beam$^{-1}$ and $\sigma_{230}=0.89$~mJy~beam$^{-1}$ at beam sizes of 4\farcs0~$\times$~2\farcs4 and 0\farcs9~$\times$~0\farcs6, respectively. These beam sizes correspond to natural weighting of visibilities. In both bands the continuum source was readily detected at flux levels listed in Table~\ref{Tab:SED}; see also Figure~\ref{fig:umon_images}. No spectral lines were detected with an rms noise level of 39.3~mJy~beam$^{-1}$ near the frequency of the CO $J$=2--1 line and at a 3.9~km~s$^{-1}$ spectral binning, which is consistent with the nondetection reported by \citet{Bujarrabal2013}. We discuss U~Mon's submillimeter emission in Section~\ref{ssec:sma_umon}.
\begin{figure*}[!ht]
\centering
\includegraphics[width=\linewidth]{fig4_umon_images_v2.png}
\caption{Top panels: XMM EPIC-pn, EPIC-MOS1, and EPIC-MOS2 images with a clear detection of X-rays for U~Mon. Bottom left panel: The EPIC-pn image annotated with the pipeline extraction vs. our compact extraction region excluding the nearby source near U~Mon. Bottom middle and right panels: U~Mon's SMA continuum images at 230 GHz (1.3~mm) and 345 GHz (870 $\mu$m), respectively. The respective beams are included in the lower left corners in dark gray.}
\label{fig:umon_images}
\end{figure*}
\subsection{{\normalfont XMM}-Newton Observations}\label{ssec:data_xray}
U Mon was observed by the X-ray Multiple Mirror (XMM-Newton) observatory on 2016 October 23 for 58.3 ks (RVb phase $\sim$0.23 in Figure \ref{fig:lc_min_var}). The XMM-Newton observations include 15 imaging exposures: three X-ray images on the EPIC-pn (55.3 ks), EPIC-MOS1 (57.0 ks), and EPIC-MOS2 (56.9 ks) detectors, and 12 exposures (each $\sim$2.2--4.4 ks) with the Optical Monitor (OM) using the UVW1, UVM2, and UVW2 filters. Representative images are shown in Figure~\ref{fig:umon_images}. U Mon is detected in all imaging exposures. Additionally, there are two grating dispersed exposures in X-ray emission on the RGS1/2 (57.2 ks), but the dispersed spectrum is not detected. The large source region used to generate the pipeline products leads to confusion and blending with a nearby source (see bottom left panel in Figure~\ref{fig:umon_images}). We re-extracted the X-ray spectra and light curves using a smaller source region with a radius of $15^{\prime\prime}$ and a source-free background region near the source. The extracted spectral products were corrected for the reduced extraction region by accounting for the encircled energy fraction in the response files.
\subsection{Spectral Energy Distribution Data}\label{sec:SED}
To construct U~Mon's SED, we downloaded available archival photometric data from VizieR. We did not take into account the phase of observations (i.e. either at maximum/minimum pulsation or long-term RVb brightness).
We included Herschel-PACS/SPIRE measurements (downloaded from the ESA Herschel Science Archive\footnote{\url{http://archives.esac.esa.int/hsa/whsa/}}; PI: C. Gielen), the $850~\micron$ flux value reported by \citet{Deruyter2005RVTau} taken with the Submillimetre Common-User Bolometer Array (SCUBA) at the James Clerk Maxwell Telescope (JCMT), and finally millimeter fluxes from \cite{Sahai2011} taken at the Combined Array for Research in Millimeter-wave Astronomy (CARMA).
\begin{figure*}[!ht]
\centering
\includegraphics[width=\linewidth,trim=10 15 10 10,clip]{fig5_sedfig_Jy.pdf}
\caption{\small SED for U Mon. Representative archival data (black points) were collected from VizieR and the Herschel archives, as well as the 850~$\mu$m measurement from \citet{Deruyter2005RVTau} and the 1.3 and 3~mm measurements from \citet{Sahai2011}. We present the new XMM-pn spectrum in teal points, the XMM-OM data in pentagons, and the SMA data in diamonds.
The scatter in the archival data is probably due to intrinsic (pulsation or RVb) variability of U~Mon at the time of observation.
Note: error bars are smaller than the symbols.}
\label{fig:UMon_SED}
\end{figure*}
\begin{table*}[!ht]
\caption{New Flux Measurements for U~Mon}
\begin{center}
\begin{tabular}{cccc}
\hline \noalign {\smallskip}
Wavelength ($\micron$) & Flux (mJy) & Date & System \\
\hline \noalign {\smallskip}
0.212 & 23.27 $\pm$ 0.038 & 2016 Oct & XMM-OM:UVM2\\
0.231 & 16.30 $\pm$ 0.030 & 2016 Oct & XMM-OM:UVW2\\
869.4 &173.3 $\pm$ 6.5 & 2018 Feb, Oct; 2019 Mar\footnote{\label{tablenote:a} Combined observations.} & SMA\\
1313 & 70.2 $\pm$ 8.5 & 2018 Feb, Oct; 2019 Mar$^{\ref{tablenote:a}}$ & SMA\\
\hline
\end{tabular}
\label{Tab:SED}
\end{center}
\end{table*}
The new observations we contribute in this paper are submillimeter and UV/X-ray from the SMA and XMM-Newton, respectively. Our SMA $870~\micron$ and 1.3~mm values are similar to those reported by \citet{Deruyter2005RVTau} and \cite{Sahai2011}, respectively. The XMM-OM magnitudes for U~Mon were converted to flux density (Jy) using the ZPs from \citet{Mason2001}. To represent the XMM EPIC-pn spectrum, we unfolded the instrumental response from the spectral data assuming the best-fit model determined using the X-ray Spectral Fitting Package \citep[{\tt XSPEC}\footnote{\url{http://heasarc.gsfc.nasa.gov/xanadu/xspec/}};][]{Arnaud1996}, dereddened the spectral data (see Section~\ref{ssec:results_SED}), and then converted from photon flux densities to Jy.
Figure~\ref{fig:UMon_SED} shows the SED for U~Mon, and Table~\ref{Tab:SED} lists our new flux measurements. In Section~\ref{ssec:results_SED} we discuss the blackbody models we used on the SED.
\section{Results}\label{sec:results}
\subsection{Secular Variations of the U~Mon Light Curve over the Past Century}\label{ssec:secular_variations}
The long-term RVb cycles of the U~Mon light curve are more salient in the AAVSO data, and we can estimate the duration of the maxima more easily than the minima. Overall, the RVb phenomenon in U~Mon changes significantly from cycle to cycle. The RVb maxima range between $\sim$3.3 and 4.8~yr, whereas the RVb minima differ even more, from several months to the longest minimum lasting $\sim$~2.5~yr (e.g., between $\sim$1977.8 and 1980.5; see Figure \ref{fig:umon_zoom_lc}).
\begin{figure*}[!ht]
\centering
\includegraphics[width=\linewidth]{fig6_umon_lc_min_var_v2.pdf}
\caption{\small Analysis of secular variations in the U~Mon light-curve data. Top: light curve binned by the fundamental period 45.74~days. The black points are DASCH data, the gray points are AAVSO data, the DASCH data overlapping two distinct RVb cycles are shown as teal squares, and the golden squares define AAVSO data sporadically overlapping DASCH data before 1951 and data corresponding to what might be times of partially obscured minima ($\sim$1997--2008). The dashed line is the mean flux at RVb maximum ($3.5\times 10^{7}$ flux units). Bottom: light curve phase-folded on the 2451-day RVb period of the system with flux normalized at unity at the base level (dashed line is the same as the top panel). See Section~\ref{ssec:secular_variations} for a discussion of the curves.}
\label{fig:lc_min_var}
\end{figure*}
In Figure~\ref{fig:lc_min_var} we explore longer-term secular changes in the light curve, enabled by unprecedented coverage spanning $\sim$130~yr. We binned both data sets by U~Mon's fundamental pulsation period (45.74~days) to focus on the RVb behavior. We plot the DASCH data in black, the AAVSO data in gray, and the DASCH data that distinctly overlap the two RVb cycles during 1975--1988 as teal squares. We also define golden squares that are AAVSO data sporadically overlapping with DASCH data before 1951, and AAVSO data corresponding to what might be times of partially obscured RVb minima ($\sim$1997--2008). The top panel is the binned light curve in time, the bottom panel is the binned light curve phased by the RVb period of 2451~days, and the flux is normalized to be unity at the base level. We fit parabolas to the data at phases of $0.5\pm 0.3$, to get a cleaner visual sense of how the RVb minima compare for the various subsets.
We find that the AAVSO data (gray points) and the DASCH overlap data (teal squares) have nearly identical RVb minima. Similarly, we see that the older DASCH data (black points) and the golden-square AAVSO data have nearly identical RVb minima, and the latter minima are about a factor of 2 less deep. In other words, it appears from the historical record that the RVb variations---which have come to be associated with a highly inclined disk shadowing the primary at certain orbital phases---have largely disappeared at least twice in the past, with a timescale of $\sim$75~yr, from the middle of the black points to the latest golden squares in Figure~\ref{fig:lc_min_var}.
Another way to examine this is shown in Figure~\ref{fig:lc_phased_60yr}, in which the disappearance of the RVb variation, due to the large scatter in the data points, is more obvious for two RVb minima in 1938 and 1944 and then again in 1998 and 2005, with a timescale between them of $\sim$60.4~yr, which we note is nine times the RVb long period. To be clear, the RVb minima are short during these cycles ($\sim$several months, as opposed to years), and they ``disappear'' relative to the large scatter/pulsation variations from the (particularly longer) maximum state at either side of RVb minima. We further discuss the interpretation of this phenomenon in Section~\ref{ssec:60yr_trend}.
\begin{figure*}[!ht]
\centering
\includegraphics[width=\linewidth]{fig7_umon_aavso-dasch_phased_60yr_lc.pdf}
\caption{\small Light curve of U~Mon phased at 60.4 yr (nine RVb cycles). The inset shows the full light curve divided in color distinguishing the two 60.4 yr cycles in the data (black points represent the first 60.4 yr cycle; gray points, second cycle). The start of a third cycle is shown as teal points. The RVb cycles highlighted in light grey include long ($\lesssim$4.5~yr) RVb maxima and short RVb minima (phase~$\sim$47~yr and $\sim$53~yr) that have durations of only several months based on the AAVSO data.}
\label{fig:lc_phased_60yr}
\end{figure*}
\subsection{Orbital Properties of the U~Mon Binary Star System} \label{ssec:orbital_properties}
From the orbit fit in Section~\ref{ssec:rv}, we obtained orbital parameters (listed in Table \ref{Tab:Properties}) that are within the errors of the values consistent with \citet{Oomen2018}, including a new periastron time of T$_{0} = 2452203 \pm 17$~days. Adding half (1225~days) of the fixed orbital period to the periastron time gives an apastron at JD $= 2453428$~days. Multiples of the orbital period with the apastron time give all apastron events (denoted as golden vertical dashed lines in Figure~\ref{fig:umon_lc}) that roughly align with the RVb minima throughout U~Mon's entire DASCH+AAVSO light curve. Mass estimates for U~Mon range from a fiducial post-AGB mass average of 0.6~M$_{\odot}$ \citep{Gezer2015,Manick2017,Weidemann1990} to most recent values of $2.00$ and 2.13~M$_{\odot}$ reported by \citet{Bodi2019}. To estimate a value for the companion's mass (M$_{2}$), we used the average value from \citet{Bodi2019} for the mass of the post-AGB (M$_{1}$ = 2.07~M$_{\odot}$) in Equation \ref{Eq:massfct} (the binary mass function for an eccentric orbit) and solved for M$_{2}$:
\begin{equation}
f(m) = \frac{P K_1^3}{2\pi G}(1-e^2)^{\frac{3}{2}} = \frac{M_2^3}{(M_1+M_2)^2} \sin^3i.
\label{Eq:massfct}
\end{equation}
Using the recalculated radial velocity curve, and keeping the orbital inclination maximum limit of $75^\circ$ \citep{Oomen2018}, we find the mass for the companion to be $2.2_{-0.75}^{+1.0}$~M$_\odot$; this mass range corresponds to an F-type or A-type star on the main sequence. Then, by using Kepler's third law, we derived a value of $5.78_{-1.4}^{+2.7}$~au for the semi-major axis. These parameters are reported in Table~\ref{Tab:Properties}.
\subsection{Spectral Energy Distribution}\label{ssec:results_SED}
We dereddened the flux data between 0.2 and 3.3~$\mu$m using functions contained in the {\tt Astropy}-affiliated package {\tt dust\_extinction}, with the F19 Milky Way R(V) dependent extinction \citep{Fitzpatrick2019} using the standard value Rv = 3.1 and total extinction: E\textsubscript{B-V}~=~0.3 \citep{Deruyter2005RVTau}.
We made use of the {\tt astropy.modeling.blackbody} routine \citep{Astropy2013} to depict scaled blackbody components on the SED (these are not fitted). We include two blackbody components with temperatures of 5000 and 1100~K to represent the stellar atmosphere \citep{Bodi2019} and the inner-disk edge (see Section~\ref{ssec:indisk_properties}), respectively. We also include a modified blackbody \citep[i.e., ``graybody'';][]{Casey2012} component with a mean temperature of 350~K to represent the dust in the extended disk. Finally, we added a 3.5~MK thermal blackbody for the X-ray spectrum.\footnote{The blackbody model is only shown for display purposes, the unfolded spectrum has a strong dependence on the best-fit 10~MK plasma model, and a blackbody model does not provide a suitable fit to the observed X-ray spectrum.} Figure \ref{fig:UMon_SED} shows the entire SED model as a black solid line, with the individual components depicted as light-gray dashed lines.
We integrated the 5000~K blackbody model (between~$\sim$0.2 and 2~$\mu$m), which gave a (dereddened) bolometric flux of F\textsubscript{bol} $\approx 1.44 \times 10^{-7}$ erg~s$^{-1}$~cm$^{-2}$, corresponding to L$_{\rm \star}$ $\sim$5556~L$_{\odot}$ and yielding a physical radius of~R$_{\rm \star}$ $\sim$100~R$_{\odot}$.
We tested how an additional blackbody component with the properties of the putative A-type companion star (9800 K and 55 L$_{\odot}$) would influence our SED model and found that such a companion would remain undetectable in the SED (even at RVb minimum) owing to the brightness of the post-AGB component (see Section~\ref{ssec:discussion_xrays}).
Assuming that the dust is optically thin at $\lambda =$ 870 $\mu$m, we used the following expression \citep[e.g.][]{Hildebrand1983} to estimate the mass of the dust in the disk:
\begin{equation}
M_{d} = \frac{F\textsubscript{$\lambda$} D^{2}}{\kappa_{\lambda}B\textsubscript{$\lambda$}(T_{d})},
\end{equation}
where B\textsubscript{$\lambda$}(T\textsubscript{d}) is the blackbody intensity at the Rayleigh-Jeans limit in the form of $\frac{2k_{SB}T}{\lambda^2}$. Using a mean dust temperature T\textsubscript{d} = 350 K, our observed SMA flux F\textsubscript{870} = 173.3 mJy, D = 1.1 kpc \citep{Bodi2019}, and a dust opacity of $\kappa_{\lambda} = 2.0$ cm$^{2}$g$^{-1}$ at 870 $\mu$m \citep{Ladjal2010}, we estimate a total dust mass of M\textsubscript{d}~$\sim 4 \times 10^{-4}$~M$_{\odot}$. For a typically assumed gas-to-dust ratio of 200 \citep[e.g.][]{Groenewegen2007,Groenewegen2009} the total dust mass suggests a gas mass of $\sim$8~$\times 10^{-2}$~M$_{\odot}$; however, this gas mass is much larger than the constraint of the molecular mass estimated by \citet{Bujarrabal2013}. The lack of molecular gas may be related to high-energy radiation that has been dissociating molecules present in the system that would support the larger distance to U~Mon. Since the primary star is not hot enough, the presence of a hotter companion or interstellar UV field could photodissociate molecules in the inner and outer disk regions \citep{Bujarrabal2013}.
\subsection{Properties of the Circumbinary Disk's Inner Edge}\label{ssec:indisk_properties}
The radius of the near-IR emission, where dust can be sublimated by stellar radiation, can also set the physical radius of the inner boundary of the disk \citep{Dullemond2001,Monnier-Millan2002,Hillen2017,Kluska2019,Lazareff2017}. Using the following luminosity--radius relation \citep[e.g.][]{Lazareff2017}, we estimated the inner-rim radius (R\textsubscript{rim}) of the CBD for U~Mon:
\begin{equation}
\label{equation:innerrim}
R_{\rm rim} = \frac{1}{2}\,\left(\frac{C_{\rm bw}}{\epsilon}\right)^{1/2}\,\left(\frac{L_{\rm \star}}{4\pi\,\sigma\,T_{\rm sub}^4}\right)^{1/2}\,
\end{equation}
where we adopted L$_{\rm \star}$ = 5480~L$_{\rm \odot}$ \citep{Bodi2019} as the stellar luminosity. C$_{\rm bw}$ is the back-warming coefficient of the inner-disk edge that ranges between $\sim$1 and 4, where C$_{\rm bw}=1$ provides a lower limit on R$_{\rm rim}$ and C$_{\rm bw}=4$ provides an upper limit on R$_{\rm rim}$ \citep{Monnier2005,Kama2009}. The cooling efficiency of the dust grains, defined by $\epsilon$ = $\kappa$(T$_{\rm dust})$/$\kappa$(T$_{\rm \star})$, is a ratio between the Planck mean opacity ($\kappa$) of the dust species at its own temperatures and that at the stellar temperature, where $\epsilon \leq 1$, and generally increases with grain size \citep{Kama2009}. For the dust in the CBD, we assume $\epsilon \sim$~1. Typically, for oxygen-rich dust species ([C/O] of 0.8 for U~Mon), the sublimation temperature (T$_{\rm sub}$) is $\sim$1100~K \citep{Bladh2013}. We note that \citet{Kluska2019} suggest a higher inner-dust temperature ($\sim 2600$~K) for U~Mon but only reach a moderate fit with their most complex model, whereas most of their sample of post-AGB binaries have near-IR circumstellar emission sublimation temperatures lower than 1200~K. Finally, the range between C$_{\rm bw} =1$ and C$_{\rm bw} = 4$ gives an R$_{\rm rim}$ between $\sim$4.5 and 9.0~au, respectively.
\subsection{Submillimeter Emission from U~Mon}
\label{ssec:sma_umon}
The SMA observations in the combined array configurations and uniform weighting of visibilities resulted in continuum maps with synthesized beam sizes of 0\farcs69 $\times$ 0\farcs37 at 1.3~mm and 0\farcs39 $\times$ 0\farcs24 at 0.87~mm, respectively. Simple fits of elliptical Gaussians to the map at 0.87~mm give a source size of (0\farcs89 $\pm$ 0\farcs34) $\times$ (0\farcs20 $\pm$ 0\farcs39) and an orientation of the longer axis at a position angle (P.A.) of 55\degr $\pm$ 24\degr. The map at 1.3~mm indicates that the source size is much smaller than the beam, i.e. with FWHM~$\lesssim$~0\farcs55.
To better constrain the size and make best use of the data, we directly tried to fit a model source to the calibrated visibilities.
Best solutions for an elliptical Gaussian and an elliptical uniform disk were found in a {\tt CASA} task {\tt uvmodelfit} and are presented in Table~\ref{Tab:sizes}.
\begin{table}[!ht]
\caption{Results of model fits to visibilities of U~Mon measured with the SMA}
\begin{center}
\begin{tabular}{ccccc}
\hline \noalign {\smallskip}
Band & Model & Major & Minor/Major & PA \\
& & (mas) & & (deg) \\
\hline \noalign {\smallskip}
1.3~mm & Gaussian & 299 $\pm$ 11 & 0.73 $\pm$ 0.04 & 83 $\pm$ 5\\
1.3~mm & Unif. disk & 483 $\pm$ 24 & 0.76 $\pm$ 0.05 & 84 $\pm$ 8\\
0.8~mm & Gaussian & 244 $\pm$ 6 & 0.86 $\pm$ 0.03 & 55 $\pm$ 7\\
0.8~mm & Unif. disk & 482 $\pm$ 64 & 1.0 $\pm$ 0.2 & $\cdots$ \\
\hline
\end{tabular}
\tablecomments{\small For Gaussian models, FWHM is given as the major-axis size.}
\label{Tab:sizes}
\end{center}
\end{table}
The model fits indicate a nearly circular source with a Gaussian FWHM smaller than about 500~mas, which corresponds to a disk diameter $\lesssim$550~au at the nominal distance of 1.1~kpc. A denser {\it uv} coverage of observations at longer baselines would be necessary to better constrain the size of the submillimeter source.
The SMA fluxes are consistent with earlier observations in the nearby bands and represent the coolest component of U~Mon's SED, which is consistent with a dusty source of T\textsubscript{d} = 350~K (see Section~\ref{ssec:results_SED}).
\subsection{X-Ray and {\normalfont UV} Emission from U Mon}
This is the first X-ray detection of an RV~Tauri variable by any X-ray telescope. The X-rays are consistent with the location of U Mon within the pointing uncertainties and the fairly broad point spread function of the XMM EPIC X-ray detectors (FWHM $\sim 12^{\prime\prime}$). There is an additional but fainter source near the position of U~Mon with a separation of $\sim20^{\prime\prime}$ (see Figure~\ref{fig:umon_images}). Given the relative brightnesses of the two sources and the radius used to extract the X-ray products, contamination by the nearby source is minimal.
The spectra were modeled using an absorbed \citep[{\tt tbabs};][]{Wilms2000}, optically thin plasma model with variable abundances \citep[{\tt vapec};][]{Smith2001,Foster2012}. Model fitting was performed with {\tt XSPEC} \citep{Arnaud1996}. No acceptable fit could be found with solar abundances, but adequate fits were found when Fe was allowed to vary. Given the limited energy resolution and moderately low count rate of the CCD spectrum, we could not obtain meaningful constraints on both the plasma properties (temperature and normalization) and the important elemental abundances in the energy range of the detected emission (C, N, O, Ne, Mg, and Fe). Instead, we fixed the Fe abundance to a previously reported value of [Fe/H]~=-0.79 and C-to-O ratio to 0.8 \citep{Giridhar2000ApJ}. The O abundance was allowed to vary along with the overlying absorption, plasma temperature, and plasma model normalization.
The best-fit parameters are presented in Table~\ref{Tab:xrayfit} along with the absorbed and intrinsic X-ray fluxes and source luminosity for a distance of 1.1~kpc \citep{Bodi2019}. The best-fit spectral model is displayed with the X-ray spectra in Figure~\ref{fig:xr_spectra}, and the unfolded spectrum is presented in the multiwavelength SED in Figure~\ref{fig:UMon_SED}; confidence level contours are shown in Figure~\ref{fig:xspec_contours}.
\begin{figure}[!ht]
\centering
\includegraphics[height=2.5in]{fig8_specfig.png}
\caption{\small U~Mon's X-ray spectra and best-fit spectral model (see Table \ref{Tab:xrayfit}).}
\label{fig:xr_spectra}
\end{figure}
\begin{table}[!ht]
\caption{X-ray Spectral Fit Results for U~Mon}
\begin{center}
\begin{tabular}{lc}
\hline \noalign {\smallskip}
Quantity & Value \\
\hline \noalign {\smallskip}
$N_{\rm H}$ ($10^{21} {\rm cm}^2$) & $1.2^{+0.6}_{-0.5}$ \\
$T_{\rm X}$ ($10^{6}$ K) & $12^{+1}_{-2}$ \\
Oxygen abundance\footnote{Abundances are number relative to solar values from \citet{1989GeCoA..53..197A}. Iron abundance has been fixed to the [Fe/H] value of \citet{Giridhar2000ApJ} after conversion to the number relative value.} & $3.4^{+3.0}_{-1.6}$ \\
Iron abundance & 0.16 \\
Model normalization ($10^{-5}$) & $2.1^{+0.3}_{-0.3}$ \\
$F_{\rm X, abs}$\footnote{Model-derived fluxes have been corrected for the encircled energy fraction of the extraction region.} ($10^{-14} {\rm ~erg~cm}^{-2}{\rm ~s}^{-1}$) & $2.2^{+0.4}_{-0.4}$ \\
$F_{\rm X, unabs}$ ($10^{-14} {\rm ~erg~cm}^{-2}{\rm ~s}^{-1}$) & $3.4^{+0.6}_{-0.6}$ \\
EM ($10^{53} {\rm ~cm}^{-3}$) & $3^{+1}_{-1}$ \\
$L_{\rm X}$ ($10^{30} {\rm ~erg~s}^{-1}$) & $5^{+1}_{-1}$ \\
\hline
\end{tabular}
\label{Tab:xrayfit}
\end{center}
\end{table}
Our analysis of the X-ray observations suggests that the emission is consistent with an iron-deficient hot ($\sim10$~MK) plasma with moderate absorption. The absorption reported in Table~\ref{Tab:xrayfit} is consistent with extinction to U Mon \citep[E$_{B-V} $$\sim$~0.3~mag;][]{Deruyter2005RVTau}. The X-ray light curve reveals no evidence of flaring activity at the 3$\sigma$ level; however, the light curves required 10 ks temporal bins to net sufficient signal-to-noise ratio. Hence we cannot evaluate activity on shorter time-scales. We discuss the origin of the X-ray emission in further detail in Section~\ref{ssec:discussion_xrays}.
The XMM Optical Monitor performed numerous exposures in three UV filters: UVW1, UVM2, and UVW2 (with effective wavelength of 0.291, 0.231, and 0.212~$\micron$, respectively). U~Mon is detected in all three bands but saturated in UVW1. The UVM2 and UVW2 measurements are included in the multiwavelength SED in Figure~\ref{fig:UMon_SED} and in Table~\ref{Tab:SED}. The UV light curves show variability consistent with Poisson noise. Given the short time frame for the UV observations, we are not very sensitive to variations or trends related to the pulsation period or the long-period variation.
\begin{figure}[!ht]
\centering
\includegraphics[height=2.5in]{fig9_spec_contoursfig.png}
\caption{\small Best-fit confidence map for the plasma temperature ($kT$) and column density ($N_{\rm H}$) of the X-ray spectral fit. The best-fit solution is depicted with a 98\% confidence range, and 68\%, 90\%, and 98\% confidence levels are depicted.}
\label{fig:xspec_contours}
\end{figure}
\section{Discussion}\label{sec:discussion}
\subsection{A 60 yr Trend in the Light Curve}\label{ssec:60yr_trend}
U~Mon's RVb behavior has been observed since before the 1900s, giving a window into the behavior of the CBD. Even with the lack of uniform monitoring in combination with various physical processes present (e.g., pulsations), an even slower trend in the light curve is still detected in the AAVSO+DASCH light curve.
In particular, three RVb cycles between 1931 and 1951 have significantly more scatter, compared to the overall light curve. They resemble three cycles between 1991 and 2011 (see Figure \ref{fig:lc_phased_60yr}). The disk-shadowing interpretation can be a possible explanation for the disappearance of the long-term variation or the lack thereof. This phenomenon can represent times when the RVb minima have become shallower and shorter, perhaps indicating some kind of precession of the binary or of disk-structure inhomogeneity at the inner-disk edge caused by binary--disk interaction \citep[e.g.][]{Kluska2018,Oomen2020}.
Furthermore, the light curve hints of a different feature of two even deeper RVb minima that occur around 1925 and more clearly around 1985 (phase $\sim$34~yr in Figure \ref{fig:lc_phased_60yr}) with the same time separation of $\sim$60.4 yr. A warp or a dense feature at the inner edge of the disk orbiting around it would only take about 6~yr (2190~days), which is comparable to the orbital period. If a warp or dense feature is farther out in a Keplerian orbit, it would have to be at a distance of 25~au to block out more of the post-AGB so that the RVb minimum becomes deeper every 60.4~yr.
As more photometric monitoring and interferometric imaging become available for post-AGB binaries, the detection of CBD warps or features may give rise to future investigations on possible second-generation planet formation such as in studies around other evolved stars \citep[e.g.][]{Hardy2016,Homan2018}.
\subsection{Circumbinary Disk Interaction}\label{ssec:discussion_bdi}
\citet{Pollard1995} and \citet{Pollard2006} found enhanced H$\alpha$ emission in U~Mon during the RVb minima. This phenomenon is reminiscent of one observed in other binary star systems with eccentric orbits within CBDs, in which matter is pulled from the inner edge of the disk and onto the stars at apastron \citep[e.g.,][]{ArtymowiczLubow1996,Basri1997}. Based on the periastron time occurring at RVb maximum, we experiment with the idea that the relevant time to explain the interaction that produces H$\alpha$ enhancement is not at periastron but rather the time when the two stars are farthest apart (apastron), which would be when the stars would presumably be most likely to interact with the inner edge of the CBD.
Based on our orbital fit for the binary, the argument of periastron ($\omega$ = $95^\circ \pm 7^\circ$) means that the longest dimension of the ellipse is oriented parallel to our line of sight. In other words, when the two stars are farthest apart, they are also along our line of sight during apastron, which happens during RVb minima. This scenario requires that the post-AGB be mostly hidden by the CBD while the companion travels in the background at an elevated angle crossing our light of sight and possibly interacting with the inner edge of the CBD, causing the enhanced H$\alpha$ emission observed by \citet{Pollard2006} during RVb minima. This is also shown from the radial velocities in Figure~\ref{fig:orbit}, which measure the motion of the post-AGB star. The orbital phase of zero is at periastron (which corresponds to RVb maximum).
The nature of the CBD--binary interaction is not yet clear in post-AGB binaries; however, processes such as photospheric chemical depletion show that it is crucial. Interestingly, U~Mon's photosphere does not show signs of depletion \citep{Giridhar2005,Gezer2015}; this could be because depletion is mostly observed in stars hotter than 5000~K \citep{Venn2014}. It is also possible that the depletion process could be interrupted by wind Roche lobe overflow onto the companion \citep{Mohamed2007}.
The recent study by \citet{Oomen2020} investigated a few disk--binary interaction mechanisms for post-AGB binaries with orbital periods of 100--2500 days and eccentricities $\gtrsim0.3$. They concluded that disk--binary interactions are unlikely to pump the eccentricity to the observed values within the evolutionary timescales for the stars they sampled. U~Mon is not very eccentric (e = 0.31); this may suggest a weak disk--binary interaction.
However, the \citet{Kluska2019} geometric image reconstruction model of U~Mon's visibility profile resulted in a fit that displayed strong azimuthal modulation, showing that this system likely has a very complex inner-disk morphology. Additionally, they showed that U~Mon has a relatively large mid-IR-to-near-IR size ratio. The authors note that it is possible that the model is not able to reproduce all the complexities, which could have an effect of on their derived size of the CBD's inner rim; thus, further observations will bring stronger constraints. Such complexities have been seen in some YSO CBD systems, which can be regarded as scaled-down versions of the CBDs around post-AGB binaries \citep[e.g.][]{Hillen2017}. For example, in the CBD of the young binary GW~Ori, imaging observations by \citet{Kraus2020} find evidence for an inner ring and a large warp in the outer part of the disk. This phenomenon occurs when the outer disk is misaligned with the orbital plane, so it wraps and breaks into precessing rings, which may provide a mechanism for planet formation.
\citet{Bollen2019} provide a detailed investigation on the possibilities for jets in explaining the H$\alpha$ emission feature observed in post-AGB systems. In their model, the H$\alpha$ feature is superimposed on the photospheric absorption from the post-AGB star and is present throughout the entire orbit. To establish whether such a jet scenario may apply to U~Mon as well would require phase-resolved, high-resolution spectroscopy of the H$\alpha$ feature.
If, instead, there is a stream outflow coming off the CBD onto the companion, always traveling around the companion and opposite of the post-AGB primary, the H$\alpha$ enhancement would have a timescale near the binary orbital period. The stream of material would be at a focus point where it would be enhanced around apastron because that would be the shortest distance the CBD material would flow onto the companion. The stability for the shortest orbit of material around a binary should be at least $\sim$3--4 times the binary separation \citep{HolmanWiegert1999}. However, it is possible that the inner edge of RV~Tauri disks may not be in long-term stable orbits. \citet{ArtymowiczLubow1996} showed hydrodynamical models for unstable inner-disk rims, so it is possible that material may be stripped by the binary, in particular the companion. This scenario may also explain the X-rays observed in U~Mon.
\subsection{Nature and Origin of X-Ray Emission from U~Mon}\label{ssec:discussion_xrays}
As a binary system with a CBD, there are a number of processes that can generate the detected X-ray emission. We consider processes from U~Mon, the companion, and interactions between the components of the system including the CBD.
If the X-ray emission originates from U Mon, the X-ray luminosity corresponds to an $L_{\rm X}/L_{\rm bol} = 10^{-7}$. Generally, such a ratio is consistent with the X-ray emission that arises from shocks caused by variable winds in O and B stars; however, the X-ray emission from U Mon does not quite follow the general trend with L$_{\rm bol}$ seen for B stars in the Carina Nebula, which would predict an X-ray luminosity up to an order of magnitude higher \citep[e.g.][]{Naze2011}. The plasma temperature ($>$10~MK) is consistent with hot components found in the Carina O and B stars; however, no fast outflows have been measured from U Mon. Indeed, the large radius of U Mon suggests an escape velocity $<$100~km~s$^{-1}$, whereas the plasma temperature requires velocities $>$800~km~s$^{-1}$ for a strong shock. Such velocities have not been measured from U Mon and are unlikely to be produced by any pulsation-induced shocks in the atmosphere of U Mon \citep{Fokin1994}, unlike that suggested for X-ray emission from the Cepheid star $\delta$ Cep \citep{2017ApJ...838...67E}. \citet{Gillet1990} measured shock amplitudes from spectroscopy for the RV~Tauri variables R~Sct and AC~Her that did not exceed $\sim$40 km s$^{-1}$.
We note that \citet{Moschou2020} recently investigated a mechanism to produce phase-dependent shocked gas X-ray emission in the pulsating atmosphere of classical Cepheids. This mechanism requires the presence of solar-like coronal plasma into which the phase-dependent shocks are driven \citep{Moschou2020}. The detected X-ray emission from U~Mon is hotter than that studied by \citet{Moschou2020}, and we lack adequate phase coverage to establish the presence of such shocks from U~Mon, as well as the necessary signal-to-noise ratio in our XMM X-ray light curve to establish the presence of such shocks from U Mon. A solar-like coronal plasma is an intriguing notion given the detection of a magnetic field (very likely dynamo generated) at the surface of this cool post-AGB star, but there is, so far, no information on any related activity such as flares, stellar spots, or dynamo variability that would be linked to X-ray emission. Furthermore, the ability to sustain a hot stellar corona on such a large evolved star is unclear \citep{Sahai2015}.
Accretion, from a compact disk or infalling material from the large CBD (as discussed in Section~\ref{ssec:discussion_bdi}), might create shocks that can heat material up to X-ray-emitting temperatures. However, given the stellar properties of U~Mon in Table~\ref{Tab:Properties}, it is unlikely that infalling material will reach sufficiently high velocities ($<$100~${\rm km}\ {\rm s}^{-1}$ at the surface of U Mon). If the companion star has a radius $\leq$2\ R$_{\odot}$, which would make the companion consistent with an A-type main-sequence star, infalling material, from the CBD or donated by U~Mon via wind Roche lobe overflow \citep{Mohamed2007}, can reach speeds in excess of $800\ {\rm km}\ {\rm s}^{-1}$. Such infalling material can either slam into the surface of the star along magnetic fields or form into an accretion disk around the companion.
An accretion disk forming around the companion could then generate outflows as seen in other post-AGB systems \citep{Gorlova2012_AAP_timespec_jet_bd46442,Gorlova2015}; however, such outflows have yet to be observed from the U~Mon system. If such an outflow exists, its velocity could be sufficient to generate shocks capable of explaining the detected X-ray emission. Detailed monitoring of the H$\alpha$ emission from U~Mon, like that performed of similar systems by \citet{Gorlova2012_AAP_timespec_jet_bd46442,Gorlova2013}, could help establish the presence of accretion-disk-driven outflows.
If the binary companion is the source of X-ray emission, in addition to accretion, coronal activity and radiative stellar winds can produce X-ray emission similar to that given in Table~\ref{Tab:xrayfit}. However, at $\sim$2.2~M$_{\odot}$ (Table~\ref{Tab:Properties}), a main-sequence companion would be consistent with an early-type A star and unable to support magnetic activity to produce coronal X-ray emission \citep{Stelzer_2006}. Furthermore, the bolometric luminosity of such a main-sequence companion is 100 times fainter than U Mon, leading to an $L_{\rm X}/L_{\rm bol}$ of $10^{-5}$ for the companion, which is an unusually high fraction for an A-type star. Such a high fraction of X-ray emission and a high-temperature plasma are comparable to what is seen from Herbig Ae/Be stars, which are intermediate-mass pre-main-sequence stars embedded in dusty disks \citep{Stelzer_2006,Stelzer_2008}. In the case of Herbig stars, the high plasma temperature suggests that radiative winds are not the emission mechanism, and higher-order binary components, namely, unseen rapidly rotating late-type companions to the A star in a Herbig system, have been suggested as a potential origin for their X-ray emission \citep{Stelzer_2006}. Since we have not observed the companion and only have a modest constraint on its mass ($2.2^{+1.0}_{-0.75}$~M$_\odot$; see Table~\ref{Tab:Properties}), we cannot exclude the remote possibility that the companion is a compact binary. In such a scenario, the detected X-ray emission would be consistent with enhanced coronal activity from rapidly rotating late-type stars in a close binary system.
We have highlighted a number of potential origins for the X-ray emission from the U~Mon system, each with strong implications for the nature and evolution of the system. Additional supporting information, such as further constraints on the companion properties and monitoring of the H$\alpha$, like that reported in \citet{Bollen2019}, and additional observations of the X-ray emission through the phases of the U~Mon pulsation and orbit, is essential to better understand the origin and influence of the X-ray emission.
\section{Conclusions}\label{sec:conclusions}
We have conducted the most comprehensive characterization yet of an RVb system, U~Mon, with observations spanning the largest range of wavelengths (X-ray to millimeter) and the longest range of time ($\sim$130~yr).
U~Mon, one of the brightest RV~Tauri variables of its class, is shown here to be a binary system comprising a 2.07~M$_\odot$ post-AGB star that has lost mass relative to a more massive unevolved 2.2~M$_\odot$ A-star companion, and is surrounded by a large CBD. We obtained new orbital parameters within the errors of the values found by \citet{Oomen2018}, by fixing the orbital period as the photometric RVb period. The new periastron value, T\textsubscript{0}~=~$2452203$~days, reveals that the apastron times align with RVb minima (Figure~\ref{fig:orbit}). Additionally, based on our orbital fit, the argument of periastron ($\omega$~=~$95^{\circ}$) translates to the long axis of the orbit being roughly aligned along our line of sight. This supports the argument that apastron occurs when the post-AGB star is occulted by the near side of the disk at RVb minimum while the A-type star companion is most revealed at the far side of the inner hole in the CBD. Moreover, we found that the binary's semimajor-axis separation is $5.78$~au, which at apastron is comparable to the size range of the inner-disk hole radius ($\sim$4.5--9.0~au).
U~Mon shows several interesting features, some not yet seen in any other post-AGB binary, such as a magnetic field and most recently X-rays. We found a $\sim$10~MK plasma model that is consistent with the X-ray spectra and considered origins from U~Mon and/or its companion. We suggest that at apastron the companion may be close enough to strip material from the CBD, creating the X-ray emission observed. This interpretation may agree with the complex morphology found in the interferometric dataset for U~Mon by \citet{Kluska2019}.
The occurrence of the enhanced H$\alpha$ at RVb minima \citep[e.g.][]{Pollard2006} also seems to correspond to when the A-type star companion is most revealed at apastron, implying that since the post-AGB star is shadowed by the CBD at this phase, the enhanced H$\alpha$ must not be coming from the post-AGB star but rather from the environment around the A-type companion, likely an accretion disk \citep[e.g.][]{Bollen2019}. Recent work has shown that the H$\alpha$ variation is due to a bipolar jet from the circumcompanion accretion disks in other post-AGB binaries. More time series of optical spectra are needed to investigate the H$\alpha$ variability in U~Mon.
We tested whether additional blackbody components with properties of the A-type star companion would influence U~Mon's SED but found that such a companion seems to only be detected in H$\alpha$ (and possibly X-rays), even at RVb minimum, when it is most revealed in its orbit. The X-ray through millimeter emission, however, is consistent with components that include a hot thermal plasma, the stellar photosphere of the post-AGB, and the inner- and extended-disk emission. The scaling of the modified blackbody for the dust emission includes parameters such as opacity and dust grain sizes that we do not explore in this paper. However, U~Mon's new SMA observation fluxes are in agreement with the slope of the Rayleigh-Jeans form of the Planck function, between the other 60--3000 $\mu$m flux measurements. Such slopes are consistent with the presence of a composition of large dust grains \citep[radius $\gtrsim$ 0.1 mm;][]{Deruyter2005RVTau}. Due to the short SMA observations, we are only able to estimate upper limits on U~Mon's disk diameter size $\lesssim$550~au (see Table \ref{Tab:sizes}).
Nevertheless, with the AAVSO+DASCH light curve, which spans more than 130~yr, we find evidence for a 60.4 yr cycle that could be due to some structure in the disk at a separation of $\sim$25~au that temporarily allows the post-AGB star to avoid being occulted at apastron/RVb minimum. Accordingly, the 60.4 yr cycle also produces an especially deeper RVb minimum (phase $\sim$54~yr in Figure \ref{fig:lc_phased_60yr}--two cycles before the ``disappearance'' of the RVb minima). Presumably this can be a corresponding feature in the disk that causes an exceptionally large occultation of the post-AGB star. More radio observations at longer baselines would constrain the size of the disk, as well as resolve features we see in U~Mon's light curve.
Furthermore, the detection of X-rays from the U~Mon system has opened up new possibilities; do all RV~Tauri variables exhibit X-ray emission? In order to answer this question, X-rays will be an important factor to consider in forthcoming RV~Tauri studies to enable constraints on their evolution \citep[e.g.,][]{Graber_Montez_2021}.
Most importantly, U~Mon now becomes an archetype for the study of binary post-AGB stellar environments that represent an important evolutionary phase, which either sets the stage for sculpting the morphology and evolution of planetary nebulae or which may represent systems that never become planetary nebulae at all, as the disk and stellar companion arrest its development.
\acknowledgments{
We acknowledge the very helpful review of the anonymous referee, which substantively improved the paper.
We acknowledge the excellent work of the NASA's Goddard Space Flight Center's Astrophysics Science Division Communications Team in the production of a NASA news feature for this research.
L.D.V. and K.G.S acknowledge the support of the NASA MUREP Harriett G. Jenkins Predoctoral Fellowship, Grant Nos. NNX15AU33H and 80NSSC19K1292.
R.M.Jr. and L.D.V. acknowledge the federal support from the Latino Initiatives Pool, administered by the Smithsonian Latino Center.
This research has made use of XMM-Newton, an ESA Science Mission with instruments and contributions directly funded by ESA Member States and the USA (NASA).
This research has made use of the Submillimeter Array, a joint project between the Smithsonian Astrophysical Observatory and the Academia Sinica Institute of Astronomy and Astrophysics, and is funded by the Smithsonian Institution and the Academia Sinica.
We acknowledge with thanks the variable star observations from the AAVSO International Database contributed by observers worldwide and used in this research.
This research has made use of the DASCH project at Harvard, which is grateful for partial support from NSF grants AST-0407380, AST-0909073, and AST-1313370.
This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France, and of the VizieR catalog access tool, CDS, Strasbourg, France (\href{https://doi.org/10.26093/cds/vizier}{{DOI: 10.26093/cds/vizier}}). The original description of the VizieR service was published in 2000 \citep{Ochsenbein2000}.}
\vspace{5mm}
\facilities{AAVSO, DASCH, SMA, XMM-Newton}
\software{
{\tt Astropy} \citep{Astropy2013},
{\tt CASA} \citep{CASA},
{\tt MIR} ({\url{http://github.com/qi-molecules/sma-mir}}),
{\tt PHOEBE} \citep{Prsa2016},
{\tt XSPEC} \citep{Arnaud1996}}
|
\section{Introduction}\label{sec:Introduction}}
\IEEEPARstart{M}{utation} testing is a white-box testing technique used to evaluate the capability of a test suite in revealing faults. Even though this technique is efficacious in evaluating and guiding the creation of good test cases, its main disadvantage is the great computational cost related to the execution of mutants~\cite{Jia2011,Papadakis2017}. Depending on the number of mutants and test cases, the creation of new test cases, evaluation of equivalent mutants, and the eventual execution of alive mutants might be very expensive, both in terms of machine computational power and human resources.
This problem can be diminished by reducing the number of mutants, and consequently reducing the mutant execution costs~\cite{Pizzoleto2019}. This can be done using mutant reduction strategies~\cite{Jia2011,Papadakis2017,Pizzoleto2019}, however, experiments reported in the literature show that no strategy is the best for all scenarios~\cite{Pizzoleto2019}. Moreover, even the simplest strategies have some parameters to be configured depending on the software under test (SUT). Given the great number of available mutant reduction strategies and the tuning they need, a human tester may find it difficult to choose and manually configure a strategy for their SUT. Let us consider the following scenario as an example: in continuous integration practice, all developers' working copies are merged to a shared mainline repository multiple times a day, and it is very common that the engineers test the software each time they commit their code to this shared repository~\cite{AlshahwanCH0MMM19}. Similarly, an open-source software under development must be rigorously tested before each commit because there are multiple contributors to the shared code. In such scenarios, performing mutation testing using all generated mutants is likely to be infeasible and it is definitively detrimental if the testers have to wait for the mutation testing to finish in order to develop further code and/or test cases. Certainly, the mutants' execution cost can be reduced by using more powerful machines and parallelism, yet the time that could be saved by using mutant reduction strategies is meaningful for nowadays programs with tens of thousands of mutants and is beneficial for the environment~\cite{Calero2015}. Moreover, an arbitrarily selected and configured strategy can result in unsatisfactory cost reduction, test effectiveness maintenance or, even worse, both. In the long run, such arbitrarily chosen strategy could end up not saving as much time as a strategy tailored for a specific testing scenario. Determining the best strategy and its best configuration for a given scenario and context usually is only possible through an experimental evaluation, which can be seen as an optimisation problem itself~\cite{Eiben2011}.
In this sense, we advocate that the usage of hyper-heuristics~\cite{Burke2010} is a viable option to automatically generate and select mutant reduction strategies. Those strategies can be optimised to reduce the mutation execution cost while maintaining the mutation score, and the hyper-heuristic itself can relieve the tester from the burden of selection and configuration tasks. To this end, we introduce Sentinel, an off-line multi-objective learning hyper-heuristic based on Grammatical Evolution (GE)~\cite{Ryan1998} to automatically generate mutant reduction strategies (low-level heuristics) for the mutation testing of Java programs. To the best of our knowledge, there is no previous work that investigates the automatic generation, nor the selection and configuration of mutant reduction strategies.
To evaluate Sentinel, we first compare it with a Random Hyper-Heuristic as a sanity check. Then, we compare the strategies generated by Sentinel with three state-of-the-art conventional strategies: Random Mutant Sampling, Random Operator Selection and Selective Mutation. This analysis has been carried out using 10 real-world open-source systems with four different versions each (for a total of 40 releases and 4,800 comparisons).
Our empirical results show that Sentinel is able to generate strategies providing statistically significantly better results than a Random Hyper-Heuristic for all the systems always with large effect size. Furthermore, Sentinel generates strategies that outperform conventional strategies proposed in previous work. For 70 out of 80 comparisons (i.e.\ 88\%), Sentinel outperforms these state-of-the-art strategies with statistically significant differences and obtains favourable large effect sizes for 227 out of 240 cases (i.e.\ 95\%).
To summarise, the main contributions of this work are:
\begin{itemize}
\item[--] the proposal of Sentinel, a GE based multi-objective hyper-heuristic to automatically generate mutant reduction strategies that are able to provide the best trade-off between mutation execution time reduction and maintenance of the global mutation score;
\item[--] an empirical study investigating 10 real-world open source systems (for a total of 40 releases), which is the largest done so far for investigating hyper-heuristic for mutation testing;
\item[--] a public repository with the data used in this work, which allows for replication and extension of our study and can be also used for other studies on mutation testing~\cite{SentinelWebsite};
\item[--] an open-source reference implementation of Sentinel in Java to allow its usage and extension~\cite{SentinelGH}.
\end{itemize}
The next sections present a background on the main topics of this work (Section~\ref{sec:Background}), the description of Sentinel (Section~\ref{sec:Sentinel}) and its empirical evaluation (Sections~\ref{sec:Empirical Study} and~\ref{sec:Empirical Study:Results}), related work (Section~\ref{sec:Related Work}), and final remarks (Section~\ref{sec:Conclusion}).
\section{Background}
\label{sec:Background}
\subsection{Mutation Testing}
\label{sec:Mutation Testing}
Mutation testing is based on the concept of mutants, where a mutant is a derivation of the original software program with a small syntactic change. Each kind of syntactic change is introduced by using a mutation operator. A test case is said to kill a mutant if the test case yields a different output when executing the original and the mutated program. If a mutant is killed by a test case, it means that the test will be able to reveal this fault. If a mutant is still alive, then the tester may use this information to create stronger test cases able to kill such a mutant.
Usually, the testing criterion for assessing the quality of a test set is the mutation score, which computes the number of non-equivalent mutants that are killed by the test set. A mutant is called equivalent if it yields the same output of the program under test. For such a mutant, there is no test data capable to distinguish it from the original program. Equation~\ref{eq:Mutation Score} shows how the mutation score is computed: \begin{equation}
\MS(T,M) = \dfrac{\DM(T,M)}{|M| - \EM(M)}
\label{eq:Mutation Score}
\end{equation}
\noindent where $M$ is a set of mutants; $T$ is a set of test cases; $\MS(T,M)$ is the mutation score obtained when executing $T$ against $M$; $\DM(T,M)$ is the number of mutants in $M$ killed by $T$; $|M|$ is the number of mutants in $M$; and $\EM(M)$ is the number of equivalent mutants in $M$.
The greater the mutation score, the better the test set is in revealing the faults represented by the mutants. A test set is called ``adequate'' when it is able to kill all the non-equivalent mutants, i.e.\ when it obtains a mutation score of $1.0$ (100\%). Ideally, a test set should kill all mutants of a program, but that is a tricky task. It is very costly (regarding both computational cost and human effort) to create test cases to kill mutants and execute all alive mutants every time a test case is created. Moreover, determining $\EM(M)$ is (generally) undecidable, and the mutation score is usually computed assuming that all mutants are non-equivalent.
\subsection{Mutant Reduction}
\label{sec:Mutant Reduction}
Given the great cost of mutation testing, one can try to reduce its cost by using mutation cost reduction strategies. These strategies can be classified into: ``do fewer'', ``do faster'' and ``do smarter''~\cite{Offutt2001,Pizzoleto2019}. ``Do faster'' strategies consist in the faster execution of mutants such as running compiled programs instead of interpreted programs, mutant schema, parallelism, and others. ``Do smarter'' strategies try to avoid the full execution of mutants, such as Weak Mutation strategies~\cite{Howden1982} that evaluate the state of the mutant right after executing the faulty instruction. ``Do fewer'' strategies try to reduce the set of mutants that need to be executed. In this work we focus on ``do fewer'' strategies.
When reducing mutant sets, the tester must assert that the effectiveness of the reduced mutant set is not compromised, otherwise the reduced set will not be sufficient to guide the selection/creation of good test cases. The mutant reduction problem can be defined as the search for a subset of mutants $M'$ derived from all mutants $M$, such that: i)~the approximation $\MS(T', M) \approx \MS(T, M)$ (herein called ``approaching'') holds when selecting a subset of test cases $T'$ from the SUT's existing test set $T$; or ii)~$\MS(T'', M)$ is maximised when creating a new test set $T''$. Either way, it is the subset of mutants $M'$ that is used to guide the attainment of test cases. If the test effectiveness is not compromised during reduction, then only $|M'|$ mutants are needed to find an adequate set of test cases $T'$ or $T''$ that can efficiently kill all killable mutants in $M$.
Our notion of test effectiveness maintenance during test case selection is similar to the test effectiveness maintenance presented by Amman et al.~\cite{Ammann2014}:\begin{quote}
``Formally, a subset of $T$, denoted $T_{\mathit{maintain}}$, maintains the mutation score with respect to $M$ (and $T$) if for every mutant $m$ in $M$, if $T$ kills $m$ then $T_{\mathit{maintain}}$ kills $m$.''
\end{quote}
This is analogous to our approaching definition $\MS(T', M) \approx \MS(T, M)$, because if $\MS(T, M) = \MS(T', M)$ then we can also state that $T'$ is $T_{\mathit{maintain}}$ according to their definition. Notice that, in the context of our work, $T$ is the complete test suite of the SUT and represents the pool of all available test cases for that SUT. It serves as a ``ground truth'' of what the subsets $T'$ and $M'$ can achieve in terms of test effectiveness. Therefore, the mutation score $\MS(T', M)$ cannot overshoot the global mutation score $\MS(T, M)$, since $T'$ is a subset of $T$. For this reason, in this work, we use the term ``approaching'' instead of ``approximating'' to describe $\MS(T', M) \approx \MS(T, M)$. Moreover, $T$ does not usually contain all possible test cases for that SUT, however, $T$ can serve as a baseline to guide the mutant reduction without the need of generating new tests or evaluating mutants to determine their equivalence.
We can find several strategies in the literature that are used to reduce the mutant set~\cite{Pizzoleto2019}. These strategies (herein called conventional) can sample a subset of mutants~\cite{Mathur1994,Wong1995,Sahinoglu1990}, apply fewer mutation operators~\cite{Offutt1993,Mathur1991,Delamaro2014}, find a set of essential mutation operators to reuse~\cite{Barbosa2001,Offutt1996,Vincenzi1999,Namin2008}, group mutants~\cite{Hussain2008,Ji2009}, or perform higher order mutation~\cite{Harman2010}. More recent strategies are based on search based algorithms~\cite{Silva2016}. Such strategies are part of the Search Based Software Engineering (SBSE) field~\cite{Harman2012}, which in turn aims at solving hard software engineering problems with the aid of search based algorithms.
The most common conventional strategies were identified from the literature~\cite{Pizzoleto2019,Jia2011,Papadakis2017} and used in the experiments of this work. These strategies are explained next.
Random Mutant Sampling (RMS)~\cite{Mathur1994,Wong1995,Sahinoglu1990} randomly selects a set of mutants from the pool of all mutants generated for a given system. With this strategy, the tester selects a subset of mutants before executing them, thus the mutants execution is supposedly faster than executing all mutants. This strategy needs to be configured with a percentage of mutants to be sampled. Usually, the more mutants selected, the better the effectiveness of the sampled mutant set, but the greater the time spent with the mutation testing.
Random Operator Selection (ROS)~\cite{Offutt1993,Mathur1991,Delamaro2014} randomly selects a subset of mutation operators to generate mutants. Instead of generating and then selecting mutants like RMS, with ROS the selection is done before any mutant is generated. The advantage of this strategy is that, not only fewer mutants are executed, but also fewer operators are applied. This strategy requires one parameter that can be a fixed number or a percentage of operators to execute. Usually, the more operators executed, the more mutants generated, but the costlier the mutation activity.
Selective Mutation (SM)~\cite{Offutt1993,Mathur1991,Delamaro2014,Barbosa2001,Offutt1996,Vincenzi1999,Namin2008} is used to select a subset of operators. ROS is similar to SM but while ROS selects operators at random, SM uses information about the mutation testing activity in order to guide the selection of operators. In this work we used the SM version that discards the $n$ operators that generate the most mutants. This can potentially discard operators that generate a large number of redundant mutants which can be killed by a single test case. However, while it is trivial to assess which operators generate the most mutants, it is not a trivial task to assess how many of those operators should be excluded from the mutation testing activity, thus the tuning of $n$ should be considered beforehand.
\subsection{Multi-Objective Evolutionary Algorithms}
\label{sec:Sentinel:Multi-Objective Optimisation}
Evolutionary Algorithm (EA) are based on the theory of evolution~\cite{Eiben2003}, in which a population evolves over several generations by means of natural selection and reproduction. In EAs, a solution is an individual in the population represented by a chromosome (genotype). Usually the representation is an array of genes (e.g.\ bit, float or integer) on which the algorithm performs perturbations to find different solutions. Over several generations, parent individuals are selected to reproduce (crossover) and generate offspring that carry their genes. Then, the offspring are mutated for diversity and the best solutions survive to the next generation. In each generation, the solutions are evaluated using fitness functions, which compute their overall quality. This continues until a stopping condition is met.
Mono-objective EAs are straightforward: the solutions are given a fitness value each, ranked accordingly, and then the ones with the best fitness are selected to reproduce and survive during the evolutionary process. At the end of the optimisation, the engineer selects the solution with the best fitness. On the other hand, Multi-Objective Evolutionary Algorithms (MOEAs)~\cite{Coello2007} optimise simultaneously two or more objectives with equal weight considering the Pareto dominance concept~\cite{Coello2007}. If all objectives $z \in Z$ of a problem are of minimisation, a solution $x$ is said to dominate a solution $y$ ($x \prec y$) if it is better or equal in all objectives and better in at least one: \begin{equation}
\begin{split}
\forall z \in Z: z(x) \leq z(y)\\
\exists z \in Z: z(x) < z(y)
\end{split}
\end{equation}
\noindent otherwise the solutions are called non-dominated. Non-dominated solutions are not easy to compare, because each one is a viable solution for the problem at hand and represents a satisfactory trade-off between objectives. It is up to the engineers to select the solution that best fits their needs.
When performing experiments and comparing the results of algorithms however, each resulting Pareto front must be evaluated as a whole so that the general performance of the algorithm can be assessed. Quality indicators~\cite{Zitzler2003} are usually applicable to quantify the quality of the results. The main goal of quality indicators is to provide a means of computing a single value (unary indicators) that represents the quality of a Pareto front, or quantifying the difference between two fronts (binary indicators). In this work we use Hypervolume (HV) and Inverted Generational Distance (IGD)~\cite{Zitzler2003}, two of the most common unary indicators in the literature also used in SBSE~\cite{Guizzo2015,Guizzo2016,Guizzo2017,Guizzo2017b,FerrucciHRS13,ZhangHJS15, SarroPH16,Sarro2017}.
HV~\cite{Zitzler2003} computes the area or volume of the objective space that is dominated by a given front in relation to a reference point (usually the worst possible point). The greater the HV value, the greater the area of the objective space a front dominates, thus the better the algorithm in terms of finding non-dominated solutions. IGD~\cite{Zitzler2003} computes the distance between the true Pareto front (the best solutions found so far) and a given front. This is done by summing the distance in the objective space from each solution of the true Pareto front to the nearest solution of the front being evaluated. Therefore, the lower the IGD value, the closer the front is to the true Pareto front, and consequently the closer the solutions are to the best found.
\subsection{Hyper-Heuristics}
\label{sec:Hyper-Heuristics}
Hyper-Heuristics are used to select or generate the best low-level heuristics instead of trying to solve a problem directly~\cite{Burke2010}. Hence, instead of acting over the search space, hyper-heuristics search the heuristic space for good heuristics that can solve the problem.
Hyper-Heuristics can be of on-line or off-line nature. On-line hyper-heuristics select or generate the low-level heuristics during the optimisation, whereas off-line hyper-heuristics apply the training before the optimisation and then reuse the selected or generated heuristics in unseen instances of the problem. A low-level heuristic is any heuristic that is being selected or generated by a hyper-heuristic, being either a construction or a perturbation heuristic. Construction low-level heuristics start with an empty solution and gradually build it, whereas perturbation low-level heuristics start with fully built solutions and change them to generate new solutions. Examples of low-level heuristics are search operators for evolutionary algorithms, local search algorithms and metaheuristics.
\section{Our Proposal: Sentinel}
\label{sec:Sentinel}
Sentinel\footnote{Name inspired by the \textit{Sentinel} character of the X--Men comics by Marvel Comics (\url{http://marvel.com/universe/X-Men}). In the Marvel universe, a Sentinel is a giant robot in charge of killing mutants.} is a multi-objective approach based on hyper-heuristics to automatically generate mutant reduction strategies for Java programs. Instead of trying to directly perform the mutant reduction, Sentinel tries to automatically find a set of optimal strategies that can reduce the cost of mutation testing while maintaining the global mutation score (i.e.\ the mutation score when considering all available mutants).
The ultimate goal of Sentinel is to generate optimal strategies that are effective in reducing mutants. An example of a strategy generated by Sentinel is: A)~retain 80\% of the available operators; B)~execute 100\% of retained operators; C)~group generated mutants by operator; D)~order the groups by number of mutants in descending order; E)~discard the first two groups of mutants; and F)~store in $M'$ 10\% of the remaining mutants at random from each group. At the end of the strategy execution, it will simply output the reduced set of mutants $M'$ to be used as a replacement to the whole set of mutants $M$. These operations -- such as the ones from the example (A--F) -- were extracted from conventional strategies and stored in a grammar used by a GE algorithm~\cite{Ryan1998} (a type of EA) to generate the strategies (explained in the next section). With the GE algorithm and the grammar file, Sentinel is able to combine operations from different kinds of conventional strategies in order to build new and unseen strategies.
The search for strategies is done during a learning phase (i.e., evolutionary process), where Sentinel's GE implementation iteratively generates new strategies by means of recombination of genetic information from parents, mutation of children genes and natural selection of best fit strategies across several subsequent generations (see Section~\ref{sec:Sentinel:Multi-Objective Optimisation}). At the end of the evolutionary process, Sentinel returns a set containing the strategies with the best trade-off between execution time minimisation and mutation score conservation.
For the generation and execution of mutants, Sentinel uses PIT 1.2.0\footnote{\url{http://pitest.org/}}~\cite{Coles2016}, a mutation tool for Java programs. We chose to use PIT because it is one of the fastest and widely used Java mutation tools publicly available~\cite{Bowes2016, Gopinath2016, Laurent2017}. PIT already uses several cost reduction techniques, such as bytecode manipulation, multi-threading, essential operators, test case prioritisation, and code coverage analysis. Furthermore, PIT uses Partial Mutation, meaning that mutants are executed until they are killed, and returning to the user the test cases that were used to kill such mutants (the reduced set of test cases $T'$ mentioned in Section~\ref{sec:Mutant Reduction}). By using Sentinel in combination with PIT, we believe that we can further improve the cost reduction without disregarding the mutation score for non-trivial programs.
In this sense, instead of executing all mutants $M$, the tester chooses a strategy generated by Sentinel and execute it to obtain a reduced set of mutants $M'$ which is used by PIT to perform the mutation analysis. Ideally, the Sentinel training process is done only once and the resulting strategies can be reused every time the tester needs to perform the mutation testing. Therefore, the advantages of Sentinel are: i)~it provides the tester with mutant reduction strategies specially tailored and optimised for their software; and ii)~it automates the generation of strategies, so that a tester does not need to select and configure existing strategies manually through experimentation. The next subsections present details about Sentinel and the generated strategies.
\subsection{Solution Representation and Genotype--Phenotype Mapping}
\label{sec:Representation}
The overall process for generating a strategy is shown in Figure~\ref{fig:Strategy GPM} and it is called Genotype--Phenotype Mapping (GPM)~\cite{Whigham2017}.
\begin{figure}[tb]
\centering
\includegraphics[width=0.8\linewidth]{images/GPM.png}
\caption{Genotype--Phenotype Mapping of Sentinel.}
\label{fig:Strategy GPM}
\end{figure}
Although the ultimate goal of Sentinel is evolving a set of executable strategies (i.e., phenotype), GE algorithms~\cite{Ryan1998} cannot work directly on the phenotype and use a chromosome (i.e., genotype) instead to evolve the solutions, similarly to other types of EAs~\cite{Coello2007}. In our case, the chromosome is encoded as an array of $n$ integers, where each integer represents a rule to be selected from a grammar. In fact, the GPM processes transforms this array into a fully executable strategy by using a set of predefined rules forming the grammar. For example, Figure~\ref{fig:Grammar} is an excerpt of the default Sentinel grammar\footnote{The complete grammar used in this study can be found on Sentinel's web page~\cite{SentinelWebsite}.}.
\begin{figure}[tb]
\centering
\begin{mdframed}
\begin{grammar}
\fontsize{8pt}{8pt}
<mutantOperation> ::= "Retain Mutants" <selectMutants> | "Discard Mutants" <selectMutants>
<selectMutants> ::= <mutantSelectionType> <quantity> | <mutantSelectionType> <percentage>
<mutantSelectionType> ::= "Select Mutants" <selectionType> <mutantSorting> | "Select Mutants by Groups" <mutantGroupSelectionType> <selectMutants>
\end{grammar}
\end{mdframed}
\caption{Excerpt of Sentinel's grammar.}
\label{fig:Grammar}
\end{figure}
Each one of the rules before $::=$ can be translated into one of the options delimited by $|$. One of the options is chosen by consuming an integer value (gene) of the chromosome: $\mathit{chosen\_option} = \mathit{mod}(\mathit{gene}, \mathit{num\_options})$. If the option has a non-terminal rule between $<$ and $>$, then more genes must be consumed to translate each of those rules into terminal ones. The grammar rules actually represent the available operations which Sentinel encompasses: i)~Retain -- retain mutants or operators, like a filter where only the selected elements remain in the pool; ii)~Discard -- discard mutants or operators; iii)~Group -- group mutants or operators in clusters to perform actions over these groups; or iv)~Execute -- execute the operators to generate mutants. At the end of the GPM process, the genotype genes of a solution are translated into a set of operations for building a strategy (i.e., phenotype). This fully assembled strategy can be seen as an execution tree, where each node is an operation with its required parameters. In the example shown in Figure~\ref{fig:Strategy GPM}, the strategy execution starts at node A, continues until it reaches a leaf node and then proceeds to execute the next branches of the tree until all nodes are executed. During this execution path, the strategy can select operators, execute them to generate mutants and then select the mutants. The final result of such execution is the reduced mutant set $M'$.
\subsection{Objective Functions}
\label{sec:Sentinel:Objective Functions}
Sentinel generates and evaluates strategies according to two objectives: \textit{TIME} and \textit{SCORE}. Both of them are based on the mean values obtained by a strategy $e$ executed for $n$ repetitions over the training instance.
The \textit{TIME} minimisation objective for a given strategy $e$ is given by Equation~\ref{eq:Time}:\begin{equation}
\label{eq:Time}
\downarrow \mathit{\textit{TIME}}(e) = \frac{\sum_{i = 1}^{n} \mathit{cpuTime}(e_{i})}{\sum_{i = 1}^{n} \mathit{cpuTime}(c_i)}
\end{equation}
\noindent where $c$ is the conventional mutation testing procedure (generating all mutants and executing all of them). This function computes the relative CPU time used for executing the strategy $e$ and the resulting mutants in comparison to the conventional mutation procedure, i.e.\ it computes the fraction of time that the strategy takes to execute in comparison to executing all operators and mutants.
Previous work on mutation testing cost reduction usually measure the cost by computing the number of mutants~\cite{Pizzoleto2019}: the fewer the mutants, the cheaper the mutation testing. However, depending on the mutation, a mutant may take more time to execute than another one~\cite{Mresa1999,Zhang2014,Derezinska2016,Derezinska2017,Gligoric2013} (i.e.\ one cannot assume all mutants have the same cost). Thus, in this work we have actually executed each of the mutants selected by a given strategy and computed the time taken to perform the mutation testing activity. It is important to note that, due to the stochastic nature of strategies (both generated by Sentinel and conventional ones from the literature) and to small fluctuations in the CPU time measurement, we adopted a repetition of $n$ strategy executions to compute an average for both functions in each fitness evaluation. The usage of CPU time instead of number of mutants, and the repetition approach are adopted to improve the accuracy of the cost and score measurements of each strategy. Furthermore, we also applied this repetition approach during the experimentation to increase the accuracy in each independent run, while also performing 30 independent runs to cater for the stochasticity and allow the use of statistical tests. The downside of such a measurement is that the experiments become very expensive, both during the training and the testing.
The \textit{SCORE} objective for a strategy $e$ is given by Equation~\ref{eq:Score}:
\begin{equation}
\label{eq:Score}
\uparrow \mathit{\textit{SCORE}}(e) = \frac{\dfrac{1}{n} \sum_{i = 1}^{n} \MS(T'_{i},M)}{\MS(T, M)}
\end{equation}
\noindent where $\MS$ is the mutation score and $T'$ is a test set (selected by PIT from all available test cases $T$) used to kill $M'$. This function computes the average relative test effectiveness obtained by the reduced mutant set, i.e.\ it computes the relative mutation score obtained by a subset of tests $T'$ on all mutants $M$. By maximising this objective, a strategy cannot obtain a mutation score higher than the original one. For instance, if the original mutation score $\MS(T, M)$ is $0.8$ and the average mutation score $\MS(T',M)$ obtained by the strategy $e$ is $0.5$, then \textit{SCORE}$(e)$ will yield $0.625$, since $0.5$ represents 62.5\% of the global mutation score $0.8$. If \textit{SCORE} is $1.0$ ($T' = T_{\mathit{maintain}}$), then it means that $M'$ requires a subset $T'$ that can kill all killable mutants in $M$, i.e\ $T'$ is as good as $T$ in killing $M$ but only $M'$ is required to obtain $T'$.
An optimal strategy is the one that finds a mutant set that is fast to execute and that maintains the global mutation score. However, these two objectives are conflicting because by optimising one, the other will probably worsen (as explained in Section~\ref{sec:Sentinel:Multi-Objective Optimisation}). In the problem of mutant reduction, a strategy that can greatly reduce the cost of the mutation activity at the cost of mutation score approaching can be seen as good as a strategy that greatly approaches the mutation score for a slower execution speed (i.e.\ both are non-dominated). It is up to the engineer to decide which strategy shall be used: the one that saves the most execution time, the one that better approaches the mutation score, or one in between with an acceptable compromise.
\subsection{Implementation}
\label{sec:Sentinel:Implementation Aspects}
Sentinel uses a GE algorithm implementation based on the Non-dominated Sorting Genetic Algorithm II (NSGA--II)~\cite{Deb2002} MOEA to guide the strategy generation. The GE implementation of Sentinel is presented in Algorithm~\ref{alg:GE}. The main structure of the algorithm is very similar to a conventional multi-objective GE, the only exception is that our algorithm uses pruning and duplication operators after the mutation of chromosomes. The former helps to avoid the mutation and crossover of useless genes, and the latter helps to minimise chromosome wraps~\cite{Ryan1998}. The GE parameters used in the experiment are shown in Table~\ref{tab:Parameters} and were chosen based on previous work~\cite{Guizzo2015,Guizzo2017,Mariani2016}.
\begin{algorithm}[tb]
\caption{GE implementation of Sentinel}
\label{alg:GE}
\fontsize{8pt}{8pt}\selectfont
\Begin{
$\mathit{chromosomes} \gets$ Randomly initialise the population\;
$\mathit{strategies} \gets$ Perform the GPM on $\mathit{chromosomes}$ using the grammar file\; \label{alg:GE:mapping}
$\mathit{time} \gets$ Compute CPU time for the next two steps\;
$\mathit{mutants} \gets$ Execute $\mathit{strategies}$\; \label{alg:GE:results}
$\mathit{score} \gets$ Execute $\mathit{mutants}$\;
Evaluate $\mathit{chromosomes}$ according to $\mathit{time}$ and $\mathit{score}$\; \label{alg:GE:evaluate}
\While{maximum fitness evaluations not reached}{
$\mathit{parents} \gets $ Select best parents in $\mathit{chromosomes}$\;
$\mathit{offspring} \gets$ Recombine $\mathit{parents}$ with crossover\;
Mutate $\mathit{offspring}$\;
$\mathit{strategies} \gets$ Perform the GPM on $\mathit{offspring}$ using the grammar file\; \label{alg:GE:mapping offspring}
$\mathit{time} \gets$ Compute CPU time for the next two steps\;
$\mathit{mutants} \gets$ Execute $\mathit{strategies}$\;
$\mathit{score} \gets$ Execute $\mathit{mutants}$\;
Evaluate $\mathit{chromosomes}$ according to $\mathit{time}$ and $\mathit{score}$\;
$\mathit{chromosomes} \gets$ Select the best individuals in $\mathit{chromosomes}$ and $\mathit{offspring}$\;
}
\Return{Non-dominated strategies in $\mathit{chromosomes}$}\;
}
\end{algorithm}
\begin{table}[tb]
\centering
\caption{Sentinel GE parameters.}
\label{tab:Parameters}
\fontsize{6pt}{6pt}\selectfont
\def1.2{1.2}
\begin{tabular}{lr}
\toprule
Parameter & Value \\ \midrule
Independent Runs & 30 \\
Strategy Repetitions & 5 \\
Maximum Evaluations & 10,000 \\
Population Size & 100 \\
Crossover Operator & Single Point Crossover \\
Crossover Probability & 100\% \\
Mutation Operator & Random Integer Mutation \\
Mutation Probability & 1\% \\
Prune Probability & 10\% \\
Duplicate Probability & 10\% \\
Lower Gene Bound & 0 \\
Upper Gene Bound & 179 \\
Maximum Chromosome Length & 100 \\
Minimum Chromosome Length & 15 \\
Maximum Wraps & 10 \\ \bottomrule
\end{tabular}
\end{table}
At the start of the algorithm execution, a set of 100 chromosomes are generated, mapped into strategies and then evaluated. For each strategy, the evaluation consists in: i)~mapping the strategy using GPM; ii)~executing the strategy over a training instance and its test suite; iii)~collecting the reduced mutant set; iv)~executing the mutant set; and v)~computing both \textit{SCORE} and \textit{TIME} for the execution of the strategy and resulting mutants. Then, the best strategies are selected to reproduce and new children are generated throughout crossover and mutation. For the replacement, i.e.\ substitution of parents for their children to survive to the next generation, the NSGA--II procedure of pruning the populations is applied considering both convergence and diversity. This is done for 100 generations and then the algorithm is terminated returning the non-dominated strategies (Pareto front) found along the generations.
In order to reduce the training cost of Sentinel and avoid the execution of all mutants during each evolutionary process, we implemented a caching process to store information about each mutant and operator. In the first step, Sentinel executes each operator and mutant separately for the system. This means that for each mutant Sentinel creates a PIT mutation pool and executes it separately in a different Java Virtual Machine (JVM) instance. Then, the execution cost and mutation score are stored in a cache file, which in turn is used as a source of information during the fitness evaluation. Instead of executing all mutants obtained by a strategy at each fitness evaluation, Sentinel uses the information stored in this cache file to compute the objective values. Even though there is an overhead for setting up a JVM instance for each mutant during the caching, this is more viable than letting the strategies execute all selected mutants. Finally, we use the default options for PIT and apply Sentinel using the seven default mutation operators as commonly done in previous work~\cite{Bowes2016, Coles2016}.
\subsection{Final Remarks on Sentinel}
Sentinel is a novel approach since, as far as we are aware, there is no previous work on automated optimal generation/selection of mutant reduction strategies. Furthermore, Sentinel uses the actual strategies execution time and the score approaching as objectives for the optimisation. It also allows the usage of other objectives, giving the testers the possibility of generating strategies tailored to their needs. Previous work on mutant reduction usually rely only on the number of mutants and mutation score~\cite{Pizzoleto2019}.
\section{Empirical Study Design}
\label{sec:Empirical Study}
This section presents the design of the empirical study conducted to evaluate the strategies generated by Sentinel. The main goal of this evaluation is to determine the capability of the strategies generated by Sentinel in reducing the overall execution time (i.e.\ cost) of mutation testing and in maintaining the mutation score. Therefore, a strategy is better than another if it is better or equal for these two objectives, and better for at least one of them, according to the Pareto dominance concept~\cite{Coello2007}.
\subsection{Research Questions}
\label{RQs}
In order to assess Sentinel's effectiveness we answer three research questions (RQs).
First of all, we carry out a sanity check to assure that the hyper-heuristic mechanism used by Sentinel effectively learns how to generate good strategies during the training phase and therefore it is better than using a random strategy:
\textbf{RQ1 -- Sanity Check:} Is Sentinel better than a random strategy generation algorithm?
To answer this question, we run both Sentinel and the random generation algorithm for 30 independent runs using 10 different real-world systems widely used in previous work~\cite{Gopinath2016,Laurent2017,Just2014,Zhang2016}. Sentinel is a multi-objective approach, thus each independent run generates an approximated Pareto front where each solution of the front is a generated strategy. Therefore, we evaluate the resulting fronts using the HV and IGD indicators~\cite{Zitzler2003} (described in Section~\ref{sec:Sentinel:Multi-Objective Optimisation}), which have been widely used in previous SBSE work~\cite{FerrucciHRS13,ZhangHJS15,SarroPH16,Sarro2017}.
Because the true Pareto front for the problem we investigate herein is unknown and infeasible to discover, we create such a front with all the non-dominated solutions found by all algorithms, as done in previous work~\cite{Guizzo2015,Mariani2016,SarroPH16,Guizzo2017,Sarro2017}.
For each experiment, we normalise the objective values for computing both indicators, thus the HV and IGD value's range may vary from one experiment to another.
We also used the Kruskal--Wallis statistical test~\cite{Kruskal1952} to assess if there is any statistical difference between the HV and IGD values achieved by the different algorithms, and the Vargha--Delaney A12 effect size~\cite{Vargha2000} to compute the magnitude of the differences. Kruskal--Wallis is a statistical test that can show if the data of two or more groups are statistically different. In this paper we assume the confidence of 95\%, hence if the computed \textit{p-value} is lower than $0.05$, then there is statistical difference between the groups. Kruskal--Wallis is actually an extension of Wilcoxon rank sum test~\cite{Wilcoxon1945} by allowing comparisons between three or more groups. The Vargha--Delaney $\hat{A}_{\mathit{12}}$ effect size gives the magnitude of the difference between two groups A and B. The Vargha--Delaney $\hat{A}_{\mathit{12}}$ value varies between $[0,1]$, where $0.5$ represents absolute no difference between the two groups, values below $0.5$ represent that group B obtains greater values than group A, and values above $0.5$ represent that A obtains greater values than B. We chose these tests because they are non-parametric, i.e.\ they must be used when one cannot assume normal distribution of data.
After assessing if the strategies generated by Sentinel are better than random guessing in RQ1, we compare these strategies to conventional strategies commonly used in the literature: Selective Mutation (SM), Random Operator Selection (ROS), and Random Mutant Sampling (RMS) (described in Section~\ref{sec:Mutation Testing}). It is important to emphasize that, while RQ1 is designed to evaluate the hyper-heuristic generation process, RQ2 focuses on evaluating the strategies themselves. Therefore, our second question is as follows:
\textbf{RQ2 -- Sentinel's Strategies vs. Conventional Strategies:} Are the strategies generated by Sentinel capable of outperforming conventional mutant reduction strategies in both time reduction and score approaching?
Each type of strategy (SM, ROS, RMS and Sentinel-generated strategies) has multiple strategy variations that differ on their configuration parameters. The conventional strategies were implemented using the same framework as Sentinel and a set of strategies with parameter variations was created for each of them. The strategies are: i)~RMS, randomly selecting from 10\% to 90\% (steps of 10\%) of mutants; ii)~ROS, selecting from 10\% to 90\% (steps of 10\%) of the operators; and iii)~SM, excluding from one to six (steps of one) of the operators that generate the most mutants. RMS is not to be confused with the random generation hyper-heuristic evaluated in RQ1. An RMS strategy selects mutants, whereas a hyper-heuristic generates strategies, thus RQ1 and RQ2 are focused on two different aspects of our experiments.
For each of the 10 systems, we used four different versions. Hence, all strategies were executed over four different versions of the 10 systems with 30 independent runs each, for a total of 1,200 independent runs for each strategy (4,800 in total). We had to limit our analysis to three types of conventional strategies and four system versions due to the high cost of the experiment including manual collection of the data and algorithms running time.
Each of the four types of strategy (Sentinel-generated and the three conventional ones) represents a set of possible solutions, where each strategy is a non-dominated solution in the Pareto front. Given the multiple Pareto fronts, one for each system and type of strategy, we applied the same indicators (HV and IGD) and statistical tests (Kruskal--Wallis and Vargha--Delaney $\hat{A}_{\mathit{12}}$) as the previous RQ to analyse the results. A positive answer to this question will confirm Sentinel's ability to automatically generate better strategies than existing ones.
Once we assessed whether Sentinel is able to provide better strategies than conventional ones, we move to investigating if the strategies learnt by Sentinel on a given version remain effective for subsequent unseen versions, as follows:
\textbf{RQ3 -- Sentinel Strategies Effectiveness Over Time:} Can the strategies generated by Sentinel on a given software version be effectively used to test subsequent versions?
The longer a model remains effective the better. This is a crucial aspect of any automated learner since software change over time and a model built at some point may not be useful later on, thus requiring us to re-train it with consequent costs. If we find that the strategies need to be generated at every new version of the code, as previously generated ones would be ineffective for new versions, then Sentinel would not be practical.
In order to answer RQ3, we investigate this scenario: a tester trains Sentinel by using the current version of a system and, when a new version of the system is available, these previously generated strategies are used (rather than re-run Sentinel on this new version to obtain new strategies). To this end, we use the same data collected for answering RQ2, we use Sentinel to generate strategies for the first version of each of the software systems and then reuse them for the three subsequent versions. We compare the Sentinel-generated strategies to the conventional ones, since the latter can be powerful enough on subsequent versions of the system and the former might not be needed as a consequence.
\subsection{Subjects}
\label{sec:Empirical Study:Subjects}
In our empirical study we used the following 10 real-world systems (also used in previous work~\cite{Gopinath2016,Laurent2017,Just2014,Zhang2016}):
\begin{itemize}
\item[--] {\it Apache Commons Beanutils}, a library for wrapping around reflection and introspection\footnote{\url{https://github.com/apache/commons-beanutils}};
\item[--] {\it Apache Commons Codec}, an encoder/decoder for several formats such as Base64 and Hexadecimal\footnote{\url{https://github.com/apache/commons-codec}};
\item[--] {\it Apache Commons Collections}, a library for collections and arrays manipulation\footnote{\url{https://github.com/apache/commons-collections}};
\item[--] {\it Apache Commons Lang}, an utility library for \textit{java.lang}\footnote{\url{https://github.com/apache/commons-lang}};
\item[--] {\it Apache Commons Validator}, a client-server validation library\footnote{\url{https://github.com/apache/commons-validator}};
\item[--] {\it JFreeChart}, a framework for manipulating charts\footnote{\url{https://github.com/jfree/jfreechart}};
\item[--] {\it JGraphT}, a library with classes and algorithms for graphs\footnote{\url{https://github.com/jgrapht/jgrapht}};
\item[--] {\it Joda--Time}, a date and time library\footnote{\url{https://github.com/JodaOrg/joda-time}};
\item[--] {\it OGNL}, an Object--Graph Navigation Language (OGNL) library to express Java object's properties\footnote{\url{https://github.com/jkuhnert/ognl}};
\item[--] {\it Wire}, a project for Protocol Buffers for Android\footnote{\url{https://github.com/square/wire}}.
\end{itemize}
Because these programs have several minor and major versions and it would be impracticable to test Sentinel on all of their versions, we decided to use only the oldest major versions of each program that could be mutated by PIT for the training phase and three subsequent versions for the testing phase. The properties of these programs are summarised in Table~\ref{tab:Programs}. The code churn is computed by summing the number of new Logical Lines Of Code (LLOC), modified LLOC, and removed LLOC.
\begin{table}[tb]
\caption{\textbf{Software versions used in our study.} \textit{LLOC} is the number of logical lines of code; \textit{Churn} is the code churn from previous version; $|T|$ is the number of test cases in the test suite; \textit{T. LLOC} is the number of logical lines of test code; \textit{Cov} is the statement and branch coverage percentages; $|M|$ is the number of mutants generated by PIT; \textit{MS} is the mutation score obtained with the available test suite.}
\label{tab:Programs}
\fontsize{6pt}{6pt}\selectfont
\def1.2{1.2}
\begin{tabular}{lrrrrrrr}
\toprule
Program & \textit{LLOC} & \textit{Churn} & $|T|$ & \textit{T. LLOC} & \textit{Cov} & $|M|$ & $\MS$ \\ \midrule
beanutils-1.8.0 & 11,279 & -- & 877 & 20,486 & 60/62 & 2,827 & 0.61 \\
beanutils-1.8.1 & 11,362 & 179 & 892 & 20,934 & 60/61 & 2,855 & 0.59 \\
beanutils-1.8.2 & 11,362 & 22 & 893 & 20,966 & 60/61 & 2,856 & 0.59 \\
beanutils-1.8.3 & 11,376 & 35 & 896 & 21,033 & 60/61 & 2,857 & 0.59 \\ \hline
codec-1.4 & 2,994 & -- & 284 & 6,846 & 99/94 & 1,587 & 0.91 \\
codec-1.5 & 3,551 & 2,631 & 380 & 8,307 & 98/93 & 1,895 & 0.91 \\
codec-1.6 & 4,554 & 1,472 & 421 & 9,034 & 96/93 & 2,196 & 0.88 \\
codec-1.11 & 8,109 & 7,858 & 875 & 11,774 & 96/90 & 3,473 & 0.85 \\ \hline
collections-3.0 & 22,842 & -- & 1,896 & 22,828 & 90/74 & 7,488 & 0.29 \\
collections-3.1 & 25,372 & 6,552 & 2,346 & 25,734 & 82/75 & 8,298 & 0.30 \\
collections-3.2 & 26,323 & 2,022 & 2,566 & 29,076 & 81/75 & 8,637 & 0.31 \\
collections-3.2.1 & 26,323 & 161 & 2,566 & 29,076 & 81/75 & 8,632 & 0.31 \\ \hline
lang-3.0 & 18,997 & -- & 1,902 & 31,008 & 94/91 & 9,072 & 0.85 \\
lang-3.0.1 & 19,495 & 758 & 1,964 & 31,804 & 92/90 & 9,328 & 0.85 \\
lang-3.1 & 19,499 & 354 & 1,976 & 32,446 & 92/90 & 9,333 & 0.85 \\
lang-3.2 & 22,532 & 12,053 & 2,390 & 38,963 & 94/90 & 10,970 & 0.86 \\ \hline
validator-1.4.0 & 5,411 & -- & 414 & 6,367 & 79/72 & 1,811 & 0.74 \\
validator-1.4.1 & 6,031 & 1,228 & 442 & 7,389 & 82/74 & 1,917 & 0.75 \\
validator-1.5.0 & 6,669 & 1,390 & 481 & 7,922 & 84/74 & 1,979 & 0.75 \\
validator-1.5.1 & 7,014 & 524 & 486 & 8,051 & 85/75 & 1,982 & 0.75 \\ \hline
jfreechart-1.0.0 & 68,796 & -- & 1,023 & 26,823 & 62/39 & 23,417 & 0.27 \\
jfreechart-1.0.1 & 68,663 & 1,816 & 1,027 & 27,016 & 62/39 & 23,490 & 0.28 \\
jfreechart-1.0.2 & 73,162 & 9,036 & 1,073 & 28,504 & 63/39 & 24,091 & 0.28 \\
jfreechart-1.0.3 & 77,621 & 12,640 & 1,234 & 32,825 & 47/40 & 26,401 & 0.28 \\ \hline
jgrapht-0.9.0 & 12,978 & -- & 188 & 7,030 & 79/72 & 2,976 & 0.62 \\
jgrapht-0.9.1 & 13,822 & 2,110 & 647 & 8,184 & 79/73 & 3,147 & 0.66 \\
jgrapht-0.9.2 & 15,661 & 8,079 & 728 & 10,046 & 80/75 & 3,825 & 0.69 \\
jgrapht-1.0.0 & 16,417 & 11,442 & 1,201 & 13,609 & 81/75 & 4,270 & 0.71 \\ \hline
joda-2.8 & 28,479 & -- & 2,967 & 54,645 & 89/81 & 10,225 & 0.62 \\
joda-2.8.1 & 28,479 & 107 & 2,967 & 54,645 & 89/81 & 10,225 & 0.62 \\
joda-2.8.2 & 28,479 & 142 & 2,967 & 54,645 & 89/81 & 10,225 & 0.62 \\
joda-2.9 & 28,624 & 366 & 2,985 & 54,985 & 89/81 & 10,321 & 0.62 \\ \hline
ognl-3.1 & 16,103 & -- & 54 & 6,229 & 69/60 & 5,650 & 0.25 \\
ognl-3.1.1 & 16,102 & 27 & 53 & 6,223 & 69/60 & 5,652 & 0.26 \\
ognl-3.1.2 & 16,103 & 9 & 56 & 6,252 & 69/60 & 5,652 & 0.26 \\
ognl-3.1.3 & 16,109 & 9 & 57 & 6,268 & 69/60 & 5,654 & 0.26 \\ \hline
wire-2.0.0 & 1,354 & -- & 70 & 1,776 & 59/53 & 513 & 0.64 \\
wire-2.0.1 & 1,354 & 0 & 70 & 1,776 & 59/53 & 513 & 0.64 \\
wire-2.0.2 & 1,353 & 3 & 70 & 1,776 & 59/53 & 513 & 0.64 \\
wire-2.0.3 & 1,405 & 226 & 71 & 1,794 & 58/53 & 524 & 0.62 \\ \bottomrule
\end{tabular}
\end{table}
\subsection{Threats to Validity}
\label{sec:Empirical Study:Threats to Validity}
We evaluated the effectiveness of the strategies generated by Sentinel using 10 different software programs having four different versions each. The initial experiments design was to use 10 systems and 10 versions each for a total of 100 projects. However, using such a large number of projects revealed to be infeasible, given the high cost of computing the CPU time of each strategy execution. Even though the total number system versions evaluated (40) is similar or even higher to the number of systems in related work~\cite{Gopinath2016,Just2014, Laurent2017,Zhang2016}, we cannot assert that this is enough to generalise the results to other systems. Another threat to the generalisation of the results is the churn between versions which, for some systems, are somewhat small. To minimise these threats, we tried to evaluate systems of different sizes, domains, test coverage, mutation score, and that had already been used in previous work.
To answer RQ2 we trained Sentinel on the first version of a program and tested the generated strategies on the same version and three unseen subsequent versions. In this way we recreated a scenario that we would expect a tester to follow when using Sentinel: train on a version of their software, and use the generated strategies while testing that same version and subsequent ones. Training and testing an approach on the same version may usually lead to overfitting. In order to minimise this threat, we consider different versions of the software and we perform an analysis using only these subsequent versions in RQ3.
Because Sentinel uses PIT, the results of the experiments depend on the configuration of PIT and its internal features, such as test case prioritisation and mutation operator execution. In this sense, using different configurations for PIT may lead to different results, specially considering that the set $T'$ used to compute the mutation score is obtained based on PIT's test case prioritisation. To minimise this threat, we left all the default configurations of PIT unchanged to all strategy types, so that they all are tested in a same environment. Furthermore, we performed 30 independent runs to cater for stochasticity.
We compared Sentinel-generated strategies to all identified variations of the SM and ROS strategies for the default operators of PIT. For RMS we had to limit the number of strategies by defining a fixed step percentage for their configuration as done in previous work~\cite{Zhang2013,Lima2016,Zhang2010,Mathur1994}. Even though the step sizes vary from one work to another, limiting the set of sampling strategies is done to allow the empirical experimentation in feasible time. Using a smaller step of 1\% instead of 10\% for RMS could yield different results. Furthermore, the strategies generated by Sentinel are tailored specifically for the system in hand. These strategies are potentially more powerful than conventional ones for the system, hence the good results may be biased towards Sentinel's strategies.
\section{Empirical Study Results}
\label{sec:Empirical Study:Results}
This section presents and discusses the results obtained in our experiment. Subsection~\ref{sec:Empirical Study:Results:RQ1} presents the results regarding RQ1, which compares Sentinel vs. a Random hyper-heuristic generation. Subsection~\ref{sec:Empirical Study:Results:RQ2} presents the results regarding RQ2, which compares the strategies generated by Sentinel with three common strategies from the literature. Subsection~\ref{sec:Empirical Study:Results:RQ3} presents the results answering RQ3, which assesses the effectiveness of Sentinel's strategies over time. Finally, in Section~\ref{sec:Empirical Study:Results:Discussion} we discuss some final observations about the study.
\subsection{RQ1 -- Sentinel vs. Random Hyper-Heuristic}
\label{sec:Empirical Study:Results:RQ1}
Table~\ref{tab:Indicator Results RQ1} shows the mean HV and IGD results (standard deviation in brackets) obtained comparing Sentinel vs. Random hyper-heuristic, together with the corresponding effect sizes. The best indicator values and \textit{p-values} lower than $0.05$ are highlighted in bold. For the HV indicator, an effect size value closer to $1$ is favourable to Sentinel (i.e.\ greater HV values more often), whereas an effect size value closer to $0$ for IGD is better for Sentinel (i.e.\ lower IGD values more often).
\begin{table}[tb]
\centering
\caption{\textbf{RQ1}: HV and IGD mean results of Sentinel vs. Random Heuristic (standard deviation in brackets). Greater HV values and lower IGD values are better (best results in bold).
}
\label{tab:Indicator Results RQ1}
\fontsize{6pt}{6pt}\selectfont
\def1.2{1.2}
\begin{tabular}{llrrrr}
\toprule
Program & Ind. & Sentinel & Random & \textit{p-value} & Effect size \\ \midrule
beanutils-1.8.0 & HV & \textbf{0.87 (0.001)} & 0.83 (0.004) & \textbf{2.87E-11} & \textbf{1 (L)} \\
& IGD & \textbf{1.98E-4 (6.59E-6)} & 3.24E-4 (3.58E-5) & \textbf{2.87E-11} & \textbf{0 (L)} \\ \hline
codec-1.4 & HV & \textbf{0.91 (0.001)} & 0.89 (0.002) & \textbf{2.87E-11} & \textbf{1 (L)} \\
& IGD & \textbf{2.58E-4 (8.78E-5)} & 3.13E-4 (4.89E-5) & \textbf{0.003} & 0.29 (M) \\ \hline
collections-3.0 & HV & \textbf{0.83 (0.004)} & 0.81 (0.003) & \textbf{2.87E-11} & \textbf{1 (L)} \\
& IGD & \textbf{2.32E-4 (4.05E-5)} & 3.31E-4 (3.78E-5) & \textbf{1.48E-9} & \textbf{0.05 (L)} \\ \hline
lang-3.0 & HV & \textbf{0.88 (0.002)} & 0.86 (0.003) & \textbf{2.87E-11} & \textbf{1 (L)} \\
& IGD & \textbf{1.83E-4 (1.49E-5)} & 3.08E-4 (3.58E-5) & \textbf{2.87E-11} & \textbf{0 (L)} \\ \hline
validator-1.4.0 & HV & \textbf{0.90 (0.002)} & 0.87 (0.005) & \textbf{2.87E-11} & \textbf{1 (L)} \\
& IGD & \textbf{2.11E-4 (1.37E-5)} & 4.12E-4 (3.60E-5) & \textbf{2.87E-11} & \textbf{0 (L)} \\ \hline
jfreechart-1.0.0 & HV & \textbf{0.92 (0.002)} & 0.89 (0.01) & \textbf{2.87E-11} & \textbf{1 (L)} \\
& IGD & \textbf{2.36E-4 (4.04E-5)} & 5.48E-4 (9.08E-5) & \textbf{3.17E-11} & \textbf{0.001 (L)} \\ \hline
jgrapht-0.9.0 & HV & \textbf{0.93 (0.001)} & 0.90 (0.05) & \textbf{2.87E-11} & \textbf{1 (L)} \\
& IGD & \textbf{2.78E-4 (7.15E-5)} & 4.96E-4 (5.47E-5) & \textbf{1.15E-10} & \textbf{0.02 (L)} \\ \hline
joda-time-2.8 & HV & \textbf{0.85 (0.002)} & 0.82 (0.004) & \textbf{2.87E-11} & \textbf{1 (L)} \\
& IGD & \textbf{1.88E-4 (1.64E-5)} & 3.62E-4 (3.46E-5) & \textbf{2.87E-11} & \textbf{0 (L)} \\ \hline
ognl-3.1 & HV & \textbf{0.97 (0.002)} & 0.95 (0.003) & \textbf{2.87E-11} & \textbf{1 (L)} \\
& IGD & \textbf{5.15E-4 (1.64E-4)} & 6.18E-4 (1.09E-4) & \textbf{0.004} & 0.29 (M) \\ \hline
wire-2.0.0 & HV & \textbf{0.92 (0.004)} & 0.90 (0.004) & \textbf{2.87E-11} & \textbf{1 (L)} \\
& IGD & \textbf{4.89E-4 (1.39E-4)} & 4.89E-4 (5.99E-5) & 0.35 & 0.43 (N) \\ \bottomrule
\end{tabular}
\end{table}
We can observe that Sentinel is able to obtain better results than the Random hyper-heuristic for all programs and for both quality indicators with the only exception of IGD for {\it wire-2.0.0} where the results are similar. Moreover, the Kruskal--Wallis statistical test confirms statistical difference with 95\% confidence (\textit{p-value} $ < 0.05$) for all comparisons and the Vargha--Delaney $\hat{A}_{\mathit{12}}$ effect size shows large differences in 17 out of 20 comparisons in favour of Sentinel, with only two medium and one negligible differences for the IGD comparisons. As mentioned, the only equivalence in these results is for IGD for the smallest system \textit{wire-2.0.0}, where Kruskal--Wallis yielded no difference with a negligible Vargha--Delaney $\hat{A}_{\mathit{12}}$ effect size.
These results show that Sentinel is capable of generating more effective strategies than a Random hyper-heuristic. Therefore, we can positively answer our first research question:
~
\noindent
\fbox{\begin{minipage}{0.98 \columnwidth}
\textbf{RQ1}: \textit{The strategies generated by Sentinel are better than the strategies generated by a random hyper-heuristic.}
\end{minipage}
}
\subsection{RQ2 -- Sentinel's Strategies vs. Conventional Strategies}
\label{sec:Empirical Study:Results:RQ2}
As discussed in Section~\ref{sec:Sentinel:Multi-Objective Optimisation}, evaluating each objective (\textit{SCORE} and \textit{TIME}) separately is not meaningful, because in a multi-objective problem, the trade-off between the objectives is what really depicts the results quality. Evaluating each objective separately can be misleading, because a strategy that is good in one objective can be significantly worse in the other\footnote{For completeness, we report the average approaching mutation score and relative CPU time for Sentinel strategies, and the conventional strategies in Table~\ref{tab:TimeandScore}. However, care should be taken when interpreting these results separately, because, as explained in Section~\ref{sec:Sentinel:Multi-Objective Optimisation}, both objectives are equally important.}. What an engineer should do in this situation is to look at the non-dominated solutions in the objective space and select the one that best fits their needs. Figure~\ref{fig:Pareto} depicts the Pareto fronts of the non-dominated strategies tested in 30 independent runs for \textit{joda-time-2.8} and \textit{jfreechart-1.0.0} (two of the biggest systems in our experiment) to exemplify this trade-off.
\begin{figure*}[tb]
\begin{subfigure}{0.5\linewidth}
\centering
\includegraphics[width=0.8\linewidth]{images/jfreechart}
\caption{{\it jfreechart-1.0.0}}
\end{subfigure}
\begin{subfigure}{0.5\linewidth}
\centering
\includegraphics[width=0.8\linewidth]{images/joda-time}
\caption{{\it joda-time-2.8}}
\end{subfigure}
\caption{Pareto fronts comparing the Mutation Score Approaching (y-axis) vs. CPU time (x-axis) of the strategies obtained by Sentinel and the conventional strategies RMS, SM, and ROS for {\it jfreechart-1.0.0} and {\it joda-time-2.8}. The global mutation score is maintained as the original during the score approaching computation, hence a score of 1.0 in the y-axis means that the strategy obtained the same mutation score as when using the whole set of mutants and test cases.}
\label{fig:Pareto}
\end{figure*}
We can observe that Sentinel strategies are more scattered along the Pareto curve and are generally closer to the top left corner of the chart, which indicates a better trade-off between the objectives and generally more non-dominated solutions\footnote{The resulting fronts for the other 38 systems are similar to this, but for space reasons we omitted them from the paper. They are available at Sentinel's website~\cite{SentinelWebsite}.}. Even though SM strategies have a lower average execution time (concentrated on the left half of the plots) and RMS have a greater mutation score approaching (top half of the plots), they are usually dominated by strategies generated by Sentinel. In other words, when using a strategy generated by Sentinel, there is a greater chance of this strategy yielding a lower execution time for the same or better mutation score approaching than conventional strategies, or a better mutation score approaching for the same or better CPU time. For instance, if we consider a minimum mutation score approaching of 0.9 when selecting a strategy for \textit{jfreechart}, the cheapest Sentinel strategy would yield a reduction of approximately 93.5\% in execution time, RMS would yield $\approx$88\%, ROS $\approx$87\%, and SM $\approx$77\%.
We can quantify this trade-off and objectively evaluate the results considering both objectives (CPU time and score approaching) at the same time by computing the HV and IGD indicators (defined in Section~\ref{sec:Sentinel:Multi-Objective Optimisation}). Using such quality indicators for comparing sets of strategies is crucial, because if we report the optimisation of only one objective, then it would not say much about the trade-off with the other. By following the standard procedure of multi-objective analysis of reporting quality indicator values, we can assess the overall performance of the algorithms. In our case, the greater the HV of a resulting set of strategies, the greater the probability the strategies in this set have in outperforming the strategies in other sets with respect to the trade-off of score and time. Moreover, a set of strategies with an IGD value close to $0$ contains strategies that are very close to the best ones found for the program.
We report the HV and IGD results in Tables~\ref{tab:HV Results} and~\ref{tab:IGD Results} respectively. The standard deviation is shown in parenthesis. The best values and \textit{p-values} lower than $0.05$ are highlighted in bold. If two or more indicator values are highlighted in bold in the same row, it means that the statistical test showed no difference between the best strategy and the other highlighted strategies for a given system.
\begin{table}[tb]
\centering
\caption{\textbf{RQ2--RQ3:} HV results of Sentinel-generated strategies vs. conventional strategies (standard deviation in brackets). Greater HV values are better (best in bold).}
\label{tab:HV Results}
\fontsize{6pt}{6pt}\selectfont
\def1.2{1.2}
\begin{tabular}{lrrrrr}
\toprule
Program & Sentinel & SM & RMS & ROS & p-value \\ \midrule
beanutils-1.8.0 & \textbf{0.81 (0.02)} & 0.79 (0.02) & 0.76 (0.03) & 0.77 (0.04) & \textbf{3.3E-10} \\
beanutils-1.8.1 & \textbf{0.79 (0.01)} & \textbf{0.79 (0.03)} & 0.73 (0.02) & 0.75 (0.04) & \textbf{1.6E-11} \\
beanutils-1.8.2 & \textbf{0.79 (0.02)} & \textbf{0.78 (0.03)} & 0.74 (0.03) & 0.76 (0.04) & \textbf{1.9E-09} \\
beanutils-1.8.3 & \textbf{0.78 (0.03)} & \textbf{0.78 (0.03)} & 0.73 (0.03) & 0.75 (0.04) & \textbf{1.9E-10} \\ \hline
codec-1.4 & \textbf{0.89 (0.01)} & 0.87 (0.02) & 0.78 (0.005) & 0.84 (0.02) & \textbf{$<$ 2.2E-16} \\
codec-1.5 & \textbf{0.90 (0.005)} & \textbf{0.89 (0.01)} & 0.80 (0.005) & 0.84 (0.02) & \textbf{$<$ 2.2E-16} \\
codec-1.6 & \textbf{0.88 (0.003)} & \textbf{0.89 (0.02)} & 0.78 (0.004) & 0.82 (0.02) & \textbf{$<$ 2.2E-16} \\
codec-1.11 & \textbf{0.91 (0.01)} & \textbf{0.89 (0.01)} & 0.83 (0.01) & 0.83 (0.03) & \textbf{$<$ 2.2E-16} \\\hline
collections-3.0 & \textbf{0.92 (0.01)} & 0.74 (0.01) & 0.86 (0.01) & 0.79 (0.03) & \textbf{$<$ 2.2E-16} \\
collections-3.1 & \textbf{0.89 (0.01)} & 0.72 (0.02) & 0.84 (0.01) & 0.77 (0.02) & \textbf{$<$ 2.2E-16} \\
collections-3.2 & \textbf{0.90 (0.01)} & 0.73 (0.02) & 0.84 (0.01) & 0.77 (0.02) & \textbf{$<$ 2.2E-16} \\
collections-3.2.1 & \textbf{0.90 (0.01)} & 0.73 (0.01) & 0.85 (0.01) & 0.77 (0.02) & \textbf{$<$ 2.2E-16} \\\hline
lang-3.0 & \textbf{0.95 (0.004)} & 0.88 (0.01) & 0.90 (0.01) & 0.85 (0.02) & \textbf{$<$ 2.2E-16} \\
lang-3.0.1 & \textbf{0.94 (0.005)} & 0.86 (0.01) & 0.90 (0.01) & 0.84 (0.02) & \textbf{$<$ 2.2E-16} \\
lang-3.1 & \textbf{0.94 (0.01)} & 0.86 (0.02) & 0.90 (0.01) & 0.84 (0.02) & \textbf{$<$ 2.2E-16} \\
lang-3.2 & \textbf{0.95 (0.01)} & 0.87 (0.02) & 0.91 (0.01) & 0.86 (0.02) & \textbf{$<$ 2.2E-16} \\\hline
validator-1.4.0 & \textbf{0.86 (0.004)} & 0.79 (0.01) & 0.75 (0.005) & 0.76 (0.02) & \textbf{$<$ 2.2E-16} \\
validator-1.4.1 & \textbf{0.87 (0.01)} & 0.80 (0.02) & 0.76 (0.01) & 0.75 (0.02) & \textbf{$<$ 2.2E-16} \\
validator-1.5.0 & \textbf{0.86 (0.004)} & 0.77 (0.01) & 0.74 (0.01) & 0.74 (0.03) & \textbf{$<$ 2.2E-16} \\
validator-1.5.1 & \textbf{0.85 (0.01)} & 0.76 (0.02) & 0.74 (0.01) & 0.74 (0.02) & \textbf{$<$ 2.2E-16} \\\hline
jfreechart-1.0.0 & \textbf{0.97 (0.001}) & 0.84 (0.001) & 0.94 (0.001) & 0.84 (0.02) & \textbf{$<$ 2.2E-16} \\
jfreechart-1.0.1 & \textbf{0.97 (0.003)} & 0.85 (0.01) & 0.94 (0.005) & 0.85 (0.02) & \textbf{$<$ 2.2E-16} \\
jfreechart-1.0.2 & \textbf{0.97 (0.004)} & 0.85 (0.01) & 0.94 (0.01) & 0.85 (0.02) & \textbf{$<$ 2.2E-16} \\
jfreechart-1.0.3 & \textbf{0.97 (0.003)} & 0.86 (0.01) & 0.94 (0.005) & 0.85 (0.02) & \textbf{$<$ 2.2E-16} \\\hline
jgrapht-0.9.0 & \textbf{0.91 (0.01)} & 0.86 (0.01) & 0.86 (0.01) & 0.84 (0.02) & \textbf{$<$ 2.2E-16} \\
jgrapht-0.9.1 & \textbf{0.91 (0.01)} & 0.87 (0.01) & 0.85 (0.01) & 0.83 (0.02) & \textbf{$<$ 2.2E-16} \\
jgrapht-0.9.2 & \textbf{0.90 (0.004)} & 0.87 (0.01) & 0.85 (0.01) & 0.87 (0.02) & \textbf{$<$ 2.2E-16} \\
jgrapht-1.0.0 & \textbf{0.89 (0.005)} & 0.87 (0.02) & 0.85 (0.01) & 0.87 (0.01) & \textbf{$<$ 2.2E-16} \\\hline
joda-time-2.8 & \textbf{0.91 (0.01)} & 0.72 (0.01) & 0.83 (0.01) & 0.80 (0.02) & \textbf{$<$ 2.2E-16} \\
joda-time-2.8.1 & \textbf{0.90 (0.005)} & 0.71 (0.01) & 0.82 (0.01) & 0.79 (0.02) & \textbf{$<$ 2.2E-16} \\
joda-time-2.8.2 & \textbf{0.90 (0.005)} & 0.71 (0.01) & 0.81 (0.01) & 0.78 (0.02) & \textbf{$<$ 2.2E-16} \\
joda-time-2.9 & \textbf{0.90 (0.01)} & 0.70 (0.01) & 0.81 (0.01) & 0.78 (0.02) & \textbf{$<$ 2.2E-16} \\\hline
ognl-3.1 & \textbf{0.99 (0.003)} & 0.98 (0.003) & 0.97 (0.01) & 0.92 (0.02) & \textbf{$<$ 2.2E-16} \\
ognl-3.1.1 & \textbf{0.99 (0.003)} & 0.97 (0.01) & 0.92 (0.01) & 0.92 (0.03) & \textbf{$<$ 2.2E-16} \\
ognl-3.1.2 & \textbf{0.99 (0.001)} & 0.97 (0.01) & 0.97 (0.002) & 0.92 (0.02) & \textbf{$<$ 2.2E-16} \\
ognl-3.1.3 & \textbf{0.98 (0.001)} & 0.96 (0.01) & 0.91 (0.002) & 0.91 (0.02) & \textbf{$<$ 2.2E-16} \\\hline
wire-2.0.0 & 0.80 (0.01) & \textbf{0.83 (0.01)} & 0.62 (0.01) & 0.80 (0.01) & \textbf{$<$ 2.2E-16} \\
wire-2.0.1 & 0.84 (0.02) & \textbf{0.87 (0.01)} & 0.65 (0.02) & 0.84 (0.02) & \textbf{$<$ 2.2E-16} \\
wire-2.0.2 & 0.83 (0.01) & \textbf{0.86 (0.01)} & 0.64 (0.01) & 0.83 (0.02) & \textbf{$<$ 2.2E-16} \\
wire-2.0.3 & 0.66 (0.01) & \textbf{0.74 (0.05)} & 0.51 (0.01) & 0.67 (0.01) & \textbf{$<$ 2.2E-16} \\ \bottomrule
\end{tabular}
\end{table}
\begin{table*}[tb]
\centering
\caption{\textbf{RQ2--RQ3:} IGD results of Sentinel-generated strategies vs. conventional strategies (standard deviation in brackets). Lower IGD values are better (best values in bold).}
\label{tab:IGD Results}
\fontsize{6pt}{6pt}\selectfont
\def1.2{1.2}
\newcolumntype{R}[1]{>{\raggedleft\arraybackslash}p{#1}}
\begin{tabular}{lrrrrr}
\toprule
Program & Sentinel & SM & RMS & ROS & p-value \\ \midrule
beanutils-1.8.0 & \textbf{1.61E-4 (6.32E-6)} & 2.05E-3 (2.45E-4) & 2.17E-3 (8.80E-5) & 2.02E-3 (5.73E-4) & \textbf{2.1E-15} \\
beanutils-1.8.1 & \textbf{1.71E-4 (7.66E-6)} & 2.25E-3 (2.86E-4) & 2.20E-3 (6.13E-5) & 2.10E-3 (5.49E-4) & \textbf{1.2E-14} \\
beanutils-1.8.2 & \textbf{1.94E-4 (1.51E-5)} & 2.24E-3 (3.25E-4) & 2.20E-3 (1.00E-4) & 2.20E-3 (4.81E-4) & \textbf{1.8E-14} \\
beanutils-1.8.3 & \textbf{1.89E-4 (1.45E-5)} & 2.22E-3 (3.05E-4) & 2.18E-3 (7.09E-5) & 2.19E-3 (5.11E-4) & \textbf{1.2E-14} \\\hline
codec-1.4 & \textbf{2.03E-4 (9.83E-6)} & 2.57E-3 (1.57E-4) & 2.95E-3 (8.15E-5) & 2.94E-3 (3.48E-4) & \textbf{$<$ 2.2E-16} \\
codec-1.5 & \textbf{1.70E-4 (6.53E-6)} & 2.78E-3 (7.13E-5) & 3.40E-3 (8.30E-5) & 3.44E-3 (3.17E-4) & \textbf{$<$ 2.2E-16} \\
codec-1.6 & \textbf{1.65E-4 (3.44E-6)} & 2.67E-3 (1.04E-4) & 3.49E-3 (6.22E-5) & 3.55E-3 (4.64E-4) & \textbf{$<$ 2.2E-16} \\
codec-1.11 & \textbf{1.57E-4 (6.48E-6)} & 2.19E-3 (4.29E-5) & 3.15E-3 (4.93E-5) & 3.03E-3 (3.93E-4) & \textbf{$<$ 2.2E-16} \\\hline
collections-3.0 & \textbf{1.73E-4 (1.45E-5)} & 1.73E-3 (2.34E-5) & 2.31E-3 (4.86E-5) & 2.70E-3 (8.29E-4) & \textbf{$<$ 2.2E-16} \\
collections-3.1 & \textbf{1.80E-4 (1.43E-5)} & 1.64E-3 (1.56E-5) & 2.25E-3 (5.11E-5) & 1.85E-3 (4.06E-4) & \textbf{$<$ 2.2E-16} \\
collections-3.2 & \textbf{1.91E-4 (1.67E-5)} & 1.56E-3 (2.51E-5) & 2.11E-3 (6.53E-5) & 1.66E-3 (5.41E-4) & \textbf{$<$ 2.2E-16} \\
collections-3.2.1 & \textbf{1.76E-4 (1.09E-5)} & 1.54E-3 (1.71E-5) & 2.11E-3 (7.23E-5) & 1.82E-3 (6.38E-4) & \textbf{$<$ 2.2E-16} \\\hline
lang-3.0 & \textbf{9.67E-5 (7.90E-6)} & 1.74E-3 (2.50E-5) & 2.41E-3 (3.57E-5) & 1.93E-3 (5.54E-4) & \textbf{$<$ 2.2E-16} \\
lang-3.0.1 & \textbf{9.68E-5 (4.41E-6)} & 1.75E-3 (1.54E-5) & 2.40E-3 (2.82E-5) & 1.80E-3 (5.08E-4) & \textbf{$<$ 2.2E-16} \\
lang-3.1 & \textbf{1.43E-4 (1.48E-5)} & 1.84E-3 (3.51E-5) & 2.39E-3 (3.35E-5) & 1.72E-3 (4.25E-4) & \textbf{$<$ 2.2E-16} \\
lang-3.2 & \textbf{1.78E-4 (3.33E-5)} & 1.85E-3 (3.45E-5) & 2.41E-3 (2.95E-5) & 1.84E-3 (4.46E-4) & \textbf{$<$ 2.2E-16} \\\hline
validator-1.4.0 & \textbf{1.72E-4 (6.38E-6)} & 1.71E-3 (7.04E-5) & 3.19E-3 (8.53E-5) & 2.37E-3 (5.94E-4) & \textbf{$<$ 2.2E-16} \\
validator-1.4.1 & \textbf{1.85E-4 (9.46E-6)} & 1.77E-3 (9.72E-5) & 3.13E-3 (1.04E-4) & 2.75E-3 (5.54E-4) & \textbf{$<$ 2.2E-16} \\
validator-1.5.0 & \textbf{1.83E-4 (5.53E-6)} & 1.82E-3 (9.32E-5) & 3.04E-3 (8.97E-5) & 2.55E-3 (8.34E-4) & \textbf{$<$ 2.2E-16} \\
validator-1.5.1 & \textbf{1.90E-4 (1.10E-5)} & 1.79E-3 (9.93E-5) & 3.07E-3 (8.16E-5) & 2.68E-3 (7.56E-4) & \textbf{$<$ 2.2E-16} \\\hline
jfreechart-1.0.0 & \textbf{1.60E-4 (1.52E-6)} & 1.74E-3 (1.96E-5) & 2.72E-3 (4.59E-5) & 1.63E-3 (2.44E-4) & \textbf{$<$ 2.2E-16} \\
jfreechart-1.0.1 & \textbf{1.50E-4 (6.29E-6)} & 1.56E-3 (3.65E-5) & 2.74E-3 (4.37E-5) & 1.54E-3 (2.98E-4) & \textbf{$<$ 2.2E-16} \\
jfreechart-1.0.2 & \textbf{1.68E-4 (7.61E-6)} & 1.58E-3 (4.08E-5) & 2.76E-3 (3.62E-5) & 1.61E-3 (3.63E-4) & \textbf{$<$ 2.2E-16} \\
jfreechart-1.0.3 & \textbf{1.51E-4 (3.52E-6)} & 1.53E-3 (3.32E-5) & 2.74E-3 (4.91E-5) & 1.65E-3 (4.29E-4) & \textbf{$<$ 2.2E-16} \\\hline
jgrapht-0.9.0 & \textbf{2.23E-4 (1.70E-5)} & 1.58E-3 (5.61E-5) & 3.72E-3 (7.56E-5) & 2.43E-3 (6.82E-4) & \textbf{$<$ 2.2E-16} \\
jgrapht-0.9.1 & \textbf{2.15E-4 (1.44E-5)} & 1.86E-3 (6.78E-5) & 4.06E-3 (8.42E-5) & 2.64E-3 (7.86E-4) & \textbf{$<$ 2.2E-16} \\
jgrapht-0.9.2 & \textbf{2.85E-4 (1.56E-5)} & 1.72E-3 (4.53E-5) & 3.72E-3 (5.88E-5) & 2.25E-3 (5.48E-4) & \textbf{$<$ 2.2E-16} \\
jgrapht-1.0.0 & \textbf{2.27E-4 (1.78E-5)} & 1.68E-3 (4.42E-5) & 3.65E-3 (7.08E-5) & 2.29E-3 (6.34E-4) & \textbf{$<$ 2.2E-16} \\\hline
joda-time-2.8 & \textbf{1.26E-4 (9.44E-6)} & 1.66E-3 (2.72E-5) & 1.66E-3 (2.43E-5) & 1.25E-3 (2.44E-4) & \textbf{$<$ 2.2E-16} \\
joda-time-2.8.1 & \textbf{1.11E-4 (5.59E-6)} & 1.70E-3 (2.36E-5) & 1.66E-3 (3.10E-5) & 1.30E-3 (3.02E-4) & \textbf{$<$ 2.2E-16} \\
joda-time-2.8.2 & \textbf{1.13E-4 (8.08E-6)} & 1.71E-3 (2.79E-5) & 1.65E-3 (3.35E-5) & 1.33E-3 (3.46E-4) & \textbf{$<$ 2.2E-16} \\
joda-time-2.9 & \textbf{1.25E-4 (6.08E-6)} & 1.68E-3 (2.02E-5) & 1.67E-3 (2.77E-5) & 1.36E-3 (3.47E-4) & \textbf{$<$ 2.2E-16} \\\hline
ognl-3.1 & \textbf{9.35E-4 (4.91E-5)} & 5.87E-3 (1.12E-4) & 6.74E-3 (1.06E-4) & 5.19E-3 (9.84E-4) & \textbf{$<$ 2.2E-16} \\
ognl-3.1.1 & \textbf{9.22E-4 (5.10E-5)} & 5.75E-3 (7.91E-5) & 6.53E-3 (8.59E-5) & 4.56E-3 (1.02E-3) & \textbf{$<$ 2.2E-16} \\
ognl-3.1.2 & \textbf{1.06E-3 (1.12E-5)} & 5.65E-3 (1.05E-4) & 6.49E-3 (1.11E-4) & 4.51E-3 (1.08E-3) & \textbf{$<$ 2.2E-16} \\
ognl-3.1.3 & \textbf{1.06E-3 (9.70E-6)} & 6.01E-3 (1.28E-4) & 6.82E-3 (7.98E-5) & 5.14E-3 (1.14E-3) & \textbf{$<$ 2.2E-16} \\\hline
wire-2.0.0 & \textbf{5.08E-4 (2.26E-5)} & 3.13E-3 (3.11E-4) & 5.77E-3 (2.82E-4) & 4.13E-3 (1.02E-3) & \textbf{$<$ 2.2E-16} \\
wire-2.0.1 & \textbf{4.68E-4 (2.34E-5)} & 3.18E-3 (2.38E-4) & 5.69E-3 (3.37E-4) & 3.99E-3 (1.07E-3) & \textbf{$<$ 2.2E-16} \\
wire-2.0.2 & \textbf{4.61E-4 (1.48E-5)} & 3.05E-3 (1.55E-4) & 5.79E-3 (3.66E-4) & 4.09E-3 (1.19E-3) & \textbf{$<$ 2.2E-16} \\
wire-2.0.3 & \textbf{4.68E-4 (1.83E-5)} & 3.03E-3 (3.49E-4) & 5.62E-3 (2.70E-4) & 3.76E-3 (1.02E-3) & \textbf{$<$ 2.2E-16} \\ \bottomrule
\end{tabular}
\end{table*}
We can observe that, overall, the strategies generated by Sentinel are able to outperform the conventional ones according to both indicators. More precisely, for 70 out of 80 statistical comparisons, Sentinel strategies outperformed the conventional ones with significant differences. Only for {\it wire} SM obtained significantly better HV results, but not for IGD. For \textit{codec-1.5/1.6/1.11} and \textit{beanutils-1.8.1/1.8.2/1.8.3}, SM showed no statistical significant difference regarding HV, but fell behind in the IGD comparison.
Table~\ref{tab:Effect Size} shows the Vargha--Delaney $\hat{A}_{\mathit{12}}$ effect size results for HV and IGD. The group of strategies generated by Sentinel is subject A and each conventional strategy group -- shown in columns from two to seven -- is subject B. Large differences in favour of Sentinel strategies are highlighted in bold.
The strategies generated by Sentinel obtained large effect size in 227 out of 240 comparisons (i.e.\ 95\% of the cases), while SM and ROS obtained favourable large effect size only in one system: SM for all versions of {\it wire} and ROS for {\it wire-2.0.3}.
\begin{table}[tb]
\centering
\caption{\textbf{RQ2--RQ3:} Effect size results (L = large, M=medium, S=small, N = negligible) of Sentinel-generated strategies vs. conventional strategies. Values greater than $0.5$ are better for Sentinel (best in bold).}
\label{tab:Effect Size}
\fontsize{6pt}{6pt}\selectfont
\def1.2{1.2}
\newcolumntype{R}[1]{>{\raggedleft\arraybackslash}p{#1}}
\begin{tabular}{lrrrrrr}
\toprule
\multirow{2}{*}{Program} & \multicolumn{3}{c|}{HV} & \multicolumn{3}{c}{IGD} \\
& SM & RMS & \multicolumn{1}{r|}{ROS} & SM & RMS & ROS \\ \midrule
beanutils-1.8.0 & \textbf{0.78 (L)} & \textbf{0.94 (L)} & \textbf{0.84 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
beanutils-1.8.1 & 0.63 (S) & \textbf{0.96 (L)} & \textbf{0.80 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
beanutils-1.8.2 & 0.70 (M) & \textbf{0.88 (L)} & \textbf{0.77 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
beanutils-1.8.3 & 0.64 (S) & \textbf{0.91 (L)} & \textbf{0.80 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\\hline
codec-1.4 & \textbf{0.88 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
codec-1.5 & 0.69 (M) & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
codec-1.6 & 0.45 (N) & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
codec-1.11 & \textbf{0.88 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\\hline
collections-3.0 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
collections-3.1 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
collections-3.2 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
collections-3.2.1 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\\hline
lang-3.0 & \textbf{1 (L)} & \textbf{0.99 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
lang-3.0.1 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
lang-3.1 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
lang-3.2 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\\hline
validator-1.4.0 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
validator-1.4.1 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
validator-1.5.0 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
validator-1.5.1 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\\hline
jfreechart-1.0.0 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
jfreechart-1.0.1 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
jfreechart-1.0.2 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
jfreechart-1.0.3 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\\hline
jgrapht-0.9.0 & \textbf{0.99 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
jgrapht-0.9.1 & \textbf{0.97 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
jgrapht-0.9.2 & \textbf{0.95 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
jgrapht-1.0.0 & \textbf{0.92 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\\hline
joda-time-2.8 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
joda-time-2.8.1 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
joda-time-2.8.2 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
joda-time-2.9 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\\hline
ognl-3.1 & \textbf{0.99 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
ognl-3.1.1 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
ognl-3.1.2 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
ognl-3.1.3 & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{1 (L)} & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\\hline
wire-2.0.0 & 0.03 (L) & \textbf{1 (L)} & 0.57 (N) & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
wire-2.0.1 & 0.04 (L) & \textbf{1 (L)} & 0.40 (S) & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
wire-2.0.2 & 0.05 (L) & \textbf{1 (L)} & 0.45 (N) & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\
wire-2.0.3 & 0 (L) & \textbf{1 (L)} & 0.19 (L) & \textbf{0 (L)} & \textbf{0 (L)} & \textbf{0 (L)} \\ \bottomrule
\end{tabular}
\end{table}
Therefore, we can conclude that the strategies generated by Sentinel perform statistically significantly better than the conventional strategies for the majority of the systems tested (i.e.\ 95\% of the cases). This positively answers our research question:
~
\noindent
\fbox{\begin{minipage}{0.98 \columnwidth}
\textbf{RQ2}: \textit{The strategies generated by Sentinel present a significantly better trade-off between execution time and score approaching than the conventional strategies RMS, ROS and SM.}
\end{minipage}
}
\subsection{RQ3 -- Sentinel Strategies Effectiveness Over Time}
\label{sec:Empirical Study:Results:RQ3}
Training any automated or manual approach to solve a problem comes with an additional cost. Such a cost is alleviated if the trained model (in our case, the strategies found) can be re-used over time to solve subsequent unseen instances of the problem. In this section, we discuss and compare the training time of Sentinel and conventional strategies, and then assess if the strategies learnt at a certain time (i.e.\ for a given version of a system) can be re-used later on for subsequent versions without the need of generating new ones (and hence to re-train Sentinel incurring in additional training costs).
Table~\ref{tab:Training Times} shows the average training time for Sentinel and conventional strategies (i.e.\ the cost of manually comparing all RMS, ROS and SM strategies in order to choose the best one). We can observe that, for six out of 10 systems, the training cost of Sentinel is lower than experimenting and comparing RMS, ROS and SM due to the caching system used\footnote{The caching mechanism is described in Section~\ref{sec:Sentinel:Implementation Aspects}. For some of the systems caching is faster than executing all conventional strategies, but for others this incurs in a costlier overhead for setting up one JVM instance for each mutant.}, yet the training cost of all these approaches is not negligible.
\begin{table}[tb]
\centering
\caption{\textbf{RQ3:} Average training time of Sentinel and conventional strategies. Time format ``hh:mm:ss''.}
\label{tab:Training Times}
\fontsize{8pt}{8pt}\selectfont
\def1.2{1.2}
\begin{tabular}{lrr}
\toprule
Program & Sentinel & Conventional Strategies \\ \midrule
beanutils & 04:12:23 & 01:36:24 \\
codec & 01:22:11 & 01:56:23 \\
collections & 07:24:22 & 02:37:29 \\
lang & 11:23:49 & 04:23:09 \\
validator & 38:44 & 01:06:21 \\
jfreechart & 07:51:33 & 11:38:16 \\
jgrapht & 01:49:51 & 04:41:22 \\
joda-time & 03:25:33 & 06:54:45 \\
ognl & 01:19:44 & 51:27 \\
wire & 01:41 & 11:58 \\ \bottomrule
\end{tabular}
\end{table}
If we had to train for each version of a software, the training cost would make their use ineffective. Fortunately, the results reported in Section~\ref{sec:Empirical Study:Results:RQ2} show that reusing strategies over time is possible without significant loss in their effectiveness. As we can observe from Tables~\ref{tab:HV Results}-\ref{tab:Effect Size}, not only the strategies generated by Sentinel are able to hold their good results across multiple versions, but they also outperform conventional strategies. For seven out of the 10 analysed software programs, the strategies generated by Sentinel showed favourable large statistical differences for all the subsequent versions considering both IGD and HV. For two of those 10 programs, the strategies generated by Sentinel showed no statistical differences to SM strategies for HV in subsequent versions, and outperformed ROS and RMS with large statistical differences. Looking at the IGD results, Sentinel strategies were able to maintain large statistical differences throughout all the evaluated versions for all the 10 systems. The systems \textit{beanutils} and \textit{codec} are the only two for which Sentinel's generated strategies did not maintain their superiority considering HV values for subsequent versions. The code churn of \textit{beanutils} is one of the smallest between the programs (Table~\ref{tab:Programs}), whereas \textit{codec}'s churn is somewhat close to the median. For systems in between those churn sizes (or in the extremities for that matter), Sentinel's strategies statistically outperformed the conventional ones in subsequent versions. Hence, we did not observe any relation between code churn and effectiveness degradation of the trained strategies.
Summarising, for 95\% of comparisons, Sentinel strategies learnt on a previous version can be reused as they are statistically better or equivalent to conventional ones when used for subsequent versions of a same software. Therefore, we can state that, overall, Sentinel strategies can be trained on a given system's version and then reused up to three subsequent versions without either affecting their mutant reduction effectiveness, or incurring in additional training costs. This answers positively our third and last research question:
~
\noindent
\fbox{\begin{minipage}{0.98 \columnwidth}
\textbf{RQ3}: \textit{The strategies generated by Sentinel can be effectively used for subsequent versions of a software.}
\end{minipage}
}
\subsection{Discussion}
\label{sec:Empirical Study:Results:Discussion}
The results presented in this section reveal that arbitrarily selecting and configuring a strategy may lead to unsatisfactory time reduction, mutation score degradation or, in the worst case, both. For instance, \textit{joda-time-2.8} has approximately 10,000 mutants, thus RMS 10\% gives us 1,000 mutants. The average result for RMS 10\% (lower cluster of black squares in Figure~\ref{fig:Pareto}) is approximately 15\% time cost and a mutation score approaching of only 55\% on average. For \textit{jfreechart-1.0.0}, there are approximately 20,000 mutants. The lower cluster of black squares of Figure~\ref{fig:Pareto}(a) represents RMS 10\%, which selects 2,000 mutants. The average result for this strategy is around 5\% CPU time cost but approaching only 75\% of the original mutation score. In other words, the same percentage of mutants in different systems yields different results in both mutation score and CPU time: while 1,000 mutants in \textit{joda-time-2.8} yields 15\% of time and 55\% of score approaching, 2,000 mutants in \textit{jfreechart-1.0.0} yields only 5\% of time and 75\% score. We also observed such a discrepancy for the other systems. With these examples in mind and based on what we observed during our experiments, we can state that choosing the right parameters and strategies is crucial for effectively reducing the mutant set and that mutant reduction strategies should be carefully tuned for each software system.
These results show that there is room for improvement of mutant reduction strategies and doing it automatically seems preferable in most cases. As far as we could observe, the strategies generated by Sentinel obtain the best trade-off between mutation score approaching and execution time, and generally maintain this best performance over subsequent software versions. Sentinel also removes from the hands of the engineer the tedious, error-prone and costly task of selecting and configuring strategies. We advocate that Sentinel should be used in two situations where the mutant reduction is more impactful: i)~when the time taken to execute all mutants is impracticable; and ii)~when the mutation testing is performed several times during the software development process. Indeed, how long is ``impracticable time'' and how many is ``several times'' depend on many factors such as environmental constraints, software testing budget and even on the required reliability level involved in the testing process. In any case, if the testers find themselves in at least one of those two situations, then the training cost of Sentinel can be a good price to pay for the significant mutant reduction trade-off provided by the generated strategies, specially considering that doing an exhaustive manual experimentation to find the best conventional strategy was more expensive than Sentinel training for six out of 10 systems (as seen in Section~\ref{sec:Empirical Study:Results:RQ3}). Moreover, Sentinel's training cost can be further reduced by running it in parallel~\cite{Sarro2020GECCO,GeronimoFMS12,FerrucciSalzaSarro,DiMartino2013,FerrucciSKS15,SalzaFS16,SalzaFS16a}.
Finally, based on our data, we observed that number of mutants is not always an accurate surrogate for cost. Thus, we advise engineers to measure execution time when comparing the cost of mutation testing for a better assessment~\cite{Guizzo2020}.
\section{Related Work}
\label{sec:Related Work}
In this section we discuss previous work on hyper-heuristics for mutation testing and on mutant reduction strategies. Comprehensive surveys on mutation testing can be found elsewhere~\cite{Jia2011, Bashir2012, Papadakis2017}. We also refer the reader to a recent systematic review on mutation cost reduction by Pizzoleto et al.~\cite{Pizzoleto2019} for a comprehensive description of cost reduction strategies used in the literature.
\subsection{Hyper-Heuristics for Mutation Testing}
\label{sec:Related Work:Hyper-Heuristics in Mutation Testing}
The use of hyper-heuristics for Software Engineering have been recently investigated providing promising results for project management~\cite{Sarro2017}, software clustering~\cite{Kumari2016}, combinatorial interaction testing~\cite{Jia2015}, integration and test order~\cite{Guizzo2015, Mariani2016, Guizzo2017} and feature models testing~\cite{Strickler2016}. Few studies have investigated hyper-heuristics for mutation testing~\cite{Ferreira2017, Strickler2016, Filho2017, Lima2017}. The work of Ferreira et al.~\cite{Ferreira2017} and Strickler et al.~\cite{Strickler2016} use the online Hyper-heuristic for the Integration and Test Order Problem (HITO)~\cite{Guizzo2015}. The idea is to automatically select mutation and crossover operators of MOEAs. Such hyper-heuristic is applied in a different problem than the one for which it was originally proposed~\cite{Guizzo2015}: the selection of products for testing Software Product Lines (SPLs). HITO achieved competitive results with respect to conventional MOEAs. Jakuboviski Filho et al.~\cite{Filho2017} also applied an offline hyper-heuristic based on GE (proposed by Mariani et al.~\cite{Mariani2016}) to automatically generate MOEAs for the same SPL problem. The generated MOEAs were compared to NSGA--II and to HITO~\cite{Guizzo2015} and showed competitive results with respect to HITO and generally better results than NSGA--II.
Such studies~\cite{Ferreira2017, Strickler2016, Filho2017} have as main goal the selection of test data (products) to satisfy the mutation testing of Feature Models. Sentinel has a distinct goal, since it is used to generate strategies to reduce the set of mutants. While Sentinel is an off-line generation hyper-heuristic for mutant reduction strategies, the hyper-heuristics used in the related work are on-line selection hyper-heuristics for adaptively selecting evolutionary operators~\cite{Ferreira2017, Strickler2016} or off-line generation of MOEAs~\cite{Filho2017}.
Lima and Vergilio~\cite{Lima2017} proposed a hyper-heuristic to automatically select crossover and mutation operators of a MOEA in order to generate HOMs. The goal of this work is different from ours, since the approach searches to find the best HOMs by optimising three objectives: i)~minimising the total number of generated HOMs; ii)~maximising the number of revealed subtle faults (i.e.\ faults that can only be revealed by the HOM); and iii)~maximising the number of strongly subsuming HOMs.
\subsection{Mutant Reduction Strategies}
\label{sec:Related Work:Mutants Reduction Strategies}
Mathur and Wong~\cite{Mathur1994} applied mutant sampling on C programs and were able to reduce the number of mutants by 90\% while losing only $0.16$ points in the mutation score. Wong et al.~\cite{Wong1995} used a similar strategy for the same testing scenario, by selecting only a percentage of mutants from each available mutation operator and showed that this strategy was equally efficient. Papadakis and Malevris~\cite{Papadakis2010} conducted a study on C programs by sampling 10\%, 20\%, 30\%, 40\%, 50\% and 60\% mutants and found that the mutation effectiveness is reduced by 26\%--6\% using such a strategy. Sahinoglu and Spafford~\cite{Sahinoglu1990} proposed a strategy based on the Bayesian Inference for sampling mutants, which was evaluated for the unit testing of C programs showing that a better mutant reduction was obtained with respect to a conventional strategy.
Hussain et al.~\cite{Hussain2008} proposed mutant clustering strategies which use clustering algorithms to select mutants from different clusters, because test cases that kill a mutant likely kill the mutants in the same cluster. The problem is that all the test cases must be executed against all mutants before applying the clustering algorithm, which is a very costly task. Another problem is how to define the number of centroids. Ji et al.~\cite{Ji2009} propose a clustering strategy that uses information about the program domain to perform the mutant clustering. The advantage of this strategy is the generation and selection of mutants before executing the test cases. The authors evaluated this strategy in the unit test of a single Java program and observed that the strategy is capable of maintaining the mutation score as high as $0.9$ while reducing in 50\% the number of mutants.
Mathur~\cite{Mathur1991} omitted the two most costly operators out of the 22 mutation operators available in the Mothra tool~\cite{King1991} used to mutate Fortran 77 programs. These two operators generate between 30\% and 40\% of all mutants. Results show a mutation score of $0.9999$ and a mutant reduction of 24\%. Offutt et al.~\cite{Offutt1993} extended the work of Mathur~\cite{Mathur1991} by omitting the four and six more costly operators. For the four operators exclusion, the mutant reduction was approximately 40\% with a $0.9984$ score, whereas for the six operator exclusion the obtained score was $0.8871$ with a 60\% reduction. Delamaro et al.~\cite{Delamaro2014} performed an empirical study using only one mutation operator for C programs. The hypothesis is that using only one powerful operator should be enough to perform the mutation testing. The mutation operator responsible for removing lines of code is the most efficient one, since it generates approximately 3.26\% of all mutants and obtains a score of approximately $0.92$.
Barbosa et al.~\cite{Barbosa2001} defined some guidelines for the selection of essential operators (a set with the best operators). The authors obtained a set of 10 essential operators among the 77 available in the Proteum tool~\cite{Delamaro2001}. This set was able to reduce by 65\% the number of mutants while maintaining a mutation score of $0.996$. Vincenzi et al.~\cite{Vincenzi1999} also used guidelines like these, but for the mutation in the integration testing phase (interface mutation). The results showed that it was possible to reduce by 73\% the number of generated mutants while keeping the mutation score as high as $0.998$. Namin et al.~\cite{Namin2008} proposed a statistical analysis procedure with a linear regression model to find essential operators. A set of 28 operators was found which generated only 8\% of all mutants and obtained an acceptable mutation score.
Previous work also compared different conventional strategies. Gopinath et al.~\cite{Gopinath2015} compared mutant sampling strategies that sample a constant number of mutants to strategies that sample a percentage, finding that strategies that use a constant number of mutants (as low as 1,000) can obtain an accurate mutation score (approximately 93\%) when compared to the whole set of mutants. Zhang et al.~\cite{Zhang2013} discovered that using 5\% random mutant sampling in combination with operator-based selection strategies can greatly reduce the number of mutants while maintaining a great mutation accuracy. Zhang et al.~\cite{Zhang2010} compared strategies that select mutation operators with mutant sampling strategies, showing that selecting operators is not superior to sampling random mutants. Gopinath et al.~\cite{Gopinath2016} observed similar results with a theoretical and empirical analysis and also concluded that the reduction limit over random mutant sampling strategies is 13\%. Lima et al.~\cite{Lima2016} compared SM, RMS, HOM generation strategies and an Evolutionary Algorithm for selecting mutants and test cases to reduce the mutation cost of C programs. The authors discovered that SM performed better than the other strategies and that the HOM strategies generated more mutants with a lower mutation score.
These works diverge on the conclusion about which strategy is the best, as shown in a recent literature review~\cite{Pizzoleto2019}. In fact, it depends on the context of the test, such as program under test, testing phase and programming language. Hence, using an automatic hyper-heuristic such as Sentinel can provide good strategies that are tailored for a specific scenario. Moreover, we observed that some mutants take longer to execute than others, e.g.\ mutants that die due to time out or mutants with a fault in memory allocation. By assessing the CPU execution time we can accurately measure the cost reduction of strategies, as opposed to using solely the number of mutants~\cite{Guizzo2020}.
Recently, Petrovic and Ivankovic~\cite{Petrovic2018} reported how mutation testing is used at Google at scale. Their approach deviates from the conventional mutation analysis by changing the way mutants are generated and presented to the developer. Their tool only generates one possible mutant per covered line. Traditional mutation tools usually also only apply mutations to covered lines, however they usually do not restrict the generation of mutants to one per line. As the authors state themselves, they do not use mutation score to measure the test effectiveness, since the mutation score relies on the whole set of possible mutants and they in fact only generate a limited set of mutants using a probabilistic technique. The second difference is that the focus of their approach is to maximise the ``usefulness'' of mutants by showing only interesting mutants to the developers during code review. If the developer judges an alive mutant as irrelevant, even if killable, the mutant is simply ignored. According to the developer's feedback, a heuristic is applied to prune the ``uninteresting'' lines (dubbed ``arid lines'') from the mutation process. Therefore, the mutant set depends on historical data obtained from the developers and on the heuristic to determine which mutants are uninteresting. In conventional mutation, all mutants are considered in order to improve the mutation score.
The third difference is on the scope of the mutation. While traditional mutation usually treats a program as a whole, their approach narrows down the mutation to diffs and commits. Those differences unveil a very interesting relation between current academic literature on mutation testing and the practice in industry. The challenges faced in each scenario might be different and should be considered in future work.
\section{Conclusion}
\label{sec:Conclusion}
This paper presented Sentinel, a hyper-heuristic based on GE to generate mutant reduction strategies. Our empirical results show that Sentinel performs statistically significantly better than a Random hyper-heuristic for all of the 10 used real-world systems with large effect size in 95\% of the comparisons. Moreover, when compared to state-of-the-art strategies, the strategies generated by Sentinel achieved statistically significantly better results for 70 out of 80 comparisons (88\%), and large effect size in 95\% of the cases. Also, we investigated if the mutation strategies generated with Sentinel on a given version can be used in subsequent versions of a same system revealing positive results in 95\% of the cases. These results show that Sentinel strategies are effective to the mutant reduction problem when considering both global mutation score approaching and CPU time minimisation.
\begin{table}[]
\centering
\caption{Average CPU time and approaching score per strategy over 30 independent runs (CPU time is given as the percentage of time used by a strategy to perform the mutation in relation to generating and executing all mutants).}
\fontsize{6pt}{6pt}\selectfont
\def1.2{1.2}
\label{tab:TimeandScore}
\begin{tabular}{lrrrrrrrr}
\toprule
\multirow{2}{*}{Program} & \multicolumn{4}{c|}{Score} & \multicolumn{4}{c}{Time} \\
& Sen. & SM & RMS & \multicolumn{1}{c|}{ROS} & Sen. & SM & RMS & ROS \\ \midrule
beanutils-1.8.0 & 0.63 & 0.48 & 0.89 & 0.82 & 0.43 & 0.27 & 0.54 & 0.49 \\
beanutils-1.8.1 & 0.63 & 0.48 & 0.89 & 0.81 & 0.46 & 0.28 & 0.56 & 0.50 \\
beanutils-1.8.2 & 0.63 & 0.48 & 0.88 & 0.82 & 0.46 & 0.28 & 0.56 & 0.52 \\
beanutils-1.8.3 & 0.63 & 0.48 & 0.88 & 0.82 & 0.46 & 0.28 & 0.55 & 0.50 \\ \hline
codec-1.4 & 0.68 & 0.80 & 0.93 & 0.90 & 0.34 & 0.27 & 0.48 & 0.41 \\
codec-1.5 & 0.65 & 0.83 & 0.93 & 0.90 & 0.31 & 0.26 & 0.47 & 0.40 \\
codec-1.6 & 0.64 & 0.82 & 0.93 & 0.90 & 0.32 & 0.27 & 0.48 & 0.42 \\
codec-1.11 & 0.67 & 0.81 & 0.93 & 0.89 & 0.25 & 0.24 & 0.43 & 0.39 \\ \hline
collections-3.0 & 0.55 & 0.56 & 0.85 & 0.81 & 0.16 & 0.14 & 0.33 & 0.35 \\
collections-3.1 & 0.54 & 0.54 & 0.85 & 0.77 & 0.16 & 0.14 & 0.33 & 0.32 \\
collections-3.2 & 0.56 & 0.53 & 0.84 & 0.76 & 0.17 & 0.14 & 0.34 & 0.33 \\
collections-3.2.1 & 0.56 & 0.53 & 0.84 & 0.76 & 0.16 & 0.14 & 0.33 & 0.32 \\ \hline
lang-3.0 & 0.59 & 0.69 & 0.90 & 0.83 & 0.13 & 0.12 & 0.31 & 0.29 \\
lang-3.0.1 & 0.59 & 0.69 & 0.90 & 0.83 & 0.12 & 0.12 & 0.29 & 0.29 \\
lang-3.1 & 0.59 & 0.68 & 0.90 & 0.82 & 0.15 & 0.15 & 0.38 & 0.35 \\
lang-3.2 & 0.59 & 0.69 & 0.90 & 0.82 & 0.15 & 0.16 & 0.36 & 0.35 \\ \hline
validator-1.4.0 & 0.63 & 0.59 & 0.93 & 0.84 & 0.35 & 0.28 & 0.52 & 0.48 \\
validator-1.4.1 & 0.64 & 0.61 & 0.93 & 0.86 & 0.35 & 0.28 & 0.53 & 0.50 \\
validator-1.5.0 & 0.63 & 0.60 & 0.92 & 0.84 & 0.36 & 0.30 & 0.56 & 0.51 \\
validator-1.5.1 & 0.62 & 0.60 & 0.92 & 0.85 & 0.36 & 0.30 & 0.55 & 0.51 \\ \hline
jfreechart-1.0.0 & 0.69 & 0.59 & 0.94 & 0.80 & 0.13 & 0.16 & 0.34 & 0.33 \\
jfreechart-1.0.1 & 0.69 & 0.60 & 0.94 & 0.79 & 0.11 & 0.13 & 0.27 & 0.25 \\
jfreechart-1.0.2 & 0.69 & 0.61 & 0.94 & 0.81 & 0.11 & 0.13 & 0.28 & 0.27 \\
jfreechart-1.0.3 & 0.68 & 0.60 & 0.94 & 0.81 & 0.11 & 0.13 & 0.27 & 0.26 \\ \hline
jgrapht-0.9.0 & 0.68 & 0.74 & 0.96 & 0.89 & 0.25 & 0.25 & 0.45 & 0.42 \\
jgrapht-0.9.1 & 0.65 & 0.74 & 0.96 & 0.88 & 0.26 & 0.26 & 0.46 & 0.42 \\
jgrapht-0.9.2 & 0.68 & 0.75 & 0.96 & 0.88 & 0.26 & 0.26 & 0.49 & 0.45 \\
jgrapht-1.0.0 & 0.66 & 0.74 & 0.95 & 0.87 & 0.25 & 0.25 & 0.46 & 0.42 \\ \hline
joda-time-2.8 & 0.65 & 0.50 & 0.86 & 0.76 & 0.26 & 0.16 & 0.39 & 0.35 \\
joda-time-2.8.1 & 0.65 & 0.50 & 0.86 & 0.76 & 0.28 & 0.18 & 0.42 & 0.36 \\
joda-time-2.8.2 & 0.65 & 0.50 & 0.86 & 0.76 & 0.28 & 0.18 & 0.41 & 0.37 \\
joda-time-2.9 & 0.64 & 0.51 & 0.86 & 0.76 & 0.27 & 0.18 & 0.40 & 0.35 \\ \hline
ognl-3.1 & 0.60 & 0.94 & 0.99 & 0.95 & 0.05 & 0.11 & 0.33 & 0.31 \\
ognl-3.1.1 & 0.61 & 0.93 & 0.99 & 0.94 & 0.04 & 0.11 & 0.32 & 0.29 \\
ognl-3.1.2 & 0.61 & 0.93 & 0.99 & 0.94 & 0.05 & 0.13 & 0.38 & 0.34 \\
ognl-3.1.3 & 0.59 & 0.93 & 0.99 & 0.95 & 0.05 & 0.13 & 0.38 & 0.36 \\ \hline
wire-2.0.0 & 0.45 & 0.81 & 0.95 & 0.90 & 0.48 & 0.51 & 0.71 & 0.61 \\
wire-2.0.1 & 0.42 & 0.81 & 0.95 & 0.91 & 0.50 & 0.52 & 0.72 & 0.62 \\
wire-2.0.2 & 0.42 & 0.80 & 0.95 & 0.91 & 0.50 & 0.52 & 0.71 & 0.62 \\
wire-2.0.3 & 0.42 & 0.79 & 0.95 & 0.90 & 0.49 & 0.49 & 0.70 & 0.61 \\ \bottomrule
\end{tabular}
\end{table}
As future work, Sentinel could be deployed as a service in the cloud in order to further reduce its training costs and facilitate its adoption in practice. Further scientific aspects of our proposal could be investigated by experimenting Sentinel with other fitness functions and by assessing whether the generated strategies can be used for cross-projects mutation testing, i.e.\ if strategies generated using a given program as training subject can obtain good results when used for mutation testing of other programs from similar domains. There are other cost reduction techniques that do not rely on mutant reduction such as parallelism, weak mutation, predictive mutation~\cite{Zhang2016}, the approach proposed by Petrovic and Ivankovic~\cite{Petrovic2018}, and test case selection techniques such as Regression Mutation Testing (RMT)~\cite{Chen2018}. Such techniques can be used in combination with mutant reduction strategies and in different phases of the testing activity. For example, RMT can be applied during regression testing, and then the mutant reduction strategies generated by Sentinel can be used to further reduce the mutant set. Their interoperability should be investigated in the future.
We have made a Java implementation of Sentinel (together with our data and results) publicly available~\cite{SentinelWebsite,SentinelGH} to allow for replication and extension of our study, and to facilitate Sentinel's adoption by both practitioners and researchers, who can tackle this rich avenue of future work.
\section*{Acknowledgements}
This work is supported by the Microsoft Azure Research Award (MS-AZR-0036P), the Brazilian funding agencies CAPES and CNPq (grants 307762/2015-7 and 305968/2018), and the ERC advanced fellowship grant no. 741278 (EPIC).
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Inverse imaging problems can often be described as
\begin{equation}\label{eq:forward_problem}
\mathbf{z} = \Ad \XX_{\mathrm{true}} + \mathbf{e}
\end{equation}
where $\XX_{\mathrm{true}}\in V^{n}$ with $V \in \{\mathbb R, \mathbb C\}$ is the object to be imaged, $\Ad: V^{n}\to V^{m}$ is a linear operator which models the data-acquisition process, $\mathbf{e}\in V^{m}$ denotes some random noise component and $\mathbf{z}\in V^{m}$ represents the measured data. The goal is to reconstruct $\XX_{\mathrm{true}}$ or at least a good enough approximation of it given the data $\mathbf{z}$. In practice, problem \eqref{eq:forward_problem} is however ill-posed for various reasons. For example, in Magnetic Resonance Imaging (MRI) which is known to suffer from long acquisition times, the measurement process is often accelerated by undersampling in the raw-data domain, the so-called $k$-space, leading to an underdetermined systems. In low-dose CT, where one reduces the radiation exposure of the patient by reducing the energy of the photons emitted from the X-ray source, the measured data is noisy. Further, different inherent properties of the operator $\Ad$ also often determine how well-posed the problem is.
Therefore, the reconstruction procedure requires the use of regularization methods to be able to obtain high quality images and particularly in medical imaging, images with diagnostic accuracy. A prominent approach is to formulate the reconstruction as a minimization problem
\begin{equation}\label{eq:reg_problem}
\underset{\XX }{\min} \, d(\Ad \XX, \mathbf{z}) + \mathcal{R}(\XX),
\end{equation}
where $d(\,\cdot \,, \, \cdot \,)$ denotes a data-discrepancy measure and $\mathcal{R}(\,\cdot\,)$ a regularization term. Typical choices for $\mathcal{R}$ vary from $\mathcal{R}(\,\cdot\,) = \| \,\cdot\,\|_2^2$ for the well-known Tikhonov regularization \cite{Tikhonov} or $\mathcal{R}(\,\cdot\,) = \| \mathbf T \,\cdot\,\|_1$ for methods enforcing sparsity in some basis \cite{Donoho_1994, Chang_2000}.
One of the most widely applied methods is the so-called Total Variation (TV) regularization \cite{ROF92, ChambolleLions}.
Remaining in the finite dimensional setting, and choosing the square of the $\ell_{2}$ norm as data discrepancy (appropriate for Gaussian noise), the reconstruction problem is formulated as
\begin{equation}\label{eq:tv_min_problem}
\underset{\XX }{\min}\; \frac{1}{2} \| \Ad \XX - \mathbf{z}\|_2^2 + \lambda \| \nabla \XX \|_1.
\end{equation}
Here $\nabla$ denotes a finite-differences operator and $\lambda>0$ is a scalar regularization parameter that balances the effect of the two terms. This means that the regularization imposed on the sought image is given by sparsity in the gradient domain of the image measured with respect to the $\ell_1$-norm. One reason for the great success of this method lies in its simple intuition, interpretability as well as its interesting mathematical properties. As a result, in the last decades, it has driven both theoretical as well as applied research fields such as biomedical engineering, inverse problems, optimization and geometric measure theory among others \cite{chambollepock2016, Burger2013, scherzer2009variational}, with the complete list of publications in which the approach is investigated for different reconstruction problems in different imaging modalities being quite extensive. In addition, there exist nowadays numerous algorithms with proven convergence guarantees \cite{chambolle2011first, wang2008new, stadler, hint_kun, chambollepock2016, Tai_augmented} as well as extensions to overcome inherent limitations of the structural properties of the solutions of the problem \eqref{eq:tv_min_problem}, e.g.\ the total generalized variation (TGV) \cite{TGV}, for solving for the well-known TV staircasing artefacts (blocky-like, piecewise constant structures).\\
A crucial aspect which impacts the quality and the usefulness of the images which can be reconstructed by solving problem \eqref{eq:tv_min_problem} is the careful choice of the parameter $\lambda$. Underestimating $\lambda$ yields poor regularization, while overestimating it yields too smooth images with artificial ``cartoon-like" appearance. Particularly in medical imaging applications, where images are at the basis of diagnostic decisions and therapy planning, a proper choice of any regularization parameter is crucial. There exist quite a few methods regarding the automatic choice of a scalar parameter $\lambda$, placed either in the TV term or in the data-discrepancy term, based on the discrepancy principle, $L$-curve methods and others, e.g. \cite{Chuan_discrepancy, Golub1979, CALVETTI2000423, Montagner2014}.\\
However, employing one single scalar parameter $\lambda$ which globally dictates the strength of the regularization over the entire image might seem sub-optimal for various obvious reasons. Depending on the application, it might be desirable to maintain locally higher data-fidelity instead of enforcing visually appealing but rather wrong image features. In that case, one can replace the parameter $\lambda\in \mathbb R_{+}$ in \eqref{eq:tv_min_problem} with a spatially varying and pixel/voxel dependent one, denoted now by $\boldsymbol{\Lambda}\in \mathbb R_{+}^{qn}$,
with $q$ denoting the number of directions for which the partial derivatives are computed. Implementation-wise that translates to a stack of diagonal operators which contain a regularization parameter for each single pixel/voxel in the respective gradient domain of the image, resulting in a problem of the form
\begin{equation}\label{eq:tv_min_problem_spatial}
\underset{\XX }{\min}\; \frac{1}{2} \| \Ad \XX - \mathbf{z}\|_2^2 +\|\boldsymbol{\Lambda} \nabla \XX\|_{1}.
\end{equation}
An automatic choice for such spatially varying regularization parameter is rather challenging, as the number of its components drastically increases. Towards that task, bilevel optimization techniques have been employed during the last years, which have the following general formulation:
\begin{equation}\label{general_bilevel}
\left \{
\begin{aligned}
&\min_{\boldsymbol{\Lambda}}\; \sum_{i=1}^{M}l(\XX^{i}(\boldsymbol{\Lambda}),\XX_{\mathrm{true}}^{i})\\
&\text{subject to }\;\;\XX^{i}(\boldsymbol{\Lambda})=\underset{\XX}{\operatorname{argmin}}\;\frac{1}{2} \| \Ad \XX - \mathbf{z}_{i}\|_2^2 + \| \boldsymbol{\Lambda}\nabla \XX \|_1, \quad i=1,\ldots,M.
\end{aligned}\right.
\end{equation}
Here, $(\mathbf{z}_{i}, \XX_{\mathrm{true}}^{i})_{i=1}^{M}$ are $M$ pairs of measured data and corresponding ground truth, and $l$ is a suitable upper level objective. For instance, in the case where $l(x_{1},x_{2})=l_{\mathrm{PSNR}}(x_{1},x_{2}):=\|x_{1}-x_{2}\|_{2}^{2}$, the bilevel problem \eqref{general_bilevel} aims to compute the parameters $\boldsymbol{\Lambda}$ which are ``on the average the best ones" (i.e.\ PSNR-maximizing), for the given $M$ data pairs. The idea is that, given some new data $\mathbf{z}_{\mathrm{test}}$
that has been measured in a similar way as $(\mathbf{z}_{i})_{i=1}^{M}$, solving \eqref{eq:tv_min_problem_spatial} (in the ``online phase") with the offline-computed $\boldsymbol{\Lambda}$ will yield a good reconstruction.
This scheme has been extensively studied both for scalar and spatially varying regularization parameters. However, in practice it has mainly been applied for image denoising (i.e.\ $\Ad=\mathbf{I}_n$) and for scalar or coarse patch-based parameters \cite{pockbilevel, bilevellearning, noiselearning, Chung_Reyes_Schoenlieb, DelosReyes2021}. An extension for learning the optimal sampling pattern in MRI \cite{Sherry_2020}, as well as extensions to non-local and higher order regularizers \cite{D_elia_2021, TGV_learning2} have been considered as well. Further, unsupervised approaches, employing upper level energies that do not depend on the ground truth $\XX_{\mathrm{true}}$, i.e., $l=l(\XX(\boldsymbol{\Lambda}))$ and $M=1$, have also been considered in a series of works \cite{bilevelconvex, bilevelTGV, bilevel_handbook, HiRa17, hintermuellerPartII}. The upper level energy considered there aims to constrain localized versions of the image residuals $\Ad \XX - \mathbf{z}$ within a certain tight corridor around the variance of the (Gaussian) noise $\mathbf{e}$, which is assumed to be known. Even though these bilevel optimization methods are typically accompanied by elegant mathematical theories, there exist limitations on the computational time they require in order to give satisfactory results. For instance, employing these methods for 2D or even 3D dynamic imaging requires a vast computation effort and as a result, these limitations pose a challenge for the application in modern medical imaging modalities and hence in the clinical routine.\\
Recently, methods that are based on neural networks (NNs) have been proposed for the task of the estimation of such regularization parameter-maps. In \cite{Afkham_2021}, the authors employ a classical supervised learning approach in order to learn the map from the data $\mathbf{z}$ to the optimal scalar regularization parameter $\lambda$. The pipeline consists again of an offline and an online phase. More precisely, given again $M$ pairs of measured data and corresponding ground truth images $(\mathbf{z}_{i}, \XX_{\mathrm{true}}^{i})_{i=1}^{M}$, during the first part of the offline phase, a corresponding family of optimal regularization parameters $(\lambda_{i})_{i=1}^{M}$ is computed, e.g.\ by employing a scheme like \eqref{general_bilevel} separately for each $i$. Then, in the second part of the offline phase, using the training data $\mathcal{D}=\{(\lambda_{i}, \mathbf{z}_{i})_{i=1}^{M}\}$, the parameters $\Theta$ of a NN $\mathcal{N}_{\Theta}$ are learned by minimizing
\begin{equation}\label{sup_learn_param_off1}
\min_{\Theta} \mathcal{L}(\Theta) := \frac{1}{M} \sum_{i=1}^{M} l( \mathcal{N}_{\Theta}(\mathbf{z}_{i}),\lambda_{i}),
\end{equation}
for a suitable loss function $l$. Once an estimate of the optimal parameters $\Theta$ has been learned, one passes to the online phase, and given some new data $\mathbf{z}_{\mathrm{test}}$, the regularization parameter is simply calculated by applying the learned network to $\mathbf{z}_{\mathrm{test}}$, i.e., $\lambda_{\Theta}=\mathcal{N}_{\Theta}(\mathbf{z}_{\mathrm{test}})$ and the classical image reconstruction problem
\begin{equation}\label{sup_learn_param_off2}
\underset{\XX }{\min}\; \frac{1}{2} \| \Ad \XX - \mathbf{z}_{\mathrm{test}}\|_2^2 + \lambda_{\Theta}\| \nabla \XX \|_1,
\end{equation}
is solved by an appropriate algorithm. The idea is that, due to the good generalizability and adaptability of NNs on unseen data, the computed regularization parameter $\lambda_{\Theta}=\mathcal{N}_{\Theta}(\mathbf{z}_{\mathrm{test}})$ will be better adapted to $\mathbf{z}_{\mathrm{test}}$ than the ``average" $\lambda$ of the (scalar parameter version of) bilevel optimization approach \eqref{general_bilevel}.
The authors in \cite{Afkham_2021} apply this pipeline to learn scalar regularization parameters for computerized tomography reconstruction and image deblurring. Nevertheless, the computational burden for computing offline the training data as well as solving \eqref{sup_learn_param_off2} for high dimensional (3D and dynamic) problems still remains. A similar approach, where the supervised learning problem \eqref{sup_learn_param_off1} is performed at the level of small image-patches, was performed in \cite{Nekhili_2022} for the image denoising problem.\\
In this work, inspired by the recent success of unrolled NNs \cite{Monga_2021}, and targeting a variety of inverse problems including dynamic ones, we apply a different strategy for the construction of the regularization parameter-maps. We construct an unrolled NN which corresponds to an implementation of an iterative scheme of finite length to approach the solution of problem \eqref{eq:tv_min_problem} assuming a \textit{fixed} regularization parameter-map. Within the unrolled NN, the regularization parameter-map is estimated from the input data and is used throughout the whole reconstruction scheme. To be more precise, given some initial estimate $\XX_{0}$ we work with an iterative scheme
\begin{equation}\label{iteraive_scheme}
\XX_{T}=S^{T}(\XX_{0},\mathbf{z}, \boldsymbol{\Lambda},\Ad), \quad T=0,1,2,\ldots
\end{equation}
for which we know that $\XX_{T}\to S^*(\mathbf{z},\boldsymbol{\Lambda},\Ad)$ as $T\to\infty$ where $S^*(\mathbf{z},\boldsymbol{\Lambda}, \Ad)$ is a solution of \eqref{eq:tv_min_problem_spatial}.
We note that sometimes, we will drop the dependence of $S^{T}$ on $\XX_{0}$, $\mathbf{z}$, $\Ad$, for notational convenience. Then, for some fixed number of iterations $T\in \mathbb{N}$, our unrolled NN reads as follows:
\begin{equation}\label{unrolled_intro}
\left \{
\begin{aligned}
\boldsymbol{\Lambda}_{\Theta}&=\mathrm{NET}_{\Theta}(\XX_0),\\
\XX_{1}&=S^{1}(\XX_{0},\mathbf{z}, \boldsymbol{\Lambda}_{\Theta},\Ad),\\
\XX_{2}&=S^{2}(\XX_{0},\mathbf{z}, \boldsymbol{\Lambda}_{\Theta},\Ad),\\
\vdots\\
\XX_{T}&=S^{T}(\XX_{0},\mathbf{z}, \boldsymbol{\Lambda}_{\Theta},\Ad).
\end{aligned}
\right.
\end{equation}
Here, $\mathrm{NET}_{\Theta}$ denotes some convolutional NN (CNN) with learnable parameters $\Theta$. We denote by $\mathcal{N}_{\Theta}^{T}$ the overall resulting network, i.e.
\[\mathcal{N}_{\Theta}^{T}(\XX_0)=S^{T}(\XX_{0},\mathbf{z}, \boldsymbol{\Lambda}_{\Theta},\Ad)
=S^{T}(\XX_{0},\mathbf{z}, \mathrm{NET}_{\Theta}(\XX_0),\Ad).\] The unrolled NN can then be end-to-end trained in a supervised manner on a set of input-target image-pairs.
This resulting network can be identified as a pipeline that combines in a sequential way
\begin{itemize}
\item the estimation of the regularization parameter-map which is adapted to the data $\mathbf{z}$ (and hence in medical imaging to the new patient) and
\item the iterative scheme that solves the image reconstruction problem.
\end{itemize}
In particular, given a new unseen input data $\mathbf{z}_{\mathrm{test}}$, the regularization parameter-map $\boldsymbol{\Lambda}_{\Theta}$ is estimated and stays fixed. Then, the reconstruction problem is solved by unrolling an appropriate algorithm. As such, the resulting method is entirely interpretable and naturally inherits all convergence properties of the initial reconstruction algorithm since the data-driven component merely lies in the choice of the parameter-map.\\
Our approach can be considered to belong to the family of recently developed image reconstruction methods that combine elements both from \textit{model-based} and \textit{data-driven} regularization approaches. This is a modern and active field of research where interpretability and convergence guarantees from the traditional variational image reconstruction approaches are combined with the flexibility and adaptability of the deep-learning based methods. These combined approaches which aim to bring together the best of both worlds can result from instance by learning the regularization functional $\mathcal{R}$ from data and embed it into a scheme like\eqref{eq:reg_problem}, see \cite{Lunz_2018, Li_2020}, by enforcing the reconstruction to be close to an output of a network via $\mathcal{R}$, e.g.\ $\mathcal{R}(\,\cdot\,) = \|\,\cdot\, - u_{\Theta}(\mathbf{s})\|_2^2$ for some network $u_{\Theta}$ with trainable parameters $\Theta$ and input $\mathbf{s}$ \cite{schlemper2017deep, hyun2018deep, Kofler_2020, Duff_2021}, by substituting proximal operators in classical iterative schemes by learned NN denoisers (in a ``plug-and-play'' fashion) \cite{Meihardt_2017, Romano_2017}, or by using learned iterative schemes \cite{Adler_2017, Adler_2018, Hammernik_2018, Kofler_2021}, see also the review papers \cite{McCann_2017, arridge_maass_oktem_schonlieb_2019, Monga_2021}. Since one of our choices for the iterative scheme \eqref{iteraive_scheme} will be the Primal-Dual Hybrid Gradient method (PDHG) of Chambolle and Pock \cite{chambolle2011first}, our approach is related to the Learned Primal-Dual method \cite{Adler_2018}, where the proximal operators in the primal and dual step of PDHG are fully substituted by learnable networks. Here, we decrease the complexity but increase the interpretability by keeping the iterations of the iterative scheme untouched and put all the power of NNs in the estimation of the input-dependent regularization parameter-map, given in the first line of \eqref{unrolled_intro}. As a result, our approach can be regarded as an intermediate approach between \cite{Afkham_2021} and \cite{Adler_2018}.
One of the main reasons we follow this approach is because, apart from the increased interpretability, we also target dynamic imaging applications and we are particularly interested in the interplay between the learned temporal and spatial regularization. As far as we are aware and in contrast to static imaging problems, there are no existing works on automatically computing regularization parameters for dynamic problems that are both spatially and temporally varying.
Furthermore, because the ``black-box'' nature of CNNs in entirely put on the estimation of the regularization parameter-maps, the probability to possibly observe instabilities of the method in the sense of \cite{Colbrook_2022} is rather small. From a theoretical point of view, at least for denoising, it can be shown that for smooth regularization parameter-maps, no artefacts (i.e.\ new discontinuities) can appear in the reconstructions and for rougher weights, any creation of new discontinuities can be controlled \cite{caselles2007discontinuity, mine_spatial, jalalzai2014discontinuities}. Moreover, even the worst-case of locally very large produced regularization weights will only result in a locally flat area in the image with controlled values.
Further, from a practical point of view, in preliminary experiments, we have observed that even fully random regularization parameter-maps yield reconstructions whose artefacts can be at worst similar to the ones which would result from a locally too low or too strong TV-regularization.
We evaluate the proposed approach on a variety of reconstruction problems such as accelerated cardiac cine MRI, quantitative MRI, dynamic image denoising and low-dose computerized tomography (CT). We show that the proposed approach significantly improves the reconstruction results which can be obtained by the respective methods using only scalar regularization values, and better preserves the fine scale details by adapting the regularization strength to the given data. We finally stress that even though here we focus on TV regularization only, the proposed framework can be in principle adapted to more sophisticated regularization methods
The rest of the paper is organized as follows. In Section \ref{sec:spat_temp_vrm} we review spatio-temporal TV-based regularization and introduce notation. In Section \ref{sec:proposed_method}, we present our proposed approach for obtaining a data/patient-adaptive spatial or spatio-temporal regularization parameter-map. We investigate theoretical aspects of the proposed approach in Section~\ref{sec:consistency_ana_scheme}, focusing on the consistency of the unrolled scheme. In Section \ref{sec:experiments}, we conduct experiments to evaluate the proposed method on different imaging problems. We conclude the work in Section \ref{sec:conclusion} by discussing some aspects of the proposed approach, its limitations and possible future research directions.
\section{Spatio-Temporal Variational Regularization Models}\label{sec:spat_temp_vrm}
In this section we introduce in more detail the different spatio-temporal regularization functionals and we review relevant works from the literature. In parallel, we also fix the different notations for the several regularization parameters (scalar and spatially/spatio-temporally varying) in relationship to the way these are computed, e.g.\ supervised, unsupervised, ground truth-based, NNs-based.
\subsection{Spatio-Temporal Total Variation}
Setting $V^{n}:=V^{n_{x}\times n_{y} \times n_{t}}$, $n_{x}, n_{y}, n_{t}\in \mathbb{N}$, we define the discrete spatio-temporal gradient operator $\nabla : V^{n}\to (V\times V\times V)^{n}$ as
\begin{equation}\label{def_spatio_temp_grad}
\nabla \XX(z)= [\nabla_x \XX(z), \nabla_y \XX(z), \nabla_t \XX(z)]^{\scriptstyle \boldsymbol{\mathsf{T}}}, \quad \XX \in V^{n},
\end{equation}
where $\nabla_x, \nabla_y, \nabla_t$ are finite difference operators along the corresponding dimension. Note that in the case $V=\mathbb{C}$, $\nabla_x \XX(z):=[\nabla_x \mathrm{Re}(\XX(z)), \nabla_x \mathrm{Im}(\XX(z))]$, with $\mathrm{Re}(\XX(z))$ and $\mathrm{Im}(\XX(z))$ denoting the real and the imaginary part of $\XX(z)\in \mathbb{C}$, respectively and similarly for $\nabla_y \XX(z)$ and $\nabla_t \XX(z)$. Here, $z\in [1, \ldots, n_{x}]\times [1, \ldots, n_{y}]\times [1, \ldots, n_{t}]:=\mathcal{I}$ denotes the set of indices. The (anisotropic) total variation of $\XX\in V^{n}$ is the $\ell_{1,1}$-norm of $\nabla \XX$
\begin{equation}\label{standard_TV}
\mathrm{TV}(\XX)=\|\nabla \XX\|_{1}=\sum_{z\in \mathcal{I}}|\nabla \XX(z)|_{1}:= \sum_{z\in \mathcal{I}} |\nabla_x \XX(z)| + |\nabla_y \XX(z)| + |\nabla_t \XX(z)|,
\end{equation}
and the isotropic version is defined analogously using the $\ell_{2,1}$-norm, i.e., with the Euclidean norm $|\cdot|_{2}$ being used instead of $|\cdot|_{1}$ in \eqref{standard_TV}. In this work we employ the anisotropic version of TV and we mention that for $V=\mathbb{C}$ we also set $|\nabla_x \XX(z)|=|\nabla_x \mathrm{Re}(\XX(z))| + |\nabla_x \mathrm{Im}(\XX(z))|$, and similarly for the $y$- and $t$-direction.
\subsection{Notations on the Different Regularization Weights and Corresponding Spatio-Temporal Total Variation Functionals}
In general, we will denote scalar and spatially (and/or temporally) varying regularization parameters with $\lambda$ and $\boldsymbol{\Lambda}$, respectively. Whenever such a parameter is the output of a NN (with weights $\Theta$), the subindex $\Theta$ will be used, i.e., $\lambda_{\Theta}$ or $\boldsymbol{\Lambda}_{\Theta}$. If such a parameter produces the best corresponding TV-reconstruction with respect to the PSNR for some given data $\mathbf{z}$, it will be denoted by $\lambda_{\mathrm{P}}$ or $\boldsymbol{\Lambda}_{\mathrm{P}}$. For instance, these would be the optimal parameters that are solutions to the bilevel scheme \eqref{general_bilevel} when the upper level energy $l_{\mathrm{PSNR}}$ is used and $M=1$. In that case, the training and the test image coincide. If the ``best" is understood as ``on average" based on some training data, i.e.\ bilevel scheme \eqref{general_bilevel} with $l_{\mathrm{PSNR}}$ and $M>1$, we denote these parameters as $\lambda_{\tilde{\mathrm{P}}}$ or $\boldsymbol{\Lambda}_{\tilde{\mathrm{P}}}$. In that case the test image is not part of the training data.
On the other hand, we use superindices to index whether the different components of the regularization parameters that correspond to the different dimensions are the same or not. For instance,
\begin{align}
\lambda^{x,y,t}&=(\lambda^{x}, \lambda^{y}, \lambda^{t}) \in \mathbb R_{+}^{3}\label{def_lambda_x_y_t},\\
\lambda^{xy,t}&=(\lambda^{xy}, \lambda^{xy}, \lambda^{t}) \in \mathbb R_{+}^{3}\label{def_lambda_xy_t},\\
\lambda^{xyt}&=(\lambda^{xyt}, \lambda^{xyt}, \lambda^{xyt}) \in \mathbb R_{+}^{3},\label{def_lambda_xyt}
\end{align}
denote parameters that weight all the components differently, weight only the spatial components equally, and weight all the components equally respectively. For instance, the use of \eqref{def_lambda_x_y_t} leads to the following version of weighted TV:
\begin{equation}\label{scalar_lambda_x_y_t_TV}
\mathrm{TV}_{\lambda^{x,y,t}}(\XX):=\|\lambda^{x,y,t}\nabla \XX\|_{1}=\sum_{z\in \mathcal{I}}\lambda^{x,y,t}|\nabla \XX(z)|_{1}:= \sum_{z\in \mathcal{I}} \lambda^{x}|\nabla_x \XX(z)| +\lambda^{y} |\nabla_y \XX(z)| + \lambda^{t}|\nabla_t \XX(z)|.
\end{equation}
In contrast, $\lambda^{xy,t}=(\lambda^{xy}, \lambda^{xy}, \lambda^{t})$ denotes a parameter where the spatial components $x$ and $y$ are weighted equally. Analogously, we define the spatio-temporally varying versions, generally denoted by $\LLambda\in \mathbb{R}_{+}^{qn}$. In particular, we define
\begin{align}
\boldsymbol{\Lambda}^{x,y,t}&=(\boldsymbol{\Lambda}^{x}, \boldsymbol{\Lambda}^{y},\boldsymbol{\Lambda}^{t}) \in (\mathbb R_{+}^{n})^{3},\label{def_Lambda_x_y_t}\\
\boldsymbol{\Lambda}^{xy,t}&=(\boldsymbol{\Lambda}^{xy}, \boldsymbol{\Lambda}^{xy}, \boldsymbol{\Lambda}^{t}) \in (\mathbb R_{+}^{n})^{3},\label{def_Lambda_xy_t}\\
\boldsymbol{\Lambda}^{xyt}&=(\boldsymbol{\Lambda}^{xyt}, \boldsymbol{\Lambda}^{xyt}, \boldsymbol{\Lambda}^{xyt}) \in (\mathbb R_{+}^{n})^{3},\label{def_Lambda_xyt}
\end{align}
with \eqref{def_Lambda_xy_t}, for instance, leading to the following version of weighted TV
\begin{align}\label{spatial_Lambda_x_y_t_TV}
\mathrm{TV}_{\boldsymbol{\Lambda}^{xy,t} }(\XX)&:=\| \boldsymbol{\Lambda}^{xy,t}\nabla \XX\|_{1}=\sum_{z\in \mathcal{I}}|\boldsymbol{\Lambda}^{xy,t}(z)\nabla \XX(z)|_{1} \\
&:=\sum_{z\in \mathcal{I}} \boldsymbol{\Lambda}^{xy}(z)|\nabla_x \XX(z)| + \boldsymbol{\Lambda}^{xy}(z)|\nabla_y \XX(z)| + \boldsymbol{\Lambda}^{t}(z)|\nabla_t \XX(z)|.
\end{align}
Here the multiplication of $\boldsymbol{\Lambda}^{xy,t}$ and $\nabla\XX$ is considered component-wise. The full notations of the type, e.g.\ $\boldsymbol{\Lambda}_{\Theta}^{xy,t}$, $\boldsymbol{\Lambda}_{\mathrm{P}}^{xy,t}$, $\lambda_{\mathrm{P}}^{xy,t}$ have the obvious meaning.
By definition, we have the following, easily checked, inequalities for the PSNRs of the corresponding reconstructed images:
\begin{align}
&\lambda_{\tilde{\mathrm{P}}}^{xy,t}\preceq \lambda_{\mathrm{P}}^{xy,t}\preceq \boldsymbol{\Lambda}_{\mathrm{P}}^{xy,t},\\
&\lambda_{\tilde{\mathrm{P}}}^{x,y,t}\preceq \lambda_{\mathrm{P}}^{x,y,t}\preceq \boldsymbol{\Lambda}_{\mathrm{P}}^{x,y,t},\\
&\lambda_{\mathrm{P}}^{xy,t}\preceq \lambda_{\mathrm{P}}^{x,y,t},\\
&\boldsymbol{\Lambda}_{\mathrm{P}}^{xy,t}\preceq \boldsymbol{\Lambda}_{\mathrm{P}}^{x,y,t}.
\end{align}
Of particular interest will be the comparison of the above to the reconstructions that correspond to the parameters $\boldsymbol{\Lambda}_{\Theta}^{x,y,t}$/$\boldsymbol{\Lambda}_{\Theta}^{xy,t}$/$\boldsymbol{\Lambda}_{\Theta}^{xyt}$ that we learn through our unrolled scheme. We also note that the quantities correspond to spatial regularization only, i.e., static imaging tasks are defined straightforwardly by omitting the temporal component $t$.
\subsection{Related Literature on Spatio-Temporal Total Variation-type Functionals}
There have been quite a few related works in the dynamic inverse problems literature that employ regularization functionals of the type \eqref{scalar_lambda_x_y_t_TV}, or higher order extensions. Even though, we will also use our approach for static tasks, we briefly review these works since the literature on computing spatio-temporal regularization parameters for dynamic problems is essentially void.
In \cite{TGV_video}, the authors use a regularization functional defined as an infimal convolution of functionals of the type \eqref{scalar_lambda_x_y_t_TV} for video denoising and decompression, an approach which splits the image sequence into two components with little change in space and time respectively. A bilevel approach for dynamic denoising is considered in \cite{BilevelDynamicDenoising}. A higher extension of the approach, applied to dynamic MRI was investigated in \cite{Schloegl_2017} and for dynamic PET in \cite{bergounioux:hal-01694064}. Regularization of the type \eqref{scalar_lambda_x_y_t_TV} has been also considered for dynamic tomographic imaging \cite{Papoutsellis_2021} and dynamic cardiac MRI \cite{wang2013compressed}. We stress however that in all these works all regularization parameters are scalar and they are manually selected.
Here we allow for better flexibility in the regularization by automatically computing regularization parameters that are both spatially and temporally dependent, and we let these parameters guide the decoupling to static and moving parts in the image sequence.
\section{Proposed Unrolled Network Structure}\label{sec:proposed_method}
As described in \eqref{unrolled_intro}, our network architecture $\mathcal{N}_{\Theta}^T$ consists of two parts.
The first part of the network is concerned with the determination of the regularization parameter-map $\boldsymbol{\Lambda}_{\Theta}$ which is subsequently fed into the second part. We describe this procedure in more detail in Section \ref{subsubsec:obtaining_lambda}. Assuming the regularization parameter-map $\LLambda_{\Theta}$ is fixed in the second module of the network, $\LLambda_{\Theta}$ is fed into an unrolled iterative scheme of length $T$ which, if run until convergence, exactly solves
\begin{equation}\label{eq:tv_min_problem_parameter_map}
\underset{\XX }{\min}\; \frac{1}{2} \|\Ad \XX - \mathbf{z} \|_2^2 + \|\LLambda_{\Theta} \nabla\XX\|_1.
\end{equation}
For the latter we choose $T$ iterations of the PDHG algorithm \cite{chambolle2011first}, which we briefly recall here.
\begin{figure}[!t]
\centering
\includegraphics[width=0.9\linewidth]{figures/approach/NET_Theta_PDHG.pdf}
\caption{\small{Illustration of the proposed network architecture for a dynamic cardiac MR image reconstruction problem. The network consists of a sub-network which estimates the regularization parameter-maps and a sub-network which reconstructs the image using the PDHG algorithm described in Algorithm \ref{algo:pdhg_algo}. First, a spatio-temporal parameter-map $ \boldsymbol{\Lambda_{\Theta}}$, here as in \eqref{def_Lambda_xy_t}, is estimated by applying $\mathrm{NET}_{\Theta}$ to an input image $\XX_0$. The regularization parameter-map is then used within the reconstruction network which assumes the parameter-map to be fixed. The regularization parameter-map is trained such that the output of the PDHG algorithm is close to a reference image. } }
\label{fig:tv_primal_dual_nn.pdf}
\end{figure}
By denoting $X=V^{n}$, $Z=V^{m}, Q=V^{qn}$, the image reconstruction problem \eqref{eq:tv_min_problem_spatial} can be equivalently formulated as
\begin{equation}\label{eq:fKx_gx}
\underset{\XX\in X }{\min}\; f(\mathbf{K}\XX) + g(\XX),
\end{equation}
with $f:Y=Z\times Q\to \mathbb{R}$ where
\begin{align}
f(\YY)=f(\mathbf{p}, \mathbf{q}) := f_1(\mathbf{p}) + f_2(\mathbf{q}) = \frac{1}{2} \| \mathbf{p} - \mathbf{z} \|_2^2 + \| \boldsymbol{\Lambda}_{\Theta} \, \mathbf{q} \|_1, \quad
\mathbf{K} := \begin{bmatrix}
\Ad\\
\nabla
\end{bmatrix},\quad
g(\XX) := \mathbf{0} \label{eq:identification3}.
\end{align}
Here, the variables $\mathbf{p}, \mathbf{q}$ belong to the finite dimensional Euclidean spaces $Z$ and $Q$ that correspond to the specificities, e.g.\ dimensions, of each problem, and $\mathbf{K}:X\to Y$.
The PDHG algorithm for solving problems of the general form \eqref{eq:fKx_gx}, i.e., with $f$ and $g$ convex as well as $\mathbf{K}$ bounded and linear is described in Algorithm \ref{algo:pdhg_algo}. Recall that, for a convex function $h$ and scalar $\sigma>0$, the proximal operator $\mathrm{prox}_{\sigma h}$ is defined as
\begin{equation}\label{eq:prox}
\mathrm{prox}_{\sigma h}(\overline{\YY}):= \underset{\YY}{\operatorname{argmin}} \, \frac{1}{2}\| \YY - \overline{\YY} \|_2^2 + \sigma h(\YY),
\end{equation}
while the convex conjugate of $h$ is defined as
\begin{equation}\label{eq:conjugate}
h^\ast(\overline{\YY}):= \underset{\YY }{\max} \, \langle \YY , \overline{\YY} \rangle - h(\YY).
\end{equation}
In order to be consistent with our purposes, we have stated the Algorithm \ref{algo:pdhg_algo} such that it terminates in $T$ iterations with an output $\XX_{T}$. However, from standard convergence analysis it holds that $\XX_{T}\to \XX^{\ast}$ as $T\to\infty$, where $\XX^{\ast}$ solves \eqref{eq:tv_min_problem_spatial}.
\begin{algorithm}[h]
\caption{Unrolled PDHG algorithm \cite{chambolle2011first}}\label{algo:pdhg_algo}
\begin{algorithmic}[1]
\INPUT $L = \| \mathbf{K}\|$,\; $\tau\sigma\le 1/L^2$, \; $\theta = 1$, \; $\text{initial guess}~\XX_0$
\PARAMETER number of iterations $T>0$
\OUTPUT reconstructed image $\XX_{T}$
\STATE $\bar{\XX_0} = \XX_0$
\STATE $\YY_0 = \mathbf{0}$
\FOR {$k < T$ }
\STATE $\YY_{k+1} = \mathrm{prox}_{\sigma f^\ast}( \YY_k + \sigma \mathbf{K} \bar{\XX}_k)$
\STATE $\XX_{k+1} = \mathrm{prox}_{\tau g}( \XX_k - \tau \mathbf{K}^{\scriptstyle \boldsymbol{\mathsf{T}}} \YY_{k+1})$
\STATE $\bar{\XX}_{k+1} = \XX_{k+1} + \theta(\XX_{k+1} - \XX_k)$
\ENDFOR
\end{algorithmic}
\end{algorithm}
\begin{remark}\label{rem:Lambda_in_algo}
Later we recall the precise form of Algorithm \eqref{algo:pdhg_algo} for the problem \eqref{eq:tv_min_problem_spatial}. Here, we only mention that the $\mathrm{prox}_{\sigma f^{\ast}}$ for $f$ as in \eqref{eq:identification3}, decouples to $\mathrm{prox}_{\sigma f_{1}^{\ast}}$ and $\mathrm{prox}_{\sigma f_{2}^{\ast}}$, with the latter acting as a pointwise projection (``clipping'') onto the bilateral set $[-\boldsymbol{\Lambda}_{i}, \boldsymbol{\Lambda}_{i}]$, $i=1,\ldots, qn$. In particular, the map $\LLambda \mapsto \mathrm{prox}_{\sigma f_{2}^{\ast}}(\mathbf{q})$ is Lipschitz with constant one, for every $\mathbf{q}$. We remark that this is the only place where the parameter $\LLambda$ appears in the version of Algorithm \eqref{algo:pdhg_algo} for the problem \eqref{eq:tv_min_problem_spatial}.
\end{remark}
\begin{remark}\label{rem:PD3O}
We mention that for the low-dose CT application which we consider in Section \ref{sec:CT}, we will be using a generalization of PDHG, namely the PD3O algorithm \cite{Yan2018} which is better adapted for the Kullback-Leibler divergence fidelity term used there. We give more details later in that section.
\end{remark}
\subsection{Obtaining the Regularization Parameter-Map Via a CNN}\label{subsubsec:obtaining_lambda}
In our set-up, $\boldsymbol{\Lambda}_{\Theta}$ is the output of a CNN with parameters $\Theta$, denoted by $\mathrm{NET}_{\Theta}$, which takes as an input an initial image $\XX_0$, i.e., $\boldsymbol{\Lambda}_{\Theta} = \mathrm{NET}_{\Theta}(\XX_0)$.
Depending on the structure of the considered imaging problem, we can explore different possibilities for the construction of the latter. For instance, for a dynamic imaging problem, i.e., 2D + time, we might prefer to attribute equal importance to the $x$- and $y$-direction, but use a different parameter-map for the temporal component resulting in
\begin{equation}\label{eq:spatio_temporal_lambda_map}
\boldsymbol{\Lambda}_{\Theta}=(\boldsymbol{\Lambda}_{\Theta}^{xy}, \boldsymbol{\Lambda}_{\Theta}^{xy},\boldsymbol{\Lambda}_{\Theta}^{t}).
\end{equation}
This choice is motivated by the later shown cardiac cine MRI reconstruction problem. There, the temporal dimension is the one which on the one hand exhibits the largest correlation to be exploited by the TV-method, but on the other hand also the one which contains the diagnostic information and therefore requires special care to ensure that important features are preserved.\\
For a 3D imaging problem, one could for example attribute equal importance to all spatial-directions or opt for a construction as in \eqref{eq:spatio_temporal_lambda_map}, if for example the $z$-direction has a different resolution than the $x$- and $y$-directions. Moreover, for complex-valued images, it seems intuitive to share the same regularization map across the real and the imaginary parts of the images.\\
The core of the overall network, denoted by $u_{\Theta}$, consists of a (sub-)CNN with high expressive capabilities such as the U-Net \cite{Ronneberger2015}.
To constrain the regularization parameter-maps to be strictly positive, we then apply a softplus activation function $\phi$ and, as last operation, we multiply the output by a positive parameter $t>0$.
Empirically, we have experienced that the network's training benefits in terms of faster convergence if the order of the scale of the output is properly set depending on the application. This can be achieved either by accordingly initializing the weights of the network $u_{\Theta}$, or in a simpler way, as we do here by scaling the output of the CNN. Summarizing, given an input image $\XX_0$, we estimate the corresponding regularization parameter-map by
\begin{equation}\label{eq:lambda_reg_map}
\boldsymbol{\Lambda}_{\Theta}=\mathrm{NET}_{\Theta}(\XX_0) = t\, \phi( u_{\Theta}(\XX_0)).
\end{equation}
We finally recall that the overall network has the form
\begin{equation}\label{eq:final_net}
\mathcal{N}_{\Theta}^{T}(\XX_0)=
S^{T}(\XX_{0},\mathbf{z}, \mathrm{NET}_{\Theta}(\XX_0),\Ad).
\end{equation}
\begin{remark}
Note that in our set-up we use the same quantity $\XX_{0}$ as the input for the CNN-block $\mathrm{NET}_{\Theta}(\XX_{0})$ as well as the initialization for the unrolled PDHG $S^{T}(\XX_{0},\ldots)$. According to our experience, this produces satisfactory results, see also the discussion in Section \ref{sec:initialization}. However this is not a hard constraint of the method and one could also further experiment with having different values for these variables.
\end{remark}
\begin{figure}[!t]
\centering
\includegraphics[width=0.8\linewidth]{figures/approach/u_Theta_v2.pdf}
\caption{\small{
The CNN-block $u_{\Theta}$ is a U-Net which we hyper-parametrize by the number of encoding stages, the number of convolutional layers and the initial number of filters applied to the input-image. The filter-sizes as well as the number of output channels of $u_{\Theta}$ depend on the considered application. The network shown here is a 3D U-Net with four encoding stages, two convolutional layers per stage and 32 initially applied filters. Here, the CNN is constructed to yield two different components of the regularization parameter-map, i.e.\ $\LLambda_{\Theta}^{xy}$ and $\LLambda_{\Theta}^t$, which are then used to construct the final spatio-temporal regularization parameter-map $\LLambda_{\Theta}^{xy,t}$ according to \eqref{def_Lambda_xy_t}, i.e.\ by $\LLambda_{\Theta}^{xy,t} = (\LLambda_{\Theta}^{xy}, \LLambda_{\Theta}^{xy}, \LLambda_{\Theta}^{t})$.}}
\label{fig:u_Theta.pdf}
\end{figure}
\subsection{Network Training}
Training the network $\mathcal{N}_{\Theta}^T$ refers to minimizing a chosen energy-function $\mathcal{L}$ over a set of input-target training pairs $\mathcal{D}=\{ (\XX_0^i, \XX_{\mathrm{true}}^i)_{i=1}^M : \XX_0^i = \Ad^\ddagger \mathbf{z}_i, \,\mathbf{z}_i:= \Ad \XX_{\mathrm{true}}^i + \mathbf{e}_i\}$. Here $\Ad^\ddagger$ denotes some reconstruction operator, e.g. the pseudo-inverse of $\Ad$, which provides the inputs $\XX_{0}^{i}$ and $\mathbf{e}_i$ are noisy terms. Using an appropriate loss function $l$ and potentially some regularization function $r$ for the weights $\Theta$, we end up with the following energy-function:
\begin{equation}\label{eq:loss_fct}
\mathcal{L}(\Theta) = \frac{1}{M} \sum_{i=1}^{M} l \big(\mathcal{N}_{\Theta}^T( \XX_0^{i})\, , \, \XX_{\mathrm{true}}^{i}\big)+ r(\Theta).
\end{equation}
Note that $r$ can also encode some constraints on $\Theta$ by being an indicator of some set.
Note that the network is trained end-to-end from the initial reconstruction to its estimate. Therefore, the set $\Theta$ is adjusted such that the estimated parameter-map $\boldsymbol{\Lambda}_{\Theta}$ is appropriate for a subsequent reconstruction using a suitable reconstruction algorithm as for example, the primal-dual method in Algorithm \ref{algo:pdhg_algo}. This conceptually highly differs from approaches as in \cite{Afkham_2021}, in which a network is trained to estimate the best scalar regularization parameter which is previously obtained by a time-consuming grid-search. First of all, in \cite{Afkham_2021} the learning procedure is entirely decoupled from the employed reconstruction algorithm. Second, opposed to our approach, the method requires access to a target regularization parameter, meaning that a generalization of \cite{Afkham_2021} to regularization parameter-maps would require access to entire target regularization parameter-maps which can typically only obtained by even more time-consuming approaches. Our approach, in contrast, allows to implicitly learn the regularization parameter-maps by unrolling the reconstruction algorithm and thus only requires access to ground truth target images.
\section{Consistency Analysis of the Unrolled Scheme}\label{sec:consistency_ana_scheme}
In addition to the practical advantages of the proposed method which will be highlighted in Section \ref{sec:experiments}, we want to discuss some of the emerging theoretical questions and in particular some consistency results when we let the number of unrolled iterations $T\to\infty$. We note that there are papers that study hyperparameter search with bilevel optimization and unrolled optimization methods, see e.g. \cite{liu2020generic, liu2022general, ochs2015bilevel}. Although some of the latter articles provide consistency analysis in different contexts, we think that none of the techniques presented there can be applied to our problem.
We will be using the space and operator notation $X, Y, Z$ and $\mathbf{K}$ as these were defined in the previous section and we will also set $\mathcal{V}:=X\times Y$. For simplicity here we work with the real-valued case, i.e.\ $V=\mathbb{R}$.
Recall that the solution of the convex variational problem \eqref{eq:tv_min_problem_spatial}
and the corresponding $T$-th iterate of the unrolled algorithm are
denoted by $\XX^{\ast}=S^*(\mathbf{z},\boldsymbol{\Lambda})$ and $\XX_{T}=S^T(\XX_{0},\mathbf{z},\boldsymbol{\Lambda})$. Recall also that for the ease of notation we sometimes suppress the dependence of $S^{T}$ on the initialization $\XX_0$ of Algorithm \ref{algo:pdhg_algo}, as well as the one of the dual variable $\YY_{0}$. We then have $S^T(\XX_0,\mathbf{z},\boldsymbol{\Lambda}) \to S^*(\mathbf{z},\boldsymbol{\Lambda})$ as $T\to\infty$.
Furthermore, for this section we consider a more general fidelity term $d$, such that $f_{1}(\cdot):=d(\cdot,\mathbf{z})$.
Let us now consider the learning framework as presented in Section \ref{sec:proposed_method}
\begin{equation}
\min_{\Theta\in \mathbb{R}^{\ell}} \mathcal{L}^T(\Theta) := \frac{1}{M} \sum_{i=1}^M l (\mathcal{N}_\Theta^T(\XX_0^i)\,,\, \XX_{\text{true}}^i) + r(\Theta), \label{eq:empirical_risk_T}
\end{equation}
as well as the corresponding training scheme where no unrolling is taking place, i.e.,
\begin{equation}
\min_{\Theta \in \mathbb{R}^{\ell}} \mathcal{L}^*(\Theta) := \frac{1}{M} \sum_{i=1}^M l (\mathcal{N}_\Theta^*(\XX_0^i)\,,\, \XX_{\mathrm{true}}^i) + r(\Theta), \label{eq:empirical_risk}
\end{equation}
where we used analogously the notation $\mathcal{N}_\Theta^*(\XX_{0}) := S^*(\mathbf{z},\LLambda_\Theta(\XX_{0}))$. Our target will be to show convergence of ($\epsilon$)-minimizers
of \eqref{eq:empirical_risk_T} to minimizers of \eqref{eq:empirical_risk} as $T\to\infty$ under appropriate conditions via a $\Gamma$-convergence argument.
Naturally, in order to guarantee existence of minimizers for the problems \eqref{eq:empirical_risk_T} and \eqref{eq:empirical_risk}, the functionals $\mathcal{L}^{T}$ and $\mathcal{L^{\ast}}$ must be coercive, in addition to the standard lower semicontinuity assumptions. However, it is not so clear if this can be achieved without imposing coercivity via the regularization function $r$, which can be the case when e.g.\ $r$ is some norm in $\mathbb{R}^{\ell}$ or an indicator function of a bounded set. Even though strictly speaking, it is not needed for our main consistency result Corollary \ref{cor:consistency}, we will assume that the minimization problems \eqref{eq:empirical_risk_T} and \eqref{eq:empirical_risk} indeed admit solutions.
Of course in deep learning practice, one does not compute minimizers for these problems, but rather it is aimed that the energy is decreased up to some degree based on a validation set, in order to guarantee generalizability. However, the analysis presented here can serve as a starting point to further show consistency in the level of stationary points and/or energy decrease using validation sets.
Below we summarize a series of assumptions which we will need next:
\begin{assumption}\label{main_assumptions}
We assume that the following hold:
\begin{enumerate}
\item[(i)] The operator $\Ad:X\to Z$ is injective.
\item [(ii)] The fidelity term $d(\cdot,\mathbf{z})$ is $\mu_{\mathbf{z}}$-strongly convex and Lipschitz continuously differentiable for every $\mathbf{z}\in Z$.
We denote by $L_{\mathbf{z}}$ the corresponding Lipschitz constant.
\item[(iii)] The parameters $\sigma, \tau>0$ in Algorithm \ref{algo:pdhg_algo} are small enough such that the matrix
\begin{equation}\label{M_st}
\mathbf{M} =
\begin{pmatrix}
\frac{1}{\tau} \Id & -\mathbf{K}^{{\scriptstyle \boldsymbol{\mathsf{T}}}} \\
-\mathbf{K} & \frac{1}{\sigma} \Id
\end{pmatrix}
\end{equation}
is symmetric, positive definite and thus defines a norm in $\mathcal{V}$.
Then there exist $c,C>0$ such that
\begin{equation}
c \| \VV \|_{2} \leq \|\VV \|_{\mathbf{M}} := \sqrt{\langle \mathbf{M} \VV , \VV \rangle} \leq C \| \VV \|_{2} \quad \text{for every $\VV \in \mathcal{V}$.}
\end{equation}
\item[(iv)] The regularization function $r:\mathbb{R}^{\ell}\to\overline{\mathbb{R}}:= \mathbb{R} \cup \lbrace + \infty \rbrace$ is proper and lower semicontinuous.
\item[(v)] The loss function $l: X \times X \to \mathbb{R}$ is continuous.
\item[(vi)] The activation functions in the U-Net $u_{\Theta}$ are continuous.
\end{enumerate}
\end{assumption}
\begin{remark}\label{rem:injectivity}
The injectivity of the operator $\Ad$,
together with the strong convexity of $d$, is used in order to ensure that $\XX\mapsto d(\Ad \XX, \mathbf{z})$ is strongly convex.
This indeed guarantees uniqueness of the solution for the variational problem, and in particular the map $\boldsymbol{\Lambda} \mapsto S^*(\mathbf{z},\boldsymbol{\Lambda})$ is well-defined and single-valued. For the applications we will consider in Section \ref{sec:experiments}, i.e., denoising, MRI with multiple receiver coils and CT with enough angular views and detectors, this injectivity assumption is satisfied. We note however it might be possible to drop this injectivity assumption following \cite{vaiter2015model}, \cite{vaiter2012robust}, or \cite{zhang2016one}.
\end{remark}
We start with the following Proposition \ref{prop:unroll_properties} regarding Lipschitz continuity and equicontinuity of the iterates $S^T(\mathbf{z},\boldsymbol{\Lambda}) $ with respect to $\boldsymbol{\Lambda}$. Note that the convergence $\boldsymbol{\Lambda}_{T}\to \LLambda$ as $T\to\infty$, in (iii) below, is merely part of a technical condition and it is not associated to the structure of our unrolled scheme where, as we have pointed out, the CNN-output $\boldsymbol{\Lambda}_{\Theta}=\mathrm{NET}_{\Theta}(\XX_0)$ remains unchanged.
\begin{proposition}\label{prop:unroll_properties}
Assuming $(i)$-$(iii)$ of Assumption \ref{main_assumptions}, the following statements hold:
\begin{enumerate}[label=(\roman*)]
\item The solution map $\boldsymbol{\Lambda} \mapsto S^*(\mathbf{z},\boldsymbol{\Lambda})$ is Lipschitz continuous for every $\mathbf{z}\in Z$. In particular the following bound holds for every $\LLambda_{1}, \LLambda_{2}\in \mathbb{R}_{+}^{qn}$,
\begin{equation}
\| S^*(\LLambda_1,\mathbf{z}) - S^*(\LLambda_2,\mathbf{z}) \|_{2} \leq \frac{2 \| \nabla \|}{\lambda_{\min}(\Ad^{\scriptstyle \boldsymbol{\mathsf{T}}} \Ad) \mu_\mathbf{z}} \|\LLambda_1 - \LLambda_2\|_{2}.
\end{equation}
\item The map $\boldsymbol{\Lambda} \mapsto S^T(\XX_0,\mathbf{z},\boldsymbol{\Lambda})$ is Lipschitz continuous for every $\mathbf{z}\in Z$, $\XX_0 \in X$ and $T\in \mathbb{N}$.
\item For $\|\LLambda\|_{2} \leq \overline{\LLambda}$ we obtain the following sub-linear rate, for $\VV_{0}:=(\XX_{0},\YY_{0})$ being the initial iterates of Algorithm \ref{algo:pdhg_algo}
\begin{equation}
\|S^{T}(\XX_{0}, \mathbf{z},\LLambda)-S^{\ast}(\mathbf{z},\LLambda)\|_{2} \leq \frac{3 C_{\mathbf{z},\Ad}}{T^{1/4}} \left( 1 +
\| \VV_0 - \VV^*(\LLambda,\mathbf{z}) \|_\mathbf{M} \right),
\end{equation}
where
\begin{equation}\label{def_C_z_A}
C_{\mathbf{z},\Ad} := \frac{\max \left( C L_\mathbf{z} \| \Ad \| \, , \, 4C \overline{\LLambda} \, ,\, 2\, ,\, \lambda_{\min}(\Ad^{\scriptstyle \boldsymbol{\mathsf{T}}} \Ad)\mu_\mathbf{z} \right)}{\lambda_{\min}(\Ad^{\scriptstyle \boldsymbol{\mathsf{T}}} \Ad) \mu_\mathbf{z}},
\end{equation}
with $\lambda_{\min}(\Ad^{\scriptstyle \boldsymbol{\mathsf{T}}} \Ad)$ denoting the smallest eigenvalue of $\Ad^{\scriptstyle \boldsymbol{\mathsf{T}}} \Ad$.\\
\item Whenever $\boldsymbol{\Lambda}_{T}\to \boldsymbol{\Lambda}$ as $T\to\infty$, it holds $S^T(\XX_0,\mathbf{z},\boldsymbol{\Lambda}_{T})\to S^*(\mathbf{z},\boldsymbol{\Lambda})$ for every $\XX_{0} \in X$, $\mathbf{z} \in Z$.
\end{enumerate}
\end{proposition}
\begin{proof}
(i) This statement is proved similarly to e.g.\ in \cite[Theorem 4.1]{Reyes_Villacis} and it is strongly based on the $\mu_{\mathbf{z}}$-strong convexity of the map $d(\cdot,\mathbf{z})$ and the injectivity of $\Ad$. The statement (ii) can also be seen easily since the only dependence of $\boldsymbol{\Lambda}$ in the unrolled PDHG scheme is via the pointwise projection onto $[-\boldsymbol{\Lambda}_{i},\boldsymbol{\Lambda}_{i}]$ which is a Lipschitz map, recall Remark \ref{rem:Lambda_in_algo}. As a result, the map $\boldsymbol{\Lambda} \mapsto S^T(\XX_{0}, \mathbf{z},\boldsymbol{\Lambda})$ is Lipschitz, as a composition of Lipschitz functions.
The proof of (iii) is more involved. We fix a ball of radius $\overline{\LLambda}$ centered at the origin, denoted by $B_{\bar{\LLambda}}\subset \mathbb{R}_{+}^{qn}$ and let $\LLambda\in B_{\overline{\LLambda}}$ be arbitrary. In what follows, we initially suppress the dependence of all variables on $\LLambda$.
Define the primal-dual gap
\begin{equation}
L(\XX,\YY) := \langle \mathbf{K} \XX, \YY \rangle - f^*(\YY) + g(\XX),
\end{equation}
and denote by $\VV_{T}:=(\XX_{T},\YY_{T})$ the iterates of the Algorithm \ref{algo:pdhg_algo} and by $\VV^{\ast}=(\XX^{\ast},\YY^{\ast})$ the corresponding limits. Then, the following estimate holds, see \cite[Corollary 1]{lu2022infimal} for a proof,
\begin{align}\label{lu2022infimal_inequality}
L(\XX_T,\YY) - L(\XX,\YY_T) \leq \frac{1}{\sqrt{T}} \left( \|\VV_0 - \VV^{*} \|_{\mathbf{M}}^{2} +
\|
\VV_0 - \VV^{*} \|_{\mathbf{M}} \|\VV - \VV^{*} \|_{\mathbf{M}} \right),
\end{align}
where $\VV=(\XX,\YY)$ is arbitrary.
We can thus take the supremum over $\YY\in \partial f(K\XX_{T})$ in both sides in \eqref{lu2022infimal_inequality} and estimate the left hand side as follows
\begin{equation}\label{ineq_1}
\sup_{\YY\in \partial f(\mathbf{K}\XX_{T})}L(\XX_T,\YY) - L(\XX,\YY_T)
\ge \sup_{\YY\in \partial f(\mathbf{K}\XX_{T})} \langle \mathbf{K}\XX_{T}, \YY \rangle - f^{\ast}(\YY) - \langle \mathbf{K}\XX,\YY_{T} \rangle+f^{\ast}(\YY_{T})\ge f(\mathbf{K}\XX_{T})-f(\mathbf{K}\XX),
\end{equation}
where we used the fact that $ \langle \mathbf{K}\XX_{T}, \YY \rangle - f^{\ast}(\YY)=f(\mathbf{K}\XX_{T})$ if and only if $\YY\in \partial f(\mathbf{K}\XX_{T})$. By setting $\XX=\XX^{\ast}$, using the $\mu_{\mathbf{z}}$-strong convexity of $f_{1}(\cdot)=d(\cdot,\mathbf{z})$, the convexity of $f_{2}$, together with $\YY^{\ast}\in \partial f (\mathbf{K}\XX^{\ast})$ we
deduce
\begin{equation}\label{ineq_2}
f(\mathbf{K}\XX_{T}) - f(\mathbf{K}\XX^{\ast})\ge \langle \mathbf{K}^{\ast}\YY^{\ast}, \XX_{T}-\XX^{\ast} \rangle + \frac{\mu_{\mathbf{z}}}{2} \| \Ad\XX_{T}-\Ad\XX^{\ast}\|_{2}^{2}.
\end{equation}
Taking into account that $\mathbf{K}^{\ast} \YY^{\ast}=0$ (taking limits at line 6 of Algorithm \ref{algo:pdhg_algo}, using the fact that $\mathrm{prox}_{\tau g}=Id$), using the injectivity of $\Ad$, we infer from \eqref{ineq_1} and \eqref{ineq_2}
\begin{equation}\label{lu2022infimal_inequality_v2}
\|\XX_{T}-\XX^{\ast}\|_{2}^{2}\le \frac{2}{\lambda_{\min}(\Ad^{{\scriptstyle \boldsymbol{\mathsf{T}}}} \Ad)\mu_{\mathbf{z}}\sqrt{T}} \sup_{\YY\in \partial f(\mathbf{K}\XX_{T})} \left( \|\VV_0 - \VV^{*} \|_{\mathbf{M}}^{2} +
\|
\VV_0 - \VV^{*} \|_{\mathbf{M}} \|\VV - \VV^{*} \|_{\mathbf{M}} \right).
\end{equation}
We proceed by estimating the last term in \eqref{lu2022infimal_inequality_v2} again making the dependence of $\LLambda$ explicit. Thus, recalling that $\VV=(\XX^{\ast}(\LLambda), \YY)$ with $(\mathbf{p},\mathbf{q})=:\YY\in \partial f(\mathbf{K}\XX_{T}(\LLambda))$ arbitrary, we have
\begin{align}
\|\VV-\VV^{\ast}(\LLambda)\|_{\mathbf{M}}
&\le C \sqrt{ \|\XX^{\ast}(\LLambda) - \XX^{\ast} (\LLambda)\|_{2}^{2} + \|\mathbf{p}-\mathbf{p}^{\ast}(\LLambda)\|_{2}^{2} + \|\mathbf{q}-\mathbf{q}^{\ast}(\LLambda)\|_{2}^{2}}\nonumber\\
&= C \sqrt{ \|\mathbf{p}-\mathbf{p}^{\ast}(\LLambda)\|_{2}^{2} + \|\mathbf{q}-\mathbf{q}^{\ast}(\LLambda)\|_{2}^{2}}\nonumber\\
& \le C \sqrt{ \|\nabla f_{1}(\Ad\XX_{T}(\LLambda)) - \nabla f_{1} (\Ad \XX^{\ast} (\LLambda)) \|_{2}^{2} + 4\overline{\LLambda}^{2}}\nonumber\\
& \le C\left ( L_{\mathbf{z}}\|\Ad\| \|\XX_{T}(\LLambda)-\XX^{\ast}(\LLambda)\|_{2} + 2 \overline{\LLambda}\right ),\label{last_bound}
\end{align}
where the last inequality used the fact that $\sqrt{a^{2}+b^{2}}\le a+ b$ for $a,b\ge 0$.
We also used the relationship $\mathbf{p}^{\ast}(\LLambda)=\nabla f_{1}(\Ad \XX^{\ast}(\LLambda))$, the Lipschitz continuity of $\nabla f_{1}$, as well as $\mathbf{q}^{\ast}(\LLambda)\in \partial f_{2} (\nabla \XX^{\ast} (\LLambda))$ which implies that $\mathbf{q}^{\ast}(\LLambda)\in B_{\overline{\LLambda}}$.
By combining \eqref{lu2022infimal_inequality_v2}, \eqref{first_bound} and \eqref{last_bound}, and by defining
\[r_{0}(\LLambda):=\|\VV_{0}-\VV^{\ast}(\LLambda)\|_{\mathbf{M}},\]
we end up to
\begin{equation*}\label{combined_bounds}
\|\XX_{T}(\LLambda)-\XX^{\ast}(\LLambda)\|_{2}^{2}\le \frac{2}{\lambda_{\min}(\Ad^{\scriptstyle \boldsymbol{\mathsf{T}}} \Ad) \mu_\mathbf{z} \sqrt{T}} \left( r_0(\LLambda)^2 + r_0(\LLambda) C L_\mathbf{z} \| \Ad \| \|\XX_{T}(\LLambda)-\XX^{\ast}(\LLambda)\|_{2} + 2 C r_0(\LLambda)\overline{\LLambda} \right).
\end{equation*}
By setting $r_T(\LLambda) := \|\XX_{T}(\LLambda)-\XX^{\ast}(\LLambda)\|_{2}$ and
\begin{equation}
C_1 := \frac{C L_\mathbf{z} \| \Ad \|}{\lambda_{\min}(\Ad^{\scriptstyle \boldsymbol{\mathsf{T}}} \Ad) \mu_\mathbf{z}}
\quad
C_2 := \frac{4 C \overline{\LLambda}}{\lambda_{\min}(\Ad^{\scriptstyle \boldsymbol{\mathsf{T}}} \Ad) \mu_\mathbf{z}}
\quad
C_3 := \frac{2}{\lambda_{\min}(\Ad^{\scriptstyle \boldsymbol{\mathsf{T}}} \Ad) \mu_\mathbf{z}},
\end{equation}
we infer
\begin{align*}
r_T(\LLambda)^2 - \frac{2 C_1}{\sqrt{T}} r_T(\LLambda)r_0(\LLambda)
+ \frac{C_1^2 r_0(\LLambda)^2}{T}
\le
\frac{2}{\sqrt{T}} \left( C_3 r_0(\LLambda)^2 + C_2r_0(\LLambda) \right) + \frac{C_1^2 r_0(\LLambda)^2}{T}.
\end{align*}
After applying the binomial formula, this yields
\begin{align}
r_T(\LLambda)
&\le
\frac{C_1 r_0(\LLambda)}{\sqrt{T}}
+
\sqrt{\frac{2}{\sqrt{T}} \left( C_3 r_0(\LLambda)^2 + C_2 r_0(\LLambda) \right) + \frac{C_1^2 r_0(\LLambda)^2}{T}} \\
&\le
\frac{C_{\mathbf{z} ,\Ad} r_0(\LLambda)}{\sqrt{T}} +
\sqrt{\frac{C_{\mathbf{z} ,\Ad}}{\sqrt{T}} \left( r_0(\LLambda)^2 + r_0(\LLambda) \right) + \frac{C_{\mathbf{z} ,\Ad}^2 r_0(\LLambda)^2}{T}}\\
&\leq
\frac{3 C_{\mathbf{z},\Ad}}{T^{1/4}} (r_0(\LLambda) + 1),
\end{align}
where the last inequality uses basic estimates, like $\sqrt{T} \leq T$, again $\sqrt{a^2 + b^2} \leq a + b$ for $a,b \geq0$ and the fact that $C_{\mathbf{z},\Ad} \geq 1$ by its definition \eqref{def_C_z_A}. This proves (iii).
To show (iv) let $\LLambda_{T}\to \LLambda$ and fix
$\overline{\LLambda}:=\sup_{T\in\mathbb{N}}\|\LLambda_{T}\|_{2}$. By (iii) we have that
\begin{equation}
\|\XX_{T}(\LLambda_T)-\XX^{\ast}(\LLambda_T)\|_{2} \leq \frac{3 C_{\mathbf{z},\Ad}}{T^{1/4}} \left(1 +
\| \VV_0 - \VV^*(\LLambda_T) \|_\mathbf{M} \right), \label{eq:bound_lambda_T}
\end{equation}
where we can further estimate by norm-equivalence
\begin{equation}\label{first_term}
\|\VV_{0}-\VV^{\ast}(\LLambda_T)\|_{\mathbf{M}}
\le C \sqrt{ \|\XX_{0}-\XX^{\ast}(\LLambda_T)\|_{2}^{2} + \|\mathbf{p}_{0}-\mathbf{p}^{\ast}(\LLambda_T)\|_{2}^{2} + \|\mathbf{q}_{0}-\mathbf{q}^{\ast}(\LLambda_T)\|_{2}^{2} }.
\end{equation}
Using again the boundedness of $(\LLambda_{T})_{T\in\mathbb{N}}$, the continuity of $\nabla f_{1}$ and the relationships $\mathbf{p}^{\ast}(\LLambda)=\nabla f_{1}(\Ad \XX^{\ast}(\LLambda))$ and $\mathbf{q}^{\ast}(\LLambda)\in \partial f_{2} (\nabla \XX^{\ast} (\LLambda))$, we conclude that there exists a constant $\hat{C}>0$ independent of $T$, such that
\begin{equation}\label{first_bound}
\|\VV_{0}-\VV^{\ast}(\LLambda_{T})\|_{\mathbf{M}}\le \hat{C}, \quad \text{ for every $T\in \mathbb{N}$}.
\end{equation}
Thus we deduce
\[\|\XX_{T}(\LLambda_{T})-\XX^{\ast}(\LLambda_{T})\|_{2}\to 0, \quad \text{as }T\to\infty.\]
We finally use the triangle inequality to obtain
\[\|\XX_{T}(\LLambda_{T})-\XX^{\ast}(\LLambda)\|_{2} \le \|\XX_{T}(\LLambda_{T})-\XX^{\ast}(\LLambda_{T})\|_{2}+\|\XX^{\ast}(\LLambda_{T})-\XX^{\ast}(\LLambda)\|_{2}\to 0,\]
as $T\to\infty$, where we have also used $(i)$.
\end{proof}
We can now proceed with our main result.
\begin{theorem}
Let the Assumption \ref{main_assumptions} hold, let the training set $\mathcal{D}$ be fixed and consider the sequence of functionals $\mathcal{L}^{T}:\mathbb{R}^{\ell} \to \overline{\mathbb{R}}$, $T\in\mathbb{N}$, as well as $\mathcal{L}^{\ast}:\mathbb{R}^{\ell} \to \overline{\mathbb{R}}$ defined as in \eqref{eq:empirical_risk_T} and \eqref{eq:empirical_risk}. Then we have that $\mathcal{L}^{T}$ $\Gamma$-converges to $\mathcal{L}^{\ast}$ as $T\to\infty$.
\end{theorem}
\begin{proof}
It suffices to check the conditions in the definition of $\Gamma$-convergence \cite{dalmasogamma}, i.e.,:
\begin{enumerate}
\item[$(i)$] For all $\Theta_{T}\to \Theta$, it holds $\mathcal{L}^{\ast}(\Theta)\le \liminf_{T\to\infty} \mathcal{L}^{T}(\Theta_{T})$.
\item[$(ii)$] For all $\Theta\in \mathbb{R}^{\ell}$, there exists $\Theta_{T}\to \Theta \text{ such that } \limsup_{T\to\infty} \mathcal{L}^{T}(\Theta_{T})\le \mathcal{L}^{\ast}(\Theta)$.
\end{enumerate}
The first condition holds due to the lower semicontinuity of $r$, the continuity of the map $\Theta\mapsto \LLambda_{\Theta}$, Proposition \ref{prop:unroll_properties} $(iv)$ and the continuity of the loss function $l$.
The fact that the map $\Theta\mapsto \LLambda_{\Theta}$ is continuous, follows by the continuity of all constituent functions, in particular, from the continuity of the activation functions of the U-Net $u_{\Theta}$. The second condition follows similarly, setting $\Theta_{T}:=\Theta$ for all $T\in\mathbb{N}$ and using the convergence of the iterative scheme, i.e. $S^{T}(\XX_{0}, \mathbf{z},\LLambda)\to S^{\ast}(\mathbf{z},\LLambda)$ as $T\to\infty$, as well as the continuity of the other involved functions.
\end{proof}
The following consistency result follows directly from the $\Gamma$-convergence, see \cite[Corollary 7.20]{dalmasogamma} for a proof.
\begin{corollary}[Consistency of the unrolled scheme]\label{cor:consistency}
Let the Assumption \ref{main_assumptions} hold, let the training set $\mathcal{D}$ be fixed and let $\epsilon_{T}\to 0$. Suppose that $\Theta_{T}$ is an $\epsilon_{T}$-minimizer of $\mathcal{L}^{T}$ i.e.\ $\mathcal{L}^{T}(\Theta_{T})\le \inf_{\Theta\in\mathbb{R}^{\ell}}\mathcal{L}^{T}(\Theta)+\epsilon_{T}$. Then, if $\Theta$ is an accumulation point of $(\Theta_{T})_{T\in\mathbb{N}}$ it is a minimizer of $\mathcal{L}^{\ast}$ and $\mathcal{L}^{\ast}(\Theta)=\limsup_{T\to\infty}\mathcal{L}^{T}(\Theta_{T})$.
\end{corollary}
\section{Applications}\label{sec:experiments}
In the following, we apply our proposed method to several different imaging problems to demonstrate its versatility. The considered imaging problems differ in terms of the operator $\Ad$ and, more importantly, on the number of dimensions, e.g.\ 2D, 3D or 2D+time as well as on the specific role the dynamic component plays in the respective problem.
All images were evaluated in terms of PSNR, normalized root mean-squared error (NRMSE), structural similarity index measure\cite{wang2004image} (SSIM) and blur effect \cite{crete2007blur}.\\
\texttt{Python} code is available at \href{https://www.github.com/koflera/LearningRegularizationParameterMaps}{\url{github.com/koflera/LearningRegularizationParameterMaps}}.
\subsection{Initialization for the Unrolled PDHG}\label{sec:initialization}
In general, an initial image for the PDHG can be directly reconstructed from the measured data by applying the adjoint of the forward operator, i.e., $\XX_0:=\Ad^{\scriptstyle \boldsymbol{\mathsf{H}}} \mathbf{z}$. Often, the set-up for realistic imaging problems is that $\Ad$ is given by a tall operator, i.e., $m > n$. Therefore, to obtain a better estimate of the unknown image from which one can estimate $\LLambda_{\Theta}$ by applying the CNN $\mathrm{NET}_{\Theta}$, one can consider the normal equation
\begin{equation}\label{eq:normal_eqs}
\Ad^{\scriptstyle \boldsymbol{\mathsf{H}}} \Ad \XX = \Ad^{\scriptstyle \boldsymbol{\mathsf{H}}} \mathbf{z},
\end{equation}
and approximately solve it. As $\Ad$ is typically constructed such that the normal operator $\Ad^{\scriptstyle \boldsymbol{\mathsf{H}}} \Ad$ is invertible, in the absence of noise, i.e., $\mathbf{z} \in \mathrm{range}(\Ad)$, solving \eqref{eq:normal_eqs} using an iterative scheme to approximate $\XX^{\dagger}:= (\Ad^{\scriptstyle \boldsymbol{\mathsf{H}}} \Ad)^{-1}\Ad^{\scriptstyle \boldsymbol{\mathsf{H}}} \mathbf{z}$ would allow for a perfect reconstruction of the ground truth image. However, in the presence of noise, early stopping is required to avoid a noise amplification during the iterations. An approximate solution of \eqref{eq:normal_eqs} can then be used as a better initial estimate for the PDHG method as well as the image from which the CNN $\mathrm{NET}_{\Theta}$ estimates the different components of the regularization map $\boldsymbol{\Lambda}_{\Theta}$.
For the case that the considered imaging problem is not overdetermined, i.e., $m\leq n$, e.g.\ for image denoising, one simply uses $\XX_0:=\Ad^{\scriptstyle \boldsymbol{\mathsf{H}}} \mathbf{z}$ as the input of $\mathrm{NET}_{\Theta}$.\\
\subsection{Dynamic Cardiac MR Image Reconstruction}\label{subsec:dyn_mri}
Here, we apply the proposed NN to a dynamic cardiac MR image reconstruction problem. The problem consists of a set of independent 2D problems from which static images of the heart can be reconstructed. By stacking the different images along time, one can obtain a sequence of images which cover the entire cardiac cycle, also referred to as cardiac cine MRI. In clinical practice, cardiac cine MRI can be used to assess the cardiac function, see e.g.\ \cite{von2017representation}. Due to the structure of the problem, the temporal dimension is the one which offers the greatest potential to exploit the sparsity of the image in its gradient domain. However, a careful choice of the regularization parameter-map is required to ensure that the cardiac motion as well as smaller diagnostic image details are well-preserved after the reconstruction.
\subsubsection{Problem Formulation}
For a complex-valued dynamic 2D MR image with vector representation $\XX \in \mathbb C^N$ with $n=n_x\cdot n_y\cdot n_t$, the forward operator in \eqref{eq:forward_problem} is given as
\begin{equation}\label{eq:operator_mri}
\Ad:= (\Id_{n_c} \otimes \Ed) \Cd,
\end{equation}
where $\Id_{n_c}$ denotes the $n_c\times n_c$-sized identity-operator with $n_c$ being the number of receiver coils used for the data acquisition. The operator $\Cd$ is a tall operator which contains the coil-sensitivity maps, i.e. $\Cd = [\Cd_1, \ldots, \Cd_{n_c}]^{\scriptstyle \boldsymbol{\mathsf{T}}}$ with $\Cd_k = \diag(\CC_k)$ and $\CC_k \in \mathbb C^n$, $k=1,\ldots,n_c$. Let $\Ed_{I_t}:=\Sd_{I_t} \Fd$ be an operator which acquires the $k$-space data of a static 2D MR image $\XX_t$ at time-point $t$ by sampling the $k$-space coefficients indexed by the set $I_t \subset J$, where $J=\{1,\ldots, n_{xy}\}$ with $n_{xy}:=n_x \cdot n_y$.
Thereby, the mask $\Sd_{I_t}\in \{0,1\}^{m_t \times n_{xy}}$ with $m_t< n_{xy}$ for all $t=1,\ldots,n_t$ models the undersampling process. Undersampling the Fourier-space data is employed in order to accelerate the data acquisition process which usually takes place during a breathhold of the patient. Finally, the encoding operator $\Ed$ is given by
\begin{equation}
\Ed:=
\begin{pmatrix}
\Ed_{I_1} & \mathbf{0}_{m_1\times n_{xy}} & \mathbf{0}_{m_1\times n_{xy}} & \cdots & \mathbf{0}_{m_1\times n_{xy}} \\
\mathbf{0}_{m_2\times n_{xy}} & \Ed_{I_2}& \mathbf{0}_{m_2\times n_{xy}} & \cdots & \mathbf{0}_{m_2\times n_{xy}} \\
\vdots & \vdots & \vdots & & \vdots & \\
\mathbf{0}_{m_{n_t}\times n_{xy}} & \mathbf{0}_{m_{n_t}\times n_{xy}} & \mathbf{0}_{m_{n_t}\times n_{xy}} & \cdots & \Ed_{I_{n_t}}
\end{pmatrix},
\end{equation}
where $\mathbf{0}_{m_t \times n_{xy} } \in \{0\}^{m_t \times n_{xy}}$ denotes a $m_t \times n_{xy}$-sized zero-matrix. Typically, the number of receiver coils $n_c$ is chosen to ensure that $m:=n_c\cdot (m_1 + \ldots + m_t)>n$, i.e.\ problem \eqref{eq:forward_problem} is overdetermined when \eqref{eq:operator_mri} is the forward model.\\
\subsubsection{PDHG for Dynamic Multi-Coil MRI}
For the sake of completeness, we briefly summarize the PDHG-algorithm based on the identification mentioned in \eqref{eq:identification3}.
Recall the definition of $f_2$ from \eqref{eq:identification3}. Since
\begin{comment}
\end{comment}
\begin{equation}\label{eq:clipping_op}
\big(\mathrm{prox}_{\tau f_2^\ast}(\mathbf{q})\big)_i =
\left \{
\begin{aligned}
- & (\LLambda_{\Theta})_i, \quad &\mathbf{q}_i &\in \big(-\infty, -(\LLambda_{\Theta})_i \big) \\
& \mathbf{q}_i, \quad &\mathbf{q}_i &\in \big[-(\LLambda_{\Theta})_i , (\LLambda_{\Theta})_i ) \big] \\
& (\LLambda_{\Theta})_i, &\mathbf{q}_i &\in \big( (\LLambda_{\Theta})_i, \infty \big)
\end{aligned}
\right. ,
\end{equation}
the proximal operator $\mathrm{prox}_{\tau f_2^\ast}$ acts by ``clipping" each entry in the vector $\mathbf{q}$ if its magnitude exceeds the corresponding entry in $\boldsymbol{\Lambda}_{\Theta}$ and we therefore abbreviate it as $\mathrm{prox}_{\tau f_2^\ast}:=\mathrm{clip}_{\boldsymbol{\Lambda}_{\Theta}}$ to emphasize its dependence on the regularization parameter-map $\LLambda_{\Theta}$. The algorithm is summarized in Algorithm \ref{algo:tv_reco_algo_mri}.
\begin{algorithm}
\caption{Unrolled PDHG algorithm for general linear operator $\Ad$ with $d(\,\cdot \, , \, \cdot \,) = \frac{1}{2}\| \cdot - \cdot\|_2^2$ and \textit{fixed} regularization parameter-map $\boldsymbol{\Lambda}_{\Theta}$ (adapted from \cite{sidky2012convex})}\label{algo:tv_reco_algo_mri}
\begin{algorithmic}[1]
\INPUT $L =\| [\Ad, \nabla]^{\scriptstyle \boldsymbol{\mathsf{T}}} \|$, \; $\tau = 1/L$, \; $\sigma = 1/L$, \; $\theta = 1$, \;$\text{initial guess}~\XX_0$
\OUTPUT reconstructed image $\XX_{\mathrm{TV}}$
\STATE $\bar{\XX_0} = \XX_0$
\STATE $\mathbf{p}_0 = \mathbf{0}$
\STATE $\mathbf{q}_0 = \mathbf{0}$
\FOR {$k < T$ }
\STATE $\mathbf{p}_{k+1} = ( \mathbf{p}_k + \sigma ( \Ad \bar{\XX}_k - \YY) / ( 1 + \sigma)$
\STATE $\mathbf{q}_{k+1} = \mathrm{clip}_{\boldsymbol{\Lambda}_{\Theta}}(\mathbf{q}_k + \sigma \nabla \bar{\XX}_k)$
\STATE $\XX_{k+1} = \XX_k - \tau \Ad^{\scriptstyle \boldsymbol{\mathsf{H}}} \mathbf{p}_{k+1} - \tau \nabla^{\scriptstyle \boldsymbol{\mathsf{T}}} \mathbf{q}_{k+1} $
\STATE $\bar{\XX}_{k+1} = \XX_{k+1} + \theta ( \XX_{k+1} - \XX_k)$
\ENDFOR
\STATE $\XX_{\mathrm{TV}} = \XX_T$
\end{algorithmic}
\end{algorithm}
\subsubsection{Experimental Set-Up}
\label{subsubsec:experiments:cardiac}
We used a set of 216 cardiac cine MR images of the study \cite{kolbitsch2014cardiac} which we split in portion of 144/36/36 for training, validation and testing. The images have shape $n_x\times n_y\times n_t = 160 \times 160 \times 30$ and a resolution of $2 \times 2$ mm$^2$ with a slice thickness of 8 mm$^2$. The number of receiver coils is $n_c=12$. We retrospectively simulated $k$-space data according to \eqref{eq:forward_problem} using the model in \eqref{eq:operator_mri} as the forward operator simulating acceleration factors of $R=4,6,8$ with complex-valued Gaussian noise with standard deviation $\sigma = 0.15, 0.30, 0.45$.\\
As described in Section \ref{subsubsec:obtaining_lambda}, we constructed $u_{\Theta}$ such that it yields two different parameter-maps. One for the spatial $x$- and $y$-directions and one for the temporal direction, i.e., $\boldsymbol{\Lambda}_{\Theta}:=(\boldsymbol{\Lambda}_{\Theta}^{xy}, \boldsymbol{\Lambda}_{\Theta}^{xy},\boldsymbol{\Lambda}_{\Theta}^{t})$.
The CNN $u_{\Theta}$ here corresponds to a 3D U-Net with two input-channels (for the real and the imaginary part of the image, respectively), three encoding stages, two convolutional layers per stage and an initial number of eight filters which are applied to the input image. As in Figure \ref{fig:u_Theta.pdf}, the last layer consists of a $1\times1\times 1$ convolution with two output channels (the first for the parameter-map for the $x$- and $y$-directions, the second for the parameter-map for the $t$-direction) and the softplus activation function $\phi$. Note that the gradients of the real and the imaginary parts of the images share the same regularization parameter-map. The scaling factor $t$ in \eqref{eq:lambda_reg_map} was set to $t=0.1$. The overall number of trainable parameters of $u_{\Theta}$ is 97\,290.
To reduce training times, the network was trained on patches of shape $n_x^\prime \times n_y^\prime \times n_t^\prime = 160 \times 160 \times 16$. The network's number of overall iterations was set to $T=256$ during training, while at test time, we used $T=4096$ iterations. The reason for the different number of iterations at training and test time is discussed later in Subsection \ref{subsec:choosing_T}. The parameters $\sigma, \tau$ and $\theta$ were trained as well and constrained to be in the intervals $(0,1/L), (0,1/L)$ and $(0,1)$, respectively, by using a sigmoid activation-function. Despite of the training, we mention that no noteworthy changes were visible after training, i.e. $\sigma \approx \tau \approx 1/L$. Not training $\sigma, \tau$ and $\theta$ also led to similar results as the ones shown later. As training routine, we used the Adam optimizer \cite{kingma2014adam} with initial learning rate of $10^{-4}$ to minimize the mean squared error (MSE) between the reconstructed image and the target image. We trained all networks for 200 epochs while evaluating the network 25 times over the entire training and validation datasets. We then used the model configuration for which the MSE on the validation set was the lowest.
\subsubsection{Results}
Figure \ref{fig:mri_results} shows an example of a single frame of the reconstructed MR image sequences for an acceleration factor of $R=6$ using several approaches. We show the reconstructions that correspond to the single scalar parameter $\lambda_{\mathrm{P}}^{xyt}$ as well as to the scalar parameter pair (one spatial and one temporal) $\lambda_{\mathrm{P}}^{xy,t}=(\lambda_{\mathrm{P}}^{xy}, \lambda_{\mathrm{P}}^{xy}, \lambda_{\mathrm{P}}^{t})$ which are the parameters that maximize the PSNR of entire cine MR image and are obtained via a grid search by making use of the corresponding ground truth image. We also show the results that correspond to the parameters $\lambda_{\tilde{\mathrm{P}}}^{xyt}$ and $\lambda_{\tilde{\mathrm{P}}}^{xy,t}$ which are respectively the single and the pair of scalar parameters that on average maximize the PSNR over the training set. These were obtained by treating the scalar regularization parameters as trainable parameters and training them by minimizing \eqref{eq:loss_fct}. We finally show the results for our estimated parameter-map $\boldsymbol{\Lambda}_{\Theta}^{xy,t}$ with the proposed method. As observed, for all choices of the regularization parameters, the error with respect to the target image was significantly reduced compared to the initial zero-filled reconstruction. Further, we can see how the use of the estimated parameter-map yields the most accurate reconstruction and the best preservation of image details.
Figure \ref{fig:box_plots_mri} summarizes the results obtained over the test set with the help of box-plots. Compared to the initial zero-filled reconstruction, an improvement is clearly visible for all choices of the regularization parameter with respect to all reported measures and for all acceleration factors. In addition we see how allowing the temporal direction to be differently regularized than the two spatial dimensions positively influences the results compared to having one global parameter $\lambda$ (orange vs blue). Last, we see how using the proposed method to estimate an entire spatio-temporal parameter-map $\boldsymbol{\Lambda}_{\Theta}$ further surpasses the scalar regularization parameter-maps (green vs orange and blue), especially in terms of SSIM.
Table \ref{tab:dyn_mri_results} lists the mean and the standard deviation of all TV-reconstructions. The results are consistent with the ones from the box-plots.
Figure \ref{fig:lambda_maps_mri} shows an example of a spatio-temporal regularization parameter-map which was estimated using the proposed approach for an acceleration factor of $R=6$. The network $u_{\Theta}$ estimates the regularization parameter-map to be pointwise relatively consistenly higher than the spatially required regularization. This result is in fact expected as the temporal dimension is the one for which the gradients of the images are the sparsest because of the high temporal correlation. Further, we see how the network consistently predicts both the spatial regularization as well as the temporal regularization to be less strong in the area where most of the movement is expected, i.e.\ in the cardiac region.
\begin{figure}
\centering
\begin{minipage}{\linewidth}
\begin{minipage}{\linewidth}
\hspace{0.8cm} PDHG $\lambda^{xyt}_{\tilde{\mathrm{P}}}$ \hspace{0.4cm} PDHG $\lambda^{xyt}_{\mathrm{P}}$ \hspace{0.4cm} PDHG $\lambda^{xy,t}_{\tilde{\mathrm{P}}}$ \hspace{0.4cm} PDHG $\lambda^{xy,t}_{\mathrm{P}}$ \hspace{0.4cm} PDHG $\LLambda^{xy,t}_\Theta$ \hspace{0.4cm} Target/ZF
\end{minipage}
\begin{minipage}{\linewidth}
\rotatebox{90}{
\begin{minipage}{0.4\linewidth}
\hspace{-4.6cm }$R=8$ \hspace{2.9cm} $R=6$ \hspace{3.cm} $R=4$
\end{minipage}
}
\begin{minipage}{\linewidth}
\begin{minipage}{\linewidth}
\resizebox{\linewidth}{!}{
\includegraphics[height=3cm]{figures/MRI/mri_results/xyt_avg_R4_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xyt_avg_R4_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/xyt_best_R4_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xyt_best_R4_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/xy_t_avg_R4_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xy_t_avg_R4_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/xy_t_best_R4_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xy_t_best_R4_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/xcnn_R4_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xcnn_R4_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/xf_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xf_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
}
\resizebox{\linewidth}{!}{
\includegraphics[height=3cm]{figures/MRI/mri_results/exyt_avg_R4_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/exyt_avg_R4_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/exyt_best_R4_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/exyt_best_R4_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/exy_t_avg_R4_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/exy_t_avg_R4_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/exy_t_best_R4_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/exy_t_best_R4_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/excnn_R4_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/excnn_R4_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/xu_R4_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xu_R4_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
}
\end{minipage}
\begin{minipage}{\linewidth}
\resizebox{\linewidth}{!}{
\includegraphics[height=3cm]{figures/MRI/mri_results/xyt_avg_R6_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xyt_avg_R6_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/xyt_best_R6_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xyt_best_R6_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/xy_t_avg_R6_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xy_t_avg_R6_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/xy_t_best_R6_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xy_t_best_R6_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/xcnn_R6_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xcnn_R6_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/xf_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xf_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
}
\resizebox{\linewidth}{!}{
\includegraphics[height=3cm]{figures/MRI/mri_results/exyt_avg_R6_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/exyt_avg_R6_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/exyt_best_R6_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/exyt_best_R6_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/exy_t_avg_R6_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/exy_t_avg_R6_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/exy_t_best_R6_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/exy_t_best_R6_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/excnn_R6_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/excnn_R6_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/xu_R6_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xu_R6_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
}
\end{minipage}
\begin{minipage}{\linewidth}
\resizebox{\linewidth}{!}{
\includegraphics[height=3cm]{figures/MRI/mri_results/xyt_avg_R8_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xyt_avg_R8_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/xyt_best_R8_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xyt_best_R8_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/xy_t_avg_R8_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xy_t_avg_R8_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/xy_t_best_R8_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xy_t_best_R8_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/xcnn_R8_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xcnn_R8_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/xf_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xf_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
}
\resizebox{\linewidth}{!}{
\includegraphics[height=3cm]{figures/MRI/mri_results/exyt_avg_R8_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/exyt_avg_R8_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/exyt_best_R8_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/exyt_best_R8_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/exy_t_avg_R8_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/exy_t_avg_R8_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/exy_t_best_R8_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/exy_t_best_R8_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/excnn_R8_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/excnn_R8_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/MRI/mri_results/xu_R8_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/MRI/mri_results/xu_R8_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
}
\end{minipage}
\end{minipage}
\end{minipage}
\end{minipage}
\caption{An example of images reconstructed with the primal-dual scheme in Algorithm \ref{algo:tv_reco_algo_mri} for different choices of regularization parameters and acceleration factors $R=4,6,8$.
Single scalar regularization parameter $\lambda_{\mathrm{P}}^{xyt}$ and $\lambda_{\tilde{\mathrm{P}}}^{xyt}$, two scalar regularization parameters for differently weighted spatial and temporal components, $\lambda_{\mathrm{P}}^{xy,t}$ and $\lambda_{\tilde{\mathrm{P}}}^{xy,t}$, the proposed spatially and temporal dependent parameter-map $\boldsymbol{\Lambda}_{\Theta}^{xy,t}$ obtained with the network $\mathcal{N}_{\Theta}^T$. The last column shows the target image and the zero-filled reconstruction. Note again that the results for $\lambda_{\mathrm{P}}^{xyt}$ and $\lambda_{\mathrm{P}}^{xy,t}$ were obtained performing a grid-search for $\lambda^{xyt}>0$ and $\lambda^{xy},\lambda^t>0$, assuming the ground truth image to be known. Therefore, the results for $\lambda_{\mathrm{P}}^{xyt}$ and $\lambda_{\mathrm{P}}^{xy,t}$ cannot be obtained in practice and merely serve for illustrating that the proposed $\boldsymbol{\Lambda}_{\Theta}^{xy,t}$ yields competitive results. }
\label{fig:mri_results}
\label{fig:my_label}
\end{figure}
\begin{figure}[!h]
\centering
\includegraphics[width=0.5\linewidth]{figures/MRI/mri_recon/legend_dyn_mri.pdf}\\
\includegraphics[width=0.325\linewidth]{figures/MRI/mri_recon/box_plots_mri_unbound_NRMSE.pdf}
\includegraphics[width=0.325\linewidth]{figures/MRI/mri_recon/box_plots_mri_unbound_PSNR.pdf}
\includegraphics[width=0.325\linewidth]{figures/MRI/mri_recon/box_plots_mri_unbound_SSIM.pdf}
\caption{Box-plots summarizing the reconstruction results in terms of PSNR, NRMSE and SSIM obtained with the PDHG for a dynamic cardiac MR image reconstruction problem for different choices of the regularization parameter. Zero-filled reconstruction (black), single scalar regularization parameter ($\lambda_{\tilde{\mathrm{P}}}^{xyt}$, blue), two scalar regularization parameters; one for the spatial $x$- and $y$-direction, one for the temporal $t$-direction ($\lambda_{\tilde{\mathrm{P}}}^{xy,t}$, blue) and the proposed spatially and temporal dependent parameter-map $\boldsymbol{\Lambda}_{\Theta}^{xy,t}$ obtained with a CNN ($\mathrm{NET}_{\Theta}$, green).}
\label{fig:box_plots_mri}
\end{figure}
\begin{table}[h]
\centering
\footnotesize\rm
\begin{tabular}{r|r|rcl|rcl|rcl}
& & \multicolumn{3}{c}{\textbf{PDHG - $\lambda_{\tilde{\mathrm{P}}}^{xyt}$}} & \multicolumn{3}{c}{\textbf{PDHG - $\lambda_{\tilde{\mathrm{P}}}^{xy,t}$}} & \multicolumn{3}{c}{\textbf{PDHG - $\boldsymbol{\Lambda}_{\Theta}^{xy,t}$}} \\[5pt]
\hline
& \textbf{SSIM} & $0.836$ & $\pm$ & $0.048$ & $0.824$ & $\pm$ & $0.059$ & \textbf{0.927} & $\pm$ & $0.016$ \\
$R=4$ & \textbf{PSNR} & $32.35$ & $\pm$ & $2.21$ & $33.19$ & $\pm$ & $2.09$ & \textbf{33.91} & $ \pm$ & $ 2.23$ \\
& \textbf{NRMSE} & $0.113$ & $\pm$ & $0.007$ & $0.106$ & $\pm$ & $0.010$ & \textbf{0.099} & $ \pm $ & $ 0.008$ \\
& \textbf{Blur} & $0.369$ & $\pm$ & $0.018$ & \textbf{0.353} & $ \pm $ & $0.017$ & $0.359$ & $\pm$ & $0.019$ \\
\hline
& \textbf{SSIM} & $0.833$ & $\pm$ & $0.038$ & $0.834$ & $\pm$ & $0.048$ & \textbf{0.915} & $ \pm $ & $ 0.018$ \\
$R=6$ & \textbf{PSNR} & $30.98$ & $\pm$ & $2.28$ & $32.28$ & $\pm$ & $2.10$ & \textbf{32.94} & $ \pm $ & $2.21$ \\
& \textbf{NRMSE} & $0.127$ & $\pm$ & $0.009$ & 0.112 & $ \pm $& $0.008$ & \textbf{0.108}& $ \pm $& $0.008$ \\
& \textbf{Blur} & $0.391$ & $\pm$ & $0.021$ & $0.369$ & $\pm$ & $0.018$ & \textbf{0.365} & $ \pm $ & $ 0.019$ \\
\hline
& \textbf{SSIM} & $0.822$ & $\pm$ & $0.035$ & $0.832$ & $\pm$ & $0.042$ & \textbf{0.902}& $ \pm $&$0.021$ \\
$R=8$ & \textbf{PSNR} & $29.95$ & $\pm$ & $2.32$ & $31.37$ & $\pm$ & $2.14$ & \textbf{32.10} & $ \pm $ & $ 2.18$ \\
& \textbf{NRMSE} & $0.141$ & $\pm$ & $0.011$ & 0.122& $ \pm $& $0.009$ & \textbf{0.117} & $ \pm $& $0.008$ \\
& \textbf{Blur} & $0.408$ & $\pm$ & $0.025$ & $0.382$ & $\pm$ & $0.020$ & \textbf{0.371}& $ \pm $& $0.020$ \\
\hline
\end{tabular}
\caption{Mean and standard deviation of the measures SSIM, PSNR and NRMSE and Blur obtained over the test set. The TV-reconstruction using the proposed spatio-temporal parameter-maps $\boldsymbol{\Lambda}_{\Theta}^{xy,t}$ improves the results especially in terms of SSIM and PSNR.}\label{tab:dyn_mri_results}
\end{table}
\begin{figure}
\centering
\begin{overpic}[width=0.96\linewidth]{figures/MRI/mri_lambda_maps/mri_lambda_maps_kz6_R6.pdf}
\put(4,18){ {\textcolor{black}{\normalsize{Zero-Filled}}}}
\put(24,18){ {\textcolor{black}{\normalsize{PDHG $\boldsymbol{\Lambda}_{\Theta}^{xy,t}$}}}}
\put(49,18){ {\textcolor{black}{\normalsize{$\boldsymbol{\Lambda}_{\Theta}^{xy}$}}}}
\put(69,18){ {\textcolor{black}{\normalsize{$\boldsymbol{\Lambda}_{\Theta}^{t}$}}}}
\put(86,18){ {\textcolor{black}{\normalsize{Target}}}}
\end{overpic}
\begin{overpic}[width=0.96\linewidth]{figures/MRI/mri_lambda_maps/mri_lambda_maps_kz12_R6.pdf}
\end{overpic}
\begin{overpic}[width=0.96\linewidth]{figures/MRI/mri_lambda_maps/mri_lambda_maps_kz28_R6.pdf}
\end{overpic}
\includegraphics[width=0.8\linewidth]{figures/MRI/mri_lambda_maps/cbar.pdf}
\caption{Different examples of reconstruction results and regularization parameter-maps for an acceleration factor of $R=6$. From left to right for each row: zero-filled reconstruction, PDHG-reconstruction with $T=4096$ obtained with the proposed CNN-based regularization spatio-temporal parameter-map $\boldsymbol{\Lambda}_{\Theta}^{xy,t}=(\boldsymbol{\Lambda}_{\Theta}^{xy}, \boldsymbol{\Lambda}_{\Theta}^{xy}, \boldsymbol{\Lambda}_{\Theta}^{t})$, spatial parameter-map $\boldsymbol{\Lambda}_{\Theta}^{xy}$, temporal parameter-map $\boldsymbol{\Lambda}_{\Theta}^{t}$ and target ground truth image. Both parameter-maps $\boldsymbol{\Lambda}_{\Theta}^{xy}$ and $\boldsymbol{\Lambda}_{\Theta}^t$ are displayed on the scale $[0,0.25]$.}
\label{fig:lambda_maps_mri}
\end{figure}
\begin{remark}
From Algorithm \ref{algo:tv_reco_algo_mri}, we see that the considered PDHG algorithm for solving problem \eqref{eq:tv_min_problem_parameter_map} involves the repeated \textit{separate} application of the forward and the adjoint operators $\Ad$ and $\Ad^{\scriptstyle \boldsymbol{\mathsf{H}}}$. Depending on the considered problem - more precisely, on the operator of the data-acquisition - this aspect can be problematic.
For example, for non-Cartesian sampling trajectories in MRI, the separate application of $\Ad^{\scriptstyle \boldsymbol{\mathsf{H}}}$ and $\Ad$ is considerably slower than the composition of $\Ad^{\scriptstyle \boldsymbol{\mathsf{H}}}\Ad$ because the latter can be efficiently approximated by the Toeplitz-kernel trick \cite{feichtinger1995efficient}, see e.g.\ \cite{mani2015fast,tamir2017t2} for applications. Thereby, the composition of the forward and the adjoint can be approximated by $\Ad^{\scriptstyle \boldsymbol{\mathsf{H}}} \Ad \approx \Fd^{\scriptstyle \boldsymbol{\mathsf{H}}} \Wd \Fd$, where $\Wd$ are Toeplitz-kernels which can be estimated depending on the sampling trajectories and $\Fd^{\scriptstyle \boldsymbol{\mathsf{H}}}$ and $\Fd$ denote efficient implementations of the FFT.
Therefore, choosing a different reconstruction algorithm that requires the application of $\Ad^{\scriptstyle \boldsymbol{\mathsf{H}}} \Ad$ rather than $\Ad^{\scriptstyle \boldsymbol{\mathsf{H}}}$ and $\Ad$ separately, e.g.\ \cite{wang2008new, goldstein2009split}, may be a viable option for non-Cartesian MRI.
\end{remark}
\subsection{Quantitative MRI Reconstruction}\label{subsec:qmri}
Here, we apply the proposed method to estimate voxel-wise regularization parameter-maps to be used in a quantitative brain MRI reconstruction problem. Similar to the previous case study, the problem consists of different decoupled 2D problems. However, the third temporal dimension contains information about the changing magnetization and thus over time, the contrast of the images changes. Moreover, the speed at which the contrast changes is voxel-depending. This suggests that, different from the previously shown dynamic MRI example, the dynamic component of the estimated regularization parameter-maps should also change over time and thus regularize each time point of the images differently.
\subsubsection{Problem Formulation}
Formally, the data-acquisition process for quantitative MRI reconstruction problems is given by
\begin{equation}\label{eq:qmri_forward_problem}
\mathbf{z} = \Ad\, q(\mathbf{u}) + \mathbf{e},
\end{equation}
where the operator $\Ad$ takes the exact form as in Subsection \ref{subsec:dyn_mri}. However, instead of acquiring the $k$-space data of a sequence of qualitative 2D images $\XX = [\XX_1, \ldots, \XX_{N_t}]^{\scriptstyle \boldsymbol{\mathsf{T}}}$ with similar image contrast, the operator $\Ad$ collects the $k$-space data of the qualitative images defined by
\begin{equation}\label{eq:q_forward_model}
\XX_t = q_t(\mathbf{u}),
\end{equation}
where $q_t:\mathbb{R}^{un} \rightarrow \mathbb{R}^n$ combines the vector containing the $u$ quantitative parameters $\mathbf{u}=[\mathbf{u}_1, \ldots, \mathbf{u}_u]^{\scriptstyle \boldsymbol{\mathsf{T}}}$ to a qualitative image by a non-linear signal-model $q_t$. In the following, we will consider the inversion recovery signal model for $T_1$-mapping given by
\begin{align}\label{eq:signal_model_qmri}
q_t: \mathbb{R}^{3 n} &\rightarrow \mathbb{C}^n \\
[\mathbf{T}_1, \operatorname{Re}(\mathbf{M}_0), \operatorname{Im}(\mathbf{M}_0) ]^{\scriptstyle \boldsymbol{\mathsf{T}}} &\mapsto q_t\left(\mathbf{T}_1, \operatorname{Re}(\mathbf{M}_0), \operatorname{Im}(\mathbf{M}_0 )\right)= \mathbf{M}_0 (1-2 e^{-t/\mathbf{T_1}}) ,
\end{align}
where the vector $\mathbf{T}_1$ denotes the longitudinal relaxation times for all pixels and $\operatorname{Re}(\mathbf{M}_0)$ and $\operatorname{Im}(\mathbf{M}_0)$ denote real and imaginary parts of the steady-state magnetization, respectively \cite{haase1990}.
Note that in quantitative MR imaging, one is ultimately interested in the quantities contained in the vector $\mathbf{u}$. However, often, qualitative images are first reconstructed (using some regularization method) as an intermediate step, from which then the vector $\mathbf{u}$ is estimated in a second step using non-linear regression methods, see for example \cite{tamir2020computational}. We can formulate the image reconstruction problem by
\begin{equation}\label{eq:qmri_recon_problem}
\left \{
\begin{aligned}
\underset{\XX}{\min}
\;\frac{1}{2} \|\Ad \XX - \mathbf{z}\|_2^2 + \| \boldsymbol{\Lambda}_{\Theta} \nabla \XX \|_1,\\
\text{ s.t. } \XX_t = q_{t_i}(\mathbf{u})\, \quad 1\leq i\leq N_t.
\end{aligned}
\right.
\end{equation}
First, we train the proposed NN to estimate appropriate pixel-dependent regularization parameter-maps $\boldsymbol{\Lambda}_{\Theta}^{xy,t}$ to solve the TV-minimization problem and in a second step, perform a pixel-wise regression to obtain the vector $\mathbf{u}$.
\subsubsection{Experimental Set-Up}
We used the BrainWeb \cite{brainweb} dataset of 20 segmented healthy human heads as a basis to generate a quantitative MRI dataset with known ground truth. The subjects were split 17/1/2 for training, validation and testing. We considered axial slices, rescaled to $192\times192$ pixels. In each axial slice, we sampled for each tissue class from uniform distributions around anatomically plausible values the complex magnetization and the longitudinal relaxation rate $R_1=1/T_1$. The phase of the magnetization was further modulated by low amplitude 2D polynomials, approximating residual phases in the acquisition model. Following the signal model \eqref{eq:signal_model_qmri}, we generated images for the inversion times 0.05\,s, 0.1\,s, 0.2\,s, 0.35\,s, 0.5\,s, 1.0\,s, 1.5\,s, 2\,s, 3\,s, 4\,s and transformed them into (undersampled) Cartesian k-space. The number of simulated receiver coils was 8. The acceleration factor $R$ was chosen from 4, 6, and 8 for comparisons. In each case, complex Gaussian noise with $\sigma$ randomly chosen from [0.04, 0.4] was added in k-space. The proposed unrolled network was used to reconstruct the (qualitative) images at different inversion time points.
Similar to Section \ref{subsec:dyn_mri}, we choose a simple 3D U-Net with two downsampling steps, two 3D convolution layers for each encoder and decoder block with LeakyReLu activation, and 8 initial filters, resulting in only 97402 parameters. We used a scaled softplus activation, $\beta \phi(x / \beta)$ with $\beta=5$, for the final activation to keep the predicted regularization strength positive. We initialized the bias of the final convolution layer with -1 (empirically chosen) to stabilize training by starting at a low regularization. We trained the network with AdamW \cite{adamw} (weight decay $10^{-4}$), cosine annealing learning rate schedule with linear warmup over one epoch with a maximum learning rate of $10^{-2}$, and a batch size of 4. The number of iterations of the unrolled PDHG is set to $T=32$ during warmup and $T=128$ for the rest of the training. Again, $\sigma$, $\tau$ and $\theta$ were trainable parameters.
Optimization was done by minimizing the MSE between the ground truth images and the obtained images after masking out non-brain regions. To find $\lambda_{\mathrm{P}}$ and $\lambda_{\tilde{\mathrm{P}}}$, grid searches, similar to those in the dynamic MRI case, were performed with a fixed number of $T=256$ iterations.
For evaluation, we used $T=256$ iterations of PDHG. We calculated the PSNR and SSIM of the reconstructed images. As a comparison method, we also performed standard iterative MRI reconstruction (CG with early stopping) without any TV regularization \cite{pruessmann2001advances}. We determined the optimal number of iterations based on the MSE to the ground truth images.
Finally, we performed a pixelwise regression on the reconstructed images $\XX$ using the Broyden–Fletcher–Goldfarb–Shanno (BFGS) algorithm, minimizing $\| |q_{t_i}(\mathbf{u})|-|\XX| \|_2^2$, to obtain $T_1$-maps and calculated the RMSE.
\subsubsection{Results}
Figure \ref{fig:qmri_y} shows examples of the quantitative (magnitude) images $\UU$ of three of the 112 simulated inversion recovery measurements in the test dataset. We also show the regularization parameter-maps for regularization along the spatial directions and along the inversion-time direction generated by the network. The mean PSNR and SSIM of our proposed method is consistently higher for all considered acceleration factors, even compared to PDHG with regularization strength along spatial and inversion-time direction chosen by grid-search with access to the ground truth images (shown in Figure \ref{fig:box_plots_qmri} and Table \ref{tab:qmri_results}).
The resulting $T_1$ parameter-maps after performing the regression on the reconstructed images are shown in Figure \ref{fig:qmri_t1}. Again, our proposed method results in the lowest RMS deviation from the ground truth images (Table \ref{tab:qmri_results}).
\begin{figure}
\hspace{0.05\textwidth}CG-SENSE\hspace{0.04\textwidth}PDHG $\lambda^{xy,t}_{\tilde{\mathrm{P}}}$\hspace{0.035\textwidth}PDHG $\lambda^{xy,t}_{\mathrm{P}}$\hspace{0.03\textwidth}PDHG $\LLambda^{xy,t}_\Theta$ \hspace{0.02\textwidth}Target/ZF \hspace{0.05\textwidth}$\LLambda^{xy}_\Theta$ / $\LLambda^{t}_\Theta$
\centering
\rotatebox[origin=c]{90}{Example 1}\,\,\begin{subfigure}[p]{\textwidth}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/34/y_cg.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/34/y_tv_grid_xy_t_constant.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/34/y_tv_grid_xy_t.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/34/y_tv_unet_acc6.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/34/y_gt.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/34/lam_xy_tv_unet_acc6.pdf}
\vskip -3pt
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/34/y_err_cg.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/34/y_err_tv_grid_xy_t_constant.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/34/y_err_tv_grid_xy_t.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/34/y_err_tv_unet_acc6.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/34/y_zf.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/34/lam_t_tv_unet_acc6.pdf}
\end{subfigure}
\vspace{0.01\textwidth}
\rotatebox[origin=c]{90}{Example 2}\,\,\begin{subfigure}[p]{\textwidth}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/20/y_cg.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/20/y_tv_grid_xy_t_constant.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/20/y_tv_grid_xy_t.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/20/y_tv_unet_acc6.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/20/y_gt.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/20/lam_xy_tv_unet_acc6.pdf}
\vskip -3pt
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/20/y_err_cg.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/20/y_err_tv_grid_xy_t_constant.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/20/y_err_tv_grid_xy_t.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/20/y_err_tv_unet_acc6.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/20/y_zf.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/20/lam_t_tv_unet_acc6.pdf}
\end{subfigure}
\vspace{0.01\textwidth}
\rotatebox[origin=c]{90}{Example 3}\,\,\begin{subfigure}[p]{\textwidth}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/58/y_cg.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/58/y_tv_grid_xy_t_constant.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/58/y_tv_grid_xy_t.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/58/y_tv_unet_acc6.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/58/y_gt.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/58/lam_xy_tv_unet_acc6.pdf}
\vskip -3pt
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/58/y_err_cg.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/58/y_err_tv_grid_xy_t_constant.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/58/y_err_tv_grid_xy_t.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/58/y_err_tv_unet_acc6.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/58/y_zf.pdf}
\includegraphics[width=0.15\textwidth]{figures/QMRI/images/acc6/58/lam_t_tv_unet_acc6.pdf}
\includegraphics[width=0.46\textwidth]{figures/QMRI/images/lam_t_colorbarH.pdf}\includegraphics[width=0.46\textwidth]{figures/QMRI/images/lam_xy_colorbarH.pdf}
\end{subfigure}
\caption{Three exemplary sets of reconstructed, qualitative magnitude images (all acceleration factor 6, $\sigma = $ 0.10 / 0.15 / 0.24 ) and absolute error compared to the ground truth (bottom rows). The last column shows spatial (top) and temporal (bottom) regularization strengths maps for the PDHG reconstruction generated by the CNN, the second to last column shows the synthetic ground truth as well as the zero-filled reconstruction (bottom rows). Qualitatively, the reconstruction was reduced by the proposed method without resulting in noticeably increased blur. The regularization parameter-maps show strong spatial regularization within areas of homogeneous tissue and increased temporal regularization at the transitions between different tissues.}
\label{fig:qmri_y}
\end{figure}
\begin{figure}
\hspace{0.16\textwidth}CG-SENSE\hspace{0.03\textwidth}PDHG $\lambda^{xy,t}_{\tilde{\mathrm{P}}}$\hspace{0.015\textwidth}PDHG $\lambda^{xy,t}_{\mathrm{P}}$\hspace{0.015\textwidth}PDHG $\LLambda^{xy,t}_\Theta$ \hspace{0.0\textwidth}Ground Truth
\centering
\def20{20}
\defacc8{acc6}
\rotatebox[origin=c]{90}{$R=6$}\,\,\begin{subfigure}[p]{0.7\textwidth}
\includegraphics[width=0.19\textwidth]{figures/QMRI/images/acc8/20/t1_cg_5.pdf}
\includegraphics[width=0.19\textwidth]{figures/QMRI/images/acc8/20/t1_tv_grid_xy_t_constant.pdf}
\includegraphics[width=0.19\textwidth]{figures/QMRI/images/acc8/20/t1_tv_grid_xy_t.pdf}
\includegraphics[width=0.19\textwidth]{figures/QMRI/images/acc8/20/t1_tv_unet_acc6.pdf}
\includegraphics[width=0.19\textwidth]{figures/QMRI/images/acc8/20/t1_gt.pdf}
\vskip -5pt
\includegraphics[width=0.19\textwidth]{figures/QMRI/images/acc8/20/t1_err_cg_5.pdf}
\includegraphics[width=0.19\textwidth]{figures/QMRI/images/acc8/20/t1_err_tv_grid_xy_t_constant.pdf}
\includegraphics[width=0.19\textwidth]{figures/QMRI/images/acc8/20/t1_err_tv_grid_xy_t.pdf}
\includegraphics[width=0.19\textwidth]{figures/QMRI/images/acc8/20/t1_err_tv_unet_acc6.pdf}
\hspace{0.19\textwidth}
\end{subfigure}
\vspace{0.005\textwidth}
\defacc8{acc8}
\rotatebox[origin=c]{90}{$R=8$}\,\,\begin{subfigure}[p]{0.7\textwidth}
\includegraphics[width=0.19\textwidth]{figures/QMRI/images/acc8/20/t1_cg_5.pdf}
\includegraphics[width=0.19\textwidth]{figures/QMRI/images/acc8/20/t1_tv_grid_xy_t_constant.pdf}
\includegraphics[width=0.19\textwidth]{figures/QMRI/images/acc8/20/t1_tv_grid_xy_t.pdf}
\includegraphics[width=0.19\textwidth]{figures/QMRI/images/acc8/20/t1_tv_unet_acc8.pdf}
\includegraphics[width=0.19\textwidth]{figures/QMRI/images/acc8/20/t1_gt.pdf}
\vskip -5pt
\includegraphics[width=0.19\textwidth]{figures/QMRI/images/acc8/20/t1_err_cg_5.pdf}
\includegraphics[width=0.19\textwidth]{figures/QMRI/images/acc8/20/t1_err_tv_grid_xy_t_constant.pdf}
\includegraphics[width=0.19\textwidth]{figures/QMRI/images/acc8/20/t1_err_tv_grid_xy_t.pdf}
\includegraphics[width=0.19\textwidth]{figures/QMRI/images/acc8/20/t1_err_tv_unet_acc8.pdf}
\includegraphics[width=0.07\textwidth]{figures/QMRI/images/acc8/20/t1_colorbar.pdf}
\includegraphics[width=0.06\textwidth]{figures/QMRI/images/acc8/20/t1_err_colorbar.pdf}
\end{subfigure}
\caption{Resulting quantitative $T_1$ parameter-maps after performing the regression on the reconstructed magnitude images and absolute errors compared to the ground truth (bottom row). Shown is \textit{Example 2} of Figure \ref{fig:qmri_y}, at acceleration 6 and 8.}
\label{fig:qmri_t1}
\end{figure}
\begin{figure}
\centering
\includegraphics{figures/QMRI/boxplots/legend.pdf}
\includegraphics[width=0.3\textwidth]{figures/QMRI/boxplots/y/psnr_image.pdf}
\includegraphics[width=0.3\textwidth]{figures/QMRI/boxplots/y/ssim_image.pdf}
\includegraphics[width=0.3\textwidth]{figures/QMRI/boxplots/t1/rmse_image.pdf}
\caption{Quantification result of the quality of the reconstructed magnitude images in terms of PSNR and SSIM, and RMSE of the $T_1$ values after performing the signal regression. We compare our proposed CNN-based generation of optimal spatial and temporally varying regularization maps for the PDHG reconstruction with a CG-SENSE reconstruction\cite{pruessmann2001advances} (early stopping for solving the normal equations), two scalar regularization values chosen on the whole test dataset by grid-search ($\lambda_{\tilde{\mathrm{P}}}^{xy,t}$), as well as two scalar regularization values that were chosen for each image of the test dataset by grid-search with access to the ground truth images ($\lambda_{\mathrm{P}}^{xy,t}$).}
\label{fig:box_plots_qmri}
\end{figure}
\begin{table}[h]
\centering
\footnotesize\rm
\begin{tabular}{r|l|rcl|rcl|rcl|rcl}
& & \multicolumn{3}{c|}{\textbf{CG-SENSE}} & \multicolumn{3}{c|}{\textbf{PDHG - $\lambda_{\tilde{\mathrm{P}}}^{xy,t}$}} & \multicolumn{3}{c|}{\textbf{PDHG - $\lambda_{\mathrm{P}}^{xy,t}$}} & \multicolumn{3}{c}{\textbf{PDHG - $\boldsymbol{\Lambda}_{\Theta}^{xy,t}$}} \\[5pt]
\hline
& \textbf{PSNR}& $24.62$&$\pm$&$3.45$& $31.89$&$\pm$&$1.70$& $32.85$&$\pm$&$2.48$& $\mathbf{34.23}$&$\pm$&$2.50$\\
$R=4$& \textbf{SSIM}& $0.654$&$\pm$&$0.095$& $0.884$&$\pm$&$0.044$& $0.902$&$\pm$&$0.042$& $\mathbf{0.930}$&$\pm$&$0.031$\\
& \textbf{RMSE} [ms]& $107$&$\pm$&$34$& $76$&$\pm$&$15$& $68$&$\pm$&$21$& $\mathbf{58}$&$\pm$&$18$\\
\hline
& \textbf{PSNR}& $24.25$&$\pm$&$2.49$& $30.19$&$\pm$&$1.58$& $30.62$&$\pm$&$1.80$& $\mathbf{32.62}$&$\pm$&$1.79$\\
$R=6$& \textbf{SSIM}& $0.639$&$\pm$&$0.077$& $0.843$&$\pm$&$0.048$& $0.859$&$\pm$&$0.044$& $\mathbf{0.914}$&$\pm$&$0.027$\\
& \textbf{RMSE} [ms]& $122$&$\pm$&$26$& $94$&$\pm$&$19$& $91$&$\pm$&$22$& $\mathbf{70}$&$\pm$&$16$\\
\hline
& \textbf{PSNR}& $23.87$&$\pm$&$1.90$& $28.70$&$\pm$&$1.42$& $28.93$&$\pm$&$1.53$& $\mathbf{31.61}$&$\pm$&$1.48$\\
$R=8$& \textbf{SSIM}& $0.623$&$\pm$&$0.063$& $0.799$&$\pm$&$0.043$& $0.810$&$\pm$&$0.043$& $\mathbf{0.897}$&$\pm$&$0.026$\\
& \textbf{RMSE} [ms]& $140$&$\pm$&$25$& $117$&$\pm$&$24$& $114$&$\pm$&$25$& $\mathbf{82}$&$\pm$&$18$\\
\hline
\end{tabular}
\caption{Mean and standard deviation of the measures PSNR and SSIM of the qualitative images and RMSE of the $T_1$ parameter-maps over the test set.}\label{tab:qmri_results}
\end{table}
\newpage
\subsection{Dynamic Image Denoising}
Here, we apply the proposed method to estimate voxel-wise dependent regularization parameter-maps to be used in a dynamic image denoising problem. An important difference to the previously considered cardiac MRI example is that, while in the latter, a clear inherent distinction between the black background and the object of interest is possible, for the next videos to be considered, this is not the case. The samples might show scenes with static camera position and only moving objects or scenes in which also the camera-position changes over time.
\subsubsection{Problem Formulation}
The real-valued noisy video samples are denoted by $\XX \in \mathbb R^n$ with $n=n_x\cdot n_y\cdot n_t$. The forward operator for the dynamic denoising problem is simply given by an $n \times n$ identity operator, i.e.\ $\Ad = \Id_{n}$.
\subsubsection{Experimental Set-Up}
For training and testing, we used video samples from the benchmark dataset for multiple object tracking \cite{MOT16}, containing both dynamic and static camera scenes. For training and validation, we scaled the resolution of the video samples by $0.5$ in each direction and extracted patches of size $n_x\times n_y\times n_t = 192 \times 192 \times 32$. During the training process, we used $1751$ patches for training and $1000$ patches from different video samples for validation. We tested the trained model on scaled resolution but the full spatial dimension with $100$ time points per test sample. Gaussian noise with a random standard deviation in the range of $\sigma = 0.1, 0.2, 0.3$ was added to the samples during training. For simplicity and increased training speed, we use a grey-scaled version of the video samples. Because the grey-scaled image data is real-valued, the CNN $u_{\Theta}$ was constructed as described in Section \ref{subsubsec:experiments:cardiac}, but with only one output-channel per output-dimension. For this example, we use the same CNN-block $u_{\Theta}$ as in Figure \ref{fig:u_Theta.pdf}. For comparison, we also trained $\lambda^{xyt}$ which holds a single value for both, the spatial and temporal dimension, and $\lambda^{xy, t}$ which holds two different values for the spatial and temporal dimension. During training we used $T=128$ network iterations, for testing we increased the number of iterations to $T=1024$. We minimized the mean squared error (MSE) between denoised and ground truth patches using the Adam optimizer \cite{kingma2014adam} with an initial learning rate of $10^{-4}$. All the training was performed for $100$ epochs, where validation was performed every second epoch.
\subsubsection{Results}
\begin{figure}[!h]
\begin{minipage}{\linewidth}
\begin{minipage}{\linewidth}
\hspace{1.6cm} PDHG $\lambda^{xyt}_{\tilde{\mathrm{P}}}$ \hspace{1.7cm} PDHG $\lambda^{xy,t}_{\tilde{\mathrm{P}}}$ \hspace{1.7cm} PDHG $\LLambda^{xy,t}_\Theta$ \hspace{1.7cm} Target/Noisy
\end{minipage}
\begin{minipage}{\linewidth}
\rotatebox{90}{
\begin{minipage}{0.4\linewidth}
\hspace{-4.2cm }Static camera \hspace{3.5cm} Moving camera
\end{minipage}
}
\begin{minipage}{\linewidth}
\centering
\resizebox{\linewidth}{!}{
\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/xyt_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/xyt_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/xy_t_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/xy_t_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/cnn_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/cnn_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/ground_truth_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/ground_truth_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
}
\resizebox{\linewidth}{!}{
\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/err_xyt_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/err_xyt_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/err_xy_t_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/err_xy_t_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/err_cnn_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/err_cnn_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/noisy_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/noisy_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
}
\resizebox{\linewidth}{!}{
\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/xyt_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/xyt_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/xy_t_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/xy_t_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/cnn_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/cnn_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/ground_truth_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/ground_truth_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
}
\resizebox{\linewidth}{!}{
\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/err_xyt_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/err_xyt_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/err_xy_t_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/err_xy_t_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/err_cnn_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/err_cnn_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/noisy_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/noisy_xy.pdf}};
\spy on (0.8, 1.5) in node [left] at (3.0, 2.5);
\end{tikzpicture}
}
\end{minipage}
\end{minipage}
\end{minipage}
\caption{An example of dynamic denoising with the PDHG from Algorithm \ref{algo:tv_reco_algo_mri} for different choices of regularization parameters at moving (\url{https://motchallenge.net/vis/MOT17-14}) and static (\url{https://motchallenge.net/vis/MOT17-01}) camera view. Single scalar regularization parameter $\lambda_{\mathrm{\tilde{P}}}^{xyt}$, two scalar regularization parameters for differently weighted spatial and temporal components $\lambda_{\mathrm{\tilde{P}}}^{xy,t}$, and the proposed spatially and temporally dependent parameter-map $\boldsymbol{\Lambda}_{\Theta}^{xy,t}$ obtained with the network $\mathcal{N}_{\Theta}^T$. The last column shows the target image and the noisy sample. The row underneath the denoised image shows the error map.}
\label{fig:dynamic_denoising:method_comparison}
\end{figure}
We compare the 2D time frames of the video samples from the test dataset to the denoised frames, regularized by $\lambda_{\tilde{\mathrm{P}}}^{xyt}$, $\lambda_{\tilde{\mathrm{P}}}^{xy,t}$ and by the spatio-temporal parameter-map $\boldsymbol{\Lambda}_{\Theta}^{xy,t}$. The metrics were calculated frame-wise for all samples at three different noise levels, characterized by the standard deviation of the Gaussian distribution. From the box-plots in Figure \ref{fig:dynamic_denoising:box_plots} we see that the PDHG reconstructions using the proposed spatio-temporal regularization parameter-map yield superior reconstructions compared to $\lambda_{\tilde{\mathrm{P}}}^{xyt}$ and $\lambda_{\tilde{\mathrm{P}}}^{xy,t}$ with respect to all measures.Table \ref{tab:dyn_denoising_results} quantitatively summarizes the results . In Figure \ref{fig:dynamic_denoising:method_comparison}, we compare two samples from the test dataset with a static camera view in the first row and a dynamic camera view in the third row. The vertical red lines in Figure \ref{fig:dynamic_denoising:method_comparison} indicate the $x$-location of the $yt$-excerpt shown to the left of each image. The second and the fourth row show the pointwise absolute errors of the respective images. For both samples, the lowest error is achieved by the $\boldsymbol{\Lambda}_{\Theta}^{xy,t}$ parameter-map. The spatial and the temporal components of the obtained regularization parameter-maps $\boldsymbol{\Lambda}_{\Theta}^{xy,t}$ are visualized in Figure \ref{fig:dynamic_denoising:parameter_maps}. Here, the noisy samples, the results obtained with PDGH using $\boldsymbol{\Lambda}_{\Theta}$, the spatially and temporally dependent $\boldsymbol{\Lambda}_{\Theta}^{xy,t}$ parameter-maps and the ground truth-images are depicted. By comparing the static and dynamic case, we see that the trained CNN is able to differentiate between the two inherently different cases. Thereby, for the video sample with the static camera position, where the background remains constant over time and only objects are changing position, the CNN imposes an overall higher temporal regularization. For the video sample where the camera position also changes over time, the CNN is able to predict the less prominent potential to exploit the temporal gradient-sparsity and thus assigns relatively low
\begin{figure}[!h]
\centering
\includegraphics[width=0.5\linewidth]{figures/DynDenoising/legend_dyn_denoising.pdf}\\
\includegraphics[width=0.325\linewidth]{figures/DynDenoising/box_plots_dynamic_nrmse.pdf}
\includegraphics[width=0.325\linewidth]{figures/DynDenoising/box_plots_dynamic_psnr.pdf}
\includegraphics[width=0.325\linewidth]{figures/DynDenoising/box_plots_dynamic_ssim.pdf}
\caption{Box-plots summarizing the results in terms of PSNR, NRMSE and SSIM obtained with the PDHG algorithm for dynamic denoising. Single scalar regularization parameter ($\lambda_{\tilde{\mathrm{P}}}^{xyt}$, blue), two scalar regularization parameters; one for the spatial $x$- and $y$-direction, one for the temporal $t$-direction ($\lambda_{\tilde{\mathrm{P}}}^{xy,t}$, orange) and the proposed spatially and temporally dependent parameter-map $\boldsymbol{\Lambda}_{\Theta}^{xy,t}$ obtained with a CNN ($\mathrm{NET}_{\Theta}$, green).}
\label{fig:dynamic_denoising:box_plots}
\end{figure}
\begin{figure}[!h]
\begin{minipage}{\linewidth}
\begin{minipage}{\linewidth}
\hspace{1.6cm} Noisy \hspace{1.4cm} PDHG $\LLambda^{xy,t}_\Theta$ \hspace{1.6cm} $\LLambda^{xy}_\Theta$ \hspace{2.2cm} $\LLambda^{t}_\Theta$ \hspace{1.8cm} Target
\end{minipage}
\begin{minipage}{\linewidth}
\rotatebox{90}{
\begin{minipage}{0.16\linewidth}
\hspace{-2.4cm }Static camera \hspace{0.1cm} Moving camera
\end{minipage}
}
\begin{minipage}{\linewidth}
\centering
\resizebox{\linewidth}{!}{
\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/noisy_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/noisy_xy.pdf}};
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/cnn_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/cnn_xy.pdf}};
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/lambda_xy_cnn_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/lambda_xy_cnn_xy.pdf}};
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/lambda_t_cnn_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/lambda_t_cnn_xy.pdf}};
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/ground_truth_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_14_sigma_02/ground_truth_xy.pdf}};
\end{tikzpicture}
}
\resizebox{\linewidth}{!}{
\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/noisy_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/noisy_xy.pdf}};
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/cnn_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/cnn_xy.pdf}};
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/lambda_xy_cnn_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/lambda_xy_cnn_xy.pdf}};
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/lambda_t_cnn_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/lambda_t_cnn_xy.pdf}};
\end{tikzpicture}
\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/ground_truth_yt.pdf}\hspace{-0.1cm}
\begin{tikzpicture}[spy using outlines={rectangle, white, magnification=2, size=1.cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[height=3cm]{figures/DynDenoising/sample_01_sigma_02/ground_truth_xy.pdf}};
\end{tikzpicture}
}
\end{minipage}
\end{minipage}
\begin{minipage}{\linewidth}
\vspace{0.2cm}
\begin{center}
$\LLambda^{xy}_\Theta$\\[0.1cm]
\includegraphics[width=0.8\linewidth]{figures/DynDenoising/cbar_005.pdf}
\end{center}
\end{minipage}
\begin{minipage}{\linewidth}
\vspace{0.2cm}
\begin{center}
$\LLambda^{t}_\Theta$\\[0.1cm]
\includegraphics[width=0.8\linewidth]{figures/DynDenoising/cbar_001.pdf}
\end{center}
\end{minipage}
\end{minipage}
\caption{Two examples with moving (top row) and static (bottom row) camera view. The columns show the noisy sample, the denoised sample, spatial and temporal dependent parameter-maps $\boldsymbol{\Lambda}_{\Theta}^{xy}$ and $\boldsymbol{\Lambda}_{\Theta}^{t}$ and the ground truth sample. By comparing the top and the bottom row, we can see that the CNN-block $\mathrm{NET}_{\Theta}$ is able to differentiate between scenes with static and dynamic camera positions as it exploits the higher temporal correlation in the first by assigning higher temporal regularization values and at the same time provides lower temporal regularization values when there is less temporal correlation to exploit due to the moving camera position. }
\label{fig:dynamic_denoising:parameter_maps}
\end{figure}
\begin{table}[h]
\centering
\footnotesize\rm
\begin{tabular}{r|r|rcl|rcl|rcl}
& & \multicolumn{3}{c}{\textbf{PDHG - $\lambda_{\tilde{\mathrm{P}}}^{xyt}$}} & \multicolumn{3}{c}{\textbf{PDHG - $\lambda_{\tilde{\mathrm{P}}}^{xy,t}$}} & \multicolumn{3}{c}{\textbf{PDHG - $\boldsymbol{\Lambda}_{\Theta}^{xy,t}$}} \\[5pt]
\hline
& \textbf{SSIM} & $0.941$ & $\pm$ & $0.017$ & $0.934$ & $\pm$ & $0.021$ & $\textbf{0.968}$ & $\pm$ & $0.011$ \\
$\sigma=0.1$ & \textbf{PSNR} & $33.26$ & $\pm$ & $1.53$ & $34.65$ & $\pm$ & $1.75$ & $\textbf{39.30}$ & $\pm$ & $2.06$ \\
& \textbf{NRMSE} & $0.044$ & $\pm$ & $0.009$ & $0.038$ & $\pm$ & $0.011$ & $\textbf{0.022}$ & $\pm$ & $0.007$ \\
\hline
& \textbf{SSIM} & $0.837$ & $\pm$ & $0.075$ & $0.914$ & $\pm$ & $0.021$ & $\textbf{0.940}$ & $\pm$ & $0.021$ \\
$\sigma=0.2$ & \textbf{PSNR} & $31.32$ & $\pm$ & $2.25$ & $33.52$ & $\pm$ & $1.63$ & $\textbf{35.537}$ & $\pm$ & $2.24$ \\
& \textbf{NRMSE} & $0.056$ & $\pm$ & $0.014$ & $0.043$ & $\pm$ & $0.011$ & $\textbf{0.035}$ & $\pm$ & $0.011$ \\
\hline
& \textbf{SSIM} & $0.649$ & $\pm$ & $0.105$ & $0.814$ & $\pm$ & $0.067$ & $\textbf{0.915}$ & $\pm$ & $0.028$ \\
$\sigma=0.3$ & \textbf{PSNR} & $28.09$ & $\pm$ & $2.54$ & $31.04$ & $\pm$ & $2.06$ & $\textbf{33.36}$ & $\pm$ & $2.33$ \\
& \textbf{NRMSE} & $0.082$ & $\pm$ & $0.024$ & $0.058$ & $\pm$ & $0.016$ & $\textbf{0.045}$ & $\pm$ & $0.014$ \\
\hline
\end{tabular}
\caption{Mean and standard deviation of the measures SSIM, PSNR and NRMSE and Blur obtained over the test set for the dynamic image denoising example. The TV-reconstruction using the proposed spatio-temporal parameter-maps $\boldsymbol{\Lambda}_{\Theta}^{xy,t}$ improves the results especially in terms of SSIM and PSNR.}\label{tab:dyn_denoising_results}
\end{table}
\subsection{Low-Dose Computerized Tomography}\label{sec:CT}
In the last section, we show an application of our proposed method to a static 2D low-dose CT reconstruction problem. Because of the different noise statistics and as a result, a different fidelity term, see \eqref{eq_CTdatafidelity} below, the problem requires the use of a reconstruction algorithm different than PDHG which shows that our proposed method can be used in conjunction with any iterative scheme. We mention however that since this fidelity term is not strongly convex, the consistency results of Section
\ref{sec:consistency_ana_scheme} cannot be applied in this case. We leave the corresponding extension of these results to the CT case for future work.
\subsubsection{Problem Formulation}
We consider the proposed NN for the low-dose Computerized Tomography (CT) setting. Here a two-dimensional parallel beam geometry is chosen and the corresponding ray transform is given by the Radon transform \cite{Radon86}. As forward operator, we then consider the discretized Radon transformation, which is a finite-dimensional linear map $\Ad\colon \mathbb R^n \to \mathbb R^m$, where $n$ is the dimension of the image space and $m$ is the product between the number of angles of the measurement and the number of the equidistant detector bins. Then we can formulate the inverse problem as
\begin{align*}
\mathbf{z}= \Ad \XX + \mathbf{e},~\text{where} ~ \mathbf{e} = -\Ad \XX - \log (\mathbf{\tilde{N}_1} / N_0) ~ \text{and} ~ \mathbf{\tilde{N}_1} \sim \text{Pois}(N_0 \exp(-\Ad \XX \mu)),
\end{align*}
where $\mu$ is a normalization constant and $N_0$ denotes the mean photon count per detector bin without attenuation. Note that here we do not have Gaussian noise, but some noise which follows the negative log-transformation of a Poisson distribution. Therefore, the data-discrepancy in \eqref{eq:tv_min_problem} is not the L2-error, and the correct term can be derived from a Bayesian viewpoint, where the data-discrepancy corresponds to the negative log-likelihood $p_{Y|X=x}$. Using that the negative log-likelihood of a Poisson distributed random variable is given by the Kullback-Leibler divergence, the resulting data-discrepancy can be written as
\begin{align} \label{eq_CTdatafidelity}
d(\Ad \XX ,\mathbf{z}) = \sum_{i=1}^m e^{-(\Ad \XX )_i \mu} N_0 - e^{-\mathbf{z}_i \mu} N_0 \big(-(\Ad \XX)_i \mu + \log(N_0) \big),
\end{align}
see e.g. \cite{ADHHMS2022,Leuschner21} for more details. Consequently, we cannot use Algorithm~\ref{algo:tv_reco_algo_mri} for reconstruction and a reformulation of Algorithm~\ref{algo:pdhg_algo} for this data-discrepancy does not lead to a closed form. This can be seen by the proximal operator of $f^*$ in line 3 of Algorithm~\ref{algo:pdhg_algo}. In our case the convex functional $f$ would be given by $f(\XX) = d(\XX,\mathbf{z})$, i.e.,
\begin{align} \label{eq:f_data_fidelity}
f(\XX) = \sum_{i=1}^m e^{-\XX_i \mu} N_0 - \tilde{\mathbf{z}} N_0 \big(-\XX_i \mu + \log(N_0) \big),
\end{align}
where we set $\tilde{\mathbf{z}} = e^{-\mathbf{z}_i \mu}$ for simplicity.
Then the convex conjugate is given by
\begin{align} \label{eq:dual_general}
f^*(\mathbf{p}) = \max_\XX \sum_{i=1}^m \XX_i \mathbf{p}_i - f(\XX) = \max_\XX \sum_{i=1}^m \XX_i \mathbf{p}_i - e^{-\XX_i \mu} N_0 - \tilde{\mathbf{z}} N_0 \XX_i \mu,
\end{align}
where we used in the second equality that $\tilde{\mathbf{z}} \log(N_0)$ is independent of $\XX$. Differentiation with respect to $\XX$ shows that for a maximizer $\hat{\XX}$ it holds
\begin{align*}
\hat{\XX}_i = -\frac{1}{\mu} \log \big(\mathbf{z}- \frac{\mathbf{z}_i}{\mu N_0} \big).
\end{align*}
Inserting this in \eqref{eq:dual_general} yields the convex conjugate of $f$
\begin{align*}
f^*(\mathbf{p}) &= \sum_{i=1}^m -\frac{1}{\mu} \log \big(\mathbf{z}- \frac{\mathbf{p}_i}{\mu N_0} \big) \mathbf{p}_i - \big(\mathbf{z}- \frac{\mathbf{p}_i}{\mu N_0} \big) + \mathbf{z}\log \big(\mathbf{z}- \frac{\mathbf{p}_i}{\mu N_0} \big) \\
&=\sum_{i=1}^m \big(\mathbf{z}- \frac{\mathbf{p}_i}{\mu} \big) \log \big(\mathbf{z}- \frac{\mathbf{p}_i}{\mu N_0} \big) -\big(\mathbf{z}- \frac{\mathbf{p}_i}{\mu N_0} \big).
\end{align*}
Then for this $f^*$ the proximal operator does not have a simple closed form.
As a remedy we consider the primal-dual algorithm PD3O \cite{Yan2018}, which is a generalization of the PDHG method. The PD3O aims to minimize the sum of proper, lower semi-continuous and convex functions
\begin{align*}
\min_\XX f(\mathbf{K}\XX) + g(\XX) + h(\XX),
\end{align*}
where $\mathbf{K}\colon V^n \to V^{\tilde{m}}$ is a bounded linear operator, $h$ is differentiable with a Lipschitz continuous gradient and for both $g$ and $f^*$ the proximal operator has a analytical solution. The general scheme of PD3O is described in Algorithm~\ref{algo:general_pd3o_algo}
\begin{algorithm}[h]
\caption{Unrolled PD3O algorithm (adapted from \cite{Yan2018})}\label{algo:general_pd3o_algo}
\begin{algorithmic}[1]
\INPUT $L = \text{Lip}(\nabla h)$,\; $\tau = 2/L$, \; $\sigma = 1/(\tau \Vert \mathbf{K} \mathbf{K}^{\scriptstyle \boldsymbol{\mathsf{T}}}\| )$, \; $\text{initial guess}~ \bar{\XX}_0$
\OUTPUT reconstructed image $\XX_{\mathrm{TV}}$
\STATE $\mathbf{p}_0 = \mathbf{\bar{x}_0}$
\STATE $\mathbf{q}_0 = \mathbf{0}$
\FOR {$k < T$ }
\STATE $\mathbf{q}_{k+1} = \mathrm{prox}_{\sigma f^{*}}(\mathbf{q}_k + \sigma \mathbf{K}\bar{x}_k)$
\STATE $\mathbf{p}_{k+1} = \mathrm{prox}_{\tau g}(\mathbf{p}_k - \tau \nabla h(\mathbf{p}_k) - \tau \mathbf{K}^{\scriptstyle \boldsymbol{\mathsf{T}}} \mathbf{q}_{k+1})$
\STATE $\bar{\XX}_{k+1} = 2\mathbf{p}_{k+1} - \mathbf{p}_k + \tau\nabla h(\mathbf{p}_k) - \tau \nabla h(\mathbf{p}_{k+1})$
\ENDFOR
\STATE $\XX_{\mathrm{TV}} = \XX_T$
\end{algorithmic}
\end{algorithm}
Note that we recover the PDHG algorithm if we set $h = 0$. For application of PD3O to our CT case we define
\begin{align*}
f(\mathbf{q}) &= \Vert \mathbf{\Lambda} \, \mathbf{q} \Vert_1, \\
g(\mathbf{p}) &= \iota_{\{\mathbf{p} \ge 0 \}} (\mathbf{p}) = \begin{cases}
0\quad &\text{if} ~\mathbf{p} \ge 0, \\
+ \infty & \text{else},
\end{cases} \\
h(\mathbf{p}) &= \sum_{i=1}^m e^{-\mathbf{p}_i \mu} N_0 - e^{-\mathbf{z}_i \mu} N_0 \big(-\mathbf{p}_i \mu + \log(N_0) \big), \\
\mathbf{K} &= \nabla.
\end{align*}
The proximal operator of $f^*$ is already given in \eqref{eq:clipping_op}, the proximal operator of $g$ is given by
\begin{align*}
\text{prox}_{\tau g}(\mathbf{z}) = \text{ReLU}(\mathbf{z}) = \begin{cases}
\mathbf{z} \quad &\text{if}~ \mathbf{z} \ge 0, \\
0 &\text{else},
\end{cases}
\end{align*}
and $\nabla h$ is given by
\begin{align*}
\nabla h(\mathbf{p}) = \mu N_0 \Ad^{\scriptstyle \boldsymbol{\mathsf{T}}} \big( - e^{-\Ad \mathbf{p} \mu} + e^{- \mathbf{z}\mu} \big).
\end{align*}
Note that $\nabla h$ is not globally Lipschitz continuous, but due to the non-negativity constraint $g$ we only have to consider $\nabla h$ for $\mathbf{p}$ with non-negative entries. Consequently, we can find an upper bound of the Lipschitz constant of $\nabla h$ by $\text{Lip}(\nabla h) \le \Vert \Ad \Vert^2 \mu^2 N_0$.
The resulting scheme we use for CT reconstruction is summarized in Algorithm~\ref{algo:tv_pd3o_algo}.
\begin{algorithm}[h]
\caption{Unrolled PD3O algorithm for general linear operator $\Ad$ with $h(\, \cdot \,) = d(\Ad \cdot \, , \mathbf{z})$, $d$ as in \eqref{eq_CTdatafidelity} and \textit{fixed} regularization parameter-map $\boldsymbol{\Lambda}$ (adapted from \cite{Yan2018})}\label{algo:tv_pd3o_algo}
\begin{algorithmic}[1]
\INPUT $L = \text{Lip}(\nabla h)$, \; $\tau = 2/L$, \; $\sigma = 1/(\tau\Vert \nabla \Vert )$, \; $\text{initial guess}~ \mathbf{\bar{x}_0}$
\OUTPUT reconstructed image $\XX_{\mathrm{TV}}$
\STATE $\mathbf{p}_0 = \mathbf{\bar{x}_0}$
\STATE $\mathbf{q}_0 = \mathbf{0}$
\FOR {$k < T$ }
\STATE $\mathbf{q}_{k+1} = \mathrm{clip}_{\boldsymbol{\Lambda}}(\mathbf{q}_k + \sigma \nabla \bar{\XX}_k)$
\STATE $\mathbf{p}_{k+1} = \text{ReLU} ( \mathbf{p}_k - \tau \mu N_0 \Ad^{\scriptstyle \boldsymbol{\mathsf{T}}} ( e^{-\mathbf{z}\mu} - e^{-\Ad\mathbf{p}_k \mu}) - \tau \nabla^{\scriptstyle \boldsymbol{\mathsf{T}}} \mathbf{q}_{k+1})$
\STATE $\bar{\XX}_{k+1} = 2 \mathbf{p}_{k+1} - \mathbf{p}_k + \tau \mu N_0 \Ad^{\scriptstyle \boldsymbol{\mathsf{T}}}(e^{-\Ad\mathbf{p}_{k+1} \mu} - e^{-\Ad\mathbf{p}_k \mu})$
\ENDFOR
\STATE $\XX_{\mathrm{TV}} = \bar{\XX}_T$
\end{algorithmic}
\end{algorithm}
\subsubsection{Experimental Set-Up}
We use the LoDoPaB dataset \cite{LoDoPaB21}\footnote{available at \url{https://zenodo.org/record/3384092##.Ylglz3VBwgM}} for low-dose CT imaging. It is based on scans of the Lung Image Database Consortium and Image Database Resource Initiative \cite{Armato11} which serve as ground truth images, while the measurements are simulated. The dataset contains 35820 training images, 3522 validation images and 3553 test images. Here the ground truth images have a resolution of $362\times 362$ on a domain of $26\text{cm}\times26\text{cm}$. We only use the first 300 training images and the first 10 validation images.
For the forward operator we consider a normalization constant $\mu = 81.35858$, the mean photon count per detector bin $N_0 = 4096$ as well as 513 equidistant detector bins and 1000 equidistant angles between 0 and $\pi$. A detailed description of the data generation process is given in \cite{LoDoPaB21}.
Following the naming convention of Figure~\ref{fig:u_Theta.pdf}, the network $u_\theta$ is a 2D U-Net\footnote{available at \url{https://jleuschn.github.io/docs.dival/_modules/dival/reconstructors/networks/unet.html}}, where the number of channels at different scales are 32, 32, 64, 64 and 128 resulting in 610673 trainable parameters. For training we use Adam optimizer \cite{kingma2014adam} with a learning rate of $10^{-4}$, a batch size of 1 and train for 50 epochs. Then we used the model configuration for which the MSE on the validation set was lowest. The number of iterations of PD3O is set to $T=512$ resulting in a training time of around 24 hours on a single NVIDIA GeForce RTX 2080 super GPU with 8 GB GPU memory. At test time, we use $T=1024$ iterations for reconstruction. The forward and the adjoint operator as well as the FBP were implemented using the publicly available library \texttt{ODL}\cite{adler2017operator}.
\subsubsection{Results}
\begin{figure}
\centering
\begin{subfigure}[t]{.24\textwidth}
\caption*{PD3O $\lambda_{\tilde{\mathrm{P}}}^{xy}$}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=3,size=2.3cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[width=\linewidth]{figures/CT/LamXY17.png}};
\spy on (1.8,.8) in node [left] at (2.33,2.49);
\end{tikzpicture}
\end{subfigure}%
\hfill
\begin{subfigure}[t]{.24\textwidth}
\caption*{PD3O $\lambda_{\mathrm{P}}^{xy}$}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=3,size=2.3cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) { \includegraphics[width=\linewidth]{figures/CT/LamXYopt17.png}};
\spy on (1.8,.8) in node [left] at (2.33,2.49);
\end{tikzpicture}
\end{subfigure}%
\hfill
\begin{subfigure}[t]{.24\textwidth}
\caption*{PD3O $\boldsymbol{\Lambda}_{\Theta}^{xy}$}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=3,size=2.3cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) { \includegraphics[width=\linewidth]{figures/CT/LamCNN17.png}};
\spy on (1.8,.8) in node [left] at (2.33,2.49);
\end{tikzpicture}
\end{subfigure}%
\hfill
\begin{subfigure}[t]{.24\textwidth}
\caption*{Target / FBP}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=3,size=2.3cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) { \includegraphics[width=\linewidth]{figures/CT/GT17.png}};
\spy on (1.8,.8) in node [left] at (2.33,2.49);
\end{tikzpicture}
\end{subfigure}%
\begin{subfigure}[t]{.24\textwidth}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=3,size=2.3cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) { \includegraphics[width=\linewidth]{figures/CT/XY17_error.png}};
\spy on (1.8,.8) in node [left] at (2.33,2.49);
\end{tikzpicture}
\end{subfigure}%
\hfill
\begin{subfigure}[t]{.24\textwidth}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=3,size=2.3cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) { \includegraphics[width=\linewidth]{figures/CT/XYopt17_error.png}};
\spy on (1.8,.8) in node [left] at (2.33,2.49);
\end{tikzpicture}
\end{subfigure}%
\hfill
\begin{subfigure}[t]{.24\textwidth}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=3,size=2.3cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) { \includegraphics[width=\linewidth]{figures/CT/CNN17_error.png}};
\spy on (1.8,.8) in node [left] at (2.33,2.49);
\end{tikzpicture}
\end{subfigure}%
\hfill
\begin{subfigure}[t]{.24\textwidth}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=3,size=2.3cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) { \includegraphics[width=\linewidth]{figures/CT/FBP17.png}};
\spy on (1.8,.8) in node [left] at (2.33,2.49);
\end{tikzpicture}
\end{subfigure}%
\caption{Reconstruction of the ground truth CT image using different choices of regularization parameters. The last column shows the ground truth and the FBP reconstruction.
\textit{Top}: full image. \textit{Bottom}: difference to the ground truth.} \label{fig:CT_img}
\end{figure}
In Figure~\ref{fig:CT_img} we compare the PD3O reconstructions (top) and their corresponding errors with respect to the ground truth (bottom) using different regularization parameter choices $\lambda_{\tilde{P}}^{xy}$, $\lambda_{P}^{xy}$ and $\boldsymbol\Lambda_{\Theta}$ for PD3O. Obviously, using the estimated parameter-map $\boldsymbol\Lambda_{\Theta}$ leads to a significant improvement of the reconstruction. In particular, sharp edges are retained, while using a constant regularizing parameter results in a significant blur. This can be also seen in Table~\ref{tab:ct_results}, where we compare the NRMSE, PSNR, SSIM and blur and evaluated on the first 100 test images of the LoDoBaP dataset. These results are visualized in Figure~\ref{fig:box_plots_CT} using box-plots. Note that the FBP seems to better than PD3O-$\lambda_{\tilde{P}}^{xy}$ in terms of the blur effect, but this can be explained by the fact that FBP reconstructions admit a lot of high-frequency artefacts leading to a small blur effect.
Further PD3O-$\boldsymbol\Lambda_{\Theta}$ reconstructions with their corresponding estimated parameter-maps $\boldsymbol\Lambda_{\Theta}$ are shown in Figure~\ref{fig:CT_img_lambdamaps}. Note that the parameter-maps are given in a logarithmic scale. As expected, the regularization is strong in constant areas and less strong on edges or finer details in order to reduce a smoothing in these regions.
\begin{figure}[!h]
\centering
\includegraphics[width=0.325\linewidth]{figures/CT/legend_ct.pdf}\\
\includegraphics[width=0.24\linewidth]{figures/CT/box_plot_NRMSE.pdf}
\includegraphics[width=0.24\linewidth]{figures/CT/box_plot_PSNR.pdf}
\includegraphics[width=0.24\linewidth]{figures/CT/box_plot_SSIM.pdf}
\includegraphics[width=0.24\linewidth]{figures/CT/box_plot_Blur.pdf}
\caption{Box-plots summarizing the reconstruction results in terms of PSNR, NRMSE, SSIM and blur effect obtained with the PD3O algorithm for a CT reconstruction problem for different choices of the regularization parameter. Filtered back-projection (black), single scalar regularization parameter ($\lambda_{\tilde{\mathrm{P}}}^{xy}$, blue) and the proposed parameter-map $\boldsymbol{\Lambda}_{\Theta}$ obtained with a CNN ($\mathrm{NET}_{\Theta}$, green).}
\label{fig:box_plots_CT}
\end{figure}
\begin{table}[h]
\centering
\footnotesize\rm
\begin{tabular}[t]{c|rcl|rcl|rcl}
& \multicolumn{3}{c}{\textbf{FBP}} & \multicolumn{3}{c}{\textbf{PD3O - $\lambda_{\tilde{\mathrm{P}}}^{xy}$}} & \multicolumn{3}{c}{\textbf{PD3O - $\LLambda_{\Theta}^{xy}$}} \\[5pt]
\hline
\textbf{PSNR} & 30.37 &$\pm$& 2.95 & 32.87 &$\pm$& 3.59 & \textbf{33.90} &$\pm$& 3.94 \\
\textbf{SSIM} & 0.739 &$\pm$& 0.141 & 0.796 &$\pm$& 0.152 & \textbf{0.809} &$\pm$& 0.157 \\
\textbf{NRMSE} & 0.101 &$\pm$& 0.028 & 0.079 &$\pm$& 0.032 & \textbf{0.071} &$\pm$& 0.033 \\
\textbf{Blur Effect} & 0.412 &$\pm$& 0.067 & 0.472 &$\pm$& 0.038 & \textbf{0.407} &$\pm$& 0.043 \\
\hline
\textbf{Training time} & & - & & & 5 h & & & 24 h \\
\textbf{Runtime} & & 0.03 s & & & 5.08 s & & & 5.08 s
\end{tabular}
\caption{Mean and standard deviation of the measures SSIM, PSNR, NRMSE and Blur effect obtained over the CT test set. The best value ist marked in bold. The TV-reconstruction using the proposed parameter-maps $\boldsymbol{\Lambda}_{\Theta}^{xy}$ improve the results in every quality measure.}\label{tab:ct_results}
\end{table}
\begin{figure}[!h]
\centering
\begin{subfigure}[t]{.24\textwidth}
\caption*{FBP}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=2,size=1.5cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) { \includegraphics[width=\linewidth]{figures/CT/FBP19.png}};
\spy on (1.75,1.) in node [left] at (1.52,2.9);
\end{tikzpicture}
\end{subfigure}%
\hfill
\begin{subfigure}[t]{.24\textwidth}
\caption*{PD3O $\boldsymbol{\Lambda}_{\Theta}^{xy}$}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=2,size=1.5cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) { \includegraphics[width=\linewidth]{figures/CT/LamCNN19.png}};
\spy on (1.75,1.) in node [left] at (1.52,2.9);
\end{tikzpicture}
\end{subfigure}%
\hfill
\begin{subfigure}[t]{.24\textwidth}
\caption*{$\boldsymbol{\Lambda_{\Theta}}^{xy}$}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=2,size=1.5cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[width=\linewidth]{figures/CT/LambdaMap19.png}};
\spy on (1.75,1.) in node [left] at (1.52,2.9);
\end{tikzpicture}
\end{subfigure}%
\hfill
\begin{subfigure}[t]{.24\textwidth}
\caption*{Target}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=2,size=1.5cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[width=\linewidth]{figures/CT/GT19.png}};
\spy on (1.75,1.) in node [left] at (1.52,2.9);
\end{tikzpicture}
\end{subfigure}%
\begin{subfigure}[t]{.24\textwidth}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=2,size=1.5cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) { \includegraphics[width=\linewidth]{figures/CT/FBP22.png}};
\spy on (1.75,.65) in node [left] at (1.52,2.9);
\end{tikzpicture}
\end{subfigure}%
\hfill
\begin{subfigure}[t]{.24\textwidth}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=2,size=1.5cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) { \includegraphics[width=\linewidth]{figures/CT/LamCNN22.png}};
\spy on (1.75,.65) in node [left] at (1.52,2.9);
\end{tikzpicture}
\end{subfigure}%
\hfill
\begin{subfigure}[t]{.24\textwidth}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=2,size=1.5cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[width=\linewidth]{figures/CT/LambdaMap22.png}};
\spy on (1.75,.65) in node [left] at (1.52,2.9);
\end{tikzpicture}
\end{subfigure}%
\hfill
\begin{subfigure}[t]{.24\textwidth}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=2,size=1.5cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[width=\linewidth]{figures/CT/GT22.png}};
\spy on (1.75,.65) in node [left] at (1.52,2.9);
\end{tikzpicture}
\end{subfigure}%
\begin{subfigure}[t]{.24\textwidth}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=1.8,size=1.5cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) { \includegraphics[width=\linewidth]{figures/CT/FBP39.png}};
\spy on (1.9,.8) in node [left] at (1.52,2.9);
\end{tikzpicture}
\end{subfigure}%
\hfill
\begin{subfigure}[t]{.24\textwidth}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=1.8,size=1.5cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) { \includegraphics[width=\linewidth]{figures/CT/LamCNN39.png}};
\spy on (1.9,.8) in node [left] at (1.52,2.9);
\end{tikzpicture}
\end{subfigure}%
\hfill
\begin{subfigure}[t]{.24\textwidth}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=1.8,size=1.5cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[width=\linewidth]{figures/CT/LambdaMap39.png}};
\spy on (1.9,.8) in node [left] at (1.52,2.9);
\end{tikzpicture}
\end{subfigure}%
\hfill
\begin{subfigure}[t]{.24\textwidth}
\begin{tikzpicture}[spy using outlines={rectangle,white,magnification=1.8,size=1.5cm, connect spies}]
\node[anchor=south west,inner sep=0] at (0,0) {\includegraphics[width=\linewidth]{figures/CT/GT39.png}};
\spy on (1.9,.8) in node [left] at (1.52,2.9);
\end{tikzpicture}
\end{subfigure}%
\includegraphics[width=0.9\linewidth]{figures/CT/cbar_ct.pdf}
\caption{Different reconstructions obtained with PD3O employing the regularization parameter-maps obtained with the proposed CNN. From left to right: initial FBP-reconstruction, PD3O-reconstruction, spatial regularization parameter-map and ground truth image. As can be seen, the network attributes higher regularization parameters to image content with smooth structures while it yields lower regularization parameters at the edges to prevent smoothing.} \label{fig:CT_img_lambdamaps}
\end{figure}
\newpage
\subsection{Choosing the Number of Iterations $T$}\label{subsec:choosing_T}
Since our proposed method to obtain the regularization parameter-map $\boldsymbol{\Lambda}_{\Theta}$ is based on unrolling an iterative algorithm as PDHG or PD3O using a fixed number of iterations $T$, questions about how to choose $T$ during training as well as at inference time are relevant.
Recall from Section \ref{sec:consistency_ana_scheme} that $\XX_T := S^T(\mathbf{z}, \LLambda)$ and $\XX^* := S^*(\mathbf{z}, \LLambda)$ denote the $T$-th iterate and the exact solution of problem \eqref{eq:tv_min_problem_spatial}, respectively.
For addressing questions about the choice of $T$ at training and testing time, here, we emphasize the dependence of the solutions $\XX_T$ and $\XX^\ast$ on the set of parameters $\Theta$, by writing $\XX_T(\Theta) := S^T(\mathbf{z}, \LLambda_{\Theta})$ and $\XX^{\ast}(\Theta) := S^\ast(\mathbf{z}, \LLambda_{\Theta})$ and by denoting $\Theta_T$ as the set of trainable parameters which is obtained by training the network which unrolls using $T$ iterations of PDHG or PD3O.
We illustrate the following considerations relying on results obtained for the dynamic cardiac MRI application shown in Subsection \ref{subsec:dyn_mri} but point out that these could be derived from the other applications examples as well.
\subsubsection{Choosing $T$ at Training Time}
Clearly, the obvious choice is to set $T$ as high as possible during training. The reason is to hope to be able to have $\XX_T(\Theta)\approx \XX^\ast(\Theta) $ and therefore to optimize $\Theta$ such that its optimal when the reconstruction algorithm given by $\mathcal{N}_{\Theta}^T$ is run until convergence. However, choosing a high $T$ increases training times as well as hardware requirements. Thus, one could on purpose choose or be forced to choose to use a lower $T$ for training and hope that the sub-network $\mathrm{NET}_{\Theta}$ is flexible enough to compensate for that.\\
Figure \ref{fig:T_training_curves} shows the validation error during training of an unrolled PDHG for the dynamic MRI example for different $T$. As can be seen, for smaller $T$, the NNs' ability to accurately reconstruct the images is clearly reduced. Further, Figure \ref{fig:lambda_maps_Theta_T} shows an example of different regularization parameter-maps which were obtained by training using a different number of iterations $T$. It shows that indeed, the obtained regularization parameter-maps vary depending on the number of iterations chosen for training, although they seem to share local features. Clearly, when $T$ is set too low, the network tends to yield higher regularization parameter-maps to try to compensate for the low number of iterations. However, from Figure \ref{fig:T_training_curves} one cannot infer whether the limited reconstruction accuracy is attributable to the too low number of iterations, a resulting sub-optimal $\Theta_T$ or combinations thereof.
\begin{figure}[h]
\centering
\includegraphics[width=0.8\linewidth]{figures/T_test_curves/legend_T8_T256.pdf}
\includegraphics[width=0.6\linewidth]{figures/T_training_curves/T_training_curves.pdf}
\caption{The validation error during training of the proposed method for the previously shown dynamic cardiac MRI example for different number of iterations $T$ which were used for PDHG. Using higher $T$ results in more accurate reconstructions.}
\label{fig:T_training_curves}
\end{figure}
\begin{figure}[!h]
\centering
\begin{overpic}[width=0.87\linewidth]{figures/T_training_curves/lambda_maps_T.pdf}
\put(-7,20){ {\textcolor{black}{\normalsize{\bf{$\boldsymbol{\Lambda}_{\Theta_T}^{xy}$ }}}}}
\put(-7,8){ {\textcolor{black}{\normalsize{\bf{$\boldsymbol{\Lambda}_{\Theta_T}^t$ }}}}}
\put(6,29){ {\textcolor{black}{\normalsize{\bf{$T=8$}}}}}
\put(22,29){ {\textcolor{black}{\normalsize{\bf{$T=16$}}}}}
\put(38,29){ {\textcolor{black}{\normalsize{\bf{$T=32$}}}}}
\put(55,29){ {\textcolor{black}{\normalsize{\bf{$T=64$}}}}}
\put(70,29){ {\textcolor{black}{\normalsize{\bf{$T=128$}}}}}
\put(86,29){ {\textcolor{black}{\normalsize{\bf{$T=256$}}}}}
\end{overpic}
\includegraphics[width=0.8\linewidth]{figures/T_training_curves/cbar_0_5.pdf}
\caption{An example of different spatio-temporal regularization parameter-maps $\boldsymbol{\Lambda}_{\Theta_T}^{xy,t}$ for the dynamic cardiac MRI example obtained by training the unrolled network $\mathcal{N}_{\Theta}^T$ using different numbers of iterations $T$. All images are shown on the scale $[0,0.5]$. Although for different $T$ the regularization parameter-maps have a similar structure - for example, the cardiac region is consistently regularized less strongly along time - for lower $T$, the values tend to be higher in general. Intuitively speaking, the network estimates higher regularization values in order to compensate for the lower $T$.}
\label{fig:lambda_maps_Theta_T}
\end{figure}
\subsubsection{Choosing $T$ at Test Time}
\begin{figure}
\centering
\resizebox{\linewidth}{!}{
\begin{overpic}[width=0.95\linewidth]{figures/T_test_curves/legend_T8_T256.pdf}
\put(14,-4){ {\textcolor{black}{\large{\bf{$R=4$}}}}}
\put(46,-4){ {\textcolor{black}{\large{\bf{$R=6$}}}}}
\put(80,-4){ {\textcolor{black}{\large{\bf{$R=8$}}}}}
\end{overpic}
}\\
\vspace{0.7cm}
\includegraphics[width=0.315\linewidth]{figures/T_test_curves/T_train_vs_T_test_R4_NRMSE.pdf}
\includegraphics[width=0.315\linewidth]{figures/T_test_curves/T_train_vs_T_test_R6_NRMSE.pdf}
\includegraphics[width=0.315\linewidth]{figures/T_test_curves/T_train_vs_T_test_R8_NRMSE.pdf}\\
\includegraphics[width=0.315\linewidth]{figures/T_test_curves/T_train_vs_T_test_R4_PSNR.pdf}
\includegraphics[width=0.315\linewidth]{figures/T_test_curves/T_train_vs_T_test_R6_PSNR.pdf}
\includegraphics[width=0.315\linewidth]{figures/T_test_curves/T_train_vs_T_test_R8_PSNR.pdf}\\
\includegraphics[width=0.315\linewidth]{figures/T_test_curves/T_train_vs_T_test_R4_SSIM.pdf}
\includegraphics[width=0.315\linewidth]{figures/T_test_curves/T_train_vs_T_test_R6_SSIM.pdf}
\includegraphics[width=0.315\linewidth]{figures/T_test_curves/T_train_vs_T_test_R8_SSIM.pdf}\\
\caption{Average NRMSE, PSNR and SSIM for different PDHG-networks $\mathcal{N}_{{\Theta}_{T_{\mathrm{train}}}}^{T_{\mathrm{test}}}$ trained and tested with different combinations of $T_{\mathrm{train}}$ and $T_{\mathrm{test}}$ shown for the accelerated dynamic cardiac MRI example for acceleration factors $R=4,6,8$. The color of the dashed lines encodes the number of iterations which were used for training, and consequently, the number of iterations used at test time for which one could expect the corresponding network to yield the best measure. We see that, however, this does not consistently hold, especially for lower $T_{\mathrm{train}}$, where choosing $T_{\mathrm{test}}>T_{\mathrm{train}}$ quite consistently improves the results up to some point with respect to all measures.}
\label{fig:T_train_vs_test_curves}
\end{figure}
Recall that at test time, the proposed method generates an input-dependent regularization parameter-map $\boldsymbol{\Lambda}_{\Theta_T}$ which is inherently dependent on the number of iterations $T$ the network was trained with. With $\boldsymbol{\Lambda}_{\Theta_T}$, we can then formulate the reconstruction problem \eqref{eq:tv_min_problem_parameter_map}. Conceptually, it might be desirable to exactly solve problem \eqref{eq:tv_min_problem_parameter_map}, i.e., to run the network $\mathcal{N}_{\Theta_T}^{T^{\prime}}$ until convergence by setting $T^{\prime}$ high enough, i.e. conceptually let\ $T^{\prime}\rightarrow \infty$. By the triangle inequality, we have
\begin{equation}\label{eq:approx_accuracy_test}
\|\XX^{\ast}(\Theta_T) - \XX_{\mathrm{true}}\|_2 \leq \| \XX^{\ast}(\Theta_T) - \XX_{T^{\prime}}(\Theta_T) \|_2 + \|\XX_{T^{\prime}}(\Theta_T)- \XX_{\mathrm{true}} \|_2,
\end{equation}
where $\| \XX^{\ast}(\Theta_T) - \XX_{T^{\prime}}(\Theta_T) \|_2 \rightarrow 0$ as $T^{\prime}\rightarrow \infty$ due to the convergence of the algorithm that the network $\mathcal{N}_{\Theta_T}^{T^{\prime}}$ implicitly defines. Note however, that the contribution of the second term is not necessarily the smallest for $T^{\prime}=T$ (see also Figure \ref{fig:T_train_vs_test_curves}, especially for lower $T_{\mathrm{test}}$), since $\Theta_T$ was obtained by training only $\Theta$ (see \eqref{eq:loss_fct}) and not $\Theta$ \textit{and} $T^{\prime}$ jointly. This means that, in general, there could exist a configuration which further improves the results, i.e.\ $\|\XX_{T}(\Theta_T)- \XX_{\mathrm{true}} \|_2 \geq \|\XX_{T^{\prime}}(\Theta_T)- \XX_{\mathrm{true}} \|_2$ for some $T^{\prime}\neq T$. This is clearly visible from Figure \ref{fig:T_train_vs_test_curves}, which shows the average NRMSE, PSNR and SSIM for different combinations of $T_{\mathrm{train}}$ and $T_{\mathrm{test}}$. Interestingly, it reveals that setting $T_{\mathrm{train}}=T_{\mathrm{test}}$ is indeed not consistently the best choice. Especially for lower $T_{\mathrm{train}}$, setting $T_{\mathrm{test}}>T_{\mathrm{train}}$ introduces further regularization and yields more accurate reconstructions. For higher $T_{\mathrm{train}}$, in contrast, we see that setting $T_{\mathrm{test}}>T_{\mathrm{train}}$ possibly introduces reconstruction errors which, however, can be entirely attributed to the regularization inherently imposed by the TV, i.e., coming from the term $\|\XX^{\ast}(\Theta_T)- \XX_{\mathrm{true}} \|_2$.\\
This means that, in general, one can view our proposed method in two different flavours. From one point of view, with the proposed method, we can generate a regularization parameter-map $\boldsymbol{\Lambda}_{\Theta_T}$ which we then use to formulate the reconstruction problem \eqref{eq:tv_min_problem_parameter_map} and which we then aim to subsequently solve \textit{exactly}, i.e., the corresponding algorithm defined by $\mathcal{N}_{\Theta_T}^{T^{\prime}}$ is run until convergence by letting $T^{\prime} \rightarrow \infty$. Thereby, at test time, we implicitly accept the inherent model-error which is made by choosing the TV-minimization as the underlying regularization method. All the results shown in the paper were generated following this strategy.
From a second, more applied perspective, we can view the proposed approach which yields regularization parameter-maps which are also tailored to the specific number of iterations the network was trained with. Therefore, assuming one was able to use a high-enough $T_{\mathrm{train}}$ for training, at test time, one simply uses $T_{\mathrm{test}}=T_{\mathrm{train}}$ or, if $T_{\mathrm{train}}$ had to be strongly compromised during training (for example, due to limited GPU-memory) one can manually adjust an appropriate $T_{\mathrm{test}}$ on a validation set to compensate for the effects seen in Figure \ref{fig:T_train_vs_test_curves}.
\section{Conclusion}\label{sec:conclusion}
We have presented a simple yet efficient data-driven approach to automatically select data/patient-adaptive spatial/spatio-temporal dependent regularization parameter-maps for the variational regularization approach focusing on TV-minimization. This constitutes a simple yet efficient and elegant way to combine variational methods with the versatility of deep learning-based approaches, yielding an entirely interpretable reconstruction algorithm which inherits all theoretical properties of the scheme the network implicitly defines. We showed consistency results of the proposed unrolled scheme and we applied the proposed method to a dynamic MRI reconstruction problem, a quantitative MRI reconstruction problem, a dynamic image denoising problem and a low-dose CT reconstruction problem. In the following, we discuss possible future research directions and we also comment on the limitations of our approach.
We can immediately identify several different components worth further investigations.
First of all, for a fixed problem formulation and choice of regularization method (i.e.\ the TV-minimization considered in this work) there exist several different reconstruction algorithms, all with their theoretical and practical advantages and limitations, see e.g.\ \cite{wang2008new, goldstein2009split, chambolle2011first, stadler}. It might be interesting to investigate whether our approach yields similar regularization maps regardless of the chosen reconstruction method and if not, to what extent they differ in.
Second, in this work, we have considered the TV-minimization as the regularization method of choice. However, also TV minimization-based methods are known to have limitations, e.g.\ in producing staircasing effects. We hypothesize that the proposed method could as well be expanded to TGV-based methods \cite{TGV} to overcome these limitations. In addition, the parameter-map learning can be applied when a combination of regularizers is considered. For example, similar to the dynamic MRI and denoising case studies, the proposed method can be used for Hyperspectral X-ray CT, where the spatial and spectral domains are regularized differently, see e.g., \cite{Warr2021,Warr2022}. Further, other regularization methods as for example Wavelet-based methods \cite{Donoho_1994, Chang_2000} could be considered as well, where instead of employing the finite differences operator $\nabla$, a Wavelet-operator $\mathbf{\Psi}$ would be the sparsity-transform of choice. Thereby, the multi-scale decomposition of the U-Net which we have used in our work also naturally fits the problem and could be utilized to estimate different parameter-maps for each different level of the Wavelet-decomposition.
Third, although we have used a plain U-Net \cite{Ronneberger2015} for the estimation of the regularization parameter-maps, there exist nowadays more sophisticated network architectures, e.g.\ transformers \cite{lu2022transformer, liang2021swinir}, which could be potentially adopted as well. Lastly, from the theoretical prospective, future work can include extension of the consistency results to stationary points instead for minimizers only as well as extension to the non-strongly convex fidelity terms in order to cover the CT case as well. It would be also interesting to investigate theoretically in what degree CNN-produced artefacts in the parameter-maps can affect or create artefacts to the corresponding reconstructions.
The main limitations of the proposed approach are the ones which are common for every unrolled NN: the large GPU-memory consumption to store intermediate results and their corresponding gradients while training on the one hand and the possibly long training times which are attributed to the need to repeatedly apply the forward and the adjoint operator during training on the other hand.
As we have seen from Figure \ref{fig:T_training_curves}, to be able to learn the regularization parameter-map with a CNN as proposed, one must be able to use a certain number of iterations $T$ for the unrolled NN to ensure that the output image of the reconstruction network has sufficiently converged to the solution of problem \eqref{eq:tv_min_problem}. How large this number needs to be depends on the considered application as well as the convergence rate of the unrolled algorithm which is used for the reconstruction.
\section*{Acknowledgments}
The authors acknowledge the support of the German Research Foundation (DFG) under Germany's Excellence Strategy – The Berlin Mathematics
Research Center MATH+ (EXC-2046/1, project ID: 390685689) as this work was initiated during the Hackathon event “Maths Meets Image”, Berlin, March 2022, which was part of the MATH+ Thematic Einstein Semester on ``Mathematics of Imaging in Real-World Challenges". This work is further supported via the MATH+ project EF3-7.
This work was funded by the UK EPSRC grants the ``Computational Collaborative Project in Synergistic Reconstruction for Biomedical Imaging" (CCP SyneRBI) EP/T026693/1; ``A Reconstruction Toolkit for Multichannel CT" EP/P02226X/1 and ``Collaborative Computational Project in tomographic imaging’' (CCPi) EP/M022498/1 and EP/T026677/1. This work made use of computational support by CoSeC, the Computational Science Centre for Research Communities, through CCP SyneRBI and CCPi.
This work is part of the Metrology for Artificial Intelligence for Medicine (M4AIM) project that is funded by the German Federal Ministry for Economic Affairs and Energy (BMWi) in the framework of the QI-Digital initiative.
\bibliographystyle{plain}
|
\section*{Supplementary Material}
|
\section{Introduction}
Remotely sensed imagery is an important means to acquire information about the Earth with wide applications in geography, land surveying, ecology, and oceanography, among others. Particularly, remote sensing (RS) scene classification is a fundamental task in RS image analytics that classifies each image into a predefined set of semantic categories. As large-scale RS image datasets become more easily accessible, it has recently been feasible to train effective Deep Convolutional Neural Networks (DCNNs) for the task.
One of the key limitations of training DCNNs for RS scene classification, however, is that a large amount of manual labeling is required, in order to train models with good generalization abilities. Commercial imaging companies are collecting hundreds of terabytes of new satellite images every day; it is virtually impossible to manually label even a fraction of the ever-expanding image data. One possible remedy to this dilemma would be the use of labels obtained through crowd-sourcing platforms. However, it would be difficult to eliminate labeling errors even with label voting or aggregation. In addition, labels obtained through other channels such as automated labeling methods or auxiliary information, such as user-generated labels could also be error-prone, as manual verification of the labels is time-consuming. In this paper, we address the problem of training DCNNs on datasets with noisy labels. Here, labels being ``noisy'' means that some of the images in the datasets are incorrectly categorized. Conventionally, training DCNNs with noisy labels would inevitably lead to a rapid degradation of model performance. From a model learning perspective, it would be ideal if we could train models that are robust to label noise. Importantly, this would suggest that the learning algorithms are more secure against deliberated attacks such as label manipulation, which is a vital consideration in the security of machine learning algorithms.
In this work, we propose a robust remote sensing scene classification method with iterative multi-view voting and entropy ranking. Our method is able to deal with noisy labels by progressively removing and rectifying labeling errors. Specifically, at each time step, we first divide the training data into disjoint parts for separate training and voting. The unanimity in the voting reveals the correctness of the labels, so that we can train a strong model with only the images with unanimous votes. In addition, we adopt entropy as an effective measure for prediction uncertainty, in order to partially recover labeling errors by ranking and selection. By performing the above steps repeatedly, the quality of the labeling and the trained models is gradually improved. The main contributions of our method are two-fold: 1) We propose a simple yet effective method for robust scene classification of remotely sensed images. The two main steps, multi-view voting and entropy ranking, are complementary and enhance each other. Specifically, multi-view voting provides a way to identify errors in labels, and entropy ranking could then partially rectify these errors based on prediction uncertainty. 2) Our method provides a general framework for robust learning, and works easily with different deep models. Also, experiments on two public datasets, WHU-RS19 and AID, demonstrate the efficacy of the proposed method.
The rest of the paper is organized as follows. Sec.~\ref{sec:related} briefly reviews recent work on deep learning-based image classification with noisy labels, as well as those for remote sensing scene classification. Sec.~\ref{sec:approach} describes the proposed method in detail, followed by experimental evaluation results in Sec.~\ref{sec:exp} and closing remarks in Sec.~\ref{sec:conclusion}.
\section{Related Work}
\label{sec:related}
Many deep learning-based methods have been proposed to solve the problem of label noise in generic image classification. For example, Wang $et\ al.$~\cite{wang2019symmetric} proved that on the ``simple'' classes, the cross-entropy learning criterion will overfit noisy labels. To address this issue, the authors proposed anti-noise reverse cross-entropy, a method of symmetric cross-entropy learning. Xia $et\ al.$~\cite{xia2020robust} approached the problem by making a distinction between critical and non-critical parameters. The effect of noisy labels is scaled down by applying different rules for these two types of parameters. Another closely related work from Kaneko $et\ al.$~\cite{kaneko2019label} proposed label-noise robust Generative Adversarial Networks (GANs), which trained a noise transition model to extract knowledge from clean labels even in the presence of noise. In addition, Huang $et\ al.$~\cite{huang2019o2u} proposed O2U-net to detect noisy labels through parameter adjustments. Specifically, the network can be transferred from the over-fitting state to the under-fitting state cyclically, during which the normalized average loss of a sample is recorded for the identification of label noise. Vahdat~\cite{vahdat2017toward} proposed a conditional random field that represents the relationship between noisy and clean labels in a semi-supervised setting. Furthermore, Cordeiro $et\ al.$~\cite{cordeiro2021longremix} divided the training data into clean and noisy sets in an unsupervised manner, followed by semi-supervised learning to minimize the empirical vicinal risk (EVR). They demonstrated the importance of the accuracy of the unsupervised classifier and the the size of the training set to minimize the EVR. Consequently, a new algorithm, termed LongReMix, was proposed. Unlike existing work, in this paper we propose to refine noisy labels in an iterative fashion by multi-view voting and entropy ranking.
In the realm of RS image classification, there have also been some recent attempts to address the problem of label noise. For instance, Kang $et\ al.$~\cite{kang2020robust} used the negative Box–Cox transformation to downplay the effect of noisy labels, and proposed a robust normalized softmax loss. Li $et\ al.$~\cite{li2020error} used multi-view CNNs to correct errors in labels, and proposed an error-tolerant deep learning method, in addition to an adaptive multi-feature collaborative representation classifier to improve classification quality. Tu $et\ al.$~\cite{tu2020robust} designed a covariance matrix representation-based noisy label model. Li $et\ al.$~\cite{li2021complementary} combined ordinary learning and complementary learning as the latter can reduce the probability of learning misinformation. Being closest to our work, Li $et\ al.$~\cite{li2019learning} proposed a robust scene classification method based on multi-view voting. Our work differs from theirs as we use entropy as an important additional predictor for label correctness. In particular, we propose a way to rectify the labels of images with non-unanimous votes, which is not possible with their method. Also, the entropy ranking method we proposed provides performance improvements in addition to multi-view voting at different label noise ratios on two public datasets.
\begin{figure}[htbp]
\centering
\includegraphics[width=\textwidth]{noisy_label_flow.pdf} \caption{Robust remote sensing scene classification via iterative multi-view voting and entropy ranking. In each iteration, the dataset with noisy labels is divided into disjoint subsets for multi-view training and voting (shown in the blue rectangle). Based on voting unanimity, the dataset is split into a strong dataset with certain labels and a weak dataset with uncertain labels. A DCNN is trained with the strong dataset and based on its prediction entropy, confident samples with lower entropy are further added into the strong dataset (shown in the gray rectangle). The above steps are carried out several times to produce a final training set. See Sec.~\ref{sec:approach} for details.}
\label{fig:overall}
\end{figure}
\section{Our Approach}
\label{sec:approach}
In this paper, we propose an approach to RS scene classification that is robust to label noise. Given an initial training set with noisy labels, we perform iterative refinement via multi-view voting and entropy ranking to produce a cleaner dataset with fewer incorrect labels. Our method is based on two key observations: 1) images with correct labels tend to receive unanimous votes when we split the original dataset into multiple subsets to perform multi-view voting, and 2) the entropy of predictions can be used as an additional predictor for label correctness. Thus, we refine our dataset in an iterative fashion with both multi-view voting and entropy ranking. The overall process is illustrated in Fig.~\ref{fig:overall}.
More specifically, we take two steps for dataset refinement in each iteration: 1) multi-view training and voting (shown in the blue rectangle in Fig.~\ref{fig:overall}), and 2) entropy ranking and selection (shown in the gray rectangle in Fig.~\ref{fig:overall}). All the cylinders in Fig.~\ref{fig:overall} represent training datasets, the green dots in the cylinders represent correctly labeled images, and the red dots represent incorrectly labeled images.
In the multi-view training and voting step, we split the training set into $n$ disjoint subsets to train $n$ DCNNs. Based on the predictions from these DCNNs, the original dataset with noisy labels can be divided into a strong dataset and a weak dataset. The samples in the strong dataset are those with consistent votes across all $n$ DCNN models, and the weak dataset contains samples with inconsistent voting results.
In the entropy ranking and selection step, a portion of noisy labels is converted into correct labels. This is done by training a DCNN on the strong dataset and evaluating its prediction entropy on the weak dataset. Here, entropy is used as an effective measure of label correctness. Samples with lower prediction entropy (with their labels converted according to the most likely class in prediction) are further added into the strong dataset, to produce a cleaner output dataset for the current iteration. After $M$ iterations, a dataset with the lowest proportion of incorrect labels will be obtained, and a DCNN is trained with this dataset to obtain a robust classification model. We summarize the main procedure of our algorithm in Algorithm 1.
The rest of this section is organized as follows. In Subsections~\ref{sec:mvv} and ~\ref{sec:er}, we describe details in the multi-view training and voting step and the entropy ranking step, respectively. Subsection~\ref{sec:ir} provides details on the iterative refinement.
\RestyleAlgo{ruled}
\begin{algorithm}[t]
\caption{The proposed algorithm with multi-view voting and entropy ranking.}
\KwIn{Dataset with noisy labels $\mathcal{D}$. Semantic classes $\mathcal{C}$.}
Iterative refinement with $M$ iterations:
\For{ $m = 1 : M$ }
{
Divide the dataset $\mathcal{D}_m$ ($\mathcal{D}_1 = \mathcal{D}$. $\mathcal{D}_m$ = $\mathcal{D}^s_{m-1}$ if $m>1$) into $n$ disjoint parts:\\
$\mathcal{S}_1\cap\mathcal{S}_2\cap\cdot\cdot\cdot\cap\mathcal{S}_n$=$\emptyset$,\ \ \
$\mathcal{D}$=$\mathcal{S}_1\cup\mathcal{S}_2\cup\cdot\cdot\cdot\cup\mathcal{S}_n$
\For{ $\mathcal{S}_j = \mathcal{S}_1 : \mathcal{S}_n$ }
{
Train a Deep Convolutional Neural Network~(DCNN) with $\mathcal{S}_j$\\
\Return $\mathit{DCNN}_j$
}
\For{ $\mathit{DCNN}_t$ = $\mathit{DCNN}_1$ : $\mathit{DCNN}_n$ }
{
\For{ $(\mathbf{x}_i,y_i)$ in $\mathcal{D}_m$ }
{
Predict the label $z_i^t = \arg \max_{c \in \mathcal{C}}p_{\mathit{DCNN}_t}(\mathbf{x}_i=c)$ for each image
}
}
\For{ $(\mathbf{x}_i,y_i)$ in $\mathcal{D}_m$ }
{
If $z_i^1$=$z_i^2$=$\cdot\cdot\cdot$=$z_i^n$:
label the image accordingly and place the image into strong dataset $ (\mathbf{x}_i, y_i) \rightarrow \mathcal{D}^s$\\
Else : place the image into weak dataset $\mathbf{x}_i \rightarrow \mathcal{D}^w$
}
Train strong ${\textit{DCNN}}_s$ with $\mathcal{D}^s$ \\
\For{ $\mathbf{x}_i$ in $\mathcal{D}^w$ }
{
Calculate the entropy of $E^w(\mathbf{x}_i)$\\
If $E^w(\mathbf{x}_i) \leq \alpha$ : predict label $\hat{y}_i$ with ${\textit{DCNN}}_s$ and place the image into strong dataset $(\mathbf{x}_i, \hat{y_i}) \rightarrow \mathcal{D}^s$ \\
}
Output dataset for the current iteration $\mathcal{D}^s_m = \mathcal{D}^s$
}
Train a final model $\textit{DCNN}_f$ with $\mathcal{D}^s_M$ \\
\Return ${DCNN}_f$
\label{alg:pseudolabel}
\end{algorithm}
\subsection{Multi-View Training and Voting}
\label{sec:mvv}
For datasets containing noisy labels, it has been shown that the negative impact of noisy labels can be effectively mitigated by splitting the dataset into several parts and training them separately. Here we follow the multi-view training and voting method outlined in~\cite{li2019learning} while removing the training of a post-hoc Support Vector Machine (SVM)~\cite{vapnik1999nature} for simplicity considerations.
More formally, denote the input RS image as $\mathbf{x} \in \mathbb{R}^{H \times W \times 3}$ and $y \in \mathcal{C}$ its corresponding semantic class label, where $\mathcal{C}=\{1 \dots C\}$ is the label space. At the beginning of training, we are given an input dataset $\mathcal{D} = \{(\mathbf{x}_i, y_i) \}_{i=1}^N$ whose labels may be partially incorrect. In addition, denote $\mathcal{D}_m$ as the input training set at the $m$-th iteration of our algorithm. The total number of iterations is $M$ (see Subsection~\ref{sec:ir}). Naturally we have $\mathcal{D}_1=\mathcal{D}$.
Apart from $\mathcal{D}_m$, the iteration number $m$ is omitted below for notation simplicity. In each iteration, we split $\mathcal{D}_m$ into $n$ disjoint parts, i.e., $\mathcal{S}_1\cap\mathcal{S}_2\cap\cdot\cdot\cdot\cap\mathcal{S}_n$=$\emptyset$,\ \ \
$\mathcal{D}_m$=$\mathcal{S}_1\cup\mathcal{S}_2\cup\cdot\cdot\cdot\cup\mathcal{S}_n$.
Next, $n$ DCNNs are trained on $\mathcal{S}_1, \mathcal{S}_2, \cdots \mathcal{S}_n$, respectively, so as to obtain $\mathit{DCNN}_1$, $\mathit{DCNN}_2$, $\cdot\cdot\cdot$, $\mathit{DCNN}_n$. We use these DCNN models to make predictions on all RS images in $\mathcal{D}_m$, and if the prediction results of all models for a given image $\mathbf{x}$ are consistent, the image is placed in the strong
dataset $\mathcal{D}^s = \{(\mathbf{x}_i,y_i)\}_{i=1}^K$. Otherwise, we remove the label corresponding to this image and place it into the weak
dataset $\mathcal{D}^w = \{\mathbf{x}_i\}_{i=K+1}^N$.
We note that, once an image is placed into the weak dataset $\mathcal{D}^w$, it shall remain
in $\mathcal{D}^w$ across iterations such that the number of images in $\mathcal{D}^s$ and
$\mathcal{D}^w$ should always add up to $N$. However, as we will show in Sec~\ref{sec:er},
the strong dataset $\mathcal{D}^s$ and consequently the quality of the model trained on $\mathcal{D}^s$ will improve over time. Therefore, images in $\mathcal{D}^w$ may be moved back to $\mathcal{D}^s$ at a later time.
In the above, $\mathbf{x}_i$ represents the $i$-th image in the strong dataset, $y_i$ represents the predicted label of the $i$-th image, $N$ represents the number of images in the entire training set.
Fig.~\ref{fig:vote} illustrates the voting process. In the top row and the bottom row are images and their corresponding labels. The trapezoidal blocks in the middle represent the trained DCNN models after splitting the dataset. The arrows between the images and the DCNNs indicates that each image will get predictions from all DCNNs, and the arrows between the DCNNs and the labels indicate the predictions by voting. If the votes are unanimous, then the image will be added into the strong dataset. Otherwise, the corresponding image will be added into the weak dataset (images shown with red border in the top row, and also with a cross by their corresponding labels in the bottom row).
\begin{figure}[htbp]
\centering
\includegraphics[width=\textwidth]{vote_flow.pdf}
\caption{Example of the multi-view voting process. All images are jointly predicted by all DCNN models. Depending on whether the votes are unanimous or non-unanimous, we sort the images into a strong dataset (images with unanimous votes) and a weak dataset (images with non-unanimous votes, marked with red borders and a cross next to the corresponding labels). Lines with different colors represent the training and prediction paths for different training images. See text for details.}
\label{fig:vote}
\end{figure}
\subsection{Entropy Ranking}
\label{sec:er}
After we obtain $\mathcal{D}^s$ with the steps outlined in the previous section, we can train a strong DCNN model $\mathit{DCNN}_s$ with $\mathcal{D}^s$. We note that this DCNN will likely be trained with cleaner labels, so that its predictions will less likely be affected by label noise. In this regard, we further evaluate the predictions of $\mathit{DCNN}_s$ on the weak dataset $\mathcal{D}^w$, in the hope that we can additionally recover some images from $\mathcal{D}^w$. Here, we use entropy as a proxy for prediction certainty, i.e., we only choose images receiving a low entropy in the prediction by $\mathit{DCNN}_s$.
Specifically, given an image from the weak dataset $\mathbf{x}_i \in \mathcal{D}^w$, denote $p(\mathbf{x}_i=c)$ as the the output probability for the $c$-th class by $\mathit{DCNN}_s$, the entropy $E^w(\mathbf{x}_i)$ can be written as:
\begin{equation}
E^w(\mathbf{x}_i)=-\sum\limits_{c \in \mathcal{C}}p(\mathbf{x}_i=c)\log p(\mathbf{x}_i=c)
\end{equation}
We then sort the entropy for all images in $\mathcal{D}^w$ in ascending order, and take a hyperparameters $\alpha$ as the entropy threshold. Specifically, we move the image from the weak dataset into the strong dataset if the prediction entropy of an image is not greater than $\alpha$. If $\alpha$ is lower, we only move images with high prediction certainty from the weak dataset to the strong dataset, potentially missing some images with correct predictions as a result. If $\alpha$ is higher, we move more images into the strong dataset, resulting in a higher chance of incorrect labels in the strong dataset. In practice, the value of $\alpha$ is chosen empirically (see discussion in Subsection~\ref{sec:id}). The process above can be written as:
\begin{equation}
\mathcal{D}^w = \mathcal{D}^w \backslash \mathbf{x}_i, \: (\mathbf{x}_i, \hat{y}_i) \rightarrow \mathcal{D}^s, \quad \textrm{if} \: E^w(\mathbf{x}_i) \leq \alpha
\end{equation}
\noindent where $\mathcal{D}^w = \mathcal{D}^w \backslash \mathbf{x}_i$ means that we remove $\mathbf{x}_i$ from $\mathcal{D}^w$ and $(\mathbf{x}_i, \hat{y}_i) \rightarrow \mathcal{D}^s$ indicates that we add $(\mathbf{x}_i, \hat{y}_i)$ into $\mathcal{D}^s$. Here, $\hat{y}_i = \arg \max_{c \in \mathcal{C}} p(\mathbf{x}_i=c)$ is the maximum a posteriori (MAP) estimate provided by $\mathit{DCNN}_s$, as the truth label $y_i$ may be noisy and hence not used.
\subsection{Iterative Refinement}
\label{sec:ir}
In practice, the multi-view voting results may be partially incorrect and that noisy labels could sometimes be included in the strong dataset $\mathcal{D}^s$. This problem can be alleviated by performing the steps in Sec.~\ref{sec:mvv} and Sec.~\ref{sec:er} in an iterative fashion and gradually improving the quality of the strong dataset. Specifically, denote $\mathcal{D}^s_{m-1}$ as the strong dataset we obtained at the end of the previous iteration. We let this strong dataset be the input dataset in the next iteration, i.e., $\mathcal{D}_{m} = \mathcal{D}^s_{m-1}$.
In our experiments, we found that the iterative refinement of $\mathcal{D}^s$ is particularly important when the proportion of noisy labels is large. We set the total number of iterations to $M=3$ empirically as it provides the largest performance gain in most cases.
\section{Experiments and Discussions}
\label{sec:exp}
In this section, we use two widely used public RS image datasets to demonstrate the efficacy of the proposed noise-tolerant robust RS scene classification method. In the following, we first describe the details of the datasets and our implementation, and then present results and discussions.
\subsection{Datasets}
\noindent \textbf{WHU-RS19 dataset~\cite{sheng2012high}.} The WHU-RS19 dataset has a total of $1,005$ images, which belong to $19$ different classes. The size of each image is 600 $\times$ 600 pixels. Following~\cite{li2019learning}, we use $50\%$ of the images for training and the remaining $50\%$ for testing.
\noindent \textbf{Aerial Image Datasets (AID) dataset~\cite{xia2017aid}.} The AID dataset has a total of $10,000$ images, which belong to $30$ different classes. The number of images per class is between $200$ and $500$, and the size of each image is 600 $\times$ 600 pixels. Again, we use $50\%$ of the images for training and the remaining $50\%$ for testing following~\cite{li2019learning}.
\subsection{Implementation Details}
\label{sec:id}
In our experiments, we introduce random errors from $10\%$ to $50\%$ (at $10\%$ interval) of the original correct labels, in order to create varying levels of label noise. We use a 50-layer ResNet~\cite{he2016deep} pretrained on ImageNet~\cite{deng2009imagenet} for all DCNNs. In the preprocessing step, all images are resized to 256 $\times$ 256 pixels.
We use Adam~\cite{kingma2014adam} as the optimizer with learning rate set to $0.001$.
Each DCNN is trained for 50 epochs with the standard cross-entropy loss.
We implement our method with PyTorch, and all the experiments were run on a single NVIDIA GeForce RTX 2080 Ti GPU.
In order to ensure fairness of the experiments, in the initial training set, the proportion of label noise in each class is consistent with the proportion of label noise in the entire training set, and the noisy labels are randomly distributed.
The hyperparameter $M$, which is the number of iterations, is chosen empirically. Specifically, we perform experiments with different noise ratios, i.e., $10\%$, $20\%$, $30\%$, $40\%$, and $50\%$. For each noise ratio, we perform $10$ different noise distribution experiments. Following~\cite{li2019learning}, two DCNN views(i.e., $n=2$) are used to validate our method, and we will try more cases in subsequent works. Finally, the average of the results is taken to choose the number of iterations $M$. We present the average accuracy results we obtained in Fig.~\ref{fig:numiter}. It is clear that $M=3$ works best and we use this value throughout our experiments. In particular, increasing $M$ beyond $3$ may lead to over-fitting to the label noise so it may even lead to a degradation of performance.
Another important hyperparameter $\alpha$, which is the entropy threshold for recovering images from the weak dataset, is chosen by grid search. We present an example in Fig.~\ref{fig:entropy} that shows a typical value of $1.5$ for $\alpha$ on the WHU-RS19 dataset.
\begin{figure}[htbp]
\centering
\begin{minipage}[t]{0.45\textwidth}
\centering
\includegraphics[width=5.5cm]{Iteratively.jpg}
\caption{Classification accuracy \textit{vs.} number of iterations $M$ on the WHU-RS19 dataset under varying label noise ratios.}
\label{fig:numiter}
\end{minipage}
\begin{minipage}[t]{0.1\textwidth}
\end{minipage}
\begin{minipage}[t]{0.45\textwidth}
\centering
\includegraphics[width=5.7cm]{the_entropy.png}
\caption{An example entropy distribution in the weak dataset on the WHU-RS19 dataset ($40\%$ noise ratio, $2$nd iteration).}
\label{fig:entropy}
\end{minipage}
\end{figure}
\subsection{Results}
We present the quantitative results of our experiments in Table~\ref{tab:exp_whu} and Table~\ref{tab:exp_aid}. Table~\ref{tab:exp_whu} shows the experimental evaluation results on the WHU-RS19 dataset. Table~\ref{tab:exp_aid} shows the experimental evaluation results on the AID dataset.
It is clear that our method, Remote Sensing Multi-View Voting and Entropy Ranking (RS-MVVER), is superior to the competing method by a clear margin. We note that while the accuracy improvements may seem small in some cases, our method provides consistent improvements at all noise ratios. In addition, even a small improvement is important for remote sensing scene classification which may involve images of critical infrastructures, and we generally want the results to be as accurate as possible. Specifically, we compare our method against two baseline methods: ResNet-50~\cite{he2016deep} and RS-ETDL~\cite{li2019learning}. We note that the latter is a strong baseline method that also uses multi-view voting for learning from noisy labels in remote sensing scene classification. The results suggest that both multi-view voting and entropy ranking can positively impact model performance.
\begin{table}[htbp]
\centering
\caption{ Overall accuracy(\%) on WHU-RS19 using different ratios of noisy labels. The average accuracy and the $95\%$ confidence interval across $10$ experiments are reported.}
\begin{tabular}{|c|c|c|c|}
\hline
Noise ratio & ResNet50 in \cite{he2016deep} & RS-ETDL in \cite{li2019learning} & RS-MVVER~(Ours)\\
\hline
10\% & 91.35$\colorpm{1.42}$ & 91.79$\colorpm{2.22}$ & {\bfseries 92.18}$\colorpm{0.72}$\\
20\% & 88.65$\colorpm{1.36}$ & 89.74$\colorpm{1.88}$ & {\bfseries 91.80}$\colorpm{1.38}$\\
30\% & 87.12$\colorpm{2.63}$ & 88.91$\colorpm{1.96}$ & {\bfseries 91.53}$\colorpm{0.31}$\\
40\% & 83.27$\colorpm{1.58}$ & 88.71$\colorpm{1.45}$ & {\bfseries 90.45}$\colorpm{1.24}$\\
50\% & 82.12$\colorpm{3.14}$ & 89.10$\colorpm{3.57}$ & {\bfseries 90.13}$\colorpm{3.09}$\\
\hline
\end{tabular}
\label{tab:exp_whu}
\end{table}
\begin{table}[htbp]
\centering
\caption{ Overall accuracy(\%) on AID using different ratios of noisy labels. The average accuracy and the $95\%$ confidence interval across $5$ experiments are reported.}
\begin{tabular}{|c|c|c|c|}
\hline
Noise ratio & ResNet50 in \cite{he2016deep} & RS-ETDL in \cite{li2019learning} & RS-MVVER~(Ours)\\
\hline
10\% & 75.45$\colorpm{0.37}$ & 75.78$\colorpm{0.32}$ & {\bfseries 76.34}$\colorpm{0.38}$\\
20\% & 71.72$\colorpm{0.31}$ & 73.93$\colorpm{0.54}$ & {\bfseries 74.60}$\colorpm{1.00}$\\
30\% & 68.89$\colorpm{0.68}$ & 71.40$\colorpm{0.85}$ & {\bfseries 72.97}$\colorpm{0.92}$\\
40\% & 66.39$\colorpm{0.39}$ & 70.34$\colorpm{0.23}$ & {\bfseries 70.46}$\colorpm{0.23}$\\
50\% & 62.37$\colorpm{0.72}$ & 65.12$\colorpm{0.56}$ & {\bfseries 66.07}$\colorpm{1.33}$\\
\hline
\end{tabular}
\label{tab:exp_aid}
\end{table}
\begin{figure}[htbp]
\centering
\includegraphics[width=\textwidth]{noisy_change_flow.pdf}
\caption{Dataset before and after $M$ rounds of iterative refinement~($40\%$ noise ratio on the WHU-RS19 dataset). Images with red borders represent the wrongly labeled ones in the original training data set, and images with green borders represent the correctly labeled ones. After $3$ rounds of iterations, we can clearly see that the noisy labels are greatly reduced.}
\label{fig:iterrefine}
\end{figure}
Finally, we present a qualitative example of the iterative refinement to the dataset. Fig.~\ref{fig:iterrefine} shows an example of the iterative refinement process of the WHU-RS19 dataset. The original dataset has $40\%$ noisy labels, and the number of noisy labels are either eliminated or greatly reduced for most classes after $3$ rounds of iterative refinement. We note that the iterative refinement is performed during training, so that it does not impact evaluation performance. In our experiments, a single training iteration takes $15$ and $60$ minutes on the WHU-RS19 and the AID dataset, respectively. This is acceptable as the training is performed offline.
\section{Conclusion}
\label{sec:conclusion}
In this paper, we propose a robust remote sensing scene classification method with multi-view voting and entropy ranking that can deal with training datasets containing noisy labels. Our method begins with splitting the original training dataset into disjoint parts and training a set of multi-view classifiers. Next, we use multi-view voting followed by entropy ranking to sort images into a strong dataset and a weak dataset. By iteratively perfoming the steps above, we can obtain a cleaner dataset with much fewer noisy labels. Experiments on two public datasets, WHU-RS19 and AID, show that our method compares favorably with competing methods. We hope that our method could provide some new insights and a better starting point for future work on this topic. In particular, we would like to explore incorporating knowledge from images in the weak dataset or additional unlabeled images into our learning objectives.
\subsubsection{Acknowledgment.}
This work is supported by NSFC (61703195), Fujian NSF (2022J011112) and Fuzhou Technology Planning Program (2021-ZD-284).
\bibliographystyle{splncs04}
|
\section{Introduction}
\label{sec:introduction}
How can we develop robots that can autonomously explore the environment, acquire knowledge, and learn skills continuously? Creating autonomous cognitive and developmental robots that can co-exist in our society has been considered an ultimate goal of cognitive and developmental robotics and artificial intelligence (AI) since the inception of these fields. Autonomous robots that can develop in the real world and collaborate with us may also be called embodied artificial general intelligence (AGI).
The recent success of artificial intelligence depends primarily on large-scale human-annotated data. However, human infants can acquire knowledge and skills from sensorimotor information through physical interactions with their environment and social interactions with others (e.g., their parents or caregivers).
Importantly, the aim is to build robots that can continuously develop through embodied interactions, their learning process must be strongly based on their own sensorimotor experiences.
This autonomous learning process that occurs throughout development is also referred to as continual or lifelong learning \cite{lungarella2003developmental,lesort2020continual,DeLange2021continual}, and is considered the foundation for the emergence of both individual and social abilities necessary for robots with adaptive and collaborative capabilities.
Recently, {\it world models} have attracted renewed attention in artificial intelligence~\cite{friston2021world,ha2018world,hafner2020mastering,wu2022daydreamer}.
Now, the term ``world'' does not indicate the objective world but rather refers to a world understood from a robot's point of view\footnote{This viewpoint may be called a robot's subjective point of view of the world. Philosophically, however, whether robots can have a ``subjective'' point of view remains controversial~\cite{kiverstein2007could}. Therefore, we describe this point of view simply as ``a robot's point of view''.}. This idea corresponds to that of \textit{Umwelt} proposed by Uexk\"{u}ll~\cite{Uexkull}. Umwelt, literally around-world, meaning environment or surroundings, refers to the self-centered world of an organism perceived through its species-specific sensors\footnote{Importantly, the relationship between Umwelt and world modeling was suggested in semiotics. Sebeok pointed out that the closest equivalent of Umwelt in English is ``model''~\cite{sebeok2001biosemiotics}. An Umwelt is created and constructed through a functional cycle, which includes 1) anticipation of a perceptual cue, 2) perception, 3) working out a relation between the perception and action (either simply executing a habit or using representation, or modeling anew), and 4) action (operation)~\cite{kull2009umwelt}.}.
Therefore, notably, the world model is different from the bird's-eye model of the world that was aimed to build in good-old-fashioned AI and criticized later~\cite{Brooks1991}\footnote{Also, notably, the world-model approach is different from behavior-based robotics~\cite{arkin1998behavior}, which does not learn world models.}.
A cognitive system learns a world model to predict its future sensory observations better and optimize its policies, also referred to as controllers. Note that although typically the term ``world model'' is used to denote the spatiotemporal dynamics of the external environment, it could also equally apply to bodily dynamics (including interoceptive signals from inside the body) and the social environment.
This world-model view entails previous ideas and results, such as the effect of behavioral feedback on sensory sampling and perceptual learning \cite{verschure2003environmentally} and the resulting acquisition of self-centered, yet efficient, representations induced by an active perception strategy on the part of an agent \cite{ognibene2014ecological}.
\textit{Predictive coding} is another related theory that recently has become more and more influential \cite{clark2013a}. It is heavily influenced by Helmholtz's early theories of perception as a process driven by learning, knowledge, and inference \cite{helmholtz1867handbuch}. Predictive coding proposes that the brain infers the external causes of sensations by continuously predicting its input through top-down signals and adapts to minimize prediction error ~\cite{Rao1999,Friston2005}. This substantiates the idea that the brain might use an adaptive world model to support perception. The free energy principle (FEP) also proposes a similar vision. It argues that our brain supports both perception (perceptual inference) and action (active inference) using a form of variational Bayesian inference; in particular, using (variational) free energy, it assesses the quality of the prediction and its conformity to prior beliefs~\cite{Parr2022}. These ideas, which are currently influential in neuroscience and cognitive science, are also used in cognitive and developmental robotics, neurorobotics~\cite{ciria2021predictive,oliver2021empirical,lanillos2021active}, and artificial intelligence to develop neurodynamics realizing adaptive behaviors and social perception~\cite{tani2016exploring}.
Although such a learning-driven world model-based approach is promising in cognitive and developmental robotics, the many applications and studies of world models tend to be limited to simulation studies or adopt an offline pretrained world model \cite{ibarz2021train}.
Meanwhile, many studies based on predictive coding have been conducted in the field of cognitive robotics and neurorobotics. However, the relationship between the world model-based approach in AI and the predictive coding-based approach in robotics has rarely been discussed in an integrated manner. We believe that clarifying the definition, relationship, current state of the art, notable research gaps in work on world models, predictive coding, free-energy principle, and active inference in the context of cognitive and developmental robotics is important for further progress in this field. Based on the current status, we elucidate the frontiers and challenges toward this holy grail in cognitive and developmental robotics.
In this survey paper, we aim to build bridges and clarify the challenges and frontiers of world models and predictive coding in cognitive robotics.
The remainder of this paper is structured as follows. Section 2 provides a working definition of each key concept. Section 3 describes prior works related to the concepts and clarifies state of the art. Section 4 describes some notable challenges.
Some additional discussion is provided in Section 5, and we conclude the work in Section 6.
\begin{figure}[tb]
\centering
\includegraphics[width=1.0\linewidth]{fig1.pdf}
\caption{Overview of challenges and relationships between topics described in this survey. A robot, similar to a human, receives \emph{sensations} $x$, infers \emph{internal states} $z$, exhibits \emph{actions} $a$, and affects \emph{causes} $E$ in the social and physical environment. The initial problem is determining the models and architecture that a robot uses to efficiently and effectively learn \emph{latent representations}. An approach to this problem uses a \emph{neuro-symbolic predictive model}, which combines neural network and symbolic models. The notion of \emph{object affordance} highlights the importance of object-centric representation learning and the coupling of action and perception. \emph{Social interaction} with other agents is also an important area of research. Developing artificial intelligence for cognitive and developmental autonomous robots based on knowledge of neuroscience, i.e., \emph{brain-inspiredworld models}, is promising. Creating \emph{cognitive architecture} and developing and sharing software frameworks for this purpose will also be an important frontier.}
\label{fig:overview}
\end{figure}
\section{Working definition}
\subsection{World model}
World models describe the {\it internal models} of an agent, which encodes how world states evolve, respond to agents' actions, and relate to a given sensory input~\cite{ha2018recurrent,friston2021world}. The term {\it world model} dates back to the beginnings of artificial intelligence and robotics ~\cite{nilsson1984shakey}. Early research in machine learning studied how an agent could independently acquire and adapt a world model to \cite{schmidhuber1990making, sutton1990integrated}. Currently, it usually refers to predictive models~\cite{hafner2019dream}, which are mainly encoded using deep neural networks.
In recent years, advancements in the studies on deep neural networks have enabled self-supervised (or unsupervised) learning\footnote{Self-supervised learning is a type of unsupervised learning that aims to accomplish a task by learning to predict or classify any part of unsupervised data from any other part. In contrast to self-supervised learning, unsupervised learning includes clustering.} of large-scale world models directly from observations (sensory inputs)~\cite{ha2018recurrent,levine2016end,van_der_Himst_2020}, and these models have been applied in various areas of artificial intelligence, including reinforcement learning (RL).
World models allow agents to perform a sample-efficient prediction of the present state of the world and enable the prediction of future states, which further enables efficient planning (equivalent to model-based reinforcement learning or control). A compact internal representation further enables planning in an efficient low-dimensional space.
The key elements of world models are \emph{prediction} and \emph{inference}\footnote{The use of these terms is sometimes incongruent in the literature on statistics and machine learning; the word {\it inference} is also used to describe prediction or substituted by other concepts, such as \emph{encoding} and \emph{decoding} in the literature on (variational) autoencoders.}. \emph{Prediction} is the probabilistic process of generating the \emph{observation} $x$ given the \emph{state} (or \emph{representation}) $z$, whereas \emph{inference} is the process of obtaining a state representation $z$ from an observation $x$ in a probabilistic manner. In real-world settings, observations $x$ are large (high-dimensional, e.g., images) and provide only partial information about the world (partial observability). At the same time, the latent representation $z$ is assumed to represent the internal state of the world.
In a static case, these can be summarized as the generative processes of probability distributions as follows.
\begin{align}
\makebox[8em][l]{Prediction:} &x\sim p(x|z) \nonumber\\
\makebox[8em][l]{Inference:} &z\sim q(z|x), \label{eq_gm}
\end{align}
where $p$ is a generative model and $q$ is an inference (or recognition) model\footnote{In classical formulations of control theory \citep{kalman1960filtering}, AI \citep{cassandra1994acting}, and probabilistic robotics \citep{thrun2005probabilistic}, the inference step is performed by exactly inverting the prediction probability $p$ using the Bayes theorem. However, this poses several computational challenges and is often intractable. For dealing with such complexity, sampling-based approximate inference can be performed using Monte Carlo methods~\cite{chen2003filtering,thrun2002particle}. In the context of world models and predictive coding, variational Bayes approaches are often preferred~\cite{Parr2022}. Variational Bayes involves the definition of an approximate inference distribution $\hat{q}$. Amortized inference allows us to approximate the $q(x)$ using a neural network, which is called an inference network, and obtain an inference model $q(x|a)$~\cite{kingma2013auto}.}. These models are considered parameterized in deep neural networks. When these models are trained simultaneously, generative approaches (e.g., variational autoencoders~\cite{kingma2013auto}) are employed~\cite{ha2018recurrent}. There are also cases where only an inference model is trained, in which case a discriminative approach (e.g., contrastive learning~\cite{laskin2020curl}) is used\footnote{However, recently, studies have also shown that contrastive learning can be interpreted as a generative approach~\cite{nakamura2022self}. Therefore, the boundary between generative and discriminative approaches is, to a certain extent, blurred.}.
In the most common conditions, the state of the environment (and agent body) and the observations evolve over time in response to the agent's actions. In such cases, the state $z$ is often assumed to satisfy Markov conditions. In turn, due to typical sensory limitations such as limited field of view or occlusions, the environment is assumed to follow a partially observable Markov decision process (POMDP)~\cite{cassandra1994acting}. That is, when the current internal state of the environment is $z_t$ (where the subscript represents a discrete time step), performing an action $a_t$ causes the internal state to transition to $z_{t+1}$ and the corresponding $x_{t+1}$ is observed. In the POMDP case, the prediction and inference models are given as follows.
\begin{align}
\makebox[12em][l]{Prediction (transition):} &z_t\sim p(z_{t}|z_{t-1}, a_{t-1}) \nonumber\\
\makebox[12em][l]{Prediction (generation):} &x_t\sim p(x_t|z_t) \nonumber\\
\makebox[12em][l]{Inference:} &z_t\sim q(z_t|x_{1:t}, a_{1:t-1}),
\end{align}
where $x_{1: t}$ denotes the set of observations from the first step ($x_1$) to step $t$ ($x_t$). To learn a state space model (SSM) on the time interval $1$ to $T$, a variational approach can be adopted that maximizes the following objective~\cite{hafner2019learning,hafner2019dream,hafner2020mastering}:
\begin{align}
&\log p(x_{1: T}|a_{1: T-1}) \nonumber\\
&\geq \sum_{t=1}^T \underbrace{\mathbb{E}_{q(z_t|x_{1: t}, a_{1: t-1})}[\log p(x_t|z_t)]}_{(Negative)\ prediction\ error} - \underbrace{\mathbb{E}_{q(z_{t-1}|x_{1: t-1}, a_{1: t-2})}[D_{\rm KL}[q(z_t|x_{1: t}, a_{1: t-1})||p(z_{t}|z_{t-1}, a_{t-1})]]}_{Regularization} \equiv \sum_{t=1}^TL_t,
\label{elbo}
\end{align}
where the first term in Eq.~\eqref{elbo} represents the prediction error (or reconstruction error) of the observation, and the second term represents the regularization for the state representation (so that the transition model and the inference model yield the same state representation).
\subsection{Predictive coding and the free-energy principle}
The original predictive coding model provided by Rao and Ballard~\cite{Rao1999} was proposed as a model of visual processing in the brain. The model assumes a hierarchically organized neural network, and top-down and bottom-up interactions at each hierarchical level are considered. In the top-down process, higher levels generate predictions about lower-level neural activities, and the lowest level generates sensory predictions. In the bottom-up process, residual errors between the predictions and actual activities (or sensory inputs) are computed and used to correct the originally generated predictions at each level. Predictive coding models learn spatial and temporal statistical regularities at each level for efficient coding and to reduce the redundancy of the predicted activity of lower levels \cite{huang2011predictive,hogendoorn2019predictive}. The main principle behind this hierarchical predictive coding cortical organization in the brain is prediction error minimization (PEM) \cite{Friston2005,friston2009predictivecoding,Friston2010}. This idea based on the principle of PEM has been extended to various cognitive processes, and this framework is usually referred to as predictive processing \cite{clark2013a,clark2015surfing,hohwy2020new}. This approach is being recently used also in machine learning to learn robust generative models of data \cite{ororbia2022neural}.
The principle of PEM can be situated within a more general principle of free-energy minimization because the amount of variational free energy, the core information measure used in the FEP, can be understood, under simplifying assumptions, as the amount of prediction error \cite{Friston2005, Friston2006b, Friston2010}.
The variational free energy $F_{t}$, which is an upper bound on the surprise $-\log p(x_t|x_{1:t-1}, a_{1:t-1})$, is the negative value of the evidence lower bound (ELBO) $L_{t}$ introduced in Eq.~\eqref{elbo} as follows.
\begin{align}
-L_{t} &= F_{t}\nonumber \\
&=\underbrace{-\mathbb{E}_{q(z_t|x_{1: t}, a_{1: t-1})}[\log p(x_t|z_t)]}_{Prediction~error}+\underbrace{\mathbb{E}_{q(z_{t-1}|x_{1: t-1}, a_{1: t-2})}[D_{\rm KL}[q(z_t|x_{1: t}, a_{1: t-1})||p(z_{t}|z_{t-1}, a_{t-1})]}_{Regularization}\nonumber \\
&=\underbrace{D_{\rm KL}[q(z_t|x_{1: t}, a_{1: t-1})||p(z_{t}|x_{1: t}, a_{1:t-1})]}_{Divergence} - \underbrace{\log p(x_t|x_{1:t-1},a_{1:t-1})}_{Evidence}\nonumber \\
&\geq \underbrace{-\log p(x_t|x_{1:t-1},a_{1: t-1})}_{Surprise}.
\label{vfe}
\end{align}
From the second line of Eq.~\eqref{vfe}, when observations are assumed to follow a Gaussian distribution with a fixed variance, minimizing the variational free energy is equivalent to minimizing the sum of mean squared errors and a regularization term.
The FEP is a mathematical formulation of how self-organizing systems, such as biological agents, brains, and cells, are able to maintain an equilibrium with their environment by means of minimizing variational free energy, or the surprise associated with sensations\footnote{In contrast to states of surprise, free-energy can be measured because it is a function of sensory input and the inferred state \cite{Friston2010}} \cite{friston2011action,friston2009predictivecoding,buckley2017a}.
In the FEP, different cognitive processes such as perception and action can be understood as different ways to minimize the variational free energy in terms of probabilistic inference called active inference \cite{Friston2010e, Friston2012f, friston2015active} as detailed in the next subsection.
\subsection{Active inference and exploration}
Active inference is a normative framework that derives from the FEP and provides a unifying account for perception, control, and learning in terms of minimization of the variational free energy in the past, present, and future. This unification is important in neuroscience as it reflects neural mechanisms and on a computational level because it offers new perspectives and the possibility of sharing algorithmic solutions between all these functions and transforming them into sophisticated robotic behaviors~\cite{lanillos2021active}.
For example, perception aims to minimize the variational free energy in the past and present by inferring the latent representations of observed sensory inputs\footnote{This perception can be regarded as a variational Bayesian version of the original predictive coding that employs a maximum a posteriori (MAP) estimation \citep{friston2009predictivecoding}.}, e.g., when an orange appears in the field of view instead of the apple as currently encoded in the internal representation, the representation state can change toward that of an orange ~\cite{Friston2005}. Conversely, actions try to minimize the variational free energy in the present by actively sampling sensory inputs, e.g., by moving the gaze away from the orange toward an apple. In addition to selecting an action in the present, agents can infer a sequence of future actions (or policy) that elicit the most plausible future states ~\cite{Friston2012f,friston2015active,kruglanski2020a,friston2012optimal} by considering the minimization of \textit{expected} free energy, as detailed below.
While active inference introduces an important perspective towards an understanding of adaptive and autonomous behaviors, an obvious behavioral imperative, the exploration-exploitation dilemma, seems in conflict with this idea because exploration, i.e., observing an uncertain aspect of the environment, would result in obtaining an unpredictable outcome \cite{friston2012free,sun2020dark}.
Indeed exploration and active perception have a central role in robot control and learning. Several tasks focus on robots’ ability to explore an unknown environment \cite{chaplot2020learning,ramakrishnan2019emergence,Ammirato2017}. Furthermore, in social contexts, unobservable factors such as others' intentions must be actively considered to allow for efficient human-robot collaboration \cite{ognibene2013towards,lee2015stare,donnarumma2017action,Wang_2020_CVPR}.
However, in \cite{friston2015active,schwartenbeck2019computational,Parr2022}, the authors showed that active inference can easily support exploratory behaviors and that it can provide an elegant formal solution for the exploration-exploitation dilemma.
In fact, we must consider that planning behaviors for an extended period of time require anticipating future data. More specifically, to infer the best action sequences (policies), one must also predict the future observations they would produce. This is realized in the active inference framework by minimizing the expected free energy over a time interval $T$.
We can express this as the sum of two terms, including \textbf{i)} the variational information gain term \cite{Denzler2002Infotheory,Sommerlade2008InfoTheory,ognibene2013towards,hafner2022DivergenceMinimization}, or \textit{epistemic} value \cite{friston2015active}, defined as the expected KL divergence between the distribution of the latent states conditioned on the expected observations $q(z_{t+1:T}|x_{t+1:T}, a_{t:T-1})$ and the prior distribution on the latent states $q(z_{t+1:T}|a_{t:T-1})$ that represents the reduction in uncertainty on the latent states $z_{t+1:T}$ provided by the expected observations $x_{t+1:T}$, and \textbf{ii)} the extrinsic or \textit{pragmatic} value $\log p(x_{t+1:T}|C)$, where $C$ denotes the agent's preferences.
This results in the following expression\footnote{Note that in the literature on active inference, a sequence of actions $a_{t:T-1}$ is referred to as a policy $\pi$. This policy is different from a policy in reinforcement learning, where it represents a statistical mapping from states to actions ($\pi(a_t|s_t)$). Using this notation of the policy $\pi$ and mean-field approximation, the general formulation of the expected free energy can be described by the following more practical formulation. $G(\pi) = -\sum_{\tau=t+1}^{T}\mathbb{E}_{q(x_{\tau}|\pi)}[D_{\rm KL}[q(z_{\tau}|x_{\tau}, \pi)||q(z_{\tau}|\pi)]]-\mathbb{E}_{q(x_{\tau}|\pi)}[\log p(x_{\tau}|C)]$, where the time step $\tau>t$ used here is a future time step.}.
\begin{align}
G(a_{t:T-1}) =-\underbrace{\mathbb{E}_{q(z_{t+1:T}, x_{t+1:T}|a_{t:T-1})}[D_{\rm KL}[q(z_{t+1:T}|x_{t+1:T}, a_{t:T-1})||q(z_{t+1:T}|a_{t:T-1})]]}_{Epistemic\ value} \nonumber \\
-\underbrace{\mathbb{E}_{q(x_{t+1:T}|a_{t:T-1})}[\log p(x_{t+1:T}|C)]}_{Pragmatic\ value}. \label{eq:expected_F}
\end{align}
The epistemic value term favors obtaining observations that disambiguate the world state such as obtaining the address for the best apple shop in town, versus observations that correspond to multiple (aliased) state such as corridors in a mall. Without the factor of variational information gain, asking the address of the shop would not be preferred to any other action that would not immediately result in obtaining an apple.
Thus, minimizing expected free energy corresponds to maximizing the sum of epistemic and pragmatic values over an extended period and defines the optimal trade-off between exploration and exploitation. The similarity between the epistemic value term in Eq.~\eqref{eq:expected_F} and the divergence term in Eq.~\eqref{vfe} with an inverted sign may be noted. This is due to the different role that observations play in expected free-energy formulation, where they comprise not observed data but expected observations. Finally, the close connection between variational free energy (Eq.~\eqref{vfe}), expected free energy (Eq.~\eqref{eq:expected_F}), used in this context to define behaviors with exploration capabilities, and ELBO (Eq.~\eqref{elbo}), used to model learning processes objectives, shows the versatility of this type of formulation, the extension and refinement of which currently a promising field of research that aims to develop an autonomous system with the ability to efficiently acquire and execute complex skills \cite{hafner2022DivergenceMinimization}. For a more advanced and detailed presentation, we refer to \cite{friston2015active,parr2019generalised,hafner2022DivergenceMinimization,Parr2022}.
Another important framework that considers behaviors as inference is planning or {\it control as inference} (CaI)~\citep{attias2003planning,toussaint2009robot,kappen2012optimal,botvinick2012planning,millidge2020relationship}. The main difference between CaI and active inference is that CaI introduces a binary optimality variable $O_t$ that represents whether an action $a_t$ in state $z_t$ is optimal (or preferred)~\cite{van2020controlled,millidge2020relationship}. If the reward for taking action $a_t$ in state $z_t$ is $r\left(z_t, a_t\right)$, the conditional distribution of the optimality variable is defined as follows.
\begin{align}
p\left(O_t=1 \mid z_t, a_t\right)\equiv\exp \left(r\left(z_t, a_t\right)\right).
\end{align}
Thus, unlike active inference, CaI can introduce the value of the reward at each time explicitly and independently of the observation's generative model\footnote{Therefore, unlike active inference, CaI does not require the assumption of POMDP.}.
CaI aims to obtain the optimal policy $p(a_t|z_t)$ for inference. If the variational inference is chosen as a solution to the intractability of exact inference (as with active inference), we seek the policy that maximizes the following ELBO\footnote{Here, MDP is assumed, i.e., state $z_t$ is an observed variable rather than a latent variable; if POMDP is assumed, inference and generative models for observations are added to this ELBO.}.
\begin{align}
\log p(O_{1:T}) \geq \mathbb{E}_{\prod_{t=1}^T p(a_t|z_t)p(z_{t}|z_{t-1},a_{t-1})}\left[\sum_{t=1}^T r\left(z_t, a_t\right)+H\left(p\left(a_t \mid z_t\right)\right)\right],
\end{align}
where $H$ represents the entropy.
This corresponds to the entropy-regularized expected reward, and reinforcement learning with this as the objective is called entropy-regularized reinforcement learning~\citep{haarnoja2018soft}.
\section{Prior works}
\subsection{World models and model-based reinforcement learning in AI and robotics}
In this section, we describe world models used in \emph{model-based reinforcement learning} in the context of artificial intelligence and robotics.
\begin{figure}[tb]
\centering
\includegraphics[width=1.0\linewidth]{worldmodels_researches.pdf}
\caption{Researches of world models. Left: Dreamer~\cite{hafner2019dream} and Dreaming~\cite{okada2021dreaming}. Right: robot control system using NewtonianVAE~\cite{Okumura2022}.
}
\label{fig:worldmodels_researches}
\end{figure}
Time-series world models conditioned on behavior have been studied for policy learning for some time.
Schmidhuber proposed learning an agent's policy (utility) via an RNN-based world model obtained by self-supervised learning~\cite{schmidhuber1990making}. Based on this idea, Ha et al. introduced a large-scale world model consisting of a VAE and an RNN that learned directly from observations (time-series images) from the external world~\cite{ha2018world}. They showed that the policies of agents trained only on this world model, which learns a game environment, can work properly in real game environments.
Since this study, research has been conducted on self-supervised learning of models of an environment directly from observations, along with ideas referred to as ``world models''. However, the authors trained spatial compression (VAE) and temporal transitions (RNN) separately; thus, the perspective of learning state representations was not considered.
Subsequently, VAE-based models that simultaneously learn time transitions and spatial compression have been proposed.
Kaiser et al. proposed a VAE-based world model with discrete latent variables designed to predict the next frame and reward from the stacked frames of the previous four steps and the current action and showed that model-based reinforcement learning using this model performed adequately in an Atari video game environment with high sample efficiency~\cite{kaiser2019model}.
One limitation of this model is that it does not include RNNs and cannot account for long-term prediction. Moreover, the measures are learned from the observation space, so the learned representation is not fully exploited.
Ke et al. showed that learning long-term transitions using a stochastic RNN-based world model contributes to high performance on tasks that require long-term prediction~\cite{ke2019learning}.
All of these models, however, are autoregressive, requiring the generation of a high-dimensional observation space every step for long-term prediction, and are unable to transition within the latent space.
Recently, models that learn transitions in latent space without requiring autoregressive generation have been widely used.
Hafner et al. introduced a recurrent state space model (RSSM) that includes RNNs in SSM and showed that it could be used for long-term prediction and model-based reinforcement learning with higher performance than model-free learning~\cite{hafner2019learning}.
While PlaNet~\cite{hafner2019learning}, the first study using RSSM, used an existing model-based planning method (cross-entropy method) for planning in the latent space\footnote{PlaNet was extended to be uncertainty-aware on the basis of Bayesian inference~\cite{okada2020planet}.}, Dreamer~\cite{hafner2019dream}, a subsequent method, explicitly modeled the policy and value function in neural networks and learned a world model through gradients in an actor-critic framework, resulting in a better performance than PlaNet.
This model has been further developed by replacing the latent variables with discrete values, which significantly outperformed model-free performance in Atari game environments (Dreamer V2~\cite{hafner2020mastering}), and by using contrastive learning instead of reconstruction, which resulted in higher performance on tasks that were difficult to reconstruct (Dreaming~\cite{okada2021dreaming}, see the left side of Fig.~\ref{fig:worldmodels_researches}). They were also combined and compared (Dreaming V2~\cite{okada2022dreamingv2}).
In terms of obtaining a good state representation for control, enforcing explicit constraints on transitions is preferable. For example, NewtonianVAE was able to form PD-controllable state space~\cite{jaques2021newtonianvae}. However, to develop such a model, what kind of state representation the world model should acquire (as a good representation for control) should be considered, which remains as yet relatively unclear (see section 4.1 for details).
These world models have been shown to be effective in learning using real robots. Okumura et al. successfully applied the NewtonianVAE to a robot and enabled it to perform a precise socket insertion task~\cite{Okumura2022}(see the right side of Fig.~\ref{fig:worldmodels_researches}). Wu et al. showed that Dreamer V2 enabled real robots to perform online learning with very high sample efficiency and performance, which includes a pipeline of acquiring data through interaction with the external world, learning a world model, and controlling the robot using the model~\cite{wu2022daydreamer}. However, all of these results are for a single environment and task, and what kinds of world models should be acquired for robot control in diverse environments and tasks remains unclear.
\subsection{Predictive coding and active inference in cognitive and developmental robotics}
In recent years, an increasing body of research has considered predictive coding models for perception and action in robotics. Recent comprehensive reviews on active inference and predictive processing in robotics can be found in~\cite{lanillos2021active,ciria2021predictive}, respectively. These ideas aim to provide a general mathematical account of behavior. Importantly, they incorporate adaptation and robustness to current methods in cognitive and developmental robotics.
Since the early works of Tani et al. using hierarchically organized RNNs~\cite{Tani2003}, a variety of methods have been proposed to exploit this idea of prediction-error-minimization or propagation. ``Higher levels'' (internal representation) generate predictions about the dynamics of the ``lower levels'' up to the sensorimotor level. Prediction errors at the sensorimotor level, given the observations, are then propagated ``upwards'' in the hierarchy correcting the internal state and thus minimizing the errors.
Extensions of Tani's approach allow multiple time scales~\cite{Yamashita2008, Nishimoto2009, Namikawa2011, Yamashita2012a} (see the left side of Fig. \ref{fig:PC_AI_robots}), stochasticity~\cite{Murata2013, Murata2015a} and stochastic latent representations~\cite{Ahmadi2019}. In particular, a precision-weighting mechanism for the PEM enabled robots to extract stochastic or fluctuating structures of temporal sensorimotor sequences and utilize the extracted structures for their action generation~\cite{Murata2015a}. Interestingly, this mechanism is related to the precision account in psychiatric disorders~\cite{lanillos2020review} (especially autism spectrum disorder~\cite{VandeCruys2014, Lawson2014}) and several works have proposed cognitive robot models based on aberrant-precision to model unusual perception and action~\cite{Idei2018, Idei2020, Idei2021a}.
\begin{figure}[tb]
\centering
\includegraphics[width=1.0\linewidth]{PC_AI_robots.pdf}
\caption{Research on predictive coding and active inference in cognitive robotics. Left: adaptation to environmental changes by prediction error minimization~\cite{Yamashita2012a}. Right: body perception and action by pixel-based deep active inference~\cite{sancaktar2020end}.
}
\label{fig:PC_AI_robots}
\end{figure}
Aside from hierarchical RNNs, active inference controllers for robotic manipulators ~\cite{pezzato2020active,meo2021multimodal} and humanoid robots~\cite{oliver2021empirical,sancaktar2020end} have also been developed based on Lanillos's initial work on predictive coding adaptive perception and learning~\cite{lanillos2018active,lanillos2018adaptive} for both low-dimensional and high-dimensional inputs (see the right side of Fig. \ref{fig:PC_AI_robots}). These methods have widespread applications such as object manipulation~\cite{Ito2006a}, imitation~\cite{Ito2006a}, language acquisition~\cite{Sugita2005}, social interaction~\cite{Chen2016a} and navigation~\cite{ccatal2021robot}.
Cognitive robots benefit from predictive coding mechanisms to infer others' actions \cite{imre2019}. The reuse of common circuits for both movement generation and action estimation seems to be a key principle in the sensorimotor organization. Recently, the authors of \cite{seker2022} proposed deep modality blending networks (DMBN) designed to create a common latent space from the multi-modal experience of a robot by blending multi-modal signals with a stochastic weighting mechanism. Using a state-of-the-art skill-encoding system referred to as Conditional Neural Movement Primitives (CNMPs) \cite{seker2019rss}, they showed that deep learning could facilitate action recognition and produce structures to sustain anatomical (mirror-like) and effect-based imitation capabilities when combined with a novel modality-blending scheme.
Current state-of-the-art research is focusing on scaling active inference in planning tasks~\cite{buckley2017a} with high-dimensional inputs~\cite{van_der_Himst_2020,tschantz2020scaling} and improving representation learning through multimodal common latent space~\cite{seker2022} or introducing structural inductive biases, such as objects~\cite{van2022object}. Whilst active inference is a promising framework for robotics~\cite{da2022active}, current works are still limited to a particular aspect of cognitive and developmental processes. Therefore, in addition to extending the scalability of computational frameworks, continual or lifelong learning for developing abilities from low-level sensorimotor skills to higher-order cognitive functions should also be considered.
\section{Frontiers and Challenges}
\subsection{Latent representations for action planning}\label{sec:latent-rep}
\label{sec:representationsForAction}
One of the most important challenges in world-model approaches of any kind is that of efficiently performing planning, in the sense of generating meaningful actions to solve a sequential task~\cite{valenzo2022grounding}. Working in the high-dimensional space of the sensorimotor manifold is very computationally expensive and provides local optima solutions~\cite{sancaktar2020end}. In fact, current approaches in planning use a compressed encoded representation of the world dynamics, which aids in the process of predicting future states and in action generation~\cite{ha2018recurrent}. In reinforcement learning, state representation learning is tied to learned tasks to achieve high performance because it depends on the actions needed to obtain the maximum expected reward~\cite{hafner2019dream}. However, this sometimes prevents generalization across tasks. Decoupled action-representation world models are an interesting work-around~\cite{laskin2020curl}. In deep active inference~\cite{lanillos2021active} amortized methods have also been considered~\cite{van_der_Himst_2020,fountas2020deep}, in addition to contrastive~\cite{mazzaglia2021contrastive} and iterative amortized inference approaches~\cite{van2022object}.
However, the key question cannot be narrowed down to that what type of architecture or method should be used. Rather, what type of information should be encoded in the latent representation and how this information is processed must be a key focus so that information is not uncoupled from the sensorimotor process, particularly from motor control, which is a key limitation of existing endeavors in robotics.
There has been considerable discussion as to what would comprise an appropriate state representation of a world model; that is, what inductive bias or prior knowledge should be given~\cite{bohmer2015autonomous, achille2018separation, lesort2018state}. Here, we list the properties of this prior knowledge we consider important.
\begin{itemize}
\item {\bf Low dimensionality.} Observations obtained from the environment are high-dimensional, and compressing this information into a low-dimensional space is critical for efficient data handling, abstraction, and planning. This approach is the most frequently considered in state representation learning. The challenge is how best to represent observations in a low-dimensional encoding while retaining the necessary task-dependant information. Recent literature focused on generative and discriminative approaches to tackle this.
\item {\bf Meaningful abstraction and disentanglement.} Low-dimensional representations should have scene-understanding and task meaning, such as objects~\cite{van2022object}, locations~\cite{stoianov2022hippocampal} and temporal events~\cite{gumbsch2021sparsely}. Representation disentanglement proposes that factors of variation with different semantics should be separated, contributing to the requirements for sufficiency and efficiency in state representation. Object-centric representation learning is related to this hypothesis, \cite{greff2019multi,van2022object}, in which every observed object is encoded independently.
\item {\bf Compositionality.} Although disentanglement aims to separate independent factors, the agent should also acquire their relationships and hierarchy. In the case of object representations, there should also be relations or implication relations among objects. Currently, methods such as those using graph neural networks are being considered, but they do not provide an essential solution. This idea of the compositionality of representation is also relevant to the neuro-symbolic approach.
\item {\bf Dynamics prediction.} These three properties are important not only for learning representations in static environments but in dynamic worlds, e.g., they consider a transition model that depends on external factors and agent actions. The best latent representation is one that allows transitions to be easily predictable for given actions. Many recent models use RNNs to learn transitions, which incorporate information on long-term dependence~\cite{hafner2019learning,hafner2019dream}. One way to make transitions more predictable is to incorporate prior knowledge of the physical world (e.g., dynamics following Newton's laws of motion~\cite{jaques2021newtonianvae}). Furthermore, by learning to separate representations that are not related to control from state representations, representations that are easier to control can be acquired~\cite{wang2022denoised}.
\item {\bf Values are sufficiently encoded.} To perform reinforcement learning on the state representation of the world model, the value of the state representative to the agent must be known. For example, a recurrent state representation learns to predict the reward from the state so that the reward is embedded in the representation~\cite{hafner2019dream}. However, because the value of the state changes depending on the task, it remains unclear whether this hypothesis should be introduced in a world model that should acquire a prediction model that is as task-independent as possible. Alternatively, in active inference approaches, the agent value function cannot be modified, and it is defined by expected free energy. Here, the challenge becomes learning the state preferences and being able to predict the transitions that may yield those preferences.
\item {\bf Task-agnostic.} Representations should be informative to solve narrow problems where the agent is trained but also sufficiently general to be reused in tasks with different kinds of variability or new tasks that the agent has never encountered.
\item{\bf Fusion of multiple-types multimodal information.} Robots inevitably face a variety of events with their multimodal sensorimotor systems. Observations given to world models are from multiple sources (e.g., social non-social, sensorimotor purely sensorial, and linguistic and non-linguistic). They can have different reliability and volatility and represent various aspects of the world. Therefore, the world models must properly encode the internal representation in a stable and efficient manner.
\end{itemize}
These are some of the elements that we identified that a latent representation should be fulfilled to provide a smooth connection with real-world interaction and provide power for solving cognitive tasks. Importantly, abstract representation and disentanglement, such as objects or events encoding, may be important to achieve efficient planning, reducing the gap for neuro-symbolic solutions. However, the connection between the low-dimensional (and hierarchical) encoding and the synchronization with the sensorimotor control remains a major challenge.
\subsection{Neuro-symbolic predictive models}\label{sec:neuro-symbolic}
In this section, we provide an overview of state-of-the-art techniques in which symbols and rules are discovered and used by robots through neuro-symbolic approaches. The term {\it symbols}, here, refers to manipulative discrete representations used in symbolic AI and cognitive science. The neuro-symbolic approach attempts to integrate conventional symbolic and modern neural network-based AIs.
Both biological and artificial agents benefit from predictive coding mechanisms for reasoning, decision-making, and planning. Predictive forward models are used to generate plans that involve a sequence of actions. For example, chimpanzees are known to generate multi-step plans that include stacking a number of boxes on top of each other, grabbing a long stick, climbing on top of a stack of boxes, and using the stick to reach the object that was initially out of reach \cite{Kohler1927,Steedman2002}. While the underlying cognitive mechanisms for high-level planning remain unknown, different specific brain regions have been shown to become active in inductive and deductive reasoning in humans \cite{Goel1998} while predicting the effects of actions \cite{Al2014}. In artificial agents, on the other hand, standard search and planning rely heavily on manually coded or learned state transitions and prediction models \cite[Ch.\ 3–6,10-11]{Russell2016}.
The seminal work of \cite{mugan2012autonomous} addressed the learning of discrete representations of predictive models, i.e., dynamic Bayesian networks, by discretizing the continuous features of the environment to plan goal-directed arm/hand control. \cite{ahmetoglu2022high} showed the units generated by slow feature analysis with the lowest eigenvalues resemble symbolic representations that highly correlate with high-level features, which might be considered precursors for fully symbolic systems. \cite{Konidaris2014,Konidaris2015} studied methods to discover useful symbols that can be directly utilized in problem and domain definition language (PDDL) for various agent settings. In simulation and the real world, the discovered symbols were directly used as predicates in the action descriptions to generate deterministic and probabilistic symbolic plans. \cite{james2019learning} learned symbols in the ego-centric frame of the agent to transfer the learned symbols into novel settings. \cite{Ugur-2015-ICRA,Ugur-2015-Humanoids} discovered symbols in the continuous perceptual space of the robots for PDDL-based manipulation planning via combining several machine learning algorithms such as X-means clustering and support vector machine (SVM) classification. Although the symbols were discovered by the robot without any human intervention, the continuous perceptual features were manually encoded by the authors. Towards an end-to-end framework, \cite{Ahmetoglu2022} used directly raw camera image and pixel values to discover symbols via a novel deep predictive coding neural architecture. In detail, they proposed a deep encoder-decoder network with a binary bottleneck layer designed to take a camera image and an action as input and output the action effects in pixel coordinates. The binary activations in the bottleneck layer encode object symbols that not only depend on the visual input but are also shaped based on action and effect. In other words, the objects that provide the same affordances \cite{Sahin2007} were automatically grouped together as object symbols. To distill the knowledge represented by the neural network into rules useful for symbolic reasoning, a decision tree was trained to reproduce its decoder function. Probabilistic rules were extracted from the effect predictor / neural decoder and encoded in the probabilistic PDDL, which can be directly used by the off-the-shelf AI planners. In follow-up work, \cite{Ahmetoglu2023} used a multi-head attention mechanism to learn symbols to encode affordances of a varying number of objects.
Asai et al. implemented a neural framework where a state autoencoder with a discrete bottleneck layer was trained first, and preconditions and effects of actions were learned next~\cite{asai2017classical}. In follow-up work, \cite{asai2020learning} combined the previous two systems and discovered action preconditions and effects together with visual symbols. These works were realized in visual environments such as 2-D puzzles and achieved visualized plan executions. An important aspect of pure visual neuro-symbolic systems and studies on neuro-symbolic robotics is that in robotics, predictive coding over object symbols takes actions and effects into account in addition to the features of objects and the environment, which facilitates the formation of symbols that are likely to capture object affordances \cite{Gibson1979,Zech2017,jamone16aff,renaudo2022computational}.
However, in the context of world models, methods to integrate prior symbolic knowledge into VAE and SSM-based world models are still being explored. The bottom-up formation of symbolic representations is closely related to disentanglement and compositionality discussed in Section 4.1. Moreover, leveraging linguistic knowledge in world modeling is a challenge in relation to neuro-symbolic predictive models.
\subsection{Affordance perception}\label{sec:affordance}
Affordance perception has often been discussed independently of world models, but in fact, it is closely related. According to the original definition provided by Gibson~\cite{gibson66aff,gibson77aff,Gibson1979}, an affordance is an action possibility offered to the agent by the environment. A stable surface may afford to be traversed; a stone may afford the possibility of being used as a hammer; a door handle may afford the possibility of being opened. The concept of affordances and affordance perception has then been further analyzed and revisited in psychology, neuroscience, cognitive science, artificial intelligence, and robotics (see \cite{jamone16aff} for a recent survey). The ability to perceive affordances is crucial for any biological or artificial agent to interact successfully with the environment.
Central to the idea of affordances is that the action possibilities depend on both the agent and the environment; the same environment would offer different action possibilities to different agents, depending on their sensorimotor capabilities. A stable surface affords the possibility of traversal to an agent that is able to locomote and whose body dimensions fit the size of the surface borders; a stone affords the possibility of being used as a hammer to an agent who is able to pick it and who has enough force to lift it; a door handle affords to open to an agent who knows how to open doors, assuming it is well-designed \cite{norman99aff}. Therefore, those affordances must be \emph{learned} autonomously by the agent. In fact, the agent must learn how to perceive them. The means by which agents perceive affordances are those of ecological perception, powerfully illustrated by Eleanor Jack Gibson \cite{gibson94aff,gibson00aff,gibson03aff}: ``narrowing down from a vast manifold of (perceptual) information to the minimal, optimal information that specifies the affordance of an event, object, or layout’’ \cite[p.284]{gibson03aff}. The agent must learn what minimal information is to be picked; this happens both through evolution and development, leveraging the sensorimotor exploration of the environment by physical interaction.
Interestingly, while exploring the action possibilities, the agent can learn the effects of those actions as well. This is crucial for biological agents and turns out to be extremely useful for artificial systems as well. In fact, most computational models of affordances in robotics rely on representations that include not only the action but also the effects (or in other terms, the goal) of the action \cite{sahin2007afford,montesano08aff,Ugur2011,krueger11aff,Szedmak2014,goncalves14aff,dehban16aff,dehban17aff,ugur17aff,sarathy18aff,strama18aff,imre2019,seker2019rss,tekden2020}; therefore, the perceived possibilities for actions (and for achieving certain effects) can be used for action planning, leading to problem-solving \cite{antunes16planning}.
Such comprehensive models of affordances are, in fact, world models; they are internal models of how the world behaves ``in the eyes'' of the learning agent, and they can be used by the agent to make predictions about how the world will change if certain actions are performed. Therefore, it is not surprising that the computational techniques used for learning affordance models often overlap with those used for learning world models \cite{dehban22aff}.
It is worth noting that, in world model approaches, a robot only receives raw sensory information and needs to extract the relevant semantics from such data flow; therefore, to successfully integrate affordance perception in these systems, the challenges of meaningful abstraction/disentanglement and object-centric representation learning, described in Section 4.1, are particularly relevant.
\subsection{Social interaction}\label{sec:social}
Robots' ``worlds'' do not consist of physical objects alone but also of social entities, i.e., people who give them social guidance and try to cooperate with them. World models should model and predict social dynamics involving people's behaviors, and infer their latent variables, e.g., intentions and emotions, to cooperate with them, that is, to control social phenomena.
Efficient and safe human-robot collaboration and interaction are some of the main research objectives of robotics and have important practical applications \cite{shen2016investigation,albrecht2018autonomous,el2019cobot,hentout2019human,magrini2020human,ognibene2022active,semeraro2023human}.
Associating beliefs, intentions, or mental states to other agents, theory of mind, or, in other words, trying to predict the internal state of another agent’s world model to understand its activities and context \cite{ognibene2013contextual}, is an essential aspect of human interaction \cite{fotopoulou2017mentalizing,veissiere2020thinking,saxe2006uniquely} and has attracted attention in robotics \cite{bianco2019tomadvantages}.
Mutual understanding using a world model in social interaction can play an important role when complex interactions are challenging the perceptual systems of the agents, inducing a mismatch between their interpretation of the current context \cite{baker2017rational,bianco2022learningtom}. It is also crucial when different levels of knowledge and expertise induce different representations of a domain, as well as different points of view, which may induce conflictual interactions \cite{bianchi2021sweat} or different support strategies~\cite{ognibene2019proactive}. For example, the perspective of an automotive mechanic and that of an ordinary user differ considerably, so collaboration may be difficult if one cannot properly infer the internal state of others.
A robots’ world model can play a crucial role in its operation and functionality \cite{heinze2004modelling}.
The recent progress in machine learning methods has resulted in substantial improvement in action recognition methodologies \cite{roggen2010collecting,zeng2014convolutional,yang2015deep,Wang_2020_CVPR}.
However, this approach has often focused on shallow and purely perceptual representations of the observed activities resulting in limited flexibility in terms of contexts, tasks, and observed actors demanding a substantial amount of difficult-to-collect data and retraining time to apply the system in relatively similar conditions \cite{lee2019model}. Approaches such as goal recognition as planning or inverse planning \cite{ramirez2009plan,baker2009action,sohrabi2016plan,albrecht2018autonomous}, that, given a model of the environment, understand others' activities by computing plans that would result in the observed actions have shown the flexibility advantage delivered in intention recognition by a world model.
Several works have extended this approach. The problem of dealing with behaviors generated under partial observability, which may require inferring both the plan and the beliefs, the mental state \cite{bianco2019tomadvantages,bianco2022learningtom}, of the observed actor, was studied with both classical planning \cite{ramirez2011goal} and Bayesian approaches \cite{baker2011bayesian,baker2017rational}. The impact of missing observations for the observer agent has also been analyzed~\cite{ramirez2011goal}. A further step has been proposed by active methods for activity recognition \cite{shvo2020active,amato2019active,ognibene2013towards} that use the same world model both to interpret others' actions as well as selecting actions that would improve the recognition process, e.g., by giving access to the most informative observations \cite{lee2015stare} and allow the completion of a joint task \cite{ognibene2019proactive}.
While even the initial formulations of this approach were computationally aware \cite{ramirez2009plan}, their efficiency is often affected by the length of the observed behavior and the environment complexity, resulting in methods that can seldom be applied online on a robot. Several models proposed a pre-compile approach that transformed the world into a form that would allow efficient plan recognition \cite{lee2019model}. The adoption of hierarchical world model representations has also been considered to constrain the computational and modeling costs of the process \cite{demiris2006perceiving,cardona2017toward,proietti2021active}. Precomputed and robust local plans, in the form of the same motor controllers that the robot uses to perform its own actions, have also been adopted to allow active perception for action recognition and prediction on humanoid robots \cite{ognibene2013towards,ognibene2013contextual}.
One of the main issues of the approach, also related to computational efficiency considerations, is relying on specific algorithms for planning that aiming for the optimal plan may misinterpret the bounded rational behaviors that collaborators may perform.
This problem was faced by using online Bayesian inference in \cite{NEURIPS2020_df3aebc6}.
The additional flexibility provided by world models in social interaction skills is likely relevant beyond activity recognition. It is easy to imagine that purely supervised models may be limited in terms of perspective-taking and the ability to reason based on the world structure may help to adapt to partners with different sensory systems \cite{johnson2005perceptual,pandey2013towards,fischer2019computational,bianco2019tomadvantages}. Similarly, world models are likely to help with imitation learning by dealing with embodiment mismatch between the observed actor and the learner \cite{torabi2019recent}. Finally, physical cooperation \cite{mortl2012role,li2016framework} and signaling \cite{pezzulo2019body,ognibene2019implicit} would also be more flexible when integrating world and partner models in the equation, for example, to account for the trust of the human cooperator towards the robot \cite{kok2020trust}. Finally, a world model may also be learned through socially rich experiences and sources of information (e.g., imitation \cite{torabi2019recent} or verbal instructions) in addition to the results of autonomous exploration. However, developing a robust, efficient, and flexible enough representation may prove to be one of the main challenges in this effort.
\subsection{Brain-inspired world models}
In cognitive science, it has long been postulated that the brain learns small-scale models of the world and uses these models for various cognitive functions, such as perception, planning, and imagination~\cite{Craik1943}. For example, theories of perception-as-inference described perception as an inferential process, which works by ``inverting'' a generative model of how the percepts are generated \cite{Helmholtz1867,Gregory1997}. As discussed above, these ideas (and others) have been recently formalized under the label of the \emph{Bayesian Brain} \cite{Doya2007} and extended by \emph{Active Inference} from the domain of perception to other domains, such as action planning and interoception~\cite{Parr2022}.
In parallel, there have been many attempts to describe mathematically and to assess the neuronal underpinnings of world models and of inference processes empirically (e.g., \cite{shiffrin2020brain}). One question that has received a great deal of attention is how the brain might encode internal world models in the neuronal substrate. Given that the brain models are often assumed to be probabilistic, various formal schemes have been proposed that describe plausible neuronal implementations of probabilistic variables and of Bayesian inference over these variables, such as, for example, probabilistic population codes \cite{Ma2006} and sampling schemes \cite{Buesing2011}. These attempts show that (probabilistic, generative) world models could be at least potentially implemented in neuronal substrate \cite{sebastian2020local,lynn2020humans} -- and even updated after statistical learning \cite{Berkes2011} -- but the specific scheme(s) that the brain might use for this remain to be fully assessed.
Another relevant question is what algorithms the brain might use to perform inference over world models. A strong candidate in neuroscience is \emph{predictive coding} \cite{Rao1999,Friston2005}. Several studies have aimed to validate its key empirical predictions, showing that under the appropriate conditions, it is possible to observe predictions \cite{Ekman2017}, prediction errors \cite{Daw2011a} and other signatures of inference in brain signals \cite{Gomez2019} and that neural activity in lower visual areas in the absence of bottom-up inputs could be explained by the top-down, feedback dynamics postulated by predictive coding \cite{Muckli2015}. These and other studies (see \cite{Walsh2020} for a recent review) lend some support for predictive coding, but the theory remains under development.
At yet another level, one may ask what the systems-level architecture that supports world models and whether different parts of the brain might model different aspects of the world is. Anatomical considerations suggest that the brain is not a monolithic entity but rather is composed of several areas and networks \cite{Bullmore2009}; however, the extent to which these areas or networks are modularized and how they exactly influence each other are heavily discussed \cite{Laird2017}. One interesting consideration is that cortical brain areas in humans and monkeys appear to be organized along principal gradients (defined by functional connectivity); in one of these gradients, heteromodal areas (e.g., prefrontal cortex) are placed at the top, and unimodal areas (e.g., primary visual area) at the bottom, recapitulating the structure of a putative hierarchical generative model \cite{Margulies2016}. Another interesting consideration is that there seems to be a ``division of labor'' between brain pathways that perform complementary computations, such as the two visual pathways for processing ``what'' and ``where'' information \cite{ungerleider94}. These anatomical and functional separations might be potentially interpreted as useful \emph{factorizations} of the brain generative models.
A whole-brain probabilistic generative model (WB-PGM) approach attempts to build a cognitive architecture for cognitive and developmental robots integrating probabilistic generative model (PGM)-based modules referring comprehensive knowledge of human and animal brain architectures and their anatomy~\cite{taniguchi2022whole}.
The above studies indicate that at a general level, both neuroscience and machine learning / AI conceive world models and inference in similar ways. However, at a more detailed level, there might be profound differences between the ways these two disciplines use the same concepts. Predictive coding and other biological schemes proposed in neuroscience exploit top-down dynamics (and recurrences) in ways that are rarely used in machine learning. Furthermore, brain information processing is heavily based on spontaneous brain dynamics, which are largely absent in machine learning systems; see \cite{Singer2021,Pezzulo2021b,gyorgy2019brain} for a detailed discussion of putative computational roles of spontaneous dynamics. Moreover, it is plausible to assume that different parts of the brain might be specialized (or might have different \emph{inductive biases}) to process different statistical regularities, rendering them able to learn and model (for example) slower or faster dynamics of the visual scenes, one's own body, the actions of other agents, or extended temporal events \cite{Pezzulo2017a}. It is worth highlighting here that, although prediction errors have a central role in learning, there are other forms of statistical learning, such as those based on Hebbian associative learning \cite{nazli2022statistical}. It remains to be understood how to best endow our more advanced machine learning systems with the ability of the brain to perform (apparently) specialized computations but also orchestrate them coherently. Finally, it is important to remember that the brain is an evolved system, and our more advanced cognitive abilities are grounded in (the neuronal mechanisms supporting) simpler sensorimotor skills \cite{Pezzulo2016,Cisek2019}. Trying to develop advanced cognitive systems without the necessary requirements for embodied interaction and ``phylogenetic refinement'' might lead to solutions that differ completely from how the brain works -- or that fail altogether.
\subsection{Cognitive architectures}
Truly cognitive and developmental robots, i.e., embodied AGI, that behave autonomously and flexibly in the real environment would have a wide range of sensors and exhibit multiple functions. That requires a large-scale world model that deals with multimodal sensory observations and multilayered state representations. Considering the discussion in Section 4.5, such word models may be factorized in a proper manner from engineering and biological viewpoints.
To realize embodied AGIs, further frameworks and architectures to factorize a total world model into cognitive modules and to integrate individual cognitive capabilities into a cognitive system are required.
The idea is related to {\it cognitive architectures}, which have been studied in cognitive science, artificial intelligence, and robotics~\cite{lieto2018role,kotseruba202040}.
In cognitive science, cognitive functionalities such as memory, perception, and decision-making are implemented as modules in the cognitive architectures studied, and the specific task can be solved by activating these modules coordinately.
ACT-R \cite{anderson2009can} and Soar \cite{laird2008extending} are representatives of cognitive architectures.
It has been shown that the model implemented by ACT-R can explain the time to solve the task by humans, and activation patterns of the brain can be predicted by activation patterns of the modules~\cite{anderson2013human}.
Furthermore, Soar has been used for controlling robots~\cite{puigbo2013controlling} and learning games~\cite{mohan2009learning}.
However, complex machine learning methods that have rapidly advanced in a decade are not introduced yet.
Sigma \cite{rosenbloom2016sigma,damgaard2022toward} is a newer cognitive architecture that introduces the generative flow graph, a generalized probabilistic graphical model.
Therefore, the model can be implemented using probabilistic programming techniques \cite{bingham2019pyro,paige2017learning,tran2016edward}.
Furthermore, the concept of the standard model of the mind is discussed through a synthesis across these three cognitive architectures \cite{Laird2017}. Particularly, cognitive architectures based on first principles, e.g., with a general computation scheme, such as free energy minimization~\cite{Friston2005}, are especially attractive.
The architecture for social cognition has also been proposed \cite{sandini2018social}.
The authors point out that these architectures explained above are incomplete in dealing with the social aspect of cognition and describe the elements of architecture for social cognition.
Clarion \cite{Clarion} is another cognitive architecture based on dual process theory \cite{kahneman2003perspective}.
In this architecture, each subsystem is composed of explicit and implicit processes, and it is shown that the interaction between implicit-explicit processes can explain psychological phenomena.
In robotics, several types of cognitive architecture have been proposed. One of them is ArmarX \cite{vahrenkamp2015robot}, which has three layers, including a middleware layer, a robot framework layer, and an application layer. This three-layered structure simplifies the development robotics software easier.
(Neuro-)Serket \cite{serket_nakamura18,taniguchi2020neuro} is another approach to integrating cognitive modules\footnote{Neuro-SERKET is an updated version of SERKET.}.
In (Neuro-)Serket, modules are described by the (deep) PGM and trained mutually by exchanging messages between modules.
To make it easy to develop large-scale models, the modules in Neuro-Serket are weakly connected through the Serket interface.
(Neuro-)SERKET is closely related to the world model-based approach because SERKET requires each module to be a PGM, i.e., a model based on prediction and inference as Eq.~\eqref{eq_gm}, and integrate modules into a large PGM.
This architecture does not provide any restrictions regarding the functionalities of modules.
Therefore, it has high flexibility but brings high dimensional design space at the same time.
To reduce the large degree of freedom in the design space, a brain-inspired approach, WBA-PGM, was proposed~\cite{taniguchi2022whole}.
In this approach, a cognitive model was constructed by connecting PGM-based modules utilizing knowledge from neuroscience.
By referring to the brain studies, WBA-PGM constrains the function of modules and their connection and reduces the design space of the cognitive model.
There are two crucial requirements for cognitive architecture for cognitive and developmental robots, which can be used along with the approach based on world models and predictive coding.
The first is the engineering aspect which is seen in (Neuro-)Serket and ArmarX. The scale of cognitive models that enables the robots to behave flexibly in the real environment is very large, and many modules must be connected and work collaboratively.
Furthermore, the model needs to introduce machine learning techniques that are not only existing as well as those will be developed in rapid progress.
The development of such a model would require a massive engineering effort, and this is considered a notable obstacle to realizing such robots.
Therefore, architecture is needed to simplify development.
Another requirement is that of the scientific aspect seen in ACT-R, Soar, WBA-PGM, and Clarion. Developing AGI, which is human-like intelligence, referring to the knowledge regarding humans obtained in cognitive science and neuroscience, can accelerate its development.
However, meeting these two aspects completely is very challenging.
All machine learning techniques and module connections might necessarily be not reasonable from the point of view of cognitive science and neuroscience.
On the other hand, entire humans are not understood yet.
Therefore, finding common ground between engineering and science aspects and developing a novel cognitive architecture is a current challenge.
Developing a large-scale cognitive architecture and overcoming the problems described in the previous subsections is also a challenge.
\section{Discussion}
As we described, world models and predictive coding are promising approaches in cognitive and developmental robotics. Before closing this paper, we will mention some remaining issues which have not been addressed in the main body sufficiently.
\textbf{Language and world models}:~Umwelts, i.e., worlds from first-person views, of biological systems are not monolithic but have some sort of structure. Notably, language and symbolic systems have syntactic structures. The interaction between high-level cognitive capabilities, e.g., language and reasoning, and low-level cognitive capabilities, e.g., perception and action, is essential in world modeling. Recently, large-scale language models (LLMs) have been replacing many natural language processing methods~\cite{devlin2019bert,brown2020language}, including reasoning tasks, which have been conducted solely by symbolic AI by the end of 2010s~\cite{wang2022self,kojima2022large}. Recently, the use of LLMs in robotics has been attempted, e.g., ~\cite{saycan2022arxiv}.
It is clear that language learning and understanding by robots is itself a frontier~\cite{tangiuchi2019survey}.
To leverage the symbolic knowledge in LLMs, integration of LLMs and world models will be an important challenge.
This shift from models of artificial symbols in conventional AI to models of natural language, i.e., a human symbol system, is resonating with the discussion in symbol emergence in cognitive and developmental systems~\cite{Steels2008,taniguchi2018symbol,Taniguchi2016SER}. An important topic is then considering not only the integration of human language into robots' world models in a top-down manner but also the bottom-up formation of symbol systems, including language in relation to world models.
\textbf{Policy representations}:~How should the policies of robots be represented? Conventionally, policies are described as feedback controllers $\pi(z_t, a_t) = p(a_t|z_t)$ in reinforcement and imitation learning. Even though one direction of world model approaches is to explore task agnostic representations (Section \ref{sec:representationsForAction}), the decomposition of world modeling and policy learning can be controversial. In a conventional approach of world models, policies ($\pi=p(a_t|z_t)$ or $a_{t:T}$) and world models ($p(z_{t+1}|z_t, a_t) $ and $p(o_t|z_t)$) are decoupled. In contrast, a series of studies about predictive coding in neuro-robotics have been intentionally entangling policies and world models and making robots directly learn $p(o_{t+1}, a_{t+1}| o_{1:t}, a_{1:t})$ and exhibiting many successful results in robotics, e.g.,\cite{jung2019goal,ito2022efficient,tani2016exploring}.
As the notion of affordance also suggests, actions and perceptions are not independent and entangled, generally. The question ``to what extent should we decouple world model and policy representations?’’ should be investigated.
\textbf{From artificial cognition to human cognition}:~Cognitive and developmental robotics are also constructive approaches to human developmental cognition. Not only learning from neuro-, cognitive and developmental sciences but also provide with scientific feedback to them is also an important mission. Building a virtuous circle between studies on human and artificial studies is a challenge.
The constructive approach may give us a novel approach to scientific and philosophical hard problems like self-awareness~\cite{hoffmann2021robot} and consciousness. The relationship between the multimodal world model and global workspace theory was suggested~\cite{juliani2022link}. Extending the discussion between world models and consciousness using robots may be an exciting challenge. Moreover, the relationship between predictive coding and emotion is worth exploring to build emotional robots and understand the emotions of biological systems~\cite{seth2013interoceptive,bauermeister2022role,hieida2022survey}.
\textbf{Software frameworks for implementation}:~To accelerate the studies on world models and predictive coding in robotics, the development framework for cognitive and developmental robotics is crucially important. In robotics, not only AI ``software’’ frameworks but also middle-ware are important. Recently, ROS has been widely used in the robotics community for bridging hardware and AI software layers. Developing and sharing such software frameworks as a community will be important, e.g., \cite{Lotfi2023sii}. Moreover, the world model involves many types of knowledge, and the knowledge can be used for achieving multiple functions via active inference. The software framework should allow the world model to efficiently organize the knowledge and to perform (cross-modal) active inference. A great initiative is the discrete state-space active inference python library~\cite{heins2022pymdp}. However, it can only be used for toy examples due to scalability issues and to be useful in cognitive and developmental robotics. It needs further development. For instance, the support for high-dimensional input observations and the possibility of combining discrete and continuous action and state representations are something that has been addressed in robotic approaches~\cite{lanillos2021active}.
\textbf{Data-efficient and autonomous learning}:~A generalist agent called GATO was developed based on Transformers and shown to be able to solve various tasks with one neural network~\cite{reed2022generalist}. Although the approach is superficially different, the approach is really related to the world models and the predictive coding approach. However, the learning system is hugely data-hungry. It is very questionable if the model can be regarded as a model of human intelligence. Moreover, to train the generalist agent, researchers need to prepare a large dataset and simulation environment.
Human children can autonomously explore their environment and acquire data through active exploration. Moreover, they use heuristics and biases in their developmental process. Learning and considering the human developmental process will give us the inspiration to build real generalist agents.
Developing a data-efficient autonomous learning architecture with world models and predictive coding at its core is the key to a truly cognitive and developmental system.
\textbf{Emergence of behaviors}:~Should an agent have a completely internal model of its world? Lastly, we raise fundamental speculation about the world model-based approach.
Behaviors are not externalization of internally designed trajectories but something to emerge through the interaction between the body and the environment.
For example, it has been proven by passive walking machines that the behavior of walking emerges only from the interaction between the body and the environment, without any computation by the brain~\cite{mcgeer1990passive}. About three decades ago, Brooks famously advocated the physical grounding hypothesis together with subsumption architecture, saying the world is its own best model~\cite{brooks1990elephants}. The robots behaved smoothly and flexibly without any explicit world models. This is also referred to as morphological computation, which means the body itself implicitly processes information dynamically~\cite{pfeifer2007self,pfeifer2009morphological}. Soft robotics emphasizes these points nowadays. Combining the viewpoints of the emergence of behaviors with complex physical dynamics and the world model-based approach is another important challenge.
\section{Conclusion}
In this survey paper, we have aimed to clarify the frontiers and challenges of world models and predictive coding in cognitive and developmental robotics.
Creating an autonomous robot that can actively explore the real environment, acquire knowledge, and learn skills continuously is the ultimate goal of cognitive and developmental robotics.
To make the robot continuously develop through active exploration, the robot's learning process should be based on sensorimotor information obtained through physical and social interactions with the physical and social environment. Following the motivation, this paper reviewed studies related to world models and predictive coding in cognitive and developmental robotics and related AI studies.
We clarified the definition of world model and predictive coding in robotics, in conjunction with those of FEP and active inference, and discussed the relationship between them.
We also introduced state-of-the-art and research gaps of studies on world models and predictive in robotics. We described six frontiers and challenges, i.e., latent representations for action planning, neuro-symbolic predictive models, affordance perception, social interaction, brain-inspired world models, and cognitive architecture.
Through the survey and clarification of challenges, we provided future directions for developing cognitive and developmental robots based on world models and predictive coding.
\section*{Acknowledgements}
This work was partially supported by JST Moonshot R\&D, Grant Number JPMJMS2033 and JPMJMS2011, JST PRESTO Grant Number JPMJPR22C9, the BAGEP Award of the Science Academy, TUBITAK ARDEB 1001 program (project number: 120E274), the European Union’s Horizon 2020 Framework Programme for Research and Innovation under Specific Grant Agreements No. 945539 (Human Brain Project SGA3), No. 952215 (TAILOR), and No. 824153 (POTION), the European Research Council under the Grant Agreement No. 820213 (ThinkAhead), the Human Brain Project Specific Grant Agreement 3 grant (ID 643945539, for the SPIKEFERENCE project), Deepself project under the Priority Programme “The Active Self” (SPP 2134), the project `COURAGE - A social media companion safeguarding and educating students' (No. 95563 and No. 9B145), and the Volkswagen Foundation inside the initiative Artificial Intelligence and the Society of the Future.
\bibliographystyle{tADR}
|
\section*{Introduction}
In Voevodsky's theory of mixed motives, the notion of $\mathbb{A}^1$-invariant sheaf played a fundamental role (see \cite{voetri}, \cite{VoeA1}, \cite{MVW} etc.).
An $\mathbb{A}^1$-invariant sheaf is a \emph{sheaf with transfers} satisfying \emph{$\mathbb{A}^1$-invariance}.
A sheaf with transfers is a Nisnevich sheaf of abelian groups on the category of \emph{finite correspondences}, denoted $\category{Cor}_k$, where $k$ is the fixed base field.
The objects of $\category{Cor}_k$ are smooth schemes over $k$, and a morphism $X \to Y$ in $\category{Cor}_k$ is given by a finite formal sum of integral closed subschemes of $X \times Y$ which are finite surjective over a connected component of $X$.
We say that a sheaf with transfers $F$ is $\mathbb{A}^1$-invariant if the map $F(X) \to F(X \times \mathbb{A}^1)$, induced by the first projection $X \times \mathbb{A}^1\to X$, is an isomorphism for any $X \in \category{Cor}_k$.
In fact, we can define the notion of $\mathbb{A}^1$-invariant sheaf not only over a field but also over \emph{any noetherian scheme}, by using the notion of \emph{relative cycle}.
Cisinski-D\'eglise worked out this idea in \cite{CD} with extreme generality, and it is now widely accepted as a standard framework of mixed motives over a general base.
On the other hand, in a series of papers \cite{KSY1}, \cite{KSY2}, Kahn-Saito-Yamazaki developed the theory of \emph{reciprocity sheaves}.
This is a vast generalization of the theory of $\mathbb{A}^1$-invariant sheaves over a field, and it captures ramification-theoretic information of invariants of schemes (see e.g. \cite{RS21}).
The class of reciprocity sheaves includes many interesting examples that are not $\mathbb{A}^1$-invariant, such as the sheaf of differential forms, the (truncated) Hodge-Witt sheaf, and all commutative algebraic groups.
The first aim of this paper is to develop the theory of reciprocity sheaves over an arbitrary noetherian scheme.
The following table shows the relationship of our work with the previous ones:
\begin{table}[htb]
\centering
\begin{tabular}{|c||c|c|} \hline
& $\mathbb{A}^1$-invariant sheaves & Reciprocity sheaves \\ \hline \hline
over a field & Voevodsky & Kahn-Saito-Yamazaki \\ \hline
over a noetherian scheme & Cisinski-D\'eglise & This paper \\ \hline
\end{tabular}
\end{table}
Let us recall what reciprocity sheaves are.
The key idea is to replace smooth schemes by {\it proper modulus pairs}.
A proper modulus pair is a pair $(X,D_X)$ of a proper $k$-scheme $X$ and an effective Cartier divisor $D_X$ on $X$ such that $X \setminus |D_X|$ is a smooth $k$-scheme.
For example, the pair $\bcube :=(\mathbb{P}^1,\{\infty\})$ is a proper modulus pair, which we call the \emph{cube}.
We can extend the notion of finite correspondence to the notion of \emph{modulus correspondence}, taking into account the information of Cartier divisors (see \cite{KMSY1}, \cite{KMSY2} for precise definitions).
We write $\category{MCor}_k$ for the resulting category.
A \emph{modulus presheaf with transfers} is an additive presheaf of abelian groups on $\category{MCor}_k$.
A modulus presheaf with transfers $F$ is said to be \emph{cube-invariant} if for any proper modulus pair $\sX = (X,D_X)$, the map
\[
F(\sX) \to F(\sX \boxtimes \bcube), \quad \sX \boxtimes \bcube := (X \times \mathbb{P}^1, \pr_1^*D_X +\pr_2^*\{\infty\}),
\]
which is induced by the first projection, is an isomorphism.
A presheaf with transfers (resp. sheaf with transfers) $F$ is said to be a reciprocity presheaf (resp. reciprocity sheaf) if it belongs to the essential image of
\[
\left(
\begin{tabular}{c}
\text{cube-invariant}\\
\text{presheaves}
\end{tabular}
\right)
\hookrightarrow
\PSh(\category{MCor}_k)
\xrightarrow{\omega_!}
\PSh(\category{Cor}_k),
\]
where $\omega_!$ is the unique colimit-preserving functor extending (through the Yoneda embedding) the forgetful functor
\[
\category{MCor}_k \to \category{Cor}_k; \quad (X,D_X) \mapsto X \setminus |D_X|.
\]
Originally, the notion of reciprocity sheaf given in \cite{KSY1} looks quite different from the above one.
However, in \cite{KSY2}, it is shown that the above definition coincides with the original one, provided that the base field $k$ is perfect.
An advantage of the above description is that we can think of a reciprocity sheaf as the ``shadow'' of a cube-invariant presheaf.
Since the definition of cube-invariant presheaves is very similar to that of $\mathbb{A}^1$-invariant sheaves, one can prove many important properties of cube-invariant presheaves by following Voevodsky's classical methods, at least partly.
In the first half of this paper, we will show that the above formulation of reciprocity sheaves works well after replacing the base field $k$ by a noetherian scheme $S$.
The notions of proper modulus pair, modulus correspondence and modulus (pre)sheaf with transfers are already introduced in \cite{Kelly-Miyazaki}, but we (re-)define all these notions over noetherian base schemes, both for the sake of completeness and for a more essential reason: we would like to allow the effective divisor $D_X$ to have \emph{rational coefficients} (see applications below for the reason why we need this generalization).
Based on the category $\category{Cor}_S$ of finite correspondences over $S$ \cite[Definition 9.1.8]{CD}, we will define the category of \emph{proper $\mathbb{Q}$-modulus pairs} $\category{QCor}_S^\mathrm{prop}$ over $S$ (Definition \ref{def_QCor}).
We can define the notion of cube-invariant presheaf and the functor $\omega_!\colon \PSh(\category{QCor}_S^\mathrm{prop})\to \PSh(\category{Cor}_S)$ as before.
Then we will introduce the notion of reciprocity sheaf over $S$, and study their behavior under categorical operations, i.e. base-change, restriction, etc.
\
In the latter half of this paper, we will give some applications of our generalized theory.
The main results include a motivic construction of the de Rham-Witt complex of Hesselholt-Madsen \cite{Hesselholt-Madsen}, which we will sketch below.
It is well-known that the multiplicative group $\mathbb{G}_m$, which is an important example of an $\mathbb{A}^1$-invariant sheaf, has a motivic presentation.
That is, there exists a canonical isomorphism $\mathbb{G}_m \simeq h_0^{\mathbb{A}^1}(\mathbb{Z}_\tr(\mathbb{A}^1\setminus\{0\})/\mathbb{Z})$ in $\PSh(\category{Cor}_k^\mathrm{aff})$, where $\category{Cor}_k^\mathrm{aff}$ is the full subcategory of $\category{Cor}_k$ consisting of affine schemes, $\mathbb{Z}_\tr\colon \category{Cor}_k\to \PSh(\category{Cor}_k)$ is the Yoneda embedding, and $h_0^{\mathbb{A}^1}$ is the $0$-th Suslin homology:
$$
h_0^{\mathbb{A}^1}F(X)=\Coker(F(X\times \mathbb{A}^1)\xrightarrow{i_0^*-i_1^*} F(X)).
$$
Moreover, the group structure on $\mathbb{G}_m$ is simply induced by
the multiplication morphism $\mathbb{A}^1\setminus\{0\} \times \mathbb{A}^1\setminus\{0\} \to \mathbb{A}^1\setminus\{0\}$.
Existence of a motivic presentation of $\mathbb{G}_m$ is fundamental in the theory of $\mathbb{A}^1$-invariant sheaves and mixed motives; it is essentially equivalent to the computation of motivic cohomology in weight one.
It is a natural idea to extend this to other invariants of schemes.
In this paper, we give a motivic presentation for the de Rham-Witt complex $W_n\Omega^q (X)$ of a scheme $X$, which is a fundamental example of an invariant of schemes which is not $\mathbb{A}^1$-invariant.
We start with a motivic presentation of the ring of big Witt vectors $\mathbb{W}_n(A)=1+tA[t]/(t^{n+1})$ of a ring $A$.
Let $S$ be a noetherian regular scheme.
For $n\geq 0$, we define a modulus presheaf with transfers $\mathbb{W}_n^+$ over $S$ by
$$
\mathbb{W}_n^+=\varinjlim_{\varepsilon>0}\mathbb{Z}_\tr(\mathbb{P}^1_S, (n+\varepsilon)\{\infty\})/\mathbb{Z},
$$
where $\mathbb{Z}_\tr\colon \category{QCor}_S^\mathrm{prop}\to \PSh(\category{QCor}_S^\mathrm{prop})$ is the Yoneda embedding.
Then there are operations $F_s, V_s$ on $\mathbb{W}_n^+$ for each $s>0$ which are induced by the morphism $t\mapsto t^s$ on $\mathbb{A}^1_S$ and its transpose.
These operations are called the Frobenius and the Verschiebung.
Moreover, the multiplication on $\mathbb{A}^1_S$ induces a multiplication on $\mathbb{W}_n^+$.
The $0$-th Suslin homology $h_0^{\mathbb{A}^1}$ generalizes to modulus presheaves with transfers:
$$
h_0^{\smallcube} F(\mathcal{X})=\Coker(F(\mathcal{X}\boxtimes\bcube)\xrightarrow{i_0^*-i_1^*} F(\mathcal{X})).
$$
We write $h_0=\omega_!h_0^{\smallcube}$.
First we prove the following result.
\begin{theorem}[Theorem \ref{thm:ab-comparison}, Proposition \ref{prop:ring-comparison}]\label{thm:intro_1}
Let $S$ be a noetherian regular scheme.
Then there is an isomorphism $h_0\mathbb{W}_n^+\xrightarrow{\sim}\mathbb{W}_n$ in $\PSh(\category{Cor}_S^\mathrm{aff})$ which preserves the multiplication, Frobenius, and the Verschiebung.
\end{theorem}
In the proof of this theorem, we use the fact that the group $h_0\mathbb{Z}_\tr(\mathbb{P}^1_S, r\{\infty\})\;(r\in \mathbb{Q}_{>0})$ depends only on $\lceil r \rceil$.
This is a consequence of the \textit{motivic Hasse-Arf theorem} which we prove in Theorem \ref{geometric Hasse-Arf for curves}.
It is an analogue of the classical Hasse-Arf theorem which states that the upper ramification group $G^{(r)}$ of an abelian extension of a local field depends only on $\lceil r \rceil$.
Theorem \ref{thm:intro_1} has a non-trivial application to reciprocity presheaves.
Imitating the construction in \cite{Miyazaki-19}, we define
$
NF(X)=\Ker(i_0^*\colon F(X\times\mathbb{A}^1)\to F(X))
$
for a reciprocity presheaf $F$.
Then we have $NF=0$ if and only if $F$ is $\mathbb{A}^1$-invariant.
Using the above theorem, we can prove that $\mathbb{W}=\varprojlim_{n\geq 0}\mathbb{W}_n$ acts on $NF$.
As a consequence, we obtain the following generalization of the result of Binda-Cao-Kai-Sugiyama \cite[Theorem 1.3]{BCKS}.
\begin{theorem}[Corollary \ref{BCKSgen}]
Let $S$ be a noetherian regular scheme, $F$ be a reciprocity presheaf over $S$ and assume that $F$ is separated for the Zariski topology.
\begin{enumerate}
\item If $S$ is a $\mathbb{Q}$-scheme and $F\otimes\mathbb{Q}=0$, then $F$ is $\mathbb{A}^1$-invariant.
\item If $S$ is a $\mathbb{Z}_{(p)}$-scheme and $F\otimes\mathbb{Z}_{(p)}=0$, then $F$ is $\mathbb{A}^1$-invariant.
\item If $S$ is an $\mathbb{F}_p$-scheme and $F$ is $p$-torsion-free, then $F$ is $\mathbb{A}^1$-invariant.
\end{enumerate}
\end{theorem}
Next we give a motivic presentation of the ring of $p$-typical Witt vectors $W_n(A)$ of a $\mathbb{Z}_{(p)}$-algebra $A$.
Let us assume that the base scheme $S$ is a noetherian regular $\mathbb{Z}_{(p)}$-scheme.
For $n\geq 0$, we define a modulus presheaf with transfers $\widehat{\mathbb{W}}_n^+$ by
$$
\widehat{\mathbb{W}}_n^+=\varinjlim_{\varepsilon>0}\mathbb{Z}_\tr(\mathbb{P}^1_S, \varepsilon\{0\}+(n+\varepsilon)\{\infty\}).
$$
As in the case of $\mathbb{W}_n^+$, we have operations $F_s,V_s$ for $s>0$ and a multiplication on $\widehat{\mathbb{W}}_n^+$.
We define $W_n^+$ to be the quotient of $\widehat{\mathbb{W}}_{p^{n-1}}^+\otimes \mathbb{Z}_{(p)}$ by the images of the idempotents $\ell^{-1}V_\ell F_\ell$ for all prime numbers $\ell\neq p$.
The operations $F_p,V_p$ descend to operations $F,V$ on $W_n^+$ called the Frobenius and the Verschiebung, and the multiplication also descends to $W_n^+$.
We prove the following
\begin{theorem}[Corollary \ref{cor:p_typical_comparison}]
Let $S$ be a noetherian regular $\mathbb{Z}_{(p)}$-scheme.
Then there is an isomorphism $h_0W_n^+\xrightarrow{\sim}W_n$ in $\PSh(\category{Cor}_S^\mathrm{aff})$ which is compatible with the Frobenius, the Verschiebung, and the multiplication.
\end{theorem}
To treat the de Rham-Witt complex, we have to assume that $p\neq 2$ and that the base scheme $S$ admits \emph{de Rham-Witt transfers}, i.e. there exists a suitable action of finite correspondences on the de Rham-Witt complex.
This is satisfied when $S\in \category{Sm}_k$ for some field $k$ of characteristic $p$.
We define a modulus presheaf with transfers $\mathbb{G}_m^+$ by
$$
\mathbb{G}_m^+=\varinjlim_{\varepsilon>0}\mathbb{Z}_\tr(\mathbb{P}^1_S, \varepsilon\{0\}+\varepsilon\{\infty\})/\mathbb{Z}.
$$
Our main result can be stated as follows:
\begin{theorem}[Theorem \ref{thm:rep-WOmega}]
Let $S$ be a noetherian regular $\mathbb{Z}_{(p)}$-scheme with $p\neq 2$.
Assume that $S$ admits de Rham-Witt transfers (e.g. $S\in \category{Sm}_k$ for some field $k$ of characteristic $p$).
Then there is an isomorphism
$$a_\mathrm{inj} a_{\Nis} h_0 (W_n^+ \otimes \mathbb{G}_m^{+\otimes q})\xrightarrow{\sim}W_n\Omega^q$$
in $\Sh_\Nis(\category{Cor}_S)$.
\end{theorem}
Here, $a_\Nis$ denotes the Nisnevich sheafification, and $a_\mathrm{inj}$ denotes the maximal quotient sheaf such that the restriction to the generic fibers is injective.
The proof goes as follows: first we prove that the left hand side admits a Witt complex structure (Theorem \ref{thm:theta}).
Since the de Rham-Witt complex is initial in the category of Witt complexes, we obtain a unique morphism from the right hand side to the left hand side.
To prove that this morphism is an isomorphism, we construct an inverse by using the existence of the transfer structure on $W_n\Omega^q$.
We note that a similar motivic presentation for $\Omega^q$ is obtained by R\"ulling-Sugiyama-Yamazaki \cite{RSY} over a perfect field of characteristic $\neq 2,3,5$, and by the first author \cite{Koi2} over a perfect field of characteristic $\neq 2$.
Also, a presentation of the big de Rham-Witt complex $\mathbb{W}_n\Omega^q$ using additive higher Chow groups is obtained by R\"ulling \cite{Rulling-thesis} for fields, and by Krishna-Park \cite{KP11} for regular semilocal algebras over a field.
We expect that these results will be connected by some comparison isomorphism between Suslin homology and additive higher Chow groups.
\
The structure of the present paper is as follows.
In \S \ref{sec:preliminiary},
we prepare preliminary results on normalization and ($\mathbb{Q}$-)Cartier divisors.
In \S \ref{sec:modpair},
we define a $\mathbb{Q}$-modulus pair to be a pair of a finitely reducible scheme and a $\mathbb{Q}$-effective $\mathbb{Q}$-Cartier divisor. We define the category of modulus correspondences over an arbitrary noetherian scheme, by taking the amalgamated sum of the constructions of \cite{KMSY1} and \cite{CD}.
In \S \ref{sec:ci-rec},
we define and study the notion cube-invariant presheaf over a general noetherian base. The category of reciprocity presheaves is defined, following \cite{KSY2}, as the image of cube invariant presheaves under the left Kan extension of the functor $(X,D_X) \mapsto X \setminus |D_X|$.
Moreover, we study the behavior of reciprocity presheaves under base change and restriction with respect to the base scheme.
In \S \ref{sec:ex-rec},
we provide several examples of reciprocity sheaves.
The transfer structure on a commutative group scheme is constructed in Appendix \ref{sec:appendix}.
In \S \ref{sec:HA},
we formulate and prove the motivic Hasse-Arf theorem, by comparing the $0$-th Suslin homology group and the Chow group of relative $0$-cycles of a modulus pair. This comparison can be seen as a generalization of a result in \cite{RY16}.
In \S \ref{sec:Witt},
we apply our machinery to give a motivic construction of the ring of Witt vectors and basic operations on it.
As an application, we give a generalization of the result of Binda-Cao-Kai-Sugiyama \cite{BCKS} on torsion and divisibility of a reciprocity sheaf.
In \S \ref{sec:dRWitt},
we give a motivic construction of the de Rham-Witt complex, assuming the existence of transfers on it.
\subsection*{Acknowledgements}
We would like to thank Shuji Saito for his interest on this work. We also appreciate his valuable comments on earlier versions of this paper, which encouraged the authors to improve the main results.
\subsection*{Notations and conventions}
\begin{itemize}
\item For a ring $A$, we write $Q(A)$ for its total ring of fractions.
For a scheme $X$, we write $\mathcal{K}_X$ for the sheaf of total ring of fractions on $X$.
For $f\in \Gamma(X,\mathcal{K}_X^\times)$, we write $\cdiv(f)$ for the Cartier divisor on $X$ defined by $f$.
\item For a scheme $X$ and $x\in X$, we write $k(x)$ for the residue field of $X$ at $x$.
We write $X^{(d)}$ for the set of points on $X$ of codimension $d$.
An element of $X^{(0)}$ is called a \emph{generic point}.
\item We say that a morphism of schemes $f\colon X\to Y$ is \emph{pseudo-dominant} if it takes generic points to generic points, i.e., $f(X^{(0)})\subset Y^{(0)}$.
\item For an integral scheme $X$, we write $k(X)$ for its function field.
If $X$ is a noetherian normal integral scheme and $f\in k(X)^\times$, we write $\div(f)$ for the Weil divisor on $X$ defined by $f$.
\item We write $\category{Sch}_S$ for the category of $S$-schemes, and $\category{Sm}_S$ for the category of smooth separated $S$-schemes of finite type.
We define $\category{Sm}_S^\mathrm{aff}$ to be the full subcategory of $\category{Sm}_S$ spanned by affine schemes.
\item For an additive category $\mathcal{C}$, we write $\PSh(\mathcal{C})$ for the category of additive functors $\mathcal{C}^\op\to \category{Ab}$.
\item For a category $\mathcal{C}$, we write $\Pro(\mathcal{C})$ for the category of pro-objects in $\mathcal{C}$.
A {\it pro-functor} $F\colon \mathcal{C}\dashrightarrow\mathcal{D}$ is a functor $F\colon \mathcal{C}\to \Pro(\mathcal{D})$.
Let $F\colon \mathcal{C}\dashrightarrow\mathcal{D}$ be a pro-functor and $G\colon \mathcal{D}\to \mathcal{C}$ be a functor.
Then $F$ is called {\it pro-left adjoint} to $G$ if there is an isomorphism
$\Hom_{\Pro(\mathcal{D})}(F(c),d)\simeq \Hom_{\mathcal{C}}(c,G(d))$ which is natural in $c\in \mathcal{C},d\in \mathcal{D}$.
\end{itemize}
\newpage
\section{Preliminaries}\label{sec:preliminiary}
\subsection{Generalities on normalization}
In this paper, we frequently use the notion of normalization.
One can define the normalization morphism $X^N \to X$ of a scheme $X$ if $X$ has only finitely many irreducible components.
Since this condition is wordy, we introduce the following terminology:
\begin{definition}\label{def:loc-finitary}
A scheme is called \emph{finitely reducible} if it has only finitely many irreducible components.
\end{definition}
For example, any noetherian scheme is finitely reducible.
Note that any finitely reducible normal scheme is a finite disjoint union of normal integral schemes.
\begin{definition}
Let $X$ be a finitely reducible scheme.
Let $Y=\coprod_{\eta \in X^{(0)}}\Spec k(\eta)$ and let $j\colon Y\to X$ be the canonical morphism.
We define the \emph{normalization} of $X$ by $X^N:=\Spec_{X}\mathcal{A}$, where $\mathcal{A}$ is the integral closure of $\mathcal{O}_X$ in $j_*\mathcal{O}_Y$.
\end{definition}
The following lemma shows that the normalization of finitely reducible schemes behaves like the normalization of algebraic varieties, except that the former is just an integral morphism, and is not a finite morphism in general.
Recall that a morphism of schemes is called \emph{integral} if it is affine and universally closed.
\begin{lemma}\label{lem:normalization}
Let $X$ be a finitely reducible scheme and $\pi :X^N \to X$ be the normalization of $X$.
Then the following assertions hold:
\begin{enumerate}
\item We have $X^N\simeq \sqcup_i X_i^N$, where $X_i$ are the irreducible components of $X$ endowed with the reduced scheme structure.
\item The scheme $X^N$ is a finite disjoint union of normal integral schemes.
\item \label{normalization_integral} The morphism $\pi :X^N \to X$ is integral, surjective, and induces a bijection on the sets of generic points.
\item \label{normalization_universal_property}(universal property of normalization) For any pseudo-dominant morphism $Z \to X$ with $Z$ a finitely reducible normal scheme, there exists a unique factorization $Z \to X^N \to X$.
\end{enumerate}
\end{lemma}
\begin{proof}
First we prove (1).
We may assume that $X$ is affine.
Write $X=\Spec A$.
Then $A$ has only finitely many minimal prime ideals $\mathfrak{p}_1,\dots,\mathfrak{p}_n$.
Let $B_i$ be the integral closure of $A/\mathfrak{p}_i$ in $k(\mathfrak{p}_i)$.
Then the integral closure of $A$ in $\prod_{i=1}^nk(\mathfrak{p}_i)$ is $\prod_{i=1}^n B_i$, so we have $X^N=\coprod_{i=1}^n\Spec B_i$.
This proves (1).
The rest of the statement follows from (1) and the case where $X$ is integral; see \cite[Proposition 12.43 and Proposition 12.44]{Goertz-Wedhorn}.
\end{proof}
\subsection{Pullback of Cartier divisors}
Pullback of Cartier divisors is defined in \cite[(21.4.2)]{EGA4-4}.
The following lemma is a well-known criterion for existence of pullback:
\begin{proposition}\label{lem:pb-criteria}
Let $f\colon Y \to X$ be a morphism of schemes. Assume one of the following:
\begin{enumerate}
\item $f$ is flat.
\item $X$ is finitely reducible, $Y$ is reduced, and $f$ is pseudo-dominant.
\end{enumerate}
Then for any Cartier divisor $D$ on $X$, the pullback $f^*D$ exists.
\end{proposition}
\begin{proof}
See \cite[(21.4.5) (iii)]{EGA4-4}.
\end{proof}
\begin{corollary}
Let $X$ be a finitely reducible scheme and $D$ be a Cartier divisor on $X$.
\begin{enumerate}
\item Let $i\colon X_\red \to X$ be the canonical closed immersion. Then $i^*D$ exists.
\item Let $\pi:X^N \to X$ be the normalization. Then $\pi^*D$ exists.
\end{enumerate}
\end{corollary}
For effective Cartier divisors, we also have the following
\begin{lemma}\label{lem:pb-criteria-3}
Let $f\colon Y\to X$ be a morphism of schemes and $D$ be an effective Cartier divisor on $X$.
Suppose that $f^{-1}(X\setminus |D|)$ is scheme-theoretically dense in $Y$.
Then $f^*D$ exists.
\end{lemma}
\begin{proof}
Since the problem is local, we may assume that $X = \Spec A$, $Y=\Spec B$ are affine, and $D$ is defined by a non-zero-divisor $d\in A$.
Let $\varphi\colon A\to B$ be the ring homomorphism corresponding to $f$.
It suffices to show that $\varphi(d)$ is a non-zero-divisor in $B$.
Our assumption implies that the canonical homomorphism $B\to B[\varphi(d)^{-1}]$ is injective.
Since $\varphi(d)$ is invertible in $B[\varphi(d)^{-1}]$, it is a non-zero-divisor in $B[\varphi(d)^{-1}]$ and hence also in $B$.
\end{proof}
\subsection{Valuative criterion for inequality of Cartier divisors}
Recall that we write $D\leq D'$ for Cartier divisors $D,D'$ if $D'-D$ is effective.
We prove a valuative criterion for inequality of Cartier divisors on finitely reducible normal schemes.
See \cite[Chapter VI]{Bourbaki-Commutative-Algebra} for basic facts on valuation rings.
\begin{lemma}\label{lem:valuative}
Let $X$ be a finitely reducible normal scheme and $D,D'$ be Cartier divisors on $X$. Then the following conditions are equivalent:
\begin{enumerate}
\item $D\leq D'$.
\item For any valuation ring $R$ and any morphism $\rho\colon \Spec R \to X$ such that $\rho^*D,\rho^*D'$ are defined, we have $\rho^*D\leq \rho^*D'$.
\item \label{val3} For any generic point $\eta \in X$, any valuation ring $R$ with fraction field $k(\eta)$, and any morphism $\rho\colon \Spec R\to X$ extending $\Spec k(\eta)\to X$, we have $\rho^*D\leq \rho^*D'$.
\end{enumerate}
\end{lemma}
\begin{proof}
(1) $\implies$ (2) $\implies$ (3) is clear.
It remains to show (3) $\implies$ (1). Since the problem is local on $X$, we may assume that $X = \Spec A$ with $A$ a local normal domain with fraction field $K$.
Let $d \in K^\times$ be an element which represents the Cartier divisor $D'-D$.
Let $B:=A[d^{-1}]\subset K$.
Then the image of $d^{-1}$ in $B/\mathfrak{m}_AB$ is invertible.
Indeed, if there is some $\mathfrak{p}\in \Spec B$ with $\mathfrak{p}\cap A=\mathfrak{m}_A$ and $d^{-1}\in \mathfrak{p}$, then we can take a valuation ring $R$ with fraction field $K$ dominating $B_\mathfrak{p}$.
Then $d^{-1}$ belongs to the maximal ideal of $R$ and hence $d\not \in R$, contradicting to our assumption (3).
Therefore $1\in d^{-1}B+\mathfrak{m}_AB$ and hence we can write
\[
1 = (a_0 + a_1 d^{-1} + \cdots + a_n d^{-n})d^{-1} + (m_0 + m_1 d^{-1} + \cdots + m_l d^{-l})
\]
for some $a_i \in A$ and $m_j \in \mathfrak{m}_A$.
Noting that $1-m_0 \in A^\times$ since $A$ is local, the above equality shows that $d$ is a solution of a monic equation with coefficients in $A$.
Therefore we get $d\in A$.
\end{proof}
\begin{corollary}\label{cor:dominance}
Let $X,Y$ be finitely reducible normal schemes.
Let $f \colon Y \to X$ be a universally closed morphism such that $f(Y^{(0)})=X^{(0)}$ and let $D,D'$ be Cartier divisors on $X$.
Then we have $D\leq D'$ if and only if $f^* D\leq f^*D'$.
\end{corollary}
\begin{proof}
The ``only if'' part is obvious. To prove the ``if'' part, we may assume that $X$ is integral. Moreover, choosing an irreducible component of $Y$ dominating $X$, we may also assume that $Y$ is integral.
Suppose that $f^* D\leq f^*D'$.
We will check that $D$ and $D'$ satisfies the condition \eqref{val3} in Lemma \ref{lem:valuative}. Let $\xi$ be the generic point of $X$.
Let $R$ be a valuation ring with fraction field $K=k(\xi)$ and $\rho \colon \Spec R \to X$ be a morphism extending the canonical morphism $\Spec k(\xi) \to X$.
Let $\eta$ be the generic point of $Y$ and set $L=k(\eta)$. Then $R$ is a local subring of $L$, so there exists a valuation ring $R'$ with fraction field $L$ dominating $R$.
By the valuative criterion for universal closedness (\cite[(7.3.8)]{EGA2}, \cite[\href{https://stacks.math.columbia.edu/tag/01KF}{Tag 01KF}, \href{https://stacks.math.columbia.edu/tag/04XU}{Tag 04XU}]{stacks-project}), we can find a morphism $\rho'\colon \Spec R'\to Y$ lifting $\rho$.
Our assumption implies $\rho'^*f^*D\leq \rho'^*f^*D'$.
Now $\rho^*D\leq \rho^*D'$ follows from $R=R'\cap K$.
\end{proof}
\subsection{Weak inequality of Cartier divisors}
The valuative criterion for inequality of Cartier divisors is not valid on non-normal schemes.
To remedy this, we introduce the ``weak inequality'' for Cartier divisors on finitely reducible schemes.
\begin{definition}
Let $X$ be a finitely reducible scheme and $\pi\colon X^N \to X$ be the normalization of $X$.
For two Cartier divisors $D,D'$ on $X$, we write $D \prec D'$ if $\pi^*D \leq \pi^*D'$.
Note that if $D\prec D' \prec D''$ then we have $D \prec D''$.
\end{definition}
\begin{remark}\label{rem:reduced}
If $D\leq D'$ then $D\prec D'$.
If $i\colon X_{\red} \to X$ is the natural closed immersion, then $D\prec D'$ if and only if $i^* D\prec i^*D'$, since $X_{\red}^N = X^N$ by definition.
\end{remark}
Here is the valuative criterion for weak inequality.
\begin{lemma}\label{lem:we-valuative}
Let $X$ be a finitely reducible scheme and $D,D'$ be Cartier divisors on $X$.
Then the following conditions are equivalent:
\begin{enumerate}
\item $D\prec D'$.
\item For any valuation ring $R$ and any morphism $\rho\colon \Spec R\to X$ such that $\rho^*D,\rho^*D'$ are defined, we have $\rho^* D\leq \rho^*D'$.
\item For any generic point $\eta \in X$, any valuation ring $R$ with fraction field $k(\eta)$, and any morphism $\rho\colon \Spec R\to X$ extending $\Spec k(\eta)\to X$, we have $\rho^* D\leq \rho^*D'$.
\end{enumerate}
\end{lemma}
\begin{proof}
First we prove (1) $\implies$ (2).
Let $\pi\colon X^N \to X$ be the normalization, $R$ be a valuation ring with fraction field $K$, and $\rho\colon \Spec R \to X$ be a morphism such that $\rho^*D,\rho^*D'$ are defined.
Then there is an algebraic extension $L$ of $K$ and a morphism $\Spec L\to X^N$ lifting $\Spec K\to X$.
We can take a valuation ring $R'$ with fraction field $L$ dominating $R$.
By the valuative criterion for universal closedness (\cite[(7.3.8)]{EGA2}, \cite[\href{https://stacks.math.columbia.edu/tag/01KF}{Tag 01KF}, \href{https://stacks.math.columbia.edu/tag/04XU}{Tag 04XU}]{stacks-project}), there is a morphism $\rho'\colon \Spec R'\to X^N$ lifting $\rho$.
If $D\prec D'$, then we have $\rho'^*\pi^*D\leq \rho'^*\pi^*D'$.
Now $\rho^*D\leq \rho^*D'$ follows from $R=R'\cap K$.
The implication (2) $\implies$ (3) is clear, and (3) $\implies$ (1) is a direct consequence of Lemma \ref{lem:valuative}.
\end{proof}
The following lemma shows that the weak inequality is preserved by pullback.
\begin{lemma}\label{lem:we-pb}
Let $f\colon Y \to X$ be a morphism between finitely reducible schemes and $D,D'$ be Cartier divisors on $X$.
Assume that $f^*D, f^*D'$ are defined.
Then the following assertions hold:
\begin{enumerate}
\item If $D\prec D'$, then $f^* D\prec f^*D'$.
\item Assume that $f$ is universally closed and $f(Y^{(0)})=X^{(0)}$. If $f^* D\prec f^*D'$, then $D\prec D'$.
\end{enumerate}
\end{lemma}
\begin{proof}
By Remark \ref{rem:reduced}, we may assume that $X$ and $Y$ are reduced.
First we prove (2). By the universal property of the normalization (Lemma \ref{lem:normalization} (\ref{normalization_universal_property})), the composite $Y^N \to Y \to X$ factors through the normalization $X^N \to X$:
\[\xymatrix{
Y^N \ar[r]^g \ar[d]_{\pi_Y} & X^N \ar[d]^{\pi_X} \\
Y \ar[r]^f & X.
}\]
The morphism $g$ induces a surjection $(Y^N)^{(0)}\to (X^N)^{(0)}$ since $f(Y^{(0)})=X^{(0)}$.
Moreover, $g$ is universally closed since $f$ is universally closed, $\pi_Y$ is integral, and $\pi_X$ is separated.
Therefore, $g$ satisfies the assumption in Corollary \ref{cor:dominance}, and hence $\pi_X^* D\leq \pi_X^*D'$ if and only if $\pi_Y^* f^* D\leq \pi_Y^* f^* D'$.
This proves (2).
Next we prove (1).
Let $\{y_1,\dots,y_n\}$ be the set of generic points of $Y$.
Set $Y' := X^N \times_X Y$, and choose a point $z_i \in Y'$ lying over $y_i$ for each $i=1,\dots,n$, which is possible because $Y' \to Y$ is a surjection.
Let $Z$ be the closure of $\{z_1,\dots,z_n\}$ in $Y'$, equipped with the reduced scheme structure.
Note that $Z$ is finitely reducible by construction.
Consider the commutative diagram
\[\xymatrix{
Z \ar[r]^{\mathrm{cl.}} \ar[rd]_{p} & Y' \ar[r] \ar[d]^{\pi'} & X^N \ar[d]^{\pi_X} \\
& Y \ar[r]^f & X,
}\]
where $\pi_X$ is the normalization of $X$.
By construction, the morphism $Z \to Y$ induces a surjection $Z^{(0)}\to Y^{(0)}$.
Moreover, $Z \to Y$ is universally closed as the composite of the integral morphism $\pi'$ and the closed immersion $Z \to Y'$.
Therefore, we can apply (2) to $Z \to Y$ and obtain that $p^*f^*D\prec p^*f^*D' \implies f^*D\prec f^*D'$.
Since the composite $f \circ p$ factors through $X^N$ and $\pi_X^* D\leq \pi_X^*D'$ by our assumption, we conclude that $f^*D\prec f^*D'$.
\end{proof}
We will need the following lemmas later.
\begin{lemma}\label{lem:ineq_domination}
Let $X$ be a quasi-compact finitely reducible scheme and $D_1, D_2, \Sigma$ be effective Cartier divisors on $X$.
Assume that $D_1 |_U \prec D_2 |_U$, where $U:=X \setminus |\Sigma|$. Then we have $D_1 \prec D_2 + n\Sigma$ for $n\gg 0$.
\end{lemma}
\begin{proof}
First we note that the normalization morphism $X^N \to X$ is affine and hence quasi-compact.
Replacing $X$ with $X^N$, we may assume that $X$ is normal.
In this case, the symbol $\prec$ in the assertion becomes $\leq$.
Moreover, by covering $X$ with finitely many affine open subsets, we are reduced to the case where $X = \Spec A$ is affine and $D_1, D_2, \Sigma$ are defined by $d_1,d_2,s \in A$, respectively.
Then the assumption says that $d_2 / d_1 \in A[s^{-1}]$, which implies $s^n \cdot d_2 / d_1 \in A$ for $n\gg 0$.
Therefore we have $D_1 \leq D_2 + n\Sigma$ for $n\gg 0$.
\end{proof}
\begin{lemma}\label{lem:ineq_limit}
Let $(X_i)_{i\in I}$ be a projective system of quasi-compact finitely reducible schemes with affine flat transition morphisms such that $X\simeq \varprojlim_{i\in I}X_i$ is integral.
Suppose that we are given effective Cartier divisors $D_i,D'_i$ on $X_i$ for each $i\in I$, and they are compatible with the transition morphisms.
We write $D,D'$ for the pullback of $D_i,D'_i$ to $X$.
If $D\prec D'$, then $D_i\prec D'_i$ holds for some $i\in I$.
\end{lemma}
\begin{proof}
We may assume that $I$ has a minimum element $i_0$.
Taking a finite affine open covering of $X_{i_0}$, we may assume that $X_{i_0} = \Spec A_{i_0}$ and that $D_{i_0},D'_{i_0}$ are defined by non-zero-divisors $d_{i_0},d'_{i_0}\in A_{i_0}$.
We can write $X_i = \Spec A_i$ and $X = \Spec A$ with $A=\varinjlim_{i\in I} A_i$.
We write $d,d'$ for the image of $d_{i_0},d'_{i_0}$ in $A$.
For $i\in I$, we write $d_i,d'_i$ for the image of $d_{i_0},d'_{i_0}$ in $A_i$.
Since the transition morphisms are flat, $d,d',d_i,d'_i$ are all non-zero-divisors.
By our assumption that $D\prec D'$, there is some $e\in Q(A)$ which is integral over $A$ such that $d'=de$.
The element $e$ satisfies some monic equation
$$
e^n+a_{n-1}e^{n-1}+\cdots+a_0 = 0
$$
where $a_0,\dots,a_{n-1}\in A$.
Multiplying by $d^n$, we get
$$
d'^n+a_{n-1}d'^{n-1}d+\cdots+a_0d^n =0
$$
in $A$.
Since $A=\varinjlim_{i\in I} A_i$, there is some $i\in I$ and $b_0,\dots,b_{n-1}\in A_i$ such that
$$
d'^n_i+b_{n-1}d'^{n-1}_i d_i+\cdots+b_0d_i^n = 0
$$
holds in $A_i$.
If $\mathfrak{p}$ is a minimal prime ideal of $A_i$, then the above equality shows that $d'_i/d_i\in k(\mathfrak{p})$ is integral over $A_i/\mathfrak{p}$.
Therefore we get $D_i\prec D'_i$.
\end{proof}
\subsection{$\mathbb{Q}$-Cartier divisors}
We introduce the notion of $\mathbb{Q}$-Cartier divisor over a general scheme, which will play a fundamental role in this paper.
\begin{definition}
For a scheme $X$, we write $\CDiv(X)$ for the group of Cartier divisors on $X$, and set
$\CDiv_{\mathbb{Q}} (X) := \CDiv (X) \otimes_{\mathbb{Z}} \mathbb{Q}$.
We call an element of $\CDiv_{\mathbb{Q}} (X)$ a \emph{$\mathbb{Q}$-Cartier divisor on $X$}.
\end{definition}
If $X$ is noetherian and normal, then the group $\CDiv (X)$ is embedded into the free abelian group of Weil divisors on $X$. In particular, $\CDiv (X)$ is a free abelian group in this case.
If $X$ is only finitely reducible and normal, then we have the following
\begin{lemma}\label{lem:Q-inj}
Let $X$ be a finitely reducible normal scheme. Then the group $\CDiv (X)$ is a torsion-free $\mathbb{Z}$-module.
Equivalently, the natural map $\CDiv (X) \to \CDiv_{\mathbb{Q}} (X)$ is injective.
\end{lemma}
\begin{remark}
The group $\CDiv (X)$ is not free in general.
For example, if $R$ is a valuation ring with value group $\mathbb{Q}$, then we have $\CDiv(\Spec R) \simeq \mathbb{Q}$.
\end{remark}
\begin{proof}
We may assume that $X$ is a normal integral scheme.
Since $\CDiv (X)$ is by definition the section over $X$ of the Zariski sheaf $\mathcal{K}_X^\times / \mathcal{O}_X^\times$, we may assume that $X = \Spec A$ with $A$ a local normal domain. Then we have $\CDiv (X) \simeq K^\times / A^\times$, where $K$ is the fraction field of $A$.
It suffices to prove that $K^\times/A^\times$ is torsion-free.
Let $f\in K^\times$.
Suppose that there is some integer $m>0$ such that $a := f^m \in A^\times$.
Then $f$ is a solution in $K$ of the monic equation $x^m - a$ with coefficients in $A$.
Since $A$ is a normal domain, this shows that $f$ belongs to $A$.
Since $f^m \in A^\times$ by assumption, we also have $f \in A^\times$.
This shows that $K^\times/A^\times$ is torsion-free.
\end{proof}
\begin{definition}
Let $f\colon Y \to X$ be a morphism of schemes and $D$ be a $\mathbb{Q}$-Cartier divisor on $X$.
We say that the pullback $f^*D$ of $D$ by $f$ \emph{exists} if there is an ordinary Cartier divisor $E$ on $X$ and $r \in \mathbb{Q}$ with $D = rE$ such that the pullback $f^* E$ exists.
In this situation, we define $f^* D := rf^*E\in \CDiv_{\mathbb{Q}} (Y)$.
This does not depend on the choice of $E$ and $r$.
\end{definition}
We extend the notion of effectiveness for $\mathbb{Q}$-Cartier divisors.
\begin{definition}
Let $X$ be a scheme and $D$ be a $\mathbb{Q}$-Cartier divisor on $X$.
We say that $D$ is \emph{$\mathbb{Q}$-effective} if there is an ordinary effective Cartier divisor $E$ on $X$ and $r\in \mathbb{Q}_{>0}$ such that $D = rE$ in $\CDiv_{\mathbb{Q}}(X)$.
For $\mathbb{Q}$-Cartier divisors $D,D'$, we write $D\leq D'$ if $D'-D$ is $\mathbb{Q}$-effective.
\end{definition}
If an ordinary Cartier divisor $D$ is effective, then it is $\mathbb{Q}$-effective.
The converse is not true in general, but it is true for normal schemes:
\begin{lemma}\label{effective_vs_q-effective}
Let $X$ be a normal scheme and $D$ be an ordinary Cartier divisor on $X$.
Then $D$ is effective if and only if $D$ is $\mathbb{Q}$-effective.
\end{lemma}
\begin{proof}
We may assume that $X=\Spec A$ with $A$ a local normal domain and $D=\div(a)$ with $a\in Q(A)^\times$.
If $D$ is $\mathbb{Q}$-effective, then there is some non-zero-divisor $b\in A$ and $n\geq 1$ such that $\div(a)=(1/n)\div(b)$ in $\CDiv_\mathbb{Q}(X)$.
By Lemma \ref{lem:Q-inj}, we have $\div(a^n)=\div(b)$ in $\CDiv(X)$.
This implies that $a^n-ub=0$ for some $u\in A^\times$, so $a\in A$ since $A$ is normal.
Therefore $D$ is effective.
\end{proof}
\begin{definition}
Let $X$ be a finitely reducible scheme and $D,D'$ be $\mathbb{Q}$-Cartier divisors on $X$.
Let $\pi\colon X^N\to X$ be the normalization.
We write $D\prec D'$ if $\pi^*D\leq \pi^*D'$.
\end{definition}
\begin{remark}
Let $X$ be a finitely reducible scheme and $D,D'$ be ordinary Cartier divisors on $X$.
Then we have $D\prec D'$ as $\mathbb{Q}$-Cartier divisors if and only if $D\prec D'$ as ordinary Cartier divisors; this follows from Lemma \ref{effective_vs_q-effective}.
\end{remark}
\begin{definition}
Let $X$ be a scheme and $D$ be a $\mathbb{Q}$-effective $\mathbb{Q}$-Cartier divisor on $X$.
Suppose that $D=rE$ in $\CDiv_\mathbb{Q}(X)$ where $E$ is an ordinary effective Cartier divisor on $X$ and $r\in \mathbb{Q}_{>0}$.
Then the support $|D|$ of $D$ is defined to be $|E|$.
This is well-defined since $|E|=|nE|$ for any ordinary effective Cartier divisor $E$ and $n\geq 1$.
\end{definition}
One can easily see that Lemma \ref{lem:we-valuative}, Lemma \ref{lem:we-pb}, Lemma \ref{lem:ineq_domination}, and Lemma \ref{lem:ineq_limit} are also true for $\mathbb{Q}$-Cartier divisors.
\subsection{Relative cycles}
In this subsection we recall the theory of relative cycles in the style of Cisinski-D\'eglise \cite[Chapter 8]{CD}.
\begin{definition}[Cycles]
A {\it cycle} (with rational coefficients) is a pair $(X,\alpha)$ where $X$ is a noetherian scheme and $\alpha$ is a formal $\mathbb{Q}$-linear combination of integral closed subschemes of $X$.
We often omit $X$ from the notation and say that $\alpha$ is a cycle on $X$.
An expression $\alpha = \sum_{i=1}^n m_i[V_i]$ is said to be in the \emph{standard form} if $V_1,\dots,V_n$ are distinct and $m_i\neq 0$.
In this case the schemes $V_1,\dots,V_n$ are called the \emph{components} of $\alpha$, and the \emph{support} of $\alpha$ is defined by $|\alpha| = \bigcup_{i=1}^n V_i$.
We say that $\alpha$ has \emph{integral coefficients} if $m_1,\dots,m_n$ are all integers.
For a closed subscheme $W$ of $X$, we define the associated cycle $[W]$ on $X$ by
$[W]=\sum_{i=1}^n \length(\mathcal{O}_{W,W_i})[W_i]$,
where $W_1,\dots,W_n$ are the irreducible components of $W$.
For a noetherian scheme $X$, we use the same symbol $X$ for the cycle $(X,[X])$.
\end{definition}
\begin{definition}[Morphisms of cycles]
A \emph{morphism of cycles} $f\colon (X,\alpha)\to (Y,\beta)$ is a morphism of schemes $f\colon X\to Y$ such that $f(|\alpha|)\subset |\beta|$.
We say that $f$ is {\it finite} if $|\alpha|\to |\beta|$ is finite.
We say that $f$ is {\it pseudo-dominant} if $f$ sends the generic points of components of $\alpha$ to those of $\beta$.
\end{definition}
There is a class of morphisms of cycles called {\it special} morphisms \cite[Definition 8.1.49]{CD}, for which the {\it pullback} is defined.
If $X\to Y$ is a morphism of noetherian schemes and $\beta$ is a cycle on $Y$, then cycles on $X$ which are special over $\beta$ form a $\mathbb{Q}$-vector space.
Consider the following diagram of cycles:
\begin{align}\label{pullback_diagram}
\xymatrix@R=10pt@C=10pt{
&(X,\alpha)\ar[d]\\
(Y,\beta)\ar[r] &(S,\sigma).
}
\end{align}
If $\alpha\to \sigma$ is special, then a cycle $\alpha\otimes_\sigma \beta$ on $X\times_SY$ called the {\it pullback} of $\alpha$ along $\beta\to \sigma$ is defined.
The cycle $\alpha\otimes_\sigma\beta$ is special over $\beta$ and the operation $\otimes_\sigma$ is $\mathbb{Q}$-bilinear.
Special morphisms are pseudo-dominant, and the class of special morphisms is closed under composition \cite[Corollary 8.2.6]{CD}.
\begin{lemma}\label{lem:pullback_lemma}
The pullback operation satisfies the following properties:
\begin{enumerate}
\item \label{lem:flat_cycles}
{\rm (Flat cycles)} Consider the situation (\ref{pullback_diagram}) with $\sigma=[S]$, $\alpha = [V]$, and $\beta = [W]$, where $V\subset X$, $W\subset Y$ are closed subschemes.
If $V$ is flat and equidimensional over $S$, then $\alpha$ is special over $S$ and we have
$\alpha\otimes_S\beta = [V\times_S W]$.
\item \label{lem:commutativity}
{\rm (Commutativity)} Consider the situation (\ref{pullback_diagram}).
Suppose that $\alpha$ and $\beta$ are both special over $S$.
Then we have
$\alpha\otimes_S\beta = \beta\otimes_S\alpha$.
\item \label{lem:associativity}
{\rm (Associativity)}
Consider the following diagram of cycles.
$$
\xymatrix@R=10pt@C=10pt{
&&\alpha\ar[d]\\
&\beta\ar[r]\ar[d]&\sigma\\
\gamma\ar[r]&\tau
}
$$
Suppose that $\alpha\to \sigma$ and $\beta\to \tau$ are special.
Then we have
$(\alpha\otimes_\sigma \beta)\otimes_\tau \gamma = \alpha \otimes_\sigma (\beta \otimes_\tau \gamma)$.
\item \label{lem:support}
{\rm (Support)}
Consider the situation (\ref{pullback_diagram}).
Suppose that $\alpha$ is special over $\sigma$.
Then $|\alpha\otimes_\sigma \beta|$ is a closed subscheme of $|\alpha|\times_{|\sigma|}|\beta|$.
\end{enumerate}
\end{lemma}
\begin{proof}
See \cite[Section 8.1.35 (P3) and Corollary 8.2.2]{CD}, \cite[Corollary 8.2.3]{CD}, \cite[Corollary 8.2.7]{CD}, and \cite[Proposition 8.2.8 and Corollary 8.2.10]{CD}, respectively.
\end{proof}
\begin{definition}[$\mathbb{Z}$-universal]
Let $\alpha,\beta$ be cycles with integral coefficients.
A morphism of cycles $\alpha\to \beta$ is called \emph{$\mathbb{Z}$-universal} if it is special and for any cycle $\gamma$ with integral coefficients and a morphism of cycles $\gamma\to \beta$, the pullback $\alpha\otimes_\beta\gamma$ also has integral coefficients.
For example, if $X\to S$ is a morphism of noetherian schemes and $V\subset X$ is a closed subscheme which is flat and equidimensional over $S$, then $[V]$ is $\mathbb{Z}$-universal over $S$ by Lemma \ref{lem:pullback_lemma} (\ref{lem:flat_cycles}).
For a morphism of noetherian schemes $X\to S$, we write $c_0(X/S)$ for the group of cycles on $X$ which are finite and $\mathbb{Z}$-universal over $S$.
\end{definition}
When $X$ is regular, the group $c_0(X/S)$ admits a simple description:
\begin{lemma}\label{lem:regular_z_universal}
Let $S$ be a noetherian scheme and $X$ be a regular $S$-scheme which is of finite type and separated over $S$.
Then $c_0(X/S)$ consists of all cycles on $X$ with integral coefficients which are finite and pseudo-dominant over $S$.
\end{lemma}
\begin{proof}
See \cite[Remark 9.1.3]{CD}.
\end{proof}
\begin{lemma}\label{lem:pullback_c0}
Cosinder the following diagram of noetherian schemes.
$$
\xymatrix@R=10pt@C=10pt{
&X\ar[d]\\
Y\ar[d]\ar[r] &S\\
T
}
$$
If $\alpha\in c_0(X/S)$ and $\beta\in c_0(Y/T)$, then $\alpha\otimes_S\beta\in c_0(X\times_SY/T)$.
\end{lemma}
\begin{proof}
The cycle $\alpha\otimes_S\beta$ is $\mathbb{Z}$-universal over $T$ by Lemma \ref{lem:pullback_lemma} (\ref{lem:associativity}).
By Lemma \ref{lem:pullback_lemma} (\ref{lem:support}), $|\alpha\otimes_S\beta|$ is a closed subscheme of $|\alpha|\times_S|\beta|$.
Since $|\alpha|$ is finite over $S$ and $|\beta|$ is finite over $T$, it follows that $|\alpha\otimes_S\beta|$ is finite over $T$.
\end{proof}
\begin{definition}[Pushforward]
If $f\colon X\to Y$ is a morphism of noetherian schemes and $\alpha$ is a cycle on $X$, then a cycle $f_*\alpha$ on $Y$ called the {\it pushforward} of $\alpha$ along $f$ is defined by
$$
[V]\mapsto
\begin{cases}
[k(V):k(W)][W] &([k(V):k(W)]<\infty) \\
0 &(\text{otherwise}),
\end{cases}
$$
where $W$ is the scheme-theoretic image of $V$ under $f$.
\end{definition}
\begin{lemma}\label{lem:push_c0}
Let $S$ be a noetherian scheme and $X,Y$ be noetherian $S$-schemes.
Suppose that $Y$ is separated and of finite type over $S$.
Then for any $S$-morphism $f\colon X\to Y$ and $\alpha\in c_0(X/S)$, we have $f_*\alpha\in c_0(Y/S)$.
\end{lemma}
\begin{proof}
See \cite[Section 9.1.1]{CD}.
\end{proof}
\subsection{Presheaves with transfers}
In this subsection we recall the definition of presheaves with transfers in the style of \cite{CD}.
We fix a noetherian scheme $S$.
Let $X,Y$ be noetherian $S$-schemes.
We set $c_S(X,Y):=c_0(X\times_SY/X)$ and call its elements {\it finite correspondences} from $X$ to $Y$ over $S$.
\begin{example}
(1) The graph $\Gamma_f$ of a morphism $f\colon X\to Y$ over $S$ gives an element $[\Gamma_f] \in c_S(X,Y)$, for which we simply write $f$.
(2) If $g\colon Y\to X$ is finite and flat, then the transpose of $\Gamma_g$ gives an element $[{}^t\Gamma_g]\in c_S(X,Y)$, for which we write ${}^tg$.
\end{example}
\begin{definition}[Composition]
Let $X,Y,Z$ be noetherian $S$-schemes.
For $\alpha\in c_S(X,Y)$ and $\beta\in c_S(Y,Z)$, we define a cycle $\beta\circ\alpha$ on $X\times_SZ$ by the formula
$\beta\circ\alpha = \pr_{13,*}(\beta\otimes_Y\alpha)$,
where $\pr_{13} \colon X\times_S Y\times_S Z\to X\times_S Z$ is the projection.
Applying Lemma \ref{lem:pullback_c0} to the diagram below, we see that $\beta\otimes_Y\alpha \in c_S(X\times_SY\times_SZ/X)$.
$$
\xymatrix@R=15pt@C=15pt{
{} &Y\times_SZ\ar[d]^-{\pr_1}\\
X\times_SY\ar[r]^-{\pr_2}\ar[d]^-{\pr_1} &Y\\
X
}
$$
If $Z$ is separated and of finite type over $S$, then it follows from Lemma \ref{lem:push_c0} that $\beta\circ \alpha \in c_S(X,Z)$.
This gives a bilinear pairing
$\circ\colon c_S(Y,Z)\times c_S(X,Y)\to c_S(X,Z)$.
\end{definition}
\begin{lemma}\label{lem:composition_formula}
Let $X,Y,Z,W$ be noetherian $S$-schemes.
Suppose that $Z,W$ are separated and of finite type over $S$.
\begin{enumerate}
\item Let $\alpha\in c_S(X,Y)$, $\beta\in c_S(Y,Z)$, and $\gamma \in c_S(Z,W)$.
Then we have $(\gamma\circ \beta)\circ \alpha = \gamma\circ (\beta\circ \alpha)$.
\item Let $f\colon X\to Y$ be an $S$-morphism and $\beta\in c_S(Y,Z)$.
Then we have $\beta\circ f = \beta\otimes_YX$.
\item Let $g\colon Y\to Z$ be an $S$-morphism and $\alpha\in c_S(X,Y)$.
Then we have $g \circ \alpha = (\id\times g)_*\alpha$.
\end{enumerate}
\end{lemma}
\begin{proof}
See \cite[Proposition 9.1.7]{CD}.
In loc.cit., it is assumed that $X,Y$ are also separated and of finite type over $S$, but this assumption is not used in the proof.
\end{proof}
\begin{definition}[Category $\category{Cor}_S$]
Using the above pairing as the composition, we can define an additive category $\category{Cor}_S$ whose objects are the same as $\category{Sm}_S$ and whose morphisms are finite correspondences.
Note that we have a faithful functor $\category{Sm}_S\hookrightarrow \category{Cor}_S$ given by $f\mapsto [\Gamma_f]$ on morphisms.
We define $\category{Cor}_S^\mathrm{aff}$ to be the full subcategory of $\category{Cor}_S$ spanned by affine schemes.
\end{definition}
Next we define a symmetric monoidal structure on $\category{Cor}_S$.
Let $X,Y,X',Y'\in \category{Sm}_S$ and $\alpha\in c_S(X,Y)$, $\alpha'\in c_S(X',Y')$.
We define a cycle $\alpha\times \alpha'$ on $X\times_S X'\times_S Y\times_S Y'$ by
$$
\alpha\times \alpha':= (\alpha\otimes_X (X\times_S X'))\otimes_{X\times_S X'}(\alpha'\otimes_{X'} (X\times_S X')).
$$
Applying Lemma \ref{lem:pullback_c0} to the diagram below, we get $\alpha\times \alpha' \in c_S(X\times_S X',Y\times_S Y')$.
$$
\xymatrix@R=15pt@C=15pt{
&X\times_SX'\times_SY\ar[d]^-{\pr_{12}}\\
X\times_SX'\times_SY'\ar[r]^-{\pr_{12}}\ar[d]^-{\pr_{12}} &X\times_SX'\\
X\times_SX'
}
$$
We have $(\beta\circ\alpha)\times (\beta'\circ \alpha') = (\beta\times \beta')\circ (\alpha\times \alpha')$ \cite[Proposition 9.2.2]{CD} and hence the fiber product over $S$ gives a monoidal structure $\times_S$ on $\category{Cor}_S$ with unit $S$.
This monoidal structure has a symmetry $X\times_S Y\simeq Y\times_S X$ by Lemma \ref{lem:pullback_lemma} (\ref{lem:commutativity}).
\begin{definition}[Presheaves with transfers]
An object of $\PSh(\category{Cor}_S)$ is called a {\it presheaf with transfers} over $S$.
We write $\mathbb{Z}_\tr(X)$ for the presheaf with transfers represented by $X\in \category{Cor}_S$.
The fiber product $\times_S$ extends by colimits to a symmetric monoidal structure $\otimes$ on $\PSh(\category{Cor}_S)$ with unit $\mathbb{Z}=\mathbb{Z}_\tr(S)$.
This symmetric monoidal structure is closed:
$$
\mathcal{H}\mathrm{om}(F,G)(X) =\Hom(\mathbb{Z}_\tr(X)\otimes F, G).
$$
\end{definition}
\begin{definition}
Let $F_1,\dots,F_n\in \PSh(\category{Cor}_S)$, $X\in \category{Sm}_S$ and $a_i\in F_i(X)\;(i=1,\dots,n)$.
We define $a_1\otimes\dots\otimes a_n\in (F_1\otimes\dots\otimes F_n)(X)$ by the composition
$$
\mathbb{Z}_\tr(X)\xrightarrow{\Delta_X}\mathbb{Z}_\tr(X^n) = \mathbb{Z}_\tr(X)^{\otimes n}\xrightarrow{a_1\otimes\dots\otimes a_n} F_1\otimes\dots\otimes F_n.
$$
\end{definition}
\begin{lemma}[Projection formula]
Let $F,G\in \PSh(\category{Cor}_S)$, $X\in \category{Sm}_S$, $Y\in \category{Sm}_X$ and let $\pi\colon Y\to X$ be the structure morphism.
Let $a\in F(X)$ and $b\in G(Y)$.
For any $\alpha\in c_X(X,Y)$, we have
$$
\alpha^*(\pi^*a\otimes b) = a \otimes (\alpha^*b).
$$
Here, we regard $\alpha$ as an element of $c_S(X,Y)$ via the closed immersion $X\times_XY\to X\times_SY$.
\end{lemma}
\begin{proof}
Let $\delta\colon Y=X\times_XY\to X\times_SY$ denote the canonical closed immersion.
It suffices to show that the following diagram in $\category{Cor}_S$ is commutative:
$$
\xymatrix{
X\ar[rr]^-{\Delta_X}\ar[d]^-{\delta_*\alpha} &&X\times_SX\ar[d]^-{\id\times\delta_*\alpha}\\
Y\ar[r]^-{\Delta_Y} &Y\times_S Y\ar[r]^-{\pi\times\id} &X\times_SY.
}
$$
By Lemma \ref{lem:composition_formula}, the composition $(\id\times \delta_*\alpha)\circ \Delta_X$ is represented by the cycle
$(\delta_*\alpha \otimes_X (X\times_SX)) \otimes_{X\times_S X} [\Delta_X]
=\delta_*\alpha \otimes_X [\Delta_X]
= (\Delta_X\times \id)_*\delta_*\alpha$,
and the composition $(\pi\times \id)\circ\Delta_Y\circ \delta_*\alpha$ is represented by the cycle $(\id\times (\pi,\id))_*\delta_*\alpha=(\pi,\pi,\id)_*\alpha$.
These cycles are clearly the same.
\end{proof}
Let $f\colon T\to S$ be a morphism of noetherian schemes.
For $X\in\category{Sm}_S$, we write $X_T:=X\times_ST\in \category{Sm}_T$.
For $X,Y\in \category{Sm}_S$ and $\alpha\in c_S(X,Y)$, we write $\alpha_T:=\alpha\otimes_XX_T\in c_T(X_T,Y_T)$.
Then we have $(\beta\circ\alpha)_T = \beta_T\circ \alpha_T$ \cite[Lemma 9.3.1]{CD}.
\begin{definition}[Base change functor]
Let $f\colon T\to S$ be a morphism of noetherian schemes.
The {\it base change functor} $f^\star\colon \category{Cor}_S\to \category{Cor}_T$ is defined by $X\mapsto X_T$.
This induces an adjunction
$$
f^*\colon \PSh(\category{Cor}_S)\rightleftarrows \PSh(\category{Cor}_T)\colon f_*.
$$
The functor $f^*$ is the extension of $f^\star$ by colimits and the functor $f_*$ is given by precomposition with $f^\star$.
\end{definition}
Let $p\colon T\to S$ be a morphism of noetherian schemes such that $T\in \category{Sm}_S$.
Then any $X\in \category{Sm}_T$ can also be regarded as an object of $\category{Sm}_S$.
Let $X,Y\in \category{Sm}_T$ and $\alpha\in c_T(X,Y)$.
Let $\delta\colon X\times_TY\to X\times_SY$ be the canonical closed immersion.
We set $\alpha^S:=\delta_*\alpha \in c_S(X,Y)$.
Then we have $(\beta\circ\alpha)^S = \beta^S\circ\alpha^S$ \cite[Lemma 9.3.5]{CD}.
\begin{definition}[Restriction functor]
Let $p\colon T\to S$ be a morphism of noetherian schemes such that $T\in \category{Sm}_S$.
The {\it restriction functor} $p_\star\colon \category{Cor}_T\to \category{Cor}_S$ is defined by $X\mapsto X$.
This gives a left adjoint of $p^\star$ since $X\times_SY\simeq X\times_TY_T$.
Let
$p_\#\colon \PSh(\category{Cor}_T)\to \PSh(\category{Cor}_S)$
be the extension of $p_\star$ by colimits.
The adjunction $p_\star\dashv p^\star$ shows that $p_\#$ is left adjoint to $p^*$:
$$
p_\#\colon \PSh(\category{Cor}_T)\rightleftarrows \PSh(\category{Cor}_S)\colon p^*.
$$
Moreover, $p^*$ is given by precomposition with $p_\star$.
\end{definition}
\subsection{Pro-smooth morphisms}
As we saw above, if $p\colon T\to S$ is a morphism of notherian schemes such that $T\in \category{Sm}_S$, then the functor $p^*\colon \PSh(\category{Cor}_S)\to \PSh(\category{Cor}_T)$ admits a left adjoint $p_\#$.
In fact, this can be generalized to \emph{pro-smooth} morphisms which we define here.
\begin{definition}[Pro-smooth morphisms]\label{def:pro_smooth}
Let $X$ be a noetherian $S$-scheme.
We say that $X\to S$ is \emph{pro-smooth} if there is a projective system $(X_i)_{i\in I}$ in $\category{Sm}_S$ such that
\begin{enumerate}
\item the transition morphisms $\rho_{ij}\colon X_j\to X_i$ are affine and flat, and
\item $X\simeq \varprojlim_{i\in I}X_i$ as an $S$-scheme.
\end{enumerate}
\end{definition}
In this situation, the canonical homomorphism
$$
\textstyle\varinjlim_{i\in I}c_S(X_i,Y)\to c_S(X,Y)
$$
is an isomorphism for all $Y\in \category{Sm}_S$ \cite[Proposition 9.3.9]{CD}.
A standard limit argument shows the following
\begin{lemma}\label{lem:psm}
Let $p\colon T\to S$ be a pro-smooth morphism of noetherian schemes.
Then any $X\in \category{Sm}_T$ is pro-smooth over $S$.
\end{lemma}
Since a pro-left adjoint can be constructed objectwise, we get the following
\begin{lemma}
Let $p\colon T\to S$ be a pro-smooth morphism of noetherian schemes.
Then $p^\star \colon \category{Cor}_S\to \category{Cor}_T$ admits a pro-left adjoint $p_\star\colon \category{Cor}_T\dashrightarrow \category{Cor}_S$ which is given by
$X\mapsto \quotprojlim_{i\in I} X_i$, where $(X_i)_{i\in I}$ is a projective system in $\category{Sm}_S$ as in Definition \ref{def:pro_smooth}, whose existence is guaranteed by Lemma \ref{lem:psm}.
\end{lemma}
\begin{definition}
We define the pro-functor
$p_\#\colon \PSh(\category{Cor}_T)\dashrightarrow \PSh(\category{Cor}_S)$
to be the extension of $p_\star$ by colimits; note that $\Pro(\PSh(\category{Cor}_S))$ is cocomplete by \cite[Proposition 11.1]{Isaksen01}.
Since $p_\star$ is pro-left adjoint to $p^\star$, $p_\#$ is pro-left adjoint to $p^*$:
$$
\xymatrix{
p_\# \colon \PSh(\category{Cor}_T) \ar@<0.5ex>@{-->}[r]&\ar@<0.5ex>[l]\PSh(\category{Cor}_S) \colon p^*.
}
$$
Moreover, $p^*$ is given by precomposition with $p_\star$.
\end{definition}
\subsection{Nisnevich sheaves}
We say that $F\in \PSh(\category{Cor}_S)$ is a Nisnevich sheaf if $F|_{\category{Sm}_S}$ is a Nisnevich sheaf.
We write $\Sh_\Nis(\category{Cor}_S)$ for the full subcategory of $\PSh(\category{Cor}_S)$ consisting of Nisnevich sheaves.
It is an abelian category by the following result:
\begin{lemma}
The inclusion functor $\Sh_\Nis(\category{Cor}_S)\to \PSh(\category{Cor}_S)$ admits a left adjoint $a_\Nis$ whose restriction to $\PSh(\category{Sm}_S)$ is the usual Nisnevich sheafification.
\end{lemma}
\begin{proof}
See \cite[Proposition 10.3.9]{CD}.
\end{proof}
Let $f\colon T\to S$ be a morphism of noetherian schemes.
Then for any $F\in \Sh_\Nis(\category{Cor}_T)$, the presheaf $f_*F$ is a Nisnevich sheaf.
If $f$ is pro-smooth, then for any $F\in \Sh_\Nis(\category{Cor}_S)$, the presheaf $f^*F$ is also a Nisnevich sheaf.
We are mainly interested in Nisnevich sheaves with injectivity properties.
\begin{definition}
Assume that $S$ is reduced and let $F\in \Sh_\Nis(\category{Cor}_S)$.
\begin{enumerate}
\item We say that $F$ has {\it global injectivity} if for any $X\in \category{Sm}_S$ and any dense open subset $U\subset X$, the homomorphism $F(X)\to F(U)$ is injective.
\item Let $j\colon \coprod_{\eta\in S^{(0)}}\Spec k(\eta)\to S$ be the canonical pro-smooth morphism.
We say that $F$ has {\it horizontal injectivity} if the unit morphism
$\textstyle F\to j_*j^*F$
is a monomorphism.
In general, we define $a_\mathrm{inj} F$ to be the image of this morphism in $\Sh_\Nis(\category{Cor}_S)$.
\end{enumerate}
\end{definition}
Clearly, global injectivity is stronger than horizontal injectivity.
By definition of $a_\mathrm{inj}$, we have the following
\begin{lemma}
Assume that $S$ is reduced and let $F\in \Sh_\Nis(\category{Cor}_S)$.
Then $a_\mathrm{inj} F$ has horizontal injectivity.
If $G\in \Sh_\Nis(\category{Cor}_S)$ has horizontal injectivity, then the canonical homomorphism
$\Hom(a_\mathrm{inj} F,G)\to \Hom(F,G)$
is an isomorphism.
\end{lemma}
\section{$\mathbb{Q}$-modulus pairs}\label{sec:modpair}
In this section we define the notion of $\mathbb{Q}$-modulus pair, which plays a central role in this paper.
We extend the theory of presheaves with transfers to $\mathbb{Q}$-modulus pairs.
\subsection{Definition of $\mathbb{Q}$-modulus pairs}
\begin{definition}[$\mathbb{Q}$-modulus pairs]
A {\it $\mathbb{Q}$-modulus pair} $\mathcal{X}$ is a pair $(X,D_X)$ where
\begin{itemize}
\item $X$ is a finitely reducible scheme, and
\item $D_X$ is a $\mathbb{Q}$-effective $\mathbb{Q}$-Cartier divisor on $X$.
\end{itemize}
We write $X^\circ := X\setminus |D_X|$ and call it the \emph{interior} of $\mathcal{X}$.
\end{definition}
\begin{definition}[Ambient morphisms]
An \emph{ambient morphism} $f\colon \mathcal{X}\to \mathcal{Y}$ of $\mathbb{Q}$-modulus pairs is a morphism $f\colon X\to Y$ of schemes such that $f(X^\circ)\subset Y^\circ$ and $D_X\succ f^*D_Y$ holds.
Such a morphism induces a morphism $f^\circ\colon X^\circ \to Y^\circ$ of schemes.
\begin{enumerate}
\item We say that $f$ is \emph{minimal} if $D_X=f^*D_Y$ holds.
\item Let P be a property of morphisms of schemes.
We say that $f$ satisfies P if the underlying morphism of schemes $f\colon X\to Y$ satisfies P.
We say that $f$ satisfies interiorly P (int-P for short) if $f^\circ\colon X^\circ\to Y^\circ$ satisfies P.
\end{enumerate}
\end{definition}
\begin{remark}
Any finitely reducible scheme $X$ can be seen as a $\mathbb{Q}$-modulus pair $(X,\emptyset)$, for which we simply write $X$.
Any morphism $f\colon X\to Y$ between such schemes can be seen as a minimal ambient morphism $f\colon (X,\emptyset)\to (Y,\emptyset)$.
\end{remark}
\begin{definition}[Category $\category{QSm}_S$]
Let $S$ be a scheme.
We write $\category{QSm}_S$ for the category whose objects are $\mathbb{Q}$-modulus pairs over $S$ which are int-smooth, of finite type, and separated over $S$, and whose morphisms are ambient morphsims over $S$.
$\category{QSm}_S$ admits a symmetric monoidal structure given by the {\it box product}
$$
\mathcal{X}\boxtimes \mathcal{Y} := (X\times_S Y, \pr_1^*D_X+\pr_2^*D_Y)
$$
with unit object $S$.
We write $\category{QSm}_S^\mathrm{prop}$ for the full subcategory of $\category{QSm}_S$ spanned by objects which are proper over $S$.
\end{definition}
\subsection{Modulus correspondences}
\begin{definition}
Let $S$ be a noetherian scheme.
Let $\mathcal{X},\mathcal{Y}\in \category{QSm}_S$ and let $V\subset X^\circ\times_S Y^\circ$ be an integral closed subscheme.
Let $\overline{V}$ denote the scheme-theoretic closure of $V$ in $X\times_S Y$.
\begin{enumerate}
\item We say that $V$ is \emph{left proper} if $\overline{V}$ is proper over $X$.
Note that this condition is vacuous when $\mathcal{Y}$ is proper over $S$.
\item We say that $V$ is \emph{admissible} if the weak inequality
$(\pr_1^*D_X)|_{\overline{V}}\succ (\pr_2^*D_Y)|_{\overline{V}}$
holds.
\end{enumerate}
A \emph{modulus correspondence} from $\mathcal{X}$ to $\mathcal{Y}$ is an element of $c_S(X^\circ,Y^\circ)$ whose components are left proper and admissible.
We write $c_S(\mathcal{X},\mathcal{Y})$ for the abelian group of modulus correspondences from $\mathcal{X}$ to $\mathcal{Y}$.
\end{definition}
Note that for any ambient morphism $f\colon \mathcal{X}\to \mathcal{Y}$ over $S$, the graph of $f^\circ\colon X^\circ\to Y^\circ$ gives a modulus correspondence from $\mathcal{X}$ to $\mathcal{Y}$, for which we simply write $f$.
\begin{remark}
The assumption that $S$ is noetherian is necessary for using the machinery of relative cycles developed by Cisinski-D\'eglise.
\end{remark}
\begin{lemma}\label{lem:admissibility_comparison}
Let $\mathcal{X},\mathcal{Y}\in \category{QSm}_S$ and let $V\subset X^\circ \times_SY^\circ$ be an integral closed subscheme.
The following conditions are equivalent:
\begin{enumerate}
\item $V$ is admissible.
\item For any valuation ring $R$ and any morphism $\rho\colon \Spec R\to X\times_SY$ which sends the generic point into $V$, we have $\rho^*\pr_1^*D_X\geq \rho^*\pr_2^*D_Y$.
\end{enumerate}
\end{lemma}
\begin{proof}
This is an immediate consequence of Lemma \ref{lem:we-valuative}.
\end{proof}
\begin{lemma}\label{lem:composition_admissible}
Let $\mathcal{X},\mathcal{Y},\mathcal{Z}\in \category{QSm}_S$ and $\alpha\in c_S(\mathcal{X},\mathcal{Y})$, $\beta\in c_S(\mathcal{Y},\mathcal{Z})$.
Then $\beta\circ \alpha\in c_S(\mathcal{X},\mathcal{Z})$.
\end{lemma}
\begin{proof}
First we show that every component of $\beta\circ\alpha$ is left proper.
Since any component of $\beta\circ\alpha$ is the image of some component of $\beta\otimes_Y \alpha$, it suffices to prove that $\overline{|\beta\otimes_Y \alpha|}\subset X\times_SY\times_SZ$ is proper over $X$.
Our assumption implies that $\overline{|\alpha|}$ is proper over $X$ and $\overline{|\beta|}$ is proper over $Y$.
Therefore $\overline{|\alpha|}\times_Y\overline{|\beta|}$ is proper over $X$.
By Lemma \ref{lem:pullback_lemma} (\ref{lem:support}), $|\beta\otimes_Y \alpha|$ is a closed subscheme of $\overline{|\alpha|}\times_Y\overline{|\beta|}$, so its closure is proper over $X$.
Next we prove that every component $V$ of $\beta\circ \alpha$ is admissible.
As we saw above, $V$ is the image of some component $V'$ of $\beta\otimes_Y \alpha$, and $\overline{V'}\to \overline{V}$ is proper.
By Lemma \ref{lem:we-pb} (2), it suffices to prove that $(\pr_1^*D_X)|_{\overline{V'}}\succ (\pr_3^*D_Z)|_{\overline{V'}}$ holds.
Let $R$ be a valuation ring and $\rho\colon \Spec R\to X\times_SY\times_SZ$ be a morphism which sends the generic point into $V'$.
By Lemma \ref{lem:we-valuative}, it suffices to check that
$\rho^*\pr_1^*D_X\geq \rho^*\pr_3^*D_Z$ holds.
Recall that we have a morphism of cycles $\beta\otimes_Y\alpha\to \alpha$.
By our hypothesis that $\alpha$ has admissible components, we can apply Lemma \ref{lem:admissibility_comparison} to the morphism
$$
\Spec R\xrightarrow{\rho}X\times_SY\times_SZ\xrightarrow{\pr_{12}}X\times_SY
$$
to get $\rho^*\pr_1^*D_X\geq \rho^*\pr_2^*D_Y$.
Similarly we have $\rho^*\pr_2^*D_Y\geq \rho^*\pr_3^*D_Z$, so we get the desired inequality.
\end{proof}
\begin{definition}\label{def_QCor}
We define $\category{QCor}_S$ to be the category whose objects are the same as $\category{QSm}_S$ and whose morphisms are modulus correspondences over $S$.
Note that we have a faithful functor $\category{QSm}_S\hookrightarrow \category{QCor}_S$ given by $f\mapsto [\Gamma_{f^\circ}]$.
We write $\category{QCor}_S^\mathrm{prop}$ for the full subcategory of $\category{QCor}_S$ spanned by objects which are proper over $S$.
\end{definition}
\begin{remark}
Let $\mathcal{X},\mathcal{Y}\in \category{QSm}_S$ and let $f\colon \mathcal{X}\to \mathcal{Y}$ be a proper minimal ambient morphism such that $f^\circ\colon X^\circ \to Y^\circ$ is an isomorphism.
Then $f$ becomes an isomorphism in $\category{QCor}_S$, since the transpose ${}^tf^\circ$ of $f^\circ$ gives an inverse of $f$.
This will be used without mentioning in the rest of the paper.
\end{remark}
The following lemma shows that the box product gives a symmetric monoidal structure $\boxtimes$ on $\category{QCor}_S$ with unit $S$.
This restricts to a symmetric monoidal structure on $\category{QCor}_S^\mathrm{prop}$.
\begin{lemma}
Let $\mathcal{X}_1,\mathcal{Y}_1,\mathcal{X}_2,\mathcal{Y}_2\in \category{QSm}_S$ and $\alpha\in c_S(\mathcal{X}_1,\mathcal{Y}_1)$, $\beta\in c_S(\mathcal{X}_2,\mathcal{Y}_2)$.
Then we have $\alpha\times \beta \in c_S(\mathcal{X}_1\boxtimes\mathcal{X}_2,\mathcal{Y}_1\boxtimes\mathcal{Y}_2)$.
\end{lemma}
\begin{proof}
First we prove that every component of $\alpha\times\beta$ is left proper.
By the construction of $\alpha\times \beta$ and Lemma \ref{lem:pullback_lemma} (\ref{lem:support}), $|\alpha\times \beta|$ is a closed subscheme of $|\alpha|\times_S|\beta|$.
Since $\overline{|\alpha|}$ is proper over $X_1$ and $\overline{|\beta|}$ is proper over $X_2$, it follows that $\overline{|\alpha\times \beta|}$ is proper over $X_1\times_S X_2$.
Next we prove that every component $V$ of $\alpha$ is admissible.
Let $R$ be a valuation ring and $\rho\colon \Spec R\to X_1\times_SX_2\times_SY_1\times_SY_2$ be a morphism which sends the generic point into $V$.
By Lemma \ref{lem:admissibility_comparison}, it suffices to check that
$$
\rho^*(\pr_1^*D_{X_1}+\pr_2^*D_{X_2})\geq \rho^*(\pr_3^*D_{Y_1}+\pr_4^*D_{Y_2}).
$$
Recall that we have morphisms of cycles $\alpha\times \beta\to \alpha$.
By our hypothesis that $\alpha$ has admissible components, we can apply Lemma \ref{lem:admissibility_comparison} to the morphism
$$
\Spec R\xrightarrow{\rho}X_1\times_SX_2\times_SY_1\times_SY_2\xrightarrow{\pr_{13}}X_1\times_SY_1
$$
to get $\rho^*\pr_1^*D_{X_1}\geq \rho^*\pr_3^*D_{Y_1}$.
Similarly we have $\rho^*\pr_2^*D_{X_2}\geq \rho^*\pr_4^*D_{Y_2}$, so we get the desired inequality.
\end{proof}
\subsection{Compactifications}
Let $S$ be a noetherian scheme.
In this subsection we embed $\category{QCor}_S$ into $\Pro(\category{QCor}_S^\mathrm{prop})$ by considering compactifications.
\begin{definition}[Compactification]
A \emph{compactification} of $\mathcal{X}\in \category{QSm}_S$ is a triple $(\overline{X},\overline{D},\Sigma)$ where $\overline{X}$ is a proper $S$-scheme and $\overline{D},\Sigma$ are $\mathbb{Q}$-effective $\mathbb{Q}$-Cartier divisors on $\overline{X}$, equipped with an isomorphism $X\simeq \overline{X}\setminus |\Sigma|$ such that $D_X=\overline{D}|_X$.
We say that a compactification $(\overline{X}_1,\overline{D}_1,\Sigma_1)$ \emph{dominates} $(\overline{X}_2,\overline{D}_2,\Sigma_2)$ if the diagonal $\Delta_{X^\circ}\subset X^\circ\times_S X^\circ$ gives a modulus correspondence from $(\overline{X}_1,\overline{D}_1+\Sigma_1)$ to $(\overline{X}_2, \overline{D}_2+\Sigma_2)$.
This defines a category $\category{Comp}(\mathcal{X})$ of compactifications of $\mathcal{X}$.
\end{definition}
\begin{lemma}\label{Comp_is_cofiltered}
For any $\mathcal{X}\in \category{QSm}_S$, the category $\category{Comp}(\mathcal{X})$ is cofiltered.
\end{lemma}
\begin{proof}
First we prove that $\category{Comp}(\mathcal{X})$ is non-empty.
Write $D_X=qE$ where $E$ is an ordinary effective Cartier divisor on $X$ and $q\in \mathbb{Q}_{>0}$.
Using Nagata's compactification theorem (\cite{Nagata-imbedding}, \cite{Conrad-Deligne}), we can take an open immersion $X\to \overline{X}$ with $\overline{X}$ proper over $S$.
Blowing up along $(\overline{X}\setminus X)_\red$ if necessary, we may assume that $\overline{X}\setminus X$ is the support of an ordinary effective Cartier divisor $\Sigma$ on $\overline{X}$.
Moreover, blowing up along the scheme-theoretic closure of $E$ in $\overline{X}$ if necesssary, we may assume that $E$ extends to an ordinary effective Cartier divisor $\overline{E}$ on $\overline{X}$.
Then $(\overline{X},q\overline{E},\Sigma)$ gives an object of $\category{Comp}(\mathcal{X})$.
Next we prove that $\category{Comp}(\mathcal{X})$ is cofiltered.
Let $(\overline{X}_1,\overline{D}_1,\Sigma_1)$ and $(\overline{X}_2,\overline{D}_2,\Sigma_2)$ be compactifications of $\mathcal{X}$.
Let $\Gamma$ be the scheme-theoretic closure of the diagonal $\Delta_X\subset X\times_SX$ in $\overline{X}_1\times_S \overline{X}_2$ and let $p_1\colon \Gamma\to \overline{X}_1$, $p_2\colon \Gamma\to \overline{X}_2$ be the projections.
Then $(\Gamma,p_i^*\overline{D}_i,p_i^*\Sigma_i)$ dominates $(\overline{X}_i,\overline{D}_i,\Sigma_i)$ for $i=1,2$, and $(\Gamma,p_1^*\overline{D}_1,n\cdot p_1^*\Sigma_1)$ dominates both of them for $n\gg 0$ by Lemma \ref{lem:ineq_domination}.
Therefore $\category{Comp}(\mathcal{X})$ is cofiltered.
\end{proof}
From the above proof we also get the following
\begin{lemma}\label{cofinal_in_Comp}
Let $\mathcal{X}\in \category{QSm}_S$ and $(\overline{X},\overline{D},\Sigma)\in \category{Comp}(\mathcal{X})$.
Then $\{(\overline{X},\overline{D},n\Sigma)\}_{n>0}$ is cofinal in $\category{Comp}(\mathcal{X})$.
\end{lemma}
\begin{lemma}\label{lem:correspondence_domination}
Let $\mathcal{X},\mathcal{Y}\in \category{QSm}_S$, $\alpha\in c_S(\mathcal{X},\mathcal{Y})$,
$(\overline{X},\overline{D},\Sigma) \in \category{Comp}(\mathcal{X})$ and
$(\overline{Y},\overline{E},\Xi) \in \category{Comp}(\mathcal{Y})$.
Then $\alpha$ gives a modulus correspondence from $(\overline{X},\overline{D}+n\Sigma)$ to $(\overline{Y},\overline{E}+\Xi)$ for $n\gg 0$.
\end{lemma}
\begin{proof}
Let $V$ be a component of $\alpha$, $\overline{V}$ be its scheme-theoretic closure in $X\times_SY$, and $\widetilde{V}$ be its scheme-theoretic closure in $\overline{X}\times_S \overline{Y}$.
Let $p\colon \widetilde{V}\to \overline{X}$ and $q\colon \widetilde{V}\to \overline{Y}$ be the canonical projections.
Since $\overline{V}$ is proper over $X$, $\widetilde{V}\setminus \overline{V}$ is contained in $|p^*\Sigma|$.
On the other hand, since $\alpha$ gives a modulus correspondence from $\mathcal{X}$ to $\mathcal{Y}$, we have $p^*\overline{D}|_{\overline{V}}\succ q^*\overline{E}|_{\overline{V}}$.
By Lemma \ref{lem:ineq_domination}, we get $p^*(\overline{D}+n\Sigma)\succ q^*(\overline{E}+\Xi)$ for $n\gg 0$.
\end{proof}
\begin{corollary}\label{cor:pre_embedding}
Let $\mathcal{X},\mathcal{Y}\in \category{QSm}_S$ and $\alpha\in c_S(\mathcal{X},\mathcal{Y})$.
Then we have
$$
c_S(\mathcal{X},\mathcal{Y}) = \bigcap_{(\overline{Y},\overline{E},\Xi)\in \category{Comp}(\mathcal{Y})}\bigcup_{(\overline{X},\overline{D},\Sigma)\in \category{Comp}(\mathcal{X})}c_S((\overline{X},\overline{D}+\Sigma),(\overline{Y},\overline{E}+\Xi))
$$
as a subgroup of $c_S(X^\circ, Y^\circ)$.
\end{corollary}
By Corollary \ref{cor:pre_embedding}, we get a fully faithful embedding
$$
\category{QCor}_S\to \Pro(\category{QCor}_S^\mathrm{prop});\quad \mathcal{X}\mapsto \quotprojlim_{(\overline{X},\overline{D},\Sigma)\in \category{Comp}(\mathcal{X})}(\overline{X},\overline{D}+\Sigma).
$$
This functor is symmetric monoidal by Lemma \ref{cofinal_in_Comp}.
Throughout the rest of the paper, we always regard $\category{QCor}_S$ as a full subcategory of $\Pro(\category{QCor}_S^\mathrm{prop})$.
\subsection{Modulus presheaves with transfers}
We fix a noetherian scheme $S$.
\begin{definition}[Modulus presheaves with transfers]
An object of the category $\PSh(\category{QCor}_S^\mathrm{prop})$ is called a \emph{modulus presheaf with transfers} over $S$.
We write $\mathbb{Z}_\tr(\mathcal{X})$ for the object of $\PSh(\category{QCor}_S^\mathrm{prop})$ represented by $\mathcal{X}\in \category{QCor}_S^\mathrm{prop}$.
The box product $\boxtimes$ extends by colimits to a symmetric monoidal structure $\otimes$ on $\PSh(\category{QCor}_S^\mathrm{prop})$ with unit $\mathbb{Z} = \mathbb{Z}_\tr(S)$.
This symmetric monoidal structure is closed:
$$
\mathcal{H}\mathrm{om}(F,G)(\mathcal{X}) =\Hom(\mathbb{Z}_\tr(\mathcal{X})\otimes F, G).
$$
\end{definition}
\begin{remark}
For any $F\in \PSh(\category{QCor}_S^\mathrm{prop})$, we can extend $F$ to a presheaf on $\Pro(\category{QCor}_S^\mathrm{prop})$ by setting
$F(\quotprojlim_{\lambda} \mathcal{X}_\lambda):=\varinjlim_{\lambda} F(\mathcal{X}_\lambda)$.
In particular, we can define $F(\mathcal{X})$ for any $\mathcal{X}\in \category{QSm}_S$.
\end{remark}
Let $f\colon T\to S$ be a morphism of noetherian schemes and $\mathcal{X}\in \category{QSm}_S^\mathrm{prop}$.
Let $X_T^*$ denote the scheme-theoretic closure of $(X^\circ)_T$ in $X_T$.
Then $(X^\circ)_T$ is scheme-theoretically dense in $X_T^*$, so the pullback $(D_X)_T$ of $D_X$ by $X_T^*\to X$ is defined (Lemma \ref{lem:pb-criteria-3}).
We set $\mathcal{X}_T:=(X_T^*,(D_X)_T)\in \category{QSm}_T^\mathrm{prop}$.
\begin{lemma}
Let $f\colon T\to S$ be a morphism of noetherian schemes and $\mathcal{X},\mathcal{Y}\in \category{QSm}_S^\mathrm{prop}$.
If $\alpha\in c_S(\mathcal{X},\mathcal{Y})$, then $\alpha_T\in c_T(\mathcal{X}_T,\mathcal{Y}_T)$.
\end{lemma}
\begin{proof}
Recall that $\alpha_T=\alpha\otimes_{X^\circ} (X^\circ)_T$ by definition.
It suffices to prove that every component $V$ of $\alpha_T$ is admissible.
Let $R$ be a valuation ring and $\rho\colon \Spec R\to X_T^*\times_TY_T^*$ be a morphism which sends the generic point into $V$.
By Lemma \ref{lem:admissibility_comparison}, it suffices to prove $\rho^*\pr_1^*(D_X)_T\geq \rho^*\pr_2^*(D_Y)_T$.
Recall that we have a morphism of cycles $\alpha_T\to \alpha$.
Since $\alpha$ has admissible components, we can apply Lemma \ref{lem:admissibility_comparison} to the morphism
$$
\Spec R\xrightarrow{\rho}X_T^*\times_TY_T^*\to X\times_SY
$$
to get the desired inequality.
\end{proof}
\begin{definition}[Base change functor]
Let $f\colon T\to S$ be a morphism of noetherian schemes.
We define the {\it base change functor} $f^\star \colon \category{QCor}_S^\mathrm{prop} \to \category{QCor}_T^\mathrm{prop}$ by $\mathcal{X}\mapsto \mathcal{X}_T$.
This induces an adjunction
$$
f^*\colon \PSh(\category{QCor}_S^\mathrm{prop})\rightleftarrows \PSh(\category{QCor}_T^\mathrm{prop})\colon f_*.
$$
The functor $f^*$ is the extension of $f^\star$ by colimits and the functor $f_*$ is given by precomposition with $f^\star$.
\end{definition}
Next we let $p\colon T\to S$ be a morphism of noetherian schemes such that $T\in \category{Sm}_S$.
Then any $\mathcal{X}\in \category{QSm}_T^\mathrm{prop}$ can also be seen as an object of $\category{QSm}_S$; note that $\mathcal{X}$ is not necessarily proper over $S$.
\begin{lemma}
Let $p\colon T\to S$ be a morphism of noetherian schemes such that $T\in \category{Sm}_S$.
Let $\mathcal{X},\mathcal{Y}\in \category{QSm}_T^\mathrm{prop}$.
If $\alpha\in c_T(\mathcal{X},\mathcal{Y})$, then $\alpha^S\in c_S(\mathcal{X},\mathcal{Y})$.
\end{lemma}
\begin{proof}
Recall that $\alpha^S=\delta_*\alpha$ where $\delta\colon X^\circ\times_TY^\circ\to X^\circ\times_SY^\circ$ is the canonical closed immersion.
If we write $\alpha = \sum_{i=1}^n m_i[V_i]$ in the standard form, then we have $\alpha^S = \sum_{i=1}^n m_i[\delta(V_i)]$.
The scheme-theoretic closure $\overline{\delta(V_i)}$ in $X\times_SY$ coincides with the scheme-theoretic image $\overline{\delta}(\overline{V_i})$, where $\overline{\delta}\colon X\times_TY\to X\times_SY$ is the canonical closed immersion.
Therefore the left properness and the admissibility of $V_i$ imply those of $\delta(V_i)$.
\end{proof}
\begin{definition}[Restriction pro-functor]
Let $p\colon T\to S$ be a morphism of noetherian schemes such that $T\in \category{Sm}_S$.
We define the {\it restriction pro-functor}
$p_\star\colon \category{QCor}_T^\mathrm{prop}\dashrightarrow \category{QCor}_S^\mathrm{prop}$
to be the composition
$\category{QCor}_T^\mathrm{prop}\to \category{QCor}_S\hookrightarrow \Pro(\category{QCor}_S^\mathrm{prop})$,
where the first functor is given by $\mathcal{X}\mapsto \mathcal{X}$ and the second is the canonical embedding.
\end{definition}
\begin{lemma}\label{lem:adj_smooth_Q}
Let $p\colon T\to S$ be a morphism of noetherian schemes such that $T\in \category{Sm}_S$.
Then the restriction pro-functor $p_\star$ is pro-left adjoint to the base change functor $p^\star$:
$$
\xymatrix{
p_\star \colon \category{QCor}_T^\mathrm{prop} \ar@<0.5ex>@{-->}[r]&\ar@<0.5ex>[l]\category{QCor}_S^\mathrm{prop} \colon p^\star.
}
$$
\end{lemma}
\begin{proof}
Recall that the isomorphism $c_S(X^\circ, Y^\circ) \simeq c_T(X^\circ, Y^\circ_T)$ is given by the canonical identification $X^\circ\times_S Y^\circ\simeq X^\circ\times_TY^\circ_T$.
It suffices to check that this restricts to an isomorphism $c_S(\mathcal{X},\mathcal{Y})\simeq c_T(\mathcal{X},\mathcal{Y}_T)$.
Since $p$ is flat, $(X^\circ)_T$ is scheme-theoretically dense in $X_T$ and hence $X_T^*=X_T$.
Therefore the claim follows from the canonical identification $X\times_SY\simeq X\times_TY_T$.
\end{proof}
\begin{definition}
Let $p\colon T\to S$ be a morphism of noetherian schemes such that $T\in \category{Sm}_S$.
We define the pro-functor
$p_\#\colon \PSh(\category{QCor}_T^\mathrm{prop})\dashrightarrow\PSh(\category{QCor}_S^\mathrm{prop})$
to be the extension of $p_\star$ by colimits.
Since $p_\star$ is pro-left adjoint to $p^\star$, $p_\#$ is pro-left adjoint to $p^*$:
$$
\xymatrix{
p_\# \colon \PSh(\category{QCor}_T^\mathrm{prop}) \ar@<0.5ex>@{-->}[r]&\ar@<0.5ex>[l]\PSh(\category{QCor}_S^\mathrm{prop}) \colon p^*.
}
$$
Moreover, $p^*$ is given by precomposition with $p_\star$.
\end{definition}
\subsection{Relation to the classical theory}
\begin{definition}[Functor $\omega$]
We define $\omega\colon \category{QCor}_S^\mathrm{prop}\to \category{Cor}_S$ by $\mathcal{X}\to X^\circ$.
This induces an adjunction
$$
\omega_!\colon \PSh(\category{QCor}_S^\mathrm{prop})\rightleftarrows \PSh(\category{Cor}_S)\colon \omega^*.
$$
The functor $\omega_!$ is the extension of $\omega$ by colimits and $\omega^*$ is given by precomposition with $\omega$.
\end{definition}
\begin{definition}[Compactification functor]
We define $\comp\colon \category{Cor}_S\hookrightarrow \Pro(\category{QCor}_S^\mathrm{prop})$ to be the composition of canonical embeddings
$\category{Cor}_S\hookrightarrow \category{QCor}_S \hookrightarrow \Pro(\category{QCor}_S^\mathrm{prop})$.
Then $\comp$ is pro-left adjoint to $\omega$.
We define a pro-functor
$\comp_!\colon \PSh(\category{Cor}_S)\dashrightarrow \PSh(\category{QCor}_S^\mathrm{prop})$
to be the extension of $\comp$ by colimits.
Then $\comp_!$ is pro-left adjoint to $\omega_!$:
$$
\xymatrix{
\comp_! \colon \PSh(\category{Cor}_S) \ar@<0.5ex>@{-->}[r]&\ar@<0.5ex>[l]\PSh(\category{QCor}_S^\mathrm{prop}) \colon \omega_!.
}
$$
Moreover, $\omega_!$ is given by precomposition with $\comp$.
\end{definition}
For $X\in \category{Sm}_S$, using Nagata's compactification theorem, we can choose $\mathcal{Y}\in \category{QSm}_S^\mathrm{prop}$ together with an isomorphism of $S$-schemes $Y^\circ \simeq X$.
Then $\{(Y,\emptyset,nD_Y)\}_{n>0}$ is cofinal in $\category{Comp}(X)$ (Lemma \ref{cofinal_in_Comp}), so $\comp(X)$ can be written as
$\comp(X)=\quotprojlim_{n>0}(Y,nD_Y)$.
Since $\omega_!$ is given by precomposition with $\comp$, we get
$\omega_!F(X)=\varinjlim_{n>0} F(Y,nD_Y)$.
We can also see from this description that $\comp$ is symmetric monoidal.
\begin{proposition}\label{omega_star_and_shriek}
The following assertions hold:
\begin{enumerate}
\item The functor $\omega^*$ is exact and fully faithful.
\item The functor $\omega_!$ is exact and symmetric monoidal.
\end{enumerate}
\end{proposition}
\begin{proof}
Both functors are clearly exact.
Moreover, $\omega\circ\comp\simeq \id$ implies $\omega_!\omega^*\simeq\id$ and hence $\omega^*$ is fully faithful.
Since $\omega$ is symmetric monoidal and $\omega_!$ is its extension by colimits, $\omega_!$ is also symmetric monoidal.
\end{proof}
\begin{lemma}\label{restriction_compatibility}
The following assertions hold:
\begin{enumerate}
\item
Let $f\colon T\to S$ be a morphism of noetherian schemes.
Then the following diagrams are commutative up to canonical isomorphism:
$$
\xymatrix{
\category{Cor}_S\ar[r]^-{\comp}\ar[d]^-{f^\star} &\Pro(\category{QCor}_S^\mathrm{prop})\ar[d]^-{\Pro(f^\star)} &\category{QCor}_S^\mathrm{prop}\ar[r]^-{\omega} \ar[d]^-{f^\star} &\category{Cor}_S\ar[d]^-{f^\star}\\
\category{Cor}_T\ar[r]^-{\comp} &\Pro(\category{QCor}_T^\mathrm{prop}) &\category{QCor}_T^\mathrm{prop}\ar[r]^-{\omega} &\category{Cor}_T.
}
$$
\item
Let $p\colon T\to S$ be a morphism of noetherian schemes such that $T\in \category{Sm}_S$.
Then the following diagram is commutative up to canonical isomorphism:
$$
\xymatrix{
\category{QCor}_T^\mathrm{prop}\ar[r]^-{\omega} \ar@{-->}[d]^-{p_\star} &\category{Cor}_T\ar[d]^-{p_\star}\\
\category{QCor}_S^\mathrm{prop}\ar[r]^-{\omega} &\category{Cor}_S.
}
$$
\end{enumerate}
\end{lemma}
\begin{proof}
The only non-trivial part is the commutativity of the left square in (1).
Let $X\in \category{Sm}_S$ and choose $\mathcal{Y}\in \category{QCor}_S^\mathrm{prop}$ together with an isomorphism $Y^\circ\simeq X$ over $S$.
Then we have
\begin{align*}
\comp(f^\star X) \simeq{}&\quotprojlim_{n>0} (Y_T^*,n(D_Y)_T)\\
{} \simeq{}&\Pro(f^\star) \quotprojlim_{n>0} (Y,nD_Y)
{} \simeq{}\Pro(f^\star) \comp(X).
\end{align*}
Hence the left square in (1) is commutative up to canonical isomorphism.
\end{proof}
\begin{lemma}\label{presheaf_restriction_compatibility}
The following assertions hold.
\begin{enumerate}
\item
Let $f\colon T\to S$ be a morphism of noetherian schemes.
Then we have canonical isomorphisms
\begin{align*}
&f_*\omega_! \simeq \omega_!f_*\colon \PSh(\category{QCor}_T^\mathrm{prop})\to \PSh(\category{Cor}_S),\\
&f_*\omega^* \simeq \omega^*f_*\colon \PSh(\category{Cor}_T)\to \PSh(\category{QCor}_S^\mathrm{prop}),\\
&f^*\omega_! \simeq \omega_!f^*\colon \PSh(\category{QCor}_S^\mathrm{prop})\to \PSh(\category{Cor}_T).
\end{align*}
\item
Let $p\colon T\to S$ be a morphism of noetherian schemes such that $T\in \category{Sm}_S$.
Then we have canonical isomorphisms
\begin{align*}
&p^*\omega^*\simeq \omega^*p^*\colon \PSh(\category{Cor}_S)\to \PSh(\category{QCor}_T^\mathrm{prop}),\\
&p_\#\omega_! \simeq \omega_!p_\# \colon \PSh(\category{QCor}_T^\mathrm{prop})\dashrightarrow \PSh(\category{Cor}_S).
\end{align*}
\end{enumerate}
\end{lemma}
\begin{proof}
(1) Since $f_*$ (resp. $\omega_!,\omega^*$) is given by precomposition with $f^\star$ (resp. $\comp,\omega$), the first two formulas follow from Lemma \ref{restriction_compatibility} (1).
The third one follows from the second one by adjunction.
(2) Since $p^*$ (resp. $\omega^*$) is given by precomposition with $p_\star$ (resp. $\omega$), the first formula follows from Lemma \ref{restriction_compatibility} (2).
The second one follows from the first one by pro-adjunction.\qedhere
\end{proof}
\subsection{Pro-smooth $\mathbb{Q}$-modulus pairs}
As we have already seen, for a morphism $p\colon T\to S$ of notherian schemes such that $T\in \category{Sm}_S$, the functor $p^*\colon \PSh(\category{QCor}_S^\mathrm{prop})\to \PSh(\category{QCor}_T^\mathrm{prop})$ admits a pro-left adjoint $p_\#$.
This can be generalized to pro-smooth morphisms.
\begin{definition}[Pro-smooth $\mathbb{Q}$-modulus pairs]\label{def:pro_modulus_pair}
Let $\mathcal{X}$ be a $\mathbb{Q}$-modulus pair over $S$ such that $X$ is noetherian.
We say that $\mathcal{X}\to S$ is \emph{pro-smooth} if there is a projective system $(\mathcal{X}_i)_{i\in I}$ in $\category{QSm}_S$ such that
\begin{itemize}
\item the transition morphisms $\rho_{ij}\colon \mathcal{X}_j\to \mathcal{X}_i$ are affine, flat, and minimal,
\item $X\simeq \varprojlim_{i\in I} X_i$ as $S$-schemes, and
\item $D_X$ is the pullback of $D_{X_i}$.
\end{itemize}
\end{definition}
\begin{lemma}
In the above situation, the canonical homomorphism
$$
\textstyle\varinjlim_{i\in I}c_S(\mathcal{X}_i,\mathcal{Y})\to c_S(\mathcal{X},\mathcal{Y})
$$
is an isomorphism for all $\mathcal{Y}\in \category{QSm}_S$.
\end{lemma}
\begin{proof}
The colimit $\varinjlim_{i\in I}c_S(\mathcal{X}_i,\mathcal{Y})$ is a subgroup of $\varinjlim_{i\in I}c_S(X_i^\circ,Y^\circ)$.
Similarly, $c_S(\mathcal{X},\mathcal{Y})$ is a subgroup of $c_S(X^\circ,Y^\circ)$.
Therefore the claim follows from the isomorphism
$$
\textstyle\varinjlim_{i\in I}c_S(X_i^\circ,Y^\circ)\simeq c_S(X^\circ,Y^\circ)
$$
\cite[Proposition 9.3.9]{CD} and Lemma \ref{lem:ineq_limit}.
\end{proof}
A standard limit argument shows the following
\begin{lemma}\label{pro_smooth_composition}
Let $p\colon T\to S$ be a pro-smooth morphism of noetherian schemes.
Then any $\mathcal{X}\in \category{QSm}_T$ is pro-smooth over $S$.
\end{lemma}
Since a pro-left adjoint can be constructed objectwise, we get the following
\begin{lemma}
Let $p\colon T\to S$ be a pro-smooth morphism of noetherian schemes.
Then the functor $p^\star \colon \category{QCor}_S^\mathrm{prop}\to \category{QCor}_T^\mathrm{prop}$ has a pro-left adjoint $p_\star\colon \category{QCor}_T^\mathrm{prop}\dashrightarrow \category{QCor}_S^\mathrm{prop}$ which is given by
$\mathcal{X}\mapsto \varprojlim_{i\in I}\mathcal{X}_i$, where $(\mathcal{X}_i)_{i\in I}$ is a projective system in $\category{QSm}_S$ as in Definition \ref{def:pro_modulus_pair}, whose existence is guaranteed by Lemma \ref{pro_smooth_composition}.
Here, we regard each $\mathcal{X}_i$ as a pro-object of $\category{QCor}_S^\mathrm{prop}$.
\end{lemma}
\begin{definition}
We define the pro-functor
$p_\#\colon \PSh(\category{QCor}_T^\mathrm{prop})\dashrightarrow \PSh(\category{QCor}_S^\mathrm{prop})$
to be the extension of $p_\star$ by colimits.
Since $p_\star$ is pro-left adjoint to $p^\star$, $p_\#$ is pro-left adjoint to $p^*$:
$$
\xymatrix{
p_\# \colon \PSh(\category{QCor}_T^\mathrm{prop}) \ar@<0.5ex>@{-->}[r]&\ar@<0.5ex>[l]\PSh(\category{QCor}_S^\mathrm{prop}) \colon p^*.
}
$$
Moreover, $p^*$ is given by precomposition with $p_\star$.
\end{definition}
\section{Cube-invariance and reciprocity}\label{sec:ci-rec}
Fix a noetherian scheme $S$.
Following \cite{KSY2}, we introduce a class of modulus presheaves with transfers called \emph{cube-invariant presheaves} which is an analogue of the class of $\mathbb{A}^1$-invariant presheaves used in the classical theory of motives.
This leads to the notion of \emph{reciprocity presheaf}, which is a presheaf with transfers that can be ``lifted'' to a cube-invariant presheaf.
\subsection{The cube}
In the classical theory of motives, the affine line $\mathbb{A}^1_S$ plays the role of the unit interval.
In the world of $\mathbb{Q}$-modulus pairs, we use a compactified version of $\mathbb{A}^1_S$:
\begin{definition}[Cube]
The $\mathbb{Q}$-modulus pair $\bcube:=(\mathbb{P}^1_S,\{\infty\})$ is called the {\it cube} over $S$.
We write $\pi\colon \bcube\to S$ for the ambient morphism given by the canonical projection $\mathbb{P}^1_S\to S$ and $i_\varepsilon\colon S\to \bcube\quad(\varepsilon = 0,1)$ for the ambient morphism given by $\varepsilon\colon S\to \mathbb{P}^1_S$.
\end{definition}
\begin{definition}[Cube-homotopy]
$\mathcal{X},\mathcal{Y}\in \category{QSm}_S$.
Two modulus correspondences $\alpha_0,\alpha_1\in c_S(\mathcal{X}, \mathcal{Y})$ are called {\it cube-homotopic} if there is some $\gamma\in c_S(\mathcal{X}\boxtimes \bcube, \mathcal{Y})$ such that $\gamma\circ (\id\boxtimes i_\varepsilon)=\alpha_\varepsilon\;(\varepsilon = 0,1)$.
In this case we write $\alpha_0\sim \alpha_1$ and call $\gamma$ a \emph{cube-homotopy} between $\alpha_0$ and $\alpha_1$.
We say that $\alpha\in c_S(\mathcal{X}, \mathcal{Y})$ is a {\it cube-homotopy equivalence} if there is some $\beta\in c_S(\mathcal{Y},\mathcal{X})$ such that $\beta\circ \alpha\sim \id_{\mathcal{X}}$ and $\alpha\circ \beta\sim \id_{\mathcal{Y}}$ hold.
We call such $\beta$ a \emph{cube-homotopy inverse} of $\alpha$.
\end{definition}
\begin{lemma}
The relation $\sim$ is an equivalence relation.
\end{lemma}
\begin{proof}
Let $\mathcal{X},\mathcal{Y}\in \category{QSm}_S$ and $\alpha_i\in c_S(\mathcal{X},\mathcal{Y})\;(i=0,1,2)$.
Let $\gamma_{01}\in c_S(\mathcal{X}\boxtimes \bcube, \mathcal{Y})$ (resp. $\gamma_{12}\in c_S(\mathcal{X}\boxtimes \bcube, \mathcal{Y})$) be a cube-homotopy between $\alpha_0$ and $\alpha_1$ (resp. $\alpha_1$ and $\alpha_2$).
Then
$\gamma_{01}+\gamma_{12}-\alpha_1\circ (\id\boxtimes \pi)\colon \mathcal{X}\boxtimes \bcube\to \mathcal{Y}$
gives a cube-homotopy between $\alpha_0$ and $\alpha_2$.
\end{proof}
\begin{lemma}\label{cube_multiplication_admissible}
Consider the multiplication map
$\mu\colon \mathbb{A}^1_S\times_S \mathbb{A}^1_S\to \mathbb{A}^1_S;\;(x,y)\mapsto xy$.
The graph of $\mu$ gives a modulus correspondence from $\bcube\boxtimes\bcube$ to $\bcube$ over $S$.
\end{lemma}
\begin{proof}
See \cite[Lemma 5.1.1]{KMSY3}.
\end{proof}
\begin{lemma}\label{cube_equivalence}
For any $\mathcal{X}\in \category{QSm}_S^\mathrm{prop}$, the modulus correspondence
$\id\boxtimes \pi \in c_S(\mathcal{X}\boxtimes \bcube, \mathcal{X})$ is a cube-homotopy equivalence.
\end{lemma}
\begin{proof}
Let us prove that $\id\boxtimes i_\varepsilon\colon \mathcal{X}\to \mathcal{X}\boxtimes\bcube\;(\varepsilon = 0,1)$ gives a cube-homotopy inverse.
The composition $(\id\boxtimes \pi)\circ (\id\boxtimes i_\varepsilon)$ is the identity.
Set
$\gamma:=\id\boxtimes \mu\colon \mathcal{X}\boxtimes \bcube\boxtimes \bcube \to \mathcal{X}\boxtimes \bcube$.
Then $\gamma\circ (\id\boxtimes i_0)=(\id\boxtimes i_\varepsilon)\circ (\id\boxtimes \pi)$ and $\gamma\circ (\id\boxtimes i_1)=\id$, so we have $(\id\boxtimes i_\varepsilon)\circ (\id\boxtimes \pi)\sim \id$.
\end{proof}
\subsection{Cube-invariance}
\begin{definition}
We say that $F\in \PSh(\category{QCor}_S^\mathrm{prop})$ is {\it cube-invariant} if the map
$(\id \boxtimes \pi)^*\colon F(\mathcal{X})\to F(\mathcal{X}\boxtimes \bcube)$
is an isomorphism for all $\mathcal{X}\in \category{QSm}_S^\mathrm{prop}$.
\end{definition}
\begin{lemma}\label{cube_invariance_characterization}
For $F\in \PSh(\category{QCor}_S^\mathrm{prop})$, the following conditions are equivalent:
\begin{enumerate}
\item $F$ is cube-invariant.
\item For any $\mathcal{X}\in \category{QSm}_S^\mathrm{prop}$, the map $(\id \boxtimes i_0)^*\colon F(\mathcal{X}\boxtimes \bcube)\to F(\mathcal{X})$ is injective.
\item For any $\mathcal{X}\in \category{QSm}_S^\mathrm{prop}$, the map $(\id \boxtimes i_0)^*-(\id\boxtimes i_1)^*\colon F(\mathcal{X}\boxtimes \bcube)\to F(\mathcal{X})$ is $0$.
\item For any two cube-homotopic morphisms $\alpha_0,\alpha_1$ in $\category{QCor}_S^\mathrm{prop}$, we have $\alpha_0^*=\alpha_1^*$ on $F$.
\end{enumerate}
\end{lemma}
\begin{proof}
(1) $\iff$ (2) follows from $\pi\circ i_0=\id$.
(1)$\implies$(3) follows from $\pi\circ i_0=\pi\circ i_1 = \id$.
Let us prove (3)$\implies$(4).
Let $\gamma\colon \mathcal{X}\boxtimes \bcube\to \mathcal{Y}$ be a cube-homotopy between $\alpha_0$ and $\alpha_1$.
Then we have $\alpha_0^*=(\id\boxtimes i_0)^*\gamma^*=(\id\boxtimes i_1)^*\gamma^*=\alpha_1^*$ on $F$.
This shows that (3)$\implies$(4).
If (4) holds, then for any cube-homotopy equivalence $\alpha\in c_S(\mathcal{X},\mathcal{Y})$ in $\category{QCor}_S^\mathrm{prop}$, the map $\alpha^*\colon F(\mathcal{Y})\to F(\mathcal{X})$ is an isomorphism.
Therefore (1) follows from Lemma \ref{cube_equivalence}.
\end{proof}
\begin{lemma}
The class of cube-invariant objects in $\PSh(\category{QCor}_S^\mathrm{prop})$ is closed under taking subobjects, quotients and extensions.
\end{lemma}
\begin{proof}
The claim for subobjects follows from the equivalence of (1) and (2) in Lemma \ref{cube_invariance_characterization}.
The remaining assertions then follow by the five lemma.
\end{proof}
There are two canonical ways to make a presheaf on $\category{QCor}_S^\mathrm{prop}$ cube-invariant: one is to take the maximal cube-invariant quotient, and the other is to take the maximal cube-invariant subobject.
The former is called the \emph{cube-localization} and the latter is called the \emph{cube-invariant part}.
\begin{definition}
The \emph{cube-localization} of $F\in \PSh(\category{QCor}_S^\mathrm{prop})$ is defined by
$$
h_0^{\smallcube} F (\mathcal{X}) := \Coker (F(\mathcal{X}\boxtimes \bcube)\xrightarrow{d}F(\mathcal{X}))
$$
where $d = (\id\boxtimes i_0)^*-(\id\boxtimes i_1)^*$.
There is a canonical epimorphism $F\twoheadrightarrow h_0^{\smallcube} F$.
For $\mathcal{X}\in \category{QSm}_S^\mathrm{prop}$, we write $h_0^{\smallcube}(\mathcal{X})$ for $h_0^{\smallcube}\mathbb{Z}_\tr(\mathcal{X})$.
\end{definition}
\begin{lemma}\label{properties_of_lower_h0}
The following assertions hold.
\begin{enumerate}
\item For any $F\in \PSh(\category{QCor}_S^\mathrm{prop})$, $h_0^{\smallcube} F$ is a cube-invariant quotient of $F$.
\item Let $F,G\in \PSh(\category{QCor}_S^\mathrm{prop})$.
If $G$ is cube-invariant, then the canonical homomorphism
$\Hom(h_0^{\smallcube} F,G)\to \Hom(F,G)$
is an isomorphism.
In other words, $h_0^{\smallcube} F$ is the maximal cube-invariant quotient of $F$.
\item For any $\mathcal{X}\in \category{QSm}_S^\mathrm{prop}$, the morphism $h_0^{\smallcube}(\mathcal{X}\boxtimes \bcube)\to h_0^{\smallcube} (\mathcal{X})$ induced by $\id\boxtimes \pi$ is an isomorphism.
\end{enumerate}
\end{lemma}
\begin{proof}
(1) follows from the equivalence of (1) and (3) in Lemma \ref{cube_invariance_characterization}.
(2) The canonical morphism $G\to h_0^{\smallcube} G$ is an isomorphism by Lemma \ref{cube_invariance_characterization}.
The claim follows from this and (1).
(3) By the Yoneda lemma, it suffices to prove that if $F\in \PSh(\category{QCor}_S^\mathrm{prop})$ is cube-invariant then
$$
\Hom(h_0^{\smallcube}(\mathcal{X}),F)\to \Hom(h_0^{\smallcube}(\mathcal{X}\boxtimes \bcube),F)
$$
is an isomorphism.
By (2), this map can be identified with $(\id\boxtimes \pi)^*\colon F(\mathcal{X})\to F(\mathcal{X}\boxtimes \bcube)$,
which is an isomorphism since $F$ is cube-invariant.
\end{proof}
\begin{definition}
The \emph{cube-invariant part} of $F\in \PSh(\category{QCor}_S^\mathrm{prop})$ is defined by
$$
h^{0,\smallcube} F (\mathcal{X}) := \Hom(h_0^{\smallcube}(\mathcal{X}), F).
$$
\end{definition}
\begin{lemma}\label{properties_of_upper_h0}
The following assertions hold.
\begin{enumerate}
\item For any $F\in \PSh(\category{QCor}_S^\mathrm{prop})$, $h^{0,\smallcube} F$ is a cube-invariant subobject of $F$.
\item Let $F,G\in \PSh(\category{QCor}_S^\mathrm{prop})$.
If $F$ is cube-invariant, then the canonical homomorphism
$\Hom(F,h^{0,\smallcube} G)\to \Hom(F,G)$
is an isomorphism.
In other words, $h^{0,\smallcube} F$ is the maximal cube-invariant subobject of $F$.
\item The functor $h^{0,\smallcube}$ is right adjoint to $h_0^{\smallcube}$.
\end{enumerate}
\end{lemma}
\begin{proof}
(1) follows from Lemma \ref{properties_of_lower_h0} (3).
(2) If $F\in \PSh(\category{QCor}_S^\mathrm{prop})$ is cube-invariant, then the canonical morphism $h^{0,\smallcube} F\to F$ is an isomorphism.
The claim follows from this and (1).
(3) follows from (2) and Lemma \ref{properties_of_lower_h0} (2).
\end{proof}
\begin{remark}
In \cite{KSY2}, $h_0^{\smallcube}$ and $h^{0,\smallcube}$ are defined to be functors taking values in the category of cube-invariant presheaves.
On the other hand, we define $h_0^{\smallcube}$ and $h^{0,\smallcube}$ as \emph{endofunctors} of $\PSh(\category{QCor}_S^\mathrm{prop})$.
This has the advantage that $h^{0,\smallcube}$ becomes right adjoint to $h_0^{\smallcube}$.
\end{remark}
\begin{lemma}\label{lem:h0_tensor}
Let $F,G\in \PSh(\category{QCor}_S^\mathrm{prop})$.
\begin{enumerate}
\item If $G$ is cube-invariant, then $\mathcal{H}\mathrm{om}(F,G)$ is also cube-invariant.
\item The canonical epimorphism $h_0^{\smallcube}(F\otimes G)\twoheadrightarrow h_0^{\smallcube}((h_0^{\smallcube} F)\otimes G)$ is an isomorphism.
\end{enumerate}
\end{lemma}
\begin{proof}
(1) For any $\mathcal{X}\in \category{QSm}_S^\mathrm{prop}$, we have
$\mathcal{H}\mathrm{om}(F,G)(\mathcal{X}) =\Hom(\mathbb{Z}_\tr(\mathcal{X})\otimes F, G)$.
Since $G$ is cube-invariant, $\mathcal{H}\mathrm{om}(F,G)$ is cube-invariant when $F$ is representable.
Since $\mathcal{H}\mathrm{om}({-},G)$ turns colimits into limits, the same is true for a general $F$.
(2) For any $H\in \PSh(\category{QCor}_S^\mathrm{prop})$, we have
\begin{align*}
\Hom(h_0^{\smallcube}(F\otimes G),H)
\simeq{}&\Hom(F,\mathcal{H}\mathrm{om}(G,h^{0,\smallcube} H))&\text{($h_0^{\smallcube}\dashv h^{0,\smallcube}$)}\\
\simeq{}&\Hom(h_0^{\smallcube} F, \mathcal{H}\mathrm{om}(G,h^{0,\smallcube} H))&\text{(by (1))}\\
\simeq{}&\Hom(h_0^{\smallcube}((h_0^{\smallcube} F)\otimes G),H).&\text{($h_0^{\smallcube}\dashv h^{0,\smallcube}$)}
\end{align*}
Therefore we get the desired result by the Yoneda lemma.\qedhere
\end{proof}
\begin{remark}
By Lemma \ref{lem:h0_tensor}, we have $h_0^{\smallcube}(F\otimes G)\simeq h_0^{\smallcube}((h_0^{\smallcube} F)\otimes (h_0^{\smallcube} G))$ for any $F,G\in \PSh(\category{QCor}_S^\mathrm{prop})$.
The latter is a quotient of $h_0^{\smallcube} F\otimes h_0^{\smallcube} G$, so we get an epimorphism
$h_0^{\smallcube} F\otimes h_0^{\smallcube} G \twoheadrightarrow h_0^{\smallcube}(F\otimes G)$.
In particular, $h_0^{\smallcube}$ admits a canonical lax symmetric monoidal structure.
\end{remark}
\begin{definition}\label{def:Lh0}
We define two functors $h_0$ and $h^0$ as follows.
\begin{enumerate}
\item We define $h_0 := \omega_! h_0^{\smallcube} \colon \PSh(\category{QCor}_S^\mathrm{prop})\to \PSh(\category{Cor}_S)$.
$h_0 F$ is called the \emph{$0$-th Suslin homology} of $F$.
Since $h_0^{\smallcube}$ is lax symmetric monoidal and $\omega_!$ is symmetric monoidal, it follows that $h_0$ is lax symmetric monoidal.
For $\mathcal{X}\in \category{QSm}_S^\mathrm{prop}$, we write $h_0(\mathcal{X})$ for $h_0\mathbb{Z}_\tr(\mathcal{X})$.
\item We define $h^0:=h^{0,\smallcube}\omega^*\colon \PSh(\category{Cor}_S)\to \PSh(\category{QCor}_S^\mathrm{prop})$.
$h^0 F$ is called the \emph{$0$-th Suslin cohomology} of $F$.
It follows from the adjunctions $h_0^{\smallcube}\dashv h^{0,\smallcube}$ and $\omega_!\dashv \omega^*$ that $h^0$ is right adjoint to $h_0$.
\end{enumerate}
\end{definition}
\begin{remark}
In \cite{KSY2}, a functor named $\omega^{\CI}$ is defined by the same formula as $h^0$, but it is regarded as a functor taking values in the category of cube-invariant presheaves rather than $\PSh(\category{QCor}_S^\mathrm{prop})$.
\end{remark}
The functors we have defined so far can be summarized as follows.
$$
\xymatrix{
\PSh(\category{QCor}_S^\mathrm{prop})
\ar@<1ex>[r]^-{h_0^{\smallcube}}
\ar@/^23pt/[rr]^-{h_0}
&
\PSh(\category{QCor}_S^\mathrm{prop})
\ar@<1ex>[r]^-{\omega_!}
\ar@<1ex>[l]^-{h^{0,\smallcube}}
\ar@{}[l]|-{\text{\rotatebox{-90}{$\dashv$}}}
&
\PSh(\category{Cor}_S)
\ar@<1ex>[l]^-{\omega^*}
\ar@{}[l]|-{\text{\rotatebox{-90}{$\dashv$}}}
\ar@/^23pt/[ll]^-{h^0}
}
$$
\begin{lemma}\label{Lh0_description}
Let $\mathcal{X}\in \category{QSm}_S^\mathrm{prop}$ and $Y\in \category{Sm}_S$.
The canonical epimorphism $\mathbb{Z}_\tr(X^\circ)\twoheadrightarrow h_0(\mathcal{X})$ induces an isomorphism
$h_0(\mathcal{X})(Y) \simeq \Coker (\mathbb{Z}_\tr(\mathcal{X})(Y\boxtimes \bcube)\xrightarrow{d} \mathbb{Z}_\tr(X^\circ)(Y))$.
\end{lemma}
\begin{proof}
This follows from the definition of $h_0$ and the isomorphism $\mathbb{Z}_\tr(\mathcal{X})(Y)\simeq \mathbb{Z}_\tr(X^\circ)(Y)$.
\end{proof}
The map $d$ in the statement of Lemma \ref{Lh0_description} can be computed easily in the following case:
\begin{lemma}\label{principal_divisor_vs_homotopy}
Let $K$ be a field, $X\in \category{Sm}_K$ and $\alpha\in \mathbb{Z}_\tr(X)(\mathbb{P}^1_K\setminus \{1\})$.
Write $\alpha = \sum_{i=1}^n m_i[V_i]$ in the standard form.
Let $p_i\colon V_i\to \mathbb{P}^1_K\setminus \{1\}$ and $q_i\colon V_i\to X$ be the canonical projections.
Then we have
$$
(i_0 - i_\infty)^*\alpha = {\textstyle \sum_{i=1}^n} m_i q_{i,*}\div(p_i).
$$
\end{lemma}
\begin{proof}
Each $V_i$ is finite and dominant over $\mathbb{A}^1_K$ and hence flat over $\mathbb{A}^1_K$.
Therefore, by Lemma \ref{lem:composition_formula} (2) and Lemma \ref{lem:pullback_lemma} (\ref{lem:flat_cycles}), we have
$i_0^*\alpha= \alpha\otimes_{\mathbb{P}^1_K\setminus\{1\}} \{0\} = {\textstyle \sum_{i=1}^n} m_iq_{i,*}[p_i^{-1}(\{0\})]$
and similarly for $i_\infty^*\alpha$.
Hence we get
\begin{align*}
(i_0 - i_\infty)^*\alpha {}={}& {\textstyle \sum_{i=1}^n} m_iq_{i,*}([p_i^{-1}(\{0\})]-[p_i^{-1}(\{\infty\})])\\
{}={}& {\textstyle \sum_{i=1}^n} m_iq_{i,*}\div(p_i).
\end{align*}
This completes the proof.
\end{proof}
\subsection{Reciprocity sheaves}
We define the notion of reciprocity sheaf following \cite{KSY2}.
\begin{definition}[Modulus]
Let $X\in \category{Sm}_S$, $F\in \PSh(\category{Cor}_S)$ and $a\in F(X)$.
Let $\mathcal{Y}\in \category{QSm}_S^\mathrm{prop}$ and suppose that $Y^\circ \simeq X$ as $S$-schemes.
We say that $\mathcal{Y}$ is a {\it modulus} for $a$ if $a\colon \mathbb{Z}_\tr(X)\to F$ factors through the canonical epimorphism
$\mathbb{Z}_\tr(X)\twoheadrightarrow h_0(\mathcal{Y})$.
This is equivalent to saying that $a\colon \mathbb{Z}_\tr(\mathcal{Y})\to \omega^*F$ factors through the cube-invariant part $h^0F$.
\end{definition}
\begin{lemma}\label{reciprocity_equivalent}
Let $F\in \PSh(\category{Cor}_S)$.
The following conditions are equivalent:
\begin{enumerate}
\item The counit morphism $h_0h^0F\hookrightarrow F$ is an isomorphism.
\item For any $X\in \category{Sm}_S$ and any $a\in F(X)$, there is some $\mathcal{Y}\in \category{QSm}_S^\mathrm{prop}$ with an isomorphism $Y^\circ \simeq X$ of $S$-schemes such that
$a\in h^0F(\mathcal{Y})$.
\item Every section of $F$ admits a modulus.
\end{enumerate}
\end{lemma}
\begin{proof}
The equivalence of (1) and (2) follows from $h_0h^0F\simeq \omega_!h^0F$ and the explicit formula for $\omega_!$.
The equivalence of (2) and (3) is clear from the definition.
\end{proof}
\begin{definition}[Reciprocity]
We say that $F$ has {\it reciprocity} or $F$ is a \emph{reciprocity presheaf} if it satisfies the equivalent conditions in Lemma \ref{reciprocity_equivalent}.
If $F$ is moreover a Nisnevich sheaf, then we say that $F$ is a \emph{reciprocity sheaf}.
We define $\category{RSC}_S$ (resp. $\category{RSC}_{S,\Nis}$) to be the full subcategory of $\PSh(\category{Cor}_S)$ consisting of reciprocity presheaves (resp. reciprocity sheaves).
\end{definition}
\begin{remark}
In \cite{KSY2} the terms ``SC-modulus" and ``SC-reciprocity" are used instead of ``modulus" and ``reciprocity".
\end{remark}
\begin{lemma}
The category $\category{RSC}_S$ is closed under taking subobjects and quotients in $\PSh(\category{Cor}_S)$.
In particular, $\category{RSC}_S$ is an abelian category and the inclusion functor $\category{RSC}_S\to \PSh(\category{Cor}_S)$ is exact.
\end{lemma}
\begin{proof}
This is obvious from the definition.
\end{proof}
\begin{lemma}
Let $k$ be a field.
Then the above definition of $\category{RSC}_k$ coincides with the one given in \cite[Definition 2.2.4]{KSY2}.
\end{lemma}
\begin{proof}
Let $\category{RSC}^{\mathrm{KSY}}_k$ denote the category defined in \cite[Definition 2.2.4]{KSY2}.
An object $F\in \PSh(\category{Cor}_k)$ lies in $\category{RSC}_k^{\mathrm{KSY}}$ if and only if for any $X\in \category{Sm}_k$, all sections of $F(X)$ admit a modulus of the form $(Y,D_Y)$, where $D_Y$ is an ordinary effective Cartier divisor on $Y$.
Therefore the inclusion $\category{RSC}_k^{\mathrm{KSY}}\subset \category{RSC}_k$ is clear.
Conversely, suppose that $F\in \category{RSC}_k$.
For any smooth $k$-scheme $X$ and $a\in F(X)$, there is a modulus $\mathcal{Y}$ for $a$.
Write $D_Y=rE_Y$ where $E_Y$ is an ordinary effective Cartier divisor on $Y$ and $r\in \mathbb{Q}_{>0}$.
Then $(Y,\ceil{r}E_Y)$ is also a modulus for $a$ since the canonical epimorphism $\mathbb{Z}_\tr(X)\twoheadrightarrow h_0(\mathcal{Y})$ factors through $h_0(Y,\ceil{r}E_Y)$.
Therefore we get $F\in \category{RSC}_k^{\mathrm{KSY}}$.
\end{proof}
\begin{lemma}\label{omega_shriek_CI_is_RSC}
If $F\in \PSh(\category{QCor}_S^\mathrm{prop})$ is cube-invariant, then $\omega_!F$ has reciprocity.
In particular, the category $\category{RSC}_S$ coincides with the essential image of $h_0$.
\end{lemma}
\begin{proof}
The composition $h_0F\xrightarrow{h_0(\text{unit})} h_0h^0h_0F\xrightarrow{\text{counit}} h_0F$ is the identity.
Since the counit morphism $h_0h^0G\to G$ is a monomorphism for any $G\in \PSh(\category{Cor}_S)$, it follows that the counit morphism $h_0h^0h_0F\to h_0F$ is an isomorphism.
Therefore $\omega_!F\simeq h_0F$ has reciprocity.
\end{proof}
Reciprocity is a generalization of $\mathbb{A}^1$-invariance:
\begin{lemma}\label{a1_invariance_implies_reciprocity}
If $F\in \PSh(\category{Cor}_S)$ is $\mathbb{A}^1$-invariant, then $F$ has reciprocity.
\end{lemma}
\begin{proof}
If $F$ is $\mathbb{A}^1$-invariant, then $\omega^*F$ is cube-invariant since $\omega^*F(\mathcal{X}\boxtimes\bcube)=F(X^\circ \times_S \mathbb{A}^1_S)\simeq F(X^\circ)=\omega^*F(\mathcal{X})$.
Therefore we have $h_0h^0F\simeq \omega_!\omega^*F\simeq F$.
\end{proof}
\begin{remark}
Let $F\in \category{RSC}_S$ and $\mathcal{X}\in \category{QSm}_S^\mathrm{prop}$.
The group $h^0 F(\mathcal{X})$ can be thought of as the subgroup of $\omega^*F(\mathcal{X})=F(X^\circ)$ consisting of elements whose ``ramification'' is bounded by $D_X$.
In the case where $S$ is the spectrum of a perfect field, this recovers several classical notions in ramification theory such as the Artin conductor or the irregularity \cite{RS21}.
For further developments of the ramification theory of reciprocity sheaves over a perfect field, see \cite{RS21}, \cite{RSram1}, \cite{RSram2}, and \cite{RSram3}.
\end{remark}
\subsection{Change of base}
\begin{lemma}\label{base_change_and_h0}
The following assertions hold.
\begin{enumerate}
\item Let $f\colon T\to S$ be a morphism of noetherian schemes.
Then we have canonical isomorphisms
\begin{align*}
&f_*h_0^{\smallcube}\simeqh_0^{\smallcube} f_*\colon \PSh(\category{QCor}_T^\mathrm{prop})\to \PSh(\category{QCor}_S^\mathrm{prop}),\\
&f_*h_0\simeq h_0 f_*\colon \PSh(\category{QCor}_T^\mathrm{prop})\to \PSh(\category{Cor}_S).
\end{align*}
\item Let $p\colon T\to S$ be a pro-smooth morphism of noetherian schemes.
Then we have canonical isomorphisms
\begin{align*}
&p^*h_0^{\smallcube} \simeq h_0^{\smallcube} p^*\colon \PSh(\category{QCor}_S^\mathrm{prop})\to \PSh(\category{QCor}_T^\mathrm{prop}),\\
&p^*h_0 \simeq h_0p^*\colon \PSh(\category{QCor}_S^\mathrm{prop})\to \PSh(\category{Cor}_T).
\end{align*}
\item Let $p\colon T\to S$ be a pro-smooth morphism of noetherian schemes.
Then we have canonical isomorphisms
\begin{align*}
&p_*h^{0,\smallcube} \simeq h^{0,\smallcube} p_*\colon \PSh(\category{QCor}_T^\mathrm{prop})\to \PSh(\category{QCor}_S^\mathrm{prop}),\\
&p_*h^0 \simeq h^0p_*\colon \PSh(\category{Cor}_T)\to \PSh(\category{QCor}_S^\mathrm{prop}).
\end{align*}
\end{enumerate}
\end{lemma}
\begin{proof}
(1) Since $f_*$ commutes with $\omega_!$ by Lemma \ref{presheaf_restriction_compatibility}, we have only to prove the first isomorphism.
Let $F\in \PSh(\category{QCor}_T^\mathrm{prop})$ and $\mathcal{X}\in \category{QSm}_S^\mathrm{prop}$.
We have
\begin{align*}
f_*h_0^{\smallcube} F(\mathcal{X})
={}& h_0^{\smallcube} F(\mathcal{X}_T)\\
\simeq{}& \Coker (F(\mathcal{X}_T\boxtimes\bcube)\xrightarrow{d} F(\mathcal{X}_T))\\
\simeq{}& \Coker (f_*F(\mathcal{X}\boxtimes\bcube)\xrightarrow{d} f_*F(\mathcal{X}))
\simeq{} h_0^{\smallcube} f_*F (\mathcal{X}).
\end{align*}
Here, we used the fact that the scheme-theoretic closure of $X^\circ_T\times\mathbb{A}^1$ in $X_T\times\mathbb{P}^1$ is $X_T^*\times\mathbb{P}^1$.
(2) Since $p^*$ commutes with $\omega_!$ by Lemma \ref{presheaf_restriction_compatibility}, we have only to prove the first isomorphism.
Let $F\in \PSh(\category{QCor}_S^\mathrm{prop})$ and $\mathcal{X}\in \category{QSm}_T^\mathrm{prop}$.
By Lemma \ref{pro_smooth_composition}, we can take a projective system $(\mathcal{X}_i)_{i\in I}$ in $\category{QSm}_S$ as in Definition \ref{def:pro_modulus_pair}.
Since $p^*$ is given by precomposition with $p_\star$, we get
\begin{align*}
p^*h_0^{\smallcube} F (\mathcal{X}) \simeq {} & \textstyle\varinjlim_{i\in I} h_0^{\smallcube} F(\mathcal{X}_i)\\
\simeq {} & \textstyle\varinjlim_{i\in I} \Coker(F(\mathcal{X}_i\boxtimes\bcube)\xrightarrow{d}F(\mathcal{X}_i))\\
\simeq {} & \textstyle\Coker(\varinjlim_{i\in I}F(\mathcal{X}_i\boxtimes\bcube)\xrightarrow{d}\varinjlim_{i\in I}F(\mathcal{X}_i))\\
\simeq {} & \textstyle\Coker(p^*F(\mathcal{X}\boxtimes\bcube)\xrightarrow{d}p^*F(\mathcal{X}))
\simeq {} h_0^{\smallcube} p^*F(\mathcal{X}).
\end{align*}
(3) This follows from (2) and adjunctions.
\end{proof}
\begin{corollary}\label{base_change_and_CI}
The following assertions hold.
\begin{enumerate}
\item Let $f\colon T\to S$ be a morphism of noetherian schemes.
Then the functor
$f_*\colon \PSh(\category{QCor}_T^\mathrm{prop})\to \PSh(\category{QCor}_S^\mathrm{prop})$
preserves cube-invariance, and the functor
$f_*\colon \PSh(\category{Cor}_T)\to \PSh(\category{Cor}_S)$
preserves reciprocity.
\item Let $p\colon T\to S$ be a pro-smooth morphism of noetherian schemes.
Then the functor
$p^*\colon \PSh(\category{QCor}_S^\mathrm{prop})\to \PSh(\category{QCor}_T^\mathrm{prop})$
preserves cube-invariance, and the functor
$p^*\colon \allowbreak \PSh(\category{Cor}_S)\to \PSh(\category{Cor}_T)$
preserves reciprocity.
\end{enumerate}
\end{corollary}
\begin{proof}
This follows from Lemma \ref{base_change_and_h0} and Lemma \ref{omega_shriek_CI_is_RSC}.
\end{proof}
\begin{remark}\label{rem:reduction_to_the_base}
Let $p\colon T\to S$ be a pro-smooth morphism of noetherian schemes.
Then for any $F\in \PSh(\category{QCor}_S^\mathrm{prop})$ and $X\in \category{Sm}_T$, we have canonical isomorphisms
$h_0F(X)\simeq p^*h_0F(X)\simeq h_0p^*F(X)$
by Lemma \ref{base_change_and_h0}.
This will be used later.
\end{remark}
As an easy application of Corollary \ref{base_change_and_CI}, we generalize the following injectivity theorem for reciprocity presheaves.
\begin{theorem}\label{inj_thm}
Let $k$ be a field and $F\in\category{RSC}_k$.
Then $a_\Nis F$ has global injectivity.
\end{theorem}
\begin{proof}
See \cite[Theorem 6]{KSY1}.
The definition of reciprocity sheaves used in \cite{KSY1} is different from ours, but it is more general than ours, as proved in \cite[Theorem 3.2.1]{KSY2}.
\end{proof}
\begin{corollary}\label{global_inj}
Let $S$ be a noetherian reduced scheme and $F\in \category{RSC}_S$.
Then $a_\mathrm{inj} a_\Nis F$ has global injectivity.
\end{corollary}
\begin{proof}
Let $X\in \category{Sm}_S$ and let $U$ be a dense open subset of $X$.
Set $T=\coprod_{\eta\in S^{(0)}}\Spec k(\eta)$ and write $j\colon T\to S$ for the canonical pro-smooth morphism.
Then by definition of $a_\mathrm{inj}$, we have the following commutative diagram:
$$
\xymatrix{
a_\mathrm{inj} a_\Nis F(X)\ar[r]\ar[d]&a_\mathrm{inj} a_\Nis F(U)\ar[d]\\
j^*a_\Nis F(X_T)\ar[r]&j^*a_\Nis F(U_T).
}
$$
Therefore it suffices to show that $j^*a_\Nis F=a_\Nis j^* F$ has global injectivity.
Since $j^*F$ is a reciprocity presheaf over $T$ by Corollary \ref{base_change_and_CI}, this follows from Theorem \ref{inj_thm}.
\end{proof}
\section{Examples of reciprocity sheaves}\label{sec:ex-rec}
\emph{Throughout this section, we fix a noetherian normal scheme $S$.}
In this section we give some examples of reciprocity sheaves.
First three examples (smooth commutative group schemes, the sheaf of \'etale torsors, and the sheaf of rank one $\ell$-adic local systems) are defined over any noetherian normal scheme $S$.
The other examples are treated only in the case where $S\in \category{Sm}_k$ for a perfect field $k$.
\subsection{Preliminary lemma}
\begin{lemma}\label{lem:ramification_generic}
Let $p\colon T\to S$ be a pro-smooth morphism of noetherian schemes.
Let $F\in \PSh(\category{Cor}_S)$ and suppose that the unit morphism $F\to p_*p^*F$ is a monomorphism.
Then for any $\mathcal{X}\in \category{QSm}_S^\mathrm{prop}$ we have
$$
h^0F(\mathcal{X})=\{a\in F(X^\circ)\mid p^*a\in h^0p^*F(\mathcal{X}_T)\}.
$$
\end{lemma}
\begin{proof}
The claim is equivalent to the following assertion: for any section $a\in F(X^\circ)$, the morphism
$a\colon \mathbb{Z}_\tr(X^\circ)\to F$
factors through $h_0(\mathcal{X})$ if and only if the morphism
$p^*a\colon \mathbb{Z}_\tr((X^\circ)_T)\to p^*F$
factors through $p^*h_0(\mathcal{X})=h_0(\mathcal{X}_T)$.
The `only if' part is obvious.
Suppose that $p^*a$ factors through $p^*h_0(\mathcal{X})$.
Then $p_*p^*a\colon p_*p^*\mathbb{Z}_\tr(X^\circ)\to p_*p^*F$ factors through $p_*p^*h_0(\mathcal{X})$.
Therefore the composition $\mathbb{Z}_\tr(X^\circ)\to F\to p_*p^*F$ factors through $h_0(\mathcal{X})$.
Since $F\to p_*p^*F$ is a monomorphism, we see that $a\colon \mathbb{Z}_\tr(X^\circ)\to F$ factors through $h_0(\mathcal{X})$.
\end{proof}
\begin{lemma}\label{lem:reciprocity_check}
Let $F\in \Sh_\Nis(\category{Cor}_S)$.
Suppose that $F$ has horizontal injectivity and $j_\eta^*F$ has reciprocity for all $\eta\in S^{(0)}$, where $j_\eta\colon \Spec k(\eta)\to S$ is the canonical morphism.
Then $F$ has reciprocity and we have
$$
h^0F(\mathcal{X})=\{a\in F(X^\circ)\mid j_\eta^*a\in h^0j_\eta^*F(\mathcal{X}_{\eta})\;(\forall \eta\in S^{(0)})\}.
$$
\end{lemma}
\begin{proof}
By Lemma \ref{base_change_and_CI} (1), $j_{\eta,*}j_\eta^*F$ has reciprocity for each $\eta\in S^{(0)}$ and hence $\bigoplus_{\eta\in S^{(0)}}j_{\eta,*}j_\eta^*F$ has reciprocity.
Since $\category{RSC}_S$ is closed under taking subobjects, we get $F\in \category{RSC}_S$.
The second assertion follows from Lemma \ref{lem:ramification_generic}.
\end{proof}
\subsection{Smooth commutative group schemes}
Let $G$ be a commutative group scheme over $S$ separated over $S$.
Then $G$ canonically extends to an object of $\PSh(\category{Cor}_S)$ by Theorem \ref{thm:app_summary}.
\begin{theorem}\label{thm:grp_sch_reciprocity}
Let $S$ be a noetherian normal scheme.
Let $G$ be a commutative group scheme over $S$ which is separated, smooth, and locally of finite type over $S$.
Then $G$ is a reciprocity sheaf over $S$ having global injectivity.
\end{theorem}
\begin{proof}
The global injectivity follows from Lemma \ref{lem:dominant_injective}.
By Lemma \ref{lem:reciprocity_check}, it suffices to check the reciprocity in the case $S=\Spec k$ where $k$ is a field.
Let $\overline{k}$ be an algebraic closure of $k$ and $f\colon \Spec \overline{k}\to \Spec k$ be the canonical morphism.
Since the canonical morphism $G\to f_*G_{\overline{k}}$ is a monomorphism, we may assume that $k=\overline{k}$.
In this case the result follows from \cite[Corollary 3.2.5 (1)]{KSY2}.
\end{proof}
\begin{example}
The group schemes $\mathbb{G}_a$, $\mathbb{G}_m$, and $\mathbb{W}_n$ (see \S \ref{sec:Witt}) are reciprocity sheaves.
If $S$ is a $\mathbb{Z}_p$-scheme, then the group scheme $W_n$ (see \S \ref{sec:Witt}) is also a reciprocity sheaf.
\end{example}
\subsection{The sheaf of \'etale torsors}
Let $A$ be a torsion abelian group.
We can regard $A$ as a constant group scheme over $S$ and hence $A\in \PSh(\category{Cor}_S)$ by Theorem \ref{thm:app_summary}.
Consequently, the \'etale cohomology presheaf $H^i_\et({-},A)\in \PSh(\category{Sm}_S)$ canonically extends to an object of $\PSh(\category{Cor}_S)$ (see \cite[Remark 10.4.4]{CD} for the corresponding result for Nisnevich cohomology).
We show that $H^1_\et({-},A)$ is a reciprocity sheaf.
\begin{lemma}\label{lem:constant_Nisnevich}
Let $X$ be an integral scheme and $F$ be a sheaf of abelian groups on $X_\Nis$.
Suppose that for any dense open immersion $\iota\colon U\to V$ between \'etale $X$-schemes, the induced homomorphism $F(\iota)\colon F(V)\to F(U)$ is an isomorphism.
Then we have
$$H^i_\Nis(X,F)=0\quad(i>0).$$
\end{lemma}
\begin{proof}
Let $\eta$ be the generic point of $X$ and $U\to X$ be a Nisnevich covering.
Then there is some generic point $\xi\in U$ such that $\xi$ maps to $\eta$ and $k(\eta)\simeq k(\xi)$.
Therefore we can find an open subscheme $V\subset U$ such that $V\to X$ is a dense open immersion.
Let $\check{C}^\bullet(U,F)$ denote the \v{C}ech complex of $F$ with respect to $U\to X$.
We define $h^i\colon \check{C}^i(U,F)\to \check{C}^{i-1}(U,F)$ to be the composition
\begin{align*}
F(U^i)\to F(U^{i-1}\times_X V)\xleftarrow{\sim}F(U^{i-1}).
\end{align*}
This gives a contraction of the cofiber of $F(X)\to \check{C}^\bullet(U,F)$, so we get
$\check{H}^i(U,F)=0\quad (i>0)$.
The claim now follows from the \v{C}ech-to-derived functor spectral sequence.
\end{proof}
\begin{lemma}\label{lem:h1_nisnevich_sheaf}
For any abelian group $A$, the presheaf $H^1_\et({-},A)$ is a Nisnevich sheaf on $\category{Sm}_S$.
\end{lemma}
\begin{proof}
Let $\varepsilon\colon (\category{Sm}_S)_\et\to (\category{Sm}_S)_\Nis$ be the change-of-topology morphism of sites.
For each $X\in \category{Sm}_S$, we have an exact sequence
$$
H^1_\Nis(X,\varepsilon_*A)\to H^1_\et(X,A)\to H^0_\Nis(X,R^1\varepsilon_*A)\to H^2_\Nis(X,\varepsilon_*A)
$$
arising from the Leray spectral sequence for $\varepsilon$.
Since $\varepsilon_*A|_X$ is constant, we get $H^i_\Nis(X,\varepsilon_*A)=0$ for $i>0$ by Lemma \ref{lem:constant_Nisnevich}.
Therefore $H^1_\et({-},A)\simeq R^1\varepsilon_*A$ is a Nisnevich sheaf.
\end{proof}
\begin{lemma}\label{lem:h1_injectivity}
For any abelian group $A$, the sheaf $H^1_\et({-},A)$ has global injectivity.
\end{lemma}
\begin{proof}
Let $X\in \category{Sm}_S$ and let $j\colon U\to X$ be a dense open immersion.
Then the Leray spectral sequence for $j$ shows that the homomorphism
$H^1_\et(X,j_*A)\to H^1_\et(U,A)$
is injective.
Since $j_*A$ is isomorphic to the constant sheaf $A$, this implies global injectivity.
\end{proof}
\begin{theorem}\label{thm:h1_reciprocity}
Let $S$ be a noetherian normal scheme and $A$ be a torsion abelian group.
Then $H^1_\et({-},A)$ is a reciprocity sheaf over $S$ having global injectivity.
\end{theorem}
\begin{proof}
The global injectivity is proved in Lemma \ref{lem:h1_injectivity}.
By Lemma \ref{lem:reciprocity_check}, it suffices to prove the reciprocity in the case $S=\Spec k$ where $k$ is a field.
Moreover, we may assume that $A$ is $p$-primary for some prime number $p$.
If $\ch (k)\neq p$, then $H^1_\et({-},A)$ is $\mathbb{A}^1$-invariant \cite[Corollary 7.7.4]{LeiFuEC} and hence is a reciprocity sheaf (Lemma \ref{a1_invariance_implies_reciprocity}).
If $\ch (k) = p$, then the Artin-Schreier-Witt exact sequence
$$
0\to \mathbb{Z}/p^n\mathbb{Z}\to W_n\xrightarrow{F-\id} W_n\to 0
$$
induces a long exact sequence
$$
\cdots\to W_n\xrightarrow{F-\id} W_n\to H^1_\et({-},\mathbb{Z}/p^n\mathbb{Z})\to H^1_\et({-},W_n)=0\to \cdots
$$
in $\PSh(\category{Cor}_k)$.
It follows that $H^1_\et({-},\mathbb{Z}/p^n\mathbb{Z})$ is a quotient presheaf of $W_n$ and hence has reciprocity.
Writing $A$ as a filtered colimit of finite abelian $p$-groups, we get the desired result.
\end{proof}
\subsection{The sheaf of rank one $\ell$-adic local systems}
This example is taken from \cite[Section 8.9]{RS21}.
Let $\ell$ be a prime number invertible on $S$.
We fix an algebraic closure $\overline{\mathbb{Q}}_\ell$ of $\mathbb{Q}_\ell$ and a compatible system of primitive roots of unity $(\zeta_n)_{n=1}^\infty$ in $\overline{\mathbb{Q}}_\ell^\times$.
For $X\in \category{Sm}_S$, we define $\Lisse(X)$ to be the group of isomorphism classes of lisse $\overline{\mathbb{Q}}_\ell$-sheaves of rank one, where the addition is given by the tensor product.
Note that we have
$$
\Lisse(X)\simeq \varinjlim_{E/\mathbb{Q}_\ell} H^1_\et(X,\mathcal{O}_E^\times) := \varinjlim_{E/\mathbb{Q}_\ell} \varprojlim_n H^1_\et(X,(\mathcal{O}_E/\mathfrak{m}_E^n)^\times),
$$
where $E$ runs over finite sub-extensions of $\overline{\mathbb{Q}}_\ell/\mathbb{Q}_\ell$ and $\mathcal{O}_E$ (resp. $\mathfrak{m}_E$) denotes the ring of integers (resp. maximal ideal of the ring of integers) of $E$.
As in the previous example, $\Lisse$ can be regarded an object of $\PSh(\category{Cor}_S)$.
\begin{theorem}
Let $S$ be a noetherian normal scheme and $\ell$ be a prime number invertible on $S$.
Then $\Lisse$ is a reciprocity sheaf over $S$ having global injectivity.
\end{theorem}
\begin{proof}
The global injectivity follows from Lemma \ref{lem:h1_injectivity}.
By Lemma \ref{lem:reciprocity_check}, it suffices to prove the reciprocity in the case $S=\Spec k$ where $k$ is a field.
Write $m_E=|\mathcal{O}_E/\mathfrak{m}_E|-1$.
Then $\zeta_{m_E}\in \mathcal{O}_E^\times$ defines an isomorphism
$$
(\mathcal{O}_E/\mathfrak{m}^n_E)^\times \simeq \mathbb{Z}/m_E\mathbb{Z}\times (U_E^{(1)}/U_E^{(n)}),
$$
where $U^{(i)}=1+\mathfrak{m}_E^i$.
By Theorem \ref{thm:h1_reciprocity}, the sheaf $H^1_\et({-},\mathbb{Z}/m_E\mathbb{Z})$ has reciprocity.
On the other hand, the sheaf
$\varprojlim_n H^1_\et({-},U_E^{(1)}/U_E^{(n)})$
is $\mathbb{A}^1$-invariant \cite[Corollary 7.7.4]{LeiFuEC} and hence has reciprocity (Lemma \ref{a1_invariance_implies_reciprocity}).
Therefore $\Lisse$ has reciprocity.
\end{proof}
\subsection{Other examples}
We recall some examples of reciprocity sheaves over a perfect field:
\begin{theorem}
Let $k$ be a perfect field.
\begin{enumerate}
\item $\Omega^i$ is a reciprocity sheaf over $k$.
\item Suppose that $\ch(k)=p>0$.
Then $W_n\Omega^i$ is a reciprocity sheaf over $k$.
\item Suppose that $\ch(k)=0$.
For $X\in \category{Sm}_k$, let $\Conn^1(X)$ (resp. $\Conn^1_\mathrm{int}$) denote the group of isomorphism classes of line bundles with connections (resp. integrable connections) on $X$.
Then $\Conn^1$ and $\Conn^1_\mathrm{int}$ are reciprocity sheaves over $k$.
\item Let $\varepsilon$ denote the change-of-topology morphism of sites $(\category{Sm}_k)_\et\to (\category{Sm}_k)_\Nis$.
Let $\mathbb{Q}/\mathbb{Z}(n)$ denote the \'etale motivic complex of weight $n$.
Then $R^i\varepsilon_*\mathbb{Q}/\mathbb{Z}(n)$ is a reciprocity sheaf over $k$.
In particular, $\Br=H^2_\et({-},\mathbb{Q}/\mathbb{Z}(1))$ is a reciprocity sheaf over $k$.
\end{enumerate}
\end{theorem}
\begin{proof}
See \cite[Corollary 3.2.5]{KSY2} for (1) and (2), \cite[Section 6.2]{RS21} for (3) and \cite[Section 11.1 (6)]{BRS} for (4).
\end{proof}
Suppose that $S\in \category{Sm}_k$ for a perfect field $k$ and let $p\colon S\to \Spec k$ be the structure morphism.
Let $F$ be a reciprocity sheaf over $k$.
Then by Lemma \ref{base_change_and_CI} (2), $p^*F$ also has reciprocity.
Moreover, Theorem \ref{inj_thm} shows that $p^*F$ has global injectivity.
In particular, we have the following
\begin{theorem}
Suppose that $S\in \category{Sm}_k$ for a perfect field $k$.
\begin{enumerate}
\item $\Omega^i$ is a reciprocity sheaf over $S$ having global injectivity.
\item Suppose that $\ch(k)=p>0$.
Then $W_n\Omega^i$ is a reciprocity sheaf over $S$ having global injectivity.
\item Suppose that $\ch(k)=0$.
Then $\Conn^1$ and $\Conn^1_\mathrm{int}$ are reciprocity sheaves over $S$ having global injectivity.
\item Let $\varepsilon$ denote the change-of-topology morphism of sites $(\category{Sm}_S)_\et\to (\category{Sm}_S)_\Nis$.
Let $\mathbb{Q}/\mathbb{Z}(n)$ denote the \'etale motivic complex of weight $n$.
Then $R^i\varepsilon_*\mathbb{Q}/\mathbb{Z}(n)$ is a reciprocity sheaf over $S$ having global injectivity.
In particular, $\Br=H^2_\et({-},\mathbb{Q}/\mathbb{Z}(1))$ is a reciprocity sheaf over $S$ having global injectivity.
\end{enumerate}
\end{theorem}
\section{Modulus curves and the motivic Hasse-Arf theorem}\label{sec:HA}
\emph{Throughout this section, we fix a noetherian regular scheme $S$.}
In this section, we prove that the $0$-th Suslin homology of a modulus curve (Definition \ref{def:modulus_curve}) over $S$ is isomorphic to the Chow group of relative $0$-cycles (Definition \ref{def:CH0}).
Using this result, we prove a motivic analogue of the Hasse-Arf theorem.
\subsection{Admissible rational functions}
\begin{definition}
Let $\mathcal{X}$ be a $\mathbb{Q}$-modulus pair.
We say that $\mathcal{X}$ is \emph{integral} if $X$ is integral as a scheme.
\end{definition}
\begin{lemma}
Let $\mathcal{X}$ be a $\mathbb{Q}$-modulus pair.
Then $\mathcal{X}$ is integral if and only if $X^\circ$ is integral.
\end{lemma}
\begin{proof}
The ``only if'' part is clear.
Suppose that $X^\circ$ is integral.
Since $X^\circ$ is a dense open subset of $X$, it follows that $X$ is irreducible.
Let us prove that $X$ is reduced.
We may assume that $X =\Spec A$ and $D_X$ is an ordinary effective Cartier divisor defined by a non-zero-divisor $d\in A$.
Then we have $X^\circ=\Spec A[d^{-1}]$.
Since $A\subset A[d^{-1}]$ and $A[d^{-1}]$ is reduced, $A$ is also reduced.
\end{proof}
\begin{definition}\label{def:admissible_functions}
Let $\mathcal{X}$ be an integral $\mathbb{Q}$-modulus pair.
We say that $f\in k(X)^\times$ is \emph{admissible} with respect to $D_X$ if there is some open neighborhood $U$ of $|D_X|$ such that
\begin{enumerate}
\item $f$ is regular and invertible on $U^N$, and
\item $\cdiv(f-1)\succ D_X$ holds on $U$ (including the case $f=1$).
\end{enumerate}
We define $\Adm(\mathcal{X})\subset k(X)^\times$ to be the subgroup consisting of rational functions which are admissible with respect to $D_X$.
\end{definition}
\begin{lemma}\label{lem:G_valuative}
Let $\mathcal{X}$ be an integral $\mathbb{Q}$-modulus pair.
Then $f\in k(X)^\times$ is admissible with respect to $D_X$ if and only if for any valuation ring $R$ with fraction field $k(X)$ and any morphism $\rho\colon \Spec R\to X$ extending $\Spec k(X)\to X$ such that $\rho^*D_X$ is non-trivial, the inequality $v(f-1)\geq v(\rho^*D_X)$ holds in $\Gamma\otimes_\mathbb{Z} \mathbb{Q}$.
Here, $v\colon k(X)^\times\to \Gamma$ denotes the valuation of $R$.
\end{lemma}
\begin{proof}
The ``only if'' part is easy.
Let us prove the ``if'' part.
Let $\pi\colon X^N\to X$ be the normalization.
For $y\in \pi^{-1}(|D_X|)$, our assumption shows that for any valuation ring $R$ with fraction field $k(X)$ dominating $\mathcal{O}_{X^N,y}$, we have $f\in R^\times$.
Therefore $f$ is invertible on some open neighborhood $V$ of $\pi^{-1}(|D_X|)$ in $X^N$.
Since $\pi$ is universally closed, there is an open neighborhood $U$ of $|D_X|$ such that $\pi^{-1}(U)\subset V$.
The condition (1) in Definition \ref{def:admissible_functions} is satisfied for this $U$.
Let us verify the condition (2).
Let $R$ be a valuation ring with fraction field $k(X)$ and $\rho\colon \Spec R\to U$ be a morphism extending $\Spec k(X)\to U$.
Let $v\colon k(X)^\times\to \Gamma$ be the valuation associated to $R$.
If $\rho^*D_X$ is trivial, then we have $\rho^*\cdiv(f-1)\geq 0=\rho^*D_X$, since $f-1$ is regular on $V$.
Otherwise, our assumption shows that $\rho^*\cdiv(f-1)\geq \rho^*D_X$.
By Lemma \ref{lem:valuative}, we get $\cdiv(f-1)\succ D_X$ on $U$.
\end{proof}
\begin{lemma}\label{higher_unit_norm}
Let $(K,v)$ be a valuation field with value group $\Gamma$ and $L/K$ be a finite extension.
Let $m>0$, $\gamma\in (\Gamma\otimes_\mathbb{Z} \mathbb{Q})_{> 0}$, $f\in L^\times$ and suppose that for every valuation $w$ on $L$ extending $v$, we have
$w(f-1)\geq \gamma$.
Then we have
$v(\Nm_{L/K}(f) - 1) \geq \gamma$.
\end{lemma}
\begin{proof}
Let $i$ denote the inseparable degree of $L/K$.
Fix an algebraic closure $\overline{K}$ of $K$.
Then $v$ can be extended to a valuation on $\overline{K}$.
Let us fix such an extension.
We have
$$
\Nm_{L/K}(f)=\textstyle \prod_{\sigma} \sigma(f)^i =\prod_{\sigma} (1+(\sigma(f)-1))^i
$$
where $\sigma$ runs over the set of distinct $K$-embeddings of $L$ into $\overline{K}$.
Now $v\circ \sigma$ gives a valuation on $L$ extending $v$ for each $\sigma$, so we have
$v(\sigma(f)-1)\geq \gamma$ by our assumption.
By the above expression for $\Nm_{L/K}(f)$, we get the desired inequality.
\end{proof}
\begin{lemma}\label{principal_divisor_push}
Let $\mathcal{X}, \mathcal{C}$ be integral $\mathbb{Q}$-modulus pairs and $q\colon \mathcal{X}\to \mathcal{C}$ be a proper dominant ambient morphism with $[k(X):k(C)]<\infty$.
Then for any $f\in \Adm(\mathcal{X})$, we have $\Nm_{k(X)/k(C)}(f)\in \Adm(\mathcal{C})$.
\end{lemma}
\begin{proof}
By Lemma \ref{lem:G_valuative}, it suffices to show that for any valuation ring $R$ with fraction field $k(C)$ and any morphism $\rho\colon \Spec R\to C$ extending $\Spec k(C)\to C$ such that $\rho^*D_C$ is non-trivial, we have $v(\Nm_{k(X)/k(C)}(f)-1)\geq v(\rho^*D_C)$.
Let $v\colon k(C)^\times\to \Gamma$ be the valuation associated to $R$.
Let $w$ be an arbitrary extension of $v$ to $k(X)$ and $R'$ be its valuation ring.
Since $q\colon X\to C$ is proper, we can find a unique dashed arrow $\varphi$ in the following diagram which makes it commute:
$$
\xymatrix{
\Spec k(X)\ar[r]\ar[d] &\Spec R'\ar@{-->}[r]^-{\varphi}\ar[d] &X\ar[d]^-q\\
\Spec k(C)\ar[r] &\Spec R\ar[r]^-{\rho} &C.
}
$$
By Lemma \ref{lem:G_valuative}, we have $w(f-1)\geq w(\varphi^*D_X)\geq w(\varphi^*q^*D_C)=v(\rho^*D_C)$.
By Lemma \ref{higher_unit_norm}, we get $v(\Nm_{k(X)/k(C)}(f)-1)\geq v(\rho^*D_C)$, as was to be shown.
\end{proof}
\subsection{Modulus curves}
Recall that $S$ is a fixed noetherian regular scheme.
\begin{definition}[Modulus curves]\label{def:modulus_curve}
An object $\mathcal{C}\in \category{QSm}_S$ is called a {\it modulus curve} if $C^\circ \to S$ is quasi-affine and has purely $1$-dimensional fibers.
\end{definition}
\begin{lemma}\label{modulus_curve_every_fiber}
Suppose that $\mathcal{C}\in \category{QSm}_S^\mathrm{prop}$ is a modulus curve.
If $V$ is a closed subscheme of $C$ contained in $C^\circ$, then $V$ is finite over $S$.
\end{lemma}
\begin{proof}
Since $C\to S$ is proper, $V\to S$ is also proper.
For any $x\in S$, the fiber $V_x$ of $V$ is a closed subscheme of $C^\circ_x$ which is proper over $\Spec k(x)$.
Since $C^\circ_x$ is quasi-affine over $\Spec k(x)$, it follows that $V_x$ is a finite $k(x)$-scheme.
Therefore $V\to S$ is quasi-finite and hence finite.
\end{proof}
\begin{lemma}\label{lem:modulus_curve_Weil}
Suppose that $\mathcal{C}\in \category{QSm}_S^\mathrm{prop}$ is a modulus curve.
Then $c_0(C^\circ/S)$ coincides with the group of Weil divisors on $C^\circ$ whose components are closed in $C$.
\end{lemma}
\begin{proof}
Since $S$ is regular, $c_0(C^\circ/S)$ consists of all cycles on $C^\circ$ which are finite and pseudo-dominant over $S$ (Lemma \ref{lem:regular_z_universal}).
Therefore it suffices to prove that an integral closed subscheme $V\subset C^\circ$ is finite and pseudo-dominant over $S$ if and only if $V$ has codimension $1$ and closed in $C$.
Suppose that $V$ is finite and pseudo-dominant over $S$.
Then $V$ is finite over $C$ and hence is closed in $C$.
Let $\xi$ be the generic point of $V$ and $\eta$ its image in $S$.
Then the fiber dimension theorem for flat morphisms \cite[Corollary 14.95]{Goertz-Wedhorn} implies
$\codim_{C^\circ} (\xi) = \codim_S (\eta) + \codim_{C^\circ_\eta} (\xi)$.
Since $V$ is finite and pseudo-dominant over $S$, we have $\codim_S (\eta) = 0$ and $\codim_{C^\circ_\eta} (\xi)=1$.
This implies $\codim_{C^\circ} (\xi)=1$.
Conversely, suppose that $V$ has codimension $1$ and is closed in $C$.
Then $V$ is finite over $S$ by Lemma \ref{modulus_curve_every_fiber}.
We have $\codim_{C^\circ} (\xi) = 1$ and $\codim_{C^\circ_\eta} (\xi)=1$, so the fiber dimension theorem implies $\codim_S(\eta) = 0$.
Therefore $V$ is pseudo-dominant over $S$.
\end{proof}
\begin{lemma}\label{pseudo-normalization}
Let $\mathcal{C}\in \category{QSm}_S^\mathrm{prop}$ be a modulus curve.
Then there are integral modulus curves $\mathcal{C}_1,\dots,\mathcal{C}_n\in \category{QSm}_S^\mathrm{prop}$ and a finite minimal ambient morphism $\pi\colon \coprod_{i=1}^n \mathcal{C}_i\to \mathcal{C}$ which induces an isomorphism $\coprod_{i=1}^n C_i^\circ\to C^\circ$.
In particular, $\pi$ is an isomorphism in $\category{QCor}_S^\mathrm{prop}$.
\end{lemma}
\begin{proof}
Let $C_1^\circ,\dots,C_n^\circ$ be the connected components of $C^\circ$, and let $C_i$ be the scheme-theoretic closure of $C_i^\circ$ in $C$.
Then we have a closed immersion $\pi_i\colon C_i\to C$.
Setting $D_{C_i}=\pi_i^*D_C$, we get the desired integral modulus curves $\mathcal{C}_i=(C_i,D_{C_i})$.
\end{proof}
\begin{definition}\label{def:CH0}
Let $\mathcal{C}\in \category{QSm}_S^\mathrm{prop}$ be an integral modulus curve.
Then for any $f\in \Adm(\mathcal{C})$, the components of the Weil divisor $\div(f)$ on $C^\circ$ is closed in $C$ and hence $\div(f)\in c_0(C^\circ/S)$ by Lemma \ref{lem:modulus_curve_Weil}.
We define the \emph{Chow group of relative 0-cycles} of $\mathcal{C}$ by
$$
\CH_0(\mathcal{C}/S):=\Coker(\Adm(\mathcal{C})\xrightarrow{\div}c_0(C^\circ/S)).
$$
If $\mathcal{C}\in \category{QSm}_S^\mathrm{prop}$ is a (not necessarily integral) modulus curve, then we define
$$
\CH_0(\mathcal{C}/S):=\textstyle \bigoplus_{i=1}^n \CH_0(\mathcal{C}_i/S),
$$
where $\mathcal{C}_1,\dots,\mathcal{C}_n$ are integral modulus curves constructed in Lemma \ref{pseudo-normalization}.
\end{definition}
\begin{remark}
When $S$ is the spectrum of a field, the definition of $\CH_0(\mathcal{C}/S)$ coincides with the definition of the \emph{relative Chow group of $0$-cycles} $\mathrm{C}(C,D_C)$ from \cite[Definition 1.6]{Kerz-Saito}.
The latter group is also defined for higher dimensional varieties, and it was used to establish ramified higher dimensional class field theory.
\end{remark}
\begin{theorem}\label{modulus_curve_suslin_homology}
Let $S$ be a noetherian regular scheme and $\mathcal{C}\in \category{QSm}_S^\mathrm{prop}$ be a modulus curve.
Then for $T\in \category{Sm}_S$, the canonical surjection $\rho\colon c_0((C^\circ)_T/T)\twoheadrightarrow \CH_0(\mathcal{C}_T/T)$ induces an isomorphism
$h_0(\mathcal{C})(T)\xrightarrow{\sim} \CH_0(\mathcal{C}_T/T)$.
\end{theorem}
\begin{remark}
When $S\in \category{Sm}_k$ for some field $k$ and $\mathcal{C}$ satisfies the assumptions of Lemma \ref{relative_Picard_comparison}, this result is proved in \cite[Theorem 1.1]{RY16}.
\end{remark}
\begin{proof}
We may assume that $S=T$ by Remark \ref{rem:reduction_to_the_base}, and that $C$ is integral.
Moreover, replacing $S$ by its connected component containing the image of $C$, we may assume that $S$ is integral.
We identify $\bcube$ with $(\mathbb{P}^1_S,\{1\})$ via the automorphism $t\mapsto \frac{t}{t+1}$ of $\mathbb{P}^1_S$.
It suffices to prove that
$$
\Im(\Adm(\mathcal{C})\xrightarrow{\div} c_0(C^\circ/S))=\Im(c_S((\mathbb{P}^1_S,\{1\}),\mathcal{C})\xrightarrow{{-}\circ (i_0-i_\infty)}c_0(C^\circ/S)).
$$
We will prove a stronger statement that there is a surjective homomorphism
$\Phi\colon c_S((\mathbb{P}^1_S,\{1\}),\mathcal{C})\twoheadrightarrow \Adm(\mathcal{C})$
which makes the following diagram commutative:
\begin{align}\label{modulus_curve_diagram}
\xymatrix{
c_S((\mathbb{P}^1_S,\{1\}),\mathcal{C})\ar@{->>}[r]^-{\Phi}\ar[rd] _-{{-}\circ(i_0-i_\infty)} &\Adm(\mathcal{C})\ar[d]^-{\div}\\
&c_0(C^\circ/S).
}
\end{align}
The group $c_S((\mathbb{P}^1_S,\{1\}),\mathcal{C})$ is freely generated by integral closed subschemes of $(\mathbb{P}_S^1\setminus \{1\})\times_SC^\circ$ which are finite surjective over $\mathbb{P}^1_S\setminus\{1\}$ and admissible.
Let $V$ be such a closed subscheme.
Let $\overline{V}$ denote the scheme-theoretic closure of $V$ in $\mathbb{P}^1_S\times_SC$ and $p\colon \overline{V}\to \mathbb{P}^1_S$, $q\colon \overline{V}\to C$ be the canonical projections.
Then the admissibility of $V$ implies $p^*\{1\}\succ q^*D_C$ and hence $p\in \Adm(\overline{V},q^*D_C)$.
We define
$$
\Phi([V])=\begin{cases}
\Nm_{k(\overline{V})/k(C)}(p) &(\overline{V}\to C\text{ is dominant}),\\
1 &(\text{otherwise}).
\end{cases}
$$
Lemma \ref{principal_divisor_push} shows that this gives a homomorphism $\Phi\colon c_S((\mathbb{P}^1_S,\{1\}),\mathcal{C})\to \Adm(\mathcal{C})$.
Let us prove that (\ref{modulus_curve_diagram}) is commutative.
Let $\eta$ be the generic point of $S$ and $K=k(\eta)$.
Since the operations appearing in (\ref{modulus_curve_diagram}) are compatible with base change to $\Spec K$ and the map $c_0(C^\circ/S)\to c_0(C^\circ_K/K)$ is injective, we may assume that $S=\Spec K$ where $K$ is a field.
In this case, the result follows from Lemma \ref{principal_divisor_vs_homotopy}.
It remains to show that $\Phi$ is surjective.
Let $f\in \Adm(\mathcal{C})$.
If $f=1$, then $\Phi(0)=f$.
Otherwise, we take an open neighborhood $U$ of $|D_C|$ satisfying the conditions (1) and (2) in Definition \ref{def:admissible_functions}.
Then $f$ is regular and invertible on the normalization $U^N$.
Define $\overline{V}$ to be the scheme-theoretic image of
$(f,\pi)\colon U^N\to \mathbb{P}^1_S\times_SC$
where $\pi\colon U^N\to C$ is the canonical morphism.
Let $p\colon \overline{V}\to \mathbb{P}^1_S$ and $q\colon \overline{V}\to C$ be the canonical projections.
Let $V\subset \overline{V}$ denote the inverse image of $\mathbb{P}^1_S\setminus\{1\}$ under $p$.
Since $f=1$ holds on $\pi^{-1}(|D_C|)$, $V$ is contained in $(\mathbb{P}^1_S\setminus\{1\})\times_SC^\circ$.
We will prove that $[V]$ gives an element of $\Phi^{-1}(f)$.
First we prove that $V$ is finite surjective over $\mathbb{P}^1_S\setminus\{1\}$.
Since $f=1$ holds on $\pi^{-1}(|D_C|)$ and $f\neq 1$, $V$ is dominant over $\mathbb{P}^1_S\setminus\{1\}$.
Applying Lemma \ref{modulus_curve_every_fiber} to the modulus curve $(\mathbb{P}^1_S\setminus\{1\})\boxtimes C$ over $\mathbb{P}^1_S\setminus\{1\}$, we see that $V$ is finite over $\mathbb{P}^1_S\setminus\{1\}$.
The admissibility of $V$ follows from the assumption that $\cdiv(f-1)\succ \pi^*D_C$ holds on $U$.
Finally the equality $\Phi([V])=f$ holds by construction.
\end{proof}
We also have the following comparison result between the Chow group of relative 0-cycles and the relative Picard group.
\begin{lemma}\label{relative_Picard_comparison}
Let $S$ be a noetherian regular scheme and $\mathcal{C}\in \category{QSm}_S^\mathrm{prop}$ be a modulus curve.
Suppose that
\begin{enumerate}
\item $C$ is normal,
\item $D_C$ is an ordinary Cartier divisor, and
\item $|D_C|$ is contained in some affine open subset of $C$.
\end{enumerate}
Then $\CH_0(\mathcal{C}/S)$ is canonically isomorphic to the relative Picard group $\Pic(C,D_C)$.
\end{lemma}
\begin{proof}
We may assume that $C$ is integral.
First note that by (1) and (2), a rational function $f\in k(X)^\times$ is admissible with respect to $D_C$ if and only if
\begin{itemize}
\item $f$ is regular and invertible on some neightborhood of $|D_C|$, and
\item $f|_{D_C}=1$.
\end{itemize}
On the other hand, $\Pic(C,D_C)$ is by definition the group of isomorphism classes of pairs $(L,\alpha)$ where $L$ is a line bundle on $C$ and $\alpha$ is a nowhere-vanishing section of $L|_{D_C}$.
By (3), any such $\alpha$ can be extended to a rational section of $L$ which is regular and invertible on some neighborhood of $|D_C|$.
For two such extensions $\widetilde{\alpha}_1,\widetilde{\alpha}_2$, the quotient $\widetilde{\alpha}_1/\widetilde{\alpha}_2$ is admissible with respect to $D_C$.
Therefore $\Pic(C,D_C)$ can be identified with the cokernel of $\cdiv\colon \Adm(\mathcal{C})\to Q$ where $Q$ is the group of Cartier divisors on $C$ whose support is disjoint from $|D_C|$.
Finally, we have $Q\simeq c_0(C^\circ/S)$ since $C^\circ$ is locally factorial.
\end{proof}
\subsection{Motivic Hasse-Arf theorem}
Let $X$ be a noetherian normal scheme.
Then the group of Cartier divisors on $X$ can be seen as a subgroup of the group of Weil divisors on $X$.
Therefore a $\mathbb{Q}$-Cartier divisor on $X$ can be seen as a Weil divisor on $X$ with coefficients in $\mathbb{Q}$.
Let $E$ be a $\mathbb{Q}$-effective $\mathbb{Q}$-Cartier divisor on $X$ and write $E=\sum_{i=1}^n r_iE_i\;(r_i\in \mathbb{Q}_{>0})$ as a Weil divisor.
We define the {\it round-up} of $E$ by
$$
\textstyle\lceil E\rceil = \sum_{i=1}^n \lceil r_i\rceil E_i.
$$
\begin{theorem}\label{geometric Hasse-Arf for curves}
Let $S$ be a noetherian regular scheme.
Let $\mathcal{C}\in \category{QSm}_S^\mathrm{prop}$ and $E$ be a $\mathbb{Q}$-effective $\mathbb{Q}$-Cartier divisor on $C^\circ$ whose components are closed in $C$.
Suppose that $(C,D_C+E)$ is a modulus curve over $S$.
Then the canonical epimorphism
$$
h_0(C,D_C+\lceil E\rceil)\twoheadrightarrow h_0(C,D_C+E)
$$
is an isomorphism.
\end{theorem}
\begin{proof}
We want to prove that for any $T\in \category{Sm}_S$, the canonical surjection
$$
h_0(C,D_C+\lceil E\rceil)(T)\twoheadrightarrow h_0(C,D_C+E)(T)
$$
is an isomorphism.
By Theorem \ref{modulus_curve_suslin_homology}, it suffices to show that
$$
\CH_0((C_T,(D_C)_T+\lceil E\rceil_T)/T)\to \CH_0((C_T,(D_C)_T+E_T)/T)
$$
is an isomorphism.
Write $E=\sum_{i=1}^n r_iE_i$ as a Weil divisor.
Then each $(E_i)_T$ is smooth over $E_i$ and hence is reduced, so we have $\lceil E\rceil_T=\lceil E_T\rceil$.
Therefore we may assume that $T=S$.
Now the claim follows from the fact that $\div(f)$ has integral coefficients for any rational function $f$ on $C^\circ$.
\end{proof}
The following corollary can be seen as a motivic analogue of the Hasse-Arf theorem.
\begin{corollary}\label{Motivic Hasse-Arf for curves}
Let $S$ be a noetherian regular scheme and $C\in \category{Sm}_S$.
Let $E$ be a $\mathbb{Q}$-effective $\mathbb{Q}$-Cartier divisor on $C$ whose components are finite over $S$.
Suppose that $(C,E)$ is a modulus curve over $S$.
Then, for any $F\in \category{RSC}_S$, the canonical inclusion
$$
h^0F(C,E)\hookrightarrow h^0F(C,\lceil E\rceil)
$$
is an isomorphism.
\end{corollary}
\begin{proof}
This from Theorem \ref{geometric Hasse-Arf for curves} by taking a colimit over compactifications.
\end{proof}
\section{Motivic construction of the ring of Witt vectors}\label{sec:Witt}
\emph{Throughout this section, we fix a noetherian regular scheme $S$.}
In this section, we present a construction of the ring of Witt vectors using $\mathbb{Q}$-modulus pairs.
\subsection{Usual construction}
First we recall the usual definition of the ring of Witt vectors; see \cite{Hes15} for details.
Let $A$ be a ring and $n\geq 0$.
The \emph{group of big Witt vectors of length $n$} of $A$ is defined by
$$\mathbb{W}_n(A):=1+tA[t]/(t^{n+1}) \subset (A[t]/(t^{n+1}))^\times.$$
For $a\in A$, we write $[a]$ for the element $1-at\in \mathbb{W}_n(A)$.
The presheaf $X\mapsto\mathbb{W}_n(\mathcal{O}(X))$ on $\category{Sm}_S$ is represented by a commutative group scheme $\mathbb{W}_n$ (whose underlying scheme is isomorphic to $\mathbb{A}^n_S$) and hence extends to an object of $\PSh(\category{Cor}_S)$ (Theorem \ref{thm:app_summary}).
There are several important morphisms in $\PSh(\category{Cor}_S)$:
\begin{align*}
U&\colon \mathbb{Z} \to \mathbb{W}_n,& \star&\colon \mathbb{W}_n\otimes\mathbb{W}_n\to \mathbb{W}_n,\\
F_s&\colon \mathbb{W}_{sn}\to \mathbb{W}_n\quad (s\geq 1),& V_s&\colon \mathbb{W}_n\to \mathbb{W}_{sn}\quad (s\geq 1).
\end{align*}
They are called the unit, the multiplication, the Frobenius, and the Verschiebung.
The unit $U$ is given by $1\mapsto [1]$, and the Verschiebung $V_s$ is induced by $t\mapsto t^s$.
The multiplication $\star$ is characterized by $[a]\star[b]=[ab]$, and the Frobenius $F_s$ is characterized by $F_s([a])=[a^s]$.
We define
$$\widehat{\mathbb{W}}_n:=\mathbb{W}_n\oplus \mathbb{G}_m\oplus \mathbb{Z} \in \PSh(\category{Cor}_S).$$
For a ring $A$ and $a\in A$, we write $[a]$ for the element $([a],a,1)\in \widehat{\mathbb{W}}_n(A)$.
We extend the natural transformations $U$, $\star$, $F_s$, and $V_s$ to $\widehat{\mathbb{W}}_n$ by setting
\begin{align*}
&U(1)=[1],\quad (\alpha,a,m)\star(\alpha',a',m')=(\alpha\star\alpha',a^{m'}{a'}^m,mm'),\\
&F_s(\alpha,a,m)=(F_s(\alpha),a^s,m),\quad V_s(\alpha,a,m)=(V_s(\alpha),a,sm).
\end{align*}
\begin{lemma}\label{lem:Witt_properties}
The following assertions hold for both $\mathbb{W}_n$ and $\widehat{\mathbb{W}}_n$:
\begin{enumerate}
\item $\star$ is commutative, associative, and unital with unit $U$.
\item $F_1=V_1=\id$, $F_sF_r=F_{sr}$, $V_sV_r=V_{sr}$.
\item $F_sV_s=s\cdot \id$. If $(s,r)=1$, then $F_sV_r=V_rF_s$.
\item $V_s\circ \star\circ (\id\otimes F_s) = \star\circ (V_s\otimes\id)$.
\item $\star\circ (F_s\otimes F_s)=F_s\circ \star$.
\end{enumerate}
\end{lemma}
\begin{proof}
The statement for $\mathbb{W}_n$ is well-known (see e.g. \cite{Hes15}), and the statement for $\widehat{\mathbb{W}}_n$ follows easily from this.
\end{proof}
\begin{lemma}
For any $n\geq 0$, $s\geq 1$, and $0\leq m\leq sn$, the morphism $V_sF_s\colon \mathbb{W}_{sn}\to \mathbb{W}_{sn}$ descends to $\mathbb{W}_{m}\to \mathbb{W}_{m}$.
The same also holds for $\widehat{\mathbb{W}}_n$.
\end{lemma}
\begin{proof}
Since $s(\lfloor m/s\rfloor+1)\geq m+1$, the morphism
$V_s\colon \mathbb{W}_{\lfloor m/s\rfloor}\to \mathbb{W}_{s\lfloor m/s\rfloor}$
lifts to
$\mathbb{W}_{\lfloor m/s\rfloor}\to \mathbb{W}_m$.
Composing with $\mathbb{W}_m\twoheadrightarrow \mathbb{W}_{s\lfloor m/s\rfloor}\xrightarrow{F_s} \mathbb{W}_{\lfloor m/s\rfloor}$, we get the desired morphism.
\end{proof}
Let $p$ be a prime number and suppose that $S$ is a $\mathbb{Z}_{(p)}$-scheme.
For a prime number $\ell$ different from $p$, we have an endomorphism
$\ell^{-1}V_\ell F_\ell$ of $\widehat{\mathbb{W}}_n\otimes \mathbb{Z}_{(p)}$.
By Lemma \ref{lem:Witt_properties}, $\ell^{-1}V_\ell F_\ell$ is idempotent and hence defines a direct summand $\Im(\ell^{-1}V_\ell F_\ell)$ of $\widehat{\mathbb{W}}_n\otimes \mathbb{Z}_{(p)}$.
For $n\geq 1$, the \emph{presheaf of $p$-typical Witt vectors of length $n$} is defined by
$$
W_n:=(\widehat{\mathbb{W}}_{p^{n-1}}\otimes \mathbb{Z}_{(p)})/\textstyle\sum_{\ell\neq p}\Im (\ell^{-1}V_\ell F_\ell) \in \PSh(\category{Cor}_S)
$$
where the sum is taken over all prine numbers different from $p$.
\begin{lemma}
The above definition of $W_n$ coincides with the usual one (e.g. \cite{Hes15}).
\end{lemma}
\begin{proof}
Since $\ell^{-1} V_\ell F_\ell(0,a,m)=(0,a,m)$, the subpresheaf $(0\oplus \mathbb{G}_m\oplus \mathbb{Z})\otimes\mathbb{Z}_{(p)}$ of $\widehat{\mathbb{W}}_{p^{n-1}}\otimes\mathbb{Z}_{(p)}$ is contained in $\Im (\ell^{-1} V_\ell F_\ell)$.
Moreover, $\mathbb{W}_{p^{n-1}}$ is a presheaf of $\mathbb{Z}_{(p)}$-modules since $S$ is a $\mathbb{Z}_{(p)}$-scheme.
Therefore we have
$$
W_n=\mathbb{W}_{p^{n-1}}/\textstyle\sum_{\ell\neq p}\Im (\ell^{-1} V_\ell F_\ell).
$$
Now the claim follows from the $p$-typical decomposition of $\mathbb{W}_{p^{n-1}}$ (see \cite[Proposition 1.10]{Hes15}).
\end{proof}
By Lemma \ref{lem:Witt_properties}, the homomorphisms $U$, $\star$, $F_p$, and $V_p$ descend to
\begin{align*}
U&\colon \mathbb{Z}_{(p)}\to W_n,&\star&\colon W_n\otimes W_n\to W_n,\\
F&\colon W_{n+1}\to W_n,&V&\colon W_n\to W_{n+1}
\end{align*}
and we have the following
\begin{lemma}\label{lem:p_Witt_properties}
The following assertions hold for $W_n$:
\begin{enumerate}
\item $\star$ is commutative, associative, and unital with unit $U$.
\item $FV=p\cdot \id$.
\item $V\circ\star\circ(\id\otimes F) = \star\circ(V\otimes\id)$.
\item $\star\circ (F\otimes F)=F\circ \star$.
\end{enumerate}
\end{lemma}
\subsection{Motivic construction}
\begin{definition}
For $n\geq 0$, we define $\mathbb{W}^+_n, \widehat{\mathbb{W}}^+_n\in \PSh(\category{QCor}_S^\mathrm{prop})$ by
$$
\mathbb{W}^+_n:= \varinjlim_{\varepsilon>0}\mathbb{Z}_\tr(\mathbb{P}^1_S,(n+\varepsilon)\{\infty\})/\mathbb{Z},\quad
\widehat{\mathbb{W}}^+_n:= \varinjlim_{\varepsilon>0} \mathbb{Z}_\tr(\mathbb{P}^1_S,\varepsilon\{0\}+(n+\varepsilon)\{\infty\})
$$
where $\mathbb{Z}$ is viewed as a direct summand of $\mathbb{Z}_\tr(\mathbb{P}^1_S,(n+\varepsilon)\{\infty\})$ via $i_0\colon S\to \mathbb{A}^1_S$.
\end{definition}
\begin{definition}\label{def:motivic_operations}
Let $n\geq 0$.
We define the unit, the multiplication, the Frobenius, and the Verschiebung on $\mathbb{W}^+_n$ as follows.
\begin{enumerate}
\item We define $U\colon \mathbb{Z}\to \mathbb{W}^+_n$ to be the morphism induced by $i_1\colon S\to \mathbb{A}^1_S$.
\item We define $\star\colon \mathbb{W}^+_n \otimes \mathbb{W}^+_n \to \mathbb{W}^+_n$ to be the morphism induced by the multiplication map $\mu \colon \AA^1_S \times_S \AA^1_S \to \AA^1_S;\; (x,y) \mapsto xy$ (see Lemma \ref{cube_multiplication_admissible}).
\item For $s\geq 1$, we define $F_s\colon \mathbb{W}^+_{sn}\to \mathbb{W}^+_n$ to be the morphism induced by $\rho_s\colon \AA^1_S \to \AA^1_S;\;x \mapsto x^s$.
\item For $s\geq 1$, we define $V_s\colon \mathbb{W}^+_{n}\to \mathbb{W}^+_{sn}$ to be the morphism induced by ${}^t\rho_s$.
\end{enumerate}
Similarly, we define the unit, the multiplication, the Frobenius, and the Verschiebung on $\widehat{\mathbb{W}}^+_n$.
For $X=\Spec A\in \category{Sm}_S^\mathrm{aff}$ and $a\in A$ (resp. $a\in A^\times$), we write $[a]$ for the element of $h_0\mathbb{W}^+_n(X)$ (resp. $h_0\widehat{\mathbb{W}}^+_n(X)$) represented by the morphism $a\colon X\to \mathbb{A}^1_S$ (resp. $a\colon X\to \mathbb{A}^1_S\setminus\{0\}$).
\end{definition}
\begin{lemma}\label{lem:motivic_Witt_properties}
The following assertions hold for both $\mathbb{W}^+_n$ and $\widehat{\mathbb{W}}^+_n$:
\begin{enumerate}
\item $\star$ is commutative, associative, and unital with unit $U$.
\item $F_1=V_1=\id$, $F_sF_r=F_{sr}$, $V_sV_r=V_{sr}$.
\item $F_sV_s=s\cdot \id$. If $(s,r)=1$, then $F_sV_r=V_rF_s$.
\item $V_s\circ\star\circ (\id\otimes F_s) = \star\circ (V_s\otimes \id)$.
\item $\star\circ(F_s\otimes F_s)=F_s\circ \star$.
\end{enumerate}
\end{lemma}
\begin{proof}
(1) follows from the corresponding properties of $\mu\colon \mathbb{A}^1_S\times_S\mathbb{A}^1_S\to \mathbb{A}^1_S$.
(2) follows from $\rho_1=\id$ and $\rho_r\circ\rho_s=\rho_{rs}$.
The first assertion in (3) follows from $\rho_s \circ {}^t\rho_s=s\cdot\id$.
The second assertion follows from the fact that $\rho_s \circ {}^t\rho_r$ and ${}^t\rho_r\circ\rho_s$ are both represented by the cycle on $\mathbb{A}^2_S$ defined by $x^s=y^r$.
To prove (4), it suffices to show that the following diagram in $\category{Cor}_S$ is commutative:
$$
\xymatrix{
\mathbb{A}^1_S\times_S\mathbb{A}^1_S\ar[rr]^-{{}^t\rho_s\times\id}\ar[d]^-{\id\times \rho_s} &&\mathbb{A}^1_S\times_S\mathbb{A}^1_S\ar[d]^-\mu\\
\mathbb{A}^1_S\times_S\mathbb{A}^1_S\ar[r]^-\mu &\mathbb{A}^1_S\ar[r]^-{{}^t\rho_s} &\mathbb{A}^1_S.
}
$$
One can easily check that both compositions are represented by the cycle on $\mathbb{A}^3_S$ defined by $xy^s=z^s$.
Finally, (5) follows from $\rho_s\circ \mu=\mu\circ(\rho_s\times\rho_s)$.
\end{proof}
\begin{lemma}
For any $n\geq 0$, $s\geq 1$ and $0\leq m\leq sn$, the morphism $V_sF_s\colon \mathbb{W}^+_{sn}\to \mathbb{W}^+_{sn}$ descends to $\mathbb{W}^+_m\to \mathbb{W}^+_m$.
The same also holds for $\widehat{\mathbb{W}}_n$.
\end{lemma}
\begin{proof}
To prove the first statement, it suffices to show that the finite correspondence ${}^t\rho_s\circ\rho_s\in c_S(\mathbb{A}^1_S,\mathbb{A}^1_S)$ is contained in $c_S((\mathbb{P}^1_S,r\{\infty\}),(\mathbb{P}^1_S,r\{\infty\}))$ for every $r\in \mathbb{Q}_{>0}$.
This follows from the fact that ${}^t\rho_s\circ\rho_s$ is represented by the cycle on $\mathbb{A}^2_S$ defined by $x^s=y^s$, which is symmetric in $x$ and $y$.
The second statement can be proved similarly.
\end{proof}
Let $p$ be a prime number and suppose that $S$ is a $\mathbb{Z}_{(p)}$-scheme.
For a prime number $\ell$ different from $p$, we have an endomorphism
$\ell^{-1}V_\ell F_\ell$ of $\widehat{\mathbb{W}}^+_{n}\otimes\mathbb{Z}_{(p)}$.
By Lemma \ref{lem:motivic_Witt_properties}, $\ell^{-1}V_\ell F_\ell$ is an idempotent and hence defines a direct summand $\Im(\ell^{-1}V_\ell F_\ell)$ of $\widehat{\mathbb{W}}^+_{n}\otimes\mathbb{Z}_{(p)}$.
For $n\geq 1$, we define
\begin{equation}\label{def:Wn+}
W^+_n:=(\widehat{\mathbb{W}}^+_{p^{n-1}}\otimes\mathbb{Z}_{(p)})/\textstyle\sum_{\ell\neq p}\Im (\ell^{-1} V_\ell F_\ell) \in \PSh(\category{QCor}_S^\mathrm{prop}).
\end{equation}
For $X=\Spec A\in \category{Sm}_S^\mathrm{aff}$ and $a\in A^\times$, we write $[a]$ for the image of $[a]\in h_0\widehat{\mathbb{W}}^+_{p^{n-1}}(X)$ in $h_0W^+_n(X)$.
By Lemma \ref{lem:motivic_Witt_properties}, the morphisms $U$, $\star$, $F_p$, and $V_p$ descend to
\begin{align*}
U&\colon \mathbb{Z}_{(p)}\to W^+_n,&\star&\colon W^+_n\otimes W^+_n\to W^+_n,\\
F&\colon W^+_{n+1}\to W^+_n,&V&\colon W^+_n\to W^+_{n+1}
\end{align*}
and we have the following
\begin{lemma}\label{lem:motivic_p_Witt_properties}
The following assertions hold for $W^+_n$:
\begin{enumerate}
\item $\star$ is commutative, associative, and unital with unit $U$.
\item $FV=p\cdot \id$.
\item $V\circ(\id\otimes F) = \star\circ (V\otimes \id)$.
\item $\star\circ(F\otimes F)=F\circ \star$.
\end{enumerate}
\end{lemma}
\subsection{Comparison}
\begin{theorem}\label{thm:ab-comparison}
Let $S$ be a noetherian regular scheme and $n\geq 0$.
\begin{enumerate}
\item There is an isomorphism
$\varphi\colon h_0\mathbb{W}^+_n\xrightarrow{\sim} \mathbb{W}_n$
in $\PSh(\category{Cor}_S^\mathrm{aff})$.
For $X=\Spec A\in \category{Sm}_S^\mathrm{aff}$, the image of $\div(x^m+a_1x^{m-1}+\dots+a_m) \in h_0\mathbb{W}^+_n(X)$ is $1+a_1t+\dots+a_mt^m \in \bW_{n}(X)$.
In particular, we have $\varphi_X([a])=[a]$ for $a\in A$.
\item There is an isomorphism
$\widehat{\varphi}\colon h_0\widehat{\mathbb{W}}^+_n\xrightarrow{\sim} \widehat{\mathbb{W}}_n$
in $\PSh(\category{Cor}_S^\mathrm{aff})$.
For $X=\Spec A\in \category{Sm}_S^\mathrm{aff}$, the image of $\div(x^m+a_1x^{m-1}+\dots+a_m)\in h_0\widehat{\mathbb{W}}^+_n(X)$ is $(1+a_1t+\dots+a_mt^m, (-1)^ma_m, m) \in \widehat{\bW}_{n} (X)$.
In particular, we have $\widehat{\varphi}_X([a])=[a]$ for $a\in A^\times$.
\end{enumerate}
\end{theorem}
\begin{proof}
We prove only (1); the proof for (2) is similar.
First we construct an isomorphism
$$
\varphi_X\colon h_0\mathbb{W}^+_n(X)\xrightarrow{\sim} \mathbb{W}_n(X)
$$
for $X=\Spec A\in \category{Sm}_S^\mathrm{aff}$.
We may assume that $X$ is connected.
For $\varepsilon \in (0,1)\cap \mathbb{Q}$ we have
\begin{align*}h_0(\mathbb{P}^1_S,(n+\varepsilon)\{\infty\})(X)&\simeq h_0(\mathbb{P}^1_S,(n+1)\{\infty\})(X) &\text{(Theorem \ref{geometric Hasse-Arf for curves})}\\
&\simeq \CH_0((\mathbb{P}^1_X,(n+1)\{\infty\})/X)&\text{(Theorem \ref{modulus_curve_suslin_homology})}\\
&\simeq \Pic(\mathbb{P}^1_X,(n+1)\{\infty\}).&\text{(Lemma \ref{relative_Picard_comparison})}
\end{align*}
The last term can be easily identified with $\mathbb{W}_n(X)\oplus \mathbb{Z}$ (see e.g. \cite[Proposition 1.1]{Koi2}).
Therefore we have $h_0\mathbb{W}^+_n(X) \simeq (\mathbb{W}_n(X)\oplus \mathbb{Z})/\mathbb{Z}\simeq \mathbb{W}_n(X)$.
It remains to show that this isomorphism is compatible with transfers.
Let $X,Y\in \category{Sm}_S^\mathrm{aff}$ and $\alpha\in c_S(X,Y)$.
It suffices to prove that the following diagram is commutative:
$$
\xymatrix{
h_0\mathbb{W}^+_n(Y)\ar[r]^-{\varphi_{Y}} \ar[d]^-{\alpha^*} &\mathbb{W}_n(Y)\ar[d]^-{\alpha^*}\\
h_0\mathbb{W}^+_n(X)\ar[r]^-{\varphi_{X}} &\mathbb{W}_n(X).
}
$$
We may assume that $X,Y$ are connected.
Let $\eta\in X$ be the generic point and $K$ be an algebraic closure of $k(\eta)$.
Let $p\colon\Spec K\to S$ be the canonical morphism.
Since the composition
$$
h_0\mathbb{W}^+_n(X) \xrightarrow{p^*}h_0\mathbb{W}^+_n(X_K)\to h_0\mathbb{W}^+_n(K)
$$
is compatible with $\varphi$ and the map $\mathbb{W}_n(X)\to \mathbb{W}_n(K)$ is injective, we may assume that $X=\Spec K$ where $K$ is an algebraically closed field.
In this case, $\alpha$ can be written as a $\mathbb{Z}$-linear combination of morphisms $\Spec K\to Y$, so the assertion is obvious.
\end{proof}
\begin{proposition}\label{prop:ring-comparison}
The isomorphisms $\varphi\colon h_0\mathbb{W}^+_n\xrightarrow{\sim} \mathbb{W}_n$ and $\widehat{\varphi} \colon h_0\widehat{\mathbb{W}}^+_n\xrightarrow{\sim} \widehat{\mathbb{W}}_n$ in $\PSh(\category{Cor}_S^\mathrm{aff})$ from Theorem \ref{thm:ab-comparison} are compatible with the unit, the multiplication, the Frobenius, and the Verschiebung.
\end{proposition}
\begin{proof}
We prove only the statement for $\varphi$; the proof for $\widehat{\varphi}$ is similar.
It suffices to prove that for each $X\in \category{Sm}_S^\mathrm{aff}$, the isomorphism $\varphi_X\colon h_0\mathbb{W}^+_n(X)\xrightarrow{\sim} \mathbb{W}_n(X)$ is compatible with these operations.
As in the proof of Theorem \ref{thm:ab-comparison}, we are reduced to the case $X=\Spec K$ where $K$ is an algebraically closed field.
In this case, $\mathbb{W}_n(K)$ is generated by elements of the form $[a]$ with $a\in K^\times$.
For $a,b\in K^\times$, we have
$$
U(1)=[1],\quad [a]\star[b]=[ab],\quad F_s([a])=[a^s],\quad V_s([a])=1-at^s
$$ in $\mathbb{W}_n(K)$ and
$$
U(1)=[1],\quad [a]\star[b]=[ab],\quad F_s([a])=[a^s],\quad V_s ([a]) = \div(x^s-a)
$$
in $h_0\mathbb{W}^+_n(K)$.
This proves the claim.
\end{proof}
\begin{corollary}\label{cor:p_typical_comparison}
Let $p$ be a prime number and $S$ be a noetherian regular $\mathbb{Z}_{(p)}$-scheme.
For $n\geq 1$, there is an isomorphism
$$
\varphi^{(p)}\colon h_0 W^+_n\xrightarrow{\sim} W_n
$$
in $\PSh(\category{Cor}_S^\mathrm{aff})$ which is compatible with the unit, the multiplication, the Frobenius, and the Verschiebung.
If $X=\Spec A\in \category{Sm}_S^\mathrm{aff}$, then we have $\varphi^{(p)}_X([a])=[a]$ for $a\in A^\times$.
\end{corollary}
\subsection{Application to torsion and divisibility of reciprocity sheaves}
In this subsection, we give an application of the motivic presentation of the ring of big Witt vectors to reciprocity sheaves.
Let $F\in \category{RSC}_S$.
Imitating the construction in \cite{Miyazaki-19},
We define $N_nF,NF\in \PSh(\category{Cor}_S)$ by
\begin{align*}
N_nF(X)&:=\varprojlim_{\varepsilon>0}\Ker(h^0F(X\boxtimes(\mathbb{P}^1_S,(n+\varepsilon)\{\infty\}))\xrightarrow{i_0^*}F(X)),\\
NF(X)&:=\Ker(F(X\times_S\mathbb{A}^1_S)\xrightarrow{i_0^*}F(X)).
\end{align*}
In other words, we define $N_nF:=h_0\mathcal{H}\mathrm{om}(\mathbb{W}^+_n,h^0F)$ and $NF:=\varinjlim_{n\geq 0}N_nF$.
The presheaf $NF$ measures how far $F$ is from being $\mathbb{A}^1$-invariant.
The composition
$$
\mathbb{W}^+_n\otimes \mathbb{W}^+_n\otimes \mathcal{H}\mathrm{om}(\mathbb{W}^+_n,h^0F)\xrightarrow{\star\otimes \id}
\mathbb{W}^+_n\otimes \mathcal{H}\mathrm{om}(\mathbb{W}^+_n,h^0F)\xrightarrow{\mathrm{ev}}
h^0F
$$
induces an action
$$
\mathbb{W}^+_n\otimes \mathcal{H}\mathrm{om}(\mathbb{W}^+_n,h^0F) \to \mathcal{H}\mathrm{om}(\mathbb{W}^+_n,h^0F)
$$
by adjunction.
Taking $h_0$ we obtain an action
$$
h_0\mathbb{W}^+_n\otimes N_nF\to N_nF.
$$
Since we have an isomorphism of rings $h_0\mathbb{W}^+_n(X)\simeq W_n(X)$ for $X\in \category{Sm}_S^\mathrm{aff}$ by Theorem \ref{thm:ab-comparison} and Proposition \ref{prop:ring-comparison}, we obtain the following
\begin{theorem}
Let $S$ be a noetherian regular scheme.
Let $F\in \category{RSC}_S$ and $X\in \category{Sm}_S^\mathrm{aff}$.
Then there is a canonical action of $\mathbb{W}_n(X)$ on $N_nF(X)$ which is natural in $X$ and $n$.
\end{theorem}
\begin{corollary}
Let $S$ be a noetherian regular scheme, $F\in \category{RSC}_S$ and $X\in \category{Sm}_S^\mathrm{aff}$.
\begin{enumerate}
\item If $p$ is a prime number invertible on $X$, then $N_nF(X)$ and $NF(X)$ are uniquely $p$-divisible.
\item If $p$ is a prime number nilpotent on $X$, then $N_nF(X)$ and $NF(X)$ are $p$-groups.
\end{enumerate}
\end{corollary}
\begin{proof}
This follows from the fact that if $p$ is invertible (resp. nilpotent) in a ring $A$, then $p$ is also invertible (resp. nilpotent) in $\mathbb{W}_n(A)$; see \cite[Lemma 1.9 and Proposition 1.10]{Hes15}.
\end{proof}
As a consequence, we obtain the following generalization of the result of Binda-Cao-Kai-Sugiyama \cite[Theorem 1.3]{BCKS}.
\begin{corollary}\label{BCKSgen}
Let $S$ be a noetherian regular scheme, $F\in \category{RSC}_S$ and assume that $F$ is separated for the Zariski topology.
\begin{enumerate}
\item If $S$ is a $\mathbb{Q}$-scheme and $F\otimes\mathbb{Q}=0$, then $F$ is $\mathbb{A}^1$-invariant.
\item If $S$ is a $\mathbb{Z}_{(p)}$-scheme and $F\otimes\mathbb{Z}_{(p)}=0$, then $F$ is $\mathbb{A}^1$-invariant.
\item If $S$ is an $\mathbb{F}_p$-scheme and $F$ is $p$-torsion-free, then $F$ is $\mathbb{A}^1$-invariant.
\end{enumerate}
\end{corollary}
\section{Motivic construction of the de Rham-Witt complex}
\label{sec:dRWitt}
\emph{Throughout this section, we fix a prime number $p\neq 2$ and a noetherian regular $\mathbb{Z}_{(p)}$-scheme $S$.}
In this section, we present a construction of the de Rham-Witt complex of smooth $S$-schemes using $\mathbb{Q}$-modulus pairs.
\subsection{De Rham-Witt complex}
First we recall the definition of the de Rham-Witt complex of a $\mathbb{Z}_{(p)}$-algebra given in \cite{Hesselholt-Madsen}.
\begin{definition}
Let $A$ be a $\mathbb{Z}_{(p)}$-algebra.
A \emph{Witt complex} over $A$ is a tuple $(E_\bullet^*,F,V, \lambda)$ where
\begin{enumerate}
\item $\cdots \to E_n^*\xrightarrow{R} E_{n-1}^*\to \cdots\to E_1^*$ is a sequence of CDGAs,
\item $F\colon E_{n+1}^*\to E_n^*$ is a graded ring homomorphism compatible with $R$,
\item $V\colon E_n^*\to E_{n+1}^*$ is a graded group homomorphism compatible with $R$,
\item $\lambda\colon W_n(A)\to E_n^0$ is a ring homomorphism compatible with $R$, $F$, and $V$,
\end{enumerate}
such that the following relations hold:
$$
V(x\star F(y))=V(x)\star y,\quad FdV=d,\quad FV=p\cdot\id, \quad F(d\lambda[a])=\lambda[a^{p-1}]\star d\lambda[a]\;(a\in A).
$$
\end{definition}
The category of Witt complexes over $A$ is known to have an initial object $W_n\Omega^*_A$ and it is called the \emph{de Rham-Witt complex} of $A$.
We have $W_n\Omega^0_A\simeq W_n(A)$.
By construction, $W_n\Omega^q_A$ is a quotient of $\Omega^q_{W_n(A)}$.
The presheaf $\Spec A\mapsto W_n\Omega^q_A$ on $\category{Sm}_S^\mathrm{aff}$ extends to an \'etale sheaf $W_n\Omega^q$ on $\category{Sm}_S$ having global injectivity.
\subsection{Motivic construction}
\begin{definition}\label{def:Gm+}
We define $\mathbb{G}_m^+\in \PSh(\category{QCor}_S^\mathrm{prop})$ by
$$
\mathbb{G}_m^+:=\varinjlim_{\varepsilon>0}\mathbb{Z}_\tr(\mathbb{P}^1_S, \varepsilon\{0\}+\varepsilon\{\infty\})/\mathbb{Z}
$$
where $\mathbb{Z}$ is viewed as a direct summand of $\mathbb{Z}_\tr(\mathbb{P}^1_S,\varepsilon\{0\}+\varepsilon\{\infty\})$ via $i_1\colon S\to \mathbb{A}^1_S\setminus\{0\}$.
\end{definition}
\begin{lemma}\label{Gm_anti_commutative}
The following assertions hold.
\begin{enumerate}
\item Let $F_s\colon \mathbb{G}_m^+\to \mathbb{G}_m^+$ be the morphism induced by $\rho_s$.
Then we have $h_0(F_s)=s\cdot \id$.
\item Let $V_s\colon \mathbb{G}_m^+\to \mathbb{G}_m^+$ be the morphism induced by ${}^t\rho_s$.
If $s$ is odd, then $h_0(V_s)=\id$.
\item For $X=\Spec A \in \category{Sm}_S^\mathrm{aff}$ and $a\in A^\times$, we write $[a]$ for the element of $h_0(\mathbb{G}_m^+)(X)$ represented by the morphism $a\colon X\to \mathbb{A}^1_S\setminus\{0\}$.
Then we have $[a]+[b]=[ab]$.
\item Let $\tau\colon \mathbb{G}_m^{+\otimes 2}\to \mathbb{G}_m^{+\otimes 2}$ be the morphism induced by
$
(\mathbb{A}^1_S\setminus\{0\})^2\to (\mathbb{A}^1_S\setminus\{0\})^2;
\; (x,y)\mapsto (y,x).
$
Then we have $h_0(\tau)=-\id$.
\item Let $\delta\colon \mathbb{G}_m^+\to \mathbb{G}_m^{+\otimes 2}$ be the morphism induced by
$
\mathbb{A}^1_S\setminus\{0\}\to (\mathbb{A}^1_S\setminus\{0\})^2;
\; x\mapsto (x,x).
$
Then we have $2\cdot h_0(\delta)=0$.
\end{enumerate}
\end{lemma}
\begin{proof}
\noindent
(1) Define $\Gamma\subset (\mathbb{A}^1_S\setminus\{0\})\times_S\mathbb{A}^1_S\times_S(\mathbb{A}^1_S\setminus\{0\})=\Spec_S\mathcal{O}_S[x^\pm,t,y^\pm]$ by the equation
$$
(1-t)(y-x^s)(y-1)^{s-1}+t(y-x)^s=0.
$$
Regarding the left hand side as a polynomial in $y$, the leading term and the constant term have invertible coefficients.
Therefore $\Gamma$ is finite locally free over $(\mathbb{A}^1_S\setminus\{0\})\times_S\mathbb{A}^1_S$ and hence defines a finite correspondence $[\Gamma]\in c_S((\mathbb{A}^1_S\setminus\{0\})\times_S\mathbb{A}^1_S, \mathbb{A}^1_S\setminus\{0\})$.
This induces a morphism $\mathbb{G}_m^+\otimes\mathbb{Z}_\tr(\bcube)\to \mathbb{G}_m^+$ whose restriction to $t=0$ (resp. $t=1$) is $F_s$ (resp. $s\cdot \id$).
This shows that $h_0(F_s)=s\cdot \id$.
\noindent
(2) Define $\Gamma\subset (\mathbb{A}^1_S\setminus\{0\})\times_S\mathbb{A}^1_S\times_S(\mathbb{A}^1_S\setminus\{0\})=\Spec_S\mathcal{O}_S[x^\pm,t,y^\pm]$ by the equation
$$
(1-t)(y^s-x)+t(y-x)(y-1)^{s-1}=0.
$$
Regarding the left hand side as a polynomial in $y$, the leading term and the constant term have invertible coefficients \emph{if $s$ is odd}.
Therefore $\Gamma$ is finite locally free over $(\mathbb{A}^1_S\setminus\{0\})\times_S\mathbb{A}^1_S$ and hence defines a finite correspondence $[\Gamma]\in c_S((\mathbb{A}^1_S\setminus\{0\})\times_S\mathbb{A}^1_S, \mathbb{A}^1_S\setminus\{0\})$.
This induces a morphism $\mathbb{G}_m^+\otimes\mathbb{Z}_\tr(\bcube)\to \mathbb{G}_m^+$ whose restriction to $t=0$ (resp. $t=1$) is $V_s$ (resp. $\id$).
This shows that $h_0(V_s)=\id$.
\noindent
(3) Define $\Gamma\subset X\times_S\mathbb{A}^1_S\times_S(\mathbb{A}^1_S\setminus\{0\})=\Spec A[t,x^\pm]$ by the equation
$$
(1-t)(x-a)(x-b)+t(x-ab)(x-1)=0.
$$
Regarding the left hand side as a polynomial in $x$, the leading term and the constant term have invertible coefficients.
Therefore $\Gamma$ is finite locally free over $\Spec A[t]$ and hence defines a finite correspondence $[\Gamma]\in c_S(X\times_S\mathbb{A}^1_S, \mathbb{A}^1_S\setminus\{0\})$.
This induces an element of $\mathbb{G}_m^+ (X \boxtimes \bcube)$ whose restriction to $t=0$ (resp. $t=1$) is $[a]+[b]$ (resp. $[ab]$).
This shows that $[a]+[b]=[ab]$.
\noindent
(4) Define $\Gamma\subset (\mathbb{A}^1_S\setminus\{0\})^2\times_S\mathbb{A}^1_S\times_S(\mathbb{A}^1_S\setminus\{0\})^2=\Spec_S\mathcal{O}_S[x^\pm,y^\pm,t,z^\pm,w^\pm]$ by the equations
$$
z+w=(1-t)(x+y)+t(xy+1),\quad zw=xy.
$$
One can check that $\Gamma$ is finite locally free over $(\mathbb{A}^1_S\setminus\{0\})^2\times_S\mathbb{A}^1_S$ and hence defines a finite correspondence $[\Gamma]\in c_S((\mathbb{A}^1_S\setminus\{0\})^2\times_S\mathbb{A}^1_S, \mathbb{A}^1_S\setminus\{0\})$.
This induces a morphism $\mathbb{G}_m^{+\otimes 2}\otimes \mathbb{Z}_\tr(\bcube)\to \mathbb{G}_m^{+\otimes 2}$ whose restriction to $t=0$ (resp. $t=1$) is $\id+\tau$ (resp. $0$).
This shows that $h_0(\id+\tau)=0$ and hence $h_0(\tau)=-\id$.
\noindent
(5) We have $\tau\circ \delta=\delta$.
Since $h_0(\tau)=-\id$ by (4), we get $-h_0(\delta)=h_0(\delta)$ and hence $2\cdot h_0(\delta)=0$.
\end{proof}
\begin{remark}
If $S=\Spec k$ for some field $k$, then we can prove that $h_0(\mathbb{G}_m^{+\otimes q})\cong K_q^M$ holds in $\PSh(\category{Cor}_k)$, where $K_q^M$ is the unramified sheaf of Milnor $K$-groups; it is essentially a corollary of \cite[Theorem 5.1]{MVW}.
We expect that this holds for an arbitrary $S$.
\end{remark}
\begin{definition}
Let $n\geq 1$.
\begin{enumerate}
\item
For $q,r\geq 0$, we define the \emph{multiplication} on $W_n^+\otimes \mathbb{G}_m^{+\otimes *}$ by
\begin{align*}
\star\colon (W^+_n\otimes\mathbb{G}_m^{+\otimes q})
\otimes(W^+_n\otimes\mathbb{G}_m^{+\otimes r})
&\xrightarrow{\sim}
W^+_n\otimes W^+_n
\otimes\mathbb{G}_m^{+\otimes q}\otimes\mathbb{G}_m^{+\otimes r}\\
&\xrightarrow{\star\otimes \id}
W^+_n\otimes\mathbb{G}_m^{+\otimes (q+r)}.
\end{align*}
For each $X\in \category{Sm}_S^\mathrm{aff}$, this makes $h_0 (W_n^+\otimes \mathbb{G}_m^{+\otimes *})(X)$ into a graded ring.
By Lemma \ref{Gm_anti_commutative} (4), this multiplication is graded commutative.
Moreover, it is a $\mathbb{Z}_{(p)}$-algebra since $h_0 W_n^+(X)\simeq W_n(X)$ is so.
\item
For $q\geq 0$, we define the \emph{Frobenius} by
$F:=F\otimes \id\colon W^+_{n+1}\otimes\mathbb{G}_m^{+\otimes *}
\to W^+_n\otimes\mathbb{G}_m^{+\otimes *}$
and the \emph{Verschiebung} by
$V:=V\otimes \id\colon W^+_n\otimes\mathbb{G}_m^{+\otimes *}
\to W^+_{n+1}\otimes\mathbb{G}_m^{+\otimes *}$.
By Lemma \ref{lem:motivic_p_Witt_properties}, $F$ is a graded ring homomorphism, and we have $V(x\star F(y))=V(x)\star y$, $FV=p$.
\item
We define $\lambda\colon W_n\xrightarrow{\sim} h_0 W^+_n$ in $\PSh(\category{Cor}_S^\mathrm{aff})$ to be the inverse of the isomorphism $\varphi^{(p)}$ given in Corollary \ref{cor:p_typical_comparison}.
This is a ring homomorphism compatible with $F$ and $V$.
\end{enumerate}
\end{definition}
\begin{definition}
We define $d\colon \widehat{\mathbb{W}}^+_n\to \widehat{\mathbb{W}}^+_n\otimes \mathbb{G}_m^+$ to be the morphism induced by
$
\Delta\colon \mathbb{A}^1_S\setminus\{0\}\to (\mathbb{A}^1_S\setminus\{0\})^2;\; x\mapsto (x,x).
$
\end{definition}
\begin{lemma}\label{delta_and_Verschiebung}
For any prime number $\ell$ different from $p$, the following diagram becomes commutative after applying $h_0$:
$$
\xymatrix{
\widehat{\mathbb{W}}^+_{p^{n-1}}\ar[r]^-{d}\ar[d]^-{\ell V_\ell F_\ell} &
\widehat{\mathbb{W}}^+_{p^{n-1}}\otimes\mathbb{G}_m^+\ar[d]^-{\ell V_\ell F_\ell\otimes \id}\\
\widehat{\mathbb{W}}^+_{p^{n-1}}\ar[r]^-{d}&
\widehat{\mathbb{W}}^+_{p^{n-1}}\otimes\mathbb{G}_m^+.
}
$$
\end{lemma}
\begin{proof}
By Lemma \ref{Gm_anti_commutative} (1), we have $h_0(\id\otimes F_\ell)=\ell\cdot \id$ on $h_0(\widehat{\mathbb{W}}^+_{p^{n-1}}\otimes\mathbb{G}_m^+)$.
Therefore it suffices to show that the following diagram is commutative:
$$
\xymatrix{
\widehat{\mathbb{W}}^+_{p^{n-1}}\ar[r]^-d\ar[d]^-{V_\ell F_\ell}&
\widehat{\mathbb{W}}^+_{p^{n-1}}\otimes\mathbb{G}_m^+\ar[r]^-{\id\otimes F_\ell}&
\widehat{\mathbb{W}}^+_{p^{n-1}}\otimes\mathbb{G}_m^+\ar[d]^-{V_\ell F_\ell\otimes \id}\\
\widehat{\mathbb{W}}^+_{p^{n-1}}\ar[r]^-d&
\widehat{\mathbb{W}}^+_{p^{n-1}}\otimes\mathbb{G}_m^+\ar[r]^-{\id\otimes F_\ell}&
\widehat{\mathbb{W}}^+_{p^{n-1}}\otimes\mathbb{G}_m^+.
}
$$
This is further reduced to the commutativity of the following diagram in $\category{Cor}_S$:
$$
\xymatrix{
\mathbb{A}^1_S\setminus\{0\}\ar[r]^-\Delta\ar[d]^-{{}^t\rho_\ell\circ\rho_\ell}&
(\mathbb{A}^1_S\setminus\{0\})^2\ar[r]^-{\id\times \rho_\ell}&
(\mathbb{A}^1_S\setminus\{0\})^2\ar[d]^-{({}^t\rho_\ell\circ\rho_\ell)\times\id}\\
\mathbb{A}^1_S\setminus\{0\}\ar[r]^-\Delta&
(\mathbb{A}^1_S\setminus\{0\})^2\ar[r]^-{\id\times\rho_\ell}&
(\mathbb{A}^1_S\setminus\{0\})^2.
}
$$
Both compositions are given by the cycle on $(\mathbb{A}^1_S\setminus\{0\})^3$ defined by $x^\ell=y^\ell=z$.
\end{proof}
\begin{definition}\label{def:motivic_operations}
Let $n\geq 1$.
We define the \emph{differential}
$$
d\colon h_0(W^+_n\otimes\mathbb{G}_m^{+\otimes q})
\to h_0(W^+_n\otimes\mathbb{G}_m^{+\otimes (q+1)})
$$
to be the morphism induced by $d\otimes \id\colon\widehat{\mathbb{W}}^+_{p^{n-1}}\otimes\mathbb{G}_m^{+\otimes q}\to \widehat{\mathbb{W}}^+_{p^{n-1}}\otimes\mathbb{G}_m^{+\otimes (q+1)}$ via Lemma \ref{delta_and_Verschiebung}.
\end{definition}
\begin{lemma}\label{delta_Leibniz}
The following diagram becomes commutative after applying $h_0$:
$$
\xymatrix{
\widehat{\mathbb{W}}_{p^{n-1}}^{+\otimes 2}\ar[r]^-{\star}\ar[d]^-{d\otimes\id+\id\otimes d} &
\widehat{\mathbb{W}}^+_{p^{n-1}}\ar[d]^-{d}\\
\widehat{\mathbb{W}}_{p^{n-1}}^{+\otimes 2}\otimes\mathbb{G}_m^+\ar[r]^-{\star\otimes\id}&
\widehat{\mathbb{W}}^+_{p^{n-1}}\otimes\mathbb{G}_m^+.
}
$$
\end{lemma}
\begin{proof}
It suffices to show that the following diagram becomes commutative after composing with the canonical epimorphism $\mathbb{Z}_\tr(\mathbb{A}^1_S\setminus\{0\})^{\otimes 2}\to h_0 (\widehat{\mathbb{W}}^+_{p^{n-1}}\otimes\mathbb{G}_m^+)$:
$$
\xymatrix{
\mathbb{Z}_\tr(\mathbb{A}^1_S\setminus\{0\})^{\otimes 2}\ar[r]^- \mu\ar[d]^-{q_1+q_2}
&\mathbb{Z}_\tr(\mathbb{A}^1_S\setminus\{0\})\ar[d]^-\Delta\\
\mathbb{Z}_\tr(\mathbb{A}^1_S\setminus\{0\})^{\otimes 3}\ar[r]^-{\mu\otimes \id}
&\mathbb{Z}_\tr(\mathbb{A}^1_S\setminus\{0\})^{\otimes 2}.
}
$$
Here, $q_1$ (resp. $q_2$) denotes the morphism $(x,y)\mapsto (x,y,x)$ (resp. $(x,y)\mapsto (x,y,y)$).
The two compositions are identified with
\begin{align*}
&\mathbb{Z}_\tr(\mathbb{A}^1_S\setminus\{0\})^{\otimes 2}\xrightarrow{\beta}
\mathbb{Z}_\tr(\mathbb{A}^1_S\setminus\{0\})^{\otimes 4}\xrightarrow{\mu\otimes\mu}
\mathbb{Z}_\tr(\mathbb{A}^1_S\setminus\{0\})^{\otimes 2},
\\
&\mathbb{Z}_\tr(\mathbb{A}^1_S\setminus\{0\})^{\otimes 2}\xrightarrow{\beta}
\mathbb{Z}_\tr(\mathbb{A}^1_S\setminus\{0\})^{\otimes 4}\xrightarrow{\mu\otimes (\pr_1+\pr_2)}
\mathbb{Z}_\tr(\mathbb{A}^1_S\setminus\{0\})^{\otimes 2}
\end{align*}
where $\beta$ is the morphism $(x,y)\mapsto (x,y,x,y)$.
Therefore it suffices to show that the two morphisms
$$
\mu,\pr_1+\pr_2\colon \mathbb{Z}_\tr(\mathbb{A}^1_S\setminus\{0\})^{\otimes 2}\to \mathbb{Z}_\tr(\mathbb{A}^1_S\setminus\{0\})
$$
coincide after composing with the canonical epimorphism $\mathbb{Z}_\tr(\mathbb{A}^1_S\setminus\{0\})\twoheadrightarrow h_0\mathbb{G}_m^+$.
This follows from Lemma \ref{Gm_anti_commutative} (3).
\end{proof}
\begin{lemma}\label{FdV}
The following diagram becomes commutative after applying $h_0$:
$$
\xymatrix{
\widehat{\mathbb{W}}^+_{p^{n-1}}\ar[r]^-{V_p}\ar[d]^-{d} &
\widehat{\mathbb{W}}^+_{p^n}\ar[d]^-{d}\\
\widehat{\mathbb{W}}^+_{p^{n-1}}\otimes\mathbb{G}_m^+&
\widehat{\mathbb{W}}^+_{p^n}\otimes\mathbb{G}_m^+\ar[l]_-{F_p\otimes \id}.
}
$$
\end{lemma}
\begin{proof}
By Lemma \ref{Gm_anti_commutative} (2) and our assumption that $p$ is odd, we have $h_0(\id\otimes V_p)=\id$ on $h_0(\widehat{\mathbb{W}}^+_{p^{n-1}}\otimes\mathbb{G}_m^+)$.
Therefore it suffices to show that the following diagram is commutative:
$$
\xymatrix{
\widehat{\mathbb{W}}^+_{p^{n-1}}\ar[rr]^-{V_p}\ar[d]^-{d} &&
\widehat{\mathbb{W}}^+_{p^n}\ar[d]^-{d}\\
\widehat{\mathbb{W}}^+_{p^{n-1}}\otimes\mathbb{G}_m^+\ar[r]^-{\id\otimes V_p}&
\widehat{\mathbb{W}}^+_{p^{n-1}}\otimes\mathbb{G}_m^+&
\widehat{\mathbb{W}}^+_{p^n}\otimes\mathbb{G}_m^+\ar[l]_-{F_p\otimes \id}.
}
$$
This is further reduced to the commutativity of the following diagram in $\category{Cor}_S$:
$$
\xymatrix{
\mathbb{A}^1_S\setminus\{0\}\ar[rr]^-{{}^t\rho_p}\ar[d]^-{\Delta} &&
\mathbb{A}^1_S\setminus\{0\}\ar[d]^-{\Delta}\\
(\mathbb{A}^1_S\setminus\{0\})^2\ar[r]^-{\id\otimes {}^t\rho_p}&
(\mathbb{A}^1_S\setminus\{0\})^2&
(\mathbb{A}^1_S\setminus\{0\})^2\ar[l]_-{F_p\otimes \id}.
}
$$
Both compositions are given by the cycle on $(\mathbb{A}^1_S\setminus\{0\})^3$ defined by $x=y=z^p$.
\end{proof}
\begin{theorem}\label{thm:theta}
Let $X=\Spec A\in \category{Sm}_S^\mathrm{aff}$.
Then the tuple $(a_\mathrm{inj} a_\Nis h_0(W^+_\bullet\otimes \mathbb{G}_m^{+\otimes *})(X), F, V, \lambda)$ is a Witt complex over $A$.
In particular, we have a unique homomorphism of Witt complexes
$$
\theta\colon W_\bullet \Omega^*_A\to a_\mathrm{inj} a_\Nis h_0(W^+_\bullet\otimes \mathbb{G}_m^{+\otimes *})(X)
$$
which extends to a morphism $\theta\colon W_n \Omega^q\to a_\mathrm{inj} a_\Nis h_0(W^+_n\otimes \mathbb{G}_m^{+\otimes q})$ in $\Sh_\Nis(\category{Sm}_S)$.
\end{theorem}
\begin{proof}
By Lemma \ref{Gm_anti_commutative} (5), we have $2d^2=0$ and hence $d^2=0$.
By Lemma \ref{delta_Leibniz} and Lemma \ref{Gm_anti_commutative} (4), we see that $d$ satisfies the Leibniz rule.
The relation $FdV=d$ holds by Lemma \ref{FdV}.
Let us show that the relation
$$
F(d\lambda[a])=\lambda[a^{p-1}]\star d\lambda[a]
$$
holds for any $a\in A$.
Since $a_\mathrm{inj} a_\Nis h_0(W^+_n\otimes \mathbb{G}_m^{+\otimes q})$ has global injectivity (Lemma \ref{global_inj}), we may assume that $a$ is invertible.
Then both sides are represented by the class of the morphism $(a^p,a)\colon \Spec A\to (\mathbb{A}^1_S\setminus\{0\})^2$ in $c_S(\Spec A,(\mathbb{A}^1_S\setminus\{0\})^2)$.
\end{proof}
\subsection{Compatibility with transfers}
In the last subsection, we have constructed a morphism $\theta\colon W_n \Omega^q\to a_\mathrm{inj} a_\Nis h_0(W^+_n\otimes \mathbb{G}_m^{+\otimes q})$ in $\Sh_\Nis(\category{Sm}_S)$.
We prove that $\theta$ is compatible with transfers, provided that $W_n\Omega^q$ admits a suitable transfer structure.
\begin{definition}
We say that $S$ admits \emph{de Rham-Witt transfers} if $W_n\Omega^q$ can be extended to an object of $\PSh(\category{Cor}_S)$ so that it is compatible with the trace maps defined in \cite[Theorem 2.6]{Rulling-thesis} for finite field extensions.
Note that this condition uniquely determines the transfer structure on $W_n\Omega^q$, since $W_n\Omega^q$ has global injectivity.
\end{definition}
\begin{example}
If $S=\Spec k$ for some perfect field $k$ of characteristic $p$, then $S$ admits de Rham-Witt transfers by \cite[Theorem B.2.1]{KSY1}; for the comptibility with the trace maps, see \cite[Section 7.9]{RS21}.
More generally, if there is a pro-smooth morphism $\pi\colon S\to \Spec k$ for some perfect field $k$ of characteristic $p$, then $\pi^*W_n\Omega^q$ gives the desired transfer structure on $W_n\Omega^q$ over $S$, so $S$ admits de Rham-Witt transfers.
For example, if $S\in \category{Sm}_k$ for some (possibly imperfect) field $k$ of characteristic $p$, then $S$ admits de Rham-Witt transfers.
\end{example}
\begin{lemma}[Projection formula]
Assume that $S$ admits de Rham-Witt transfers.
Let $X\in \category{Sm}_S$, $Y\in \category{Sm}_X$ and let $\pi\colon Y\to X$ be the structure morphism.
Let $a\in W_n\Omega^q(X)$ and $b\in W_n\Omega^q(Y)$.
For any $\alpha\in c_X(X,Y)$, we have
$$
\alpha^*(\pi^*a\star b)=a \star \alpha^*b.
$$
\end{lemma}
\begin{proof}
Since the transfer structure of $W_n\Omega^q$ is assumed to be compatible with the trace maps defined in \cite[Theorem 2.6]{Rulling-thesis}, the claim follows from the projection formula for the trace maps.
\end{proof}
\begin{lemma}\label{theta_transfer_1}
Assume that $S$ admits de Rham-Witt transfers.
Let $X=\Spec A\in \category{Sm}_S$.
For any $\alpha\in c_A(\Spec A, \Spec A[t])$, the following diagram is commutative:
$$
\xymatrix{
W_n\Omega^q_{A[t]}\ar[r]^-{\theta}\ar[d]^-{\alpha^*} &a_\mathrm{inj} a_\Nis h_0(W_n^+\otimes\mathbb{G}_m^{+\otimes q})(A[t])\ar[d]^-{\alpha^*}\\
W_n\Omega^q_A\ar[r]^-{\theta} &a_\mathrm{inj} a_\Nis h_0(W_n^+\otimes\mathbb{G}_m^{+\otimes q})(A).
}
$$
\end{lemma}
The following proof is inspired by the proof of \cite[Proposition 5.19]{KP11}.
\begin{proof}
We prove by induction on $q$.
For $q=0$, this follows from Corollary \ref{cor:p_typical_comparison}.
Suppose that $q\geq 1$.
We prove by induction on $n$.
We say that an element $\omega\in W_n\Omega^q_{A[t]}$ is \emph{traceable} if $\alpha^*\theta(\omega)=\theta(\alpha^*\omega)$ holds.
Since $F$, $V$, and $d$ are compatible with $\alpha^*$, if $\omega$ is traceable then so is $F(\omega)$, $V(\omega)$, and $d\omega$.
The group $W_n\Omega^q_{A[t]}$ is a quotient of $\Omega^q_{W_n(A[t])}$ and hence generated by elements of the form
$$
\omega = V^{j_0}([a_0])\star dV^{j_1}([a_1])\star \dots\star dV^{j_q}([a_q])
$$
where $a_0,\dots,a_q\in A[t]$.
Let us prove that $\omega$ is traceable.
\noindent
(1) If $j_0=\dots=j_q=0$, then we can write $\omega$ as a $\mathbb{Z}$-linear combination of elements of the form
\begin{align*}
\eta &= [c_0t^{k_0}]\star d[c_1]\star \dots\star d[c_q],\\
\xi &= [c_0t^{k_0}]\star d[t]\star d[c_1]\star \dots\star d[c_{q-1}]
\end{align*}
where $c_0,\dots,c_q\in A$.
First we show that $\eta$ is traceable.
By the projection formula, it suffices to show that $[c_0t^{k_0}]$ is traceable.
This follows from the case $q=0$.
Next we show that $\xi$ is traceable.
We write $k_0+1=mp^e$ with $p\nmid m$.
Then we have
\begin{align*}
F^e(d[t^m])&=F^{e-1}([t^{m(p-1)}]\star d[t^m])\\
&=F^{e-2}([t^{mp(p-1)}]\star [t^{m(p-1)}]\star d[t^m])\\
&=\cdots=[t^{m(p^e-1)}]\star d[t^m]\\
&=m[t^{mp^e-1}]\star d[t] = m[t^{k_0}]\star d[t]
\end{align*}
and hence
$$
\xi = m^{-1}F^e (d[t^m])\star [c_0]\star d[c_1]\star \dots\star d[c_{q-1}].
$$
By the projection formula, it suffices to show that $m^{-1}F^e (d[t^m])$ is traceable.
This follows from the case $q=0$.
\noindent
(2) If $j_0\geq 1$, then we have $\omega = V(\omega')$ for some $\omega'\in W_{n-1}\Omega^q_{A[t]}$, so the induction hypothesis on $n$ shows that $\omega$ is traceable.
\noindent
(3) If $j_0=0$ and $j_1\geq 1$, then we have
$$
[a_0]\star dV^{j_1}([a_1])=d([a_0]\star V^{j_1}([a_1])) - V^{j_1}([a_1]) \star d([a_0]).
$$
We set $\omega_0=[a_0]\star V^{j_1}([a_1])$, $\omega_1=V^{j_1}([a_1]) \star d([a_0])$, and $\omega_2=dV^{j_2}([a_2])\star \dots\star dV^{j_q}([a_q])$, so that $\omega=d(\omega_0\star \omega_2)-\omega_1\star \omega_2$.
Then $d(\omega_0\star \omega_2)$ is traceable by the induction hypothesis on $q$, and $\omega_1\star\omega_2$ is traceable by (2).
\noindent
(4) If $j_0=0$ and $j_i\geq 1$ for some $2\leq i\leq q$, then $\omega$ is traceable by the same argument as in (3).
\end{proof}
\begin{lemma}\label{theta_transfer_2}
Assume that $S$ admits de Rham-Witt transfers.
Let $X=\Spec A\in \category{Sm}_S^\mathrm{aff}$ and set $A_N=A[t_1,\dots,t_N]$.
For any $\alpha\in c_A(\Spec A, \Spec A_N)$, the following diagram is commutative:
$$
\xymatrix{
W_n\Omega^q_{A_N}\ar[r]^-{\theta}\ar[d]^-{\alpha^*} &a_\mathrm{inj} a_\Nis h_0(W_n^+\otimes\mathbb{G}_m^{+\otimes q})(A_N)\ar[d]^-{\alpha^*}\\
W_n\Omega^q_A\ar[r]^-{\theta} &a_\mathrm{inj} a_\Nis h_0(W_n^+\otimes\mathbb{G}_m^{+\otimes q})(A).
}
$$
\end{lemma}
\begin{proof}
Since the target of $\theta$ has global injectivity, we may assume that $A=K$ is a field pro-smooth over $S$.
We proceed by induction on $N$, the case $N=1$ being Lemma \ref{theta_transfer_1}.
We may assume that $\alpha = [x]$ where $x$ is a closed point of $\Spec K_N$.
Let $\pi\colon \Spec K_N\to \Spec K_{N-1}$ be the canonical projection and set $y= \pi(x)$.
Then $[y]$ defines an element of $c_K(\Spec K, \Spec K_{N-1})$.
Since $x$ is a closed point of $\mathbb{A}^1_{k(y)}$, it is defined by some monic equation
$$
t_N^m+a_{m-1}t_N^{m-1}+\dots+a_0=0
$$
where $a_0,\dots,a_{m-1}\in k(y)$.
Let $b_i$ be a lift of $a_i$ to $K_{N-1}$ for $i=0,\dots,m-1$.
Define $\Gamma\subset \Spec K_N$ by the equation
$$
t_N^m+b_{m-1}t_N^{m-1}+\dots+b_0=0.
$$
Then $[\Gamma]$ defines an element of $c_{K_{N-1}}(K_{N-1},K_N)$.
Noting that $\Gamma$ is flat over $\Spec K_{N-1}$, we have $[x]=[\Gamma]\circ[y]$ by Lemma \ref{lem:pullback_lemma} (\ref{lem:flat_cycles}).
Therefore the claim follows from the induction hypothesis and Lemma \ref{theta_transfer_1}.
\end{proof}
\begin{lemma}\label{theta_transfer_3}
Assume that $S$ admits de Rham-Witt transfers.
Let $X,Y\in \category{Sm}_S$.
For any $\alpha\in c_S(X,Y)$, the following diagram is commutative:
$$
\xymatrix{
W_n\Omega^q(Y)\ar[r]^-{\theta}\ar[d]^-{\alpha^*} &a_\mathrm{inj} a_\Nis h_0(W_n^+\otimes\mathbb{G}_m^{+\otimes q})(Y)\ar[d]^-{\alpha^*}\\
W_n\Omega^q(X)\ar[r]^-{\theta} &a_\mathrm{inj} a_\Nis h_0(W_n^+\otimes\mathbb{G}_m^{+\otimes q})(X).
}
$$
In other words, $\theta$ is a morphism in $\Sh_\Nis(\category{Cor}_S)$.
\end{lemma}
\begin{proof}
The finite correspondence $\alpha$ can be factored as
$$
X\xrightarrow{(\id,\alpha)}X\times_S Y\xrightarrow{\pr_2} Y.
$$
Therefore we may assume that $Y\in \category{Sm}_X$ and $\alpha\in c_X(X,Y)$.
Since $\theta$ is a morphism of Nisnevich sheaves, we may assume that $X=\Spec A$, $Y=\Spec B$.
Take a closed immersion $\Spec B\to \Spec A_N$ where $A_N=A[t_1,\dots,t_N]$.
Since $W_n\Omega_{A_N}\to W_n\Omega_B$ is surjective, we may assume that $B=A_N$, so the claim follows from Lemma \ref{theta_transfer_2}.
\end{proof}
\subsection{Comparison}
\begin{definition}
Assume that $S$ admits de Rham-Witt transfers.
Consider the sections
$$
[t]\colon \mathbb{Z}_\tr(\mathbb{A}^1_S\setminus\{0\})\to W_n,\quad t\colon \mathbb{Z}_\tr(\mathbb{A}^1_S\setminus\{0\})\to \mathbb{G}_m
$$
where $t$ is the coordinate function of $\mathbb{A}^1_S\setminus\{0\}$.
One can easily see that these morphisms descend to
$$
[t]\colon \omega_!W^+_n\to W_n, \quad t\colon \omega_!\mathbb{G}^+_m \to \mathbb{G}_m.
$$
Taking a tensor product and composing with
$$
W_n\otimes\mathbb{G}_m^{\otimes q}\to W_n\Omega^q;\quad x\otimes a_1\otimes\dots\otimes a_q\mapsto x\star\dlog[a_1]\star\dots\star\dlog[a_q],
$$
we get a morphism
$$
\eta\colon \omega_!(W_n^+\otimes \mathbb{G}_m^{+\otimes q})\to W_n\Omega^q
$$
in $\PSh(\category{Cor}_S)$.
In other words, $\eta$ is the morphism induced by the section
$[t_0]\star \dlog[t_1]\star \cdots\star \dlog[t_q]$ of $W_n\Omega^q$ on $(\mathbb{A}^1_S\setminus\{0\})^{q+1}$, where $t_0,\dots,t_q$ are the coordinate functions of $(\mathbb{A}^1_S\setminus\{0\})^{q+1}$.
\end{definition}
\begin{lemma}
Assume that $S$ admits de Rham-Witt transfers.
Then the morphism $\eta\colon \omega_!(W_n^+\otimes \mathbb{G}_m^{+\otimes q})\to W_n\Omega^q$ descends to $h_0 (W_n^+\otimes \mathbb{G}_m^{+\otimes q})$.
\end{lemma}
Note that $\eta$ further factors through $a_\mathrm{inj} a_\Nis h_0 (W_n^+\otimes \mathbb{G}_m^{+\otimes q})$ since $W_n\Omega^q$ is a Nisnevich sheaf having horizontal injectivity.
\begin{proof}
We identify $\eta$ with the corresponding section of $W_n\Omega^q$ on $(\mathbb{A}^1_S\setminus\{0\})^{q+1}$.
Fix $\varepsilon\in \mathbb{Q}_{>0}$.
Let $\mathcal{W}=(\mathbb{P}^1_S,\varepsilon\{0\}+(p^{n-1}+\varepsilon)\{\infty\})$ and $\mathcal{G}=(\mathbb{P}^1_S,\varepsilon\{0\}+\varepsilon\{\infty\})$.
It suffices to show that for any connected $X\in \category{Sm}_S$ and
$$
\alpha\in c_S(X\boxtimes (\mathbb{P}^1_S,\{1\}), \mathcal{W}\boxtimes \mathcal{G}^{\boxtimes q}),
$$
we have $\partial(\alpha)^*\eta=0 \in W_n\Omega^q(X)$, where $\partial(\alpha)=\alpha\circ(\id\boxtimes i_0-\id\boxtimes i_\infty)$.
Since the homomorphism $W_n\Omega^q(X)\to W_n\Omega^q(k(X))$ is injective, we may assume that $X=\Spec K$ where $K$ is a field pro-smooth over $S$.
Moreover, we may assume that $\alpha=[\Gamma]$ for some integral closed subscheme $\Gamma\subset (\mathbb{P}^1_K\setminus\{1\})\times_K(\mathbb{A}^1_K\setminus\{0\})^{q+1}$ which is finite surjective over $\mathbb{P}^1_K\setminus\{1\}$.
Let $C$ denote the normalization of $\Gamma$.
Let $g\colon C\to \mathbb{P}^1_K\setminus\{1\}$ and $(f_0,\dots,f_q)\colon C\to (\mathbb{A}^1_K\setminus\{0\})^{q+1}$ denote the canonical projections.
Then by Lemma \ref{principal_divisor_vs_homotopy}, $\partial(\alpha)$ is the composition of $(f_0,\dots,f_q)$ and $\div(g)\in c_K(\Spec K,C)$.
Therefore it suffices to show that
$$
\textstyle\sum_{x\in C^{(1)}} n_x\Tr_{K(x)/K}([f_0]\star\dlog[f_1]\star\dots\star \dlog[f_q])=0
$$
where $\div(g)=\sum_{x\in C^{(1)}} n_x\{x\}$.
Using the residue map (\cite{Rulling-thesis}, \cite{Rulling-erratum}), we can rewrite the left hand side as
$$
\textstyle\sum_{x\in C^{(1)}} \Res_{C/K,x}([f_0]\star\dlog[f_1]\star\dots\star \dlog[f_q]\star \dlog[g]).
$$
By the residue theorem (\cite[Theorem 2.19]{Rulling-thesis}, \cite[Theorem 2]{Rulling-erratum}), the above sum is zero if we replace $C$ with the regular projective model $\overline{C}$ of $C$.
Therefore it suffices to show that $[f_0]\star\dlog[f_1]\star\dots\star \dlog[f_q]\star \dlog[g]$ is regular at every $x\in \overline{C}\setminus C$.
Let $\pi$ be a local parameter at $x$ and write $f_i=u_i\pi^{m_i}$ and $g=1+u\pi^m$ where $u_i,u\in \mathcal{O}_{\overline{C},x}^\times$.
Note that we always have $m>0$.
The admissibility of $\Gamma$ implies
$$
m\geq \varepsilon \max\{m_0,0\}+(p^{n-1}+\varepsilon)\max\{-m_0,0\}+\varepsilon (|m_1|+\cdots+|m_q|)>-p^{n-1}m_0.
$$
Hence the result follows from the next lemma.
\end{proof}
\begin{lemma}
Let $(K,v)$ be a henselian discrete valuation field, $\mathcal{O}_K$ be its valuation ring and $\mathfrak{m}_K$ be the maximal ideal of $\mathcal{O}_K$.
Suppose that $\mathcal{O}_K$ is a $\mathbb{Z}_{(p)}$-algebra.
Let $f_0,\dots,f_q\in K^\times$, $a\in \mathfrak{m}_K$ and assume that $v(a)>-p^{n-1}v(f_0)$.
Then we have $[f_0]\star\dlog[f_1]\star\cdots\star\dlog[f_q]\star\dlog[1+a]\in W_n\Omega^q(\mathcal{O}_K)$.
\end{lemma}
\begin{proof}
Let $\pi$ be a uniformizer of $K$.
Write $f_i=u_i\pi^{m_i}$ and $a=u\pi^m$ where $u_i, u\in \mathcal{O}_K^\times$.
Then we have $m>-p^{n-1}m_0$.
By \cite[Lemma 3.4]{Rulling-thesis}, we can write
\begin{align*}
[1+a]&=[1]+\textstyle\sum_{i=0}^{n-1}V^i([\pi^m]\star c_i)
\end{align*}
where $c_i\in W_n(\mathcal{O}_K)$.
Therefore it suffices to show that
$$
[f_0]\star\dlog[f_1]\star\cdots\star\dlog[f_q]\star dV^i([\pi^m]\star c_i)
$$
is regular for $i=0,1,\dots,n-1$.
By the Leibniz rule, it suffices to show that
\begin{enumerate}
\item $[f_0]\star\dlog[f_1]\star\cdots\star\dlog[f_q]\star V^i([\pi^m]\star c_i)$,
\item $d[f_0]\star\dlog[f_1]\star\cdots\star\dlog[f_q]\star V^i([\pi^m]\star c_i)$
\end{enumerate}
are regular.
For (1), we have
\begin{align*}
&[f_0]\star\dlog[f_1]\star\cdots\star\dlog[f_q]\star V^i([\pi^m]\star c_i)\\
=&V^i(F^i[f_0]\star\dlog[f_1]\star\cdots\star\dlog[f_q]\star [\pi^m]\star c_i)\\
=&V^i([u_0^{p^i}\pi^{m+p^im_0}]\star\dlog[f_1]\star\cdots\star\dlog[f_q]\star c_i)
\end{align*}
and hence the claim follows from $m>-p^im_0$.
For (2), we have
$d[f_0]=[f_0]\star \dlog [f_0]$
and hence the claim follows from the regularity of (1).
\end{proof}
\begin{lemma}\label{eta_compatibility}
Assume that $S$ admits de Rham-Witt transfers.
Then the morphism $\eta\colon h_0 (W^+_n\otimes \mathbb{G}_m^{+\otimes q})\to W_n\Omega^q$ is compatible with $\star$, $d$, $F$, $V$, and $\lambda$.
\end{lemma}
\begin{proof}
The claim for $\star,F,V,\lambda$ follows from Proposition \ref{prop:ring-comparison}.
Let us prove the claim for $d$.
Define $\delta\colon (\mathbb{A}^1_S\setminus\{0\})^{q+1}\to (\mathbb{A}^1_S\setminus\{0\})^{q+2}$ by
$(t_0,\dots,t_q)\mapsto (t_0,t_0,t_1,\dots,t_q)$.
Then we have the following commutative diagram:
$$
\xymatrix{
\mathbb{Z}_\tr(\mathbb{A}^1_S\setminus\{0\})^{\otimes (q+1)}\ar[r]^-\delta\ar@{->>}[d] &\mathbb{Z}_\tr(\mathbb{A}^1_S\setminus\{0\})^{\otimes (q+2)}\ar@{->>}[d]\\
h_0 (W_n^+\otimes \mathbb{G}_m^{+\otimes q})\ar[r] ^-d &h_0 (W_n^+\otimes \mathbb{G}_m^{+\otimes (q+1)}).
}
$$
Therefore it suffices to show that $\delta^*\eta = d\eta$ holds as a section of $W_n\Omega^{q+1}$ on $(\mathbb{A}^1_S\setminus\{0\})^{q+1}$.
This follows from the equality $[t]\star \dlog[t]=d[t]$ in $W_n\Omega^1(\mathbb{A}^1_S\setminus\{0\})$.
\end{proof}
\begin{lemma}\label{lem:theta_of_eta}
Assume that $S$ admits de Rham-Witt transfers.
Then we have
$$
\theta(\eta) = \id
$$
where the right hand side denotes the section of $a_\mathrm{inj} a_\Nis h_0(W_n^+\otimes\mathbb{G}_m^{+\otimes q})$ on $(\mathbb{A}^1_S\setminus\{0\})^{q+1}$ represented by the identity morphism.
\end{lemma}
\begin{proof}
Taking an affine open covering of $S$, we may assume that $S=\Spec A$.
The section $\eta$ can be written as
$$
\eta = [t_0t_1^{-1}\cdots t_q^{-1}]\star d[t_1]\star\dots\star d[t_q]
$$
where $t_0,\dots,t_q$ are the coordinate functions of $(\mathbb{A}^1_A\setminus\{0\})^{q+1}$.
Since $\theta$ is a morphism of Witt complexes, $\theta(\eta)$ is given by the same formula in $h_0(W_n^+\otimes \mathbb{G}_m^{+\otimes q})((\mathbb{A}^1_A\setminus\{0\})^{q+1})$.
The element $[t_0t_1^{-1}\cdots t_q^{-1}]$ is represented by the morphism
$$
t_0t_1^{-1}\cdots t_q^{-1}\colon (\mathbb{A}^1_A\setminus\{0\})^{q+1}\to \mathbb{A}^1_A\setminus\{0\}
$$
and the element $d[t_i]$ is represented by the morphism
$$
(\mathbb{A}^1_A\setminus\{0\})^{q+1}\xrightarrow{t_i}\mathbb{A}^1_A\setminus\{0\}\xrightarrow{\Delta} (\mathbb{A}^1_A\setminus\{0\})^2.
$$
By definition of the multiplication on $h_0(W_n^+\otimes \mathbb{G}_m^{+\otimes *})$, we see that $\theta(\eta)$ is represented by the identity morphism.
\end{proof}
\begin{theorem}\label{thm:rep-WOmega}
Let $S$ be a noetherian regular $\mathbb{Z}_{(p)}$-scheme with $p\neq 2$.
Assume that $S$ admits de Rham-Witt transfers (e.g. $S\in \category{Sm}_k$ for some field $k$ of characteristic $p$).
Then there is an isomorphism $$\eta \colon a_\mathrm{inj} a_\Nis h_0 (W^+_n\otimes \mathbb{G}_m^{+\otimes q}) \xrightarrow{\sim} W_n\Omega^q$$ in $\Sh_\Nis(\category{Cor}_S)$, whose inverse is given by $\theta$.
\end{theorem}
\begin{proof}
Lemma \ref{lem:theta_of_eta} shows that $\theta\circ \eta=\id$.
Let $X=\Spec A\in \category{Sm}_S^\mathrm{aff}$.
Since $W_n\Omega^q_A$ is initial in the category of Witt complexes over $A$, and $\eta\circ \theta$ gives an endomorphism of $W_n\Omega^q_A$ as a Witt complex over $A$ by Lemma \ref{eta_compatibility}, we get $\eta\circ \theta=\id$.
\end{proof}
\begin{corollary}\label{cor:rep-WOmega-field}
Let $k$ be a field of characteristic $p\neq 2$.
Then there is an isomorphism
\[
\eta\colon a_{\Nis} h_0 (W_n^+ \otimes \mathbb{G}_m^{+\otimes q})\xrightarrow{\sim} W_n\Omega^q
\]
in $\Sh_\Nis(\category{Cor}_k)$, whose inverse is given by $\theta$.
\end{corollary}
|
\section{Introduction}
\label{sec:intro}
Granular materials have been paid more and more attention recently, such as sand, grains, snow, et al. Compared to the molecular particles modeled with elastic collision, the granular gases behave quite differently due to the energy dissipation during collisions. Therefore, most theories for the elastically colliding spheres fail to describe the granular gases. The Boltzmann equation, an important model in elastic theory, can also be extended to describe the behavior of granular gases. Recently, the inelastic Boltzmann equation has also been applied to model the social and biological systems \cite{Pareschi2014}.
Owing to the energy loss, the inelastic collision operator is fundamentally different from the elastic operator. Both analytical and numerical theories in this field are still at an early stage, and we refer the readers to a recent review \cite{Villani2006} for some related results and open questions. Numerically, some methods have been proposed to solve the inelastic Boltzmann equation. The Direct Simulation Monte Carlo (DSMC) method \cite{Bird}, which is initially proposed for the elastic Boltzmann equation, has recently been extended to the inelastic case \cite{Gamba2005, Astillero2005}. It can efficiently compute the highly rarefied gas flow but does not work well in low-speed and unsteady flows. In recent years, deterministic methods have made significant progress in kinetics. For example, the Fourier spectral method \cite{Hu2016, Mouhot} has been applied to simulate the Boltzmann equation, and then successfully extended to the inelastic case \cite{Hu2019, Wu2015}. Besides, the Petrov--Galerkin spectral method has been proposed for the inelastic Boltzmann equation \cite{Hu2020}, and a unified gas-kinetic scheme is also adopted to deal with the inelastic collision of granular gases \cite{Liu2019}.
For inelastic gas flows, people are always concerned with the behavior of some macroscopic variables, especially temperature. Therefore, we focus on the Hermite spectral method, which can express the important macroscopic variables with the expansion coefficients of the first few orders. The history of the Hermite spectral method can be traced back to Grad's work \cite{Grad} in 1949, which is known as the moment method. It is based on the idea to take the steady state, Maxwellian, as the weight function. Then, the distribution function is expanded using the orthogonal polynomials with the weight function, and these polynomials are just Hermite polynomials. In the past few years, great progress has been made in the application of the Hermite spectral method. An algorithm to approximate the general quadratic Boltzmann collision operator is first derived in \cite{Approximation2019}. Soon, the method is verified with the success in the simulation of rarefied gas flow \cite{ZhichengHu2019}. Later, it has been modified and extended to the field of collisional plasma \cite{FPL2020} and the multi-species Boltzmann equation \cite{multi2022}.
In this paper, we develop a numerical algorithm based on the Hermite spectral method for the inelastic Boltzmann equation. Even though the Maxwellian is no longer the steady state of inelastic collisions, the Hermite spectral method maintains the advantage of easy expressions for the macroscopic variables. We first derive the algorithm of the inelastic quadratic collision term under the framework of the Hermite spectral method, where the complex calculation of the expansion coefficients for the quadratic collision term can be greatly reduced. For the VHS model, these coefficients can even be derived exactly with several summations. Next, to balance the accuracy and computational cost, a new collision model is built by combining the quadratic collision model and a linearized inelastic collision model modified from the work in \cite{Filbet2013, Astillero2005}. Following the similar lines of \cite{ZhichengHu2019, multi2022}, the Strang splitting method is utilized to separate the convection and collision parts. The finite volume method is applied to solve the convection term similarly to \cite{ZhichengHu2019}. The collision term can be efficiently computed with the new collision model, which consumes much less computational cost than the original quadratic term but maintains reliable numerical accuracy.
In the numerical experiments, we first implement two important spatially homogeneous experiments in the granular gas flow, including the heating source problem\cite{Noije1998} and Haff's cooling law \cite{Haff1983}. Then the tests of one-dimensional benchmark problems are carried out, including the Couette flow and Fourier heat transfer. Finally, a two-dimensional periodic diffusion is simulated to further validate the accuracy and efficiency of the method.
An excellent agreement can be observed between the numerical solution and the reference solution of the direct simulation Monte Carlo (DSMC) method.
The rest of this paper is organized as follows: in Sec. \ref{sec:pre}, we introduce the inelastic collision operator and the general framework of the Hermite spectral method. The algorithm to discretize the collision term and the special simplification for the VHS model are given in Sec. \ref{sec:method}. The whole numerical scheme is completed in Sec. \ref{sec:numerical} with numerical experiments presented in Sec. \ref{sec:experiment}. The paper ends with some concluding remarks in Sec. \ref{sec:conclusion} and several supplementary contents in the Appendix.
\section{Numerical scheme}
\label{sec:numerical}
In this section, the numerical scheme to solve the moment equations \eqref{eq:moment} is introduced. Strang-splitting is utilized to split the moment equation into the convection step and the collision step. Precisely, the numerical scheme to solve the convection step is proposed in Sec. \ref{sec:con} and the specially designed algorithm to solve the collision step is discussed in Sec. \ref{sec:scheme_coll}.
For convenience, we first consider the numerical scheme for spatially one-dimensional cases, which indicates
\begin{equation}
\label{eq:explain1D}
\pd{\cdot}{x_2}=\pd{\cdot}{x_3}=0.
\end{equation}
Therefore, the Boltzmann equation is split into
\begin{itemize}
\item Convection step
\begin{equation}
\label{eq:con}
\pd{f}{t} + v_1 \pd{f}{x_1} = 0.
\end{equation}
\item Collision step
\begin{equation}
\label{eq:col_step}
\pd{f}{t} = \frac{1}{{\rm Kn}}\mQ[f,f](\bv).
\end{equation}
\end{itemize}
\subsection{Convection step}
\label{sec:con}
Before introducing the numerical scheme to solve the convection step, we should choose the expansion center $[\ou, \oT]$ in the expansion \eqref{eq:Her-expan}. Following \cite{ZhichengHu2019, multi2022}, spatially and temporally constant $[\ou, \oT]$ is chosen in the convection step and the specific number is problem-dependent. Then, let $\bdf^{[\ou, \oT]}$ be a column vector with all $f_{\alpha}^{[\ou, \oT]}, |\alpha| \leqslant M$ as its components. Thus, the moment equations \eqref{eq:moment} of the convection step can be rewritten in the matrix-vector form as
\begin{equation}
\label{eq:vec_moment}
\pd{\bdf^{[\ou, \oT]}}{t}+\bA_1^{[\ou, \oT]}\pd{\bdf^{[\ou, \oT]}}{x_1}= 0,
\end{equation}
where $\bA_1^{[\ou, \oT]}$ is a constant matrix and can be diagonalized. We refer the readers to \cite{ZhichengHu2019} for more details.
Then, we propose the numerical scheme to solve the convection term. Suppose a spatial domain $\Omega\subset\bbR$ is discretized by a uniform grid with cell size $\Delta x$ and cell centers $\{x_j\}$. We denote $\left(\bdf_j^{[\ou, \oT]}\right)^n$ as the approximation of the average of $\bdf^{[\ou, \oT]}$ in the $j$th grid cell $[x_j-\frac12\Delta x, x_j+\frac12\Delta x]$ at time $t^n$. The finite element method is utilized to solve the convection part, and the system can be solved by the forward-Euler method with a time step size $\Delta t$ as follows:
\begin{equation}
\label{eq:scheme_con}
\begin{split}
&\left(\bdf^{[\ou, \oT]}\right)_j^{n+1, \ast}=\left(\bdf^{[\ou, \oT]}\right)_j^n-\frac{\Delta t}{\Delta x}\left(\bF_{j+1/2}^n-\bF_{j-1/2}^n\right), \\
\end{split}
\end{equation}
where $\bF_{j+1/2}^n$ is the numerical flux chosen according to HLL scheme \cite{HLL}
\begin{equation}
\label{eq:HLL_flux}
\bF^n_{j+1/2}=
\left\{
\begin{array}{ll}
\bA_1\bdf^{n,L}_{j+1/2},& \la^L\geqslant 0, \\
\frac{\la^R\bA_1\bdf^{n,L}_{j+1/2}-
\la^L\bA_1\bdf^{n,R}_{j+1/2}+\la^R\la^L
\left(\bdf^{n,R}_{j+1/2}-\bdf^{n,L}_{j+1/2}\right)}
{\la^R-\la^L},& \la^L<0<\la^R, \\[2mm]
\bA_1\bdf^{n,R}_{j+1/2},& \la^R\leqslant 0.
\end{array}
\right.
\end{equation}
where the superscript $[\ou, \oT]$ on $\bdf$ is omitted for neatness in \eqref{eq:HLL_flux}. $\lambda^L$ and $\lambda^R$ are the characteristic velocity
\begin{equation}
\la^L=\overline{u}_1 - C_{M+1}\sqrt{\bT},\quad \la^R=\overline{u}_1 + C_{M+1}\sqrt{\bT},
\end{equation}
which are the minimum and maximum eigenvalues of $\bA_1$, and $C_{M+1}$ is the largest root of the Hermite polynomial of degree $M+1$. In \eqref{eq:HLL_flux}, $\bdf^{n,L}_{j+1/2}$ and $\bdf^{n,R}_{j+1/2}$ are computed with WENO reconstruction \cite{Weno}, and the details are given in App. \ref{app:WENO}. In addition, the time step size must be chosen to satisfy the CFL condition
\begin{equation}
\label{eq:CFL}
{\rm CFL}\triangleq \Delta t \frac{|\overline{u}_1| + C_{M+1}\sqrt{\bT}}{\Delta x} < 1.
\end{equation}
Now we complete the numerical scheme of spatially one-dimensional case. This scheme can be naturally extended to spatially three-dimensional situations.
\subsection{Collision step}
\label{sec:scheme_coll}
For the collision step, as is stated, the computational cost to compute the collision term is still quite expensive as $\mO(M^9)$. Therefore, a new collision model is built to reduce cost following the idea in \cite{Approximation2019}. Here, we will introduce the new collision model, and then introduce the numerical scheme in the collision step.
\subsubsection{Building new collision model}
\label{sec:new_col}
To build the new collision model, the quadratic collision term and a simplified operator such as the BGK operator are combined, where the quadratic operator is utilized to deal with the low-order terms in the new collision model, while the high-order terms will be approximated with the simplified collision operator to save memory and computational cost. In the inelastic case, the linear model with a similar form to \eqref{eq:linear} is utilized as
\begin{equation}
\label{eq:new-linear}
\mQ_{L}[f](\bv)=\nu_1\rho\left[f_{G}(\bv)-f(\bv)\right]+\nu_2\rho\nabla_{\bv}\cdot\left[(\bv-\bu)f(\bv)\right].
\end{equation}
Different from \eqref{eq:linear}, there is a $\rho$ here in each term to match the quadratic form of distribution functions in the original collision model. $\nu_1$ and $\nu_2$ are some constant parameters, which will be discussed later.
We first discuss the Hermite expansion of $\mQ_L(\bv)$. Same as the BGK operator in the classical case \cite{ZhichengHu2019, multi2022}, the expansion center is chosen as the local macroscopic velocity and temperature as
\begin{equation}
\label{eq:Her-expan-local}
f(t,\bx,\bv) \approx\sum_{|\alpha| \leqslant M}f^{[\bu,\theta]}_{\alpha}(t,\bx)
\mH_{\alpha}^{[\bu, \theta]}(\bv),
\end{equation}
with which the computational cost for expanding the collision term \eqref{eq:new-linear} can be greatly reduced. Thus, the expansion coefficients of $f_G$ can be computed with \cite{CaiPHD}
\begin{equation}
\label{eq:Her-ESBGK}
f_{G,\alpha}^{[\bu, \theta]}=\left\{\begin{array}{ll}
\rho, & \alpha=0, \\
0, & |\alpha|=1, \\
\frac{1-1/\Pr}{\alpha_i \rho}\sum_{k=1}^3\sigma_{ik}f_{G, \alpha-e_i-e_k}^{[\bu, \theta]}, & |\alpha|\geqslant 2, \; i\in\{1,2,3\} \text{ s.t. } \alpha_i>0.
\end{array}
\right.
\end{equation}
Here we emphasize that the last relationship in \eqref{eq:Her-ESBGK} holds for any $i\in\{1,2,3\}$ subject to $\alpha_i>0$.
With $f_0^{[\bu, \theta]} = \rho$, the energy loss term $\rho \nabla_{\bv} \cdot [(\bv - \bu) f(\bv)]$ is expanded as
\begin{equation}
\label{eq:new-linear-2}
\rho \nabla_{\bv} \cdot [(\bv - \bu) f(\bv)] \approx \sum_{|\alpha|\leqslant M} f_0^{[\bu, \theta]}\left(|\alpha|f_{\alpha}^{[\bu, \theta]}+\sum_{d=1}^3f_{\alpha-2e_d}^{[\bu, \theta]}\right) \mH_{\alpha}^{\bu, \theta}(\bv),
\end{equation}
where $f_{\alpha}^{[\bu, \theta]}$ is regarded as $0$ in \eqref{eq:Her-ESBGK} and \eqref{eq:new-linear-2} if $\alpha$ contains any negative index.
let $\bdf^{[\bu, \theta]}$ be a column vector with all $f_{\alpha}^{[\bu, \theta]}, |\alpha| \leqslant M$ as its components. Combining \eqref{eq:Her-ESBGK} and \eqref{eq:new-linear-2}, the linear collision model $\mQ_L(\bv)$ can be expanded as
\begin{equation}
\label{eq:coeff_QL}
\mQ_L[f](\bv) \approx \sum_{|\alpha|\leqslant M} Q_{L, \alpha}[\bdf^{[\bu, \theta]}] \mH_{\alpha}^{\bu, \theta}(\bv),
\end{equation}
with
\begin{equation}
\label{eq:coe_QL_1}
Q_{L, \alpha}[\bdf^{[\bu, \theta]} ] = \nu_1f_0^{[\bu, \theta]}\left(
f_{G,\alpha}^{[\bu, \theta]} - f_{\alpha}^{[\bu, \theta]}\right) - \nu_2 f_0^{[\bu, \theta]}\left(|\alpha|f_{\alpha}^{[\bu, \theta]}+\sum_{d=1}^3f_{\alpha-2e_d}^{[\bu, \theta]}\right).
\end{equation}
Next, following similar lines in \cite{Approximation2019, multi2022}, we will build the new collision model that is combined with the quadratic collision term \eqref{eq:expan_coll} and linearized operator \eqref{eq:coeff_QL} as
\begin{equation}
\label{eq:new_coll}
\mQ_{\rm new}[f](\bv) = \sum_{|\alpha| \leqslant M} Q_{{\rm new}, \alpha}[\bdf^{[\bu, \theta]}, \bdf^{[\bu, \theta]}] \mH^{[\bu, \theta]}_{\alpha}(\bv),
\end{equation}
with
\begin{equation}
\label{eq:coe_new_coll}
Q_{{\rm new}, \alpha}[\bdf^{[\bu, \theta]}, \bdf^{[\bu, \theta]}]=\left\{
\begin{array}{ll}
\sum\limits_{|\lambda|, |\kappa| \leqslant M_0}\Aalk^{[\bu, \theta]} f^{[\bu, \theta]}_{\lambda}
f^{[\bu, \theta]}_{\kappa}, & |\alpha| \leqslant M_0, \\[4mm]
Q_{L,\alpha}[\bdf^{[\bu, \theta]}], & M_0 < |\alpha|\leqslant M.
\end{array}
\right.
\end{equation}
\begin{remark}
In \eqref{eq:coe_new_coll}, the coefficient $\Aalk^{[\bu, \theta]}$ is derived through $\Aalk$ with Proposition \ref{thm:change_center}. The parameter $M_0$ is problem-dependent, which is always decided empirically. Based on our experience, $M_0 = 10$ is enough for most problems.
\end{remark}
Then let us go back to \eqref{eq:coe_QL_1} and decide the parameters $\nu_1$ and $\nu_2$ to complete the collision model. Since $\nu_1$ indicates the damping rate of high-order terms, we follow the same idea in \cite[Sec. 3.3.2]{Approximation2019} to decide this parameter, where it is expected that the high-order terms decay faster but do not introduce any gap in the damping rate between the terms of $|\alpha|\leqslant M_0$ and $|\alpha|>M_0$. Thus, $\Aalk$ is regarded as a matrix with respect to $\lambda$ and $\kappa$ for each fixed $\alpha$, and $\nu_1$ is set to be the minimum of all eigenvalues of $\Aalk$ among $|\alpha|\leqslant M_0$ (i.e. the spectral radius of damping rate in quadratic part). We refer the readers to \cite[Sec. 3.3.2]{Approximation2019} for more details. As to $\nu_2$, it is borrowed from the cooling rate in \cite[(2.16)]{Astillero2005} that
\begin{equation}
\label{eq:nu2}
\nu_2=\frac{2}{3\sqrt{\pi}}(1-e^2).
\end{equation}
So far, we have derived the new collision model. The numerical scheme to solve the collision step using this new collision model is discussed in the next section.
\subsubsection{Numerical scheme to update the collision step}
\label{sec:num_col}
Based on the new collision model, the numerical scheme to update the collision step is presented in this section. In the framework of the Hermite spectral method, the governing equation in the collision step \eqref{eq:col_step} is reduced into
\begin{equation}
\label{eq:moment_col}
\pd{\bdf^{[\bu, \theta]}_{j}}{t} = \frac{1}{{\rm Kn}} \bQ_{\rm new}[\bdf^{[\bu, \theta]}, \bdf^{[\bu, \theta]}],
\end{equation}
where $\bQ_{\rm new}[\bdf^{[\bu, \theta]}, \bdf^{[\bu, \theta]}]$ is a column vector with all $ Q_{{\rm new}, \alpha}[\bdf^{[\bu, \theta]}, \bdf^{[\bu, \theta]}], |\alpha|\leqslant M$ as its components.
After the convection step at time $t^{n+1}$, $\left(\bdf^{[\ou, \oT]}\right)_j^{n+1, \ast}$ is obtained. Then we derive the expansion coefficients $\left(\bdf^{[\bu^{n+1, \ast}_j, \theta^{n+1, \ast}_j]}\right)_j^{n+1, \ast}$ under the expansion center $[\bu^{n+1, \ast}_j, \theta^{n+1, \ast}_j]$ with the projection algorithm in Theorem \ref{thm:project}. The expansion center $[\bu^{n+1, \ast}_j, \theta^{n+1, \ast}_j]$ is the macroscopic velocity and temperature after the convection step at time $t^{n+1}$, which is obtained by \eqref{eq:macro_f}. Next, the forward Euler scheme is adopted to update \eqref{eq:moment_col} as
\begin{equation}
\label{eq:scheme_col}
\left(\bdf^{\ast}\right)_j^{n+1}=\left(\bdf^{\ast}\right)_j^{n+1, \ast}
+\Delta t \bQ_{\rm new}\left[\left(\bdf^{\ast}\right)^{n+1, \ast}_j, \left(\bdf^{\ast}\right)^{n+1, \ast}_j\right],
\end{equation}
where $\bdf^{\ast} \triangleq \bdf^{[\bu^{n+1, \ast}_j, \theta^{n+1, \ast}_j]}$ for abbreviation.
Finally, the projection algorithm in Theorem \ref{thm:project} is utilized once again to obtain $\left(\bdf^{[\ou, \oT]}\right)_j^{n+1}$ based on $\left(\bdf^{\ast}\right)_j^{n+1}$, which finishes the collision part and enters the next time step.
\subsection{Outline of the numerical algorithm}
The overall numerical scheme is summarized in Algorithm \ref{algo:inhomo}.
\begin{algorithm}[htbp]
\caption{Numerical algorithm}
\label{algo:inhomo}
\begin{algorithmic}[1]
\item Preparation: calculate and store $\Aalk$ in \eqref{eq:Aalk} with the algorithm in Sec. \ref{sec:method}.
\item Let $n=0$, and determine an expansion center $[\ou, \oT]$ for the convection step. Calculate the initial value of $\left(\bdf^{[\ou, \oT]}\right)_{j}^0$.
\item Set the time step $\Delta t^n$ with the CFL condition \eqref{eq:CFL}.
\item Solve the convection step \eqref{eq:scheme_con} to obtain
$\left(\bdf^{[\ou, \oT]}\right)_j^{n+1, \ast}$.
\item Obtain the macroscopic velocity and temperature $\bu^{n+1, \ast}_{j}, \theta^{n+1, \ast}_{j}$ of $\left(\bdf^{[\ou, \oT]}\right)_j^{n+1, \ast}$ using \eqref{eq:macro_f}.
\item Project $\left(\bdf^{[\ou, \oT]}\right)_j^{n+1, \ast}$ to the function space of expansion center $[\bu^{n, \ast}_{j}, \theta^{n, \ast}_{j}]$ with Theorem \ref{thm:project}, and obtain $\left(\bdf^{[\bu_{j}^{n+1, \ast}, \theta_j^{n+1, \ast}]}\right)_j^{n+1, \ast}$.
\item Update $\left(\bdf^{[\bu_{j}^{n+1, \ast}, \theta_j^{n+1, \ast}]}\right)_j^{n+1}$ with \eqref{eq:scheme_col}.
\item Project $\left(\bdf^{[\bu_{j}^{n+1, \ast}, \theta_j^{n+1, \ast}]}\right)_j^{n+1}$ to the function space of expansion center $[\ou, \oT]$, and obtain $\left(\bdf^{[\ou, \oT]}\right)_j^{n+1}$ .
\item Let $n\leftarrow n+1$, and return to step 3.
\end{algorithmic}
\end{algorithm}
\section{Inelastic Boltzmann equation and Hermite spectral method}
\label{sec:pre}
In this section, we will first give a brief review of the Boltzmann equation and the inelastic collision model, and then introduce the general framework for solving the Boltzmann equation using Hermite spectral method.
\subsection{Inelastic Boltzmann equation}
\label{sec:Boltzmann}
The inelastic Boltzmann equation is always utilized to describe the granular gas flow, whose nondimensionalized form \cite{Brilliantov} is
\begin{equation}
\label{eq:Boltzmann}
\pd{f}{t}+\bv\cdot\nabla_{\bx} f=\frac{1}{{\rm Kn}}\mQ[f,f](\bv),
\end{equation}
where $f(t, \bx, \bv)$ is the distribution function, which depends on time $t\in\bbR^+$, physical space $\bx\in\bbR^3$ and particle velocity $\bv\in\bbR^3$. ${\rm Kn}$ is the Knudsen number indicating the ratio of the mean free path to $\mQ$ is the quadratic inelastic collision operator, which takes the form
\begin{equation}
\label{eq:operator}
\mQ[f,f](\bv)=\int_{\mR^3}\int_{S^2}\left[B(|\widetilde{\bg}|,\sigma)
f(\widetilde{\bv})f(\widetilde{\bv}_{\ast})J-B(|\bg|,\sigma)f(\bv)f(\sv)\right]\rd\sigma \rd \sv.
\end{equation}
This collision operator describes the inelastic collision between two particles. Suppose two particles with velocity pair $(\bv,\sv)$ are going to collide. Under the assumption of
inelastic collision, the post-collision velocity pair $(\bv',\sv')$ can be expressed as \cite{Brilliantov}
\begin{equation}
\label{eq:post-velo}
\left\{
\begin{array}{l}
\bv'=\frac{\bv+\sv}{2}+\frac{1-e}4
(\bv-\sv)+\frac{1+e}{4}|\bv-\sv|\sigma,\\
\sv'=\frac{\bv+\sv}{2}-\frac{1-e}4
(\bv-\sv)-\frac{1+e}{4}|\bv-\sv|\sigma,\\
\end{array}
\right.
\end{equation}
where $\sigma$ is a unit vector in $S^2$,
$e\in[0,1]$ is the restitution coefficient indicating the inelasticity.
During the collisions, we can derive the conservation of momentum
\begin{equation}
\label{eq:conserv}
\bv+\sv=\bv'+\sv',
\end{equation}
and the dissipation of energy
\begin{equation}
\label{eq:energy_loss}
|\bv|^2+|\sv|^2-(|\bv'|^{2}+|\sv'|^{2})=\frac{1-e^2}{4}|\bg|(|\bg|-\bg\cdot\sigma),
\end{equation}
where $\bg=\bv-\sv$ is the relative velocity. Define the pre-collisional velocities $(\widetilde{\bv},\widetilde{\bv}_{\ast})$ as the velocity pair which will turn into $(\bv,\sv)$ following the rule in \eqref{eq:post-velo} with the collision of parameter $\sigma$. Here, we want to emphasize that, unlike the elastic case, $(\widetilde{\bv}, \widetilde{\bv}_{\ast})$ does not coincide with $(\bv', \sv')$, and the collisions are not reversible. We define $\tilde{\bg} = \widetilde{\bv} - \widetilde{\bv}_{\ast}$ as the relative velocity of the pre-collisional velocities, and $J=\left\vert \pd{(\widetilde{\bv},\widetilde{\bv}_{\ast})}{(\bv,\sv)}\right\vert$ the determinant of the Jacobian. In \eqref{eq:operator}, $B$ is the collision kernel which depends on the type of interactions. The most commonly used form for the inelastic case is the variable hard sphere (VHS) model \cite{Bird}:
\begin{equation}
\label{eq:VHS}
B(|\bg|,\sigma)=C_{\varpi}|\bg|^{2(1-\varpi)},
\end{equation}
where $C_{\varpi}>0$ and $0.5\leqslant\varpi\leqslant 1$ are some constants.
Especially, the Maxwell molecules and hard sphere model (HS) are corresponding to $\varpi = 0.5$ and $\varpi=1$ respectively.
To derive the weak form of \eqref{eq:coe_new_coll}, multiplying a test function $\phi(\bv)$
and taking integral on $\bv$, it holds that \cite{Hu2019}
\begin{equation}
\label{eq:weak}
\begin{split}
\int_{\mR^3}\mQ[f,f](\bv)\phi(\bv)\rd \bv
= \int_{\mR^3}\int_{\mR^3}\int_{S^2}B(|\bg|,\sigma)f(\bv)f(\sv)
\big[\phi(\bv')-\phi(\bv)\big]\rd\sigma\rd\bv\rd \sv,
\end{split}
\end{equation}
which has a more symmetric form as
\begin{equation}
\label{eq:weak_1}
\begin{split}
&\int_{\mR^3}\mQ[f,f](\bv)\phi(\bv)\rd \bv\\
&\qquad =\frac12\int_{\mR^3}\int_{\mR^3}\int_{S^2}B(|\bg|,\sigma)f(\bv)f(\sv)
\big[\phi(\bv')+\phi(\bv'_{\ast})-\phi(\bv)-\phi(\bv_{\ast})\big]\rd\sigma\rd\bv\rd \sv.
\end{split}
\end{equation}
We refer \cite{Hu2019} for the derivation of this weak form. Specifically, $\phi(\bv)$ can be chosen as any polynomial of $\bv$, which is quite important in the framework of the Hermite spectral method to discretize the collision operator. Especially, let $\phi = |\bv|^2$, From \eqref{eq:weak_1} and \eqref{eq:energy_loss}, we can derive that the inelastic collision does not conserve the total energy as
\begin{equation}
\label{eq:non-conser-E}
\int_{\mR^3}\mQ[f,f](\bv)|\bv|^2\rd\bv < 0.
\end{equation}
Due to the complex form of the quadratic collision operator, several simplified collision models are proposed. Since the total energy is not conserved during the inelastic collision, the BGK collision model \cite{BGK} does not work well in the inelastic case. In \cite[(3.12)]{Filbet2013}, a linearized collision term with energy dissipation is constructed with
\begin{equation}
\label{eq:linear}
\mQ_{\rm lin}(\bv)= \nu_1
\tau(P)\left[f_{G}(\bv)-f(\bv)\right]+
\nu_2 \nabla_{\bv}\cdot\left[(\bv-\bu)f(\bv)\right],
\end{equation}
where the parameters $\nu_1$ and $\nu_2$ is problem dependent. In \cite{Filbet2013}, $\nu_1 = \tau(P)$, where $\tau(\cdot)$ is a given function depending on the kinetic pressure $P:=\rho\theta$ and $\nu_2 = C \rho \theta^{1/2} $ with $C$ a given constant. A similar model is also proposed in \cite[(2.20)]{Astillero2005}, which contains the same energy loss term, but with different parameters $\nu_2$.
The first term in \eqref{eq:linear} is the ellipsoidal statistical BGK (ES-BGK) operator, which has the form \cite{Holway}
\begin{equation}
\label{eq:ESBGK}
\begin{split}
f_{G}&=\frac{\rho}{|2\pi \Lambda|}\exp\left(-\frac12 (\bv-\bu)^T\Lambda (\bv-\bu)\right), \\
\Lambda&=(\lambda_{ij})\in \bbR^{3\times 3}, \quad \lambda_{ij}=\frac{1}{\Pr}\theta\delta_{ij}+\left(1-\frac{1}{\Pr}\right)\frac{\sigma_{ij}+\delta_{ij}\rho\theta}{\rho},
\end{split}
\end{equation}
where $\delta_{ij}$ is the Kronecker delta, $\Pr$ is the Prandtl number, which takes the value $\frac23$ for monotonic gas. $\rho, \bu, \theta$ and $\boldsymbol{\sigma}$ are the density, macroscopic velocity, temperature and stress tensor, respectively, whose relationships with the distribution function are
\begin{equation}
\label{eq:macro}
\begin{aligned}
&\rho(t,\bx) = \int_{\mR^3}f(t,\bx,\bv)\rd\bv, \qquad
\bu(t,\bx)=\frac{1}{\rho}\int_{\mR^3}\bv f(t,\bx,\bv)\rd\bv, \\
&3\theta(t,\bx)\rho(t, \bx)= \int_{\mR^3}|\bv-\bu|^2f(t,\bx,\bv)
\rd\bv, \\
&\boldsymbol{\sigma}(t,\bx)=\int_{\mR^3}\left[(\bv-\bu)\otimes(\bv-\bu)-\frac13|\bv-\bu|^2I\right] f(t,\bx,\bv)\rd\bv.
\end{aligned}
\end{equation}
Moreover, the heat flux $\boldsymbol{q}$ could be derived from the distribution function as
\begin{equation}
\label{eq:q}
\boldsymbol{q}(t,\bx) = \frac12 \int_{\mR^3}(\bv-\bu)|\bv-\bu|^2 f(t,\bx,\bv)\rd\bv.
\end{equation}
For now, we have introduced the inelastic Boltzmann equation. Due to its high dimensionality and the complex collision term, it is always a challenge to solve it numerically. Several numerical methods have been proposed to deal with the inelastic Boltzmann equation, such as the DSMC method \cite{Gamba2005, Astillero2005}, the Fourier spectral method \cite{Hu2019, Wu2015} and the Petrov--Galerkin spectral method \cite{Hu2020}. In this work, we will develop a numerical scheme based on the Hermite spectral method, which may handle the evolution of temperature more accurately than other methods.
\subsection{Hermite spectral method}
\label{sec:Hermite}
This subsection presents the general framework for solving the inelastic Boltzmann equation using the Hermite spectral method. We begin from the local Maxwellian as
\begin{equation}
\label{eq:maxwellian}
\mM_{\bu, \theta}(\bv) = \frac{\rho}{(2\pi \theta)^{\frac32}}
\exp\left(-\frac{|\bv-\bu|^2}{2 \theta}\right),
\end{equation}
which is the steady-state solution to the elastic case ($e=1$ in \eqref{eq:post-velo}). Here, $\rho, \bu, \theta$ are the density, macroscopic velocity and temperature defined in \eqref{eq:macro}, respectively. The Hermite spectral method is successfully utilized to solve the elastic Boltzmann equation \cite{ZhichengHu2019, Approximation2019}, where the local Maxwellian \eqref{eq:maxwellian} is adopted as the weight function to generate orthogonal polynomials(i.e. Hermite polynomials), which are utilized as the basis polynomials. Precisely, assuming the expansion center is $[\ou, \oT] \in \bbR^3 \times \bbR^+$, then the weight function has the form as
\begin{equation}
\label{eq:maxwellian_expan}
\omega_{\ou, \oT}(\bv) = \frac{1}{(2\pi \oT)^{\frac32}}
\exp\left(-\frac{|\bv-\ou|^2}{2 \oT}\right),
\end{equation}
and the corresponding Hermite polynomials are defined as
\begin{definition}[Hermite Polynomials]
\label{def:Her}
For $\alpha=(\al_1,\al_2,\al_3) \in \bbN^3$, the three-dimensional
Hermite polynomial $H_{\alpha}^{\ou,\oT}(\bv)$ is defined as
\begin{equation}
\label{eq:Hermite}
H_{\alpha}^{\ou,\oT}(\bv)=\frac{(-1)^{|\alpha|}\oT^{\frac{|\alpha|}{2}}}
{\omega_{\ou,\oT}(\bv)}
\dfrac{\partial^{|\alpha|}}{\partial \bv^{\alpha}}
\omega_{\ou,\oT}(\bv),
\end{equation}
with $|\alpha|=\al_1+\al_2+\al_3$ and $\partial \bv^{\alpha} = \pa v_1^{\al_1}\pa v_2^{\al_2}\pa v_3^{\al_3}$. The Hermite polynomials have several useful properties when approximating the complex collision term, which is listed in Appendix \ref{app:Her}.
\end{definition}
Then following the similar routine in \cite{Approximation2019}, the distribution function $f(t, \bx, \bv)$ is approximated as
\begin{equation}
\label{eq:Her-expan}
f(t,\bx,\bv) \approx \sum_{|\alpha| \leqslant M}f^{[\ou,\oT]}_{\alpha}(t,\bx)
\mH\aut(\bv),
\end{equation}
where $\mH\aut(\bv)=H\aut(\bv)\omega_{\ou, \oT}(\bv)$ are the basis functions. $|\alpha| = \alpha_1 + \alpha_2 + \alpha_3$ and $M \in \mathbb{N}$ is the expansion order. Here, we want to emphasize that with a properly chosen expansion center, the computational cost may be greatly reduced. In practice, the expansion center is
often chosen as a rough average of the whole domain to ensure numerical accuracy. With the orthogonality of the basis function \eqref{eq:orth}, it holds
\begin{equation}
\label{eq:falpha}
f^{[\ou,\oT]}_{\alpha}(t,\bx)=\frac{1}{\alpha!}\int_{\mR^3}f(t,\bx,\bv) H\aut(\bv)
\rd\bv.
\end{equation}
With the expansion \eqref{eq:Her-expan}, the macroscopic variables in \eqref{eq:macro}, \eqref{eq:q} could be expressed with $f^{[\ou,\oT]}_{\alpha}$ as
\begin{equation}
\label{eq:macro_f}
\begin{split}
&\rho = f^{[\ou,\oT]}_{0},\quad u_k= \overline{u}_k+\frac{\sqrt{\oT}}{\rho}f^{[\ou,\oT]}_{e_k},
\quad \theta=\frac{2\oT}{3\rho}\sum_{k=1}^3f^{[\ou,\oT]}_{2e_k}+\oT-\frac{1}{3\rho}|\bu-\ou|^2, \\
&\sigma_{kl}=(1+\delta_{kl})\oT f^{[\ou,\oT]}_{e_i+e_j}+\delta_{kl}\rho\left(\oT - \theta\right)-
\rho\left(\overline{u}_k-u_k\right)\left(\overline{u}_l-u_l\right), \\
&q_k=2\oT^{\frac32}f^{[\ou,\oT]}_{3e_k}+(\overline{u}_k-u_k)\oT f^{[\ou,\oT]}_{2e_k}+|\ou-\bu|^2\sqrt{\oT} f^{[\ou,\oT]}_{e_k} +\\
& \qquad \qquad\sum_{l=1}^3\left[\oT^{\frac32}f^{[\ou,\oT]}_{2e_l+e_k}+\left(\overline{u}_l-u_l\right)\oT f^{[\ou,\oT]}_{e_l+e_k}+
\left(\overline{u}_k-u_k\right)\oT f^{[\ou,\oT]}_{2e_l}\right], \qquad k, l = 1, 2, 3.
\end{split}
\end{equation}
Moreover, if the expansion center is chosen as the local macroscopic velocity and temperature as $[\ou, \oT] = [\bu, \theta]$, then \eqref{eq:macro_f} is reduced into
{\small
\begin{equation}
\label{eq:macro_f_1}
\begin{aligned}
f^{[\bu, \theta]}_{e_k} = 0, \qquad \sum_{k=1}^3 f^{[\bu, \theta]}_{2e_k} = 0, \qquad
\sigma_{kl} = (1+\delta_{kl}) \theta f^{[\bu, \theta]}_{e_i + e_j}, \qquad
q_k = \theta^{\frac32}\left[2f^{[\bu, \theta]}_{3e_k}+ \sum_{l=1}^3 f^{[\bu, \theta]}_{2e_l + e_k}\right].
\end{aligned}
\end{equation}
}Therefore, the macroscopic quantities could be easily obtained under the framework of the Hermite spectral method. This is to say, we are able to govern the evolution of some important macroscopic variables with a small $M$.
For the collision term, it is also expanded using the same basis functions and approximated as
\begin{equation}
\label{eq:expan_coll}
\mQ[f,f](\bv)\approx\sum_{|\alpha|\leqslant M}Q^{[\ou,\oT]}_{\alpha}(t, \bx)\mH\aut(\bv),\qquad Q^{[\ou,\oT]}_{\alpha}(t, \bx)=\frac{1}{\alpha!}\int_{\mR^3}\mQ[f,f](\bv)
H\aut(\bv)\rd \bv.
\end{equation}
Then, substituting the expansion \eqref{eq:Her-expan}, \eqref{eq:expan_coll} into \eqref{eq:Boltzmann} and matching the coefficients on both sides, the moment equations could be derived as
\begin{equation}
\label{eq:moment}
\pd{}{t}f^{[\ou,\oT]}_{\alpha}+\sum_{d=1}^3\pd{}{x_d}\left((\alpha_d+1)\sqrt{\ot}f^{[\ou,\oT]}_{\alpha+e_d}+\overline{u}_df^{[\ou,\oT]}_{\alpha}+\sqrt{\ot}f^{[\ou,\oT]}_{\alpha-e_d}\right)=Q^{[\ou,\oT]}_{\alpha},\quad |\alpha|\leqslant M,
\end{equation}
where the recurrence relationship \eqref{eq:recur} is utilized to deal with the convection term. In \eqref{eq:moment}, $f^{[\ou,\oT]}_{\alpha}$ is regarded as $0$ if $\alpha$ contains any negative index or $|\alpha|>M$.
For now, we have derived the moment equations for the Boltzmann equation. The main difficulty to solve \eqref{eq:moment} lies in the approximation of $Q^{[\ou,\oT]}_{\alpha}(t, \bx)$ in \eqref{eq:expan_coll}, which will be discussed in detail in the following section.
\section{Proof of Theorem \ref{thm:step1}}
\label{app:thm1}
In order to prove Thm. \ref{thm:step1}, we first introduce the lemma below:
\begin{lemma}
\label{lemma:tran-Hermite}
Define $\bv=\bh+\frac12\bg$, $\bw=\bh-\frac12\bg$, then
\begin{equation}
\label{eq:tran-Hermite}
\Hl(\bv)\Hk(\bw)=\sum_{\ka'+\la'=\ka+\la}\ca1\ca2\ca3
H_{\la'}(\sqrt{2}\bh)H_{\ka'}\left(\frac1{\sqrt2}\bg\right),
\end{equation}
where the coefficients $\ca{d}$ are defined in \eqref{eq:ca}.
\end{lemma}
The proof of Lemma \ref{lemma:tran-Hermite} can be referred to \cite[Lemma 3]{Approximation2019}. Besides, we can derive the corollary
\begin{corollary}
\label{corollary:tran-Hermite1}
Define $\bv=\bh+\frac12\bg$, then it holds that
\begin{equation}
\label{tran-Hermite1}
\Hl(\bv)=2^{-\frac{|\la|}{2}}\sum_{\ka'+\la'=\la}
\frac{\la!}{\ka'!\la'!}
H_{\ka'}(\sqrt{2}\bh)H_{\la'}\left(\frac1{\sqrt2}\bg\right).
\end{equation}
\end{corollary}
The proof is straightforward by letting $\ka=\boldsymbol{0}$ in \eqref{eq:tran-Hermite}. Then we can present the proof of Theorem \ref{thm:step1}.
\begin{proof}[Proof of Theorem \ref{thm:step1}]
First, we rewrite \eqref{eq:Aalk} as
\begin{equation}
\label{eq:Aalk2}
\begin{split}
A\alk=&\frac{1}{\alpha!}\int_{\mR^3}\int_{\mR^3}\int_{S^2}B
(|\bg|,\sigma)
\Hl(\bv)\Hk(\sv)[\Ha(\bv')-\Ha(\bv)]\omega(\bv)\omega(\sv)\rd\sigma \rd\bv \rd\sv.
\end{split}
\end{equation}
Define $\bh=\frac{\bv+\sv}2=\frac{\bv'+\sv'}2$ and notice $\bg=\bv-\sv$ is the relative velocity. Besides, with $\bg'$ defined in \eqref{eq:coe_g}, we have
\begin{equation}
\label{eq: dd_con}
\begin{split}
&\bv=\bg+\frac12\bh, \qquad \sv=\bg-\frac12\bh, \qquad \bv'=\bg'+\frac12\bh,\\
&\rd \bg \rd \bh=\rd \bv \rd \bv_{\ast}, \qquad \omega(\bv)\omega(\sv)=\omega\left(\frac1{\sqrt2}\bg\right)\omega(\sqrt2 \bh).
\end{split}
\end{equation}
Combining Lemma \ref{lemma:tran-Hermite}, Corollary \ref{corollary:tran-Hermite1} and \eqref{eq: dd_con}, we can transform \eqref{eq:Aalk2} into an integral with respect to $\bg$ and $\bh$:
\begin{equation}
\label{eq:Aalk3}
\Aalk=2^{\frac{|\al|}{2}}
\sum_{\la'+\ka'=\la+\ka}\sum_{\bi+\bj=\al}
\frac{1}{\bi!\bj!}\ca1\ca2\ca3
\gamma_{\ka'}^{\bj}\eta_{\la'}^{\bi},
\end{equation}
where $\gamma_{\ka'}^{\bj}$ is the integral of $\bg$ defined in \eqref{eq:gamma}, and $\eta_{\la'}^{\bi}$ is the integral of $\bh$ defined by
\begin{equation}
\label{eq:eta}
\eta_{\la'}^{\bi}=\int_{\mR^3}H_{\la'}(\sqrt{2}\bh)H_{\bi}
(\sqrt{2}\bh)
\omega(\sqrt{2}\bh)\rd\bh=\frac{\la'!}{2^{\frac32}}\de_{\la',\bi},
\end{equation}
which can be calculated with the orthgonality \eqref{eq:orth}. The proof is completed by substituting \eqref{eq:eta} into \eqref{eq:Aalk3}.
\end{proof}
\section{The proof for VHS kernel}
\label{app:VHS}
In this section, the proof to Lemma \ref{thm:int_Her}, Proposition \ref{thm:coe_D_psi} and \ref{thm:Maxwell} is proposed.
{\renewcommand\proofname{Proof of Lemma \ref{thm:int_Her}}
\begin{proof}
From the recurrence relationship \eqref{eq:recur} of Hermite polynomials, the one-dimensional standard Hermite polynomial can be expanded as
\begin{equation}
\label{eq:coef_Her}
\begin{split}
&H_{2n}(x)=\sum_{k=0}^n\frac{(2n-1)!!}{(2n-2k-1)!!}
(-1)^kC_n^kx^{2n-2k}, \\
&H_{2n-1}(x)=\sum_{k=0}^{n-1}\frac{(2n-1)!!}{(2n-2k-1)!!}
(-1)^kC_{n-1}^kx^{2n-2k-1}.
\end{split}
\end{equation}
Substituting \eqref{eq:coef_Her} into \eqref{eq:int_poly}, and let $\mathcal{C}(\alpha_i,j_i), i = 1, 2,3$ be the coefficient of $x^{j_i}$ in $H_{\alpha_i}(x)$, it holds that
\begin{equation}
\label{eq:int_poly_1}
\mathcal{V}(\kappa, \alpha, \mu) = \sum_{\sk{\bj\pq\alpha \\ 2|(\alpha-\bj)}}\mathcal{C}(\alpha_1, j_1)\mathcal{C}(\alpha_2, j_2)\mathcal{C}(\alpha_3, j_3) \int_{\bbR^3} \bv^{\kappa} \bv^{\bj}|\bv|^{\mu}\omega(\bv) \rd \bv.
\end{equation}
With the spherical coordinate transform $\bv=(r\chi_1,r\chi_2,r\chi_3)$, where $r\in \bbR^+$ and $\chi=(\chi_1, \chi_2, \chi_3)\in S^2$, it holds that
\begin{equation}
\label{eq:mV}
\begin{split}
\mathcal{V}(\kappa, \alpha, \mu)
=&(2\pi)^{-\frac32}
\sum_{\sk{\bj\pq\alpha \\ 2|(\alpha-\bj)}}
\mathcal{C}(\alpha,\bj)
\int_0^{\infty}r^{2+\mu+|\bj|+|\kappa|}\exp\left(-\frac{r^2}{2}\right)\rd r
\int_{S^2} \chi_1^{j_1+\kappa_1}\chi_2^{j_2+\kappa_2}
\chi_3^{j_3+\kappa_3}\rd\chi.
\end{split}
\end{equation}
With Lemma \ref{thm:sp_int} and the property of the Gamma function, \eqref{eq:mV} is simplified as
\begin{equation}
\label{eq:mV_final}
\begin{split}
\mathcal{V}(\kappa, \alpha, \mu)
=(2\pi)^{-\frac32}\sum_{\sk{\bj\pq\alpha \\ 2|(\alpha-\bj)}}
\mathcal{C}(\alpha, \bj)2^{\frac{1+\mu+|\bj|+|\kappa|}2} \Gamma\left(\frac{3+\mu+|\bj|+|\kappa|}2\right)\mathcal{S}(\bj+\kappa),
\end{split}
\end{equation}
Which finishes the proof.
\end{proof}
}
{\renewcommand\proofname{Proof of Proposition \ref{thm:coe_D_psi}}
\begin{proof}
To prove \eqref{eq:lemma_D}, we first expand $H_{\alpha}(\bg')$ using \eqref{eq:coef_Her} as
\begin{equation}
\label{eq:expan_Hg'}
H_{\alpha}(\bg')=\sum_{\sk{\lambda\pq\alpha\\ 2|(\alpha-\lambda)}}\mathcal{C}(\al_1,\la_1)\mathcal{C}(\al_2,\la_2)\mathcal{C}(\al_3,\la_3)(\bg')^{\lambda}.
\end{equation}
The binomial expansion of $(\bg')^{\lambda}$ holds that
\begin{equation}
\label{eq:coe_D_1}
(\bg')^{\lambda} = \sum_{\sk{\kappa\pq\la }} C_{\lambda}^{\kappa}
\left(\frac{1-e}{2}\right)^{|\kappa|}\left(\frac{1+e}{2}\right)
^{|\la|-|\kappa|} \bg^{\kappa} |\bg|^{|\lambda|-|\kappa|} \sigma^{\lambda - \kappa}.
\end{equation}
From Lemma \ref{thm:sp_int}, we can derive that
\begin{equation}
\label{eq:coe_D_2}
\int_{S^2} \sigma^{\lambda - \kappa} \neq 0, \quad \text{ if and only if } 2|(\lambda - \kappa).
\end{equation}
Combining \eqref{eq:coe_D}, \eqref{eq:coe_D_1} and \eqref{eq:coe_D_2}, we can directly derive \eqref{eq:lemma_D}. For \eqref{eq:lemma_psi}, the proof is straightforward with the expansion of $H_{\alpha}(\bg)$ and $\int_{S^2}1\rd\sigma=4\pi$.
\end{proof}
}
{\renewcommand\proofname{Proof of Proposition \ref{thm:Maxwell}}
\begin{proof}
To prove this proposition, we begin from \eqref{eq:coe_D} and \eqref{eq:coe_psi}. From \eqref{eq:coe_D_1} and \eqref{eq:coe_D_2}, we can deduce that $\bg^{\kappa}|\bg|^{|\lambda| - |\kappa|}$ is a polynomial of $\bg$ with degree $|\lambda|$. With the orthogonality of Hermite polynomials, it holds that
$$D(\bj, \kappa, 0)=0, \quad \text{ if } |\bj|<|\kappa|. $$
When $\mu = 0$, it is obvious that $\psi(\bj,\kappa,0)$ can be nonzero only when $\bj = \kappa$. Thus, one can see from \eqref{eq:cal_gamma} that
$$\gamma_{\kappa}^{\bj} =0, \quad \text{ if } \varpi = 1~\text{and}~ |\bj|<|\kappa|. $$ Finally, when $|\alpha| < |\lambda| + |\kappa|$ in $A\alk$, it can be observed that $|\kappa'|-|\bj|=|\kappa|+|\lambda|-|\alpha|>0$ in the summation \eqref{eq:theorem_A}. We complete the proof.
\end{proof}
}
\section{Projection operator}
\label{app:project}
In this section, we present the theorem of the projection operator between different expansion centers. We refer the readers to \cite[Theorem 3.1]{ZhichengHu2019} for the related proof and details of this projection algorithm.
\begin{theorem}
\label{thm:project}
Suppose $f(\bv) \in \mF$ is expanded with two
different expansion centers $[\ou\UP[1],\oT\UP[1]]$ and
$[\ou\UP[2],\oT\UP[2]]$. From \eqref{eq:falpha}, we can compute these two sets of expansion coefficients as
\begin{equation}
\label{eq:twoexpan}
\begin{split}
& f^{[\ou\UP[1],\oT\UP[1]]}_{\al}=\frac{1}{\al!}\int H_{\alpha}^{\ou\UP[1],\oT\UP[1]}(\bv)f(\bv)\rd\bv \\
& f^{[\ou\UP[2],\oT\UP[2]]}_{\al}=\frac{1}{\al!}\int H_{\alpha}^{\ou\UP[2],\oT\UP[2]}(\bv)f(\bv)\rd\bv \\
\end{split}
\end{equation}
Then we have
\begin{equation}
\label{eq:project_expan}
f^{[\ou\UP[2],\oT\UP[2]]}_{\al}=\Big(\oT\UP[2]\Big)^{-\frac{|\al|}2}\sum_{l=0}^{|\al|}\phi_{\alpha}\UP[l],
\end{equation}
where $\phi_{\alpha}\UP[l]$ is defined recursively by
\begin{equation}
\label{eq:def_phi}
\phi_{\alpha}\UP[l]=\left\{
\begin{array}{ll}
\Big(\oT\UP[1]\Big)^{\frac{|\alpha|}2}f^{[\ou\UP[1],\oT\UP[1]]}_{\al}, & l=0, \\
\frac{1}{l}\sum_{d=1}^3\left[\Big(\ou\UP[2]-\ou\UP[1]\Big)\phi_{\al-e_d}\UP[l-1]+\frac12 \Big(\oT\UP[2]-\oT\UP[1]\Big)\phi_{\al-2e_d}\UP[l-1]\right], & 1\leqslant l \leqslant |\al|.
\end{array}
\right.
\end{equation}
In \eqref{eq:def_phi}, terms with any negative index are
regarded as $0$.
\end{theorem}
\section{WENO reconstruction}
\label{app:WENO}
In this section, the WENO reconstruction for $\bdf$ is listed, and the specific reconstruction coefficients are as below
\begin{equation}
\label{eq:WENO}
\begin{split}
& \bdf^{L,1}=\frac32\bdf^n_{j}-\frac12\bdf^n_{j-1}, \quad
\bdf^{L,2}=\frac12\bdf^n_{j}+\frac12\bdf^n_{j+1}, \\
& \bdf^{R,1}=\frac32\bdf^n_{j}-\frac12\bdf^n_{j+1}, \quad
\bdf^{R,2}=\frac12\bdf^n_{j}+\frac12\bdf^n_{j-1}, \\
& \omega_{L,1}=\frac{\ga_1}{\Big[\varepsilon+(\bdf^n_{j}-\bdf^n_{j-1})^2\Big]^2}, \quad
\omega_{L,2}=\frac{\ga_2}{\Big[\varepsilon+(\bdf^n_{j+1}-\bdf^n_{j})^2\Big]^2}, \\
& \omega_{R,1}=\frac{\ga_1}{\Big[\varepsilon+(\bdf^n_{j+1}-\bdf^n_{j})^2\Big]^2}, \quad
\omega_{R,2}=\frac{\ga_2}{\Big[\varepsilon+(\bdf^n_{j}-\bdf^n_{j-1})^2\Big]^2}, \\
&\bdf^{n,L}_{j+1/2}=\frac{\omega_{L,1}\bdf^{L,1}+\omega_{L,2}\bdf^{L,2}}
{\omega_{L,1}+\omega_{L,2}}, \quad
\bdf^{n,R}_{j-1/2}=\frac{\omega_{R,1}\bdf^{R,1}+\omega_{R,2}\bdf^{R,2}}
{\omega_{R,1}+\omega_{R,2}}, \\
& \varepsilon=10^{-6},\quad \ga_1=\frac13, \quad \ga_2=\frac23,
\end{split}
\end{equation}
where the square of $\bdf$ in \eqref{eq:WENO} indicates squaring by each element and the superscript $[\ou, \oT]$ on $\bdf$ is omitted.
\section{Nondimensionalization}
\label{app:nondim}
In this section, we provide the nondimensionalization to scale the variables as
\begin{equation}
\label{eq:nondim}
\hat{\bx} = \frac{\bx}{x_0}, \qquad \hat{\bv} = \frac{\bv}{u_0}, \qquad \hat{t}=\frac{t}{x_0/u_0}, \qquad \hat{m} = \frac{m}{m_0}, \qquad \hat{f} = \frac{f}{\rho_0 / (m_0 u_0^3)}, \qquad \hat{B}=\frac{B}{B_0},
\end{equation}
where $x_0, \rho_0$, and $m_0$ are the characteristic length, density and mass. Besides, $u_0$ is the character velocity defined as $u_0 = \sqrt{k_B \theta_0 / m_0}$ with $\theta_0$ the characteristic temperature. Besides, $B_0=\sqrt2u_0\pi d_{\rm ref}^2$ is adopted to rescale the HS collision kernel, where $d_{\rm ref}$ is the reference diameter.
The characteristic parameters utilized for Argon and the HS model are listed in Tab. \ref{table:character}, where $d_{\rm ref}$ is derived with \cite[Eq. (4.62)]{Bird}.
\begin{table}[!hptb]
\centering
\def1.5{1.5}
{\footnotesize
\begin{tabular}{ll}
\hline
Characteristic parameters for Argon: & \\
Characteristic mass $m_0$ ($\times 10^{-26}$kg) & 6.63 \\
Characteristic length $x_0$ (m) & $10^{-3}$ \\
Characteristic velocity $u_0$ (m/s) & 238.377 \\
Characteristic temperature $\theta_0$ (K) & 273 \\
\hline
Paramerters for HS model: & \\
Molecular mass: $m$ ($\times 10^{-26}$kg) & 6.63 \\
Ref. viscosity: $\mu_{\rm ref}$ ($\times 10^{-5}$Pa s) & 2.117 \\
Viscosity index: $\varpi$ & 0.5 \\
Scattering parameter: $\alpha$ & 1 \\
Ref. diameter: $d_{\rm ref}$ ($\times 10^{-10}m$) & 3.63 \\
Ref. temperature: $T_{\rm ref}$ (K) & 273 \\
\hline
\end{tabular}
}
\caption{(Nondimensionalization in App. \ref{app:nondim}) Characteristic parameters in inhomogeneous tests.}\label{table:character}
\end{table}
\section{Properties of Hermite polynomials}
\label{app:Her}
For the Hermite polynomials \eqref{eq:Hermite}, several important properties are listed below
\begin{property}
(Orthogonality)
\end{property}
\begin{equation}
\label{eq:orth}
\int_{\mR^3}H\aut(\bv)H\but(\bv)\omega_{\ou,\ot}(\bv)\rd\bv=\al_1!\al_2!\al_3!
\delta_{\alpha,\beta}.
\end{equation}
\begin{property}
(Transitivity)
\end{property}
\begin{equation}
\label{eq:tran}
H\aut(\bv)=H_{\alpha}^{\boldsymbol{0},\zeta}\left(\sqrt{\frac{{\zeta}}{{\;\bT\;}}}(\bv-\ou)\right).
\end{equation}
\begin{property}
\label{property:recur}
(Recurrence)
\end{property}
\begin{equation}
\label{eq:recur}
\begin{split}
&H^{\ou,\oT}_{\alpha+e_d}(\bv)=\frac{v_d-u_d}{\sqrt{\oT}}H^{\ou,\oT}
_{\alpha}(\bv)-\al_dH^{\ou,\oT}_{\alpha-e_d}(\bv), \\
& v_dH^{\ou,\oT}_{\alpha}=\sqrt{\oT}H^{\ou,\oT}_
{\alpha+e_d}+
u_dH^{\ou,\oT}_{\alpha}+\al_d\sqrt{\oT}H^{\ou,\oT}_{\alpha-e_d}.
\end{split}
\end{equation}
\begin{property}(Differential of Hermite polynomial)
\label{property:deri}
\begin{equation}
\label{eq:diff_Her}
\frac{\pa}{\pa v_d}H\aut(\bv)=\frac{\al_d}{\sqrt{\oT}}H_{\alpha-e_d}^{\ou,\bT}
(\bv).
\end{equation}
\end{property}
The property of transitivity can be directly derived from the definition \eqref{eq:Hermite}, and the proof of the other properties can be found in \cite{Abramowitz1964}.
\end{appendix}
\section{Approximation of the collision terms}
\label{sec:method}
In this section, we introduce the approximation of the collision term in the framework of the Hermite spectral method.
The discretization of the quadratic term is discussed in Sec. \ref{sec:coll}, and the simplification of the VHS model is presented in Sec. \ref{sec:VHS}.
\subsection{Series expansion of general collision terms}
\label{sec:coll}
We first discuss the algorithm to calculate the expansion coefficients $Q^{[\ou,\oT]}_{\alpha}(t, \bx)$ of the quadratic collision term in \eqref{eq:expan_coll}. Without loss of generality, we suppose the expansion center $[\ou, \ot]=[\boldsymbol{0}, 1]$ and omit the superscript $\ou, \ot$ for simplicity as
{\small
\begin{equation}
\label{eq:short}
f_{\alpha}(t, \bx) = f^{[\boldsymbol{0}, 1]}_{\alpha}(t,\bx), \qquad Q_{\alpha}(t,\bx) = Q^{[\boldsymbol{0}, 1]}_{\alpha}(t,\bx), \qquad H_{\alpha}(\bv) = H^{\boldsymbol{0}, 1}_{\alpha}(\bv), \qquad \omega(\bv) = \omega_{\boldsymbol{0}, 1}(\bv).
\end{equation}
}
With the weak form \eqref{eq:weak}, we can simplify \eqref{eq:expan_coll} as
\begin{equation}
\label{eq:simp1_Q}
Q_{\alpha}(t, \bx)
=\int_{\mR^3}\int_{\mR^3}\int_{S^2}
B(|\bg|,\sigma)f(\bv)f(\sv)\big[\Ha(\bv')-\Ha(\bv)
\big]\rd\sigma\rd\bv\rd \sv.
\end{equation}
Substituting the expansion of the distribution function \eqref{eq:Her-expan} into \eqref{eq:simp1_Q}, we can deduce that
\begin{equation}
\label{eq:simp2_Q}
Q_{\alpha}(t, \bx) = \sum_{|\lambda|\leqslant M}\sum_{|\kappa| \leqslant M}A\alk f_{\lambda}f_{\kappa},
\end{equation}
with
\begin{equation}
\label{eq:Aalk}
\begin{split}
A\alk=&\frac{1}{\alpha!}\int_{\mR^3}\int_{\mR^3}\int_{S^2}B
(|\bg|,\sigma)
\mHl(\bv)\mHk(\sv)[\Ha(\bv')-\Ha(\bv)]
\rd\sigma \rd\bv \rd\sv.
\end{split}
\end{equation}
With the properties of Hermite polynomials, the calculation of $A\alk$ can be greatly simplified and the result is listed in Theorem \ref{thm:step1}.
\begin{theorem} \label{thm:step1}
The coefficients $A\alk$ have the form below:
\begin{equation}
\label{eq:theorem_A}
A\alk=\frac{1}{2^{\frac{|\alpha|+3}{2}}}
\sum_{\lambda'\pq\alpha,\lambda'\pq\kappa+\lambda}\frac{1}{\bj!}
\ca1\ca2\ca3\ga_{\kappa'}^{\bj},
\end{equation}
where $\kappa'=\kappa+\lambda-\lambda'$, $\bj=\alpha-\lambda'$, and the symbol `$\pq$' means
$$(i_1,i_2,i_3) \pq (j_1,j_2,j_3) \Leftrightarrow
i_s\leqslant j_s \quad\text{for}\quad s=1,2,3.$$
Besides, the coefficients $\ca{d}$ and $\ga_{\kappa}^{\bj}$ in \eqref{eq:theorem_A} are given by
\begin{equation}
\label{eq:ca}
\ca{d}=2^{-\frac{l'_d+k'_d}{2}}
\sum_{s\in \mZ}C_{l_d}^sC_{k_d}^{l'_d-s}
(-1)^{k_d-l'_d+s},
\end{equation}
and
\begin{equation}
\label{eq:gamma}
\begin{split}
\ga_{\kappa}^{\bj}=\int_{\mR^3}\int_{S^2}
&\left[H_{\bj}\left({{\frac{\bg'}{\sqrt2}}}\right)-H_{\bj}
\left({{\frac{\bg}{\sqrt2}}}\right)\right]
\Hk\left({{\frac{\bg}{\sqrt2}}}\right)B(|\bg|,\sigma)
\omega\left({{\frac{\bg}{\sqrt2}}}\right) \rd \sigma \rd \bg,
\end{split}
\end{equation}
where from \eqref{eq:post-velo}, it hold that for $\bg'$ that
\begin{equation}
\label{eq:coe_g}
\bg'\triangleq \bv'-\sv'=\frac{1-e}{2}\bg+\frac{1+e}{2}|\bg|\sigma.
\end{equation}
\end{theorem}
The proof of Theorem \ref{thm:step1} is similar to \cite[Theorem 1]{Approximation2019}. For the completeness of this work, we put it in App. \ref{app:thm1}.
Different from the classical case, $|\bg'|$ does not equal $|\bg|$ in the inelastic model, therefore, $\ga_{\kappa}^{\bj}$ could not be further simplified as in \cite{Approximation2019}. But for the special collision kernel such as the VHS kernel, \eqref{eq:gamma} could be calculated exactly, which will be discussed in the next section.
\subsection{Simplification of VHS model}
\label{sec:VHS}
For the VHS kernel \eqref{eq:VHS}, which does not depend on the collision parameter $\sigma$, the coefficient \eqref{eq:gamma} could be calculated exactly. We will begin from two lemmas below as
\begin{lemma}
\label{thm:sp_int}
Assuming $\sigma = (\sigma_1, \sigma_2, \sigma_3)$ is a unit vector. $\kappa = (\kappa_1, \kappa_2, \kappa_3) \in \bbN^3$. Then the integral in the unit sphere holds
\begin{equation}
\label{eq:int_sigma}
\mathcal{S}(\kappa) \triangleq \int_{S^2}\si_1^{\kappa_1}\si_2^{\kappa_2}\si_3^{\kappa_3}\rd \sigma=
\left\{
\begin{array}{ll}
4\pi\frac{(\kappa-1)!!}{(|\kappa|+1)!!}, & 2| \kappa, \\
0, & \text{otherwise},
\end{array}
\right.
\end{equation}
where $(-1)!!$ is regarded as $1$, $\kappa !! = \kappa_1 !! \kappa_2!!\kappa_3 !!$, and $2|\kappa$ means all the components of $\kappa$ are even.
\end{lemma}
{\renewcommand{\proofname}{Proof of Lemma \ref{thm:sp_int}}
\begin{proof}
The proof can be completed with a spherical coordinate transform.
\end{proof}
}
\begin{lemma}
\label{thm:int_Her}
For the Hermite polynomial $H_{\alpha}(\bv)$ and the weight function $\omega(\bv)$ defined in \eqref{eq:short}, it holds that
\begin{equation}
\label{eq:int_poly}
\begin{split}
\mathcal{V}(\kappa, \alpha, \mu) &\triangleq \int_{\mR^3} \bv^{\kappa} \Ha(\bv)|\bv|^{\mu}\omega(\bv)\rd \bv\\
&=(2\pi)^{-\frac32}\sum_{\sk{\bj\pq\alpha \\ 2|(\alpha-\bj)}}
\mathcal{C}(\alpha, \bj)2^{\frac{1+\mu+|\bj|+|\kappa|}2} \Gamma\left(\frac{3+\mu+|\bj|+|\kappa|}2\right)\mathcal{S}(\bj+\kappa),
\end{split}
\end{equation}
where $\alpha, \kappa \in \bbN^3$ and $\mu\in \bbR^+$. $\Gamma(\cdot)$ is the Gamma function and
\begin{equation}
\label{eq:coe_mC}
\mathcal{C}(\alpha, \beta) = \mathcal{C}(\alpha_1, \beta_1)\mathcal{C}(\alpha_2, \beta_2)\mathcal{C}(\alpha_3, \beta_3)
\end{equation}
with $\mathcal{C}(n,k), n, k\in\bbN$ the coefficient of $x^k$ in $H_n(x)$.
\end{lemma}
The proof of Lemma \ref{thm:int_Her} is listed in App. \ref{app:VHS}. Defining the coefficients $D(\alpha, \beta, \mu)$ and $\psi(\alpha, \beta, \mu)$ as
\begin{align}
\label{eq:coe_D}
&D(\alpha, \beta, \mu)=\int_{\mR^3}\int_{S^2}\Ha(\bg')H_{\beta}(\bg)|\bg|^{\mu}\omega(\bg)\rd \sigma\rd \bg, \\
\label{eq:coe_psi}
&\psi(\alpha, \beta, \mu)=\int_{\mR^3}\int_{S^2}\Ha(\bg)H_{\beta}(\bg)|\bg|^{\mu}\omega(\bg)\rd \sigma\rd \bg,
\end{align}
then the following theorem holds for the VHS model as
\begin{theorem}
For the VHS kernel $B(|\bg|, \sigma) = C |\bg|^{2(1-\varpi)}$, \eqref{eq:gamma} could be calculated exactly as
\label{thm:VHS}
\begin{equation}
\label{eq:cal_gamma}
\gamma_{\kappa}^{\bj} = C 2^{\frac{5}{2}-\varpi}\left[D\Big(\bj, \kappa, 2(1-\varpi)\Big)-\psi\Big(\bj, \kappa, 2(1-\varpi)\Big)\right].
\end{equation}
\end{theorem}
\begin{proof}[Proof of Theorem \ref{thm:VHS}]
Applying the change of variables $\bg\rightarrow \sqrt2 \bg$, \eqref{eq:gamma} can be simplified into \eqref{eq:cal_gamma} with \eqref{eq:coe_D} and \eqref{eq:coe_psi}.
\end{proof}
For now, the point left is how to calculate $D(\alpha, \beta, \mu)$ and $\psi(\alpha, \beta, \mu)$ in \eqref{eq:coe_D} and \eqref{eq:coe_psi}, which is proposed in the Proposition below, and the proof is listed in App. \ref{app:VHS}.
\begin{proposition}
The coefficients $D(\alpha, \beta, \mu)$, and $\psi(\alpha, \beta, \mu)$ in \eqref{eq:coe_D} and \eqref{eq:coe_psi} can be calculated exactly as
\label{thm:coe_D_psi}
\begin{equation}
\label{eq:lemma_D}
\begin{split}
D(\alpha, \beta, \mu)
=\sum_{\sk{\lambda\pq\alpha \\ 2|(\alpha-\lambda)}}\mathcal{C}(\al,\la)
\sum_{\sk{\kappa\pq\la \\ 2|(\la-\kappa)}} C_{\lambda}^{\kappa}
\left(\frac{1-e}{2}\right)^{|\kappa|}\left(\frac{1+e}{2}\right)
^{|\la|-|\kappa|} \mathcal{S}(\lambda-\kappa) \mathcal{V}(\kappa, \beta, |\lambda|-|\kappa|+\mu),
\end{split}
\end{equation}
where
\begin{equation}
\label{eq:com_num}
C_{\lambda}^{\kappa} = C_{\la_1}^{\ka_1}C_{\la_2}^{\ka_2}C_{\la_3}^{\ka_3}, \qquad C_{m}^n = \frac{m!}{n!(m-n)!},
\end{equation}
and
\begin{equation}
\label{eq:lemma_psi}
\psi(\alpha, \beta, \mu)=4\pi \sum_{\sk{\lambda\pq\alpha \\ 2|(\alpha-\lambda)}}\mathcal{C}(\al,\la)\mathcal{V}(\lambda, \beta, \mu).
\end{equation}
\end{proposition}
For Maxwell molecules ($\varpi=1$), since the collision kernel $B$ does not depend on $\bg$, there are special sparsity for the coefficient $A\alk$. Thus, we have the following proposition of $\Aalk$, and the proof is listed in App. \ref{app:VHS}.
\begin{proposition}
\label{thm:Maxwell}
For the Maxwell molecules, it holds for the coefficients $A\alk$ that
$A\alk=0$ when $|\alpha|<|\lambda|+|\kappa|$.
\end{proposition}
\begin{table}[ht]
\centering
\def1.5{1.5}
{\footnotesize
\begin{tabular}{llll}
\hline
Coefficients & Formula & Used in & Computational cost \\
\hline
$\mathcal{S}(\kappa)$ & \eqref{eq:int_sigma} & \eqref{eq:int_poly}, \eqref{eq:lemma_D} & $\mO(M^3)$ \\
$\mathcal{V}(\kappa, \alpha, \mu)$ & \eqref{eq:int_poly} & \eqref{eq:lemma_D}, \eqref{eq:lemma_psi} & $\mO(M^{10})$ \\
$D(\alpha, \beta, \mu)$ & \eqref{eq:coe_D} & \eqref{eq:cal_gamma} & $\mO(M^{9})$ \\
$\psi(\alpha, \beta, \mu)$ & \eqref{eq:coe_psi} & \eqref{eq:cal_gamma} & $\mO(M^{9})$ \\
$\ga_{\kappa}^{\bj}$ & \eqref{eq:gamma} & {\eqref{eq:theorem_A}} & $\mO(M^{6})$ \\
$\ca{d}$ & \eqref{eq:ca} & {\eqref{eq:theorem_A}} & $\mO(M^4)$ \\
$A\alk$ & \eqref{eq:theorem_A} & {\eqref{eq:simp1_Q}} & $\mO(M^{12})$ \\
\hline
\end{tabular}
}
\caption{The formula and computational cost to obtain $A\alk$ in the VHS model and related coefficients.}
\label{table:cost1}
\end{table}
Consequently, the eight-dimensional integral in \eqref{eq:Aalk} is reduced into merely a series of summations for the VHS model. The computational cost for all related coefficients is listed in Tab. \ref{table:cost1}. We can find that the computational cost for all $\Aalk$ is $\mO(M^{12})$, but it will not cause much trouble since $\Aalk$ can be pre-computed offline and stored for the simulation. If the expansion center $[\ou, \oT]$ is chosen different from $[\boldsymbol{0}, 1]$, the coefficient $\Aalk^{[\ou, \oT]}$ can be derived from $\Aalk$ through a linear transform which is stated in the proposition below.
\begin{proposition}
\label{thm:change_center}
In the calculation of $\Aalk^{[\ou, \oT]}$, we emphasize that the expansion center $[\boldsymbol{0}, 1]$ is not unique. Under arbitrary expansion center $[\ou,\oT]$, the coefficients $\Aalk^{[\ou, \oT]}$ satisfies that
\begin{equation}
\label{eq:VHS_A}
\Aalk^{[\ou,\oT]}=\oT^{1-\varpi}\Aalk.
\end{equation}
We refer the readers to \cite[Sec. 3.1]{multi2022} for more details. Hence, we only need to compute and store $\Aalk$ with expansion center $[\boldsymbol{0},1]$.
\end{proposition}
Nevertheless, it is still quite expensive to solve the inelastic Boltzmann equation using Hermite spectral method. For example, the memory required to store $\Aalk$ is $\mO(M^9)$ \cite{Approximation2019}, which is too large in practice. Moreover, the computational cost for each collision term is also $\mO(M^9)$, and it will become much larger in spatially inhomogeneous tests, which is unacceptable for a large $M$. Thus, following the strategy in \cite{Approximation2019, ZhichengHu2019, multi2022}, a special design on the numerical algorithm is adopted to reduce the numerical cost, which will be introduced in the next section.
\section{Conclusion}
\label{sec:conclusion}
In this paper, we have developed a numerical scheme of the inelastic Boltzmann equation based on the Hermite spectral method, which makes us even capable to compute two-dimensional periodic model problems. The expansion coefficients of the quadratic collision model are computed utilizing the properties of the Hermite basis functions, which can be calculated exactly for the VHS model. To balance the accuracy and the computational cost, a new collision model is built with a combination of the quadratic collision term and a linearized collision operator. Finally, several benchmark problems in the granular flow are implemented to validate the numerical method. Even for two-dimensional cases, this method can capture the behavior of inelastic gas flow well with high efficiency.
\section*{Acknowledgements}
We thank Prof. Jingwei Hu from UW for her valuable suggestions. We thank Prof. Lei Wu from SUSTC for his help with the DSMC code.
The work of Yanli Wang is partially supported by the National Natural Science Foundation of China (Grant No. 12171026, U2230402 and 12031013).
\section{Numerical experiments}
\label{sec:experiment}
In this section, several numerical experiments are carried out to validate this Hermite spectral method for the inelastic Boltzmann equation. Firstly, two homogeneous cases will be tested for Maxwell molecules and hard sphere (HS) collision kernel. Next, two spatially one-dimensional problems and a spatially two-dimensional problem will be studied with the HS collision kernel.
\subsection{Homogeneous experiments}
\label{eq:sec:homo_exp}
In this section, two homogeneous problems, the heating source problem, and Haff's cooling law are studied, where the Maxwell molecular model is adopted in the heating source problem, and the HS model is utilized in Haff's cooling law.
\subsubsection{Heating source problem}
\label{sec:heat}
In this section, the heating source problem is studied, which is first introduced in \cite{Noije1998}, and similar studies can also be found in \cite{Hu2016, Wu2015}. Its governing equation is
\begin{equation}
\label{eq:HeatBoltz}
\pd{f}{t}-\varepsilon\Delta_v f=\frac{1}{{\rm Kn}}\mQ[f,f](\bv),
\end{equation}
where the second term represents the effect of the heating source with
the diffusion coefficient $\varepsilon \ll 1$. In this test, the Maxwell molecular model is utilized, and the Knudsen number is chosen such that $\frac{1}{{\rm Kn}}B = \frac{1}{4\pi}$. First, taking $\phi = 1, \bv$ in \eqref{eq:weak_1}, the conservation of mass and momentum can be derived as
\begin{equation}
\label{eq:heating_con}
\rho \equiv \rho_0, \qquad \bu = \bu_0,
\end{equation}
where $\rho_0$ and $\bu_0$ are the initial density and macroscopic velocity, respectively.
Without loss of generality, we let $\rho_0 \equiv 1$ and $\bu_0 \equiv \boldsymbol{0}$. Moreover, multiplying $\phi = |\bv|^2/3$ on both sides of \eqref{eq:HeatBoltz} and using the weak form in \eqref{eq:weak_1}, we can derive the governing equation of the temperature $\theta$ as \cite{Hu2016}
\begin{equation}
\label{eq:temp_evolve}
\begin{split}
\pd{\theta}{t}-2\varepsilon \rho_0
&=-\frac{1-e^2}{4}\theta,
\end{split}
\end{equation}
whose exact solution is
\begin{equation}
\label{eq:solu_temp}
\theta(t)=\left(\theta(0)-\frac{8\varepsilon}{1-e^2}\right)\exp\left(-\frac{1-e^2}{4}t
\right)+\frac{8\varepsilon}{1-e^2}.
\end{equation}
In the numerical simulation, the expansion center is selected as $[\ou,\bT]=[\boldsymbol{0},1]$, then the moment system of \eqref{eq:HeatBoltz} is derived as
\begin{equation}
\label{eq:HeatMoment}
\frac{\rd f_{\alpha}}{\rd t}-\varepsilon\sum_{d=1}^3f_{\alpha-2e_d}=Q_{\rm new, \alpha}, \qquad |\alpha| \leqslant M,
\end{equation}
where $Q_{\rm new, \alpha}$ is obtained in \eqref{eq:coe_new_coll}.
\begin{figure}[!hptb]
\centering
\subfloat[$e=0.2$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Heatsource/e=0.2.png}}\hfill
\subfloat[$e=0.5$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Heatsource/e=0.5.png}}\hfill
\subfloat[$e=0.8$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Heatsource/e=0.8.png}}\hfill
\caption{(Heating Problem in Sec. \ref{sec:heat}) Evolution of temperature in the heating source problem. The restitution coefficients are $e = 0.2, 0.5$ and $0.8$, successively. The blue line is the numerical solution, while the red line is the exact solution in \eqref{eq:solu_temp}.}
\label{fig:Heat}
\end{figure}
Since there is no analytical solution to this heating source problem, the solution of temperature always serves as the reference solution in this numerical test. In the simulation, the parameter $\epsilon$ is set as $\epsilon = 0.01$, and the initial condition is
\begin{equation}
\label{eq:ini_ex1}
f(0, \bv)= \mM_{0, 1}(\bv) = \frac{1}{(2\pi)^{3/2}} \exp\left(-\frac{|\bv|^2}{2}\right).
\end{equation}
Thus, the initial condition for temperature $\theta$ is $\theta(0) = 1$. Moreover, the length of the quadratic collision and the total expansion order are chosen as $[M_0, M] = [10, 10]$. The restitution coefficient $e = 0.2, 0.5, 0.8$ from $t = 0$ to $t = 5$ are tested, and the time step length is $\Delta t = 0.01$. The evolution of temperature $\theta$ is displayed in Fig. \ref{fig:Heat}, which implies that the numerical solution matches well with the analytical solution for the temperature.
\begin{remark}
Here, we want to emphasize that as is shown in Proposition \ref{thm:Maxwell}, the coefficients $A\alk$ can be nonzero only when
$|\alpha|\geqslant |\kappa|+|\lambda|$. Thus, \eqref{eq:HeatMoment} is naturally closed for any $M_0$ and $M$. That is to say, the higher-order moments will not affect the lower-order ones through the collision term, and $\theta$ can be represented by $M = 2$. Therefore, we can obtain the evolution of temperature $\theta$ through the moment system \eqref{eq:HeatMoment} with $[M_0, M] = [2, 2]$, and the computational cost can be greatly reduced.
\end{remark}
\subsubsection{Haff's cooling law}
\label{sec:haff}
In this section, we will try to observe Haff's cooling law numerically, which is first proposed by Haff in \cite{Haff1983}. The governing equation for Haff's cooling law is the same as the heating source problem as \eqref{eq:HeatBoltz} with $\epsilon = 0$. Haff's law implies that for the gas composed of inelastic hard spheres, the temperature in the spatially homogeneous problem will evolve as
\begin{equation}
\label{eq:Haff}
\theta(t)\approx \frac{\theta(0)}{(1+\gamma_0 t)^2}.
\end{equation}
Different from the Maxwell molecules, the decay speed here is $\mO(t^{-2})$. Here $\gamma_0$ is a positive constant depending on the value of $e$. We refer \cite{Hu2019, Filbet2013} for more details of this numerical test.
The nondimensionalized HS collision model has the form
\begin{equation}
\label{eq:HS}
B = \frac{1}{4\sqrt{2}\pi} |\bg|.
\end{equation}
Meanwhile, the same initial condition as \eqref{eq:ini_ex1} is adopted, and the Knudsen number is set to be ${\rm Kn}=1/sqrt{2}$. The length of the quadratic collision and the total expansion order are chosen as $[M_0, M] = [10, 40]$, and the time step length is $\Delta t = 0.01$. The evolution of the temperature with $e = 0.2, 0.5, 0.8$ from $t = 0$ to $t = 5$ is shown in Fig. \ref{fig:Haff}, where the reference solution is derived by estimating $\gamma_0$ in \eqref{eq:Haff} with the least square fitting. From Fig. \ref{fig:Haff}, it can be clearly seen that even in the case $e = 0.2$, the numerical solution matches well with the reference solution. Besides, we want to emphasize that Proposition \ref{thm:Maxwell} does not hold for the HS collision model, therefore the expansion number can not be set as $[2, 2]$ to capture the evolution of temperature.
\begin{figure}[!hptb]
\centering
\subfloat[$e=0.2$, $\gamma_0=0.364$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Haff/e=0.2.png}}\hfill
\subfloat[$e=0.5$, $\gamma_0=0.283$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Haff/e=0.5.png}}\hfill
\subfloat[$e=0.8$, $\gamma_0=0.135$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Haff/e=0.8.png}}\hfill
\caption{(Haff's cooling law in Sec. \ref{sec:haff})
Time evolution of temperature in Haff's cooling law. The restitution coefficients are $e = 0.2, 0.5$ and $0.8$, successively. The blue line is the numerical solution, and the red line is the reference solution.}
\label{fig:Haff}
\end{figure}
\subsection{Inhomogeneous experiments}
\label{sec:inhomo}
In this section, two spatially one-dimensional problems, the Couette flow and Fourier heat transfer, and one spatially two-dimensional problem the periodic diffusion are studied. In all the tests, Argon is taken as the working gas, and the HS model is utilized as the collision model.
The Knudsen number is calculated with
\begin{equation}
\label{eq:express_Kn}
{\rm Kn}=\frac{m_0}{\sqrt{2}\pi \rho_0 d_{\rm ref}^2 x_0}.
\end{equation}
Here, the parameters in \eqref{eq:express_Kn} are the nondimensionalization parameters of Argon and HS collision kernel, which are shown in Tab. \ref{table:character}, with the method of nondimensionalization listed in App. \ref{app:nondim}. The DSMC method provided in \cite{Astillero2005} for the HS collision kernel is utilized to provide the reference solution.
\subsubsection{Couette flow}
\label{sec:couette}
\begin{figure}[!hptb]
\centering
\subfloat[Density, $\rho$ (kg$\cdot$ m$^{-3}$)]
{\includegraphics[width=0.45\textwidth, height=0.36\textwidth,
clip]{image/Couette/case3/Density.png}}\hfill
\subfloat[$y$-component velocity, $u_2$ (m/s)]
{\includegraphics[width=0.45\textwidth, height=0.36\textwidth,
clip]{image/Couette/case3/Velocity.png}}\\
\subfloat[Temperature, $\theta$ (K)]
{\includegraphics[width=0.45\textwidth, height=0.36\textwidth,
clip]{image/Couette/case3/Temperature.png}}\hfill
\subfloat[Heat flux, $q_1$ (kg $\cdot$ s$^{-3}$) ]
{\includegraphics[width=0.45\textwidth, height=0.36\textwidth,
clip]{image/Couette/case3/Heatflux.png}}\hfill
\caption{(Couette flow in Sec. \ref{sec:couette}): Numerical solution of the Couette flow for ${\rm Kn}=0.2$
with $e = 1, 0.95$ and $0.9$. Lines correspond to numerical solutions and symbols denote the reference solutions from DSMC.}
\label{fig:Couette1}
\end{figure}
In this section, we will consider the 1D Couette flow, which is a benchmark problem also tested in \cite{Wu2015, ZhichengHu2019}. There are two infinite parallel plates with a distance of $10^{-3}$m. Both plates are purely diffusive and have a temperature of $273$K. They move in the opposite direction along with the plate with the speed $\bu^w= (0, \mp50, 0)$ m/s. Argon is filled in these two plates with velocity $\bu = \boldsymbol{0}$m/s, and temperature $\theta = 273$K. Two densities are considered as $\rho = 5.662\times 10^{-4}$kg$\cdot$m$^{-3}$ and $1.132 \times 10^{-4}$kg$\cdot$m$^{-3}$, which correspond to ${\rm Kn} = 0.2$ and $1$ respectively.
\begin{figure}[!hptb]
\centering
\subfloat[Density, $\rho$ (kg$\cdot$ m$^{-3}$)]
{\includegraphics[width=0.45\textwidth, height=0.36\textwidth,
clip]{image/Couette/case2/Density.png}}\hfill
\subfloat[$y$-component velocity, $u_2$ (m/s)]
{\includegraphics[width=0.45\textwidth, height=0.36\textwidth,
clip]{image/Couette/case2/Velocity.png}}\\
\subfloat[Temperature, $\theta$ (K)]
{\includegraphics[width=0.45\textwidth, height=0.36\textwidth,
clip]{image/Couette/case2/Temperature.png}}\hfill
\subfloat[Heat flux, $q_1$ (kg $\cdot$ s$^{-3}$) ]
{\includegraphics[width=0.45\textwidth, height=0.36\textwidth,
clip]{image/Couette/case2/Heatflux.png}}\hfill
\caption{(Couette flow in Sec. \ref{sec:couette}) Numerical solution of the Couette flow for ${\rm Kn}=1.0$
with $e = 1, 0.95$ and $0.9$. Lines correspond to numerical solutions and symbols denote the reference solutions from DSMC.}
\label{fig:Couette2}
\end{figure}
In the simulation, a uniform grid with $50$ cells and WENO reconstruction is utilized for the spatial discretization and the CFL condition number is set as ${\rm CFL} = 0.3$. The length for the quadratic collision term and the total expansion number are chosen as $[M_0, M] =[10, 40]$. The expansion center $[\ou, \oT]$ in the convection step is $[\boldsymbol{0}, 1]$ and the restitution coefficients $e=1, 0.95$, and $0.9$ are implemented. The density $\rho$, macroscopic velocity in the $y$-direction $u_2$, temperature $\theta$, and the heat flux $q_1$ in the $x$-direction at the steady state are studied, and the numerical results for ${\rm Kn} = 0.2$ and $1$ are illustrated in Fig. \ref{fig:Couette1} and \ref{fig:Couette2}, respectively. When ${\rm Kn} = 0.2$, all the numerical solutions coincide well with the reference solution. For the case ${\rm Kn} = 1$, the velocity $u_2$, temperature $\theta$ and heat flux $q_1$ agree well with the reference solution, while there is a small discrepancy in the density $\rho$, but the largest relative error is less than $1\%$. We also want to emphasize that there are some oscillations in the reference results, while the numerical solution keeps smooth.
\subsubsection{Fourier heat transfer}
\label{sec:fourier}
\begin{figure}[!hptb]
\centering
\subfloat[Density, $\rho$ (kg$\cdot$ m$^{-3}$)]
{\includegraphics[width=0.45\textwidth, height=0.36\textwidth,
clip]{image/Fourier/case3/Density.png}}\hfill
\subfloat[Temperature, $\theta$ (K)]
{\includegraphics[width=0.45\textwidth, height=0.36\textwidth,
clip]{image/Fourier/case3/Temperature.png}}\\
\subfloat[Stress tensor, $\sigma_{11}$ (kg$\cdot$ m$^{-1}$ $\cdot$ s$^{-2}$)]
{\includegraphics[width=0.45\textwidth, height=0.36\textwidth,
clip]{image/Fourier/case3/Stress.png}}\hfill
\subfloat[Heat flux, $q_1$ (kg $\cdot$ s$^{-3}$) ]
{\includegraphics[width=0.45\textwidth, height=0.36\textwidth,
clip]{image/Fourier/case3/Heatflux.png}}\hfill
\caption{(Fourier heat transfer in Sec. \ref{sec:fourier}) Numerical solutions of the Fourier flow for ${\rm Kn}=0.2$ with $e = 1, 0.95$ and $0.9$. Lines correspond to numerical solutions and symbols denote the reference solutions from DSMC.}
\label{fig:Fourier1}
\end{figure}
Fourier heat transfer is another widely studied problem in kinetic theory, which is also considered in \cite{Wu2015}. Similar to the Couette flow, we consider the gas of Argon between two infinitely large parallel plates. The distance between the plates is still $10^{-3}$m and both boundaries are purely diffusive. Different from Couette flow, the two plates are stationary but have a distinction in temperature. Here, the temperatures of two walls are set as $\theta_l=223$K and $\theta_r=323$K. The initial condition for Argon is velocity $\bu = \boldsymbol{0}$m/s, and temperature $\theta = 273$K. Two densities are considered as $\rho = 5.662\times 10^{-4}$kg$\cdot$m$^{-3}$ and $1.132 \times 10^{-4}$kg$\cdot$m$^{-3}$, which correspond to ${\rm Kn} = 0.2$ and $1$ respectively.
The same numerical setting such as grid, CFL number, expansion center, etc. as the Couette flow \ref{sec:couette} is utilized. The numerical results for ${\rm Kn} = 0.2$ and $1$ are plotted in Fig. \ref{fig:Fourier1}, and \ref{fig:Fourier2}, respectively, where the density $\rho$, temperature $\theta$, the shear stress $\sigma_{11}$, and heat flux $q_1$ are illustrated. We can find for both ${\rm Kn}$, the numerical solution match well with the reference, and the largest relative deviation is less than $0.5\%$ in all cases. Besides, different from the reference solution by DSMC, the numerical results keep smooth.
\begin{figure}[!hptb]
\centering
\subfloat[Density, $\rho$ (kg$\cdot$ m$^{-3}$)]
{\includegraphics[width=0.45\textwidth, height=0.36\textwidth,
clip]{image/Fourier/case2/Density.png}}\hfill
\subfloat[Temperature, $\theta$ (K)]
{\includegraphics[width=0.45\textwidth, height=0.36\textwidth,
clip]{image/Fourier/case2/Temperature.png}}\\
\subfloat[Stress tensor, $\sigma_{11}$ (kg$\cdot$ m$^{-1}$ $\cdot$ s$^{-2}$)]
{\includegraphics[width=0.45\textwidth, height=0.36\textwidth,
clip]{image/Fourier/case2/Stress.png}}\hfill
\subfloat[Heat flux, $q_1$ (kg $\cdot$ s$^{-3}$) ]
{\includegraphics[width=0.45\textwidth, height=0.36\textwidth,
clip]{image/Fourier/case2/Heatflux.png}}\hfill
\caption{(Fourier flow in Sec. \ref{sec:fourier}) Numerical solutions of the Fourier flow for ${\rm Kn}=1.0$ with $e = 1, 0.95$ and $0.9$. Lines correspond to numerical solutions and symbols denote the reference solutions from DSMC.}
\label{fig:Fourier2}
\end{figure}
\subsubsection{2D case: periodic diffusion}
\label{sec:diffuse}
\begin{figure}[!hptb]
\centering
\subfloat[$\rho$ (kg$\cdot$ m$^{-3}$), $t=0.05$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Diffuse/0.9/Density_t=0.05.png}}\hfill
\subfloat[$\theta$ (K), $t=0.05$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Diffuse/0.9/Temperature_t=0.05.png}}\hfill
\subfloat[$\sigma_{12}$ (kg$\cdot$ m$^{-1}$ $\cdot$ s$^{-2}$), $t=0.05$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Diffuse/0.9/Stress_t=0.05.png}}\\
\subfloat[$\rho$ (kg$\cdot$ m$^{-3}$), $t=0.1$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Diffuse/0.9/Density_t=0.1.png}}\hfill
\subfloat[$\theta$ (K), $t=0.1$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Diffuse/0.9/Temperature_t=0.1.png}}\hfill
\subfloat[$\sigma_{12}$ (kg$\cdot$ m$^{-1}$ $\cdot$ s$^{-2}$), $t=0.1$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Diffuse/0.9/Stress_t=0.1.png}}\hfill
\caption{(2D case: periodic diffusion in Sec. \ref{sec:diffuse}, Example 1) Solution of the periodic diffuse for $e=0.9$ with the initial condition \eqref{eq:ini2D}. Blue contours: Numerical solution. Red contours: Reference solution by DSMC.}
\label{fig:Diffuse1}
\end{figure}
In this section, we consider a two-dimensional test with periodic boundary conditions. Argon is also taken as the working gas in the square region $\Omega=[0,L]\times [0,L]$, where $L=10^{-3}$m. To validate the efficiency of this Hermite method, two examples with different initial conditions are tested.
\paragraph{Example 1}
In the first example, the initial velocity and temperature are set as $\bu=\boldsymbol{0}$ and $\theta=273K$ in the whole area, while the initial density $\rho(x,y)$ is
{\small
\begin{equation}
\label{eq:ini2D}
\rho(x,y)=1.132\times 10^{-4}\times \left[1+0.5\sin\left(2\pi\frac{x}{L}\right)\sin\left(2\pi\frac{y}{L}\right)\right] \text{kg}\cdot\text{m}^{-3}.
\end{equation}
}
In the classical case, the distribution function will diffuse to reach global equilibrium. From the macroscopic view, the macroscopic variables will finally be spatially uniform. However, in the inelastic case, due to the dissipation of total energy, the evolution of non-equilibrium macroscopic variables, especially temperature, will become much more complicated, and will approach zero with time increasing.
In the simulation, a uniform grid with $100\times 100$ cells and WENO reconstruction is adopted for spatial discretization. With this initial condition, the corresponding Knudsen number is ${\rm Kn} = 1$. The quadratic length and total expansion number are set as $[M_0, M] = [10, 30]$. The expansion center for the convection step is $[\ou, \bT] = [\boldsymbol{0}, 1]$.
The restitution coefficient $e=0.9$ is implemented in the simulation. The macroscopic variables such as the density $\rho$, the temperature $\theta$, and the stress tensor $\sigma_{12}$ are studied. The numerical results are illustrated in Fig. \ref{fig:Diffuse1}, where the solution by DSMC is utilized as the reference solution. We can see that the numerical solution coincides well with the reference solution.
Moreover, there are some oscillations in the reference solution, while the numerical solution keeps smooth.
\paragraph{Example 2}
To further validate the efficiency of this method, a more complicated initial condition is studied, where the variation period for the density is smaller, and some disturbance is added in the temperature as
{\small
\begin{equation}
\label{eq:ini2D2}
\begin{split}
&\rho(x,y)=1.132\times 10^{-4}\times \left[1+0.5\sin\left(2\pi\frac{x}{L}\right)\sin\left(4\pi\frac{y}{L}\right)\right] \text{kg}\cdot\text{m}^{-3}, \\
&\theta(x,y)=273\times \left[1+0.05\sin\left(2\pi\frac{x}{L}\right)\sin\left(2\pi\frac{y}{L}\right)\right] \text{K}.
\end{split}
\end{equation}
}
The numerical settings here such as the mesh, expansion order, et al. are the same as \textbf{ Example 1}, but a smaller restitution coefficient $e = 0.8$ is studied. The numerical solution of the density $\rho$, the temperature $\theta$, and the stress tensor $\sigma_{12}$ at $t = 0.05$ and $t = 0.1$ is plotted in Fig. \ref{fig:Diffuse2}. We can see that for this complex initial condition, the numerical solutions still agree well with the reference solution. The variation trend of these macroscopic variables is also similar to \textbf{ Example 1}, while the behavior of the temperature seems more complicated.
To explore a longer behavior of this example, the numerical solution at $t = 0.2$ is displayed in Fig. \ref{fig:Diffuse3}, where we find that these three macroscopic variables are becoming spatially uniform, and the temperature is globally decreasing. We want to emphasize that the reference solution by DSMC is filled with oscillations, which can not describe this behavior of the example with a longer time, while the numerical solution of the Hermite spectral method is still smooth.
\begin{figure}[!hptb]
\centering
\subfloat[$\rho$ (kg$\cdot$ m$^{-3}$), $t=0.05$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Diffuse/0.8/Density_t=0.05.png}}\hfill
\subfloat[$\theta$ (K), $t=0.05$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Diffuse/0.8/Temperature_t=0.05.png}}\hfill
\subfloat[$\sigma_{12}$ (kg$\cdot$ m$^{-1}$ $\cdot$ s$^{-2}$), $t=0.05$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Diffuse/0.8/Stress_t=0.05.png}}\\
\subfloat[$\rho$ (kg$\cdot$ m$^{-3}$), $t=0.1$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Diffuse/0.8/Density_t=0.1.png}}\hfill
\subfloat[$\theta$ (K), $t=0.1$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Diffuse/0.8/Temperature_t=0.1.png}}\hfill
\subfloat[$\sigma_{12}$ (kg$\cdot$ m$^{-1}$ $\cdot$ s$^{-2}$), $t=0.1$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Diffuse/0.8/Stress_t=0.1.png}}
\caption{(2D case: periodic diffusion in Sec. \ref{sec:diffuse}, Example 2) Solution of the periodic diffuse for $e=0.8$ with the initial condition \eqref{eq:ini2D2}. Blue contours: Numerical solution. Red contours: Reference solution by DSMC.}
\label{fig:Diffuse2}
\end{figure}
\begin{figure}[!hptb]
\centering
\subfloat[$\rho$ (kg$\cdot$ m$^{-3}$), $t=0.2$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Diffuse/0.8/Density_t=0.2.png}}\hfill
\subfloat[$\theta$ (K), $t=0.2$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Diffuse/0.8/Temperature_t=0.2.png}}\hfill
\subfloat[$\sigma_{12}$ (kg$\cdot$ m$^{-1}$ $\cdot$ s$^{-2}$), $t=0.2$]
{\includegraphics[width=0.3\textwidth, height=0.24\textwidth,
clip]{image/Diffuse/0.8/Stress_t=0.2.png}}
\caption{(2D case: periodic diffusion in Sec. \ref{sec:diffuse}, Example 2) Solution of the periodic diffuse for $e=0.8$ with the initial condition \eqref{eq:ini2D2} at $t = 0.2$. Blue contours: Numerical solution. Red contours: Reference solution by DSMC.}
\label{fig:Diffuse3}
\end{figure}
\paragraph{Efficiency test}
To show the efficiency of this method quantitatively, the computational time for these two examples is studied. The simulations are done on the CPU model Intel Xeon E5-2697A V4 @ 2.6GHz, and $8$ threads are utilized. The total CPU time and wall time, as well as the CPU time of each time step and each grid, are provided in Tab. \ref{table:diffuse_time}.
It shows that the total time for $e = 0.9$ end $e = 0.8$ is almost the same, which indicates that the effect of the restitution coefficient on the computational time is almost negligible. Moreover, the total CPU time is almost $8$ times of the elapsed time, which indicates the amazing parallel efficiency of this Hermite spectral method. Besides, the total degrees of freedom (DOF) in the microscopic velocity space is
\begin{equation}
\label{eq:dof}
{\rm DOF} = \frac{(M+2)(M+1)M}{6}.
\end{equation}
Then, the total DOF in this 2D problem is $4960$ and the CPU time per DOF per grid given in Tab. \ref{table:diffuse_time} is at $\mO(10^{-6})$. All these results reveal the high efficiency of this Hermite spectral method, and it will work well with parallel computing in large-scale problems.
\begin{table}[!hptb]
\centering
\def1.5{1.5}
{\footnotesize
\begin{tabular}{lll}
\hline
& Example 1 & Example 2 \\
\hline
$[M_0,M]$ & $[10,30]$ & $[10,30]$ \\
Restitution coefficient $e$ & 0.9 & 0.8 \\
End time $t$ & 0.1 & 0.1 \\
\hline
Run-time data: & & \\
Total CPU time $T_{\rm CPU}$ (s): & 79980 & 81454 \\
Elapsed time (Wall time) $T_{\rm Wall}$ (s): & 10575.8 & 10856.3\\
Parallel efficiency: & $94.53\%$ & $93.79\%$ \\
CPU time per time step (s): & 242.36 & 246.83 \\
Degree of freedom & $4960$ & $4960$ \\
CPU time per DOF per grid (s) & $4.9\times 10^{-6}$ & $5.0\times10^{-6}$ \\
\hline
\end{tabular}
}
\caption{(2D case: periodic diffusion in Sec. \ref{sec:diffuse}) Run-time data for two-dimensional periodic diffusion of $t=0.1$.}
\label{table:diffuse_time}
\end{table}
|
\section{Introduction} \label{sec:intro}
When autonomous machines act in their environment or humans interact with computers, the necessary input data is often streamed continuously. In some settings, the input streams can be easily transformed into control signals based on simple physical models. However, in more advanced scenarios, it is necessary to develop a more complex, data-driven model and use its predictions to control the machine. An important example of the latter scenario are brain-computer interfaces (BCIs), which record the user's brain activity and decode control signals based on the measured data.
Most statistical models require an input of fixed dimension and a common approach is to extract windows of a fixed size with a fixed step size from the continuous data stream. These \textit{sliding windows} are then used to predict the desired control signal with one classification per window. The advantage of this approach is that only the most recent data is used for predictions, but it comes at cost: the signal of interest might occur at any time point in the extracted window. Thus, any classifier of the sliding windows needs to be ``shift invariant'' in the sense that it detects the desired signal independently of its position in the window.
In the context of BCIs, more specifically for the analysis of data measured via electroencephalography (EEG), traditional methods are based on carefully selected features that are calculated from the data. Commonly applied techniques include the principal component analysis of EEG data in the time domain, or features based on the power spectrum in the frequency domain \citep{azlan2014, boubchir2017, boonyakitanont2020}. Due to recent advances in the field of deep learning, different architectures of neural networks were proposed that avoid a manual feature extraction and seem to outperform more traditional methods. For example the neural network \textit{EEGNet} was proposed to support multiple BCI paradigms and is often referred to as benchmark model in the field \cite{lawhern2018}. In a clinical setting, some variant of the VGG16 neural network was used to detect signals associated with epilepsy \cite{dasilva2021}. In general, deep learning has been applied successfully to a variety of tasks related to EEG data \citep{craik2019, roy2019}.
Inspired by their successes in computer vision and natural language processing, common neural networks used for the classification of EEG data are based on convolutions. Convolutional neural networks are often considered to be shift invariant and are therefore a good choice in the given context. However, they do not take the specific structure of EEG data into account. Similar to most physical processes, the electrical activity, that is recorded on the user's scalp by the EEG, can be considered as smooth \citep{ramsay2005}.
To reflect this additional structure, it is more appropriate to model the data as a (discretized) sample of an underlying smooth function.
The latter paradigm is the basis of \textit{functional data analysis} (FDA), a branch of statistics that received more and more attention throughout the last decades and remains an active area of research. Most concepts from multivariate statistics have been extended to functional data \citep{ramsay2005, kokoszka2017}. For example, many functional versions of the principal component analysis have been proposed in literature \citep{shang2014}. Different generalizations of the linear model to functional covariates and/or functional responses have been introduced \citep{cardot1999, cuevas2002}.
Finally, Portmanteau-type tests for detecting serial correlation have been proposed for functional time series \citep{gabrys2007, bucher2020}. This functional approach allows it to extract previously unavailable information from the data in form of derivatives of the continuous signal.
As methods from FDA take the functional structure of physical processes into account, they would be suitable classifiers. However, classic methods from FDA are in general not shift invariant and require the signal of interest to be at a fixed point in time. In some applications it is possible to ``register'' the functions through a suitable transformation of time \citep{sakoe1978, kneip1992, gasser1995, ramsay1998}. However, in the context of sliding windows curve registration is often not feasible and methodology that requires previous registration cannot be applied reliably.
In the present work, we propose a framework that combines the advantages of neural networks (particularly CNNs) and FDA: functional neural networks (FNNs). On one side, these networks are shift invariant, and on the other side, they are able to model the functional structure of their input. FNNs have several advantages over scalar-valued neural networks. They are independent of the sample frequency of the input data, as long as the input can be rescaled to a certain interval. Further, they allow to predict smooth outputs. And finally, they are more transparent to some extent due to smoothness constraints.
We summarize our contribution as follows:
\begin{itemize}
\item We propose extensions of fully-connected and convolutional layers to functional data.
\item We present architectures of functional neural networks based on these extensions.
\item We show that the proposed methodology works through a simulation study and real data experiments.
\end{itemize}
Whereas multi-layer perceptrons (MLPs) are not shift invariant, the introduced functional convolutional layers allow the construction of shift invariant functional convolutional neural networks. This makes FNNs helpful in any scenario where sliding windows based on a (possibly multivariate) continuous data stream are classified, and they can be employed in a variety of applications.
\section{Related Work} \label{sec:rel_work}
To the best of our knowledge, the combination of functional data and (convolutional) neural networks is only discussed in a handful of papers and the proposed methodology extends previous results. In early works MLPs with functional inputs and neurons, that transform the functional data to scalar values in the first layer, were introduced \citep{Rossi2002, rossi2004, rossi2005}. \cite{zhao2012} proposed an algorithm to train similar MLPs with inputs from a real Hilbert space. Subsequently, \cite{wang2019} proposed to use functional principal components for the transformation of the functional inputs to scalar values in the first layer. \cite{wang2020} added another layer based on functional principal components to transform the scalar-valued output of the MLP back to functional data in the last layer. More recently, fully functional neurons were proposed \citep{rao2021scalar, Rao2021}.
\section{Mathematical Preliminaries} \label{sec:math}
Let us assume, we observe $d$ quantities at $T$ time instants for $N\in\mathbb{N}$ individuals, providing us with matrices of observations
$$ \mathbf{X}^{(n)} = \left(\begin{array}{ccc}
X_{1,1}^{(n)} & \cdots & X_{1,T}^{(n)} \\
\vdots & \ddots & \vdots \\
X_{d, 1}^{(n)} & \cdots & X_{d,T}^{(n)}
\end{array}\right),
$$
for $n=1,\dots, N$, and jointly with $\mathbf{X}^{(n)}$ their corresponding ``labels'' $\mathbf{Y}^{(n)}$ which might be vectors in $\mathbb{R}^c$ or matrices
$$ \mathbf{Y}^{(n)} = \left(\begin{array}{ccc}
Y_{1,1}^{(n)} & \cdots & Y_{1,T}^{(n)} \\
\vdots & \ddots & \vdots \\
Y_{c, 1}^{(n)} & \cdots & Y_{c,T}^{(n)},
\end{array}\right)
$$
where $c$ denotes the number of quantities that we observe for $\mathbf{Y}^{(n)}$. Further assume the observed quantities to be noisy versions of an underlying smooth signal, i.\,e.,
\begin{equation}\label{loc-scale-model}
X_{i,t}^{(n)}=f_i^{(n)}\big(\tfrac tT \big) + {\varepsilon}_{i,t}^{(n)},
\end{equation}
for smooth functions $f_i^{(n)}$ and centered errors ${\varepsilon}_{i,t}^{(n)}$, for $i=1,\dots, d, t=1,\dots, T$ and $n=1,\dots, N$. Note that the degree of smoothness might vary for different applications, which leads to slight modifications in the model. This representation suggests the use of methods from functional data analysis, which consider the intrinsic structure of the data.
Throughout this work, we only require the functions $f_i^{(n)}$ to be square-integrable, i.\,e., $f_i^{(n)} \in L^2([0, 1])= \{ f:[0, 1] \to \mathbb{R} : \int_0^1 f(x)^2 \diff x < \infty \}$. Similarly, in case of matrix-valued labels $\mathbf{Y}^{(n)}$, we assume their entries to be discretized versions of some underlying functions $g_i^{(n)} \in L^2([0,1])$, more specifically $Y_{i,t}^{(n)} = g_i^{(n)}(t/T)$.
Our aim is to approximate the functional $F:\big(L^2([0,1])\big)^d \to \mathcal{Y}$, which maps an observation $\mathbf{X}$ to its corresponding label $\mathbf{Y}$, where $\mathcal{Y} = \mathbb{R}^c$ for vector-valued labels $\mathbf{Y}^{(n)}$ and $\mathcal{Y} = \big(L^2([0,1])\big)^c$ for matrix-valued labels. Formally, the functional $F$ corresponds to the conditional expectation $\mathbb{E}[\mathbf{Y}|\mathbf{X}]$.
In case of classification problems, each coordinate $F_i(\mathbf{X})$ of the functional $F$ can be interpreted as the probability of $\mathbf{X}$ belonging to class $i \in \{1,\dots, c\}$.
\subsection{Preprocessing}
Before putting observations into a neural network, it is often helpful to preprocess them by applying certain filters or normalization. In our case, we work with noisy functional data, observed at discrete time points. In functional data analysis, a common first step for this kind of data is \textit{smoothing}, which helps to reduce the errors and extends the observations from discrete time points to a continuous interval. Another preprocessing step frequently used for neural networks, is some form of normalization to ensure that the data is of a similar magnitude. We employ \textit{local linear estimation} for smoothing the data and \textit{standardization} for its normalization as described below.
\subsubsection{Smoothing}\label{sec:preprocessing}
In the literature, there exists a variety of smoothing procedures from Fourier series to expansions based on B-splines or wavelets \citep{ramsay2005}. We use local polynomial regression to estimate the functions $f_i^{(n)}$ and their first derivative(s) \citep{fan1996}.
For the sake of clarity, we omit some indices and rewrite \eqref{loc-scale-model} as $X_{t}=f\big(\tfrac tT \big) + {\varepsilon}_{t}$ for a moment. Then, if $f$ is $p+1$ times differentiable with bounded derivatives, we can define the local polynomial estimator as
\begin{align*}
&(\hat{f}(x), \widehat{f'}(x), \dots, \widehat{f^{(p)}}(x))\\ &= \argmin_{\beta_0, \dots, \beta_p} \sum_{t=1}^{T} \bigg(X_t - \sum_{j=0}^{p} \beta_j \Big(\frac{t}{T}-x\Big)^j\bigg)^2 K_h\big(\frac{t}{T}-x\big)
\end{align*}
to estimate $f$ and its first $p$ derivatives. Here $K$ denotes a kernel function, $h$ the bandwidth of the estimator and $K_h(\cdot)=K(\tfrac{\cdot}{h})$. In the following, we assume $K:\mathbb{R}\to\mathbb{R}$ to be a symmetric, twice differentiable function, supported on the interval $[-1, 1]$ and satisfying $\int_{[-1, 1]}K(x)\diff x = 1$.
From the above definition, explicit formulas can be derived for the estimators by calculating the derivatives of the right-hand side with respect to $\beta_j~(j=1,\dots, p)$ and a Taylor expansion. The result can be rewritten as a convolution of the signal and a filter depending on the kernel $K$ and the bandwidth $h$.
To simplify the notation, we will refer to the estimators of the functions $f_i^{(n)}$ and their derivatives as $h_{i, 1}^{(n)}$, thus, we obtain estimators $(h_{i, 1}^{(n)}, (h_{i, 1}^{(n)})', \dots, (h_{i, 1}^{(n)})^{(p)})$ for each $f_i^{(n)}$, $i=1,\dots, d, n=1,\dots, N$.
The choice of the bandwidth is crucial in order to obtain a good estimate of the underlying functions. If the bandwidth is chosen too small, the estimator will overfit the data, whereas a large bandwidth leads to over-smoothing \citep{silverman2018}. Oftentimes it is a good idea to use cross validation to select a bandwidth that minimizes a certain error measure, such as the mean squared error. Generally, the estimation of higher derivatives requires larger bandwidths than the estimation of the function itself.
\subsubsection{Normalization}
When neural networks are trained via some form of gradient descent, it is crucial to ensure that the input data is of a similar size, which is done through prior normalization. There are many different normalization methods and the most useful choice depends on the specific application. In the following, we will standardize the data by subtracting the mean and dividing by the standard deviation across a suitable range of the data. As we did not make any assumptions about the relation between the signals $f_i^{(n)}$ and $f_j^{(n)}$, we standardize each smoothed signal $h_{i, 1}^{(n)}$ (and its derivatives) separately, i.\,e., we calculate
$$h_{i, 2}^{(n)} = \frac{h_{i, 1}^{(n)} - \int_0^1 h_{i, 1}^{(n)}(x)\diff x}{\bigg(\int_0^1 \Big(h_{i, 1}^{(n)}(x) - \int_0^1 h_{i, 1}^{(n)}(y)\diff y\Big)^2 \diff x\bigg)^{1/2}}. $$
After this transformation, the signals are of a similar magnitude, for each observation $\mathbf{X}^{(n)}$.
\section{Functional Layers} \label{sec:dev1}
\subsection{Functional Multilayer Perceptrons}
\label{sec:f_mlp}
Once the data is smoothed and prepared to be analyzed as functional data, it is not clear how to design neural networks that take this additional structure into account. The simplest form of an artificial neural network with scalar input $(h_1, \dots, h_d)$ is the multilayer perceptron, that consists of $L$ layers with $J_1, J_2, \dots, J_L$ neurons each. The value at neuron $k$ in the $\ell$-th layer is then calculated as
\begin{equation*}
H_{(k)}^{(\ell)} = \sigma\bigg( b_{(k)}^{(\ell)} + \sum_{j=1}^{J_{\ell-1}} w_{(j,k)}^{(\ell)} H_{(j)}^{(\ell-1)}\bigg),
\end{equation*}
where $H_{(k)}^{(0)}=h_k$ denotes the network's input, $H_{(k)}^{(L)}$ its output, $b_{(k)}^{(\ell)}$ the $k$th neuron's bias and $w_{(j,k)}^{(\ell)}$ the weight between the $k$th neuron in layer $\ell$ and the $j$th neuron in layer $\ell-1$. The function $\sigma:\mathbb{R}\to\mathbb{R}$ is referred to as \textit{activation function} and enables the network to reflect non-linear dependencies.
When the input data is not scalar, but functional, \cite{Rao2021} propose to replace the scalar biases by functional biases and the weights between neurons by integral kernels, finally defining the neurons' values as
\begin{equation}\label{eq:neurons_rao}
H_{(k)}^{(\ell)}(s) = \sigma\bigg( b_{(k)}^{(\ell)}(s) + \sum_{j=1}^{J_{\ell-1}} \int w_{(j,k)}^{(\ell)}(s, t) H_{(j)}^{(\ell-1)}(t) \diff t \bigg).
\end{equation}
While this extension allows to model rather general relations between the model's input and its desired output, this flexibility makes the network's training difficult because we need to find optimal weight functions for any connection between two neurons.
Starting from the fully-connected multilayer perceptron, many advances in deep learning are due to more specific architectures, which reduce the number of model parameters to mitigate the curse of dimensionality. For instance, convolutional neural networks can be interpreted as multi-layer perceptrons, where most weights vanish and the remaining connections between neurons share a smaller set of weights.
In a similar fashion, we propose to simplify the neuron model in \eqref{eq:neurons_rao} by using simple weight functions $w_{(j,k)}^{(\ell)}:[0, 1]\to\mathbb{R}$ rather than integral kernels in $L^2([0,1]^2)$. This adaptation leads to neurons defined via
\begin{equation}\label{eq:neurons}
H_{(k)}^{(\ell)}(t) = \sigma\bigg( b_{(k)}^{(\ell)}(t) + \sum_{j=1}^{J_{\ell-1}} w_{(j,k)}^{(\ell)}(t) H_{(j)}^{(\ell-1)}(t) \bigg).
\end{equation}
The above defined neurons are fully functional in the sense that both their input and output are functions. If we try to predict scalar-valued labels in $\mathbb{R}^c$, we need to summarize the information contained in the functions. We propose to calculate the scalar product of the weights and their corresponding inputs, leading to
\begin{equation}\label{eq:neurons_last_layer}
H_{(k)}^{(\ell)} = \sigma\bigg( b_{(k)}^{(\ell)} + \sum_{j=1}^{J_{\ell-1}} \int w_{(j,k)}^{(\ell)}(t) H_{(j)}^{(\ell-1)}(t) \diff t \bigg).
\end{equation}
With this definition of a functional multilayer perceptron (F-MLP), we simplified the training and need to optimize functional weights of one variable. The theoretical framework to train the model through backpropagation based on Fréchet derivatives is provided by \citep{Rossi2002, Olver2016, Rao2021}.
The computation of Fréchet derivatives becomes tedious and computationally expensive. An efficient approach to simplify computations and simultaneously reduce the dimension of the weights' space, is to replace the weights $w_{(j,k)}^{(\ell)}(t)$ by linear combinations of a finite set of base functions. Therefore, let $\{\phi_i\}_{i=0}^q$ be a set of suitable functions, such as Legendre polynomials, wavelets or the first $q/2$ sine-cosine pairs of the Fourier basis, and consider the linear combination
\begin{equation}\label{eq:base_rep}
w_{(j,k)}^{(\ell)}(t) = \sum_{i=0}^{q} w_{(j,k)}^{(\ell, i)} \phi_i(t),
\end{equation}
for some scalar weights $w_{(j,k)}^{(\ell, i)}$. With this representation, the fully functional neural network can be described through scalar weights and we are able to use the standard scalar backpropagation.
\subsection{Functional Convolutional Neural Networks}
The functional MLP is particularly useful if the input functions are aligned (or can be aligned via a suitable transformation of time) and the signals of interest happen at the same time instants for all measurements. However, under the sliding window paradigm, for high-noise data such as speech or EEG signals, it is not possible (or at least not useful) to previously register the curves, as the signal of interest may occur at any arbitrary time instant. In this case, MLPs are impractical as they would require many parameters to model complex patterns
For scalar input, alternative network architectures have been developed that are shift invariant and therefore capable to detect certain signals independently of their position. One type of neural network that is considered as ``translation invariant'' are CNNs, which we can extend to functional data as well.
Similarly to \eqref{eq:neurons_rao}, we can define a functional convolutional layer by setting $w_{(j, k)}^{(\ell)}(s, t) = u_{(j, k)}^{(\ell)}(s - t)$ for some filter (or kernel) function $u_{(j, k)}^{(\ell)}:\mathbb{R}\to\mathbb{R}$ with support on $[-b, b]$ and bandwidth $b \in (0, 1)$, ultimately leading to
\begin{equation*}
H_{(k)}^{(\ell)}(s) = \sigma\bigg( b_{(k)}^{(\ell)}(s) + \sum_{j=1}^{J_{\ell-1}} \int u_{(j, k)}^{(\ell)}(s - t) H_{(j)}^{(\ell-1)}(t) \diff t \bigg),
\end{equation*}
where the functions $H_{(k)}^{(\ell)}$ are extended to the interval $[-b, 1 + b]$ by defining them as zero outside of the interval $[0, 1]$.
These functional convolutional layers are shift invariant in the sense that a filter, which is capable of detecting a certain signal, would detect it independently of its position in the interval $[0, 1]$. Once again, we reduce the dimension of the optimization problem by representing the filters as linear combinations of a set of base functions as in \eqref{eq:base_rep}.
\subsection{Architecture}\label{sec:architectures}
With functional versions of fully connected and convolutional layers at hand, we can define arbitrary architectures of functional neural networks (FNNs). Figure \ref{fig:architecture} displays FNNs with scalar and functional outputs, respectively. In both architectures, the first layer uses a local linear estimator to smooth the input and estimate derivatives of the smoothed signals, while the second layer standardizes the input across each signal. Following are two functional convolutional layers. For the FNN with scalar output, the last layer is a functional fully connected layer, while for the FNN with functional output, the last layer is a third functional convolutional layer.
\begin{figure}[ht]
\vskip 0.1in
\begin{center}
\centerline{\input{architectures}}
\caption{Left: Neural network architecture with functional input and scalar output. Right: Neural network architecture with functional input and output.}
\label{fig:architecture}
\end{center}
\vskip -0.2in
\end{figure}
\section{Empirical Results} \label{sec:emp_res}
We now show that the proposed methodology works with simulated and real data and compare it to benchmark models.
\subsection{Simulation Study}
\label{sec:sim_study}
\subsubsection*{Setup} Inspired by brain activity measured through electroencephalography (EEG), we generate two data
sets for the simulation study. In both cases, we simulate two-dimensional samples that belong to one of three classes.
For data set (I), we draw independently frequencies $\alpha_n\sim \mathcal{U}_{[8, 12]}, \beta_n\sim \mathcal{U}_{[13, 30]}$, time shifts $t_1^{(n)}, t_2^{(n)} \sim \mathcal{U}_{[0,1]}$, class labels $c_n\sim\mathcal{U}_{\{1, 2, 3\}}$ and standard normally distributed errors ${\varepsilon}_{i,t}^{(n)}\sim\mathcal{N}(0, 1)$. Based on these random quantities, we construct the continuous signals
\begin{align*}
f_i^{(n)}(x) = &(1-\gamma_i(c_n)) \cdot \sin(2\pi \alpha_n (x + t_i^{(n)})) \\&+ \gamma_i(c_n) \cdot \sin(2\pi \beta_n (x + t_i^{(n)}))
\end{align*}
with class dependent coefficients $\gamma(1) = (0, 0), \gamma(2) = (0.8, 0.4)$ and $\gamma(3) = (0.4, 0.8)$
and finally define the discretized, noisy samples $X_{i, t}^{(n)} = f_i^{(n)}(t/T) + {\varepsilon}_{i,t}^{(n)} $. Examples of each class are displayed in Figure \ref{fig:dataset1} of Appendix \ref{app:sim_data}.
For data set (II), we draw independently scaling factors $w_n\sim \mathcal{U}_{[0.05, 0.1]}$, time points $t_n\sim \mathcal{U}_{[0,1]}$, class labels $c_n\sim\mathcal{U}_{\{1, 2, 3\}}$ and standard normally distributed errors ${\varepsilon}_{i,t}^{(n)}\sim\mathcal{N}(0, 1)$. Based on the scaling factors $w_n$ and time points $t_n$, we construct continuous signals $$f^{(n)}(x) = \max\Big\{ - \tfrac{4}{w_n^2} (x-t_n)^2 + 3 , 0 \Big\},$$
which resemble spikes, as displayed in Figure \ref{fig:dataset2} of Appendix \ref{app:sim_data}. Again, we define discretized, noisy samples $X_{i, t}^{(n)} = \gamma_i(c) \cdot f^{(n)}(t/T) + {\varepsilon}_{i,t}^{(n)}$, where the class dependent coefficients are $\gamma(1) = (0, 0), \gamma(2) = (1, 0)$ and $\gamma(3) = (0, 1)$.
For both data sets, we vary the sample size $N \in \{1000, 2000, 3000, 4000, 5000\}$, while keeping $T=250$ fixed.
As baseline models, we use $k$-nearest neighbors (KNN) for a varying number of neighbors $k \in \{1, 2, ..., 19\}$. Before feeding the data into the model, we smooth it by applying a local linear estimator as described in Section \ref{sec:preprocessing}.
To show that functional neural networks work, and indeed surpass the performance of the baseline models, we use an FNN as described in Section \ref{sec:architectures} with two functional convolutional and one functional dense layer. For the local linear estimation, we use the quartic kernel $K(x)=\tfrac{15}{16}(1-x^2)^2$ with support $[-1, 1]$ and the bandwidth $h=5$ for the estimation of the smooth function and $h=10$ for the estimation of its derivative. For each functional layer, we used the first 5 Legendre polynomials as base functions, i.e. $\varphi_0(x)=1, \varphi_1(x)=x, \varphi_2(x) = \tfrac{1}{2}(3x^2-1), \varphi_3(x)=\tfrac{1}{2}(5x^3-3x)$ and $\varphi_4(x)=\tfrac{1}{8}(35x^4-30x^2+3)$. Further, we used 20 and 10 filters of size 25 for the two convolutional layers. As activation function we chose the \textit{exponential linear unit} (ELU), which is defined as $\sigma(x) = x \cdot \mathds{1}(x\ge 0) + (\exp(x) - 1) \cdot \mathds{1}(x<0)$. As loss function, we used the categorical crossentropy.
We trained each model 100 times, while generating a new data set for each trial. The FNNs were trained with 5 epochs.
\subsubsection*{Results}
The results of the benchmark model for both data sets with a varying number of samples $N$ and neighbors are displayed in Figures \ref{fig:simulation_results_0} and \ref{fig:simulation_results_1}. In Table \ref{tab:simulation_results}, the results of the benchmark model with the best choice of neighbors $k$ are compared with the results of the FNN. In all cases, the classifications of the functional neural network are more reliable than those of the $k$-nearest neighbors classifier. The FNN achieved an accuracy above 99.6\% in all cases, whereas the KNN classifier achieved between 93.0\% and 99.3\% for data set (I) and between 76.9\% and 86.4\% for data set (II). As expected, the shift invariance of the FNN makes it particularly helpful for data set (II), where the signal of interest may occur at any point in the observed interval. More specifically, for data set (II) the FNN's accuracy is at least 13.5\% higher than the corresponding accuracy of the KNN classifier.
\begin{figure}[ht]
\vskip 0.2in
\begin{center}
\centerline{\includegraphics[width=\columnwidth]{figures/dataset0_results.pdf}}
\caption{Mean accuracy in percent (y-axis) of the $k$-nearest neighbors classifiers for data set (I) and a varying number of neighbors (x-axis).}
\label{fig:simulation_results_0}
\end{center}
\vskip -0.2in
\end{figure}
\begin{figure}[ht]
\vskip 0.2in
\begin{center}
\centerline{\includegraphics[width=\columnwidth]{figures/dataset0_results.pdf}}
\caption{Mean accuracy in percent (y-axis) of the $k$-nearest neighbors classifiers for data set (II) and a varying number of neighbors (x-axis).}
\label{fig:simulation_results_1}
\end{center}
\vskip -0.2in
\end{figure}
\begin{table}[t]
\caption{Mean accuracy of the classifiers for the respective data sets.}
\label{tab:simulation_results}
\vskip 0.15in
\begin{center}
\begin{small}
\begin{tabular}{ c | c c | c c }
\toprule
& \multicolumn{2}{c|}{\quad{\sc Data set (I)}} & \multicolumn{2}{c}{\quad{\sc Data set (II)}} \\
$N$ & KNN & FNN & KNN & FNN \\
\midrule
1000 & 93.0\% & \textbf{99.6\%} & 76.9\% & \textbf{99.6\%} \\
2000 & 97.3\% & \textbf{99.8\%} & 81.4\% & \textbf{99.8\%} \\
3000 & 98.5\% & \textbf{99.8\%} & 83.6\% & \textbf{99.8\%} \\
4000 & 99.1\% & \textbf{99.7\%} & 85.5\% & \textbf{99.9\%} \\
5000 & 99.3\% & \textbf{99.8\%} & 86.4\% & \textbf{99.9\%} \\
\bottomrule
\end{tabular}
\end{small}
\end{center}
\vskip -0.1in
\end{table}
\subsection{Real Data Experiments}
\subsubsection*{Setup}
The BCI Competition IV Dataset 2A \citep{tangermann2012} is a common benchmark for evaluating the performance of a new method for the analysis of EEG data. To test our method, we used the 9 openly available, labeled recordings of approximately 45 minutes each.
According to the documentation, participants were asked to imagine movements of their left hand (class 1), right hand (class 2), both feet (class 3) and tongue (class 4). During each session, every imaginary movement was repeated 72 times, yielding a total of 288 trials. Each trial took approximately 8 seconds. At the beginning of each trial ($t=0\,s$), a short acoustic signal and a fixation cross on a black screen appeared. Two seconds later ($t=2\,s$), a visual cue appeared to indicate the movement, which should be imagined. The imaginary movement can be assumed to start approximately half a second after the cue ($t=2.5\,s$) and end when the fixation cross disappeared ($t=6\,s$). Each trial was followed by a short break to separate it from subsequent trials. The participants' brain activity was measured through a 22-channel EEG with 3 additional EOG (\textit{Electrooculography}) channels at a sampling rate of 250\,Hz.
For this data set, the \textit{classic approach} to benchmark a new method is to cut windows from each trial, e.\,g., between 2.5\,s and 4.5\,s after trial onset, which is feasible since the trial and cue onsets are known. However, if we move beyond externally triggered actions, we need another approach. This is particularly important in the case of brain-computer interfaces where devices should be controlled continuously. In this case, a common approach is to use \textit{sliding windows}, i.\,e., to use overlapping windows of a fixed length with a fixed step size.
We tested the proposed functional neural network, as described in Section \ref{sec:architectures} with the same specifications as in the simulation study, and compared it to the EEGNet with its default choice of hyperparameters as suggested by \cite{lawhern2018}. However, to account for the different degrees of complexity of the EEG data, we chose to use different numbers of filters in the convolutional layers. For the classic approach with 4 classes (corresponding to the 4 different imaginary movements), we tested two models with 5 + 10 and 3 + 12 filters, respectively, and denote these models by FNN(5, 10) and FNN(3, 12). These models have 2,344 and 1,564 trainable weights, which is slightly less than the 2,526 trainable parameters of the EEGNet. For the sliding windows approach with 7 classes, we increased the number of filters to 40 + 20 and 5 + 10. In this setting, we get models with 19,767 and 2,497 trainable weights compared to 2,783 for the EEGNet (with 7 classes).
For the classic approach, we used windows between the cue onset ($t=2\,s$) and the disappearance of the fixation cross ($t=6\,s$). We split each recording into 80\% train and 20\% test data, which corresponds to 230 and 58 windows each. We trained the proposed FNN and the EEGNet with 250 batches of size 32 to distinguish between the four classes (left hand, right hand, feet, tongue). In total 8000 samples were used, which means that each of the 230 training windows was used approximately 35 times.
For the sliding windows approach, we split each recording into 80\% train and 20\% test data, which corresponds to approximately 36 and 9 minutes respectively. We used windows of 1\,s and a step size of 0.016\,s which led to more than 125,000 sliding windows. These windows might coincide with a break between trials (class 1), the time between trial and cue onset (class 3), the time between cue onset and imagined movement (class 2) or one of the four imagined movements (classes 4 - 7). The windows at the transition between two classes were labeled with the most frequent class. This problem is substantially more complex than the classic approach and we have more varied data. Thus, we trained both models with 4000 batches of size 32, which is close to the number of sliding windows.
We trained each model 10 times for each of the 9 recordings.
\subsubsection*{Results}
The results for both approaches are displayed in Tables \ref{tab:real_data_results_classic} and \ref{tab:real_data_results_sliding}. As before, the accuracy represents the ratio of correctly classified windows to all windows. To account for the class imbalance, the mean recall and precision over all categories were added: the recall of a binary classifier is defined as the ratio of correctly classified positive to all true positive samples, whereas the precision of a binary classifier is defined as the ratio of correctly classified positive to all positively classified samples. These quantities for binary classifiers were extended to the multiclass problem by first calculating the respective quantity per class and then averaging the calculated quantities over all classes.
In both cases, the proposed FNNs outperformed the benchmark model. For the classic approach, the accuracy of the smaller FNN(3, 12) with 1,564 parameters had a 2.4\% higher accuracy and the larger FNN(5, 10) with 2,344 trainable parameters had a 3.47\% higher accuracy compared to the benchmark model with 2,526 parameters. The respective confusion matrices are displayed in Figure \ref{cm:classic} of Appendix \ref{app:real_data}.
For the sliding window approach, the FNN(5, 10) with a comparable number of parameters had a 1.97\% higher accuracy compared to the benchmark model, whereas the larger FNN(40, 20) outperformed the benchmark model by 4.69\% accuracy. Note that classes 4 - 7 are generally more difficult to detect. Yet, it can be seen from the confusion matrices in Figure \ref{cm:sliding_win} of Appendix \ref{app:real_data}, that the classifications of the FNN(40, 20) are particularly better for those classes, which is also reflected in the recall and precision of the classifiers.
Both the FNNs and the (default) EEGNet are relatively simple models. It can be expected that the accuracies improve for both types of models if the hyperparameters are tuned carefully. Further improvements might be possible by changing the FNN's architecture or simply using more layers.
\begin{table}[t]
\caption{Comparison of the models' quality under the classic approach.}
\label{tab:real_data_results_classic}
\vskip 0.15in
\begin{center}
\begin{small}
\begin{tabular}{lcccr}
\toprule
Model & Accuracy & Recall & Precision \\
\midrule
EEGNet & 58.13 & 58.30 & 58.59 \\
FNN(5, 10) & \textbf{61.60} & \textbf{61.81} & \textbf{61.49} \\
FNN(3, 12) & 60.53 & 60.80 & 60.56 \\
\bottomrule
\end{tabular}
\end{small}
\end{center}
\vskip -0.1in
\end{table}
\begin{table}[t]
\caption{Comparison of the models' quality under the sliding windows approach.}
\label{tab:real_data_results_sliding}
\vskip 0.15in
\begin{center}
\begin{small}
\begin{tabular}{lcccr}
\toprule
Model & Accuracy & Recall & Precision \\
\midrule
EEGNet & 46.00 & 35.46 & 40.52\\
FNN(40, 20) & \textbf{50.69} & \textbf{42.47} & \textbf{45.50}\\
FNN(5, 10) & 47.97 & 37.42 & 42.60\\
\bottomrule
\end{tabular}
\end{small}
\end{center}
\vskip -0.1in
\end{table}
\subsubsection*{Fully Functional Predictions}
With the proposed methodology it is not only possible to predict scalar-valued labels and use the model for classification, but it is also possible to predict functional labels. With the sliding windows as before, we can try to predict the class label for each time point rather than one label for the whole window. This is particularly useful at the transition from one state to another because these transitions cannot be represented by a simple classification.
We trained a functional neural network with three functional convolutional layers as depicted in the bottom of Figure \ref{fig:architecture} to predict labels for each time instant of the sliding windows. The overall accuracy was similar to the accuracy reported for the classifier above. In Figures \ref{fig:fully_func_pred_1_3} and \ref{fig:fully_func_pred_3_2} are the true and predicted labels of two windows at the transition from an inter-trial break (class 1) to the interval after a trial onset (class 3) and from the interval after a trial onset (class 3) to the interval after a cue onset (class 2). It can be seen from both figures that the predictions do not match the true labels perfectly and that the confidences at the border region are generally lower, but overall the predictions match the true labels.
\begin{figure}[ht]
\vskip 0.2in
\begin{center}
\centerline{\includegraphics[width=\columnwidth]{figures/classification_1_3.pdf}}
\caption{True and predicted class labels for a window of one second at the transition from class 1 (inter-trial break) to class 3 (time after trial onset).}
\label{fig:fully_func_pred_1_3}
\end{center}
\vskip -0.2in
\end{figure}
\begin{figure}[ht]
\vskip 0.2in
\begin{center}
\centerline{\includegraphics[width=\columnwidth]{figures/classification_2_3.pdf}}
\caption{True and predicted class labels for a window of one second at the transition from class 3 (time after trial onset) to class 2 (time after cue onset).}
\label{fig:fully_func_pred_3_2}
\end{center}
\vskip -0.2in
\end{figure}
\section{Conclusions} \label{sec:con}
In this work, we presented a new framework to analyze functional data with scalar- or functional-valued targets. We combined advantages of convolutional neural networks with those of functional data analysis. More specifically, we proposed a neural network that can be considered as shift invariant while taking the intrinsic functional structure of the data into account.
We showed that even shallow models with only two convolutional and one dense layer are more powerful than the functional k-nearest neighbors algorithm for the simulated data. Further, the results of our case study suggest that FNNs with a similar amount of trainable weights outperform EEGNet, the de facto standard model for the classification of EEG data.
The results of this paper suggest that functional neural networks are a relevant area for future research. First, it could be tested if FNNs work well with other types of time series and functional data, such as stock prices or temperature curves. It might be interesting to investigate if the methodology can be expanded to other types of data like images (considered as functions of the two variables height and width) or videos (considered as functions of the three variables time, height and width). Further, the choice of base is crucial for the performance of the network. Prior simulation studies showed that the Fourier base and Legendre polynomials lead to good results, but other bases might further improve the predictions. In FDA it is common to find roughness penalties as regularizers. Although a preliminary simulation study suggested that a base representation of the weight functions leads to better results, it would be interesting to study if the weight functions in the neural network can be learned directly while their smoothness would be ensured via corresponding roughness penalties. Finally, the proposed framework could be extended to other types of neural networks, such as recurrent neural networks or transformers.
\section*{Acknowledgment}
This work is supported by the Ministry of Economics, Innovation, Digitization and Energy of the State of North Rhine-Westphalia and the European Union, grant IT-2-2-023 (VAFES).
|
\section{Introduction}
Conversational Agents (CAs) or chatbots are gaining a wide popularity. Applications such as Apple Siri~\cite{AppleSiriHealth}, Amazon Alexa~\cite{AmazonAlexa}, and Google Assistant~\cite{GoogleAssistant} are becoming ubiquitous. Suggested by its name, CAs interact with people in natural languages, thereby providing customer services~\cite{cui2017superagent} and companionship~\cite{ta2020user}. Compared to GUIs, CAs have shallower learning curves and may even form a relationship with people through conversations~\cite{dingler2021use}. The digital health domain particularly benefits from the use of CAs: chatbots can be deployed to collect self-reports and provide personalized coaching to different individuals~\cite{mitchell2022examining}. Studies have found that people are willing to engage with chatbots and provide valuable information, such as self-reports~\cite{wei2021understanding} and survey responses~\cite{celino2020submitting}, to chatbots~\cite{xiao2020if, xiao2020tell}.
Many digital health applications that are designed to promote behavioral changes and health interventions require people's long-term adherence. Chatbots' ability to ``converse'' naturally, as distinguished from the GUI-based systems~\cite{luger2016like}, has the potential to help people develop long-term adoptions for health monitoring~\cite{dingler2021use, fitzpatrick2017delivering, mitchell2022examining}. However, existing commercial chatbot frameworks, such as Dialogflow\cite{DialogFlowDev} and Amazon Alexa~\cite{AlexaDev}, predominantly only support building rule-based and scripted chatbots~\cite{wei2021understanding, luo2020tandemtrack}. Lacking flexible flows, these chatbots usually appear robotic and unnatural~\cite{mctear2018conversational} and may cause boredom in long-term deployments. On the other hand, implementing chatbots that can have more diverse and dynamic conversations requires large and specific domain datasets~\cite{mitchell2022examining}. For example, an open-domain chatbot Meena was trained on 341GB of dialogue sessions~\cite{adiwardana2020towards}. Creating such large datasets is costly so the datasets are often proprietary and inaccessible to public. Furthermore, most demonstrations of open-ended chatbots focus on performing free-form conversation in general topics and do not support end-user customizations. Little research has been done to explore low-effort bootstrapping ways to build chatbots that can effectively perform pre-defined tasks, such as inquire people about their health information, and carry on naturalistic conversations at the same time.
Recent Large Language Models (LLMs; \textit{e.g.}, GPT-3~\cite{brown2020language}, PaLM~\cite{Chowdhery2022PALM}, OPT~\cite{Zhang2022OPT}, HyperCLOVA~\cite{Kim2021HyperCLOVA}), with billions of parameters pre-trained on a large amount of language corpora, provide a new way to build chatbots that have the ability to not only effectively inquire people for specified information but also converse naturally and flexibly. LLMs work differently from traditional task-specific models as they accept \textit{prompts} written in natural language as input to perform various tasks~\cite{brown2020language, Liu2021PretrainPromptAndPredict}.
Taking prompts that include the conversation history, LLMs can generate the following responses accordingly without any training data, thereby functioning as a chatbot. Compared to other frameworks, LLMs show great potential in scaffolding chatbots that are sensible of contexts and even respond to off-topic user messages~\cite{volum2022craft}. Further, since LLMs operate on natural language inputs, people can have the opportunity to personalize or even build their own chatbots. This feature can be especially useful for people without programming experiences, such as medical practitioners~\cite{dingler2021use}.
Despite these potentials, it is yet fully understood how LLMs \textit{read} the prompt and \textit{use} pretrained knowledge~\cite{brown2020language, Liu2021RetrievalGPT3}, the development of prompts is usually conducted through iterative trial and error~\cite{liu2022design}. While the HCI community have actively explored the use of LLMs in various domains (\textit{e.g.},~\cite{chung2022talebrush, lee2022coauthor, wu2022ai}), research that leverages LLMs for powering chatbots, particularly task-oriented ones~\cite{mehri2022lad, bae2022building, volum2022craft}, is still sparse. Due to the inherent characteristics of LLMs, LLM-driven chatbots may be error-prone~\cite{korngiebel2021considering} or digress from their tasks~\cite{volum2022craft}. Designing robust prompts is crucial for ``restricting'' chatbots to conduct desired tasks.
In this study, we investigate how LLMs can power chatbots to collect user self-reports while carrying on naturalistic conversations. Towards this aim, we built a set of chatbots (\autoref{fig:teaser}) that run on GPT-3~\cite{brown2020language} and converse to collect self-report data in four health-related topics---sleep, food intake, work and productivity, and exercise. We chose GPT-3 as an underlying LLM because it is one of the mainstream LLMs that are publicly available via commercial APIs. We formulate the model prompt to include the information slots (\textit{i.e.}, information properties of a topic) that we intend the chatbot to collect and the job identities (e.g., sleep expert for the topic sleep) to help drive the conversations. We investigate how two design factors in prompts---information specification format and personality modifier---impact the slot filling ability and the conversation style of chatbots. In total, we created 16 chatbots (4~topics $\times$ 2~formats $\times$ 2~personality modifiers) with different prompts.
To evaluate the performance of different prompts in steering chatbots, we conducted an online study ($N$ = 48) with our chatbots on a web interface. All participants talked to chatbots of the four topics but each of them experienced chatbots run on the same prompt design.
To the best of our knowledge, our work is the first to explore the usability of LLMs for building chatbots for collecting self-report data. We found that our zero-shot prompts, without either example dialogues or fine-tuning, covered 79\% of the desired information slots among all dialogues. Through conversation analysis, we found that the information specification format as well as the use of personality modifier can impact the chatbots' slot filling ability and conversation styles. Also, the chatbots generally reacted to participants' self-reported answers in an empathetic way, appreciating their accomplishments as well as sympathizing participants for the negative outcomes. Consequently, some participants perceived these chatbots to be understanding and take into account their messages when responding and others indicated that they were surprised to find the chatbots' responses were accurate and detailed. Research contributions from this work are threefold:
\begin{enumerate}
\item Empirical results from an online study ($N=48$), demonstrating the \textit{feasibility} of chatbots powered by LLMs in not only carrying on conversations to collect specified information but also exhibiting abilities in maintaining context, state-tracking, and providing off-topic suggestions.
\item Examination on how different prompt designs and other factors impact the chatbots' behaviors, providing insights for future researchers to easily scaffold chatbots for data collection with LLMs.
\item Implications on how future LLMs-based chatbot platforms can improve the conversation quality, drawn on the analysis of the dialogue errors.
\end{enumerate}
\section{Related Work}
In this section, we cover related work in the areas of (1) self-report data collection through chatbots, (2) design considerations for chatbots, (3) chatbot platforms, and (4) designing LLM prompts for chatbots.
\subsection{Self-Report Data Collection through Chatbots}
Personal informatics systems have commonly incorporated data collection techniques to track personal health and activity~\cite{Li10StageBasedModel, Choe2017SemiAutomatedTracking}.
While various physiological or physical activity data---such as step count, heart rate, and sleep duration---can be captured automatically by sensors and wearable devices~\cite{laborde2020older}, various types of personal data still demand \textit{self-reporting} by the person who self-tracks~\cite{Choe2017SemiAutomatedTracking}. For example, food intake (\textit{e.g.}, ~\cite{cordeiro2015barriers}) or work tasks (\textit{e.g.},~\cite{Kim2019UnderstandingProductivity}) are not reliably captured by sensors and thus require manual inputs. In addition, reflective questions (\textit{e.g.}, Why did you eat this food?~\cite{luo2021foodscrap}) and subjective measurements (\textit{e.g.}, Sleep quality) inherently require to be captured manually.
A majority of digital self-tracking tools that involved manual data capture inherited the traditional concepts of self-monitoring or journaling and provide form-based GUIs such as a list of checkboxes and text fields~\cite{Kim2017OmniTrack, Jeon2016MsThesis}. However, repeated manual input on a computer or smartphone screen is burdensome and may gradually disengage people from tracking~\cite{choe2014understanding, Choe2017SemiAutomatedTracking}.
As an input modality to lower the capture burden and enhance the richness of the captured information, natural language has recently gained interest~\cite{luo2021foodscrap, Silva2021Food}. Prior research found that when people are allowed to insert data in free-form natural language, they tend to provide detailed answers with surrounding contexts~\cite{luo2021foodscrap, Kim2022MyMove}.
Going further, conversational interaction, where a system and a user communicate in natural languages, has become one emerging interface for collecting self-reports.
Chatbots are considered easier to use and more accessible than GUIs as they minimize the use of graphical widgets employ the intuitive conversational interaction. Regarding data collection, a plethora of research has explored the use of chatbots in place of traditional form-based surveys (\textit{e.g.},~\cite{xiao2020tell, bemmann2021chatbots, celino2020submitting, kim2019comparing}). For example, studies with surveys with close-ended questions found that chatbots can collect the same quality, if not higher, user responses as GUIs~\cite{celino2020submitting, kim2019comparing}. \citet{xiao2020tell} built a chatbot to conduct interviews with open-ended questions. Compared to the traditional web survey, their participants showed higher engagement and provided higher-quality responses when talking to the chatbot. Further, incorporating more humanized traits, such as casual conversation styles~\cite{kim2019comparing}, self-introduction, and echoing~\cite{rhim2022application}, led to not only a higher level of user engagement and satisfaction but also more self-disclosure in responses.
With more focus on self-reported data, prior studies leveraged chatbots to collect self-reports such as emotion (\textit{e.g.},~\cite{bemmann2021chatbots}), pain level (\textit{e.g.},~\cite{wrzus2015lab}), and food intake (\textit{e.g.},~\cite{mitchell2022examining}). For example, \citet{bemmann2021chatbots} combined a chatbot with the experience sampling method (ESM, \cite{larson2014experience}) and found that personalized chatbots have the potential to collect data on sensitive or personal topics.
\citet{mitchell2022examining} compared fully-scripted, rule-based, and retrieval-based chatbots for collecting food nutrition. They found the better fulfillment of data collection is not necessarily associated with the higher perceived quality of the chatbot as a diet coach, suggesting the importance of conversational content in user experience.
This work expands the line of research on chatbots that collect self-reports. In contrast to prior studies that involved predefined conversation logic or retrieval model training on domain-specific datasets,
we explore the potential of LLMs in bootstrapping chatbots that can collect self-reports through conversations on four health topics---sleep, food, work, and exercise.
\subsection{Design Considerations for Chatbots}
Prior works in HCI explored user behaviors with chatbots and proposed suggestions to improve user experience with them. For example, ~\citet{luger2016like} found that people restricted their language uses when interacting with CAs. ~\citet{jain2018evaluating} revealed that many first-time chatbot users had disappointment and frustration with the selected chatbots: most chatbots lacked the ability to fully comprehend user messages or intentions. Since conversation breakdowns are still common~\cite{ashktorab2019resilient}, several studies have explored repair strategies, such as apologies, compensation, and providing option~\cite{lee2010gracefully}. ~\citet{ashktorab2019resilient} also evaluated other strategies, such as confirmation, repeat, keywords highlight \& explanation, and recommended that chatbots should acknowledge misunderstanding in simple terms, explain model limitation in natural ways, and adapt individualized strategies. Although existing chatbot frameworks also have error recovery features, their features are not only limited but often cannot allow quick repairs~\cite{wei2022could, myers2018patterns, mctear2018conversational}.
Another key to improving the user experience is to make chatbots more playful and human-like~\cite{liao2018all}. The level of empathy~\cite{celino2020submitting, rashkin2018towards} and the repetitive rate~\cite{see2019makes} are two commonly used metrics of human-likeness. For example, the playful interactions (\textit{e.g.}, telling jokes) or humorous responses enabled many people to start using CAs~\cite{luger2016like} and it is crucial for chatbots to support sustainable playfulness~\cite{serenko2008model}. Also, human-like features and fun personalities are found to make chatbots more enjoyable to interact~\cite{jain2018evaluating}. Even for work-related chatbots, some people still preferred chatbots that were human-like~\cite{liao2016can}, and ~\citet{liao2018all} envisioned that a reusable conversational module including common chit-chats and social attributes could be developed. In other words, future chatbot platforms should allow developers to easily build personalized chatbots with different personalities~\cite{volkel2020developing} and conversation styles~\cite{braun2019your, krenn2014effects}. Lastly, developers should aim to improve chatbots' ability to maintain contexts to support smoother and natural conversations~\cite{ashktorab2019resilient}. In this work, we investigate whether LLMs can steer chatbots that have social attributes and can resolve conversation breakdowns.
\subsection{Chatbot Platforms}
Building chatbots is challenging and time-consuming, and many design suggestions discussed above are difficult to implement. Many open-domain chatbots that engage and entertain people socially are all dependent on large datasets~\cite{zhou2020design, adiwardana2020towards}. In the HCI community, rule-based dialogue systems are widely used. ~\citet{celino2020submitting} built their survey chatbot with pre-defined conversation flows as they intended to avoid disappointments caused by the chatbot's inability to understand certain utterances~\cite{luger2016like}. Although rule-based chatbots are unlikely to cause breakdowns, the resulted rigid conversations can make people lose interest in the long term. On the other hand, ~\citet{xiao2020tell} built their survey chatbot using Juji~\cite{Jujiio}, which automatically equip chatbots with rich existing conversational skills. Using the Juji GUI to add questions is relatively simple, but it is unclear whether developers can modify the chatbot's expressed personality. Lastly, other commercial chatbot frameworks, such as Dialogflow~\cite{GoogleDialogFlow} and IBM Watson~\cite{IBMWatson}, also allow developers to build rule-based chatbots with GUIs~\cite{thorat2020review}. However, creating more dynamic conversations usually requires programming skills. Even for professional developers, it is challenging to create well-designed conversational flows and pre-define user intents and chatbot messages~\cite{mctear2018conversational}. Using LLMs to power chatbots is a new way to build chatbots~\cite{volum2022craft, brown2020language}. LLMs accept natural language prompts so that people without any knowledge of programming but are interested in building chatbots for data collection can create prompts~\cite{Kim2021HyperCLOVA}. Hence, compared to prior methods that approach personalization by building data-driven user profiles~\cite{banovic2018warming, early2016test}, customizing prompts in natural languages essentially hands off the control to each individual who builds chatbots. As such, it becomes more straightforward to scaffold personalized chatbots (\textit{e.g.}, assigning a preferred personality) by revising prompts accordingly. Nevertheless, it is unclear how to design prompts for LLMs to steer chatbots that can effectively ask questions around desired information and have different conversational styles.
\subsection{Designing LLM Prompts for Chatbots}
Prompts are natural language texts to LLMs to produce desired outputs. With proper prompt inputs, GPT-3 can be used to translate texts, answer questions, write essays, and generate dialogues without any fine-tuning~\cite{brown2020language}. While the mechanism enabling such few-shot abilities behind LLMs is still veiled~\cite{brown2020language, min2022rethinking}, some prompting techniques are found to improve the model performance. One technique that surprisingly improves the generation quality is by conditioning the prompt with an identity. For example, by inserting the statement ``\texttt{You are an expert Python programmer}'' into prompts, models can generate higher quality codes~\cite{austin2021program}, and similarly, ``\texttt{I'm a math tutor}'' is suggested to improve models' performance in solving math problems~\cite{austin_2022}. In HCI, researchers are also interested in investigating the design guidelines for prompts. ~\citet{liu2022design} explored different design factors, such as phrasing, styles, subjects, and random seeds of prompts for text-to-image generative models. They found that the subject and style keywords are more important than the connecting words and the initializing seeds can significantly impact the quality of generations. Similarly, ~\citet{lee2022coauthor} discovered that the randomness impacted writing collaboration and new identity generations. In terms of dialogues, some studies designed prompts for LLMs to generate single-turn responses~\cite{Wang2021TheraphyChatbot} and examined whether these artificial responses are comparable to human responses~\cite{tack2022ai}.
So far, there exist limited guidelines for LLM prompts for steering chatbots. OpenAI's website provides an example prompt\footnote{\url{https://beta.openai.com/docs/guides/completion/conversation}} for open-domain chatbots. For task-oriented chatbots, ~\citet{volum2022craft} built interactive Non-Player Characters (NPCs) through \textit{few-shot} prompting. By appending every user input to the initial prompt, the resulted new prompt is fed to the Codex model to generate responses in both code \& language. Their chatbots could also answer user requests, hold multi-turn conversations and retain contexts. Their results show that an initial ``seed prompt'' can enable LLMs to steer interactive chatbots. However, these Q\&A NPCs were not designed to actively ask questions to the player and collect information. Our study, on the other hand, aims to design prompts for LLMs to power chatbots that can conduct self-report collection conversations.
In summary, LLMs provide an easy and quick alternative to create chatbots~\cite{brown2020language}; however, we have only sparse guidelines on how to leverage them effectively. In this study, we set out to address this gap and provide empirical insights into prompt designs for chatbots with different ``job identities'' that are dedicated to collecting self-reports. Further, we investigate whether LLM-driven chatbots can exhibit important features such social attributes~\cite{liao2018all}, employ empathy~\cite{rashkin2018towards}, as well as have the ability to handle breakdowns~\cite{ashktorab2019resilient}.
\section{Method}
To examine how different prompt design factors impact LLM-driven chatbots' performance, we implemented a web-based chatbot interface and conducted an online user study to collect dialogue data from people. We describe our prompt designs and the experimental design of the user study.
\input{tables/slots.tex}
\subsection{The LLM-driven Chatbot Framework}
LLMs power chatbots through generating the next response based on the input prompt. \autoref{fig:teaser} describes the mechanism of our chatbot running on an LLM. At each turn, the back-end system combines a prefix (\circledigit{A} in \autoref{fig:teaser}) and the history of the ongoing conversation (\circledigit{B} in \autoref{fig:teaser}) into a single prompt and feeds it into the LLM (\circledigit{C} in \autoref{fig:teaser}). Then the LLM generates the following agent's response. The prefix, treated as an instruction for an LLM, consists of the description of persona and the specification of the desired information slots that the chatbot should capture.
\subsubsection{Prompt Designs for Chatbots}
In this study, we envisioned a self-monitoring scenario where chatbots proactively interact with people and inquire them about their retrospective health-related behaviors of the previous day through natural conversations. Towards this goal, we first chose four health-related topics---\textbf{Sleep}~\cite{min2014toss}, \textbf{Food} intake~\cite{luo2021foodscrap, Epstein2016Taking5}, \textbf{Work} and productivity~\cite{li2011understanding, Kim2019UnderstandingProductivity, Kim2016TimeAware}, and \textbf{Exercise}~\cite{luo2020tandemtrack}---that retain interests not only by the research community and are also personal use in daily life~\cite{williams2006simply}. Then, we defined the behavioral data (slots) that we intended to have chatbots to collect. \autoref{tab:slots} summarizes the information slots for each topic, which the chatbots should capture in each conversation session. Another aspect that we intended to explore was whether we could enable LLMs to steer chatbots to exhibit different conversation styles as different individual's experience with chatbots can be improved with personalizations~\cite{braun2019your}.
Currently, there is a lack of prompt designs for chatbots. While our prompt engineering for chatbots was mostly ``trial and error'', we considered the following rationals and revised prompts in each iteration. Initially, we started with OpenAI's example prompt, which defines an ``AI assistant'' with several characteristics(\textit{e.g.}, ``\texttt{The following is a conversation with an AI assistant. The assistant is helpful, creative, clever, and very friendly.}''). We modified the example prompt by including questions of interests (\textit{e.g.}, ``\texttt{How's your sleep last night?}'') as well as by adding the list of information slots (\textit{e.g.}, ``...\texttt{The assistant asks questions about human's sleep time, sleep duration, and sleep quality.}''). However, we found that these modifications tended to make conversations digress and the resulted chatbots converse more similar to generic customer service chatbots.
Findings from research~\cite{Kojima2022ZeroShotReasoners, wei2022chain, austin2021program} and anecdotal evidences from social medias~\cite{austin2021program} suggest that specifying identities (\textit{e.g.}, ``Python programmer'') can improve GPT-3 model performance. Inspired by this, we substituted the ``AI assistant'' with more specific job identities (\textit{e.g.}, ``fitness coach'') and found that the resulted chatbots appeared to have more domain knowledge and are less likely to digress in our tests. As such, we picked different job identities for each topic---\textit{sleep expert} (Sleep), \textit{dietitian} (Food), \textit{life coach} (Work), and \textit{fitness coach} (Exercise).
With the above modifications, we designed prompts by first describing a chatbot with a specific identity and then including a list of slots (\textit{e.g.}, ``\texttt{I'm inquiring about what they had for breakfast, lunch, dinner, and snacks...}''; see \autoref{fig:chatbot_prompt}, right). Through multiple trials, this design worked better than generic one despite still had occasional digressions. Hence, we experimented with another prompt format inspired by the state-tracking technique in task-oriented dialogues~\cite{Lee2021Dialogue}. Instead of being described literally, the slots are structured into a form (\textit{e.g.}, ``\texttt{Meals and snacks from yesterday: Breakfast -> [placeholder] Lunch -> [placeholder] ...}''; we used an empty string as a placeholder; see \autoref{fig:chatbot_prompt}, left). Both designs performed similarly in our limited internal testings, hence we aimed to investigate the performance of two formats in the user study.
In terms of manipulating conversational styles, we introduced the use of a modifier in prompts--``\texttt{who always shows empathy and engages my customer in conversations},'' to the prompt (See \autoref{fig:chatbot_prompt}, right). We hypothesize that with this modifier, the chatbot is more likely to express empathy in conversations and have a higher level of interactivity---\textit{i.e.}, use more emphatic expressions and be more responsive to user responses. Conversely, without the modifier (See \autoref{fig:chatbot_prompt}, left), we expect the chatbot to be more neutral, formally exchanging messages with users and appear less empathetic. Lastly, during our trials, we found that GPT-3 had the tendency to ask multiple questions in one turn. To restrict this behavior, we added ``\texttt{I only ask one question at a time.}'' to the prompt.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{img/prompts.pdf}
\caption{Prompt design combining two factors, information format and personality modifier, in the Food intake topic.}
\label{fig:chatbot_prompt}
\end{figure*}
\subsubsection{Model and Parameters}
To power chatbots with above prompts, we used \texttt{davinci-text-002}, the largest and most capable model of GPT-3 as of June 2022, publicly accessible via OpenAI's API~\cite{openai}.
This model accepts 4,000 byte-pair encoding tokens at maximum in a prompt per request. Our prompt templates in the initial state were encoded into around only 120 tokens (3\% of the limit) and allowed sufficient room for the appended conversation history.
For all chatbots, we uniformly applied the same generative parameters: temperature as 0.9, the presence penalty as 0.6, and the frequency penalty as 0.5. We kept the temperature and the presence penalty unchanged based on OpenAI's suggestions and increased the frequency penalty to reduce the re-use of words.
\subsubsection{The Web Chat Interface}
We implemented a web interface to host our LLM-driven chatbots, following a typical chat interface design (See Appendix \ref{appendix:interface}). The webpage was written in TypeScript~\cite{TypeScript} on React~\cite{React} and runs on the Node.js~\cite{nodejs} server. The server communicates with GPT-3 leveraging OpenAI's API~\cite{openai}. To simplify the conversation flow, we disabled people to submit multiple utterances in a row. Correspondingly, the chatbots also delivered one utterance at a time. When a user submitted an utterance, the server appended the current dialog history at the end of the prompt template and fed it to GPT-3 to generate the following response.
\needtocheck{
The source code for the chatbot framework and the web interface is available at \urlstyle{tt}\url{https://naver-ai.github.io/llm-chatbot}.}
\subsection{Online Study}
\subsubsection{Experimental Conditions}
Combining the two design factors, we created four designs of prompts: SP (\textbf{S}tructured format with \textbf{P}ersonality modifier), SN (\textbf{S}tructured format with \textbf{N}o personality modifier), DP (\textbf{D}escriptive format with \textbf{P}ersonality modifier), and DN (\textbf{D}escriptive format with \textbf{N}o personality modifier). Each participant was assigned to one prompt design and engaged in conversations of all four topics. (Refer to the supplementary material for all 16 variations of GPT-3 prompts created for combinations on topic and condition.) To mitigate the ordering effect among topics, half of the participants conversed in the order of Work--Food--Exercise--Sleep, and the other half in the order of Exercise--Sleep--Work--Food. Additionally, for each topic, we requested participants to engage with the chatbot twice: one in the \textbf{Positive} path (\textit{e.g.}, report high-quality sleep) and one in the \textbf{Negative} path (\textit{e.g.}, report poor sleep). Refer to Appendix \ref{appendix:paths} for an exhaustive list of paths and hints by topic provided by us to guide participants to compose their answers for each path accordingly.
\subsubsection{Web Chat Session}
After signing an electronic consent form on the study website, participants went through eight conversations ($4~topics * 2~paths$). On the web chat interface (See Appendix \ref{appendix:interface}), we put guidelines including the instructions and the conversation path that participants should follow (See Blue text in Appendix \ref{appendix:interface}, right). Since we did not incorporate ending detection algorithms, we asked participants to click the `Next' button to proceed to the next conversation when they thought the conversation was naturally over or the chatbot kept sending repetitive messages. The completed dialogues were stored in our server.
\subsubsection{Exit Survey}
After completing eight conversations, the web page automatically redirected participants to an online survey. The survey consisted of three 5-point Likert scale questions and one open-ended feedback textfield. The Likert scale questions were: (1) ``\textit{Do you think the chatbot understands your answers?}'' (2) ``\textit{Do you think the chatbot takes into account of your answers when responding?}'' and (3) ``\textit{Do you think the chatbot talks more like a human who shows more empathy or more like a robot who behaves mechanically?}'' The open-ended feedback question stated, ``\textit{If you have any other comments or thoughts about the chatbot (\textit{e.g.}, things that you've liked or disliked), please share with us.}''
The first two questions can measure whether participants think the chatbot acknowledges their answers and respond accordingly and the third question is an overall measure of whether the chatbot is perceived as being empathetic. With participants' subjective evaluations, we hope to see whether the personality modifier can impact the chatbot's way of talking.
\subsection{Participants}
We recruited participants by word-of-mouth and posting advertisements at a large tech company, social media, and online forums in local universities. We sent the link to our study website to 83 people who filled out a screener and met our inclusion criteria: (1) aged 19 or older; (2) fluent English speaker; and (3) have the experience in talking to chatbots of any kind. 54 people completed the online study session and submitted an exit survey. The entire study lasted less than 20 minutes and all participants received e-gift cards (equivalent to \$5 USD) after they completed the study.
We excluded six people's data from analysis; one made significant amount of grammatical errors and the rest completed less than half dialogues. \autoref{tab:demographics} summarizes the demographic of the final 48 participants (aged 19 to 56, 18 females). Fourteen out of 48 (29\%) participants were native/bilingual and 22 out of 48 (46\%) participants had never heard of or used LLMs. Each prompt design condition included 12 participants.
\input{tables/participants}
\subsection{Data Analysis}
We collected rich dialogue data and valuable user subjective evaluation feedbacks. We performed both quantitative and qualitative analysis to examine chatbots' conversation styles, the slot filling performance, and participants' experiences with our chatbots. For each dialogue, we calculated commonly used descriptive metrics such as the number of turns and the average word counts per turn, which we report in \autoref{sec:results:descriptive}.
\ipstart{Slot Filling Performance} Our study aimed to investigate whether LLMs can drive chatbots to effectively ask defined questions and collect desired information specified in Table~\ref{tab:slots}. To calculate the amount of information that can be obtained by our chatbots, one researcher \textit{manually} inspected and determined whether each of the pre-defined information slots could be extracted from collected dialogues. More specifically, for \textit{sleep quality} and \textit{productivity rate}, which were specified as a scale of 1 to 10, we marked the slot as filled only if a numerical value (\textit{e.g.}, 9) rather than a vague phrase (\textit{e.g.}, good sleep) was given. For \textit{feelings after eating} in Food, we treated the slot to be filled if feelings regarding one or more meals were covered. We report the analysis of slot filling rate in \autoref{sec:results:slot}. Based on the binary coding, we calculated the \textbf{slot filling rate}: the ratio of the number of information slots extracted from the dialogue against the total number of slots. We use the slot filling rate to infer the data collection performance of chatbots.
\ipstart{Dialogue Acts and User/Chatbot Behaviors} To understand the conversational behaviors of the chatbots, we coded \textit{dialogue act} for each turn of conversations. Referring to some existing taxonomies of dialogue acts~\cite{shi2020effects, traum200020, welivita2020taxonomy}, three researchers independently coded one participant's dialogues (132 turns; 1.8\%) to identify emerging dialogue acts. Additionally, researchers labeled chatbot turns that did not fit in the conversation context or originated from the inherent artifacts of an LLM. We resolved discrepancies in coding and developed the first version of codebook with three dimensions of codes: (1)~essential acts and (2)~empathy \& engagement behaviors, and (3)~problematic chatbot turns. Then two researchers reiterated the independent coding of four other participants' dialogues (1 participant from each condition, 32 dialogues in total) with the codebook. The two researchers resolved discrepancies through multiple sessions of discussion until their inter-rater reliability (Cohen's Kappa) reached 0.96 for essential acts and 0.935 for empathy \& engagement behaviors. Compared to these dialogue acts, the occurrence of errors was sparse. Hence, the two researchers discussed the entire problematic turns coded by each other together and reached the full agreement. With the finalized codebooks~(See \autoref{tab:essential_components}, \ref{tab:empathy}, and \ref{tab:errors}), the first author coded the rest of the data.
As a result, each turn was classified as one of the essential acts---\textit{greeting}, \textit{task opening}, \textit{required question/answers (RQ/RA)}, \textit{secondary question/answers (SQ/SA)}, \textit{statement}, and \textit{closing}. We assigned the most prominent act to turns consisting of multiple sentences. Independent of essential acts, we multi-coded each turn with the empathy \& engagement behaviors described in \autoref{tab:empathy}. For example, to a general compliment \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {}~``\textit{That's great}'' (\textbf{Statement}), we assigned only the \textbf{Appreciating} behavior, whereas we also treated \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {}~``\textit{That's great to hear that your legs are feeling stronger!}'' (\textbf{Statement}) to be both \textbf{Acknowledging} and \textbf{Appreciating} as the compliment directly addressed to the user input. We were interested in such acknowledging behaviors because \textit{specificity} was an important indicator of the capability of open-domain chatbots~\citet{adiwardana2020towards}.
\ipstart{Statistical Analysis} To understand the impact of the study factors, including prompt design, conversation topic, and the conversation path, to the chatbots' slot filling performance and conversational flows, we used \textit{mixed-effect models} because these models can handle unbalanced data repeatedly measured from the same participants~\cite{Pinheiro2000MixedEffects}. For each dialogue metric we want to assess, we fitted a mixed-effect model that predicts the metric, treating each dialogue as a data point. Starting from a full model containing participants as a random effect and the four main study factors--information format, personality modifier, topic, and path--and their interactions as fixed effects, we performed the step-wise backward elimination removing variables not significantly contributing the model, through Maximum-likelihood tests. For significant variables, we performed post-hoc pairwise comparisons of the least-squared means (LSM) of the metric using \texttt{emmeans}~\cite{emmeans} package in R.
\ipstart{Subjective Feedback}
To assess the difference among the experimental conditions, we conducted Kruskal-Wallis tests over the four rating questions. We also referenced the open-ended feedback from when interpreting the participants' reactions to specific phenomena of the conversations.
\section{Results}
In this section, we report the results of our study in six parts.
In \autoref{sec:results:descriptive}, we provide an overview of the dialogue dataset we collected.
In \autoref{sec:results:slot}, we report the data collection performance of our chatbots and factors that impact the performance.
In \autoref{sec:results:essential}, we report the types of the essential dialogue acts and assess how the prompt design and other factors impact the dialogue acts and, in turn, the data collection performance.
In \autoref{sec:results:empathy}, we report the types of the empathetic and engaging behaviors of chatbots and assess how the prompt design and other factors impact such behaviors of the chatbots.
In \autoref{sec:results:errors}, we explore the problematic chatbot utterances mainly caused by the erroneous behaviors of a large language model. Lastly, in \autoref{sec:results:subjective}, we report on participants' subjective evaluation from the exit surveys.
\subsection{Descriptive Statistics}\label{sec:results:descriptive}
\input{tables/descriptive-stats.tex}
From 48 participants, we collected 374 dialogues (7,442 turns in total): 91 from SP; 96 from SN; 95 from DP, and 91 from DN. Regarding the conversation topic, 94, 91, 94, and 95 dialogues were from Sleep, Work, Exercise, and Food Intake, respectively. Eight participants missed one dialogue per each and one missed two, mainly due to temporary server issues or accidental skips
Prompt designs impacted the word lengths and the number of turns of chatbots. \autoref{tab:descriptive_stats} summarizes the number of turns and word counts by prompt design. The average number of turns per dialogue is around 20 with more average turns under the two descriptive conditions (DP and DN). The maximum number of dialogue turns is 75 under the DP condition (only 1 dialogue). In terms of word counts, dialogues under the descriptive conditions (DP, DN) had more words than those under the structured conditions (SP, SN): both chatbots and participants uttered more words under the descriptive conditions. The DP condition, in particular, leads to the most number of words of dialogues.
\subsection{Slot Filling Rate}\label{sec:results:slot}
\input{tables/slot-rate.tex}
\begin{figure*}[t]
\begin{flushleft}
\sffamily\footnotesize{\textbf{***}$p$<.001; \textbf{**}$p$<.01; \textbf{*}$p$<.05}
\end{flushleft}
\centering
\includegraphics[width=0.8\textwidth]{img/fig_slot_emmeans.pdf}
\labelphantom{fig:slot_emmeans:condition}
\labelphantom{fig:slot_emmeans:topic}
\labelphantom{fig:slot_emmeans:path}
\caption{95\% confidence intervals of slot filling rate by variables with a significant effect: (a) the combination of the information format and personality modifier represented as study condition; (b) topic; and (c) the conversation path. The asterisks with arms indicate significance between the connected categories. (Refer to Appendix~\ref{appendix:stats:slot} for model details and statistics.)}
\label{fig:slot_emmeans}
\end{figure*}
Prompt designs significantly impacted the slot filling performance of chatbots. \autoref{tab:completion_rate} summarizes the average slot filling rates of chatbots by conditions and topics. On average, all chatbots have reached over 70\% slot filling rates. The dialogues in the SP-Exercise condition had the highest rate (93\%) and those in the SN-Work condition had the lowest rate (64\%).
The maximum-likelihood test revealed that there was no significant random effect of participants, indicating that participants have little impact on chatbots' data collection performance. On the other hand, there were significant random effects of the topics($p<.0001$), the conversation paths ($p=.01$), and the interaction between the information formats and personality modifiers ($p<.001$).
\autoref{fig:slot_emmeans} shows the significance over the 95\% confidence intervals of the slot filling rate in each category of the significant variables. The dialogues in DP condition had significantly lower rates than those in SP ($p=.01$) and DN ($p=.008$). This suggests that the personality modifier impacted chatbots differently: with the modifier, chatbots with the structured prompt yield higher rates whereas chatbots with the descriptive format yield higher rates without the modifier~(See \autoref{fig:slot_emmeans:condition}). In terms of topic, Exercise dialogues had the highest rate of 88.4\%, which was significantly higher than those in the other three topics: Sleep ($p=.01$), Work ($p<.0001$), and Food ($p=.02$) (See \autoref{fig:slot_emmeans:topic}). Lastly, dialogues in the Positive path had significantly higher rates than those in the Negative path ($p=.01$) (See \autoref{fig:slot_emmeans:path}).
\begin{figure*}[h]
\includegraphics[width=\textwidth]{img/fig_slot_by_question.pdf}
\caption{Breakdowns of the percentage of filled slots by the order of questions for each topic. Work and Exercise consist of four slots.}
\label{fig:slot_by_question}
\end{figure*}
As seen in \autoref{fig:slot_by_question}, there is a general trend that slots specified earlier in prompts were more likely to be covered by chatbots. For example, the first slots in all topics were covered in 90.3\% of the dialogues, but the last specified slots in Sleep (sleep quality) and Work (what went well) were omitted around 40\% of the dialogues. Interestingly, the last specified slot of Food (feelings after eating) was diligently covered: chatbots often asked how participants felt after talking about each meal rather than asking their feelings once towards the end.
\subsection{Essential Dialogue Acts}\label{sec:results:essential}
\input{tables/acts-essential}
To further understand how chatbots powered by different prompt designs talk, we categorized conversation turns into dialogue acts. We provide the summary of essential dialogue acts and their distributions in \autoref{tab:essential_components}
Here, we report chatbots' essential acts regarding question/answering and non-question statements.
\subsubsection{Required and Secondary Questions}
We identified two types of questions that the chatbots asked: required questions (RQ) and secondary questions (SQ). The RQs were directly related to the specified information slots, whereas SQs were not directly related to the information slots but rather follow-up details or elaboration. Despite being relevant to the conversation topic, SQs sometimes caused the conversation to digress. Although not very common (95 out of 1,029 SQ turns in total; 9.3\%), participants also asked questions to the chatbot, which were all categorized as SQ/SA. The majority of collected dialogues consisted of question/answering: Overall, 4,879 out of 7,442 turns ($avg.$ 64.72\% of turns per dialogue; $min=10.67\%$, $max=93.54\%$) were classified as RQ, RA, SQ, or SA (See \autoref{tab:essential_components}).
We first investigated the impact of prompt designs on the chatbot-spoken RQ and SQ turn ratios using two mixed-effect models with each turn ratio as a dependent variable, respectively.
\autoref{fig:questions_emmeans} shows the 95\% confidence intervals of the chatbot-spoken RQ and SQ turn ratios by four study factors (information format, personality modifier, topic, and path). The structured format significantly increased the RQ turn ratio ($p=.03$; see \autoref{fig:questions_emmeans:rq_format}) but decreased the SQ turn ratio ($p=.002$; see \autoref{fig:questions_emmeans:sq_format}). On the other hand, the personality modifier did not impact either RQ ($p=.94$) nor SQ ($p=.43$) turn ratios and made no difference within the same information format (See \autoref{fig:questions_emmeans:rq_condition} and \ref{fig:questions_emmeans:sq_condition}).
In terms of the conversation path, we find that, overall, the Positive path led to a higher RQ turn ratio ($p=.01$; see \autoref{fig:questions_emmeans:rq_path}) and a lower SQ turn ratio ($p=.002$; see \autoref{fig:questions_emmeans:sq_path}). However, under different topics, the conversation path had different impacts on RQ and SQ ratios. As seen in \autoref{fig:questions_emmeans:rq_topicpath}, the Positive path increased the RQ turn ratio only in the Work ($p=.02$) and Exercise ($p=.003$) dialogues and also decreased the SQ turn ratio in the same topics (Work: $p=.003$ and Exercise: $p<.001$).
As discussed above, prompt designs, topics and conversation paths have significant impacts on chatbots' question-asking behaviors. The RQ and SQ ratios further impacted the slot filling rate. We ran the maximum-likelihood tests with two mixed-effect models fitting the slot filling rate, one with the chatbot-spoken RQ turn ratio (\textit{i.e.}, the ratio of the turns classified as RQ in a dialogue) as a fixed effect and the other with the SQ turn ratio, both with participants as a random effect. We found that the RQ turn ratio was positively correlated with the slot filling rate, whereas the SQ turn ratio was negatively correlated with it: $\beta=1.08$, $SE=0.12$, $t(347.25)=8.80$, $p < .0001$ for RQ and $\beta=-0.77$, $SE=0.13$, $t(370.06)=-6.12$, $p < .0001$ for SQ.
\begin{figure*}[t]
\begin{flushleft}
\sffamily\footnotesize{\textbf{***}$p$<.001; \textbf{**}$p$<.01; \textbf{*}$p$<.05}
\end{flushleft}
\centering
\includegraphics[width=\textwidth]{img/fig_questions_emmeans.pdf}
\labelphantom{fig:questions_emmeans:rq_format}
\labelphantom{fig:questions_emmeans:rq_condition}
\labelphantom{fig:questions_emmeans:rq_path}
\labelphantom{fig:questions_emmeans:rq_topicpath}
\labelphantom{fig:questions_emmeans:sq_format}
\labelphantom{fig:questions_emmeans:sq_condition}
\labelphantom{fig:questions_emmeans:sq_path}
\labelphantom{fig:questions_emmeans:sq_topicpath}
\begin{flushright}
\footnotesize{$^1$Excluded pairs across different topics from the pairwise comparison.}
\end{flushright}
\caption{95\% confidence intervals of the turn ratios of RQ (top; a--d) and SQ (bottom; e--h) by variables with a significant effect: The asterisks with arms indicate significance between the connected categories. Note that for (d) and (h) we did not display the significance across topics. (Refer to Appendices \ref{appendix:stats:rq} and \ref{appendix:stats:sq} for model details and statistics.)}
\label{fig:questions_emmeans}
\end{figure*}
In summary, our results suggest that chatbots with a Descriptive information format tend to ask more secondary questions, and negative answers of participants also naturally elicit more secondary questions (\textit{e.g.}, \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {}~``\textit{I'm sorry to hear that you didn't workout yesterday. May I ask why?}'').
\subsubsection{Non-question Statements}
Statements from chatbots usually consist of chitchats, comments and suggestions that do not fall within the question/answering (see \autoref{tab:essential_components} for references) and participants also tend to respond to statements with other statements. Usually, the participants' statements did not contain direct clues about the desired information slots. Overall, 1,482 out of 7,442 turns ($avg.$ 18.45\% per dialogue) were classified as Statement (See \autoref{tab:essential_components}). Among the four conditions, dialogues of DP have the highest ratio of statements (24.99\%) and SP had the lowest ratio (15.32\%).
\subsection{Empathy \& Engagement Behaviors}\label{sec:results:empathy}
\input{tables/acts-empathy}
\begin{figure*}[t]
\begin{flushleft}
\sffamily\footnotesize{\textbf{***}$p$<.001; \textbf{**}$p$<.01; \textbf{*}$p$<.05}
\end{flushleft}
\centering
\includegraphics[width=0.9\textwidth]{img/fig_empathy_emmeans.pdf}
\labelphantom{fig:empathy_emmeans:ack:format}
\labelphantom{fig:empathy_emmeans:ack:personality}
\labelphantom{fig:empathy_emmeans:ack:condition}
\labelphantom{fig:empathy_emmeans:ack:path}
\labelphantom{fig:empathy_emmeans:appr:format}
\labelphantom{fig:empathy_emmeans:appr:personality}
\labelphantom{fig:empathy_emmeans:appr:condition}
\labelphantom{fig:empathy_emmeans:appr:path}
\labelphantom{fig:empathy_emmeans:symp:format}
\labelphantom{fig:empathy_emmeans:symp:personality}
\labelphantom{fig:empathy_emmeans:symp:condition}
\labelphantom{fig:empathy_emmeans:symp:path}
\caption{95\% confidence intervals of the chatbot turn ratios for Acknowledging (a--d), Appreciating (e--h), and Sympathizing (i--l) behaviors by information format, personality modifier, study condition (combinations of format and personality modifier), and conversation path. Variables that are not significant are marked as `NS.' (Refer to Appendix \ref{appendix:stats:empathy} for model details and statistics.)}
\label{fig:empathy_emmeans}
\end{figure*}
\autoref{tab:empathy} summarizes the empathy \& engagement categories and their turn ratios by experimental condition. The majority of these behavior categories were coded to the chatbot turns---1,992 chatbot turns and 403 user turns were coded with one or more behavior categories---partly because participants uttered less words than chatbots (see \autoref{tab:descriptive_stats}) and chatbots usually led the conversation while participants simply responded. \textbf{Acknowledging} was the most common empathy behavior (see \autoref{tab:empathy}) as chatbots often referred to what participants previously said in generated messages. Also, we found that our chatbots often \textbf{appreciated} participants' accomplishment (\textit{e.g.}, taking good sleep, managed to exercise) or \textbf{sympathized} participants when they reported negative outcomes (\textit{e.g.}, poor sleep quality, failed at work).
To investigate how the four study factors impact the empathetic behaviors of chatbots, we analyzed three mixed-effect models with the chatbot turn ratios of Acknowledging, Appreciating, and Sympathizing behaviors as a dependent variable, respectively.
\autoref{fig:empathy_emmeans} shows the 95\% confidence intervals of turn ratios of the three behavior categories estimated against the study factors. The information format significantly influenced the Acknowledging and Appreciating turn ratios: Dialogues in the Descriptive format had higher ratios of the Acknowledging ($p<.0001$; see \autoref{fig:empathy_emmeans:ack:format}) and Appreciating ($p=.01$; see \autoref{fig:empathy_emmeans:appr:format}) turns. Personality modifier did not solely impact these two behaviors but it influenced in conjunction with the information format (See \autoref{fig:empathy_emmeans:ack:condition} and \ref{fig:empathy_emmeans:appr:condition}). However, the personality modifier in the prompt led chatbots to produce significantly more Sympathizing turns ($p=.002$; see \autoref{fig:empathy_emmeans:symp:personality}). Besides the prompt design, the conversation path strongly influenced all three empathetic behaviors: The Positive path led to higher turn ratio of Appreciating ($p<.0001$; see \autoref{fig:empathy_emmeans:ack:path}) whereas The Negative path led to higher Acknowledging ($p<.001$; see \autoref{fig:empathy_emmeans:appr:path}) and Sympathizing ($p<.0001$; see \autoref{fig:empathy_emmeans:symp:path}) turn ratios.
\subsection{Problematic Chatbot Turns and User Responses}
\label{sec:results:errors}
\input{tables/error-category.tex}
In total, 6.7\% of the chatbot turns (257 out of 3,916) were tagged erroneous and the four categories of erroneous turns are summarized in \autoref{tab:errors}. These erroneous turns sometimes led to the unorganic termination of the conversation (\textit{i.e.}, participants ended the conversation before or without natural Closing messages). In the following, we cover these error types in detail.
\ipstart{Incorrect phrases} In structured prompts (SP, SN), we used a symbol ``\texttt{->},'' a commonly-used delimiter for key-value pairs in LLMs, to specify the information slots. This caused GPT-3 to expose such an information structure to the output as an artifact in 16 turns of messages, all of which were generated by the structured format. In the example from~\autoref{tab:errors}, GPT-3 even incorrectly ``predicted'' the slots (45 minutes, example (1) in~\autoref{tab:errors}) altogether with symbols. Other times, GPT-3 also erroneously predicted the answers for participants (example (2)~\autoref{tab:errors}). Particularly, we identified 6 instances where GPT-3 predicted the user response and appended an extra \raisebox{-2.1pt}{\includegraphics[width=9.5pt]{img/icon_human.pdf}} {}turn to the generated chatbot turns (3 turns in descriptive groups and 3 turns in structured groups).
\ipstart{Self-talk} GPT-3 sometimes generated turns in a first-person narrative or not directed to participants, which looked quite similar to the ``self-talk'' of humans. In \autoref{tab:self_error}, for example, Turn 03 is obviously not directed to participants. In our dataset, less than 1.1\% of chatbot turns were self-talk (SP: 8, SN: 27, DP: 1, DN: 9). Participants who encountered these ``self-talk'' commented that these messages were ``awkward'' (P42), ``strange'' (P24), and ``confusing''(P23). However, we found that participants always attempted to continue the conversation by following the self-talk and tried to resolve the errors (See Turn 04 in~\autoref{tab:self_error}).
\begin{fragment}
\def1.3{1.1}\setlength{\tabcolsep}{0.35em}
\small\sffamily\centering
\caption{Dialogue snippets of self-talk from P43.}
\label{tab:self_error}
\begin{tabularx}{\columnwidth}{lll}
\hline
01 &
\raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {} & \textit{Do you have any fitness concerns that you would like to address?} \\
\arrayrulecolor{tablegrayline}\hline
02 &
\raisebox{-2.1pt}{\includegraphics[width=9.5pt]{img/icon_human.pdf}} {} &
\textit{nope not at the moment.} \\\hline
03 &
\raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {} &
\begin{tabular}[c]{@{}l@{}}\textcolor{magenta}{\textit{\textbf{Based on the customer's responses, it seems that they}}}\\\textcolor{magenta}{\textbf{\textit{did not have a very active day yesterday. I would}}} \\ \textcolor{magenta}{\textit{\textbf{encourage them to try and get some form of physical}}}\\ \textcolor{magenta}{\textit{\textbf{activity in today, even if it is just a short walk.}}}\end{tabular} \\\hline
04 &
\raisebox{-2.1pt}{\includegraphics[width=9.5pt]{img/icon_human.pdf}} {} &
\textit{yes I should probably try it tomorrow.}
\\\arrayrulecolor{black}\hline
\end{tabularx}
\end{fragment}
\ipstart{Repetition}
We found that GPT-3 was susceptible to generate repetitive messages, either identically or linguistically repeating the previous chatbot turns. In total, 147 turns (3.8\% of chatbot turns) were labeled to be repetitive (SP: 11, SN: 39, DP: 48, DN: 50). Identically repetitive messages occurred to 23 participants (SP: 3, SN:5, DP: 6, DN: 9) in 31 dialogues (SP: 3, SN: 9, DP: 8, DN:10). Among the four topics, work (14 dialogues) tended to have more identically repetitive messages and exercise (2 dialogues) tended have fewer identically repetitive messages. However, these messages usually served as SQ/SA in conversations, hence they rarely influenced slot filling and data collection. Linguistically repetitive messages usually share similar wording or phrase structures. \autoref{tab:repetition_eg} presents an example dialogue snippet. At Turn 01, 03, 05, and 07, the chatbot always started with a similar phrase (\textit{i.e.}, \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {}``\textit{That's great ...}'') to compliment the participant then asked the participant a question started with ``\textit{can you}''. These linguistically repetitive messages were semantically correct and \textit{58.3\%} took place in the Acknowledging turns where the chatbot rephrased what the participant said progressed the conversation organically. However, linguistically phrased messages looked too similar, and they negatively impacted the user experience. A few participants suggested that the chatbots seemed to actually understand their responses, yet were using ``a sentence template'' (P29) to respond in a ``predefined ways'' (P13). On the other hand, repetition could also lead to ``dead loops'' of conversations.
\begin{fragment}
\def1.3{1.0}\setlength{\tabcolsep}{0.35em}
\small\sffamily\centering
\caption{Dialogue snippet of linguistically repetitive utterances from P37.}
\begin{tabularx}{\columnwidth}{lll}
\hline
01 & \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {} & \begin{tabular}[c]{@{}l@{}}\textit{That's a specific and achievable goal. Can you tell me more}\\\textit{about what libraries you need to install?}\end{tabular} \\\arrayrulecolor{tablegrayline}\hline
02 & \raisebox{-2.1pt}{\includegraphics[width=9.5pt]{img/icon_human.pdf}} {} & \textit{I don't know yet which libraries I need, but probably Pytorch, ...} \\\hline
03 & \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {} & \begin{tabular}[c]{@{}l@{}}\textit{\textcolor{magenta}{\textbf{That's a great goal. Can you}} research what libraries you need}\\\textit{and then install them?}\end{tabular} \\\hline
04 & \raisebox{-2.1pt}{\includegraphics[width=9.5pt]{img/icon_human.pdf}} {} & \textit{The libraries I need is provided in a file.} \\\hline
05 & \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {} & \begin{tabular}[c]{@{}l@{}}\textit{\textcolor{magenta}{\textbf{That's great.}} So you have a list of the libraries you need.}\\\textit{\textcolor{magenta}{\textbf{Can you install them?}}}\end{tabular} \\\hline
06 & \raisebox{-2.1pt}{\includegraphics[width=9.5pt]{img/icon_human.pdf}} {} & \textit{I need to run a script that download and install these libraries.} \\\hline
07 & \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {} & \textit{\textcolor{magenta}{\textbf{That's a great goal. Can you}} walk me through the script?} \\\arrayrulecolor{black}\hline
\end{tabularx}
\label{tab:repetition_eg}
\end{fragment}
\ipstart{Miscellaneous} There were 37 problematic chatbot turns that did not fall into the above categories. Among these turns, 19 turns were tagged to have context errors (\textit{i.e.}, chatbots did not grasp the context at all), 10 had semantic errors (\textit{i.e.}, messages that are not of human common sense), and one had both context and semantic errors. In \autoref{tab:context_error}, for example, the chatbot entirely missed that the participant said \raisebox{-2.1pt}{\includegraphics[width=9.5pt]{img/icon_human.pdf}} {}``\textit{\textbf{1 hour} of cardio}.'' But the chatbot also attempted to resolve the contextual misunderstanding by apologizing after the participant corrected it. While chatbots in our study appeared to understand people pretty well in most cases, their responses with semantic errors could be quite wrong and amusing. For instance, when being asked about workout yesterday, one participant told the chatbot that they skipped. Instead of considering the participant did not workout, the chatbot considered skipping as a jumping workout and responded \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {}``\textit{Skipping is a type of cardiovascular exercise that can help to improve your heart health and endurance.}'' Besides, there were two instances when the chatbot failed to detect the ending of the conversation and restarted with the first slot question again, which, of course, led participants to abandon the conversation. Lastly, our system went offline 5 times and caused chatbots to output empty messages, which was caused by the over-frequent API calls to OpenAI.
\ipstart{Terminating Conversations}
With the Closing turns, we found that 75.7\% of the conversations were organically ended. Among the four conditions, SP had the lowest percentage of naturally ended conversations (71.4\%) than other three conditions. For 91 non-organically ended conversations, participants abandoned 48.3\% of those conversations without encountering any obvious problematic errors. 19.8\% of prematurely ended conversations were caused by identically repetitive and another 19.8\% of were caused by linguistically repetitive messages. Lastly, the self-talk led to 3 conversations to end early, and context errors and system-offline caused the rest incomplete conversations.
\begin{fragment}
\small\sffamily\centering
\caption{Dialogue snippet of the context error from P6.}
\label{tab:context_error}
\begin{tabularx}{\columnwidth}{lll}
\hline
01 & \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {} & \textit{What workout did you do yesterday?} \\\arrayrulecolor{tablegrayline}\hline
02 & \raisebox{-2.1pt}{\includegraphics[width=9.5pt]{img/icon_human.pdf}} {} & \textit{I did \textbf{1 hour} of cardio yesterday.} \\\hline
03 & \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {} & \textit{\textcolor{magenta}{\textbf{How long did your workout last?}}} \\\hline
04 & \raisebox{-2.1pt}{\includegraphics[width=9.5pt]{img/icon_human.pdf}} {} & \begin{tabular}[c]{@{}l@{}}\textit{Come on dude. I just told you I did 1 hour of cardio.}\\\textit{How long is 1 hour?}\end{tabular}
\\\hline
05 & \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {} & \begin{tabular}[c]{@{}l@{}}\textit{I'm sorry if I came across as rude. I was just trying to get}\\\textit{some information from you about your workout yesterday.}\end{tabular} \\\arrayrulecolor{black}\hline
\end{tabularx}
\end{fragment}
\begin{figure*}[b]
\centering
\includegraphics[width = \textwidth]{img/ratings.pdf} \caption{Distributions of the subjective ratings from three scale questions in the exit survey, with breakdowns by the information format and the personality modifier.}
\label{fig:subjective_evaluation}
\end{figure*}
\subsection{Subjective Evaluation}
\label{sec:results:subjective}
\autoref{fig:subjective_evaluation} summarizes the distribution of participants for rating (1) the ability to understand, (2) the ability to acknowledge user input, and (3) the level of empathy. The Kruskal-Wallis tests showed that there were no significant differences among conditions for all three questions. In general, most participants highly rated for Q1 and Q2 that the chatbots could understand them as well as acknowledge their messages: Fifteen (31.3\%) participants rated 5 and 24 (50.0\%) rated 4 on Q1; 11 (22.9\%) rated 5 and 31 (64.6\%) rated 4 on Q2. Participants showed mixed perception for the level of empathy question with a median of 3.
Some participants gave positive feedback in the open-ended question. P25 who frequently used LLMs commented, ``\textit{I was surprised to see how accurate and detailed the bot's responses were.}'' P36 who did not have any experience with LLMs gave a similar comment---``\textit{I found it quite responsive and surprisingly considerate of my answers.}'' Despite the errors we presented above, P21 still complimented the chatbots: ``\textit{I liked/was satisfied of how the chatbot precisely gave info when I asked for it, and I felt that the relevance of the answer is very high and that it caught my point of question sharply.}'' and even suggested that, ``\textit{I felt keeping chatbot as a companion would be awesome. To regularly make casual conversation and be provided light insights about my daily life based on my casual chats.}''
\section{Discussion}
Our results showed that our zero-shot chatbots achieved great abilities in asking desired questions and understanding user responses, despite also having drawbacks. Here, we reflect on the performance of our chatbots and discuss opportunities, ethical considerations, and limitations for future work.
\subsection{Designing Effective Prompts for Chatbots that Collect Self-Reports}
Our study showed that defining a first-person job identity as well as specifying information slots in prompts was an effective method to bootstrap chatbots that ask health-related questions. However, the slot filling performance and chatbots' behaviors were sensitive to prompt design, topic, and conversation paths. We provide the following prompt design suggestions based on our findings:
\ipstart{Combine Information Format and Personality Modifier Wisely}
Although the information format and personality modifier did not consistently impact slot filling rates individually, how they were combined had different impacts on slot filling rates. The information format affected chatbots' question-asking behaviors: structured formats lead to more RQs and fewer SQs and vice versa for descriptive ones. In other words, structured formats steer chatbots to ask direct questions about the specified slots whereas descriptive ones focus more on eliciting surrounding context or additional details.
The personality modifier had a synergy with descriptive formats: Chatbots in DP had the lowest slot filling rates and the ratio of RQs, but had the highest number of acknowledging messages. Referring to ~\autoref{tab:completion_rate}, SP and DN have comparable slot filling rates.
Therefore, to build chatbots that can show a higher level of understanding through acknowledgment, using the Descriptive format without personality modifier could be the best option. But chatbots with more direct acknowledgment may be at the risk of being awkward and too robotic. Hence, when designing prompts to power chatbots for data collection, using structured format with personality modifier would be more desirable.
\ipstart{Evaluate chatbot for conversation topic and path}
There is certain discrepancy of slot filling rates between topics. One reason for such difference could be the nature of the topic. For example, the topic \textit{work} tended to be the most open-ended topic as people report different types of work, which could lead to more subject switches and digressions than others. We suspect that GPT-3 is more suitable to steer chatbots that are of less divergent topics and collect self-reports that are more structured. Also, considering the path of conversations also impacts the data collection rate, researchers may consider clearly specify different slots for both positive and negative paths. For example, developers can add ``\texttt{if the customer did not workout yesterday, I would ask them what workout they did in the past week}'' to the prompt for chatbots in the exercise topic.
\ipstart{Composition of Slots Matters}
The number and data types of slots also impact the chatbots' performance in collecting slots. With the conversation goes longer, chatbots have a tendency to miss information slots that appear later in prompts. Also, while we collected a certain amount of numerical rates for sleep quality and productivity rate, it is not guaranteed that the chatbots would cover a slot definition (\textit{e.g.}, numerical scale) as intended. Sometimes, the chatbot would simply ask \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {}``\textit{Would you say you had a good night's sleep?}'' or \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {}``\textit{Overall, how do you feel about your work and productivity yesterday?}''. Hence, if to use chatbots powered by GPT-3 to facilitate data collection, we suggest that important slots be put earlier in the prompt and the number of questions of specific data type be limited. If more data slots need to be collected, multi-stage prompts~\cite{wu2022ai} can be considered.
\subsection{Opportunities of LLM-driven Chatbots}
From the study, we learned that LLM-driven chatbots are advantageous compared with traditional chatbot platforms in multiple aspects. Here we cover some noteworthy aspects drawing on the results.
\ipstart{Versatile Responses and Follow-up Questions}
Compared to chatbots with pre-defined dialogues, chatbots in our study can deliver a great number of versatile phrases. For example, for these scale questions, GPT-3 can output phrases such as \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {}``\textit{Would you say that your sleep quality yesterday was a 10/10, 9/10, 8/10...?}'' GPT-3 can even provide clarifications for questions and ask follow-up questions which supplement the topic. However, these SQs were still on-topic and directly addressed to user inputs (See SQ/SA in~\autoref{tab:essential_components}). Follow-up questions are commonly used in human-administered interviews to increase interactivity~\cite{roulston2018qualitative} and many studies suggest that chatbots that can ask on-topic follow-up questions are considered more human-like~\cite{luger2016like, svenningsson2019artificial}. Although current chatbot frameworks (\textit{e.g.}, Amazon Alexa~\cite{AmazonAlexa})~\cite{mctear2018conversational} support follow-up/extended questions, developers need to specify both the expected slots and the follow-up phrases~\cite{schuetzler2018investigation}. On the contrary, GPT-3 could naturally ask follow-up questions, equipping chatbots with proper common sense on the topic. For example, our chatbot mapped ``\textit{Bulgogi, rice, and kimchi}'' to ``\textit{a very traditional Korean meal}'' in its response to the participant. Such response engages people through showing a level of ``understanding.''
\ipstart{Social Attributes}
Given the importance of social features such as chit-chat for positive user experience~\cite{serenko2008model, liao2018all}, our results show that we can easily equip GPT-3 with such social aspects. For example, our chatbots could respond naturally to the questions about their ``personal life''---\textit{e.g.}, \raisebox{-2.1pt}{\includegraphics[width=9.5pt]{img/icon_human.pdf}} {}``\textit{Do you workout yourself?}'' \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {}``\textit{Yes, I work out regularly myself. I find that it helps me to stay energized and focused throughout the day.}''
Further, our chatbots were also able to give suggestions relevant to the topic. While mostly originated from common sense,
some of the suggestions were in-depth and tailored. In one time, one participant asked two questions in a row (probably due to system error) were quite surprised to find that the chatbot provided a well-written response (See Turn 04 in Fragment~\ref{tab:advice}). This participant even commented that ``\textit{I know a small bit about NLP but not much when it comes to generating responses. I find it fascinating that (it) can give such in-depth answers to specific topics as I find it hard to be able to train an AI to every kind of case involving that.}''
\ipstart{Error Recovery}
Task-oriented chatbots usually have limited number of pre-defined user intents to accomplish a specific goal. For instance, a banking chatbot can provide services such as currency-exchange conversion and introduction of credit cards~\cite{li2020conversation}. However, such chatbots are usually unable to handle user messages that are out of the pre-defined intents (\textit{e.g.}, a user attempts to have small talk with the banking chatbot)~\cite{li2020conversation}. Also, they may even mis-recognize in-scope messages due to the complexity of natural language~\cite{ashktorab2019resilient}. Strategies like highlighting keywords and switching topics~\cite{ashktorab2019resilient} can help resolve conversation breakdowns at the price of making chatbots less human-like. In our case, LLM-driven chatbots could handle the out-of-scope conversations relatively well, since they could do improvisation actions relying on the ability of LLMs instead of defining intents intensively. In \autoref{tab:out_of_context}, a work chatbot with the job identity ``life coach'' handled the off-topic request (``\textit{wake me up at 6 am}'') by the participant smoothly and even provided tips on sleep. Even when misunderstanding occurred, chatbots sometimes attempted to resolve it. In \autoref{tab:context_error}, for example, the chatbot apologized for its misunderstanding and in\ref{tab:error_recovery}, GPT-3 resolved an empty message error (due to system offline) by making up an excuse for its absence.
\begin{fragment}
\small\sffamily\centering
\caption{Dialogue snippet of the bot continuing off-topic conversations from P34.}
\label{tab:out_of_context}
\begin{tabularx}{\columnwidth}{lll}
\hline
01 & \raisebox{-2.1pt}{\includegraphics[width=9.5pt]{img/icon_human.pdf}} {} & \textit{\textcolor{magenta}{\textbf{Can you wake me up at 6 am?}}} \\\arrayrulecolor{tablegrayline}\hline
02 & \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {} & \begin{tabular}[c]{@{}l@{}}\textit{\textcolor{magenta}{\textbf{I'm sorry, but I cannot physically wake you up at 6 am.}}}\\
\textit{\textcolor{magenta}{\textbf{However, I can give you some tips on how to wake up}}}
\\\textcolor{magenta}{\textbf{\textit{earlier and be productive in the morning}}}.\end{tabular} \\\hline
03 & \raisebox{-2.1pt}{\includegraphics[width=9.5pt]{img/icon_human.pdf}} {} & \textit{Yeah, please.} \\\hline
04 & \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {} & \begin{tabular}[c]{@{}l@{}}\textit{Some possible tips to help you wake up earlier in the morning}\\\textit{are to set your alarm clock for a specific time and place it}\\\textit{across the room ... [omitted]}\end{tabular}\\\arrayrulecolor{black}\hline
\end{tabularx}
\end{fragment}
\begin{fragment}
\small\sffamily\centering
\caption{Dialogue snippet of an error recovery from P2.}
\label{tab:error_recovery}
\begin{tabularx}{\columnwidth}{lll}
\hline
01 & \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {} & \textit{That's great! What time did you wake up this morning?} \\\arrayrulecolor{tablegrayline}\hline
02 & \raisebox{-2.1pt}{\includegraphics[width=9.5pt]{img/icon_human.pdf}} {} & \textit{around 9 am.} \\\hline
03 & \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {} & \textit{\textcolor{magenta}{\textbf{(...no response...) }}} \\\hline
04 & \raisebox{-2.1pt}{\includegraphics[width=9.5pt]{img/icon_human.pdf}} {} & \textit{hello?}
\\\hline
05 & \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {} & \begin{tabular}[c]{@{}l@{}}\textit{Sorry, I'm just checking to see if you're still there.}\\\textit{Yeah, so 9 am is a pretty good wake up time.} \\ \textit{How would you rate your sleep quality from last night?}\end{tabular} \\\arrayrulecolor{black}\hline
\end{tabularx}
\end{fragment}
\ipstart{Context Tracking}
Context is a key part in human conversations that connects multiple turns~\cite{brennan1990conversation}. Previous studies have suggested that chatbots should aim to sustain contexts to improve the dialogue efficiency~\cite{jain2018evaluating}. Current conversational interfaces such as Google Assistant and Amazon Alexa shows certain abilities in maintaining contexts~\cite{mctear2018conversational}; however, most of them are still criticized for not detecting contextual details~\cite{beneteau2019communication, cho2020role}. With Dialogflow, developers can define some contexts to be maintained within 5 turns\footnote{\url{https://cloud.google.com/dialogflow/es/docs/contexts-input-output}}; however, it has yet achieved truly flexible conversations through this approach. In our case, the chatbots have shown impressive abilities in sustaining some contexts without dedicated mechanisms for managing contexts. P27 noted, ``\textit{I feel like it could keep track of the context well between sentences during the conversation.}'' Through dialogue snippets presented in \autoref{tab:error_recovery} and ~\ref{tab:advice}, we can see that the context was maintained across 5 turns and 2 turns, respectively. In particular, the context (\textit{intimidation}) would be difficult to specify with most chatbot frameworks. Further, one pattern that emerged in the dataset is that chatbots liked to give a summary of all the user input in the end of conversations, which usually covered the past conversation history and maintained contexts longer than 5 turns.
\begin{fragment}
\small\sffamily\centering
\caption{Dialogue snippet of the chatbot tracking context from P33.}
\label{tab:advice}
\begin{tabularx}{\columnwidth}{lll}
\hline
01 & \raisebox{-2.1pt}{\includegraphics[width=9.5pt]{img/icon_human.pdf}} {} & \textit{Those are good. How about intimidation at the gym?} \\\arrayrulecolor{tablegrayline}\hline
02 & \raisebox{-2.1pt}{\includegraphics[width=9.5pt]{img/icon_human.pdf}} {} & \textit{Oh sorry, was that too vague?} \\\hline
03 & \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {} & \begin{tabular}[c]{@{}l@{}}\textit{No, that's a great question. A lot of people feel intimidated} \\ \textit{when they first start going to the gym. My best advice would} \\ \textit{be to just take things slow and focus on your own workout.}\\\textit{Don't worry about what other people are doing or how they're}\\\textit{looking at you. Just focus on yourself and your own goals.} \end{tabular} \\\hline
04 & \raisebox{-2.1pt}{\includegraphics[width=9.5pt]{img/icon_human.pdf}} {} & \begin{tabular}[c]{@{}l@{}}\textit{That\'s solid advice. You're surprisingly eloquent}\\\textit{for these responses!}\end{tabular}\\\arrayrulecolor{black}\hline
\end{tabularx}
\end{fragment}
\ipstart{Low-effort Bootstrapping}
We show several opportunities of LLMs in powering chatbots above. Indeed, chatbots that provide many of the above functions, including chitchats, suggestions, and context perseverance, can be trained with rich datasets. However, collecting such dataset is challenging, and training models on big datasets is costly and often inaccessible~\cite{bae2022building}. In terms of utilizing mainstream chatbot platforms to build voice applications, it is of great human effort to come up with user intents and example phrases as well as design conversation flows and logic~\cite{dingler2021use, luo2021foodscrap}. In particular, the error handling is tricky to design as conversation breakdowns can be common and even unexpected in the wild~\cite{cho2020role, myers2018patterns, wei2022could}. On the other hand, LLM's in-context learning capability enables us to skip collecting training dataset or configuring dialogue flows to create functional chatbots. Further, our results show that simple alterations of prompts can significantly influence the conversation styles of chatbots. With robust prompt designs, it is possible that people without background in AI can directly personalize chatbots using natural language.
\subsection{Drawbacks of LLM-driven Chatbots}
Although LLMs showed great potentials in steering chatbots, we also encountered several drawbacks of LLM-driven chatbots. Here we cover the two noteworthy drawbacks and strategies to overcome them.
\ipstart{Repetitiveness}
As LLMs tend to detect latent patterns in the prompts~\cite{brown2020language}, the user messages accumulated in the prompt (See \circledigit{B} in \autoref{fig:teaser}) may unintentionally trigger patternized behaviors, making chatbots produce repetitive (although not always identical) turns. For example, many of the messages generated by GPT-3 start with ``\textit{It sounds like you...}'', \textit{parroting} user responses and providing direct acknowledgment.
Although such behaviors made many participants rate the chatbots to be ``understanding'', participants also criticized the awkwardness of parroting. We suspect that such repetition was partially caused by a well-known problem of LLMs: they tend to generate repetitive messages ~\cite{Welleck2020Unlikelihood}. In worse cases, chatbots stuck in ``dead loops'' and could not progress the conversation further. P20 even responded ``please enough'' to the chatbot's repetitive questions.
\ipstart{Randomness}
LLMs predict the most probable text followed by prompts. Therefore, the generations inherently exhibit a certain level of randomness~\cite{liu2022design}, which is hard to explain or anticipate.
Such randomness might have led to erroneous responses of our chatbots. For example, sometimes the chatbots ``self-talked''~\cite{shuster2021retrieval} or exposed machine representations as responses. What makes it worse is that as the conversation history is accumulated, erroneous responses stay in the prompt and lead to other erroneous ones.
Also, the chatbots sometimes did not consistently react to the same user input. When being told the participant skipped breakfast, one chatbot under the condition DN responded, \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {}``\textit{That's not ideal. Skipping breakfast can make it harder to concentrate and can cause you to overeat later in the day},'' while the other chatbot under the same condition replied, \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {}``\textit{That's okay! Some people choose to skip breakfast}''. It is not explainable whether such inconsistencies randomly happened or were caused by prior user inputs.
The stochastic nature of LLMs does not guarantee that they would comply with all natural language instructions in prompts. As such, compared to rule-based chatbots that can almost 100\% ask pre-defined scripts~\cite{wei2022understanding}, we can see that not all specified information slots were asked by our chatbots during the study.
Despite drawbacks discussed above, LLMs-based chatbots can become a valuable and scalable tool for researchers to collect data for personal informatics~\cite{homewood2020removal}. Reflecting on our findings, we propose strategies to mitigate the erroneous behaviors of the chatbots.
GPT-3 tends to generate long responses, which may make chatbots to appear more robot-like. We suggest that researchers consider intentionally slowing down the responding delays. A longer gap may not only help create a more human-like chatbot~\cite{gnewuch2018faster} but also create time for the system to run filters and algorithms to pick better messages.
Drawing on problems identified from our analysis, we envision a chatbot system that generates three responses each turn (if the budget allows). Then, a repetition filter can be used to filter out identically repetitive messages. In terms of linguistically repetitive messages, the system can pick the message with the least linguistic similarity to the chatbot's last turn. The filter could also easily remove messages that have the self-talk errors or symbols. When the conversation is too long, a parallel prompt can be made to detect if the conversation is in a dead loop or a simple ending detection algorithm can deployed to end the conversation and improve the user experience. All these filters are cost-efficient to implement and could resolve many problems. For example, around 80\% of errors occurred in SP are repetitive messages, self-talk, and system-offline, all of which could potentially be resolved with simple filters. Lastly, we acknowledge that running LLMs is always accompanied with uncertainty and the resultant chatbots may not be able to fulfill defined tasks every time.
Hence, we recommend researchers conducting intensive testing of LLMs-powered chatbots to identify errors, understand the slot filling performance, and customize filters accordingly.
\subsection{Ethical consideration}
LLMs are trained on an existing corpus that may contain biased and problematic information~\cite{gehman2020realtoxicityprompts, park2022socialsimulacra}.
Many have also suggested that cautions should be taken when using LLMs, particularly in the field of healthcare delivery~\cite{korngiebel2021considering}. In our study, we intentionally used hints to guide participants to compose their answers when conversing as we were unsure whether inappropriate content would be generated. We did not see any biased, harmful or dangerous messages from GPT-3 in our dataset. All the chatbots appeared to give conservative suggestions. For example, one participant tried to ask diet suggestions for weight loss, but the chatbot with the job identity as a \textit{fitness coach} suggested that ``it's always best to speak with a doctor or registered dietitian before starting any weight loss plan.'' However, we also found some instances where chatbots failed to detect participants' ``teasing and nonsensical'' questions and gave advice that could potentially be dangerous to follow. For example, to a participant who said \raisebox{-2.1pt}{\includegraphics[width=9.5pt]{img/icon_human.pdf}} {}``\textit{I want to gain 50 kg of pure fat by the end of the year. How many snicker bars should I eat to complete that goal?}'', the chatbot responded with a semantically problematic message: \raisebox{-2.3pt}{\includegraphics[width=10pt]{img/icon_agent.pdf}} {}``\textit{If you're trying to eat 3000 calories a day and you're only burning 2000, then eating 3 snickers bars a day (each bar has around 1000 calories) could help you reach your goal}.'' This message not only contains the incorrect fact (\textit{i.e.}, the calories of a snicker bar) but also is irrational.
This example suggests the importance of giving precautions to users that the chatbots' messages do not guarantee medical or professional soundness~\cite{korngiebel2021considering}.
\subsection{Study Limitation}
Due to the limited number of participants, we did not perfectly counterbalance the order of topics. Fatigue effects may not be fully mitigated for Food and Sleep topics which always came after Work and Exercise, respectively. Similarly, participants always conversed in the Positive conversation path before the Negative one, although we believe that having consistent path orders would cause less confusion and mistakes. Also, while we instructed participants to follow given conversation paths, some participants might not perfectly comply with the guides, possibly affecting significance of the pairwise comparisons.
The targeted information slots consisted of only time, scale, binary, and open-ended data types. Incorporating other types of questions such as multiple choices may influence the chatbots' performance. In addition, slots in each topic had different composition of data types, so any differences among topics might be influenced by both the lexicon of the topic and the composition of data types.
Also, we did not control the conversation style of participants. Since user inputs also become part of the prompts, their linguistic patterns may affect GPT-3's generations and in turn the slot filling performance or the conversation style of chatbots itself.
We chose GPT-3 as the underlying LLM for our chatbots as it is mainstream and publicly accessible via a commercial API. Although the model we used shows overall state-of-the-art performance in accuracy, robustness, and fairness (\textit{c.f.}, ~\cite{Liang2022holisticlm}), given that LLMs can be sensitive to prompt designs~\cite{Liu2021PretrainPromptAndPredict}, we reckon that our proposed prompts may not yield similar performance on other LLMs due to the differences in the training corpora and the model architecture. For example, newer LLMs that are improved to follow instructions in a prompt~(\textit{e.g.}, \texttt{text-davinci-003}~\cite{davinci003news}) or optimized for dialogues (\textit{e.g.}, ChatGPT~\cite{chatgpt}) may be more diligent in filling slots. Therefore, future work may consider powering chatbots on other LLMs, with our proposed prompts as a starting point.
\subsection{Future Work}
Future work can explore ways to improve the performance of LLM-driven chatbots. In our study, we adopted zero-shot prompts. Researchers can try augmenting our prompts with few-shot learning by providing example dialogues~\cite{brown2020language}, which may make chatbots have more robust question-asking abilities and can handle negative paths better~\cite{bae2022building}.
Measuring chatbots requires great human efforts so more future research into the effects of these parameters on prompts is needed to provide guidance for the development of better and more robust chatbots. Researchers can also investigate multi-stage prompting~\cite{wu2022ai, wei2022chain} (\textit{i.e.}, designing several prompts for different questions in one dialogue session) if they intend to collect more than 5 slots of information. Such approaches will require incorporating dialog state tracking techniques (\textit{e.g.},~\cite{lin2021zero}) for automated slot filling. Lastly, we hope future research can investigate the user perceptions of LLM-driven chatbots, or even voice-based ones like smart speakers~\cite{wei2021understanding}.
In this study, we focused exploring the chatbots' performance and behaviors rather than the user experience. Several participants were impressed by some of the chatbots' responses but were disappointed with repetitive messages at the same time. Hence, we are interested in seeing how people will interact with an improved version of our chatbots and whether their mental models of chatbots will change along with the advancement of chatbots~\cite{liao2018all}. In addition, comparing user perception of LLM-driven chatbots with other mainstream chatbot frameworks (\textit{c.f.}, \cite{mitchell2022examining}) would provide holistic design implications for self-reporting chatbots with balanced data collection performance and user perception.
\section{Conclusion}
In this study, we explored how we can use GPT-3 for powering chatbots that can reliably ask people health-related questions through natural conversations. In an empirical user study, we found that, simply through prompting, LLMs-based chatbots could effectively deliver questions and collect desired self-reports. Particularly, we evaluated how two prompt design factors---format and personality modifier--impacted the resulted chatbots' ability in slot filling and conversation styles. While LLMs can be a promising tool to build chatbots, we also discuss problematic messages occurred in our dataset. Reflecting on our results, we provide insights into the prompt design for chatbots and give suggestions on how to improve future LLMs-based chatbots. In closing, we hope this work can inform and inspire other researchers in the fields of HCI and Personal Informatics, to effectively leverage LLMs to power enjoyable chatbots for robust data collection.
\section{Appendices}
\subsection{The Web Chatbot Interface}\label{appendix:interface}
\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{img/chat_interface.png}
\label{fig:web_interface}
\Description{A screenshot of the web chat interface of GPT-3-based chatbots used in our study. The screen consists of two panels in a horizontal layout. The left panel is a chat interface where participants can see the chatbot's and their messages, and submit a new message. The right panel includes the list of guidelines and instruction. A 'Next' button is located at the bottom of the right panel so that participants can terminate the current conversation and proceed to the next conversation.}
\end{figure}
\subsection{Hint Texts for Conversation Paths}\label{appendix:paths}
The hint texts are displayed in \textcolor[HTML]{1f6adb}{blue} (positive) or \textcolor[HTML]{db411f}{red} (negative) in the instruction (See Appendix \ref{appendix:interface}, right).
\begin{table}[H]
\small\sffamily
\def1.3{1.5}\setlength{\tabcolsep}{0.5em}
\flushleft
\begin{tabular}{|l|l!{\color{lightgray}\vrule}l|}
\hline
\rowcolor{tableheader} & \textbf{Positive Path} & \textbf{Negative Path} \\ \hline
\textbf{Sleep} & You had \textit{\textbf{good sleep}} last night. & You had \textbf{\textit{bad sleep}} last night. \\ \arrayrulecolor{tablegrayline}\hline
\textbf{Work} & You were \textit{\textbf{productive}} yesterday. & You \textit{\textbf{did nothing}} yesterday. \\ \hline
\textbf{Food Intake} & You \textit{\textbf{had all three meals}} yesterday. & You \textit{\textbf{skipped one meal}} yesterday. \\ \hline
\textbf{Exercise} & You \textit{\textbf{exercised}} yesterday. & You \textit{\textbf{did not exercise}} yesterday. \\ \arrayrulecolor{black}\hline
\end{tabular}%
\label{tab:hints}
\end{table}
\newpage
\subsection{Least-Squared Mean Statistics for Slot Filling Rate}
Estimation based on a significant linear model formula:
\newline
$SlotFillingRate = Personality+Format+Topic+Path+Personality \times Format$
\newline
$F(7,366)=6.493, p<.0001$***
\begin{table}[H]
\small\sffamily
\def1.3{1.2}\setlength{\tabcolsep}{0.5em}
\centering
\caption{Estimated least-squared means with 95\% confidence intervals (CI) of slot filling rate by condition.}
\begin{tabular}{|l|lllll|}
\hline
\rowcolor{tableheader} \textbf{Condition} & \textbf{Est. Mean} & \textbf{\textit{SE}} & \textbf{df} & \textbf{Lower CI} & \textbf{Upper CI} \\
\hline
\textbf{SP} & 83.03\% & 2.60 & 366 & 77.91 & 88.16 \\ \arrayrulecolor{tablegrayline}\hline
\textbf{SN} & 75.99\% & 2.54 & 366 & 71.00 & 80.98 \\ \hline
\textbf{DP} & 71.90\% & 2.55 & 366 & 66.89 & 76.91 \\ \hline
\textbf{DN} & 83.43\% & 2.59 & 366 & 78.33 & 88.52 \\
\arrayrulecolor{black}\hline
\end{tabular}
\end{table}
\begin{table}[H]
\small\sffamily
\def1.3{1.2}\setlength{\tabcolsep}{0.5em}
\centering
\caption{Estimated least-squared means with 95\% confidence intervals (CI) of slot filling rate by topic.}
\begin{tabular}{|l|lllll|}
\hline
\rowcolor{tableheader} \textbf{Topic} & \textbf{Est. Mean} & \textbf{\textit{SE}} & \textbf{df} & \textbf{Lower CI} & \textbf{Upper CI} \\
\hline
\textbf{Sleep} & 77.35\% & 2.56 & 366 & 72.31 & 82.39 \\ \arrayrulecolor{tablegrayline}\hline
\textbf{Work} & 70.69\% & 2.61 & 366 & 65.57 & 75.81 \\ \hline
\textbf{Food} & 77.91\% & 2.55 & 366 & 72.89 & 82.92 \\ \hline
\textbf{Exercise} & 88.40\% & 2.56 & 366 & 83.36 & 93.44 \\ \arrayrulecolor{black} \hline
\end{tabular}
\end{table}
\begin{table}[H]
\small\sffamily
\def1.3{1.2}\setlength{\tabcolsep}{0.5em}
\centering
\caption{Estimated least-squared means with 95\% confidence intervals (CI) of slot filling rate by conversation path.}
\begin{tabular}{|l|lllll|}
\hline
\rowcolor{tableheader} \textbf{Path} & \textbf{Est. Mean} & \textbf{\textit{SE}} & \textbf{df} & \textbf{Lower CI} & \textbf{Upper CI} \\
\hline
\textbf{Positive} & 81.88\% & 1.82 & 366 & 78.31 & 85.45 \\ \arrayrulecolor{tablegrayline}\hline
\textbf{Negative} & 75.30\% & 1.82 & 366 & 71.72 & 78.87 \\\arrayrulecolor{black} \hline
\end{tabular}
\end{table}
\iffalse
\begin{table}[H]
\small\sffamily
\def1.3{1.2}\setlength{\tabcolsep}{0.5em}
\centering
\caption{Estimated least-squared means with 95\% confidence intervals (CI) of slot filling rate by condition.}
\begin{tabular}{|l|lllll|}
\hline
\rowcolor{tableheader}
\textbf{Contrast} & \textbf{Effect size} & \textbf{\textit{SE}} & \textbf{df} & \textbf{\textit{t}-statistic} & \textbf{\textit{p}-value} \\ \hline
\textbf{DN - DP} & 11.53 & 3.63 & 366 & 3.17 & 0.00888172 \\ \arrayrulecolor{tablegrayline}\hline
\textbf{DN - SN} & 7.44 & 3.62 & 366 & 2.05 & 0.17105564 \\ \hline
\textbf{DN - SP} & 0.39 & 3.67 & 366 & 0.11 & 0.99955672 \\ \hline
\textbf{DP - SN} & -4.09 & 3.60 & 366 & -1.14 & 0.66648961 \\ \hline
\textbf{DP - SP} & -11.13 & 3.64 & 366 & -3.06 & 0.0128534 \\ \hline
\textbf{SN - SP} & -7.04 & 3.64 & 366 & -1.94 & 0.21405601 \\ \arrayrulecolor{black}
\hline
\end{tabular}
\end{table}
\fi
\subsection{Least-Squared Mean Statistics for Required and Secondary Question Turn Ratios}
\subsubsection{Turn Ratio of Required Questions}
Estimation based on a mixed-effect model fitted:
\newline
$RQTurnRatio = Personality+Format+Topic+Path+Personality \times Format + Topic \times Path$
\newline
with significant random effect of participants ($p<.0001$***)
\begin{table}[H]
\small\sffamily
\def1.3{1.2}\setlength{\tabcolsep}{0.5em}
\centering
\caption{Estimated least-squared means with 95\% confidence intervals (CI) of RQ turn ratio by information format.}
\begin{tabular}{|l|lllll|}
\hline
\rowcolor{tableheader} \textbf{Format} & \textbf{Est. Mean} & \textbf{\textit{SE}} & \textbf{df} & \textbf{Lower CI} & \textbf{Upper CI} \\
\hline
\textbf{Specific} & 22.44\% & 1.03 & 43.99 & 20.36 & 24.51 \\ \arrayrulecolor{tablegrayline}\hline
\textbf{Descriptive} & 19.21\% & 1.03 & 43.97 & 17.14 & 21.29 \\\arrayrulecolor{black} \hline
\end{tabular}
\end{table}
\begin{table}[H]
\small\sffamily
\def1.3{1.2}\setlength{\tabcolsep}{0.5em}
\centering
\caption{Estimated least-squared means with 95\% confidence intervals (CI) of RQ turn ratio by condition.}
\begin{tabular}{|l|lllll|}
\hline
\rowcolor{tableheader} \textbf{Condition} & \textbf{Est. Mean} & \textbf{\textit{SE}} & \textbf{df} & \textbf{Lower CI} & \textbf{Upper CI} \\
\hline
\textbf{SP} & 23.93\% & 1.47 & 44.94 & 20.98 & 26.88 \\ \arrayrulecolor{tablegrayline}\hline
\textbf{SN} & 20.94\% & 1.45 & 43.05 & 18.02 & 23.86 \\ \hline
\textbf{DP} & 17.63\% & 1.45 & 43.41 & 14.70 & 20.56 \\ \hline
\textbf{DN} & 20.80\% & 1.46 & 44.54 & 17.85 & 23.74 \\\arrayrulecolor{black} \hline
\end{tabular}
\end{table}
\begin{table}[H]
\small\sffamily
\def1.3{1.2}\setlength{\tabcolsep}{0.5em}
\centering
\caption{Estimated least-squared means with 95\% confidence intervals (CI) of RQ turn ratio by conversation path.}
\begin{tabular}{|l|lllll|}
\hline
\rowcolor{tableheader} \textbf{Path} & \textbf{Est. Mean} & \textbf{\textit{SE}} & \textbf{df} & \textbf{Lower CI} & \textbf{Upper CI} \\
\hline
\textbf{Positive} & 21.92\% & 0.86 & 82.65 & 20.22 & 23.63 \\ \arrayrulecolor{tablegrayline}\hline
\textbf{Negative} & 19.73\% & 0.86 & 82.53 & 18.02 & 21.43 \\\arrayrulecolor{black} \hline
\end{tabular}
\end{table}
\begin{table}[H]
\small\sffamily
\def1.3{1.2}\setlength{\tabcolsep}{0.5em}
\centering
\caption{Estimated least-squared means with 95\% confidence intervals (CI) of RQ turn ratio by topic and conversation path.}
\begin{tabular}{|ll|lllll|}
\hline
\rowcolor{tableheader} \textbf{Topic} & \textbf{Path} & \textbf{Est. Mean} & \textbf{\textit{SE}} & \textbf{df} & \textbf{Lower CI} & \textbf{Upper CI} \\
\hline
\multirow{2}{*}{\textbf{Sleep}} & \textbf{Positive} & 22.73\% & 1.38 & 295.42 & 20.01 & 25.45 \\
\arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{Negative} & 23.78\% & 1.41 & 301.42 & 21.00 & 26.55 \\
\arrayrulecolor{gray}\hline
\multirow{2}{*}{\textbf{Work}} & \textbf{Positive} & 19.13\% & 1.44 & 307.50 & 16.30 & 21.96 \\
\arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{Negative} & 15.04\% & 1.40 & 298.34 & 12.29 & 17.78 \\
\arrayrulecolor{gray}\hline
\multirow{2}{*}{\textbf{Food}} & \textbf{Positive} & 24.03\% & 1.38 & 295.42 & 21.31 & 26.75 \\
\arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{Negative} & 23.54\% & 1.40 & 298.34 & 20.79 & 26.28 \\
\arrayrulecolor{gray}\hline
\multirow{2}{*}{\textbf{Exercise}} & \textbf{Positive} & 21.81\% & 1.40 & 298.44 & 19.06 & 24.56 \\
\arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{Negative} & 16.55\% & 1.40 & 298.44 & 13.81 & 19.30 \\
\arrayrulecolor{black} \hline
\end{tabular}
\vspace{5mm}
\end{table}
\subsubsection{Turn Ratio of Secondary Questions}
Estimation based on a significant linear model formula:
\newline
$SQTurnRatio = Personality+Format+Topic+Path + Topic \times Path$
\newline
$F(9,364)=10.04, p<.0001$***
\begin{table}[H]
\small\sffamily
\def1.3{1.2}\setlength{\tabcolsep}{0.5em}
\centering
\caption{Estimated least-squared means with 95\% confidence intervals (CI) of SQ turn ratio by information format.}
\begin{tabular}{|l|lllll|}
\hline
\rowcolor{tableheader} \textbf{Format} & \textbf{Est. Mean} & \textbf{\textit{SE}} & \textbf{df} & \textbf{Lower CI} & \textbf{Upper CI} \\
\hline
\textbf{Specific} & 9.64\% & 0.68 & 364 & 8.30 & 10.97 \\ \arrayrulecolor{tablegrayline}\hline
\textbf{Descriptive} & 12.55\% & 0.68 & 364 & 11.22 & 13.88 \\\arrayrulecolor{black} \hline
\end{tabular}
\end{table}
\begin{table}[H]
\small\sffamily
\def1.3{1.2}\setlength{\tabcolsep}{0.5em}
\centering
\caption{Estimated least-squared means with 95\% confidence intervals (CI) of SQ turn ratio by condition.}
\begin{tabular}{|l|lllll|}
\hline
\rowcolor{tableheader} \textbf{Condition} & \textbf{Est. Mean} & \textbf{\textit{SE}} & \textbf{df} & \textbf{Lower CI} & \textbf{Upper CI} \\
\hline
\textbf{SP} & 9.26\% & 0.84 & 364 & 7.62 & 10.91 \\ \arrayrulecolor{tablegrayline}\hline
\textbf{SN} & 10.01\% & 0.82 & 364 & 8.40 & 11.63 \\ \hline
\textbf{DP} & 12.18\% & 0.82 & 364 & 10.56 & 13.80 \\ \hline
\textbf{DN} & 12.93\% & 0.83 & 364 & 11.29 & 14.57 \\\arrayrulecolor{black} \hline
\end{tabular}
\end{table}
\begin{table}[H]
\small\sffamily
\def1.3{1.2}\setlength{\tabcolsep}{0.5em}
\centering
\caption{Estimated least-squared means with 95\% confidence intervals (CI) of RQ turn ratio by conversation path.}
\begin{tabular}{|l|lllll|}
\hline
\rowcolor{tableheader} \textbf{Path} & \textbf{Est. Mean} & \textbf{\textit{SE}} & \textbf{df} & \textbf{Lower CI} & \textbf{Upper CI} \\
\hline
\textbf{Positive} & 9.59\% & 0.68 & 364 & 8.26 & 10.92 \\ \arrayrulecolor{tablegrayline}\hline
\textbf{Negative} & 12.60\% & 0.68 & 364 & 11.27 & 13.93 \\\arrayrulecolor{black} \hline
\end{tabular}
\end{table}
\begin{table}[H]
\small\sffamily
\def1.3{1.2}\setlength{\tabcolsep}{0.5em}
\centering
\caption{Estimated least-squared means with 95\% confidence intervals (CI) of SQ turn ratio by topic and conversation path.}
\begin{tabular}{|ll|lllll|}
\hline
\rowcolor{tableheader} \textbf{Topic} & \textbf{Path} & \textbf{Est. Mean} & \textbf{\textit{SE}} & \textbf{df} & \textbf{Lower CI} & \textbf{Upper CI} \\
\hline
\multirow{2}{*}{\textbf{Sleep}} & \textbf{Positive} & 7.71\% & 1.34 & 364 & 5.08 & 10.33 \\
\arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{Negative} & 7.17\% & 1.36 & 364 & 4.49 & 9.85 \\
\arrayrulecolor{gray}\hline
\multirow{2}{*}{\textbf{Work}} & \textbf{Positive} & 14.23\% & 1.39 & 364 & 11.49 & 16.98 \\
\arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{Negative} & 20.03\% & 1.35 & 364 & 17.38 & 22.68 \\
\arrayrulecolor{gray}\hline
\multirow{2}{*}{\textbf{Food}} & \textbf{Positive} & 9.66\% & 1.34 & 364 & 7.04 & 12.29 \\
\arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{Negative} & 9.48\% & 1.35 & 364 & 6.83 & 12.13 \\
\arrayrulecolor{gray}\hline
\multirow{2}{*}{\textbf{Exercise}} & \textbf{Positive} & 6.76\% & 1.35 & 364 & 4.11 & 9.42 \\
\arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{Negative} & 13.70\% & 1.35 & 364 & 11.05 & 16.35 \\
\arrayrulecolor{black} \hline
\end{tabular}
\vspace{5mm}
\end{table}
\subsection{Least-Squared Mean Statistics for Chatbot Behavioral Turn Ratios}
\subsubsection{Chatbot's Acknowledging Turn Ratio}
\section{Appendices}
\subsection{The Web Chatbot Interface}\label{appendix:interface}
\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{img/chat_interface.png}
\label{fig:web_interface}
\Description{A screenshot of the web chat interface of GPT-3-based chatbots used in our study. The screen consists of two panels in a horizontal layout. The left panel is a chat interface where participants can see the chatbot's and their messages, and submit a new message. The right panel includes the list of guidelines and instruction. A 'Next' button is located at the bottom of the right panel so that participants can terminate the current conversation and proceed to the next conversation.}
\end{figure}
\subsection{Hint Texts for Conversation Paths}\label{appendix:paths}
The hint texts are displayed in \textcolor[HTML]{1f6adb}{blue} (positive) or \textcolor[HTML]{db411f}{red} (negative) in the instruction (See Appendix \ref{appendix:interface}, right).
\begin{table}[H]
\small\sffamily
\def1.3{1.5}\setlength{\tabcolsep}{0.5em}
\flushleft
\begin{tabular}{|l|l!{\color{lightgray}\vrule}l|}
\hline
\rowcolor{tableheader} & \textbf{Positive Path} & \textbf{Negative Path} \\ \hline
\textbf{Sleep} & You had \textit{\textbf{good sleep}} last night. & You had \textbf{\textit{bad sleep}} last night. \\ \arrayrulecolor{tablegrayline}\hline
\textbf{Work} & You were \textit{\textbf{productive}} yesterday. & You \textit{\textbf{did nothing}} yesterday. \\ \hline
\textbf{Food Intake} & You \textit{\textbf{had all three meals}} yesterday. & You \textit{\textbf{skipped one meal}} yesterday. \\ \hline
\textbf{Exercise} & You \textit{\textbf{exercised}} yesterday. & You \textit{\textbf{did not exercise}} yesterday. \\ \arrayrulecolor{black}\hline
\end{tabular}%
\label{tab:hints}
\end{table}
\subsection{Least-Squared Mean Statistics for Slot Filling Rate}\label{appendix:stats:slot}
Estimation based on a significant linear model formula:
\newline
$SlotFillingRate = Personality+Format+Topic+Path+Personality \times Format$
\newline
$F(7,366)=6.493, p<.0001$***
\begin{table}[H]
\small\sffamily
\def1.3{1}\setlength{\tabcolsep}{1em}
\centering
\caption{Estimated least-squared means with 95\% confidence intervals (CI) of slot filling rate by variables.}
\begin{tabular}{|ll|lllll|}
\hline
\rowcolor{tableheader} \textbf{Variable} & \textbf{Category} & \textbf{Est. Mean} & \textbf{\textit{SE}} & \textbf{df} & \textbf{Lower CI} & \textbf{Upper CI} \\
\hline
\multirow{4}{*}{\textbf{Condition}} & \textbf{SP} & 83.03\% & 2.60 & 366 & 77.91 & 88.16 \\ \arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{SN} & 75.99\% & 2.54 & 366 & 71.00 & 80.98 \\ \cline{2-7}
& \textbf{DP} & 71.90\% & 2.55 & 366 & 66.89 & 76.91 \\ \cline{2-7}
& \textbf{DN} & 83.43\% & 2.59 & 366 & 78.33 & 88.52 \\
\arrayrulecolor{black}\hline
\multirow{4}{*}{\textbf{Topic}} & \textbf{Sleep} & 77.35\% & 2.56 & 366 & 72.31 & 82.39 \\ \arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{Work} & 70.69\% & 2.61 & 366 & 65.57 & 75.81 \\ \cline{2-7}
& \textbf{Food} & 77.91\% & 2.55 & 366 & 72.89 & 82.92 \\ \cline{2-7}
& \textbf{Exercise} & 88.40\% & 2.56 & 366 & 83.36 & 93.44 \\ \arrayrulecolor{black} \hline
\multirow{2}{*}{\textbf{Path}} & \textbf{Positive} & 81.88\% & 1.82 & 366 & 78.31 & 85.45 \\ \arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{Negative} & 75.30\% & 1.82 & 366 & 71.72 & 78.87 \\\arrayrulecolor{black} \hline
\end{tabular}
\end{table}
\subsection{Least-Squared Mean Statistics for Required and Secondary Question Turn Ratios}
\subsubsection{Turn Ratio of Required Questions}\label{appendix:stats:rq}
Estimation based on a mixed-effect model fitted:
\newline
$RQTurnRatio = Personality+Format+Topic+Path+Personality \times Format + Topic \times Path$
\newline
with significant random effect of participants ($p<.0001$***)
\begin{table}[H]
\small\sffamily
\def1.3{1}\setlength{\tabcolsep}{1em}
\centering
\caption{Estimated least-squared means with 95\% confidence intervals (CI) of RQ turn ratio by variables.}
\begin{tabular}{|lll|lllll|}
\hline
\rowcolor{tableheader} \textbf{Variable} & \multicolumn{2}{l|}{\textbf{Category}} & \textbf{Est. Mean} & \textbf{\textit{SE}} & \textbf{df} & \textbf{Lower CI} & \textbf{Upper CI} \\
\hline
\multirow{2}{*}{\textbf{Format}} & \textbf{Specific} & & 22.44\% & 1.03 & 43.99 & 20.36 & 24.51 \\ \arrayrulecolor{tablegrayline}\cline{2-8}
& \textbf{Descriptive} & & 19.21\% & 1.03 & 43.97 & 17.14 & 21.29 \\\arrayrulecolor{black} \hline
\multirow{4}{*}{\textbf{Condition}} & \textbf{SP} & & 23.93\% & 1.47 & 44.94 & 20.98 & 26.88 \\ \arrayrulecolor{tablegrayline}\cline{2-8}
& \textbf{SN} & & 20.94\% & 1.45 & 43.05 & 18.02 & 23.86 \\ \cline{2-8}
& \textbf{DP} & & 17.63\% & 1.45 & 43.41 & 14.70 & 20.56 \\ \cline{2-8}
& \textbf{DN} & & 20.80\% & 1.46 & 44.54 & 17.85 & 23.74 \\\arrayrulecolor{black} \hline
\multirow{2}{*}{\textbf{Path}} & \textbf{Positive} & & 21.92\% & 0.86 & 82.65 & 20.22 & 23.63 \\ \arrayrulecolor{tablegrayline}\cline{2-8}
& \textbf{Negative} & & 19.73\% & 0.86 & 82.53 & 18.02 & 21.43 \\\arrayrulecolor{black} \hline
\multirow{8}{*}{\textbf{Topic $\times$ Path}} & \multirow{2}{*}{\textbf{Sleep}} & \textbf{Positive} & 22.73\% & 1.38 & 295.42 & 20.01 & 25.45 \\
\arrayrulecolor{tablegrayline}\cline{3-8}
& & \textbf{Negative} & 23.78\% & 1.41 & 301.42 & 21.00 & 26.55 \\
\arrayrulecolor{gray}\cline{2-8}
& \multirow{2}{*}{\textbf{Work}} & \textbf{Positive} & 19.13\% & 1.44 & 307.50 & 16.30 & 21.96 \\
\arrayrulecolor{tablegrayline}\cline{3-8}
& & \textbf{Negative} & 15.04\% & 1.40 & 298.34 & 12.29 & 17.78 \\
\arrayrulecolor{gray}\cline{2-8}
& \multirow{2}{*}{\textbf{Food}} & \textbf{Positive} & 24.03\% & 1.38 & 295.42 & 21.31 & 26.75 \\
\arrayrulecolor{tablegrayline}\cline{3-8}
& & \textbf{Negative} & 23.54\% & 1.40 & 298.34 & 20.79 & 26.28 \\
\arrayrulecolor{gray}\cline{2-8}
& \multirow{2}{*}{\textbf{Exercise}} & \textbf{Positive} & 21.81\% & 1.40 & 298.44 & 19.06 & 24.56 \\
\arrayrulecolor{tablegrayline}\cline{3-8}
& & \textbf{Negative} & 16.55\% & 1.40 & 298.44 & 13.81 & 19.30 \\
\arrayrulecolor{black} \hline
\end{tabular}
\end{table}
\subsubsection{Turn Ratio of Secondary Questions}\label{appendix:stats:sq}
Estimation based on a significant linear model formula:
\newline
$SQTurnRatio = Personality+Format+Topic+Path + Topic \times Path$
\newline
$F(9,364)=10.04, p<.0001$***
\begin{table}[H]
\small\sffamily
\def1.3{1}\setlength{\tabcolsep}{1em}
\centering
\caption{Estimated least-squared means with 95\% confidence intervals (CI) of SQ turn ratio by variables.}
\begin{tabular}{|lll|lllll|}
\hline
\rowcolor{tableheader} \textbf{Variable} & \multicolumn{2}{l|}{\textbf{Category}} & \textbf{Est. Mean} & \textbf{\textit{SE}} & \textbf{df} & \textbf{Lower CI} & \textbf{Upper CI} \\
\hline
\multirow{2}{*}{\textbf{Format}} & \textbf{Specific} & & 9.64\% & 0.68 & 364 & 8.30 & 10.97 \\ \arrayrulecolor{tablegrayline}\cline{2-8}
& \textbf{Descriptive} & & 12.55\% & 0.68 & 364 & 11.22 & 13.88 \\\arrayrulecolor{black} \hline
\multirow{4}{*}{\textbf{Condition}} & \textbf{SP} & & 9.26\% & 0.84 & 364 & 7.62 & 10.91 \\ \arrayrulecolor{tablegrayline}\cline{2-8}
& \textbf{SN} & & 10.01\% & 0.82 & 364 & 8.40 & 11.63 \\ \cline{2-8}
& \textbf{DP} & & 12.18\% & 0.82 & 364 & 10.56 & 13.80 \\ \cline{2-8}
& \textbf{DN} & & 12.93\% & 0.83 & 364 & 11.29 & 14.57 \\\arrayrulecolor{black} \hline
\multirow{2}{*}{\textbf{Path}} & \textbf{Positive} & & 9.59\% & 0.68 & 364 & 8.26 & 10.92 \\ \arrayrulecolor{tablegrayline}\cline{2-8}
& \textbf{Negative} & & 12.60\% & 0.68 & 364 & 11.27 & 13.93 \\\arrayrulecolor{black} \hline
\multirow{8}{*}{\textbf{Topic $\times$ Path}} & \multirow{2}{*}{\textbf{Sleep}} & \textbf{Positive} & 7.71\% & 1.34 & 364 & 5.08 & 10.33 \\
\arrayrulecolor{tablegrayline}\cline{3-8}
& & \textbf{Negative} & 7.17\% & 1.36 & 364 & 4.49 & 9.85 \\
\arrayrulecolor{gray}\cline{2-8}
& \multirow{2}{*}{\textbf{Work}} & \textbf{Positive} & 14.23\% & 1.39 & 364 & 11.49 & 16.98 \\
\arrayrulecolor{tablegrayline}\cline{3-8}
& & \textbf{Negative} & 20.03\% & 1.35 & 364 & 17.38 & 22.68 \\
\arrayrulecolor{gray}\cline{2-8}
& \multirow{2}{*}{\textbf{Food}} & \textbf{Positive} & 9.66\% & 1.34 & 364 & 7.04 & 12.29 \\
\arrayrulecolor{tablegrayline}\cline{3-8}
& & \textbf{Negative} & 9.48\% & 1.35 & 364 & 6.83 & 12.13 \\
\arrayrulecolor{gray}\cline{2-8}
& \multirow{2}{*}{\textbf{Exercise}} & \textbf{Positive} & 6.76\% & 1.35 & 364 & 4.11 & 9.42 \\
\arrayrulecolor{tablegrayline}\cline{3-8}
& & \textbf{Negative} & 13.70\% & 1.35 & 364 & 11.05 & 16.35 \\
\arrayrulecolor{black} \hline
\end{tabular}
\end{table}
\newpage
\subsection{Least-Squared Mean Statistics for Chatbot Behavioral Turn Ratios}\label{appendix:stats:empathy}
\subsubsection{Chatbot's Acknowledging Turn Ratio}\label{appendix:stats:acknowledging}
Estimation based on a significant linear model formula:
\newline
$AcknowledgingTurnRatio = Personality+Format+Topic+Path+Personality \times Format$
\newline
$F(7,366)=6.715, p<.0001$***
\begin{table}[H]
\small\sffamily
\def1.3{1}\setlength{\tabcolsep}{1em}
\centering
\caption{Estimated least-squared means with 95\% confidence intervals (CI) of Acknowledging turn ratio by variables.}
\begin{tabular}{|ll|lllll|}
\hline
\rowcolor{tableheader} \textbf{Variable} & \textbf{Category} & \textbf{Est. Mean} & \textbf{\textit{SE}} & \textbf{df} & \textbf{Lower CI} & \textbf{Upper CI} \\
\hline
\multirow{2}{*}{\textbf{Format}} & \textbf{Specific} & 17.44\% & 0.74 & 366 & 15.98 & 18.90 \\ \arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{Descriptive} & 22.56\% & 0.74 & 366 & 21.10 & 24.02 \\\arrayrulecolor{black} \hline
\multirow{2}{*}{\textbf{Personality Modifier}} & \textbf{With Modifier} & 20.85\% & 0.74 & 366 & 19.39 & 22.32 \\ \arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{No Modifier} & 19.15\% & 0.74 & 366 & 17.69 & 20.61 \\\arrayrulecolor{black} \hline
\multirow{4}{*}{\textbf{Condition}} & \textbf{SP} & 17.20\% & 1.06 & 366 & 15.11 & 19.30 \\ \arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{SN} & 17.68\% & 1.04 & 366 & 15.64 & 19.72 \\ \cline{2-7}
& \textbf{DP} & 24.51\% & 1.04 & 366 & 22.46 & 26.55 \\ \cline{2-7}
& \textbf{DN} & 20.62\% & 1.06 & 366 & 18.54 & 22.70 \\\arrayrulecolor{black} \hline
\multirow{2}{*}{\textbf{Path}} & \textbf{Positive} & 17.99\% & 0.74 & 366 & 16.52 & 19.45 \\ \arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{Negative} & 22.02\% & 0.74 & 366 & 20.56 & 23.48 \\
\arrayrulecolor{black} \hline
\end{tabular}
\end{table}
\subsubsection{Chatbot's Appreciating Turn Ratio}\label{appendix:stats:appreciating}
Estimation based on a significant linear model formula:
\newline
$AppreciatingTurnRatio = Personality+Format+Topic+Path+Topic \times Format$
\newline
$F(9,364)=8.903, p<.0001$***
\begin{table}[H]
\small\sffamily
\def1.3{1}\setlength{\tabcolsep}{1em}
\centering
\caption{Estimated least-squared means with 95\% confidence intervals (CI) of Appreciating turn ratio by variables.}
\begin{tabular}{|ll|lllll|}
\hline
\rowcolor{tableheader} \textbf{Variable} & \textbf{Category} & \textbf{Est. Mean} & \textbf{\textit{SE}} & \textbf{df} & \textbf{Lower CI} & \textbf{Upper CI} \\
\hline
\multirow{2}{*}{\textbf{Format}} & \textbf{Specific} & 7.61\% & 0.59 & 364 & 6.44 & 8.78 \\ \arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{Descriptive} & 9.64\% & 0.59 & 364 & 8.47 & 10.81 \\\arrayrulecolor{black} \hline
\multirow{2}{*}{\textbf{Personality Modifier}} & \textbf{With Modifier} & 9.36\% & 0.60 & 364 & 8.19 & 10.53 \\ \arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{No Modifier} & 7.89\% & 0.59 & 364 & 6.73 & 9.06 \\\arrayrulecolor{black} \hline
\multirow{4}{*}{\textbf{Condition}} & \textbf{SP} & 8.34\% & 0.73 & 364 & 6.89 & 9.78 \\ \arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{SN} & 6.87\% & 0.72 & 364 & 5.46 & 8.29 \\ \cline{2-7}
& \textbf{DP} & 10.38\% & 0.72 & 364 & 8.95 & 11.80 \\ \cline{2-7}
& \textbf{DN} & 8.91\% & 0.73 & 364 & 7.47 & 10.35 \\\arrayrulecolor{black} \hline
\multirow{2}{*}{\textbf{Path}} & \textbf{Positive} & 11.33\% & 0.59 & 364 & 10.16 & 12.50 \\ \arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{Negative} & 5.92\% & 0.59 & 364 & 4.75 & 7.09 \\
\arrayrulecolor{black} \hline
\end{tabular}
\end{table}
\subsubsection{Chatbot's Sympathizing Turn Ratio}\label{appendix:stats:sympathizing}
Estimation based on a significant linear model formula:
\newline
$SympathizingTurnRatio = Personality+Format+Topic+Path+Topic \times Path$
\newline
$F(9,364)=8.219, p<.0001$***
\begin{table}[H]
\small\sffamily
\def1.3{1}\setlength{\tabcolsep}{1em}
\centering
\caption{Estimated least-squared means with 95\% confidence intervals (CI) of Sympathizing turn ratio by variables.}
\begin{tabular}{|ll|lllll|}
\hline
\rowcolor{tableheader} \textbf{Variable} & \textbf{Category} & \textbf{Est. Mean} & \textbf{\textit{SE}} & \textbf{df} & \textbf{Lower CI} & \textbf{Upper CI} \\
\hline
\multirow{2}{*}{\textbf{Format}} & \textbf{Specific} & 3.18\% & 0.39 & 364 & 2.41 & 3.96 \\ \arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{Descriptive} & 3.03\% & 0.39 & 364 & 2.26 & 3.81 \\\arrayrulecolor{black} \hline
\multirow{2}{*}{\textbf{Personality Modifier}} & \textbf{With Modifier} & 3.97\% & 0.40 & 364 & 3.19 & 4.75 \\ \arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{No Modifier} & 2.25\% & 0.39 & 364 & 1.47 & 3.02 \\\arrayrulecolor{black} \hline
\multirow{4}{*}{\textbf{Condition}} & \textbf{SP} & 4.04\% & 0.49 & 364 & 3.08 & 5.00 \\ \arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{SN} & 2.32\% & 0.48 & 364 & 1.38 & 3.27 \\ \cline{2-7}
& \textbf{DP} & 3.89\% & 0.48 & 364 & 2.95 & 4.84 \\ \cline{2-7}
& \textbf{DN} & 2.17\% & 0.49 & 364 & 1.22 & 3.13 \\\arrayrulecolor{black} \hline
\multirow{2}{*}{\textbf{Path}} & \textbf{Positive} & 1.16\% & 0.40 & 364 & 0.39 & 1.94 \\ \arrayrulecolor{tablegrayline}\cline{2-7}
& \textbf{Negative} & 5.05\% & 0.39 & 364 & 4.28 & 5.83 \\
\arrayrulecolor{black} \hline
\end{tabular}
\end{table}
|
\section{Introduction}
\IEEEPARstart{S}{moothing} algorithms are important techniques for estimating the state of dynamic systems, which have been applied in various industrial applications such as target tracking \cite{9648188}, system identification \cite{LIU2021106995}, positioning \cite{CHIANG20112633}, and many others \cite{8571180,BALENZUELA2022110218,wang2016variational}. The Rauch-Tung-Striebel (RTS) smoother \cite{rauch1965maximum} is important for estimating the state of linear estimation with Gaussian noise. For nonlinear systems, some classic Gaussian smoothers are developed such as the extended RTS (ERTS) \cite{sarkka2013bayesian}, the unscented RTS (URTS) \cite{4484208}, the cubature Kalman \cite{ARASARATNAM20112245} smoothers and some iterative strategy-based smoothers \cite{bell1994iterated}.
These aforementioned smoothers perform well with the presence of both the Gaussian process and measurement noises. However, a class of non-Gaussian measurement noise with heavy-tailed distributions is common in practical applications \cite{8540327,WANG2022108394,WANG20205058,ARAVKIN201763}. Non-Gaussian noise will dramatically decrease the performance of the smoothers that are initially devised for Gaussian noises. To address this issue, some strategies have been developed. Robust statistics, especially M-estimations \cite{WANG2019115,karlgaard2015nonlinear}, is one of the popular strategies to develop robust smoother for non-Gaussian noises. Another popular approach is to use the heavy-tailed distributions to model the non-Gaussian process and measurement noises, such as the Laplace \cite{5746504} distribution, Student's t distribution \cite{7416166,7812899,9549725}, and generalized Gaussian scale mixture distribution \cite{8706940}. In addition, the robust smoothers based on ${H_\infty }$ \cite{5762390} and Gaussian mixture model have also been studied. In recent, the maximum correntropy criterion (MCC) \cite{4355325}, in information theoretic learning (ITL) \cite{5952087}, receives a great deal of attention, and some RTS-type smoothers are developed for nonlinear systems, such as maximum correntropy RTS (MC-RTS) \cite{wang2020maximum} smoother, maximum mixture correntropy nonlinear smoother (MMC-ORNS) \cite{LU2021108215}, and robust cubature Kalman smoother (RCKS) \cite{8540327}.
Although the MCC criterion is a good option for suppressing the effects of the heavy-tailed non-Gaussian noises \cite{8736038,9354208}, it may not perform as well for more complicated noise \cite{LU2021108215}, for example noises with multimodal distribution. In ITL, the minimum error entropy (MEE) criterion \cite{erdogmus2002error,HE2023109188}, as an important cost function, has been applied in adaptive filtering \cite{WANG2022108410,HE20221362,WANG2021107836,HE20223434}, Kalman filter (KF) \cite{9745084,CHEN2017,FENG2022108806}, and others \cite{silva2006error,santamaria2002entropy}.
In addition, unscented KF \cite{dang2020robust} and cubature KF \cite{dang2021cubature} are also combined with the MEE criterion to enhance the estimation performance of the nonlinear systems with non-Gaussian noise. However, the MEE-based smoother has not been investigated. Moreover, compared with the KF-type algorithms, the smoothers based on the RTS framework has an additional step called backward pass, and this additional step gives the smoother an improved performance than the corresponding KF-type algorithms. Hence, a novel RTS smoother using the MEE criterion as a cost function is investigated, which is expected to obtain superior performance.
In this study, a RTS-type smoother based the MEE criterion (MEE-RTS smoother) is proposed for the state estimation of the linear systems. Furthermore, an extended RTS smoother based on the MEE criterion is also developed for the nonlinear systems. Several theoretical analyses of the MEE-RTS smoother are presented, such as the mean error behavior, mean square error behavior, and computational complexity. Numerical simulations indicate the proposed smoothers outperform several existing robust smoothers in the mean-square deviation (MSD).
The remaining parts of this article are structured as follows. In Section \ref{background}, the problem formulation on the proposed smoothers are given. Two new smoothers on the basis of the MEE criterion are presented in Section \ref{mee_smoother}. A few theoretical analysis of the MEE-RTS smoother are presented in Section \ref{tanluysis}. Simulations, conclusions, and acknowledgments are presented in Section \ref{simulation}, Section \ref{conclusion}, and Section \ref{acknowledgments}.
\section{Problem formulation}\label{background}
A discrete-time linear systems, in this paper, is consider, and it is described using the following state-space model
\begin{align}\label{1qtrtxtj1fH}
\left\{ \begin{gathered}
{{\boldsymbol{x}}_t} = {\boldsymbol{F}_t}{{\boldsymbol{x}}_{t - 1}} + {{\boldsymbol{q}}_{t}}, \hfill \\
{{\boldsymbol{y}}_t} = {\boldsymbol{H}_t}{{\boldsymbol{x}}_t} + {{\boldsymbol{r}}_t}, \hfill \\
\end{gathered} \right.
\end{align}
where ${{\boldsymbol{x}}_t} \in {\mathbb{R}^{n \times 1}}$ represents the state of the linear systems at time $t$; ${{\boldsymbol{y}}_t} \in {\mathbb{R}^{m \times 1}}$ denotes the measurement of interest; ${\boldsymbol{F}_t}$ and ${\boldsymbol{H}_t}$ denote the state transition matrix and measurement matrix; ${{\boldsymbol{q}}_{t}}$ and ${{\boldsymbol{r}}_t}$ represent, respectively, the process and measurement noises. Nominally, both ${{\boldsymbol{q}}_{t}}$ and ${{\boldsymbol{r}}_t}$ follow zero-mean Gaussian distributions, and their covariance matrices are ${{\boldsymbol{Q}}_t}$ and ${{\boldsymbol{R}}_t}$, respectively. However, due to the impulse disturbance, outliers, or other factors, the distributions of ${{\boldsymbol{q}}_{t}}$ and ${{\boldsymbol{r}}_t}$ are generally no longer Gaussian, and reveal the heavy-tail properties. Such non-Gaussian distributions will degrade the performance of the existing RTS smoothers since they are initially devised under the Gaussian assumptions. To deal with the performance degradation, in this work, a robust smoother is developed to estimate the states ${{\boldsymbol{x}}_{1:T}} \triangleq \left\{ {{{\boldsymbol{x}}_1}, \cdots ,{{\boldsymbol{x}}_T}} \right\}$ utilizing the measurements ${{\boldsymbol{y}}_{1:T}} \triangleq \left\{ {{{\boldsymbol{y}}_1}, \cdots ,{{\boldsymbol{y}}_T}} \right\}$ with the presence of the heavy-tailed non-Gaussian noise. Specifically, we integrate the MEE criterion into the conventional RTS smoother framework to dampen the negative effect of the non-Gaussian noises.
\section{MEE based robust smoother}\label{mee_smoother}
\subsection{Minimum error entropy criterion}
The quadratic Renyi's entropy \cite{chen2019minimum}, also known as the error entropy, is the definition of
\begin{align}\label{kekrvgolrhj1}
{H_2}\left( e \right) = - \log {V_2}\left( e \right),
\end{align}
where error $e\triangleq X-Y$ denotes the error made between two random variables ${X}$ and ${Y}$, and the information potential ${V_2}\left( e \right)$ is given by
\begin{equation}\label{zkhxkhe1jrfpED}
\begin{split}
{V_2}\left( e \right) = \int {{p^2}\left( e \right)} de = {\text{E}}\left[ {p\left( e \right)} \right],
\end{split}
\end{equation}
where $p\left( e \right)$ represent the probability density function (PDF) on ${e}$, and ${{\text{E}}\left[ \cdot \right]}$ is an expectation operator. The PDF ${p\left( e \right)}$ can be estimated via the Parzen window strategy employing the error samples:
\begin{equation}\label{xkhiejxxksgmg}
\begin{split}
\hat p\left( e \right) = \frac{1}{L}\sum\limits_{i = 1}^L G \left( {e - {e_i}} \right),
\end{split}
\end{equation}
where $G(e) = \left( {1/\sqrt {2\pi } \sigma } \right)\exp \left( { - {e^2}/2{\sigma ^2}} \right)$ is the Gaussian kernel function, and parameter ${\sigma }$ is the kernel bandwidth of $G\left( \cdot \right)$, and ${\left\{ {{e_i}} \right\}_{i = 1}^L}$ denote $L$ independent and identically distributed (i.i.d.) error samples. With the estimated PDF in \eqref{xkhiejxxksgmg}, one can obtain an estimate of ${{V_2}\left( e \right)}$ via \eqref{zkhxkhe1jrfpED} as
\begin{equation}\label{gsigeiejjiij}
{\hat V_2}\left( e \right) = \frac{1}{L}\sum\limits_{i = 1}^L {\hat p\left( e \right)} = \frac{1}{{{L^2}}}\sum\limits_{i = 1}^L {\sum\limits_{j = 1}^L {G\left( {{e_i} - {e_j}} \right)} } ,
\end{equation}
and the detailed derivation process can be found in \cite{5952087}.
In ITL, minimizing the error entropy can be used as a criterion for optimization, i.e., the MEE criterion, which is given by
\begin{align}
\min H_2(e) = \max \log V_2(e).
\end{align}
Due to the fact that $\max \log V_2(e)$ is equivalent to $\max V_2(e)$ and $V_2(e) $ can be estimated via (5) with samples, the MEE criterion can be further expressed as
\begin{align}
\max {V_2}(e) \approx \max \frac{1}{{{L^2}}}\sum\limits_{i = 1}^L {\sum\limits_{j = 1}^L G } \left( {{e_i} - {e_j}} \right).
\end{align}
The MEE criterion has unique advantages for handling complicated non-Gaussian noise \cite{chen2019minimum}, which motivates us to develop a new smoother based on the MEE criterion.
\subsection{The MEE based RTS smoother}
\subsubsection{Forward filtering procedure}
The forward filtering process is the foundation of the backward smoothing process, and its output $\left\{ {{{{\boldsymbol{\tilde x}}}_{t|t}},{{\boldsymbol{P}}_{t|t}}} \right\}$ are the input of the backward smoothing process. In the linear-regression-based KF solution, the measurement equation and filter update be recast as a regression problem \cite{karlgaard2015nonlinear}. Denote ${{\boldsymbol{x}}_t}$ the real state of the system, and hence the state prediction error can be expressed as ${{\boldsymbol{\delta }}_t} = {{\boldsymbol{\tilde x}}_{t|t - 1}} - {{\boldsymbol{x}}_t}$, where ${{\boldsymbol{\tilde x}}_{t|t - 1}}$ is the predicted state. With ${{{\boldsymbol{y}}_t}}$ and ${{{{\boldsymbol{\tilde x}}}_{t|t - 1}}}$, the regression problem with the form of
\begin{align}\label{gongshiXtjblh}
\left[ {\begin{array}{*{20}{c}}
{{{\boldsymbol{y}}_t}} \\
{{{{\boldsymbol{\tilde x}}}_{t|t - 1}}}
\end{array}} \right] = \left[ {\begin{array}{*{20}{c}}
{{{\boldsymbol{H}}_t}{{\boldsymbol{x}}_t}} \\
{{{\boldsymbol{x}}_t}}
\end{array}} \right] + \left[ {\begin{array}{*{20}{c}}
{{{\boldsymbol{r}}_t}} \\
{{{\boldsymbol{\delta }}_t}}
\end{array}} \right].
\end{align}
Defining
\begin{align}
{{\boldsymbol{M}}_t} = \left[ {\begin{array}{*{20}{c}}
{{\boldsymbol{R}}_t^{ - \tfrac{1}{2}}}&{\boldsymbol{0}} \\
{\boldsymbol{0}}&{{\boldsymbol{P}}_{t|t - 1}^{ - \tfrac{1}{2}}}
\end{array}} \right],
\end{align}
and multiplying ${{\boldsymbol{M}}_t}$ on both sides of \eqref{gongshiXtjblh} will transform the linear regression problem in \eqref{gongshiXtjblh} as
\begin{align}\label{8etjiatxTzD}
{{\boldsymbol{d}}_t} = {{\boldsymbol{Z}}_t}{{\boldsymbol{x}}_t} + {{\boldsymbol{e}}_t},
\end{align}
where
\begin{align}\label{ytRpxttj1bla}
{{\boldsymbol{d}}_t} = \left[ {\begin{array}{*{20}{c}}
{{\boldsymbol{R}}_t^{ - \tfrac{1}{2}}{{\boldsymbol{y}}_t}} \\
{{\boldsymbol{P}}_{t|t - 1}^{ - \tfrac{1}{2}}{{{\boldsymbol{\tilde x}}}_{t|t - 1}}}
\end{array}} \right],
\end{align}
\begin{align}\label{zTdzkrhrttsZk}
{{\boldsymbol{Z}}_t} = \left[ {\begin{array}{*{20}{c}}
{{\boldsymbol{R}}_t^{ - \tfrac{1}{2}}{{\boldsymbol{H}}_t}} \\
{{\boldsymbol{P}}_{t|t - 1}^{ - \tfrac{1}{2}}}
\end{array}} \right],
\end{align}
and
\begin{align}
{{\boldsymbol{e}}_t} = \left[ {\begin{array}{*{20}{c}}
{{\boldsymbol{R}}_t^{ - \tfrac{1}{2}}}&{{{\boldsymbol{0}}_n}} \\
{{{\boldsymbol{0}}_n}}&{{\boldsymbol{P}}_{t|t - 1}^{ - \tfrac{1}{2}}}
\end{array}} \right]\left[ {\begin{array}{*{20}{c}}
{{{\boldsymbol{r}}_t}} \\
{{{\boldsymbol{\delta }}_t}}
\end{array}} \right].
\end{align}
From \eqref{8etjiatxTzD}, the augmented error can be written as
\begin{align}\label{xtzTjTddeyTe}
{{\boldsymbol{e}}_t} = {{\boldsymbol{d}}_t} - {{\boldsymbol{Z}}_t}{{\boldsymbol{x}}_t},
\end{align}
and hence the optimal estimation of ${{\boldsymbol{\tilde x}}_{t|t}}$ can be gained by the MEE criterion, which is formulated as
\begin{align}\label{etfijitfeGsig}
{{{\boldsymbol{\tilde x}}}_{t|t}} &= \arg \mathop {\max }\limits_{{{\boldsymbol{x}}_t}} {J_{MEE}}\left( {{{\boldsymbol{x}}_t}} \right)\notag\\
&= \arg \mathop {\max }\limits_{{{\boldsymbol{x}}_t}} \frac{1}{{{N^2}}}\sum\limits_{i = 1}^N {\sum\limits_{j = 1}^N {G\left( {{e_{t;i}} - {e_{t;j}}} \right)} } .
\end{align}
where ${e_{t;i}} = {d_{t;i}} - {{\boldsymbol{z}}_{t;i}}{{\boldsymbol{x}}_t}$, ${e_{t;i}}$ and ${d_{t;i}}$ are the $i$th element of ${{\boldsymbol{e}}_t}$ and ${{\boldsymbol{d}}_t}$, ${{\boldsymbol{z}}_{t;i}}$ represents the $i$th row of ${{\boldsymbol{Z}}_t}$. The variable $N$ is equal to $m + n$.
Set the gradient of the cost function in \eqref{etfijitfeGsig} on ${{{\boldsymbol{e}}_t}}$ will lead to
\begin{align}\label{zerodetetFai}
\frac{{\text{2}}}{{{N^2}{\sigma ^2}}}{\boldsymbol{Z}}_t^{\text{T}}{{\boldsymbol{\Psi }}_t}{{\boldsymbol{e}}_t} - \frac{{\text{2}}}{{{N^2}{\sigma ^2}}}{\boldsymbol{Z}}_t^{\text{T}}{{\boldsymbol{\Phi }}_t}{{\boldsymbol{e}}_t} = 0,
\end{align}
where the $\left( {i,j} \right)$th element of ${{\boldsymbol{\Psi }}_t}$ and ${{\boldsymbol{\Phi }}_t}$ are, respectively, given by
${\left[ {{{\boldsymbol{\Phi }}_t}} \right]_{ij}} = G\left( {{e_{t;i}} - {e_{t;j}}} \right)$ and
\begin{align*}
{\left[ {{{\boldsymbol{\Psi }}_t}} \right]_{ij}} = \left\{ {\begin{array}{*{20}{l}}
{\sum\limits_{j = 1}^N {G\left( {{e_{t;i}} - {e_{t;j}}} \right),} } \\
{0,i \ne j,}
\end{array}} \right.
\end{align*}
Substitute \eqref{xtzTjTddeyTe} into \eqref{zerodetetFai} and one can reach at
\begin{align}\label{DtOUtZtTj1zktZ}
{{\boldsymbol{x}}_t} = {\left[ {{\boldsymbol{Z}}_t^{\text{T}}{{\boldsymbol{\Omega }}_t}{{\boldsymbol{Z}}_t}} \right]^{ - 1}}{\boldsymbol{Z}}_t^{\text{T}}{{\boldsymbol{\Omega }}_t}{{\boldsymbol{d}}_t},
\end{align}
where ${{\boldsymbol{\Omega }}_t}$ is expressed as
\begin{align}\label{outfaizhzTjtfaizh}
{{\boldsymbol{\Omega }}_t} &= {\boldsymbol{\Psi }}_t^{\text{T}}{{\boldsymbol{\Psi }}_t} + {\boldsymbol{\Phi }}_t^{\text{T}}{{\boldsymbol{\Phi }}_t} \notag\\
&= \left[ {\begin{array}{*{20}{c}}
{{{\boldsymbol{\Omega }}_{t;y}}}&{{{\boldsymbol{\Omega }}_{t;xy}}} \\
{{{\boldsymbol{\Omega }}_{t;yx}}}&{{{\boldsymbol{\Omega }}_{t;x}}}
\end{array}} \right].
\end{align}
Define ${\left[ {{\boldsymbol{Z}}_t^{\text{T}}{{\boldsymbol{\Omega }}_t}{{\boldsymbol{Z}}_t}} \right]^{ - 1}}{\boldsymbol{Z}}_t^{\text{T}}{{\boldsymbol{\Omega }}_t}{{\boldsymbol{d}}_t} = l\left( {{{\boldsymbol{x}}_t}} \right)$, and thus Eq. \eqref{DtOUtZtTj1zktZ} becomes a function on ${{\boldsymbol{x}}_t}$, so the estimated ${{{{\boldsymbol{\tilde x}}}_{t|t}}}$ can be obtained via the fixed-point iteration (FPI) method
\begin{align}
{\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right)_{k + 1}} = l\left[ {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right)}_k}} \right] = {\left[ {{\boldsymbol{Z}}_t^{\text{T}}{{\left( {{{{\boldsymbol{\hat \Omega }}}_t}} \right)}_k}{{\boldsymbol{Z}}_t}} \right]^{ - 1}}{\boldsymbol{Z}}_t^{\text{T}}{\left( {{{{\boldsymbol{\hat \Omega }}}_t}} \right)_k}{{\boldsymbol{d}}_t},
\end{align}
where variables with superscript $ \wedge $ denote variables that are affected by the introduction of the FPI method, e.g. ${{{{\boldsymbol{\hat \Omega }}}_t}}$, where the subscript $k$ is the number of the FPI method.
Combining \eqref{ytRpxttj1bla}, \eqref{zTdzkrhrttsZk}, and \eqref{outfaizhzTjtfaizh} yields
\begin{align}\label{17ZtOtZtTDtOUtT}
\left\{ \begin{gathered}
{\boldsymbol{Z}}_t^{\text{T}}{\left( {{{{\boldsymbol{\hat \Omega }}}_t}} \right)_k}{{\boldsymbol{Z}}_t} = {\boldsymbol{H}}_t^{\text{T}}{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^y} \right)_k}{{\boldsymbol{H}}_t} + {\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{yx}} \right)_k}{{\boldsymbol{H}}_t} \hfill \\
+ {\boldsymbol{H}}_t^{\text{T}}{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{xy}} \right)_k} + {\left( {{\boldsymbol{\hat P}}_{t|t - 1}^x} \right)_k}, \hfill \\
{\boldsymbol{Z}}_t^{\text{T}}{\left( {{{{\boldsymbol{\hat \Omega }}}_t}} \right)_k}{{\boldsymbol{d}}_t} = \left[ {{\boldsymbol{H}}_t^{\text{T}}{{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^y} \right)}_k} + {{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{yx}} \right)}_k}} \right]{{\boldsymbol{y}}_t} \hfill \\
+ \left[ {{\boldsymbol{H}}_t^{\text{T}}{{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{xy}} \right)}_k} + {{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^x} \right)}_k}} \right]{{{\boldsymbol{\tilde x}}}_{t|t - 1}}. \hfill \\
\end{gathered} \right.
\end{align}
with
\begin{align}\label{PPPPxy}
\left\{ \begin{gathered}
{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^x} \right)_k} = {\left[ {{\boldsymbol{P}}_{t|t - 1}^{ - \tfrac{1}{2}}} \right]^{\text{T}}}{\left( {{{{\boldsymbol{\hat \Omega }}}_{t;x}}} \right)_k}{\boldsymbol{P}}_{t|t - 1}^{ - \tfrac{1}{2}}, \hfill \\
{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{xy}} \right)_k} = {\left[ {{\boldsymbol{R}}_t^{ - \tfrac{1}{2}}} \right]^{\text{T}}}{\left( {{{{\boldsymbol{\hat \Omega }}}_{t;xy}}} \right)_k}{\boldsymbol{P}}_{t|t - 1}^{ - \tfrac{1}{2}}, \hfill \\
{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{yx}} \right)_k} = {\left[ {{\boldsymbol{P}}_{t|t - 1}^{ - \tfrac{1}{2}}} \right]^{\text{T}}}{\left( {{{{\boldsymbol{\hat \Omega }}}_{t;yx}}} \right)_k}{\boldsymbol{R}}_t^{ - \tfrac{1}{2}}, \hfill \\
{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^y} \right)_k} = {\left[ {{\boldsymbol{R}}_t^{ - \tfrac{1}{2}}} \right]^{\text{T}}}{\left( {{{{\boldsymbol{\hat \Omega }}}_{t;y}}} \right)_k}{\boldsymbol{R}}_t^{ - \tfrac{1}{2}}. \hfill \\
\end{gathered} \right.
\end{align}
Substituting \eqref{17ZtOtZtTDtOUtT} into \eqref{DtOUtZtTj1zktZ} yields the update of ${\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right)_{k + 1}}$ as
\begin{align}\label{XtKtYtjHx}
{\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right)_{k + 1}} = {{{\boldsymbol{\tilde x}}}_{t|t - 1}} + {\left( {{{{\boldsymbol{\hat K}}}_t}} \right)_k}\left( {{{\boldsymbol{y}}_t} - {{\boldsymbol{H}}_t}{{{\boldsymbol{\tilde x}}}_{t|t - 1}}} \right),
\end{align}
where the Kalman gain is
\begin{align}\label{KgainKj1yjZoTzK}
{\left( {{{{\boldsymbol{\hat K}}}_t}} \right)_k} = {\left[ {{\boldsymbol{Z}}_t^{\text{T}}{{\left( {{{{\boldsymbol{\hat \Omega }}}_t}} \right)}_k}{{\boldsymbol{Z}}_t}} \right]^{ - 1}}\left[ {{{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{yx}} \right)}_k} + {\boldsymbol{H}}_t^{\text{T}}{{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^y} \right)}_k}} \right].
\end{align}
When the condition ${{\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right)}_{k + 1}} - {{\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right)}_k}} \right\|} \mathord{\left/
{\vphantom {{\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right)}_{k + 1}} - {{\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right)}_k}} \right\|} {\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right)}_k}} \right\|}}} \right.
\kern-\nulldelimiterspace} {\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right)}_k}} \right\|}} \leqslant \tau $ is satisfied, the FPI algorithm stops, and ${{\boldsymbol{\hat K}}_t} = {\left( {{{{\boldsymbol{\hat K}}}_t}} \right)_k}$.
The covariance matrix ${{\boldsymbol{P}}_{t|t}}$ can be updated in the following way
\begin{align}\label{pTtIrKtTjTjIdanwei}
{{\boldsymbol{P}}_{t|t}} = \left( {{\boldsymbol{I}} - {{{\boldsymbol{\hat K}}}_t}{{\boldsymbol{H}}_t}} \right){{\boldsymbol{P}}_{t|t - 1}}{\left( {{\boldsymbol{I}} - {{{\boldsymbol{\hat K}}}_t}{{\boldsymbol{H}}_t}} \right)^{\text{T}}} + {{\boldsymbol{\hat K}}_t}{{\boldsymbol{R}}_t}{\boldsymbol{\hat K}}_t^{\text{T}}.
\end{align}
\subsubsection{Backward smoothing procedure}
The backward smoothing procedure is performed after the forward filtering procedure. The objective of the backward smoothing procedure is the smoothing state ${{\boldsymbol{\tilde x}}_{t|T}}\left( {t = T - 1,T - 2, \cdots ,1} \right)$ and covariance
matrices ${{\boldsymbol{P}}_{t|T}}\left( {t = T - 1,T - 2, \cdots ,1} \right)$.
Similar to the forward pass, the backward smoothing problem can be reformulated as a linear regression problem based on the smoothing state ${{{\boldsymbol{\tilde x}}}_{t + 1|T}}$ and the filtering state ${{{\boldsymbol{\tilde x}}}_{t|t}}$. The regression problem takes the form
\begin{align}\label{routxtFtjgat}
\left[ {\begin{array}{*{20}{c}}
{{{{\boldsymbol{\tilde x}}}_{t + 1|T}}} \\
{{{{\boldsymbol{\tilde x}}}_{t|t}}}
\end{array}} \right] = \left[ {\begin{array}{*{20}{c}}
{{{\boldsymbol{F}}_t}{{\boldsymbol{x}}_t}} \\
{{{\boldsymbol{x}}_t}}
\end{array}} \right] + \left[ {\begin{array}{*{20}{c}}
{{{\boldsymbol{\rho }}_t}} \\
{{{\boldsymbol{\gamma }}_t}}
\end{array}} \right],
\end{align}
where ${{\boldsymbol{\gamma }}_t} = {{{\boldsymbol{\tilde x}}}_{t|t}} - {{\boldsymbol{x}}_t}$ denotes the estimation error of ${{{\boldsymbol{\tilde x}}}_{t|t}}$, and ${{\boldsymbol{\rho }}_t} = {{{\boldsymbol{\tilde x}}}_{t + 1|T}} - {{\boldsymbol{F}}_t}{{\boldsymbol{x}}_t}$ represents the prediction error of the state.
Define
\begin{align}
{{\boldsymbol{Y}}_t} = \left[ {\begin{array}{*{20}{c}}
{{\boldsymbol{Q}}_t^{ - \tfrac{1}{2}}}&{\boldsymbol{0}} \\
{\boldsymbol{0}}&{{\boldsymbol{P}}_{t|t}^{ - \tfrac{1}{2}}}
\end{array}} \right].
\end{align}
Both sides of \eqref{routxtFtjgat} are simultaneously multiplied together by ${{\boldsymbol{Y}}_t}$, and we can get
\begin{align}\label{thetadeyuwtxtjia}
{{\boldsymbol{\Theta }}_t} = {{\boldsymbol{W}}_t}{{\boldsymbol{x}}_t} + {{\boldsymbol{\varepsilon }}_t}
\end{align}
with
\begin{align}\label{14xtxiT-12}
{{\boldsymbol{\Theta }}_t} = \left[ {\begin{array}{*{20}{c}}
{{\boldsymbol{Q}}_t^{ - \tfrac{1}{2}}{{{\boldsymbol{\tilde x}}}_{t + 1|T}}} \\
{{\boldsymbol{P}}_{t|t}^{ - \tfrac{1}{2}}{{{\boldsymbol{\tilde x}}}_{t|t}}}
\end{array}} \right],
\end{align}
\begin{align}\label{FP12fqt}
{{\boldsymbol{W}}_t} = \left[ {\begin{array}{*{20}{c}}
{{\boldsymbol{Q}}_t^{ - \tfrac{1}{2}}{{\boldsymbol{F}}_t}} \\
{{\boldsymbol{P}}_{t|t}^{ - \tfrac{1}{2}}}
\end{array}} \right],
\end{align}
and
\begin{align}
{{\boldsymbol{\varepsilon }}_t} = \left[ {\begin{array}{*{20}{c}}
{{\boldsymbol{Q}}_t^{ - \tfrac{1}{2}}}&{\boldsymbol{0}} \\
{\boldsymbol{0}}&{{\boldsymbol{P}}_{t|t}^{ - \tfrac{1}{2}}}
\end{array}} \right]\left[ {\begin{array}{*{20}{c}}
{{{\boldsymbol{\rho }}_t}} \\
{{{\boldsymbol{\gamma }}_t}}
\end{array}} \right].
\end{align}
From \eqref{thetadeyuwtxtjia}, the augmented error can be written as
\begin{align}\label{thetajwtxt}
{{\boldsymbol{\varepsilon }}_t} = {{\boldsymbol{\Theta }}_t} - {{\boldsymbol{W}}_t}{{\boldsymbol{x}}_t}.
\end{align}
Based on the MEE criterion, the optimal estimation of ${{{\boldsymbol{x}}_t}}$ is obtained via
\begin{align}\label{etijetjbbjG}
\begin{gathered}
{{{\boldsymbol{\tilde x}}}_{t|T}} = \arg \mathop {\max }\limits_{{{\boldsymbol{x}}_t}} {J_{MEE}}\left( {{{\boldsymbol{x}}_t}} \right) \hfill \\
= \arg \mathop {\max }\limits_{{{\boldsymbol{x}}_t}} \frac{1}{{{L^2}}}\sum\limits_{i = 1}^L {\sum\limits_{j = 1}^L {G\left( {{\varepsilon _{t;i}} - {\varepsilon _{t;j}}} \right)} } , \hfill \\
\end{gathered}
\end{align}
where ${\varepsilon _{t;i}} = {\theta _{t;i}} - {{\boldsymbol{w}}_{t;i}}{{\boldsymbol{x}}_t}$, ${\varepsilon _{t;i}}$ and ${\theta _{t;i}}$ are the $i$th element of ${{{\boldsymbol{\varepsilon }}_t}}$ and ${{{\boldsymbol{\Theta }}_t}}$, ${{\boldsymbol{w}}_{t;i}}$ represents the $i$th row of ${{\boldsymbol{W}}_t}$, and $L = 2n$ is the dimension of ${{{\boldsymbol{\varepsilon }}_t}}$.
The gradient of the cost function in \eqref{etijetjbbjG} can be expressed as:
\begin{align}\label{16tijbgtTWTLI}
\frac{{\partial {J_{MEE}}\left( {{{\boldsymbol{x}}_t}} \right)}}{{\partial {{\boldsymbol{x}}_t}}} = \frac{{\text{2}}}{{{L^2}{\sigma ^2}}}\sum\limits_{i = 1}^L {\sum\limits_{j = 1}^L {\left[ {{\boldsymbol{w}}_{t;j}^{\text{T}}{{G}}\left( {{\varepsilon _{t;j}} - {\varepsilon _{t;i}}} \right){\varepsilon _{t;j}}} \right]} } \notag\\
- \frac{{\text{2}}}{{{L^2}{\sigma ^2}}}\sum\limits_{i = 1}^L {\sum\limits_{j = 1}^L {\left[ {{\boldsymbol{w}}_{t;j}^{\text{T}}{{G}}\left( {{\varepsilon _{t;j}} - {\varepsilon _{t;i}}} \right){\varepsilon _{t;i}}} \right].} }
\end{align}
To further simplify, Eq. \eqref{16tijbgtTWTLI} can be further written as
\begin{align}\label{17edtbesL}
\frac{{\partial {J_{MEE}}\left( {{{\boldsymbol{x}}_t}} \right)}}{{\partial {{\boldsymbol{x}}_t}}}{\text{ = }}\frac{{\text{2}}}{{{L^2}{\sigma ^2}}}{\boldsymbol{W}}_t^{\text{T}}{{\boldsymbol{\Gamma }}_t}{{\boldsymbol{\varepsilon }}_t} - \frac{{\text{2}}}{{{L^2}{\sigma ^2}}}{\boldsymbol{W}}_t^{\text{T}}{{\boldsymbol{\Lambda }}_t}{{\boldsymbol{\varepsilon }}_t}
\end{align}
with
\begin{align}\label{FFigmatrix}
{\left[ {{{\boldsymbol{\Gamma }}_t}} \right]_{ij}} = \left\{ {\begin{array}{*{20}{l}}
{\sum\limits_{j = 1}^L {{{G}}\left( {{\varepsilon _{t;i}} - {\varepsilon _{t;j}}} \right),i = j,} } \\
{0,{\text{ }}i \ne j,}
\end{array}} \right.
\end{align}
and
\begin{align}\label{Getigjetij}
{\left[ {{{\boldsymbol{\Lambda }}_t}} \right]_{ij}} = {{G}}\left( {{\varepsilon _{t;i}} - {\varepsilon _{t;j}}} \right),
\end{align}
where ${\left[ {{{\boldsymbol{\Gamma }}_t}} \right]_{ij}}$ and ${\left[ {{{\boldsymbol{\Lambda }}_t}} \right]_{ij}}$ are the elements of the matrices ${{\boldsymbol{\Gamma }}_t}$ and ${{\boldsymbol{\Lambda }}_t}$, respectively.
To calculate ${{\boldsymbol{\tilde x}}_{t|T}}$ using \eqref{etijetjbbjG}, Eq. \eqref{17edtbesL} is set to zero, and a numerically stable solution \cite{wang2021numerically} can be obtained
\begin{align}\label{xtbwbfgdfd}
\left\{ {\begin{array}{*{20}{c}}
{{{\boldsymbol{\Gamma }}_t}{{\boldsymbol{\varepsilon }}_t}{\text{ = }}{{\boldsymbol{\Gamma }}_t}\left[ {{{\boldsymbol{\Theta }}_t} - {{\boldsymbol{W}}_t}{{\boldsymbol{x}}_t}} \right]{\text{ = 0,}}} \\
{{{\boldsymbol{\Lambda }}_t}{{\boldsymbol{\varepsilon }}_t}{\text{ = }}{{\boldsymbol{\Lambda }}_t}\left[ {{{\boldsymbol{\Theta }}_t} - {{\boldsymbol{W}}_t}{{\boldsymbol{x}}_t}} \right]{\text{ = 0.}}}
\end{array}} \right.
\end{align}
It is obvious that \eqref{xtbwbfgdfd} can be further written as
\begin{align}\label{wbgsfsdtDBT}
\left[ {\begin{array}{*{20}{c}}
{{{\boldsymbol{\Gamma }}_t}} \\
{{{\boldsymbol{\Lambda }}_t}}
\end{array}} \right]{{\boldsymbol{\Theta }}_t}{\text{ = }}\left[ {\begin{array}{*{20}{c}}
{{{\boldsymbol{\Gamma }}_t}} \\
{{{\boldsymbol{\Lambda }}_t}}
\end{array}} \right]{{\boldsymbol{W}}_t}{{\boldsymbol{x}}_t}.
\end{align}
Multiply both sides of \eqref{wbgsfsdtDBT} using ${\boldsymbol{W}}_t^{\text{T}}\left[ {\begin{array}{*{20}{c}}
{{\boldsymbol{\Gamma }}_t^{\text{T}}}&{{\boldsymbol{\Lambda }}_t^{\text{T}}}
\end{array}} \right]$, and we can get
\begin{align}
{\boldsymbol{W}}_t^{\text{T}}\left( {{\boldsymbol{\Gamma }}_t^{\text{T}}{{\boldsymbol{\Gamma }}_t} + {\boldsymbol{\Lambda }}_t^{\text{T}}{{\boldsymbol{\Lambda }}_t}} \right){{\boldsymbol{\Theta }}_t} = {\boldsymbol{W}}_t^{\text{T}}\left( {{\boldsymbol{\Gamma }}_t^{\text{T}}{{\boldsymbol{\Gamma }}_t} + {\boldsymbol{\Lambda }}_t^{\text{T}}{{\boldsymbol{\Lambda }}_t}} \right){{\boldsymbol{W}}_t}{{\boldsymbol{x}}_t}.
\end{align}
Further, the state of the system ${{\boldsymbol{x}}_t}$ can be obtained by
\begin{align}\label{theasanttTwSAN}
{{\boldsymbol{x}}_t} = {\left( {{\boldsymbol{W}}_t^{\text{T}}{{\boldsymbol{\Xi }}_t}{{\boldsymbol{W}}_t}} \right)^{ - 1}}{\boldsymbol{W}}_t^{\text{T}}{{\boldsymbol{\Xi }}_t}{{\boldsymbol{\Theta }}_t}
\end{align}
with
\begin{align}\label{ffAATT}
{{\boldsymbol{\Xi }}_t} = {\boldsymbol{\Gamma }}_t^{\text{T}}{{\boldsymbol{\Gamma }}_t} + {\boldsymbol{\Lambda }}_t^{\text{T}}{{\boldsymbol{\Lambda }}_t}.
\end{align}
Since ${\left( {{\boldsymbol{W}}_t^{\text{T}}{{\boldsymbol{\Xi }}_t}{{\boldsymbol{W}}_t}} \right)^{ - 1}}{\boldsymbol{W}}_t^{\text{T}}{{\boldsymbol{\Xi }}_t}{{\boldsymbol{\Theta }}_t}$ is a function of ${{\boldsymbol{x}}_t}$, Eq. \eqref{theasanttTwSAN} can be further written as
\begin{align}
{{\boldsymbol{x}}_t} = {\boldsymbol{g}}\left( {{{\boldsymbol{x}}_t}} \right){\text{ = }}{\left( {{\boldsymbol{W}}_t^{\text{T}}{{\boldsymbol{\Xi }}_t}{{\boldsymbol{W}}_t}} \right)^{ - 1}}{\boldsymbol{W}}_t^{\text{T}}{{\boldsymbol{\Xi }}_t}{{\boldsymbol{\Theta }}_t}.
\end{align}
To obtain the numerical solution of ${{\boldsymbol{x}}_t}$, two common methods are often used. One solution in \cite{wang2020maximum,WANG20178659} is to use an approximate replacement method (ARM) that replaces ${{\boldsymbol{x}}_t}$ contained in ${\varepsilon _{t;i}}$ by ${{\boldsymbol{\tilde x}}_{t|t}}$. The other solution is to use the FPI method to calculate the smoothing estimation ${{{\boldsymbol{\tilde x}}}_{t|T}}$, and one can obtain
\begin{align}
{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)_k}{\text{ = }}{\left[ {{\boldsymbol{W}}_t^{\text{T}}{{\left( {{{{\boldsymbol{\hat \Xi }}}_t}} \right)}_k}{{\boldsymbol{W}}_t}} \right]^{ - 1}}{\boldsymbol{W}}_t^{\text{T}}{\left( {{{{\boldsymbol{\hat \Xi }}}_t}} \right)_k}{{\boldsymbol{\Theta }}_t},
\end{align}
where $k$ denotes the number of the FPI method, and the initial value of the FPI is ${\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)_{k = 0}} = {{{\boldsymbol{\tilde x}}}_{t|t}}$. When only one FPI loop is performed, the FPI turns into the ARM in \cite{wang2020maximum}, which indicates that the ARM is a special case of the FPI method.
Whether FPI or ARM is used, ${{\boldsymbol{x}}_t}$ in (14) will be replaced by ${{{\boldsymbol{\tilde x}}}_{t|t}}$. The superscript $ \wedge $ is used to indicate the variable that has changed as a result of this substitution such as ${\hat \varepsilon _{t;i}}$, ${{\boldsymbol{\hat \Xi }}_t}$ etc.
To simplify the following derivation, ${\left( {{{{\boldsymbol{\hat \Xi }}}_t}} \right)_k}$ is written in the form of a block matrix as follows
\begin{align}\label{ssssttt11122122}
{\left( {{{{\boldsymbol{\hat \Xi }}}_t}} \right)_k}{\text{ = }}\left[ {\begin{array}{*{20}{c}}
{{{\left( {{{{\boldsymbol{\hat \Xi }}}_{t;{x_1}}}} \right)}_k}}&{{{\left( {{{{\boldsymbol{\hat \Xi }}}_{t;{x_1}{x_2}}}} \right)}_k}} \\
{{{\left( {{{{\boldsymbol{\hat \Xi }}}_{t;{x_2}{x_1}}}} \right)}_k}}&{{{\left( {{{{\boldsymbol{\hat \Xi }}}_{t;{x_2}}}} \right)}_k}}
\end{array}} \right],
\end{align}
where ${{\boldsymbol{\hat \Xi }}_{t;{x_1}}} \in {\mathbb{R}^{n \times n}}$, ${{\boldsymbol{\hat \Xi }}_{t;{x_1}{x_2}}} \in {\mathbb{R}^{n \times n}}$, ${{\boldsymbol{\hat \Xi }}_{t;{x_2}{x_1}}} \in {\mathbb{R}^{n \times n}}$, and ${{\boldsymbol{\hat \Xi }}_{t;{x_2}}} \in {\mathbb{R}^{n \times n}}$.
Substitute \eqref{FP12fqt} and \eqref{ssssttt11122122} into ${\boldsymbol{W}}_t^{\text{T}}{({{{\boldsymbol{\hat \Xi }}}_t})_k}{{\boldsymbol{W}}_t}$ yields
\begin{align}\label{30wwswtT}
\begin{gathered}
{\boldsymbol{W}}_t^{\text{T}}{\left( {{{{\boldsymbol{\hat \Xi }}}_t}} \right)_k}{{\boldsymbol{\Theta }}_t} = \left[ {{\boldsymbol{F}}_t^{\text{T}}{{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_1}}} \right)}_k} + {{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_2}{x_1}}} \right)}_k}} \right] \hfill \\
\times {{{\boldsymbol{\tilde x}}}_{t + 1|T}} + \left[ {{\boldsymbol{F}}_t^{\text{T}}{{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_1}{x_2}}} \right)}_k} + {{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_2}}} \right)}_k}} \right]{{{\boldsymbol{\tilde x}}}_{t|t}}. \hfill \\
\end{gathered}
\end{align}
with
\begin{align}\label{27PPPPx12}
\left\{ \begin{gathered}
{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_1}}} \right)_k} = {\left( {{\boldsymbol{Q}}_t^{ - \tfrac{1}{2}}} \right)^{\text{T}}}{\left( {{{{\boldsymbol{\hat \Xi }}}_{t;{x_1}}}} \right)_k}{\boldsymbol{Q}}_t^{ - \tfrac{1}{2}}, \hfill \\
{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_2}}} \right)_k} = {\left( {{\boldsymbol{P}}_{t|t}^{ - \tfrac{1}{2}}} \right)^{\text{T}}}{\left( {{{{\boldsymbol{\hat \Xi }}}_{t;{x_2}}}} \right)_k}{\boldsymbol{P}}_{t|t}^{ - \tfrac{1}{2}}, \hfill \\
{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_1}{x_2}}} \right)_k} = {\left( {{\boldsymbol{Q}}_t^{ - \tfrac{1}{2}}} \right)^{\text{T}}}{\left( {{{{\boldsymbol{\hat \Xi }}}_{t;{x_1}{x_2}}}} \right)_k}{\boldsymbol{P}}_{t|t}^{ - \tfrac{1}{2}}, \hfill \\
{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_2}{x_1}}} \right)_k} = {\left( {{\boldsymbol{P}}_{t|t}^{ - \tfrac{1}{2}}} \right)^{\text{T}}}{\left( {{{{\boldsymbol{\hat \Xi }}}_{t;{x_2}{x_1}}}} \right)_k}{\boldsymbol{Q}}_t^{ - \tfrac{1}{2}}. \hfill \\
\end{gathered} \right.
\end{align}
Similar to the process of obtaining \eqref{30wwswtT}, Eq. \eqref{DbtStTWbt} can be obtained by substituting \eqref{14xtxiT-12}, \eqref{FP12fqt}, and \eqref{ssssttt11122122} into ${\boldsymbol{W}}_t^{\text{T}}{({{{\boldsymbol{\hat \Xi }}}_t})_k}{{\boldsymbol{\Theta }}_t}$
\begin{align}\label{DbtStTWbt}
\begin{gathered}
{\boldsymbol{W}}_t^{\text{T}}{\left( {{{{\boldsymbol{\hat \Xi }}}_t}} \right)_k}{{\boldsymbol{\Theta }}_t} = \left[ {{\boldsymbol{F}}_t^{\text{T}}{{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_1}}} \right)}_k} + {{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_2}{x_1}}} \right)}_k}} \right] \hfill \\
\times {{{\boldsymbol{\tilde x}}}_{t + 1|T}} + \left[ {{\boldsymbol{F}}_t^{\text{T}}{{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_1}{x_2}}} \right)}_k} + {{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_2}}} \right)}_k}} \right]{{{\boldsymbol{\tilde x}}}_{t|t}}. \hfill \\
\end{gathered}
\end{align}
The matrix inversion lemma \cite{alexander2012adaptive} is utilized, and the following definitions
\begin{align}
\left\{ \begin{gathered}
{{\boldsymbol{A}}_k} = \left[ {{{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_2}}} \right)}_k} + {\boldsymbol{F}}_t^{\text{T}}{{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_1}{x_2}}} \right)}_k}} \right], \hfill \\
{{\boldsymbol{B}}_k} = \left[ {{\boldsymbol{F}}_t^{\text{T}}{{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_1}}} \right)}_k} + {{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_2}{x_1}}} \right)}_k}} \right], \hfill \\
{\boldsymbol{C}} = {\boldsymbol{I}}, \hfill \\
{\boldsymbol{D}} = {{\boldsymbol{F}}_t}. \hfill \\
\end{gathered} \right.
\end{align}
${\left[ {{\boldsymbol{W}}_t^{\text{T}}{{\left( {{{{\boldsymbol{\hat \Xi }}}_t}} \right)}_k}{{\boldsymbol{W}}_t}} \right]^{ - 1}}$ can be further written as
\begin{align}\label{nfAKBTjajltb}
{\left[ {{\boldsymbol{W}}_t^{\text{T}}{{\left( {{{{\boldsymbol{\hat \Xi }}}_t}} \right)}_k}{{\boldsymbol{W}}_t}} \right]^{ - 1}} = {\left( {{\boldsymbol{L}}_t^b} \right)_k} = {\boldsymbol{A}}_k^{ - 1} - {\left( {{\boldsymbol{\hat K}}_t^b} \right)_k}{{\boldsymbol{F}}_t}{\boldsymbol{A}}_k^{ - 1}
\end{align}
with
\begin{align}\label{FABJIBAk}
{\left( {{\boldsymbol{\hat K}}_t^b} \right)_k} = {\boldsymbol{A}}_k^{ - 1}{{\boldsymbol{B}}_k}{\left( {{\boldsymbol{I}} + {{\boldsymbol{F}}_t}{\boldsymbol{A}}_k^{ - 1}{{\boldsymbol{B}}_k}} \right)^{ - 1}},
\end{align}
where ${\left( {{\boldsymbol{\hat K}}_t^b} \right)_k}$ is defined as the smoothing gain.
From \eqref{FABJIBAk}, one can obtain ${\left( {{\boldsymbol{\hat K}}_t^b} \right)_k}\left( {{\boldsymbol{I}} + {{\boldsymbol{F}}_t}{\boldsymbol{A}}_k^{ - 1}{{\boldsymbol{B}}_k}} \right){\text{ = }}{\boldsymbol{A}}_k^{ - 1}{{\boldsymbol{B}}_k}$. Multiply ${{\boldsymbol{B}}_k}$ on both sides of \eqref{nfAKBTjajltb}, and we can obtain
\begin{align}\label{31KbtABFK}
\begin{gathered}
{\left( {{\boldsymbol{L}}_t^b} \right)_k}{{\boldsymbol{B}}_k} = {\boldsymbol{A}}_k^{ - 1}{{\boldsymbol{B}}_k} - {\left( {{\boldsymbol{\hat K}}_t^b} \right)_k}{{\boldsymbol{F}}_t}{\boldsymbol{A}}_k^{ - 1}{{\boldsymbol{B}}_k} \hfill \\
= {\left( {{\boldsymbol{\hat K}}_t^b} \right)_k} + {\left( {{\boldsymbol{\hat K}}_t^b} \right)_k}{{\boldsymbol{F}}_t}{\boldsymbol{A}}_k^{ - 1}{{\boldsymbol{B}}_k} - {\left( {{\boldsymbol{\hat K}}_t^b} \right)_k}{{\boldsymbol{F}}_t}{\boldsymbol{A}}_k^{ - 1}{{\boldsymbol{B}}_k} \hfill \\
= {\left( {{\boldsymbol{\hat K}}_t^b} \right)_k}. \hfill \\
\end{gathered}
\end{align}
Combining \eqref{nfAKBTjajltb}, \eqref{DbtStTWbt}, and \eqref{31KbtABFK} leads to the updated strategy of ${{{\boldsymbol{\tilde x}}}_t}$
\begin{align}\label{38xtTxmao}
\begin{gathered}
{{\boldsymbol{x}}_{t|T}} = {\left[ {{\boldsymbol{W}}_t^{\text{T}}{{\left( {{{{\boldsymbol{\hat \Xi }}}_t}} \right)}_k}{{\boldsymbol{W}}_t}} \right]^{ - 1}}{\boldsymbol{W}}_t^{\text{T}}{\left( {{{{\boldsymbol{\hat \Xi }}}_t}} \right)_k}{{\boldsymbol{\Theta }}_t} \hfill \\
= {\left( {{\boldsymbol{L}}_t^b} \right)_k}{{\boldsymbol{B}}_k}{{{\boldsymbol{\tilde x}}}_{t + 1|T}} + {\left( {{\boldsymbol{L}}_t^b} \right)_k}{{\boldsymbol{A}}_k}{{{\boldsymbol{\tilde x}}}_{t|t}} \hfill \\
= {\left( {{\boldsymbol{\hat K}}_t^b} \right)_k}{{{\boldsymbol{\tilde x}}}_{t + 1|T}} + \left[ {{\boldsymbol{A}}_k^{ - 1} - {{\left( {{\boldsymbol{\hat K}}_t^b} \right)}_k}{{\boldsymbol{F}}_t}{\boldsymbol{A}}_k^{ - 1}} \right]{{\boldsymbol{A}}_k}{{{\boldsymbol{\tilde x}}}_{t|t}} \hfill \\
= {{{\boldsymbol{\tilde x}}}_{t|t}} + {\left( {{\boldsymbol{\hat K}}_t^b} \right)_k}\left( {{{{\boldsymbol{\tilde x}}}_{t + 1|T}} - {{{\boldsymbol{\tilde x}}}_{t + 1|t}}} \right). \hfill \\
\end{gathered}
\end{align}
where ${\left( {{\boldsymbol{\hat K}}_t^b} \right)_k}$ is
\begin{align}\label{gainmeesmoother}
\begin{gathered}
{\left( {{\boldsymbol{\hat K}}_t^b} \right)_k} = {\left[ {{\boldsymbol{W}}_t^{\text{T}}{{\left( {{{{\boldsymbol{\hat \Xi }}}_t}} \right)}_k}{{\boldsymbol{W}}_t}} \right]^{ - 1}} \hfill \\
\times \left[ {{\boldsymbol{F}}_t^{\text{T}}{{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_1}}} \right)}_k} + {{\left( {{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_2}{x_1}}} \right)}_k}} \right]. \hfill \\
\end{gathered}
\end{align}
When condition ${{\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)}_{k + 1}} - {{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)}_k}} \right\|} \mathord{\left/
{\vphantom {{\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)}_{k + 1}} - {{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)}_k}} \right\|} {\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)}_k}} \right\|}}} \right.
\kern-\nulldelimiterspace} {\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)}_k}} \right\|}} \leqslant \tau $ is satisfied, the the FPI algorithm stops, and ${\boldsymbol{\hat K}}_t^b = {\left( {{\boldsymbol{\hat K}}_t^b} \right)_k}$.
The error covariance of smoothed estimate ${{{\boldsymbol{\tilde x}}}_{t|T}}$ can be obtained by
\begin{align}\label{tKTtPtJ1IjN}
{{\boldsymbol{P}}_{t|T}} = {{\boldsymbol{P}}_{t|t}} + {\boldsymbol{K}}_t^b\left( {{{\boldsymbol{P}}_{t + 1|N}} - {{\boldsymbol{P}}_{t + 1|t}}} \right){\left( {{\boldsymbol{K}}_t^b} \right)^{\text{T}}}.
\end{align}
The detailed derivation of \eqref{tKTtPtJ1IjN} is shown in \ref{deadapin}.
According to the above derivation, the pseudo-code of the MEE-RTS smoother is summarized in Algorithm \ref{MEERTS-Smoother}.
\begin{algorithm}[t]
\caption{RTS smoother based on the MEE} \label{MEERTS-Smoother}
\LinesNumbered
\KwIn{state transition matrix ${\boldsymbol{F}_t}$, measurement matrix ${\boldsymbol{H}_t}$, covariance matrices ${\boldsymbol{Q}_t}$, ${\boldsymbol{R}_t}$, and kernel bandwidth $\sigma$}
\KwOut{${{\boldsymbol{\tilde x}}_{t|T}}$}
\textbf{Initialization}: ${{\boldsymbol{\tilde x}}_{0|0}} = {{\boldsymbol{x}}_0}$, ${{\boldsymbol{P}}_{0|0}} = {{\boldsymbol{P}}_0}$\;
\textbf{Forward Pass:}\
\For{$t\leftarrow $1$ $ \KwTo $T$}{
Calculate ${{\boldsymbol{\tilde x}}_{t|t - 1}}$ and ${{\boldsymbol{P}}_{t|t - 1}}$ using
\begin{subequations}\label{prespttj1dqt}
\begin{numcases}{}
{{{\boldsymbol{\tilde x}}}_{t|t - 1}} = {{\boldsymbol{F}}_t}{{{\boldsymbol{\tilde x}}}_{t - 1|t - 1}},\\\label{tr71qkj1axkjej}
{{\boldsymbol{P}}_{t|t - 1}} = {{\boldsymbol{F}}_t}{{\boldsymbol{P}}_{t - 1|t - 1}}{\boldsymbol{F}}_t^{\text{T}} + {{\boldsymbol{Q}}_t};\label{vkjxkCdYk}
\end{numcases}
\end{subequations}
\While{${{\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right)}_{k + 1}} - {{\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right)}_k}} \right\|} \mathord{\left/
{\vphantom {{\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right)}_{k + 1}} - {{\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right)}_k}} \right\|} {\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right)}_k}} \right\|}}} \right.
\kern-\nulldelimiterspace} {\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right)}_k}} \right\|}} \geqslant \tau $}{Update ${{{\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right)}_{k + 1}}}$ using \eqref{xtzTjTddeyTe}, \eqref{outfaizhzTjtfaizh}, \eqref{PPPPxy}, \eqref{XtKtYtjHx}, and \eqref{KgainKj1yjZoTzK};}
Update ${{\boldsymbol{P}}_{t|t}}$ using \eqref{pTtIrKtTjTjIdanwei}\;
}
\textbf{Backward Pass:}\
\For{$t\leftarrow $T - 1$ $ \KwTo $1$}{
Calculate ${{{{\boldsymbol{\tilde x}}}_{t + 1|t}}}$ and ${{{{\boldsymbol{P}}_{t + 1|t}}}}$ utilizing \eqref{prespttj1dqt}\;
Let $k=1$ and ${\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)_{k = 0}} = {{{\boldsymbol{\tilde x}}}_{t|t}}$\;
\While{${{\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)}_{k + 1}} - {{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)}_k}} \right\|} \mathord{\left/
{\vphantom {{\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)}_{k + 1}} - {{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)}_k}} \right\|} {\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)}_k}} \right\|}}} \right.
\kern-\nulldelimiterspace} {\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)}_k}} \right\|}} \geqslant \tau $}{
Update ${\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)_{k + 1}}$ using \eqref{38xtTxmao}
with \eqref{gainmeesmoother}, \eqref{ssssttt11122122}, \eqref{27PPPPx12}, and \eqref{thetajwtxt};
}
Calculate ${{\boldsymbol{P}}_{t|T}}$ utilizing \eqref{tKTtPtJ1IjN}\;
}
\end{algorithm}
\begin{algorithm}[t]
\caption{Extended RTS smoother based on the MEE} \label{MEERTS-ExtendedSmoother}
\LinesNumbered
\KwIn{${\boldsymbol{f}}\left( \cdot \right)$, ${\boldsymbol{h}}\left( \cdot \right)$, ${\boldsymbol{Q}_t}$, ${\boldsymbol{R}_t}$, and $\sigma$}
\KwOut{${{\boldsymbol{\tilde x}}_{t|T}}$}
\textbf{Initialization}: ${{\boldsymbol{\tilde x}}_{0|0}} = {{\boldsymbol{x}}_0}$, ${{\boldsymbol{P}}_{0|0}} = {{\boldsymbol{P}}_0}$\;
\textbf{Forward Pass:}\
\For{$t\leftarrow $1$ $ \KwTo $T$}{
These steps of the MEE-EKF algorithm are presented in \cite{chen2019minimum}.
}
\textbf{Backward Pass:}\
\For{$t\leftarrow $T - 1$ $ \KwTo $1$}{
Predicting the state of the system using ${{\boldsymbol{\tilde x}}_{t + 1|t}} = {\boldsymbol{f}}\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right)$\;
Calculate the Jacobian matrix and covariance matrix using \eqref{yjkbjuzhenF} and ${{\boldsymbol{P}}_{t|t - 1}} = {{\boldsymbol{F}}_{t - 1}}{{\boldsymbol{P}}_{t - 1|t - 1}}{\boldsymbol{F}}_{t - 1}^{\text{T}} + {\boldsymbol{Q}}$ \;
Let $k=1$ and ${\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)_{k = 0}} = {{{\boldsymbol{\tilde x}}}_{t|t}}$\;
\While{${{\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)}_{k + 1}} - {{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)}_k}} \right\|} \mathord{\left/
{\vphantom {{\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)}_{k + 1}} - {{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)}_k}} \right\|} {\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)}_k}} \right\|}}} \right.
\kern-\nulldelimiterspace} {\left\| {{{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)}_k}} \right\|}} \geqslant \tau $}{
Update ${\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)_{k + 1}}$ using ${\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)_{k + 1}} = {{\boldsymbol{\tilde x}}_{t|t}} + {\boldsymbol{\hat K}}_t^b\left( {{{{\boldsymbol{\tilde x}}}_{t + 1|T}} - {{{\boldsymbol{\tilde x}}}_{t + 1|t}}} \right)$
with $\left\{ \begin{gathered}
{\boldsymbol{\hat K}}_t^b = {\left( {{\boldsymbol{W}}_t^{\text{T}}{{{\boldsymbol{\hat \Xi }}}_t}{{\boldsymbol{W}}_t}} \right)^{ - 1}}\left( {{\boldsymbol{F}}_t^{\text{T}}{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_1}} + {\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_2}{x_1}}} \right), \hfill \\
{{{\boldsymbol{\hat \Xi }}}_t} = {\boldsymbol{\hat \Gamma }}_t^{\text{T}}{{{\boldsymbol{\hat \Gamma }}}_t} + {\boldsymbol{\hat \Lambda }}_t^{\text{T}}{{{\boldsymbol{\hat \Lambda }}}_t},{\left[ {{{{\boldsymbol{\hat \Lambda }}}_t}} \right]_{ij}} = {\text{G}}\left[ {\hat \varepsilon _{t;i}^l - \hat \varepsilon _{t;j}^l} \right], \hfill \\
{\left[ {{{{\boldsymbol{\hat \Gamma }}}_t}} \right]_{ij}} = \left\{ {\begin{array}{*{20}{l}}
{\sum\limits_{j = 1}^L {{{\text{G}}_\sigma }\left[ {\hat \varepsilon _{t;i}^l - \hat \varepsilon _{t;j}^l} \right],i = j,} } \\
{0,i \ne j,}
\end{array}} \right. \hfill \\
{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_1}} = {\left( {{\boldsymbol{Q}}_t^{ - \tfrac{1}{2}}} \right)^{\text{T}}}{{{\boldsymbol{\hat \Xi }}}_{t;{x_1}}}{\boldsymbol{Q}}_t^{ - \tfrac{1}{2}}, \hfill \\
{{\hat \varepsilon }_{t;i}} = {\theta _{t;i}} - {{\boldsymbol{w}}_{t;i}}{\left( {{{{\boldsymbol{\tilde x}}}_{t|T}}} \right)_k}, \hfill \\
{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_2}{x_1}} = {\left( {{\boldsymbol{P}}_{t|t}^{ - \tfrac{1}{2}}} \right)^{\text{T}}}{{{\boldsymbol{\hat \Xi }}}_{t;{x_2}{x_1}}}{\boldsymbol{Q}}_t^{ - \tfrac{1}{2}}, \hfill \\
\end{gathered} \right.$
}
Update covariance matrix using
\begin{align}
{{\boldsymbol{P}}_{t|T}} \approx {{\boldsymbol{P}}_{t|t}} + {\boldsymbol{\hat K}}_t^b\left( {{{\boldsymbol{P}}_{t + 1|N}} - {{\boldsymbol{P}}_{t + 1|t}}} \right){\left( {{\boldsymbol{\hat K}}_t^b} \right)^{\text{T}}}
\end{align}
}
\end{algorithm}
\subsection{The MEE based extended RTS smoother}
For state estimation of the nonlinear systems, the process and measurement equations can be described as
\begin{align}\label{nonlinear}
\left\{ \begin{gathered}
{{\boldsymbol{x}}_t} = {\boldsymbol{f}}\left( {{{\boldsymbol{x}}_{t - 1}}} \right) + {{\boldsymbol{q}}_{t - 1}}, \hfill \\
{{\boldsymbol{y}}_t} = {\boldsymbol{h}}\left( {{{\boldsymbol{x}}_t}} \right) + {{\boldsymbol{r}}_t}. \hfill \\
\end{gathered} \right.
\end{align}
where mapping ${\boldsymbol{f}}\left( \cdot \right)$ denotes a known nonlinear state transition function, mapping ${\boldsymbol{h}}\left( \cdot \right)$ is a known nonlinear measurement function. The ERTS is a frequently utilized tool for the state smoothing problems for the nonlinear systems. In the ERTS, the nonlinear functions ${\boldsymbol{f}}\left( \cdot \right)$ is linearized by the first-order Taylor series expansion at ${{{{\boldsymbol{\tilde x}}}_{t - 1|t - 1}}}$, i.e.,
\begin{align}\label{34Htxjxmtstj1}
{\boldsymbol{f}}\left( {{{\boldsymbol{x}}_{t - 1}}} \right) \approx {\boldsymbol{f}}\left( {{{{\boldsymbol{\tilde x}}}_{t - 1|t - 1}}} \right) + {{\boldsymbol{F}}_{t - 1}}\left( {{{\boldsymbol{x}}_{t - 1}} - {{{\boldsymbol{\tilde x}}}_{t - 1|t - 1}}} \right),
\end{align}
where the Jacobian matrix of the nonlinear function ${\boldsymbol{f}}\left( \cdot \right)$ is obtained via
\begin{align}\label{yjkbjuzhenF}
{{\boldsymbol{F}}_{t - 1}} = {\left. {\frac{{\partial {\boldsymbol{f}}\left( {{{\boldsymbol{x}}_{t - 1}}} \right)}}{{\partial {\boldsymbol{x}}}}} \right|_{{\boldsymbol{x}} = {{{\boldsymbol{\tilde x}}}_{t - 1|t - 1}}}}.
\end{align}
Combining \eqref{nonlinear} and \eqref{34Htxjxmtstj1} yields:
\begin{align}\label{xqxhhy}
{{\boldsymbol{x}}_t} \approx {\boldsymbol{f}}\left( {{{{\boldsymbol{\tilde x}}}_{t - 1|t - 1}}} \right) + {{\boldsymbol{F}}_{t - 1}}\left( {{{\boldsymbol{x}}_{t - 1}} - {{{\boldsymbol{\tilde x}}}_{t - 1|t - 1}}} \right) + {{\boldsymbol{q}}_{t - 1}}.
\end{align}
Substitute \eqref{xqxhhy} into \eqref{nonlinear} and the augmented model for the MEE-ERTS can be written as
\begin{align}
\begin{gathered}
{\boldsymbol{\varepsilon }}_t^l = \left[ {\begin{array}{*{20}{c}}
{{\boldsymbol{Q}}_t^{ - \tfrac{1}{2}}}&{{{\boldsymbol{0}}_n}} \\
{{{\boldsymbol{0}}_n}}&{{\boldsymbol{P}}_{t|t}^{ - \tfrac{1}{2}}}
\end{array}} \right]\left[ {\begin{array}{*{20}{c}}
{{{{\boldsymbol{\tilde x}}}_{t + 1|T}} - {\boldsymbol{f}}\left( {{{\boldsymbol{x}}_t}} \right)} \\
{{{{\boldsymbol{\tilde x}}}_{t|t}} - {{\boldsymbol{x}}_t}}
\end{array}} \right] \hfill \\
= \left[ {\begin{array}{*{20}{c}}
{{\boldsymbol{Q}}_t^{ - \tfrac{1}{2}}}&{{{\boldsymbol{0}}_n}} \\
{{{\boldsymbol{0}}_n}}&{{\boldsymbol{P}}_{t|t}^{ - \tfrac{1}{2}}}
\end{array}} \right]\left[ {\begin{array}{*{20}{c}}
\begin{gathered}
{{{\boldsymbol{\tilde x}}}_{t + 1|T}} - {\boldsymbol{f}}\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right) - \hfill \\
{{\boldsymbol{F}}_t}\left( {{{\boldsymbol{x}}_t} - {{{\boldsymbol{\tilde x}}}_{t|t}}} \right) \hfill \\
\end{gathered} \\
{{{{\boldsymbol{\tilde x}}}_{t|t}} - {{\boldsymbol{x}}_t}}
\end{array}} \right]. \hfill \\
\end{gathered}
\end{align}
and we also obtain
\begin{align}\label{DbEtWbExtjebt}
{\boldsymbol{\hat \varepsilon }}_t^l = {\boldsymbol{\Theta }}_t^l - {\boldsymbol{W}}_t^l{\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right)_k}
\end{align}
with
\begin{align}\label{47xmaottgPje21}
{\boldsymbol{\Theta }}_t^l = \left[ {\begin{array}{*{20}{c}}
{{\boldsymbol{Q}}_t^{ - \tfrac{1}{2}}\left[ {{{{\boldsymbol{\tilde x}}}_{t + 1|T}} - {\boldsymbol{f}}\left( {{{{\boldsymbol{\tilde x}}}_{t|t}}} \right) + {{\boldsymbol{F}}_t}{{{\boldsymbol{\tilde x}}}_{t|t}}} \right]} \\
{{\boldsymbol{P}}_{t|t}^{ - \tfrac{1}{2}}{{{\boldsymbol{\tilde x}}}_{t|t}}}
\end{array}} \right]
\end{align}
and
\begin{align}\label{4884pttgfue12jP}
{\boldsymbol{W}}_t^l = \left[ {\begin{array}{*{20}{c}}
{{\boldsymbol{Q}}_t^{ - \tfrac{1}{2}}{{\boldsymbol{F}}_t}} \\
{{\boldsymbol{P}}_{t|t}^{ - \tfrac{1}{2}}}
\end{array}} \right].
\end{align}
It is apparent that \eqref{DbEtWbExtjebt} have a similar form as the one in \eqref{thetajwtxt}. Hence, the derivation process of the backward pass of the MEE-ERTS smoother is similar to that of the backward pass of the MEE-RTS smoother. In order to avoid the repetition, the derivation process of the MEE-ERTS smoother is omitted and its pseudo-code is presented in Algorithm \ref{MEERTS-ExtendedSmoother}.
\begin{remark}
It is notable that the threshold $\tau$, in both Algorithm \ref{MEERTS-Smoother} and Algorithm \ref{MEERTS-ExtendedSmoother}, is a vital parameter to control the number of the FPI method. The larger the threshold is, the fewer the number of FPI are required, and therefore the less computational burden of the algorithms, which in turn reduces the performance of the proposed smoothers in steady-state errors to some extent, whereas a smaller threshold means the opposite.
\end{remark}
\section{Theoretical analysis}\label{tanluysis}
\subsection{Mean error behavior}
In this part, ${{\boldsymbol{\xi }}_{s;t}} = {{\boldsymbol{x}}_t} - {{\boldsymbol{\tilde x}}_{t|T}}$ is employed to present the estimation error of smoother, and ${{\boldsymbol{\xi }}_{f;t}} = {{\boldsymbol{x}}_t} - {{\boldsymbol{\tilde x}}_{t|t}}$ denotes the estimation error of filter. According to \eqref{31KbtABFK}, ${{\boldsymbol{\xi }}_{s;t}}$ can be rewritten as:
\begin{align}\label{kyplslongktb}
{{\boldsymbol{\xi }}_{s;t}} &= {{\boldsymbol{x}}_t} - {{{\boldsymbol{\tilde x}}}_{t|T}}\notag\\
&= {{\boldsymbol{\xi }}_{f;t}} - {\boldsymbol{\hat K}}_t^b\left( {{{{\boldsymbol{\tilde x}}}_{t + 1|T}} - {{{\boldsymbol{\tilde x}}}_{t + 1|t}}} \right).
\end{align}
According to ${{{\boldsymbol{\tilde x}}}_{t|t - 1}} = {{\boldsymbol{F}}_t}{{{\boldsymbol{\tilde x}}}_{t - 1|t - 1}}$, ${{\boldsymbol{\xi }}_{f;t - 1}} = {{\boldsymbol{x}}_{t - 1}} - {{{\boldsymbol{\tilde x}}}_{t - 1|t - 1}}$, and ${{\boldsymbol{x}}_t} = {{\boldsymbol{F}}_t}{{\boldsymbol{x}}_{t - 1}} + {{\boldsymbol{q}}_t}$, one can obtain
\begin{align}\label{51sftjj1jianQ2}
{{{\boldsymbol{\tilde x}}}_{t + 1|t}} = {{\boldsymbol{x}}_{t + 1}} - {{\boldsymbol{q}}_{t + 1}} - {{\boldsymbol{F}}_t}{{\boldsymbol{\xi }}_{f;t}}.
\end{align}
Substituting \eqref{51sftjj1jianQ2} into \eqref{kyplslongktb} yields
\begin{align}\label{50qtjekbtssa}
{{\boldsymbol{\xi }}_{s;t}} = \left( {{\boldsymbol{I}} - {\boldsymbol{\hat K}}_t^b{{\boldsymbol{F}}_t}} \right){{\boldsymbol{\xi }}_{f;t}} + {\boldsymbol{\hat K}}_t^b{{\boldsymbol{\xi }}_{s;t + 1}} - {\boldsymbol{\hat K}}_t^b{{\boldsymbol{q}}_{t + 1}}.
\end{align}
Since the process and measurement noises are zero mean, the expectation of estimation error ${{\boldsymbol{\xi }}_{s;t}}$ can be expressed as
\begin{align}\label{51sanstj1fenha}
{\text{E}}\left[ {{{\boldsymbol{\xi }}_{s;t}}} \right] = \left( {{\boldsymbol{I}} - {\boldsymbol{\hat K}}_t^b{{\boldsymbol{F}}_t}} \right){\text{E}}\left[ {{{\boldsymbol{\xi }}_{f;t}}} \right]{\text{ + }}{\boldsymbol{\hat K}}_t^b{\text{E}}\left[ {{{\boldsymbol{\xi }}_{s;t + 1}}} \right].
\end{align}
From \eqref{27PPPPx12}, it is obvious that ${{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_1}}}$ and ${{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_2}}}$ are both strictly diagonally dominant. One can obtain
\begin{align}\label{52KbtFyudengyu}
{\boldsymbol{\hat K}}_t^b{{\boldsymbol{F}}_t} \approx {\left( {{\boldsymbol{F}}_t^{\text{T}}{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_1}}{{\boldsymbol{F}}_t} + {\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_2}}} \right)^{ - 1}}\left( {{\boldsymbol{F}}_t^{\text{T}}{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_1}}{{\boldsymbol{F}}_t}} \right).
\end{align}
Substituting \eqref{52KbtFyudengyu} into \eqref{51sanstj1fenha} yields
\begin{align}\label{Essfhtyudyfni}
\begin{gathered}
{\text{E}}\left[ {{{\boldsymbol{\xi }}_{s;t}}} \right] \approx \left[ {{\boldsymbol{I}} - {{\left( {{\boldsymbol{F}}_t^{\text{T}}{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_1}}{{\boldsymbol{F}}_t} + {\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_2}}} \right)}^{ - 1}}\left( {{\boldsymbol{F}}_t^{\text{T}}{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_1}}{{\boldsymbol{F}}_t}} \right)} \right] \hfill \\
\times {\text{E}}\left[ {{{\boldsymbol{\xi }}_{f;t}}} \right] + \left[ {{{\left( {{\boldsymbol{F}}_t^{\text{T}}{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_1}}{{\boldsymbol{F}}_t} + {\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_2}}} \right)}^{ - 1}}\left( {{\boldsymbol{F}}_t^{\text{T}}{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_1}}{{\boldsymbol{F}}_t}} \right)} \right] \times \hfill \\
{\boldsymbol{F}}_t^{ - 1}{\text{E}}\left[ {{{\boldsymbol{\xi }}_{s;t + 1}}} \right]. \hfill \\
\end{gathered}
\end{align}
From \eqref{Essfhtyudyfni}, one can obtain that ${\text{E}}\left[ {{{\boldsymbol{\xi }}_{s;t}}} \right]$ is the linear combination of ${\text{E}}\left[ {{{\boldsymbol{\xi }}_{f;t}}} \right]$ and ${\text{E}}\left[ {{{\boldsymbol{\xi }}_{s;t + 1}}} \right]$. Assuming that ${{\boldsymbol{F}}_t}$ is stable, one can conclude that both ${\boldsymbol{F}}_t^{\text{T}}{\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_1}}{{\boldsymbol{F}}_t}$ and ${\boldsymbol{\hat P}}_{t|t - 1}^{b;{x_2}}$ are positive semi-definite, which will lead to the fact that \eqref{52KbtFyudengyu} is stable. Further, \eqref{Essfhtyudyfni} is also stable, which means that the proposed MEE-RTS smoother operates stably.
\subsection{Mean square error behavior}
According to \eqref{50qtjekbtssa}, the error covariance matrix of the state can be written as
\begin{align}\label{EsansanQFBT}
\begin{gathered}
{\text{E}}\left[ {{{\boldsymbol{\xi }}_{s;t}}{\boldsymbol{\xi }}_{s;t}^{\text{T}}} \right] = \hfill \\
\left( {{\boldsymbol{I}} - {\text{E}}\left[ {{\boldsymbol{\hat K}}_t^b} \right]{{\boldsymbol{F}}_t}} \right){\text{E}}\left[ {{{\boldsymbol{\xi }}_{f;t}}{\boldsymbol{\xi }}_{f;t}^{\text{T}}} \right]{\left( {{\boldsymbol{I}} - {\text{E}}\left[ {{\boldsymbol{\hat K}}_t^b} \right]{{\boldsymbol{F}}_t}} \right)^{\text{T}}} \hfill \\
+ {\text{E}}\left[ {{\boldsymbol{\hat K}}_t^b} \right]{\text{E}}\left[ {{{\boldsymbol{\xi }}_{s;t + 1}}{\boldsymbol{\xi }}_{s;t + 1}^{\text{T}}} \right]{\text{E}}\left[ {{{\left( {{\boldsymbol{\hat K}}_t^b} \right)}^{\text{T}}}} \right] \hfill \\
{\text{ + E}}\left[ {{\boldsymbol{\hat K}}_t^b} \right]{{\boldsymbol{Q}}_t}{\text{E}}\left[ {{{\left( {{\boldsymbol{\hat K}}_t^b} \right)}^{\text{T}}}} \right]. \hfill \\
\end{gathered}
\end{align}
A convergent gain of smoother ${\boldsymbol{K}}_t^b$ can be obtained utilizing the fixed-point algorithm, and the expectation of gain ${\text{E}}\left[ {{\boldsymbol{K}}_t^b} \right]$ can be obtained by using the adaptive solution
\begin{align}
{\text{E}}\left[ {{\boldsymbol{K}}_t^b} \right] = \left( {1 - \iota } \right){\text{E}}\left[ {{\boldsymbol{K}}_{t - 1}^b} \right] + \iota {\boldsymbol{K}}_t^b,
\end{align}
where $\iota $ is the forgotten factor.
Formula \eqref{EsansanQFBT} can be simplified as
\begin{align}\label{56NtOtOTt}
{{\boldsymbol{{\rm N}}}_t}{\text{ = }}{{\boldsymbol{\Omega }}_t}{{\boldsymbol{{\rm N}}}_{t - 1}}{\boldsymbol{\Omega }}_t^{\text{T}}{\text{ + }}{{\boldsymbol{Y}}_t}
\end{align}
with
\begin{align}
\left\{ \begin{gathered}
{{\boldsymbol{{\rm N}}}_t} = {\text{E}}\left[ {{{\boldsymbol{\xi }}_{s;t}}{\boldsymbol{\xi }}_{s;t}^{\text{T}}} \right], \hfill \\
{{\boldsymbol{\Omega }}_t} = {\text{E}}\left[ {{\boldsymbol{\hat K}}_t^b} \right], \hfill \\
{{\boldsymbol{Y}}_t} = \left( {{\boldsymbol{I}} - {\text{E}}\left[ {{\boldsymbol{\hat K}}_t^b} \right]{{\boldsymbol{F}}_t}} \right){\text{E}}\left[ {{{\boldsymbol{\xi }}_{f;t}}{\boldsymbol{\xi }}_{f;t}^{\text{T}}} \right]{\left( {{\boldsymbol{I}} - {\text{E}}\left[ {{\boldsymbol{\hat K}}_t^b} \right]{{\boldsymbol{F}}_t}} \right)^{\text{T}}} \hfill \\
{\text{ + E}}\left[ {{\boldsymbol{\hat K}}_t^b} \right]{{\boldsymbol{Q}}_t}{\text{E}}\left[ {{{\left( {{\boldsymbol{\hat K}}_t^b} \right)}^{\text{T}}}} \right]. \hfill \\
\end{gathered} \right.
\end{align}\
From the matrix vector operator:
\begin{align}\label{vectorpoerator}
\left\{ \begin{gathered}
{\text{vec}}\left( {{\boldsymbol{SUV}}} \right) = \left( {{{\boldsymbol{V}}^{\text{T}}} \otimes {\boldsymbol{S}}} \right){\text{vec}}\left( {\boldsymbol{U}} \right), \hfill \\
{\text{vec}}\left( {{\boldsymbol{S}} + {\boldsymbol{V}}} \right) = {\text{vec}}\left( {\boldsymbol{S}} \right) + {\text{vec}}\left( {\boldsymbol{V}} \right), \hfill \\
\end{gathered} \right.
\end{align}
where $ \otimes $ represents the Kronecker product and ${\text{vec}}\left( \cdot \right)$ denotes vectorization operation.
The closed-form solution of \eqref{56NtOtOTt} can be written as
\begin{align}
\mathop {\lim }\limits_{t \to \infty } {\text{vec}}\left( {{{\boldsymbol{{\rm N}}}_t}} \right) = \mathop {\lim }\limits_{t \to \infty } {\left( {{\boldsymbol{I}} - {{\boldsymbol{\Omega }}_t} \otimes {{\boldsymbol{\Omega }}_t}} \right)^{ - 1}}{\text{vec}}\left( {{{\boldsymbol{Y}}_t}} \right),
\end{align}
where ${\text{vec}}\left( {{{\boldsymbol{Y}}_t}} \right)$ can be calculated by using \eqref{vectorpoerator}, and one can be obtain
\begin{align}\label{QVECtSEgZZTFtS}
{\text{vec}}\left( {{{\boldsymbol{Y}}_t}} \right) = {{\boldsymbol{\xi }}_t}{\text{vec}}\left[ {{\text{E}}\left[ {{{\boldsymbol{\xi }}_{f;t}}{\boldsymbol{\xi }}_{f;t}^{\text{T}}} \right]} \right] + {{\boldsymbol{\varsigma }}_t}{\text{vec}}\left( {{{\boldsymbol{Q}}_t}} \right)
\end{align}
with
\begin{align}
\left\{ \begin{gathered}
{{\boldsymbol{\xi }}_t}{\text{ = }}\left( {{\boldsymbol{I}} - {\text{E}}\left[ {{\boldsymbol{\hat K}}_t^b} \right]{{\boldsymbol{F}}_t}} \right) \otimes \left( {{\boldsymbol{I}} - {\text{E}}\left[ {{\boldsymbol{\hat K}}_t^b} \right]{{\boldsymbol{F}}_t}} \right), \hfill \\
{{\boldsymbol{\varsigma }}_t} = {\text{E}}\left[ {{\boldsymbol{\hat K}}_t^b} \right] \otimes {\text{E}}\left[ {{\boldsymbol{\hat K}}_t^b} \right]. \hfill \\
\end{gathered} \right.
\end{align}
The error covariance matrix ${{\text{E}}\left[ {{{\boldsymbol{\xi }}_{f;t}}{\boldsymbol{\xi }}_{f;t}^{\text{T}}} \right]}$, in \eqref{QVECtSEgZZTFtS}, has been calculated in \cite{wang2021numerically}.
\subsection{Computational complexity}
In this section, the computational burden of the proposed MEE-RTS smoother is compared with that of the RTS smoother and MC-RTS smoother for the linear systems. It is noted that the MC-RTS smoother in \cite{wang2020maximum} is developed to estimate the state of the nonlinear systems, and the scheme for using MC-RTS smoother to handle state estimation of linear systems is not given. The pseudo-code for the MC-RTS smoother used to estimate the state of the linear systems (MC-RTSL) is summarised in Algorithm \ref{MCCRTS-linear}, which is used as the baseline in terms of computational burden and steady-state error. In Algorithm \ref{MCCRTS-linear}, ${{{\boldsymbol{K}}_{t;MC}}}$ and ${\boldsymbol{K}}_{t;MC}^b$ are gains of forward and backward pass.
From Algorithm \ref{MCCRTS-linear}, the computational burden of the forward pass and backward pass of the MC-RTSL smoother are
\begin{align}
\left\{ \begin{gathered}
{S_{MC\_F}} = 8{n^3} + 10{n^2}m + 6n{m^2} + 2{m^2} - {n^2} \hfill \\
+ 4mn - n + m + 10 + 3O\left( {{m^3}} \right) + O\left( {{n^3}} \right), \hfill \\
{S_{MC\_B}} = 12{n^3} + 7{n^2} + n + 10 + 4O\left( {{n^3}} \right), \hfill \\
\end{gathered} \right.
\end{align}
where ${S_{MC\_F}}$ and ${S_{MC\_B}}$ represent the computational complexity of the forward pass and backward pass. Hence, the computational complexity of MC-RTSL smoother is
\begin{align}\label{ccmc-rtslbn3}
\begin{gathered}
{S_{MC - RTSL}} = 20{n^3} + 10{n^2}m + 6n{m^2} + 2{m^2} + \hfill \\
6{n^2} + 4mn + m + 20 + 5O\left( {{n^3}} \right) + 3O\left( {{m^3}} \right). \hfill \\
\end{gathered}
\end{align}
The computational complexity of the forward pass of the MEE-RTS smoother is
\begin{align}\label{meeforwardcb}
\begin{gathered}
{S_{R - MEEKF}} = \left( {7{M_f} + 8} \right){n^3} + 7{M_f}{m^3} - {n^2} \hfill \\
+ \left( {19{M_f} + 6} \right){n^2}m + \left( {15{M_f} + 2} \right)n{m^2} \hfill \\
+ {M_f}O\left( {{n^3}} \right) + {M_f}m + \left( {5{M_f} - 1} \right)n \hfill \\
+ \left( {7{M_f} - 1} \right)nm + 3{M_f}O\left( {{m^3}} \right). \hfill \\
\end{gathered}
\end{align}
where ${{M_f}}$ is the number of the FPI method of the forward pass. The main formulas for the backward pass and their computational complexity are shown in Table \ref{ccotmeerts}, and the computational burden of the backward pass of the MEE-RTS smoother is
\begin{align}\label{meeforwardbp}
\begin{gathered}
{S_{MEE\_B}} = \left( {48{M_b} + 8} \right){n^3} + \left( {1 - 6{M_b}} \right){n^2} \hfill \\
+ \left( {1 - 9{M_b}} \right)n + 3{M_b}O({n^3}) + 24{M_b}, \hfill \\
\end{gathered}
\end{align}
where ${M_b}$ denotes the number of the FPI method of the backward pass. Combining \eqref{meeforwardcb} and \eqref{meeforwardbp}, the computational complexity of the MEE-RTS smoother is
\begin{align}\label{mee-rtsccb24m3}
\begin{gathered}
{S_{MEE - RTS}} = \left( {7{M_f} + 48{M_b} + 16} \right){n^3} + 7{M_f}{m^3} \hfill \\
+ \left( {19{M_f} + 6} \right){n^2}m - 6{M_b}{n^2} + \left( {15{M_f} + 2} \right)n{m^2} \hfill \\
+ \left( {{M_f} + 3{M_b}} \right)O\left( {{n^3}} \right) + {M_f}m + \left( {5{M_f} - 9{M_b}} \right)n \hfill \\
+ \left( {7{M_f} - 1} \right)nm + 3{M_f}O\left( {{m^3}} \right) + 24{M_b}. \hfill \\
\end{gathered}
\end{align}
From \eqref{ccmc-rtslbn3} and \eqref{mee-rtsccb24m3}, we can conclude that the computational complexity of the MEE-RTS smoother is slightly higher than that of MC-RTSL smoother. The main reasons for this phenomenon are the adoption of the MEE criterion and the FPI method of the R-MEEKF algorithm. In the MEE-RTS smoother, the MEE criterion is introduced into the forward and backward pass, and the MEE criterion with double summation is more computationally expensive than the MCC criterion with a single summation. The increase in computational burden due to the introduction of the MEE is inevitable. In addition, the computational burden of the FPI method is higher than or equal to the computational complexity of ARM.
From the above discussion, one can be deduced that reducing the number of the FPI method (the number can be controlled by the parameter $\tau$ in \cite{chen2019minimum}) or using ARM can significantly lower the computational complexity of the MEE-RTS algorithm. However, these two solutions also decrease the performance of the MEE-RTS smoother in steady-state error. Hence, the choice of parameter $\tau$ should therefore trade-off the steady-state error and the computational complexity of the MEE-RTS smoother, the detailed selection scheme of the parameter $\tau$ is present in Section \ref{para_perfro_mee}.
\begin{algorithm}[t]
\caption{MC-RTS smoother for linear systems} \label{MCCRTS-linear}
\LinesNumbered
\KwIn{state transition matrix ${\boldsymbol{F}}$, measurement matrix ${\boldsymbol{H}}$, covariance matrices ${\boldsymbol{Q}}$, ${\boldsymbol{R}}$, and kernel bandwidth $\sigma$}
\KwOut{${{\boldsymbol{\tilde x}}_{t|T}}$}
\textbf{Initialization}: ${{\boldsymbol{\tilde x}}_{0|0}} = {{\boldsymbol{x}}_0}$, ${{\boldsymbol{P}}_{0|0}} = {{\boldsymbol{P}}_0}$\;
\textbf{Forward Pass:}\
\For{$t\leftarrow $1$ $ \KwTo $T$}{
Calculate ${{\boldsymbol{\tilde x}}_{t|t - 1}}$ and ${{\boldsymbol{P}}_{t|t-1}}$ using \eqref{prespttj1dqt}\;
Calculate ${{\boldsymbol{K}}_{t;MC}}$ and ${\phi _{\boldsymbol{R}}}$ using
\begin{align}
\left\{ \begin{gathered}
{{\boldsymbol{K}}_{t;MC}} = {\left( {{{\boldsymbol{P}}_{t|t - 1}} + {{\boldsymbol{H}}^{\text{T}}}{\phi _{\boldsymbol{R}}}{{\boldsymbol{R}}^{ - 1}}{\boldsymbol{H}}} \right)^{ - 1}}{{\boldsymbol{H}}^{\text{T}}}{\phi _{\boldsymbol{R}}}{{\boldsymbol{R}}^{ - 1}}, \hfill \\
{\phi _{\boldsymbol{R}}}{\text{ = G}}\left( {{{\left\| {{{\boldsymbol{y}}_t} - {\boldsymbol{H}}{{\boldsymbol{x}}_{t|t - 1}}} \right\|}_{\boldsymbol{R}}}} \right); \hfill \\
\end{gathered} \right.
\end{align}\
Calculate ${{\boldsymbol{\tilde x}}_{t|t}}$ and ${{\boldsymbol{P}}_{t|t}}$ using\
\begin{align}
\left\{ \begin{gathered}
{{{\boldsymbol{\tilde x}}}_{t|t}} = {{{\boldsymbol{\tilde x}}}_{t|t - 1}}{\text{ + }}{{\boldsymbol{K}}_{t;MC}}\left( {{{\boldsymbol{y}}_t} - {\boldsymbol{H}}{{{\boldsymbol{\tilde x}}}_{t|t - 1}}} \right), \hfill \\
{{\boldsymbol{P}}_{t|t}} = {{\boldsymbol{K}}_{t;MC}}{{\boldsymbol{R}}_t}{\boldsymbol{K}}_{t;MC}^{\text{T}} + \hfill \\
\left( {{{\boldsymbol{I}}_n} - {{\boldsymbol{K}}_{t;MC}}{\boldsymbol{H}}} \right){{\boldsymbol{P}}_{t|t - 1}}{\left( {{{\boldsymbol{I}}_n} - {{\boldsymbol{K}}_{t;MC}}{\boldsymbol{H}}} \right)^{\text{T}}}; \hfill \\
\end{gathered} \right.
\end{align}
}
\textbf{Backward Pass:}\
\For{$t\leftarrow $T - 1$ $ \KwTo $1$}{
Calculate ${\phi _{\boldsymbol{Q}}}$ and ${\boldsymbol{K}}_{t;MC}^b$ using
\begin{align}
\left\{ \begin{gathered}
{\phi _{\boldsymbol{Q}}}{\text{ = G}}\left( {{{\left\| {{{{\boldsymbol{\tilde x}}}_{t + 1|T}} - {\boldsymbol{F}}{{{\boldsymbol{\tilde x}}}_{t|t}}} \right\|}_{\boldsymbol{Q}}}} \right), \hfill \\
{\boldsymbol{K}}_{t;MC}^b = {\left( {{\boldsymbol{P}}_{t|t}^{ - 1} + {{\boldsymbol{F}}^{\text{T}}}{\phi _{\boldsymbol{Q}}}{{\boldsymbol{Q}}^{ - 1}}{\boldsymbol{F}}} \right)^{ - 1}}{{\boldsymbol{F}}^{\text{T}}}{\phi _{\boldsymbol{Q}}}{{\boldsymbol{Q}}^{ - 1}}; \hfill \\
\end{gathered} \right.
\end{align}
Calculate ${{\boldsymbol{\tilde x}}_{t|T}}$ and ${{\boldsymbol{P}}_{t|T}}$ using\
\begin{align}
\left\{ \begin{gathered}
{{{\boldsymbol{\tilde x}}}_{t|T}} = {{{\boldsymbol{\tilde x}}}_{t|t}} + {\boldsymbol{K}}_{t;MC}^b\left( {{{{\boldsymbol{\tilde x}}}_{t + 1|T}} - {\boldsymbol{F}}{{{\boldsymbol{\tilde x}}}_{t|t}}} \right), \hfill \\
{{\boldsymbol{P}}_{t|T}} = {\boldsymbol{K}}_{t;MC}^b\left( {{{\boldsymbol{P}}_{t + 1|T}} - {{\boldsymbol{P}}_{t + 1|t}}} \right){\left( {{\boldsymbol{K}}_{t;MC}^b} \right)^{\text{T}}} \hfill \\
+ {{\boldsymbol{P}}_{t|t}}; \hfill \\
\end{gathered} \right.
\end{align}
}
\end{algorithm}
\begin{table}
\centering
\caption{Computational complexities of the MEE-RTS algorithm.}\label{ccotmeerts}
\begin{tabular}{llll}
\hline
& & & matrix inversion,\\
& ${ \times }$ & ${ + {\text{ }}and{\text{ }} - }$ & Cholesky decomposition, \\
& & & ,and exponentiation \\ \hline
\eqref{tr71qkj1axkjej} & ${n^2}$ & ${n^2} - n$ & $0$ \\
\eqref{vkjxkCdYk} & $2{n^3}$ & $2{n^3} - {n^2}$ & ${0}$ \\
\eqref{38xtTxmao} & ${n^2}$ & ${n^2} + n$ & ${0}$ \\
\eqref{gainmeesmoother} & $8{n^3}$ & $8{n^3} - 4{n^2}$ & $3O({n^3})$ \\
\eqref{ffAATT} & $16{n^3}$ & $16{n^3} - 4{n^2}$ & ${0}$ \\
\eqref{FFigmatrix} & $2n + 7$ & $2n + 1$ & $4$ \\
\eqref{Getigjetij} & $2n + 7$ & $2n + 1$ & $4$ \\
\eqref{tKTtPtJ1IjN} & $2{n^3}$ & $2{n^3}$ & ${0}$ \\ \hline
\end{tabular}
\end{table}
\section{Simulation}\label{simulation}
In this section, several numerical simulations are performed to demonstrate the performance of the proposed algorithms. In each run, 1000 samples are used to calculate the MSD of the MEE-RTS and MEE-ERTS smoothers. Simulation results are averaged over 300 independent Monte Carlo runs. The MSD \cite{HE2023108787} is defined as
\begin{equation}
\begin{split}
MSD = 10{\log _{10}}\left( {{{\left\| {{{\boldsymbol{x}}_t} - {{{\boldsymbol{\tilde x}}}_{t|T}}} \right\|}^2}} \right).
\end{split}
\end{equation}
\subsection{Noise model}
A few non-Gaussian noise models are presented below before performing the simulation, such as mixed-Gaussian noise, $\alpha$-stable noise and Rayleigh noise.
\begin{itemize}
\item[1] Modeling mixed-Gaussian noise \cite{HE20221362} as follows:
\begin{align}\label{mixed-Gaussian}
r \sim \lambda \mathcal{N}\left( {{a_1},{\mu _1}} \right){\text{ + }}\left( {1 - \lambda } \right)\mathcal{N}\left( {{a_2},{\mu _2}} \right),0 \leqslant \lambda \leqslant 1,
\end{align}
where parameter $\lambda $ presents the mixing factor. The noise with a mixed-Gaussian distribution can be represented as $r \sim M\left( {\lambda ,{a_1},{a_2},{\mu _1},{\mu _2}} \right)$.
\item[2] The characteristic function of the ${\alpha}$-stable noise \cite{peng2017constrained} is defined as:
\begin{align}
\psi \left( t \right) = \exp \left\{ {j\vartheta t - \gamma {{\left| t \right|}^{{a_3}}}\left[ {1 + jb\operatorname{sgn} \left( t \right)S\left( {t,{a_3}} \right)} \right]} \right\},
\end{align}
with
\begin{align}
S\left( {t,{a_3}} \right) = \left\{ {\begin{array}{*{20}{l}}
{\tan \left( {\frac{{{a_3}\pi }}{2}} \right),{\text{ }}{a_3} \ne {\text{1,}}} \\
{\frac{2}{\pi }\log \left| t \right|,{\text{ }}{a_3}{\text{ = 1}}{\text{,}}}
\end{array}} \right.
\end{align}
where ${{a_3}}$ is the characteristic factor, $b$ represents the symmetry parameter, ${\gamma}$ and $\vartheta $ are the dispersion parameter and location parameter. The noise with ${\alpha}$-stable distribution is expressed as $r \sim S\left( {{a_3},b,\gamma ,\vartheta } \right)$.
\item[3] The Rayleigh distribution can be represented by the probability density function shown below:
\begin{align}
r\left( t \right) = \frac{t}{{{\sigma ^2}}}\exp \left( { - \frac{{{t^2}}}{{2{\sigma ^2}}}} \right).
\end{align}
The noise with the Rayleigh distribution is written as ${v \sim R\left( \sigma \right)}$.
\end{itemize}
\subsection{Example 1}
In this sub-section, a constant acceleration motion model \cite{CHEN2017} is considered. The state-space model are given by
\begin{align}
{{\boldsymbol{x}}_t} = \left[ {\begin{array}{*{20}{c}}
1&{\Delta T}&{{{\Delta {T^2}} \mathord{\left/
{\vphantom {{\Delta {T^2}} 2}} \right.
\kern-\nulldelimiterspace} 2}} \\
0&1&{\Delta T} \\
0&0&1
\end{array}} \right]{{\boldsymbol{x}}_{t - 1}} + {{\boldsymbol{q}}_{t - 1}},
\end{align}
and
\begin{align}
{{\boldsymbol{y}}_t} = \left[ {\begin{array}{*{20}{c}}
{\begin{array}{*{20}{c}}
1&0&0
\end{array}} \\
{\begin{array}{*{20}{c}}
0&1&0
\end{array}}
\end{array}} \right]{{\boldsymbol{x}}_t} + {{\boldsymbol{r}}_t}.
\end{align}
Here, $\Delta T = 0.1sec$ represents sampling interval of the model, ${{\boldsymbol{x}}_t} = {\left[ {\begin{array}{*{20}{c}}
{{x_{1;t}}}&{{x_{2;t}}}&{{x_{3;t}}}
\end{array}} \right]^{\text{T}}}$ with ${{x_{1;t}}}$, ${{x_{2;t}}}$, and ${{x_{3;t}}}$ denote position, velocity, and acceleration respectively.
Five scenarios with different noise models are considered. The specific forms of process and measurement noises are shown below:
\begin{align}
\left\{ {\begin{array}{*{20}{l}}
{{q_{t;i}} \sim M\left( {0.9,0,0,0.01,25} \right),} \\
{{r_{t;i}} \sim \mathcal{N}\left( {0,0.01} \right),}
\end{array}} \right.
\end{align}
where ${{q_{t;i}}}$ and ${{r_{t;i}}}$ represent the $i$th element of ${{{\boldsymbol{q}}_t}}$ and ${{{\boldsymbol{r}}_t}}$. The second scenario is that process and measurement noises are both mixed-Gaussian noise, which takes the following form
\begin{align}
\left\{ \begin{gathered}
{q_{t;i}} \sim M\left( {0.9,0,0,0.01,25} \right), \hfill \\
{r_{t;i}} \sim M\left( {0.7,0,0,0.01,900} \right). \hfill \\
\end{gathered} \right.
\end{align}
The third scenario is that both the process noise is mixed-Gaussian noise, and the measurement noise is a mixture of ${\alpha}$-stable noise and Gaussian noise with the following form
\begin{align}
\left\{ \begin{gathered}
{q_{t;i}} \sim M\left( {0.9,0,0,0.01,25} \right), \hfill \\
{r_{t;i}} \sim 0.9S\left( {1.25,1,0,0.5} \right){\text{ + 0}}{\text{.1}}\mathcal{N}\left( {0,900} \right). \hfill \\
\end{gathered} \right.
\end{align}
where $0.9$ denotes the probability that the noise ${{r_{t;i}}}$ obeys the distribution of $S\left( {1.25,1,0,0.5} \right)$, and the representation below has a similar meaning.
The fourth simulation scenario has the following form:
\begin{align}
\left\{ {\begin{array}{*{20}{l}}
{{q_{t;i}} \sim M\left( {0.9,0,0,0.01,25} \right),} \\
{{r_{t;i}} \sim 0.7R\left( 2 \right) + 0.3\mathcal{N}\left( {0,900} \right).}
\end{array}} \right.
\end{align}
The fifth scenario is that the process noise obeys a mixed-Gaussian distribution and the measurement noise is a multimodal distribution of the following form
\begin{align}
\left\{ {\begin{array}{*{20}{l}}
{{q_{t;i}} \sim M\left( {0.95,0,0,0.01,25} \right),} \\
{{r_{t;i}} \sim M\left( {0.6,2, - 2,0.01,100} \right).}
\end{array}} \right.
\end{align}
The initial state ${{\boldsymbol{x}}_0}$, initial estimated value ${{\boldsymbol{\tilde x}}_{0|0}}$, and initial covariance ${{\boldsymbol{P}}_{0|0}}$ of the above mentioned five scenarios are given by:
\begin{align}\label{x0bolapxx00}
\left\{ \begin{gathered}
{{\boldsymbol{x}}_0} \sim \mathcal{N}\left( {0,{{\boldsymbol{I}}_n}} \right), \hfill \\
{{{\boldsymbol{\tilde x}}}_{0|0}} \sim \mathcal{N}\left( {{{\boldsymbol{x}}_0},{\boldsymbol{Q}}} \right), \hfill \\
{{\boldsymbol{P}}_{0|0}} = {{\boldsymbol{I}}_n}. \hfill \\
\end{gathered} \right.
\end{align}
In these five scenarios, we compare the performance of the MEE-RTS smoother with that of the KF, RTS smoother \cite{rauch1965maximum}, MCKF \cite{CHEN2017}, MC-RTS smoother \cite{wang2020maximum}, R-MEEKF \cite{wang2021numerically} algorithms. These results are shown in Fig. \ref{two_mixture}, Fig. \ref{mixalphamsd}, and Table \ref{MSD_different_noise}.
First, the performance of the MEE-RTS smoother in a heavy-tailed noise environment is studied, and the simulation results are shown in Fig. \ref{two_mixture} and Fig. \ref{mixalphamsd}. The convergence curves of MSD of $x_1$, $x_2$, and $x_3$ are presented in Figs. \ref{Gaussianmixnoisepos}, \ref{Gaussianmixnoisevel}, and \ref{Gaussianmixnoisea}. Fig. \ref{Gaussianmixnoiseaxishuu} shows the variation curve of the MSD of ${\boldsymbol{x}}$ with the mixture coefficient $\lambda$ of measurement noise ${{r_t}}$. Some parameters of the considered algorithms are also presented in Fig. \ref{two_mixture}. It is obvious that the proposed MEE-RTS smoother performs best in the second scenario. From Fig. \ref{two_mixture}, one can witness that the performance of these algorithms degrades as the number of outliers increases. For the third scenario, the corresponding results and the parameter settings of these considered algorithms are shown in Fig. \ref{mixalphamsd}, and the effect of the mixture coefficient $\lambda$ of the observed noise $r_i$ with respect to the performance of the MEE-RTS smoother was studied. Similar results to the second scenario can be obtained from Fig. \ref{mixalphamsd}.
Then the robustness of the MEE-RTS smoother in different scenarios is investigated, and the simulation results are presented in Table \ref{MSD_different_noise}. The elements in Table \ref{MSD_different_noise} represent the MSD of $x_1$, $x_2$, and $x_3$ for the different algorithms in the considered scenarios, for example (-20.1, -20.1, -20.1) represents the MSD of $x_1$, $x_2$, and $x_3$ of the KF algorithm in the first scenario. In particular, $2.0$ in $0.9 (2.0)$ denotes the kernel width of the R-MEE and RTS algorithms in the fifth scenario. From Table \ref{MSD_different_noise}, we can get that 1): the MEE-RTS smoother performs best with the non-Gaussian noise; 2): in a Gaussian noise environment, the MEE-RTS algorithm performs nearly as well as the RTS algorithm. These two points show that the MEE-RTS smoother is robust in different scenarios.
\begin{figure*}[!t]
\centering
\subfloat[]{\includegraphics[width=3.5in]{Gaussian_Mixture_Pos.eps}%
\label{Gaussianmixnoisepos}}
\hfil
\subfloat[]{\includegraphics[width=3.5in]{Gaussian_Mixture_Vel.eps}%
\label{Gaussianmixnoisevel}}
\hfil
\subfloat[]{\includegraphics[width=3.5in]{Gaussian_Mixture_A.eps}%
\label{Gaussianmixnoisea}}
\hfil
\subfloat[]{\includegraphics[width=3.5in]{Gaussian_Mixture_p.eps}%
\label{Gaussianmixnoiseaxishuu}}
\caption{MSD of different algorithms in second scenario. (a) MSD of $x_1$. (b) MSD of $x_2$. (c) MSD of $x_3$. (d) The MSD with different $\lambda $.}
\label{two_mixture}
\end{figure*}
\begin{figure*}[!t]
\centering
\subfloat[]{\includegraphics[width=3.5in]{Com_Alpha_Stable_Pos.eps}%
\label{figuremixalpha}}
\hfil
\subfloat[]{\includegraphics[width=3.5in]{Com_Alpha_Stable_Vels.eps}%
\label{alphanoisevel}}
\hfil
\subfloat[]{\includegraphics[width=3.5in]{Com_Alpha_Stable_A.eps}%
\label{alphanoisea}}
\hfil
\subfloat[]{\includegraphics[width=3.5in]{Gaussian_Mixture_p_alpha.eps}%
\label{alphanoisea1}}
\caption{The MSD of these considered algorithms in third scenario. (a) MSD of $x_1$. (b) MSD of $x_2$. (c) MSD of $x_3$. (d) The MSD with different $\lambda $.}
\label{mixalphamsd}
\end{figure*}
\begin{table*}[htbp]
\centering
\caption{The steady-state MSD (dB) of different algorithms in different scenarios}\label{MSD_different_noise}
\resizebox{\textwidth}{!}
{
\begin{tabular}{lllllll}
\hline
Algorithm & $\sigma$ & First scenario & Second scenario & Third scenario & Fourth scenario & Fifth scenario \\ \hline
KF & N/A & (-20.1,-20.1,-20.1) & (11.5,14.8,12.0) & (15.8,13.3,9.0) & (13.3,14.8,14.4) & (9.0,8.6,8.2)\\
RTS & N/A & (\textbf{-20.1,-20.2,-20.2}) & (14.4,11.4,10.9) & (15.3,12.9,8.8) & (14.3,13.8,12.9) & (8.3,7.9,7.6)\\
MCKF & 2.0 & (-19.2,-19.2,-19.2) & (14.0,11.2,10.9) & (12.5,11.7,8.6) & (14.0,13.7,12.9) & (8.8,8.6,7.9)\\
MC-RTS & 2.0 & (-19.3,-19.3,-19.3) & (13.5,10.3,10.1) & (12.0,11.2,8.5) & (13.4,13.1,12.4) & (8.1,7.8,7.3)\\
R-MEEKF & 0.9(2.0) & (-19.0,-19.0,-19.0) & (10.5,10.2,10.5) & (9.0,9.4,8.8) & (12.2,12.6,12.2) & (8.5,8.5,7.7)\\
MEE-RTS & 0.9(2.0) & (-19.2,-19.1,-19.0) & (\textbf{9.5,9.3,9.7}) & (\textbf{8.4,8.5,7.9}) & (\textbf{10.8,11.0,11.5}) & (\textbf{7.9,7.7,7.1})\\ \hline
\end{tabular}
}
\end{table*}
\begin{table*}[htbp]
\centering
\caption{The steady-state MSD (dB) of different algorithms with different kernel bandwidth}\label{MSD_kernel}
\resizebox{\textwidth}{!}
{
\begin{tabular}{lllllll}
\hline
Algorithm & $\sigma$ & First scenario & Second scenario & Third scenario & Fourth scenario & Fifth scenario\\ \hline
RTS & N/A & (\textbf{-20.1,-20.2,-20.2}) & (14.4,11.4,10.9) & (15.3,12.9,8.8) & (14.3,13.1,12.4) & (8.3,7.9,7.6)\\
MC-RTS & 2 & (-19.3,-19.3,-19.3) & (13.5,10.3,10.1) & (12.0,11.2,8.5) & (13.4,13.1,12.4) & (8.1,7.8,7.3)\\
MEE-RTS & 0.1 & (-19.2,-19.1,-19.0) & (17.1,14.6,13.3) & (18.3,17.3,15.4) & (18.3,17.2,15.5) & (13.7,13.2,12.1) \\
MEE-RTS & 0.5 & (-19.2,-19.1,-19.0) & (13.1,11.4,11.3) & (11.7,11.9,12.3) & (11.9,11.7,12.1) & (13.1,11.9,11.2)\\
MEE-RTS & 0.9 & (-19.2,-19.1,-19.0 & (\textbf{9.5,9.3,9.7}) & (\textbf{8.4,8.5,7.9}) & (\textbf{10.8,11.0,11.5}) & (11.5,10.4,10.2)\\
MEE-RTS & 2.0 & (-19.3,-19.3,-19.1) & (13.3,9.5,9.2) & (11.9,12.0,11.8) & (12.0,11.9,11.7) & (\textbf{7.9,7.7,7.1})\\
MEE-RTS & 5.0 & (-19.4,-19.4,-19.3) & (14.6,9.4,10.0) & (13.9,13.6,12.8) & (14.0,13.5,12.8) & (8.9,8.7,8.5)\\
MEE-RTS & 50.0 & (-19.5,-19.6,-19.7) & (17.9,11.9,11.3) & (14.6,14.3,13.3) & (14.7,14.3,13.4) & (9.1,9.1,8.8)\\
MEE-RTS & 100.0 & (-19.9,-19.9,-19.8) & (18.3,13.5,12.1) & (15.2,15.1,15.1) & (16.5,15.1,14.7) & (9.2,9.5,9.8)\\
\hline
\end{tabular}
}
\end{table*}
\subsection{Example 2}
In this part, a vehicle tracking model \cite{9923771} is considered to verify the performance of the MEE-ETRS smoother, and the dataset \cite{citweb} of a vehicle is obtained from the the measurements of lidar and radar sensors. The state equation can be expressed as
\begin{align}
{{\boldsymbol{x}}_t} = \left[ {\begin{array}{*{20}{c}}
1&0&{\Delta T}&0 \\
0&1&0&{\Delta T} \\
0&0&1&0 \\
0&0&0&1
\end{array}} \right]{{\boldsymbol{x}}_{t - 1}} + {{\boldsymbol{q}}_{t - 1}},
\end{align}
and the measurement equation of a radar is
\begin{align}
{{\boldsymbol{y}}_{t;radar}} = \left[ {\begin{array}{*{20}{c}}
{\sqrt {x_{1;t}^2 + x_{2;t}^2} } \\
{\arctan \left( {\frac{{{x_{2;t}}}}{{{x_{1;t}}}}} \right)} \\
{\frac{{{x_{1;t}}{x_{3;t}} + {x_{2;t}}{x_{4;t}}}}{{\sqrt {x_{1;t}^2 + x_{2;t}^2} }}}
\end{array}} \right] + {{\boldsymbol{r}}_{t;radar}},
\end{align}
and of a lidar is given by
\begin{align}
{{\boldsymbol{y}}_{t;lidar}} = \left[ {\begin{array}{*{20}{c}}
{\begin{array}{*{20}{c}}
1&0&0&0
\end{array}} \\
{\begin{array}{*{20}{c}}
0&1&0&0
\end{array}}
\end{array}} \right]{{\boldsymbol{x}}_t} + {{\boldsymbol{r}}_{t;lidar}},
\end{align}
for lidar.
${{\boldsymbol{r}}_{t;radar}}$ and ${{\boldsymbol{r}}_{t;lidar}}$ are measurement noises of radar and lidar, and they obey distributions ${{\boldsymbol{r}}_{t;lidar}} \sim M\left( {\begin{array}{*{20}{c}}
{0.9}&0&{0.01{{\boldsymbol{I}}_3}}&{9{{\boldsymbol{r}}_{radar}}}
\end{array}} \right)$ and ${{\boldsymbol{r}}_{t;lidar}} \sim M\left( {\begin{array}{*{20}{c}}
{0.9}&0&{0.01{{\boldsymbol{I}}_2}}&{9{{\boldsymbol{r}}_{lidar}}}
\end{array}} \right)$. Here, the matrices ${{\boldsymbol{r}}_{radar}}$ and ${{\boldsymbol{r}}_{lidar}}$ are $diag\left[ {\begin{array}{*{20}{c}}
1&{0.01}&{0.01}
\end{array}} \right]$ and $diag\left[ {\begin{array}{*{20}{c}}
1&{0.01}
\end{array}} \right]$.
The state of vehicle can be represented as ${{\boldsymbol{x}}_t} = {\left[ {{x_{1;t}},{x_{2;t}},{x_{3;t}},{x_{4;t}}} \right]^{\text{T}}}$. ${{x_{1;t}}}$ and ${{x_{2;t}}}$ represent the position of $x$, $y$ axis, ${{x_{3;t}}}$ and ${{x_{4;t}}}$ denote the velocity of $x$, $y$ axis. The sampling interval is $\Delta T = 0.1\sec $, the covariance matrix of process noise is
\begin{align*}
{{\boldsymbol{Q}}} = \left[ {\begin{array}{*{20}{c}}
{\tfrac{{\Delta {T^2}}}{4}}&0&{\tfrac{{\Delta {T^3}}}{2}}&0 \\
0&{\tfrac{{\Delta {T^2}}}{4}}&0&{\tfrac{{\Delta {T^3}}}{2}} \\
{\tfrac{{\Delta {T^3}}}{2}}&0&{\Delta {T^2}}&0 \\
0&{\tfrac{{\Delta {T^3}}}{2}}&0&{\Delta {T^2}}
\end{array}} \right].
\end{align*}
The initial values of ${{\boldsymbol{x}}_0}$, ${{\boldsymbol{\tilde x}}_{0|0}}$, and ${{\boldsymbol{P}}_{0|0}}$ are the same as \eqref{x0bolapxx00}. The covariance matrix of measurement noise is set to ${{\boldsymbol{R}}_t} = \left[ {0.09,0.05,0.09} \right]$.
The variation curves of the MSD of different algorithms are presented in Fig. \ref{tracking}. According to Fig. \ref{tracking}, one can obtain that the MEE-ERTS smoother outperforms several algorithms in terms of steady-state error.
\begin{figure}
\centerline{\includegraphics[width=0.5\textwidth]{tracking_data1.eps}}
\caption{MSD of different algorithms.}\label{tracking}
\end{figure}
\subsection{Parameter selection} \label{para_perfro_mee}
In this part, the influence of the critical parameters $\sigma$ and $\tau$ on the performance of the MEE-RTS smoother is investigated, and guidance is given on the choice of parameters.
The influence of the kernel bandwidth $\sigma$ on the performance of the MEE-RTS smoother is investigated in the above five scenarios, and the simulation results are displayed in Table \ref{MSD_kernel} and Fig. \ref{MEERTSsigma}. As seen in Table \ref{MSD_kernel}, one can be inferred that the performance of the MEE-RTS smoother first increases and then decreases in the increase of the parameter $\sigma$ with the presence of non-Gaussian noise. The MEE-RTS smoother can yield the required performance when $\sigma$ is close to 0.9 or 2.0. In a Gaussian noise environment, the performance of the MEE-RTS smoother improves with the increasing $sigma$ but does not exceed that of the RTS smoother and KF algorithms. The relationship between algorithm performance and kernel bandwidth guides the selection of kernel bandwidth. The selection scheme about $\sigma$ is as follows: 1) a larger $\sigma$ should be set in a Gaussian noise environment; 2) the optimal kernel width is different in different heavy-tailed noise scenarios, and suitable $\sigma$ is 0.9 or 2.0 in the considered scenarios.
The influence of threshold $\tau$ with respect to the performance of the MEE-RTS algorithm is also investigated in the presence of non-Gaussian noise. The simulation results are presented in Fig. \ref{meertstau} and Table \ref{MSD_tau} with different $\tau$. In the second scenario, Fig. \ref{meertstau} presents the MSD of MEE-RTS smoother with different $\tau$ and $\sigma = 0.9$. The MSD, computational time, and the number $M$ of the FPI method of RTS, MC-RTSL, and MEE-RTS smoothers are presented in Table \ref{MSD_tau}. These smoothers are measured using MATLAB 2020a, which works on an i5-8400 and a 2.80GHz CPU. From Fig. \ref{MEERTSTau} and Table \ref{MSD_tau}, the performance of the MEE-RTS smoother in steady-state MSD decreases to some extent as the threshold decreases in non-Gaussian noise environments, however, the cost (the computational complexity and computational time) of such a small increase in performance is expensive. Hence, the choice of threshold should trade off the steady-state MSD and cost, with smaller thresholds being chosen to improve the performance of the algorithm where real-time requirements are low, and larger thresholds or ARM being used in the opposite direction.
\begin{figure*}[!t]
\centering
\subfloat[]{\includegraphics[width=3.5in]{MEE_RTS_sigma.eps}%
\label{MEERTSsigma}}
\hfil
\subfloat[]{\includegraphics[width=3.5in]{MEE_RTS_Tau_2.eps}%
\label{meertstau}}
\caption{MSD of MEE-RTS smoother with different $\sigma$. (a) MSD of MEE-RTS smoother with different $\sigma$. (b) MSD of the MEE-RTS smoother with different $\tau $.}
\label{MEERTSTau}
\end{figure*}
\begin{table*}
\centering
\caption{The steady-state MSD, computational time, and the number of the FPI method of different algorithms with different $\tau$}\label{MSD_tau}
\resizebox{\textwidth}{!}
{
\begin{tabular}{lllllllllll}
\hline
& $\tau$ & \multicolumn{3}{c}{Second scenario} & \multicolumn{3}{c}{Third scenario} & \multicolumn{3}{c}{Fourth scenario} \\ \hline
Algorithm & & MSD & time (sec) & M & MSD & time (sec) & M & MSD & time (sec) & M \\
RTS & N/A & 13.29 & ${\text{1}}{\text{.685}} \times {10^{ - 5}}$ & N/A & 12.94 & ${\text{1}}{\text{.643}} \times {10^{ - 5}}$ & N/A & 13.52 & ${\text{1}}{\text{.691}} \times {10^{ - 5}}$ & N/A \\
MC-RTSL & N/A & 11.07 & ${\text{7}}{\text{.499}} \times {10^{ - 5}}$ & N/A & 10.89 & ${\text{1}}{\text{.679}} \times {10^{ - 5}}$ & N/A & 12.81 & ${\text{1}}{\text{.656}} \times {10^{ - 5}}$ & N/A \\
MEE-RTS & ${10^{ - 1}}$ & 9.52 & ${\text{2}}{\text{.322}} \times {10^{ - 4}}$ & 1.0249 & 8.99 & ${\text{2}}{\text{.302}} \times {10^{ - 4}}$ & 1.0225 & 10.93 & ${\text{2}}{\text{.337}} \times {10^{ - 4}}$ & 1.0101 \\
MEE-RTS & ${10^{ - 2}}$ & 9.51 & ${\text{2}}{\text{.562}} \times {10^{ - 4}}$ & 1.2128 & 8.94 & ${\text{2}}{\text{.563}} \times {10^{ - 4}}$ & 1.1962 & 10.91 & ${\text{2}}{\text{.530}} \times {10^{ - 4}}$ & 1.1893 \\
MEE-RTS & ${10^{ - 3}}$ & 9.49 & ${\text{3}}{\text{.155}} \times {10^{ - 4}}$ & 1.6599 & 8.93 & ${\text{3}}{\text{.130}} \times {10^{ - 4}}$ & 1.6423 & 10.89 & ${\text{3}}{\text{.104}} \times {10^{ - 4}}$ & 1.6098 \\
MEE-RTS & ${10^{ - 4}}$ & 9.48 & ${\text{3}}{\text{.419}} \times {10^{ - 4}}$ & 1.9380 & 8.91 & ${\text{3}}{\text{.349}} \times {10^{ - 4}}$ & 1.9413 & 10.88 & ${\text{3}}{\text{.324}} \times {10^{ - 4}}$ & 1.8310 \\
MEE-RTS & ${10^{ - 5}}$ & 9.47 & ${\text{3}}{\text{.849}} \times {10^{ - 4}}$ & 2.3892 & 8.90 & ${\text{3}}{\text{.530}} \times {10^{ - 4}}$ & 2.4786 & 10.87 & ${\text{3}}{\text{.687}} \times {10^{ - 4}}$ & 2.2041 \\
MEE-RTS & ${10^{ - 6}}$ & 9.46 & ${\text{4}}{\text{.453}} \times {10^{ - 4}}$ & 2.9323 & 8.89 & ${\text{4}}{\text{.543}} \times {10^{ - 4}}$ & 3.0513 & 10.87 & ${\text{4}}{\text{.155}} \times {10^{ - 4}}$ & 2.6500 \\
MEE-RTS & ${10^{ - 7}}$ & 9.45 & ${\text{4}}{\text{.931}} \times {10^{ - 4}}$ & 3.4424 & 8.88 & ${\text{5}}{\text{.177}} \times {10^{ - 4}}$ & 3.6535 & 10.86 & ${\text{7}}{\text{.701}} \times {10^{ - 4}}$ & 3.1446 \\
MEE-RTS & ${10^{ - 8}}$ & 9.45 & ${\text{5}}{\text{.468}} \times {10^{ - 4}}$ & 3.9648 & 8.88 & ${\text{5}}{\text{.740}} \times {10^{ - 4}}$ & 4.2352 & 10.86 & ${\text{5}}{\text{.076}} \times {10^{ - 4}}$ & 3.4870 \\ \hline
\end{tabular}
}
\end{table*}
\section{Conclusion}\label{conclusion}
In this study, two new RTS-type smoothers based on the MEE criterion were developed for estimating the state of the linear and nonlinear systems with non-Gaussian noise. Some theoretical analysis of the MEE-RTS smoother were given, and the results of these analyses have indicated that the MEE-RTS smoother is stable. Numerical simulation results have indicated that the proposed smoothers perform better than several existing smoothers with non-Gaussian noise.
The choice of threshold $\tau$ should trade off the steady-state MSD and computational complexity of the MEE-RTS smoother, with smaller threshold being chosen to improve the performance of the algorithm where real-time requirements are low, and larger thresholds or ARM being used in the opposite direction. Furthermore, the kernel width $\sigma$ has a significant influence on the steady-state MSD. Kernel width that are either too large or too small can degrade the performance of the MEE-RTS smoother and should be adjusted empirically.
The MEE-ERTS smoother is also developed for nonlinear systems. However, retaining only the first-order Taylor expansion term inevitably introduces errors due to linearization. To extend the application, there is a need to develop an RTS-type smoother based on the MEE criterion with higher accuracy for estimating the state of nonlinear systems, which is our future work.
\section{Acknowledgments}\label{acknowledgments}
This work was funded by the NNSFC with Nos. 51975107 and 62103083, Sichuan Science and Technology Major Project No.2019ZDZX0020, and Sichuan Science and Technology Program, No. 2022YFG0343. The first two authors contributed equally to this study.
|
\section{Introduction}
State estimation in wireless sensor networks (WSNs) has been a subject of considerable interest recently\cite{YANG2019156}, and it has been applied widely, such as in static or dynamic target positioning and tracking\cite{8769946}, indoor positioning\cite{9031316}, vehicle navigation\cite{8613889}, and smart grids\cite{8187698}. However, a number of challenges affecting the state estimation performance in WSNs remain to be fully addressed, such as those due to data packet drops and non-Gaussian noise, which are both very common in practical applications of automatic control and target tracking.
The application of Kalman filters to the state estimation problem in WSNs with intermittent observations due to data packet drops has undergone significant development recently\cite{1272646,1333199,KARIMI201949,9123558,8920146}. The development of distributed Kalman filter (DKF) approaches has demonstrated particularly great progress in this regard\cite{YANG2019156,wang2014information,7417484,8409298,alam2014distributed,6374678}. In a DKF approach, each sensor in a WSN computes local state estimation based on the information obtained from its own observations and the observations sent from its neighbors. In addition, algorithms based on the consensus approach have been proposed to address random link failure in WSNs due to data packet drops\cite{7417484}. This issue has also been addressed by developing a novel state estimation algorithm based on a distributed agent dynamical system\cite{alam2014distributed}. Zhou \emph{et al.}\cite{8409298} considered the existence of data packet drops when designing a DKF for WSNs, and the performance of a stationary DKF and a Kalman consensus filter (KCF) were also evaluated. However, all of the above-discussed algorithms have accounted for the impact of data packet drops in the state estimation process under a Gaussian noise assumption.
A number of studies have focused on conducting state estimation in WSNs under a non-Gaussian noise environment. For example, cost functions based on information theoretic learning (ITL) have been proposed\cite{wang2018complex,liu2007correntropy,wang2018switching,chen2017maximum,8424469,WANG20178659}. In addition, approaches based on the maximum correntropy (MC) criterion have demonstrated superior state estimation performance under conditions of impulsive noise\cite{liu2007correntropy,wang2018switching,8718566}. Moreover, a proposed DFK algorithm based on MC (i.e., the DMCKF algorithm) has been demonstrated to significantly improve the state estimation performance under non-Gaussian noise conditions\cite{wang2019distributed}. However, all of the above-discussed algorithms fail to take account of data packet drops in the state estimation process of WSNs.
The present study addresses the above-discussed issues by proposing a novel DMCKF algorithm accounting for data packet drops (i.e., the DMCKF-DPD algorithm) for conducting state estimation in WSNs. In contrast to conventional DMCKF algorithms, the proposed algorithm is based on a fixed-point iterative algorithm to update the posterior estimates of the WSN node states. The effectiveness and feasibility of the algorithm are verified by simulations conducted in a WSN with intermittent observations due to data packet drops under a non-Gaussian noise environment. Moreover, the computational complexity of the DMCKF-DPD algorithm is demonstrated to be moderate compared with that of a conventional stationary DKF\cite{8409298}, and we provide a sufficient condition to ensure the convergence of the proposed algorithm.
The remainder of this paper is organized as follows. In Section \ref{section:Preliminaries}, the MC criterion and state-space model are briefly reviewed. In Section \ref{section:derivation}, the derivation, computational complexity, and convergence issue of the DMCKF-DPD algorithm are presented. The simulation examples are provided in Section \ref{section:simulations}, and, finally, the conclusion is given in Section \ref{section:Conclusion}.
\section{Preliminaries} \label{section:Preliminaries}
\subsection{Correntropy}
The concept of correntropy, first proposed by Principe \emph{et al.}\cite{liu2007correntropy}, is a very practical method for evaluating the generalized similarity between random variables ${X,Y \in R}$ with the same dimensions. Here, correntropy is defined as
\begin{equation}
\begin{split}
\operatorname{V} \left( {X,Y} \right) = \operatorname{E} \left[ {\kappa \left( {X,Y} \right)} \right] = \int {\kappa \left( {x,y} \right)} \operatorname{d} {\operatorname{F} _{XY}}\left( {x,y} \right),
\end{split}
\end{equation}
where ${{\text{E[}}{\text{.]}}}$ is the expectation operator, ${\operatorname{V} (.)}$ represents the information potential, ${{\operatorname{F} _{XY}}(x,y)}$ represents the probability distribution function (PDF) with respect to variables ${X}$ and ${Y}$, and ${\kappa \left( { \cdot {\rm{ }},{\rm{ }} \cdot } \right)}$ is the shift-invariant Mercer kernel. Here, we employ the Gaussian kernel, which is given as
\begin{equation}\label{equ:epxeeg}
\begin{split}
\kappa \left( {x,y} \right) = {\operatorname{G} _\sigma }\left( e \right) = \exp \left( { - \frac{{{e^2}}}{{2{\sigma ^2}}}} \right),
\end{split}
\end{equation}
where ${e = x - y}$ represents the error between elements ${x}$ and ${y}$, and ${\sigma > 0}$ represents the kernel bandwidth (or kernel size) of the Gaussian kernel function.
However, only a limited amount of data related to the variables ${X}$ and ${Y}$ can be obtained in realistic scenarios, and the PDF ${{\operatorname{F} _{XY}}(x,y)}$ is usually unknown. Under these conditions, a sample estimator can be used to calculate the correntropy as follows:
\begin{equation}
\begin{split}
\hat V\left( {X,Y} \right) = \frac{1}{N}\sum\limits_{i = 1}^N {{\operatorname{G} _\sigma }} \left( {{e^i}} \right),
\end{split}
\end{equation}
where
\begin{equation}
\begin{split}
{e^i} = {x^i} - {y^i},\left( {{x^i},{y^i} \in \left\{ {{x^i},{y^i}} \right\}_{i = 1}^N} \right),
\end{split}
\end{equation}
and ${N}$ samples are employed to define ${{\operatorname{F} _{XY}}(x,y)}$.
Applying a Taylor expansion to the Gaussian kernel function yields the following:
\begin{equation}\label{equ:n2yjxkke}
\begin{split}
\hat V\left( {X,Y} \right) = \sum\limits_{n = 0}^\infty {\frac{{{{\left( { - 1} \right)}^n}}}{{{2^n}{\sigma ^{2n}}n!}}\operatorname{E} \left[ {{{\left( {X - Y} \right)}^{2n}}} \right]} .
\end{split}
\end{equation}
We can infer from (\ref{equ:n2yjxkke}) that the correntropy is the weighted sum of all even-order moments of relation ${X - Y}$. Compared with other similarity measurement schemes, such as the mean-square error (MSE) criterion, correntropy contains all even-order moments, and is therefore useful for nonlinear and signal processing applications in non-Gaussian noise environments.
\subsection{State-space model}
The state-space model of the DKF over the nodes of a WSN under conditions of data packet drops is the first problem that must be solved to estimate the state of a target node. Here, we adopt a powerful state-space model \cite{8409298}, which is described as follows.
Consider a WSN with ${N}$ sensors. The states and observations are as follows:
\begin{equation}
\begin{split}
\begin{gathered}
{{\boldsymbol{x}}_k} = {{\boldsymbol{A}}_k}{{\boldsymbol{x}}_{k - 1}} + {{\boldsymbol{q}}_k}, \hfill \\
{\boldsymbol{y}}_k^i = {{\boldsymbol{C}}^i}{\boldsymbol{x}}_k^i + {\boldsymbol{v}}_k^i,{\text{ }}i = 1,2, \cdots ,N,{\text{and }}i \in \Omega , \hfill \\
\end{gathered}
\end{split}
\end{equation}
where ${{{\boldsymbol{x}}_k},{{\boldsymbol{x}}_{k - 1}} \in {\mathbb{R}^{n \times 1}}}$ represent the states of the dynamical system at instants ${k}$ and ${k - 1}$, ${\Omega }$ represents the set of all sensors in the network, ${{\boldsymbol{y}}_k^i \in {\mathbb{R}^{{m_i} \times 1}}({m_i} = \operatorname{rank} ({{\boldsymbol{C}}_i})\forall i \in \Omega )}$ denotes the observations obtained by node ${i}$ at instant ${k}$, ${{{\boldsymbol{A}}_k}}$ and ${{{\boldsymbol{C}}^i}}$ represent the state transition matrix and observation matrix of the system, respectively, and ${{{\boldsymbol{q}}_k}}$ and ${{\boldsymbol{v}}_k^i}$ are the mutually uncorrelated process noise and measurement noise, respectively, with means of zero and covariances respectively given as follows:
\begin{equation}
\begin{split}
{\text{E}}\left[ {{{\boldsymbol{w}}_k}{\boldsymbol{w}}_k^{\text{T}}} \right] = {\boldsymbol{P}},{\text{ E}}\left[ {{\boldsymbol{v}}_k^i{{({\boldsymbol{v}}_k^i)}^{\text{T}}}} \right] = {{\boldsymbol{R}}^i}.
\end{split}
\end{equation}
We then define ${{\Omega _i} \subset \Omega }$ to represent the set of all neighboring sensors of the ${i}$th sensor. Therefore, set ${{\mho _i} = {\Omega _i} \cup \{ i\}}$ contains node ${i}$ itself and all its neighboring nodes. According to the properties of a WSN, the process whereby node ${i}$ receives the observations from its neighboring node ${j}$ at instant ${k}$ can be modeled as follows:
\begin{equation}
\begin{split}\label{equ:iosjoosjkj}
{\boldsymbol{s}}_k^{i,j} = \gamma _k^{i,j}{\boldsymbol{y}}_k^j,{\text{ }}i \in \Omega ,{\text{ }}j \in {\mho _i}.
\end{split}
\end{equation}
Here, the term ${\gamma _k^{i,j}}$ is used to indicate the existence of data packet drops over the WSN. We employ the Bernoulli model for modeling the data packet loss process and assume that the process has independent and identically distributed properties for ${\gamma _k^{i,j}}$ with the probability
\begin{equation}
\begin{split}
\operatorname{P} \left\{ {\gamma _k^{i;j} = 1} \right\} = p_k^{i;j} > 0{\text{ }}\forall k \geqslant 0,
\end{split}
\end{equation}
and
\begin{equation}
\begin{split}
\begin{gathered}
\operatorname{E} \left[ {\gamma _k^{i;j} = 1} \right] = p_k^{i;j}, \hfill \\
{\left( {\mu _k^{i;j}} \right)^2} = \operatorname{E} \left[ {{{\left( {\gamma _k^{i;j} - p_k^{i;j}} \right)}^2}} \right] = p_k^{i;j} - {\left( {p_k^{i;j}} \right)^2}. \hfill \\
\end{gathered}
\end{split}
\end{equation}
We assume that variables ${\gamma _k^{i,j}}$ and ${\gamma _k^{j,i}}$ are mutually independent (i.e., ${\forall i \ne j}$), but the condition ${{p_{i,j}} = {p_{j,i}}}$ is allowed, and ${\gamma _k^{i,j}}$ is independent of the process noise, measurement noise, and initial state of the dynamical system. According to (\ref{equ:iosjoosjkj}), if sensor ${i}$ receives the observations from its neighboring node ${j}$ at instant ${k}$, then ${\gamma _k^{i,j} = 1}$ and ${\gamma _k^{i,j} = 0}$ when all the components of ${{{\boldsymbol{y}}_k^j}}$ are lost. We assume that node ${i}$ can receive all observations from itself at any time, so ${\gamma _k^{i,j} \equiv 1}$.
Based on the above analysis, the various factors of a DKF model that must take data packet drops into account can be expressed as
\begin{equation}
\begin{split}
\left\{ \begin{gathered}
{\boldsymbol{y}}_k^{{\mho _i}} = {\text{vec}}{\left\{ {{\boldsymbol{y}}_k^j} \right\}_{j \in {\mho _i}}}{\text{,}} \hfill \\
{\boldsymbol{v}}_k^{{\mho _i}}{\text{ }} = {\text{vec}}{\left\{ {{\boldsymbol{v}}_k^j} \right\}_{j \in {\mho _i}}}, \hfill \\
{\boldsymbol{s}}_k^{{\mho _i}} = {\text{vec}}{\left\{ {{\boldsymbol{s}}_k^j} \right\}_{j \in {\mho _i}}}, \hfill \\
{{\boldsymbol{C}}^{{\mho _i}}} = {\text{col}}{\left\{ {{{\boldsymbol{C}}^j}} \right\}_{j \in {\mho _i}}}, \hfill \\
{{\boldsymbol{R}}^{{\mho _i}}} = \operatorname{diag} {\left\{ {{{\boldsymbol{R}}^j}} \right\}_{j \in {\mho _i}}}, \hfill \\
{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}} = \operatorname{diag} {\left\{ {\gamma _k^{i,j}{{\boldsymbol{I}}_{{m_j}}}} \right\}_{j \in {\mho _i}}}. \hfill \\
\end{gathered} \right.
\end{split}
\end{equation}
Here, ${{\text{vec}}\left\{ \cdot \right\}}$, ${{\text{col}}\left\{ \cdot \right\}}$, and ${\operatorname{diag} \left\{ \cdot \right\}}$ denote the vectorization operation, the columnization operation, and the (block) diagonalization operation, respectively, and ${{{\boldsymbol{I}}_{{m_j}}}(j \in {\mho _i})}$ represents an identity matrix of dimension ${{m_j} \times {m_j}}$. According to the above discussion, the state-space model of the ${i}$th sensor in the case of data packet drops can be written in a compact form as
\begin{equation}\label{equ:kiqj1jkisaa}
\begin{split}
{\boldsymbol{x}}_k^i = {\boldsymbol{Ax}}_{k - 1}^i + {\boldsymbol{q}}_k^i,
\end{split}
\end{equation}
\begin{equation}\label{equ:aa}
\begin{split}
{\boldsymbol{y}}_k^{{\mho _i}} = {{\boldsymbol{C}}^{{\mho _i}}}{\boldsymbol{x}}_k^i + {\boldsymbol{v}}_k^{{\mho _i}},
\end{split}
\end{equation}
\begin{equation}\label{equ:kiyikd}
\begin{split}
{\boldsymbol{s}}_k^{{\mho _i}} = {\boldsymbol{D}}_k^{\gamma ;{\mho _i}}{\boldsymbol{y}}_k^{{\mho _i}}.
\end{split}
\end{equation}
In addition, the covariance of ${{\boldsymbol{v}}_k^{{\mho _i}}}$ is
\begin{equation}
\begin{split}
{{\boldsymbol{R}}^{{\mho _i}}} = \operatorname{diag} {\left\{ {{{\boldsymbol{R}}^j}} \right\}_{j \in {\mho _i}}}.
\end{split}
\end{equation}
\section{Proposed DMCKF-DPD algorithm}\label{section:derivation}
\subsection{Algorithm derivation}
We first derive the stationary DMCKF in the case of data packet drops based on the model given by (\ref{equ:kiqj1jkisaa})-(\ref{equ:kiyikd}). The state prediction error of the algorithm can be written as
\begin{equation}
\begin{split}
{\boldsymbol{\varepsilon }}_{k|k - 1}^i = {\boldsymbol{x}}_k^i - {\boldsymbol{\hat x}}_{k|k - 1}^i.
\end{split}
\end{equation}
Combining this with the above state-space model yields the following augmented model:
\begin{equation}\label{equ:kigjkixkio}
\begin{split}
\left[ {\begin{array}{*{20}{c}}
{{\boldsymbol{\hat x}}_{k|k - 1}^i} \\
{{\boldsymbol{s}}_k^{{\mho _i}}}
\end{array}} \right] = \left[ {\begin{array}{*{20}{c}}
{{{\boldsymbol{I}}_n}} \\
{{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}{{\boldsymbol{C}}^{{\mho _i}}}}
\end{array}} \right]{\boldsymbol{x}}_k^i{\text{ + }}{\boldsymbol{g}}_k^i,
\end{split}
\end{equation}
where ${{{{\boldsymbol{I}}_n}}}$ represents an ${n \times n}$ identity matrix and ${{\boldsymbol{g}}_k^i}$ represents the augmented noise vector containing the state and measurement errors of the dynamical system, which is defined as
\begin{equation}
\begin{split}
{\boldsymbol{g}}_k^i = \left[ {\begin{array}{*{20}{c}}
{ - {\boldsymbol{\varepsilon }}_{k|k - 1}^i} \\
{{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}{\boldsymbol{v}}_k^{{\mho _i}}}
\end{array}} \right].
\end{split}
\end{equation}
Assuming that the covariance matrix of the augmented vector ${{\text{E}}[{\boldsymbol{g}}_k^i{({\boldsymbol{g}}_k^i)^{\text{T}}}]}$ is positive definite yields the following:
\begin{equation}\label{equ:tkikbkikb}
\begin{split}
\begin{gathered}
{\text{E}}\left[ {{\boldsymbol{g}}_k^i{{\left( {{\boldsymbol{g}}_k^i} \right)}^{\text{T}}}} \right] = \left[ {\begin{array}{*{20}{c}}
{{\boldsymbol{P}}_{k|k - 1}^i}&0 \\
0&{{\boldsymbol{D}}_{p;k}^{{\mho _i}}{{\boldsymbol{R}}^{{\mho _i}}}{\boldsymbol{D}}_{p;k}^{{\mho _i}}}
\end{array}} \right] \hfill \\
= \left[ {\begin{array}{*{20}{c}}
{{\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i{{\left( {{\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i} \right)}^{\text{T}}}}&0 \\
0&{{\boldsymbol{B}}_{R;k}^i{{\left( {{\boldsymbol{B}}_{R;k}^i} \right)}^{\text{T}}}}
\end{array}} \right] \hfill \\
= {\boldsymbol{B}}_k^i{\left( {{\boldsymbol{B}}_k^i} \right)^{\text{T}}}, \hfill \\
\end{gathered}
\end{split}
\end{equation}
with
\begin{equation}
\begin{split}
{\text{E}}\left[ {{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}} \right] = \operatorname{diag} {\left\{ {p_k^{i;j}{{\boldsymbol{I}}_{{m_j}}}} \right\}_{j \in {\mho _i}}} = {\boldsymbol{D}}_{p;k}^{{\mho _i}}.
\end{split}
\end{equation}
Here, ${{\boldsymbol{B}}_k^i}$ and ${{({\boldsymbol{B}}_k^i)^{\text{T}}}}$ can be obtained by the Cholesky decomposition of ${{\text{E}}[{\boldsymbol{g}}_k^i{({\boldsymbol{g}}_k^i)^{\text{T}}}]}$, and matrix ${{\boldsymbol{D}}_{p;k}^{{\mho _i}}}$ represents the expectation of ${{{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}}}$. Left multiplying each term of (\ref{equ:kigjkixkio}) by $ {{({\boldsymbol{B}}_k^i)^{ - 1}}}$ yields the following:
\begin{equation}
\begin{split}
{\boldsymbol{D}}_k^i = {\boldsymbol{W}}_k^i{\boldsymbol{x}}_k^i + {\boldsymbol{e}}_k^i,
\end{split}
\end{equation}
where
\begin{equation}\label{equ:jkiociuf}
\begin{split}
\left\{ {\begin{array}{*{20}{l}}
{{\boldsymbol{D}}_k^i = {{\left( {{\boldsymbol{B}}_k^i} \right)}^{ - 1}}\left[ {\begin{array}{*{20}{c}}
{{\boldsymbol{\hat x}}_{k|k - 1}^i} \\
{{\boldsymbol{s}}_k^{{\mho _i}}}
\end{array}} \right],} \\
{{\boldsymbol{W}}_k^i = {{\left( {{\boldsymbol{B}}_k^i} \right)}^{ - 1}}\left[ {\begin{array}{*{20}{c}}
{\boldsymbol{I}} \\
{{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}{{\boldsymbol{C}}^{{\mho _i}}}}
\end{array}} \right],} \\
{{\boldsymbol{e}}_k^i = {{\left( {{\boldsymbol{B}}_k^i} \right)}^{ - 1}}{\boldsymbol{g}}_k^i.}
\end{array}} \right.
\end{split}
\end{equation}
According to the above derivation, we propose the following cost function based on the MC criterion:
\begin{equation}
\begin{split}
{J_L}\left( {x_k^i} \right) = \frac{1}{L}\sum\limits_{h = 1}^L {{\operatorname{G} _\sigma }\left( {d_k^{i;h} - {\boldsymbol{w}}_k^{i;h}{\boldsymbol{x}}_k^i} \right)},
\end{split}
\end{equation}
where ${d_k^{i;h}}$ represents the ${h}$th element of ${{\boldsymbol{D}}_k^i}$, ${{\boldsymbol{w}}_k^{i;h}}$ represents the ${h}$th row of ${{\boldsymbol{W}}_k^i}$, and
\begin{equation}
\begin{split}
L = n + {m_{i;a}},\left( {{{\left( {{m_{i;a}}} \right)}_{i \in \Omega }} = \sum\limits_{j \in {\Omega _i}} {{m_j}} } \right)
\end{split}
\end{equation}
denotes the number of elements of ${{\boldsymbol{D}}_k^i}$. Then, the objective function of the optimal state estimation ${{\boldsymbol{x}}_k^i}$ based on the MC criterion is
\begin{equation}
\begin{split}
{\boldsymbol{\hat x}}_k^i = \mathop {\max }\limits_{{\boldsymbol{x}}_k^i} {J_L}\left( {{\boldsymbol{x}}_k^i} \right) = \mathop {\max }\limits_{{\boldsymbol{x}}_k^i} \sum\limits_{h = 1}^L {{\operatorname{G} _\sigma }} \left( {e_k^{i;h}} \right),
\end{split}
\end{equation}
where ${e_k^{i;h} = d_k^{i;h} - {\boldsymbol{w}}_k^{i;h}{\boldsymbol{x}}_k^i}$ and ${\left( {h = 1,2, \cdots L} \right)}$ is the ${h}$th element of ${e_k^{i}}$. Finally, the optimal state estimation of ${{\boldsymbol{\hat x}}_k^i}$ is achieved by maximizing the information potential ${\operatorname{V} (.)}$. To that end, we set the gradient of the cost function ${{{J_L}({\boldsymbol{x}}_k^i)}}$ regarding ${{{\boldsymbol{x}}_k^i}}$ to zero:
\begin{equation}
\begin{split}
\begin{gathered}
\frac{{\partial {J_L}\left( {{\boldsymbol{x}}_k^i} \right)}}{{\partial {\boldsymbol{x}}_k^i}} \hfill \\
= \frac{1}{{{\sigma ^2}}}\sum\limits_{h = 1}^L {{\operatorname{G} _\sigma }\left( {e_k^{i;h}} \right)} {\left( {{\boldsymbol{w}}_k^{i;h}} \right)^{\text{T}}}\left( {d_k^{i;h} - {\boldsymbol{w}}_k^{i;h}{\boldsymbol{x}}_k^i} \right) \hfill \\
= 0, \hfill \\
\end{gathered}
\end{split}
\end{equation}
and obtaining the optimal state of ${{{\boldsymbol{x}}_k^i}}$ is relatively easy:
\begin{equation}\label{equ:khfidtkkhf}
\begin{split}
\begin{gathered}
{\boldsymbol{x}}_k^i = {\left[ {\sum\limits_{h = 1}^L {{\operatorname{G} _\sigma }\left( {e_k^{i;h}} \right){{\left( {{\boldsymbol{w}}_k^{i;h}} \right)}^{\text{T}}}} {\boldsymbol{w}}_k^{i;h}} \right]^{ - 1}} \times \hfill \\
\sum\limits_{h = 1}^L {{\operatorname{G} _\sigma }\left( {e_k^{i;h}} \right){{\left( {{\boldsymbol{w}}_k^{i;h}} \right)}^{\text{T}}}} d_k^{i;h}. \hfill \\
\end{gathered}
\end{split}
\end{equation}
Because ${e_k^{i;h} = d_k^{i;h} - {\boldsymbol{w}}_k^{i;h}{\boldsymbol{x}}_k^i}$ is a function of ${{\boldsymbol{x}}_k^i}$, the optimal state estimation in (\ref{equ:khfidtkkhf}) is a fixed-point iterative equation of ${{\boldsymbol{x}}_k^i}$, and can be rewritten in the following form:
\begin{equation}\label{equ:dkixkkfddkix}
\begin{split}
{\boldsymbol{x}}_k^i = \operatorname{f} \left( {{\boldsymbol{x}}_k^i} \right),
\end{split}
\end{equation}
where
\begin{equation}
\begin{split}
\begin{gathered}
\operatorname{f} \left( {{\boldsymbol{x}}_k^i} \right) = {\left[ {\sum\limits_{h = 1}^L {{\operatorname{G} _\sigma }\left( {d_k^{i;h} - {\boldsymbol{w}}_k^{i;h}{\boldsymbol{x}}_k^i} \right)} {{\left( {{\boldsymbol{w}}_k^{i;h}} \right)}^{\text{T}}}{\boldsymbol{w}}_k^{i;h}} \right]^{ - 1}} \hfill \\
\sum\limits_{h = 1}^L {{\operatorname{G} _\sigma }\left( {d_k^{i;h} - {\boldsymbol{w}}_k^{i;h}{\boldsymbol{x}}_k^i} \right)} {\left( {{\boldsymbol{w}}_k^{i;h}} \right)^{\text{T}}}{\boldsymbol{d}}_k^{i;h}. \hfill \\
\end{gathered}
\end{split}
\end{equation}
According to the above derivation, the iterative equation in (\ref{equ:dkixkkfddkix}) can be written as follows:
\begin{equation}
\begin{split}
{\left( {{\boldsymbol{\hat x}}_k^i} \right)_{t + 1}} = \operatorname{f} \left[ {{{\left( {{\boldsymbol{x}}_k^i} \right)}_t}} \right].
\end{split}
\end{equation}
Here, ${{({\boldsymbol{\hat x}}_k^i)_{t + 1}}}$ represents the result of ${{\boldsymbol{x}}_k^i}$ at the fixed-point iteration ${t+1}$, and (\ref{equ:khfidtkkhf}) can be further written in the form of matrix multiplication:
\begin{equation}\label{equ:kidkiatkwi}
\begin{split}
{\boldsymbol{x}}_k^i = {\left[ {{{\left( {{\boldsymbol{W}}_k^i} \right)}^{\text{T}}}{\boldsymbol{\Lambda }}_k^i{\boldsymbol{W}}_k^i} \right]^{ - 1}}{\left( {{\boldsymbol{W}}_k^i} \right)^{\text{T}}}{\boldsymbol{\Lambda }}_k^i{\boldsymbol{D}}_k^i,
\end{split}
\end{equation}
where
\begin{equation}\label{equ:jkktjsikesg}
\begin{split}
\begin{gathered}
{\boldsymbol{\Lambda }}_k^i = \left[ {\begin{array}{*{20}{c}}
{{\boldsymbol{\Lambda }}_{x;k}^i}&0 \\
0&{{\boldsymbol{\Lambda }}_{y;k}^i}
\end{array}} \right], \hfill \\
{\boldsymbol{\Lambda }}_{x;k}^i = {\text{diag}}\left[ {{\operatorname{G} _\sigma }\left( {e_k^{i;1}} \right), \cdots ,{\operatorname{G} _\sigma }\left( {e_k^{i;n}} \right)} \right], \hfill \\
{\boldsymbol{\Lambda }}_{y;k}^i = {\text{diag}}\left[ {{\operatorname{G} _\sigma }\left( {e_k^{i;n + 1}} \right), \cdots ,{\operatorname{G} _\sigma }\left( {e_k^{i;n + {m_{i;a}}}} \right)} \right]. \hfill \\
\end{gathered}
\end{split}
\end{equation}
According to (\ref{equ:jkiociuf}), (\ref{equ:kidkiatkwi}), and (\ref{equ:jkktjsikesg}), we obtain the following:
\begin{equation}
\begin{split}
\begin{gathered}
{\left[ {{{\left( {{\boldsymbol{W}}_k^i} \right)}^{\text{T}}}{\boldsymbol{\Lambda }}_k^i{\boldsymbol{W}}_k^i} \right]^{ - 1}} \hfill \\
= {\left\{ \begin{gathered}
{\left[ {{{\left( {{\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i} \right)}^{ - 1}}} \right]^{\text{T}}}{\boldsymbol{\Lambda }}_{x;k}^i{\left( {{\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i} \right)^{ - 1}} \hfill \\
+ {\left( {{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}{{\boldsymbol{C}}^{{\mho _i}}}} \right)^{\text{T}}} \times \hfill \\
{\left[ {{{\left( {{\boldsymbol{B}}_{R;k}^i} \right)}^{ - 1}}} \right]^{\text{T}}}{\boldsymbol{\Lambda }}_{y;k}^i{\left( {{\boldsymbol{B}}_{R;k}^i} \right)^{ - 1}}\left( {{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}{{\boldsymbol{C}}^{{\mho _i}}}} \right) \hfill \\
\end{gathered} \right\}^{ - 1}}. \hfill \\
\end{gathered}
\end{split}
\end{equation}
We then apply the following matrix inversion lemma \cite{chen2017maximum}
\begin{equation}
\begin{split}
\begin{gathered}
{\left( {{\boldsymbol{G}} + {\boldsymbol{BCD}}} \right)^{ - 1}} \hfill \\
= {{\boldsymbol{G}}^{ - 1}} - {{\boldsymbol{G}}^{ - 1}}{\boldsymbol{B}}{\left( {{{\boldsymbol{C}}^{ - 1}} + {\boldsymbol{D}}{{\boldsymbol{G}}^{ - 1}}{\boldsymbol{B}}} \right)^{ - 1}}{\boldsymbol{D}}{{\boldsymbol{G}}^{ - 1}} \hfill \\
\end{gathered}
\end{split}
\end{equation}
with the identifications
\begin{equation}
\begin{split}
\left\{ \begin{gathered}
{\boldsymbol{G}} = {\left[ {{{\left( {{\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i} \right)}^{ - 1}}} \right]^{\text{T}}}{\boldsymbol{\Lambda }}_{x;k}^i{\left( {{\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i} \right)^{ - 1}}, \hfill \\
{\boldsymbol{B}} = {\left( {{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}{{\boldsymbol{C}}^{{\mho _i}}}} \right)^{\text{T}}}, \hfill \\
{\boldsymbol{C}} = {\left[ {{{\left( {{\boldsymbol{B}}_{R;k}^i} \right)}^{ - 1}}} \right]^{\text{T}}}{\boldsymbol{\Lambda }}_{y;k}^i{\left( {{\boldsymbol{B}}_{R;k}^i} \right)^{ - 1}}, \hfill \\
{\boldsymbol{D}} = {\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}{{\boldsymbol{C}}^{{\mho _i}}}, \hfill \\
\end{gathered} \right.
\end{split}
\end{equation}
and obtain the expression (\ref{equ:jt1jkpkib}) given below.
\newcounter{mytempeqncnt}
\begin{figure*}[!t]
\normalsize
\setcounter{mytempeqncnt}{\value{equation}}
\begin{equation}\label{equ:jt1jkpkib}
\begin{gathered}
\begin{gathered}
{\left[ {{{\left( {{\boldsymbol{W}}_k^i} \right)}^{\text{T}}}{\boldsymbol{\Lambda }}_k^i{\boldsymbol{W}}_k^i} \right]^{ - 1}} = {\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i{\left( {{\boldsymbol{\Lambda }}_{x;k}^i} \right)^{ - 1}}{\left( {{\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i} \right)^{\text{T}}} - {\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i{\left( {{\boldsymbol{\Lambda }}_{x;k}^i} \right)^{ - 1}}{\left( {{\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i} \right)^{\text{T}}}{\left( {{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}{C^{{\mho _i}}}} \right)^{\text{T}}} \hfill \\
{\left[ {{\boldsymbol{B}}_{R;k}^i{{\left( {{\boldsymbol{\Lambda }}_{y;k}^i} \right)}^{ - 1}}{{\left( {{\boldsymbol{B}}_{R;k}^i} \right)}^{\text{T}}} + \left( {{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}{{\boldsymbol{C}}^{{\mho _i}}}} \right){\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i{{\left( {{\boldsymbol{\Lambda }}_{x;k}^i} \right)}^{ - 1}}{{\left( {{\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i} \right)}^{\text{T}}}{{\left( {{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}{C^{{\mho _i}}}} \right)}^{\text{T}}}} \right]^{ - 1}} \hfill \\
\times \left( {{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}{{\boldsymbol{C}}^{{\mho _i}}}} \right){\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i{\left( {{\boldsymbol{\Lambda }}_{x;k}^i} \right)^{ - 1}}{\left( {{\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i} \right)^{\text{T}}} \hfill \\
\end{gathered}
\end{gathered}
\end{equation}
\hrulefill
\vspace*{4pt}
\end{figure*}
According to (\ref{equ:jkiociuf}), (\ref{equ:kidkiatkwi}), and (\ref{equ:jkktjsikesg}), we obtain the following:
\begin{equation}\label{equ:kious1fkrikb}
\begin{split}
\begin{gathered}
{\left( {{\boldsymbol{W}}_k^i} \right)^{\text{T}}}{\boldsymbol{\Lambda }}_k^i{\boldsymbol{D}}_k^i \hfill \\
= {\left[ {{{\left( {{\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i} \right)}^{ - 1}}} \right]^{\text{T}}}{\boldsymbol{\Lambda }}_{x;k}^i{\left( {{\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i} \right)^{ - 1}}{\boldsymbol{\hat x}}_{k|k - 1}^i \hfill \\
+ {\left( {{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}{{\boldsymbol{C}}^{{\mho _i}}}} \right)^{\text{T}}}{\left[ {{{\left( {{\boldsymbol{B}}_{R;k}^i} \right)}^{ - 1}}} \right]^{\text{T}}}{\boldsymbol{\Lambda }}_{y;k}^i{\left( {{\boldsymbol{B}}_{R;k}^i} \right)^{ - 1}}{\boldsymbol{s}}_k^{{\mho _i}}. \hfill \\
\end{gathered}
\end{split}
\end{equation}
Substituting formulas (\ref{equ:jt1jkpkib}) and (\ref{equ:kious1fkrikb}) into (\ref{equ:kidkiatkwi}) yields
\begin{equation}\label{equ:tijksxiug}
\begin{split}
{\boldsymbol{\hat x}}_{k|k}^i = {\boldsymbol{\hat x}}_{k|k - 1}^i + {\boldsymbol{\bar K}}_k^i\left[ {{\boldsymbol{s}}_k^{{\mho _i}} - {\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}{{\boldsymbol{C}}^{{\mho _i}}}{\boldsymbol{\hat x}}_{k|k - 1}^i} \right],
\end{split}
\end{equation}
where
\begin{equation}
\begin{split}
\left\{ \begin{gathered}
{\boldsymbol{\bar K}}_k^i = {\boldsymbol{\bar P}}_{k|k - 1}^i{\left( {{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}{{\boldsymbol{C}}^{{\mho _i}}}} \right)^{\text{T}}} \hfill \\
\left[ {{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}{{\boldsymbol{C}}^{{\mho _i}}}{\boldsymbol{\bar P}}_{k|k - 1}^i{{\left( {{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}{{\boldsymbol{C}}^{{\mho _i}}}} \right)}^{\text{T}}} + {\boldsymbol{\bar R}}_k^i} \right], \hfill \\
{\boldsymbol{\bar P}}_{k|k - 1}^i = {\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i{\left( {{\boldsymbol{\Lambda }}_{x;k}^i} \right)^{ - 1}}{\left( {{\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i} \right)^{\text{T}}}, \hfill \\
{\boldsymbol{\bar R}}_k^i = {\boldsymbol{B}}_{R;k}^i{\left( {{\boldsymbol{\Lambda }}_{y;k}^i} \right)^{ - 1}}{\left( {{\boldsymbol{B}}_{R;k}^i} \right)^{\text{T}}}. \hfill \\
\end{gathered} \right.
\end{split}
\end{equation}
\textbf{Remark 1.} Equation (\ref{equ:tijksxiug}) is the optimal solution of ${{\boldsymbol{x}}_k^i}$, and it depends on the prior estimate ${{\boldsymbol{\hat x}}_{k|k - 1}^i}$ and available observation information ${{{\boldsymbol{s}}_k^{{\mho _i}}}}$. The observation information obtainable by the ${i}$th node is determined by the matrix ${{{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}}}$.
According to the above derivations, we summarize the steps of the proposed DMCKF-DPD algorithm as follows.
\begin{enumerate}
\item Initialize the values of the ${\sigma }$, ${\varepsilon }$ (a small positive number), ${{\boldsymbol{\hat x}}_{0|0}^i}$, and ${{\boldsymbol{P}}_{0|0}^i}$.
\item Apply (\ref{equ:tkikbkikb}) to calculate ${{\boldsymbol{B}}_k^i}$, and calculate ${{\boldsymbol{\hat x}}_{k|k - 1}^i}$ and ${{\boldsymbol{P}}_{k|k - 1}^i}$ as follows:
\begin{equation}\label{equ:kixaki}
\begin{split}
\begin{gathered}
{\boldsymbol{x}}_k^i = {\boldsymbol{Ax}}_{k - 1}^i, \hfill \\
{\boldsymbol{P}}_{k|k - 1}^i = {\boldsymbol{AP}}_{k - 1|k - 1}^i{{\boldsymbol{A}}^{\text{T}}} + {\boldsymbol{Q}}. \hfill \\
\end{gathered}
\end{split}
\end{equation}
\item Update the parameters ${{\boldsymbol{D}}_k^i = {\boldsymbol{W}}_k^i{\boldsymbol{x}}_k^i + {\boldsymbol{e}}_k^i}$ in the new system.
\item Utilize the fixed-point iterative algorithm to estimate the state of system as follows:
\begin{equation} \label{equ:kkiick}
\begin{split}
\begin{gathered}
{\left( {{\boldsymbol{\hat x}}_{k|k}^i} \right)_{t + 1}} \hfill \\
= {\boldsymbol{\hat x}}_{k|k - 1}^i + {\boldsymbol{\bar K}}_k^i\left[ {{\boldsymbol{s}}_k^{{\mho _i}} - {\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}{{\boldsymbol{C}}^{{\mho _i}}}{\boldsymbol{\hat x}}_{k|k - 1}^i} \right], \hfill \\
\end{gathered}
\end{split}
\end{equation}
where
\begin{equation} \label{equ:kirptikc}
\begin{split}
\begin{gathered}
{\boldsymbol{\tilde K}}_k^i = {\boldsymbol{\tilde P}}_{k|k - 1}^i{\left( {{{\boldsymbol{C}}^{{\mho _i}}}} \right)^{\text{T}}}{\left( {{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}} \right)^{\text{T}}} \times \hfill \\
\left[ {{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}{{\boldsymbol{C}}^{{\mho _i}}}{\boldsymbol{\bar P}}_{k|k - 1}^i{{\left( {{{\boldsymbol{C}}^{{\mho _i}}}} \right)}^{\text{T}}}{{\left( {{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}} \right)}^{\text{T}}} + {\boldsymbol{\tilde R}}_k^i} \right], \hfill \\
\end{gathered}
\end{split}
\end{equation}
\begin{equation} \label{equ:trkibyi}
\begin{split}
{\boldsymbol{\tilde P}}_{k|k - 1}^i = {\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i{\left( {{\boldsymbol{\tilde \Lambda }}_{x;k}^i} \right)^{ - 1}}{\left( {{\boldsymbol{B}}_{P\left( {k|k - 1} \right)}^i} \right)^{\text{T}}},
\end{split}
\end{equation}
\begin{equation} \label{equ:tkrikyi}
\begin{split}
{\boldsymbol{\tilde R}}_k^i = {\boldsymbol{B}}_{R;k}^i{\left( {{\boldsymbol{\tilde \Lambda }}_{y;k}^i} \right)^{ - 1}}{\left( {{\boldsymbol{B}}_{R;k}^i} \right)^{\text{T}}},
\end{split}
\end{equation}
\begin{equation} \label{equ:kinegkie}
\begin{split}
{\boldsymbol{\tilde \Lambda }}_{x;k}^i = \operatorname{diag} \left[ {{\operatorname{G} _\sigma }\left( {\tilde e_k^{i;1}} \right), \cdots ,{\operatorname{G} _\sigma }\left( {\tilde e_k^{i;n}} \right)} \right],
\end{split}
\end{equation}
\begin{equation} \label{equ:kngegknig}
\begin{split}
{\boldsymbol{\tilde \Lambda }}_{y;k}^i = \operatorname{diag} \left[ {{\operatorname{G} _\sigma }\left( {\tilde e_k^{i;n + 1}} \right), \cdots ,{\operatorname{G} _\sigma }\left( {\tilde e_k^{i;n + {m_{i;a}}}} \right)} \right],
\end{split}
\end{equation}
\begin{equation} \label{equ:tkkikihw}
\begin{split}
\tilde e_k^{i;h} = d_k^{i;h} - {\boldsymbol{w}}_k^{i;h}{\left( {{\boldsymbol{\hat x}}_{k|k}^{i;h}} \right)_t}.
\end{split}
\end{equation}
\item Compare the state values obtained by the fixed-point iterative algorithm at ${t}$ and ${t+1}$. If these two state estimates satisfy the following criterion:
\begin{equation} \label{equ:sgmxydys}
\begin{split}
\frac{{\left\| {{{\left( {{\boldsymbol{\hat x}}_{k|k}^i} \right)}_{t + 1}} - {{\left( {{\boldsymbol{\hat x}}_{k|k}^i} \right)}_t}} \right\|}}{{{{\left( {{\boldsymbol{\hat x}}_{k|k}^i} \right)}_t}}} \leqslant \varepsilon ,
\end{split}
\end{equation}
where ${\varepsilon }$ is the termination condition of the fixed-point iterative algorithm, set the optimal estimate of ${{\boldsymbol{\hat x}}_{k|k}^i}$ equal to ${{({\boldsymbol{\hat x}}_{k|k}^i)_{t + 1}}}$, and continue to step 6). Otherwise, set ${t \leftarrow t + 1}$ and return to step 4).
\item Update the covariance matrix as follows:
\begin{equation}\label{equ:dikikbkoikri}
\begin{split}
\begin{gathered}
{\boldsymbol{P}}_{k|k}^i = \left( {{\boldsymbol{I}} - {\boldsymbol{\tilde K}}_k^i{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}{{\boldsymbol{C}}^{{\mho _i}}}} \right){\boldsymbol{P}}_{k|k - 1}^i \hfill \\
\left[ {{\boldsymbol{I}} - {\boldsymbol{\tilde K}}_k^i{{\left( {{{\boldsymbol{C}}^{{\mho _i}}}} \right)}^{\text{T}}}{{\left( {{\boldsymbol{D}}_{\gamma ;k}^{{\mho _i}}} \right)}^{\text{T}}}} \right] + {\boldsymbol{\tilde K}}_k^i{\boldsymbol{R}}_k^{{\mho _i}}{\left( {{\boldsymbol{\tilde K}}_k^i} \right)^{\text{T}}}, \hfill \\
\end{gathered}
\end{split}
\end{equation}
set ${k \leftarrow k + 1}$, and return to step 2).
\end{enumerate}
\textbf{Remark 2.}
The kernel bandwidth ${\sigma }$ is an important parameter in the MC criterion. Broadly speaking, the convergence rate of the algorithm increases with increasing ${\sigma }$, but the accuracy of the algorithm decreases with increasing ${\sigma }$.
\subsection{Computational complexity}
The computational complexity of the proposed DMCKF-DPD algorithm can be compared with that of the stationary DKF \cite{8409298} in terms of the equations and operations employed by the two algorithms, which are listed in Table 1.
The stationary DKF algorithm mainly includes (\ref{equ:aa}), (\ref{equ:kiyikd}), and (\ref{equ:kixaki}) cited in the present paper, and (5) and (6) cited within\cite{8409298}. Therefore, we can evaluate the computational complexity of the stationary DKF as
\begin{equation}\label{equ:iugoiug}
\begin{split}
\begin{gathered}
{S_{SDKF}} = 11{n^3} + 12m_{i;a}^2n + 10{m_{i;a}}{n^2} + 4m_{i;a}^2 \hfill \\
- 2{m_{i;a}}n - {n^2} - 2n - {m_{i;a}} + 2O(m_{i;a}^3). \hfill \\
\end{gathered}
\end{split}
\end{equation}
The proposed algorithm mainly includes (\ref{equ:aa}) and (\ref{equ:kiyikd}), (\ref{equ:kixaki}), (\ref{equ:epxeeg}), and (\ref{equ:kkiick})-(\ref{equ:tkkikihw}). Accordingly, the computational complexity of the DMCKF-DPD algorithm can be defined based on an average number of fixed-point iterative algorithm iterations ${T}$ as
\begin{equation}\label{equ:ijgotno}
\begin{split}
\begin{gathered}
{S_{{\text{SDMCKF}}}} \hfill \\
{\text{ = 6}}Tm_{i;a}^3 + 6T{n^3} + {\text{16}}Tm_{i;a}^2n + 10T{m_{i;a}}{n^2} \hfill \\
+ (2 - 3T)m_{i;a}^2 + 2{n^2} + (2 - 3T){m_{i;a}}n + \hfill \\
({\text{6}}T - 1){m_{i;a}} + ({\text{6}}T - 1)n + T{\text{O(}}{n^3}{\text{)}} + T{\text{O(}}m_{i;a}^3{\text{)}}. \hfill \\
\end{gathered}
\end{split}
\end{equation}
\begin{table}\label{table1}
\caption{Computational complexities of the proposed DMCKF-DPD algorithm and the stationary DKF algorithm\cite{8409298}}
\begin{tabular*}{8.9cm}{lll}
\hline
Equation & Addition/subtraction & Division,\\
& and multiplication & matrix inversion,\\
& & Cholesky \\
& & decomposition,\\
& & and exponentiation\\
\hline
(\ref{equ:aa}) & ${2n{m_{i;a}}}$ & 0\\
(\ref{equ:kiyikd}) & ${2m_{i;a}^2 - {m_{i;a}}}$ & 0\\
(\ref{equ:kixaki}) & ${2{n^2} - n}$ & 0\\
{(5)} in\cite{8409298} & ${\begin{gathered}
9{n^3} - 4{n^2} + 6{m_{i;a}}{n^2} + \hfill \\
4m_{i;a}^2n - 3{m_{i;a}}n + m_{i;a}^2 \hfill \\
\end{gathered} }$ & ${{\text{O(}}m_{i;a}^3{\text{)}}}$\\
{(6a)} in\cite{8409298} & ${{\text{2}}m_{i;a}^2n + 3{m_{i;a}}n - n + 2{n^2}}$ & 0\\
{(6b)} in\cite{8409298} & ${\begin{gathered}
{\text{2}}{n^3} - {n^2} + {\text{4}}{m_{i;a}}{n^2} + \hfill \\
{\text{6}}m_{i;a}^2n - {\text{4}}{m_{i;a}}n + m_{i;a}^2{\text{ }} \hfill \\
\end{gathered} }$ & ${{\text{O(}}m_{i;a}^3{\text{)}}}$\\
(\ref{equ:epxeeg}) & ${2n + 4}$ & 2\\
(\ref{equ:kkiick}) & ${2m_{i;a}^2n + {\text{3}}{m_{i;a}}n}$ & 0\\
(\ref{equ:kirptikc}) & ${\begin{gathered}
2m_{i;a}^3 + 8m_{i;a}^2n + 4{m_{i;a}}{n^2} - \hfill \\
5{m_{i;a}}n - m_{i;a}^2 \hfill \\
\end{gathered} }$ & 0\\
(\ref{equ:trkibyi}) & ${4{n^3} + {\text{4}}n}$ & ${2n + {\text{O}}({n^3}){\text{ }}}$\\
(\ref{equ:tkrikyi}) & ${2{m_{i;a}}n + 4{m_{i;a}} + 4m_{i;a}^3 - 2m_{i;a}^2}$ & ${2{m_{i;a}} + {\text{O}}(m_{i;a}^3)}$\\
(\ref{equ:kinegkie}) & ${2{n^2} + {\text{4}}n}$ & ${2n}$\\
(\ref{equ:kngegknig}) & ${2{m_{i;a}}n + {\text{4}}{m_{i;a}}}$ & ${2{m_{i;a}}}$\\
(\ref{equ:tkkikihw}) & ${2n{\text{ }}}$ & 0\\
\hline
\end{tabular*}
\end{table}
We can infer from this discussion that the computational complexity of the DMCKF-DPD algorithm is moderate compared with that of the stationary DKF, provided that the value of ${T}$ is small, which is indeed the case, as will be demonstrated later in Section \ref{section:simulations}.
\subsection{Convergence issue}
A full analysis of the convergence behavior of the DMCKF-DPD algorithm based on the fixed-point iterative algorithm is very complicated. Therefore, we give only a sufficient condition that ensures the convergence of the fixed-point iterative algorithm. However, a detailed proof process is not presented here because the convergence condition is similar to the analysis presented in an earlier work\cite{chen2017maximum}, which can be consulted for additional details.
\textbf{Theorem 1.} First, we assume the conditions ${{\beta _i} > {\zeta _i} = \frac{{\sqrt n \sum\limits_{h = 1}^L {{{\left\| {{{\left( {{\boldsymbol{w}}_k^{i;h}} \right)}^{\text{T}}}} \right\|}_1}\left| {d_k^{i;h}} \right|} }}{{{\lambda _{\min }}\sum\limits_{h = 1}^L {{{\left( {{\boldsymbol{w}}_k^{i;h}} \right)}^{\text{T}}}{\boldsymbol{w}}_k^{i;h}} }}}$ and ${{\sigma _i} \ge \max {\rm{\{ }}\sigma _i^*,\sigma _i^\diamondsuit {\rm{\} }}}$. Here, ${\sigma _i^*}$ is the optimal result of the equation ${\phi ({\sigma _i}) = {\beta _i}}$, where
\newcounter{mytempeqncnt1}
\begin{figure*}[!t]
\normalsize
\setcounter{mytempeqncnt1}{\value{equation}}
\begin{equation}\label{equ:wlindsisikc}
\psi \left( {{\sigma _i}} \right) = \frac{{\sqrt n \sum\limits_{h = 1}^L {\left[ {\left( {{\beta _i}{{\left\| {{\boldsymbol{w}}_k^{i;h}} \right\|}_1} + \left| {d_k^{i;h}} \right|} \right){{\left\| {{\boldsymbol{w}}_k^{i;h}} \right\|}_1} \times \left( {{\beta _i}{{\left\| {{{\left( {{\boldsymbol{w}}_k^{i;h}} \right)}^{\text{T}}}{\boldsymbol{w}}_k^{i;h}} \right\|}_1} + {{\left\| {{{\left( {{\boldsymbol{w}}_k^{i;h}} \right)}^{\text{T}}}d_k^{i;h}} \right\|}_1}} \right)} \right]} }}{{\sigma _i^2{\lambda _{\min }}\left[ {\sum\limits_{h = 1}^L {{\operatorname{G} _{{\sigma _i}}}\left( {{\beta _i}{{\left\| {{\boldsymbol{w}}_k^{i;h}} \right\|}_1} + \left| {d_k^{i;h}} \right|} \right){{\left( {{\boldsymbol{w}}_k^{i;h}} \right)}^{\text{T}}}{\boldsymbol{w}}_k^{i;h}} } \right]}},{\sigma _i} \in (0,\infty )
\end{equation}
\vspace*{4pt}
\end{figure*}
\begin{equation}\label{equ:kddokz}
\begin{split}
\begin{gathered}
\phi \left( {{\sigma _i}} \right) = \frac{{\sqrt n \sum\limits_{h = 1}^L {{{\left\| {{{\left( {{\boldsymbol{w}}_k^{i;h}} \right)}^{\text{T}}}} \right\|}_1}\left| {d_k^{i;h}} \right|} }}{{{\lambda _{\min }}\left[ {\sum\limits_{h = 1}^L {{\operatorname{G} _{{\sigma _i}}}\left( {{\beta _i}{{\left\| {{\boldsymbol{w}}_k^{i;h}} \right\|}_1} + \left| {d_k^{i;h}} \right|} \right){{\left( {{\boldsymbol{w}}_k^{i;h}} \right)}^{\text{T}}}{\boldsymbol{w}}_k^{i;h}} } \right]}}, \hfill \\
{\sigma _i} \in \left( {0,\infty } \right), \hfill \\
\end{gathered}
\end{split}
\end{equation}
and ${\sigma _i^\diamondsuit }$ is the result of the equation ${\psi ({\sigma _i}) = {\alpha _i}(0 < {\alpha _i} < 1)}$, where ${\psi ({\sigma _i})}$is given in (\ref{equ:wlindsisikc}) below. Accordingly, it holds that ${||\operatorname{f} ({\boldsymbol{x}}_k^i)|{|_1} \leqslant {\beta _i}}$ and ${||{\nabla _{{\boldsymbol{x}}_k^i}}\operatorname{f} ({\boldsymbol{x}}_k^i)|{|_1} \leqslant {\alpha _i}}$ for all ${{\boldsymbol{x}}_k^i \in \{ {\boldsymbol{x}}_k^i \in {\mathbb{R}^n}:||{\boldsymbol{x}}_k^i|{|_1} \leqslant {\beta _i}\} }$. Here, the ${n \times n}$ Jacobian matrix ${\operatorname{f} ({\boldsymbol{x}}_k^i)}$ is given as follows:
\begin{equation}
\begin{split}
{\nabla _{{\boldsymbol{x}}_k^i}}\operatorname{f} \left( {{\boldsymbol{x}}_k^i} \right) = \left[ {\frac{\partial }{{\partial {\boldsymbol{x}}_k^{i;1}}}\operatorname{f} \left( {{\boldsymbol{x}}_k^i} \right) \cdots \frac{\partial }{{\partial {\boldsymbol{x}}_k^{i;n}}}\operatorname{f} \left( {{\boldsymbol{x}}_k^i} \right)} \right],
\end{split}
\end{equation}
where the terms are defined in (\ref{equ:kddokzkkhiw}) below
\newcounter{mytempeqncnt2}
\begin{figure*}[!t]
\normalsize
\setcounter{mytempeqncnt2}{\value{equation}}
\begin{equation}\label{equ:kddokzkkhiw}
\frac{\partial }{{\partial {\boldsymbol{x}}_k^{i;j}}}\operatorname{f} \left( {{\boldsymbol{x}}_k^i} \right) = - {{\boldsymbol{T}}_g}\frac{1}{{{\sigma ^2}}}\sum\limits_{i = 1}^L {\left[ {e_k^{i;h}w_k^{i;h;j}{\operatorname{G} _\sigma }\left( {e_k^i} \right){{\left( {{\boldsymbol{w}}_k^{i;h}} \right)}^{\text{T}}}{\boldsymbol{w}}_k^{i;h}} \right]} \operatorname{f} \left( {{\boldsymbol{x}}_k^i} \right) + {{\boldsymbol{T}}_g}\frac{1}{{{\sigma ^2}}}\sum\limits_{i = 1}^L {\left[ {e_k^{i;h}w_k^{i;h;j}{\operatorname{G} _\sigma }\left( {e_k^i} \right){{\left( {{\boldsymbol{w}}_k^{i;h}} \right)}^{\text{T}}}d_k^{i;h}} \right]}
\end{equation}
\hrulefill
\vspace*{4pt}
\end{figure*}
with
\begin{equation}
\begin{split}
{{\boldsymbol{T}}_g} = {\left[ {\sum\limits_{h = 1}^L {{\operatorname{G} _\sigma }\left( {d_k^{i;h} - {\boldsymbol{w}}_k^{i;h}{\boldsymbol{x}}_k^i} \right){{\left( {{\boldsymbol{w}}_k^{i;h}} \right)}^{\text{T}}}{\boldsymbol{w}}_k^{i;h}} } \right]^{ - 1}},
\end{split}
\end{equation}
and ${{w_k^{i;h;j}}}$ is the ${j}$th element of the vector ${{{\boldsymbol{w}}_k^{i;h}}}$.
According to Theorem 1, we obtain the following conditions:
\begin{equation}
\begin{split}
\left\{ {\begin{array}{*{20}{l}}
{{{\left\| {\operatorname{f} \left( {{\boldsymbol{x}}_k^i} \right)} \right\|}_1} \leqslant \beta ,} \\
{{{\left\| {{\nabla _{x_k^i}}\operatorname{f} \left( {{\boldsymbol{x}}_k^i} \right)} \right\|}_1} \leqslant \alpha < 1.}
\end{array}} \right.
\end{split}
\end{equation}
if the kernel bandwidth ${\sigma }$ is sufficiently large (e.g., greater than ${\max {\text{\{ }}\sigma _i^*,\sigma _i^\diamondsuit {\text{\} }}}$). According to the Banach fixed-point theorem the fixed-point iterative algorithm in DMCKF-DPD will surely converge to a unique fixed point in the range ${{\boldsymbol{x}}_k^i \in \{ {\boldsymbol{x}}_k^i \in {\mathbb{R}^n}: ||{\boldsymbol{x}}_k^i|{|_1} \leqslant {\beta _i}\} }$ provided that the initial state of the system meets the condition ${||{({\boldsymbol{x}}_k^i)_0}|| \leqslant \beta }$ and ${\sigma }$ is sufficiently large.
Theorem 1 demonstrates that the kernel bandwidth of the Gaussian kernel function has a important influence on the convergence of the DMCKF-DPD algorithm. Here, reducing the kernel bandwidth can improve the accuracy of state estimation, but this will also decrease the convergence rate of the algorithm or make it diverge. Conversely, increasing the kernel width will increase the convergence rate of the algorithm, but will often yield poor estimation performance under impulsive noise conditions. In practice, the kernel bandwidth can be selected by trial and error in accordance with the desired estimation accuracy and convergence rate of the algorithm.
\section{simulations}\label{section:simulations}
We compared the state estimation performances obtained by the proposed DMCKF-DPD algorithm and the stationary DKF algorithm\cite{8409298} when applying them to a classic WSN example \cite{5411741} under data packet drop and impulsive noise conditions. The network includes 20 sensor nodes (${N=20}$), and the nodes and connections are illustrated in Fig. 1. We consider the state estimation problem with the following state and observation equations:
\begin{equation}
\begin{split}
\left[ {\begin{array}{*{20}{c}}
{x_{k + 1}^{i;1}} \\
{x_{k + 1}^{i;2}} \\
{x_{k + 1}^{i;3}}
\end{array}} \right] = \left[ {\begin{array}{*{20}{c}}
1&{\Delta T}&{\Delta {T^2}/2} \\
0&1&{\Delta T} \\
0&0&1
\end{array}} \right]\left[ {\begin{array}{*{20}{c}}
{x_k^{i;1}} \\
{x_k^{i;2}} \\
{x_k^{i;3}}
\end{array}} \right] + {\boldsymbol{q}}_k^i,
\end{split}
\end{equation}
\begin{equation}
\begin{split}
{\boldsymbol{y}}_k^i = \left[ {\begin{array}{*{20}{c}}
0&1&0
\end{array}} \right]\left[ {\begin{array}{*{20}{c}}
{x_k^{i;1}} \\
{x_k^{i;2}} \\
{x_k^{i;3}}
\end{array}} \right] + {\boldsymbol{v}}_k^i.
\end{split}
\end{equation}
Here, ${\Delta T{\text{ = 0}}{\text{.1}}}$ s represents the measurement time interval, ${{\boldsymbol{x}}_k^i = {[x_k^{i;1}{\text{ }}x_k^{i;2}{\text{ }}x_k^{i;3}]^{\text{T}}}}$ represent the position, velocity, and acceleration of the target, respectively, and ${{\boldsymbol{q}}_k^i}$ and ${{\boldsymbol{v}}_k^i}$ respectively represent mutually uncorrelated process noise and measurement noise (impulsive noise) with the following distributions:
\begin{equation}
\begin{split}
\begin{gathered}
{\boldsymbol{q}}_k^{i;1} \sim 0.9N(0,{\text{ }}0.01) + 0.1N(0,{\text{ }}1), \hfill \\
{\boldsymbol{q}}_k^{i;2} \sim 0.9N(0,{\text{ }}0.01) + 0.1N(0,{\text{ }}1), \hfill \\
{\boldsymbol{q}}_k^{i;3} \sim 0.9N(0,{\text{ }}0.01) + 0.1N(0,{\text{ }}1), \hfill \\
{{\boldsymbol{v}}_k} \sim 0.9N(0,{\text{ }}0.01) + 0.1N(0,{\text{ }}100). \hfill \\
\end{gathered}
\end{split}
\end{equation}
In the simulations, we assume that only the velocity of the target can be observed. In addition, we set ${\varepsilon = {10^{ - 6}}}$, and the initial values of the true state (state of the target), the estimated state, and the covariance matrix are respectively given as follows:
\begin{equation}
\begin{split}
\begin{gathered}
{\boldsymbol{x}}_0^i = {\left[ {\begin{array}{*{20}{c}}
0&0&1
\end{array}} \right]^{\text{T}}}, \hfill \\
{\boldsymbol{\hat x}}_{0|0}^i = {\left[ {\begin{array}{*{20}{c}}
0&0&1
\end{array}} \right]^{\text{T}}} + N(0,{\text{ 0}}{\text{.01}}) \times {\left[ {\begin{array}{*{20}{c}}
1&1&1
\end{array}} \right]^{\text{T}}}, \hfill \\
{\boldsymbol{P}}_{0|0}^i = 0.01 \times \operatorname{diag} \left\{ {\begin{array}{*{20}{c}}
1&1&1
\end{array}} \right\}. \hfill \\
\end{gathered}
\end{split}
\end{equation}
The state estimation performance was measured using the mean square deviation (MSD), which can be defined for node ${i}$ at instant ${k}$ as follows:
\begin{equation}
\begin{split}
{\text{MSD}}_k^i: = E\left\{ {{{\left\| {x_k^i - \hat x_k^i} \right\|}^2}} \right\}.
\end{split}
\end{equation}
The results listed in the tables are the averages of 100 independent trials, and each trial consists of 1000 iterations.
\begin{figure}
\label{Fig1}
\centerline{\includegraphics[width=\columnwidth]{fig1}}
\caption{Network topology of the example WSN with ${N = 20}$ nodes}
\end{figure}
\begin{table}
\centering
\caption{Average number of iterations for every time step of the proposed DMCKF-DPD algorithm with different values of ${\sigma }$}
\begin{tabular*}{8.0cm}{lccc}
\hline
Average number & ${p_k^{i;j} = 0.9}$ & ${p_k^{i;j} = 0.8}$ &${p_k^{i;j} = 0.7}$ \\
of iterations \\
\hline
${\sigma = 0.4}$ & 3.4760 & 3.9280 & 4.2500 \\
${\sigma = 0.6}$ & 2.3850 & 2.6710 & 2.9460 \\
${\sigma = 1.0}$ & 1.8620 & 1.9650 & 2.1730 \\
${\sigma = 4.0}$ & 1.1640 & 1.2140 & 1.2480 \\
${\sigma = 8.0}$ & 1.0690 & 1.0870 & 1.1020 \\
\hline
\end{tabular*}
\end{table}
\begin{table*}
\centering
\caption{Average MSD values obtained under impulsive noise conditions for the distributed network}
\begin{tabular*}{13.5cm}{lccccccc}
\hline
Node & 16 & 5 & 4 & 2 & 8 & 9 & 7 \\
\hline
Number of neighbors & 1 & 2 & 3 & 4 & 5 & 6 & 7 \\
Stationary DKF algorithm (${\sigma = 2}$) & -2.1291 & -2.2715 & -2.2672 & -2.2812 & -2.2907 & -2.2857 & -2.3007 \\
DMCKF-DPD algorithm (${\sigma = 2}$) & -3.8419 & -3.9370 & -4.0695 & -4.0463 & -4.0989 & -4.2063 & -4.1700 \\
DMCKF-DPD algorithm (${\sigma = 5}$) & -3.8033 & -3.9126 & -4.0546 & -4.0387 & -4.0596 & -4.0722 & -4.1064 \\
DMCKF-DPD algorithm (${\sigma = 10}$) & -3.7800 & -3.9042 & -4.0376 & -4.0358 & -4.0417 & -4.0331 & -4.0926 \\
\hline
\end{tabular*}
\end{table*}
\begin{figure}
\label{fig2}
\centerline{\includegraphics[width=\columnwidth]{fig2}}
\caption{MSD values obtained by the state estimation algorithms under data packet drop and impulsive noise conditions for the distributed network (${\sigma = 2}$).}
\end{figure}
The MSD values obtained by the two different algorithms with ${\sigma = 2}$ are presented in Fig. 2 as a function of the number of iterations. Here, the labels Stationary DKF (0.8) and Stationary DKF (0.7) correspond to the stationary DKF algorithm with parameters ${p_k^{i;j} = 0.8}$ and ${p_k^{i;j} = 0.7}$, respectively. Similarly, the labels DMCKF-DPD (0.8) and DMCKF-DPD (0.7) correspond to the proposed algorithm with parameters ${p_k^{i;j} = 0.8}$ and ${p_k^{i;j} = 0.7}$, respectively. It is obvious that the DMCKF-DPD algorithm outperformed the conventional stationary DKF algorithm by about 1.5 dB when considering data packet drops and impulsive noise.
The average number of iterations required by the proposed DMCKF-DPD algorithm for convergence to an optimal solution at every time step are listed in Table 2 for different values of ${\sigma }$. We note that the convergence to the optimal solution is very fast because the initial value of the fixed-point iterative algorithm is set as (${{({\boldsymbol{\hat x}}_{k|k}^i)_0} = {\boldsymbol{\hat x}}_{k|k - 1}^i}$). It is also obvious that the number of iterations required by the fixed-point iterative algorithm decreases with increasing ${\sigma }$, and the convergence rate of the DMCKF-DPD algorithm correspondingly increases.
The average MSD values obtained at select nodes of the distributed system by the DMCKF-DPD and stationary DKF algorithms under impulsive noise conditions (${p_k^{i;j} = 0.8}$) are listed in Table 3. The proposed algorithm demonstrates better state estimation performance than the stationary DKF algorithm at all kernel bandwidths considered. We also note that the performance of the proposed algorithm is increasingly superior to that of the conventional algorithm as the number of neighbors for a given node in the network increases.
\section{Conclusion}\label{section:Conclusion}
This paper presented a novel distributed Kalman filter algorithm for state estimation in WSNs that accounts for data packet drops under non-Gaussian noise conditions. The approach requires each node in the WSN to recursively compute local state estimates in a collaborative manner with its neighboring nodes. The computational complexity of the DMCKF-DPD algorithm was demonstrated to be moderate compared to that of the conventional stationary DKF. In addition, a sufficient condition to ensure the convergence of the fixed-point iterative algorithm was presented. Simulations conducted with a 20-node WSN clearly demonstrated that the DMCKF-DPD algorithm provides significantly more accurate state estimation performance than the stationary DKF, particularly under non-Gaussian noise conditions. As future work, we intend to investigate the state estimation performance of a distributed Kalman filter based on the minimum error entropy (MEE) criterion for WSNs under data packet drop conditions.
|
\section{Introduction}
In-memory key-value stores (IMKVSes) are widely used in real-world applications, especially online services (e.g., e-commerce and social network), because of their ultra-fast query processing speed. For example, Memcached~\cite{memcached}, Redis~\cite{redis}, KeyDB~\cite{keydb} and their variants~\cite{chandramouli2018faster, fan2013memc3, harris2010distributed, bailey2013exploring} have been deployed in production environments of big internet companies such as Facebook, Amazon and Twitter. As all data resides in memory, the data persistent function is a key feature of IMKVSes for data backup.
A common data persistent approach is to take a point-in-time snapshot of the in-memory data with the system call {\it fork} and dump the snapshot into the file system. Figure~\ref{fig:relatedwork}(a) gives an example of the fork-based snapshot method. In the beginning, the storage engine (the parent process) invokes \emph{fork} to create a child process. As \emph{fork} creates a new process by duplicating the parent process, the child process will hold the same data as the parent. Thus, we can ask the child process to write the data into a file in the background but keep the parent process continuous processing queries. Although the storage engine delegates the heavy IO task to the child process, the fork-based snapshot method can incur latency spikes~\cite{trivago, li2018consistent}. Specifically, queries arriving during the period of taking a snapshot (from the start of \emph{fork} to the end of persisting data) can have a long latency because the storage engine runs into the kernel mode and is out-of-service for queries. For example, the query arrives in time $T_0$ in Figure \ref{fig:relatedwork}(a). For brevity, queries arriving during the period of taking a snapshot are called \emph{snapshot queries}, while the others are called \emph{normal queries}. In general, IMKVSes take a snapshot periodically (e.g., Redis by default takes a snapshot every 60 seconds if at least 10000 records are modified~\cite{redis-default-rdb, zhao2021demand}). Consequently, latency spikes for snapshot queries are not rare.
\begin{figure*}[!t]
\centering
\includegraphics[width=2.0\columnwidth]{imgs/relatedwork.pdf}
\caption{The workflow of the parent and child process with (a) default fork, (b) shared page table (SPT)-based fork, (c) the proposed Async-fork in the snapshot procedure.}
\label{fig:relatedwork}
\end{figure*}
To solve the problem, researchers proposed a variety of snapshot algorithms such as Copy-on-Update~\cite{liedes2006siren, cao2013fault}, Zigzag~\cite{cao2011fast} and Ping-Pong~\cite{cao2011fast}. These algorithms focus on lowering the cost of taking a snapshot by optimizing the occasion of copying data and reducing the amount of data copied. However, a recent study~\cite{li2018consistent} finds that the performance of the fork-based method is generally competitive with these advanced methods, and even outperforms them for write-intensive workloads. Additionally, the implementation of the fork-based method is very simple and requires a small engineering effort. As such, industrial IMKVSes do not adopt these snapshot algorithms from academia and keep the fork-based as the built-in data persistent approach. As a result, IMKVSes still have latency spikes caused by the fork operation, which can harm the service quality of online applications.
In contrast to existing research focusing on optimizing snapshot algorithms, we propose a straightforward approach to address the problem. Specifically, we optimize the fork operation to reduce the long latency incurred by the fork-based snapshot from the operating system (OS) level, while keeping the data persistent mechanism in IMKVSes unchanged.
We first conduct an in-depth study to reveal the impact of the fork operation on query latency. Our profiling results on the default fork operation show that the overhead of \emph{fork} results in a long tail latency (up to hundreds of milliseconds) for snapshot queries, and copying the page table dominates the execution time of \emph{fork}. This motivates us to further investigate the impact of advanced techniques~\cite{hugepage,THP, mccracken2003sharing, dong2016shared, zhao2021demand} that can accelerate the fork operation. We consider two optimization strategies, the huge page~\cite{hugepage,THP} and the shared page table~\cite{mccracken2003sharing, dong2016shared, zhao2021demand}. As the huge page can degrade the performance of IMKVSes~\cite{redisTroubles, keyDBTrobules}, our profiling only involves the shared page table-based {\it fork} (SPT-based {\it fork})~\cite{zhao2021demand}, which is the latest method. The SPT-based {\it fork} proposed to share the page table between the parent process and the child process in a copy-on-write (CoW) manner to reduce the cost of the fork operation. However, the CoW can frequently interrupt the parent process as shown in Figure~\ref{fig:relatedwork}(b). Our profiling results show that frequent interruption can incur non-negligible overhead for snapshot queries although the SPT-based {\it fork} significantly reduces the tail latency compared with the default fork operation. Moreover, our analysis finds that the shared page table introduces the data leakage vulnerability, which potentially leads to an inconsistent snapshot. Thus, both the huge table and shared page table techniques cannot be applied in this scenario.
Motivated by these findings, we propose Async-fork to mitigate the latency spikes for snapshot queries by optimizing the fork operation. Figure \ref{fig:relatedwork}(c) demonstrates the general idea. As copying the page table dominates the cost of the fork operation, Async-fork offloads this workload from the parent process to the child process to reduce the duration that the parent process runs into the kernel mode. This design also ensures that both the parent and child processes have an exclusive page table to avoid the data leakage vulnerability caused by the shared page table.
However, it is far from trivial to achieve in design since the asynchronization operations of the two processes on the page table can result in an inconsistent snapshot, i.e., the parent process may modify the page table, while the copy operation of the child process is in process. To address the problem, we design the proactive synchronization technique. This technique enables the parent process to detect all modifications (including that triggered by either users or OS) to the page table. If the parent process detects that some page table entries will be modified and these entries are not copied, then it will proactively copy them to the child process. Otherwise, these entries must have been copied to the child process and the parent process will directly modify them. In this way, the proactive synchronization technique keeps the snapshot consistent and reduces the number of interruptions to the parent process compared with SPT-based {\it fork}. Additionally, we parallelize the copy operation of the child process to further accelerate Async-fork.
We implement Async-fork in the Linux kernel (both x86 and ARM64).
Async-fork is integrated into the OS and transparent to IMKVSes.
The technique is also deployed in the online Redis databases in public Clouds\footnote{\url{https://www.alibabacloud.com/product/apsaradb-for-redis}. Last accessed on 2022/11/13.}. Despite that, we conduct experiments on our local machine for the purpose of test flexibility. In the experiments, we select two popular IMKVSes, Redis and KeyDB, and use the Redis benchmark~\cite{redis-bench} and Memtier benchmark~\cite{memtier}. The database instance size is varied from 1GB, 2GB, and 4GB … to 64GB. Although the SPT-based {\it fork}~\cite{zhao2021demand} may lead to data leakage, our experiment involves this method for comparison purposes because of its efficiency. For Redis, our experiment results show that 1) compared with the default {\it fork}, Async-fork reduces the 99\%-ile latency of snapshot queries by 17.57\% (from 0.074ms to 0.061ms) on 1GB instance, 81.76\% (from 0.435ms to 0.079ms) on 8GB instance and 99.84\% (from 991.9ms to 1.5ms) on 64GB instance; and 2) compared with the latest SPT-based {\it fork}, Async-fork reduces the 99\%-ile latency of snapshot queries by 2.87\% on 1GB instance, 39.73\% on 8GB instance and 61.97\% on 64GB instance. We obtain similar results on KeyDB. These results demonstrate the efficacy of the technique proposed in this paper, especially for the large instances that can lead to long latency.
In summary, we make the following contributions in this paper.
\begin{itemize}
\item We conduct an in-depth study of the impact of the fork operations on the latency of snapshot queries in IMKVSes.
\item We propose Async-fork that can mitigate the long latency of snapshot queries from the OS level, which is orthogonal to existing research on the problem.
\item The technique is implemented in the Linux kernel (both x86 and ARM64) and deployed in the online Redis database in public clouds.
\item We conduct extensive experiments with Redis and KeyDB to evaluate the efficacy of the proposed techniques.
\end{itemize}
\section{Background} \label{sec:background}
In this section, we first introduce the preliminaries and then discuss the related work.
\subsection{Preliminary}\label{sec:os_basics}
We first briefly review two operating system concepts, \emph{virtual memory} and \emph{fork} that are closely related to this work. As our technique is implemented and deployed in Linux, we introduce these concepts in the context of Linux. Then, we discuss the use cases of \emph{fork} in databases.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{imgs/background_new.pdf}
\caption{\sun{The organization of the page table.}}
\label{fig:page_table}
\end{figure}
\textbf{Virtual Memory.} As an effective approach to managing hardware memory resources, virtual memory is widely used in modern operating systems. A process has its own virtual memory space, which is organized into a set of virtual memory areas (VMA). Each VMA describes a continuous area in the virtual memory space. The page table is the data structure used to map the virtual memory space to the physical memory. It consists of a collection of page table entries (PTE), each of which maintains the virtual-to-physical address translation information and access permissions. A VMA corresponds to multiple PTEs.
\sun{Figure \ref{fig:page_table} shows an example of the page table.
To reduce the memory cost, the page table is stored as a multi-level radix tree in which PTEs locate in leaf nodes (i.e., PTE Level in Figure~\ref{fig:page_table}). The part in the area marked with the dashed line corresponds to VMA$_0$. The tree at most has five levels. From top to bottom, they are the page global directory (PGD) level, the P4D level, the page upper directory (PUD) level, the page middle directory (PMD) level and the PTE level. As P4D is generally disabled, we focus on the other four levels in this paper. Except for the PTE level, an entry stores the physical address of a page while this page is used as the next-level node (table). With the page size setting to 4KB, a table in each level contains 512 entries. Given a VMA, “VMA’s PTEs” refers to the PTEs corresponding to the VMA and “VMA’s PMDs” is the set of PMD entries that are parents of these PTEs in the tree. For example,
PMD$_0$ and PMD$_1$ belong to VMA$_0$'s PMDs in Figure \ref{fig:page_table}.}
\textbf{Fork Operation.} \emph{Fork} is a system call that creates a new process by duplicating the calling process~\cite{fork2}. Both processes have separate memory spaces. The new (resp. calling) process is called the child (resp. parent) process. To accelerate the operation, Linux implements \emph{fork} with the copy-on-write (CoW) strategy. Specifically, while invoking \emph{fork}, the parent process runs in the kernel mode and copies the metadata (e.g, VMAs, the page table, file descriptors, and signals) to the child process. The PTEs of both parent and child are set to write-protected. After that, the process that first modifies a write-protected page triggers a page fault, which leads to the copy of the page. In a word, benefiting from CoW, \emph{fork} copies the metadata only.
\sun{\textbf{Use Cases of \emph{fork} in Database}. \emph{Fork} has a number of database use cases because it can easily and efficiently create a snapshot of in-memory data, the consistency of which is guaranteed by the OS. In general, these cases can be categorized into two classes based on the usage of the snapshot. First, use \emph{fork} to
delegate dedicated tasks, which have expensive IO or computation costs,
to a child process without blocking the service of the parent process.
MDC~\cite{park2020memory} uses \emph{fork} to record checkpoints for
in-memory databases. Redis uses \emph{fork} to conduct \emph{log rewriting}~\cite{redis-persist} that optimizes the Append Only File (AOF). FlurryDB~\cite{mior2011flurrydb} proposes to create replica based on \emph{fork} in distributed environments. Second,
use \emph{fork} to create snapshots to support concurrent transaction processing because
\emph{fork} provides the snapshot isolation between processes. HyPer~\cite{kemper2011hyper} proposes to evaluate hybrid OLTP and OLAP queries based on snapshots created by \emph{fork}. AnKer~\cite{sharma2018} designs a fine-grained snapshot mechanism to support MVCC. In particular, AnKer takes a partial snapshot of in-memory data by co-designing the database engine and the system call {\it fork}. Different from our research on accelerating the fork operation from the OS level, AnKer focuses on optimizing which in-memory data should be captured by the snapshot.}
\sun{All these use cases can potentially benefit from Async-fork because 1) they can encounter the query latency spike problem incurred by the fork operation; and 2) Async-fork can accelerate the snapshot creation. This paper focuses on the
scenario that uses \emph{fork} to take a point-in-time snapshot of in-memory data to persist the data~\cite{redis-persist, keydb-persist}. In particular, the storage engine calls \emph{fork} to create a child process that holds the same data as it. Then, the child process writes the data to the hard drive, while the storage engine can continue to serve users’ queries. Although the storage engine delegates the data dump task to the child process, it will be out-of-service for queries during the invocation of \emph{fork} because it runs into the kernel mode. We are particularly interested in
this scenario because the IMKVS is one of the most important services in public cloud and popular IMKVSes (e.g., Redis and KeyDB)~\cite{db-engine}
use this mechanism to persist data. Consequently, these stores encounter serious a query
latency spike problem (see Section \ref{sec:moti}), while they
are generally used in mission-critical applications that have a rigid latency constraint.} \pp{We also evaluate the effectiveness of Async-fork on log rewriting in Redis~\cite{redis-persist}. The experiment results are presented in the technical report~\cite{techreport}.}
\sun{\emph{Remarks.} Instead of developing a general-purpose solution to replace the default {\it fork} in the OS, the goal of Async-fork is to provide an efficient fork operation for the scenarios (especially for IMKVSes) where 1) the applications are memory-intensive, and 2) the parent process is latency-sensitive. Our design allows Async-fork and the default {\it fork} to run in parallel in the OS. Users can easily choose the fork method used in applications (see Section \ref{sec:implementation_details}).}
\subsection{Related Work}\label{sec:related_work}
Consistent snapshot is essential for in-memory databases to support backup and disaster recovery~\cite{zhang2015memory, li2018consistent, diaconu2013hekaton, redis-persistence-demystified}. Some consistent snapshot mechanisms have been proposed to trade off throughput, latency, and memory footprint~\cite{bronevetsky2006recent, cao2013fault, cao2011fast, liedes2006siren, li2018consistent}. Naive snapshot~\cite{bronevetsky2006recent} blocked the storage engine until a deep copy of all the in-memory data is created, which is not suitable for IMKVSes in which the latency is critical. There are also some non-blocking snapshot mechanisms. Copy-on-Update~\cite{liedes2006siren, cao2013fault} proposed to create a shadow copy of the in-memory data; the storage engine is free to access any data but create a deep copy when updates it for the first time.
Note that, the fork-based snapshot is a Copy-on-Update variant that leverages the operating system. Some other mechanisms used multi-version concurrency control (MVCC)~\cite{bernstein1987concurrency} to keep multiple versions of in-memory data. Zigzag~\cite{cao2011fast} maintained another untouched copy of the in-memory data and introduced metadata bits to indicate which copy the store engine should read from or write to. Based on Zigzag, Ping-Pong~\cite{cao2011fast} maintained three versions of the data to lower the cost of managing metadata bits. Hourglass and Piggyback~\cite{li2018consistent} were developed by combining Zigzag and Ping-Pong.
Although fork-based snapshot results in long latency during snapshot process, popular industrial IMKVSes (Redis and KeyDB) still adopts fork-based snapshot for two reasons: 1) {\it fork} provides a simple engineering implementation for consistent snapshot, while it requires great efforts to integrate the above approaches into the IMKVS. 2) None of the above approaches completely outperform the fork-based snapshot in write-intensive workloads~\cite{li2018consistent}. For example, Ping-Pong and Hourglass can mitigate the latency spikes, while Ping-Pong incurs 3x memory footprint and Hourglass results in higher latency during normal operation. This work resolves the latency spikes of the fork-based snapshot, while keeping its original superiorities.
Previous work~\cite{liedes2006siren, park2020memory} noted that the memory footprint increases during snapshot process due to the CoW strategy. MDC~\cite{park2020memory} proposed to release the pages that have been persisted as soon as possible. AnKer~\cite{sharma2018} introduced a fine-grained version of {\it fork} to take partial snapshot when databases do not need to persist all the data. \sun{CCoW~\cite{electronics11030461} optimized the CoW mechanism based on the spatial locality of memory access. It prioritizes the copy for high-locality memory regions to improve the performance on write-intensive workloads.} Async-fork is orthogonal and complementary to them.
There is also other approach to persist data in IMKVSes. Redis and KeyDB use Append Only File (AOF)~\cite{redis-persist, keydb-persist} to log every write operation received by the storage engine, that will be played again after the database reboots to reconstruct the original dataset. The snapshot and AOF are complementary, and it is recommended to enable both of them simultaneously in IMKVSes.
\section{Motivation}\label{sec:moti}
In this section, we present our profiling results to demonstrate the impact of the fork operation on the query latency. We first evaluate the performance of {\it fork} for taking the snapshot to pinpoint the key performance factors. We then reveal the impact of the fork operation on query latency. In addition to the default {\it fork} in Linux, our profiling involves the state-of-the-art approach~\cite{zhao2021demand} of optimizing the fork operation. Lastly, we summarize our findings according to the profiling results.
\textbf{Profiling Setting.} We use Redis benchmark~\cite{redis-bench} to study the performance in the experiments. The detailed configuration of the test machine is introduced in Section~\ref{sec:experimental-set}.
The experiment reports the latency of a query, that is the elapsed time between the time point that the client issues the query and that the client receives the response.
In the experiment, we enhance the benchmark by generating queries in the open-loop mode~\cite{schroeder2006open, zhang2016treadmill}. This enhancement sends commands to the server without waiting for replies to previous queries to simulate real-world environments. The database instance size is varied from 1GB to 64GB.
By default, Redis takes one snapshot per 60 seconds if at least 10000 keys changed. In order to measure the impact of the fork operation accurately, we execute the BGSAVE command to trigger the operation of taking a snapshot. We classify queries into two groups, \emph{normal} and \emph{snapshot}, based on their arrival time.
The snapshot queries are the queries arriving during the period of taking the snapshot (i.e., from the invocation of {\it fork} until the end of persisting the in-memory data), while the others are normal queries. We measure the 99\%-ile (p99) latency and the maximum latency of normal and snapshot queries, respectively. The two latencies greatly impact user experiences and are often used to measure the performance of user-facing databases~\cite{kasture2016, li2018consistent}. We repeat each experiment five times and report the average value.
\subsection{Performance of \emph{fork} for Taking a Snapshot}\label{sec:fork_time}
\begin{figure}
\centering
\includegraphics[width=.9\columnwidth]{imgs/characterization_fork_execution_time.pdf}
\caption{The execution time of {\it fork} and the percentage of the time of copying the page table in {\it fork.}}
\label{fig:characterization_fork}
\end{figure}
Figure~\ref{fig:characterization_fork} presents the execution time of \emph{fork} and the time of copying the page table in the fork operation. We can see that the execution time grows roughly linearly with the instance size increasing from 1GB to 64GB. The 1GB instance takes less than 10 ms, while the 64GB instance takes more than 600 ms. We can also see that the time of the page table copy dominates the execution time. Particularly, the copy operation takes over 97\% percentage of the execution time on all test cases and up to 99.93\% percentage on the 64 GB instance. Without loss of generality, we measure the detailed metrics of copying the page table on the 8 GB instance to further study the copy operation.
In the experiment, the page table of the 8GB instance has one PGD entry, eight PUDs, $2 ^ {12}$ PMDs and $2 ^ {21}$ PTEs. Overall, \emph{fork} copies the table level-by-level from top to bottom along the radix tree. The copy of one PGD/PUD/PMD entry requires to apply for a page to store its children and initialize the page. The operation takes around 500 ns. Thus, the copy of the $ 2 ^ {12}$ PMDs takes around 2 ms, while the overhead of copying PGDs and PUDs is trivial because there are only a few entries. The rest time (around 70 ms) is spent on copying PTEs. Based on the results, we have the following observation.
\emph{\textbf{Observation 1.} For the fork operation, the execution time dramatically grows with the instance size increasing, and the page table copy dominates the cost. For the page table copy, the overhead of copying PGDs and PUDs is trivial, while that of copying PMDs and PTEs is non-negligible.}
\subsection{Impact of Fork Operations on Latency}\label{sec:problem}
We reveal the impact of the fork operation on the query latency in this subsection. Before presenting the results, we first introduce existing optimization approaches to the fork operation.
\emph{Huge Page}~\cite{hugepage,THP}. In the operating system (OS), we can increase the page size to reduce the number of pages used by a process, for example, setting the page size to 2MB instead of 4KB. The large page size can reduce the number of PTEs and accelerate the fork operation. \sun{However, previous study found that the page fault latency can increase from 3.6$\mu s$ to 378$\mu s$ after enabling huge page because compacting and zeroing memory in the page fault incurs higher overhead on huge page than that on regular page~\cite{kwon2016coordinated}. Moreover, Redis consumes much more memory space after enabling the huge page because applications do not always fully utilize the big page (e.g., in the experiment of~\cite{kwon2016coordinated}, the memory consumption of Redis increased from 12.2GB to 20.7GB). Additionally, allocating huge page can lead to many fragments in the physical memory because it requires consecutive physical memory areas to build a huge page. Consequently, the kernel needs to perform a heavy defragment operation which leads to high CPU utilization~\cite{panwar2018making}. The experiment of~\cite{panwar2018making} shows that the benchmark {\it milc} in SPEC CPU 2006 spends 343 seconds (37\% of its overall execution time) in kernel mode to perform the defragment operation when the memory is highly fragmented.}
\begin{table}[t!]
\caption{Migrating a page from ``X'' to ``Y'' when the parent and child process share the page table.\label{tab:TLB}}
\begin{tabular}{|c|l|l|l|}
\hline
\textbf{Step} & \multicolumn{1}{c|}{\textbf{Operation}} & \multicolumn{1}{c|}{\textbf{Parent(P)}} & \multicolumn{1}{c|}{\textbf{Child(C)}} \\ \hline
1 & Initial state & \begin{tabular}[c]{@{}l@{}}TLB: V $\rightarrow$ X\\ PTE: V $\rightarrow$ X\end{tabular} & \begin{tabular}[c]{@{}l@{}}TLB: V $\rightarrow$ X\\ PTE: V $\rightarrow$ X\end{tabular} \\ \hline
2 & \begin{tabular}[c]{@{}l@{}}P: Set PTE $\rightarrow$\\ None present\end{tabular} & \begin{tabular}[c]{@{}l@{}}TLB: V $\rightarrow$ X\\ PTE: V $\rightarrow$ N\end{tabular} & \begin{tabular}[c]{@{}l@{}}TLB: V $\rightarrow$ X\\ PTE: V $\rightarrow$ N\end{tabular} \\ \hline
3 & P: Flush TLB & \begin{tabular}[c]{@{}l@{}}TLB: N/A\\ PTE: V $\rightarrow$ N\end{tabular} & \begin{tabular}[c]{@{}l@{}}TLB: V $\rightarrow$ X\\ PTE: V $\rightarrow$ N\end{tabular} \\ \hline
4 & \begin{tabular}[c]{@{}l@{}}C: Skipped \\ because N!=X\end{tabular} & \begin{tabular}[c]{@{}l@{}}TLB: N/A\\ PTE: V $\rightarrow$ N\end{tabular} & \begin{tabular}[c]{@{}l@{}}TLB: V $\rightarrow$ X\\ PTE: V $\rightarrow$ N\end{tabular} \\ \hline
5 & P: Update PTE & \begin{tabular}[c]{@{}l@{}}TLB: N/A\\ PTE: V $\rightarrow$ Y\end{tabular} & \begin{tabular}[c]{@{}l@{}}\textcolor{black}{TLB: V $\rightarrow$ X}\\ \textcolor{black}{PTE: V $\rightarrow$ Y}\end{tabular} \\ \hline
6 & P\&C: Access V & \begin{tabular}[c]{@{}l@{}}TLB: V $\rightarrow$ Y\\ PTE: V $\rightarrow$ Y\end{tabular} & \begin{tabular}[c]{@{}l@{}}\textcolor{black}{TLB: V $\rightarrow$ X}\\ \textcolor{black}{PTE: V $\rightarrow$ Y}\end{tabular} \\ \hline
\end{tabular}
\end{table}
\sun{Consequently, huge page is recommended to be disabled in many databases (e.g., Couchbase, MongoDB, KeyDB and Redis~\cite{couchbase, mongodb, keyDBTrobules, redisTroubles}). In particular, two No-SQL databases, Couchbase and MongoDB, recommend users to disable the technique because huge page performs poorly with random memory accesses in workloads~\cite{couchbase, mongodb}. KeyDB and Redis, which are two IMKVSes, recommend users to disable huge page because the technique can incur a big latency penalty and big memory usage~\cite{keyDBTrobules, redisTroubles}. Specifically, if we enable the huge page, the parent and child processes share huge pages after calling \emph{fork} to persist on disk. In a busy instance, a few event loops in either of the two processes will cause to target thousands of pages and trigger the copy operation of a large amount of process memory because of the copy-on-write mechanism in OS. Consequently, this leads to a big latency and a big memory usage.} As such, our profiling does not involve this technique.
\emph{Shared Page Table}~\cite{mccracken2003sharing, dong2016shared, zhao2021demand}. This technique proposed to share the page table between parent and child in a copy-on-write (CoW) manner. Specifically, the fork operation returns immediately after copying the metadata except for the page table. The page table will be copied in a CoW manner.
\sun{However, we find that the shared page table design introduces the data leakage problem, the working set size estimation problem and the NUMA problem. First, the inconsistency between the shared page table and the translation lookaside buffer (TLB) of the child process can lead to the data leakage problem. Second, we cannot accurately estimate the working set size, which indicates the memory usage of each process and is important for cloud resource management~\cite{zhang2020ursa}. This is because the usage is calculated based on the states in the page table~\cite{wss}, while the table is shared by multiple processes. Third, the shared page table and the corresponding processes can locate on different NUMA nodes, which increases the TLB miss overhead~\cite{achermann2020mitosis, panwar2021fast}. Moreover, the NUMA balance mechanism cannot work as expected due to the shared page table. Due to space limit, we discuss the working set size problem and the NUMA problem in the technical report~\cite{techreport}. In the following, we use the example in Table \ref{tab:TLB} to demonstrate the data leakage problem, which is the most serious one among the three problems.
We also write a test program\footnote{\pp{\url{https://doi.org/10.5281/zenodo.7189585}, Last accessed on 2022/11/13.}} to reproduce the example in practice.}
TLB is the hardware to accelerate the translation from a virtual address to a physical address by caching recent translation results.
Initially, the virtual address “V” is mapped to the physical address “X” in PTE and the mapping is cached in TLB. Note that PTE is shared between parent and child, whereas the two processes have their own TLB entries. Suppose that the memory management mechanism (e.g., memory compaction~\cite{kcompactd, kcompactd-code}, swap~\cite{swap} and NUMA balance~\cite{copypagerange}) of OS mitigates the page from “X” to a new page frame “Y” in the parent process. Then, OS sets the mapping from “V->X” to “V->N” (None Present) to indicate that the mapping is invalid in Step 2 and flushes the parent’s TLB entry in Step 3.
\sun{For other processes, the OS loops over each of them to check whether its PTEs contain
"V->X"; if so then set the value to "V->N" and flush the TLB; otherwise, skip the process. This works well if each process has a private page table. However, ODF uses the shared page table design. When the OS checks the child process, the PTE has been set to "V->N" in the parent process because the PTE is shared between the parent and child processes and therefore the OS cannot find any PTE with the value "V->X" in the child process. Thus, the OS skips the update to the child in Step 4.} In Step 5, the parent updates PTE to map “V” to “Y”. Although PTE has the correct mapping, the child’s TLB entry is inconsistent with PTE. Consequently, the future access to “V” in the child can lead to a data leakage problem. Despite that, we study the performance of the shared page table-based {\it fork} in our experiments for comparison purposes. \sun{As ODF \cite{zhao2021demand} is the latest work and the only one that is publicly available\footnote{\pp{\url{https://github.com/rssys/on-demand-fork}, Last accessed on 2022/11/13.}} among the methods \cite{mccracken2003sharing, dong2016shared, zhao2021demand} adopting the shared page table design, our experiments focus on ODF.}
\textbf{Experiment Results.} Figures \ref{fig:characterization_p99} and \ref{fig:characterization_max} present the 99\%-ile latency and the maximum latency of normal queries and snapshot queries, respectively. Snapshot-ODF denotes the results of using the On-Demand-Fork~\cite{zhao2021demand} to take the snapshot, which is the latest shared page table-based fork method. Snapshot-DEF denotes the results of using the default {\it fork} in Linux. As shown in the figures, the latency of normal queries slightly increases with the instance size varying from 1GB to 64GB. In contrast, the value of snapshot queries grows sharply. The shared page table technique dramatically reduces the latency of the default {\it fork}, especially for the large instance. For example, on the 64GB instance size, the optimization reduces the 99\%-ile latency from 911.95ms to 3.96ms and the maximum latency from 1204.78ms to 59.28ms. The latency of Snapshot-ODF is higher than that of Normal because the CoW of the page table frequently interrupts the engine process. For example, the engine process is interrupted over 7000 times on the 16GB instance, which leads to frequent out-of-service for queries. According to the analysis of existing optimization methods and the experiment results, we have the following observation.
\begin{figure}[t]
\centering
\includegraphics[width=.9\columnwidth]{imgs/characterization_p99_3bar.pdf}
\caption{The 99\%-ile latencies of normal queries and snapshot queries in Redis.}
\label{fig:characterization_p99}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=.9\columnwidth]{imgs/characterization_max_3bar.pdf}
\caption{The maximum latencies of normal queries and snapshot queries in Redis.}
\label{fig:characterization_max}
\end{figure}
\emph{\textbf{Observation 2.} The fork operation has a significant impact on the latency of snapshot queries, and the tail latency of the default function is up to hundreds of milliseconds. Although the shared page table technique reduces the latency of the default fork operation, the overhead incurred by frequent interruptions is non-negligible and the shared page table introduces the potential data leakage problem.}
\subsection{Summary}
Based on the observations, we find that copying the page table dominates the execution time of the default {\it fork} in Linux, especially, for large instances. The overhead of \emph{fork} results in a long latency (up to hundreds of milliseconds) for queries issued during the invocation of \emph{fork}. Although several optimization methods~\cite{hugepage,THP,mccracken2003sharing, dong2016shared, zhao2021demand} of the fork operation have been proposed, they either lead to poor performance of IMKVSes, or have a data leakage problem, which potentially generates an inconsistent snapshot. Thus, these optimizations cannot be used in the fork-based snapshot mechanism in IMKVSes. As IMKVSes have a rigid requirement on the latency to serve online scenarios, a high-performance {\it fork} is required to reduce the long latency of snapshot queries.
\section{Design of Async-Fork} \label{sec:design}
In this section, we introduce the design of Async-fork, an operating system-based solution that effectively reduces the long latency of snapshot queries without incurring extra vulnerabilities.
\subsection{General Idea}
\begin{figure}[t]
\centering
\includegraphics[width=1\columnwidth]{imgs/inconsistent_new.pdf}
\caption{The challenges in Async-fork.}
\label{fig:inconsistent}
\end{figure}
Figure~\ref{fig:relatedwork}(c) shows the general idea of {\bf Async-fork}.
While the parent process is responsible to copy page tables during snapshot in the default {\it fork}, as shown in the figure, Async-fork offloads the work of copying page table to the idle child process while keeping other steps in the fork unchanged.
In this way, the parent process is able to handle queries while the child process copies the page table from the parent process simultaneously.
However, it is non-trivial to achieve the design of Async-fork, as a snapshot may be inconsistent due to the asynchronous operations on the page table.
The inconsistency happens when the parent process modifies a PTE before the child process has copied it.
Take Figure~\ref{fig:inconsistent} as an example. The IMKVS takes a snapshot at time T$_0$, and the in-memory data is $\{(k_0, v_0), (k_1, v_1)\}$. Suppose a user query that sets a new KV pair $(k_2, v_2)$ arrives at time T$_2$ (\ding{172} in Figure~\ref{fig:inconsistent}), and the child process is copying PTEs from time T$_1$ to T$_3$.
The parent process handles this query,
and PTE$_2$ is modified to point to a new page (Page$_2$) that contains $(k_2, v_2)$.
If the child process has not copied the original PTE$_2$ before the modification,
it would copy the modified PTE$_2$, and owns the new pair $(k_2, v_2)$ in its memory space.
In this way, the data $(k_2, v_2)$ is persisted, and inconsistency happens (the key-value pair $(k_2, v_2)$ does not exist when the snapshot is taken at time $T_0$).
Async-fork resolves the above inconsistency problem by using the parent process to proactively synchronize the modified PTEs to the child process. We explain the detailed steps in Section~\ref{sec:consistency}.
Two main challenges have to be resolved in this solution.
\begin{algorithm}[t!]
\footnotesize
\caption{The framework of Async-fork}\label{algo:async-fork}
\tcp{The parent process}
\For {each VMA$_i$ in (VMA$_0$, VMA$_1$, VMA$_2$...)} {
Copy VMA$_i$ to child process\;
Copy VMA$_i$'s PGDs/PUDs to child process\;
Set all VMA$_i$'s PMDs to be write-protected\;
}
Put the child process on a CPU\ to run it\;
\While {true} {
\If {PTE modification is detected and this PTE's PMD is write-protected} {
Copy PMD to child process\;
Copy 512 PTEs of this PMD to child process\;
Set the PMD to be writeable\;
Set the 512 PTEs to be write-protected\;
}
}
\tcp{The child process}
\For {each VMA$_i$ in (VMA$_0$, VMA$_1$, VMA$_2$...)} {
\For {each PMD$_j$ of VMA$_i$} {
\If {PMD$_j$ is write-protected} {
Copy PMD$_j$ from parent process\;
Copy 512 PTEs of PMD$_j$ from parent process\;
Set the PMD to be writeable\;
Set the 512 PTEs to be write-protected\;
}
}
}
\end{algorithm}
Firstly, it is necessary to detect all the PTE modifications. However, besides of the user operations, many inherent memory management operations in the operating system also cause PTE modifications.
For instance, the OS periodically migrates pages among NUMA nodes~\cite{copypagerange}, causing the involved PTEs to be modified as inaccessible (\ding{173} in Figure~\ref{fig:inconsistent}). We describe the method to detect the PTE modifications in Section~\ref{sec:detect}.
Secondly, errors may occur during Async-fork. For instance, the child process may fail to copy an entry due to out of memory (\ding{174} in Figure~\ref{fig:inconsistent}).
In this case, error handling is necessary, as we should restore the process to the state before it calls Async-fork.
We present how errors are handled in Section~\ref{sec:exception}.
By resolving the above challenges, the time of the parent process used on {\it fork} is greatly reduced, and the latency of the snapshot queries can be reduced in consequence. Moreover, after the child process finishes copying the page table, the two processes have their complete private page table. \sun{Therefore, the Async-fork does not introduce the data leakage vulnerability.}
In the following sections, we use the terminology in x86 Linux to explain our design. We also implement Async-fork on Arm Linux using similar ideas, while similar results are achieved compared with x86. Algorithm~\ref{algo:async-fork} describes how the parent process (Line 1 to 14) and the child process (Line 15 to 24) work in Async-fork.
\subsection{Proactive Synchronization}\label{sec:consistency}
Before introducing the proactive synchronization in detail, we first introduce how we offload the work of copying page table to the child process in Async-fork.
{\bf Copying Page Table Asynchronously.} In the default {\it fork}, the parent process traverses all its VMAs and copies the corresponding parts of the page table to the child process. The page table is copied from top to bottom. In Async-fork, the parent process roughly follows the above process, but only copies PGD, P4D (if exists) and PUD entries to the child process (Lines 1 to 3 in Algorithm~\ref{algo:async-fork}). After that, the child process starts to run, and the parent process returns to the user mode to handle queries (Lines 7 to 14).
The child process then traverses the VMAs, and copies PMD entries and PTEs from the parent process.
Figure~\ref{fig:vmbasics} shows an example of the
asynchronous page table copy.
In the figure, the PGD/PUD entries have been copied by the parent process, but PMD entries and PTEs have not yet (some PMD entries point to ``null''). The child process copies PMD entries and PTEs from the parent process (e.g., PMD$_0$ and its PTEs).
\begin{figure}[t!]
\centering
\includegraphics[width=1\columnwidth]{imgs/vmbasics.pdf}
\caption{An example of copying page table in Async-fork. ``RW=1'' represents writable and ``RW=0'' represents write-protected.}
\label{fig:vmbasics}
\end{figure}
We offload the work of copying PMD entries and PTEs to the child process because the overhead of copying them is non-negligible, as analyzed in Section~\ref{sec:fork_time}.
Meanwhile, we keep that the parent process copies PGD/PUD entries. This is because
the overhead of copying PGD/PUD entries is trivial,
and it is more robust to minimize the change to the Linux kernel.
{\bf Synchronizing Modified PTEs Proactively.} When the child process is responsible for copying PMD entries and PTEs, it is possible that the PTEs are modified by the parent process before they are actually copied. Note that only the parent process is aware of the modifications (the way to detect the PTE modification will be introduced in Section~\ref{sec:detect}).
In general, there are two ways to copy the to-be-modified PTEs for the consistency. 1) the parent process proactively copies the PTEs to the child process; 2) the parent process notifies the child process to copy the PTEs and waits until the copying is finished. As both ways result in the same interruption in the parent process, we choose the former way (Lines 8 to 12 in Algorithm~\ref{algo:async-fork}).
More specifically, when a PTE is modified during snapshot, the parent process copies not only this PTE but also all the other PTEs of a same PTE table (512 PTEs in total), as well as the parent PMD entry to the child process proactively.
For instance, when PTE$_1$ in Figure~\ref{fig:vmbasics} is modified, the parent process proactively copies PMD$_2$, PTE$_0$ and PTE$_1$ to the child process.
We choose to copy the entire PTE table because we can quickly detect a range of PTEs that will be modified, but accurately identifying which one will be modified is expensive in practice.
{\bf Eliminating Unnecessary Synchronizations.} Always letting the parent process copy the modified PTEs is unnecessary, as it is possible that the to-be-modified PTEs have already been copied by the child process.
We identify if the PMD entries and PTEs have been copied by the child process, to avoid unnecessary synchronizations. A flag is required to track this status. \sun{We reuse the R/W flag of the PMD entry to record the status. Since the R/W flag is only used when the PMD entry points to a huge page in the x86 Linux kernel, the flag tracks the status correctly. We reuse the R/W flag because 1) this design can avoid adding new fields to the kernel data structures; and 2) the popular databases (e.g., Redis, KeyDB, MongoDB, and Couchbase) recommend disabling the huge page to improve the performance~\cite{redisTroubles, keyDBTrobules, couchbase, mongodb}. Additionally, it is unnecessary to use Async-fork if applications use the huge page because the applications with huge page do not require PTEs but a small number of PGD/PUD/PMD entries (i.e., the page table is small).}
\sun{An alternative approach is to use an unused bit in the \emph{struct page} as the flag. Specifically, each PMD entry points to a PTE table, while OS maintains a data structure (\emph{struct page}) for each PTE table. The \emph{struct page} has bits that are not used by the current Linux kernel. Previous research uses these bits as flags, for example, ODF uses some bits in \emph{struct page} as a reference counter. However, this approach requires further modification to the kernel to initialize the bit. Therefore, we do not adopt the design using the \emph{struct page}.}
If a PMD entry and its 512 PTEs have not been copied to the child process, the PMD entry will be set as write-protected (e.g., PMD$_1$ in Figure~\ref{fig:vmbasics}). Note that, it does not break the CoW strategy of {\it fork} since it still triggers the page fault when the corresponding page is written on x86~\cite{guide2011intel}. Once the PMD/PTEs have been copied to the child process (e.g., PMD$_0$), the PMD entry is changed to be writable (the PTEs are changed to be write-protected to maintain the CoW strategy). Since both parent and child processes lock the page of the PTE table with {\it trylock\_page()} when they are copying PMD entries and PTEs, they will not copy PTEs pointed by the same PMD entry at the same time.
\subsection{Detecting Modified PTEs}\label{sec:detect}
The operations that modify PTEs in the OS can be divided into two categories: 1) {\it VMA-wide modification.} Some operations act on specific VMAs, including creating, merging, deleting VMAs and so on. The modification of a VMA may also cause the VMA's PTEs be modified. For example, the user sends queries to delete lots of KV pairs. The IMKVS (parent process) then reclaims the corresponding virtual memory space by {\it munmap}. Some VMAs are hence split or deleted while the VMAs' PTEs are deleted as well. A VMA is usually large because the operating system always tends to merge adjacent VMAs. It means that the VMA-wide modification usually causes extensive PTE modifications, while many VMA's PMD entries are involved.
2) {\it PMD-wide modification.} Other operations modify the PTE directly. For example, the page of parent process can be reclaimed by the out of memory (OOM) killer. In this case, one PMD entry is involved. Note that swapping or migrating a 4KB page will change the PTE but the data will not be changed, so we will not handle it. \sun{Due to limited space, we summarize the locations where operations in the OS modify VMAs/PTEs as checkpoints in the technical report~\cite{techreport}.}
We implement the detection by hooking the checkpoints. Once a checkpoint is reached, the parent process checks whether the involved PMD entries and PTEs have been copied (by checking the R/W flag of the PMD entry). For a VMA-wide modification, all the PMD entries of this VMA are checked, while only one PMD entry is checked for a PMD-wide modification. The uncopied PMD entries and PTEs will be copied to the child process before modifying them.
If a VMA is large, the parent process may take a relatively long time to check all PMD/PTE entries by looping over each of them. \sun{We therefore introduce a two-way pointer, which helps the parent process quickly determine whether all entries of a VMA have been copied to the child, to reduce the cost. Each VMA has a two-way pointer, which is initialized by the parent process during the invocation of the Async-fork function. The pointer in the VMA of the parent process (resp. the child process) points to the corresponding VMA of the child process (resp. the parent process). In this way, the two-way pointer maintains a connection between the VMAs of the parent and the child. The connection will be closed after all PMDs/PTEs of the VMA are copied to the child. Specifically, if no VMA-wide modification happens during the copy of PMDs/PTEs of the VMA, then the child closes the connection by setting the pointers in the VMAs of both the parent and child to \emph{null} after the copy operation. Otherwise, the parent will synchronize the modification (i.e., copying the uncopied PMDs/PTEs to the child), and close the connection by setting the pointers to \emph{null} after the copy operation. As both parent and child processes can access the two-way pointers, the pointers are protected by locks to keep the state consistent. When a VMA-wide modification occurs, the parent process can quickly determine whether all PMDs/PTEs of a VMA have been copied to the child by checking the pointer's value, instead of looping over all these PMDs. Besides, the pointer is also used in handling errors (see Section~\ref{sec:exception}).}
\subsection{Handling Errors}\label{sec:exception}
Since copying the page table involves memory allocation, some errors may occur during both the default {\it fork} and Async-fork.
For instance, a process may fail to initialize a new PTE table due to out of memory.
Such error may only happen in the parent process in the default {\it fork} and has a standard way to handle the error.
However, the copying of page table is offloaded to the child process in Async-fork, such error may happen in the child process, and a method is required to handle such errors. Specifically, we should restore the parent process to the state before it calls Async-fork, to ensure that the parent process will not crash in the future.
As Async-fork may modify the R/W flags of the PMD entries of the parent process, we roll back these entries to be writable when errors occur in Async-fork.
Errors may occur 1) when the parent process copies PGD/PUD entries, 2) when the child process copies PMD/PTEs, and 3) during a proactive synchronization.
In the first case, the parent process rolls back all the write-protected PMD entries. In the second case, the child process rolls back all the remaining uncopied PMD entries. After that, we send a signal ({\it SIGKILL}) to the child process. The child process will be killed when it returns to the user mode and receive the signal.
In the third case, the parent process only rolls back the PMD entries of the VMA containing the PMD entry that is being copied. The purpose is to avoid contending for the PMD entry lock with the child process. An error code is then stored into the two-way pointer of the VMA. \sun{Before (and after) copying PMDs/PTEs of a VMA, the child process will check the pointer to see whether there are errors.} If so, then it stops copying PMD/PTE entries immediately and performs the rollback operations that are already described in the second case.
\section{Optimization and Implementation} \label{sec:opt_imp}
In this section, we present the optimization that further improves the performance of Async-fork, and the way to implement Async-fork in Linux.
\subsection{Accelerating Page Table Copy}\label{sec:acceleration}
The parent process is still interrupted when a proactive PTE synchronization is triggered.
\sun{A straightforward way to reduce the cost of the proactive PTE synchronization is to let the child process to first copy the PTEs potentially modified before other PTEs. However, this method is not practical because the data accessed by user queries are relatively random and the child process cannot determine which PTEs will be modified when the parent invokes Async-fork.}
We therefore propose to minimize the number of proactive PTE synchronizations by reducing the duration of copying the page table.
As VMAs are independent, the kernel threads can totally perform the copy in parallel and obtain near-linear speedup. Therefore, the child process may launch multiple kernel threads to copy PMD/PTEs in multiple VMAs in parallel, so that the copy completes faster. It can effectively reduce the number of proactive PTE synchronization because the synchronization only happens during the period that the child process is copying PMD/PTEs. The probability of triggering a proactive PTE synchronization gets lower when the period becomes shorter. The experiment in Section~\ref{sec:detail_EXP} shows the efficiency of this optimization.
Multiple kernel threads consume CPU cycles. These threads periodically check whether they should be preempted and give up CPU resources by calling {\it cond\_resched()}, in order to reduce the interference on other normal processes.
\subsection{Implementation of Async-fork}\label{sec:implementation_details}
\begin{figure}
\centering
\includegraphics[width=1.0\columnwidth]{imgs/acf_implementation_new.pdf}
\caption{\sun{The implementation of Async-fork.}}
\label{fig:acf_implementation}
\end{figure}
\sun{Figure~\ref{fig:acf_implementation} shows the implementation of Async-fork with a modular design. We encapsulate the code ingested to the kernel into
a hook function, which is instantiated in a kernel module. We can insert/remove the module in/from the kernel as necessary. The hook function is inserted to the call path of default {\it fork} and memory subsystem in the Linux kernel (version 4.19). Users can determine the usage of the default \emph{fork} and Async-fork with the parameter $F$.}
\sun{The hook function is enhanced from {\it copy\_page\_range()}, which accepts a flag (Fast or Slow) to control the copying of the page table. }
With ``Fast'' flag, the function copies PGD/PUD entries and sets PMD entries to be write-protected.
With ``Slow'' flag, the function copies the write-protected PMD entries and PTEs; when the copying finishes, it sets the PMD entries to be writable and sets the PTEs to be write-protected. When Async-fork is called, the parent process copies the page table to the child process using the function with ``Fast'' flag.
At the end of Async-fork's invocation, the parent process puts the child process into the {\it runqueue} of a CPU and returns to the user mode. Before the child process returns to the user mode, it copies the page table using the function with ``Slow'' flag. The parent process proactively copies the uncopied PMD/PTEs (using the function with ``Fast'' flag) to the child process before modifying them.
{\bf Flexibility.}
For the IMKVS workload that has small memory footprint, the page table copy is already short. In this case, Async-fork brings small benefit. For these workloads, we provide an interface in {\it memory cgroup} to control whether Async-fork is enabled (as well as the number of kernel threads used to speed up copying PMD/PTEs in the child process). \sun{Specifically, when users add a process to a \emph{memory cgroup}, they can pass a parameter to enable/disable Async-fork at run time (i.e., the parameter {\it F} in Figure~\ref{fig:acf_implementation}). As shown in the figure, if the parameter value is 0, then the process will use the default \emph{fork}. Otherwise, Async-fork is enabled. As such, users can determine which fork operation the process uses as necessary and use Async-fork without any modification in the source code of applications. The process will use the default {\it fork} if no parameter is passed in.}
{\bf Memory overhead.} The only memory overhead of Async-fork comes from the added pointer (8B) in each VMA of a process. In the case, the memory overhead in a process is the number of VMAs times $8B$. Considering a machine with 512GB main memory while 400 processes run simultaneously, there are roughly 760,000 VMAs according to our statistics. In this case, the memory overhead will be 760000$\times$8B $\approx$ 6MB. This overhead is generally negligible.
{\bf Support for ARM64.} The design of Async-fork can also be implemented on ARM64. Specifically, we use the {\it APTable[1:0]}~\cite{arm64} in the table descriptor of the PMD entry as the R/W flag. Async-fork can also be implemented on other architectures that support hierarchical attributes in the page table.
{\bf Consecutive Snapshots.} It is possible that the parent process starts the next snapshot using Async-fork before the previous child process finishes copying PMD/PTEs, as the parent process returns to the user mode before the child process.
In the current implementation, we will not block the next Async-fork call but keep a VMA's page table be copied by only one child process at any time. When the parent process copies a VMA to the child process, it checks the two-way pointer to identify whether there exists a previous child process copying the page table of this VMA. If exists, it proactively copies the whole page table of this VMA to the previous child process. \sun{Async-fork adopts this design because supporting concurrent fork operations need to track all child processes
that are copying page tables. This is hard because the kernel does not provide the information
and we need to inject new data structures into the kernel to record the states and synchronize these
child processes.}
\sun{As Async-fork cannot support concurrent fork operations in a process (the same
as the default {\it fork} in the OS), Async-fork cannot support the cases where
the parent process needs to conduct the fork operation in an ultra-high
frequency (e.g., in milliseconds). In spite of the limitation, Async-fork works well for database use cases because a storage engine does not create many child processes from the parent process simultaneously in practice. Specifically, conducting the fork operation frequently leads the parent to frequently turn into the kernel mode, which hurts the service quality. Moreover, many processes executing in parallel will degrade the performance due to the resource contention (e.g., IO bandwidth and CPU resources).
Therefore, IMKVSes (e.g., Redis and KeyDB) do not recommend ultra-frequent data snapshots (generally 60 seconds~\cite{redis-default-rdb}) and have no cases requiring the parent process to invoke Async-fork in milliseconds to our knowledge. For HyPer, which uses the fork operation to support concurrent transaction processing, Async-fork can work well. This is because the parent process handles OLTP (a single updater) that has a rigid requirement on the latency, whereas the child process executes OLAP which generally has a long execution time and is more tolerant of the latency than OLTP. Moreover,
HyPer notices the cost of the fork operation and designs a novel mechanism,
which makes multiple OLAP queries to share a snapshot, to improve the performance.}
\section{Evaluation} \label{sec:evalution}
\label{sec:evaluation}
In this section, we evaluate the effectiveness of Async-fork.
\subsection{Experimental Setup} \label{sec:experimental-set}
We evaluate Async-fork on a machine with two Intel Xeon Platinum 8163 processors, each of them has 24 physical cores (48 logical cores).
The machine has 384GB memory, and 1TB NVMe hard drive.
In terms of software setup, the experimental platform runs CentOS 7.9 with Linux 4.19.
Except for the experiments in Section~\ref{sec:ECS}, all the other experiments are conducted within a single machine as described above.
\textbf{Benchmarks.} We use Redis (version 5.0.10)~\cite{redis} and KeyDB (version 6.2.0)~\cite{keydb} compiled with gcc 6.5.1 as the representative IMKVS servers, and use Redis benchmark~\cite{redis-bench} as well as Memtier benchmark~\cite{memtier} to be the workload generators.
The benchmarks reveal the scenario where multiple clients send requests to the IMKVS server simultaneously. Similar to prior work~\cite{chen2019parties}, we enhance Redis benchmark to generate queries in an open-loop mode for measuring the latency accurately~\cite{schroeder2006open, zhang2016treadmill}.
By default, the experiments are conducted with the following settings: 1) 50 clients (default settings) are used in Redis benchmark, while 50 clients are used in Memtier benchmark for consistency.
2) The key range is set to $2\times 10^{8}$, the key size of 8B and the value size of 1024B. 3) Each experiment is repeated by five times and the average results are reported. 4) In Async-fork, the child process launches 7 additional kernel threads (together with the child process itself, there are 8 threads in total) to help it copy PMDs/PTEs faster. Our experiment in Figure~\ref{fig:no-multi-threaded-copy} shows that Async-fork still outperforms the state-of-the-art solution with a single child process.
5) The KeyDB server is configured with 4 threads.
\textbf{Metrics.} We launch a large number of queries ($5\times 10^{6}$) to the IMKVS and record the latency of the queries that arrive in the snapshot process (start from the parent process calls {\it fork} until the child process persists all in-memory data). \sun{Specifically, we measure the 99\%-ile latency because latency-sensitive services generally provide an SLA on some percentile. Moreover, we report the maximum latency because IMKVS is often used for demanding use cases that have a rigid requirement for the worst-case latency~\cite{redis-maxlatency}. For example, the increase of the maximum latency of Redis can lead to read error on connections~\cite{trivago}. Therefore, in the production environment~\cite{tairpmem}, the maximum latency is an important indicator of system stability. Due to limited space, we report the experiment results of query processing
throughput and total out-of-service time of the parent process in the technical report~\cite{techreport}. In the report, we also discuss the method of tuning IMKVSes to further improve the performance.}
\textbf{Baselines.} We compare Async-fork with On-Demand-Fork (denoted by {\it ODF} in short)~\cite{zhao2021demand}, the state-of-the-art shared page table-based fork. In ODF, each time a shared PTE is modified by a process, not only one PTE but 512 PTEs located on the same PTE table will be copied at the same time.
We do not report the results of the default {\it fork} in this section since it results in
10X higher latency compared with both Async-fork and ODF in most cases (already presented in Section~\ref{sec:problem}). To accurately measure the performance of the IMKVSes in the experiments, we trigger the snapshot operation manually using the BGSAVE command.
\subsection{Overall Evaluation}\label{sec:exp_write_intensive}
We first evaluate Async-fork using write-intensive workloads that require frequent snapshots for data persistence.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{imgs/write-intensive-p99-2subplot.pdf}
\caption{The 99\%-ile latency of snapshot queries.}
\label{fig:write-p99}
\end{figure}
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{imgs/write-intensive-max-2subplot.pdf}
\caption{The maximum latency of snapshot queries.}
\label{fig:write-max}
\end{figure}
{\bf Latency Results.} In this experiment, we use the Redis benchmark~\cite{redis-bench} to generate the write-intensive workload by issuing {\it SET} queries to the IMKVSes, and configure the clients to send 50,000 such queries in a second.
Since the IMKVS is often used in the context of demanding use cases, there are usually strict requirements on both the 99\%-ile latency and the worst case latency~\cite{redis-maxlatency, keydb-maxlatency}.
Figure~\ref{fig:write-p99} shows the 99\%-ile latencies of snapshot queries in Redis and KeyDB, with ODF and Async-fork.
As observed, Async-fork outperforms ODF in all the cases, and the performance gap increases when the instance size gets larger.
For instance, operating on a 64GB IMKVS instance, the 99\%-ile latency of the snapshot queries is 3.96ms (Redis) and 3.24ms (KeyDB) with ODF, while the 99\%-ile latency reduces to 1.5ms (Redis, 61.9\% reduction) and 1.03ms (KeyDB, 68.3\% reduction) with Async-fork.
Figure~\ref{fig:write-max} shows the maximum latency of the snapshot queries.
As observed, Async-fork greatly reduces the maximum latency of the benchmarks compared with ODF, even if the instance size is small.
For a 1GB IMKVS instance, the maximum latencies of the snapshot queries are 13.93ms (Redis) and 10.24ms (KeyDB) respectively with ODF, while the maximum latencies are decreased to 5.43ms (Redis, 60.97\% reduction) and 5.64ms (KeyDB, 44.95\% reduction) with Async-fork.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{imgs/pmd_new.pdf}
\caption{The frequency of interruptions in the parent process during the snapshot process.}
\label{fig:pmd_copy}
\end{figure}
{\bf Deep Diving.} \sun{The interruption of Async-fork is caused by the proactive synchronization, whereas the interruption of ODF is caused by the CoW of the page table.
We measure the interruptions of the parent process during the snapshot to understand the reason that Async-fork outperforms ODF. Specifically, the parent process turns into kernel mode and is out of service for queries when executing the \emph{copy\_pmd\_range()} function. The overhead of \emph{copy\_pmd\_range()} dominates the cost of executing in the kernel mode.
In order to examine the number of interruptions and the out-of-service time, we use the \emph{bcc} tool~\cite{bcc} to count the number of \emph{copy\_pmd\_rage()} invocations and measure the execution time of each invocation. The result of \emph{bcc} is a histogram in which the bucket is the time duration and the frequency is the number of invocations whose execution time falls into the bucket. The categories [16$us$, 31$us$] and [32$us$, 63$us$] are two default buckets of \emph{bcc}. In our experiments, all invocations fall into the two buckets.}
\sun{Figure~\ref{fig:pmd_copy} shows the frequency of the interruptions within [16us, 31us] and [32us, 63us]. We can see that Async-fork significantly reduces
the frequency of interruptions. For example, Async-fork reduces the frequency of interruptions from 7348 to 446 on the 16GB instance.} Async-fork greatly reduces the interruptions because
the interruptions happen only when the child process is copying PMD/PTEs (the required time is within 600ms as in Figure~\ref{fig:multi-threaded-copy}).
However, the interruption can happen until all data is persisted by the child process in ODF, while the data persistence operation requires tens of seconds (e.g., persisting 8GB in-memory data takes about 40s).
Under the same workload, the parent process is more vulnerable to interruption when using ODF.
\subsection{Detailed Evaluation}{\label{sec:detail_EXP}}
{\bf Sensitivities to the Read/Write Patterns.}
Figure~\ref{fig:rw-mix-max} shows the results using four workloads with different read-write patterns generated with Memtier~\cite{memtier}.
In the figure, ``1:1 (Uni.)'' represents the workload with 1:1 Set:Get Ratio and the uniform random access pattern, while ``1:10 (Gau.)'' is the workload with 1:10 Set:Get ratio and the Gaussian distribution access pattern.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{imgs/rw-mix-max-2subplot.pdf}
\caption{The 99\%-ile latency and maximum latency of snapshot queries under different workloads in an 8GB IMKVS.}
\label{fig:rw-mix-max}
\end{figure}
Observed from Figure~\ref{fig:rw-mix-max}, Async-fork still outperforms ODF.
The benefit is smaller for the workload with more {\it GET} queries.
This is because the serving (parent) process only copy a small number of PTEs for the {\it GET}-intensive workloads.
Moreover, the modified memory is smaller in the experiment with the Gaussian distribution compared with the uniform random access pattern.
With random pattern, the key-value pairs in the IMKVS have the same probability of being accessed, and parts of key-value pairs may be accessed repeatedly with the Gaussian Distribution access pattern.
Since shared PTEs are copied only when they are modified for the first time in ODF, the parent process is interrupted fewer times
with the Gaussian distribution access pattern.
In general, Async-fork works better for write-intensive workloads. The larger the modified memory is, the better Async-fork performs. \sun{Integrating Async-fork with CCoW~\cite{electronics11030461} to improve the performance on write-intensive workloads is an interesting research direction because Async-fork can utilize CCoW to copy the PTEs of high-locality memory pages in advance to reduce the number of proactive synchronizations.}
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{imgs/write-intensive-client-2subplot.pdf}
\caption{The 99\%-ile latency and the maximum latency of snapshot queries under different numbers of clients in an 8GB IMKVS.}
\label{fig:write-intensive-client}
\end{figure}
{\bf The Impact of the Number of Clients.}
In this experiment, we change the number of clients in Redis-benchmark while keeping sending 50, 000 {\it SET} queries every second to an 8GB IMKVS server.
Figure~\ref{fig:write-intensive-client} shows the results of 99\%-ile and maximum latency with 10, 50, 100 and 500 clients. As observed, Async-fork outperforms ODF, while the performance gap increases as the number of clients increases.
This is because more requests arrive at the IMKVS at the same time when the number of clients increases.
As a result, more PTEs may be modified at the same time, and the duration of one interruption to the parent process may become longer.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{imgs/multi-thread_p99_max_3bar.pdf}
\caption{The 99\%-ile and maximum latency of snapshot queries when Async-fork uses 1 or 8 threads.
}
\label{fig:no-multi-threaded-copy}
\end{figure}
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{imgs/multi-threaded-copy.pdf}
\caption{(a) The time that the child process takes to copy PMDs/PTEs in Async-fork. (b) The 99\%-ile and maximum latency in an 8GB Redis instance.}
\label{fig:multi-threaded-copy}
\end{figure}
{\bf The Impact of the Number of Threads In the Child Process.}
In Async-fork, multiple kernel threads may be used to copy PMD/PTEs in parallel in the child process.
Figure~\ref{fig:no-multi-threaded-copy} shows the 99\%-ile and maximum latency of snapshot queries under different Redis instance sizes.
Async-fork\#$i$ represents the results of using $i$ threads in total to copy the PMDs/PTEs.
Observed from Figure~\ref{fig:no-multi-threaded-copy},
Async-fork\#1 (the child process itself) still brings shorter latency than ODF.
The maximum latency of the snapshot queries is decreased by 34.3\% on average, compared with ODF.
We can also find that using more threads (Async-fork\#8) can further decrease the maximum latency.
This is because the sooner the child process finishes copying PMDs/PTEs, the lower probability the parent process is interrupted to proactively synchronize PTEs.
In more detail, Figure~\ref{fig:multi-threaded-copy}(a) shows the time that the child process takes to copy PMDs/PTEs with different numbers of kernel threads, while Figure~\ref{fig:multi-threaded-copy}(b) shows the corresponding 99\%-ile and maximum latency in an 8GB Redis instance.
As we can see, launching more kernel threads effectively reduces the time of copying PMDs/PTEs in the child process. The shorter the time is, the lower the latency becomes.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\columnwidth]{imgs/ecs-2subplot.pdf}
\caption{The 99\%-ile latency and maximum latency of snapshot queries with Async-fork in production.}
\label{fig:ecs}
\end{figure}
\subsection{Evaluation in Production Environment}\label{sec:ECS}
Async-fork has been deployed in our production Redis Clouds. We rent a Redis database from our public cloud to evaluate Async-fork in the production environment. The rented Redis server has 16GB memory and 80GB SSD. We also rent another virtual machine with 4 vCPU cores and 16GB memory from the same cloud to run the Redis benchmark as the client. The client settings are the same as previous experiments. The network bandwidth between the IMKVS server and the client is 3Gbps.
We compare Async-fork with the default {\it fork} in this subsection because ODF is not available in our cloud. Figure~\ref{fig:ecs} shows the 99\%-ile latency and maximum latency of the snapshot queries with the default {\it fork} and Async-fork.
As observed, with an 8GB instance, the 99\%-ile latency of the snapshot queries is reduced from 33.29ms to 4.92ms, and the maximum latency is reduced from 169.57ms to 24.63ms. With a 16GB instance, the former is reduced from 155.69ms to 5.02ms, while the latter is reduced from 415.19ms to 40.04ms.
Before Async-fork is deployed on our Redis Clouds,
we received a large number of complaints from tenants about the long latency when taking the snapshot.
Some tenants even have to disable the snapshot function, and they cannot reboot their Redis servers, otherwise the data lost.
After we deploy Async-fork, no more latency complaints have been received.
\section{Conclusion} \label{sec:conclusion}
In this paper, we study the latency spikes incurred by the fork-based snapshot mechanism in IMKVSes and address the problem from the operating system level. In particular, we conduct an in-depth study to reveal the impact of the fork operation on the latency spikes. According to the study, we propose Async-fork.
It optimizes the fork operation by offloading the workload of copying the page table from the parent process to the child process. To guarantee data consistency between the parent and the child, we design the proactive synchronization strategy.
Async-fork is implemented in the Linux kernel and deployed in production environments. Extensive experiment results show that the technique proposed in this paper can significantly reduce the tail latency of queries arriving during the snapshot period.
\section{Issues Incurred by Shared Page Table Design}
\sun{We initially consider using ODF in our cloud to solve the query latency spike problem. However, we find that the shared page table design adopted in ODF~\cite{zhao2021demand} (as well as previous studies~\cite{mccracken2003sharing, dong2016shared}) incurs several issues, which block its usage in the production environment.}
\sun{\textbf{Data Leakage Problem.} As presented in Section \ref{sec:problem}, ODF has the data leakage problem due to the inconsistency between the page table and the TLB. A potential fixing method for the data leakage vulnerability works as follows: 1) for each PTE, record the processes that share it; and 2) if the PTE of a page is set to be invalid in the parent process, then the OS notifies the other processes sharing the PTE to flush the TLB entry. However, it is not easy to record all the processes that share a PTE; this problem is similar to the \emph{reverse mapping} problem, i.e., mapping a physical page back to the PTEs that correspond to it. Although the reverse mapping concept looks simple and easy to implement, it is hard in practice due to the space and access efficiency problem. As introduced in the blogs\footnote{\url{https://lwn.net/Articles/23732}. Last accessed on 2022/11/13.}\vphantom{d}$^{\text{,}}$\footnote{\url{https://lwn.net/Articles/383162}. Last accessed on 2022/11/13.} by Jonathan Corbet, the author of Linux Device Drivers, each reverse mapping method requires complicated data structures and an update mechanism. Unfortunately, for each PTE, ODF records the number of processes that share it, but has no reverse mapping information. To that end, ODF cannot support the fixing method.}
\sun{\textbf{WSS Estimation Problem.} The Linux kernel cannot accurately estimate the working set size (WSS) of a process using ODF, while WSS is an important metric, which cloud providers utilize to improve the resource utilization. Specifically, WSS is the memory size that a process actually requires in run time. A process can populate a memory space that is much bigger than the actual usage (e.g., apply for a memory pool with a fixed size at startup but do not shrink in run time). The Linux kernel estimates WSS of a process by counting how many pages the process has visited (based on the accessed bit of the PTE~\cite{wss}). However, the shared page table design in ODF prevents the kernel from distinguishing the visitor of a page. Consequently, the Linux kernel cannot give an accurate estimation on WSS. In practice, it is very important for public clouds to improve resource utilization because tenants generally would like to apply for more resources than their actual demands (e.g., 68.4\% of memory space is wasted in our clouds~\cite{zhang2020ursa}).}
\sun{\textbf{NUMA Problem.} We find that the NUMA balancing mechanism does not work as expected with ODF. Specifically, when the NUMA balancing mechanism decides to change the page location for the child process, it will set PTEs to be PROT\_NONE (i.e., accessing the pages in the future triggers page faults) to ask the child process to handle the faults. However, both the parent and child processes can capture the faults due to the shared page table design. Additionally, the overhead of TLB miss increases when the processes sharing the page table run on different NUMA nodes because there is only one copy of PTEs, and some processes and the PTEs can locate into different nodes. The overhead of TLB miss can increase due to page table walk on the remote node~\cite{achermann2020mitosis, panwar2021fast}. Although our experiments do not consider NUMA environments, we mention this issue to show the gap between ODF and practice.}
\sun{Fixing these issues is far from trivial and requires significant efforts (e.g., adding new data structures and further modifying the kernel). This motivates us to develop a new fork mechanism that can work in the production environment. Different from the shared page table design, Async-fork has the same design as the default fork operation in which both parent and child processes have a private page table. The parent's PTE table is locked if a PTE (belongs to this table) will be updated in the parent process. Therefore, when the PTE is updated in the parent process, either the PTE has been copied by the child process before the update or not. For the first case, both child and parent processes have their own PTEs whereby Async-fork has no data leakage issue. For the second case, the child process will copy the PTE after the update because of the lock on the parent's PTE table. We use the example in Table~\ref{tab:TLB_FCM} to demonstrate this case. The parent's PTE table is locked during the page migration (Step 2 to Step 4). The PTE can be accessed by the child process after Step 4. Then, the child process will copy the updated PTE. As such, Async-fork does not suffer from the data leakage problem.}
\sun{Nevertheless, the idea that
shares a page table among multiple processes and uses CoW to keep consistent is promising
because it can reduce the workload of copying the page table. It is an interesting
and challenging future research to address all the above issues and develop a fork method based
on the shared page table design.}
\begin{table}[]
\caption{\sun{Migrating a page from ``X'' to ``Y'' before the child process copies the PTE (“V->X”) in Async-fork.\label{tab:TLB_FCM}}}
\begin{tabular}{|c|l|l|l|}
\hline
\textbf{Step} & \multicolumn{1}{c|}{\textbf{Operation}} & \multicolumn{1}{c|}{\textbf{Parent(P)}} & \multicolumn{1}{c|}{\textbf{Child(C)}} \\ \hline
1 & Initial state & \begin{tabular}[c]{@{}l@{}}TLB:V-\textgreater{}X\\ PTE:V-\textgreater{}X\end{tabular} & \begin{tabular}[c]{@{}l@{}}TLB:N/A\\ PTE:N/A\end{tabular} \\ \hline
2 & \begin{tabular}[c]{@{}l@{}}P: Set PTE $\rightarrow$\\ None present\end{tabular} & \begin{tabular}[c]{@{}l@{}}TLB:V-\textgreater{}X\\ PTE:V-\textgreater{}N\end{tabular} & \begin{tabular}[c]{@{}l@{}}TLB:N/A\\ PTE:N/A\end{tabular} \\ \hline
3 & P: Flush TLB & \begin{tabular}[c]{@{}l@{}}TLB:N/A\\ PTE:V-\textgreater{}N\end{tabular} & \begin{tabular}[c]{@{}l@{}}TLB:N/A\\ PTE:N/A\end{tabular} \\ \hline
4 & P: Update PTE & \begin{tabular}[c]{@{}l@{}}TLB:N/A\\ PTE:V-\textgreater{}Y\end{tabular} & \begin{tabular}[c]{@{}l@{}}TLB:N/A\\ PTE:N/A\end{tabular} \\ \hline
5 & C: Copy PTE & \begin{tabular}[c]{@{}l@{}}TLB:N/A\\ PTE:V-\textgreater{}Y\end{tabular} & \begin{tabular}[c]{@{}l@{}}TLB:N/A\\ PTE:V-\textgreater{}Y\end{tabular} \\ \hline
6 & P\&C: Access V & \begin{tabular}[c]{@{}l@{}}TLB:V-\textgreater{}Y\\ PTE:V-\textgreater{}Y\end{tabular} & \begin{tabular}[c]{@{}l@{}}TLB:V-\textgreater{}Y\\ PTE:V-\textgreater{}Y\end{tabular} \\ \hline
\end{tabular}
\end{table}
\begin{table*}[t!]
\centering
\caption{Operations that may modify VMAs/PTEs.\label{tab:modify}}
\footnotesize
\begin{tabular}{|c|c|c|l|c|c|c|}
\hline
\textbf{VMA-wide} & \textbf{Description} & \textbf{Checkpoint} & \multicolumn{1}{c|}{\textbf{}} & \textbf{PMD-wide} & \textbf{Description} & \textbf{Checkpoint} \\ \cline{1-3} \cline{5-7}
\textit{\begin{tabular}[c]{@{}c@{}}mmap(2),munmap(2),\\ madvise(2),mprotect(2),\\ mlock(2),munlock(2),\\ mremap(2)\end{tabular}} & \begin{tabular}[c]{@{}c@{}}User uses system calls\\ proactively, casuing\\ the properties of VMAs to\\ be modified or VMAs to be\\ deleted, split, extended\\ or merged.\end{tabular} & \textit{\begin{tabular}[c]{@{}c@{}}vma\_merge( ),\_\_split\_vma( ),\\ detach\_vmas\_to\_be\_unmapped( ),\\ madvise\_vma( ),\\ do\_mprotect\_pkey( ),\\ mlock\_fixup( ),\\ vma\_to\_resize( )\end{tabular}} & & page fault & \begin{tabular}[c]{@{}c@{}}A virtual address is\\ accessed for the first\\ time and OS allocates a\\ page that is mapped by it.\end{tabular} & \textit{\_\_handle\_mm\_fault( )} \\ \cline{1-3} \cline{5-7}
\textit{expand\_stack( )} & \begin{tabular}[c]{@{}c@{}}User/OS expands the\\ address space of stack.\end{tabular} & \textit{\begin{tabular}[c]{@{}c@{}}expand\_upwards( ),\\ expand\_downwards( ),\end{tabular}} & & out of memory & \begin{tabular}[c]{@{}c@{}}OOM killer reclaims\\ pages.\end{tabular} & \textit{zap\_pmd\_range( )} \\ \cline{1-3} \cline{5-7}
NUMA balance & \begin{tabular}[c]{@{}c@{}}OS periodically migrates\\ pages among NUMA nodes.\\ PTEs are modified to\\ be inaccessible.\end{tabular} & \textit{change\_prot\_numa( )} & & get user page & \begin{tabular}[c]{@{}c@{}}Direct I/O, VFIO pin\\ get pages and\\ modifies them.\end{tabular} & \textit{follow\_page\_pte( )} \\ \hline
\end{tabular}
\end{table*}
\section{Implementation Details}
We summarize the operations that modify VMAs/PTEs in the OS as checkpoints.
Table~\ref{tab:modify} lists the functions (i.e., checkpoints) that can modify a VMA or a PTE, through a comprehensive analysis on the Linux kernel. As shown in the table, some operations are caused by users' queries, while others are caused by OS inherent memory management operations. Different kernel versions show quite stable call paths. \sun{Our analysis shows that only one function related to Async-fork has ever been modified from Linux kernel 2.6 to 6.0 (see Table~\ref{tab:modify-life}). Thus, the kernel related to our work is very stable.}
\sun{In practice, updating the kernel is a critical task for both cloud providers and users. They tend to stay with a kernel version for a relatively long time to keep the stability of services. Therefore, it is common for cloud providers to customize the OS kernel for important applications or specific internal needs in the production environment even the modification to the kernel cannot be upstreamed. For example, Facebook uses a customized disk cache component\footnote{\url{https://engineering.fb.com/2013/10/09/core-data/flashcache-at-facebook-from-2010-to-2013-and-beyond}. Last accessed on 2022/11/13.} in the kernel to speed up MySQL, and Google adds specific rules\footnote{\url{https://lwn.net/Articles/871195}. Last accessed on 2022/11/13.} for the out-of-memory killer in their kernel. The implementation of Async-fork adopts the modular design, which can be easily inserted in (or removed from) the kernel.}
\begin{table}[t!]
\centering
\caption{\sun{The lifecycle and location of related functions.}\label{tab:modify-life}}
\footnotesize
\begin{tabular}{|c|c|c|}
\hline
\textbf{Function} & \textbf{Lifecycle} & \textbf{Location} \\ \hline
vma\_merge() & v2.6.12 - v6.0 & mm/mmap.c \\ \hline
\_\_split\_vma( ) & v2.6.33 - v6.0 & mm/mmap.c \\ \hline
detach\_vmas\_to\_be\_unmapped( ) & v2.6.12 - v6.0 & mm/mmap.c \\ \hline
madvise\_vma( ) & v2.6.12 - v5.16.20 & mm/madvise.c \\ \hline
do\_mprotect\_pkey( ) & v4.9 - v6.0 & mm/mprotect.c \\ \hline
mlock\_fixup( ) & v2.6.12 - v6.0 & mm/mblock.c \\ \hline
vma\_to\_resize( ) & v2.6.33 - v6.0 & mm/mremap.c \\ \hline
expand\_upwards( ) & v2.6.15 - v6.0 & mm/mmap.c \\ \hline
expand\_downwards( ) & v2.6.23 - v6.0 & mm/mmap.c \\ \hline
change\_prot\_numa( ) & v3.8 - v6.0 & mm/mempolicy.c \\ \hline
\_\_handle\_mm\_fault( ) & v3.12 - v6.0 & mm/memory.c \\ \hline
zap\_pmd\_range( ) & v2.6.12 - v6.0 & mm/memory.c \\ \hline
follow\_page\_pte( ) & v3.16 - v6.0 & mm/gup.c \\ \hline
\end{tabular}
\end{table}
\section{Supplement Experiments}
In this section, we first report the experiment results of query processing throughput and total out-of-service time for the parent process during the snapshot process. We then report the experiment results of using Async-fork in log rewriting. Additionally, we report the execution time of Async-fork and ODF. Lastly, we discuss the method of tuning IMKVSes to further improve the performance.
{\bf Throughput Results.}
This experiment measures the processing throughput of the IMKVS when taking a snapshot of the in-memory data. To do so, we use Redis benchmark~\cite{redis-bench} to launch a great many queries and report the number of completed queries on the server side as the processing throughput.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{imgs/write-intensive-throughput.pdf}
\caption{The changes of throughput during snapshot process in a 16GB Redis Server.}
\label{fig:write-16gb-throughput}
\end{figure}
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{imgs/write-intensive-throughput-keydb.pdf}
\caption{\pp{The changes of throughput during snapshot process in a 16GB KeyDB Server.}}
\label{fig:write-16gb-throughput-keydb}
\end{figure}
Figures~\ref{fig:write-16gb-throughput} and \ref{fig:write-16gb-throughput-keydb} show the processing throughput on a 16GB Redis instance and a 16GB KeyDB instance, respectively. The throughput is reported for every 50ms.
\sun{Since KeyDB handles queries with multiple threads, we measure the throughput of KeyDB as follows: 1) each thread independently measures its throughput for every 50ms and records the value with a timestamp; and 2) for each time interval (50ms), estimate the throughput of the engine by summing up the values of all threads whose timestamps fall into the interval.}
As we can see, the throughput drops rapidly when the snapshot is taken while it increases gradually after that.
For example, In the worst case, the throughput of Redis server drops to 17,592 queries per second (QPS) and 42,980 QPS with ODF and Async-fork respectively.
Besides, the throughput increases to the normal level much faster with Async-fork than with ODF. \sun{The throughput of KeyDB is higher than that of Redis because KeyDB uses four
threads to process queries in our experiments.}
The experiments with other instance sizes of Redis and KeyDB show similar results.
Figure~\ref{fig:write-low-throughput} further shows the minimum processing throughput with different instance sizes of Redis and KeyDB.
Async-fork increases the minimum throughput by 2.44x on average (up to 2.9x) compared with ODF in Redis. \sun{The minimum throughput of KeyDB is increased by 1.6x on average (up to 2.69x).} Once the load exceeds the processing ability of the server, some queries queue up and suffer from a long latency.
The lower the processing throughput is, the longer the latency will be, which is consistent with the results of Section~\ref{sec:exp_write_intensive}.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{imgs/write-intensive-low-throughput.pdf}
\caption{\pp{The minimum processing throughput during the snapshot process.}}
\label{fig:write-low-throughput}
\end{figure}
\textbf{Total Out-of-service Time.} As discussed in Section~\ref{sec:exp_write_intensive}, the parent process is out-of-service for queries during the invocation of {\it copy\_pmd\_range()}. We report the total out-of-service time (the sum of the duration of each {\it copy\_pmd\_range()} invocation), as shown in Figure~\ref{fig:pmd-copy-total}. Compared with Async-fork, the parent process is out-of-service for a longer time with ODF.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{imgs/interrupt.pdf}
\caption{\pp{The total out-of-service time of Redis server.}}
\label{fig:pmd-copy-total}
\end{figure}
\sun{\textbf{Impact of Async-fork on Log Rewriting Queries.}
In addition to taking the snapshot, Redis can also persist data with Append Only File (AOF). Redis logs every write operation received by the server in a file, while these operations can then be replayed again at server startup to reconstruct the original dataset. However, the AOF log file gets bigger and bigger as write operations are performed. Consequently, Redis conducts the {\it log rewriting} to optimize the AOF log file~\cite{redis-persist}. Specifically, the store engine calls {\it fork} to create a child process. The child process rebuilds the AOF file into the shortest sequence of commands needed to rebuild the current dataset in memory, while the parent process continues to serve the user's queries. Due to the usage of {\it fork}, latency spikes will also be incurred during the log rewriting process.}
\sun{In the experiments, we use the Redis benchmark~\cite{redis-bench} to generate the workload and execute the BGREWRITEAOF command to trigger the log rewriting operation in Redis. We record the latency of the queries arriving during the period of conducting log rewriting (from the start of the fork operation to the end of log rewriting). We call these queries {\it log rewriting queries} for brevity.}
\sun{Figure~\ref{fig:aof} presents the 99\%-ile latency and maximum latency of log rewriting queries with different fork methods. Compared with the default {\it fork}, Async-fork reduces the 99\%-ile latency of log rewriting queries by 71.81\% (from 11.53ms to 3.25ms) on 1GB instance, 90.29\% (from 84.03ms to 8.16ms) on 8GB instance and 97.66\% (from 1093.35ms to 25.59ms) on 64GB instance. Compared with ODF, Async-fork reduces the 99\%-ile latency of log rewriting queries by 39.7\% (from 5.39ms to 3.25ms) on 1GB instance, 43.92\% (from 14.55ms to 8.16ms) on 8GB instance and 71.09\% (from 88.51ms to 25.59ms) on 64GB instance. The latency of log rewriting queries is higher than that of snapshot queries because the overall performance of Redis degrades when enabling AOF. In particular, the storage engine enabling AOF needs to synchronize the log into the disk~\cite{redis-persist} in addition to processing queries.
We observe the performance degradation of normal queries as well. When AOF is disabled, the 99\%-ile latency and maximum latency of normal queries on a 16GB instance are 0.079ms and 5.78ms, respectively. In contrast, the values increase to 1.56ms and 11.3ms when AOF is enabled. Nevertheless, Async-fork significantly reduces the latency of log rewriting queries as shown in Figure \ref{fig:aof}. The results demonstrate the effectiveness of Async-fork.}
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{imgs/rewrite_with_fork.pdf}
\caption{\sun{The 99\%-ile and maximum latency of log rewriting queries in Redis.}}
\label{fig:aof}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.9\columnwidth]{imgs/fork_time.pdf}
\caption{The execution time of Async-fork and ODF.}
\label{fig:forktime}
\end{figure}
{\bf The Execution Time of Fork.}
The duration of the fork operation greatly impacts the latency of the snapshot queries.
Figure~\ref{fig:forktime} shows the time of the parent process returns from calling Async-fork and ODF at different instance sizes.
Compared with the default {\it fork} (Figure~\ref{fig:characterization_fork}), both Async-fork and ODF reduce the execution time of invoking the fork operation effectively. In the large 64GB instance, the parent process completes Async-fork in 0.61ms, and completes ODF in 1.1ms respectively.
The time is reduced because they remove the most time-consuming phase from the default {\it fork}. We can also observe that Async-fork is slightly faster than ODF. One possible reason is that ODF introduces per-page counters to support the CoW sharing, incurring extra time cost for initialization.
\textbf{Tuning IMKVSes.} In our experiments, we set Redis and KeyDB to their default settings. Tuning the hyperparameters of IMKVSes may further reduce the latency of snapshot queries. In the following, we discuss the problem, while leaving it to the future studies.
By reducing the PTE modifications, the number of proactive PTE copies in Async-fork can be further reduced.
The PTE modifications can be reduced by configuring the memory allocator appropriately.
In general, by allocating a large enough virtual memory space for storing user data and meta information, the required {\it mmap} operations at runtime can be greatly reduced. In addition, it is better to retain the unused virtual memory for late use, rather than free them by calling {\it munmap}. Both the {\it mmap} and {\it munmap} operations cause PTE modifications. The default memory allocator {\it jemalloc}\footnote{\url{https://github.com/jemalloc/jemalloc}. Last accessed on 2022/11/13.} of Redis and KeyDB supports the above configuration: 1) Pre-allocating virtual memory space through {\it zmalloc/zfree} without populating it (so that no extra physical memory is consumed). 2) Configuring the built-in property `retain' to be true.
|
\section{Introduction}
Let $k$ be a number field and let $\mathcal G$ be a commutative algebraic group defined over $k$. We denote by $M_k$ the set of places of $k$ and by $k_v$ the completion of $k$ at $v$.
In 2001, Dvornicich and Zannier, motivated by a particular case of the Hasse principle for binary quadratic forms, stated the following problem, which is known as the \emph{Local-global divisibility problem} (see \cite{DZ1}).
\begin{Problem}[Dvornicich and Zannier, \cite{DZ1}]\label{prob}
Let $q$ be a fixed positive integer. If we assume that the point $P \in \mathcal G(k)$ has the following property: for all but finitely many $v\in M_k$ there exists $D_v \in \mathcal G(k_v)$ such that $P = qD_v$; can we conclude that there exists $D \in \mathcal G(k)$ such that $P = qD$?
\end{Problem}
Clearly it is sufficient to answer the problem when $q$ is a power of a prime.
The classical case of the multiplicative group $\mathcal G = \mathbb{G}_m$ has a complete answer: positive for $q$ odd (see for example \cite{DZ1,AT}) and negative for $q = 2^m$ with $m\geq 3$ (see \cite{Trost}).
For a general commutative algebraic group, Dvornicich and Zannier gave in \cite{DZ1} a cohomological interpretation of the problem and sufficient conditions to answer the question (see also \cite{DZ3}). In particular, they show that such an answer is linked to the triviality of a subgroup of ${\rm H}^1({\rm Gal}\left(k\left(\mathcal G[q]\right)/k\right), \mathcal G[q])$, where $k\left(\mathcal G[q]\right)$ is the field obtained by adding the coordinates of the $q$--torsion points of $\mathcal G$ to $k$, called the \emph{first local cohomology group} (see Section \ref{prelim} for further details).
By using these tools, it was possible to give criteria to answer the local-global divisibility problem for several algebraic groups.
The problem for elliptic curves was long studied: for $p=q$ there is an affirmative answer over any number field $k$ \cite {Cas62, DZ1, Wong}, for powers $p^n$, with $n\geq 1$ there is an affirmative answer for all $p\geq 5$ over $\mathbb Q$ \cite{PRV3} and for all $p> (3^{[k:\mathbb Q] /2} + 1 )^2$ over a number field $k$ that does not contain $\mathbb Q({\zeta}_p+{\zeta}_p^{-1})$, where ${\zeta}_p$ is a $p$--th root of unity \cite{PRV2}. This bound on $p$ is sharp; indeed for any power $2^m$ and $3^m$, with $m \geq 2$, there exist counterexamples over $\mathbb Q$ \cite{Creu, Creu2, DZ2, GRgl, Pal3}. These counterexamples imply a negative answer to the problem for powers $2^m$ and $3^m$, with $m\geq 2$, in all but finitely many number fields. While there are no explicit counterexamples for $p\geq 5$ over a number field $k$, in \cite{Ran} necessary conditions on ${\rm Gal}(k\left({\mathcal{E}}[p]\right)/k)$ that have to be satisfied in order to have local-global divisibility by $p^l$ for a prime $p\geq 5$ are given. In addition, for elliptic curves an effective version of the hypotheses of Problem \ref{prob} is produced in \cite{DP2}. For principally polarized abelian surfaces in \cite{GRab} the authors proved sufficient conditions for the local-global divisibility by any prime power $p^l$, while in \cite{GRgl2} they generalized these conditions in order to answer the case of ${\rm GL}_2$--type varieties (see also \cite{GRgl}).
Furthermore, in \cite{Pal4} the third author produced conditions for the local-global $p$--divisibility for a general commutative algebraic group.
In the case of abelian varieties, the problem is also linked to a classical question posed by Cassels in 1962 on the $p$--divisibility of the Tate-Shafarevich group (see \cite{Cas62, CS, Creu2}).
For a detailed survey on the topic, see \cite{DP}.
\bigskip
In this work we focus on algebraic tori. In \cite{DZ1} Dvornicich and Zannier proved that the local-global divisibility by a prime $p$ holds for tori of dimension $r \leq \max\{3,2(p-1)\}$, but fails for a torus of dimension $r = p^4-p^2+1$. In this last case they produced a $k$--rational point which is almost always locally $p$--divisible, but not globally.
In \cite{Ill} Illengo improved the condition $r\leq 2(p-1)$ given by Dvornicich and Zannier with the weaker one $r<3(p-1)$. He also proved that this bound is best possible, by building an example with $r=3(p-1)$ for which the local-global divisibility by $p$ fails.
In this work we prove that the local-global divisibility by any power $p^n$, with $n\geq 2$, holds for an algebraic torus of dimension $r<p-1$. Clearly, the condition $r<3(p-1)$ must hold, because the $p^n$--divisibility implies the $p$--divisibility. We show that in general we have to lower this bound to $r<p-1$. In particular, we prove that if $r<p-1$ we have an affirmative answer to Problem \ref{prob}, while for $r$ greater than or equal to $p-1$ the local-global divisibility by $p^n$ is no longer assured. For the latter, we construct a counterexample of dimension $p-1$ for the local-global $p^2$--divisibility. We remark that, starting from this construction, one can build a counterexample for any dimension $r \geq p-1$, by taking the product of the torus that we build in Lemma \ref{costr toro} with the split torus of dimension $r-(p-1)$. This shows that $r<p-1$ must hold also for the local-global $p^n$--divisibility since again the $p^n$--divisibility implies the $p^2$--divisibility for any $n\geq 2$. We can summarize these results in the following theorem, that we prove in Section \ref{sec:dim1}:
\begin{theorem}\label{thm:1}
Let $p$ be an odd prime.
\begin{itemize}
\item[(a)] Let $k$ be a number field and let $T$ be a torus defined over $k$. If $T$ has dimension less than $p-1$, then the local-global divisibility by any power $p^n$ holds for $T(k)$.
\item[(b)] For each $r\geq p-1$ there exists a torus defined over $\mathbb Q$ of dimension $r$ and a finite extension $L/\mathbb Q$ such that the local-global divisibility by $p^n$ does not hold for $T(L)$ for any $n \geq 2$.
\end{itemize}
\end{theorem}
Nevertheless, under certain condition on the base field $k$, e.g.\ by adding a primitive $p^n$--th root of unity to $k$, we can say something more when $p-1 \leq \dim(T)<3(p-1)$. Let $k(T[p^n])$ be the field obtained by adding the coordinates of the $p^n$--torsion points of $T$ to $k$, in Section \ref{sec:dim2} we prove the following.
\begin{theorem}\label{thm:2}
Suppose that $T$ is a torus defined over $k$ with $p-1\leq \dim(T) < 3(p-1)$ and $p$ does not divide the degree $[k\left(T[p^n]\right) \cap k\left({\zeta}_{p^n}\right) : k]$, where ${\zeta}_{p^n}$ is a $p^n$--th root of unity. Then the local-global divisibility by $p^n$ holds for $T(k)$.
\end{theorem}
\noindent{\bf Acknowledgements.} The authors warmly thank Davide Lombardo for his precious help during the preparation of this manuscript and Umberto Zannier for useful discussions. The first and third authors are members of INdAM-GNSAGA. Part of this research was done while the first author was visiting the third author; the first author is grateful to the Department of Mathematics and Computer Science of the University of Calabria for the kind hospitality and to the University of L'Aquila for the relative funding.
\section{Preliminary Results}\label{prelim}
As mentioned above, a useful method introduced by Dvornicich and Zannier in \cite{DZ1} in dealing with this problem is to translate it into cohomological terms. Let us recall some definitions and results.
Let $\mathcal G$ be a commutative algebraic group, defined over a number field $k$. Given a positive integer $q$, we denote by $\mathcal G[q]$ the set of $q$-torsion points of $\mathcal G(\overline k)$. It is isomorphic to $\left({\mathbb Z}/q{\mathbb Z}\right)^n$, for some $n$ depending only on $\mathcal G$ (see for example \cite[Section 2]{DZ1}).
Let $K := k(\mathcal G[q])$ be the number field generated by adding to $k$ the coordinates of the $q$-torsion points of $\mathcal G$. It is a Galois extension of $k$ and we denote by $G$ its Galois group. Let $\Sigma$ be the set of places $v$ unramified in $K$ and let $G_v = {\rm Gal}(K_w/k_v)$, where $w$ is a place of $K$ extending $v \in \Sigma$.
In \cite{DZ1}, the authors introduce a subgroup of ${\rm H}^1(G,\mathcal G[q])$ called the \emph{first local cohomology group}:
\begin{equation}\label{def h1loc}
{\rm H}_{{\rm{loc}}}^1(G,\mathcal G[q]) = \bigcap_{v \in \Sigma} \ker \left({\rm H}^1(G,\mathcal G[q]) \xrightarrow{{\rm res}_v} {\rm H}^1(G_v, \mathcal G[q])\right).
\end{equation}
By \v{C}hebotarev Density Theorem, the Galois groups $G_v$ run over all cyclic subgroups of $G$, as $v$ varies in $\Sigma$. Therfore a cocycle $\{Z_\sigma\}_{\sigma \in G}$ in ${\rm H}^1(G,\mathcal G[q])$ is an element in ${\rm H}_{{\rm{loc}}}^1(G,\mathcal G[q])$ if and only if for every $\sigma \in G$ there exists $W_{\sigma} \in \mathcal G[q]$ such that $Z_\sigma = (\sigma-1)W_\sigma$ (see \cite{DZ1} and \cite{DP} for further details). This justifies the following, more general, definition.
\begin{definition}[Dvornicich and Zannier \cite{DZ1}]
Let $G$ be a group and let $M$ be a $G$-module. We say that a cocycle $ \{Z_g\}_{g \in G}$ of $G$ with values in $M$ \emph{satisfies the local conditions} if there exist $W_g \in M$ such that $Z_g= (g-1)W_g$ for all $g\in G$. We denote by ${\rm H}^1_{\rm{loc}}(G,M)$ the subgroup of ${\rm H}^1(G,M)$ of the classes of these cocycles.
\end{definition}
As mentioned, it is sufficient to consider the case when $q$ is a prime power, thus from now on let $q= p^n$ for some prime $p$. By \cite[Proposition 2.1]{DZ1} the triviality of the first local cohomology group gives a sufficient condition for an affirmative answer to Problem \ref{prob}:
\begin{proposition}[Dvornicich and Zannier \cite{DZ1}]\label{suffcond}
Assume that ${\rm H}^1_{{\rm{loc}}}({\rm Gal}(K/k), \mathcal G[q]) = 0$. Let $P \in \mathcal G(k)$ be a rational point with the following property: for all but finitely many primes $v$ of $k$, there exists $D_v \in \mathcal G(k_v)$ such that $P = qD_v$. Then there exists $D \in \mathcal G(k)$ such that $P = qD$.
\end{proposition}
Furthermore, by \cite[Theorem 2.5]{DZ1} it is enough to prove that ${\rm H}^1_{{\rm{loc}}}(G_p, \mathcal G[q]) = 0$ for a $p$--Sylow subgroup $G_p$ of $G$. Moreover, the converse of Proposition \ref{suffcond} is also true over a finite extension of $k$, as the following theorem by Dvornicich and Zannier shows.
\begin{theorem}[{\cite[Theorem 3]{DZ3}}]\label{neccond}
Suppose that ${\rm H}^1_{\rm{loc}}(G, \mathcal G[q])$ is not trivial. Then there exists a number field $L$ such that $L \cap K = k$ and a point $P \in \mathcal G(L)$ which is divisible by $q$ in $\mathcal G(L_w)$ for all places $w$ of $L$ but is not divisible by $q$ in $\mathcal G(L)$.
\end{theorem}
We remark that if in \eqref{def h1loc} we take all valuations, instead of almost all, then we get a group isomorphic to the Tate-Shafarevich group $\Sha(k,\mathcal G[q])$. Thus the vanishing of ${\rm H}^1_{\rm{loc}}(G,\mathcal{G}[q])$ implies the vanishing of $\Sha(k,\mathcal G[q])$, which is a sufficient condition to give an affirmative answer to the local-global divisibility problem in the case when $v$ runs over all valuations of $k$ (see for instance \cite{Creu2}).
\medskip
Let us now introduce some notation for algebraic tori that we will use in the next sections. We adopt the same notation as in \cite[Section 4]{DZ1}. Let $T$ be an algebraic torus defined over a number field $k$, of dimension $r<3(p-1)$.
There exists an isomorphism of algebraic groups (defined over $\overline k$) $\phi:T \longrightarrow \mathbb{G}^r_m$. Let $G_k$ be the absolute Galois group of $k$ and consider the following homomorphism
\[\begin{matrix}
\psi: & G_k & \longrightarrow & \rm{Aut}(\mathbb{G}^r_m) & \simeq {\rm GL}_r({\mathbb Z})\\
& \sigma & \longmapsto & \phi\circ\sigma\circ\phi^{-1}.\\
\end{matrix}
\]
The isomorphism $\phi$ is defined over some number field, thus $\ker\psi$ has finite index in $G_k$ and we can identify $\Delta:= \psi(G_k)$ with a finite subgroup of ${\rm GL}_r({\mathbb Z})$. We denote by $L$ the field fixed by $\ker\psi$; it is a normal extension of $k$ and $\Delta \simeq {\rm Gal}(L/k)$. The field $L$ is also the splitting field of the torus $T$.
Let ${\zeta} :={\zeta}_{p^n}$ be a primitive $p^n$--th root of unity and let $\chi$ be the cyclotomic character
\[
\begin{matrix}
\chi: & G_k & \longrightarrow & \left({\mathbb Z}/p^n{\mathbb Z}\right)^{\times}\\
& \sigma & \longmapsto & j_\sigma,\\
\end{matrix}
\]
where $j_\sigma$ is such that $\sigma({\zeta}) = {\zeta}^{j_\sigma}$.
Let $T[p^n]$ be the group of the $p^n$--torsion points of $T$. We have $T[p^n] = T(\overline k)[p^n] = \left\{ \left({\zeta}^{j_1},\dots,{\zeta}^{j_r}\right) \in \left(\overline k^{\times}\right)^r \mid j_h \in \left({\mathbb Z}/p^n{\mathbb Z} \right) \right\}$ and we fix the following isomorphism
\begin{equation}\label{iso}
\begin{matrix}
T[p^n] & \longrightarrow & \left({\mathbb Z}/p^n{\mathbb Z}\right)^r\\
({\zeta}^{j_1},\dots,{\zeta}^{j_r}) & \longmapsto & (j_1,\dots,j_r).\\
\end{matrix}
\end{equation}
By this isomorphism, the natural action of $G_k$ on $T[p^n]$ induces the following action on $\left({\mathbb Z}/p^n{\mathbb Z}\right)^r$: $\sigma \cdot v = j_\sigma \widetilde{\psi(\sigma)} v$ for all $v \in \left({\mathbb Z}/p^n{\mathbb Z}\right)^r$, where the tilde denotes the reduction ${\rm{mod} \hspace{0.1cm} } p^n$. Therefore we have the homomorphism
\begin{equation}\label{xi}
\begin{matrix}
\xi: & G_k & \longrightarrow & {\rm GL}_r({\mathbb Z}/p^n{\mathbb Z})\\
& \sigma & \longmapsto & j_\sigma\widetilde{\psi(\sigma)}.\\
\end{matrix}
\end{equation}
It is easy to check that the field fixed by $\ker\xi$ is $K := k\left(T[p^n]\right)$, thus the image $G$ of $\xi$ in ${\rm GL}_r({\mathbb Z}/p^n{\mathbb Z})$ is a finite subgroup isomorphic to $G_k/\ker\xi \simeq {\rm Gal}(K/k)$. Let $G_{k\left({\zeta}\right)}$ be the subgroup ${\rm Gal}\left(\overline{k}/k\left({\zeta}\right)\right)$ of $G_k$. We have $L = \overline{k}^{\ker\psi}$ and $K = \overline{k}^{\ker\xi}$, since $\ker\xi \supseteq \ker\psi\cap G_{k\left({\zeta}\right)}$ we also get $K \subseteq L\left({\zeta}\right)$.
The kernel of the restriction of $\xi$ to $G_{k\left({\zeta}\right)}$ is contained both in $\ker\xi$ and in $\ker\psi$. It follows that the image of $G_{k\left({\zeta}\right)}$ via $\xi$ is a normal subgroup $G'$ of $G$, which is also a normal subgroup of the reduction $\widetilde\Delta$ modulo $p^n$ of $\Delta$. In particular, we have the following tower of extensions:
\begin{figure}[H]
\centering
\begin{tikzpicture}
\node (a) at (0,0) {$\overline k ^{\ker \xi_{|G_{k({\zeta})}}}$};
\node (b) at (-2,-2) {$K$};
\node (c) at (2,-2) {$k({\zeta})$};
\node (d) at (0,-4) {$K\cap k({\zeta})$};
\node (e) at (0,-6) {$k$.};
\draw (a) -- (b);
\draw (a) -- (c);
\draw (b) -- (d);
\draw (c) -- (d);
\draw (d) -- (e);
\node (f) at (2,-0.7) {$G'$};
\node (g) at (0,-2.7) {$G'$};
\node (h) at (-2,-5) {$G$};
\draw[bend left] (a) edge (c);
\draw[bend left] (b) edge (d);
\draw[bend right] (b) edge (e);
\end{tikzpicture}
\caption*{}
\end{figure}
We have obtained that $G$ and $\widetilde\Delta$ have a common normal subgroup $G'$, with $[G:G'] \mid p^{n-1}(p-1)$ and $[\widetilde\Delta:G']\mid p^{n-1}(p-1)$. In \cite{DZ1} the authors can easily conclude that $G$ and $\widetilde\Delta$ also have the same $p$--Sylow subgroups (since in their situation both $[G:G']$ and $[\widetilde\Delta:G']$ are coprime with $p$) and thus they reduce themselves to study only the $p$--Sylow subgroups in $\widetilde\Delta$. Instead, in our general setting for the divisibility by $p^n$ problem, we have to distinguish two cases: either $p \mid [G:G']$ or $p \nmid [G:G']$.
By {\cite[Chapter 3, Theorem 3.37]{Mac}} the $p$-Sylow subgroups of a quotient are images of $p$-Sylow subgroups through the canonical projection homomorphism. The converse also is obviously true. Therefore we have the following statement.
\begin{theorem}\label{sylow quot}
A subgroup of a quotient is a $p$-Sylow subgroup if and only if it is the image through the canonical projection homomorphism of a $p$-Sylow subgroup.
\end{theorem}
In the proof of Theorem \ref{thm:2} we will also need the following result whose proof we include for the reader's convenience (for a more general result see \cite{Serre}).
\begin{lemma}\label{lemma:inj}
Let $p$ be an odd prime and let $\pi :{\rm GL}_n({\mathbb Z})\longrightarrow {\rm GL}_r({\mathbb Z}/p{\mathbb Z})$ be the reduction modulo $p$. Then $\pi$ is injective on finite subgroups of ${\rm GL}_r({\mathbb Z})$.
\end{lemma}
\begin{proof}
It is enough to show that if $A\in \ker\pi$ and $A$ has finite order $m$, then $A={\rm Id}$. Suppose that instead $A\neq {\rm Id}$. If $p\nmid m$, then write $A = 1 + p^k B$, with $k\geq 1$ and $B\in {\rm Mat}_r({\mathbb Z})$ such that $p$ does not divide at least one of the entries of $B$. We have
\begin{equation*}
1 = A^m = \sum_{j=0}^m \binom{m}{j}p^{jk}B^j = 1 + mp^kB + \sum_{j=2}^m \binom{m}{j}p^{kj}B^j.
\end{equation*}
Thus $mp^kB = -p^{k+1}C$, for some $C \in {\rm Mat}_r({\mathbb Z})$ and so $mB = -pC$, which is a contradiction. On the other hand, if $p \mid m$, then $\overline A := A^{p/m}$ has order $p$ and lies in $\ker \pi$. We have $\overline A = 1 + p^k\overline B$ with $k\geq 1$ and $\overline B\in {\rm Mat}_r({\mathbb Z})$ such that $p$ does not divide at least one of the entries of $\overline B$. We have
\begin{equation*}
1 = \overline A^p = \sum_{j=0}^p \binom{p}{j}p^{jk}\overline B^j = 1 + p^{k+1}\overline B + \sum_{j=2}^{p-1}{\binom{p}{j}p^{kj}\overline B^j} + p^{pk}\overline B^p.
\end{equation*}
Since $p\neq 2$ and $k \geq 1$, we get $p^{k+1}\overline B = -p^{k+2} \overline C$, for some $\overline C \in {\rm Mat}_r({\mathbb Z})$, so $B = -p\overline C$ and we have again a contradiction.
\qedhere
\end{proof}
\noindent Finally, we state the following Lemma by Illengo; this is a key result for the local-global divisibility in algebraic tori and we will repeatedly use it in our proofs.
\begin{lemma}[{\cite[Lemma 4]{Ill}}]\label{illengo}
Let $p$ be a prime and let $\Gamma$ be a $p$--group of matrices in ${\rm SL}_r(\mathbb Q)$. If $r<p(p-1)$ then $\Gamma$ is isomorphic to $({\mathbb Z}/p{\mathbb Z})^b$, for some $b \leq r/(p-1)$.
\end{lemma}
\section{Proof of Theorem \ref{thm:1}}\label{sec:dim1}
In this section we prove Theorem \ref{thm:1}. With the notation as above, we show that:
\begin{itemize}
\item[(a)] for any algebraic torus $T$ of dimension $r<p-1$ we have ${\rm H}^1_{\rm{loc}}(G_p, T[p^n]) = 0$ and then by \cite[Theorem 2.5]{DZ1} Problem \ref{prob} has affirmative answer;
\item[(b)] there exists a torus $T$ defined over $\mathbb Q$ of dimension $p-1$ such that ${\rm H}^1_{{\rm{loc}}}\left(G_p, T[p^2]\right) \neq 0$, thus by Theorem \ref{neccond}, there exists a number field in which the local-global divisibility by $p^2$ does not hold in $T$.
\end{itemize}
For any $n\geq 2$ the $p^n$--divisibility implies the $p^2$--divisibility, therefore the counterexample in (b) shows that the bound on the dimension of $T$ is best possible.
Since point (b) requires more efforts we start by showing a few results that we will use for its proof. The first step is building a torus such that $G_p$ is isomorphic to $\left({\mathbb Z}/p{\mathbb Z}\right)^2$.
\begin{lemma}\label{costr toro}
Let $p$ be an odd prime and let ${\zeta}$ be a primitive $p^2$--th root of unity. There exists an algebraic torus $T$ of dimension $r = p-1$ defined over $\mathbb Q$ such that $G \simeq {\rm Gal}\left(\mathbb Q\left(T[p^2]\right)/\mathbb Q\right)$ is isomorphic to ${\mathbb Z}/p{\mathbb Z}\times \left({\mathbb Z}/p^2{\mathbb Z}\right)^\times$. In particular, the $p$--Sylow subgroup of $G \subseteq {\rm GL}_r({\mathbb Z}/p^2{\mathbb Z})$ is generated by
\[
\gamma_1 =\begin{pmatrix}
0 & & & & -1\\
1 & 0 & & & -1\\
& \ddots & \ddots & & \vdots\\
& & 1 & 0 & -1\\
& & & 1 & -1\\
\end{pmatrix}
\quad \text{and} \quad \gamma_2 =
\begin{pmatrix}
p+1 & & \\
& \ddots &\\
&& p+1\\
\end{pmatrix}.
\]
\end{lemma}
\begin{proof}
By Dirichlet's Theorem about primes in arithmetic progressions, one can easily see that there exists a number field $L$ such that $L\cap \mathbb Q\left({\zeta}\right) = \mathbb Q$ and $L/\mathbb Q$ is a cyclic extension of degree $p$.
Let $\sigma$ be a generator of ${\rm Gal}(L/\mathbb Q)$.
\noindent Consider the split torus $\mathbb{G}_m = \mathbb{G}_{m,L}$ defined over $L$. We denote the group $\mathbb{G}_m[p^2]$ by $V\simeq {\mathbb Z}/p^2{\mathbb Z}$ using additive notation. The Galois action on $V$ is given by the cyclotomic character $\chi: G_\mathbb Q \longrightarrow ({\mathbb Z}/p^2{\mathbb Z})^{\times}$. Notice that we can restrict $\chi$ to ${\rm Gal}(\overline\mathbb Q/L)$.
Let $X = R_{L/\mathbb Q}\mathbb{G}_m$ be the Weil restriction of $\mathbb{G}_m$. It is an algebraic torus defined over $\mathbb Q$ of dimension $p$, split over $L$. The group $X[p^2]$ of the $p^2$--torsion points of $X$ is a free ${\mathbb Z}/p^2{\mathbb Z}$-module of rank $p$. The Galois representation of $G_{\mathbb Q}$ on these points is induced from $V$ and as representations we have $$X[p^2] = {\rm Ind}_{G_L}^{G_{\mathbb Q}} V = \bigoplus_{j=0}^{p-1} \sigma^j V.$$
Now let $T$ be the norm $1$ subtorus of $X$, that is $T = R^{(1)}_{L/\mathbb Q} \mathbb{G}_m = \ker (R_{L/\mathbb Q}\mathbb{G}_m \stackrel{N_{L/\mathbb Q}}{\longrightarrow} \mathbb{G}_m)$, the kernel of the (generalized) norm map on $X$. It is an algebraic torus over $\mathbb Q$ of dimension $r = p-1$, split over $L$. We are going to show that $\mathbb Q\left(T[p^2]\right) = L\left({\zeta}\right)$.
Through the isomorphism \eqref{iso}, we can regard $T[p^2]$ as the submodule $W$ of $X[p^2]$ of vectors $(x_0,x_1\dots,x_{p-1})$ such that the sum of all coordinates is equal to zero (we are using the additive notation here). The group ${\rm Gal}(L/\mathbb Q)$ acts on $X[p^2]$, and hence on $W$, by permuting the coordinates: by considering the indexes of the coordinates modulo $p$, for $h \in \{0,1,\dots,p-1\}$ we have
\[ \sigma^h \cdot (x_0,x_1\dots,x_{p-1}) = (x_{-h},x_{-h+1},\dots,x_{-h-1}). \]
The Galois action on $T[p^2]$ is given by $\xi: G_{\mathbb Q} \longrightarrow \rm{Aut}(W)$, i.e. by the action on the points of $X[p^2]$ that lie in $W$ (see \eqref{xi}). We have that $\mathbb Q\left(T[p^2]\right) = \overline\mathbb Q^{\ker\xi}$.
Thus, in order to determine this field, we need to find the elements of $G_\mathbb Q$ that act trivially on $W$. Let $g$ be an element of $G_\mathbb Q$. We can write $g = \tau\sigma^h$, with $\tau \in G_L$ and $h \in \{0,1,\dots,p-1\}$. The action of $g$ on an element of $W$ is given by
\begin{align}\label{xi action}
g \cdot (x_0,x_1\dots,x_{p-1}) & = \tau\sigma^h \cdot (x_0,x_1\dots,x_{p-1}) \nonumber \\
& = (\chi(\tau)x_{-h},\chi(\tau)x_{-h+1},\dots,\chi(\tau)x_{-h-1}) \\
& = \chi(\tau) (x_{-h},x_{-h+1},\dots,x_{-h-1}). \nonumber
\end{align}
Since $p\geq 3$, the automorphism $g = \tau\sigma^h$ fixes every $(x_0,x_1\dots,x_{p-1})$ in $W$ if and only if $h = 0$ and $\chi(\tau) = 1$, i.e. if and only if $g \in G_L \cap G_{\mathbb Q({\zeta})}$.
We therefore conclude that $\mathbb Q\left(T[p^2]\right) = L({\zeta})$.
Since $L$ and $\mathbb Q({\zeta})$ are linearly disjoint over $\mathbb Q$, the extension $\mathbb Q\left(T[p^2]\right)/\mathbb Q$ has Galois group ${\rm Gal}\left(\mathbb Q\left(T[p^2]\right)/\mathbb Q\right) \simeq {\rm Gal}\left(L/\mathbb Q\right) \times {\rm Gal}\left(\mathbb Q({\zeta})/\mathbb Q\right)$, which is isomorphic to ${\mathbb Z}/p{\mathbb Z} \times \left({\mathbb Z}/p^2{\mathbb Z}\right)^{\times}$. In particular, it has only one $p$--Sylow subgroup isomorphic to $\left({\mathbb Z}/p{\mathbb Z}\right)^2$. We can take $\sigma$, the automorphism generating ${\rm Gal}(L/\mathbb Q)$, and the automorphism $\eta$ sending ${\zeta}$ to ${\zeta}^{p+1}$ as generators of the $p$--Sylow subgroup. By choosing the basis $v = (1,-1, 0, \ldots, 0,0),\,\sigma(v),\,\sigma^2(v),\ldots,\sigma^{p-2}(v)$ of $W$, the homomorphisms $\sigma$ and $\eta$, viewed as elements of ${\rm GL}_{p-1}({\mathbb Z}/p^2{\mathbb Z})$, may be represented by the matrices
\[
\gamma_1 =\begin{pmatrix}
0 & & & & -1\\
1 & 0 & & & -1\\
& \ddots & \ddots & & \vdots\\
& & 1 & 0 & -1\\
& & & 1 & -1\\
\end{pmatrix}
\quad \text{and, respectively,} \quad \gamma_2 =
\begin{pmatrix}
p+1 & & \\
& \ddots &\\
&& p+1\\
\end{pmatrix}.
\]
\end{proof}
\begin{remark}
In the proof, we have explicitly described the homomorphism $\xi$ in \eqref{xi} of Section \ref{prelim} in the particular case where $T$ is the norm torus.
\end{remark}
Let $G_p$ be the $p$--Sylow subgroup of the group $G$ that we just constructed. Now we show that ${\rm H}^1_{{\rm{loc}}}\left(G_p,\left({\mathbb Z}/p^2{\mathbb Z}\right)^{p-1}\right) \neq 0$. As seen, thanks to the isomorphism $T[p^2] \simeq \left({\mathbb Z}/p^2{\mathbb Z}\right)^{p-1}$ this is equivalent to proving that the first local cohomology group ${\rm H}^1_{{\rm{loc}}}\left(G_p, T[p^2]\right)$ is trivial.
\begin{proposition}\label{h1loc}
There exists a (unique) extension of
\[
\gamma_1 \longmapsto v_1 = \begin{pmatrix} p-1\\0\\ \vdots\\0\\1\\\end{pmatrix}, \quad
\gamma_2 \longmapsto v_2 = \begin{pmatrix} p \\ \vdots\\p\\0\\\end{pmatrix}\\
\]
to a cocycle in ${\rm H}^1\left(G_p,\left({\mathbb Z}/p^2{\mathbb Z}\right)^{p-1}\right)$ and it is a non-trivial element of ${\rm H}_{{\rm{loc}}}^1\left(G_p, \left({\mathbb Z}/p^2{\mathbb Z}\right)^{p-1}\right)$.
\end{proposition}
\begin{proof}
To check that the assigned vectors define a cocycle, we have to prove that in $\left({\mathbb Z}/p^2{\mathbb Z}\right)^{p-1}$
\begin{align}
(1+\gamma_1+\dots+\gamma_1^{p-1})v_1 = \underline{0} \label{gamma1}\\
(1+\gamma_2+\dots+\gamma_2^{p-1})v_2 = \underline{0} \label{gamma2} \\
(1-\gamma_2)v_1 + (\gamma_1-1)v_2 = \underline{0} \label{gamma12}
\end{align}
where $\underline{0}$ is the vector with all coordinates equal to zero; indeed these conditions must be true since $\gamma_1^p=\gamma_2^p=\gamma_1\gamma_2\gamma_1^{-1}\gamma_2^{-1} = 1$ in $G_p$.
A lifting of $\gamma_1$ to ${\rm GL}_{p-1}(\mathbb Q)$ is the matrix $\gamma_1$ itself. It solves the polynomial $x^p-1$ and it is easy to see that $1$ is not an eigenvalue. Thus the minimal polynomial of $\gamma_1$ is $x^{p-1}+x^{p-2} + \dots + x+1$ and so \eqref{gamma1} holds. Since $\gamma_2 = (p+1){\rm Id}$, we have $(1+\gamma_2+\dots+\gamma_2^{p-1})v_2 = (1 +(p+1) + \dots + (p+1)^{p-1})v_2 \equiv \underline{0} \mod p^2$, hence \eqref{gamma2}. For \eqref{gamma12}, some simple calculations lead to:
\[
(1-\gamma_2)v_1 + (\gamma_1-1)v_2 \equiv \begin{pmatrix} p \\ 0 \\ \vdots \\ 0 \\ \end{pmatrix} + \begin{pmatrix} -p \\ 0 \\ \vdots \\ 0 \\ \end{pmatrix} \equiv \underline{0} \mod p^2.
\]
Thus we can extend $v_1$ and $v_2$ to a cocycle $Z = \{Z_{\gamma}\}_{\gamma\in G_p}$, with $Z_{\gamma_1}=v_1$ and $Z_{\gamma_2} = v_2$. We now prove that $Z$ is not trivial. Suppose that it is a coboundary, then there exists $w \in \left({\mathbb Z}/p^2{\mathbb Z}\right)^{p-1}$ such that $v_1 = (\gamma_1-1)w$ and $v_2 = (\gamma_2-1)w$. We denote with $w^{(i)}$ the $i$--th coordinate of $w$. From $v_2 = (\gamma_2-1)w$ we have $v_2 = p w$, so that $pw^{(p-1)} = 0$. On the other hand, from $v_1 = (\gamma_1-1)w$ we have
\begin{equation}\label{condition1}
v_1 =
\begin{pmatrix}
p-1\\
0\\
\vdots\\
0\\
1\\
\end{pmatrix} =
\begin{pmatrix}
-1 & & & & -1\\
1 & -1 & & & -1\\
& \ddots & \ddots & &\vdots\\
& & 1 &-1 & -1\\
& & & 1 & -2 \\
\end{pmatrix}
\begin{pmatrix}
w^{(1)}\\
w^{(2)}\\
\vdots\\
\vdots\\
w^{(p-1)}\\
\end{pmatrix}.
\end{equation}
Thus we obtain the system
\begin{equation}\label{system}
\begin{cases*}
-w^{(1)} - w^{(p-1)} = p-1\\
w^{(1)} - w^{(2)} - w^{(p-1)} = 0\\
\vdots\\
w^{(p-3)}-w^{(p-2)}-w^{(p-1)} = 0\\
w^{(p-2)}-2w^{(p-1)} = 1.\\
\end{cases*}
\end{equation}
By adding the equations in \eqref{system} we get $p = -pw^{(p-1)}$, in contradiction with $pw^{(p-1)} = 0$. Therefore $Z$ is not a trivial cocycle in ${\rm H}^1\left(G_p,\left({\mathbb Z}/p^2{\mathbb Z}\right)^{p-1} \right)$.\\
We are left to show that $Z$ satisfies the local conditions. The elements $\gamma_2$ and $\gamma_1\gamma_2^h$, for $h = 0,1,\dots,p-1$ are generators of the cyclic subgroups of $G_p$. So it is enough to show that there exist $W_{\gamma_2}$ and $W_{\gamma_1\gamma_2^h}$ in $({\mathbb Z}/p^2{\mathbb Z})^{p-1}$ such that $Z_{\gamma_2} = (\gamma_2-1) W_{\gamma_2}$ and $Z_{\gamma_1\gamma_2^h} = (\gamma_1\gamma_2^h - 1)W_{\gamma_1\gamma_2^h}$ for all $h \in \{0,1,\dots,p-1\}$.
First, since $\gamma_2-1 = p{\rm Id}$, we have that the image of $\gamma_2-1$ is the submodule $M$ of $({\mathbb Z}/p^2{\mathbb Z})^{p-1}$ of vectors with each coordinate divisible by $p$. The vector $v_2$ satisfies this condition, so there exists $W_{\gamma_2} \in \left({\mathbb Z}/p^2{\mathbb Z}\right)^{p-1}$ such that $v_2 = Z_{\gamma_2} = (\gamma_2-1)W_{\gamma_2}$.
Now we fix $h \in \{0,1,\dots, p-1\}$ and define
\[V = \left\{ \begin{pmatrix} v^{(1)}\\ \vdots\\ v^{(p-1)}\\ \end{pmatrix} \in \left({\mathbb Z}/p^2{\mathbb Z}\right)^{p-1}\, \Bigg\vert \, \sum_{j = 1}^{p-1} v^{(j)} \equiv 0 \mod p \right\} \subseteq \left({\mathbb Z}/p^2{\mathbb Z}\right)^{p-1}.\]
We claim that the image of $\gamma_1\gamma_2^h - 1$ is equal to $V$.
Since $\gamma_2^h = (p+1)^h {\rm Id} \equiv (1+ph) {\rm Id} \mod p^2$, we can rewrite
\[\gamma_1\gamma_2^h -1 = (1+ph)\gamma_1 -1=
\begin{pmatrix}
-1 & & & & -1-ph\\
1+ph & -1 & & &-1-ph\\
& \ddots & \ddots & & \vdots\\
& & \ddots & -1& -1-ph\\
& & & 1+ph &-2-ph\\
\end{pmatrix}
\]
and thus ${\rm Im}(\gamma_1\gamma_2^h-1) \subseteq V$.\\
With easy calculations one can check that the determinant of $\gamma_1\gamma_2^h-1$ is equal to $p$ modulo $p^2$.
Take a lifting of $\gamma_1\gamma_2^h-1$ to an integer matrix; this integer matrix has still determinant equal to $p$ modulo $p^2$. Being ${\mathbb Z}$ a PID, we can consider its Smith normal form ${\rm diag}(\alpha_1,\alpha_2,\ldots,\alpha_{p-1})$,
and we get that $p \mid \alpha_{p-1}$ while $p^2 \nmid \alpha_{p-1}$ and $p \nmid \alpha_j$ for every $j =1, \ldots, p-2$. Its projection ${\rm diag}(\widetilde \alpha_1,\widetilde \alpha_2, \ldots, \widetilde \alpha_{p-1})$ modulo $p^2$ is such that $\widetilde \alpha_j$ is invertible, for $j = 1,\ldots, p-2$, and $\widetilde \alpha_{p-1}\neq 0$ is equal to $0$ modulo $p$. Therefore, up to basis changes in $({\mathbb Z}/p^2{\mathbb Z})^{p-1}$, the map $\gamma_1\gamma_2^h-1$ is
\[
\begin{pmatrix}
1 & & & \\
& \ddots & & \\
& & 1& \\
& & & p\\
\end{pmatrix}.
\]
It follows that ${\rm Im}(\gamma_1\gamma_2^h-1)$ has index equal to $p$ in $\left({\mathbb Z}/p^2{\mathbb Z}\right)^{p-1}$. The submodule $V$ has also index equal to $p$ in $\left({\mathbb Z}/p^2{\mathbb Z}\right)^{p-1}$ and so, from the inclusions ${\rm Im}(\gamma_1\gamma_2^h-1) \subseteq V \subseteq \left({\mathbb Z}/p^2{\mathbb Z}\right)^{p-1}$, we get the equality $V = {\rm Im}(\gamma_1\gamma_2^h-1)$.
To conclude that $Z$ satisfies the local conditions it only remains to verify that $Z_{\gamma_1\gamma_2^h}$ lies in $V$. We have $Z_{\gamma_1\gamma_2^h} = v_1 + \gamma_1Z_{\gamma_2^h} = v_1 + \gamma_1(1+\gamma_2+\dots+\gamma_2^{h-1})v_2$, with $v_1 \in V$ and $v_2 \in M$ (and also $\gamma_1(1+\gamma_2+\dots+\gamma_2^{h-1})v_2 \in M$). Since $M$ is contained in $V$, we get that $Z_{\gamma_1\gamma_2^h} \in V$.
\qedhere
\end{proof}
We conclude this section with the proof of Theorem \ref{thm:1}.
\begin{proof}[Proof of Theorem \ref{thm:1}.]
\par (a) Let $T$ be an algebraic torus of dimension $r<p-1$ defined over a number field $k$. With the notation of Section \ref{prelim}, since the inclusion $K\subseteq L\left({\zeta}\right)$ holds, it is clear that $G$ is isomorphic to the quotient ${\rm Gal}(L\left({\zeta}\right)/k)/{\rm Gal}(L\left({\zeta}\right)/K)$. Being $p$ an odd prime, any $p$--Sylow subgroup of the group $\Delta \simeq {\rm Gal}(L/k)$ is contained in ${\rm SL}_r(\mathbb Q)$; hence by the condition $r<p-1$ and Lemma \ref{illengo}, we have that $\Delta$ has no nontrivial $p$--Sylow subgroups. Since
\[ {\rm Gal}(L({\zeta})/k) \longhookrightarrow {\rm Gal}(L/k) \times {\rm Gal}(k({\zeta})/k),
\]
we get that any $p$--Sylow subgroup of ${\rm Gal}(L({\zeta})/k)$ is isomorphic to a subgroup of ${\rm Gal}(k({\zeta})/k)$. The latter is isomorphic to a subgroup of $\left({\mathbb Z}/p^n{\mathbb Z}\right)^{\times}$, hence it is cyclic. Thus ${\rm Gal}(L({\zeta})/k)$ contains only one cyclic $p$--Sylow subgroup. Now let $G_p$ be a $p$--Sylow subgroup of $G$. By Theorem \ref{sylow quot}, it is the image through the projection to the quotient of the $p$--Sylow subgroup of ${\rm Gal}(L({\zeta})/k)$, hence it is cyclic too and ${\rm H}^1_{{\rm{loc}}}(G_p, T[p^n])$ is trivial.
\par (b) As mentioned in the Introduction, it is enough to prove it for $r = p-1$ and $n=2$. By Lemma \ref{costr toro} and Proposition \ref{h1loc} there exists an algebraic torus defined over $\mathbb Q$ of dimension $p-1$ such that the $p$--Sylow $G_p$ of ${\rm Gal}(\mathbb Q(T[p^2]/\mathbb Q)$ is isomorphic to $\left({\mathbb Z}/p{\mathbb Z}\right)^2$ and ${\rm H}^1_{{\rm{loc}}}(G_p,T[p^2]) \neq 0$. By Theorem \ref{neccond}, there exists a number field $L$ such that $L\cap \mathbb Q\left(T[p^2]\right) = \mathbb Q$ and the local-global divisibility by $p^2$ does not hold for $T(L)$.
\qedhere
\end{proof}
\section{Proof of Theorem \ref{thm:2}} \label{sec:dim2}
We are going to prove Theorem \ref{thm:2}. In the following proof, we will show that ${\rm H}^1_{{\rm{loc}}}(G, T[p^n]) = 0$; in particular this proves that the local-global divisibility by $p^n$ holds.
\begin{proof}[Proof of Theorem \ref{thm:2}.]
We proceed by induction on $n \geq 1$ proving that ${\rm H}^1_{{\rm{loc}}}(G, T[p^n]) = 0$. The base of the induction $n=1$ is proven in \cite{Ill}. Thus suppose $n\geq 2$ and ${\rm H}^1_{{\rm{loc}}}(G,T[p^m]) = 0$, for every $m\leq n-1$.
With the notation of Section \ref{prelim}, let $F = K \cap k({\zeta})$. We have that $[F:k] = [G:G']$, thus it is coprime with $p$ by the assumptions of the theorem.
Therefore the $p$--Sylow subgroups of $G$ are contained in $G'$, hence in $\widetilde\Delta$.
For any $j = 1,\dots,n-1$, using the same construction as in Section \ref{prelim}, we can define a group $G^{(j)} \subseteq {\rm GL}_r({\mathbb Z}/p^j{\mathbb Z})$ that is isomorphic to ${\rm Gal}(k\left(T[p^j]\right)/k)$.
Let $\pi_j: {\rm GL}_r({\mathbb Z}/p^n{\mathbb Z}) \longrightarrow {\rm GL}_r({\mathbb Z}/p^j{\mathbb Z})$ be the reduction modulo $p^j$ and let $H^{(j)}$ be the intersection of $G$ with $\ker \pi_j$. It is easy to prove that $G^{(j)}= \pi_j(G) \simeq {\rm Gal}(k\left(T[p^j]\right)/k)$ and $H^{(j)} \simeq {\rm Gal}(K/k\left(T[p^j]\right))$. We claim that $H^{(1)}$ is trivial. Every $h\in H^{(1)}$ can be written as $h = 1 + pA$, for some $A \in {\rm Mat}_r({\mathbb Z}/p^n{\mathbb Z})$, so we have $h^{p^{n-1}} \equiv {\rm Id} \mod \it{p^n}$. Hence the subgroup $H^{(1)}$ is a $p$--group and thus it is contained in a $p$--Sylow subgroup $G_p$ of $G$.
Since every $p$--Sylow subgroup of $G$ is contained in $G'$, we have that $H^{(1)}$ is contained in a $p$--Sylow subgroup $G'_p$ of $G'$ too. Thus $H^{(1)}\subseteq \widetilde\Delta_p$, where $\widetilde\Delta_p$ is a $p$--Sylow subgroup of $\widetilde\Delta$. Let $H \subset {\rm GL}_r({\mathbb Z})$ be a subgroup of $\Delta$ such that its image via the reduction modulo $p^n$ is $H^{(1)}$. Consider $\pi:{\rm GL}_r({\mathbb Z}) \longrightarrow {\rm GL}_r({\mathbb Z}/p{\mathbb Z})$ the reduction modulo $p$. We have that $H$ is contained in $\ker(\pi) \cap \Delta$. So, by using Lemma \ref{lemma:inj} we find that $H$ and, consequently, $H^{(1)}$ are trivial. Our claim is proved and we get that also every $H^{(j)}$ is trivial since they are all contained in $H^{(1)}$. Therefore $G \simeq G^{(n-1)} \simeq \dots \simeq G^{(1)}$ via the relative projections.
Consider the following short exact sequence
\[ 1 \longrightarrow T[p]\stackrel{\iota}{\longrightarrow} T[p^n] \stackrel{\varepsilon}{\longrightarrow} T[p^{n-1}] \longrightarrow 1, \]
where $\iota$ is the inclusion and $\varepsilon$ is the $p$--power map (here we are using the multiplicative notation for $T[p]$, $T[p^{n-1}]$ and $T[p^n]$).
The group $G$ acts on $T[p^n]$ and, via the projection, on $T[p]$ and $T[p^{n-1}]$ and by these actions the above short exact sequence is a sequence of $G$--modules. Thus we have the following long exact sequence:
\[ 1 \rightarrow T[p]^{G} \rightarrow T[p^n]^{G} \rightarrow T[p^{n-1}]^{G} \rightarrow {\rm H}^1(G,T[p]) \rightarrow {\rm H}^1(G,T[p^n]) \rightarrow {\rm H}^1(G,T[p^{n-1}]) \rightarrow \dots \]
\noindent Let $C$ be a cyclic subgroup of $G$ and for $i= 1, n-1, n$ let ${\rm res}_i$ be the restriction ${\rm H}^1(G,T[p^i])\longrightarrow {\rm H}^1(C,T[p^i])$. We have the following diagram with exact rows
\begin{figure}[H]
\centering
\begin{tikzpicture}\color{black}
\node(a) at (-3,2) {$\ker({\rm res}_1)$};
\node(b) at (0,2) {$\ker({\rm res}_n)$};
\node(c) at (3,2) {$\ker({\rm res}_{n-1})$};
\node (d) at (-3,0) {${\rm H}^1(G,T[p])$};
\node(e) at (0,0) {${\rm H}^1(G,T[p^n])$};
\node(f) at (3,0) {${\rm H}^1(G,T[p^{n-1}])$};
\node (g) at (-3,-2) {${\rm H}^1(C,T[p])$};
\node (h) at (0,-2) {${\rm H}^1(C,T[p^n])$};
\node(i) at (3,-2) {${\rm H}^1(C,T[p^{n-1}])$};
\node at (-2.3,-1) {${\rm res}_1$};
\node at (0.5,-1) {${\rm res}_n$};
\node at (3.7, -1) {${\rm res}_{n-1}$};
\draw[-stealth] (a) --(b);
\draw[-stealth] (b) --(c);
\draw[-stealth] (a) --(d);
\draw[-stealth] (b) --(e);
\draw[-stealth] (c) --(f);
\draw[-stealth] (d) --(e);
\draw[-stealth] (e) --(f);
\draw[-stealth] (d) --(g);
\draw[-stealth] (e) --(h);
\draw[-stealth] (f) --(i);
\draw[-stealth] (g) --(h);
\draw[-stealth] (h) --(i);
\end{tikzpicture}
\end{figure}
\noindent where the central row is given by the long exact sequence above, the lower row is obtained from the same exact sequence by restriction to the subgroup $C$, while the upper one is induced by the commutativity of the diagram given by the last two rows.
Since the group $G$ is isomorphic to $G^{(1)}$ and to $G^{(n-1)}$, we have that the cyclic subgroups of $G^{(1)}$ and $G^{(n-1)}$ are the images of the projections of the cyclic subgroups of $G$. Therefore, by taking the intersection over all the cyclic subgroups of $G$, from the first row of the diagram we have the exact sequence
\[ {\rm H}^1_{{\rm{loc}}}(G,T[p]) \longrightarrow {\rm H}^1_{{\rm{loc}}}(G,T[p^n]) \longrightarrow {\rm H}^1_{{\rm{loc}}}(G,T[p^{n-1}]). \]
Since by inductive hypothesis $ {\rm H}^1_{\rm{loc}}(G,T[p]) = {\rm H}^1_{{\rm{loc}}}(G,T[p^{n-1}])=0$, we also have ${\rm H}^1_{{\rm{loc}}}(G,T[p^n]) =0$.
\qedhere
\end{proof}
\begin{remark}
We remark that in Lemma \ref{costr toro} we have $k = \mathbb Q$ and $F := k\left(T[p^2]\right) \cap k({\zeta}) = k({\zeta})$, then $[F:k] = [\mathbb Q({\zeta}):\mathbb Q] = p(p-1)$ is divisible by $p$. Moreover, we notice that the automorphism $\eta$ in the proof of Lemma \ref{costr toro} is a generator of the subgroup $H^{(1)} \simeq {\rm Gal}\left(k\left(T[p^2]\right)/k\left(T[p]\right)\right)$, defined in the proof of Theorem \ref{thm:2} above. In particular, for the torus defined in Lemma \ref{costr toro} we have that $H^{(1)}$ is not trivial.
\end{remark}
\begin{remark}
Notice that it is always possible to construct an algebraic torus $T$, not split over $k$, that satisfies the conditions of Theorem \ref{thm:2}. An example is obtained by taking a number field $k$ that contains a $p^n$--th root of unity, any finite extension $L/k$ of degree $d<3(p-1)$ and considering $T = R_{L/k}\mathbb{G}_{m,L}$, the Weil restriction of the split torus $\mathbb{G}_{m,L}$ over $L$. The torus $T$ is defined over $k$, it is split over $L$ and it has dimension $d$. Thus $F = K\cap k({\zeta})$ is equal to $k$ and the hypotheses of the theorem are satisfied.
\end{remark}
|
\section{Introduction}
One of the important features of quantum mechanics is the quantum statistics:
Quantum particles are indistinguishable from one another, and there exist two possible types of particles, \emph{boson} and \emph{fermion},%
\footnote{%
There exists another type of particles, called \emph{anyon}, uniquely in $(2+1)$-dimensional systems.
}
obeying the Bose--Einstein statistics~\cite{Bose:1924} and the Fermi--Dirac statistics~\cite{Fermi:1999ncp,Dirac:1926jz}.
\emph{Supersymmetry} is a symmetry between bosons and fermions.
In the context of supersymmetric quantum field theory, it is formulated as a space-time symmetry based on the superfield formalism,%
\footnote{%
The first appearance of supersymmetry was as an internal symmetry introduced to describe both mesons and baryons in a unified way based on the unitary supergroup~\cite{Miyazawa:1966mfa,Miyazawa:1968zz}.
}
which provides various theoretical frameworks to understand non-perturbative aspects of quantum field theory.
Although it is typically considered as a global symmetry, one may consider a local version of supersymmetry, which inevitably involves gravitational degrees of freedom with supersymmetry, a.k.a., supergravity
In addition to the framework to describe the fundamental interactions in the nature, supersymmetry has so far provided various applications based on its mathematical structure:
In the context of condensed-matter physics, the method of supersymmetry is applied to discuss disorder systems as an alternative approach to the replica trick~\cite{Efetov:1996,Wegner:2016ahw};
The lattice model involving both hopping and spin interactions, called the $t$-$J$ model, realizes supersymmetry by tuning the hopping parameter $t$ and the spin coupling constant $J$~\cite{Sutherland:1975vr,Schlottmann:1987zz};
It has been argued that the critical point of the statistical model with the random field disorder involves emergent supersymmetry (Parisi--Sourlas supersymmetry) together with dimensional reduction~\cite{Parisi:1979ka,Parisi:1982ud};
The critical point of two-dimensional tricritical Ising model is described as a minimal model of super-Virasoro (and also the ordinary Virasoro) algebra of central charge $c = 7/10$~\cite{Friedan:1984rv};
The use of supergroup is also proposed in the context of functional renormalization group to consider a gauge invariant regulator~\cite{Arnone:2000qd}.
The notion of symmetry is well described in terms of group theory, and \emph{supergroup} is an extension of the ordinary groups involving both bosonic and fermionic degrees of freedom.
The main object that we discuss in this article is \emph{supergroup gauge theory}, which is a gauge theory with local supergroup gauge symmetry.
Regarding the particle statistics, Fierz and Pauli formulated a connection with the particle spins, which is nowadays known as the \emph{spin-statistics theorem}.
\begin{itembox}{Spin-statistics theorem~\cite{Fierz:1939,Pauli:1940zz}}
Integer-spin particles are bosons, while odd-half-integer–spin particles are fermions.
\end{itembox}
This theorem is obtained under the following conditions:
\begin{itemize}
\item Lorentz covariance and relativistic causality
\item Positive energies and positive norms in the Hilbert space
\end{itemize}
Therefore, for example, this theorem is not applied to the ghost particle appearing in the gauge fixing process, which is a spin-0 fermionic particle.
The fundamental degrees of freedom of supergroup gauge theory are spin-1 boson and fermion, and thus it is not compatible with the spin-statistics theorem.
In fact, the spectrum of supergroup gauge theory is not bounded and there appear negative energy states.
Even in such a situation, one may still apply the method of Lefschetz thimble to evaluate the path integral~\cite{Witten:2010cx,Cristoforetti:2012su} through the analytic continuation.
Furthermore, it has been also known that $\mathrm{U}(N|M)$ gauge theory is not distinguishable with $\mathrm{U}(N-M|0) = \mathrm{U}(N-M)$ theory to all orders in perturbation theory~\cite{Alvarez-Gaume:1991ozb,Yost:1991ht,Berkovits:1999im,Vafa:2001qf,Vafa:2014iua,Dijkgraaf:2016lym}.
However, instability of vacua implies that the perturbative analysis is not reliable and we need a proper non-perturbative treatment of supergroup gauge theory.
In spite of unphysical natures, as discussed in this article, we can discuss various non-perturbative aspects of supergroup gauge theory as a natural extension of the ordinary gauge theory, and it indicates a chance of non-perturbative completion.
\subsection*{Organization of the article}
A purpose of this article is to provide a self-contained overview of supergroup gauge theory.
For this purpose, we start in Sec.~\ref{sec:supermath} with basic notions of supermathematics, including the introduction of Grassmann algebras, supervector space, superalgebra, supermatrix, and Lie supergroup, which could be skipped by experienced readers on this subject.
In Sec.~\ref{sec:super_matrix}, we discuss the supermatrix model, which can be viewed as a zero-dimensional supergroup gauge theory.
In fact, the supermatrix model plays a role of a toy model in the study of supergroup gauge theory, which exhibits various similar properties to higher-dimensional theory discussed in the latter part of this article.
After introducing the eigenvalue integral form of the partition function, in particular we study the asymptotic behavior based on the Coulomb gas method.
We then study the operator formalism that we call the free field realization of the (super)matrix model, and explore the underlying infinite dimensional algebraic structure.
In Sec.~\ref{sec:supergroup_gauge_theory}, we introduce the supergroup gauge theory and discuss its basic properties.
We discuss realizations of supergroup theory from non-supergroup theory through analytic continuation to unphysical regime, and we show the construction from string/M-theory perspective providing the Seiberg--Witten geometry for supergroup gauge theory.
In Sec.~\ref{sec:instanton}, we explore instantons in supergroup gauge theory, which plays an important role in the study of non-perturbative aspects.
After considering the ADHM construction of instantons, a systematic approach to construct the instanton solutions, for supergroup theory, we study the instanton moduli space and apply the equivariant localization formalism to derive the instanton partition function.
We obtain the instanton partition function in the three-fold way, the equivariant index formula, the contour integral formula, and the combinatorial formula, for supergroup gauge theory.
In Sec.~\ref{sec:non-perturbative_study}, we explore non-perturbative aspects of supergroup gauge theory based on the instanton partition function obtained in advance.
We demonstrate that the instanton partition function is also obtained in the framework of topological string involving both positive and negative branes.
For this purpose, we introduce the negative brane analog of the topological vertex that we call the anti-veretx to compute the supergroup partition function.
We then study the non-perturbative Schwinger--Dyson equation associated with the instanton partition function, which gives rise to doubly quantum Seiberg--Witten geometry, and discuss the free field realization associated with the underlying algebraic structure.
In this case, we show that the instanton partition function obeys a $q$-analog of the Virasoro constraint.
We also explore a connection with the quantum integrable system, called the Bethe/gauge correspondence, for supergroup gauge theory, and discuss its implications on the quantum integrable system side.
We then discuss realizations of supergroup gauge theory in physical setups.
We in particular study the codimension-two surface defect operators and show that the supergroup structure emerges from intersecting defects.
\subsection*{Notations}
For $N \in \mathbb{N}$, we define the set
\begin{align}
[N] = \{ 1,\ldots,N \}
\, .
\label{eq:set_N}
\end{align}
Throughout the article, the vector bundle $\mathbf{X}$ has the Chern character,
\begin{align}
\operatorname{ch} \mathbf{X} = \sum_{i \in [\operatorname{rk} \mathbf{X}]} \mathrm{e}^{x_i}
\, ,
\end{align}
where $\operatorname{rk} \mathbf{X}$ is the rank of the bundle $\mathbf{X}$.
We denote the one-dimensional bundle by $\mathbf{x}$ with $\operatorname{ch} \mathbf{x} = \mathrm{e}^{x}$.
We denote the alternating sum of anti-symmetrizations of the bundle $\mathbf{X}$ by
\begin{align}
\wedge_y \mathbf{X} = \sum_{i =0}^{\operatorname{rk} \mathbf{X}} (-y)^i \wedge^i \mathbf{X}
\, .
\end{align}
In particular, we apply the notation, $\wedge \mathbf{X} = \wedge_1 \mathbf{X}$.
\paragraph{Special functions}
We define the $q$-shifted factorial ($q$-Pochhammer symbol)
\begin{align}
(z;q)_n = \prod_{m \in [n]} (1 - z q^{m-1})
\, ,
\label{eq:q-factorial}
\end{align}
and the theta function with the elliptic norm $q$,
\begin{align}
\theta(z;q) = (z;q)_\infty (qz^{-1};q)_\infty
\, .
\end{align}
\section{Supermathematics}\label{sec:supermath}
In this section, we summarize the basic properties of supermathematics used in this article.
See, e.g.,~\cite{Berezin:1987,DeWitt:1992,Frappat:1996pb,Varadarajan:2004yz,Freund:1986ws,Quella:2013oda,Wegner:2016ahw} for general introductions to supermathematics.
\subsection{Grassmann algebras}
The starting point of supermathematics is the Grassmann algebra, which involves anti-commuting variables (Grassmann variables),
\begin{align}
\theta_i \theta_j = - \theta_j \theta_i
\, .
\end{align}
This anti-commutativity immediately leads to nilpotency of the Grassmann variables,
\begin{align}
\theta_i^2 = 0
\, .
\end{align}
In general, a product of even number of Grassmann variables is commutative (Grassmann even), while a product of odd number of Grassmann variables is anti-commutative (Grassmann odd).
We remark that the Grassmann even variable is commutative, but still nilpotent: For example, we have $(\theta_1 \theta_2) \theta_3 = \theta_3 (\theta_1 \theta_2)$, while $(\theta_1 \theta_2)^2 = 0$.
In the definition of the complex conjugation for the Grassmann variables, we need a modification compared with the ordinary variables:
Applying the conjugation operator twice, we have an extra sign factor,
\begin{align}
\bar{\bar{\theta}} = - \theta
\, .
\label{eq:Grassmann_var_double_conj}
\end{align}
From this definition, the norm of the Grassmann variable becomes self-conjugate,%
\footnote{%
Another convention is also applied in the literature:
No sign factor for the double conjugation $\bar{\bar{\theta}} = \theta$, while the conjugation of the product is given by $\overline{\theta_1 \theta_2 \cdots \theta_n} = \bar{\theta}_n \cdots \bar{\theta}_2 \bar{\theta}_1$.
In this convention, we still have the relation $\overline{\bar{\theta} \theta} = \bar{\theta} \theta$.
}
\begin{align}
\overline{\bar{\theta} \theta} = - \theta \bar{\theta} = \bar{\theta} \theta \, ,
\end{align}
and the Hermitian scalar product behaves as
\begin{align}
\overline{\bar{\theta}_1 \theta_2} = - \theta_1 \bar{\theta}_2 = \bar{\theta}_2 \theta_1 \, .
\end{align}
\subsubsection*{Derivative and integral}
The derivative and the integral for the Grassmann variable are defined as follows,
\begin{align}
\dv{\theta_i}{\theta_j} = \delta_{ij}
\, , \qquad
\int \mathrm{d}{\theta}_i \theta_j = \delta_{ij}
\, , \qquad
\int \mathrm{d}{\theta}_i 1 = 0
\, .
\end{align}
Hence, these operations are essentially equivalent.
Since the Grassmann variables are anti-commutative, we have to be careful of the ordering of the derivative and the integral operation.
Under the linear transformation $\theta_i \to \theta'_i = (A \theta)_i = \sum_j A_{ij} \theta_j$, the integral measure behaves as follows
\begin{align}
\int \prod_i \mathrm{d}{\theta}'_i = \int \prod_{i} \mathrm{d}{\theta}_i (\det A)^{-1}
\, ,
\label{eq:Jacobian_Grassmann}
\end{align}
which is opposite to the measure behavior of the commutative variables involving the determinant factor in the numerator.
\subsection{Supervector space}
A (complex) supervector is an element of the (complex) supervector space, which is a $\mathbb{Z}_2$-graded vector space of dimension $(N|M)$,
\begin{align}
\Psi = (z_i, \theta_j)_{i \in [N], j \in [M]} \in \mathbb{C}^{N|M} = \mathbb{C}_0^N \oplus \mathbb{C}_1^M \, ,
\label{eq:supervec_def}
\end{align}
where the index of $\mathbb{C}_\sigma$, $\sigma = 0,1$ denotes the Grassmann parity.
In general, a supervector space consists of even and odd subspaces,
\begin{align}
V = V_0 \oplus V_1
\, .
\end{align}
We denote the parity of the element $x$ by $|x| = 0$ if $x \in V_0$ and $|x| = 1$ if $x \in V_1$.
We define the parity flip operator $\Pi$, such that $(\Pi V)_0 = V_1$ and $(\Pi V)_1 = V_0$.
The superdimension of $V$ is defined as
\begin{align}
\operatorname{sdim} V = \sum_{\sigma = 0, 1} (-1)^\sigma \operatorname{dim} V_\sigma = \operatorname{dim} V_0 - \operatorname{dim} V_1
\, ,
\label{eq:sdim_def}
\end{align}
which may take a negative integer value.
\subsubsection*{Superalgebra}
Superalgebra is a $\mathbb{Z}_2$-graded algebra that consists of the even part and the odd part,
\begin{align}
\mathfrak{A} = \mathfrak{A}_0 \oplus \mathfrak{A}_1
\, .
\end{align}
We define the supercommutator for $x,y \in \mathfrak{A}$ as follows,
\begin{align}
[x,y] = xy - (-1)^{|x||y|} yx
\, ,
\label{eq:supercommutator_def}
\end{align}
which is skew-symmetric,
\begin{align}
[x,y] = - (-1)^{|x||y|} [y,x]
\, .
\end{align}
Then, we define the Lie superalgebra that obeys the super analog of the Jacobi identity,
\begin{align}
[x,[y,z]] = [[x,y],z] + (-1)^{|x||y|} [y,[x,z]]
\, .
\label{eq:Jacobi_id}
\end{align}
\subsection{Supermatrix}\label{sec:supermatrix}
For supervector spaces $V$ and $W$, we define a linear map, which is described using a supermatrix,
\begin{align}
M & =
\begin{pmatrix}
M_{00} & M_{01} \\ M_{10} & M_{11}
\end{pmatrix}
\in \operatorname{Hom}(V,W) = \bigoplus_{\sigma,\sigma' = 0,1} \operatorname{Hom}(V_{\sigma'},W_\sigma)
\, ,
\label{eq:smat_def}
\end{align}
where each block of the matrix is given by
\begin{align}
M_{\sigma \sigma'} & \in \operatorname{Hom}(V_{\sigma'},W_\sigma)
\, .
\end{align}
Since $M_{01}$ and $M_{10}$ change the Grassmann parity, they consists of the Grassmann odd variables.
In particular, if $V = W$, we have $M \in \operatorname{End}(V) = \bigoplus_{\sigma,\sigma' = 0, 1} \operatorname{Hom}(V_\sigma,V_{\sigma'})$, which is invertible if both $M_{00}$ and $M_{11}$ are invertible.
The set of invertible supermatrices defines a general linear supergroup, $\mathrm{GL}(V) = \mathrm{GL}(V_0|V_1)$.
We remark that the definition of supermatrix is not unique to~\eqref{eq:smat_def}.
For example, denoting even and odd variables by $x_{ij}$ and $\xi_{ij}$, we have the following possibilities for $V = W = \mathbb{C}^{2|1}$,
\begin{align}
M =
\begin{pmatrix}
x_{11} & x_{12} & \xi_{13} \\
x_{21} & x_{22} & \xi_{23} \\
\xi_{31} & \xi_{32} & x_{33}
\end{pmatrix}
\, , \qquad
\begin{pmatrix}
x_{11} & \xi_{12} & x_{13} \\
\xi_{21} & x_{22} & \xi_{23} \\
x_{31} & \xi_{32} & x_{33}
\end{pmatrix}
\, , \qquad
\begin{pmatrix}
x_{11} & \xi_{12} & \xi_{13} \\
\xi_{21} & x_{22} & x_{23} \\
\xi_{31} & x_{32} & x_{33}
\end{pmatrix}
\, ,
\end{align}
which correspond to the convention of supervector, $\Psi = (z_1,z_2,\theta)$, $(z_1,\theta,z_2)$, $(\theta,z_1,z_2)$.
This is related to the ambiguity of the root system in Lie superalgebra.
See a related discussion in Sec.~\ref{sec:Hany-Witten_const}.
\subsubsection*{Supertrace}
We define the supertrace operation for the supermatrix,%
\footnote{%
Not to be confused with the symmetrized trace.
We will introduce one-parameter deformation of the supertrace in~\eqref{eq:supertrace_b}.
}%
\begin{align}
\str M = \sum_{\sigma = 0,1} (-1)^\sigma \tr_\sigma M
\end{align}
where $\tr_\sigma$ means the trace with respect to the subspace $V_\sigma$: $\tr_\sigma M = \tr M_{\sigma\sigma}$.
Compared with the definition of the superdimension~\eqref{eq:sdim_def}, the supertrace of the identity supermatrix provides the superdimension of the corresponding supervector space, $\str \mathbbm{1}_V = \operatorname{sdim} V$.
An important property of the supertrace is the cyclicity,
\begin{align}
\str M_1 M_2 = \str M_2 M_1
\, ,
\label{eq:str_cyclic}
\end{align}
which is analogous to the cyclic property of the ordinary matrix.
\subsubsection*{Supertransposition}
We define the supertransposition operation for the supermatrix,%
\begin{align}
\begin{pmatrix}
A & B \\ C & D
\end{pmatrix}^\text{st}
=
\begin{pmatrix}
A^\text{t} & C^\text{t} \\ -B^\text{t} & D^\text{t}
\end{pmatrix}
\, ,
\end{align}
where we denote the ordinary transposition by $A^\text{t}$, etc.
This supertransposition shows an analogous property to the ordinary case, $(M_1 M_2)^\text{st} = M_2^\text{st} M_1^\text{st}$, while it is not an involution, $(M^\text{st})^\text{st} \neq M$ in general.
The Hermitian conjugation is defined as follows,
\begin{align}
M^\dag = \overline{M}^\text{st}
\, ,
\end{align}
which is then an involution, $(M^\dag)^\dag = M$.
\subsubsection*{Superdeterminant}
We define the superdeterminant, which is also called the Berezinian,
\begin{align}
\sdet \begin{pmatrix}
A & B \\ C & D
\end{pmatrix}
= \frac{\det (A - B D^{-1} C)}{\det D}
= \frac{\det A}{\det (D - C A^{-1} B)}
\, .
\end{align}
This is analogous to the determinant formula for the block matrix,
\begin{align}
\det \begin{pmatrix}
A & B \\ C & D
\end{pmatrix}
= \det (A - B D^{-1} C) \det D
= \det A \det (D - C A^{-1} B)
\, .
\end{align}
We have the multiplicative property for the superdeterminant, $\sdet (M_1 M_2) = \sdet M_1 \sdet M_2$.
We remark an identity
\begin{align}
\str \log M = \log \str M
\, .
\end{align}
Recalling the behavior of the Grassmann variable measure~\eqref{eq:Jacobian_Grassmann}, the integral measure denoted by $\mathrm{d}{\Psi} = \mathrm{d}{z}_1 \cdots \mathrm{d}{z}_N \mathrm{d}{\theta}_1 \cdots \mathrm{d}{\theta}_M$ behaves under the linear transform for the supervector~\eqref{eq:supervec_def}, $\Psi' = M \Psi$, as follows,
\begin{align}
\mathrm{d}{\Psi'} = \mathrm{d}{\Psi} (\sdet M)
\, .
\end{align}
\subsection{Supergroup}
\subsubsection*{Unitary supergroup}
For the complex supervector space element \eqref{eq:supervec_def}, $\Psi \in \mathbb{C}^{N|M}$, we consider the squared norm as follows,
\begin{align}
|\Psi|^2 & = \sum_{i \in [N]} |z_i|^2 + \sum_{j \in [M]} \bar{\theta}_j \theta_j
= \sum_{i \in [N]} |z_i|^2 - \sum_{j \in [M]} \theta_j \bar{\theta}_j
= \str (\Psi \Psi^\dag)
\, .
\end{align}
Then, we define the unitary supergroup $\mathrm{U}(N|M)$ as the isometry group with respect to the supervector space $\mathbb{C}^{N|M}$, such that,
\begin{align}
|\Psi|^2 = |U \Psi|^2
\, , \qquad
U^\dag = U^{-1}
\, .
\end{align}
We remark that $|U \Psi|^2 = \str (U \Psi \Psi^\dag U^\dag) = \str (\Psi \Psi^\dag)$.
Hence, we have
\begin{align}
\mathrm{U}(N|M) = \{ U \in \mathrm{GL}(\mathbb{C}^{N|M}) \mid U^\dag = U^{-1} \}
\, .
\end{align}
The even part of the unitary supergroup $\mathrm{U}(N|M)$ is given by $\mathrm{U}(N) \times \mathrm{U}(M)$, and the odd part is given by $N \times \overline{M}$ and $\overline{N} \times M$ representations of the even part with the dimension $2NM$.
\subsubsection*{Orthosymplectic supergroup}
We consider the real supervector space $\mathbb{R}^{N|2M}$ and the squared norm of $\Psi=(x_1,\ldots,x_N,\theta_1,\ldots,\theta_{2M}) \in \mathbb{R}^{N|2M}$ as follows,
\begin{align}
|\Psi| = \sum_{i \in [N]} x_i^2 + 2 \sum_{j \in [M]} \theta_{2j-1} \theta_{2j} = \Psi^\text{t} \Omega \Psi = \str \left( \Omega \Psi \Psi^\text{t} \right)
\, ,
\end{align}
where we define the skew-symmetric form for the Grassmann odd sector,
\begin{align}
\Omega =
\begin{pmatrix}
\mathbbm{1}_N & 0 \\ 0 & \mathbbm{1}_M \otimes \mathbf{j}
\end{pmatrix}
\, , \qquad
\mathbf{j} =
\begin{pmatrix}
0 & 1 \\ -1 & 0
\end{pmatrix}
\, .
\end{align}
We remark that the standard bilinear form does not work for the Grassmann variables due to its nilpotency, $\sum_{j \in [M]} \theta_j \theta_j = 0$.
The isometry group for the real supervector space $\mathbb{R}^{N|2M}$ is given by
\begin{align}
|\Psi|^2 = |U \Psi|^2
\, , \qquad
U^\text{st}
\Omega
U = \Omega
\, .
\end{align}
We see that $|U \Psi|^2 = \str(\Omega U \Psi \Psi^\text{t} U^\text{st}) = \str(\Omega \Psi \Psi^\text{t})$.
The orthosymplectic supergroup for general field $\mathbb{K}$ is given as follows,
\begin{align}
\mathrm{OSp}(N|2M,\mathbb{K}) =
\{ U \in \mathrm{GL}(\mathbb{K}^{N|2M}) \mid U^\text{st} \Omega U = \Omega \}
\, .
\end{align}
Precisely speaking, the orthosymplectic group realized as the isometry group of the real supervector space is a compact supergroup, which is in fact a subgroup of the unitary supergroup, denoted by $\mathrm{UOSp}(N|M) = \mathrm{OSp}(N|2M,\mathbb{C}) \cap \mathrm{U}(N|2M)$.
This is analogous to the notation of the compact symplectic group $\mathrm{USp}(2n) = \mathrm{Sp}(2n,\mathbb{C}) \cap \mathrm{U}(2n)$, which is also understood as the quaternionic unitary group $\mathrm{U}(n,\mathbb{H})$.
In this article, we use the notation, $\mathrm{OSp}(N|M) = \mathrm{UOSp}(N|2M)$ and $\mathrm{Sp}(n) = \mathrm{USp}(2n)$ unless it causes confusion.%
\footnote{%
In this notation, we have the isomorphisms at the level of Lie algebra, $\mathfrak{sp}_1 = \mathfrak{su}_2$, $\mathfrak{sp}_2 = \mathfrak{so}_5$.
}
The even part of $\mathrm{OSp}(N|M)$ is thus given by $\mathrm{O}(N) \times \mathrm{Sp}(M)$, and the odd part is given by $N \times 2M$ representation of the even part with the dimension $2NM$.
Another situation that leads to the orthosymplectic supergroup is a subsector of the supervector space $\mathbb{C}^{2N|2M}$,
\begin{align}
\Psi =
\begin{pmatrix}
z_1 & z_2 \\
- \bar{z}_2 & \bar{z}_1 \\
z_3 & z_4 \\
\vdots & \vdots \\
- \bar{z}_{2N} & \bar{z}_{2N-1} \\
\theta_1 & \bar{\theta}_1 \\
\vdots & \vdots \\
\theta_M & \bar{\theta}_M
\end{pmatrix}
\, , \qquad
\Psi^\dag =
\begin{pmatrix}
\bar{z}_1 & - z_2 & \bar{z}_3 & \cdots & - z_{2N} & \bar{\theta}_1 & \cdots & \bar{\theta}_M \\
\bar{z}_2 & z_1 & \bar{z}_4 & \cdots & z_{2N-1} & - \theta_1 & \cdots & - \theta_{M}
\end{pmatrix}
\end{align}
where the two-by-two block in the bosonic part is identified with a quaternion
\begin{align}
x_i =
\begin{pmatrix}
z_{2i-1} & z_{2i} \\ - \bar{z}_{2i} & \bar{z}_{2i-1}
\end{pmatrix}
\in \mathbb{H}
\, .
\end{align}
Hence, this is an element in $\mathbb{H}^N \oplus \mathbb{R}^{2M} \subset \mathbb{C}^{2N|2M}$.
We denote the norm of $x \in \mathbb{H}$ by $|x|$, such that $\bar{x} x = |x|^2 \mathbbm{1}$ where $\mathbbm{1}$ is the quaternion identity.
The norm of this supervector is given by
\begin{align}
|\Psi|^2 = \tr_{\mathbb{H}} \Psi^\dag \Psi = 2 \sum_{i \in [N]} |x_i|^2 + 2 \sum_{j \in [M]} \bar{\theta}_j \theta_j = \str (\Psi \Psi^\dag)
\, ,
\end{align}
where we denote the trace operation over the quaternion by $\tr_\mathbb{H}$, i.e., for the quaternion units, $\mathbbm{1}$, $\mathbf{i}$, $\mathbf{j}$, $\mathbf{k}$, we have $\tr_\mathbb{H} \mathbbm{1} = 2$, $\tr_\mathbb{H} \mathbf{i} = \tr_\mathbb{H} \mathbf{j} = \tr_\mathbb{H} \mathbf{k} = 0$.
The isometry group of this supervector is thus given by the orthosymplectic supergroup $\mathrm{U}(2N|2M) \supset \mathrm{OSp}(M|N) \supset \mathrm{Sp}(N) \times \mathrm{O}(M)$.
\subsubsection*{Analytic continuation}
We discuss how to obtain the supergroups considered above through the analytic continuation.
First of all, the dimension of each classical group is given by%
\footnote{%
The orthogonal group is further classified by the parity of the rank, $\dim \mathrm{O}(2N) = N(2N-1)$, $\dim \mathrm{O}(2N+1) = N(2N+1)$.
We remark that $\dim \mathrm{O}(2N+1) = \dim \mathrm{Sp}(N)$.
This dimension formula is also understood from the construction of the adjoint representations for these classical groups~\eqref{eq:O_SP_adj_rep}.
}
\begin{align}
\dim \mathrm{U}(N) = N^2
\, , \qquad
\dim \mathrm{O}(N) = \frac{N(N-1)}{2}
\, , \qquad
\dim \mathrm{Sp}(N) = N(2N+1)
\, ,
\end{align}
from which we observe the following relations,
\begin{align}
\dim \mathrm{U}(-N) = \dim \mathrm{U}(N)
\, , \qquad
\dim \mathrm{O}(-2N) = \dim \mathrm{Sp}(N)
\, .
\end{align}
Such a relation is discussed also at the level of their irreducible representations~\cite{King:1971CJM}.
From this point of view, these supergroups are obtained from the ordinary classical groups through the analytic continuation,
\begin{subequations}
\begin{gather}
\mathrm{U}(N+M)
\ \stackrel{M \leftrightarrow -M}{\longleftrightarrow} \
\mathrm{U}(N|M)
\, , \\
\mathrm{O}(N+2M)
\ \stackrel{M \leftrightarrow -M}{\longleftrightarrow} \
\mathrm{OSp}(N|M)
\, , \qquad
\mathrm{Sp}(N+M)
\ \stackrel{M \leftrightarrow -M}{\longleftrightarrow} \
\mathrm{OSp}(2M|N)
\, .
\end{gather}
\end{subequations}
In particular, we have the relations between the superdimension and the dimension of the classical groups,
\begin{subequations}
\begin{align}
\operatorname{sdim} \mathrm{U}(N|M) & = (N^2 + M^2) - 2 NM = (N-M)^2 = \operatorname{dim} \mathrm{U}(N+M)\Big|_{M \to -M}
\, , \\
\operatorname{sdim} \mathrm{OSp}(N|M) & = \left( \frac{N(N-1)}{2} + M(2M+1) \right) - 2 NM = 2 \left( \frac{N}{2} - M \right)^2 - \left( \frac{N}{2} - M \right)
\nonumber \\
& = \operatorname{dim} \mathrm{O}(N + 2M)\Big|_{M \to -M} = \operatorname{dim} \mathrm{Sp}\left(\frac{N}{2} + M\right)\Bigg|_{N \to -N}
\, .
\end{align}
\end{subequations}
This interpretation based on the analytic continuation seems also reasonable from the relation between the ordinary trace and the supertrace.
\section{Supermatrix model}\label{sec:super_matrix}
The matrix model, just given by a matrix integral, is thought of as a zero-dimensional reduction of quantum field theory.
It sounds rather a simple toy model, but it has been playing a role to understand various non-perturbative aspects of quantum field theory.
In this section, we explore the supermatrix model as a toy model that exhibits supergroup symmetry.
See also \cite{Efetov:1996,Guhr:2010nh} for details on this subject.
\subsection{Hermitian supermatrix model: superunitary ensemble}\label{sec:Hermitian_supermatrix_model}
Let $H$ be an $(N|M)$-dimensional Hermitian supermatrix, $H^\dag = H$.
We define the partition function of Hermitian supermatrix model of rank $(N|M)$ as follows~\cite{Alvarez-Gaume:1991ozb,Yost:1991ht},
\begin{align}
Z_{N|M} = \int \mathrm{d}{H} \mathrm{e}^{-\frac{1}{g} \str V(H)}
\, ,
\end{align}
with a polynomial potential function of degree $(d+1)$,
\begin{align}
V(x) = \sum_{n = 1}^{d+1} \frac{t_n}{n} \, x^n
\, .
\label{eq:pot_fn_supermatrix}
\end{align}
This integral is invariant under superunitary transformation, $H \to U H U^\dag$, $U \in \mathrm{U}(N|M)$:
The potential part is invariant due to the cyclic property of the supertrace~\eqref{eq:str_cyclic}, $\str V(U H U^\dag) = \str V(H)$, and the measure part is due to the unitarity of the superdeterminant $|\sdet U| = 1$.
From this point of view, this model is also called the \emph{superunitary ensemble} as a supermatrix generalization of the unitary ensemble of the ordinary random matrices.
See, e.g.,~\cite{Mehta:2004RMT,Forrester:2010,Eynard:2015aea}.
In fact, this symmetry is interpreted as a remnant of the supergroup gauge symmetry.
Similarly to the ordinary Hermitian matrix, we can diagonalize the Hermitian supermatrix via the superunitary transform,%
\footnote{%
Eigenvalues of Hermitian supermatrices are ordinary commutative numbers (Recall that the diagonal blocks of supermatrices are ordinary matrices).
The matrix model involving Grassmann odd eigenvalues is known as the super-eigenvalue model~\cite{Alvarez-Gaume:1991vno}.
}
\begin{align}
H = U Z U^\dag
\, , \qquad
U \in \mathrm{U}(N|M)
\, ,
\label{eq:SU_diaginalization}
\end{align}
where we denote the diagonal supermatrix by
\begin{align}
Z =
\begin{pmatrix}
X & 0 \\ 0 & Y
\end{pmatrix}
\, , \qquad
X = \operatorname{diag}(x_1,\ldots,x_N)
\, , \
Y = \operatorname{diag}(y_1,\ldots,y_M)
\, .
\label{eq:diagonal_Z}
\end{align}
We remark that the choice of $U$ is not unique in the process of diagonalization: (i) The eigenvalues can be permuted $(x_i,y_j)_{i\in[N], j\in[M]} \to (x_{\sigma(i)},y_{\sigma'(j)})_{i\in[N], j\in[M]}$ where $\sigma \in \mathfrak{S}_N$, $\sigma' \in \mathfrak{S}_M$.
(ii) Since the diagonal superunitary matrix commutes with the diagonal supermatrix, $U_\text{diag} Z U_\text{diag}^\dag = Z$, where $U_\text{diag} \in \mathrm{U}(1)^{N+M}$, the decomposition is invariant under redefinition $U \to U U_\text{diag}$.
Taking the derivative of the relation above, we have the following expression,
\begin{align}
U^\dag \mathrm{d}{H} U = \mathrm{d}{Z} + [U^\dag\mathrm{d}{U}, Z]
\, .
\label{eq:UdHU}
\end{align}
We remark that $U^\dag\mathrm{d}{U}$ is the Maurer--Cartan one-form with respect to the supergroup $\mathrm{U}(N|M)$, and the Hermitian supermatrix takes a value in the Lie superalgebra $\operatorname{Lie} \mathrm{U}(N|M)$ (up to the imaginary unit).
Then, the right hand side of this equation can be also written using the covariant derivative in the adjoint representation, $D = \mathrm{d}{} + U^\dag\mathrm{d}{U}$.
Denoting $I_0 = \{1,\ldots,N\}$ and $I_1 = \{ N+1, \ldots, N+M \}$, and $\mathsf{i}(i) = i$ for $i \in I_0$ and $\mathsf{i}(i) = i - N$ for $i \in I_1$, each component of \eqref{eq:UdHU} is given by
\begin{subequations}\label{eq:UdHU_component}
\begin{align}
(U^\dag \mathrm{d}{H} U)_{ii} & =
\begin{cases}
\mathrm{d}{x}_{\mathsf{i}(i)} & (i \in I_0) \\
\mathrm{d}{y}_{\mathsf{i}(i)} & (i \in I_1)
\end{cases}
\\
(U^\dag \mathrm{d}{H} U)_{ij} & = (U^\dag \mathrm{d}{U})_{ij} \times
\begin{cases}
x_{\mathsf{i}(j)} - x_{\mathsf{i}(i)} & (i, j \in I_0) \\
y_{\mathsf{i}(j)} - y_{\mathsf{i}(i)} & (i, j \in I_1) \\
y_{\mathsf{i}(j)} - x_{\mathsf{i}(i)} & (i \in I_0, j \in I_1) \\
x_{\mathsf{i}(j)} - y_{\mathsf{i}(i)} & (i \in I_1, j \in I_0)
\end{cases}
\end{align}
\end{subequations}
Hence, the supermatrix measure is given in terms of the eigenvalues and the eigenvector components as follows,
\begin{align}
\mathrm{d}{H} = |\Delta_{N|M}(X|Y)|^2 \mathrm{d}{X} \mathrm{d}{Y} \mathrm{d}{U}
\end{align}
where we define
\begin{align}
\mathrm{d}{X} = \prod_{i \in [N]} \mathrm{d}{x}_i
\, , \qquad
\mathrm{d}{Y} = \prod_{i \in [M]} \mathrm{d}{y}_i
\, , \qquad
\mathrm{d}{U} = \prod_{1 \le i<j \le N+M} \operatorname{Re} (U^\dag \mathrm{d}{U})_{ij} \operatorname{Im} (U^\dag \mathrm{d}{U})_{ij}
\, .
\end{align}
In particular, $\mathrm{d}{U}$ is the Haar measure on the unitary supergroup $\mathrm{U}(N|M)$.
The Jacobian part is given by the Cauchy determinant
\begin{align}
\Delta_{N|M}(X|Y) = \prod_{i < j}^N (x_j - x_i) \prod_{i < j}^M (y_j - y_i) \prod_{i \in [N], j \in [M]} (x_i - y_j)^{-1}
\, ,
\end{align}
which has a determinantal formula for $N \ge M$ as
\begin{align}
\Delta_{N|M}(X|Y) = \det_{\substack{i \in [N],\, j \in [M] \\ k \in [N-M]}}
\begin{pmatrix}
x_i^{k-1} & \displaystyle \frac{1}{x_i - y_j}
\end{pmatrix}
\, .
\end{align}
In the limit $M \to 0$, this is reduced to the Vandermonde determinant,
\begin{align}
\Delta_{N|0}(X) = \Delta_N(X) = \prod_{i<j}^N (x_j - x_i) \, .
\end{align}
Based on these diagonalization process, we obtain the eigenvalue integral form of the partition function,
\begin{align}
{Z}_{N|M} = \frac{\operatorname{vol}(\mathrm{U}(N|M))}{N! M!} \int \mathrm{d}{\mu(X)} \mathrm{d}{\mu(Y)}
|\Delta_{N|M}(X|Y)|^2
\, ,
\end{align}
where the integral measure is given by
\begin{align}
\mathrm{d}{\mu(X)} = \prod_{i \in [N]} \frac{\mathrm{d}{x}_i}{2 \pi} \mathrm{e}^{-\frac{1}{g}V(x_i)}
\, , \qquad
\mathrm{d}{\mu(Y)} = \prod_{i \in [M]} \frac{\mathrm{d}{y}_i}{2 \pi} \mathrm{e}^{+\frac{1}{g}V(y_i)}
\, .
\end{align}
Constant factors are understood as follows:
The factorial terms are the volumes of $\mathfrak{S}_N$ and $\mathfrak{S}_M$, which are the Weyl group of $\mathrm{U}(N)$ and $\mathrm{U}(M)$, and $(2\pi)^{N+M} = \operatorname{vol}(\mathrm{U}(1)^{N+M})$, the volume of the maximal Cartan torus of $\mathrm{U}(N|M)$.
We have several remarks on this formula:
\begin{enumerate}
\item The signatures of the potential term for the $x$-variables and the $y$-variables are opposite.
Hence, we should consider a complex contour to obtain a converging integral.
For example, for the Gaussian case $V(x) = \frac{1}{2} x^2$, the $x$-integral is taken along the real axis, $- \infty \to + \infty$, while the $y$-integral should be taken along the imaginary axis, $- \mathrm{i} \infty \to + \mathrm{i} \infty$, or vice versa.
\item The denominator contribution in the Cauchy determinant is singular in the limit $x_i \to y_j$.
If there is an intersection of the $x$-contour and the $y$-contour, such a singularity should be regularized using the principal value prescription.
\item The volume of the unitary supergroup becomes zero if $NM \neq 0$ due to the Grassmann variable integral (Berezin's theorem. See e.g., \cite{Voronov:2015cya}).
Hence, we consider the partition function formally normalized by this zero-volume, $\mathcal{Z}_{N|M} := Z_{N|M}/\operatorname{vol}(\mathrm{U}(N|M))$.%
\footnote{%
It would be possible that the zero-volume factor cancels the diverging behavior of the eigenvalue integral to give a finite value in the end.
We do not discuss details of this issue any further in this article.
}
\end{enumerate}
Taking into account these points, we obtain the eigenvalue integral form of the supermatrix partition function.
\begin{itembox}{Hermitian supermatrix model}
The eigenvalue integral form of the regularized partition function of the Hermitian supermatrix model is given as follows,
\begin{align}
\mathcal{Z}_{N|M} =
\frac{1}{N! M!} \Xint-_{\gamma_x^N \times \gamma_y^M} \mathrm{d}{\mu(X)} \mathrm{d}{\mu(Y)}
|\Delta_{N|M}(X|Y)|^2
\, ,
\end{align}
where we denote the principal value integral by $\Xint- \mathrm{d}{x} f(x)$, and $\gamma_x$ and $\gamma_y$ are the integration contours on the complex plane that provide a converging integral.
\end{itembox}
\subsection{Real-quaternion supermatrix model: super orthosymplectic ensemble}\label{sec:real-quaternion_supermatrix_model}
We consider an $(N|M)$-dimensional real-quaternion self-conjugate supermatrix,
\begin{align}
H =
\begin{pmatrix}
A & B \\ C & D
\end{pmatrix}
\end{align}
where $A$ is an $N$-dimensional real symmetric matrix, and $D$ is an $M$-dimensional quaternion self-dual matrix (realized as a $2M$-dimensional Hermitian matrix).
$B$ is a real Grassmann matrix of size $N \times 2M$ and, in order that $H$ is self-conjugate $H^\dag = H$, we have $C = - B^\text{t}$.
In this case, similarly to the Hermitian case~\eqref{eq:SU_diaginalization}, we can diagonalize the supermatrix via the orthosymplectic transformation,
\begin{align}
H = U Z U^\dag
\, , \qquad
U \in \mathrm{OSp}(N|M)
\, .
\label{eq:OSp_diaginalization}
\end{align}
The diagonal supermatrix $Z$ is given by
\begin{align}
Z =
\begin{pmatrix}
X & 0 \\ 0 & Y \otimes \mathbbm{1}
\end{pmatrix}
\, ,
\end{align}
where we denote the identity element in quaternion by $\mathbbm{1}$.
Namely, if we use the two-by-two matrix realization of quaternion, it is given by the identity matrix of rank two.
Hence, the supertrace is in this case given by
\begin{align}
\str Z = \sum_{i \in [N]} x_i - 2 \sum_{j \in [M]} y_j
\, .
\end{align}
We should be careful of the multiplicity in the quaternionic sector, which corresponds to the Kramers doublet.
We will define a deformed supertrace operation respecting this multiplicity (see \eqref{eq:supertrace_b}).
Applying the same argument to the unitary case, we have the relation~\eqref{eq:UdHU} where the corresponding components are given as in~\eqref{eq:UdHU_component}.
In this case, we should be careful of that the $y$-variables are doubly degenerated.
Hence, we should count twice for the mixing terms $y_j - x_i$ and four ($= 2 \times 2$) times for the quaternion part $y_j - y_i$.
Therefore, the real-quaternion supermatrix measure is given as follows,
\begin{align}
\mathrm{d}{H} = |\Delta_{N|M}^{(1|4)}(X|Y)| \mathrm{d}{X} \mathrm{d}{Y} \mathrm{d}{U}
\end{align}
where we define
\begin{align}
\mathrm{d}{X} = \prod_{i \in [N]} \mathrm{d}{x}_i
\, , \qquad
\mathrm{d}{Y} = \prod_{i \in [M]} \mathrm{d}{y}_i
\, , \qquad
\end{align}
and the corresponding Haar measure $\mathrm{d}{U}$ of supergroup $\mathrm{OSp}(N|M)$.
We use the following notation for the Jacobian part,
\begin{align}
\Delta_{N|M}^{(\beta|\beta')}(X|Y) = \frac{\Delta_N(X)^\beta \Delta_M(Y)^{\beta'}}{\prod_{i \in [N], j \in [M]} (x_i - y_j)^2}
\end{align}
In this notation, we have $\Delta_{N|M}(X|Y)^2 = \Delta_{N|M}^{(2|2)}(X|Y)$.
Collecting all the contributions, we obtain the real-quaternion supermatrix.
\begin{itembox}{Real-quaternion supermatrix model}
The eigenvalue integral form of the regularized partition function of real-quaternion supermatrix model is given as follows,
\begin{align}
\mathcal{Z}_{N|M} := \frac{Z_{N|M}}{\operatorname{vol}(\mathrm{OSp}(N|M))} =
\frac{1}{N! M!} \Xint-_{\gamma_x^N \times \gamma_y^M} \mathrm{d}{\mu(X)} \mathrm{d}{\mu(Y)}
|\Delta_{N|M}^{(1|4)}(X|Y)|
\, .
\end{align}
\end{itembox}
We study several aspects of the supermatrix models in the following part.
\subsection{Coulomb gas analysis}
In the context of matrix model, we are in particular interested in the asymptotic behavior in the large size limit of the matrix model.
We study such an asymptotic limit of the supermatrix model based on the Coulomb gas analysis.
See also~\cite{Babinet:2022} for details in this part.
We start with the partition function of the $\beta$-deformed supermatrix model,
\begin{align}
\mathcal{Z}_{N|M} = \frac{1}{N! M!} \int \mathrm{d}{\mu(X)} \mathrm{d}{\mu(Y)}
|\Delta_{N|M}^{(\beta|\beta')}(X|Y)|
\, , \qquad
\beta \beta' = 4
\, ,
\end{align}
where the integral measure is given by
\begin{align}
\mathrm{d}{\mu(X)} = \prod_{i \in [N]} \frac{\mathrm{d}{x}_i}{2 \pi} \mathrm{e}^{-\frac{b}{g} V(x_i)}
\, , \qquad
\mathrm{d}{\mu(Y)} = \prod_{i \in [M]} \frac{\mathrm{d}{y}_i}{2 \pi} \mathrm{e}^{+\frac{b^{-1}}{g}V(y_i)}
\, , \qquad
b = \sqrt{\frac{\beta}{2}}
\, .
\end{align}
In this notation, the measure factor is given by
\begin{align}
\Delta_{N|M}^{(\beta|\beta')}(X|Y) = \frac{\Delta_N(X)^{2 b^2} \Delta_M(Y)^{2 b^{-2}}}{\prod_{i \in [N], j \in [M]} (x_i - y_j)^2}
\, .
\end{align}
For the $\beta$-deformed case, it is convenient to define the $b$-deformed supertrace,
\begin{align}
\str_b
\begin{pmatrix}
A & B \\ C & D
\end{pmatrix}
= b \tr_0 A - b^{-1} \tr_1 D
\, .
\label{eq:supertrace_b}
\end{align}
With this operation, the potential factor is concisely written as
\begin{align}
\str_b V(Z) = b \sum_{i \in [N]} V(x_i) - b^{-1} \sum_{j \in [M]} V(y_j)
\, ,
\end{align}
where the diagonal supermatrix $Z$ is given as~\eqref{eq:diagonal_Z}.
Such a deformation is discussed in the context of symmetric polynomial associated with the Lie superalgebra root system.
See, e.g.,~\cite{Sergeev:2001JNMP,Sergeev:2002TMP,Sergeev:2005AM,Hallnas:2009CA,Desrosiers:2012S}, and also discussions in Secs.~\ref{sec:Bethe/gauge} and \ref{sec:intersecting_defects}.
\subsubsection{Saddle point equation}
To study the asymptotic behavior of the supermatrix model, we rewrite the partition function in the following form,
\begin{align}
\mathcal{Z}_{N|M} \approx \int \mathrm{d}{X} \mathrm{d}{Y} \mathrm{e}^{-\frac{1}{g^2} S(X|Y)}
\end{align}
where the integral measure is given by
\begin{align}
\mathrm{d}{X} = \prod_{i \in [N]} \mathrm{d}{x}_i
\, , \qquad
\mathrm{d}{Y} = \prod_{j \in [M]} \mathrm{d}{y}_j
\, ,
\end{align}
and we define the effective action
\begin{align}
S(X|Y) & =
b g \sum_{i \in [N]} V(x_i) - b^{-1} g \sum_{j \in [M]} V(y_j)
\nonumber \\
& \quad
- 2 b^2 g^2 \sum_{i < j}^N \log (x_i - x_j)
- 2 b^{-2} g^2 \sum_{i < j}^M \log (y_i - y_j)
+ 2 g^2 \sum_{i \in [N], j \in [M]} \log (x_i - y_j)
\, .
\label{eq:eff_action_supermatrix}
\end{align}
Then, introducing two parameters,
\begin{align}
\mathsf{t}_0 = bgN
\, , \qquad
\mathsf{t}_1 = b^{-1}gM
\, ,
\end{align}
we consider the following asymptotic limit ('t Hooft limit) of the supermatrix model,
\begin{align}
g \ \longrightarrow \ 0
\, , \qquad
N, M \ \longrightarrow \ \infty
\, , \qquad
\mathsf{t}_0 , \mathsf{t}_1 = O(1)
\, .
\end{align}
In the 't Hooft limit, the eigenvalue integral localizes on the configuration that obeys the following saddle point equations,
\begin{subequations}\label{eq:saddle_pt_supermatrix}
\begin{align}
0 & = \frac{\partial S}{\partial x_i} =
+ b g V'(x_i)
- 2 b^2 g^2 \sum_{j \in [N]\backslash\{i\}} \frac{1}{x_i - x_j}
+ 2 g^2 \sum_{j \in [M]} \frac{1}{x_i - y_j}
\, , \\
0 & = \frac{\partial S}{\partial y_i} =
- b^{-1} g V'(y_i)
- 2 b^{-2} g^2 \sum_{j \in [M]\backslash\{i\}} \frac{1}{y_i - y_j}
+ 2 g^2 \sum_{j \in [N]} \frac{1}{y_i - x_j}
\, .
\end{align}
\end{subequations}
We introduce the auxiliary functions,
\begin{subequations}
\begin{align}
W_0 (x) & = bg \sum_{i \in [N]} \frac{1}{x - x_i} = bg \tr_0 \left( \frac{1}{x - X} \right)
\, , \\
W_1 (x) & = b^{-1}g \sum_{i \in [M]} \frac{1}{x - y_i} = b^{-1} g \tr_1 \left( \frac{1}{x - Y} \right)
\, , \\
P_0 (x) & = bg \sum_{i \in [N]} \frac{V'(x) - V'(x_i)}{x - x_i} = bg \tr_0 \left( \frac{V'(x) - V'(X)}{x - X} \right)
\, , \\
P_1 (x) & = b^{-1}g \sum_{i \in [M]} \frac{V'(x) - V'(y_i)}{x - y_i} = b^{-1}g \tr_1 \left( \frac{V'(x) - V'(Y)}{x - Y} \right)
\, .
\end{align}
\end{subequations}
The auxiliary functions $W_\sigma(x)$ are in particular called the resolvents that involve a pole singularity at $x \in \{x_i\}_{i \in [N]}$ and $x \in \{ y_j \}_{j \in [M]}$, respectively.
Although the other functions $P_\sigma(x)$ look a similar form, they are polynomial functions having no pole singularity.
Recalling that the potential function is given as~\eqref{eq:pot_fn_supermatrix}, the asymptotic behaviors of these auxiliary functions are given by
\begin{align}
W_\sigma(x) \ \xrightarrow{x \to \infty} \ \frac{\mathsf{t}_\sigma}{x}
\, , \qquad
P_\sigma(x) \ \xrightarrow{x \to \infty} \ \mathsf{t}_\sigma t_{d+1} x^{d-1}
\, .
\end{align}
Using these auxiliary functions, we may rewrite the saddle point equation~\eqref{eq:saddle_pt_supermatrix} as follows,
\begin{subequations}
\begin{align}
0 & = - P_0(x) + V'(x) W_0(x) - W_0(x)^2 - bg W_0'(x) + 2 g^2 \sum_{i \in [N], j \in [M]} \frac{1}{(x - x_i)(x_i - y_j)}
\, , \\
0 & = + P_1(x) - V'(x) W_1(x) - W_1(x)^2 - b^{-1}g W_1'(x) + 2 g^2 \sum_{i \in [N], j \in [M]} \frac{1}{(x - y_j)(y_j - x_i)}
\, .
\end{align}
\end{subequations}
Moreover, we define the supertrace analog of the auxiliary functions,
\begin{subequations}
\begin{align}
\mathsf{W}(x) & = W_0(x) - W_1(x) = g \str_b \left( \frac{1}{x - Z} \right)
\, , \\
\mathsf{P}(x) & = P_0(x) - P_1(x) = g \str_b \left( \frac{V'(x) - V'(Z)}{x - Z} \right)
\, ,
\end{align}
\end{subequations}
which show the following asymptotic behavior,
\begin{align}
\mathsf{W}(x) \ \xrightarrow{x \to \infty} \ \frac{\mathsf{t}_0 - \mathsf{t}_1}{x}
\, , \qquad
\mathsf{P}(x) \ \xrightarrow{x \to \infty} \ (\mathsf{t}_0 - \mathsf{t}_1) t_{d+1} x^{d-1}
\, .
\end{align}
The total resolvent $\mathsf{W}(x)$, that we call the superresolvent, has poles with the residue $+bg$ for $x \in \{x_i\}$ and $-b^{-1}g$ for $x \in \{y_i\}$, while $\mathsf{P}(x)$ is again a polynomial function.
Then, combining the two equations, we obtain
\begin{align}
0 & = \mathsf{P}(x) - V'(x) \mathsf{W}(x) + \mathsf{W}(x)^2 + g (b W_0'(x) + b^{-1} W_1'(x))
\, .
\label{eq:saddle_pt_supermatrix_finite}
\end{align}
We study this equation in detail in Sec.~\ref{sec:sp_curve}.
We remark that the two saddle point equations \eqref{eq:saddle_pt_supermatrix} are written as a single equation using the superresolvent,
\begin{align}
0 = V'(x) - 2 \mathsf{W}_\text{reg}(x)
\, , \qquad
x \in \{ x_i, y_j \}
\, ,
\label{eq:saddle_pt_supermatrix_single_form}
\end{align}
where we define the regularized one by
\begin{align}
\mathsf{W}_\text{reg}(x) =
\begin{cases}
\displaystyle
bg \sum_{j \in [N]\backslash\{i\}} \frac{1}{x - x_j} - W_1(x) & (x = x_i) \\
\displaystyle
W_0(x) - b^{-1} g \sum_{j \in [M]\backslash\{i\}} \frac{1}{x - y_j} - W_1(x) & (x = y_i)
\end{cases}
\end{align}
We remark that the expression of the saddle point equation~\eqref{eq:saddle_pt_supermatrix_single_form} is identical to the standard matrix model by replacing the superresolvent with the original one.
However, its analytic property should be different since the superresolvent may contain both positive and negative residues, while the original resolvent only involves positive one.
\subsubsection{Functional method}
Let us discuss an alternative approach based on the functional method.
We define the density functions $\rho_\sigma(x)$ for $\sigma = 0,1$.
Then, we rewrite the effective action~\eqref{eq:eff_action_supermatrix} using these density functions,
\begin{align}
S[\rho_{0,1}] & = \sum_{\sigma = 0, 1} \left[ (-1)^\sigma \mathsf{t}_\sigma \int \mathrm{d}{x} \rho_\sigma (x) V(x) - 2 \mathsf{t}_\sigma^2 \int_{x < y} \mathrm{d}{x} \mathrm{d}{y} \rho_\sigma(x) \rho_\sigma(y) \log |x - y| \right]
\nonumber \\
& \qquad +
2 \mathsf{t}_0 \mathsf{t}_1 \Xint- \mathrm{d}{x} \mathrm{d}{y} \rho_0(x) \rho_1(y) \log |x - y|
+ \sum_{\sigma = 0,1} \sum_{i = 1}^{m_\sigma} \mathsf{t}_\sigma \ell_{\sigma,i} \left( \epsilon_{\sigma,i} - \int_{\mathcal{C}_{\sigma,i}} \mathrm{d}{x} \rho_\sigma(x) \right)
\, ,
\end{align}
where we consider the $m_\sigma$-cut solution for each sector, $\sigma = 0, 1$:
We added the Lagrange multiplier that imposes the condition
\begin{align}
\int_{\mathcal{C}_{\sigma,i}} \mathrm{d}{x} \rho_\sigma(x) = \epsilon_{\sigma, i}
\, ,
\end{align}
where we denote the cut by
\begin{align}
\mathcal{C}_\sigma = \bigsqcup_{i \in [m_\sigma]} \mathcal{C}_{\sigma,i}
\, , \qquad
\sigma = 0, 1
\, ,
\end{align}
and the corresponding filling fraction $(\epsilon_{\sigma,i})_{\sigma = 0, 1, i \in [m_\sigma]}$ with $\displaystyle \sum_{i \in [m_\sigma]} \epsilon_{\sigma,i} = 1$.
Taking the functional derivative of the effective action, we obtain
\begin{align}
x \in \mathcal{C}_{\sigma,i} \ : \quad
\mathsf{t}_\sigma^{-1} \frac{\delta S[\rho_{0,1}]}{\delta \rho_\sigma(x)} & =
(-1)^\sigma \left( V(x) - 2 \Xint- \mathrm{d}{y} \bar{\rho}(y) \log |x-y| \right) - \ell_{\sigma,i}
\nonumber \\
& =:
(-1)^\sigma V_\text{eff}(x) - \ell_{\sigma,i}
\, ,
\end{align}
where we define the effective potential $V_\text{eff}(x)$ and the effective density functions,
\begin{align}
\bar{\rho}(x) = \mathsf{t}_0 \rho_0(x) - \mathsf{t}_1 \rho_1(x)
\, .
\end{align}
The functional version of the saddle point equation~\eqref{eq:saddle_pt_supermatrix_single_form} is obtained by the derivative of the effective potential,
\begin{align}
\dv{V_\text{eff}(x)}{x} = V'(x) - 2 \Xint- \mathrm{d}{y} \frac{\bar{\rho}(y)}{x - y}
\, .
\end{align}
Writing the integral form of the superresolvent,
\begin{align}
\mathsf{W}(x) = \int \mathrm{d}{y} \frac{\bar{\rho}(y)}{x - y}
\, ,
\end{align}
the regularized one is given by the principal value integral
\begin{align}
\mathsf{W}_\text{reg}(x) & = \Xint- \mathrm{d}{y} \frac{\bar{\rho}(y)}{x - y}
\nonumber \\
& = \operatorname{Re} \mathsf{W}(x \pm \mathrm{i} 0) := \lim_{\epsilon \to 0^+} \frac{\mathsf{W}(x + \mathrm{i} \epsilon) + \mathsf{W}(x - \mathrm{i} \epsilon)}{2}
\, .
\end{align}
Hence, we obtain the functional version of \eqref{eq:saddle_pt_supermatrix_single_form} as follows,
\begin{align}
\dv{V_\text{eff}(x)}{x} = 0
\ \implies \
V'(x) - 2 \operatorname{Re} \mathsf{W}(x \pm \mathrm{i} 0) = 0
\, , \qquad
x \in \mathcal{C}_\sigma
\, .
\end{align}
\subsubsection{Spectral curve and quantization}\label{sec:sp_curve}
We have seen that the saddle point equation of the supermatrix model gives rise to the relation among the resolvents as shown in \eqref{eq:saddle_pt_supermatrix_finite}.
Further taking the limit $g \to 0$, the relation~\eqref{eq:saddle_pt_supermatrix_finite} is written in a closed form of the superresolvent,
\begin{align}
0 = \mathsf{W}(x)^2 - V'(x) \mathsf{W}(x) + \mathsf{P}(x)
\, ,
\end{align}
which defines the spectral curve of the supermatrix model:
\begin{itembox}{Spectral curve of supermatrix model}
Given the potential function $V'(x)$ and the polynomial function $\mathsf{P}(x)$, the spectral curve of the supermatrix is given as follows,
\begin{align}
\Sigma = \{ (x,y) \in \mathbb{C} \times \mathbb{C} \mid \mathcal{H}(x,y) = 0 \}
\, , \qquad
\mathcal{H}(x,y) = y^2 - V'(x) y + \mathsf{P}(x)
\, .
\label{eq:sp_curve_supermatrix}
\end{align}
\end{itembox}
This is formally identical to the spectral curve of the standard matrix model (see, e.g.~\cite{Eynard:2015aea}), but as we mentioned before, we should be careful of its analytic property.
While the spectral curve is based on the closed equation for the superresolvent, the saddle point equation~\eqref{eq:saddle_pt_supermatrix_finite} itself is not written as a closed form.
In order to discuss an alternative form, we rewrite the resolvents as follows,
\begin{align}
W_0(x) = bg \dv{}{x} \log \psi_0(x)
\, , \qquad
W_1(x) = b^{-1}g \dv{}{x} \log \psi_1(x)
\end{align}
where we define the wave functions (characteristic polynomials),
\begin{align}
\psi_0(x) = \prod_{i \in [N]} (x - x_i)
\, , \qquad
\psi_1(x) = \prod_{i \in [M]} (x - y_i)
\, .
\end{align}
Then, the superresolvent is given by the logarithmic derivative
\begin{align}
\mathsf{W}(x) = g \dv{}{x} \log \frac{\psi_0(x)^b}{\psi_1(x)^{b^{-1}}}
\, .
\end{align}
Together with these wave functions, we can recast the saddle point equation~\eqref{eq:saddle_pt_supermatrix_finite} in the following form,
\begin{align}
\left[ D_x^{(b)2} - V'(x) D_x^{(b)} + \mathsf{P}(x) \right] \psi_0 \cdot \psi_1 = 0
\, ,
\label{eq:quantum_curve_supermatrix}
\end{align}
where we define the $b$-deformed Hirota derivative,
\begin{align}
D_x^{(b)} \psi \cdot \phi
= g \left( b \pdv{}{x} - b^{-1} \pdv{}{x'} \right) \psi(x) \phi(x')\Big|_{x = x'}
= g b \psi'(x) \phi(x) - g b^{-1} \psi(x) \phi'(x)
\, .
\end{align}
The standard Hirota derivative corresponds to the case $b = 1$ ($\beta = 2$): $D_x = D_x^{(1)}$.
In fact, this bilinear equation is interpreted as a quantization of the spectral curve~\eqref{eq:sp_curve_supermatrix}:
\begin{itembox}{Quantum curve for $\beta$-supermatrix model}
Based on the two-variable function $\mathcal{H}(x,y)$ that defines the spectral curve~\eqref{eq:sp_curve_supermatrix}, we have the quantum curve for $\beta$-supermatrix model involving the Hirota derivative,
\begin{align}
\mathcal{H}(\hat{x},\hat{y}) \psi_0 \cdot \psi_1 = 0
\, , \qquad
\hat{x} = x
\, , \
\hat{y} = g D_x^{(b)}
\, .
\end{align}
\end{itembox}
Recalling the definition of the Hirota derivative, the canonical commutation relation is given by
\begin{align}
[\hat{y}, \hat{x}] =
\begin{cases}
+ g b & (\text{for} \ \psi_0) \\
- g b^{-1} & (\text{for} \ \psi_1) \\
\end{cases}
\end{align}
From this point of view, we call the bilinear equation~\eqref{eq:quantum_curve_supermatrix} the quantum curve for the supermatrix model.
We now have the both positive and negative quantum parameters (Planck constants) for quantization of the supermatrix spectral curve, corresponding to that the superresolvent has both positive and negative residues.
Moreover, we now relate the supermatrix parameters to the so-called $\Omega$-background parameters (see Sec.~\ref{sec:instanton}),
\begin{align}
(\epsilon_1, \epsilon_2) = \left(g, - \frac{g}{b^2} \right)
\, .
\label{eq:Omega_background_matrix_parameter}
\end{align}
In this notation, we have $b^2 = - \epsilon_1 / \epsilon_2$, and thus the condition $b = 1$ is equivalent to $\epsilon_1 + \epsilon_2 = 0$.
Then, the $b$-Hirota derivative is rewritten using $\epsilon_{1,2}$ as follows,
\begin{align}
D_x^{(b)} \psi \cdot \phi & = b \epsilon_1 \psi'(x) \phi(x) + b \epsilon_2 \psi(x) \phi'(x) =: b D_x^{(\epsilon_1, \epsilon_2)} \psi \cdot \phi
\, ,
\end{align}
where $D_x^{(\epsilon_1, \epsilon_2)}$ is the $(\epsilon_1,\epsilon_2)$-deformed Hirota derivative defined in~\cite{Nakajima:2003pg}.
This operator is reduced to the ordinary derivative in the limit $\epsilon_1 \to 0$ or $\epsilon_2 \to 0$, which is called the Nekrasov--Shatashvili (NS) limit,
\begin{align}
D_x^{(\epsilon_1, \epsilon_2)} \psi \cdot \phi
\ \longrightarrow \
\begin{cases}
\epsilon_1 \psi'(x) \phi(x) & (\epsilon_2 \to 0) \\
\epsilon_2 \psi(x) \phi'(x) & (\epsilon_1 \to 0)
\end{cases}
\end{align}
See also Sec.~\ref{sec:Bethe/gauge} for a related discussion.
\subsubsection{Gaussian model}
We consider the simplest example with the quadratic potential $V(x) = \frac{1}{2} x^2$, which is called the Gaussian matrix model.
In this case, the quantum curve is given by
\begin{align}
\left[ D_x^{(b)2} - x D_x^{(b)} + (\mathsf{t}_0 - \mathsf{t}_1) \right] \psi_0 \cdot \psi_1 = 0
\, .
\end{align}
Hence, in particular for the unitary case $b = 1$ ($\beta = 2$), we have
\begin{align}
\left[ D_\xi^2 - \xi D_\xi + (N - M) \right] \psi_0 \cdot \psi_1 = 0
\, , \qquad
x = \sqrt{g} \xi
\, .
\end{align}
This bilinear equation is known to be (a part of) the bilinear equations for the $\tau$-functions in the symmetric form of the Painlevé IV equation~\cite{Noumi:1999NMJ}.
In this case, the polynomial solution is given by the generalized Hermite polynomial, which is given through specialization of Schur functions.
In the NS limit, as mentioned above, the Hirota derivative is reduced to the ordinary derivative, where this bilinear equation is accordingly reduced to the differential equation for the Hermite polynomial.
\subsection{Free field realization}\label{sec:free_field_realization}
We turn to discuss algebraic aspects of supermatrix model.
In particular, we show that the supermatrix partition function has a realization in terms of the chiral boson fields (free field realization).
\subsubsection{Operator formalism}\label{sec:op_formalism}
In order to discuss the free field formalism, we consider the ordinary $\beta$-deformed matrix model of rank $N$,
\begin{align}
\mathcal{Z}_N = \int \mathrm{d}{X} \mathrm{e}^{-\frac{b}{g} \tr V(X)} \Delta_N(X)^{2 b^2}
\, .
\end{align}
In this case, the matrix moment (the power-sum average of the eigenvalues) is given as follows,
\begin{align}
\left< \tr X^n \right> = \frac{1}{\mathcal{Z}_N} \int \mathrm{d}{X} \left( \tr X^n \right) \mathrm{e}^{-\frac{b}{g} \tr V(X)} \Delta_N(X)^{2 b^2}
= - \frac{b}{g} n \pdv{}{t_n} \log \mathcal{Z}_N
\, .
\end{align}
Hence, the derivative with the coupling constant $\{t_n\}$ plays a similar role to the multiplication of the matrix power $\{ \tr X^n \}$.
This is because the potential factor $\mathrm{e}^{-\frac{b}{g} \tr V(X)}$ plays a role of the plane wave factor $\mathrm{e}^{\mathrm{i} p x}$ in the Fourier transform (FT).
In order to have this correspondence for all $n \in \mathbb{N}$, we consider the potential with infinitely many coupling constants, $V(x) = \sum_{n=1}^\infty \frac{t_n}{n} x^n$.
For the supermatrix model defined by
\begin{align}
\mathcal{Z}_{N|M} = \frac{1}{N! M!} \int \mathrm{d}{X} \mathrm{d}{Y} \mathrm{e}^{-\frac{1}{g} \str_b V(Z)} \frac{\Delta_N(X)^{2b^2} \Delta_M(Y)^{2b^{-2}}}{\prod_{i \in [N], j \in [M]}(x_i - y_j)^2}
\, ,
\end{align}
the same argument is applied for $\langle \str_b X^n \rangle \leftrightarrow - \frac{1}{g} n \pdv{}{t_n}$.
Then, we define oscillator operators,
\begin{align}
a_n = \sqrt{2} g n \pdv{}{t_n} \left( \stackrel{\text{FT}}{\longleftrightarrow} \ - \frac{1}{\sqrt{2}b} \tr X^n \right)
\, , \qquad
a_{-n} = \frac{1}{\sqrt{2}g} t_n
\, , \qquad
(n > 0)
\end{align}
which obeys the commutation relation of the Heisenberg algebra,
\begin{align}
[a_n, a_m] = n \delta_{n+m,0}
\, .
\end{align}
In addition, we also add the zero modes $(a_0,\bar{a}_0)$ with the commutation relation,
\begin{align}
[a_n, \bar{a}_0] = \delta_{n,0}
\, ,
\end{align}
where we interpret $a_0 \stackrel{\text{FT}}{\longleftrightarrow} -(\tr X^0)/\sqrt{2}b = - N /\sqrt{2}b$.
In this formalism, there exist infinitely many operators $\{ a_n \}_{n \in \mathbb{Z}}$.
They are independent operators if the matrix size is taken to be infinite $N \to \infty$.
For example, if $N = 1$, we have relations among these operators, $\tr X^n = (\tr X)^n$.
We define the vacuum state, which is annihilated by the positive modes,
\begin{align}
a_{n} | 0 \rangle = 0
\, , \qquad
n \ge 0
\, .
\end{align}
In this sense, the positive modes are the annihilation operators, and the negative modes are the creation operators.
We also define the charged vacuum using the zero mode,
\begin{align}
| \alpha \rangle = \mathrm{e}^{\alpha \bar{a}_0} | 0 \rangle
\, , \qquad
a_0 | \alpha \rangle = \alpha | \alpha \rangle
\, .
\end{align}
Based on the operators defined above, we define the operators, called the \emph{chiral boson} and the \emph{U(1) current}, as follows,
\begin{subequations}
\begin{align}
\phi(x) & = - \sum_{n \in \mathbb{Z}_{\neq 0}} \frac{a_n}{n} x^{-n} + a_0 \log x + \bar{a}_0
& \implies \ - \frac{1}{\sqrt{2}b} \left[ \tr \log (x - X) - \frac{b}{g} V(x) \right]
\, , \\
J(x) & = \partial \phi(x) = \sum_{n \in \mathbb{Z}} \frac{a_n}{x^{n+1}}
& \implies \ - \frac{1}{\sqrt{2}b} \left[ \tr \left( \frac{1}{x - X} \right) - \frac{b}{g} V'(x) \right]
\, .
\end{align}
\end{subequations}
Furthermore, we also define the energy-momentum tensor
\begin{align}
T(x) = \frac{1}{2} {:JJ:}(x) + \rho \partial J(x) =: \sum_{n \in \mathbb{Z}} \frac{L_n}{x^{n+2}}
\, ,
\qquad
\rho = \sqrt{2}(b - b^{-1})
\, ,
\end{align}
where we denote the normal ordering symbol by ${: - :}$, the annihilation operators are placed to right, the creation operators are placed to left in this symbol.
Recalling the operator product expansion (OPE) of the current operators,
\begin{align}
J(x) J(x') = \frac{1}{(x - x')^2} + \text{regular}
\, ,
\end{align}
the normal ordering is given as follows,
\begin{align}
{:JJ:}(x) = \lim_{x' \to x} \left[ J(x) J(x') - \frac{1}{(x - x')^2} \right]
\, .
\end{align}
It turns out that the generators $\{L_n\}_{n \in \mathbb{Z}}$ written in terms of the oscillators,
\begin{align}
L_n = \frac{1}{2} \sum_{m \in \mathbb{Z}} {: a_m a_{n-m} :} - \rho (n+1) a_n
\, ,
\label{eq:Virasoro_gen_free_field}
\end{align}
obey the algebraic relation of the Virasoro algebra,
\begin{align}
[L_n, L_m] = (n-m) L_{n+m} + \frac{c}{12} n (n^2 - 1) \delta_{n+m,0}
\, ,
\end{align}
with the central charge
\begin{align}
c = 1 - 6 (b - b^{-1})^2
=
\begin{cases}
1 & (\beta = 2) \\
- 2 & (\beta = 1, 4)
\end{cases}
\, .
\end{align}
Hence, the energy-momentum tensor plays a role of the generating current of the Virasoro algebra, and the construction of the energy-momentum tensor as a bilinear form of the current operator is called the Sugawara construction.
\subsubsection{Vertex operators}
We define the vertex operator from the chiral boson,
\begin{align}
\mathsf{V}_\alpha (x) = {: \mathrm{e}^{\alpha \phi(x)} :}
\, .
\end{align}
Then, the OPE with the energy-momentum tensor is given by
\begin{align}
T(x) \mathsf{V}_\alpha (x') = \frac{\Delta_\alpha}{(x - x')^2} \mathsf{V}_\alpha (x') + \frac{1}{x - x'} \pdv{}{x'} \mathsf{V}_\alpha (x') + \text{regular}
\end{align}
where the coefficient called the conformal weight $\Delta_\alpha$ is given by
\begin{align}
\Delta_\alpha = \frac{1}{2} \alpha (\alpha + \rho)
\, .
\end{align}
From the Virasoro algebra point of view, the conformal weight is given by the eigenvalue of the operator $L_0$.
Moreover, the operator annihilated by $L_{n>0}$ is called the primary operator, and the vertex operator $\mathsf{V}_\alpha (x)$ is actually primary.
We remark that there are two possibilities to provide $\Delta_\alpha = 1$,
\begin{align}
\alpha_0 = - \sqrt{2} b
\, , \qquad
\alpha_1 = \sqrt{2} b^{-1}
\quad \implies \quad
\Delta_{\alpha_{0,1}} = 1
\, .
\end{align}
Hence, defining the screening current having the conformal weight one,
\begin{align}
S_{\sigma}(x) = {: \mathrm{e}^{\alpha_{\sigma} \phi(x)} :}
\, , \qquad
\sigma = 0, 1
\, ,
\end{align}
the singular part of the OPE with the energy-momentum tensor is written as a total derivative,
\begin{align}
T(x) S_\sigma (x') = \frac{1}{(x - x')^2} S_\sigma (x') + \frac{1}{x - x'} \pdv{}{x'} S_\sigma (x') + \text{regular}
= \pdv{}{x'} \left[ \frac{1}{x - x'} S_\sigma (x') \right] + \text{regular}
\, .
\end{align}
This implies that the screening charge defined by
\begin{align}
Q_\sigma = \oint \mathrm{d}{x} S_\sigma(x)
\, ,
\end{align}
does not provide a singular contribution in the OPE, and thus it commutes with the energy-momentum tensor
\begin{align}
[T(x), Q_\sigma] = 0
\, .
\end{align}
This is a crucial property that characterizes the Virasoro algebra:
In fact, the Virasoro algebra is defined as a commuting sub-algebra of the Heisenberg algebra, and such a characterization can be applied to more generalized situations (W-algebras).
See, e.g.,~\cite{Bouwknegt:1992wg}.
\subsubsection{Construction of matrix model}
Let us discuss how to construct the matrix model based on the operator formalism discussed above.
Recalling that the vertex operator product is given by
\begin{align}
\frac{\mathsf{V}_\alpha (x) \mathsf{V}_{\alpha'} (x')}{: \mathsf{V}_\alpha (x) \mathsf{V}_{\alpha'} (x') :} = (x - x')^{\alpha \alpha'}
\, ,
\end{align}
the screening charge product is given by
\begin{align}
Q_0^N & = \oint_{|x_1| < \cdots < |x_N|} \mathrm{d}{X} \Delta_N(X)^{2b^2} {: \prod_{i\in[N]} S_0(x_i) :}
\nonumber \\
& = \frac{1}{N!} \oint \mathrm{d}{X} \Delta_N(X)^{2b^2} {: \prod_{i \in [N]} S_0(x_i) :}
\, ,
\end{align}
where the integration contour is initially taken in the radial ordering, and then analytically continuated to obtain the second expression.
Hence, defining the $\mathcal{Z}$-state
\begin{align}
| \mathcal{Z}_N \rangle = Q_0^N | 0 \rangle
\, ,
\end{align}
and the modified dual charged vacuum,%
\footnote{%
This modified vacuum is realized as a coherent state with respect to the Heisenberg algebra.
}
\begin{align}
\langle \alpha; d+1 | a_{-n} =
\begin{cases}
\displaystyle
\langle \alpha; d+1 | \alpha & (n = 0) \\ \displaystyle
\langle \alpha; d+1 | \frac{\mathfrak{t}_n}{\sqrt{2}g} & (n \in [d+1])
\\
0 & (\text{otherwise})
\end{cases}
\end{align}
we obtain the matrix model partition function as a correlation function of the vertex operators,
\begin{align}
\langle N \alpha_0; d+1 | \mathcal{Z}_N \rangle = \langle N \alpha_0; d+1 | Q_0^N | 0 \rangle = \frac{1}{N!} \oint \mathrm{d}{X} \mathrm{e}^{-\frac{b}{g} \tr V(X)} \Delta_N(X)^{2b^2}
\, ,
\end{align}
where the potential function is now given by a finite polynomial function, $V(x) = \sum_{n=1}^{d+1} \frac{\mathfrak{t}_n}{n} x^n$.
Considering both screening charges $Q_{0,1}$, we instead obtain
\begin{align}
| \mathcal{Z}_{N|M} \rangle & := Q_0^N Q_1^M | 0 \rangle
\nonumber \\
& =
\frac{1}{N! M!} \oint \mathrm{d}{X} \mathrm{d}{Y} \frac{\Delta_N(X)^{2 b^2} \Delta_M(Y)^{2 b^{-2}}}{\prod_{i \in [N], j \in [M]} (x_i - y_j)^2}
{: \prod_{i \in [N]} S_0(x_i) \prod_{j \in [M]} S_1(y_j) :}
\, ,
\end{align}
which gives rise to the supermatrix model partition function.
\begin{itembox}{Free field realization of supermatrix partition function}
The eigenvalue integral form of supermatrix model partition function is realized using two types of screening charges,
\begin{align}
&
\langle N\alpha_0 + M\alpha_1; d+1 | \mathcal{Z}_{N|M} \rangle
\nonumber \\
& =
\frac{1}{N! M!} \oint \mathrm{d}{X} \mathrm{d}{Y} \mathrm{e}^{-\frac{b}{g} \tr V(X)} \mathrm{e}^{+\frac{b^{-1}}{g}\tr V(Y)} \frac{\Delta_N(X)^{2 b^2} \Delta_M(Y)^{2 b^{-2}}}{\prod_{i \in [N], j \in [M]} (x_i - y_j)^2}
\, .
\end{align}
\end{itembox}
We remark that such an integral form of the correlation function is known as the \emph{Dotsenko--Fateev integral formula} in the context of two-dimensional CFT.
\subsubsection{Virasoro constraint}
Let us consider the action of the energy-momentum tensor on the vacuum state,
\begin{align}
T(x) | 0 \rangle = \sum_{n \in \mathbb{Z}} \frac{L_n}{x^{n+2}} | 0 \rangle
\, .
\end{align}
Requiring the regularity at $x = 0$, we should have
\begin{align}
L_n | 0 \rangle = 0
\, , \qquad
n \ge -1
\, .
\label{eq:Virasoro_const_vac}
\end{align}
This condition is rephrased as follows:
The vacuum is interpreted as the primary field with the conformal weight zero, hence $L_{n \ge 0}$ annihilates the vacuum.
Furthermore, $L_{-1}$ is a generator of the translation realized by the derivative $\pdv{}{x}$.
Hence, $L_{-1}$ annihilates the vacuum since it is translation invariant.
Recalling that the $\mathcal{Z}$-state is constructed from the vacuum with the screening charges, that commute with the energy-momentum tensor (and thus the Virasoro generators).
This implies that the $\mathcal{Z}$-state shows the same behavior as in the case of the vacuum \eqref{eq:Virasoro_const_vac}.
\begin{itembox}{Virasoro constraint}
The $\mathcal{Z}$-state constructed by the screening charges obeys the following relation,
\begin{align}
L_n | \mathcal{Z}_{N|M} \rangle = 0
\, , \qquad
n \ge -1
\, .
\label{eq:Virasoro_const}
\end{align}
This is called the Virasoro constraint for the (super)matrix model.
\end{itembox}
We see how this Virasoro is obtained in the context of the matrix model.
We start with the following identity,
\begin{align}
0 & = \int \mathrm{d}{X} \sum_{\ell \in [N]} \pdv{}{x_\ell} \left[ x_\ell^k \mathrm{e}^{-\frac{b}{g} \tr V(X)} \Delta_N(X)^{2b^2} \right]
\nonumber \\
& = \int \mathrm{d}{X} \sum_{\ell \in [N]} \left[ k x_\ell^{k-1} - \frac{b}{g} V'(x_\ell) x_\ell^k + 2b^2 \sum_{j (\neq \ell)} \frac{x_\ell^k}{x_\ell - x_j} \right] \mathrm{e}^{-\frac{b}{g} \tr V(X)} \Delta_N(X)^{2b^2}
\, .
\end{align}
Recalling the identity
\begin{align}
\sum_{i \in [N]} \sum_{j (\neq i)} \frac{2 x_i^k}{x_i - x_j} & = \sum_{i \neq j} \frac{x_i^k - x_j^k}{x_i - x_j}
\nonumber \\
& = \sum_{m=0}^{k-1} \tr X^m \tr X^{k-m-1} - k \tr X^{k-1}
\, ,
\end{align}
we obtain the following relation among the expectation values of the matrix moments, called the \emph{loop equation},
\begin{align}
\sum_{m=0}^{k-1} \langle \tr X^m \tr X^{k-m-1} \rangle + (b^{-2} - 1) k \langle \tr X^{k-1} \rangle - \frac{b^{-1}}{g} \langle \tr X^k V'(X) \rangle = 0
\, .
\end{align}
In the operator formalism discussed in Sec.~\ref{sec:op_formalism}, we may write this relation as
\begin{align}
L_{k-1} \mathcal{Z}_N = 0
\, , \qquad
k \ge 0
\, ,
\end{align}
where we apply the free field realization of the Virasoro generator \eqref{eq:Virasoro_gen_free_field}.
See~\cite{Desrosiers:2009pz} for the loop equation of the supermatrix model in the case $b = 1$.
\section{Supergroup gauge theory}\label{sec:supergroup_gauge_theory}
In this section, we introduce supergroup gauge theory, gauge theory having supergroup gauge symmetry, and discuss fundamental perspectives.
\subsubsection*{Differential forms}
Let $G$ be a Lie supergroup, and the corresponding Lie superalgebra $\mathfrak{g} = \operatorname{Lie} G$.
Let $M$ be a $d$-dimensional space-time manifold.
The fundamental degrees of freedom of supergroup gauge theory is the one-form connection that takes a value in $\mathfrak{g}$, $A \in \Omega^1(M,\mathfrak{g})$.
Then, we define the covariant derivative $D = d + A$ and the curvature two-form given by $D^2 = dA + A^2 \in \Omega^2(M,\mathfrak{g})$.
The $G$-gauge transformation is given by
\begin{align}
G \ : \
D \longmapsto g D g^{-1}
\, , \qquad
F \longmapsto g F g^{-1}
\, , \qquad
A \longmapsto g A g^{-1} + g d g^{-1}
\, , \qquad
g \in G
\, .
\label{eq:gauge_transf}
\end{align}
Formally, these expressions are parallel with the ordinary (non-supergroup) gauge theory.
In the $\mathbb{Z}_2$-graded situation, the connection is then called the \emph{superconnection}~\cite{Quillen:1985vya,Mathai:1986tc}.
Let us discuss the supermatrix realization of these differential forms.
Let $G = \mathrm{U}(n_0|n_1)$ for the moment.
In this case, the connection $A$ is given by an anti-Hermitian supermatrix,
\begin{align}
A =
\begin{pmatrix}
A^{(0)} & \psi \\ \psi^\dag & A^{(1)}
\end{pmatrix}
\quad \text{with} \quad
A^{(\sigma)\dag} = - A^{(\sigma)}
\quad \implies \quad
A^\dag = - A
\, .
\end{align}
We recall that $\bar{\bar{\theta}} = -\theta$ for the Grassmann variable (see~\eqref{eq:Grassmann_var_double_conj}).
We remark that $A^{(\sigma)}$ is in the adjoint representation of $\mathrm{U}(n_\sigma)$, and $\psi$ is in the bifundamental representation, $\mathrm{U}(n_0) \times \overline{\mathrm{U}(n_1)}$.
Moreover, each component is given by
\begin{align}
A^{(\sigma)} = A_\mu^{(\sigma)} dx^\mu
\, , \qquad
\psi = \psi_\mu dx^\mu
\, .
\end{align}
Hence, $\psi_\mu$ is a spin-1 fermionic degree of freedom, that is not compatible with the spin-statistics theorem.
The curvature two-form is given by
\begin{align}
F = dA + A \wedge A =
\begin{pmatrix}
dA^{(0)} + A^{(0)} \wedge A^{(0)} + \psi \wedge \psi^\dag & d \psi + A \wedge \psi + \psi \wedge B \\
d \psi^\dag + \psi^\dag \wedge A + B \wedge \psi^\dag & dA^{(1)} + A^{(1)} \wedge A^{(1)} + \psi^\dag \wedge \psi
\end{pmatrix}
\, .
\end{align}
We may write the wedge product using the anti-symmetrization symbol,
\begin{align}
a \wedge b = \frac{1}{2} \left(a_\mu b_\nu - a_\nu b_\mu \right) dx^\mu dx^\nu = \frac{1}{2} a_{[\mu} b_{\nu]} dx^\mu dx^\nu
\, .
\end{align}
Therefore, the component of the curvature two-form is given by
\begin{align}
F = \frac{1}{2} F_{\mu \nu} dx^\mu dx^\nu
\end{align}
where we have
\begin{align}
F_{\mu \nu} & =
\begin{pmatrix}
F_{\mu\nu}^{(0)} + \psi_{[\mu} \psi^\dag_{\nu]} & \partial_{[\mu} \psi_{\nu]} + A^{(0)}_{[\mu} \psi_{\nu]} + \psi_{[\mu} A^{(1)}_{\nu]} \\
\partial_{[\mu} \psi_{\nu]}^\dag + \psi_{[\mu}^\dag A^{(0)}_{\nu]} + A^{(1)}_{[\mu} \psi^\dag_{\nu]} & F_{\mu\nu}^{(1)} + \psi_{[\mu}^\dag \psi_{\nu]}
\end{pmatrix}
\, .
\label{eq:super_curvature}
\end{align}
We denote the curvature associated with the bosonic subgroup of $G$ by $F^{(\sigma)} = d A^{(\sigma)} + A^{(\sigma)} \wedge A^{(\sigma)} = \frac{1}{2} F^{(\sigma)}_{\mu\nu} dx^\mu dx^\nu$, e.g., for $G = \mathrm{U}(n_0|n_1)$, $F^{(\sigma)} \in \Omega^2(M,\mathfrak{u}_{n_\sigma})$.
In general, we define the commutator for the differential forms as follows:
Let $X = X^a \otimes t^a \in \Omega^p(M,\mathfrak{g})$ and $Y = Y^a \otimes t_b \in \Omega^q(M,\mathfrak{g})$ with the generators of the Lie superalgebra denoted by $(t^a)_{a \in [\operatorname{dim} \mathfrak{g}]}$.
Recalling $X^a \wedge Y^b = (-1)^{pq} (-1)^{|a||b|} Y^b \wedge X^a$, denoted by $|a| = |X_a|$, etc, the commutator for the differential forms is defined by
\begin{align}
[X, Y] & = t^a t^b \otimes X^a \wedge Y^b - (-1)^{pq} t^b t^a \otimes Y^b \wedge X^a
\nonumber \\
& = (t^a t^b - (-1)^{|a||b|} t^b t^a) \otimes X^a \wedge Y^b
\nonumber \\
& = [t^a, t^b] \otimes X^a \wedge Y^b
\, .
\end{align}
The commutator in the last line is the supercommutator with respect to the superalgebra $\mathfrak{g}$ defined in~\eqref{eq:supercommutator_def}.
For the one-form connection, $A \in \Omega^1(M,\mathfrak{g})$, we have
\begin{align}
A \wedge A = \frac{1}{2} [A,A] = \frac{1}{2} [t^a, t^b] \otimes A^a \wedge A^b
\, .
\end{align}
Therefore, the component of the curvature two-form is given in the same way as the ordinary case,
\begin{align}
F_{\mu\nu} = \partial_\mu A_\nu - \partial_\nu A_\mu + [A_\mu, A_\nu]
\, ,
\end{align}
which takes a value in the Lie superalgebra $\mathfrak{g}$.
\subsection{Supergroup Yang--Mills theory}
We consider the Yang--Mills (YM) action for the supergroup theory,
\begin{align}
S_\text{YM} = - \frac{1}{g^2} \int_M \str (F \wedge \star F)
\, ,
\end{align}
where we define the Hodge star operator, $\star$ : $\Omega^p(M) \to \Omega^{d-p}(M)$.
Hence, we have the volume form, $F \wedge \star F = \frac{1}{4} F_{\mu\nu} F^{\mu\nu} d\!\operatorname{vol}(M)$.
In order to impose the invariance under the $G$-gauge transform~\eqref{eq:gauge_transf}, we replace the ordinary trace with the supertrace.
From the expression of the two-form curvature~\eqref{eq:super_curvature}, we can explicitly write down the YM action in terms of the gauge fields, $(A^{(\sigma)}, \psi)$, which would be complicated.
Hence, for the moment, we just write the leading contributions,
\begin{align}
S_\text{YM} = - \frac{1}{g^2} \int_M \tr_0 (F^{(0)} \wedge \star F^{(0)}) + \frac{1}{g^2} \int_M \tr_1 (F^{(1)} \wedge \star F^{(1)}) + \cdots
\, ,
\label{eq:S_YM_leading}
\end{align}
which are the standard YM actions of the bosonic subgroups $\mathrm{U}(n_\sigma)$ in $G = \mathrm{U}(n_0|n_1)$.
Now we observe that the kinetic term signatures are opposite due to the supertrace property, that means that the spectrum of supergroup YM theory is not bounded.
Such a property was also found in the supermatrix model discussed in Sec.~\ref{sec:super_matrix}, and thus it seems to be a universal behavior in supergroup theory.
Due to this unboundness, the notion of vacua is not well-defined in this case, and thus we would need a non-perturbative completion for supergroup theory:%
\footnote{%
Such a situation resembles unstable vacua.
In this case, we should find true (stable) vacua non-perturbatively.
}
In fact, it has been known that supergroup theory is pertubatively equivalent to the ordinary gauge theory through the analytic continuation, but there would be an essential difference in the non-perturbative regime.
Even though there are no bounded vacua, one can still consider the equation of motion with respect to the YM action, which is given by
\begin{align}
D(\star F) = 0
\, .
\end{align}
This is a second order non-linear PDE on $M$, which is in general difficult to solve.
In the four-dimensional case, we have a special class of solutions, called the (anti-)instanton, given by the solution of the (anti-)self-dual ((A)SD) YM equation,
\begin{align}
\star F = \mp F
\, .
\label{eq:ASDYMeq}
\end{align}
One can check that the instanton provides a solution of the equation of motion using the Bianchi identity,
\begin{align}
D(\star F) \stackrel{\text{(A)SD}}{=} \mp D F \stackrel{\text{Bianchi}}{=} 0
\, .
\end{align}
We remark that the Bianchi identity still holds for the supergroup case due to the Jacobi identity~\eqref{eq:Jacobi_id}.
In fact, the instanton plays an essential role in non-perturbative aspects of supergroup gauge theory.
We will discuss the details of instantons in Sec.~\ref{sec:instanton}.
\subsection{Quiver gauge theory realization}\label{sec:quiver_realization}
As observed in~\eqref{eq:S_YM_leading}, the YM action of supergroup gauge theory consists of two ordinary YM actions.
In fact, supergroup gauge theory has a realization as a quiver gauge theory through analytic continuation.
Quiver gauge theory is a class of gauge theories involving multiple gauge degrees of freedom: The gauge group is given by a product form, $G = \prod_i G_i$.
In addition to the gauge field, there is another degree of freedom, called the \emph{bifundamental matter}, that connects different gauge fields, transforming in the bifundamental representation of the connecting gauge groups, $G_i \times G_j$, $G_i \times \overline{G}_j$, etc.
In the case of $\mathrm{U}(n_0|n_1)$ supergroup theory, in addition to the gauge fields of $\mathrm{U}(n_\sigma)$ subgroup, there are also fermionic degrees of freedom that transform in $\mathrm{U}(n_0) \times \overline{\mathrm{U}(n_1)}$ and the conjugate.
Hence, these fermionic fields are interpreted as bifundamental matters from this point of view.
Since we have two such fields $(\psi,\psi^\dag)$, this theory is identified with $\widehat{A}_1$ quiver gauge theory.%
\footnote{%
This classification is based on the identification of quiver diagram and (affine) Dynkin diagram.
$\widehat{A}_1$ quiver consists of two nodes and two connecting edges.
}
An important feature of supergroup gauge theory is the signature of the kinetic term.
In order to realize this situation based on $\widehat{A}_1$ quiver theory, we have to assign the coupling constants as follows,
\begin{align}
\left( \frac{1}{g_0^2} \, , \, \frac{1}{g_1^2} \right)
\ \longrightarrow \
\left( + \frac{1}{g^2} \, , \, - \frac{1}{g^2} \right)
\, .
\label{eq:coupling_tuning_A1}
\end{align}
Otherwise, we impose the condition $\frac{1}{g_0^2} + \frac{1}{g_1^2} = 0$.
This assignment is in fact unphysical since we have a negative coupling.
Therefore, the supergroup gauge theory is realized in unphysical parameter regime, which could be interpreted as the analytic continuation of physical $\widehat{A}_1$ quiver gauge theory~\cite{Dijkgraaf:2016lym}.
\subsubsection*{Orthosymplectic theory}
From this point of view, the orthosymplectic supergroup theory is realized similarly by $\widehat{A}_1$ quiver with O and Sp gauge nodes.
In fact, the combinations of O $\times$ O and Sp $\times$ Sp are not compatible with their flavor symmetry, and O $\times$ Sp is a unique choice among O and Sp theories.
\subsubsection*{Chern--Simons theory}
A similar construction is also available for Chern--Simons theory.
In this case, the Chern--Simons level plays a role of the coupling constant.
In fact, the supersymmetric Chern--Simons--matter theory, a.k.a., Aharony--Bergman--Jafferis--Maldacena (ABJM) theory~\cite{Aharony:2008ug,Aharony:2008gk}, completely fits these conditions:
It is a quiver gauge theory involving two nodes and two bifundamental matters, and two couplings (levels) with opposite signs.
Actually, the partition function of ABJM theory obtained through the localization formalism supports its connection with supergroup theory:
The partition function of $\mathrm{U}(N)_k \times \mathrm{U}(M)_{-k}$ theory takes a form~\cite{Drukker:2009hy,Marino:2009jd},
\begin{align}
\mathcal{Z}_{N|M} = \frac{1}{N!M!} \int \mathrm{d}{X} \mathrm{d}{Y} \mathrm{e}^{-\frac{1}{2g} \tr X^2 + \frac{1}{2g} \tr Y^2} \frac{\prod_{i < j}^{N} \sinh (x_i - x_j)^2 \prod_{i < j}^{M} \sinh (y_i - y_j)^2}{\prod_{j \in [M]}^{i \in [N]} \cosh(x_i - y_j)^2}
\, ,
\end{align}
where we have the pure imaginary coupling $g = 2 \pi \mathrm{i} / k$.
This is interpreted as a trigonometric analog of the supermatrix model discussed in Sec.~\ref{sec:super_matrix}.
We remark that since the Chern--Simons theory is a topological theory, a negative coupling (negative level) does not imply unphysical behavior.
See also \cite{Horne:1989ue,Bourdeau:1991hn,Rozansky:1992zt} \cite{Mikhaylov:2014aoa,Mikhaylov:2015nsa} \cite{Okazaki:2017sbc} \cite{Aghaei:2018cbn} \cite{Cassia:2021uly} for related works on supergroup Chern--Simons theory.
\subsubsection*{Supergroup quiver gauge theory}
We can generalize this argument for quiver gauge theory of supergroups.
In this case, since each gauge node is of supergroup, we need copies of $\widehat{A}_1$ quiver theory.
For $\Gamma$-quiver supergroup gauge theory, the total structure is given by $(\widehat{A}_1, \Gamma)$ quiver.
Such a theory characterized by a pair of quivers is called the double quiver theory, which has a natural geometric origin in eight dimensions.
See \cite{Kimura:2022zsm}.
\subsection{String/M-theory perspective}\label{sec:string_theory_perspective}
String/M-theory provides various insights on non-perturbative aspects of gauge theory.
In particular, considering a stack of D-branes, non-Abelian gauge theory is realized as a low-energy effective theory.
Open strings ending on the brane provide matrix degrees of freedom as homomorphism of the Chan--Paton vector space associated with the boundary condition of open string:
$n$ stack of branes gives rise to $n$-dimensional vector space.
From this point of view, we require two different types of branes to realize a $\mathbb{Z}_2$-graded vector space to construct supergroup gauge theory.
A natural candidate is an anti-brane.
Even though a brane--anti-brane system exhibits similar properties, it has been known that this configuration does not yield supergroup gauge theory:
Open strings connecting brane--anti-brane give rise to the \emph{tachyon}, that is a bosonic degree of freedom transforming in the bifundamental representation of two gauge groups associated with branes and anti-branes.
See, e.g.,~\cite{Witten:1998cd,Kraus:2000nj,Takayanagi:2000rz,Alishahiha:2000du} for details.
Therefore, we need a different type of branes required for realizing a fermionic degree of freedom.
Such an object is known as the \emph{negative brane} (also called the \emph{ghost brane})~\cite{Vafa:2001qf,Okuda:2006fb,Vafa:2014iua,Dijkgraaf:2016lym}, and it has been shown that a stack of brane--negative-brane actually yields a supergroup gauge theory.
More recently, it has been pointed out that such a negative brane plays an essential role in the resurgence~\cite{Marino:2022rpz,Schiappa:2023ned}.
Comparing the anti-brane and the negative brane, while the anti-brane has a positive tension (positive energy density; source of gravity), the negative brane has a negative tension (negative energy density; source of anti-gravity) although both have negative RR charges.
From this point of view, the negative brane is associated with an unphysical open string boundary condition.
On the other hand, an advantage of negative branes is that a bound state with branes does not violate further supersymmetry, which is still BPS, while that for anti-branes is not BPS.
This property also plays an important role to discuss the instanton solution in supergroup gauge theory.
See Sec.~\ref{sec:instanton}.
\subsubsection{Hanany--Witten construction}\label{sec:Hany-Witten_const}
A stack of brane--negative-brane yields sixteen supercharges supergroup gauge theory similarly to the configuration of ordinary branes.
In order to reduce the supersymmetries, one can consider the Hanany--Witten setup~\cite{Hanany:1996ie} that involves NS5 branes and D4 branes suspended between them in the type IIA setup.
Adding the negative D4 branes, we can realize four-dimensional $\mathcal{N} = 2$ supergroup gauge theory that preserves eight supercharges.
Each brane is extended in the following directions:%
\footnote{%
We denote the ordinary $p$-brane (positive brane) by D$p$ or D$p^+$, the negative brane by D$p^-$, the anti-brane by $\overline{\text{D}p}$ or $\overline{\text{D}p}^+$, the anti-negative brane by $\overline{\text{D}p}^-$.
}
\begin{align}
\begin{tabular*}{.8\textwidth}{@{\extracolsep{\fill}}ccccccccccc}\toprule
&0&1&2&3&4&5&6&7&8&9 \\\midrule
NS5 &--&--&--&--&--&--&&& \\
D4$^\pm$ &--&--&--&--&&&--&& \\\toprule
\end{tabular*}
\end{align}
In this configuration, the positions of D4$^\pm$ branes in 45-direction are identified with the Coulomb moduli of $\mathcal{N}=2$ gauge theory (two coordinates are combined into a single complex coordinate) and the distance between NS5 branes in 6-direction $L$ is interpreted as the gauge coupling constant:
\begin{align}
\begin{tikzpicture}[scale=1.5,baseline=(current bounding box.center)]
\begin{scope}[shift={(-3,.5)},scale=.6]
\draw[-latex] (0,0) -- ++(1,0) node [right] {6};
\draw[-latex] (0,0) -- ++(0,1) node [above] {45};
\draw[-latex] (0,0) -- ++(30:1.) node [above] {789};
\end{scope}
\draw[thick] (-1,-1) -- ++(0,2) node [above] {NS5};
\draw[thick] (1,-1) -- ++(0,2) node [above] {NS5};
\draw[latex-latex,thick,blue] (-1.,-.8) -- ++(1,0) node [below,black] {$L \propto \frac{1}{g^2}$} -- ++(1,0);
\draw (-1,.2) -- ++(2,0);
\draw (-1,.4) -- ++(2,0);
\draw (-1,.6) -- ++(2,0);
\draw [dotted,thick] (-1,-.2) -- ++(2,0);
\draw [dotted,thick] (-1,-.4) -- ++(2,0);
%
\draw [thick,decorate,decoration={brace,amplitude=4pt,mirror,raise=4pt},yshift=0pt] (1.1,.1) -- ++(0,.6) node [black,midway,xshift=2.6em] {$n_0$ D4$^+$};
\draw [thick,decorate,decoration={brace,amplitude=4pt,mirror,raise=4pt},yshift=0pt] (1.1,-.5) -- ++(0,.4) node [black,midway,xshift=2.6em] {$n_1$ D4$^-$};
\end{tikzpicture}
\end{align}
Let us remark an ambiguity of the diagram presented above.
Since we have two different types of branes, we have possibly different configurations, for example, for U($2|1$) theory, as follows,
\begin{align}
\begin{tikzpicture}[baseline=(current bounding box.center),scale=1.]
\draw[thick] (0,0) -- ++(0,2);
\draw[thick] (1.8,0) -- ++(0,2);
\draw (0,1.5) -- ++(1.8,0);
\draw (0,1) -- ++(1.8,0);
\draw[dotted,thick] (0,.5) -- ++(1.8,0);
\draw (3,1.25) -- ++(0,-.5);
\draw (-.5,1.65) -- ++(.5,0);
\draw (1.8,1.35) -- ++(.5,0);
\draw (-.5,1.15) -- ++(.5,0);
\draw (1.8,.85) -- ++(.5,0);
\draw[dotted,thick] (-.5,.65) -- ++(.5,0);
\draw[dotted,thick] (1.8,.35) -- ++(.5,0);
\filldraw[fill=white,draw=black] (3,1.25) circle (.15);
\filldraw[fill=white,draw=black] (3,.75) circle (.15);
\draw (3,.75)++(135:.15) -- ++(-45:.3);
\draw (3,.75)++(45:.15) -- ++(-135:.3);
\begin{scope}[shift={(5,0)}]
\draw[thick] (0,0) -- ++(0,2);
\draw[thick] (1.8,0) -- ++(0,2);
\draw (0,1.5) -- ++(1.8,0);
\draw[dotted,thick] (0,1) -- ++(1.8,0);
\draw (0,.5) -- ++(1.8,0);
\draw (3,1.25) -- ++(0,-.5);
\draw (-.5,1.65) -- ++(.5,0);
\draw (1.8,1.35) -- ++(.5,0);
\draw[dotted,thick] (-.5,1.15) -- ++(.5,0);
\draw[dotted,thick] (1.8,.85) -- ++(.5,0);
\draw (-.5,.65) -- ++(.5,0);
\draw (1.8,.35) -- ++(.5,0);
\filldraw[fill=white,draw=black] (3,1.25) circle (.15);
\filldraw[fill=white,draw=black] (3,.75) circle (.15);
\draw (3,1.25)++(135:.15) -- ++(-45:.3);
\draw (3,1.25)++(45:.15) -- ++(-135:.3);
\draw (3,.75)++(135:.15) -- ++(-45:.3);
\draw (3,.75)++(45:.15) -- ++(-135:.3);
\end{scope}
%
\begin{scope}[shift={(10,0)}]
\draw[thick] (0,0) -- ++(0,2);
\draw[thick] (1.8,0) -- ++(0,2);
\draw[dotted,thick] (0,1.5) -- ++(1.8,0);
\draw (0,1) -- ++(1.8,0);
\draw (0,.5) -- ++(1.8,0);
\draw (3,1.25) -- ++(0,-.5);
\draw[dotted,thick] (-.5,1.65) -- ++(.5,0);
\draw[dotted,thick] (1.8,1.35) -- ++(.5,0);
\draw (-.5,1.15) -- ++(.5,0);
\draw (1.8,.85) -- ++(.5,0);
\draw (-.5,.65) -- ++(.5,0);
\draw (1.8,.35) -- ++(.5,0);
\filldraw[fill=white,draw=black] (3,1.25) circle (.15);
\filldraw[fill=white,draw=black] (3,.75) circle (.15);
\draw (3,1.25)++(135:.15) -- ++(-45:.3);
\draw (3,1.25)++(45:.15) -- ++(-135:.3);
\end{scope}
\end{tikzpicture}
\label{eq:HW_U(2|1)}
\end{align}
This ambiguity corresponds to that for the simple roots of Lie superalgebra.
We also present the corresponding Dynkin diagrams aside the brane diagrams.
See also Sec.~\ref{sec:supermatrix} for a related discussion.
\subsubsection{Gauging trick}\label{sec:gauging_trick}
Starting with the Hanany--Witten configuration for supergroup gauge theory, we in addition include infinitely extended D4$^+$ branes.
Since they have infinite length, their couplings are zero (non-dynamical).
Then, we consider gauging this configuration:
The middle part of D4$^+$ branes can be moved in 45-direction, while the external parts are still frozen, that plays a role of the flavor brane.
Further tuning the position of D4$^+$ branes, one can remove D4$^-$ branes via pair annihilation between D4$^+$ and D4$^-$ branes, and the resulting configuration does not involve D4$^-$ branes any longer:
\begin{align}
\begin{tikzpicture}[scale=1.2,baseline=(current bounding box.center)]
%
\draw[thick] (-1,-1) -- ++(0,2) node [above] {}
\draw[thick] (1,-1) -- ++(0,2) node [above] {}
%
\draw (-1,.2) -- ++(2,0);
\draw (-1,.4) -- ++(2,0);
\draw (-1,.6) -- ++(2,0);
%
\draw [dotted,thick] (-1,-.1) -- ++(2,0);
\draw [dotted,thick] (-1,-.3) -- ++(2,0);
%
\draw (-1.5,-.6) -- ++(3,0);
\draw (-1.5,-.8) -- ++(3,0);
%
\draw[-latex,blue,very thick] (2,0) -- ++(.5,0) node [above] {} -- ++(.5,0);
%
\begin{scope}[shift={(5,0)}]
\draw[thick] (-1,-1) -- ++(0,2) node [above] {}
\draw[thick] (1,-1) -- ++(0,2) node [above] {}
%
\draw (-1,.2) -- ++(2,0);
\draw (-1,.4) -- ++(2,0);
\draw (-1,.6) -- ++(2,0);
%
\draw [dotted,thick] (-1,-.1) -- ++(2,0);
\draw [dotted,thick] (-1,-.3) -- ++(2,0);
%
\draw (-1.5,-.6) -- ++(.5,0);
\draw (-1.5,-.8) -- ++(.5,0);
\draw (1,-.6) -- ++(.5,0);
\draw (1,-.8) -- ++(.5,0);
%
\draw (-1,-.5) -- ++(2,0);
\draw (-1,-.7) -- ++(2,0);
\draw[-latex,blue,very thick] (2,0) -- ++(.5,0) node [above] {} -- ++(.5,0); \end{scope}
%
\begin{scope}[shift={(10,0)}]
\draw[thick] (-1,-1) -- ++(0,2) node [above] {}
\draw[thick] (1,-1) -- ++(0,2) node [above] {}
%
\draw (-1,.2) -- ++(2,0);
\draw (-1,.4) -- ++(2,0);
\draw (-1,.6) -- ++(2,0);
%
\draw (-1.5,-.6) -- ++(.5,0);
\draw (-1.5,-.8) -- ++(.5,0);
\draw (1,-.6) -- ++(.5,0);
\draw (1,-.8) -- ++(.5,0);
\end{scope}
\end{tikzpicture}
\end{align}
This configuration is identical to U$(n_0)$ gauge theory with $n_F = 2n_1$ flavor degrees of freedom~\cite{Dijkgraaf:2016lym}.
This process implies that the (Coulomb branch of) moduli space of vacua of supergroup gauge theory has an intersection with that for the ordinary $\mathcal{N}=2$ gauge theory with flavors (SQCD).
\subsubsection{$\widehat{A}_1$ quiver realization}
As discussed in Sec.~\ref{sec:quiver_realization}, supergroup gauge theory has a realization as $\widehat{A}_1$ quiver gauge theory.
In fact, $\widehat{A}_1$ quiver theory is similarly realized in this framework.
In this case, we compactify 6-direction and put D4 branes in this direction.
Then, we have two domains suspended by NS5 branes, that realize two distinct gauge nodes.
This configuration can be converted to the previous configurations through analytic continuation as follows.
We recall that supergroup gauge theory can be obtained by tuning the couplings as in~\eqref{eq:coupling_tuning_A1} in $\widehat{A}_1$ quiver theory, and in this setup, each coupling $1/g_\sigma^2$ is interpreted as length of D4 branes between NS5 branes, $L_\sigma \propto 1/g_\sigma^2$.
From this point of view, a negative brane would be interpreted as a brane with negative length, that would be thought of as analytic continuation:
\begin{align}
\begin{tikzpicture}[baseline=(current bounding box.center)]
\draw (0,0) arc [x radius = 1.5, y radius = 1, start angle = 120, end angle = -60];
\draw (0,-.5) arc [x radius = 1.5, y radius = 1, start angle = 120, end angle = -60];
\draw (0,-1) arc [x radius = 1.5, y radius = 1, start angle = 120, end angle = -60];
\draw (0,-.25) arc [x radius = 1.5, y radius = 1, start angle = 120, end angle = 300];
\draw (0,-.75) arc [x radius = 1.5, y radius = 1, start angle = 120, end angle = 300];
%
\draw[blue,latex-latex,thick] (0,-1.5) arc [x radius = 1.5, y radius = 1, start angle = 120, end angle = -60];
\draw[blue,latex-latex,thick] (0,-1.5) arc [x radius = 1.5, y radius = 1, start angle = 120, end angle = 300];
\draw[thick] (0,-1.75) -- ++(0,2) node [above] {NS5};
\draw[thick] (0,-1.75)++(1.5,-1.73) -- ++(0,2);
%
\node at (1.8,.3) [right] {D4};
\node at (2.5,-2.) [right] {$L_0$};
\node at (-1,-2.7) [left] {$L_1$};
%
\draw[very thick,magenta,latex-latex] (4,-1.5) -- ++(2,0);
%
\begin{scope}[shift = {(8,0)}]
\draw (0,0) arc [x radius = 1.5, y radius = 1, start angle = 120, end angle = -60];
\draw (0,-.5) arc [x radius = 1.5, y radius = 1, start angle = 120, end angle = -60];
\draw (0,-1) arc [x radius = 1.5, y radius = 1, start angle = 120, end angle = -60];
\draw[dashed] (0,-.25) arc [x radius = 1.5, y radius = 1, start angle = 120, end angle = -60];
\draw[dashed] (0,-.75) arc [x radius = 1.5, y radius = 1, start angle = 120, end angle = -60];
%
\draw[blue,latex-latex,thick] (0,-1.5) arc [x radius = 1.5, y radius = 1, start angle = 120, end angle = -60];
\draw[thick] (0,-1.75) -- ++(0,2) node [above] {NS5};
\draw[thick] (0,-1.75)++(1.5,-1.73) -- ++(0,2);
%
\node at (1.8,.3) [right] {D4$^\pm$};
\node at (2.5,-2.) [right] {$L_0$};
%
\end{scope}
\end{tikzpicture}
\end{align}
Let us comment a possible connection to the gauging trick discussed above.
Recalling that the supergroup condition~\eqref{eq:coupling_tuning_A1}, namely $L_0 + L_1 = 0$, it would correspond to the shrinking limit in 6-direction.
In fact, this is the strong coupling limit, that also implies necessity of non-perturbative treatment in supergroup gauge theory.
Then, applying the T-dual in 6-direction, this direction is infinitely extended, and the gauge coupling of either first or second gauge node has to be zero.
This means that the resulting configuration is a single node gauge theory with a flavor node, which is consistent with the gauging trick.
This argument is extended to supergroup quiver gauge theory in general.
See~\cite{Kimura:2019msw,Kimura:2020jxl} for details.
\subsubsection{Seiberg--Witten theory}\label{sec:SW_theory}
Seiberg--Witten theory provides a geometric description of the moduli space of supersymmetric vacua of four-dimensional $\mathcal{N}=2$ gauge theory.
In fact, one can concisely extract the Seiberg--Witten geometry from the Hanany--Witten brane configuration~\cite{Witten:1997sc}.
In this setup, we specify the positions of D4 branes and NS5 branes using two complex variables $(x,y) \in \mathbb{C} \times \mathbb{C}^\times$ as follows,
\begin{align}
y + \frac{\mathfrak{q}}{y} = \det(x - \Phi)
\, ,
\end{align}
where we denote the exponentiated complexified gauge coupling by $\mathfrak{q} = \exp (2 \pi \mathrm{i} \tau)$, $\tau = \frac{\theta}{2\pi} + \frac{4 \pi \mathrm{i}}{g^2}$, with the $\theta$-angle, and $\Phi$ is the adjoint complex scalar field in the vector multiplet that parametrizes the Coulomb branch of the moduli space.
In order to obtain the Seiberg--Witten geometry of supergroup gauge theory, we replace the characteristic polynomial with the supercharacteristic function,
\begin{align}
y + \frac{\mathfrak{q}}{y} = \sdet(x - \Phi)
\, .
\end{align}
Hence, we have the Seiberg--Witten curve for $\mathrm{U}(n|m)$ pure SYM theory as follows.
\begin{itembox}{Seiberg--Witten curve for $\mathrm{U}(n|m)$ theory}
The Seiberg--Witten curve for $\mathrm{U}(n|m)$ pure SYM theory is given by the supercharacteristic function of the complex adjoint scalar $\Phi \in \operatorname{Lie} \mathrm{U}(n|m)$ in the vector multiplet,
\begin{align}
\Sigma = \{ (x,y) \in \mathbb{C} \times \mathbb{C}^\times \mid y + \frac{\mathfrak{q}}{y} = \sdet(x - \Phi) \}
\, .
\end{align}
\end{itembox}
In terms of the eigenvalues of $\Phi$, $\{ a_\alpha^\sigma \}_{\sigma=0,1,\alpha \in [n_\sigma]}$ for $G = \mathrm{U}(n_0|n_1)$, we may write it as follows,
\begin{align}
y + \frac{\mathfrak{q}}{y} = \frac{T_0(x)}{T_1(x)}
\, , \qquad
T_\sigma(x) = \prod_{\alpha \in [n_\sigma]} (x - a_\alpha^\sigma)
\, ,
\end{align}
which can be further rewritten as
\begin{align}
T_1(x) y + \mathfrak{q} \frac{T_1(x)}{y} = T_0(x)
\, .
\end{align}
This algebraic equation characterizes the Seiberg--Witten curve of U$(n_0)$ gauge theory with $2 n_1$ flavors, which is consistent with the gauging trick.
This algebraic equation shows the Riemann surface of genus $g = n_0 - 1$ with $2 n_1$ punctures.
Imposing the special unitary condition, $\sum_{\alpha \in [n_1]} a_\alpha^1 = 0$, the last two cycles are not independent, hence we have $(n_0 - 1)$ $A$ and $B$ cycles and $2(n_1 - 1)$ cycles associated with the punctures.
The Euler characteristics of this Riemann surface is given by $\chi = 2 - 2 (n_0 - 1) + 2 (n_1 - 1) = 2 - 2 (n_0 - n_1)$, where we have the superdimension $\operatorname{sdim} \mathbb{C}^{n_0|n_1} = n_0 - n_1$.
This relation is also obtained from $\widehat{A}_1$ quiver point of view~\cite{Dijkgraaf:2016lym}.
We can also incorporate the superflavor factor by imposing the supercharacterisitic function with respect to $G_F = \mathrm{U}(n_0^F|n_1^F)$.
We will derive the Seiberg--Witten geometry for supergroup gauge theory from the microscopic instanton counting in Sec.~\ref{sec:instanton}.
\section{Supergroup instanton counting}\label{sec:instanton}
As in the case of the ordinary non-supergroup gauge theory, the instanton plays an important role in the study of non-perturbative aspects of supergroup gauge theory.
In this section, we explore several aspects of instantons and their applications to study dynamics of supergroup gauge theory.
\subsection{Instanton moduli space}
The instanton is a solution of the ASDYM equation~\eqref{eq:ASDYMeq}.
Writing the SD/ASD part of the curvature two-form as
\begin{align}
F_\pm = \frac{F \pm \star F}{2}
\, ,
\end{align}
such that $F = F_+ + F_-$, the ASDYM equation is rewritten as
\begin{align}
F_+ = 0
\, ,
\end{align}
which would be thought of as a ``half'' flat connection, and hence the corresponding moduli space is expected to have a rich mathematical structure.
Let $M = \mathbb{C}^2$ and $G = \mathrm{U}(n)$.
In order that the YM action becomes finite, we require that the curvature behaves as $F \to 0$ as $x \to \infty$.
This implies that the connection approaches to the pure gauge at the boundary, $A \xrightarrow{x \to \infty} g d g^{-1}$, where we obtain a map, $g$: $\partial \mathbb{C}^2 = \mathbb{S}^3 \to G$.
Knowing that $\pi_3(G) = \mathbb{Z}$ for any compact simple Lie groups, we can apply a classification on the gauge field with respect to the topological charge
\begin{align}
k = \frac{1}{8 \pi^2} \int_M \tr F \wedge F = c_2[M] \in \mathbb{Z}
\, ,
\end{align}
which is called the instanton number, given by the integral of the second Chern class over $M$.%
\footnote{%
Precisely speaking, the four-manifold $M$ should be compact to have $c_2[M] \in \mathbb{Z}$.
In this case, $\mathbb{C}^2 \simeq \mathbb{R}^4$ can be mapped to a four-sphere $\mathbb{S}^4$ through the stereographic projection.
This map does not violate the ASD property since it is a conformal map.
Under this map, the boundary of $\mathbb{C}^2$ is mapped to the north pole (marked point) on $\mathbb{S}^4$, where we fix the gauge as $A = gdg^{-1}$ (called the framing).
}
Here, we focus on the ASD instanton solution having a positive charge $k \ge 0$.
The SD anti-instanton solution is similarly obtained by changing the orientation of the manifold.
Hence, the moduli space of instantons has a decomposition
\begin{align}
\mathfrak{M}_G = \bigsqcup_{k=0}^\infty \mathfrak{M}_{G,k}
\, ,
\end{align}
where each topological sector of the moduli space is given by
\begin{align}
\mathfrak{M}_{G,k} = \{ A \in \Omega^1(M,\mathfrak{g}) \mid F_+ = 0, c_2[M] = k \} / \mathcal{G}
\, .
\label{eq:mod_sp_def_naive}
\end{align}
We remark that the $\mathcal{G}$-quotient means the quotient with respect to the equivalent class under $G$-gauge transformation.
\subsection{ADHM construction of instanton}\label{sec:ADHM_const}
Although the definition of the moduli space~\eqref{eq:mod_sp_def_naive} is conceptually reasonable, it is still difficult to analyze in practice.
In order to discuss such an instanton moduli space, we can apply the ADHM construction, which is a systematic approach to obtain instanton solutions~\cite{Atiyah:1978ri}.
See, e.g.,~\cite{Dorey:2002ik} for an extended review on this subject.
For the $k$-instanton solution in $\mathrm{U}(n)$-YM theory, we define two vector spaces,
\begin{align}
N = \mathbb{C}^n
\, , \qquad
K = \mathbb{C}^k
\, .
\end{align}
The fundamental degrees of freedom in this construction are the linear maps associated with these vector spaces,
\begin{align}
X & = \operatorname{Hom}(K,K) \oplus \operatorname{Hom}(K,K) \oplus \operatorname{Hom}(N,K) \oplus \operatorname{Hom}(K,N)
\nonumber \\
& \ni (B_1, B_2, I, J)
\, ,
\end{align}
where we impose the so-called ADHM equations,
\begin{subequations}\label{eq:ADHMeqs}
\begin{align}
0 & = \mu_\mathbb{R} := [B_1, B_1^\dag] + [B_2, B_2^\dag] + II^\dag - J^\dag J
\, , \\
0 & = \mu_\mathbb{C} := [B_1, B_2] + IJ
\, .
\end{align}
\end{subequations}
We remark that these equations are invariant under the $\mathrm{U}(k)$ action,
\begin{align}
g \cdot (B_1, B_2, I, J) = (g B_1 g^{-1}, g B_2 g^{-1}, g I, J g^{-1})
\, , \qquad
g \in \mathrm{U}(k)
\, .
\end{align}
We call $\mu_{\mathbb{R},\mathbb{C}}$ the \emph{moment maps} that take a value in the dual of Lie algebra, $\mu_{\mathbb{R},\mathbb{C}} :$ $X \to \mathfrak{u}_k^\vee \otimes \mathbb{R}^3$.
Based on these degrees of freedom, we have an alternative description of the instanton moduli space,
\begin{align}
\mathfrak{M}_{G,k} = \{ (B_1,B_2,I,J) \in X \mid \mu_\mathbb{R} = 0, \mu_\mathbb{C} = 0 \} /\!\!/\!\!/ \mathrm{U}(k)
\, .
\end{align}
The triple slash means the hyper-Kähler quotient with respect to three moment maps, $(\mu_\mathbb{R}, \operatorname{Re} \mu_\mathbb{C}, \operatorname{Im} \mu_\mathbb{C})$.
We show that the ASD connection can be constructed from these ADHM variables.
We define the zero-dimensional Dirac operator,
\begin{align}
D^\dag =
\begin{pmatrix}
B_1 - z_1 & B_2 - z_2 & I \\
- B_2^\dag + \bar{z}_2 & B_1^\dag - \bar{z}_1 & - J^\dag
\end{pmatrix}
\ : \ K \otimes \mathbb{C}^2 \oplus N \to K \otimes \mathbb{C}^2
\, .
\end{align}
We remark that the space-time dependence in the Dirac operator is associated with the quaternion structure
\begin{align}
\begin{pmatrix}
z_1 & z_2 \\ - \bar{z}_2 & \bar{z}_1
\end{pmatrix}
=
\begin{pmatrix}
x_4 + \mathrm{i} x_3 & x_2 + \mathrm{i} x_1 \\ - x_2 + \mathrm{i} x_1 & x_4 - \mathrm{i} x_3
\end{pmatrix}
= x \cdot \sigma
\, , \quad
\end{align}
where we define the quarternion basis,
\begin{align}
\sigma = (\mathrm{i} \vec{\sigma}, \mathbbm{1}_{\mathbb{C}^2})
\, , \qquad
\bar{\sigma} = (-\mathrm{i} \vec{\sigma}, \mathbbm{1}_{\mathbb{C}^2})
\, .
\end{align}
Their product is given by
\begin{align}
\sigma_\mu \bar{\sigma}_\nu = \delta_{\mu\nu} + \mathrm{i} \eta_{\mu\nu}^{(+)}
\, , \qquad
\bar\sigma_\mu {\sigma}_\nu = \delta_{\mu\nu} + \mathrm{i} \eta_{\mu\nu}^{(-)}
\, ,
\end{align}
where $\eta_{\mu\nu}^{(\pm)}$ is an (A)SD tensor, called the 't Hooft symbol, $\star \eta_{\mu\nu}^{(\pm)} = \pm \eta_{\mu\nu}^{(\pm)}$.
The ADHM equations~\eqref{eq:ADHMeqs} are equivalent to the condition such that the Dirac operator squared is diagonal with respect to $\mathbb{C}^2$,
\begin{align}
D^\dag D = \Delta \otimes \mathbbm{1}_{\mathbb{C}^2}
\, ,
\end{align}
where $\Delta$ : $K \to K$ is explicitly given by
\begin{align}
\Delta & =
(B_1 - z_1)(B_1^\dag - \bar{z}_1) + (B_2 - z_2)(B_2^\dag - \bar{z}_2) + II^\dag
\nonumber \\
& =
(B_1^\dag - \bar{z}_1)(B_1 - z_1) + (B_2^\dag - \bar{z}_2)(B_2 - z_2) + J^\dag J
\, .
\end{align}
Therefore, the asymptotic behavior is given by $\Delta \xrightarrow{z \to \infty} |z|^2 \mathbbm{1}_K$.
Since the Dirac operator is a rectangular matrix of a rank $2k$, we consider the complementary space, which is given by the set of normalized zero modes (the Dirac operator kernel),
\begin{align}
\operatorname{Ker} D^\dag = \{ \Psi : N \to K \otimes \mathbb{C}^2 \oplus N \mid D^\dag \Psi = 0, \Psi^\dag \Psi = \mathbbm{1}_N \}
\, .
\end{align}
We define the projector from $K \otimes \mathbb{C}^2 \oplus N$ to $N$ using the zero modes,
\begin{align}
P := \Psi \Psi^\dag = \mathbbm{1}_{K \otimes \mathbb{C}^2 \oplus N} - D (\Delta^{-1} \otimes \mathbbm{1}_{\mathbb{C}^2}) D^\dag
\, ,
\end{align}
where we remark $PD = \Psi \Psi^\dag D = 0$.
Then, we obtain a connection from the zero modes,
\begin{align}
A = \Psi^\dag d \Psi
\, .
\end{align}
Due to the normalization condition, it turns out to be anti-Hermitian, $A^\dag = - A$.
The curvature two-form constructed from this connection is given as follows,
\begin{align}
F = dA + A \wedge A = d \Psi^\dag (\mathbbm{1}_{K \otimes \mathbb{C}^2 \oplus N} - \Psi \Psi^\dag) d \Psi = \Psi^\dag (d D) (\Delta^{-1} \otimes \mathbbm{1}_{\mathbb{C}^2}) (d D^\dag) \Psi
\, .
\end{align}
Recalling
\begin{align}
d D =
\begin{pmatrix}
- \mathbbm{1}_K \otimes \bar{\sigma} \\ 0
\end{pmatrix}
\, , \qquad
d D^\dag =
\begin{pmatrix}
- \mathbbm{1}_K \otimes \sigma & 0
\end{pmatrix}
\, ,
\end{align}
it turns out that the curvature is ASD,
\begin{align}
F = \Psi^\dag
\begin{pmatrix}
\Delta^{-1} \otimes 2 \mathrm{i} \eta^{(-)} & 0 \\ 0 & 0
\end{pmatrix}
\Psi
\, ,
\end{align}
where we define the (A)SD two-form $\eta^{(\pm)} = \frac{1}{2} \eta^{(\pm)}_{\mu\nu} dx^\mu dx^\nu$, such that $\star \eta^{(\pm)} = \pm \eta^{(\pm)}$.
In addition, applying Osborn's formula~\cite{Osborn:1978rn}, we see that the instanton number is given as follows,
\begin{align}
\frac{1}{8\pi^2} \int_M \tr F \wedge F = \frac{1}{16 \pi^2} \int_M \mathrm{d}[4]{x} \partial^2 \partial^2 \tr_K \log \Delta^{-1} = \tr_K \mathbbm{1}_K = k
\, .
\label{eq:Osborn_formula}
\end{align}
In this calculation, we use the asymptotic behavior of $\Delta \xrightarrow{x \to \infty} |x|^2 \mathbbm{1}_K$ and evaluate the boundary contribution.
\subsubsection{String theory perspective}\label{sec:ADHM_string_perspective}
As discussed in Sec.~\ref{sec:string_theory_perspective}, one can obtain gauge theory from a stack of D-branes.
In particular, $k$-instanton configuration in $\mathrm{U}(n)$-YM theory is realized as D$p$-D$(p-4)$ brane system.
Let us put $p = 4$.
In this context, the two vector spaces $(N,K)$ are identified with the Chan--Paton spaces of $n$ D4 and $k$ D0 branes, and thus the ADHM variables $(B_1,B_2,I,J)$ are interpreted as degrees of freedom of open strings connecting D0-D0, D4-D0, and D0-D4 branes, respectively.
The ADHM equations are obtained as the BPS equations of this configuration.
We remark that the anti-instanton is realized by $\overline{\text{D}0}$ brane that violates the BPS condition.
\subsubsection{Regularization of moduli space}
We would study the moduli space defined here, however it has been known that it is non-compact and singular that should be reguralized for the latter purpose.
For the sake of compactification of the instanton moduli space, we add the point-like instantons, $\bigsqcup_{\ell=0}^k \mathfrak{M}_{n,k-\ell} \times \operatorname{Sym}^\ell \mathbb{C}^2$, which is called the \emph{Uhlenbeck compactification} (see, e.g.,~\cite{Donaldson:1997}).
This is analogous to the compactification of $\mathbb{R}^4$ to $\mathbb{S}^4$ by adding the infinity.
Another regularization that we need is resolution of singularity.
This can be done by modifying the ADHM equations, $(\mu_\mathbb{R}, \mu_\mathbb{C}) = (\zeta_{\neq 0} \mathbbm{1}_K, 0)$, and the corresponding moduli space is given by~\cite{Nakajima:1993jg}
\begin{align}
\mathfrak{M}_{n,k}^\zeta = \{ (B_1, B_2, I, J) \in X \mid \mu_\mathbb{R} = \zeta_{\neq 0} \mathbbm{1}_K, \mu_\mathbb{C} = 0 \}/\!\!/\!\!/\mathrm{U}(k)
\, .
\end{align}
This modification has been also discussed in the context of the instanton moduli space on non-commutative $\mathbb{C}^2$~\cite{Nekrasov:1998ss}.
Another moduli space is for the rank-$n$ framed torsion free sheaves on $\mathbb{CP}^2$ given by the geometric invariant theory (GIT) quotient (see, e.g.,~\cite{Nakajima:1999}),
\begin{align}
\widetilde{\mathfrak{M}}_{n,k} = \{ (B_1, B_2, I, J) \in X \mid \mu_\mathbb{C} = 0, \text{(co-)stability} \}/\!\!/\mathrm{GL}(K)
\, ,
\end{align}
which is known to be isomorphic to the resolved moduli space $\mathfrak{M}_{n,k}^\zeta$.
The (co-)stability condition is given as follows,
\begin{align}
K =
\begin{cases}
\mathbb{C}[B_1, B_2] I(N) & (\zeta > 0;\ \text{stability}) \\
\left( \mathbb{C}[B_1, B_2] J(N) \right)^\vee & (\zeta < 0;\ \text{co-stability}) \\
\end{cases}
\end{align}
One can obtain the (co-)stability condition from the real part of the ADHM equation, $\mu_\mathbb{R} = \zeta \mathbbm{1}_K$, depending on the sign of the deformation parameter $\zeta$.
See, e.g.,~\cite{Kimura:2020jxl} for details.
\subsection{ADHM construction of super instanton}
We turn to the construction of instantons in supergroup gauge theory.
Let $G = \mathrm{U}(n_0|n_1)$.
Recalling that the ADHM moduli space is given by the $\mathrm{U}(k)$-quotient for the $k$-instanton sector, it seems natural to consider the supergroup quotient, i.e., $\mathrm{U}(k_0|k_1)$-quotient in the supergroup setup.
This implies that the topological sector is characterized by $k = (k_0|k_1)$ in this case.
From this point of view, we replace the vector spaces $(N,K)$
with the supervector spaces,
\begin{align}
N = N_0 \oplus N_1 = \mathbb{C}^{n_0|n_1} = \mathbb{C}^{n_0} \oplus \mathbb{C}^{n_1}
\, , \qquad
K = K_0 \oplus K_1 = \mathbb{C}^{k_0|k_1} = \mathbb{C}^{k_0} \oplus \mathbb{C}^{k_1}
\, .
\label{eq:N_K_svec_sp}
\end{align}
Then, the ADHM variables are obtained as supermatrices representing the linear maps for these supervector spaces.
The remaining steps are formally parallel with the standard case discussed in Sec.~\ref{sec:ADHM_const}.
In this case, Osborn's formula~\eqref{eq:Osborn_formula} yields
\begin{align}
\frac{1}{8\pi^2} \int_M \str F \wedge F = \str_K \mathbbm{1}_K = k_0 - k_1
\, ,
\end{align}
which implies that $k_0$ and $k_1$ count the positive and negative charge instantons.
This is reasonable from string theory perspective discussed in Sec.~\ref{sec:ADHM_string_perspective}:
The current $(k_0|k_1)$-instanton configuration in $\mathrm{U}(n_0|n_1)$-YM theory would be realized as $k_0$ D0$^+$, $k_1$ D0$^-$, $n_0$ D4$^+$, and $n_1$ D4$^-$ brane system, which relate the supervector spaces~\eqref{eq:N_K_svec_sp} to the Chan--Paton spaces.
In contrast to the brane--anti-brane system, the current situation does not violate the ASD property, which is compatible with the BPS equation.
Meanwhile, the resolved moduli space is given for $(n,k) = (n_0|n_1,k_0|k_1)$ by
\begin{align}
\mathfrak{M}_{n,k}^\zeta = \{ (B_1, B_2, I, J) \in X \mid \mu_\mathbb{R} = (+\zeta \mathbbm{1}_{K_0}) \oplus (-\zeta \mathbbm{1}_{K_1}), \mu_\mathbb{C} = 0 \}/\!\!/\!\!/\mathrm{U}(k_0|k_1)
\, .
\end{align}
We remark that the deformation parameter is assigned with opposite signs for the positive and negative sectors.
Physically speaking, this is because the positive and negative instantons (positive and negative D0 branes) are oppositely charged under the flux yielding the non-commutativity of the space-time manifold via the Seiberg--Witten map~\cite{Seiberg:1999vs}.
Hence, fixing $\zeta > 0$, this moduli space is isomorphic to the following super-GIT quotient,
\begin{align}
\widetilde{\mathfrak{M}}_{n,k} = \{ (B_1, B_2, I, J) \in X \mid \mu_\mathbb{C} = 0, \text{stability for $K_0$, co-stability for $K_1$} \}/\!\!/\mathrm{GL}(K)
\, .
\end{align}
For $\zeta < 0$, the stability and co-stability conditions are exchanged.
\subsection{Equivariant localization}
Based on the description of the instanton moduli space, we apply the equivariant localization formalism to compute the partition function of supergroup gauge theory.
See, e.g., \cite{Pestun:2016zxk} for details of the localization calculus.
In the Euclidean path integral formalism, the partition function of $G$-gauge theory is given as follows,
\begin{align}
Z = \int \mathfrak{D}\!A \, \mathrm{e}^{-S[A]}
\, ,
\end{align}
where we consider the case $M = \mathbb{R}^4$, and the action consists of the YM action and the $\theta$-term.
Writing the one-form connection in the form of $A = A_k + \delta A$, where we denote the $k$-instanton configuration by $A_k$ and the deviation by $\delta A$ (we assume no $k$-dependence), we have the following decomposition of the path integral measure, $\mathfrak{D} A = \sum_{k=0}^\infty \mathfrak{D} A_k \mathfrak{D}(\delta A)$.
In this decomposition, the weight with the action is written by $\mathrm{e}^{-S[A]} = \mathfrak{q}^k \mathrm{e}^{-S_\text{dev}[\delta A]}$, where $\mathfrak{q} = \exp (\mathrm{i} \theta - \frac{8 \pi^2}{g^2}) =: \exp (2 \pi \mathrm{i} \tau)$, and the deviation part $S_\text{dev}[\delta A]$ starts with $O(\delta A^2)$ since $A_k$ gives a solution to the equation of motion.
Noticing that the integral over $A_k$ can be identified with the integral over the moduli space $\mathfrak{M}_{G,k}$, we have the following form of the partition function,%
\footnote{%
The anti-instanton partition function shall be obtained by the complex conjugate of the instanton partition function:
The weight in this case is given by $\bar{\mathfrak{q}}^k = \exp k\left(-\mathrm{i} \theta - \frac{8 \pi^2}{g^2} \right)$, and the moduli space integral will be $\overline{Z_k}$ since it has an opposite orientation compared with the original one.
}
\begin{align}
Z = Z_\text{inst} Z_\text{pert}
\, , \qquad
Z_\text{inst} = \sum_{k = 0}^\infty \mathfrak{q}^k Z_k
\, ,
\end{align}
where $Z_\text{pert}$ is the deviation (perturbative) part, while $Z_k$ is called the \emph{instanton partition function},
\begin{align}
Z_k
= \int_{\mathfrak{M}_{G,k}} 1 = \operatorname{vol} (\mathfrak{M}_{G,k})
\, .
\label{eq:Zk_as_volume}
\end{align}
As mentioned in Sec.~\ref{sec:ADHM_const}, we should replace the moduli space $\mathfrak{M}_{G,k}$ with the regularized version.
Moreover, noticing that there exist several group actions on the moduli space, the integral is performed as the equivariant integral, and thus $\operatorname{vol} (\mathfrak{M}_{G,k})$ is understood as the equivariant volume, which can be computed based on the equivariant localization formula,
\begin{align}
\int_{\mathfrak{M}_{G,k}} \alpha = \sum_{\lambda \in \mathfrak{M}_{G,k}^\textsf{T}} \frac{\iota^* \alpha}{e(T_\lambda \mathfrak{M}_{G,k})}
\, ,
\label{eq:localization_formula}
\end{align}
where $\alpha$ is an equivariant cohomology class, $\iota^*$ is a pull-back of the inclusion map, $\iota$ : $o \times \mathfrak{M}_{G,k} \hookrightarrow \mathbb{C}^2 \times \mathfrak{M}_{G,k}$, $e$ is the equivariant Euler class, $\mathfrak{M}_{G,k}^\mathsf{T}$ is a set of the equivariant $\mathsf{T}$-fixed points, and $T_\lambda \mathfrak{M}_{G,k}$ is the tangent bundle to the moduli space at the fixed point $\lambda$.
The instanton partition function corresponds to the trivial insertion $\alpha = 1$.
In the presence of the fundamental matter, we add the equivariant Euler class of the matter bundle $\mathsf{M}$, whose fiber is the space of virtual zero modes of the Dirac operator in the instanton background.
For $\mathcal{N}=2^*$ theory involving the adjoint matter, we insert the shifted tangent bundle, $\mathsf{M}_\text{adj} \otimes T\mathfrak{M}_{G,k}$, where $\mathsf{M}_\text{adj}$ is a line bundle whose Chern root is identified with the adjoint mass parameter.%
\footnote{%
In the five-dimensional (equivariant K-theory) convention discussed below, this case corresponds to Hirzebruch's $\chi_y$-genus of the instanton moduli space with identifying $y = \mathrm{e}^m$.
}
\subsubsection{Index functor}
The formalism discussed above is generalized to five and six dimensional theories by replacing the equivariant Euler class with the corresponding multiplicative index functor:
For the vector bundle $\mathbf{X}$ with the character, $\operatorname{ch} \mathbf{X} = \sum_{i=1}^{\operatorname{rk} \mathbf{X}} n_i \mathrm{e}^{x_i}$ with multiplicity $n_i \in \mathbb{Z}$, we define
\begin{align}
\mathbb{I}[\mathbf{X}] = \prod_{i \in [\operatorname{rk} \mathbf{X}]} [x_i]^{n_i}
\, ,
\end{align}
where we apply the notation,%
\footnote{%
Not to be confused with the notation~\eqref{eq:set_N}.
}
\begin{align}
[x] =
\begin{cases}
x & (4d) \\
(1 - \mathrm{e}^{-x}) & (5d) \\
\theta(e^{-x};p) & (6d)
\end{cases}
\label{eq:index_notation}
\end{align}
The instanton partition function is then given by summation over the topological sectors,
\begin{align}
Z_k = \sum_{\lambda \in \mathfrak{M}_{G,k}^\mathsf{T}} Z_\lambda
\, , \qquad
Z_\lambda = \mathbb{I}[- T_\lambda \mathfrak{M}_{G,k}]
\, .
\end{align}
The five and six dimensional conventions correspond to the equivariant K-theory and the equivariant elliptic cohomology, respectively.
The five-dimensional index is given by the character of alternating sum of anti-symmetrizations of the dual bundle,%
\footnote{%
For the vector bundle $\mathbf{X}$ with the character, $\operatorname{ch} \mathbf{X} = \sum_{i=1}^{\operatorname{rk} \mathbf{X}} n_i \mathrm{e}^{x_i}$, the dual bundle is defined to have the character, $\operatorname{ch} \mathbf{X}^\vee = \sum_{i=1}^{\operatorname{rk} \mathbf{X}} n_i \mathrm{e}^{-x_i}$.
}
\begin{align}
\mathbb{I}[\mathbf{X}] \ \stackrel{5d}{=} \ \operatorname{ch} \wedge \mathbf{X}^\vee = \sum_{i=0}^{\infty} (-1)^i \operatorname{ch} \wedge^i \mathbf{X}^\vee
\, .
\end{align}
In this case, we have the K-theoretic instanton partition function as an equivariant Euler characteristic over the instanton moduli space~\cite{Nakajima:2003pg,Nakajima:2005fg},
\begin{align}
Z_k = \sum_{i=0}^\infty (-1)^i \operatorname{ch} H^i(\mathfrak{M}_{G,k},\mathcal{O})
\, ,
\label{eq:K-theoretic_partition_function}
\end{align}
where we denote the structure sheaf of the moduli space by $\mathcal{O}$.
From the point of view of the localization formula~\eqref{eq:localization_formula}, we may rewrite the partition function as follows,%
\footnote{%
The $\widehat{A}$ genus is typically used for the five-dimensional convention instead of the Todd genus.
See, e.g.,~\cite{Nekrasov:2002qd}.
}
\begin{align}
Z_k = \int_{\mathfrak{M}_{G,k}} \operatorname{td}(T\mathfrak{M}_{G,k})
\, ,
\end{align}
where we define the Todd class,
\begin{align}
\operatorname{td}(\operatorname{\mathbf{X}}) = \prod_{i=1}^{\operatorname{rk} \mathbf{X} }\frac{x_i}{1 - \mathrm{e}^{- x_i}} = e(\mathbf{X}) \mathbb{I}[\mathbf{X}]^{-1} = e(\mathbf{X}) \mathbb{I}[-\mathbf{X}]
\, .
\end{align}
Recalling the Hirzebruch--Riemann--Roch formula for the holomorphic Euler characteristic of a holomorphic vector bundle $E$ on $\mathfrak{M}$,%
\footnote{%
We may write the K-theoretic partition function~\eqref{eq:K-theoretic_partition_function} as an alternating sum of higher direct images and the pushforward of the projection map from the moduli space to the point (derived pushforward), which can be understood as the Grothendieck--Hirzebruch--Riemann--Roch formula.
}
\begin{align}
\chi(\mathfrak{M},E) = \int_\mathfrak{M} \operatorname{ch} E \, \operatorname{td}(T\mathfrak{M})
\, ,
\end{align}
the K-theoretic partition function is understood as an integral of ``1'' analogously to the cohomological version~\eqref{eq:Zk_as_volume}.
We remark that the equivariant Euler class contribution in the denominator of the localization formula~\eqref{eq:localization_formula} is canceled with that in the numerator of the Todd class.
See also~\cite{Pestun:2016qko} for details.
For the six-dimensional case, the top form insertion is necessary to be compatible with the modular property.
Hence, for $\mathrm{U}(n)$ gauge theory, we need to incorporate $2n$ fundamental matters, or a single adjoint matter, which correspond to the superconformal matter content in the four-dimensional convention.
We remark that these index formulations are understood as the Witten index of the supersymmetric ADHM quiver quantum mechanics on the circle $\mathbb{S}^1$ and the elliptic genus of the ADHM quiver gauge theory on the elliptic curve $\mathcal{E}$ with the nome $p \in \mathbb{C}^\times$.
\subsubsection{Equivariant fixed points}
We turn to more details on the instanton moduli space.
We denote the fiber of the cotangent bundle at the marked point $o \in \mathbb{C}^2$ by
\begin{align}
Q = T_o^\vee \mathbb{C}^2 = Q_1 \oplus Q_2
\end{align}
with the character $\operatorname{ch} Q_i = q_i = \mathrm{e}^{\epsilon_i}$ ($i=1,2$).
We use the notation
\begin{subequations}
\begin{align}
Q_{12} = Q_1 \otimes Q_2 = \det Q
\, , \qquad &
q_{12} = q_1 q_2
\quad (\epsilon_{12} = \epsilon_1 + \epsilon_2)
\, , \\
P_i = \wedge Q_i = 1 - Q_i \quad (i = 1,2)
\, , \qquad &
P_{12} = P_1 P_2
\, .
\end{align}
\end{subequations}
Then, the equivariant group actions on the moduli space are given by
\begin{subequations}
\begin{align}
g \cdot (B_1, B_2, I, J) & = (g B_1 g^{-1}, g B_2 g^{-1}, g I, J g^{-1})
\, , &&
g \in \mathrm{GL}(K)
\, , \\
h \cdot (B_1, B_2, I, J) & = (B_1, B_2, I h^{-1}, h J )
\, , &&
h \in \mathrm{GL}(N)
\, , \\
(q_1,q_2) \cdot (B_1, B_2, I, J) & = (q_1^{-1} B_1, q_2^{-1} B_2, I, q_{12}^{-1} J )
\, , &&
(q_1,q_2) \in \mathsf{T}_Q \subset \mathrm{GL}(Q)
\, .
\end{align}
\end{subequations}
Parametrizing these group elements as $g = \mathrm{e}^\phi$, $h = \mathrm{e}^a$, $q_i = \mathrm{e}^{\epsilon_i}$ ($i=1,2$), namely $\phi \in \operatorname{Lie} \mathrm{GL}(K)$, $h \in \operatorname{Lie} \mathrm{GL}(N)$, $(\epsilon_1,\epsilon_2) \in \operatorname{Lie} \mathsf{T}_Q$, the fixed point equations are given as follows,
\begin{subequations}\label{eq:fixed_pt_eq}
\begin{align}
[\phi, B_i] - \epsilon_i B_i & = 0
\, ,
\qquad (i = 1, 2)
\label{eq:fixed_point_B}
\\
\phi I - I a & = 0
\, , \\
- J \phi + (a - \epsilon_{12}) J & = 0
\, .
\end{align}
\end{subequations}
We apply the basis diagonalizing $a = \bigoplus_{\alpha \in [n]} a_\alpha$, which corresponds to $N = \bigoplus_{\alpha \in [n]} N_\alpha$, and $(a_\alpha)_{\alpha \in [n]} \in \operatorname{Lie} \mathsf{T}_N$.
We have the (left/right) eigenvalue equations, $\phi I_\alpha = a_\alpha I_\alpha$ and $J_\alpha \phi = J_\alpha (a_\alpha - \epsilon_{12})$.
Moreover, using~\eqref{eq:fixed_point_B}, we obtain
\begin{subequations}
\begin{align}
\phi B_1^{i-1} B_2^{j-1} I_\alpha
& = (a_\alpha + (i-1)\epsilon_1 + (j-1) \epsilon_2 ) B_1^{i-1} B_2^{j-1} I_\alpha
\, , \\
J_\alpha B_1^{i-1} B_2^{j-1} \phi & = J_\alpha B_1^{i-1} B_2^{j-1} (a_\alpha - i \epsilon_1 - j \epsilon_2 )
\, .
\end{align}
\end{subequations}
Applying the stability condition, $K = \mathbb{C}[B_1, B_2] I(N)$, and recalling $\dim K = k$, there exist only $k$ eigenvectors in the form of $B_1^{i-1} B_2^{j-1} I_\alpha$, which are parametrized by $n$-tuple partition, $\lambda = (\lambda_\alpha)_{\alpha \in [n]} \in \mathfrak{M}_{n,k}^\mathsf{T}$ with $|\lambda| = \sum_{\alpha \in [n]} |\lambda_\alpha| = k$, and hence $B_1^{i-1} B_2^{j-1} I_\alpha = 0$ for $(i, j) \not\in \lambda_\alpha$ and also $J_\alpha = 0$, $\alpha \in [n]$.
Namely, we have $K = \operatorname{Span}\{ B_1^{i-1} B_2^{j-1} I_\alpha \}_{\alpha \in [n], (i,j) \in \lambda_\alpha}$.
Starting with the co-stability condition, we have the left eigenvectors in the form of $J_\alpha B_1^{i-1} B_2^{j-1}$ with $I_\alpha = 0$.
We remark that together with the ADHM equation, $\mu_\mathbb{C} = [B_1, B_2] + IJ = 0$, we have $[B_1, B_2] = 0$ at the fixed point.
Therefore, we obtain the characters of the vector spaces $(N,K)$ as follows,
\begin{align}
\operatorname{ch} K\Big|_\lambda = \sum_{\alpha \in [n]} \operatorname{ch} K_\alpha\Big|_{\lambda_\alpha}
\, , \qquad
\operatorname{ch} N = \sum_{\alpha \in [n]} e^{a_\alpha}
\, ,
\end{align}
where we have
\begin{align}
\operatorname{ch} K_\alpha\Big|_{\lambda_\alpha} = \sum_{(i,j) \in \lambda_\alpha}
\begin{cases}
\mathrm{e}^{a_\alpha} q_1^{i-1} q_2^{j-1} & (\text{with stability}) \\
\mathrm{e}^{a_\alpha} q_1^{-i} q_2^{-j} & (\text{with co-stability}) \\
\end{cases}
\end{align}
This structure has a connection with the ideal $\mathsf{I}_\lambda$ generated by all monomials outside the partition, $\{ z_1^{i-1} z_2^{j-1} \}_{(i,j) \not\in \lambda}$ as follows,
\begin{align}
K_\alpha\Big|_{\lambda_\alpha} = N_\alpha \otimes
\begin{cases}
\mathsf{I}_\emptyset / \mathsf{I}_{\lambda_\alpha}
& (\text{stability}) \\
Q_{12}^\vee (\mathsf{I}_\emptyset / \mathsf{I}_{\lambda_\alpha})^\vee
& (\text{co-stability})
\end{cases}
\, .
\label{eq:K_sp_ideal_rep}
\end{align}
We remark that $\mathsf{I}_\emptyset = \mathbb{C}[z_1,z_2]$.
\subsubsection{Tangent bundle}\label{sec:tangent_bundle}
We next describe the tangent bundle to the moduli space.
From the ADHM construction, we have the following chain complex,
\begin{equation}
\begin{tikzcd}
& \operatorname{Hom}(K,K) \arrow[r,"d_1"] & \substack{\displaystyle \operatorname{Hom}(K,K) \otimes Q^\vee \\[.5em] \displaystyle \oplus \\[.5em] \displaystyle \operatorname{Hom}(N,K) \\[.5em] \displaystyle \oplus \\[.5em] \displaystyle \operatorname{Hom}(K,N) \otimes Q_{12}^\vee } \arrow[r,"d_2"] & \operatorname{Hom}(K,K) \otimes Q_{12}^\vee & \\
( &
C_0 \arrow[r] & C_1 \arrow[r]& C_2
& )
\end{tikzcd}
\end{equation}
where $C_{0,1,2}$ correspond to GL$(K)$ action, ADHM variables, and the moment map $\mu_\mathbb{C}$, and we define
\begin{align}
d_1(\xi) =
\begin{pmatrix}
[\xi, B_1] \\ [\xi, B_2] \\ \xi I \\ - J \xi
\end{pmatrix}
\, , \qquad
d_2
\begin{pmatrix}
b_1 \\ b_2 \\ i \\ j
\end{pmatrix}
= [B_1, b_2] + [b_1, B_2] + Ij + iJ
\, .
\end{align}
These maps are the infinitesimal GL$(K)$ action parametrized by $\xi \in \operatorname{Lie} \mathrm{GL}(K)$ and the differential of the moment map $\mu_\mathbb{C}$.
We can check that $d_2 \circ d_1(\xi) = [\xi, \mu_\mathbb{C}] = 0$ due to the ADHM equation, $\mu_\mathbb{C} = 0$.
Then, the tangent bundle is identified with the middle cohomology with respect to this sequence,
\begin{align}
T\mathfrak{M}_{n,k} = \operatorname{Ker} d_2 / \operatorname{Im} d_1 = C_1 - C_0 - C_2 = N^\vee K + Q_{12}^\vee K^\vee N - P_{12}^\vee K^\vee K
\, .
\end{align}
There exists another construction of the tangent bundle based on $\mathsf{Y}_o$ the observable sheaf on $\mathfrak{M}_{n,k} \times o$ obtained from $\mathsf{Y}_{\mathbb{C}^2}$ the universal sheaf on $\mathfrak{M}_{n,k} \times \mathbb{C}^2$ via the localization,
\begin{align}
\mathsf{Y} \equiv \mathsf{Y}_o = N - P_{12} K
\, , \qquad
\mathsf{Y}_{\mathbb{C}^2} = \frac{\mathsf{Y}}{P_{12}}
\, ,
\end{align}
where we use the notation that $(N,K)$ also means a bundle whose fiber is given by the vector space $N$ and $K$, respectively.
Several equivalent expressions are available for the character of the observable sheaf at the fixed point $\lambda \in \mathfrak{M}^\mathsf{T}_{n,k}$.
The first expression is given by
\begin{align}
\operatorname{ch} \mathsf{Y}\Big|_\lambda = \operatorname{ch} (P_1 \mathsf{X})
\end{align}
where we define the partial reduction of the universal sheaf $\mathsf{X} = \mathsf{Y}/P_1 = \mathsf{Y}_{\mathbb{C}^2} P_2$, and the character
\begin{align}
\operatorname{ch} \mathsf{X}
= \sum_{x \in \mathcal{X}_\lambda} x
\, , \qquad
\mathcal{X}_\lambda = \{ x_{\alpha,i} = \mathrm{e}^{a_\alpha} q_1^{i-1} q_2^{\lambda_{\alpha,i}} \}_{\alpha \in [n], i \in \mathbb{N}}
\, .
\label{eq:obs_sheaf_def}
\end{align}
In order to obtain this expression, we need the condition $|q_1| < 1$ to have $\sum_{i = 1}^\infty q_1^{i-1} = 1/(1 - q_1)$.
If $|q_2| < 1$, we may apply another expression based on the transposed partition of $\lambda$ denoted by $\check{\lambda}$,
\begin{align}
\operatorname{ch} \mathsf{Y}\Big|_\lambda = \operatorname{ch} (P_2 \check{\mathsf{X}})
\, , \qquad
\operatorname{ch} \check{\mathsf{X}}
= \sum_{x \in \check{\mathcal{X}}_\lambda} x
\, , \qquad
\check{\mathcal{X}}_\lambda = \{ x_{\alpha,i} = \mathrm{e}^{a_\alpha} q_1^{\check{\lambda}_{\alpha,j}} q_2^{j-1} \}_{\alpha \in [n], j \in \mathbb{N}}
\, .
\end{align}
Applying the co-stability condition, we instead obtain
\begin{align}
\operatorname{ch} \mathsf{Y}\Big|_\lambda =
\begin{cases}
\displaystyle
\operatorname{ch} (P_1^\vee \mathsf{X})
\, , \quad
\operatorname{ch} \mathsf{X}
= \sum_{x \in \mathcal{X}_\lambda} x
\, , \quad
\mathcal{X}_\lambda = \{ x_{\alpha,i} = \mathrm{e}^{a_\alpha} q_1^{-i+1} q_2^{-\lambda_{\alpha,i}} \}_{\alpha \in [n], i \in \mathbb{N}}
& (|q_1|>1) \\
\displaystyle
\operatorname{ch} (P_2^\vee \check{\mathsf{X}})
\, , \quad
\operatorname{ch} \check{\mathsf{X}}
= \sum_{x \in \check{\mathcal{X}}_\lambda} x
\, , \quad
\check{\mathcal{X}}_\lambda = \{ x_{\alpha,i} = \mathrm{e}^{a_\alpha} q_1^{-\check{\lambda}_{\alpha,j}} q_2^{-j+1} \}_{\alpha \in [n], j \in \mathbb{N}}
& (|q_2|>1)
\end{cases}
\, .
\end{align}
Another expression is given as a finite sum,
\begin{align}
\operatorname{ch} \mathsf{Y}\Big|_\lambda =
\begin{cases}
\displaystyle
\sum_{x \in \mathcal{X}_{\partial_+ \lambda}} x - \sum_{x \in \mathcal{X}_{\partial_- \lambda}} x q_{12}
& (\text{stability})
\\ \displaystyle
\sum_{x \in \mathcal{X}_{\partial_+ \lambda}} x q_{12} - \sum_{x \in \mathcal{X}_{\partial_- \lambda}} x
& (\text{co-stability})
\end{cases}
\label{eq:obs_sheaf_comb}
\end{align}
where we define the addable and removable boundary of the partition,
\begin{align}
\partial_+ \lambda = \{ (i,\lambda_{\alpha,i}+1) \mid \lambda_{\alpha,i-1} > \lambda_{\alpha,i} \}_{\alpha \in [n], i \in \mathbb{N}}
\, , \quad
\partial_- \lambda = \{ (i,\lambda_{\alpha,i}) \mid \lambda_{\alpha,i+1} < \lambda_{\alpha,i} \}_{\alpha \in [n], i \in \mathbb{N}}
\end{align}
where we put $\lambda_{\alpha,0} = \infty$, and we define
\begin{subequations}
\begin{align}
\mathcal{X}_{\partial_+ \lambda} & =
\begin{cases}
\displaystyle
\{ \mathrm{e}^{a_\alpha} q_1^{i-1} q_2^{j-1} \mid (i,j) \in \partial_+ \lambda \}
& (\text{stability})
\\ \displaystyle
\{ \mathrm{e}^{a_\alpha} q_1^{-i} q_2^{-j} \mid (i,j) \in \partial_+ \lambda \}
& (\text{co-stability})
\end{cases}
\, , \\
\mathcal{X}_{\partial_- \lambda} & =
\begin{cases}
\displaystyle
\{ \mathrm{e}^{a_\alpha} q_1^{i-1} q_2^{j-1} \mid (i,j) \in \partial_- \lambda \}
& (\text{stability})
\\ \displaystyle
\{ \mathrm{e}^{a_\alpha} q_1^{-i} q_2^{-j} \mid (i,j) \in \partial_- \lambda \}
& (\text{co-stability})
\end{cases}
\, .
\end{align}
\end{subequations}
The tangent bundle is then obtained from the observable sheaf as follows,
\begin{align}
\mathsf{V} := \frac{\mathsf{Y}^\vee \mathsf{Y}}{P_{12}} = - T\mathfrak{M}_{n,k} + \frac{N^\vee N}{P_{12}}
\, .
\end{align}
The last term $N^\vee N/P_{12}$ is understood as the perturbative contribution.
The combination of $\mathsf{Y}^\vee \mathsf{Y}$ corresponds to the adjoint representation of GL group.
In order to construct the SO/Sp gauge theory, we consider the anti-symmetric and symmetric tensor product, which yield the adjoint representation of SO/Sp group~\cite{Marino:2004cn,Nekrasov:2004vw,Shadchin:2005mx},
\begin{align}
\mathsf{V}_\text{SO} = \frac{1}{2} \frac{\mathsf{Y}^2 - \mathsf{Y}^{[2]}}{P_{12}}
\, , \qquad
\mathsf{V}_\text{Sp} = \frac{1}{2} \frac{\mathsf{Y}^2 + \mathsf{Y}^{[2]}}{P_{12}}
\, ,
\label{eq:O_SP_adj_rep}
\end{align}
where we denote the degree-$p$ Adams operation to $X$ by $X^{[p]}$.
Hence, the tangent bundle at the fixed point is given by
\begin{align}
\mathsf{V}_\lambda = - T_\lambda\mathfrak{M}_{n,k} + \frac{N^\vee N}{P_{12}} = \frac{\mathsf{Y}^\vee \mathsf{Y}}{P_{12}}\Bigg|_{\lambda} = \frac{P_1^\vee}{P_2} \mathsf{X}^\vee \mathsf{X}
\, .
\end{align}
Denoting the total fixed point configuration space by $\mathfrak{M}^\mathsf{T} = \bigsqcup_{k=0}^\infty \mathfrak{M}_{n,k}^\mathsf{T}$, the full partition function is given as follows.
\begin{itembox}{Equivariant index formula}
The full partition function of pure SYM gauge theory is given by summation over the fixed point contributions,
\begin{align}
Z = \sum_{\lambda \in \mathfrak{M}^\mathsf{T}} \mathfrak{q}^{|\lambda|} \, \mathbb{I}[ \mathsf{V}_\lambda ]
\ \stackrel{(5d)}{=} \
\sum_{\lambda \in \mathfrak{M}^\mathsf{T}} \mathfrak{q}^{|\lambda|} \prod_{(x,x') \in \mathcal{X}_\lambda \times \mathcal{X}_\lambda}
\frac{(x'/x;q_2^{-1})_\infty}{(q_{1} x'/x;q_2^{-1})_\infty}
\, , \qquad
|q_2| > 1
\, ,
\label{eq:full_part_fn_pureSYM_5d}
\end{align}
where $(z;q)_\infty$ is the $q$-factorial~\eqref{eq:q-factorial} interpreted as the $q$-deformation of the gamma function.
\end{itembox}
For four and six-dimensional cases, we obtain a similar infinite product formula based on the gamma and elliptic gamma functions, respectively.
See, e.g.,~\cite{Kimura:2020jxl}.
\subsubsection{Combinatorial formula}\label{sec:combinatorial_formula}
We have seen that the partition function is obtained by applying the index functor to the tangent bundle.
The full partition function is written in the closed form, whereas it involves infinite product as in~\eqref{eq:full_part_fn_pureSYM_5d}.
In fact, the infinite product contribution originates from the perturbative part, and one can subtract the finite rational contribution to be identified with the instanton partition function.
For this purpose, we evaluate the character of the tangent bundle with the fix point configuration (see, e.g.~\cite{Nakajima:1999}),
\begin{align}
\operatorname{ch} T_\lambda \mathfrak{M}_{n,k} =
\sum_{\alpha, \beta \in [n]} \mathrm{e}^{a_{\alpha \beta}} \Xi(\lambda_\alpha, \lambda_\beta; q_1, q_2)
\, ,
\end{align}
where we denote $a_{\alpha \beta} = a_\alpha - a_\beta$.
The combinatorial factor is given by
\begin{align}
\Xi(\lambda_\alpha, \lambda_\beta; q_{1,2}) & = \operatorname{ch} \left( \mathsf{I}_\emptyset/\mathsf{I}_{\lambda_\beta} + Q_{12}^\vee (\mathsf{I}_\emptyset/\mathsf{I}_{\lambda_\alpha})^\vee - P_{12}^\vee (\mathsf{I}_\emptyset/\mathsf{I}_{\lambda_\alpha})^\vee (\mathsf{I}_\emptyset/\mathsf{I}_{\lambda_\beta}) \right)
\nonumber \\
& = \sum_{s \in \lambda_\alpha} q_1^{\mathbf{l}_\beta(s)} q_2^{- \mathbf{a}_\alpha(s) - 1} + \sum_{s \in \lambda_\beta} q_1^{-\mathbf{l}_\alpha(s) - 1} q_2^{\mathbf{a}_\beta(s)}
\, ,
\label{eq:combinatorial_Xi}
\end{align}
where we define the arm and leg lengths,
\begin{align}
\mathbf{a}_\alpha(i,j) = \lambda_{\alpha,i} - j
\, , \qquad
\mathbf{l}_\alpha(i,j) = \check{\lambda}_{\alpha,j} - i
\, .
\end{align}
We remark that the same combinatorial factor is obtained starting from the co-stability condition.
We observe the following relation,
\begin{align}
\Xi(\lambda_\alpha, \lambda_\beta; q_{1,2})
= \Xi(\check{\lambda}_\alpha, \check{\lambda}_\beta; q_{2,1})
= q_{12}^{-1} \Xi(\lambda_\beta, \lambda_\alpha; q_{1,2}^{-1})
\, .
\end{align}
Hence, applying the index functor, we obtain the $k$-instanton partition function, a.k.a., \emph{Nekrasov partition function}.
\begin{itembox}{Nekrasov partition function (combinatorial formula)~\cite{Nekrasov:2002qd,Nekrasov:2003rj}}
The instanton partition function has the following combinatorial formula (Nekrasov partition function),
\begin{align}
Z_k = \sum_{\lambda \in \mathfrak{M}_{n,k}^\mathsf{T}} \mathbb{I}[-T_\lambda\mathfrak{M}_{n,k}] = \sum_{\lambda \in \mathfrak{M}_{n,k}^\mathsf{T}} \prod_{\alpha,\beta \in [n]} \frac{1}{\mathsf{N}_{\lambda_\alpha\lambda_\beta}(a_{\alpha\beta};\epsilon_{1,2})}
\, ,
\end{align}
where we define the combinatorial factor,
\begin{align}
\mathsf{N}_{\lambda_\alpha\lambda_\beta}(z;\epsilon_{1,2}) = \prod_{s \in \lambda_\alpha }[z + \epsilon_1 \mathbf{l}_\beta(s) - \epsilon_2 (\mathbf{a}_\alpha(s)+1)] \prod_{s \in \lambda_\beta }[z - \epsilon_1 (\mathbf{l}_\beta(s)+1) + \epsilon_2 \mathbf{a}_\alpha(s)]
\, .
\label{eq:Nekrasov_factor}
\end{align}
\end{itembox}
\subsubsection{Contour integral formula}
It has been known that the instanton partition function can be described in the form of contour integral, a.k.a., LMNS (Losev--Moore--Nekrasov--Shatashvili) formula.
For this purpose, we use the vector bundles on the moduli space without substituting the fixed point value, i.e., $\operatorname{ch} K = \sum_{I \in [k]} \mathrm{e}^{\phi_I}$.
Then, we compute the index functor of the tangent bundle to the $k$-instanton moduli space, which gives rise to the contour integral over the maximal Cartan torus $\mathsf{T}_K \subset \mathrm{GL}(K)$.
\begin{itembox}{LMNS formula (contour integral formula)~\cite{Losev:1997tp,Moore:1997dj,Lossev:1997bz}}
The instanton partition function has the following contour integral formula (LMNS formula),
\begin{align}
Z_k = \mathbb{I}[-T\mathfrak{M}_{n,k}] = \frac{1}{k!} \frac{[-\epsilon_{12}]^k}{[-\epsilon_{1,2}]^k} \oint_{\mathsf{T}_K} \mathrm{d}{\underline{\phi}} \frac{1}{P(\underline{\phi}) \widetilde{P}(\underline{\phi}+\epsilon_{12})} \prod_{I \neq J}^k \mathscr{S}(\phi_{IJ})^{-1}
\, ,
\end{align}
where we define the gauge polynomials
\begin{align}
P(\underline{\phi}) = \mathbb{I}[N^\vee K] = \prod_{I \in [k], \alpha \in [n]} [\phi_I - a_\alpha]
\, , \qquad
\widetilde{P}(\underline{\phi}) = \mathbb{I}[K^\vee N] = \prod_{I \in [k], \alpha \in [n]} [- \phi_I + a_\alpha]
\end{align}
and the $\mathscr{S}$-function,
\begin{align}
\mathscr{S}(z) = \frac{[z - \epsilon_{1,2}]}{[z][z - \epsilon_{12}]}
\, .
\end{align}
\end{itembox}
We use the notation, $[z - \epsilon_{1,2}] = [z - \epsilon_{1}][z - \epsilon_{2}]$.
The integral measure is defined as $\mathrm{d}{\underline{\phi}} = \prod_{I \in [k]} \mathrm{d}{\phi}_I / 2 \pi \mathrm{i}$, which is understood as the index of the zero mode appearing from $K^\vee K$.
We have the Weyl group volume, $k! = |\mathfrak{S}_k|$.
In fact, the poles of the integrand are consistent with the fixed points, such that we take the residues at $\phi_{I} = \phi_{I-1} + \epsilon_{1,2}$ or $\phi_I = a_\alpha$, which yields the map, $\{\phi_I\}_{I \in [k]} \to \{a_\alpha + (i - 1) \epsilon_1 + (j-1) \epsilon_2\}_{\alpha \in [n], (i,j) \in \lambda_{alpha}}$.
This multi-variable contour integral is also understood as the Jeffrey--Kirwan residue prescription~\cite{Jeffrey:1995}.
See, e.g.,~\cite{Benini:2013xpa,Hori:2014tda,Hwang:2014uwa,Nakamura:2015zsa} for details.
\subsubsection{Quiver gauge theory}\label{sec:quiver_gauge_theory}
Using the observable sheaf, we can similarly obtain the hypermultiplet contribution to the partition function.
For the matter bundle $\mathsf{M}$ and $\widetilde{\mathsf{M}}$, we have the (anti)fundamental hypermultiplet contribution as follows,
\begin{align}
\mathsf{H} = - \frac{\mathsf{M}^\vee \mathsf{Y}}{P_{12}}
\, , \qquad
\widetilde{\mathsf{H}} = - \frac{\mathsf{Y}^\vee \widetilde{\mathsf{M}}}{P_{12}}
\, ,
\label{eq:fund_hyp_contribution}
\end{align}
where the corresponding Chern roots are identified with the fundamental mass parameters,
\begin{align}
\operatorname{ch} \mathsf{M} = \sum_{f \in [n^F]} \mathrm{e}^{m_f}
\, , \qquad
\operatorname{ch} \widetilde{\mathsf{M}} = \sum_{f \in [\tilde{n}^F]} \mathrm{e}^{\tilde{m}_f}
\, .
\end{align}
For quiver gauge theory of type $\Gamma = (\Gamma_0,\Gamma_1)$ with $\Gamma_0 = \{ \text{nodes} \}$ and $\Gamma_1 = \{ \text{edges} \}$,%
\footnote{%
Since we are considering eight supercharge gauge theories, we also have the dual edge for each edge $e : i \to j$ given by $e^\vee : j \to i$, and the set of the dual edges denoted by $\Gamma_1^\vee = \{ e^\vee \}$.
In order to realize the oriented multiple arrow to construct non-simply-laced quiver, we need to properly assign the $\Omega$-background parameters~\cite{Kimura:2017hez}.
}
we have the instanton moduli space for each node, and the total moduli space is given by $\mathfrak{M}_{\underline{G},\underline{k}} = \bigsqcup_{i \in \Gamma_0} \mathfrak{M}_{G_i,k_i}$~\cite{Nekrasov:2012xe,Nekrasov:2013xda}.
In this case, we have the observable sheaf for each node, $(\mathsf{Y}_i)_{i \in \Gamma_0}$, and the vector multiplet and the bifundamental hypermultiplet contributions are given by
\begin{align}
\mathsf{V}_i = \frac{\mathsf{Y}_i^\vee \mathsf{Y}_i}{P_{12}}
\, , \qquad
\mathsf{H}_{e:i\to j} = - \mathsf{M}_e \frac{\mathsf{Y}_i^\vee \mathsf{Y}_j}{P_{12}}
\, ,
\end{align}
where $\mathsf{M}_e$ is a line bundle assigned to the edge $e \in \Gamma_1$ with the Chern root identified with the bifundamental mass parameter $m_e$.
For a generic quiver (except for the loop/cyclic cases), we can put $m_e$ to be zero by shift of Coulomb moduli parameters.
The total tangent bundle is thus given by~\cite{Kimura:2015rgi}
\begin{align}
- T\mathfrak{M}_{\underline{G},\underline{k}} = \sum_{i \in \Gamma_0} \mathsf{V}_i + \sum_{e: i \to j} \mathsf{H}_{e:i\to j}
= \sum_{i,j \in \Gamma_0} \frac{\mathsf{Y}_i^\vee c_{ij}^+ \mathsf{Y}_j}{P_{12}}
\, ,
\label{eq:total_tangent_bundle}
\end{align}
where we define the half $q$-Cartan matrix,
\begin{align}
c_{ij}^+ = \delta_{ij} - \sum_{e:i \to j} \mathsf{M}_e
\, .
\label{eq:half-q-Cartan}
\end{align}
We remark that this total tangent bundle involves also the perturbative contributions in addition to the instanton part.
\subsubsection*{Contour integral}
Given the tangent bundle, we obtain the contour integral form of the instanton partition function.
Let us discuss the examples.
\paragraph{$A_p$ quiver theory}
We consider the linear quiver that consists of $p$ gauge nodes:
\dynkin[label,labels={1,2,p-1,p}]{A}{}\\
In this case, we can put all the bifundamental mass parameters to be zero without loss of generality.
The bifundamental hypermultiplet contribution is given by
\begin{align}
\mathsf{H}_{i \to i+1} = - \frac{\mathsf{Y}_i^\vee \mathsf{Y}_{i+1}}{P_{12}} = - \frac{\mathsf{N}_i^\vee \mathsf{N}_{i+1}}{P_{12}} + \mathsf{N}_i^\vee \mathsf{K}_{i+1} + Q_{12}^\vee \mathsf{K}_i^\vee \mathsf{N}_{i+1} - P_{12}^\vee \mathsf{K}_i^\vee \mathsf{K}_{i+1}
\, ,
\end{align}
where the first term is interpreted as the perturbative contribution.
Hence, applying the index functor to the instanton part, we obtain the contour integral formula for the instanton partition function,
\begin{align}
Z_{\underline{k}} = \frac{1}{\underline{k}!} \frac{[-\epsilon_{12}]^{k}}{[-\epsilon_{1,2}]^{k}} \oint_{\mathsf{T}_K} \mathrm{d}{\underline{\phi}} \frac{\prod_{i \in [p-1]} P_i(\underline{\phi}_{i+1})\widetilde{P}_{i+1}(\underline{\phi}_i + \epsilon_{12})}{\prod_{i \in [p]} P_i(\underline{\phi}_i) \widetilde{P}_i(\underline{\phi}_i + \epsilon_{12})}
\frac{\prod_{i \in [p-1]} \prod^{I \in [k_{i+1}]}_{J \in [k_{i}]} \mathscr{S}(\phi_{i+1,I} - \phi_{i,J})}{\prod_{i \in [p]} \prod_{I \neq J}^{k_i} \mathscr{S}(\phi_{i,I} - \phi_{i,J})}
\, ,
\end{align}
where we denote $\underline{k}! = \prod_{i \in \Gamma_0} k_i!$ and $k = \sum_{i \in \Gamma_0} k_i$.
The gauge polynomials are given by $P_i(z) = \prod_{\alpha \in [n_i]} [z - a_{i,\alpha}]$ and $\widetilde{P}_i(z) = \prod_{\alpha \in [n_i]} [ a_{i,\alpha} - z]$.
\paragraph{$\widehat{A}_{p}$ quiver theory}
The next example is a cyclic quiver $\widehat{A}_{p}$ with $p+1$ nodes:
\dynkin[label,labels={0,1,2,p-1,p},extended]{A}{}\\
In this case, we can put all the bifundamental mass parameters to be the same, $m_{i \to i+1} = m$.
Then, the contour integral form of the instanton partition function is given by
\begin{align}
Z_{\underline{k}} = \frac{1}{\underline{k}!} \frac{[-\epsilon_{12}]^{k}}{[-\epsilon_{1,2}]^{k}} \oint_{\mathsf{T}_K} \mathrm{d}{\underline{\phi}} \prod_{i \in \mathbb{Z}_{p+1}} \left[ \frac{P_i(\underline{\phi}_{i+1} + m)\widetilde{P}_{i+1}(\underline{\phi}_i - m + \epsilon_{12})}{P_i(\underline{\phi}_i) \widetilde{P}_i(\underline{\phi}_i + \epsilon_{12})}
\frac{\prod^{I \in [k_{i+1}]}_{J \in [k_{i}]} \mathscr{S}(\phi_{i+1,I} - \phi_{i,J} + m)}{\prod_{I \neq J}^{k_i} \mathscr{S}(\phi_{i,I} - \phi_{i,J})} \right]
\, .
\end{align}
The case $p = 0$ involves a single node with a loop edge describing the hypermultiplet in the adjoint representation, which corresponds to four-dimensional $\mathcal{N} = 2^*$ theory.
The instanton partition function is given by
\begin{align}
Z_k = \frac{1}{k!} \left( \frac{[-\epsilon_{12}]}{[-\epsilon_{1,2}]} \mathscr{S}(m) \right)^k \oint_{\mathsf{T}_K} \mathrm{d}{\underline{\phi}} \frac{P(\underline{\phi} + m) \widetilde{P}(\underline{\phi} - m + \epsilon_{12})}{P(\underline{\phi}) \widetilde{P}(\underline{\phi} + \epsilon_{12})}
\prod_{I \neq J}^k \frac{\mathscr{S}(\phi_{IJ}+m)}{\mathscr{S}(\phi_{IJ})}
\, ,
\end{align}
where $m$ is the adjoint mass parameter.
\subsubsection*{Fixed points}
The fixed points are similarly described for quiver gauge theory.
For $\underline{G} = \prod_{i \in \Gamma_0} \mathrm{U}(n_i)$, we have a set of partitions characterizing the fixed points,
\begin{align}
\lambda = (\lambda_{i,\alpha})_{i \in \Gamma_0, \alpha \in [n_i]} \in \mathfrak{M}^\mathsf{T}
\, ,
\end{align}
and we define
\begin{align}
\mathcal{X}_\lambda = \bigsqcup_{i \in \Gamma_0} \mathcal{X}_{\lambda_i}
\, , \qquad
\mathcal{X}_{\lambda_i} = \{ x_{i,\alpha,k} = \mathrm{e}^{a_{i,\alpha}} q_1^{k-1} q_2^{\lambda_{i,\alpha,k}} \}_{\alpha \in [n], k \in \mathbb{N}}
\, ,
\end{align}
under the stability condition and $|q_1| < 1$.
Then, the total tangent space at the fixed point $\lambda \in \mathfrak{M}^\mathsf{T}$ is given by
\begin{align}
- T_\lambda \mathfrak{M}_{\underline{G},\underline{k}} = \sum_{i,j \in \Gamma_0} \frac{P_1^\vee}{P_2} \mathsf{X}_i^\vee c_{ij}^+ \mathsf{X}_j
\, , \qquad
\operatorname{ch} \mathsf{X}_i = \sum_{x \in \mathcal{X}_{\lambda_i}} x
\, ,
\end{align}
and thus the full partition function is given as follows,
\begin{align}
Z = \sum_{\lambda \in \mathfrak{M}^\mathsf{T}} \underline{\mathfrak{q}}^{|\lambda|} \, \mathbb{I}[- T_\lambda \mathfrak{M}_{\underline{G},\underline{k}}]
\end{align}
where we denote the instanton counting parameter,
\begin{align}
\underline{\mathfrak{q}}^{|\lambda|} = \prod_{i \in \Gamma_0} \mathfrak{q}_i^{|\lambda_i|}
\, .
\end{align}
\subsubsection{Supergroup gauge theory}
We apply the formalism discussed above to supergroup gauge theory.
The discussion of this part is mainly based on~\cite{Kimura:2019msw}.
First of all, we consider the supercharacter of the supervector spaces instead of the ordinary characters,
\begin{subequations}
\begin{align}
\operatorname{sch} K & = \operatorname{ch} K_0 - \operatorname{ch} K_1 = \sum_{I \in [k_0]} \mathrm{e}^{\phi_I^0} - \sum_{I \in [k_1]} \mathrm{e}^{\phi_I^1}
\, , \\
\operatorname{sch} N & = \operatorname{ch} N_0 - \operatorname{ch} N_1 = \sum_{\alpha \in [n_0]} \mathrm{e}^{a_\alpha^0} - \sum_{\alpha \in [n_1]} \mathrm{e}^{a_\alpha^1}
\, .
\end{align}
\end{subequations}
In order to apply the localization formula, we then analyze the fixed points in the moduli space $\mathfrak{M}_{n,k}$ under the equivariant actions.
In the case of supergroup gauge theory, we still have the same form of the fixed point equations~\eqref{eq:fixed_pt_eq}.
Recalling that we should assign the both stability and co-stability conditions,
\begin{align}
K_0 = \mathbb{C}[B_1, B_2] I(N_0)
\, , \qquad
K_1^\vee = \mathbb{C}[B_1, B_2] J(N_1)
\, ,
\end{align}
the fixed points are parametrized by $\lambda = (\lambda^\sigma_\alpha)_{\sigma = 0, 1, \alpha \in [n_\sigma]}$ with $|\lambda^\sigma| = \sum_{\alpha \in [n_\sigma]} |\lambda_\alpha^\sigma| = k_\sigma$.
Hence, each subspace is given by $K_0 = \operatorname{Span}\{ B_1^{i-1} B_2^{j-1} I_{\alpha^0} \}_{\alpha \in [n_0],(i,j) \in \lambda^0_\alpha}$ and $K_1 = \operatorname{Span}\{ J_{\alpha^1} B_1^{i-1} B_2^{j-1} \}_{\alpha \in [n_1],(i,j) \in \lambda^1_\alpha}$ at the fixed point $\lambda \in \mathfrak{M}_{n,k}^\mathsf{T}$.
Therefore, the supercharacter is given by
\begin{align}
\operatorname{sch} K\Big|_{\lambda} = \sum_{\alpha \in [n_0]} \sum_{(i,j) \in \lambda_\alpha^0} \mathrm{e}^{a_\alpha^0} q_1^{i-1} q_2^{j-1} - \sum_{\alpha \in [n_1]} \sum_{(i,j) \in \lambda_\alpha^1} \mathrm{e}^{a_\alpha^1} q_1^{-i} q_2^{-j}
\, .
\end{align}
Namely, we assign the GL($Q$)-equivariant parameters $(\epsilon_{1,2})$ for the positive node and $(-\epsilon_{1,2})$ for the negative node.
This assignment is similarly understood as the analytic continuation to obtain supergroup theory as discussed in Sec.~\ref{sec:quiver_realization}.
\subsubsection*{Contour integral formula}
We construct the observable sheaf as before, $\mathsf{Y} = N - P_{12} K$.
Obtaining the vector multiplet contribution, associated with the tangent bundle,
\begin{align}
\mathsf{V} & = \frac{\mathsf{Y}^\vee \mathsf{Y}}{P_{12}} = - T\mathfrak{M}_{n,k} + \frac{\mathsf{N}^\vee \mathsf{N}}{P_{12}}
\nonumber \\
& \xrightarrow{\text{sch}}
\sum_{\sigma,\sigma' = 0, 1} (-1)^{\sigma + \sigma'}
\left[ \frac{N_\sigma^\vee N_{\sigma'}}{P_{12}} - N_\sigma^\vee K_{\sigma'} - Q_{12}^\vee K_{\sigma}^\vee N_{\sigma'} + P_{12} K_{\sigma}^\vee K_{\sigma'} \right]
\, ,
\end{align}
we have the contour integral form of the instanton partition function.
\begin{itembox}{LMNS formula for supergroup gauge theory}
The instanton partition function of $\mathrm{U}(n_0|n_1)$ gauge theory is given by the following contour integral over the Cartan torus,
\begin{align}
Z_k & = \mathbb{I}[-T\mathfrak{M}_{n,k}]
\nonumber \\
& = \frac{1}{k_{0,1}!} \frac{[-\epsilon_{12}]^{k_{01}}}{[-\epsilon_{1,2}]^{k_{01}}} \oint_{\mathsf{T}_K} \mathrm{d}{\underline{\phi}} \, \frac{P_{0,1}(\underline{\phi}^{1,0}) \widetilde{P}_{0,1}(\underline{\phi}^{1,0}+\epsilon_{12})}{P_{0,1}(\underline{\phi}^{0,1}) \widetilde{P}_{0,1}(\underline{\phi}^{0,1}+\epsilon_{12})}
\frac{\prod^{I \in [k_0]}_{J \in [k_1]} \mathscr{S}(\phi^0_I - \phi^1_J) \mathscr{S}(\phi^1_J - \phi^0_I) }{\prod_{I \neq J}^{k_{0}} \mathscr{S}(\phi^0_{IJ}) \prod_{I \neq J}^{k_{1}} \mathscr{S}(\phi^1_{IJ})}
\, ,
\label{eq:LMNS_formula_supergroup}
\end{align}
where we define $k_{01} = k_0 + k_1$.
\end{itembox}
This formula agrees with the contour integral formula of $\widehat{A}_1$ quiver gauge theory obtained in Sec.~\ref{sec:quiver_gauge_theory}, which is consistent with the quiver gauge theory realization of supergroup gauge theory presented in Sec.~\ref{sec:quiver_realization}.
In this case, we should take the residues of the following poles, $\phi_{I}^0 = \phi_{I-1}^0 + \epsilon_{1,2}$ or $\phi_I^0 = a_\alpha^0$ and $\phi_{I}^1 = \phi_{I-1}^1 - \epsilon_{1,2}$ or $\phi_I^1 = a_\alpha^1 - \epsilon_{12}$, to be compatible with the stability and the co-stability conditions.
The formula~\eqref{eq:LMNS_formula_supergroup} is originally given as an integral over $\mathrm{GL}(K)$, which is evaluated by the maximal torus $\mathsf{T}_K$.
Hence, it should be understood as a normalized integral by the supergroup volume as discussed for the supermatrix models in Secs.~\ref{sec:Hermitian_supermatrix_model} and \ref{sec:real-quaternion_supermatrix_model}.
\subsubsection*{Equivariant index formula}
Recalling that we assign the stability and the co-stability conditions for the even and the odd nodes, we obtain for $|q_1| < 1$ and $|q_2| > 1$ as
\begin{align}
\mathsf{Y}\Big|_\lambda & = (P_1 X_0)_0 \oplus (P_2^\vee \check{X}_1)_1
\xrightarrow{\text{sch}} (1 - q_1) \sum_{x \in \mathcal{X}_{\lambda}^0} x - (1 - q_2^{-1}) \sum_{x \in \check{\mathcal{X}}_{\lambda}^1} x
\end{align}
where we define the sets,
\begin{align}
\mathcal{X}_\lambda^0 = \{ \mathrm{e}^{a^0_\alpha} q_1^{i-1} q_2^{\lambda_{\alpha,i}^0} \}_{\alpha \in [n_0], i \in \mathbb{N}}
\, , \qquad
\check{\mathcal{X}}_\lambda^1 = \{ \mathrm{e}^{a^1_\alpha} q_1^{- \check{\lambda}_{\alpha,j}^1} q_2^{-j+1} \}_{\alpha \in [n_0], j \in \mathbb{N}}
\, .
\end{align}
Imposing the fixed point configuration, the vector multiplet contribution is given as follows,
\begin{align}
\mathsf{V}_{\lambda} =
\left( \frac{P_1^\vee}{P_2} X_0^\vee X_0 + \frac{P_2^\vee}{P_1} \check{X}_1^\vee \check{X}_1 \right)_0 \oplus \left( Q_{12}^\vee X_0^\vee \check{X}_1 + \check{X}_1^\vee X_0 \right)_1
\, .
\end{align}
Applying the index functor with taking care of the conditions $|q_1| < 1$ and $|q_2| > 1$, we obtain the partition function.
\begin{itembox}{Equivariant index formula for supergroup partition function}
The partition function of $\mathrm{U}(n_0|n_1)$ gauge theory is given by summation over the fixed point contributions,
\begin{align}
Z = \sum_{\lambda \in \mathfrak{M}^\mathsf{T}} \mathfrak{q}^{|\lambda^0| - |\lambda^1|} \, Z_\lambda
\, ,
\end{align}
where the contribution at each fixed point $\lambda \in \mathfrak{M}^\mathsf{T}$ is given in the five-dimensional convention for $|q_1|<1$ and $|q_2|>1$ by
\begin{align}
Z_\lambda & =
\prod_{(x,x') \in \mathcal{X}^0_\lambda \times \mathcal{X}^0_\lambda} \frac{(x'/x;q_2^{-1})_\infty}{(q_1 x'/x;q_2^{-1})_\infty}
\prod_{(x,x') \in \check{\mathcal{X}}^1_\lambda \times \check{\mathcal{X}}^1_\lambda} \frac{(q_{12} x'/x;q_1)_\infty}{(q_1 x'/x;q_1)_\infty}
\nonumber \\
& \qquad \times
\prod_{(x,x') \in {\mathcal{X}}^0_\lambda \times \check{\mathcal{X}}_\lambda^1}
\left( 1 - \frac{x'}{x} \right)^{-1} \left( 1 - q_{12} \frac{x}{x'} \right)^{-1}
\, .
\end{align}
\end{itembox}
We remark a similarity of this partition function to the supermatrix model discussed in Sec.~\ref{sec:super_matrix}.
We will see further similarities from the underlying algebraic point of view.
Sec.~\ref{sec:quiver_W-algebra}.
\subsubsection*{Combinatorial formula}
We consider the combinatorial formula, which gives rise to the finite rational expression,
\begin{align}
\operatorname{sch} T_\lambda\mathfrak{M}_{n,k} = \sum_{\sigma,\sigma'=0,1} (-1)^{\sigma + \sigma'} \sum_{\alpha \in [n_\sigma], \beta \in [n_{\sigma'}]} \mathrm{e}^{a^\sigma_\alpha - a^{\sigma'}_\beta} \Xi_{\sigma\sigma'}(\lambda_\alpha,\lambda_\beta;q_{1,2})
\, ,
\end{align}
where the diagonal combinatorial factors are given by
\begin{align}
\Xi_{00}(\lambda_\alpha,\lambda_\beta;q_{1,2}) = \Xi(\lambda_\alpha,\lambda_\beta;q_{1,2})
\, , \qquad
\Xi_{11}(\lambda_\alpha,\lambda_\beta;q_{1,2}) = \Xi(\lambda_\beta,\lambda_\alpha;q_{1,2})
\, .
\end{align}
This is clear from the original expression of $\Xi$ in terms of the polynomial ideals~\eqref{eq:K_sp_ideal_rep} and \eqref{eq:combinatorial_Xi}.
See Sec.~\ref{sec:combinatorial_formula}.
The off-diagonal factors are given as follows (see~\cite{Kimura:2019msw,Kimura:2020jxl}),
\begin{subequations}
\begin{align}
\Xi_{01}(\lambda_\alpha,\lambda_\beta;q_{1,2}) & = \operatorname{ch} \left( Q_{12}^\vee (\mathsf{I}_\emptyset/\mathsf{I}_{\lambda_\beta})^\vee + Q_{12}^\vee (\mathsf{I}_\emptyset/\mathsf{I}_{\lambda_\alpha})^\vee - P_{12}^\vee Q_{12}^\vee (\mathsf{I}_\emptyset/\mathsf{I}_{\lambda_\alpha})^\vee (\mathsf{I}_\emptyset/\mathsf{I}_{\lambda_\beta})^\vee \right)
\nonumber \\
& =
- \sum_{i=1}^{\check\lambda_{\alpha,1}} \sum_{j'=1}^{\lambda_{\beta,1}}
\left[
q_1^{-\check\lambda_{\beta,j'}-i} q_2^{-\lambda_{\alpha,i}-j'}
- q_1^{-i} q_2^{-j'}
\right]
\nonumber \\
& \qquad
+ \sum_{(i,j) \in \lambda_\alpha} q_1^{-i} q_2^{-\lambda_{\beta,1}-j}
+ \sum_{(i',j') \in \lambda_\beta} q_1^{-\check\lambda_{\alpha,1}-i'} q_2^{-j'}
\, , \\
\Xi_{10}(\lambda_\alpha,\lambda_\beta;q_{1,2}) & = \operatorname{ch} \left( \mathsf{I}_\emptyset/\mathsf{I}_{\lambda_\beta} + \mathsf{I}_\emptyset/\mathsf{I}_{\lambda_\alpha} - P_{12}^\vee (\mathsf{I}_\emptyset/\mathsf{I}_{\lambda_\alpha}) (\mathsf{I}_\emptyset/\mathsf{I}_{\lambda_\beta}) \right)
\nonumber \\
& = - \sum_{i=1}^{\check\lambda_{\alpha,1}} \sum_{j'=1}^{\lambda_{\beta,1}}
\left[
q_1^{\check\lambda_{\beta,j'}+i-1} q_2^{\lambda_{\alpha,i}+j'-1}
- q_1^{i-1} q_2^{j'-1}
\right]
\nonumber \\
& \qquad
+ \sum_{(i,j) \in \lambda_\alpha} q_1^{i-1} q_2^{\lambda_{\beta,1}+j-1}
+ \sum_{(i',j') \in \lambda_\beta} q_1^{\check\lambda_{\alpha,1}+i'-1} q_2^{j'-1}
\, .
\end{align}
\end{subequations}
In contrast to the diagonal part $\Xi_{00(11)}$, further simplification does not occur for these off-diagonal ones.
We remark that these off-diagonal factors are symmetric under $\lambda_\alpha \leftrightarrow \lambda_\beta$,
\begin{align}
\Xi_{01}(\lambda_\alpha,\lambda_\beta;q_{1,2}) = \Xi_{01}(\lambda_\beta,\lambda_\alpha;q_{1,2})
\, , \qquad
\Xi_{10}(\lambda_\alpha,\lambda_\beta;q_{1,2}) = \Xi_{10}(\lambda_\beta,\lambda_\alpha;q_{1,2})
\, ,
\end{align}
and we have the relation,
\begin{align}
q_{12} \, \Xi_{01}(\lambda_\alpha,\lambda_\beta;q_{1,2})
= \Xi_{10}(\lambda_\alpha,\lambda_\beta;q_{1,2}^{-1})
\, .
\end{align}
Applying the index functor to this expression, we obtain the instanton partition function.
In addition to the Nekrasov factor for the diagonal part~\eqref{eq:Nekrasov_factor}, we also define the off-diagonal combinatorial factor,
\begin{align}
\mathsf{N}^{01}_{\lambda_\alpha \lambda_\beta}(z;\epsilon_{1,2}) & =
\prod_{i=1}^{\check\lambda_{\alpha,1}} \prod_{j'=1}^{\lambda_{\beta,1}} \frac{[z - i \epsilon_1 - j' \epsilon_2]}{[z - (\check{\lambda}_{\beta,j'} + i) \epsilon_1 - (\lambda_{\alpha,i} + j')]}
\nonumber \\
& \qquad \times
\prod_{(i,j) \in \lambda_\alpha} [z - i \epsilon_1 - (\lambda_{\beta,1} + j)\epsilon_2]
\prod_{(i',j') \in \lambda_\alpha} [z - (\check{\lambda}_{\alpha,1} + i') \epsilon_1 - j' \epsilon_2]
\, , \\
\mathsf{N}^{10}_{\lambda_\alpha \lambda_\beta}(z;\epsilon_{1,2}) & =
\prod_{i=1}^{\check\lambda_{\alpha,1}} \prod_{j'=1}^{\lambda_{\beta,1}} \frac{[z + (i-1) \epsilon_1 + (j'-1) \epsilon_2]}{[z + (\check{\lambda}_{\beta,j'} + i - 1) \epsilon_1 + (\lambda_{\alpha,i} + j' - 1)]}
\nonumber \\
& \qquad \times
\prod_{(i,j) \in \lambda_\alpha} [z + (i-1) \epsilon_1 + (\lambda_{\beta,1} + j - 1)\epsilon_2]
\prod_{(i',j') \in \lambda_\alpha} [z + (\check{\lambda}_{\alpha,1} + i' - 1) \epsilon_1 + (j' - 1) \epsilon_2]
\, .
\end{align}
The combinatorial form of the $k$-instanton partition function for $\mathrm{U}(n_0|n_1)$ supergroup gauge theory is given as follows.
\begin{itembox}{Nekrasov partition function for $\mathrm{U}(n_0|n_1)$ gauge theory}
The instanton partition function of $\mathrm{U}(n_0|n_1)$ gauge theory has the following combinatorial expression,
\begin{align}
Z_k = \sum_{\lambda \in \mathfrak{M}_{n,k}^\mathsf{T}} \frac{\prod_{\alpha \in [n_0], \beta \in [n_1]} \mathsf{N}^{01}_{\lambda_\alpha^0 \lambda_\beta^1}(a_\alpha^0 - a_\beta^1;\epsilon_{1,2}) \mathsf{N}^{10}_{\lambda_\beta^1 \lambda_\alpha^0}(a_\beta^1-a_\alpha^0;\epsilon_{1,2})}{\prod_{\sigma = 0, 1} \prod_{\alpha,\beta \in [n_\sigma]} \mathsf{N}_{\lambda_\alpha^\sigma \lambda_\beta^\sigma}(a_\alpha^\sigma - a_\beta^\sigma;\epsilon_{1,2})}
\, .
\end{align}
\end{itembox}
We see that the denominator factors are the standard gauge node contributions, whereas the numerator factors are analogous, but not identical to the bifundamental hypermultiplet contributions.
\section{Non-perturbative aspects of supergroup gauge theory}\label{sec:non-perturbative_study}
In this section, we explore non-perturbative aspects of supergroup gauge theory based on the instanton partition function obtained in the previous section.
\subsection{Topological string theory approach}
We have seen in Sec.~\ref{sec:string_theory_perspective} that string/M-theory provides a geometric description of $\mathcal{N}=2$ gauge theory.
Choosing the Calabi--Yau (CY) three-fold constructed from resolution of specific singularities, one can obtain the corresponding Seiberg--Witten geometry.
This process is known as geometric engineering~\cite{Katz:1996fh,Katz:1997eq}.
This description can be further pursued at the level of the microscopic instanton counting:
The A-model topological string amplitude of the corresponding non-compact toric CY three-fold computes the instanton partition function~\cite{Iqbal:2003ix,Iqbal:2003zz,Eguchi:2003sj}.
The primary example is the local $\mathbb{P}^1 \times \mathbb{P}^1$ geometry, which corresponds to pure SU(2) SYM theory.
In this case, we have the following web diagram, which is dual to the toric diagram,
\begin{align}
\begin{tikzpicture}[baseline=(current bounding box.center),scale = 1]
\draw (-.5,-.5) rectangle (.5,.5);
\draw (.5,.5) -- ++(45:.5);
\draw (-.5,.5) -- ++(135:.5);
\draw (.5,-.5) -- ++(-45:.5);
\draw (-.5,-.5) -- ++(225:.5);
%
\draw[very thick,blue,latex-latex] (1.5,0) -- ++(2,0);
%
\node at (0,-1.75) {web diagram};
%
\begin{scope}[shift={(5.5,0)}]
\draw[dotted] (0,1) -- (1,0) -- (0,-1) --(-1,0) -- cycle;
\draw[dotted] (-1,0) -- ++(2,0);
\draw[dotted] (0,-1) -- ++(0,2);
\filldraw (0,0) circle (.05);
\filldraw (1,0) circle (.05);
\filldraw (0,1) circle (.05);
\filldraw (0,-1) circle (.05);
\filldraw (-1,0) circle (.05);
%
\node at (0,-1.75) {toric diagram};
%
\end{scope}
\end{tikzpicture}
\end{align}
This web diagram can be identified with the configuration of $(p,q)$-brane in type IIB theory obtained from the T-dual of the Hanany--Witten configuration discussed in Sec.~\ref{sec:Hany-Witten_const}.
In order to compute the topological string amplitude, we may apply a systematic approach, called the topological vertex~\cite{Aganagic:2003db}.
The idea is to describe the toric CY three-fold by gluing the local $\mathbb{C}^3$ patches, corresponding to each trivalent vertex in the web diagram.
We denote the topological vertex function with generic boundary conditions parametrized by the partitions by
\begin{align}
C_{\lambda \mu \nu}(q) = q^{\frac{\kappa(\mu)}{2}} s_{\check{\nu}}(q^{-\rho}) \sum_{\eta} s_{\check{\lambda}/\eta}(q^{-\nu - \rho})
s_{\mu/\eta}(q^{-\check{\nu} - \rho}) = \quad
\begin{tikzpicture}[baseline=(current bounding box.center),scale=.7]
\draw[-latex] (0,0) -- ++(1,0) node [right] {$\mu$};
\draw[-latex] (0,0) -- ++(0,1) node [above] {$\lambda$};
\draw[-latex] (0,0) -- ++(225:1) node [left] {$\nu$};
\end{tikzpicture}
\end{align}
where we denote the skew Schur function by $s_{\lambda/\eta}(X) = \sum_{\mu} c_{\eta\mu}^\lambda s_\mu(X)$ with the Littlewood--Richardson coefficient $c_{\eta\mu}^\lambda$.
The combinatorial factor is defined by $\kappa(\lambda)/2 = \sum_{(i,j) \in \lambda} (j-i)$.
We denote the Weyl vector by $\rho$, such that $q^{-\lambda - \rho} = \{ q^{-\lambda_i + i - \frac{1}{2}} \}_{i \in \mathbb{N}}$.
The string coupling $g_s$ with $q = \mathrm{e}^{g_s}$ would be identified with the $\Omega$-background parameters at the self-dual point (unrefined situation), $g_s = \epsilon_2 = - \epsilon_1$.
In order to reproduce generic $\epsilon_{1,2}$, we need the refinement of the topological vertex~\cite{Awata:2005fa,Iqbal:2007ii}.
From this point of view, recalling that the web diagram describes $(p,q)$-branes, we introduce the anti-topological vertex (anti-vertex for short) involving the negative coupling to realize supergroup gauge theory~\cite{Vafa:2001qf,Kimura:2020lmc},
\begin{align}
\overline{C}_{\lambda \mu \nu} (q) = {C}_{\lambda \mu \nu} (q^{-1})
= \quad
\begin{tikzpicture}[baseline=(current bounding box.center),scale=.7]
\draw[-latex,dashed] (0,0) -- ++(1,0) node [right] {$\mu$};
\draw[-latex,dashed] (0,0) -- ++(0,1) node [above] {$\lambda$};
\draw[-latex,dashed] (0,0) -- ++(225:1) node [left] {$\nu$};
\end{tikzpicture}
\end{align}
We can check that this vertex--anti-vertex formalism reproduces the instanton partition function of supergroup gauge theory~\cite{Kimura:2020lmc}.
Moreover, this vertex formalism exhibits further interesting features.
As discussed in Sec.~\ref{sec:Hany-Witten_const}, the brane realization is not unique for supergroup gauge theory.
In fact, we can see that the same instanton partition function can be obtained from different configurations corresponding to the possible Dynkin diagrams~\eqref{eq:HW_U(2|1)}.
Another remark is the relation to non-supergroup gauge theory.
From the combinatorial property of the Schur function, flipping the coupling $q \to q^{-1}$ yields the transposition of the partition $\lambda \to \check{\lambda}$, etc, which leads to the relation
\begin{align}
\begin{tikzpicture}[baseline=(current bounding box.center),scale=.7]
\draw[-latex] (0,0) -- ++(1,0) node [right] {$\mu$};
\draw[-latex] (0,0) -- ++(0,1) node [above] {$\lambda$};
\draw[-latex] (0,0) -- ++(225:1) node [left] {$\nu$};
\end{tikzpicture}
\quad \simeq \quad
\begin{tikzpicture}[baseline=(current bounding box.center),xscale=-1,scale=.7]
\draw[-latex,dashed] (0,0) -- ++(1,0) node [left] {$\mu$};
\draw[-latex,dashed] (0,0) -- ++(0,1) node [above] {$\lambda$};
\draw[-latex,dashed] (0,0) -- ++(225:1) node [right] {$\nu$};
\end{tikzpicture}
\end{align}
up to the framing factor.
We remark that such a reflection relation holds only for the unrefined case.
Hence, we have realization of $\mathrm{U}(1|1)$ theory in terms of the ordinary vertices,
\begin{align}
\begin{tikzpicture}[baseline=(current bounding box.center)]
\draw (0,0) -- (1,0);
\draw (0,0) -- ++(225:.5);
\draw (1,0) -- ++(-45:.5);
\draw[dashed] (225:.5) -- ++(225:.5) -- ++(2.414,0) -- ++(135:.5);
\draw (0,0) -- ++(90:.5);
\draw (1,0) -- ++(90:.5);
\draw[dashed] (1,0)++(-45:1) -- ++(-30:.5);
\draw[dashed] (225:1) -- ++(210:.5);
\end{tikzpicture}
\qquad \simeq \qquad
\begin{tikzpicture}[baseline=(current bounding box.center)]
\draw (0,0) -- (1,0);
\draw (0,0) -- ++(225:1);
\draw (1,0) -- ++(-45:1);
\draw (0,0) -- ++(90:.5);
\draw (1,0) -- ++(90:.5);
\draw (1,0)++(-45:1)++(.5,0) -- ++(-.5,0) -- ++(0,-.5);
\draw (225:1)++(-.5,0) -- ++(.5,0) -- ++(0,-.5);
\end{tikzpicture}
\end{align}
This configuration gives rise to $\mathrm{U}(1)$ theory with two fundamental matters, which is consistent with the gauging trick discussed in Sec.~\ref{sec:gauging_trick}.
Let us comment on a algebraic interpretation of the anti-vertex.
It has been known that the refined topological vertex is understood as the intertwiner of quantum troidal algebra of $\mathfrak{gl}_1$~\cite{Awata:2011ce}, known as Ding--Iohara--Miki algebra~\cite{Ding:1996mq,Miki:2007JMP}.
In this context, the anti-vertex is realized as the intertwiner involving the negative level~\cite{Bourgine:2018fjy,Noshita:2022dxv}.
See also~\cite{Rapcak:2019wzw} for a related realization of superalgebra from the CY geometry.
\subsection{Non-perturbative Schwinger--Dyson equation}\label{sec:SD_eq}
The instanton partition function is given by summation over the fixed point configurations, which can be related to each other by the process to add/remove instantons.
We discuss the behavior of instanton partition function under such a non-perturbative process of adding/removing instantons, which gives rise to functional relations that we call the \emph{non-perturbative Schwinger--Dyson equations}~\cite{Kanno:2012hk,Kanno:2013aha,Bourgine:2015szm,Nekrasov:2015wsu,Kimura:2015rgi}.
\subsubsection{Adding/removing instantons}
We consider instanton-adding process by shifting the vector space, $K \to K + V$ with $V = \mathbb{C}^v$, where $v$ is the number of instantons that we add to the configuration.
Under this shift, the tangent bundle at the fixed point $\lambda \in \mathfrak{M}_{n,k}^\mathsf{T}$ behaves as follows,
\begin{align}
\delta_V T_\lambda \mathfrak{M}_{n,k}
& = \mathsf{Y}^\vee V + Q_{12}^\vee V^\vee \mathsf{Y} - P_{12}^{-1} V^\vee V
\, .
\label{eq:tangent_bundle_shift+}
\end{align}
Hence, this shift defines a map, $\delta_V$ : $\mathfrak{M}_{n,k} \to \mathfrak{M}_{n,k+v}$.
Applying the index functor, we obtain
\begin{align}
\delta_V Z_\lambda = \mathbb{I}[-\delta_V T_\lambda \mathfrak{M}_{n,k}] = \frac{Z_{k+v}}{Z_\lambda} = \frac{1}{v!} \frac{[-\epsilon_{12}]^v}{[-\epsilon_{1,2}]^v} \oint \mathrm{d}{\underline{\phi}} \frac{1}{\mathscr{Y}_\lambda(\underline{\phi}) \widetilde{\mathscr{Y}}_\lambda(\underline{\phi}+\epsilon_{12})} \prod_{I \neq J}^v \mathscr{S}(\phi_{IJ})^{-1}
\, ,
\end{align}
where we define the $\mathscr{Y}$-function with the line bundle $\mathsf{x}$ whose Chern root is given by $x$,
\begin{subequations}\label{eq:Y-func_def}
\begin{align}
\mathscr{Y}_\lambda(x) & = \mathbb{I}[\mathsf{Y}^\vee \mathbf{x}] \stackrel{\eqref{eq:obs_sheaf_def}}{=} \prod_{x' \in \mathcal{X}_\lambda^{[\log]}} \frac{[x - x']}{[x - x' - \epsilon_1]}
\stackrel{\eqref{eq:obs_sheaf_comb}}{=} \frac{\prod_{x' \in \mathcal{X}_{\partial_+\lambda}^{[\log]}} [x - x']}{\prod_{x' \in \mathcal{X}_{\partial_-\lambda}^{[\log]}} [x - x' - \epsilon_{12}]}
\, , \\
\widetilde{\mathscr{Y}}_\lambda(x) & = \mathbb{I}[ \mathbf{x}^\vee \mathsf{Y}] \stackrel{\eqref{eq:obs_sheaf_def}}{=} \prod_{x' \in \mathcal{X}_\lambda^{[\log]}} \frac{[x' - x]}{[x' - x + \epsilon_1]}
\stackrel{\eqref{eq:obs_sheaf_comb}}{=} \frac{\prod_{x' \in \mathcal{X}_{\partial_+\lambda}^{[\log]}} [x' - x]}{\prod_{x' \in \mathcal{X}_{\partial_-\lambda}^{[\log]}} [x' - x + \epsilon_{12}]}
\, ,
\end{align}
\end{subequations}
with the set
\begin{align}
\mathcal{X}_\bullet^{[\log]} = \{ \log x \mid x \in \mathcal{X}_\bullet \}
\, .
\end{align}
We have the relation for these two $\mathscr{Y}$-functions,
\begin{align}
\mathscr{Y}(z) =
\begin{cases}
(-1)^n \widetilde{\mathscr{Y}}(z) & (4d) \\
(-1)^n \mathrm{e}^{- n z} \mathrm{e}^{\sum_{\alpha \in [n]} a_\alpha} \widetilde{\mathscr{Y}}(z) & (5d \ \& \ 6d) \\
\end{cases}
.
\label{eq:Y-fn_convert}
\end{align}
We denote $\mathscr{Y}_\lambda(\underline{\phi}) = \prod_{I \in [v]} \mathscr{Y}_\lambda(\phi_I)$ and $\mathrm{d}{\underline{\phi}} = \prod_{I \in [v]} \mathrm{d}{\phi}_I/2 \pi \mathrm{i}$.
We focus on the simplest case, $v = 1$.
In this case, we evaluate the contour integral surrounding the poles of $1/\mathscr{Y}_\lambda(\phi)$ based on the finite product form of the $\mathscr{Y}$-function,
\begin{align}
\delta_V Z_\lambda & = \frac{[-\epsilon_{12}]}{[-\epsilon_{1,2}]} \oint \frac{\mathrm{d}{\phi}}{2 \pi \mathrm{i}} \frac{1}{\mathscr{Y}_\lambda({\phi}) \widetilde{\mathscr{Y}}_\lambda({\phi}+\epsilon_{12})}
\nonumber \\
& = \frac{1}{[-\epsilon_{1,2}]} \sum_{x \in \mathcal{X}_{\partial_+ \lambda}^{[\log]}} \frac{\prod_{x' \in \mathcal{X}_{\partial_- \lambda}^{[\log]}} [x - x' - \epsilon_{12}][x' - x]}{\prod_{x' \in \mathcal{X}_{\partial_+ \lambda}^{[\log]}\backslash\{x\}} [x - x'][x' - x - \epsilon_{12}]}
= \sum_{x \in \mathcal{X}_{\partial_+ \lambda}^{[\log]}} \frac{-1}{\mathscr{Y}_\lambda(x) \widetilde{\mathscr{Y}}_{\lambda'}(x+\epsilon_{12})}
\, ,
\end{align}
where the configuration $\lambda'$ is obtained by shift (adding an instanton).
The last equality is shown as follows,
\begin{align}
\frac{1}{[-\epsilon_{1,2}]}
\frac{\prod_{x' \in \mathcal{X}_{\partial_- \lambda}^{[\log]}} [x - x' - \epsilon_{12}][x' - x]}{\prod_{x' \in \mathcal{X}_{\partial_+ \lambda}^{[\log]}\backslash\{x\}} [x - x'][x' - x - \epsilon_{12}]}
& =
\frac{1}{\mathscr{Y}_\lambda(x) \widetilde{\mathscr{Y}}_{\lambda}(x+\epsilon_{12})} \times \lim_{x'' \to x} \frac{[x - x''][x'' - x - \epsilon_{12}]}{[-\epsilon_{1,2}]}
\nonumber \\
& =
\frac{1}{\mathscr{Y}_\lambda(x) \widetilde{\mathscr{Y}}_{\lambda}(x+\epsilon_{12})}
\nonumber \\
& \qquad \times \lim_{x'' \to x} \frac{[x - x''][x'' - x - \epsilon_{12}][x'' - x - \epsilon_{1,2}]}{[-\epsilon_{1,2}][x'' - x][x'' - x - \epsilon_{12}]}
\nonumber \\
& = \frac{-1}{\mathscr{Y}_\lambda(x) \widetilde{\mathscr{Y}}_{\lambda'}(x+\epsilon_{12})}
\, .
\end{align}
\if0
\begin{align}
\prod_{x' \in \mathcal{X}_\lambda^{[\log]}\backslash\{x\}} \frac{[x - x' - \epsilon_1][x' - x - \epsilon_2]}{[x - x'][x' - x - \epsilon_{12}]}
& = \frac{1}{\mathscr{Y}_\lambda(x) \widetilde{\mathscr{Y}}_{\lambda}(x+\epsilon_{12})} \times \lim_{x'' \to x} \frac{[x - x''][x'' - x - \epsilon_{12}]}{[x -x'' - \epsilon_1][x'' - x - \epsilon_{2}]}
\nonumber \\
& = \frac{1}{\mathscr{Y}_\lambda(x) \widetilde{\mathscr{Y}}_{\lambda'}(x+\epsilon_{12})} \times \lim_{x'' \to x} \frac{[x - x''][x'' - x - \epsilon_{1}]}{[x -x'' - \epsilon_1][x'' - x]}
\nonumber \\
& = \frac{-1}{\mathscr{Y}_\lambda(x) \widetilde{\mathscr{Y}}_{\lambda'}(x+\epsilon_{12})}
\, .
\end{align}
Although there exist formally infinitely many poles in the integrand, we have poles at the place where we can add a box to the partition $\lambda$.
\fi
From this expression, we identify $Z_{\lambda'}/Z_\lambda = -1/\mathscr{Y}_\lambda(x) \widetilde{\mathscr{Y}}_{\lambda'}(x+\epsilon_{12})$, from which we obtain
\begin{align}
\res_{x' \to x} \left[ Z_{\lambda'} \widetilde{\mathscr{Y}}_{\lambda'}(x'+\epsilon_{12}) + Z_\lambda \frac{1}{{\mathscr{Y}}_{\lambda}(x')} \right] = 0
\, .
\end{align}
Summing up all the configurations, we see that
\begin{align}
\mathscr{T}(x) := \left< \widetilde{\mathscr{Y}}(x+\epsilon_{12}) + \frac{\mathfrak{q}}{\mathscr{Y}(x)} \right>
\end{align}
is a pole-free regular function in $x$, where we define the gauge theory average of the observable by $\left< \mathcal{O} \right> = \frac{1}{Z} \sum_{\lambda \in \mathfrak{M}^\mathsf{T}} Z_\lambda \mathcal{O}_\lambda$.
The combination of the $\mathscr{Y}$-functions yielding a pole-free function is called the (average of) $qq$-character (of $A_1$ quiver in this case)~\cite{Nekrasov:2015wsu}.%
\footnote{%
The character of the fundamental representation of SL(2) is given by $y + y^{-1}$.
In general, the Seiberg--Witten geometry of quiver gauge theory is described using the fundamental representation characters of the Lie group associated with the quiver structure~\cite{Nekrasov:2012xe}.
Considering the situation $(\epsilon_1, \epsilon_2) = (\hbar,0)$, this character is promoted to the $q$-character~\cite{Nekrasov:2013xda} of Yangian/quantum affine algebra~\cite{Knight:1995JA,Frenkel:1998}.
The $qq$-character is a double quantum deformation of the character appearing in the presence of generic $\Omega$-background parameters.
Physically, the $qq$-character is realized as a codimension-four defect operator~\cite{Kim:2016qqs,Agarwal:2018tso,Haouzi:2019jzk}.
}
Indeed, the Seiberg--Witten curve is obtained in the classical limit $\epsilon_{1,2} \to 0$ of this relation, $y + \mathfrak{q}/y = \det(x - \phi)$ by identifying the characteristic polynomial by $\mathscr{T}(x)$.%
\footnote{%
Precisely speaking, we should convert $\widetilde{\mathscr{Y}}$ to $\mathscr{Y}$ using the relation~\eqref{eq:Y-fn_convert} providing an additional polynomial factor, which can be interpreted as the Chern--Simons term contribution in five dimensions.
}
We remark that in this limit, we can apply the saddle point analysis, so that the observable average can be replaced by the on-shell value with the saddle point configuration $\lambda_*$, $\left< \mathscr{Y}(x) \right> \to \mathscr{Y}_{\lambda_*}(x) = y(x)$ and $\left< 1/ \mathscr{Y}(x) \right> \to 1/\mathscr{Y}_{\lambda_*}(x) = 1/y(x)$.
We may consider the removing process in a similar way.
In this case, we consider the shift of the form, $K \to K - V$.
The tangent bundle behaves as
\begin{align}
\delta_{-V} T_\lambda \mathfrak{M}_{n,k}
& = - \mathsf{Y}^\vee V - Q_{12}^\vee V^\vee \mathsf{Y} - P_{12}^{-1} V^\vee V
\, ,
\end{align}
where $\delta_{-V}$ : $\mathfrak{M}_{n,k} \to \mathfrak{M}_{n,k-v}$, and the index functor yields
\begin{align}
\delta_{-V} Z_\lambda = \mathbb{I}[-\delta_{-V} T_\lambda \mathfrak{M}_{n,k}] = \frac{Z_{k-v}}{Z_\lambda} = \frac{1}{v!} \frac{[-\epsilon_{12}]^v}{[-\epsilon_{1,2}]^v} \oint \mathrm{d}{\underline{\phi}} \mathscr{Y}_\lambda(\underline{\phi}) \widetilde{\mathscr{Y}}_\lambda(\underline{\phi}+\epsilon_{12}) \prod_{I \neq J}^v \mathscr{S}(\phi_{IJ})^{-1}
\, .
\end{align}
In the case of $v = 1$, we obtain
\begin{align}
\delta_{-V} Z_\lambda & = \frac{[-\epsilon_{12}]}{[-\epsilon_{1,2}]} \oint \frac{\mathrm{d}{\phi}}{2 \pi \mathrm{i}} \mathscr{Y}_\lambda({\phi}) \widetilde{\mathscr{Y}}_\lambda({\phi}+\epsilon_{12})
= \sum_{x \in \mathcal{X}_{\partial_- \lambda}^{[\log]}} - \mathscr{Y}_{\lambda'}(x) \widetilde{\mathscr{Y}}_{\lambda}(x+\epsilon_{12})
\, ,
\end{align}
where we denote the configuration obtained by removing an instanton by $\lambda'$.
We can similarly obtain the $qq$-character from this expression as well.
\subsubsection{Supergroup analysis}
We consider the adding/removing instanton process for supergroup gauge theory.
Denoting the observable sheaf $\mathsf{Y} = \mathsf{Y}_0 \oplus \mathsf{Y}_1$, the character of even and odd part is given as follows,
\begin{align}
\operatorname{ch} \mathsf{Y}_0\Big|_{\lambda} =
\sum_{x \in \mathcal{X}_{\partial_+ \lambda^0}} x - \sum_{x \in \mathcal{X}_{\partial_- \lambda^0}} x q_{12}
\, , \qquad
\operatorname{ch} \mathsf{Y}_1\Big|_{\lambda} =
\sum_{x \in \mathcal{X}_{\partial_+ \lambda^1}} x q_{12} - \sum_{x \in \mathcal{X}_{\partial_- \lambda^1}} x
\end{align}
where we define
\begin{align}
\mathcal{X}_{\partial_\pm \lambda^0} = \{ \mathrm{e}^{a_\alpha^0} q_1^{i-1} q_2^{j-1} \mid (i,j) \in \partial_\pm \lambda^0 \}
\, , \qquad
\mathcal{X}_{\partial_\pm \lambda^1} = \{ \mathrm{e}^{a_\alpha^1} q_1^{-i} q_2^{-j} \mid (i,j) \in \partial_\pm \lambda^1 \}
\, .
\end{align}
The $\mathscr{Y}$-function is then defined as follows,
\begin{align}
\mathscr{Y}_\lambda(z) = \frac{\mathscr{Y}_{0,\lambda^0}(z)}{\mathscr{Y}_{1,\lambda^1}(z)}
\, , \qquad
\widetilde{\mathscr{Y}}_\lambda(z) = \frac{\widetilde{\mathscr{Y}}_{0,\lambda^0}(z)}{\widetilde{\mathscr{Y}}_{1,\lambda^1}(z)}
\, ,
\end{align}
where each factor is given by
\begin{subequations}
\begin{align}
\mathscr{Y}_{0,\lambda^0}(z) = \frac{\prod_{x \in \mathcal{X}_{\partial_+\lambda^0}^{[\log]}} [z - x]}{\prod_{x \in \mathcal{X}_{\partial_-\lambda^0}^{[\log]}} [z - x - \epsilon_{12}]}
\, , \qquad &
\mathscr{Y}_{1,\lambda^1}(z) = \frac{\prod_{x \in \mathcal{X}_{\partial_+\lambda^1}^{[\log]}} [z - x - \epsilon_{12}]}{\prod_{x \in \mathcal{X}_{\partial_-\lambda^1}^{[\log]}} [z - x]}
\, , \\
\widetilde{\mathscr{Y}}_{0,\lambda^0}(z) = \frac{\prod_{x \in \mathcal{X}_{\partial_+\lambda^0}^{[\log]}} [x - z]}{\prod_{x \in \mathcal{X}_{\partial_-\lambda^0}^{[\log]}} [x - z + \epsilon_{12}]}
\, , \qquad &
\widetilde{\mathscr{Y}}_{1,\lambda^1}(z) = \frac{\prod_{x \in \mathcal{X}_{\partial_+\lambda^1}^{[\log]}} [x - z + \epsilon_{12}]}{\prod_{x \in \mathcal{X}_{\partial_-\lambda^1}^{[\log]}} [x - z]}
\, .
\end{align}
\end{subequations}
Under the shift $K \to K + V$ by $V = \mathbb{C}^{v_0|v_1}$, we have the same expression for the tangent bundle as before~\eqref{eq:tangent_bundle_shift+}.
The resulting contour integral is given as follows,
\begin{align}
\delta_V Z_\lambda = \frac{1}{v_{0,1}!} \frac{[-\epsilon_{12}]^{v_{01}}}{[-\epsilon_{1,2}]^{v_{01}}} \oint \mathrm{d}{\underline{\phi}} \frac{1}{\mathscr{Y}_\lambda(\underline{\phi}) \widetilde{\mathscr{Y}}_\lambda(\underline{\phi}+\epsilon_{12})} \frac{\prod^{I \in [v_0]}_{J \in [v_1]} \mathscr{S}(\phi^0_I - \phi^1_J) \mathscr{S}(\phi^1_J - \phi^0_I) }{\prod_{I \neq J}^{v_{0}} \mathscr{S}(\phi^0_{IJ}) \prod_{I \neq J}^{v_{1}} \mathscr{S}(\phi^1_{IJ})}
\, .
\end{align}
The one-dimensional cases $v = (1|0)$ and $(0|1)$ are simultaneously formulated by
\begin{align}
\delta_V Z_\lambda & = \frac{[-\epsilon_{12}]}{[-\epsilon_{1,2}]} \oint \frac{\mathrm{d}{\phi}}{2 \pi \mathrm{i}} \frac{1}{\mathscr{Y}_\lambda({\phi}) \widetilde{\mathscr{Y}}_\lambda({\phi}+\epsilon_{12})}
= \frac{[-\epsilon_{12}]}{[-\epsilon_{1,2}]} \oint \frac{\mathrm{d}{\phi}}{2 \pi \mathrm{i}} \frac{\mathscr{Y}_{1,\lambda^1}({\phi}) \widetilde{\mathscr{Y}}_{1,\lambda^1}({\phi}+\epsilon_{12})}{\mathscr{Y}_{0,\lambda^0}({\phi}) \widetilde{\mathscr{Y}}_{0,\lambda^0}({\phi}+\epsilon_{12})}
\nonumber \\ &
= \frac{[-\epsilon_{12}]}{[-\epsilon_{1,2}]} \oint \frac{\mathrm{d}{\phi}}{2 \pi \mathrm{i}} \frac{\prod_{x \in \mathcal{X}_{\partial_-\lambda^0}^{[\log]}} [\phi - x - \epsilon_{12}][x - \phi]}{\prod_{x \in \mathcal{X}_{\partial_+\lambda^0}^{[\log]}} [\phi - x][x - \phi - \epsilon_{12}]}
\frac{\prod_{x \in \mathcal{X}_{\partial_+\lambda^1}^{[\log]}} [\phi - x - \epsilon_{12}][x - \phi]}{\prod_{x \in \mathcal{X}_{\partial_-\lambda^1}^{[\log]}} [\phi - x][x - \phi - \epsilon_{12}]}
\, .
\end{align}
From this expression, we observe that adding an instanton to the positive node (positive instanton) is equivalent to removing an instanton from the negative node (negative instanton).
Similarly, removing a positive instanton is equivalent to adding a negative instanton from this point of view.
Therefore, we obtain the same $qq$-character expression in terms of the total $\mathscr{Y}$-functions,
\begin{align}
\mathscr{T}(x)
= \left< \widetilde{\mathscr{Y}}(x+\epsilon_{12}) + \frac{\mathfrak{q}}{\mathscr{Y}(x)} \right>
= \left< \frac{\widetilde{\mathscr{Y}}_0(x+\epsilon_{12})}{\widetilde{\mathscr{Y}}_1(x+\epsilon_{12})} + \mathfrak{q} \frac{\mathscr{Y}_1(x)}{\mathscr{Y}_0(x)} \right>
\, ,
\end{align}
which is again consistent with $\widehat{A}_1$ quiver realization in Sec.~\ref{sec:quiver_realization} by identifying $\mathscr{Y}_{0,1}(x)$ with the $\mathscr{Y}$-functions of $\widehat{A}_1$ quiver.
Identifying the $\mathscr{T}$-function with the supercharacteristic function of the adjoint scalar, $\mathscr{T}(x) = \sdet(x - \Phi)$, we reproduce the Seiberg--Witten curve for $\mathrm{U}(n_0|n_1)$ theory discussed in Sec.~\ref{sec:SW_theory} in the classical limit $\epsilon_{1,2} \to 0$.
\subsubsection{Geometry of $qq$-character}
Let us comment on geometric representation theoretical perspectives of the $qq$-character.
We reconsider the shift $K \to K + V$ in the presence of the $\mathscr{Y}$-functions,
\begin{align}
\delta_V \left( \mathsf{Y}^\vee W - T\mathfrak{M}_{n,k} \right)
& =
- \mathsf{Y}^\vee V - Q_{12}^\vee V^\vee \mathsf{Y} - P_{12}^\vee V^\vee W + P_{12}^\vee V^\vee V
\, ,
\end{align}
where we define $W = \mathbb{C}^w$ with the character $\operatorname{ch} W = \sum_{\alpha \in [w]} \mathrm{e}^{\xi_\alpha}$, such that $\mathbb{I}[\mathsf{Y}^\vee W] = \prod_{\alpha \in [w]} \mathscr{Y}(\xi_\alpha)$.
The generic $qq$-character is then obtained by the following index formula,
\begin{align}
\mathsf{T}_w = \sum_{v = 0}^\infty \mathfrak{q}^v \, \mathsf{T}_{w,v}
\, ,
\end{align}
where each contribution is given by
\begin{align}
\mathsf{T}_{w,v}
& = \mathbb{I}[ \mathsf{Y}^\vee (W - (1 + Q_{12}^\vee) V) - P_{12}^\vee V^\vee W + P_{12}^\vee V^\vee V ]
\nonumber \\
& = \frac{1}{v!} \frac{[-\epsilon_{12}]^v}{[-\epsilon_{1,2}]^v} \oint \mathrm{d}{\underline{\phi}} \frac{\mathscr{Y}(\underline{\xi})}{\mathscr{Y}(\underline{\phi}) \mathscr{Y}(\underline{\phi}-\epsilon_{12})}
\mathscr{S}(\underline{\xi} - \underline{\phi}) \prod_{I \neq J}^v \mathscr{S}(\phi_{IJ})^{-1}
\, .
\label{eq:T_wv_A1}
\end{align}
We denote $\mathscr{Y}(\underline{\xi}) = \prod_{\alpha \in [w]} \mathscr{Y}(\xi_\alpha)$, $\mathscr{S}(\underline{\xi} - \underline{\phi}) = \prod_{\alpha \in [w], I \in [v]} \mathscr{S}(\xi_\alpha - \phi_I)$, etc.
We remark that the dual function $\widetilde{\mathscr{Y}}$ is converted to $\mathscr{Y}$ for convenience.
Taking the pole at $\phi_I = \xi_\alpha$ in the contour integral, we obtain the result~\cite{Nekrasov:2015wsu,Kimura:2015rgi},
\begin{align}
\mathsf{T}_{w,v} & = \sum_{\substack{\mathsf{I} \bigsqcup \mathsf{J} = [w] \\ |\mathsf{J}| = v}} \prod_{\alpha \in \mathsf{I}, \beta \in \mathsf{J}} \mathscr{S}(\xi_\alpha - \xi_\beta) \frac{\prod_{\alpha \in \mathsf{I}} \mathscr{Y}(\xi_\alpha)}{\prod_{\beta \in \mathsf{J}} \mathscr{Y}(\xi_\beta - \epsilon_{12})}
\, .
\end{align}
There are ${w \choose v}$ contributions in $\mathsf{T}_{w,v}$ for given $(w,v)$, which correspond to the fixed points in the cotangent bundle of the Grassmannian $\mathrm{Gr}(v,w)$ given as a quiver variety of type $A_1$.
From the representation theoretical point of view, this corresponds to the degree-$w$ tensor product of the fundamental representation of quantum affine algebra $U_q(\widehat{\mathfrak{sl}_2})$ associated with $A_1$ quiver~\cite{Nakajima:1994nid,Nakajima:1998DM,Nakajima:1999JAMS}.
In order to obtain the $qq$-character of the irreducible representation, we need to specialize the parameters $(\mathrm{e}^{\xi_\alpha})_{\alpha \in [w]} \to (\mathrm{e}^\xi, \mathrm{e}^\xi q_1, \ldots, \mathrm{e}^\xi q_1^{w-1})$, known as the $q$-segment condition~\cite{Chari:1991CMP,Chari:1994pf}.
See also~\cite{Kimura:2022spi}.
The contribution to the $qq$-character $\mathsf{T}_{w,v}$ shown in~\eqref{eq:T_wv_A1} has more direct interpretation in terms of the quiver variety of type $A_1$, which we denote by $\mathfrak{M}_{w,v} = \{ (I,J) \in \operatorname{Hom}(W,V) \oplus \operatorname{Hom}(V,W) \mid IJ = 0 \} /\!\!/ \operatorname{GL}(V)$ with the dimension, $\operatorname{dim} \mathfrak{M}_{w,v} = 2 v(w - v)$.
Hence, it becomes empty when $v > w$.
Applying the same argument to Sec.~\ref{sec:tangent_bundle}, the tangent bundle is given by $T\mathfrak{M}_{v,w} = W^\vee V + Q_{12}^\vee V^\vee W - (1 + Q_{12}^\vee) V^\vee V$.
Denoting $c = 1 + Q_{12}^\vee$ ($q$-Cartan matrix of type $A_1$, see the definition~\eqref{eq:full-q-Cartan}), we obtain a geometric formula in the five-dimensional convention~\cite{Nekrasov:2015wsu,KPfractional},
\begin{align}
\mathsf{T}_{w,v} = q_2^{-\frac{1}{2} \operatorname{dim} \mathfrak{M}_{w,v}}\int_{\mathfrak{M}_{w,v}} \frac{\operatorname{ch} \wedge \mathsf{Y} W^\vee}{\operatorname{ch} \wedge \mathsf{Y} c^\vee V^\vee} \operatorname{ch} \wedge_{q_2} T^\vee\mathfrak{M}_{w,v} \operatorname{td} (T\mathfrak{M}_{w,v})
\, .
\label{eq:qq-ch_geom}
\end{align}
Moreover, we may rewrite this formula as follows,
\begin{align}
\mathsf{T}_{w,v} = \int_{\mathfrak{M}_{w,v}} \frac{\operatorname{ch} \wedge \mathsf{Y} W^\vee}{\operatorname{ch} \wedge \mathsf{Y} c^\vee V^\vee} \widehat{X}_y (T\mathfrak{M}_{w,v})
\, ,
\label{eq:qq-ch_geom_nom}
\end{align}
where we define another genus, an analog of $\widehat{A}$ genus (see, e.g.,~\cite{Gottsche:2015AG}), by
\begin{align}
\widehat{X}_y(\mathbf{X}) = \prod_{i \in [\operatorname{rk} \mathbf{X}]} \frac{x_i ( \mathrm{e}^{\frac{x_i}{2}} y^{
-\frac{1}{2}} - \mathrm{e}^{-\frac{x_i}{2}} y^{\frac{1}{2}} )}{\mathrm{e}^{\frac{x_i}{2}} - \mathrm{e}^{-\frac{x_i}{2}}}
\, .
\end{align}
We remark that this genus is reduced to the $L$-genus when we take $y^{\frac{1}{2}} = \mathrm{i}$, up to the normalization $(-1)^{\frac{1}{2} \operatorname{rk} \mathbf{X}}$.
From this point of view, it is clear that, if there is no $\mathscr{Y}$-function insertion, the $qq$-character is reduced to the (normalized) $\chi_{q_2}$-genus of the quiver variety $\mathfrak{M}_{w,v}$, and thus the $q$-character (the limit $q_2 \to 1$ of $qq$-character) is reduced to the Euler characteristics.
On the other hand, another deformation of the $q$-character, a.k.a., the $t$-analog of $q$-character, is the generating function of the Betti numbers (analogous to the Poincaré polynomial) of the fixed point set of the quiver variety under the $S^1$ action~\cite{Nakajima:2001PC,Nakajima:2004AM}, which is also interpreted as another kind of deformation of Euler characteristics.
\paragraph{Generic quiver}
We can apply this formalism to generic quiver gauge theory.
In this case, we start with the total tangent bundle~\eqref{eq:total_tangent_bundle}, and consider the shift $K_i \to K_i + V_i$ ($i \in \Gamma_0$) with generic $\mathscr{Y}$-function insertions,
\begin{align}
\delta_V\left( \sum_{i \in \Gamma_0} \mathsf{Y}_i^\vee W_i - T\mathfrak{M}_{\underline{n},\underline{k}} \right)
& = \sum_{i,j \in \Gamma_0} \left( - \mathsf{Y}_i^\vee c_{ij}^+ V_j - Q_{12}^\vee V_i^\vee c_{ij}^+ \mathsf{Y}_j + P_{12}^\vee V_i^\vee c_{ij}^+ V_j - P_{12}^\vee V_i^\vee \delta_{ij} W_j \right)
\, ,
\end{align}
where $\operatorname{ch} W_i = \sum_{\alpha \in [w_i]} \mathrm{e}^{\xi_{i,\alpha}}$, such that $\mathbb{I}\left[\sum_{i \in \Gamma_0} \mathsf{Y}_i^\vee W_i\right] = \prod_{i \in \Gamma_0} \prod_{\alpha \in [w_i]} \mathscr{Y}_i(\xi_{i,\alpha})$.
Defining the full $q$-Cartan matrix from the half one~\eqref{eq:half-q-Cartan} as
\begin{align}
c_{ij} = c_{ij}^+ + c_{ij}^- = (1 + Q_{12}^\vee) \delta_{ij} - \sum_{e:i \to j} \mathsf{M}_e - \sum_{e:j \to i} Q_{12}^\vee \mathsf{M}_e^\vee
\, , \qquad
c_{ij}^- = Q_{12}^\vee c_{ji}^{+\vee}
\, ,
\label{eq:full-q-Cartan}
\end{align}
and converting $\widetilde{\mathscr{Y}}$ to $\mathscr{Y}$ as before, we have the contour integral form of the $qq$-character,
\begin{align}
\mathsf{T}_{\underline{w}} = \sum_{\underline{v}} \underline{\mathfrak{q}}^{\underline{v}} \, \mathsf{T}_{\underline{w},\underline{v}}
\, ,
\end{align}
where each contribution is given by
\begin{align}
\mathsf{T}_{\underline{w},\underline{v}} & =
\mathbb{I} \left[ \mathsf{Y}_i^\vee (W_i - c_{ij} V_j) - P_{12}^\vee V_i^\vee W_i + P_{12}^\vee V_i^\vee c_{ij}^+ V_j \right]
\nonumber \\
& =
\frac{1}{\underline{v}!} \frac{[-\epsilon_{12}]^v}{[-\epsilon_{1,2}]^v} \oint \mathrm{d}{\underline{\phi}} \prod_{i \in \Gamma_0} \frac{\mathscr{Y}_i(\underline{\xi}_i)}{\mathscr{A}_i(\underline{\phi}_i)} \mathscr{S}(\underline{\xi}_i - \underline{\phi}_i) \prod_{I \neq J}^{v_i} \mathscr{S}(\phi_{i,IJ})^{-1} \prod_{e : i \to j} \mathscr{S}(\underline{\phi}_j - \underline{\phi}_i + m_e)
\end{align}
and we define the $\mathscr{A}$-function for $\operatorname{ch} \mathsf{x} = \mathrm{e}^x$,
\begin{align}
\mathscr{A}_i (x) = \mathbb{I}\left[ \sum_{j \in \Gamma_0} \mathsf{Y}_j^\vee c_{ji} \mathsf{x} \right] =
\frac{\mathscr{Y}_i(x) \mathscr{Y}_i(x - \epsilon_{12})}{\prod_{e:j \to i} \mathscr{Y}_j(x + m_e) \prod_{e:i \to j} \mathscr{Y}_j(x - m_e - \epsilon_{12})}
\, .
\end{align}
Therefore, the $\mathscr{Y}$-function and the $\mathscr{A}$-function are interpreted as (exponentiated) weight and root vectors, and their operator analogs are discussed in the construction of $q$-deformed W-algebras~\cite{Frenkel:1997,Kimura:2015rgi}.
We can obtain the geometric formulas, \eqref{eq:qq-ch_geom} and \eqref{eq:qq-ch_geom_nom}, for generic quiver by replacing the $q$-Cartan matrix for generic one~\eqref{eq:full-q-Cartan}~\cite{Nekrasov:2015wsu,KPfractional}.
\paragraph{Supergroup case}
The geometric formalism presented above is also applicable to supergroup case.
The representation of quiver $\Gamma$ consists of the vector spaces assigned to each node and the linear maps between them, and we denote the category of representations of quiver $\Gamma$ by $\operatorname{Rep}(\Gamma)$ (see, e.g.,~\cite{Kirillov:2016}).
Then, the \emph{super-representation of quiver} is similarly constructed by replacing the ordinary vector spaces with the supervector spaces~\cite{Thind:2010,Bovdi:2020IJAC}.
In this case, we obtain a contour integral formula for the $qq$-character similarly to the supergroup LMNS formula~\eqref{eq:LMNS_formula_supergroup}.
\subsection{Free field realization}\label{sec:quiver_W-algebra}
Similarly to the matrix model as discussed in Sec.~\ref{sec:free_field_realization}, the gauge theory partition function has a similar free field realization.
Let us discuss an operator formalism of gauge theory in this part.
\subsubsection{Holomorphic deformation}
In the context of four-dimensional $\mathcal{N}=2$ gauge theory, the holomorphic function, called the \emph{prepotential}, plays a central role to characterize the supersymmetric Lagrangian.
The ordinary SYM theory corresponds to the quadratic prepotential $\mathscr{F} = \tr \Phi^2$ at UV.
We now consider generic holomorphic deformation of the prepotential, $\mathscr{F} \to \mathscr{F} + \sum_{n=1}^\infty t_n \tr \Phi^n$.
Even after the deformation, we can still apply the localization computation, and the partition function is given as follows~\cite{Nakajima:2003uh,Marshakov:2006ii},
\begin{align}
Z(t) = \sum_{\lambda \in \mathfrak{M}^\mathsf{T}} Z_\lambda(t)
\, , \qquad
Z_\lambda(t) = Z_\lambda Z^\text{pot}_\lambda(t)
\end{align}
where we define the potential term
\begin{align}
Z^\text{pot}_\lambda(t) = \exp \left( \sum_{n=1}^\infty t_n \mathcal{O}_{n,\lambda} \right)
\, .
\end{align}
We denote the $\lambda$-fixed point contribution of the chiral ring operator by $\mathcal{O}_{n,\lambda} = \tr \Phi^n|_\lambda$ for 4d, and 5d and 6d analogues are obtained by the Wilson loop and the Wilson surface extending on the circle and the torus.
This $t$-deformed partition function is a generating function of the chiral ring operators,
\begin{align}
\left< \mathcal{O}_n \right> = \pdv{}{t_n} \log Z(t)\Big|_{t \to 0}
\, .
\end{align}
From this point of view, the $t$-dependent part plays a similar role to the potential function in the matrix model as discussed in Sec.~\ref{sec:free_field_realization}, and we see the Heisenberg algebra,
\begin{align}
[\partial_n, t_m] = \delta_{n,m}
\end{align}
on the Fock space, $\mathsf{F} = \mathbb{C} \llbracket \partial_n, t_n \rrbracket \ket{0}$, such that the vacuum state is defined as $\partial_n \ket{0} = 0$ ($t$-constant).
The dual vacuum is then defined by $\bra{0} t_n = 0$.
\subsubsection{$Z$-state}
In the operator formalism, the deformation parameter behaves as an operator acting on the Fock space.
From this point of view, the $t$-extended partition function is also seen as an operator, and through the operator-state correspondence, we define the $Z$-state as follows,
\begin{align}
\ket{Z} = Z(t) \ket{0}
\, .
\end{align}
We also define $\ket{Z_\lambda} = Z_\lambda(t) \ket{0}$, such that the total $Z$-state is given by $\ket{Z} = \sum_{\lambda \in \mathfrak{M}^\mathsf{T}} \ket{Z_\lambda}$.
From this $Z$-state, the undeformed partition function can be obtained as follows,
\begin{align}
Z = \braket{0}{Z} = \bra{0} Z(t) \ket{0}
\, .
\label{eq:Z-correlator}
\end{align}
Namely, the partition function is given as a correlator (of vertex operators as we will see soon) in this formalism.
Such a connection between the BPS observables and the formalism of CFT-like theory is called the BPS/CFT correspondence~\cite{Nekrasov:2004UA}, and the realization of the partition function as a correlator is interpreted as a consequence of such a correspondence.
We remark that the expression~\eqref{eq:Z-correlator} implies that the partition function is realized as a correlator on a sphere.
One can similarly consider a torus correlator (a character on the Fock module), and it turns out that such a torus correlator computes the six-dimensional partition function in this context~\cite{Kimura:2016dys}.
In the formalism of class $\mathcal{S}$ theory~\cite{Gaiotto:2009we}, on the other hand, the torus correlator is associated with the loop structure in quiver gauge theory, typically found in affine quiver gauge theory.
These two realizations are related to each other through the duality exchanging base and fiber of Seiberg--Witten geometry, and S-duality in type IIB string theory~\cite{Katz:1997eq,Aharony:1997bh}.
\subsubsection{Vertex operators}
Let us consider the vertex operator realization of the $Z$-state.
We focus on the five-dimensional convention.
See \cite{Nieri:2019mdl} and \cite{Kimura:2016dys} for four and six dimensional cases.
For this purpose, we define the screening currents,
\begin{align}
S_\sigma(x) = {: \exp \left( s_0^\sigma \log x + \tilde{s}_0^\sigma + \sum_{n \in \mathbb{Z}_{\neq 0}} s_n^\sigma x^{-n} \right) :}
\, , \qquad
\sigma = 0, 1
\, ,
\end{align}
with the commutation relations,
\begin{subequations}
\begin{align}
[s_n^0, s_m^0] = - \frac{1}{n} \frac{1 - q_1^n}{1 - q_2^{-n}} (1 + q_{12}^{-n}) \delta_{n+m,0}
\, , \qquad &
[s_n^1, s_m^1] = - \frac{1}{n} \frac{1 - q_2^n}{1 - q_1^{-n}} (1 + q_{12}^{-n}) \delta_{n+m,0}
\, , \\
[s_n^0, s_m^1] = \frac{1}{n} (q_1^n + q_2^{-n}) \delta_{n+m,0}
\, , \qquad &
[s_n^1, s_m^0] = \frac{1}{n} (q_1^{-n} + q_2^{n}) \delta_{n+m,0}
\, , \\
[\tilde{s}_0^0 , s_n^0] = - 2 b^2 \delta_{n,0}
\, , \qquad
[\tilde{s}_0^1 , s_n^1] = - 2 b^{-2} \delta_{n,0}
\, , \qquad &
[\tilde{s}_0^0 , s_n^1] = [\tilde{s}_0^1 , s_n^0] = 2 \delta_{n,0}
\, ,
\end{align}
\end{subequations}
with $b^2 = - \epsilon_1/\epsilon_2$.
This parametrization is taken to be compatible with the matrix model notation~\eqref{eq:Omega_background_matrix_parameter}.
For $|q_1| < 1$, $|q_2| > 1$, we have the following OPEs,
\begin{subequations}
\begin{align}
\frac{S_0(x) S_0(x')}{:S_0(x) S_0(x'):} & =
\frac{(x'/x;q_2^{-1})_\infty (q_{12}^{-1} x'/x;q_2^{-1})_\infty}{(q_1 x'/x;q_2^{-1})_\infty (q_{2}^{-1} x'/x;q_2^{-1})_\infty}
x^{\prime -2 b^2}
\, , \\
\frac{S_1(x) S_1(x')}{:S_1(x) S_1(x'):} & =
\frac{(x'/x;q_1)_\infty (q_{12} x'/x;q_1)_\infty}{(q_1 x'/x;q_1)_\infty (q_{2}^{-1} x'/x;q_1)_\infty}
x^{\prime -2 b^{-2}}
\, , \\
\frac{S_0(x) S_1(x')}{:S_0(x) S_1(x'):} & =
\frac{(-q_2 xx')}{(1 - q_1 x'/x)(1 - q_2 x/x')}
\, .
\end{align}
\end{subequations}
Based on the screening current, we have the following realization for the $\lambda$-fixed point contribution for U($n$) gauge theory~\cite{Kimura:2015rgi},%
\footnote{%
In this case, we need infinitely many screening currents to construct five-dimensional gauge theory partition function.
A similar construction is known for three-dimensional gauge theory, which involves a finite number of the screening currents~\cite{Aganagic:2013tta,Aganagic:2014oia,Aganagic:2015cta,Nedelin:2016gwu}.
These theories are related through the Higgsing process.
See Sec.~\ref{sec:intersecting_defects}.
}
\begin{align}
Z_\lambda(t) = \prod^\succ_{x \in \mathcal{X}_\lambda} S_0(x)
\end{align}
The product $\prod^\succ$ is the radial ordering product with the ordering in the set $\mathcal{X}_\lambda$, $\prod^\succ S_0(x) = S_0(x_1) S_0(x_2) \cdots$ for $|x_1| > |x_2| > \cdots$ with $|q_1| \ll |q_2^{-1}| < 1$ and $\mathrm{e}^{a_\alpha} \sim 1$.
In order to discuss the total partition function obtained by summation over the fixed point contributions, we define the screening charge from the screening current,
\begin{align}
Q_0(x) = \sum_{k \in \mathbb{Z}} S_0(x q_2^k)
\, , \qquad
Q_1(x) = \sum_{k \in \mathbb{Z}} S_1(x q_1^k)
\, .
\end{align}
Then, the $Z$-state is constructed from the screening charge,%
\footnote{%
We also have the free field realization of the contour integral formula based on similar vertex operators.
See \cite{Kimura:2019hnw,Kimura:2022zsm} for details.
}
\begin{align}
\ket{Z} = Z(t) \ket{0}
\, , \qquad
Z(t) = \prod^\succ_{x \in \mathcal{X}_\emptyset} Q_0(x)
\, .
\end{align}
We remark that in the expansion of the screening charge product, there are contributions of pseudo-fixed points, $\lambda \in \mathfrak{M}^\mathbb{Z}\backslash \mathfrak{M}^\mathsf{T}$, where $\mathfrak{M}^\mathbb{Z} = \{ \mathbb{Z} \ni \lambda_{\alpha,i} \xrightarrow{i \gg 1} 0, \alpha \in [n], i \in \mathbb{N} \}$.
However, we have $Z_\lambda(t) = 0$ for $\lambda \in \mathfrak{M}^\mathbb{Z}\backslash \mathfrak{M}^\mathsf{T}$, and hence it does not contribute to the partition function.
We have a similar realization of supergroup gauge theory partition function.
In this case, we may use both the screening currents to obtain the $Z$-state~\cite{Kimura:2021ngu,Noshita:2022dxv}.
\begin{itembox}{$Z$-state for supergroup gauge theory}
The $Z$-state for supergroup gauge theory is constructed by two types of screening charges,
\begin{align}
\ket{Z} = \prod^\succ_{x \in \mathcal{X}_\emptyset^0} Q_0(x) \prod^\succ_{x \in \check{\mathcal{X}}_\emptyset^1} Q_1(x) \ket{0}
\, .
\end{align}
\end{itembox}
In fact, this expression is analogous to the operator formalism of the supermatrix model discussed in Sec.~\ref{sec:free_field_realization}.
\subsubsection{$q$-Virasoro constraint}
Similarly to the Virasoro algebra discussed in Sec.~\ref{sec:free_field_realization}, one can construct an operator that commutes with the screening charges in this case as well.
We define the vertex operator, called the $\mathsf{Y}$-operator, which yields the $\mathscr{Y}$-function in gauge theory as follows,
\begin{align}
\mathscr{Y}(x) = \bra{0} \mathsf{Y}(x) \ket{Z}
\, , \qquad
\widetilde{\mathscr{Y}}(x) = \bra{Z} \mathsf{Y}(x) \ket{0}
\, .
\end{align}
Then, we define the $\mathsf{T}$-operator using the $\mathsf{Y}$-operator,
\begin{align}
\mathsf{T}(x) = \mathsf{Y}(x) + \mathsf{Y}^{-1}(x q_{12}^{-1})
\, .
\end{align}
Namely, this is an operator analog of the $qq$-character introduced in Sec.~\ref{sec:SD_eq}.
We see that this $\mathsf{T}$-operator commutes with the screening charges, $[\mathsf{T}(x), Q_\sigma(x')] = 0$.
Having the mode expansion, $\mathsf{T}(x) = \sum_{n \in \mathbb{Z}} T_n x^{-n}$, they obey the following algebraic relation,
\begin{align}
[T_n, T_{m}] = - \sum_{k=1}^\infty f_k ( T_{n-k} T_{m+k} - T_{m-k} T_{n+k} ) - \frac{(1 - q_1^n)(1 - q_2^n)}{1 - q_{12}} (q_{12}^n - q_{12}^{-n}) \delta_{n+m,0}
\, ,
\label{eq:qVirasoro_rel}
\end{align}
where the coefficients $\{ f_k \}_{k \in \mathbb{N}}$ are defined from the structure function,
\begin{align}
f(x) = \exp\left( \sum_{n = 1}^\infty \frac{(1 - q_1)(1 - q_2)}{n (1 + q_{12}^n)} x^n \right) = \sum_{k = 0}^\infty f_k x^k
\, .
\end{align}
This structure function comes from the OPE between the $\mathsf{Y}$-operators, $\mathsf{Y}(x) \mathsf{Y}(x')/{:\mathsf{Y}(x) \mathsf{Y}(x'):} = f^{-1}(x'/x)$.
The algebraic relation~\eqref{eq:qVirasoro_rel} is equivalent to the following relation for the $\mathsf{T}$-operator,
\begin{align}
f\left(\frac{x'}{x}\right) \mathsf{T}(x) \mathsf{T}(x') - f\left(\frac{x}{x'}\right) \mathsf{T}(x') \mathsf{T}(x) = - \frac{(1 - q_1)(1 - q_2)}{1 - q_{12}} \left( \delta\left(\frac{x'}{x} q_{12}\right) - \delta\left(\frac{x}{x'} q_{12}\right) \right)
\, ,
\label{eq:fTT_rel}
\end{align}
where we define the multiplicative $\delta$-function, $\delta(z) = \sum_{n \in \mathbb{Z}} z^n$.
This is called the \emph{quadratic relation} (or the $fTT$ relation) for the generating current of the $q$-deformed Virasoro algebra constructed in~\cite{Shiraishi:1995rp}.
We have remarks on this relation.
Taking the limit $q_1 \to 1$ or $q_2 \to 1$, the quadratic relation is simply reduced to $[\mathsf{T}(x), \mathsf{T}(x')] = 0$.
Hence, the $\mathsf{T}$-operator becomes a commuting operator, which is then identified with the $q$-character, and also with the transfer matrix of the corresponding quantum integrable system.
From this point of view, we have a connection with quantum integrable system in this limit.
See Sec.~\ref{sec:Bethe/gauge} for a related discussion.
Another remark is that, interpreting the $\mathsf{T}$-operator as the $qq$-character, the quadratic relation corresponds to the anti-symmetric tensor product of the fundamental representation of $A_1$ algebra, namely $\mathfrak{sl}(2)$.
Therefore, the right hand side of~\eqref{eq:fTT_rel} is interpreted as a contribution of the trivial representation.
We can see more structures in generic tensor product.
See \cite{Kimura:2022spi} for more details.
From this point of view, the gauge theory average of $qq$-character is identified with the $\mathsf{T}$-operator correlator, $\mathscr{T}(x) = \bra{0} \mathsf{T}(x) \ket{Z}$.
In order to obtain a degree-$n$ polynomial average, we need a modification of the vacuum state.
We introduce the degree-$n$ vacuum, such that $\mathsf{T}_m \ket{n} = 0$ for $m>n$, and define the modified $Z$-state, $\ket{Z} = Z(t) \ket{n}$.
Recalling that the $\mathsf{T}$-operator commutes with the screening charges (hence with $Z(t)$), the $qq$-character average, $\mathscr{T}(x) = \bra{0} \mathsf{T}(x) \ket{Z}$, becomes a degree-$n$ polynomial in $x$.
This is an analogue of the Virasoro constraint that we call the $q$-Virasoro constraint.
See also \cite{Kanno:2012hk,Kanno:2013aha,Nedelin:2015mio,Nedelin:2016gwu,Lodin:2018lbz} for related discussions.
From this point of view, we obtain the same $q$-Virasoro constraint for supergroup gauge theory since the $\mathsf{T}$-operator commutes with the both screening charges, $Q_\sigma(x)$.
Considering generic quiver gauge theory, we will obtain the constraint with quiver W-algebra.
See~\cite{Kimura:2015rgi} for details.
\subsection{Bethe/gauge correspondence}\label{sec:Bethe/gauge}
Although we have focused on gauge theory on $\mathbb{C}^2$, we may discuss its implication to further low dimensional theory.
Taking the limit $\epsilon_2 \to 0$, while keeping $\epsilon_1$ finite, a.k.a., \emph{Nekrasov--Shatashvili (NS) limit}, the partition function asymptotically behaves as $Z \approx \exp \left( \frac{1}{\epsilon_2} \widetilde{\mathscr{W}} + \cdots \right)$, where $\widetilde{\mathscr{W}}$ is identified with the twisted superpotential of the corresponding two-dimensional $\mathcal{N}=(2,2)$ theory.
Then, the saddle point equation in this limit yields the twisted $F$-term condition,
\begin{align}
\exp \left( \pdv{\widetilde{\mathscr{W}}}{\sigma} \right) = 1
\, ,
\end{align}
where we denote the scalar field in the twisted chiral multiplet by $\sigma$.
For example, for four-dimensional U($n$) gauge theory with $2n$ flavors ($n$ fundamental and $n$ anti-fundamental hypermultiplets), the saddle point equation is given by
\begin{align}
- \mathfrak{q} \frac{a(x) {d}(x + \epsilon_{12})}{\mathscr{Y}(x) {\mathscr{Y}}(x + \epsilon_{12})} = 1
\, ,
\end{align}
where $a(x)$ and ${d}(x)$ are the matter polynomials, $a(x) = \prod_{f \in [n]} [x - m_f]$ and ${d}(x) = \prod_{f \in [n]} [x - \widetilde{m}_f]$.
Recalling the definition of the $\mathscr{Y}$-function~\eqref{eq:Y-func_def}, we write $\mathscr{Y}(x) = \mathscr{Q}(x) / \mathscr{Q}(x - \epsilon_1)$ with redefinition, $\mathscr{Y}(x) \to d(x) \mathscr{Y}(x)$.
Then, the saddle point equation is written in the following form in the limit $\epsilon_2 \to 0$,%
\footnote{%
In this case, the $\mathscr{Q}$-function involves infinitely many roots.
We need to impose the Higgsing condition to obtain a finite polynomial $\mathscr{Q}$-function.
See, e.g.,~\cite{Dorey:2011pa,Chen:2011sj} and also Sec.~\ref{sec:intersecting_defects} for a related discussion.
}
\begin{align}
\frac{a(x)}{d(x)} = - \mathfrak{q} \frac{\mathscr{Q}(x + \epsilon_1)}{\mathscr{Q}(x - \epsilon_1)}
\, ,
\end{align}
which can be identified with the Bethe equation of $\mathfrak{sl}(2)$-spin chain (choice of the index convention~\eqref{eq:index_notation} corresponds to XXX/XXZ/XYZ chain) of length $L = n$ and the roots of the $\mathscr{Q}$-function identified with the Bethe roots.
Denoting the mass parameter by $m_\alpha = \nu_\alpha + \mathrm{i} s_\alpha$ and put $\widetilde{m}_\alpha = \bar{m}_\alpha$, the parameters $(\nu_\alpha,s_\alpha)$ are identified with the inhomogeneity and the spin of the site $\alpha \in [L]$.
In this context, the gauge coupling $\tau$ with $\mathfrak{q} = \exp(2 \pi \mathrm{i} \tau)$ is identified with the twist boundary condition parameter.
Applying this formalism to supergroup gauge theory, we obtain the Bethe equation of the form,
\begin{align}
\frac{a_0(x)}{a_1(x)} \frac{d_1(x)}{d_0(x)} = - \mathfrak{q} \frac{\mathscr{Q}_0(x + \epsilon_1)}{\mathscr{Q}_0(x - \epsilon_1)} \frac{\mathscr{Q}_1(x + \epsilon_1)}{\mathscr{Q}_1(x - \epsilon_1)}
\, ,
\end{align}
which implies positive and negative magnons carrying positive and negative excitations and the sites involving positive and negative spins~\cite{Kimura:2019msw,Chen:2020rxu}.
We remark that the situation that we now consider is different from the spin chain with superalgebra symmetry.
In order to realize such a superspin chain, we need to consider quiver gauge theory that corresponds to the Dynkin diagram of the Lie superalgebra~\cite{Orlando:2010uu,Nekrasov:2018gne,Zenkevich:2018fzl,Ishtiaque:2021jan}.
The spin chain model is not a unique example to be discussed in this framework.
For example, it has been known that the Seiberg--Witten curve of $G$-SYM theory is given by the spectral curve of $\widehat{^LG}$-Toda chain~\cite{Martinec:1995by}.
In fact, one can obtain quantum Toda chain from pure SYM theory by imposing the codimension-two surface defect.
In the presence of such a defect, we need a modification of the instanton moduli space to the so-called (affine) Laumon space, which is equivalent to consider instantons on the partial orbifold space, $\mathbb{C} \times \mathbb{C}/\mathbb{Z}_n$~\cite{Feigin:2011SM,Finkelberg:2010QDZ,Kanno:2011fw}.
In this context, the $\Omega$-background parameter $\epsilon_1$ plays a role of the quantum parameter, and when $\epsilon_2$ is also finite, we obtain the non-stationary quantum integrable system, which is reduced to the stationary system in the NS limit, $\epsilon_2 \to 0$.
In this case, one can construct the quantum Toda Hamiltonian from the $qq$-character ($q$-character in the NS limit) in the presence of the surface defect.
Applying this formalism to supergroup gauge theory, we then obtain a super-Toda chain, which is associated with the root system of the corresponding Lie superalgebra~\cite{vanderLende:1994JMP} (In general, one can construct the integrable Toda chain from the root system data).
Further incorporating the hypermultiplet in the adjoint representation, namely $\mathcal{N} = 2^*$ theory, the Toda chain is promoted to the Calogero--Moser--Sutherland (CMS) system, and we obtain the ``double'' CMS system associated with the Lie superalgebra~\cite{Kerov:1998IMRN,Sergeev:2001JNMP,Sergeev:2002TMP} from supergroup gauge theory with the surface defect.
See~\cite{Chen:2020rxu} for details.
\subsection{Higgsing and intersecting defects}\label{sec:intersecting_defects}
As we discussed throughout this article, supergroup gauge theory inevitably violates the unitarity, and thus its physical realization seems not to be straightforward.
In this part, we would demonstrate that supergroup theory can be engineered from physical setups by imposing the defect operators.
In the context of supersymmetric gauge theory, there exist two branches in the moduli space of supersymmetric vacua, called the Coulomb and Higgs branches (In general, one can also consider the mixed branch).
Although Seiberg--Witten theory of four-dimensional $\mathcal{N}=2$ gauge theory describes the Coulomb branch, it has been known that there exists a root of Higgs branch locus, which is reached by tuning the fundamental mass parameter to the Coulomb moduli parameter as $m_f = a_\alpha$~\cite{Dorey:1998yh,Dorey:1999zk}.
In the presence of the $\Omega$-background, this condition is ``quantized'' to be $m_f = a_\alpha + n_{\alpha,1} \epsilon_1 + n_{\alpha,2} \epsilon_2$, and under this condition, the site $(n_{\alpha,1} + 1, n_{\alpha,2} + 1)$ cannot be included by the $\alpha$-th partition, $(n_{\alpha,1} + 1, n_{\alpha,2} + 1) \not\in \lambda_\alpha$ (the pit condition~\cite{Bershtein:2018SM}).%
\footnote{%
This can be seen as follows:
The fundamental hypermultiplet contribution is given by \eqref{eq:fund_hyp_contribution}, whose instanton part reads $\mathsf{H}_\text{inst} = \mathsf{M}^\vee \mathsf{K}$.
At the fixed point $\lambda \in \mathfrak{M}^\mathsf{T}$, we have the character $\operatorname{ch} \mathsf{H}_\text{inst} = \sum_{\alpha \in [n]} \sum_{(i,j) \in \lambda_{\alpha}} q_1^{i-1-n_{\alpha,1}} q_2^{j-1-n_{\alpha,2}}$.
Hence, if $(n_{\alpha,1} + 1, n_{\alpha,2} + 1) \in \lambda_\alpha$, we have a zero mode, and the corresponding contribution to the partition function becomes zero.
}
Namely, we have a restriction on the instanton configuration.
In fact, such a restriction on a partition is also discussed in the context of representation theory of supergroup:
Partitions with the pit condition parameterize irreducible representations of U$(n_{\alpha,1}|n_{\alpha,2})$.
In particular, for $n_{\alpha,2} = 0$, partitions have at most $n_{\alpha,1}$ rows, which yield irreducible representations of U$(n_{\alpha,1})$ group.
From physical point of view, $n_{\alpha,1,2}$ are interpreted as the numbers of fluxes or vortex defects in $\mathbb{C}_1$ and $\mathbb{C}_2$ direction, respectively.
Hence, the situation with both $n_{\alpha,1,2}$ realizes intersecting defects crossing at the origin~\cite{Dimofte:2010tz,Bonelli:2011fq,Gomis:2016ljm,Pan:2016fbl,Gorsky:2017hro,Nieri:2017ntx}.
See also~\cite{Jeong:2021rll}.
Computing the partition function of this configuration, it is given in the form of (deformation of) supermatrix model, which implies that we have an emerging supergroup structure originated from physical (non-supergroup) theory by considering the intersecting defects.
A similar realization of supergroup theory can be discussed in a higher-dimensional setup.
Starting with an eight-dimensional setup on $\mathbb{C}^4$, called the gauge origami, we incorporate D-branes extended in $\mathbb{C}_1 \times \mathbb{C}_2$ and $\mathbb{C}_2 \times \mathbb{C}_3$~\cite{Nekrasov:2016ydq,Nekrasov:2017rqy}.
In this case, we obtain the non-stationary double CMS system at finite $\epsilon_2$, and the stationary one in the limit $\epsilon_2 \to 0$~\cite{Nekrasov:2017gzb,Chen:2019vvt}.
This implies that supergroup structure emerges from intersecting defects at the $\mathbb{C}_2$-plane.
\subsection*{Acknowledgement}
I would like to thank
Nicolas Babinet,
Heng-Yu Chen,
Norton Lee,
Fabrizio Nieri,
Go Noshita,
Vasily Pestun,
Yuji Sugimoto
for collaborations and fruitful discussions on this subject.
I am in particular grateful to Go Noshita for careful reading of the manuscript and valuable comments.
This work was in part supported by ``Investissements d'Avenir'' program, Project ISITE-BFC (No.~ANR-15-IDEX-0003), EIPHI Graduate School (No. ANR-17-EURE-0002), and Bourgogne-Franche-Comté region.
\section{Mathematical backgrounds}
\bibliographystyle{amsalpha_mod}
|
\chapter{Possible Universes Coupled to the Hypermultiplets of $\mathcal{N} = 2$ Supergravity}
\label{RDE}
We continue our study started in the previous chapter. To study the current era of the universe, we have to add a cosmological
constant term ($\Lambda$) that represents the dark energy responsible for the accelerating expansion of the universe. So here we study a brane filled with all radiation, dust and energy that represents our present 4- dimensional universe.
When studying the time evolution of such brane, we have found that the moduli $G_{i\bar{j}} \dot z^{i} \dot z^{\bar{j}}$
with $\Lambda$ (the 3-brane cosmological constant) can induce a 3-brane cosmic evolution same as our world, and even in many
cases (initial conditions) when removing $\Lambda$ and let only the bulk cosmological constant ($\tilde{\Lambda}$)
the late time acceleration of the brane- universe is kept.
Means that the whole cosmological history of our own universe,
whether- early time inflationary epoch or late time accelerated expansion-
may be interpreted only by the effects of a bulk (extra dimension). Moreover studying different initial conditions (ICs) enables to
find out other possible universes that may be exist with our known universe. Finally our model expects
possible scenarios of the future behaviour of the brane- world depending on the ICs, $\Lambda$ and $\tilde{\Lambda}$.
\section{Brane embedding and the fields equations}
Consider adding the brane cosmological constant term $\Lambda(t)$ in addition to the bulk cosmological constant term
$\tilde{\Lambda}(t)$ to the Einstein equtions $G_{\mu\nu} + \Lambda g_{\mu\nu} = T^{Brane}_{\mu\nu}+T^{Bulk}_{yy} $,
and $G_{yy} + \tilde{\Lambda} g_{yy} = T^{Bulk}_{yy} $. The matter desnsity and pressure of
a brane filled with radition, dust and pure energy are:
\begin{equation}
\rho(t)= \frac{1}{a^3}+ \frac{1}{a^4}, ~~~~~~ p(t)= \frac{1}{3a^4}.
\label{all}
\end{equation}
Then the Friedmann-like equations become:
\begin{eqnarray}
3\left[ {\left( {\frac{{\dot a}}{a}} \right)^2 + \left( {\frac{{\dot a}}{a}} \right)\left( {\frac{{\dot b}}{b}} \right)} \right] &=& G_{i\bar j} \dot z^i \dot z^{\bar j} + \frac{{1 }}{{a^3 }} + \frac{1}{{a^4 }} + \Lambda\left(t\right) \nonumber\\
2\frac{{\ddot a}}{a} + \left( {\frac{{\dot a}}{a}} \right)^2 + \frac{{\ddot b}}{b} + 2\left( {\frac{{\dot a}}{a}} \right)\left( {\frac{{\dot b}}{b}} \right) &=& \Lambda\left(t\right) - \frac{1}{3a^4} - G_{i\bar j} \dot z^i \dot z^{\bar j} \nonumber\\
3\left[ {\frac{{\ddot a}}{a} + \left( {\frac{{\dot a}}{a}} \right)^2 } \right] &=& \tilde\Lambda \left(t\right) - G_{i\bar j} \dot z^i \dot z^{\bar j} ,
\label{EE3}
\end{eqnarray}
Analysis of the fields equations are performed numerically following the initial conditions schemes outlined in table (\ref{tab2})
and (\ref{tabb}) . This is broken down into two major categories: branes with vanishing initial conditions on the scale factors, and those with non-vanishing values but initially decreasing in velocity.
The studies are performed in two cases: first for constant $\Lambda$ and $\tilde \Lambda$ while
solving for the brane scale factor $a$, the bulk scale factor $b$ and the moduli norm $G_{i\bar j} \dot z^i \dot z^{\bar j}$,
second for constant $b$ and $\tilde \Lambda$ while solving for the brane scale factor $a$, $\Lambda$
and $G_{i\bar j} \dot z^i \dot z^{\bar j}$.
For all cases for IC set \#1 I demonstrate all the results for $a$, $b$, $G_{i\bar j} \dot z^i \dot z^{\bar j}$
and the hypermultiplets fields, while for the last of ICs, we demonstrate the different universes (ICs) only according to
$a$, $b$, the moduli and $k$.
Anyways, the behavior of the scalar fields are similar, albeit with different rates of change.
\begin{table}[H]
\centering
\caption{The eighteen sets of initial conditions (IC) used in the constant $\Lambda$ computations.}
\label{tab2}
\vspace{0.5cm}
\begin{adjustbox}{max width=\textwidth}
\begin{tabular}{|c|c|c|c|c|c|c|l|}
\hline
\textbf{IC Set \#} & \multicolumn{1}{l|}{$a\left(0\right)$} & $b\left(0\right)$ & $\dot a\left(0\right)$ & $\dot b\left(0\right)$ & $\Lambda$ & $\tilde \Lambda$ & \textbf{Description} \\ \hline
\textbf{1} & \multirow{9}{*}{0} & \multirow{9}{*}{0} & \multirow{9}{*}{0} & \multirow{9}{*}{0} & 0 & 0 & \multirow{9}{*}{Big bang-like IC with vanishing initial velocities} \\ \cline{1-1} \cline{6-7}
\textbf{2} & & & & & 1 & 0 & \\ \cline{1-1} \cline{6-7}
\textbf{3} & & & & & 0 & 1 & \\ \cline{1-1} \cline{6-7}
\textbf{4} & & & & & 1 & 1 & \\ \cline{1-1} \cline{6-7}
\textbf{5} & & & & & -1 & 0 & \\ \cline{1-1} \cline{6-7}
\textbf{6} & & & & & 0 & -1 & \\ \cline{1-1} \cline{6-7}
\textbf{7} & & & & & -1 & -1 & \\ \cline{1-1} \cline{6-7}
\textbf{8} & & & & & 1 & -1 & \\ \cline{1-1} \cline{6-7}
\textbf{9} & & & & & -1 & 1 & \\ \hline\hline
\textbf{10} & \multirow{9}{*}{1} & \multirow{9}{*}{-0.2} & \multirow{9}{*}{1} & \multirow{9}{*}{-0.2} & 0 & 0 &
\multirow{9}{*}{Non-singular IC with initial negative velocities} \\ \cline{1-1} \cline{6-7}
\textbf{11} & & & & & 1 & 0 & \\ \cline{1-1} \cline{6-7}
\textbf{12} & & & & & 0 & 1 & \\ \cline{1-1} \cline{6-7}
\textbf{13} & & & & & 1 & 1 & \\ \cline{1-1} \cline{6-7}
\textbf{14} & & & & & -1 & 0 & \\ \cline{1-1} \cline{6-7}
\textbf{15} & & & & & 0 & -1 & \\ \cline{1-1} \cline{6-7}
\textbf{16} & & & & & -1 & -1 & \\ \cline{1-1} \cline{6-7}
\textbf{17} & & & & & 1 & -1 & \\ \cline{1-1} \cline{6-7}
\textbf{18} & & & & & -1 & 1 & \\ \hline
\end{tabular}
\end{adjustbox}
\end{table}
\subsection{Constant $\Lambda$}
The cosmological constants $\Lambda$ and $\tilde \Lambda$ are both taken to be either vanishing, positive, or negative. The calculation is performed over all possible permutations of these values.
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC1/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve. The curve for $b$ is scaled up by a factor of 10.}
\label{abzz00001constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC1/a_dot_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve. The curve for $\dot b$ is scaled up by a factor of 10.}
\label{adotbdot00001constantL}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC1/a_ddot_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve. The curve for $\ddot b$ is scaled up by a factor of 10.}
\label{addotbddot00001constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC1/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), and $\dot k\left(0\right)=-1$ (red curve). While $\dot k\left(0\right)=0$ diverges.}
\label{k00001constantL}
\end{subfigure}
\caption{Initial conditions set number 1 for constant $\Lambda$.}
\label{Fig1}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC1/sigma}
\caption{The dilaton $\sigma$; same for all three $\dot k\left(0\right)$.}
\label{sigma00001constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC1/sigma_dot}
\caption{The dilatonic field strength $\dot\sigma$.}
\label{sigmadot00001constantL}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC1/phi}
\caption{The universal axion $\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve). The solution diverges for $\dot k\left(0\right)=0$.}
\label{phi00001constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC1/phi_dot}
\caption{The axionic field strength $\dot\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve).}
\label{phidot00001constantL}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC1/xx}
\caption{$ \langle \Xi | \dot{\Xi} \rangle$ for $\dot{k}(0)= 1$ (blue), and $\dot{k}(0) = -1$ (red).}
\label{xx00001constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC1/Omega}
\caption{$\Omega$ at $\dot k\left(0\right) = 1$ and $ \dot\sigma \left(0\right) =0 $.}
\label{omega00001constantL}
\end{subfigure}
\vspace{0.3cm}
\caption{Initial conditions set number 1 for constant $\Lambda$ (continued) .}
\label{Fig2}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC2/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{abzz00002constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC2/a_dot_b_dot_EARLY}
\caption{The early time expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve. The curve for $\dot b$ is scaled up by a factor of 10.}
\label{adotbdotEARLY00002constantL}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC2/a_dot_b_dot_LATE}
\caption{The late time expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve.}
\label{adotbdotLATE00002constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC2/a_ddot_b_ddot_EARLY}
\caption{The early time accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve. The curve for $\ddot b$ is scaled up by a factor of 10.}
\label{addotbddotEARLY00002constantL}
\end{subfigure}
\caption{Initial conditions set number 2 for constant $\Lambda$.}
\label{Fig3}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC2/a_ddot_b_ddot_LATE}
\caption{The late time accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve.}
\label{addotbddotLATE00002constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC2/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), and $\dot k\left(0\right)=-1$ (red curve). While $\dot k\left(0\right)=0$ diverges.}
\label{k00002constantL}
\end{subfigure}
\caption{Initial conditions set number 2 for constant $\Lambda$ (continued).}
\label{Fig33}
\end{figure}
\vspace{3cm}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC3/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve. The curve for $b$ is scaled up by a factor of 10.}
\label{abzz00003constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC3/a_dot_b_dot_EARLY}
\caption{The early time expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve. The curve for $\dot b$ is scaled up by a factor of 10.}
\label{adotbdotEARLY00003constantL}
\end{subfigure}
\caption{Initial conditions set number 3 for constant $\Lambda$.}
\label{Fig5dash}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC3/a_dot_b_dot_LATE}
\caption{The late time expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve. The curve for $\dot b$ is scaled up by a factor of 50.}
\label{adotbdotLATE00003constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC3/a_ddot_b_ddot_EARLY}
\caption{The early time accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve. The curve for $\dot b$ is scaled up by a factor of 10.}
\label{addotbddotEARLY00003constantL}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC3/a_ddot_b_ddot_LATE}
\caption{The late time accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve. The curve for $\ddot b$ is scaled up by a factor of 50.}
\label{addotbddotLATE00003constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC3/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), and $\dot k\left(0\right)=-1$ (red curve). While $\dot k\left(0\right)=0$ diverges.}
\label{k00003constantL}
\end{subfigure}
\caption{Initial conditions set number 3 for constant $\Lambda$ (continued).}
\label{Fig5dash}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC4/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{abzz00004constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC4/a_dot_b_dot_EARLY}
\caption{The early time expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve. The curve for $\dot b$ is scaled up by a factor of 10.}
\label{adotbdotEARLY00004constantL}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC4/a_dot_b_dot_LATE}
\caption{The late time expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve.}
\label{adotbdotLATE00004constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC4/a_ddot_b_ddot_EARLY}
\caption{The early time accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve. The curve for $\ddot b$ is scaled up by a factor of 10.}
\label{addotbddotEARLY00004constantL}
\end{subfigure}
\caption{Initial conditions set number 4 for constant $\Lambda$.}
\label{Fig7}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC4/a_ddot_b_ddot_LATE}
\caption{The late time accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve.}
\label{addotbddotLATE00004constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC4/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k00004constantL}
\end{subfigure}
\caption{Initial conditions set number 4 for constant $\Lambda$ (continued).}
\label{Fig77}
\end{figure}
\vspace{3cm}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC5/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve. The curve for $b$ is scaled up by a factor of 10.}
\label{abzz00005constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC5/a_dot_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve. The curve for $\dot b$ is scaled up by a factor of 10.}
\label{adotbdotEARLY00005constantL}
\end{subfigure}
\caption{Initial conditions set number 5 for constant $\Lambda$ .}
\label{Fig9}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC5/a_ddot_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve.}
\label{addotbddotEARLY00005constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC5/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), and $\dot k\left(0\right)=-1$ (red curve). While $\dot k\left(0\right)=0$ diverges.}
\label{k00005constantL}
\end{subfigure}
\caption{Initial conditions set number 5 for constant $\Lambda$ (continued).}
\label{Fig10}
\end{figure}
\vspace{3cm}
\vspace{-0.7cm}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC6/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{abzz00006constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC6/a_dot_b_dot_EARLY}
\caption{The early time expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve. The curve for $\dot b$ is scaled up by a factor of 10.}
\label{adotbdotEARLY00006constantL}
\end{subfigure}
\caption{Initial conditions set number 6 for constant $\Lambda$.}
\label{Fig112}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC6/a_dot_b_dot_LATE}
\caption{The late time expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve. The curve for $\dot b$ is scaled up by a factor of 10.}
\label{adotbdotLATE00006constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC6/a_ddot_b_ddot_EARLY}
\caption{The early time accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve. The curve for $\ddot b$ is scaled up by a factor of 10.}
\label{addotbddotEARLY00006constantL}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC6/a_ddot_b_ddot_LATE}
\caption{The late time accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve. The curve for $\ddot a$ is scaled up by a factor of 10.}
\label{addotbddotEARLY00006constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC6/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), and $\dot k\left(0\right)=-1$ (red curve). While $\dot k\left(0\right)=0$ diverges.}
\label{k00006constantL}
\end{subfigure}
\caption{Initial conditions set number 6 for constant $\Lambda$ (continued).}
\label{Fig12}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC7/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while ${G_{i\bar j} \dot z^i \dot z^{\bar j}}$ (not the absolute value) by the green curve. The curve for $b$ is scaled up by a factor of 10.}
\label{abzz00007constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC7/a_dot_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve. The curve for $\dot b$ is scaled up by a factor of 10.}
\label{adotbdot00007constantL}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC7/a_ddot_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve. The curve for $\ddot b$ is scaled up by a factor of 10.}
\label{addotbddot00007constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC7/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k00007constantL}
\end{subfigure}
\caption{Initial conditions set number 7 for constant $\Lambda$.}
\label{Fig13}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC8/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{abzz00008constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC8/a_dot_b_dot_EARLY}
\caption{The early time expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve. The curve for $\dot b$ is scaled up by a factor of 10.}
\label{adotbdotEARLY00008constantL}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC8/a_dot_b_dot_LATE}
\caption{The late time expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve.}
\label{adotbdotLATE00008constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC8/a_ddot_b_ddot_EARLY}
\caption{The early time accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve.}
\label{addotbddotEARLY00008constantL}
\end{subfigure}
\caption{Initial conditions set number 8 for constant $\Lambda$.}
\label{Fig161}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC8/a_ddot_b_ddot_LATE}
\caption{The late time accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve.}
\label{addotbddotLATE00008constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC8/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k00008constantL}
\end{subfigure}
\caption{Initial conditions set number 8 for constant $\Lambda$ (continued).}
\label{Fig16}
\end{figure}
\vspace{-0.1cm}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC9/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{abzz00009constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC9/a_dot_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve. The curve for $\dot b$ is scaled up by a factor of 10.}
\label{adotbdot00009constantL}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC9/a_ddot_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve. The curve for $\ddot b$ is scaled up by a factor of 10.}
\label{addotbddot00009constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/0000/IC9/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), and $\dot k\left(0\right)=-1$ (red curve). While $\dot k\left(0\right)=0$ diverges.}
\label{k00009constantL}
\end{subfigure}
\caption{Initial conditions set number 9 for constant $\Lambda$.}
\label{Fig17}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC10/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{abzz1-02-1-021constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC10/a_dot_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve.}
\label{adotbdot1-02-1-021constantL}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC10/a_ddot_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve.}
\label{addotbddot1-02-1-021constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC10/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k1-02-1-021constantL}
\end{subfigure}
\caption{Initial conditions set number 10 for constant $\Lambda$.}
\label{Fig19}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC11/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{abzz1-02-1-022constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC11/a_dot_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve.}
\label{adotbdot1-02-1-022constantL}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC11/a_ddot_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve.}
\label{addotbddot1-02-1-022constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC11/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k1-02-1-022constantL}
\end{subfigure}
\caption{Initial conditions set number 11 for constant $\Lambda$.}
\label{Fig22}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC12/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{abzz1-02-1-023constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC12/a_dot_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve.}
\label{adotbdot1-02-1-023constantL}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC12/a_ddot_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve.}
\label{addotbddot1-02-1-023constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC12/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k1-02-1-023constantL}
\end{subfigure}
\caption{Initial conditions set number 12 for constant $\Lambda$.}
\label{Fig24}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC13/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{abzz1-02-1-024constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC13/a_dot_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve.}
\label{adotbdot1-02-1-024constantL}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC13/a_ddot_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve.}
\label{addotbddot1-02-1-024constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC13/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k1-02-1-024constantL}
\end{subfigure}
\caption{Initial conditions set number 13 for constant $\Lambda$.}
\label{Fig26}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC14/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{abzz1-02-1-025constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC14/a_dot_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve.}
\label{adotbdotEARLY1-02-1-025constantL}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC14/a_ddot_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve.}
\label{addotbddotEARLY1-02-1-025constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC14/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k1-02-1-025constantL}
\end{subfigure}
\caption{Initial conditions set number 14 for constant $\Lambda$.}
\label{Fig27}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC15/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{abzz1-02-1-026constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC15/a_dot_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve.}
\label{adotbdot1-02-1-026constantL}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC15/a_ddot_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve.}
\label{addotbddot1-02-1-026constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC15/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k1-02-1-026constantL}
\end{subfigure}
\caption{Initial conditions set number 15 for constant $\Lambda$.}
\label{Fig30}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC16/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while ${G_{i\bar j} \dot z^i \dot z^{\bar j}}$ (not the absolute value) by the green curve.}
\label{abzz1-02-1-027constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC16/a_dot_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve.}
\label{adotbdot1-02-1-027constantL}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC16/a_ddot_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve.}
\label{addotbddot1-02-1-027constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC16/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k1-02-1-027constantL}
\end{subfigure}
\caption{Initial conditions set number 16 for constant $\Lambda$.}
\label{Fig32}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC17/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{abzz1-02-1-028constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC17/a_dot_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve.}
\label{adotbdot1-02-1-028constantL}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC17/a_ddot_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve.}
\label{addotbddot1-02-1-028constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC17/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k1-02-1-028constantL}
\end{subfigure}
\caption{Initial conditions set number 17 for constant $\Lambda$.}
\label{Fig34}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC18/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{abzz1-02-1-029constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC18/a_dot_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve.}
\label{adotbdot1-02-1-029constantL}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC18/a_ddot_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve.}
\label{addotbddot1-02-1-029constantL}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/1_constant_lambda/1-02-1-02/IC18/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k1-02-1-029constantL}
\end{subfigure}
\caption{Initial conditions set number 18 for constant $\Lambda$.}
\label{Fig35}
\end{figure}
The conclusions one may draw from all of these cases are straightforwardly obtained by inspection. The scale factors $a$ of these `universes' are strongly correlated to the values of the moduli flow norm $ G_{i\bar j} \dot z^i \dot z^{\bar j} $; large values of the later correspond to small values of the former, maximum values of the later correspond to minimum values of the former, and so on. Particularly important is that the flow norm always starts at large, almost infinite, values and decays rapidly, inducing an inflationary behavior for $a\left(t\right)$ \emph{without} the unnecessary introduction of an inflaton \cite{Emam:2020oyb}. It is also interesting that later time accelerations do not necessarily arise because of the presence of a cosmological constant $\Lambda$. For instance in IC set \#3 there exists a late time accelerating behavior of $a$ that is \emph{not} induced by the presence of $\Lambda$, but rather seems to be related to $\tilde \Lambda$ mediated by the moduli flow, and similarly for IC set \#6. This last is particularly rich in that it exhibits inflation, deceleration, acceleration, and then continues to an almost constant value for $a$. If taken seriously from a cosmological perspective it seems to predict a universe that at later times oscillates and then decays to a constant `size.' If this represents our universe then our current time may, perhaps, be between $t=4$ and $t=5$ on the plot's scale. Another interesting behavior is exhibited by the bulk scale factor $b$ in IC sets \#3, 5, 7, and 9. It starts at positive values then smoothly switches to negative! This `spacelike to timelike' behavior of metric components indicates that there is a causal discontinuity (a horizon if you will) in the bulk at a certain critical time $t_c$; where $b$ vanishes. In general, our own universe may be described by the solutions following IC sets \#2, 3, 4, 5 (which has a future big rip despite a negative $\Lambda$), 6, 8, and 9 (also with a big rip).
If we look at IC set \#1 particularly, we can describe the behaviour of the remaining fields. For example the dilaton is known to be proportional to the volume of the underlying Calabi-Yau metric. In most solutions this initially drops, in correlation with the inflationary epoch, then expands mostly to constant values and in some cases (correlated with the big rip) to infinity. This is consistent with the behavior of the scale factors in that all eleven dimensions behave similarly. The axionic values are mostly related to the initial inflation as one might expect.
\subsection{Constant $b$}
Analysis here follows the initial conditions schemes outlined in table (\ref{tabb}). Once again this is broken down into the same two major categories with varying values for $\tilde \Lambda$. The first six sets are for constant $\tilde \Lambda$ allowing for varying $\Lambda$, while the second set of six are for constant $\Lambda$ allowing for varying $\tilde\Lambda$.
\begin{table}[H]
\centering
\caption{The twelve sets of initial conditions (IC) used in the constant $b$ computations.}
\label{tabb}
\vspace{0.5cm}
\begin{adjustbox}{max width=\textwidth}
\begin{tabular}{|c|c|c|c|l|}
\hline
\textbf{IC Set \#} & $a\left(0\right)$ & $\dot a\left(0\right)$ & $\tilde \Lambda$ & \textbf{Description} \\ \hline
\textbf{1} & \multirow{3}{*}{0} & \multirow{3}{*}{0} & 0 & \multirow{3}{*}{Big bang-like IC} \\ \cline{1-1} \cline{4-4}
\textbf{2} & & & 1 & \\ \cline{1-1} \cline{4-4}
\textbf{3} & & & -1 & \\ \hline
\textbf{4} & \multirow{3}{*}{1} & \multirow{3}{*}{-0.2} & 0 & \multirow{3}{*}{Non- singular IC with negative initial velocity} \\ \cline{1-1} \cline{4-4}
\textbf{5} & & & 1 & \\ \cline{1-1} \cline{4-4}
\textbf{6} & & & -1 & \\ \hline\hline
\textbf{IC Set \#} & $a\left(0\right)$ & $\dot a\left(0\right)$ & $\Lambda$ & \textbf{Description} \\ \hline
\textbf{7} & \multirow{3}{*}{0} & \multirow{3}{*}{0} & 0 & \multirow{3}{*}{Big bang-like IC} \\ \cline{1-1} \cline{4-4}
\textbf{8} & & & 1 & \\ \cline{1-1} \cline{4-4}
\textbf{9} & & & -1 & \\ \hline
\textbf{10} & \multirow{3}{*}{1} & \multirow{3}{*}{-0.2} & 0 & \multirow{3}{*}{Non- singular IC with negative initial velocity} \\ \cline{1-1} \cline{4-4}
\textbf{11} & & & 1 & \\ \cline{1-1} \cline{4-4}
\textbf{12} & & & -1 & \\ \hline
\end{tabular}
\end{adjustbox}
\end{table}
\clearpage
\newpage
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC1/a_L_zz}
\caption{The scale factor $a$ is represented by the blue curve, $\Lambda$ by the red curve, while $ {G_{i\bar j} \dot z^i \dot z^{\bar j}} $ by the green curve.}
\label{aLzz00001constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC1/a_dot_L_dot}
\caption{The expansion rates: $\dot a$ is represented by the blue curve, and $\dot \Lambda$ by the red curve. The curve for $\dot a$ is scaled up by a factor of 3000.}
\label{adotLdot00001constantb}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC1/a_ddot_L_ddot}
\caption{The accelerations: $\ddot a$ is represented by the blue curve, and $\ddot \Lambda$ by the red curve. The curve for $\dot a$ is scaled up by a factor of 10000.}
\label{addotLddot00001constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC1/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k00001constantb}
\end{subfigure}
\caption{Initial conditions set number 1 for constant $b$.}
\label{Fig38}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC1/sigma}
\caption{The dilaton $\sigma$; same for all three $\dot k\left(0\right)$.}
\label{sigma00001constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC1/sigmadot}
\caption{The dilatonic field strength $\dot\sigma$.}
\label{sigmadot00001constantb}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC1/phi}
\caption{The universal axion $\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve). The solution diverges for $\dot k\left(0\right)=0$.}
\label{phi00001constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC1/phidot}
\caption{The axionic field strength $\dot\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve).}
\label{phidot00001constantb}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC1/xx}
\caption{$ \langle \Xi | \dot \Xi \rangle$ for $\dot{k}(0)= 1$ (blue), and $\dot{k}(0) = -1$ (red).}
\label{xx00001constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC1/Omega}
\caption{$\Omega$ at $\dot k\left(0\right) = 1$ and $ \dot\sigma \left(0\right) =0 $.}
\label{omega00001constantb}
\end{subfigure}
\vspace{0.3cm}
\caption{Initial conditions set number 1 for constant $b$ (continued).}
\label{Fig37}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC2/a_L_zz}
\caption{The scale factor $a$ is represented by the blue curve, $\Lambda$ by the red curve, while $ {G_{i\bar j} \dot z^i \dot z^{\bar j}} $ by the green curve.}
\label{aLzz00002constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC2/a_dot_L_dot_EARLY}
\caption{The early time expansion rates: $\dot a$ is represented by the blue curve, and $\dot \Lambda$ by the red curve. The curve for $\dot a$ is scaled up by a factor of 3000.}
\label{adotLdotEARLY00002constantb}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC2/a_dot_L_dot_LATE}
\caption{The late time expansion rates: $\dot a$ is represented by the blue curve, and $\dot \Lambda$ by the red curve.}
\label{adotLdotLATE00002constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC2/a_ddot_L_ddot_EARLY}
\caption{The early time accelerations: $\ddot a$ is represented by the blue curve, and $\ddot \Lambda$ by the rred curve. The curve for $\dot a$ is scaled up by a factor of 10000.}
\label{addotLddotEARLY00002constantb}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC2/a_ddot_L_ddot_LATE}
\caption{The late time accelerations: $\ddot a$ is represented by the blue curve, and $\ddot \Lambda$ by the red curve.}
\label{addotLddotlate00002constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC2/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k00002constantb}
\end{subfigure}
\caption{Initial conditions set number 2 for constant $b$.}
\label{Fig391}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC3/a_L_zz}
\caption{The scale factor $a$ is represented by the blue curve, $\Lambda$ by the red curve, while $ {G_{i\bar j} \dot z^i \dot z^{\bar j}} $ by the green curve.}
\label{aLzz00003constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC3/a_dot_L_dot}
\caption{The expansion rates: $\dot a$ is represented by the blue curve, and $\dot \Lambda$ by the red curve.}
\label{adotLdot00003constantb}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC3/a_ddot_L_ddot}
\caption{The accelerations: $\ddot a$ is represented by the blue curve, and $\ddot \Lambda$ by the red curve.}
\label{addotLddot00003constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC3/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k00003constantb}
\end{subfigure}
\caption{Initial conditions set number 3 for constant $b$.}
\label{Fig42}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC4/a_L_zz}
\caption{The scale factor $a$ is represented by the blue curve, $\Lambda$ by the red curve, while $ {G_{i\bar j} \dot z^i \dot z^{\bar j}} $ by the green curve.}
\label{aLzz1-02-1-024constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC4/a_dot_L_dot}
\caption{The expansion rates: $\dot a$ is represented by the blue curve, and $\dot \Lambda$ by the red curve.}
\label{adotLdot1-02-1-024constantb}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC4/a_ddot_L_ddot}
\caption{The accelerations: $\ddot a$ is represented by the blue curve, and $\ddot \Lambda$ by the red curve.}
\label{addotLddot1-02-1-024constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC4/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k1-02-1-024constantb}
\end{subfigure}
\caption{Initial conditions set number 4 for constant $b$.}
\label{Fig44}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC5/a_L_zz}
\caption{The scale factor $a$ is represented by the blue curve, $\Lambda$ by the red curve, while $ {G_{i\bar j} \dot z^i \dot z^{\bar j}} $ by the green curve.}
\label{aLzz1-02-1-025constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC5/a_dot_L_dot}
\caption{The expansion rates: $\dot a$ is represented by the blue curve, and $\dot \Lambda$ by the red curve.}
\label{adotLdot1-02-1-025constantb}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC5/a_ddot_L_ddot}
\caption{The accelerations: $\ddot a$ is represented by the blue curve, and $\ddot \Lambda$ by the red curve.}
\label{addotLddot1-02-1-025constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC5/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k1-02-1-025constantb}
\end{subfigure}
\caption{Initial conditions set number 5 for constant $b$.}
\label{Fig46}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC6/a_L_zz}
\caption{The scale factor $a$ is represented by the blue curve, $\Lambda$ by the red curve, while $ {G_{i\bar j} \dot z^i \dot z^{\bar j}} $ by the green curve.}
\label{aLzz1-02-1-026constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC6/a_dot_L_dot}
\caption{The expansion rates: $\dot a$ is represented by the blue curve, and $\dot \Lambda$ by the red curve.}
\label{adotLdot1-02-1-026constantb}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC6/a_ddot_L_ddot}
\caption{The accelerations: $\ddot a$ is represented by the blue curve, and $\ddot \Lambda$ by the red curve.}
\label{addotLddot1-02-1-026constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC6/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k1-02-1-026constantb}
\end{subfigure}
\caption{Initial conditions set number 6 for constant $b$.}
\label{Fig47}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC7/a_zz_LamT}
\caption{The scale factor $a$ is represented by the blue curve, $\tilde\Lambda$ by the red curve, while $ \left|{G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{aLTzz00007constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC7/a_dot_LamT_dot}
\caption{The expansion rates: $\dot a$ is represented by the blue curve, and $\dot \tilde\Lambda$ by the red curve. The curve for $\dot a$ is scaled up by a factor of 10000.}
\label{adotLTdot00007constantb}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC7/a_ddot_LamT_ddot}
\caption{The accelerations: $\ddot a$ is represented by the blue curve, and $\ddot \tilde\Lambda$ by the red curve. The curve for $\ddot a$ is scaled up by a factor of 10000.}
\label{addotLTddot00007constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC7/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), and $\dot k\left(0\right)=-1$ (red curve). While $\dot k\left(0\right)=0$ diverges.}
\label{k00007constantb}
\end{subfigure}
\caption{Initial conditions set number 7 for constant $b$.}
\label{Fig50}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC8/a_zz_LamT}
\caption{The scale factor $a$ is represented by the blue curve, $\tilde\Lambda$ by the red curve, while $ \left|{G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{aLTzz00008constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC8/a_dot_LamT_dot}
\caption{The expansion rates: $\dot a$ is represented by the blue curve, and $\dot \tilde\Lambda$ by the red curve. The curve for $\dot a$ is scaled up by a factor of 10000.}
\label{adotLTdot00008constantb}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC8/a_ddot_LamT_ddot}
\caption{The accelerations: $\ddot a$ is represented by the blue curve, and $\ddot \tilde\Lambda$ by the red curve. The curve for $\ddot a$ is scaled up by a factor of 10000.}
\label{addotLTddot00008constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC8/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k00008constantb}
\end{subfigure}
\caption{Initial conditions set number 8 for constant $b$.}
\label{Fig51}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC9/a_zz_LamT}
\caption{The scale factor $a$ is represented by the blue curve, $\tilde\Lambda$ by the red curve, while $ {G_{i\bar j} \dot z^i \dot z^{\bar j}} $ by the green curve.}
\label{aLTzz00009constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC9/a_dot_LamT_dot}
\caption{The expansion rates: $\dot a$ is represented by the blue curve, and $\dot \tilde\Lambda$ by the red curve. The curve for $\dot a$ is scaled up by a factor of 10000.}
\label{adotLTdot00009constantb}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC9/a_ddot_LamT_ddot}
\caption{The accelerations: $\ddot a$ is represented by the blue curve, and $\ddot \tilde\Lambda$ by the red curve. The curve for $\ddot a$ is scaled up by a factor of 10000.}
\label{addotLTddot00009constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC9/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k00009constantb}
\end{subfigure}
\caption{Initial conditions set number 9 for constant $b$.}
\label{Fig54}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC10/a_zz_LamT}
\caption{The scale factor $a$ is represented by the blue curve, $\tilde\Lambda$ by the red curve, while $ \left|{G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{aLTzz000010constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC10/a_dot_LamT_dot}
\caption{The expansion rates: $\dot a$ is represented by the blue curve, and $\dot \tilde\Lambda$ by the red curve.}
\label{adotLTdot000010constantb}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC10/a_ddot_LamT_ddot}
\caption{The accelerations: $\ddot a$ is represented by the blue curve, and $\ddot \tilde\Lambda$ by the red curve.}
\label{addotLTddot000010constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC10/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k000010constantb}
\end{subfigure}
\caption{Initial conditions set number 10 for constant $b$.}
\label{Fig55}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC11/a_zz_LamT}
\caption{The scale factor $a$ is represented by the blue curve, $\tilde\Lambda$ by the red curve, while $ \left|{G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{aLTzz000011constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC11/a_dot_LamT_dot}
\caption{The expansion rates: $\dot a$ is represented by the blue curve, and $\dot \tilde\Lambda$ by the red curve.}
\label{adotLTdot000011constantb}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC11/a_ddot_LamT_ddot}
\caption{The accelerations: $\ddot a$ is represented by the blue curve, and $\ddot \tilde\Lambda$ by the red curve.}
\label{addotLTddot000011constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC11/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k000011constantb}
\end{subfigure}
\caption{Initial conditions set number 11 for constant $b$.}
\label{Fig58}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC12/a_zz_LamT}
\caption{The scale factor $a$ is represented by the blue curve, $\tilde\Lambda$ by the red curve, while $ {G_{i\bar j} \dot z^i \dot z^{\bar j}} $ by the green curve.}
\label{aLTzz000012constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC12/a_dot_LamT_dot}
\caption{The expansion rates: $\dot a$ is represented by the blue curve, and $\dot \tilde\Lambda$ by the red curve.}
\label{adotLTdot000012constantb}
\end{subfigure}
\\[5em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC12/a_ddot_LamT_ddot}
\caption{The accelerations: $\ddot a$ is represented by the blue curve, and $\ddot \tilde\Lambda$ by the red curve.}
\label{addotLTddot000012constantb}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics2/2_constant_b/IC12/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{k000012constantb}
\end{subfigure}
\caption{Initial conditions set number 12 for constant $b$.}
\label{Fig57}
\end{figure}
Once again the correlation between the scale factor $a$ and the moduli flow norm is clear. For the first set of six IC's, it is interesting to note that a constant value of the bulk scale $b$ seems to necessarily lead to mostly negative values $\Lambda\left(t\right)$. As such one cannot seriously consider those from a cosmological perspective. However, IC sets \#2, and 5 are clear exceptions. In fact these are the only two cases that may be thought of as possible histories of our universe. There are two interesting cases of periodical universes; these are IC sets \# 3 and 6, corresponding to negative $\tilde \Lambda$. For the second set of six initial conditions only set\#8 can be a model of our universe, although a trivial one since this is a set with a positive $\Lambda$. Oscillating universes with negative $\tilde \Lambda$ appear in sets 9 and 12. Finally the behavior of the remaining fields is similar to those studied in the constant $\Lambda$ cases.
\section{A brane- world depends on the bulk extra dimension}
Consider a case where all the fields in the bulk are dependent on time and the bulk extra dimension ($y$) \cite{Emam2013}.
The 3- brane and the bulk cosmological constants $\Lambda$ and $\tilde{\Lambda}$ are static.
The brane- universe contains dust and radiation with density and pressure independent of time, and it's a de-Sitter- like spacetime with $p=-\rho$. the energy stress tensors of the bulk, then will be:
\begin{eqnarray}
\nonumber
T_{\mu\nu}^{Bulk} &=& G_{i\bar{j}} \partial_\mu z^i \partial_\nu z^{\bar{j}} - \frac{1}{2} g_{\mu\nu} G_{i\bar{j}} \partial_\alpha z^i \partial^\alpha z^{\bar{j}}, \\
T_{yy} &=& G_{i\bar{j}} \partial_y z^i \partial_y z^{\bar{j}} - \frac{1}{2} g_{yy} G_{i\bar{j}} \partial_\alpha z^i \partial^\alpha z^{\bar{j}}.
\end{eqnarray}
Solving the Einstein's equations:
\begin{equation}
\begin{aligned}
G_{\mu\nu}+\Lambda g_{\mu\nu} =& T_{\mu\nu}^{Bulk} + T_{\mu\nu}^{3b} \\
G_{yy}+\tilde{\Lambda} g_{yy} =& T_{yy}, ~~~~~~~~ G_{yt}= T_{yt}=0,
\end{aligned}
\end{equation}
gives:
\begin{eqnarray}
\nonumber
G_{tt} - \Lambda &=& \frac{1}{2} G_{i\bar{j}} ~\dot z^{i} ~ \dot z^{\bar{j}} + \frac{1}{2} b^2 G_{i\bar{j}} ~z^{i'} ~ z^{\bar{j}'} + \rho, \\ \nonumber
G_{rr}+ a^2 \Lambda & =& \frac{1}{2} a^2 G_{i\bar{j}} ~\dot z^{i} ~ \dot z^{\bar{j}} - \frac{1}{2} a^2 b^2 G_{i\bar{j}} ~z^{i'} ~ z^{\bar{j}'} + a^2 p \\
G_{yy} + b^2 \tilde{\Lambda} &=& \frac{1}{2} b^2 G_{i\bar{j}} ~\dot z^{i} ~ \dot z^{\bar{j}} + \frac{1}{2} G_{i\bar{j}} ~z^{i'} ~ z^{\bar{j}'} .
\end{eqnarray}
The components of the Einstein tensor are:
\begin{eqnarray}
\label{GG2}
\nonumber G_{tt}&=& 3\left[ {\left( {\frac{{\dot a}}{a}} \right)^2 + \left( {\frac{{\dot a}}{a}} \right)\left( {\frac{{\dot b}}{b}} \right)} \right] ,\\ \nonumber
G_{rr} &=& - \left( 2 \ddot a a + \dot a^2 \right) - \frac{{a}}{b} \left( \ddot b a + 2 \dot a \dot b \right) , \\
G_{yy} &=&
-3 \Big(\frac{b}{a}\Big)^2 (\ddot a a + \dot a^2) .
\end{eqnarray}
These equations admit multiple branes located at various value of $y=y_I (I=1,...,N \in \mathbb{Z})$, such that $y \to \sum\limits_{I=1}^N |y-y_I|$. However we study the fields effectively at an arbitrary $y$ value or in other words as observed from our brane. For a static extra dimension, the 3-brane's Hubble parameter $H=\frac{\dot{a}}{a}$, and the bulk's scale factor $b=1$. Thus, the modified Friedmann equations become:
\begin{eqnarray}
3 H^2 &=& \dot z {\dot{\bar z}} +z^\prime \bar{z}^\prime + \rho+ \Lambda , \label{lam1} \\
- \Big( 2 \frac{\ddot a}{a} + H^2 \Big) &=& - \Lambda - z^\prime \bar{z}^\prime + \dot z {\dot{\bar z}} + p , \label{lam2} \\
-3 \Big( \frac{\ddot a}{a} + H^2 \Big) &=& z^\prime \bar{z}^\prime + \dot z {\dot{\bar z}} - \tilde{\Lambda}. \label{lam3}
\end{eqnarray}
Define $\Lambda_{eff}= z^\prime \bar{z}^\prime + \Lambda$. Where $\Lambda$ is interpreted as
the cosmological constant related to the value of the quantum mechanical vacuum energy, while $\Lambda_{eff}$ is the observed value of the dark energy density based on the brane universe rate of acceleration. While the norm of the moduli variation with respect to the bulk fifth dimension acts to reduce the first to the second. $z' \bar{z}' \equiv G_{i\bar{j}} ~z^{i'} ~ z^{\bar{j}'}$ and $\dot z {\dot{\bar z}} \equiv G_{i\bar{j}} ~\dot z^{i} ~ \dot z^{\bar{j}} $. It is known that the scalar quantity
$G_{i\bar{j}} ~z^{i'} ~ z^{\bar{j}'}$ has a negative value.
Now substitute by Equ. (\ref{lam1}) in Equ. (\ref{lam2}), then substitute by both Equ. (\ref{lam1}), and Equ. (\ref{lam2}) into Equ. (\ref{lam3}), the equations become:
\begin{eqnarray}
\nonumber H^2 &=& \frac{1}{3} \Big(\Lambda_{eff} + \dot z {\dot{\bar z}} +\rho \Big),
\\ \nonumber \frac{\ddot a}{a} &=& \frac{1}{3} \Lambda_{eff} - \frac{2}{3} \dot z {\dot{\bar z}} - \frac{1}{6} (\rho+ 3 p),
\\ \tilde{\Lambda} &=& 2 \Lambda_{eff} + z' \bar{z}' + \frac{1}{2} (\rho- 3 p) .
\label{fr1}
\end{eqnarray}
For a dynamical extra dimension: $\dot{b} \neq 0 $,
let $\tilde{H} =\frac{\dot{b}}{b}$ that defines the Hubbel parameter of the bulk, then the modified Friedmann equations become:
\begin{eqnarray}
\nonumber 3 \left[ H^2 + H \tilde{H} \right] &=& \dot z {\dot{\bar z}} + b^2 z^\prime \bar{z}^\prime + \rho+ \Lambda , \\ \nonumber
- \Big( 2 \frac{\ddot a}{a} + \frac{\ddot b}{b} + H^2 + H \tilde{H} \Big) &=& - \Lambda - b^2 z^\prime \bar{z}^\prime + \dot z {\dot{\bar z}} + p , \\
-3 b^2 \Big( \frac{\ddot a}{a} + H^2 \Big) &=& z^\prime \bar{z}^\prime + b^2 \dot z {\dot{\bar z}} - b^2 \tilde{\Lambda}.
\end{eqnarray}
Let $\Lambda_{eff}= b^2 z^\prime \bar{z}^\prime + \Lambda$
\begin{eqnarray}
\nonumber H^2 + H \tilde{H} &=& \frac{1}{3} \Big( \Lambda_{eff} + \dot z {\dot{\bar z}} + \rho \Big), \\ \nonumber
\frac{\ddot a}{a} + \frac{1}{2} \frac{\ddot b}{b} &=& \frac{1}{3} \Lambda_{eff} -\frac{2}{3} \dot z {\dot{\bar z}}
- \frac{1}{6} (\rho+ 3 p) , \\
\tilde{\Lambda} &=& 2 \Lambda_{eff} + z' \bar{z}' -\frac{\ddot b}{b} - H \tilde{H} + \frac{1}{2} (\rho- 3 p) .
\label{fr2}
\end{eqnarray}
For a de-Sitter-like spacetime $p=-\rho$. From Equ.(\ref{fr1}), and Equ.(\ref{fr2}) we notice that the fifth extra dimension
whether a de-Sitter-like if it is static with a constant scale factor, or it can be an anti de-Sitter-like if it expands with time and
$\frac{\ddot b}{b}$ and $H \tilde{H}$ are large enough. Also in this case the bulk scale factor $b(t)$ contributes in reducing the 3-brane's cosmological constant value. In this way the cosmological constant problem can be resolved analytically in a different way than the previous solutions have been made in this chapter. Where in this section what contributes in enhancing the brane's cosmological constant to get the observed value is the moduli's variation with respect to the fifth extra dimension.
\chapter{Conclusion}
We have studied a 3-brane embedded in higher dimensional bulk. Studying higher dimensions and brane- theories have manifested themselves in recent years as they lead to understand superstring theories, the dimensional reduction and the topological spaces where these theories exist. Brane- cosmology then has been emerged - supersymmetric, non-supersymmetric - or within the landscape of string theory \cite{sus, st} as modeling the universe as expanding branes in different evolution stages like inflation , accelerating or decelerating expansions.
Indeed our aim is that usually the early times of the universe like inflation or radiation dominant eras are investigated solely by general relativity theories, while the current accelerating expansion era is investigated by quantum field theories like supersymmetric theories. On the other hand, there is no such distinction in the universe time evolution
through our study framework.
We have continued previous works began before \cite{Emam2013} and \cite{Emam:2015laa}. Where in \cite{Emam2013} a 3-brane embedded in
$\mathcal{N}=2$ five dimensional supergravity with hypermultiplets in the bulk have been studied. All the hypermultiplets fields depend on the bulk’s fifth extra dimension y as well as time. The brane- universe considered a de-Sitter space filled only with energy ( cosmological constant $\Lambda$ ). In \cite{Emam:2015laa} a BPS condition has been imposed on the 3-brane which was totally vacuous, while the hypermultiplets taken to be very near from the brane- world, i.e., the y- dependence vanished. It has shown in the last that the moduli of the complex structure of the underlying Calabi–Yau (CY) space act as a possible agent for the various cosmological stages of the considered brane. However that wasn’t a pretty realistic model because the brane was empty of any matter, radiation or energy. So may be questions raised that how far the results there connect to our universe. Also there was assumptions made about the brane scale factor $a(t)$ and the velocity norm of the moduli $G_{i\bar{j}} \dot{z}^i \dot{z}^{\bar{j}}$ to evolve with time similarly as our universe’s time evolution.
So we have studied a BPS 3-brane embedded in $\mathcal{N}=2$ $\mathcal{D}=5$ supergravity. The branes studied once filled with radiation ensemble the early times of our baby universe after the big bang, once filled with matter ensemble the later on times and once filled with matter, radiation and energy to ensemble the current accelerated expansion era. All these cases investigated at different initial conditions, three of which are big bang/ initial singularity- like conditions. That yields different brane- universes for whose history are similar to
our own.
Confirming to the previous results, we have found that velocity norm of the Calabi–Yau complex structure moduli is strongly correlated to brane scale factor $a(t)$. However we didn’t need to make any assumptions, we got our results directly by solving the Einstein’s field equations numerically. $G_{i\bar{j}} \dot{z}^i \dot{z}^{\bar{j}}$ starts large and quickly decays to zero, `causing` the expansion of the
brane- universe. This is consistent with arguments that our universe’s big bang must have produced large, but unstable, quantities of moduli that decayed very quickly \cite{Bodeker:2006ij, Dine:2006ii}. Most solutions have an early, and very short, period of positive acceleration for the brane’s scale factor; strongly suggesting a causality between the decaying moduli velocity norms and an early inflationary period.
Also we show here that the moduli norm not only related to $a(t)$, but it’s also related to the bulk’s scale factor $b(t)$ which points to the dimensions of the bulk itself and related to the remaining hypermultiplet fields which exhibit interesting early time behavior, then mostly vanish asymptotically following the behavior of $G_{i\bar{j}} \dot{z}^i \dot{z}^{\bar{j}}$. Moreover, we have found that in many cases the brane-universes exhibit late time acceleration with introducing a brane cosmological constant $\Lambda$ or at vanishing $\Lambda$ and introducing an extra bulk’s cosmological constant $\tilde{\Lambda}$ . Declares that the problem of the dark energy can be interpreted by only the effects of a higher dimensional bulk.
Some of the 3-branes solutions describe universes that show very early inflation, late time deceleration and current time accelerated expansion just exactly like our own universe . Even within these solutions we can predict various scenarios about the universe’s destiny. We studied the moduli's variation with respect to the fifth extra dimension, where the cosmological constant problem can be resolved analytically.
These results shed light on the role of the topology of the Calabi-Yau manifold. In other words what are the deep correlations that make
$G_{i\bar{j}} \dot{z}^i \dot{z}^{\bar{j}}$ derives the whole time evolution of the brane- world and the bulk as well as the other hypermultiplet fields.
That made us go in the last chapter further towards the Calabi–Yau manifold topological structure to answer questions like, how the moduli fields $z^i$ vary with time ? what are the time and space dependence of the CY complex structure space shown in its metric $G_{i\bar{j}}$ ? what about the dimensions of CY manifold? it’s related to the dilaton field which evolve in time, so does the volume of the CY manifold really depend on time? we studied analytically the moduli and the metric component G in case the dimension of the CY complex structure space $\mathcal{M}_c$, $h_{(2,1)}=1$, i.e., $i,\bar{j} =1$. For instance, in case when the Calabi-Yau manifold is considered as a quintic 3-fold, $h_{(2,1)}=101$, however here we do not need to make this assumption because we already know many information about $G_{i\bar{j}} \dot{z}^i \dot{z}^{\bar{j}}$. Then we have found analytic solutions to the the potential of the complex structure space and the volume of the Calabi–Yau manifold which as shown change with time even with a rare rate.
In summary, we outlined previous studies about 3brane- universes embedded in five dimensional bulk of
$\mathcal{N}=2$ supergravity. Which gives a comprehensive interpretation about our universe’s cosmological aspects since its birth till the current epoch. That may has not been provided in such way before, i.e., a complete explanation of the universe’s history and future predictions.
We have taken into account the study of the topology of the CY manifold where the theory itself live. We paid an attention that may be there are other multiverses have been evolved with our own . Because there are no any reasons that the very tuned initial conditions that created our universe are special or unique.
\chapter{Theory of Everything Towards Unification }
The standard model (SM) agrees with all
the existing particle physics experiments, and the general relativity is consistent with most of the astrophysical and cosmological observations. However both the models have some drawbacks as the following:
\begin{itemize}
\item The general relativity is a classical theory, not valid at the quantum level.
\item The standard model involves a large number of parameters, some of which are unnaturally small.
\item Neutrino oscillations and the convergence of gauge coupling constants point to the existence
of a heavy mass scale, the Grand Unification scale, $M_U \sim 10^{16}$ GeV, and/or the string
scale $M_{str} \sim 10^{17}$ GeV, not explained by the SM.
\item Many features of the standard cosmological model such as the very nature of dark matter
and dark energy, the asymmetry between matter and antimatter, the initial Big Bang
singularity (still present in inflationary models), the quantum structure of black holes and of
space- time itself, are not yet well understood.
\end{itemize}
Supersymmetry, in its spontaneously broken phase, is the leading proposal for new physics
beyond the Standard Model, subject to experimental tests in modern colliders \cite{Endo:2021}, and is the
leading paradigm for the early universe and inflationary cosmology. Supersymmetry is naturally
realized in the framework of superstrings, which is the only known candidate for
a consistent quantum theory of gravity.
Furthermore, superstring theory (1984-1994) has successfully
passed a series of nontrivial tests \cite{Kounnas}:
\begin{itemize}
\item Gravity and the basic ingredients of the Standard Model emerge very naturally.
\item Superstring theory is free of ultraviolet divergences and anomalies.
\item It incorporates in a very beautiful and natural way many of the major theoretical ideas
such as supersymmetry, Grand Unification and the possible existence of large extra dimensions.
\item It has explained the microscopic origin of the entropy and other thermodynamical properties of black holes.
\end{itemize}
In 1995, the physicist Edward Witten discovered the mother of all string theories. He found various indications that the perturbative string theories fit together into a coherent nonperturbative theory, which he dubbed M-theory. M-theory looks like each of the string theories in different physical contexts but does not itself have limits on its regime of validity - a major requirement for the theory of everything.
So though Witten didn't write down complete equations of M-theory, why, then, string/M-theory have given the edge of theoretical physics searches for this decade and most probably for the next decade?
Fig. (\ref{Mom}) shows how M- theory unifies all the higher dimensional theories in a beautiful and an elegant way.
In this chapter we will introduce the basics of superstring theories, also we will give a general overview on supergravity, especially the eleven dimensional theory, in such way we can directly deal with in chapter 3.
\section{String Basics}
\begin{figure}[!t]
\centering
\includegraphics[scale=0.3]{./Fig/mother.pdf}
\caption{The current state of string theory: The inner star gives the five known
superstring theories in ten dimensions. The outer circuit represents the three known
supergravity theories in ten dimensions in addition to the eleven dimensional theory. }
\label{Mom}
\end{figure}
The string is one dimensional object, as it moves, it sweeps out a two dimensional surface in space-time called a worldsheet. To describe the dynamics of the string we use the simplest Poincar\'e invariant action, the Nambu-Goto action
\begin{eqnarray}
\nonumber S_{NG} &=&- \frac{1}{2\pi\alpha'}~ \int d\tau d\sigma~ \sqrt{- det~ h_{ab}} \\
& =& - \frac{1}{2\pi\alpha'}~ \int d\tau d\sigma~ \sqrt{- |(\partial_a X^\mu) (\partial_b X^\nu) \eta_{\mu\nu}|},
\label{ng}
\end{eqnarray}
this generalize the relativistic action of a point particle. It describes a string worldsheet of a proper time $\tau$ and a proper length $\sigma$, propagating in flat D- dimensional space-time background $X^\mu(\tau,\sigma)$, $\mu=0,...,D-1$,
Fig. (\ref{sheet}) \cite{Nambuac}. $h_{ab}= \partial_a X^\mu \partial_b X^\nu ~ \eta_{\mu\nu} $
\footnote{\em Where our convention for the signature of the Lorentzian metrices is $(-,+,+,+)$.} is the induced metric on the world sheet, defining the pullback of space-time on the world sheet as an object embedded in space-time. By analogue to the point particle case, the constant $\frac{1}{2 \pi \alpha’}$ is the tension of the string or the mass per unit length.$\alpha’$ is called Regge slope and it’s 1 for bosonic string \footnote{ \em Look for instance at Refs. \cite{BRegge, Kir:2007}.}.
It’s useful however to write the action in a form which removes the square root \footnote{\em For quantization.}. Polyakov showed that this can be achieved by using the formalism of general relativity. Introducing a two dimensional metric $\gamma^{ab}$ . Then the invariant action is
\begin{equation}
S_P = - \frac{1}{4\pi\alpha'}~ \int d\tau d\sigma~ \sqrt{-\gamma} \gamma^{ab} ~(\partial_a X^\mu) (\partial_b X^\nu) \eta_{\mu\nu},
\label{sp}
\end{equation}
besides Poincar\'e symmetry, this action has diffeomorphism invariance under the worldsheet coordinates transformation and Weyl invariant under the rescaling of the worldsheet metric $ \gamma_{ab} \to \gamma'_{ab} = e^{2 \omega(\tau,\sigma)} \gamma_a$ .
\begin{figure}[!t]
\centering
\includegraphics[scale=0.35]{./Fig/NGo.pdf}
\caption
{A string's world sheet. The function $X^\mu(\tau,\sigma)$ embeds the world sheet,
parameterized by $(\tau,\sigma)$ into space-time coordinates $X^\mu$}
\label{sheet}
\end{figure}
Varying the action with respect to $\gamma_{ab}$ yields a two-dimensional energy momentum tensor
\begin{equation}
T^{ab}(\tau,\sigma)=-\frac{4\pi}{\sqrt{-\gamma}}\frac{\delta S}{\delta \gamma_{ab}} =
- \frac{1}{\alpha'} \{\partial^a X_\mu \partial^b X^\mu - \frac{1}{2} \gamma^{ab} \gamma_{cd}
\partial^c X_\mu \partial^d X^\mu \}.
\end{equation}
While varying the action with respect to $X^\mu$ yields the string equation of motion
\begin{equation}
\nabla^2 X^\mu =0,
\label{eom}
\end{equation}
where the Laplacian is with respect to the worldsheet coordinates. This is a single
string propagating without sources. Since the canonical momentum to $X^\mu$ is $(\partial_a X^\mu)$,
we find for an open string we are forced to impose Neumann boundary conditions Equ. (\ref{bco}) in order to avoid momentum flow off the ends of the string:
\begin{equation}
\left.\begin{aligned}
X^{' \mu}(\tau,0) &= 0\\
X^{' \mu}(\tau,\pi)&= 0
\end{aligned} \right\} \mbox{ Open String Neumann BCS}
\label{bco}
\end{equation}
While for a closed string the following periodic boundary condition have been chosen:
\begin{equation}
\left.
\begin{aligned}
X^{' \mu}(\tau,0) &= X^{' \mu}(\tau,\pi)\\
X^{\mu}(\tau,0)&= X^{\mu}(\tau,\pi)\\
\gamma_{ab}(\tau,0)& = \gamma_{ab} (\tau,\pi)\\
\end{aligned} \right\} \text{ Closed String Periodic BCS}
\label{bcc}
\end{equation}
\subsection{Spectrum}
Solving the string wave equation (\ref{eom}) and imposing the boundary conditions Equ. (\ref{bco}), gives the most general solutions for the open string
\begin{equation}
X^\mu (\sigma,\tau)=x^\mu + 2 \alpha' p^\mu \tau + i \sqrt{2\alpha'} \sum^{n=\infty}_{\substack{n=-\infty\\
n\neq0~~~}} \frac{1}{n} \alpha^\mu_n e^{-in\tau}~ cos (n\sigma). \label{FM}
\end{equation}
While imposing the boundary conditions Equ. (\ref{bcc}) gives the most general solutions for
the closed string, wherein it sweeps out a cylinder in space-time as it propagates. A closed string has a couple of solutions corresponding to right moving and left moving waves
\begin{equation}
X^\mu(\tau, \sigma)= X^\mu_R(\sigma^-)+ X^\mu_L(\sigma^+)
\end{equation}
\begin{equation}
\begin{aligned}
X^\mu_R(\sigma^-) &= \frac{x^\mu}{2} + \alpha' p^\mu \sigma^- + i \sqrt{\frac{\alpha'}{2}} \sum^{n=\infty}_{\substack{n=-\infty\\
n\neq0~~~}} \frac{1}{n} \alpha^\mu_n e^{-2in\sigma^-}\\
X^\mu_L(\sigma^+) &= \frac{x^\mu}{2} + \alpha' p^\mu \sigma^+ + i \sqrt{\frac{\alpha'}{2}} \sum^{n=\infty}_{\substack{n=-\infty\\
n\neq0~~~}} \frac{1}{n} \alpha^\mu_n e^{-2in\sigma^+}.
\end{aligned}
\end{equation}
$x^\mu$ and $p^\mu$ are the position and momentum vectors of the center of mass of
the string respectively, and $\alpha^\mu_n$ are arbitrary Fourier modes, which in the quantum
theory are interpreted as annihilation and creation operators in the Hilbert space of
the string’s oscillations. The masses of open string states are
\begin{equation}
m^2=\frac{1}{\alpha'} (N-1),
\end{equation}
this implies a condition on the space-time dimension where the string theory lives that the dimension should equal 26.
The ground state (N = 0) is tachyonic, having negative mass-squared. Since the
potential for a scalar field is $\frac{1}{2} m^2 \phi^2$, this means that the ground state is unstable. One way
to think about this is that we are using the bosonic string only as a toy model, and are indeed expanding around an unstable
state. Tachyonic state can be removed by extending space-time symmetry further than Poincar\'e symmetry
and going to the superstring theory.
The masses of closed strings are
\begin{equation}
m^2=\frac{2}{\alpha'} (N-\tilde{N}-2).
\end{equation}
The string states presumed to yield the observed elementary particles in lower
dimensions are the ones corresponding to $N = \tilde{N} = 1$, representing massless particles.
An important such state is a particle with spin two. This, naturally, is interpreted as
the graviton, confirming that string theory necessarily contains gravity as a solution.
Since the harmonics on the string can go arbitrarily high, we find that there is an
infinite tower of massive states allowed on the string. The masses of these ‘particles’
are of the order of the Planck mass, making them highly inaccessible to us. It is
assumed that the particles we observe today are the result of the massless string
states, acquiring mass via spontaneous symmetry breaking.
\subsection{Interactions}
In string theory interactions arise very elegantly because they are described by processes
in which strings join and split. In these processes, the worldsheets of free strings combine
to form a single worldsheet. The worldsheets of interacting strings have the
construction of Riemann surfaces, that are some of the most interesting two-dimensional surfaces. They are,
roughly speaking, surfaces where the two coordinates make up a complex variable. In equivalent Riemann surfaces can be distinguished by parameters called moduli
\footnote{\em We will see later how these moduli play a vital rule in the topology of the theory.}.
Finding a way to construct all Riemann surfaces with their associated moduli is a
difficult mathematical problem.
To include these interaction, another terms are added to Polyakov action Equ. (\ref{sp}). Indeed, that action is the simplest version for D-dimensional bosonic fields $X^\mu(\tau, \sigma)$. By the analogy with GR, one can add the Einstein-Hilbert action
\begin{equation}
\chi = \frac{1}{4\pi} \int \sqrt{g} \mathcal{R},
\end{equation}
$\mathcal{R}$ again being the worldsheet curvature. This is not merely diff $\times$ Weyl invariant, it is
topologically invariant. For worldsheets with boundaries, as
in open string theory, (diff $\cdot$ Weyl) invariance requires also a boundrey term:
\begin{equation}
\chi = \frac{1}{4\pi} \int d\sigma d\tau \sqrt{\gamma}~ \mathcal{R} +
\frac{1}{2\pi} \int_{\mbox{boundary}} ds~~ k,
\end{equation}
where k is the extrinsic curvature on the boundary \cite{polchinski1994}.
The quantity $\chi$ is called Eular characteristic and it depends only on the topology of the world sheet and so will only matter when comparing world sheets of different topology. Now, the full string action
that resembles two-dimensional gravity coupled to D bosonic "matter" fields $X^\mu$ is given by
\begin{equation}
S= S_P + \lambda~ \chi,
\end{equation}
where $\lambda$ is an arbitrary parameter \footnote{ \em It will turn out that $\lambda$ is not a free parameter. In the full string theory, it has dynamical meaning, and will be equivalent to the expectation value of one of the massless fields - the dilaton - described by the string.}.
So what is $\lambda$'s role $?$ The path integral defining the string theory:
\begin{equation}
Z_P = \int dX d\gamma ~ e^{-S}.
\end{equation}
For a closed surface with h bundles $\chi= 2 - 2 \mbox{h}$. At closed string emission or reabsorbtion,
as in Fig. (\ref{int}), the resulting amplitudes will be weighted by a factor $e^{-\lambda \chi}$. So while this term does not affect anything local like the worldsheet equations of motion, it does affect the relative weights of surfaces of
different topologies. The quantity $g_c = e^\lambda$ is called the closed string coupling.
For open string amplitudes analogue to Fig. (\ref{int}) the Euler number is $2 - 2\mbox{h} - \mbox{b}$, where b are
the boundaries. Adding a strip increases b by one and so the path integral weight changes by $e^{\lambda}$. So that
$g_c= g_o^2$, where $g_o$ is the open string coupling.
\begin{figure}[!t]
\centering
\includegraphics[scale=0.25]{./Fig/int.pdf}
\caption{Closed strings scattering by the exchange of (a) one string, and (b) two strings. }
\label{int}
\end{figure}
Treated as a two dimensional conformal field theory, string
theory in a flat 26-dimensional background can be reformulated as if it were in a
curved background. Roughly speaking, a theory of strings on a curved background is
equivalent to one embedded in a flat metric with a background teaming with coherent
gravitonic string states. This enables us to rewrite the Polyakov action in terms of
a curved background metric $G_{\mu\nu}(X)$. We further generalize this by including other
massless background fields; namely a totally antisymmetric tensor field $B_{\mu\nu}(X)$ and
a scalar field $\phi$ known as the dilaton. So, even though these are stringy states, one
can treat them as independent fields that couple to the string. The Polyakov action
then becomes
\begin{equation}
S_\sigma =\frac{1}{4\pi\alpha'}~ \int d\tau d\sigma ~\sqrt{\gamma} [(\gamma^{ab} G_{\mu\nu}+i\varepsilon^{ab}B_{\mu\nu}) (\partial_a X^\mu)
(\partial_b X^\nu) +\alpha' R \phi],
\end{equation}
this action is known as the sigma model \cite{polchinski1994} . In addition
to these fields, we will see in the next section that certain p-form gauge fields arise in
superstring theory. Again, these are really massless string states which we organize in
terms of classical background fields. In other words, the spectrum of massless states
in string theory is identical to that of a free field theory with such fields.
What about the free parameters in string theory ? Indeed , the string theory has no free parameters.
The strength of interactions is parameterized by the string coupling g which is a dimensionless number determined by the expectation value of the dilaton field. The Regge slope $\alpha'$ is dimensionful, but the string coupling g, together with the
Regge slope $\alpha'$, determined by the value of Newton’s constant and determines the tension of D-branes.
So that $\alpha'$ is not a free parameter and even can simply be absorbed in the definition of $X^\mu$.
\section{Superstrings}
The bosonic string theory that have been previewed so far contains important bosonic particles, such as the photon and the graviton. Non- Abelian gauge bosons, needed to transit the strong and weak forces , also arise in string theory as we will see later. However this theory suffers from many drawbacks. First, to be consistent and anomaly free, it has to live in 26 space-time dimensions. Second, the spectrum of both closed and open strings does not contain fermions. Third, the theory ground state is a tachyon. Tachyons are unphysical states because they assemble an instability of the vaccum. Indeed all these obstacles can be resolved on the way of making the theory a realistic one that describing the nature by including the fermions in its spectrum.
In string theory, the incorporation of fermions turns out to require symmetry farther than Poincar\'e symmetry. It requires supersymmetry, a symmetry that relates fermions and bosons, and the resulting theory is called superstring. In superstring theory, there is an equal number of fermions and bosons at every mass level. If supersymmetry exists in nature it must be spontaneously broken because we do not observe degeneracies between fermions and bosons.
Anyway supersymmetry is considerd one of most presumably scenarios beyond the Standard Model of particle physics.
To extend the Polyakov action to be supersymmetric invariant, the world sheet bosonic fields $X^\mu(\sigma,\tau)$ are paired up with internal degrees of freedom describing fermionic partener $\psi^\mu(\sigma,\tau)$ . The new fields $\psi^\mu(\sigma,\tau)$ are two component Majorana spinors and vectors under Lorentz rotations $SO(D-1,1)$ . The final worldsheet action is obtained by adding the standard Dirac of $D$ freely massless fermions to the free theory of $D$ massless bosons
\begin{equation} S= - \frac{1}{2\pi} \int d^2 \sigma (\partial_\alpha X^\mu \partial^\alpha X^\nu + \bar{\psi}^\mu \partial_\alpha \gamma^\alpha \psi^\nu ) \eta_{\mu\nu},
\label{susyA}
\end{equation}
here $\gamma^\alpha$ , with $\alpha =0,1$. In two dimensions, a particular simple choice for the $\gamma$- matrices is:
\begin{equation}
\gamma^0 = \sigma_2 , ~~~~~~ \gamma^1 = i \sigma_1,
\label{gamma}
\end{equation}
Dirac matrices obey the Clifford algebra $\{ \gamma_\alpha,\gamma_\beta \} =2 \eta_{\alpha\beta}$. The fermionic worldsheet $\psi^\mu$ are made of Grassmann variables , so they obey the anticommutation relations
\begin{equation} \{ \psi^\mu ,\psi^\nu \}= 0. \end{equation}
Since the Dirac matrices are purely real , Equ. (\ref{gamma}) is a Majorana representation, with eigenfunctions
\begin{equation}
\psi = \begin{pmatrix}
\psi_- \\
\psi_+
\end{pmatrix}.
\end{equation}
The Dirac conjugate of a spinor is $\bar{\psi} = \psi^\dagger \beta$, $\beta=i\gamma_0$, which for a Majorana spinor $\psi^T \beta$ . In this notation the fermionic part of the action is \cite{POLCHINSKI:2005, Dine2007}
\begin{equation}
S_f = \frac{i}{\pi} \int d^2 \sigma (\psi_- \partial_+ \psi_- + \psi_+ \partial_- \psi_+),
\label{sf}
\end{equation}
where $\psi_\pm = \frac{1}{2} (\partial_\tau\pm \partial_\sigma)$ refer to the worldsheet coordinates defined by $\sigma^\pm = \tau\pm \sigma$. The equation of motion for the two spinor components is the Dirac equation , which now take the form
\begin{equation}
\partial_+ \psi_- = 0, ~~~~~~~~~~~ \partial_- \psi_+ = 0,
\label{eom2}
\end{equation}
these equations describe left-moving and right-moving waves. Thus in two dimensions,
the fields $\psi_\pm$ are Majorana- Weyl spinors.
\section{Model expansions and types of superstring theories}
The possible boundary conditions for the fermionic fields $\psi^\mu$ are exactly the same as for the case of the bosonic string theory $X^\mu$ . The equations of motion (\ref{eom2}) guarantee that the action Equ. (\ref{sf}) is stationry . Requiring the vanishing of the action variation at the boundaries is fulfilled by admitting that
\begin{equation}
\begin{aligned}
\psi(\tau,\pi) &= + \psi (\tau, -\pi) ~~~~~ \mbox{Ramond boundary condition}\\
\psi(\tau,\pi) &= - \psi (\tau, -\pi) ~~~~~ \mbox{Neveu-Schwarz boundary condition}.
\end{aligned}
\end{equation}
So as the superstring moves in space , if we considered it moves through an interval $\sigma \in [-\pi,\pi]$ \footnote{\em Equivalently $\sigma \in [0,2\pi]$ in the sense of closed string .}, the full superstring state space breaks into two subspaces or, two sectors : a Ramond (R) sector which contains periodic fermions , and
Neveu–Schwarz (NS) sector which contains antiperiodic fermions.
Correspondingly, the fields mode expansion:
$\bullet$ For open strings: In the R sector
\begin{equation}
\psi_-^\mu(\tau,\sigma) = \frac{1}{\sqrt{2}} \sum_{n\in \mathbb{Z}} d^\mu_n e^{-in(\tau-\sigma)} , ~~~~~~~~~ \psi_+^\mu(\tau,\sigma) = \frac{1}{\sqrt{2}} \sum_{n\in \mathbb{Z}} d^\mu_n e^{-in(\tau+\sigma)}
\end{equation}
In the NS sector
\begin{equation}
\psi_-^\mu(\tau,\sigma) = \frac{1}{\sqrt{2}} \sum_{r\in \mathbb{Z}+ 1/2} b^\mu_r e^{-i r (\tau-\sigma)} , ~~~~~~~~~ \psi_+^\mu(\tau,\sigma) = \frac{1}{\sqrt{2}} \sum_{r\in \mathbb{Z}+1/2} b^\mu_r e^{-ir(\tau+\sigma)}
\end{equation}
The Majorana condition requires these expansions to be real, and hence $d^\mu_{-n}= d^{\mu \dagger}_{n}$.
$\bullet$ For closed strings: In the R sector
\begin{equation}
\psi_-^\mu(\tau,\sigma) = \sum_{n\in \mathbb{Z}} d^\mu_n e^{-2 i n (\tau-\sigma)} , ~~~~~~~~~ \psi_+^\mu(\tau,\sigma) = \sum_{n\in \mathbb{Z}} \tilde{d}^\mu_n e^{- 2 i n (\tau+\sigma)},
\end{equation}
In the NS sector
\begin{equation}
\psi_-^\mu(\tau,\sigma) = \sum_{r\in \mathbb{Z}+ 1/2} b^\mu_r e^{-2 i r (\tau-\sigma)} , ~~~~~~~~~ \psi_+^\mu(\tau,\sigma) = \sum_{r\in \mathbb{Z}+1/2} \tilde{b}^\mu_r e^{-2 i r (\tau+\sigma)}.
\end{equation}
So that an open superstring has two sectors (NS and R), while closed
string sectors corresponding to the different pairings of the left- and right-movers there are four distinct closed-string sectors: (NS, NS), (NS, R), (R, NS), (R,R).
In closed superstring theories spacetime bosons arise from the (NS,NS) sector and also from the (R,R) sector, since this sector is “doubly” fermionic. The spacetime fermions arise from the (NS,R) and (R, NS)
sectors. In order to get a closed string theory with supersymmetry we must truncate the four sectors above. A consistent truncation arises if we use truncated left and right sectors to begin with \cite{Zwi:2009}. Suppose we take, for example,
\begin{equation}
\mbox{left sector}:~~
\begin{Bmatrix}
\mbox{NS+}& \\
\mbox{R-}&
\end{Bmatrix}, ~~~~~~~~
\mbox{right sector}:~~
\begin{Bmatrix}
\mbox{NS+}& \\
\mbox{R+}&
\end{Bmatrix}.
\end{equation}
Combining these sectors multiplicatively we find the four sectors of the type \RNum{2}A superstring:
\begin{center}
Type \RNum{2}A : (NS+, NS+), (NS+,R+), (R-, NS+), (R-,R+).
\end{center}
The (NS+, NS+) sector is the one responsible for the graviton $G_{\mu\nu}$ ,
the antisymmetric $B_{\mu\nu}$ field (called Kalb- Ramond) and the dilaton $\phi$.
A different theory, a type \RNum{2}B superstring, arises if the chosen Ramond sectors are of the same type:
\begin{equation}
\mbox{left }:~~
\begin{Bmatrix}
\mbox{NS+}& \\
\mbox{R-}&
\end{Bmatrix}, ~~~~~~~~
\mbox{right sector}:~~
\begin{Bmatrix}
\mbox{NS+}& \\
\mbox{R-}&
\end{Bmatrix}.
\end{equation}
We then get
\begin{center}
Type \RNum{2}B : (NS+, NS+), (NS+,R-), (R-, NS+), (R-,R-).
\end{center}
While the (NS+, NS+) bosons of type \RNum{2}A and type \RNum{2}B theories are the same
\footnote{\em This sector is universal in all five superstring theories.}, the R-R bosons are rather different:
\begin{center}
R–R massless fields in type \RNum{2}A : $A_\mu, A_{\mu\nu\rho},$
\end{center}
\begin{center}
R–R massless fields in type \RNum{2}B: $A$, $A_{\mu\nu}$, $A_{\mu\nu\rho\sigma}$ ,
\end{center}
where $A_\mu$ is Maxwell field, $A_{\mu\nu\rho}$ is a 3-form gauge field, $A$ is a scalar field and $A_{\mu\nu}$ is a Kalb–Ramond field. The above R–R fields are deeply related to the existence of stable
solutions of string theory, known as branes. The index \RNum{2} refers to the theories being
$\mathcal{N}=2$ supersymmetric.
Actually the previous spectrum truncating (or projecting) is made in a very specific way that eliminates tachyons
appearance and leads to a supersymmetric theory in ten-dimensional space-time
\footnote{ \em This projection is called the
GSO projection, since it was introduced by Gliozzi, Scherk and Olive. For more details look for example at \cite{Katrin}.}.
In addition to the type \RNum{2} theories, there are also two heterotic superstring theories. They
are also closed string theories, but while a type \RNum{2} closed superstring arises by combining
together left moving and right moving copies of open superstrings, in the heterotic
string we combine a left moving open bosonic string with a right moving open superstring.
Out of the 26 left moving bosonic coordinates of the bosonic fields only ten of them are
matched by the right moving bosonic coordinates of the superstrings. As a result, this
theory effectively lives in ten-dimensional spacetime.
Heterotic strings come in two versions:
$E_8 \times E_8$ type and $SO(32)$ type. These labels characterize the groups of symmetries
that exist in the theories. $E_8$ is a group, in fact, it is the largest exceptional group (the E is
for exceptional). The group $SO(32)$ is the group generated by 32-by-32 matrices that are
orthogonal and have unit determinant.
Finally, in addition to both type \RNum{2} and heterotic theories, there is the type \RNum{1} theory.
This is a supersymmetric theory of open and closed unoriented strings. A string theory is
unoriented, if the states of the theory are invariant under
an operation that reverses the orientation of the strings. Whereas type \RNum{2} theories and the
heterotic theories are theories of oriented closed strings.
Thus the complete list of ten-dimensional supersymmetric string theories is therefore
\begin{itemize}
\item Type \RNum{2} A,
\item Type \RNum{2} B,
\item $E_8 \times E_8$ heterotic,
\item $SO(32)$ heterotic,
\item Type \RNum{1}.
\end{itemize}
These five theories have all been known since the middle of the 1980s. Some relationships (known as dualities )
between them were found soon after their discovery, but a clearer picture emerged
only in the late 1990s. The limit of type \RNum{2}A theory as the string coupling is taken to infinity
was shown to give a theory in eleven dimensions. This theory is called M-theory.
\footnote{\em With the meaning of M to be decided when the nature of the theory becomes clear. Also sometimes it is refered by
the mother of all string theories.}
It is known, however, that M-theory is not a string theory. M-theory contains membranes (2-branes) and
5-branes. However, the discovery of many other relationships between the five
string theories and M-theory has made it clear that we really have just one
theory. In conclusion: M-theory is a unique theory, and the five superstrings and
are different limits of this theory. For now days the edges of this physics subject have not been reached yet.
\section{Supergravity}
The first development of supergravity theory was in the seventies and constructed quite separately from string theory.
The $\mathcal{N}=1$ SUGRA in four spacetime dimensions as the minimal possible supersymmetric extension of GR. Today it is well known supergravity theories are the low
energy limits of string theory as we will show later. Basically one expands the string action in $\alpha'$ and drops
all terms of $\mathcal{O}(\alpha'^2)$ and higher. Yet that is was not of course the case when SUGRA has been
discovered. Historically, it was expected since in $\mathcal{N}=1~ D=4$ SUGRA there is more symmetry than in pure gravity,
the high–energy (short distance) behavior will improve. Although this is true, still the (super)symmetry is not enough to
cancel all divergences in the theory. That led to extend the supersymmetry itself for ($\mathcal{N} > 1$) in higher dimensions ($D > 4$) to get for the time being vast realm of supergravity theories \cite{theME}.
However $\mathcal{N}=1~ D=11$ has earned its salient features because it is a unique theory, and its soltions are the solutions of M-theory as well. In the following we will present a whole figure about SUGRA, then we will speak solely about $\mathcal{N}$=1~ D=11 SUGRA, starting from how its action is built, a brief review about Calab- Yau 3 fold, and then how it is reduced over CY to
$\mathcal{N}=2~ D=5$ SUGRA, the theory that we introduce solutions for it after that.
Generally speaking there are several ways to build a supergravity theory:
\begin{itemize}
\item By Brute force. By supersymmetrizing the Einstein-Hilbert action, i.e. adding fermionc degrees of freedom (D.O.F.) as the superpartner of the graviton $g_{\mu\nu}$, then finding out the SUSY transformations that keep the new action Lorentz invariant. Unfortunately, on quantization this theory is not renormalizable exactly as GR. Hence this way is not so accurate or, at best, the theory is low energy effective limit of a better idea, which is indeed what SUGRA turned out to be.
\item To begin with a supersymmetric Yang-Mills action without gravity and then make the supersymmetric
transformations local, i.e. become spacetime dependent. This directly implies the existence of gravity in the theory.
In te next section we will give more details about this procedure, but keeping in mind this minimal SUGRA is not perturbative.
\item Pertubative SUGRA in ten spacetime dimensions is a low energy limit of perturbative superstring theories in $D=10$. As illustrated in Fig. (\ref{Mom}), type \RNum{2}A string theory reduces to type \RNum{2}A SUGRA, type \RNum{2}B string theory reduces to type \RNum {2}B SUGRA, type \RNum{1} string theory reduces to type \RNum{1} SUGRA and the heterotic string theories both reduce to type \RNum{1} SUGRA.
\item Different SUGRA models in lower spacetime dimensions is obtained form the compactifiction of higher dimensional SUGRA models over various types of manifolds. Indeed this way is very important because all the trials now to detect any signal for the higher dimensional theories within our four dimensions world. we will talk bout the dimensional reduction in next pages.
\end{itemize}
\subsection{Generating SUGRA from SUSY}
In general new fields should be added when a global symmetry becomes local. Consider the gauge symmetry, for example the Dirac action \cite{Freedman1976, Cerdeno:1998hs}
\begin{equation} S= i \int d^4 x ~ \bar{\psi} \gamma^\mu \partial_\mu \psi \end{equation}
is invariant under the global symmetry $\psi'= e^{i\alpha} \psi$. If we make this symmetry local $\psi' = e^{i\alpha(x)} \psi$, the action becomes
\begin{equation} S'= i \int d^4 x ~ e^{i\zeta(x)} \bar{\psi} \gamma^\mu \partial_\mu (e^{-i\zeta(x)} \psi), \end{equation}
\begin{equation} S'= S+
i \int d^4 x \bar{\psi} \gamma^\mu \psi (-i) (\partial_\mu \zeta), \end{equation}
adding a new term:
\begin{equation}
i A_\mu \bar{\psi} \gamma^\mu \psi e^{i\zeta(x)},
\end{equation}
now;
\begin{equation} S'= S+
i \int d^4 x \bar{\psi} \gamma^\mu \psi (A_\mu +\partial_\mu \zeta) e^{i\zeta(x)}, \end{equation}
\begin{equation} A'_\mu=A_\mu+\partial_\mu \zeta. \end{equation}
Second example, consider the action of a scalar field in flat space- time
\begin{equation}
S= \int d^4 x \Big[ \frac{1}{2} \eta^{\mu\nu} \partial_\mu \phi \partial_\nu \phi - V(\phi) \Big],
\end{equation}
where $\eta^{\mu\nu}$ is the Minkowski metric. This action is invariant under Poincar\'e symmetry
\begin{equation}
x^\mu \to \Lambda^\mu_\nu x^\nu + a^\mu,
\end{equation}
to make this transformation local, we replace the Lorentz four vector $\mathbf{\Lambda}$ by "some thing" depends on space- time $\Lambda^\mu_\nu=\frac{\partial x^\mu}{\partial x^\nu}$ which is the generalized coordinate transformations in special relativity. Now to keep the action invariant we have to change to
\begin{equation}
S= \int d^4 x \sqrt{-|g|} ~\Big[ \frac{1}{2} g^{\mu\nu} \partial_\mu \phi \partial_\nu \phi - V(\phi) + \mathcal{R} \Big] ,
\end{equation}
where $g^{\mu\nu}$ is the metric. From the view point of the particle physics it is a spin- 2 particle called the graviton. $\mathcal{R} $ is the Ricci scalar. In the case of supersymmetry, consider the free Wess–Zumino Lagrangian
\begin{equation} \mathcal{L} = \partial_\mu \phi^* \partial^\mu\phi + \chi^\dagger i \bar{\sigma}^\mu (\partial_\mu \chi). \end{equation}
The corresponding action is invariant under the SUSY transformations
\begin{equation} \delta \phi = \zeta. \chi= \zeta^a \chi_a= \zeta^\dagger (-i \sigma_2) \chi ,~~~~~~~~ \delta \chi=-i \sigma^\mu (i\sigma_2 \zeta^*) \partial_\mu \phi, \end{equation}
together with their hermitian conjugates. When $\zeta \to \zeta(x)$
\begin{equation} \delta \mathcal{L} = \mbox{terms seen before} + \chi^\dagger (i\sigma_2) (\partial^\mu \phi ) (\partial_\mu \eta^*)
- \chi^\dagger~ i\bar{\sigma}^\mu~ i \sigma^\nu~ (\partial_\mu \phi) (\partial_\nu \bar{\chi}), \end{equation}
where
\begin{equation} \delta \mathcal{L}_{\mbox{Extra}} \sim K_\mu (\partial^\mu \zeta). \end{equation}
To cancel these extra terms, we add to the Lagrangian
\begin{equation} \mathcal{L}_{\mbox{Extra}}= K_\mu \Psi^\mu_a,\end{equation}
provided that:
\begin{equation} \Psi'^\mu = \Psi^\mu + \partial^\mu \zeta,\end{equation}
$\Psi^\mu$ is a spin $\frac{3}{2}$ particle called the gravitino. So that the Lagrangian becomes
\begin{equation} \mathcal{L} = \partial_\mu \phi^* \partial^\mu\phi - \frac{1}{2} \bar{\psi} \gamma^\mu \partial_\mu \psi +K_\mu \Psi^\mu, \end{equation}
where in the second term we used Dirac 4-spinor formula. The variation of $\Psi^\mu $ cancels $\delta \mathcal{L}_{\mbox{Extra}}$. However it has been found that there are still extra terms
\begin{equation}
\delta \mathcal{L} \sim - K_\mu (\partial^\mu \zeta) - \bar{\Psi} \gamma^\nu \zeta(x) \left( (\partial_\mu \phi^*) (\partial_\nu \phi) + \frac{1}{2} \bar{\psi} \gamma_\mu \gamma_\nu \psi \right).
\end{equation}
The last two terms are defined as the energy-momentum tensor which is given by
\begin{equation}
T_{\mu\nu} = \frac{1}{\sqrt{-|g|}} \frac{\partial \mathcal{L}}{\partial \eta^{\mu\nu}},
\end{equation}
so to be canceled , another extra term is added which might be now contracted with $T_{\mu\nu}$
\begin{equation} \mathcal{L}_{Extra~\RNum{2}} \sim g_{\mu\nu} T^{\mu\nu}, \end{equation}
provided that
\begin{equation} \delta g_{\mu\nu} = \bar{\psi}_\mu \gamma_\nu \zeta(x), \end{equation}
where $g_{\mu\nu}$ is the graviton. In $\mathcal{N}=1$ supersymmetry the gravitino $\Psi_\mu$ is the superpartener of the graviton.
Some thing can be concluded when gravity arises naturally in a locally supersymmetric theory. That is
supersymmetry is not just an arbitrary restriction on nature, but rather a maximal symmetry of spacetime with far reaching implications.
\subsection{SUGRA in $11$- dimensions}
This theory is the most general SUGRA and it is completely unique if we required that we can not never have particles have higher spin more than 2. Then 11 dimension is the highest possible number of spacetime dimension that we can construct supergravity theory \cite{freedman:2012}.
Note that, since D=11 is the highest dimension where supersymmetry can live, it was expected when D=11 SUGRA dimensionaly reduced, SUSY and Einstein's relativity are obtained. But unfortunately this is not the case. At dimensional reduction, there are many extra fields produced not exist in the original theories, that makes the theoreticians till nowadays keep wondering what is the right relation between D=11 SUGRA and SUSY and GR.
Consider for instance D=3 $\mathcal{N}$=1 supergravity, on shell, $e^a_\mu$ has $\Big (\frac{(d-1)(d-2)}{2}-1 \Big)$ D.O.F. and $\psi^a_\mu$ has $ \Big ( (d-3) 2^{\frac{d-1}{2}}/2 \Big) $. So on shell $\{e^a_\mu,\psi^a_\mu\}$ has no D.O.F. Off-shell however, we have for $e^a_\mu$, $\frac{d(d-1)}{4}=3$ D.O.F. and for $\psi^a_\mu$ , $(d-1) 2^{\frac{d-1}{2}}=4$. So we need to add an auxiliary scalar S \cite{Intro1}.
Now consider D=11 $\mathcal{N}$=1 supegravity. On sell $e^{\hat{\alpha}}_\mu$ has 44 D.O.F. and $\psi^a_\mu$ has 128 D.O.F. ($\mu=0,...,10$). So we need extra (84) bosonic D.O.F. That will be a 3-form gauge field
\begin{equation} A_3=\frac{1}{3!}~ A_{\mu\nu\rho}~ dx^\mu \wedge dx^\nu \wedge dx^\rho. \end{equation}
The field strength will be a 4-form defined by
\begin{equation}
F=dA=\frac{1}{4!}~ F_{\mu\nu\rho\lambda}~ dx^\mu \wedge dx^\nu \wedge dx^\rho \wedge dx^\lambda,
\end{equation}
where
\begin{equation}
F_{\mu\nu\rho\lambda} = \partial_\mu A_{\nu\rho\lambda} - \partial_\nu A_{\mu\rho\lambda} + \partial_\rho A_{\mu\nu\lambda}
- \partial_\lambda A_{\mu\nu\rho}.
\end{equation}
The Lagrangian
\begin{equation}
2~ k_{11}^2~ \mathcal{L} = |e|~ \mathcal{R} -2~ i ~|e|~ \bar{\psi}_\mu \gamma^{\mu\nu\rho} (\bar{\nabla}_\nu \psi_\rho)- \frac{1}{2}~ \frac{|e|}{4!}~
F_{\mu\nu\rho\lambda} F^{\mu\nu\rho\lambda} + \mbox{additional Ferminoic terms}, \label{11lag}
\end{equation}
where the greek symbols run over $0,...,10$, $|e|=\sqrt{-|g|}$, $g= \mbox{det}~ g_{\mu\nu}$ , $\gamma^{\mu\nu\rho}\sim \gamma^{[\mu\nu}~~ \gamma^{\rho]}$, $\gamma^{\mu\nu}\sim[\gamma^\mu,\gamma^\nu]$, and $\gamma^\mu=e^\mu_{\hat{\alpha}} \gamma^{\hat{\alpha}}$. The Lagrngian is invariant under the following transformations:
\begin{equation}
\delta e^{\hat{\alpha}}_\mu = i \bar{\zeta} \gamma^{\hat{\alpha}} \psi_\mu,
\end{equation}
\begin{equation}
\delta \psi_\mu = \bar{\nabla}_\mu \zeta.
\end{equation}
To find out the transformation of $A_{\mu\nu\rho}$ which keep the Lagrangian invariant, we should add extra term
$\frac{1}{6~ 3! (4!)^2}~ \varepsilon_{\mu_1,..\mu_{11}} A^{\mu_1,..\mu_3} F^{\mu_4,..\mu_7} F^{\mu_7,..\mu_{11}}$. Focusing only on the bosonic terms, and write the Lagrangian in the vielbein formula \footnote{See Appendix (\ref{Df}).}:
\begin{equation}
2~ k_{11}^2~ \mathcal{L} = |e|~ \Big[ \mathcal{R} - \frac{1}{48} F^2\Big] - \frac{1}{6} A \wedge F \wedge F ,
\label{Lag111}
\end{equation}
which now is invariant under
\begin{equation}
\delta e^{\hat{\alpha}}_\mu =- \frac{i}{2} \bar{\zeta} \gamma^{\hat{\alpha}} \psi_\mu,
\end{equation}
\begin{equation}
\delta \psi_\mu = 2 \bar{\nabla}_\mu \zeta + \frac{1}{144} F_{\nu\rho\lambda\sigma} \Big[ \gamma_\mu^{\nu\rho\lambda\sigma}-
8 \delta^\nu_\mu \gamma^{\rho\lambda\sigma}\Big]\zeta,
\end{equation}
\begin{equation}
\delta A_{\mu\nu\rho} = \frac{3}{2} \bar{\zeta} \gamma_{[\mu\nu} ~ \psi_{\rho]}.
\end{equation}
The next step to write the equation of motions of the gauge field strength . Returning to Maxwell's equations in 4- dimensions, $\partial_\mu F^{\mu\nu}= j^\nu$. They can be written as $ d * F = J$, which gives Gauss’s law and Ampere’s law and
$ d F =0$, Bianchi identity that gives Faraday's law. Similarly in 11-dimensions, the source free E.O.M of the 4- form field strength are
\begin{equation}
\begin{aligned}
d * F + \frac{1}{2} ~ F \wedge F =0 ,
~~~~~~d F =0.
\end{aligned}
\end{equation}
If there is a source
\begin{equation}
\begin{aligned}
\int (d * F + \frac{1}{2} F \wedge F) = Q_E,
~~~~\int d F = Q_M,
\end{aligned}
\end{equation}
where $Q_E$ and $Q_B$ are constants and calling them with these names has nothing to do with the electric and the magnetic charges, but that’s just an analogy with the Maxwell’s equations.
The question now, in eleven dimensions, what are these constants ? and what is the source of these charges ?
For instance , in Maxwell’s equations , the source of electric charges is the electrons. Can we have here point- like particles with $Q_E$ and $Q_M$?
In 4- dimensions: ~~~~~~ $\mathcal{L}_{int} \propto A_\mu J^\mu, $
where $J^\mu$ is the current of a point- like particle and the index $\mu$ describes the motion of a particle that has a world line in 4 dimensions space- time. However
In 11- dimensions : ~~~~~ \begin{equation} \nonumber \mathcal{L}_{int} \propto A_{\mu\nu\rho} J^{\mu\nu\rho}, \end{equation}
$J^{\mu\nu\rho}$ describes a hyperplane ( some thing with higher dimensions ) propagates in time, but it has two space dimensions , i.e., it’s a 3-dimensional hyperplane , two spatial propagate in time , imagine it as “ a paper”.
It’s a generalization of the world line and it’s called a supermembrane or (2Brane). Two here refers it’s a two dimensional brane propagating in time .
When the 2Brane lives in 11- dimensions is wrapped a round 1- dimension of radius $R$, then $R \to 0$
limit is taken, this mean we go from 11-dimensions lower to 10-dimensions . This’s called klein kaluza compactification . Indeed it has been found that superstrings in 10- dimensions come from membranes in 11- dimensions when wrapped around a circle . The membrane has an electric $Q_E$, exactly like a point particle in 4- dimensions.
So we can see how the extra term added to the Lagrangian that has a gauge field is strongly related to the presence of a super 2Brane in 11- dimensions. So what do we have now?
a gravitational object carries an electric charge ( you can call the coupling constant of Lagrangian term that has $A$ ), has a mass and propagates gravitationaly in space- time ( so it can be described by a metric ).
These discoveries were incredible in supergravity, because it were discovered separately in superstigns in the same time.
Turns now to the magnetic charges in 11- dimensions. The first question we start by, do the electric and magnetic charges couple to the same branes or no ?
In Maxwell’s equations $Q_E=\int d * F$ which means $Q_E$ is related to $d*F$, while $Q_M=\int d F$ which means $Q_M$ is related to $d F$. $F$ is a two form, while $* F$ is a $(D-2)$ form, i.e., $*F $ is a two form . So that in Electromagnetism (EM) $F_{\mu\nu}$ is a $4 \times 4$ matrix, and its dual $F^{\mu\nu}$ is a $4 \times 4 $ too, so in 4- dimensions the point particle can carry a magnetic and an electric charges .
In 11- dimensions. however $F$ is a 4 form , while $* F$ is a 7- form. $*F$ has a seven indices $F^{\mu\nu\rho\lambda\sigma\delta\zeta}$ and it’s the field strength of a six form gauge field, $F=d A$, that is responsible for the magnetic charge. And as the 3- form gauge field couples to two dimensional brane (called M2-brane) , the 6- form brane couples to 5- dimensional brane ( 5- dimensional surface propagating in time that’s called M5- brane).
M2-brane and M5-brand are considered the simplest solutions of supergravity in 11- dimensions, and there are proofs that you can not construct any other branes dimensions. However a combinations of these two branes can be constructed because they are gravitational objects, hence there can be multiple branes. But any ways M2 and M5 are the fundamental branes. When the M5- brane is wrapped over a 1- dimension, we get a 4- dimensions brane just like when we wrap the M2-brane over 1- dimension. and get a string, and it is found there are 4- dimensions branes in string theory. Indeed that is the goal from the beginning to start by the most possible high dimension and then compictify to lower dimensions and see what we will have. So as bold lines, there are no point particles in 11- dimensions, they can not exist , you can consider Maxwell’s equations in EM as a special case of higher gauge field theories, and finally M-branes ( M2- M5) solutions of supergravity are non perturbative.
The anti-commutations relations of the SUSY generators now become
\begin{equation}
\{Q,\bar{Q}\} = \gamma^\mu P_\mu + \gamma^{\mu\nu} Z_{\mu\nu} + \gamma^{\mu\nu\rho\lambda\sigma} Z_{\mu\nu\rho\lambda\sigma}.
\end{equation}
It has been found that these ``Zs`` are topological charges of the M-branes, $Z_{\mu\nu}$ is the topological charge of M2-brane and $Z_{\mu\nu\rho\lambda\sigma}$ is the topological charge of M5-brane.
\subsection{Branes in string theory}
To manifest that the origin of branes in string theory is completely different than in supergravity.
As previously pointed out in string theory branes raised to conserve momentum at the end points boundary
conditions for the open strings. So open strings always bounded to dynamical objects that have extended dimensions, i.e.,
strings end points always attached to some hyperplanes which is called p-branes. Where if $p=0$ it is a point particle, if $p=1$ it is a string, if $p=2$ it’s a membrane and so on.
However, although branes in string theory are in 10- dimensions and branes in supergravity are in 11- dimensions , theoreticians found how they relate to each other via Kaluza Klein reduction as will be mentioned next.
\subsection{Branes as solution for SUGRA}
The question now how these branes are written as solution for SUGRA?
We have admitted before that they are gravitational objects, i.e., they have an energy density and so they gravitate in addition they have a coupling with the form gauge \cite{theME}. Let us write M2-brane. First of all as these branes exist "if they exist" , they break SUSY because they define a preferred direction of space. So Poincar\'e symmetry is broken, i.e., supersymmetry is broken. In the case of M2-brane , it breaks the space into two pieces , i.e., M2-brane breaks SUSY to two halfs.
The simplest form of M2-branes happens when the fermions vanish, and we get pure bosonic solutions.
The M2-brane metric is then
\begin{equation}
ds^2_{11}= ds^2_{M2} + ds^2_{\mbox{transverse}},
\end{equation}
it is 11- dimensional metric, but 11- dimensions written in two terms , a term explains the world volume of the brane and the other term explains the remaining dimensions. So $ds^2_{\mbox{transverse}}$ is the metric of all dimensions that perpendicular to my brane. Writing general matrices for each piece
\begin{equation}
ds^2_{11}= g_{ab}~ dx^a~ dx^b + g_{\mu\nu}~ dx^\mu~ dx^\nu,
\label{M2}
\end{equation}
where $a,b=0,1,2$ and $\mu,\nu=3,...,10$. Take into account that we are assuming a symmetry in space, like a rotation symmetry for these dimensions. For instance , I need to describe the 8- dimensions by $SO(8)$ , so the metrices $g_{ab}$ and $g_{\mu\nu}$ most likely will depend on a radius r where $r^2 = x^2_3+ ...+ x^2_{10}$ is a radial distance in 8-dimensions.
Take all of these and apply on the equations of motion, and in the Lagrangian Equ.(\ref{Lag111}) that yielded from
Einstein field equations (EFE) \footnote{ \em Appendix (\ref{GR}).} including the Energy-moment tensor. Also $\delta \psi_\mu=0$ can be checked (there is a theorem states that this condition implies EFE). At the end we get the metric of M2-brane as
\begin{equation}
ds^2_{11} = H^{-3/2}(r) ~\eta_{ab}~ dx^a dx^b + H^{3/2}(r)~ \delta_{\mu\nu} ~dx^\mu dx^\nu,
\label{M232}
\end{equation}
where $H(r)$ is a radial function , $\eta_{ab}$ is the Minkowski metric. $H^{3/2}$ is a radial direction of the transverse space. Remember that the M2-brane has an electric charge under the gauge field A
\begin{equation}
A_{012}= \pm H^{-1},
\end{equation}
or in general
\begin{equation}
A_{\mu\nu\rho}= \pm \varepsilon{\mu\nu\rho} H^{-1}.
\end{equation}
From EFE, H(r) satisfies the Laplace equation
\begin{equation}
\nabla_8^2 H=0 ~~\to \delta^{\mu\nu} \partial_\mu \partial_\nu H=0,
\end{equation}
that yields
\begin{equation}
H= 1 +\frac{M_2}{r^6}, \label{GP}
\end{equation}
this is the equation of the gravitational potential. Where $M_2$ is an arbitrary integration constant proportional to the energy density of the brane. To figure out this formula of $H$, Equ. (\ref{GP}), let us return to the Laplace equation in 3- dimensions $\nabla^2 \phi =0$ with the spherically symmetric solution $\phi \sim 1 + \frac{q}{r}$, with the Electric field $E\sim \frac{1}{r^2}$. So we say the point particle is surrounded by a two dimensional spherical charge with volume $r^3$ (Gauss law).
Now in 11- dimensions what is the surface should surround the M2-brane ? it’s a surface having a volume $\sim r^8$, and an Electric field $\frac{1}{r^7}$ and so that the potential $\propto \frac{1}{r^6}$. Another case, if you make similar calculus for instance for a black hole (BH) , you will get $H \sim \frac{1}{r}$ ( similar for Schwarzschild radius).
\bigskip
\textbf{M5-brane}:
In the same manner, let us construct a metric that has a symmetry in the transverse directions
\begin{equation}
dS^2_{11} = g_{ab}~ dx^a dx^b + g_{\mu\nu}~ dx^\mu dx^\nu,
\end{equation}
where $a,b=(0,1,2,3,4,5)$ and $\mu,\nu=(6,7,8,9,10)$.
\begin{equation}
ds^2_{11} = H^{-3/2} ~ \eta_{ab} ~ dx^a dx^b + H^{3/2} ~ \eta_{\mu\nu} ~ dx^\mu dx^\nu ,
\label{M5}
\end{equation}
where the first part is the world volume of the M5-brane and the second part is transverse direction. The 6- form gauge filed is given by
\begin{equation}
A_{012345} = \pm H^{-1}(r),
\end{equation}
where $H(r)$ is given by $\nabla^2_5 H(r)=0 ~ \to ~ H(r)= 1 + \frac{M_5}{r^3}.$
Now to figure out what are $M_2$ and $M_5$ and how they are related to the "electric" and the "magnetic" charges , literally the 3- form and the 6- form gauge fields, let us recall the familiar GR in 4- dimensions with an electric field. So we will have the Einstein equation with the Maxwell’s equations as
\begin{equation}
\mathcal{L} = |e| [ \mathcal{R} - \frac{1}{4}~ F_{\mu\nu} F^{\mu\nu} ] ,
\end{equation}
take the Reissner–Nordström solution which represents electrically charged black hole with a metric
\begin{equation}
ds^2= - f(r) dt^2 + f^{-1} (r) dr^2+ r^2 d \Omega^2,
\end{equation}
where $d\Omega^2= d\theta^2 + \sin^2 \theta d\phi^2$ and
\begin{equation}
f(r) = 1 - \frac{2M}{r} + \frac{Q^2}{r^2} ,
\end{equation}
where Q is the ordinary electric charge of the black hole. Note that can be the Schwarzschild black hole solution except for the last term which has Q. Here the event horizon of the black hole depends on two parameters, it’s charge as well as its mass, so simply if $Q \to 0 $ we will get Schwarzschild BH. The electric field specified by $A_0 = \frac{Q}{r}$ ( just a Coulomb field). To get the BH’s event horizon we set $f(r)=0$ and solve for r and since the equation is quadratic in $r$, we get two event horizons
\begin{equation}
r_\pm = M \pm \sqrt{M^2 - Q^2} ,
\end{equation}
which means we have certain conditions.
Case : $M=Q \to $ there is a single horizon (maximal case) \\
Case: $M^2 < |Q| \to $ no horizon ( a naked singularity).
There are hypothesis that called "cosmic censorship" states that the naked singularities are not allowed by nature . So whatever a nature try to make a naked singularity , it doesn't work. So physically speaking, M should be greater than $|Q|$, $M \geq |Q|$. This’s called BPS (Bogomol’nyi-Prasad-Sommerfield) condition and when $M=|Q|$ it’s called saturated BPS condition.
Now let us rewrite the RN solution at the maximal case $M=|Q|$, change the coordinates $r \to (r+M)$, then the metric becomes
\begin{equation}
ds^2_{M=|Q|} = -H^{-2} dt^2 + H^2 \delta_{ij}~ dx^i dx^j ,
\end{equation}
where $i,j=1,2,3$ and $H=1+\frac{M}{r}$ as a solution of $\nabla^2_3 H(r)=0$. And the gauge field $A_0 = H^{-1}$. Now notice the similarity between this case and the M2 or M5 branes case Equ. (\ref{M232}), and Equ. (\ref{M5}). Therefore M2 and M5 are interpreted as generalized Maxwell's BPS black holes (or black branes) with one event horizon . They behave as just multi-dimensions black holes.
Also we can add more branes
\begin{equation}
H = 1 + \sum_{i=1}^k \frac{M_i}{r_i},
\end{equation}
so we have $k$ 2M-branes at $r=({\textbf r}-{\textbf r_I})$. This is a stable configuration, i.e., these BHs are static, they do not move because the electric repulsion exactly cancels the gravitational attraction. But this is only when the BPS condition is realized, so if BPS didn't saturated, there will be instability , the BHs start to move.
\section{Compactificaton; An overview}
\label{compact}
Consider Kaluza- Klein compactificaton. It is have been made in 1922, that there was a hope to unify GR and EM. It assumed that we live in a 5- dimensional space- time with only gravity.
But what about the fifth dimension that we can not see. It was assumed that this dimension is very small, it can be wrapped it over a circle
\begin{equation}
0 \leq x_5 \leq 2\pi R,
\end{equation}
and assume we live where R is very small, so we can divide the metric as
the extra 4 degrees of freedom will be a gauge field $A_\mu$ , that can be the gauge field of EM. But what about thee extra diagonal scalar term that doesn't belong to the GR nor the EM. Nowadays this scalar is called the dilaton. But this model has been ruled out because it predicted a scalar field which was not known at these times.
However upon the discovery of D=11 supergravity and trying to dimensionaly reduced, Kaluza- Klein reduction has been bring to light again. But every time we go lower in dimension , we get extra fields. For instance down to 10- dimensions, the 11- dimensions metric will be
divided as:
\begin{equation}
ds^2_{11} = (e^{2\alpha} g_{\mu\nu} + e^{2\alpha} A_\mu A_\nu ) dx^\mu dx^\nu + 2 e^{2\alpha} A_\mu dx^\mu~ dy+ e^{2\alpha} dy^2,
\end{equation}
where $\mu,\nu=0,...,9$, the first term is the solution of the EFE in 10 dimensions. So the graviton breaks down to a new gauge field $A_\mu$ and the dilaton $e^{2\phi}$. The 11- dimensions gauge field and the field strength also break down
\begin{equation}
A_{LMN}^{(3)} \to A_{\mu\nu\rho}^{(3)} + A_{\mu\nu y} \equiv B_{\mu\nu}^{(2)},
\end{equation}
\begin{equation}
F_{LMNP} \to F_{\mu\nu\rho\lambda} + G_{\mu\nu\kappa} ~~~ (G=dB),
\end{equation}
where $y$ is the extra dimension we wrapping $0 \leq y \leq 2\pi R$. Also the spinors will split into two. So we get $\mathcal{N}=2$ supergravity in 10- dimensions, which called type \RNum{2}A SUGRA.
This wrapping over a circle can be made many times down to 4- dimensions.
In each time we reduce a dimension, we add new field breaks SUSY into half's . So this way in reduction is not preferable because SUSY will be broken too much than we phenomenologicaly need.
We can make dimension reduction over manifolds. But what are the type of manifolds can be used to reduce from higher to lower dimensions while SUSY is preserved. For instance we can reduce over Calabi-Yau manifold to go from D=10 to D=4 with preserving SUSY. The 10- dimensions metric will then be
\begin{equation}
ds^2_{10}= g_{\mu\nu} ~ dx^\mu dx^\nu + g_{ab} ~ dx^a dx^b,
\end{equation}
where the first part is the 4- dimensions we want $\mu,\nu=0,...,3$ and the second part is the 6- dimensions manifold $a,b=4,...,10$ . There is also what is called G-Calibrated manifold which is 7- dimensional, so we can reduce over it form D=11 directly to D=4.
Calabi-Yau manifolds has many characteristics \footnote{\em As we will see in the next chapter.}, like they always have even dimensions, and usually written in complex number. For instance the 6- dimensions Calabi-Yau can be written as
\begin{eqnarray}
\nonumber z_1&=&x_1+ix_2, \\
\nonumber z_2&=&x_3+ix_4, \\
z_3 &=&x_5+ix_6,
\end{eqnarray}
so instead of 6- dimensions real manifold, now we have only 3- dimensions complex manifold. Anyhow , although the Calabi- Yau manifold's properties are well known, till these days no body knows what it looks like. That because Calabi- Yau manifold has to be closed manifold not opened and unfortunately no one could construct yet a closed Calabi- Yau manifold, but we know it just exist due to Yau's theorem, which proves the existence of a metric , even though we do not know yet how this metric looks like.
\section{Low energy supergravity }
Let us see how \RNum{2}A superstring theory arises from the dimensional reduction of the eleven-dimensional supergravity
\cite{POLCHINSKI:2005}
As shown before, the eleven-dimensional supergravity has two bosonic fields, the metric $G_{MN}$ and a 3-form potential $A_{MNP} \equiv A_3$ with field strength $F_4$. The bosonic part of the action is given by
\begin{equation}
2 \kappa_{11}^2 S_{11} = \int d^{11} x ~ ( - G )^{1/2} ~ \Big ( R - \frac{1}{2} |F_4|^2 \Big ) - \frac{1}{6} \int A_3 \wedge F_4 \wedge F_4,
\label{s11}
\end{equation}
now dimensionaly reduce . The general metric that is invariant under translations in the 10-direction is
\begin{eqnarray}
\nonumber ds^2 &= & G_{MN}^{11} (x^\mu ) ~ dx^M ~ d x^N\\
&=& G_{\mu\nu}^{10} (x^\mu ) ~ dx^\mu ~ d x^\nu + \text{exp} (2 \sigma (x^\mu )) [ dx^{10} + A_\nu (x^\mu) dx^\nu]^2 ,
\end{eqnarray}
where $M,~ N$ run from 0 to 11 and $\mu,\nu$ from 0 to 9. The eleven-dimensional metric reduces to a ten-dimensional metric, a gauge field $A_1$, and a scalar $\sigma$. The potential $A_3$ reduces to two $A_3$ and $A_2$ . The three terms in Equ.(\ref{s11}) become
\begin{eqnarray}
\nonumber S_1 &=& \frac{1}{2\kappa_{10}^2} \int ~ d^{10} x ~ (-G)^{1/2} ~\Big ( e^\sigma R - \frac{1}{2} e^{3\sigma} |F_2|^2 \Big ), \\ \nonumber
S_2 &=& - \frac{1}{4\kappa_{10}^2} \int ~ d^{10} x ~ (-G)^{1/2} ~\Big ( e^{-\sigma} |F_3|^2 + e^\sigma |\tilde{F_4}|^2 \Big) , \\
S_3 &=& - \frac{1}{4\kappa_{10}^2} \int A_2 \wedge F_4 \wedge F_4 =
- \frac{1}{4\kappa_{10}^2} \int A_3 \wedge F_3 \wedge F_4, \label{s10}
\end{eqnarray}
where
\begin{equation}
\tilde{F_4}= d A_3 - A_1 \wedge F_3,
\end{equation}
the fields of the reduced theory are the same as the bosonic fields of the \RNum{2}A string, as they must be. In particular the scalar $\sigma$ must be the dilaton $\phi$ up to some field redefinition. The terms in the action have a variety of $\sigma$- dependency.
Since we have arrived at the action Equ. ( \ref{s10}) without reference to string theory, we have no idea as yet how these fields are related to those in the world-sheet action. We will proceed by guesswork, and then explain the result in world-sheet terms. First redefine
\begin{equation}
G_{\mu\nu} = e^{-\sigma} G_{\mu\nu} ~ (\text{new}), ~~~ \sigma =\frac{2\phi}{3}, \label{dil}
\end{equation}
substitute by the new metric. Then
\begin{eqnarray}
\nonumber S_{\RNum{2}A} &=& S_{NS} + S_R + S_{CS} , \\ \nonumber
S_{NS} &=& \frac{1}{2\kappa_{10}^2} \int ~ d^{10} x ~ (-G)^{1/2}
e^{-2 \phi} ~\Big ( R - 4 \partial_\mu \Phi \partial^\mu \Phi - \frac{1}{2} |H_3|^2 \Big) , \\ \nonumber
S_R&=& - \frac{1}{4\kappa_{10}^2} \int ~ d^{10} x ~ (-G)^{1/2} ~\Big ( |F_2|^2 + |\tilde{F_4}|^2 \Big) , \\
S_{CS} &=& - \frac{1}{4\kappa_{10}^2} \int B_2 \wedge F_4 \wedge F_4 , \label{ss}
\end{eqnarray}
note that $R \to e^\sigma R + ....$, that $(-G)^{1/2} \to e^{-5\sigma} ~ (- G)^{1/2}$.
The NS action now involves the dilaton in standard form. Equ. (\ref{dil}) is the unique redefinition that does this. The R action does not have the expected factor of $e^{-2\phi}$, but can be brought to this form by the further redefinition
\begin{equation}
C_1 = e^{-\Phi} C’_1,
\end{equation}
after which
\begin{equation}
\int d^{10} x ~ (-G)^{1/2}~ |F_2|^2 = \int d^{10} x ~ (-G)^{1/2}~ e^{-2\Phi} ~|F’_2|^2,
\end{equation}
\begin{equation}
F’_2 \equiv dC’_1- d\Phi \wedge C’_1,
\end{equation}
and similarly for $F_3$ and $C_3$. With the Bianchi identity and gauge transformations
\begin{equation}
dF'_2=d\Phi\wedge F'_2,~~~~~~~~~~~ \delta C'_I = d\lambda'_0 - \lambda'_0 d\Phi.
\label{B-gt}
\end{equation}
Let us now make contact with string theory and see why the background R–R fields appearing in the world-sheet action have the more complicated properties Equ. (\ref{B-gt}). We work at the linearized level, in terms of the vertex operators
\begin{equation}
\eta_\alpha \tilde{\eta}_\beta ( C \Gamma^{\mu_1...\mu_p} )_{\alpha \beta} e_{\mu_1...\mu_p} (X),
\label{vo}
\end{equation}
here $\eta_\alpha$ is the R ground state vertex operator and $\Gamma^{\mu_1...\mu_p}= \Gamma^{[\mu_1...\Gamma^{\mu_p}]}$ .
Amount to two Dirac equations, one acting on the left spinor index and one on the right:
\begin{equation}
\Gamma^\nu \Gamma^{\mu_1...\mu_p} \partial_\nu e_{\mu_1....\mu_p} (X) =
\Gamma^{\mu_1...\mu_p} \Gamma^\nu \partial_\nu e_{\mu_1....\mu_p} (X) =0,
\end{equation}
this Dirac equation is equivalent to:
\begin{equation}
d e_p =d \star e_p =0,
\end{equation}
these equations have the same form as the field equation and Bianchi identity for a $p$-form field strength. Thus the function $e_{\mu_1...\mu_p} (X) $ appearing in the vertex operator as the R–R field strength rather than potential. To confirm this, observe that in the \RNum{2}A theory the spinors in the R–R vertex operator, Equ. (\ref{vo}) have opposite chirality and so their product contains forms of even rank, the same as the \RNum{2}A R–R field strengths.
\section{Experimental signatures for string theory}
Detecting superstring theory has many perspective and can be in many different ways, as:
\begin{itemize}
\item Finding new supersymmetric particles by the high energy accelerators like the Large Hadron Collider (LHC) at CERN will prove supersymmetry and enhance string theory situation . On the other side, finding standard model- like particles but at a greater mass will be a discovery of what’s called Kaluza-Klein recurrences which may suggest the existence of extra dimensions, one of the predominant predictions of string theories.
\item As gravity is thought to be a force able to probe these extra dimensions, another way in which we may be find evidence for string theory is through the disappearance of gravitons, the hypothesized carrier of gravity, into these other dimensions. The particle might be carried away without a trace, but it would leave behind an imbalance in momentum and energy. So careful analysis would be needed to decide if a graviton has traveled to another dimension or we have instead produced a SUSY particle,
whose decay products would also be invisible to our detector.
\item Through its impact on the earliest, most extreme moments of the universe, the physics of string theory may have left faint cosmological signatures—for example, in the form of gravitational waves or a particular pattern of temperature variations in the cosmic microwave background radiation \cite{Erik2011}.
\end{itemize}
\chapter{$\mathcal{D} = 5~~ \mathcal{N}= 2$ supergravity with hypermultiplets}
\section{Dimensional reduction}
The unique supersymmetric gravity theory in eleven dimensions has the following bosonic action:
\begin{equation}
S_{11}= \frac{1}{2\kappa_{11}^2} \int~ d^{11} x~~~~ \Big[ \sqrt{-G}~ \Big( \mathcal{R} -
\frac{1}{48} \mathcal{F}^2 \Big) + \frac{1}{6~ 3! (4!)^2}~ \epsilon_{\mu_1,..\mu_{11}} \mathcal{A}^{\mu_1,..\mu_3} \mathcal{F}^{\mu_4,..\mu_7} \mathcal{F}^{\mu_7,..\mu_{11}} \Big].
\end{equation}
Write this in the differential forms language, \footnote{Appendix (\ref{Df}).},
the Einstein-Hilbert
term
\begin{equation}
d^{11} x~ |e|~ \mathcal{R} \to ~ \mathcal{R} \star 1,
\end{equation}
the Maxwell's terms
\begin{equation}
\begin{aligned}
d^{11} x~ \frac{|e|}{48} \mathcal{F}^2 \to ~~ \frac{1}{2} \mathcal{F} \wedge \star \mathcal{F}, \\
\mbox{and the third term} \to - \frac{1}{6} \mathcal{A} \wedge \mathcal{F} \wedge \mathcal{F},
\end{aligned}
\end{equation}
So that;
\begin{equation}
S_{11}= \int_{11} \Big( \mathcal{R} \star 1 - \frac{1}{2} \mathcal{F} \wedge \star \mathcal{F} - \frac{1}{6} \mathcal{A} \wedge \mathcal{F} \wedge \mathcal{F} \Big),
\end{equation}
$\mathcal{R}$ is the $D = 11$ Ricci scalar, $\mathcal{A}$ is the 3-form gauge potential and
$\mathcal{F} = d \mathcal{A}$. Make the dimensional reduction:
\begin{equation}
\begin{aligned}
ds^2 &= g_{MN} ~ dx^M dx^N \\
& = e^{\frac{2}{3}\sigma}~g_{\mu\nu} ~ dx^\mu dx^\nu + e^{\frac{\sigma}{3}}~ k_{ab} ~ dx^a dx^b, \\
& ~~~ M,N = 0,..., 10 ~~~ \mu,\nu = 0,..., 4 ~~~ a,b = 1,...,6, \label{metric}
\end{aligned}
\end{equation}
where the metric $g_{\mu\nu}$ is the five dimensional metric, which will eventually contain our
2-brane solutions. $k_{ab}$ is a metric on the six dimensional compact subspace $\mathcal{M}$, the dilaton $\sigma$ is a function in $x^\mu$ only and the warp factors are chosen to give the conventional coefficients in five dimensions, guaranteeing that the gravitational term in the action
will have the standard Einstein-Hilbert form. We choose a complex structure on $\mathcal{M}$ such that
\begin{equation}
k_{ab} dx^a dx^b = k_{mn} dx^m dx^n + k_{\bar{m}\bar{n}} dx^{\bar{m}} dx^{\bar{n}} + 2 k_{m\bar{n}} dx^{m} dx^{\bar{n}},
\end{equation}
where the holomorphic and antiholomorphic indices ($m,n; \bar{m}, \bar{n}$) are three dimensional on $\mathcal{M}$. The
Hermiticity condition demands that $k_{mn} = k_{\bar{m}\bar{n}}=0$, while the Ricci tensor for
$\mathcal{M}$ is set to zero
as dictated by Yau’s theorem. Furthermore, we consider the case where only the complex structure
is deformed, which requires $\delta k_{m\bar{n}}=0$, and $(\delta k_{mn},\delta k_{\bar{m}\bar{n}} \neq 0 )$.
Now, the flux compactification of the gauge field is done by expanding $\mathcal{A}$ into two forms, one
is the five dimensional gauge field A while the other contains the components of $\mathcal{A}$
on $\mathcal{M}$ written
in terms of the cohomology forms $(\alpha_I, \beta^I)$ as follows:
\begin{eqnarray}
\nonumber
\mathcal{A} &=& A + \sqrt{2} (\zeta^I \alpha_I + \tilde{\zeta}_I \beta^I), ~~~ ( I=0,..., h_{2,1}) \\ \nonumber
\mathcal{F} &=& d \mathcal{A} = \mathcal{F} + \mathcal{F}_\mu dx^\mu, \\ \nonumber
F_\mu dx^\mu &=& \sqrt{2} [ (d\zeta^I) \wedge \alpha_I + (d\tilde{\zeta}_I) \wedge \beta^I ] \\
&=& \sqrt{2} [ (\partial_\mu \zeta^I) \alpha_I + (\partial_\mu \tilde{\zeta}_I) \beta^I ] \wedge dx^\mu,
\label{af}
\end{eqnarray}
where $A\equiv \frac{1}{3!} A_{\mu\nu\rho} dx^\mu dx^\nu dx^\rho$. The coefficients $\zeta^I$ and $\tilde{\zeta}_I$ appear as axial scalar fields in the lower dimensional theory. We also note that $A$ in five dimensions is dual to a scalar
field which we will call $a$ (also known as the universal axion). The set ($a$ , $\sigma,~ \zeta^0,~ \tilde{\zeta}_0$ ) as the universal hypermultiplet \cite{CYM2-2,theME}.
The rest of the hypermultiplets are ($ z^i,~ z^{\bar{i}},~ \zeta^i,~ \tilde{\zeta}_i:~ i=1,...,h_{2,1}$). Also included in the hypermultiplets are the fermionic partners of the hypermultiplet scalars known as the hyperini
(singular: hyperino). However, in what follows, we will only discuss the bosonic part of the action.
The hyperini, as well as the gravitini, will make their appearance in the SUSY variation equations
later. The parts one needs for the dimensional reduction are as follows:
1. The metric components
\begin{equation}
\begin{aligned}
G_{\mu\nu} &= e^{\frac{2}{3}\sigma} g_{\mu\nu}, ~~~ G^{\mu\nu} = e^{-\frac{2}{3}\sigma} g^{\mu\nu} \\
G_{ab} &= e^{-\frac{\sigma}{3}} k_{ab}, ~~~ G^{ab} = e^{\frac{\sigma}{3}} k^{ab} \\
G &= \mbox{det} G_{MN} = e^{\frac{4}{3}\sigma} gk, \\
g &= \mbox{det} g_{\mu\nu}, ~~~ k= \mbox{det} k_{ab}.
\end{aligned}
\end{equation}
2. The Christoffel symbols
\begin{equation}
\begin{aligned}
\Gamma^\mu_{\nu\rho} &= \tilde{\Gamma}^\mu_{\nu\rho} [g] + \frac{1}{3} [ \delta^\mu_\nu (\delta_\rho \sigma) +
\delta^\mu_\rho (\delta_\nu \sigma) - \delta_{\nu\rho} \delta^{\mu \kappa} (\partial_\kappa \sigma)] \\
\Gamma^\mu_{ab} &= \frac{1}{6} e^{-\sigma} k_{ab} (\partial^\mu \sigma) - \frac{1}{2} e^{-\sigma} (\partial^\mu k_{ab}) \\
\Gamma^a_{\mu b} &= \frac{1}{2} k^{ac} (\partial_\mu k_{cb}) - \frac{1}{6} \delta^a_b (\partial_\mu \sigma) \\
\Gamma^a_{bc} &= \hat{\Gamma}^a_{bc} [k],
\end{aligned}
\end{equation}
where the (\~{}) and the ($\verb!^!$) refer to the purely five and six dimensional components respectively. Now, calculating the eleven dimensional Ricci scalar based on this gives:
\begin{equation}
\sqrt{|G|} \mathcal{R} = \sqrt{|gk|} \Big [ \mathcal{R}[g] - \frac{1}{2} (\partial_\mu \sigma) (\partial^\mu \sigma)
- \frac{1}{4} k^{m\bar{n}} k^{r\bar{p}} (\partial_\mu k_{mr}) (\partial_\mu k_{\bar{n}\bar{p}})\Big],
\end{equation}
where we have used $\hat{\mathcal{R}}_{ab}=0$ since $\mathcal{M}$ is Ricci-flat,
as well as dropped all total derivatives and terms
containing $k_{mn}$, $k_{\bar{m}\bar{n}}$, and $\delta k_{m\bar{n}}$. Using Equ. (\ref{41-ch2})
\begin{equation}
\int_{11}~~ \mathcal{R} \star 1 = \int d^5 x~ \sqrt{|g|}~ \Big[ R -\frac{1}{2} (\partial_\mu \sigma) (\partial^\mu \sigma)
- G_{i\bar{j}} (\partial_\mu z^i) (\partial^\mu z^{\bar{j}}) \Big],
\end{equation}
where we have normalized the volume of the compact space to $V_{CY}= \sqrt{dw}^6 |k| = 1$. Next, the Maxwell term is:
\begin{equation}
- \frac{1}{2} \mathcal{F} \wedge \star \mathcal{F} \to - \frac{1}{2} \frac{1}{4!} \mathcal{F}_{LMNP} \mathcal{F}^{LMNP} = - \frac{1}{48} \Big(e^{-2\sigma} F_{\mu\nu\rho\sigma} F^{\mu\nu\rho\sigma}
+ e^{\frac{2}{3}\sigma} F_\mu F^\mu \Big).
\end{equation}
Substituting, we get:
\begin{equation}
\begin{aligned}
- \frac{1}{2} \int_{11} \mathcal{F} \wedge \star \mathcal{F} &= - \frac{1}{48}~ \int d^5 x~ \sqrt{|g|} e^{-2\sigma}~ F_{\mu\nu\rho\sigma} F^{\mu\nu\rho\sigma} \\
&- \int d^5 x ~ \sqrt{|g|} e^{\sigma}~ \Big [ (\partial_\mu \zeta^I) (\partial^\mu \zeta^J) \int_\mathcal{M} \alpha_I \wedge \star \alpha_J \\
&+ (\partial_\mu \zeta^I) (\partial^\mu \tilde{\zeta}_J) \int_\mathcal{M} \alpha_I \wedge \star \beta^J +
+(\partial_\mu \tilde{\zeta}_I) (\partial^\mu \zeta^J) \int_\mathcal{M} \beta^I \wedge \star \alpha_J \\
&+ (\partial_\mu \tilde{\zeta}_I) (\partial^\mu \tilde{\zeta}_J) \int_\mathcal{M} \beta^I \wedge \star \beta^J \Big ],
\end{aligned}
\end{equation}
where the Hodge star on the right hand side is with respect to $\mathcal{M}$. Now, using Equ. (\ref{dual-coh}), and Equ. (\ref{hodge-dual})
we end up with:
\begin{equation}
\begin{aligned}
- \frac{1}{2} \int_{11} \mathcal{F} \wedge \star \mathcal{F} = &- \int d^5 x~ \sqrt{|g|} \Big\{ \frac{1}{48}~ e^{-2\sigma}~ F_{\mu\nu\rho\sigma} F^{\mu\nu\rho\sigma} \\
& + e^\sigma \Big[(\gamma_{IJ} + \gamma^{KL} \theta_{IK} \theta_{JL})(\partial_\mu \zeta^I) (\partial^\mu \zeta^J) +
\gamma^{IJ} (\partial_\mu \tilde{\zeta}_I) (\partial^\mu \tilde{\zeta}_J) \\
&+ 2 \gamma^{IK} \theta_{JK} (\partial_\mu \zeta^J) (\partial^\mu \tilde{\zeta}_J) \Big] \Big\}
\end{aligned}
\end{equation}
Finally, the Chern-Simons term gives:
\begin{eqnarray}
\nonumber
- \frac{1}{6} \int_{11} \mathcal{A} \wedge \mathcal{F} \wedge \mathcal{F} &=& - \int_5 \Big[ \zeta^I F \wedge d \tilde{\zeta}_J \int_{\mathcal{M}} \alpha_I \wedge \beta^J
- \tilde{\zeta}_I F \wedge d \zeta^J \int_{\mathcal{M}} \alpha_J \wedge \beta^I \Big] \\
&=& - \int_5 F \wedge (\zeta^I d \tilde{\zeta}_I - \tilde{\zeta}_I d\zeta^I).
\end{eqnarray}
To sum up, the ungauged five dimensional $\mathcal{N} = 2$ supergravity bosonic action with vanishing
vector multiplets is \cite{CYM2-2} :
\begin{equation}
\begin{aligned}
S_5 =& \int~ d^5 x |g|~ \Big\{ R -\frac{1}{2} (\partial_\mu \sigma) (\partial^\mu \sigma) -
G_{i\bar{j}} \partial_\mu z^i \partial^\mu z^{\bar{j}} - \frac{1}{2} e^{-2\sigma} F^2 \\
-& e^\sigma \Big[(\gamma_{IJ} + \gamma^{KL} \theta_{IK} \theta_{JL})(\partial_\mu \zeta^I) (\partial^\mu \zeta^J) +
\gamma^{IJ} (\partial_\mu \tilde{\zeta}_I) (\partial^\mu \tilde{\zeta}_J)
+ 2 \gamma^{IK} \theta_{JK} (\partial_\mu \zeta^J) (\partial^\mu \tilde{\zeta}_J) \Big] \Big\} \\
-& \int_5 F ~(\zeta^I \partial_\mu \tilde{\zeta}_I - \tilde{\zeta}_I \partial_\mu \zeta^I) \wedge dx^\mu,
\end{aligned}
\end{equation}
Or:
\begin{eqnarray}
\nonumber
S_5 &=& \int_5 \Big[ \mathcal{R} \star 1 - \frac{1}{2} d\sigma \wedge \star d\sigma - G_{i\bar{j}} dz^i \wedge \star dz^{\bar{j}} -
F \wedge (\zeta^I d \tilde{\zeta}_I - \tilde{\zeta}_I d\zeta^I)
\\ &-& \frac{1}{2} e^{-2\sigma} F \wedge \star F -e^\sigma ~ X \Big],
\label{act}
\end{eqnarray}
where
\begin{equation}
X = \Big[ (\gamma_{IJ} + \gamma^{KL} \theta_{IK} \theta_{JL}) d \zeta^I \wedge \star d \zeta^J
+ \gamma^{IJ} d \tilde{\zeta}_I \wedge \star d\tilde{\zeta}_J +2 \gamma^{IK} \theta_{JK} d \zeta^J \wedge \star d\tilde{\zeta}_I \Big],
\label{xx}\end{equation}
To complete the picture, we vary the action and present the field equations of $\sigma$, ($z^i, z^{\bar{i}}$), A and
($\zeta^I, \tilde{\zeta}_I$) respectively:
\begin{equation}
(\Delta \sigma) \star 1 - e^\sigma X + e^{-2\sigma} F \wedge \star F = 0
\end{equation}
\begin{equation}
\begin{aligned}
(\Delta z^i) \star 1 + \Gamma^i_{jk} dz^j \wedge \star dz^k - \frac{1}{2} e^\sigma G^{i\bar{j}} (\partial_{\bar{j}} X) \star 1 &=0 \\
(\Delta z^{\bar{i}}) \star 1 + \Gamma^{\bar{i}}_{\bar{j}\bar{k}} dz^{\bar{j}} \wedge \star dz^{\bar{k}} - \frac{1}{2} e^\sigma G^{\bar{i}j} (\partial_{\bar{j}} X) \star 1 &=0
\end{aligned}
\label{zEOM}
\end{equation}
\begin{equation}
d^\dagger \Big[ e^{-2\sigma} F + \star (\zeta^I d\tilde{\zeta}_I - \tilde{\zeta}_I d \zeta^I ) \Big] =0
\end{equation} \label{AA}
\begin{equation}
\begin{aligned}
d^\dagger \Big[ e^{\sigma} \gamma^{IK} \theta_{JK} d\zeta^J + e^\sigma \gamma^{IJ} d\tilde{\zeta}_J + \zeta^I \star F \Big] &=0 \\
d^\dagger \Big[ e^{\sigma} ( \gamma_{IJ} + \gamma^{KL} \theta_{IK} \theta_{JL}) d\zeta^J + e^\sigma \gamma^{JK}
\theta_{IK} d\tilde{\zeta}_J - \tilde{\zeta}_I \star F \Big] &=0 \label{zeta}
\end{aligned}
\end{equation}
From a five dimensional perspective, the moduli ($z^i$ , $z^{\bar{i}}$) behave as scalar fields.
We recall, however, that the behavior of the other fields is dependent on the moduli, i.e. they are functions in them.
Hence it is possible to treat Equ. (\ref{zEOM}) as constraints that can be used to reduce the degrees of freedom of the other field
equations.
Equations (\ref{AA}), and (\ref{zeta}) are clearly the statements that the forms:
\begin{equation}
\begin{aligned}
\jmath_2 &= e^{-2\sigma} F + \star (\zeta^I d \tilde{\zeta}_I - \tilde{\zeta}_I d \zeta^I) \\
\jmath_5 &= e^\sigma \gamma^{IK} \theta_{JK} d\zeta^J + e^\sigma \gamma^{IJ} d\tilde{\zeta}_J + \zeta^I \star F \\
\tilde{\jmath}_{5|I} &= e^\sigma (\gamma_{IJ} + \gamma^{KL} \theta_{IK} \theta_{JL} ) d\zeta^J + e^\sigma \gamma^{JK} \theta_{IK}
d\tilde{\zeta}_J - \tilde{\zeta}_I \star F
\end{aligned}
\end{equation}
From a five dimensional perspective, they can be thought of as the result of the invariance of the action
under particular infinitesimal shifts of $A$ and ($\zeta$, $\tilde{\zeta}$).
The charge densities corresponding to them can then be found in the usual way by:
\begin{equation}
\wp_2 = \int \jmath_2, ~~~~~ \wp_5^I = \int \jmath_5^I, ~~~~~ \tilde{\wp}_{5|I} = \int \tilde{\jmath}_{5|I}.
\end{equation}
The geometric way of understanding these charges is noting that they descend from the eleven
dimensional electric and magnetic M-brane charges, hence the ($2, 5$) labels. M2-branes wrapping
special Lagrangian cycles of $\mathcal{M}$ generate $\wp_2$ while the wrapping of M5-branes excite
($\wp_5^I$, $\tilde{\wp}_{5|I}$).
Finally, for completeness sake we also give $da$, where $a$ is the universal axion dual to $A$. Since
Equ. (\ref{AA}) is equivalent to $d^2 a = 0$, we conclude that
\begin{equation}
da = e^{-2\sigma} \star F - (\zeta^I d\tilde{\zeta}_I - \tilde{\zeta}_I d\zeta^I), \label{aa1}
\end{equation}
where $a$ is governed by the field equation
\begin{equation}
d^\dagger \Big[ e^{2\sigma} da + e^{2\sigma} \Big(\zeta^I d \tilde{\zeta}_I - \tilde{\zeta}_I d\zeta^I \Big) \Big] = 0;
\label{aa2} \end{equation}
as a consequence of $dF = 0$. Both terms involving $F$ in Equ. (\ref{act}) could then be replaced by the single expression
\begin{equation}
S_a = \frac{1}{2} \int e^{2\sigma} \Big[ da + \Big(\zeta^I d \tilde{\zeta}_I - \tilde{\zeta}_I d\zeta^I \Big) \Big]
\wedge \star \Big[ da + \Big(\zeta^I d \tilde{\zeta}_I - \tilde{\zeta}_I d\zeta^I \Big) \Big].
\label{aa3}\end{equation}
\section{The theory in manifestly symplectic form}
For the sake of completeness, we also give a recently proposed form of the $\mathcal{N} = 2$ theory,
clearly highlighting its symplectic structure \cite{CYM2-2}. Since the action is invariant under rotations in $\mathbf{Sp}$,
then it is clear that $R,~ d\sigma,~ dz$ and $F$ are themselves symplectic invariants. The axion fields ($\zeta,\tilde{\zeta}$),
however, can be thought of as components of an $\mathbf{Sp}$ ‘axions vector’. If we define:
\begin{equation}
|Xi \rangle = \begin{pmatrix} \zeta^I \\ -\tilde{\zeta}_I \end{pmatrix} , ~~~~
|d Xi \rangle = \begin{pmatrix} d \zeta^I \\ -d \tilde{\zeta}_I \end{pmatrix}
\end{equation}
then
\begin{equation}
\langle \Xi | d \Xi \rangle = \zeta^I d \tilde{\zeta}_I - \tilde{\zeta}_I d \zeta^I,
\end{equation}
as well as:
\begin{equation}
\begin{aligned}
& \langle \partial_\mu \Xi | \Lambda | \partial^\mu \Xi \rangle = \\
& - (\gamma_{IJ} + \gamma^{KL} \theta_{IK} \theta_{JL}) (\partial_\mu \zeta^I) (\partial^\mu \zeta^J) -
\gamma^{IJ} (\partial_\mu \tilde{\zeta}_I) (\partial^\mu \tilde{\zeta}_J)
- 2 \gamma^{IK} \theta_{JK} (\partial_\mu \zeta^J) (\partial^\mu \tilde{\zeta}_I),
\end{aligned}
\end{equation}
such that Equ. (\ref{xx}) becomes
\begin{equation}
\begin{aligned}
X &= (\gamma_{IJ} + \gamma^{KL} \theta_{IK} \theta_{JL}) d \zeta^I \wedge \star d \zeta^J
+ \gamma^{IJ} d \tilde{\zeta}_I \wedge \star d\tilde{\zeta}_J +2 \gamma^{IK} \theta_{JK} d \zeta^J \wedge \star d\tilde{\zeta}_I,\\
&= - \langle \partial_\mu \Xi | \Lambda | \partial^\mu \Xi \rangle \star 1.
\end{aligned}
\end{equation}
As a consequence of this language, the field expansion Equ. (\ref{af}) could be rewritten
\begin{equation}
\begin{aligned}
\mathcal{A} &= A + \sqrt{2} \langle \Theta | \rangle, \\
\mathcal{F} &= d \mathcal{A} = F + \sqrt{2} \langle \Theta \mathop{|}_\Lambda d \Xi \rangle.
\end{aligned}
\end{equation}
The bosonic action in manifest symplectic covariance is hence:
\begin{equation}
\begin{aligned}
S_5 = \int_5 &\Big[ \mathcal{R} \star 1 - \frac{1}{2} d\sigma \wedge \star d\sigma - G_{i\bar{j}} dz^i \wedge \star dz^{\bar{j}}
- F \wedge \langle \Xi | d \Xi \rangle
\\&- \frac{1}{2} e^{-2\sigma} F \wedge \star F +e^\sigma ~ \langle \partial_\mu \Xi | \Lambda | \partial^\mu \Xi \rangle \star 1 \Big].
\end{aligned} \label{act-f}
\end{equation}
The equations of motion are now
\begin{equation}
(\Delta \sigma) \star 1 + e^\sigma \langle \partial_\mu \Xi | \Lambda | \partial^\mu \Xi \rangle \star 1 + e^{-2\sigma} F \wedge \star F = 0
\label{sigma}
\end{equation}
\begin{equation}
\begin{aligned}
(\Delta z^i) \star 1 + \Gamma^i_{jk} dz^j \wedge \star dz^k + \frac{1}{2} e^\sigma G^{i\bar{j}} \partial_{\bar{j}} \langle \partial_\mu \Xi | \Lambda | \partial^\mu \Xi \rangle \star 1 &=0 \\
(\Delta z^{\bar{i}}) \star 1 + \Gamma^{\bar{i}}_{\bar{j}\bar{k}} dz^{\bar{j}} \wedge \star dz^{\bar{k}} + \frac{1}{2} e^\sigma G^{\bar{i}j} \partial_{\bar{j}} \langle \partial_\mu \Xi | \Lambda | \partial^\mu \Xi \rangle \star 1 &=0 \label{zzz}
\end{aligned}
\end{equation}
\begin{equation}
d^\dagger \Big[ e^{-2\sigma} F + \star \langle \Xi | d \Xi \rangle \Big] =0 \label{AA-f}
\end{equation}
\begin{equation}
d^\dagger \Big[ e^{\sigma} | \Lambda d\Xi \rangle + \star F | \Xi \rangle \Big] =0. \label{zeta-ff}
\end{equation}
Note that, as is usual for Chern-Simons actions, the explicit appearance of the gauge potential $| \Xi \rangle$
in Equ. (\ref{AA-f}), and Equ. (\ref{zeta-ff}) does not have an effect on the physics since:
\begin{equation}
\begin{aligned}
d^\dagger \star \langle\Xi|d\Xi\rangle &\to d \langle\Xi|d\Xi\rangle = \langle d \Xi \mathop{|}_\Lambda d\Xi\rangle \\
d^\dagger \star F|\Xi\rangle &\to d[F|\Xi\rangle]= F \wedge |d\Xi\rangle.
\end{aligned}
\end{equation}
The Noether currents and charges become
\begin{equation}
\begin{aligned}
\jmath_2 &= e^{-2\sigma} F + \star \langle \Xi | d \Xi \rangle \\
|\jmath_5 \rangle &= e^\sigma | \Lambda d\Xi \rangle + \star F |\Xi \rangle
\wp_2 = \int \jmath_2, ~~~~~ |\wp_5\rangle = \int |\jmath_5 \rangle.
\end{aligned}
\end{equation}
The equations of the universal axion Equ. (\ref{aa1}), Equ. (\ref{aa2}), and Equ. (\ref{aa3}) are now
\begin{equation}
da = e^{-2\sigma} \star F - \langle \Xi | d \Xi \rangle,
\end{equation}
\begin{equation}
d^\dagger \Big[ e^{2\sigma} da + e^{2\sigma} \langle \Xi | d \Xi \rangle \Big] = 0 ~~~~~~~~ \mbox{and}
\end{equation}
\begin{equation}
S_a = \frac{1}{2} \int e^{2\sigma} \Big[ da + \langle \Xi | d \Xi \rangle \Big]
\wedge \star \Big[ da + \langle \Xi | d \Xi \rangle \Big].
\end{equation}
So Equ. (\ref{AA-f}), and Equ. (\ref{zeta-ff}) can also written as:
\begin{equation}
d^\dagger \{ e^\sigma | \mathbf{\Lambda} d\Xi \rangle - e^{2\sigma} [ d a + \langle \Xi | d \Xi \rangle] |\Xi \rangle \} =0,
\label{v-f}
\end{equation}
\begin{equation}
d^\dagger [ e^{2\sigma} d a + e^{2\sigma} \langle \Xi | d \Xi \rangle] =0.
\label{k-f}
\end{equation}
The full action is symmetric under the following SUSY transformations:
\begin{eqnarray}
\delta _\epsilon \psi ^1 &=& D \epsilon _1 + \frac{1}{4}\left\{ {i {e^{\sigma } \left[ {d\phi + \left\langle {\Xi }
\mathrel{\left | {\vphantom {\Xi {d\Xi }}}
\right. \kern-\nulldelimiterspace} {{d\Xi }} \right\rangle } \right]}- Y} \right\}\epsilon _1 - e^{\frac{\sigma }{2}} \left\langle {{\bar V}}
\mathrel{\left | {\vphantom {{\bar V} {d\Xi }}} \right. \kern-\nulldelimiterspace} {{d\Xi }} \right\rangle\epsilon _2 \nonumber\\
\delta _\epsilon \psi ^2 &=& D \epsilon _2 - \frac{1}{4}\left\{ {i {e^{\sigma } \left[ {d\phi + \left\langle {\Xi }
\mathrel{\left | {\vphantom {\Xi {d\Xi }}} \right. \kern-\nulldelimiterspace}
{{d\Xi }} \right\rangle } \right]}- Y} \right\}\epsilon _2 + e^{\frac{\sigma }{2}} \left\langle {V}
\mathrel{\left | {\vphantom {V {d\Xi }}} \right. \kern-\nulldelimiterspace} {{d\Xi }} \right\rangle \epsilon _1, \label{SUSYGraviton}
\end{eqnarray}
\begin{eqnarray}
\delta _\epsilon \xi _1^0 &=& e^{\frac{\sigma }{2}} \left\langle {V}
\mathrel{\left | {\vphantom {V {\partial _\mu \Xi }}} \right. \kern-\nulldelimiterspace} {{\partial _\mu \Xi }} \right\rangle \Gamma ^\mu \epsilon _1 - \left\{ {\frac{1}{2}\left( {\partial _\mu \sigma } \right) - \frac{i}{2} e^{\sigma } \left[ {\left(\partial _\mu \phi\right) + \left\langle {\Xi }
\mathrel{\left | {\vphantom {\Xi {\partial _\mu \Xi }}} \right. \kern-\nulldelimiterspace}
{{\partial _\mu \Xi }} \right\rangle } \right]} \right\}\Gamma ^\mu \epsilon _2 \nonumber\\
\delta _\epsilon \xi _2^0 \nonumber &=& e^{\frac{\sigma }{2}} \left\langle {{\bar V}}
\mathrel{\left | {\vphantom {{\bar V} {\partial _\mu \Xi }}} \right. \kern-\nulldelimiterspace} {{\partial _\mu \Xi }} \right\rangle \Gamma ^\mu \epsilon _2 + \left\{ {\frac{1}{2}\left( {\partial _\mu \sigma } \right) + \frac{i}{2} e^{\sigma } \left[ {\left(\partial _\mu \phi\right) + \left\langle {\Xi }
\mathrel{\left | {\vphantom {\Xi {\partial _\mu \Xi }}} \right. \kern-\nulldelimiterspace}
{{\partial _\mu \Xi }} \right\rangle } \right]} \right\}\Gamma ^\mu \epsilon
_1, \\ \label{SUSYHyperon1}
\end{eqnarray}
and
\begin{eqnarray}
\delta _\epsilon \xi _1^{\hat i} &=& e^{\frac{\sigma }{2}} e^{\hat ij} \left\langle {{U_j }}
\mathrel{\left | {\vphantom {{U_j } {\partial _\mu \Xi }}} \right. \kern-\nulldelimiterspace} {{\partial _\mu \Xi }} \right\rangle \Gamma ^\mu \epsilon _1 - e_{\,\,\,\bar j}^{\hat i} \left( {\partial _\mu z^{\bar j} } \right)\Gamma ^\mu \epsilon _2 \\
\delta _\epsilon \xi _2^{\hat i} &=& e^{\frac{\sigma }{2}} e^{\hat i\bar j} \left\langle {{U_{\bar j} }}
\mathrel{\left | {\vphantom {{U_{\bar j} } {\partial _\mu \Xi }}} \right. \kern-\nulldelimiterspace} {{\partial _\mu \Xi }} \right\rangle \Gamma ^\mu \epsilon _2 + e_{\,\,\,j}^{\hat i} \left( {\partial _\mu z^j } \right)\Gamma ^\mu \epsilon _1,\label{SUSYHyperon2}
\end{eqnarray}
where $\left(\psi ^1, \psi ^2\right)$ are the two gravitini and $\left(\xi _1^I, \xi _2^I\right)$ are the hyperini. The quantity $Y$ is defined by:
\begin{equation}
Y = \frac{{\bar Z^I N_{IJ} {d Z^J } -
Z^I N_{IJ} {d \bar Z^J } }}{{\bar Z^I N_{IJ} Z^J
}},\label{DefOfY}
\end{equation}
where $N_{IJ} = \mathfrak{Im} \left({\partial_IF_J } \right)$. The $e$'s are the beins of the special K\"{a}hler metric $G_{i\bar j}$, the $\epsilon$'s are the five-dimensional $\mathcal{N}=2$ SUSY spinors and the $\Gamma$'s are the usual Dirac matrices. The covariant derivative $D$ is given by $D=dx^\mu\left( \partial _\mu + \frac{1}{4}\omega _\mu^{\,\,\,\,\hat \mu\hat \nu} \Gamma _{\hat \mu\hat \nu}\right)\label{DefOfCovDerivative}$ as usual, where the $\omega$'s are the spin connections and the hatted indices are frame indices in a flat tangent space.
\chapter{Inflationary Brane-Worlds Coupled to the Calabi-Yau
Complex Structure Moduli}
\label{ch4}
Still there are many quests have not been resolved yet about the Cosmological evolution. For instance, around the first seconds of
our universe birth. Namely, what happened before the Big Bang ?
What do drive the universe as its way till now times? What were the universe initial conditions (ICs) ?
And how it rapidly exapanded " in size "
in a very short period of time during the inflation era ? Or what does the cosmological constant really represent ?
The term has been added to Einstein field equations to account for the accelerated expansion on the time being epoch.
\footnote{See Appendix (\ref{GR}) for more information about the cosmological constant.}
Finally how the universe will end up ?
What are the most probable scenarios for our universe future?
There are many theories try to interpret all these mystries, like quantum gravity, see for example \cite{Hertog:2013, James:2018},
or higher curvature theories of gravity
\footnote{Including string theory itself.}, as \cite{Baumann:2009, McAllister:2010}.
However, we try to look at these from other perspective through modeling the universe as a 3-brane imbedded in a higher five dimensional spacetime called the bulk.
Indeed modeling the universe as a 3-brane embedded in a higher dimensional bulk, initially proposed by \cite{Randall:1999ee} and further developed by many authors such as \cite{Brax:2003fv}, \cite{Maartens:2010ar}, and \cite{Roane:2007zz}. While our work
is an extension to a previous studies have been initiated before \cite{Emam:2015laa}.
In this chapter we manifest the correlation has been deduced between
the moduli of the complex structure of the underlying Calabi-Yau manifold and the dynamic behavior of the brane-world . We numerically solve the field equations in case of dust filled and radiation filled branes.
Also we study the time dependence of the $\mathcal{N} = 2$ hypermultiplets scalars. Although these fields are restricted to the bulk, the dilaton, for instance, is related to the volume of CY manifold, so it’s important to see how it behaves at different universes (ICs).
\section{Brane dynamics}
We consider a metric of the form
\begin{equation}%
ds^2= - e^{2\alpha(t,y)} dt^2 + e^{2\beta(t,y)} (dr^2 + r^2 d \Omega^2 ) + e^{2\gamma(t,y)} dy^2,
\label{metricE}
\end{equation}%
where $d\Omega^2= d\theta^2+\sin^2(\theta) d\phi^2.$ This metric may be interpreted as representing a single 3-brane located at $y = 0$ in the transverse space, it may also represent a stack of N branes located at various values of $y = y_I (I = 1, ...., N \in \mathbb{Z} ) $ where the warp functions $\alpha$, $\beta$ and $\gamma$ are rewritten such that
$y \to \sum_{I=1}^N |y-y_I|$. Either way, we will eventually focus on the four dimensional $(t, r, \theta, \phi)$ dynamics, effectively evaluating the warp functions at a specific, but arbitrary, $y$ value. We also
note that a metric of the form Equ. (\ref{metricE}) was shown in
\cite{Kallosh:2001} to be exactly the type needed for a consistent BPS cosmology .
The components of the Einstein tensor are
\begin{eqnarray} \nonumber G_{tt} &=&3 (\dot\beta^2+\dot\beta \dot\gamma)-3 e^{2(\alpha-\gamma)} (\beta^{\prime \prime} + 2 \beta^{\prime 2}-\beta^{\prime} \gamma^{\prime}) \\
\nonumber G_{rr} &=& - e^{2(\beta-\alpha)}[ 2 \ddot\beta+3\dot\beta^2+\ddot\gamma+ \dot\gamma^2 + 2\dot\beta(\dot\gamma-\dot\alpha) -\dot\alpha \dot\gamma)]
\\ \nonumber &+& e^{2(\beta-\gamma)}[ 2\beta^{\prime \prime}+3\beta^{\prime 2}+\alpha^{\prime \prime} + \alpha^{\prime 2} + 2\beta^{\prime} (\alpha^{\prime}-\gamma^{\prime}) -\alpha^{\prime} \gamma^{\prime})]
\end{eqnarray}
\newpage
\begin{eqnarray}
\nonumber G_{(\theta\theta,~ \phi\phi)} &=& e^{2(\beta-\gamma)} ~ (r^2,~ r^2 sin^2\theta)~
[3\beta^{\prime 2} - 2 \beta^{\prime} (\gamma^{\prime}-\alpha^{\prime})- \gamma^{\prime}\alpha^{\prime}+\alpha^{\prime 2}+2\beta^{\prime \prime}+\alpha^{\prime \prime}] \\ \nonumber &-& e^{2(\beta-\alpha)} ~ ( r^2, ~ r^2 \sin^2\theta)
[3 \dot\beta^2+ 2 \dot\beta( \dot\gamma- \dot\alpha )+ \dot\gamma^2 - \dot\gamma \dot\alpha+ 2 \ddot\beta+\ddot\gamma], \\
\nonumber G_{yy} &=& 3 (\beta^{ \prime 2} + \beta' \alpha')-3 e^{2(\gamma-\alpha)} (\ddot\beta + 2 \dot\beta^2-\dot\beta \dot\alpha) \\
G_{ty} &=& 3( \alpha^{\prime} \dot\beta-\beta^{\prime} \dot\beta+ \beta^{\prime} \dot\gamma - \dot\beta^{\prime}) ,
\label{GG}
\end{eqnarray}
where a prime is a derivative with respect to $y$ and a dot is a derivative with respect to $t$. We consider only the fields' dependence on time, so that $G_{ty}$ totally vanishes.
The energy momentum tensor is given by \cite{BookME} :
\begin{equation}%
T_{\mu\nu} = -2 \frac{\partial \mathcal{L}}{\partial g^{\mu\nu}}+ g_{\mu\nu} \mathcal{L}.
\end{equation}%
ln addition to the usual perfect fluid stress tensor on the brane:
\begin{equation}%
T_{\mu\nu} =( \rho + p ) U_\mu U_\nu + g_{\mu\nu} p
\end{equation}%
From (\ref{act-f});
\begin{multline}
\mathcal{L}= - R + \frac{g^{\alpha\beta}}{4} \partial_\alpha \sigma \partial_\beta \sigma + \frac{g^{\alpha\beta}}{2} G_{i\bar{j}} \partial_\alpha z^i \partial_\beta z^{\bar{j}} - \frac{g^{\alpha\beta}}{2}e^\sigma [\partial_\alpha \Xi|\Lambda|\partial_\beta \Xi ] \\+
\frac{e^{2\sigma}}{2} [\partial_\alpha \phi +\langle \Xi|\partial_\alpha \Xi\rangle ] ~ [ \partial_\beta \phi+ \langle \Xi|\partial_\beta \Xi\rangle ]~~~~~~~~~~~~~~~~~~~~~~~~~~~~
\end{multline}
\begin{equation}%
T_{\mu\nu} = T_{\mu\nu}^{Bulk} + T_{\mu\nu}^{3b},
\end{equation}%
\begin{equation}%
T_{yy} = T_{yy}^{Bulk}, ~~~~~~~~~~ T_{yt}=0.
\end{equation}%
\begin{multline}
-2\frac{\partial \mathcal{L}}{\partial g^{\mu\nu}}
= -\frac{1}{2} \partial_\mu \sigma \partial_\nu \sigma - G_{i\bar{j}} \partial_\mu z^i \partial_\nu z^{\bar{j}} + e^\sigma [\partial_\mu \Xi|\Lambda|\partial_\nu \Xi ] \\ - \frac{e^{2\sigma}}{2} [\partial_\mu \phi + \langle \Xi|\partial_\mu \Xi\rangle ] ~ [ \partial_\nu \phi+ \langle \Xi|\partial_\nu \Xi\rangle ],
~~~~~~~~~~~~~~~~~~~
\end{multline}
so that:
\begin{eqnarray} \nonumber
T_{\mu\nu}^{Bulk} &=& -\frac{1}{2} \partial_\mu \sigma \partial_\nu \sigma - G_{i\bar{j}} \partial_\mu z^i \partial_\nu z^{\bar{j}} + e^\sigma [\partial_\mu \Xi|\Lambda|\partial_\nu \Xi ]
\\\nonumber &-& \frac{e^{2\sigma}}{2} [\partial_\mu \phi + \langle \Xi|\partial_\mu \Xi\rangle ] ~ [ \partial_\nu \phi+ \langle \Xi|\partial_\nu \Xi\rangle ] + \frac{g_{\mu\nu}}{4} \partial_\alpha \sigma \partial^\alpha \sigma
+ \frac{g_{\mu\nu}}{2} G_{i\bar{j}} \partial_\alpha z^i \partial^\alpha z^{\bar{j}}
\\\nonumber &-& \frac{g_{\mu\nu}}{2} e^\sigma[\partial_\alpha \Xi|\Lambda|\partial^\alpha \Xi ]
+ g_{\mu\nu} \frac{e^{2\sigma}}{4} [\partial_\alpha \phi + \langle \Xi|\partial_\alpha \Xi\rangle ] ~ [ \partial^\alpha \phi+ \langle \Xi|\partial^\alpha \Xi\rangle ]
,\\
\label{munutensor}
\end{eqnarray}
and:
\begin{eqnarray}
\nonumber T_{yy} &=& -\frac{1}{2} \partial_y \sigma \partial_y \sigma - G_{i\bar{j}} \partial_y z^i \partial_y z^{\bar{j}} + e^\sigma [\partial_y \Xi|\Lambda|\partial_y \Xi ]
\\\nonumber
&-& \frac{e^{2\sigma}}{2} [\partial_y \phi + \langle \Xi|\partial_y \Xi \rangle ] ~ [ \partial_y \phi+ \langle \Xi|\partial_y \Xi\rangle ]
+ \frac{g_{yy}}{4} \partial_y \sigma \partial^y \sigma
+\frac{g_{yy}}{2} G_{i\bar{j}} \partial_y z^i \partial^y z^{\bar{j}}\\\nonumber
&-& \frac{g_{yy}}{2} e^\sigma [\partial_y \Xi|\Lambda|\partial^y \Xi ] + g_{yy} \frac{e^{2\sigma}}{4} [\partial_y \phi + \langle \Xi|\partial_y \Xi\rangle ] ~ [ \partial^y \phi+ \langle \Xi|\partial^y \Xi \rangle ]
.\\
\end{eqnarray}
Then:
\begin{equation}%
T^{3b}_{tt} = \rho(t), ~~~~~~~
T^{3b}_{rr} = e^{2\beta} p(t), ~~~~~~~
T^{3b}_{\theta\theta} = e^{2\beta} r^2 p(t), ~~~~~
T^{3b}_{\phi\phi} = e^{2\beta} \sin{\theta}^2 p(t), ~~~~~~~
\end{equation}%
where $\rho$ and $p$ depend on time. We are interested in bosonic configurations that preserve some supersymmetry,
so the stress tensor (\ref{munutensor})
can be considerably simplified by considering the vanishing of the supersymmetric variations
(\ref{SUSYHyperon1}, \ref{SUSYHyperon2}), that leads to the BPS condition \cite{Emam:2015laa, Emam2013}
\begin{equation} e^\sigma \langle \Xi|\mathbf {\Lambda}| \star d\Xi \rangle
= \frac{1}{2} d\sigma \wedge \star d\sigma + \frac{1}{2} e^{2\sigma} [d\phi + \langle\Xi|~d\Xi \rangle] \wedge \star [d\phi + \langle \Xi|~d\Xi\rangle] + 2 |\dot z|^2,
\label{bps}
\end{equation}
Substitute by (\ref{bps}) in (\ref{munutensor}), leaving the dynamics to depend only on the complex
structure moduli ($z^i,z^{\bar{i}}$) :
\begin{equation}%
T_{\mu\nu}= G_{i\bar{j}} \partial_\mu z^i \partial_\nu z^{\bar{j}} - \frac{1}{2} g_{\mu\nu} G_{i\bar{j}} \partial_\alpha z^i \partial^\alpha z^{\bar{j}}.
\end{equation}%
Then:
\begin{eqnarray}
\label{EFE1}
\nonumber G_{tt} &=& \frac{1}{2}G_{i\bar{j}} {\dot z^i} {\dot z^{\bar{j}}}+ \frac{1}{2} e^{2(\alpha-\gamma)} G_{i\bar{j}} z^{`i} z^{`\bar{j}}+e^{2\alpha}\rho, \\\nonumber
G_{rr} &=& \frac{1}{2}G_{i\bar{j}} e^{2(\beta-\alpha)} {\dot z^i} {\dot z^{\bar{j}}}- \frac{1}{2} e^{2(\beta-\gamma)} G_{i\bar{j}} z^{`i} z^{`\bar{j}}+e^{2\beta} p, \\\nonumber
G_{\theta\theta} &=& \frac{1}{2} G_{i\bar{j}} e^{2(\beta-\alpha)} r^2 {\dot z^i} {\dot z^{\bar{j}}} - \frac{1}{2}G_{i\bar{j}}r^2 e^{2(\beta-\gamma)} z^{`i} z^{`\bar{j}} + e^{2\beta} r^2 p, \\\nonumber
G_{\phi\phi} &=& \frac{1}{2} G_{i\bar{j}} e^{2(\beta-\alpha)} r^2 \sin{\theta}^2 {\dot z^i} {\dot z^{\bar{j}}} - \frac{r^2 \sin{\theta}^2}{2}G_{i\bar{j}} e^{2(\beta-\gamma)} z^{`i} z^{`\bar{j}} + e^{2\beta} r^2 \sin{\theta}^2 p, \\
\end{eqnarray}
and;
\begin{equation}%
\label{EFE2}
G_{yy} = \frac{1}{2} G_{i\bar{j}} z^{'i} z^{\bar{'j}}+\frac{1}{2} e^{2(\gamma-\alpha)} G_{i\bar{j}} \dot z^{i} \dot z^{\bar{j}},
\end{equation}
The components of the Einstein tensor are given by Equ. (\ref{GG}).
The universal axion equations (\ref{v-f}) and (\ref{k-f}) can be integrated to give
\begin{equation}
d\phi + \langle\Xi|d\Xi\rangle = n e^{-2\sigma} dh,
\label{phiEOM}
\end{equation}
where $h$ is harmonic in $(t,y)$, i.e. satisfies $\Delta h=0$, and n $\in \mathbb{R}$. Similarly:
\begin{equation} e^\sigma|\Lambda d\Xi\rangle- n dh |\Xi\rangle = s|dK\rangle ~~~~~
\mbox{where} ~~~~~ |\Delta K \rangle =0, ~~~
s \in \mathbb{R}.
\label{uniAx}
\end{equation}
Using the hyperini transformations (\ref{SUSYHyperon1}) and (\ref{SUSYHyperon2}) with (\ref{uniAx}) and make the assumption that
$\epsilon_1 = \pm \epsilon_2$ we find the following for the axions \cite{Emam:2015laa}:
\begin{equation}
|d\Xi\rangle = e^{-\frac{\sigma}{2}}~ Re ~ [
( e^{-\sigma} dh-id\sigma) |V\rangle + 2i |U_i\rangle dz^i
],
\label{xi1}
\end{equation}
\begin{equation}
|\Xi\rangle dh = e^{-\frac{\sigma}{2}}~ Re ~ [
( d\sigma+i e^{-\sigma} dh) |V\rangle + 2 |U_i\rangle dz^i],
\label{xi2}
\end{equation}
at $n=1$ and $s=0$. On the other hand, the equation of the harmonic function $h$, that arises from $\Delta h=0$ is given by:
\begin{equation}
e^{\alpha-\gamma} [h'' + (\alpha' + 3 \beta' - \gamma') h' ] =
e^{\gamma-\alpha} [\ddot{h} - (\dot{\alpha} - 3 \dot{\beta} - \dot{\gamma}) \dot{h} ]
\label{harmo}
\end{equation}
\section{The cosmology of a single brane}
The warp functions as well as $h$ can be considered as separable as follows:
\begin{equation}
\begin{split}
e^{\beta(t,y)} &= a(t)~F(y) \\
e^{\gamma(t,y)} &= b(t)~ K(y) \\
e^{\alpha(t,y)} &= c(t)~ N(y) \\
h(t,y) &= k(t)~ M(y)
\end{split}
\end{equation}
Our interest is the dynamics of a single brane out of an infinite number of possible 3-branes along $y$, so we will evalutate the functions $F (y),~ K (y),~ N (y)$, and $M (y)$ near the brane of interestand normalize the result to unity, i.e. $F (0) = 1$ and so on, where the brane under study is located at $y = 0$.
So that:
\begin{equation}
\begin{aligned}
& e^{2\alpha} = c^2(t) \to 1, ~~~~~~ \mbox{(RW-like metric)}, ~~~~~ \dot\alpha = 0, \\
& e^{2\gamma} = b^2(t), ~~~~~~~~~\dot\gamma = \frac{\dot b}{b}, \\
& e^{2\beta} = a^2(t), ~~~~~~~~ \dot\beta= \frac{\dot a}{a}.
\end{aligned}
\end{equation}
Equations (\ref{GG}) and (\ref{harmo}) are the basic
equations governing the dynamics of the multi-brane spacetime. Einstein equations
reduce to the Friedmann-like form
\begin{eqnarray}
\label{eee}
3\left[ {\left( {\frac{{\dot a}}{a}} \right)^2 + \left( {\frac{{\dot a}}{a}} \right)\left( {\frac{{\dot b}}{b}} \right)} \right] &=& G_{i\bar j} \dot z^i \dot z^{\bar j} + \rho(t) \nonumber\\
2\frac{{\ddot a}}{a} + \left( {\frac{{\dot a}}{a}} \right)^2 + \frac{{\ddot b}}{b} + 2\left( {\frac{{\dot a}}{a}} \right)\left( {\frac{{\dot b}}{b}} \right) &=& - G_{i\bar j} \dot z^i \dot z^{\bar j} - p(t) \nonumber\\
3\left[ {\frac{{\ddot a}}{a} + \left( {\frac{{\dot a}}{a}} \right)^2 } \right] &=& - G_{i\bar j} \dot z^i \dot z^{\bar j}.
\end{eqnarray}
It is clear that the quantity $G_{i\bar j} \dot z^i \dot z^{\bar j}$ plays an important role in the solution since
as it will be shown to be the source of the dynamics of the brane. It is interpreted as the norm of the moduli's flow velocity.
Now the equation of the harmonic function (\ref{harmo}) becomes:
\begin{equation}
\ddot k + \Big[3 \frac{\dot a}{a} +\frac{\dot b}{b} \Big] \dot k= 0
\label{keq}
\end{equation}
We tend to make $k$ and all the hypermultiplet bulk fields depend only on the scale factor $a$. So by using
(\ref{EFE1}) and (\ref{GG}) in (\ref{keq}) we get for $k$:
\begin{equation}
\ddot k - \Big(3 \frac{\dot a}{a} +3\frac{\ddot a}{\dot a} - \frac{1}{a^2 \dot a} \Big) \dot k= 0
\end{equation}
For the dilaton, from (\ref{sigma}), (\ref{AA-f}), and (\ref{zeta-ff}) we get
the time dependence equation of $\sigma$:
\begin{equation}
\ddot \sigma + \frac{1}{2} \dot \sigma^2 -
\frac{e^{-2\sigma}}{2} \dot k^2 - 6 \Big( \frac{\ddot a}{a} + \frac{\dot a^2}{a^2} \Big)=0.
\end{equation}
To get the time dependence of $\phi$ (\ref{phiEOM}) we should to get $\langle\Xi|\dot \Xi\rangle$ first. From
(\ref{xi1}) and (\ref{xi2})
\footnote{Within the text some times we use $|\dot{z}|^2 \equiv G_{i\bar{j}} d\dot{z}^i d\dot{z}^{\bar{j}}$.}
\begin{equation}
\langle \Xi|\dot {\Xi}\rangle = \frac{e^{-\sigma}}{2 \dot k} \Big( e^{-2\sigma} \dot k^2 + \dot \sigma^2 + 4 |\dot z|^2 \Big),
\end{equation}
so that
\begin{equation}
\dot \phi =e^{-2\sigma} {\dot k} -\frac{e^{-\sigma}}{2 \dot k}
\Big[ e^{-2\sigma} \dot k^2 + \dot \sigma^2 - 12 \Big( \frac{\ddot a}{a} + \frac{\dot a^2}{a^2} \Big) \Big]
\end{equation}
Finally, while again using the assumption $\epsilon_1=\pm\epsilon_2$, the vanishing of the gravitini equations (\ref{SUSYGraviton}) gives the following for the near-brane spinors
\begin{equation}\label{Spinor}
\epsilon\left( t\right) = e^{\frac{\sigma }{2} + \frac{3}{4}in\Omega k - \Upsilon } \hat \epsilon,
\end{equation}
where $\hat \epsilon$ is an arbitrary constant spinor and the functions $\Upsilon$ and $\Omega$ are solutions of $\dot \Upsilon = Y$ and
\begin{equation}\label{Omega}
\frac{d}{{dt}}\left( {\Omega k} \right) = e^{ - \sigma } \dot k
\end{equation}
respectively. This and all the field equations are the same in case of dust- filled brane, radiation- filled brane and
radiation, dust with energy filled brane.
The equations derived here are much too complicated to solve analytically.
We then present a numerical solution for the time dependence of the quantities involved.
\section{A dust-filled brane}
\label{DD}
For a dust- filled brane
\begin{equation}\label{Density1}
T_{tt}^{\textit {Brane}} = \rho\left(t\right) = \frac{1}{{a^3 }}, ~~~~~~~~~~ p(t) =0,
\end{equation}
so that the Einstein equations (\ref{eee}) become
\begin{eqnarray}
3\left[ {\left( {\frac{{\dot a}}{a}} \right)^2 + \left( {\frac{{\dot a}}{a}} \right)\left( {\frac{{\dot b}}{b}} \right)} \right] &=& G_{i\bar j} \dot z^i \dot z^{\bar j} + \frac{1}{{a^3 }} \nonumber\\
2\frac{{\ddot a}}{a} + \left( {\frac{{\dot a}}{a}} \right)^2 + \frac{{\ddot b}}{b} + 2\left( {\frac{{\dot a}}{a}} \right)\left( {\frac{{\dot b}}{b}} \right) &=& - G_{i\bar j} \dot z^i \dot z^{\bar j} \nonumber\\
3\left[ {\frac{{\ddot a}}{a} + \left( {\frac{{\dot a}}{a}} \right)^2 } \right] &=& - G_{i\bar j} \dot z^i \dot z^{\bar j}.
\end{eqnarray}
We numerically solve the given equations for the scale factors $a$ and $b$. We also find $G_{i\bar j} \dot z^i \dot z^{\bar j}$ and note that all three quantities are clearly related; confirming \cite{Emam:2015laa}. In fact this work takes this relation further than was found in the previous work; wherein the form of $a$ had to be assumed to fit the \emph{assumption} of an expanding universe. In our case we find all three results \emph{without} having to make assumptions about any of them, and show that the expansion of the brane-world is inevitable in the presence of decaying moduli. The fact that we find this without any imposed assumptions is ascribed to the presence of the constraints (\ref{Density1}). To complete the picture, we also solve for $k$, $\sigma$, $\phi$, $\left\langle {\Xi } \mathrel{\left | {\vphantom {\Xi {\dot \Xi }}} \right. \kern-\nulldelimiterspace} {{\dot \Xi }} \right\rangle $, as well as $\Omega$. Clearly we do not, and in fact \emph{cannot}, have a complete solution because the field equations for the moduli themselves cannot be solved without explicit knowledge of a metric on the Calabi-Yau submanifold and its exact topology. In fact, we take this work to be one step closer a possible deeper understanding of these topics; namely what the Calabi-Yau submanifold exactly looks like. Finally, let's note that because of the nature of numerical analysis, we find and present all results using a variety of initial conditions, some of which pertain to the possible use of this result to model our universe (\emph{i.e.} a big bang-like singularity as an initial condition), while the remaining ones are given for completeness and additional confirmation of the causal relation between the moduli and the scale factors. These ICs are listed in tabel (\ref{112-4}), where
for the first five initial conditions we find that $b\propto a$ up to an arbitrary scaling constant determined by the initial conditions. As such and without loss of generality we choose a scale of unity.
Consequently $a$, $b$, and their derivatives exactly coincide in the dust plots of IC1 through IC5. In the sixth case different initial conditions are chosen for $a$ and $b$, and thus they no longer coincide. In most solutions in this subsection the norm $G_{i\bar j} \dot z^i \dot z^{\bar j}$ is negative; except IC4 Fig. (\ref{33}), however we plot its absolute value for ease of reference.
\begin{table}[H]
\centering
\caption{The six sets of initial conditions (IC) used in the dust and radiation filled branes computations.}
\label{112-4}
\vspace{1cm}
\begin{adjustbox}{max width=\textwidth}
\begin{tabular}{|c|c|c|c|c|l|}
\hline \rule{0pt}{1cm}
\large{\textbf{IC Set Number}} & $a$ & $b$ & $\dot a$ & $\dot b$ & \large{\textbf{Description}} \\ \hline \rule{0pt}{1cm}
\textbf{1} & 0 & 0 & 0 & 0 & \large{Big bang-like IC with vanishing initial velocities} \\\hline \rule{0pt}{1cm}
\textbf{2} & 1 & 0 & 1 & 0 & \large{Non-singular IC with vanishing initial velocities} \\ \hline \rule{0pt}{1cm}
\textbf{3} & 0 & 1 & 0 & 1 & \large{Big bang-like IC with initial positive velocities} \\ \hline \rule{0pt}{1cm}
\textbf{4} & 1 & 1 & 1 & 1 & \large{Non-singular IC with initial positive velocities} \\ \hline \rule{0pt}{1cm}
\textbf{5} & 1 & -0.2 & 1 & -0.2 & \large{Non-singular IC with initial negative velocities} \\ \hline \rule{0pt}{1cm}
\textbf{6} & 0 & 0 & 1 & 0 & \large{Initial singularity in $a$ only with vanishing initial velocities} \\ \hline
\end{tabular}
\end{adjustbox}
\end{table}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic1/a_b_zz}
\caption{The scale factors $a$ and $b$; represented by the blue curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ represented by the red curve.}
\label{1-4}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic1/a_b_dot}
\caption{The expansion rates of the scale factors. Both $\dot a$ and $\dot b$ are represented by the shown curve.}
\label{2}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic1/a_b_ddot}
\caption{The accelerations of the scale factors. Both $\ddot a$ and $\ddot b$ are represented by the shown curve.}
\label{3}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic1/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{4}
\end{subfigure}
\caption{Dust-filled brane world with initial conditions set number 1 .}
\label{Fig1}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic1/sigma}
\caption{The dilaton $\sigma$; same for all three $\dot k\left(0\right)$.}
\label{5}
\end{subfigure}
\qquad%
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic1/sigma_dot}
\caption{The dilatonic field strength $\dot\sigma$.}
\label{6}
\end{subfigure}
\qquad
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic1/Zoom_sigma_1}
\caption{A zoom-in on the dilaton for $\dot k\left(0\right)=0$.}
\label{7}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic1/Zoom_sigma_2}
\caption{A zoom-in on the dilaton for $\dot k\left(0\right)=-1$.}
\label{8}
\end{subfigure}
\caption{Dust-filled brane world with initial conditions set number 1 (continued).}
\label{Fig2}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic1/phi}
\caption{The universal axion $\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve). The solution diverges for $\dot k\left(0\right)=0$.}
\label{9}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic1/phi_dot}
\caption{The axionic field strength $\dot\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve).}
\label{10}
\end{subfigure}
\qquad
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic1/xxx}
\caption{$\langle \Xi | \dot{\Xi} \rangle$ for $\dot{k}(0)= 1$ (blue), and $\dot{k}(0) = -1$ (red).}
\label{11}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic1/omega}
\caption{$\Omega$ at $\dot k\left(0\right) = 1$ and $ \dot\sigma \left(0\right) =0 $.}
\label{12}
\end{subfigure}%
\caption{Dust-filled brane world with initial conditions set number 1 (continued).}
\label{Fig22}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic2/a_b_zz}
\caption{The scale factors $a$ and $b$; represented by the blue curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ represented by the red curve.}
\label{13}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic2/a_b_dot}
\caption{The expansion rates of the scale factors. Both $\dot a$ and $\dot b$ are represented by the shown curve.}
\label{14}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic2/a_b_ddot}
\caption{The accelerations of the scale factors. Both $\ddot a$ and $\ddot b$ are represented by the shown curve.}
\label{15}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic2/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{16}
\end{subfigure}
\caption{Dust-filled brane world with initial conditions set number 2.}
\label{Fig4}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic2/sigma}
\caption{The dilaton $\sigma$; same for all three $\dot k\left(0\right)$.}
\label{17}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic2/sigma_dot}
\caption{The dilatonic field strength $\dot\sigma$.}
\label{18}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic2/phi}
\caption{The universal axion $\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve). The solution diverges for $\dot k\left(0\right)=0$.}
\label{19}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic2/phi_dot}
\caption{The axionic field strength $\dot\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve).}
\label{20}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic2/xxx}
\caption{$\langle \Xi | \dot{\Xi} \rangle$ for $\dot{k}(0)= 1$ (blue), and $\dot{k}(0) = -1$ (red).}
\label{21}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic2/omega}
\caption{$\Omega$ at $\dot k\left(0\right) = 1$ and $ \dot\sigma \left(0\right) =0 $.}
\label{22}
\end{subfigure}
\caption{Dust-filled brane world with initial conditions set number 2 (continued).}
\label{Fig44}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic3/a_b_zz}
\caption{The scale factors $a$ and $b$; represented by the blue curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ represented by the red curve.}
\label{23}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic3/a_b_dot}
\caption{The expansion rates of the scale factors. Both $\dot a$ and $\dot b$ are represented by the shown curve.}
\label{24}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic3/a_b_ddot}
\caption{The accelerations of the scale factors. Both $\ddot a$ and $\ddot b$ are represented by the shown curve.}
\label{25}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic3/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{26}
\end{subfigure}
\caption{Dust-filled brane world with initial conditions set number 3.}
\label{Fig6}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic3/sigma}
\caption{The dilaton $\sigma$; same for all three $\dot k\left(0\right)$.}
\label{27}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic3/sigma_dot}
\caption{The dilatonic field strength $\dot\sigma$.}
\label{28}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic3/phi}
\caption{The universal axion $\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve). The solution diverges for $\dot k\left(0\right)=0$.}
\label{29}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic3/phi_dot}
\caption{The axionic field strength $\dot\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve).}
\label{30}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic3/xxx}
\caption{$ \langle \Xi | \dot{\Xi} \rangle$ for $\dot{k}(0)= 1$ (blue), and $\dot{k}(0) = -1$ (red).}
\label{31}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic3/omega}
\caption{$\Omega$ at $\dot k\left(0\right) = 1$ and $ \dot\sigma \left(0\right) =0 $.}
\label{32}
\end{subfigure}
\caption{Dust-filled brane world with initial conditions set number 3 (continued).}
\label{Fig66}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic4/a_b_zz}
\caption{The scale factors $a$ and $b$; represented by the blue curve, while ${G_{i\bar j} \dot z^i \dot z^{\bar j}}$ (already positive) represented by the red curve.}
\label{33}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic4/a_b_dot}
\caption{The expansion rates of the scale factors. Both $\dot a$ and $\dot b$ are represented by the shown curve.}
\label{34}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic4/a_b_ddot}
\caption{The accelerations of the scale factors. Both $\ddot a$ and $\ddot b$ are represented by the shown curve.}
\label{35}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic4/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{36}
\end{subfigure}
\caption{Dust-filled brane world with initial conditions set number 4.}
\label{Fig8}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic4/sigma}
\caption{The dilaton $\sigma$; same for all three $\dot k\left(0\right)$.}
\label{37}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic4/sigma_dot}
\caption{The dilatonic field strength $\dot\sigma$.}
\label{38}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic4/phi}
\caption{The universal axion $\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve). The solution diverges for $\dot k\left(0\right)=0$.}
\label{39}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic4/phi_dot}
\caption{The axionic field strength $\dot\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve).}
\label{40}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic4/xxx}
\caption{$ \langle \Xi | \dot{\Xi} \rangle$ for $\dot{k}(0)= 1$ (blue), and $\dot{k}(0) = -1$ (red).}
\label{41}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic4/omega}
\caption{$\Omega$ at $\dot k\left(0\right) = 1$ and $ \dot\sigma \left(0\right) =0 $.}
\label{42}
\end{subfigure}
\caption{Dust-filled brane world with initial conditions set number 4 (continued).}
\label{Fig88}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic5/a_b_zz}
\caption{The scale factors $a$ and $b$; represented by the blue curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ represented by the red curve.}
\label{43}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic5/a_b_dot}
\caption{The expansion rates of the scale factors. Both $\dot a$ and $\dot b$ are represented by the shown curve.}
\label{44}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic5/a_b_ddot}
\caption{The accelerations of the scale factors. Both $\ddot a$ and $\ddot b$ are represented by the shown curve.}
\label{45}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic5/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{46}
\end{subfigure}
\caption{Dust-filled brane world with initial conditions set number 5.}
\label{Fig10}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic5/sigma}
\caption{The dilaton $\sigma$; same for all three $\dot k\left(0\right)$.}
\label{47}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic5/sigma_dot}
\caption{The dilatonic field strength $\dot\sigma$.}
\label{48}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic5/phi}
\caption{The universal axion $\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve). The solution diverges for $\dot k\left(0\right)=0$.}
\label{49}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic5/phi_dot}
\caption{The axionic field strength $\dot\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve).}
\label{50}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic5/xxx}
\caption{$ \langle \Xi | \dot{\Xi} \rangle$ for $\dot{k}(0)= 1$ (blue), and $\dot{k}(0) = -1$ (red).}
\label{51}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic5/omega}
\caption{$\Omega$ at $\dot k\left(0\right) = 1$ and $ \dot\sigma \left(0\right) =0 $.}
\label{52}
\end{subfigure}
\caption{Dust-filled brane world with initial conditions set number 5 (continued).}
\label{Fig101}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic6/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve. The curve for $b$ is scaled down by a factor of 40 to fit in the graph.}
\label{103}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic6/a_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve. The curve for $\dot b$ is scaled down by a factor of 40 to fit in the graph.}
\label{104}
\end{subfigure}
\\[9em]
\begin{subfigure}[b]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic6/a_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve (scaled down 40-fold).}
\label{105}
\end{subfigure}
\qquad
\begin{subfigure}[b]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic6/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{106}
\end{subfigure}
\caption{Dust-filled brane world with initial conditions set number 6.}
\label{Fig21}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic6/sigma}
\caption{The dilaton $\sigma$; same for all three $\dot k\left(0\right)$.}
\label{107}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic6/sigma_dot}
\caption{The dilatonic field strength $\dot\sigma$.}
\label{108}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic6/phi}
\caption{The universal axion $\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve). The solution diverges for $\dot k\left(0\right)=0$.}
\label{109}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic6/phi_dot}
\caption{The axionic field strength $\dot\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve).}
\label{110}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic6/xxx}
\caption{$ \langle \Xi | \dot{\Xi} \rangle$ for $\dot{k}(0)= 1$ (blue), and $\dot{k}(0) = -1$ (red).}
\label{111}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Dust/ic6/omega}
\caption{$\Omega$ at $\dot k\left(0\right) = 1$ and $ \dot\sigma \left(0\right) =0 $.}
\label{112}
\end{subfigure}
\caption{Dust-filled brane world with initial conditions set number 6 (continued).}
\label{Fig21}
\end{figure}
Comparing the graphs for the various initial conditions we conclude the following: It is clear that the behavior of the scaling factors $a$ and $b$ is directly related to the moduli. The norm $G_{i\bar j} \dot z^i \dot z^{\bar j}$ always starts at a large value then decays. Even when it increases a bit as in IC5 Fig. (\ref{43}) it then peaks and decays. As a consequence the brane inevitably expands; \emph{even} when given contracting initial conditions (IC5). The larger the norm's initial value the smaller the starting size of the brane. The norm asymptotes to zero as the brane settles in its expected negative acceleration after an initial short phase of rapid accelerated expansion in most cases (except IC4). This last is particularly suggestive if one considers it in cosmological terms; as it implies an early and short inflationary epoch that coincides with the initial dynamics of the hypermultiplet fields. The next layer of initial conditions belongs to the harmonic function $k$; for which we have chosen positive, negative and zero initial velocities. The vanishing initial velocity case leads to a vanishing $k$ for all times, which is problematic for $\phi$, $\dot\phi$ and $\langle \Xi | \dot{\Xi} \rangle$ and thus we take as trivial, while the positive and negative values lead to a well-defined behavior for the $\phi$, $\dot\phi$ and $\langle \Xi | \dot{\Xi} \rangle$ fields. It is clear that for most initial conditions (the exception is IC5) $k$ expands rapidly then asymptotes to a constant value which leads to a similar behavior for $\phi$, $\dot\phi$ and $\langle \Xi | \dot{\Xi} \rangle$. The dilaton $\sigma$ (with initial conditions $\sigma=\dot \sigma=0$ at $t=0$) and its field strength $\dot\sigma$ do not seem to be sensitive to the $k$ initial conditions and neither does the function $\Omega$. In short we conclude that an initial phase of large moduli decay coincides with a rapid change in the rest of the hypermultiplet fields; leading to inflation. While the large $t$ behavior for all fields is a Friedmann-like negatively accelerated expansion, and vanishing hypermultiplet field strengths. If we consider our own universe as such a brane, then all of this certainly makes perfect sense.
\section{A radiation-filled brane}
\label{rr}
For a radiation- filled brane
\begin{equation}
T_{tt}^{{\emph Brane}} = \rho\left(t\right) = \frac{1}{{a^4 }}, ~~~~~~~~~~ p(t) =\frac{1}{{3a^4 }},
\label{Rad}
\end{equation}
where $p\left( t \right)$ is the pressure of the fluid related to the density via the equation of state $p = \frac{1}{3}\rho$.
This leads to the Friedmann-like equations:
\begin{eqnarray}
3\left[ {\left( {\frac{{\dot a}}{a}} \right)^2 + \left( {\frac{{\dot a}}{a}} \right)\left( {\frac{{\dot b}}{b}} \right)} \right] &=& G_{i\bar j} \dot z^i \dot z^{\bar j} + \frac{1}{{a^4 }} \nonumber\\
2\frac{{\ddot a}}{a} + \left( {\frac{{\dot a}}{a}} \right)^2 + \frac{{\ddot b}}{b} + 2\left( {\frac{{\dot a}}{a}} \right)\left( {\frac{{\dot b}}{b}} \right) &=& - G_{i\bar j} \dot z^i \dot z^{\bar j} - \frac{1}{{3a^4 }}\nonumber\\
3\left[ {\frac{{\ddot a}}{a} + \left( {\frac{{\dot a}}{a}} \right)^2 } \right] &=& - G_{i\bar j} \dot z^i \dot z^{\bar j}.
\end{eqnarray}
The majority of the solutions in this section the norm $G_{i\bar j} \dot z^i \dot z^{\bar j}$ is negative for all times and we plot its absolute value, \emph{except} for IC4 where it is positive from the start.
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic1/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve (flat on the $t$ axis), while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{53}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic1/a_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve (flat on the $t$ axis).}
\label{54}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic1/a_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve (flat on the $t$ axis).}
\label{55}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic1/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{56}
\end{subfigure}
\caption{Radiation-filled brane world with initial conditions set number 1.}
\label{Fig11}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic1/sigma}
\caption{The dilaton $\sigma$; same for all three $\dot k\left(0\right)$.}
\label{57}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic1/sigma_dot}
\caption{The dilatonic field strength $\dot\sigma$.}
\label{58}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic1/phi}
\caption{The universal axion $\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve). The solution diverges for $\dot k\left(0\right)=0$.}
\label{59}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic1/phi_dot}
\caption{The axionic field strength $\dot\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve).}
\label{60}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic1/xxx}
\caption{$ \langle \Xi | \dot{\Xi} \rangle$ for $\dot{k}(0)= 1$ (blue), and $\dot{k}(0) = -1$ (red).}
\label{61}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic1/omega}
\caption{$\Omega$ at $\dot k\left(0\right) = 1$ and $ \dot\sigma \left(0\right) =0 $.}
\label{62}
\end{subfigure}
\caption{Radiation-filled brane world with initial conditions set number 1 (continued).}
\label{Fig12}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic2/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{63}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic2/a_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve (flat on the $t$ axis).}
\label{64}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic2/a_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve (flat on the $t$ axis).}
\label{65}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic2/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{66}
\end{subfigure}
\caption{Radiation-filled brane world with initial conditions set number 2.}
\label{Fig13}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic2/sigma}
\caption{The dilaton $\sigma$; same for all three $\dot k\left(0\right)$.}
\label{67}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic2/sigma_dot}
\caption{The dilatonic field strength $\dot\sigma$.}
\label{68}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic2/phi}
\caption{The universal axion $\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve). The solution diverges for $\dot k\left(0\right)=0$.}
\label{69}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic2/phi_dot}
\caption{The axionic field strength $\dot\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve).}
\label{70}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic2/xxx}
\caption{$ \langle \Xi | \dot{\Xi} \rangle$ for $\dot{k}(0)= 1$ (blue), and $\dot{k}(0) = -1$ (red).}
\label{71}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic2/omega}
\caption{$\Omega$ at $\dot k\left(0\right) = 1$ and $ \dot\sigma \left(0\right) =0 $.}
\label{72}
\end{subfigure}
\caption{Radiation-filled brane world with initial conditions set number 2 (continued).}
\label{Fig14}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic3/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve (almost flat on the $t$ axis), while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{73}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic3/a_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve.}
\label{74}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic3/a_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve.}
\label{75}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic3/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{76}
\end{subfigure}
\caption{Radiation-filled brane world with initial conditions set number 3.}
\label{Fig15}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic3/sigma}
\caption{The dilaton $\sigma$; same for all three $\dot k\left(0\right)$.}
\label{77}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic3/sigma_dot}
\caption{The dilatonic field strength $\dot\sigma$.}
\label{78}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic3/phi}
\caption{The universal axion $\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve). The solution diverges for $\dot k\left(0\right)=0$.}
\label{79}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic3/phi_dot}
\caption{The axionic field strength $\dot\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve).}
\label{80}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic3/xxx}
\caption{ $ \langle \Xi | \dot{\Xi} \rangle$ for $\dot{k}(0)= 1$ (blue), and $\dot{k}(0) = -1$ (red).}
\label{81}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic3/omega}
\caption{$\Omega$ at $\dot k\left(0\right) = 1$ and $ \dot\sigma \left(0\right) =0 $.}
\label{82}
\end{subfigure}
\caption{Radiation-filled brane world with initial conditions set number 3 (continued).}
\label{Fig16}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic4/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $ {G_{i\bar j} \dot z^i \dot z^{\bar j}} $ (already positive) by the green curve.}
\label{83}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic4/a_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve.}
\label{84}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic4/a_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve.}
\label{85}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic4/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{86}
\end{subfigure}
\caption{Radiation-filled brane world with initial conditions set number 4.}
\label{Fig17}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic4/sigma}
\caption{The dilaton $\sigma$; same for all three $\dot k\left(0\right)$.}
\label{87}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic4/sigma_dot}
\caption{The dilatonic field strength $\dot\sigma$.}
\label{88}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic4/phi}
\caption{The universal axion $\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve). The solution diverges for $\dot k\left(0\right)=0$.}
\label{89}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic4/phi_dot}
\caption{The axionic field strength $\dot\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve).}
\label{90}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic4/xxx}
\caption{$ \langle \Xi | \dot{\Xi} \rangle$ for $\dot{k}(0)= 1$ (blue), and $\dot{k}(0) = -1$ (red).}
\label{91}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic4/omega}
\caption{$\Omega$ at $\dot k\left(0\right) = 1$ and $ \dot\sigma \left(0\right) =0 $.}
\label{92}
\end{subfigure}
\caption{Radiation-filled brane world with initial conditions set number 4 (continued).}
\label{Fig18}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic5/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{93}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic5/a_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve.}
\label{94}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic5/a_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve.}
\label{95}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic5/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{96}
\end{subfigure}
\caption{Radiation-filled brane world with initial conditions set number 5.}
\label{Fig19}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic5/sigma}
\caption{The dilaton $\sigma$; same for all three $\dot k\left(0\right)$.}
\label{97}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic5/sigma_dot}
\caption{The dilatonic field strength $\dot\sigma$.}
\label{98}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic5/phi}
\caption{The universal axion $\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve). The solution diverges for $\dot k\left(0\right)=0$.}
\label{99}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic5/phi_dot}
\caption{The axionic field strength $\dot\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve).}
\label{100}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic5/xxx}
\caption{$ \langle \Xi | \dot{\Xi} \rangle$ for $\dot{k}(0)= 1$ (blue), and $\dot{k}(0) = -1$ (red).}
\label{101}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic5/omega}
\caption{$\Omega$ at $\dot k\left(0\right) = 1$ and $ \dot\sigma \left(0\right) =0 $.}
\label{102}
\end{subfigure}
\caption{Radiation-filled brane world with initial conditions set number 5 (continued).}
\label{Fig20}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic6/a_b_zz}
\caption{The scale factor $a$ is represented by the blue curve, $b$ by the red curve, while $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ by the green curve.}
\label{113}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic6/a_b_dot}
\caption{The expansion rates of the scale factors: $\dot a$ is represented by the blue curve, and $\dot b$ by the red curve (flat on the $t$ axis).}
\label{114}
\end{subfigure}
\\[9em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic6/a_b_ddot}
\caption{The accelerations of the scale factors: $\ddot a$ is represented by the blue curve, and $\ddot b$ by the red curve (flat on the $t$ axis).}
\label{115}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic6/k}
\caption{The harmonic function $k$ using: $\dot k\left(0\right)=1$ (blue curve), $\dot k\left(0\right)=0$ (green line), and $\dot k\left(0\right)=-1$ (red curve).}
\label{116}
\end{subfigure}
\caption{Radiation-filled brane world with initial conditions set number 6.}
\label{Fig23}
\end{figure}
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic6/sigma}
\caption{The dilaton $\sigma$; same for all three $\dot k\left(0\right)$.}
\label{117}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic6/sigma_dot}
\caption{The dilatonic field strength $\dot\sigma$.}
\label{118}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic6/phi}
\caption{The universal axion $\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve). The solution diverges for $\dot k\left(0\right)=0$.}
\label{119}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic6/phi_dot}
\caption{The axionic field strength $\dot\phi$ for $\dot k\left(0\right) = 1$ (blue curve), and $\dot k\left(0\right) = -1$ (red curve).}
\label{120}
\end{subfigure}
\\[4em]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic6/xxx}
\caption{$ \langle \Xi | \dot{\Xi} \rangle$ for $\dot{k}(0)= 1$ (blue), and $\dot{k}(0) = -1$ (red).}
\label{121}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=0.7\columnwidth]{./Graphics/Radiation/ic6/omega}
\caption{$\Omega$ at $\dot k\left(0\right) = 1$ and $ \dot\sigma \left(0\right) =0 $.}
\label{122}
\end{subfigure}
\caption{Radiation-filled brane world with initial conditions set number 6 (continued).}
\label{Fig24}
\end{figure}
\newpage
The first thing we notice in this solution is that the bulk's scale factor $b$ does not match the brane's scale factor $a$ as it generally did for the dust case, even when they have the same initial conditions. It either stays constant at its initial value, or quickly asymptotes to some specific number. Comparing with the dust case, it is also clear that $a$ expands faster for the radiation filled brane, which is expected. It is also obvious here that the norm ${G_{i\bar j} \dot z^i \dot z^{\bar j}} $ is more \emph{strongly} coupled to $a$ than it is to $b$. The behavior of the remaining fields is very similar to the dust case, albeit with different rates of change.
\chapter{General Relativity notes}
\label{GR}
In this appendix, we will review some basic concepts in general relativity \cite{MUKHANOV2005}, \cite{Hobson2006}, \cite{Wald:1984}, \cite{Gary}, \cite{Frieman2015}, \cite{Terzic2008}, \cite{Bellido-lec}, \cite{Emam2018}.
\section*{The equivalence principle}
The effect of gravity can be cancelled locally by using free falling frame.\\
\section*{The geodesic equation}
The world line of a particle freely falling under gravity is given in general by the geodesic equation
\begin{equation}%
\frac{d^2 x^\mu}{d \tau^2} + \Gamma^\mu_{\nu\sigma} \frac{d x^\nu}{d \tau} \frac{d x^\sigma}{d\tau} = 0,
\end{equation}%
where $\Gamma^\mu_{\nu\sigma}$ are called the Christoffel symbol, or the affine connection
\begin{equation}%
\Gamma^\lambda_{\mu\nu}= \frac{1}{2} g^{\lambda\kappa} [ (\partial_\mu g_{\nu\kappa}) + (\partial_\nu g_{\mu\kappa}) - ( \partial_\kappa g_{\mu\nu}) ] .
\end{equation}%
\section*{The curvature tensor}
The Riemann tensor is defined by:
\b
R^{\sigma}_{~\mu\rho\nu}= (\partial_\rho \Gamma^\sigma_{\mu\nu} ) - (\partial_\mu \Gamma^\sigma_{\rho\nu} ) + \Gamma^\alpha_{\mu\nu}
\Gamma^\sigma_{\alpha\rho} - \Gamma^\alpha_{\rho\nu} \Gamma^\sigma_{\alpha\mu}
\end{equation}%
It is related to the curvature of a manifold in such that at flat region of a manifold, we may choose coordinates where the line element takes the form:
\begin{equation}%
ds^2=\epsilon_1 (dX^1)^2 + \epsilon_2 (dX^2)^2 + … + \epsilon_N (d X^N)^2,
\nonumber%
\end{equation}%
where $\epsilon_a=\pm 1$
throughout the region. In these coordinates $\Gamma^\lambda_{\mu\nu}$ and its derivatives are zero, and hence
\begin{equation}%
R^{~~~~\sigma}_{\mu\rho\nu}=0
\nonumber%
\end{equation}%
While for the manifold the line element is given by:
\begin{equation}%
ds^2= g_{\mu\nu} (x) dx^\mu dx^\nu ,
\nonumber%
\end{equation}%
and the covariant derivative of a vector field takes the form:
\begin{equation}%
\nabla_\nu v_\mu = \partial_\nu v_\mu - \Gamma^\lambda_{\mu\nu} v_\lambda.
\nonumber%
\end{equation}%
The Ricci tensor is given by :
\begin{equation}%
R_{\mu\nu} =R^{~~~~\rho}_{\mu\rho\nu},
\nonumber%
\end{equation}%
while the Ricci scalar
\begin{equation}%
R= g^{\mu\nu} R_{\mu\nu},
\nonumber%
\end{equation}%
and the Einstein tensor
\begin{equation}%
G^{;\mu\nu} \equiv R^{\mu\nu}- \frac{1}{2} g^{\mu\nu} R.
\end{equation}%
\section*{The gravitational field equations}
In an IRF ( instantaneous rest frame ) the components of the energy–momentum tensor of a perfect fluid are given by
\b
[T^{\mu\nu}]= \left(\begin{array}{cccc}
\rho c^2 & 0 & 0 & 0 \\
0 & p & 0 & 0 \\
0 & 0 & p & 0 \\
0 & 0 & 0 & p \\
\end{array}
\right),
\end{equation}
so that
\begin{equation}%
T^{\mu\nu}=(\rho + p/c^2) u^\mu u^\nu - p \delta^{\mu\nu} ,
\label{EMT}
\end{equation}%
where $u^\mu=(c,0,0,0)$.
Einstein field equations are given by:
\begin{equation}
G_{\mu\nu} = \kappa T_{\mu\nu},
\label{EFEA1}
\end{equation}
or:
\begin{equation}%
R_{\mu\nu}- \frac{1}{2} g_{\mu\nu} R = -\kappa T_{\mu\nu},\label{FE}%
\end{equation}%
where $\kappa= 8\pi G/c^4$. Or in another form
\begin{equation}%
R_{\mu\nu}= - \kappa(T_{\mu\nu}-\frac{1}{2} T g_{\mu\nu}),
\end{equation}%
where $T \equiv T^\mu_\mu$.
\section*{The Schwarzschild geometry}
Schwarzschild explored the metric $g_{\mu\nu}$
representing the static spherically symmetric gravitational field in the empty space surrounding some massive spherical object such as a star.
The Schwarzschild metric for the empty spacetime outside a spherical body of mass M is
\begin{equation}%
ds^2 = c^2 \bigg( 1- \frac{2\mu}{ r} \bigg) dt^2- \bigg( 1 - \frac{2\mu}{r} \bigg)^{-1}
dr^2 - r^2 d\theta^2 - r^2 \sin^2 \theta d\phi^2 ,
\label{sch}
\end{equation}%
where $\mu=GM/c^2$.
Clearly, the metric functions are infinite at $r=2\mu$ which is known as the Schwarzschild radius.
\newpage
\section*{The Friedmann–Robertson–Walker geometry}
\vspace{-0.1em}
(The maximal symmetric 3-dimensional space)
\subsection*{The cosmological principle}
At any particular time the universe looks the same from all positions in space at a particular time and all directions in space at any point are equivalent.
Or: on the very largest scales, to high accuracy the universe is isotropic and homogenous. The line element takes the form
\begin{equation}%
ds^2= c^2 dt^2 - g_{ij} dx^i dx^j ,
\nonumber%
\end{equation}%
where the $g_{ij}$ are functions of the coordinates $(t,x^i)$. However to incorporate the time independence , the time can enter the $g_{ij}$ only through a common factor , so that the ratios of small distances are the same at all times. Hence the metric must take the form
\begin{equation}%
ds^2= c^2 dt^2 - S^2(t) h_{ij} dx^i dx^j ,
\end{equation}%
where S(t) is a time-dependent scale factor and the $h_{ij}$ are functions of the coordinates $x^i$.
Accordingly after some calculations (\ref{sch}) becomes
\begin{equation}%
ds^2= c^2 dt^2 - a^2(t) \Bigg[\frac{dr^2}{1-kr^2} + r^2 (d\theta^2 +sin^2 \theta d\phi^2)\Bigg], \label{FRW-m2}
\end{equation}%
which's called Friedmann–Robertson–Walker (FRW) metric. a(t) is a scale function. $k$ takes the values −1, 0, or 1 depending on whether the spatial section has negative (pseudo-sphere, $a^2 < 0$ ) , zero (2- dimension flat space) or positive ( sphere, $a^2 > 0$) curvature respectively. It is also clear that the coordinates r appearing in the FRW metric are still comoving, i.e. the worldline of a galaxy, ignoring any peculiar velocity, has fixed values of $(r,\theta,\phi)$. Another more convenient form
\begin{equation}%
ds^2 = c^2 dt^2 - a^2(t) \big[d\chi^2+ S^2 (\chi) (d\theta^2+ \sin^2 \theta d\phi^2)\big]
\end{equation}%
where the function $S(\chi) =( \sin \chi, \chi, \sinh \chi)$ if k= ( 1,0,-1) , respectively. \\
\large{The Hubble parameter}
\begin{equation}%
H(t)=\frac{\dot a(t)}{a(t)}.
\end{equation}%
\large{The deceleration parameter}
\begin{equation}%
q(t) = - \frac{\ddot a(t) a(t)}{\dot a^2(t)}.
\nonumber%
\end{equation}%
{\large Hubble’s law}
\begin{equation}%
v= H_0 d,
\label{HL}
\end{equation}%
The galaxies appear to recede from us with a recession speed proportional to their distance from us , (the expansion of the universe- 1929).
\vspace{0.5cm}
\textbf {The cosmological field equations (FRW equations):}
Solving the gravitational field equations \eqref{FE} with the FRW metric \eqref{FRW-m2} yields:
\begin{equation}%
\nonumber%
\dot a^2 = \frac{8\pi G}{3}\rho a^2 -c^2 k,
\end{equation}%
\begin{equation}%
\ddot a = - \frac{4 \pi G}{3} \bigg(\rho + \frac{3p}{c^2}\bigg) a.
\label{fr}
\end{equation}%
These two differential equations determine the time evolution of the scale factor a(t)( literally control the time evolution of the universe , for instances they tell the universe can‘t be static) and are known as the Friedmann–Lemaître equations.
\section*{The Continuity equation}
The energy density is conserved according to
\begin{equation}%
\dot \rho = - 3 ( \rho + p ) H,
\end{equation}%
for dust p $<< \rho $, $\dot \rho =-3 \rho H$.
\bigskip
- Each component of the cosmological fluid is modelled as a perfect fluid with an equation of state of the form
\begin{equation}%
p_i = w_i \rho_i c^2,
\end{equation}%
where $w_i=0$ for pressure less ‘dust’, $w_i=1/3$ for radiation and $w_i=-1$ for the vacuum.
\section*{Cosmological parameters}
The density parameters or simply densities, are dimensionless parameters defined by
\begin{equation}%
\Omega_i(t)\equiv \frac{8\pi G}{3H^2(t)} \rho_i(t),
\end{equation}%
where i denotes 'm', 'r' or $'\Lambda'$, and $\rho$ is the mass density. Define the total density in the universe
\begin{equation}%
\Omega= \Omega_m+ \Omega_r + \Omega_\Lambda =1-\Omega_k,
\end{equation}%
where $\Omega_k(t)= - \frac{c^2 k}{H^2(t)a^2(t)}$ is called the curvature density parameter. So it’s clear that $\Omega$ determine the spatial curvature of the universe, see Fig. (\ref{dp}).
\begin{figure}[!t]
\centering
\includegraphics[scale=0.55]{./Fig/dp.pdf}
\caption{Possible shapes of the universe depending on the density parameter $\Omega$.}
\label{dp}
\end{figure}
\section*{A variational approach to general relativity}
The action S for a set of fields defined on some general four-dimensional spacetime manifold should take the form of an integral of the Lagrangian density, of the fields $\phi^a$ and their first ( and possibly higher) derivatives over some four-dimensional region $R$ of the spacetime %
\begin{equation}%
S = \int_\mathcal{R} \mathcal{L} (\Phi^a, \partial_\mu \Phi^a, \partial_\mu \partial_\nu \Phi^a, ..) d^4 x.
\end{equation}%
Define the field Lagrangian L
\begin{equation}%
\mathcal{L} = L \sqrt{- g},
\nonumber
\end{equation}%
where g is the determinant of the metric tensor in that coordinate system. The Euler- Lagrange (EL) equations can be found easily by the principle of least action $\delta S=0$
\begin{equation}%
\frac{\partial \mathcal{L}}{\partial \Phi^a} - \partial_\mu \bigg [\frac{\partial \mathcal{L}}{\partial (\partial_\mu \Phi^a)} \bigg] = 0.
\end{equation}%
\section*{The Einstein–Hilbert action}
\begin{equation}%
S_{EH}= \int_\mathcal{R} R \sqrt{-g} ~ d^4 x.
\label{EHA}
\end{equation}%
By demanding that $\delta S_{SH}=0$ and using the fact that the variation $\delta g^{\mu \nu}$ is arbitrary, we thus recover Einstein’s field equations in vacuo:%
\begin{equation}%
G_{\mu\nu} \equiv g_{\mu\nu}-\frac{1}{2} g_{\mu\nu} R =0.
\nonumber
\end{equation}%
In the presence of matter
\begin{equation}%
S= \frac{1}{2\kappa} S_{EH} + S_M =
\int_\mathcal{R} \bigg( \frac{1}{2\kappa} \mathcal{L}_{EH}+\mathcal{L}_M \bigg)~ d^4 x,
\nonumber
\end{equation}%
where $\kappa=8\pi G/c^4$. Let us consider the varying of the action with respect to the (inverse) metric , to obtain
\begin{equation}%
\frac{1}{2\kappa} \frac{\delta \mathcal{L}_{EH}}{\delta g^{\mu\nu}} + \frac{\delta \mathcal{L}_M}{\delta g^{\mu\nu}} =0.
\nonumber%
\end{equation}%
But
\begin{equation}%
\frac{\delta \mathcal{L}_{EH}}{\delta g^{\mu\nu}} = \sqrt{-g} G_{\mu \nu},
\nonumber%
\end{equation}%
Define the energy - momentum tensor of the non-gravitational fields (or ‘matter’) by
\begin{equation}%
T_{\mu\nu}=\frac{2}{\sqrt{-g}} \frac{\delta \mathcal{L}_M}{\delta g^{\mu\nu}}.
\nonumber%
\end{equation}%
Then we recover the full Einstein Equations
\begin{equation}%
G_{\mu\nu} = - \kappa T_{\mu\nu}.
\nonumber
\end{equation}%
\section*{The cosmological constant \protect\footnote{\cite{BookME}.}}
During the discovery of the general relativity by Albert Einstein in 1915, it was thought that the universe is static. While by looking at the first equation of Friedmann equations (\ref{fr}), we will find for a flat universe $k=0$, the scale factor $\left(\frac{\dot{a}}{a}\right)$ vanishes only if the universe is empty from any matter or energy $\rho=0$. In 1917 Einstein added the
cosmological constant term to the right- hand side of the first Friedmann equation to cancel the effect of the density, and to guarantee a vanishing $\dot{a}$. The complete Einstein field equations (\ref{EFEA1}) become:
\begin{equation}
G_{\mu\nu} +\Lambda g_{\mu\nu} = \kappa T_{\mu\nu}.
\end{equation}
The cosmological constant $\Lambda$ is assumed to be a positive real number. It can be accounted as a part of the matter/energy content of the universe by moving it to the right- hand side:
\begin{equation}
G_{\mu\nu} = \kappa T_{\mu\nu}-\Lambda g_{\mu\nu}.
\end{equation}
As we show this term generates a negative pressure, and thus acts as a `repulsive` force, counteracting the attractive gravitational force in the universe. Later, in 1929 Edwin Hubble formulated Hubble's law (\ref{HL}) which states that the universe is expanding ($\dot{a} >0$). Einstein returned and removed the cosmological constant from his equations, calling it his `greatest blunder`, because he predicted the expansion of the universe theoretically. In 1998 the cosmological society was astonished by the discovery the speed of the expansion of the universe is indeed accelerating with time. A discovery that earned the $2011$ Nobel prize in physics
\cite{Lambda-Art, Lambda-A}. The research teams was observing type $Ia$ supernovae \cite{supernova}, and found that over 50 distant supernova whose light was weaker than expected.
This was a strong sign that the universe has an accelerated expansion ($\ddot{a} >0 $).
Nowadays \footnote{The age of the universe is 14 billion years since the big bang.} the cosmological constant term is added to Einstein equations to account for this late- time accelerated expansion. However, it's not clear yet what is the nature of $\Lambda$, sometimes it's referred by dark energy, some kind of unkonwn energy who has a repulsive effect and drive the universe's acceleration. On the other hand the cosmological constant term is related to the vacuum energy, consider for instance the energy- momentum tensor of a scalar field
\begin{equation}%
T^\alpha_\beta = \frac{1}{2}\partial^\alpha \phi \partial_\beta \phi - \bigg( \frac{1}{2} \partial^\gamma \phi \partial_\gamma \phi - V(\phi) \bigg) \delta^\alpha_\beta,
\end{equation}%
It can be in the form of a perfect fluid
\begin{equation}%
T^{\mu\nu}=(\rho + p/c^2) u^\mu u^\nu - p \delta^{\mu\nu} ,
\label{EMT}
\end{equation}%
by defining
\begin{equation}
\rho \equiv \frac{1}{2} \dot \phi^2 + V(\phi), ~
p \equiv\frac{1}{2} \dot \phi^2 - V(\phi), ~
u^\alpha \equiv \partial^\alpha/ \sqrt{\partial^\gamma \phi \partial_\gamma \Phi}.
\end{equation}%
If the potential $V(\phi)$ has a local minimum at some point $\phi_0$ then
\begin{equation}
\rho = - p = - V(\phi_0) ,
~~~ T^\alpha_\beta = V(\phi_0) \delta^\alpha_\beta.
\end{equation}%
Substituting in the Einstein equations imitates a cosmological term
\begin{equation}
G_{\mu\nu} - \Lambda g_{\mu\nu} = - \kappa V(\phi_0) g_{\mu\nu} ,
\end{equation}
means that $\Lambda= \kappa V(\phi_0)$. So that
the cosmological term can always be interpreted as the contribution of
vacuum energy to the Einstein equations. The value of the theoretical vacuum energy density
arises from the quantum field theory is given by
\begin{equation}
\rho^{(th)}_{\Lambda} = (10^{18} ~ \text{GeV})^4 ,
\end{equation}
while the cosmological observations imply
\begin{equation}
\rho^{(obs)}_{\Lambda} \leq (10^{-12} ~ \text{GeV})^4.
\end{equation}
This huge discrepancy between the theoretical and observational values is called the cosmological constant problem.
\chapter[Differential forms]{Differential forms\protect\footnote{\cite{CYM2-2}.}}
\label{Df}
\section{Differential forms on manifolds}
Consider a D-dimensional Riemannian/Lorentzian manifold $\mathcal{M}$ . A differential form $\omega$ or
$\omega_p$ of order p on $\mathcal{M}$, also known as a p-form, is a totally antisymmetric tensor of type (0, p).
It may be defined in terms of the differentials $dx^\mu$ , for instance
\begin{equation} \nonumber \text{1- form} : \omega_{(1)}=\omega_0~ dx, \end{equation}
\begin{equation} \nonumber \text{2- form} : \omega_{(2)} = \omega_0~ dx ~ dy.\end{equation}
But any surface has a direction upward and downward, like the flux. So define the wedge product:
\begin{equation} \nonumber dx \wedge dy = - dy \wedge dx, \end{equation}
So that:
\begin{equation} \nonumber \text{3-form} : \omega_{(3)}= \omega_0~ dx \wedge dy \wedge dz, \end{equation}
\begin{equation} \nonumber \text{4-forms} : \omega_{(4)}=\omega_0~ dt \wedge dx \wedge dy \wedge dz. \end{equation}
The differential form component should be antisymmetric, because one can define
\begin{equation} \nonumber \omega_{(2)}= \omega_{12}~ dx \wedge dy= w_{21}~ dy \wedge dx, \end{equation}
leads to $\omega_{12}= - \omega_{21}$, in general
\begin{equation} \nonumber \omega_{(i)} = \frac{1}{2}\omega_{ij} dx^i \wedge dx^j, \end{equation}
\begin{equation} \nonumber \omega_{(3)}= \frac{1}{3!} \omega_{ijk}~ dx^i \wedge dx^j \wedge dx^k, \end{equation}
In the standard way, 1-forms themselves acting as basis
\begin{equation} \nonumber \omega_p = \frac{1}{p!} ~ \omega_{p_i ……p_p}~ dx^{\mu_i} \wedge …………\wedge dx^{\mu_p}. \end{equation}
The so-called wedge product $\wedge$ is defined such that:
\begin{equation} \nonumber dx^{\mu_i} \wedge dx^{\mu_i}=0. \end{equation}
\begin{equation} \nonumber dx^{\mu_i} \wedge dx^{\mu_j}= - dx^{\mu_j} \wedge dx^{\mu_i}. \end{equation}
\subsection*{The exterior derivative}
The so-called exterior derivative $d= dx^\nu~ \partial_\nu$ is defined as an operator that maps p-forms
into (p + 1)-forms for instance:
\begin{equation}
d \omega_{(3)} = \frac{1}{3!} (\partial_\alpha w_{\mu\nu\rho}) ~ dx^\alpha \wedge dx^\mu \wedge dx^\nu \wedge dx^\rho = w_{(4)},\end{equation}
in general:
\begin{equation}
\lambda_{p+1}= d \omega_p = dx^\nu~ \partial_\nu~ \omega_p =\frac{1}{p!} ~ ( \partial_\nu~ \omega_{\mu_1 ... \mu_p}) ~~ dx^\nu \wedge dx^{\mu_1} \wedge ... \wedge dx^{\mu_p} ,
\end{equation}
satisfying the product rule
\begin{equation}
d (\omega_p \wedge \eta_r) = (d \omega_p) \wedge \eta_r + (-1)^p \omega_p \wedge (d\eta_r),
\end{equation}
as well as
\begin{equation}
d^2 = 0.
\label{cc}
\end{equation}
An operator satisfying (\ref{cc}) is called nilpotent. Differential forms that can be written
as exterior derivatives of other forms, such as $\lambda = d \omega$, are called exact, while forms whose
exterior derivative vanishes, e.g. $d \lambda = 0$, are called closed. Because of (\ref{cc}), exact forms are
always closed, while the converse is not necessarily true. For example;
\begin{eqnarray} \nonumber A &=& A_\mu dx^\mu = A_0 dt +A_i dx_i ,\\ \nonumber
dA &=& \frac{1}{2} (\partial_\alpha A_\mu ) ~ dx^\alpha \wedge dx^\mu, \\ \nonumber &=&
(\partial_\alpha A_\mu) dx^\alpha \wedge dx_ \mu + (\partial_\mu A_\alpha) dx^\mu \wedge dx^\alpha ,\\ \nonumber F&=&( \partial_\alpha A_\mu - \partial_\mu A_\alpha) dx^\alpha \wedge dx^\mu, \\ \nonumber
dF &=& d^2 A =0.
\end{eqnarray}
\subsection*{Hodge operator}
Define the Hodge-duality operator $\star$, mapping p-forms into
(D - p)- forms, for instance:
\begin{equation}
\star( dx \wedge dy) = dt \wedge dz,
\end{equation}
\begin{equation}
\star F_{\mu\nu} = F^{\mu\nu} ,
\end{equation}
so that
\begin{equation}
\mathcal{L}_{EM} = \frac{1}{4} F \wedge \star F,
\end{equation}
also in this language, the volume form is the Hodge dual of the identity,
\begin{equation} \star 1 = \sqrt{|g|} ~ dx^1 \wedge ...\wedge dx^D.
\end{equation}
\section{The vielbein formulation of the general relativity}
Consider a Riemannian manifold with a line element $ds^2 = g_{\mu\nu} dx^\mu dx^\nu$ ,
where $g_{\mu\nu}$ is a gauge field, and the curvature tensor $R^{\sigma}_{~\mu\rho\nu}$ as the field strength.
Consider this manifold is Lorentzian, which means every point on this curved manifold is Lorantz invariant, i.e., every point is locally flat.
Take a tangent plane to every point, and define on it little vectors called beins, in which they can transform under local special relativistic transformations, and under global general coordinates transformations, because they can move on the curved manifold, so they have two indices $e_\mu^{\hat{\alpha}}$. We can define a $1$- form
\begin{equation} e^{\alpha} = e^{\hat{\alpha}}_\mu ~ dx^\mu, \end{equation}
in this language:
\begin{equation} g_{\mu\nu} = e_\mu^{\hat{\alpha}} ~ e_\nu^{\hat{\beta}} ~ \eta_{\hat{\alpha} \hat{\beta}}, \end{equation}
\begin{equation} \nonumber ds^2 = e_\mu^{\hat{\alpha}} ~ e_\nu^{\hat{\beta}} ~\eta_{\hat{\alpha} \hat{\beta}}~ dx^\mu dx^\nu, \end{equation}
\begin{equation} ds^2 = e^{\hat\alpha} e^{\hat\beta} \eta_{\hat\alpha \hat\beta}, \end{equation}
with
\begin{equation} e^{\hat\alpha}_\mu= e^{\hat\alpha}_\nu~~ \frac{\partial x^\nu}{\partial x^\mu}= e^{\hat\beta}_\mu~~ \Lambda^{\hat \alpha}_{\hat \beta}. \end{equation}
Define the $1$-form spin connection:
\begin{equation}
\omega^{\hat{\alpha} \hat{\beta}} = \omega^{\hat{\alpha}\hat{\beta}}_\mu dx^\mu,
\end{equation}
\begin{equation} \omega_\mu^{\hat\alpha\hat\beta}= e^{\hat\alpha \rho} (\Delta_\mu e^{\hat \beta}_\rho ) = e^{\hat\alpha\rho} (\partial_\mu e^{\hat\beta}_\rho - \Gamma^\nu_{\mu \rho} e^{\hat\beta}_\nu),
\end{equation}
roughly speaking the Bein is the square root of the metric and the spin connection is the square of the Christoffel symbol.
Spin connection couples spinors to the GR. Now the Ricci tensor is given by
\begin{eqnarray}
\nonumber R^{\hat \alpha \hat \beta} &=&
d \omega^{\hat \alpha \hat \beta} + \omega^{\hat\alpha}_{\hat\rho} \wedge \omega^{\hat\rho \hat \beta},\\
R^{\hat \alpha \hat \beta} &=& R^{\hat \alpha \hat \beta }_{\mu\nu} ~ dx^\mu \wedge dx^\nu,
\end{eqnarray}
and the Riemann tensor:
\begin{equation}
R_{\mu\nu\rho r} =
e_{\hat\alpha \mu } e_{\hat \beta \nu } R^{\hat \alpha \hat \beta}_{\rho r}.
\end{equation}
For the torsion $\tau_{\mu\nu}^\rho = \Gamma^\rho_{\mu\nu}-\Gamma^\rho_{\nu\mu}=0,$ define the 2-form
\begin{equation} \tau^{\hat\alpha}~= d e^{\hat\alpha} + \omega^{\hat\alpha}_{\hat r} e^{\hat r}, \end{equation}
where
\begin{equation} \omega^{\hat\alpha}_{\hat{r}} = \eta_{\hat r \hat \rho} \omega^{\hat \alpha\hat\rho}.
\end{equation}
\begin{eqnarray} \nonumber
\tau^{\hat \alpha}&=& \tau^{\hat \alpha}_{\mu\nu} dx^\mu \wedge dx^\nu ,\\
\tau^\rho_{\mu\nu} &=& e^\rho_{\hat \alpha} \tau^{\hat \alpha}_{\mu\nu}.
\end{eqnarray}
Since the gravitino has two indices $\Psi^{\mu a}$, it acts as the gauge field which has to be introduced to make local SUSY. Define the covariant derivative that couples $\Psi$ to gravity
\begin{equation} \tilde{\Delta}_\mu \Psi =
\partial_\mu \Psi+ \frac{1}{4} \omega^{\hat\alpha\hat\beta}_\mu \gamma_{\hat\alpha\hat\beta} \Psi,
\end{equation}
which means the spinors now transform in a curved background. The EH action becomes (Palatini action)
\begin{eqnarray} \nonumber S &=& \int d^4 x \sqrt{-|g|} R, \\
&=& \int d^4 x ~ |e| ~ e^\mu_{\hat \mu} e^\nu_{\hat \nu}~ R^{\hat \mu \hat \nu}_{\mu\nu}.
\end{eqnarray}
\chapter[Topics on geometery and topology]{Topics on geometery and topology \protect\footnote{\cite{He2018, Berglund1994}.}}
\label{cy}
\section{Calabi-Yau 3-fold as a hypersurface}
The hypersurface is a manifold of dimension $(n-1)$ embedded in an ambient space of dimension $n$, generally
an Euclidean space or a projective space $\mathbb{CP}$. Hypersurfaces can then defined by a single implicit
equation, at least locally (near every point) and sometimes globally.
\subsection*{The projective space}
Is attained by adding a " point at infinity " to an affine space. Hence, affine complex coordinates $(x,y)$
$\in \mathbb{C}^2$ is promoted to projective coordinates $(x,y,z)$ $\in \mathbb{C}^3$ with the extra identification
that $(x,y,z) \sim \lambda (x,y,z)$ for any non-zero $\lambda \in \mathbb{C}^x$. In other words, the elliptic curve is
cubic in $\mathbb{CP}^2$ with homogenous coordinates $[x:y:z]$.
Note that, some times the complex algebraic geometery is refered by $\mathbb{P}^n$. The Complex projective space is defined by
\begin{eqnarray}
\nonumber \mathbb{P}^n_{[z_1:z_2:...:z_n]} :=& \mathbb{C}^{n+1}_{(x_0,x_1,...,x_n)} / \sim ~ ;
\\ & (x_0,x_1,...,x_n) \sim \lambda (x_0,x_1,...,x_n).
\end{eqnarray}
Now let us see how a polynomial embedded in the above $\mathbb{P}^n$
\subsection*{The quintic}
So we can construct a 1- dimentional Calabi- Yau manifold like the torus a cubic in 3 dimensions in $\mathbb{P}^2 (n=2)$.
Note that:
\begin{itemize}
\item $1= 2-1$ : the dimension of the CY manifold is 1 less than that of the ambient space since it's defined by a single
polynomial; an example of a hypersurface.
\item $3= 2+1$ : the degree of the polynomial exceeds the dimension of the ambient projective space by 1,
and is thus equal to the number of the homogenous coordintes ; this is CY condition.
\end{itemize}
General speaking:
\begin{description}
\item [Calabi- Yau (n-1)- fold ($\mathcal{M}$):] can be constructed as a hypersurface in $\mathbb{P}^n$, i.e.,
as a single homogenous polynomial of degree d in $n+1$ projective coordinates. Litrally:
\item [ Proposition:] A homogenous degree $d=n+1$ polynomial (in $n+1$ projective coordinates)
as a hypersurface in $\mathbb{P}^n$ defines a Calabi-Yau (n-1)- fold.
\end{description}
The simplest example is the quintic $Q$ : It's a CY 3-fold with degree 5 and a polynomial in $\mathbb{P}^4$.
A generic quintic hypersurface in $\mathbb{P}^4(5)_{(z_0:...:z_4)}$ can be writtin as
\begin{equation}
Q:= \sum_{i=0}^4 z_i^5 - \psi \mathop{\Pi}^4_{i=0} z_i, ~~~~ \psi \in \mathbb{C},
\end{equation}
This is called "Fermat form" , with $\psi$ an explicit complex parameter.
\begin{equation}
\psi = h^{(2,1)} (Q) = 101,
\end{equation}
complex space deformation parameters, while K\"{a}hler class has only one parameter $h^{(1,1)} (Q)=1$.
\chapter{The Topology of the Calabi-Yau Manifold}
In the previous chapters we have solved Einstein field equations in case of world- branes filled with dust, radiation, or both,
and dust with radiation and dark energy ($\Lambda$). These solutions shows us the time evolution of the quantity
$G_{i\bar j} \dot z^i \dot z^{\bar j}$ \footnote{\textit {Which before we roughly called the moduli .}} at different initial conditions.
In this chapter, we will study each of the moduli $z^i$ and the metric $G_{i\bar j}$ dependence on time.
We take the dimension of the Calabi-Yau complex structure space $\mathcal{M}_c$, $h_{(2,1)}=1$, which means
we have a toy model with only one moduli $z$ and a single $\mathcal{M}_c$ metric component $G$.
Although this seems a simple case, as we will see it will get us closer for a deeper understanding
of the complex structure of the CY manifold.
Indeed there are many studies about the geometry of $\mathcal{M}_c$, like \cite{Candelas:355} and \cite{Candelas:359}, while in \cite{Belavin:2018}, \cite {Candelas:3551}, and \cite{Belavin:2017} the Calabi-Yau 3-fold is considered as a quintic threefold in the $\mathbb{P}^4$ projection space
\footnote{\em See Appendix (\ref{cy}).}. Here however we do not need to make the last assumption because we already know the time behavior of
$G_{i\bar j} \dot z^i \dot z^{\bar j}$, also the field equations of the moduli $z^i$ and its complex conjugate are well known through
$\mathcal{N}=2$ $\mathcal{D}=5$ supergravity, Equ. (\ref{zzz}) .
Since all solutions that we have introduced till now for the $\mathcal{N}=2$ $\mathcal{D}=5$ supergravity are numeric, in the last of the chapter, we solve
Friedmann- like equations (\ref{eee}) analytically in the case of $b=1$ and show how both solutions manifest $G_{i\bar j} \dot z^i \dot z^{\bar j}$ correlations to the brane scale factor $a(t)$.
\section{The moduli field equations}
Starting by the field equations of the moduli $z^i$ and its complex conjugate $z^{\bar{i}}$
\begin{eqnarray} \nonumber
& (\Delta z^i) \star 1 + \Gamma^i_{jk} dz^j \wedge \star dz^k + \frac{1}{2} e^\sigma G^{i\bar{j}} \partial_{\bar{j}} \langle d \Xi | \underset{\Lambda}{\mathbf {\Lambda}} | \star d\Xi \rangle =0 \\
& (\Delta z^{\bar{i}}) \star 1 + \Gamma^{\bar{i}}_{\bar{j}\bar{k}} dz^{\bar{j}} \wedge \star dz^{\bar{k}} + \frac{1}{2} e^\sigma G^{\bar{i}j} \partial_{j} \langle d \Xi | \underset{\Lambda}{\mathbf \Lambda} | \star d \Xi \rangle =0
\end{eqnarray}
But from the BPS condition Equ. (\ref{bps}), The moduli field equations become:
\begin{eqnarray} \nonumber
& (\Delta z^i) \star 1 + \Gamma^i_{jk} dz^j \wedge \star dz^k + G^{i\bar{j}} (\partial_{\bar{j}} G_{l\bar{k}}) dz^l \wedge \star dz^{\bar{k}} =0 \\
& (\Delta z^{\bar{i}}) \star 1 + \Gamma^{\bar{i}}_{\bar{j}\bar{k}} dz^{\bar{j}} \wedge \star dz^{\bar{k}} + G^{\bar{i}j} (\partial_j G_{l\bar{k}}) dz^l \wedge \star dz^{\bar{k}} =0.
\end{eqnarray}
Dropping the differential forms formulation, we get:
\begin{eqnarray} \nonumber
& \nabla^2 z^i + \Gamma^i_{jk} \partial_\mu z^j \partial^\mu z^k + G^{i\bar{j}} (\partial_{\bar{j}} G_{l\bar{k}})\partial_\mu z^l \partial^\mu z^{\bar{k}} =0 \\
& \nabla^2 z^{\bar{i}} + \Gamma^{\bar{i}}_{\bar{j}\bar{k}} \partial_\mu z^{\bar{j}} \partial^\mu z^{\bar{k}} +
G^{\bar{i}j} (\partial_{j} G_{l\bar{k}})\partial_\mu z^l \partial^\mu z^{\bar{k}} =0.
\label{zz23}
\end{eqnarray}
The connections or the Christoffel symbols are related to the metric by \cite{CYM2-2}:
\begin{equation}
\Gamma^i_{jk} = G^{i\bar{p}} \partial_j G_{k\bar{p}}, ~~~~~~~~ \Gamma^{\bar{i}}_{\bar{j}\bar{k}} = G^{p\bar{i}} \partial_{\bar{j}}
G_{\bar{k}p},
\end{equation}
substitute in Equ. (\ref{zz23}), we get:
\begin{eqnarray} \nonumber
& \nabla^2 z^i + G^{i\bar{p}} \partial_j G_{k\bar{p}} ~ \partial_\mu z^j \partial^\mu z^k + G^{i\bar{j}} (\partial_{\bar{j}} G_{l\bar{k}})\partial_\mu z^l \partial^\mu z^{\bar{k}} =0 \\
& \nabla^2 z^{\bar{i}} + G^{p\bar{i}} \partial_{\bar{j}}
G_{\bar{k}p}~ \partial_\mu z^{\bar{j}} \partial^\mu z^{\bar{k}} +
G^{\bar{i}j} (\partial_{j} G_{l\bar{k}})\partial_\mu z^l \partial^\mu z^{\bar{k}} =0
\label{zz34}\end{eqnarray}
The moduli are independent of the 3- spatial dimensions. And consider the Hodge number $h_{2,1} =1$, which means we have only one moduli $z$, its complex conjugate $z^*$, a single K\"{a}hler metric component $G$ and the dimension of the moduli space $\mathcal{M}_C$ is unity. So that
Equ. (\ref{zz34}) simplify to:
\begin{eqnarray} \nonumber
& g^{tt} \ddot{z} +\frac {1}{G} (\partial_z G) ~ \dot {z}^2 + \frac {1}{G} (\partial_{z^*} G) \dot{z} \dot{z}^* =0 \\
& g^{tt} \ddot{z}^* + \frac {1}{G^*}(\partial_{z^*} G^*) ~ \dot {z^*}^2 + \frac {1}{G^*} (\partial_z G) \dot{z} \dot{z}^* =0
\end{eqnarray}
From the Robertson-Walker like metric Equ. (\ref{metricE}), the moduli field equations become:
\begin{equation}
\ddot{z} +\frac {1}{G} (\partial_z G) ~ \dot {z}^2 + \frac {1}{G} (\partial_{z^*} G) \dot{z} \dot{z}^* =0
\label{z}
\end{equation}
\begin{equation}
\ddot{z}^* + \frac {1}{G^*}(\partial_{z^*} G^*) ~ \dot {z^*}^2 + \frac {1}{G^*} (\partial_z G) \dot{z} \dot{z}^* =0
\label{zstar}
\end{equation}
We can solve the equation of $\ddot{z}$ or $\ddot{z^*}$ with
the numeric solution obtained for the moduli velocity norm
$G_{i\bar{j}} \dot z^i \dot z^{\bar{j}}$ in chapters (\ref{ch4}) and (\ref{RDE}).
Where for instance, in chapter (\ref{RDE}) EFE have been solved in case
the total density equals the dust plus radiation densities $\rho = \rho_r+ \rho_m \propto 1/a^4 +1/a^3$,
the total pressure equals to the radiation pressure $p=p_r=\rho_r/3$, $\Lambda=1$ (de Sitter space), and
$\tilde{\Lambda}=0$. That yields the brane's and the bulk's scale factors, and $\left| {G_{i\bar j} \dot z^i \dot z^{\bar j}} \right|$ as functions in time. Using suitable fitting functions, we get the solution of the velocity norm of the complex structure moduli given by
\begin{equation}
{G_{i\bar j} \dot z^i \dot z^{\bar j}} (t) \simeq -0.4 (t + 0.004)^{-0.9}.
\label{zsol}
\end{equation}
The brane's scale factor varies with time exponentially $a(t) \sim e^{0.2~t}$ which means
the brane- universe undergoes an inflationary expansion. While the bulk scale factor is given by
$b(t) \sim 0.06 e^{0.87~t}$.
Solving the moduli field equation Equ. (\ref{z}) with Equ. (\ref{zsol}), gives
the moduli's variation with time \cite{salem2022}:
\begin{equation}
z(t) \simeq C + \frac{0.001 +0.25~ t}{(1 + 250~ t)^{0.6}},
\label{z-t}
\end{equation}
for $\dot{z}[0]=0.1$. C is the integration constant, for $z[0]=1$, $C \sim 1$. We have made a further approximation here by considering the moduli real. In Fig. (\ref{zz}- left) and (\ref{zz}- right) the moduli and the moduli velocity are plotted versus time, respectively.
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=1.0\columnwidth]{./Fig/z.pdf}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=1.0\columnwidth]{./Fig/zdot.pdf}
\end{subfigure}
\caption{(Left panel): The moduli is plotted versus time.
(Right panel): The moduli velocity is plotted versus time. For $C =1$, and in case
of radiation, dust and $\Lambda$ filled brane with $\Lambda=1$ and $\tilde{\Lambda}=0$. }
\label{zz}
\end{figure}
The K\"{a}hler metric can be directly obtained by substituting $\dot{z}$ solution in Equ. (\ref{zsol}). In Fig. (\ref{G}- left)
one component of the metric $G_{i\bar{j}}$ multiplied by a factor $10^{-2}$ is plotted versus time.
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=1.0\columnwidth]{./Fig/g.pdf}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=1.0\columnwidth]{./Fig/kappa.pdf}
\end{subfigure}
\caption{(Left panel): One component of the K\"{a}hler metric multiplied by a factor $10^{-2}$ is
plotted versus time. (Right panel):
The K\"{a}hler potential of $\mathcal{M}_C$ is plotted versus time.}
\label{G}
\end{figure}
We will use Equ. (\ref{pot22}) to get $\kappa$ as a function in time.
According to our approximation it can be written as:
\begin{equation}
G = \partial_z \partial_{z^*} \kappa.
\label{Gk}
\end{equation}
Let all fields depend on time, it becomes:
\begin{equation}
G(t) = \frac{\partial}{\partial t} \left( \frac{1}{\dot{z}} \frac{\dot{\kappa}}{\dot{z^*}} \right).
\end{equation}
Solving for the K\"{a}hler potential, yields:
\begin{equation}
\kappa (t) = - 0.298~ t- 0.0056 ~ t^2 .
\end{equation}
Fig. (\ref{G}- right) shows the absolute value of the potential plotted versus time.
Since the volume of the CY manifold $\mathcal{M}$ is related to the K\"{a}hler potential by $\text{Vol}(\mathcal{M})=e^{-\kappa}$
(\ref{veve}),
we have plotted in Fig. (\ref{vol}- left) the volume of the Calabi-Yau manifold versus time. As seen it increases with time.
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=1.0\columnwidth]{./Fig/vol.pdf}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=1.0\columnwidth]{./Fig/z-z.pdf}
\end{subfigure}
\caption{(Left panel): The volume of the Calabi-Yau manifold plotted versus time.
The Hodge number $h_{2,1}=1$.
(Right panel):
Numeric and analytic $G_{i\bar{j}} \dot z^i \dot z^{\bar{j}}$ versus time in red and blue curves, respectively.
The Hodge number $h_{2,1}=1$.}
\label{vol}
\end{figure}
For the sake of comparison, Fig. (\ref{vol}- right) shows $G_{i\bar{j}} \dot z^i \dot z^{\bar{j}}$ plotted versus time as it is obtained directly from the numeric solution of the field equations (\ref{EE3}) without any approximations a long a side as it is obtained when solving Equ. (\ref{zsol}) with Equ. (\ref{z-t}).
\section{Analytic solutions for the field equations}
Modeling a 3 brane- world in $\mathcal{D}=5$ $\mathcal{N}=2$ supergravity has been studied in many earlier papers. In the first \cite{Emam:2015laa} an `empty' brane-world was analytically studied, and a correlation between the dynamics of the brane, \emph{i.e.} the time evolution of the scale factors, was found.
This was further explored in chapter (\ref{ch4}) where a couple of brane cases was considered: A dust filled brane and a radiation filled brane.
In all cases not only was the previous correlation confirmed, leading to an expanding brane brought about by the complex structure moduli, but we have also found that the early decay of the norm of the moduli's flow velocity $G_{i\bar j} \dot z^i \dot z^{\bar j}$ directly leads to a period of rapid accelerated expansion of the brane. That means the complex structure moduli can play the role of the inflaton.
The studies in chapter (\ref{ch4}) were only numeric. In this section we will
solve the field equations analytically to manifest the moduli
correlations.
Indeed $G_{i\bar{j}} \dot z_i \dot z_{\bar{j}}^*$
drives the dynamics of all the theory fields, whether in the bulk or on the
brane- world. This may be understood from a topological perspective , since
$G_{i\bar{j}}$ is just the mertic of the Calabi- Yau manifold complex structure space $\mathcal{M}_C$ and the moduli
can be treated as the complex coordinates of $\mathcal{M}_C$ and its number
($h_{2,1}$) is the dimension of that space.
Anyhow we would like to show here analytic formulas for the brane- world scale factor.
Then the analytic moduli can be obtained directly from the field equations.
This study is different than what has been done in \cite{Emam:2015laa} where
an analytic formula for $a(t)$ has been assumed in case of a vacuous brane-world.
To solve analytically we will take a constant bulk scale factor $b=1$, then the field equations reduce to:
\begin{eqnarray}
\nonumber 3 \left( \frac{\dot a}{a} \right)^2 &=& G_{i\bar j} \dot z^i \dot z^{\bar j} + \rho, \\ \nonumber
2 \left( \frac{\ddot a}{a} \right) + \left(\frac{\dot a}{a} \right)^2 &=& - G_{i\bar j} \dot z^i \dot z^{\bar j} - p, \\
-\left[ {\frac{{\ddot a}}{a} + \left( {\frac{{\dot a}}{a}} \right)^2 } \right] &=& - G_{i\bar j} \dot z^i \dot z^{\bar j}, \label{FE-consb}
\end{eqnarray}
where we consider $\rho$ and $p$ are constants, and $\Lambda=\tilde{\Lambda}=0$. Solving these eqautions together gives:
\begin{equation}
a(t)= c_2 \cosh \left[\sqrt{3}~ \sqrt{p} ~~(c_1+t)\right]^{1/3},
\end{equation}
where $c_1$, and $c_2$ are the constants of integration. While the energy density is given by:
\begin{equation}
\rho=p+ \frac{2 \ddot{a}}{a}+\frac{4 \dot{a}^2}{a^2}.
\label{ED}
\end{equation}
In Fig. (\ref{atA}), the brane's scale factor is plotted versus time by the blue curve at $c_1=0$ and $c_2=0.1$ for initial conditions $a(0)=0.1$ and $\dot{a}(0)=0$. $p=0.1$, and from (\ref{ED}) $\rho \sim 3 p =0.3$ . While the scale factor of $\Lambda$CDM model \footnote{\em Lambda cold matter model: The standard model of big bang cosmology.} \cite{Planck:2018} is plotted by the dashed pink curve at current matter density parameter $\Omega_{m0}= 0.31$, current radiation density parameter $\Omega_{r0}= 9.2 \times 10^{-5}$, and $\Omega_{\Lambda 0} = 0.7$. The current value of the Hubble parameter is given by $H_0= 0.0686751 ~[\text{Gyr}^{-1}]$. The vertical grid line is at the present time of the universe $t_0= 13.842~ [\text{Gyr}]$, and the horizontal grid line is at the current value of our universe scale factor, parameterized at $a_0=1$.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\columnwidth]{./Fig/atA.pdf}
\caption{The blue curve: The brane's scale factor is plotted versus time
at $c_1=0$ and $c_2=0.1$ for initial conditions $a(0)=0.1$ and $\dot{a}(0)=0$. $p=0.1$. The pink dashed curve: The scale factor of $\Lambda$CDM model at $\Omega_{m0}= 0.31$, $\Omega_{r0}= 9.2 \times 10^{-5}$, and $\Omega_{\Lambda 0}=0.7$. The current universe age since the big bang is given by the vertical gird line $t_0= 13.842~ [\text{Gyr}]$, and the current value of the our universe's scale factor is given by horizontal gird line $a_0=1$. }
\label{atA}
\end{figure}
For a vacuous brane- universe $\rho=p=\Lambda=0$, the field equations become:
\begin{eqnarray}
\nonumber 3 \left( \frac{\dot a}{a} \right)^2 &=& G_{i\bar j} \dot z^i \dot z^{\bar j}, \\
\nonumber 2 \left( \frac{\ddot a}{a} \right) + \left(\frac{\dot a}{a} \right)^2 &=& - G_{i\bar j} \dot z^i \dot z^{\bar j}, \\
-\left[ {\frac{{\ddot a}}{a} + \left( {\frac{{\dot a}}{a}} \right)^2 } \right] &=& - G_{i\bar j} \dot z^i \dot z^{\bar j}, \label{FE-consb2}
\end{eqnarray}
Solving together yields:
\begin{equation}
a(t)= c_2 (3 t - c_1)^{1/3},
\label{atv}
\end{equation}
while the moduli $G_{i\bar j} \dot z^i \dot z^{\bar j}$ can be given by any of the three equations in (\ref{FE-consb2}). The most fitted values of $c_1$ and $c_2$ for the $\Lambda$CDM scale factor are given by $c_1= -0.04$, and $c_2 = 0.29$ at IC $a(0)=0.1$ and $\dot{a}(0)=2.4$. This shown in Fig. (\ref{atAv}-Left) where the brane's scale factor (\ref{atv}) is plotted versus time by the blue curve, while $\Lambda$CDM scale factor is plotted by the dashed pink curve. Fig. (\ref{atAv}-Right) shows the correlation between the moduli and the brane's scale factor where both are plotted versus time.
\begin{figure}[H]
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=1.0\columnwidth]{./Fig/atAv.pdf}
\end{subfigure}
\qquad
\begin{subfigure}[t]{.5\linewidth}
\centering
\includegraphics[width=1.0\columnwidth]{./Fig/azAv.pdf}
\end{subfigure}
\caption{(Left panel): The brane's scale factor is plotted versus time by the blue curve $\rho=p=\Lambda=0$. $c_1= -0.04,~ c_2 = 0.29 $, $a(0)=0.1,~ \dot{a}(0)=2.4 $. While $\Lambda$CDM scale factor is plotted by the pink curve. (Right panel): The moduli $G_{i\bar j} \dot z^i \dot z^{\bar j}$ and the brane's scale factor are plotted versus time.}
\label{atAv}
\end{figure}
\chapter{Manifolds}
The space-time in string theory and though in supergravity is often described by means of a mathematical object called manifold.
So in order to understand the dimensional reduction of $D=11$ SUGRA to $D=5$ SUGRA over Calabi-Yau manifold
as we will show in the next chapter, we should to understand pretty well what are CY manifolds as we will illustrate in this chapter. CY 3-fold is a complex special K$\ddot{\mbox{a}}$hler manifold. Before talking about this class of manifolds, we will give first a brief overview on the different types of manifolds, starting with elementary definitions and properties \cite{Ooguri2015}.
\section{Riemannian geometry}
The $n$- dimensional manifold $\mathcal{M}$ is a topological space that each point $x \in \mathcal{M}$ has an open set $U$ $\subseteq$ $\mathcal{M}$ such that $U$ has a mapping $\phi$ to $\mathbb{R}^n$. The pair $(U, \phi)$ is called coordinate chart, while $\{(U_i,\phi_i)\}$ is called atlas, where
$n \in \mathbb {Z}$. A smooth manifold is called \textbf{A Riemannian manifold} if it admits a symmetric positive-definite metric
$ g= g_{\mu\nu} (x) dx^\mu dx^\nu$. A manifold is called \textbf{Lorentzian} if its metric
is the Minkowski metric $\eta= \mbox{diag}(-1,1,...,1)$, i.e. behaves locally like
$\mathbb{R}^{1,n-1}$ . A Levi-Civita connection may be chosen, leading to the usual
expressions for the Christoffel symbols, the Riemann and Ricci tensors and the Ricci scalar:
\begin{eqnarray}
\Gamma^\lambda_{\mu\nu} &=& \frac{1}{2} g^{\lambda\kappa} [(\partial_\mu g_{\nu\kappa}) + (\partial_\nu g_{\mu\kappa})-
(\partial_\kappa g_{\mu\nu}) ],\\
R_{\mu\rho\nu}^{~~~\sigma} &=& (\partial_\rho \Gamma^\sigma_{\mu\nu}) - (\partial_\mu \Gamma^\sigma_{\rho\nu})
+ \Gamma^\alpha_{\mu\nu} \Gamma^\sigma_{\alpha\rho} - \Gamma^\alpha_{\rho\nu} \Gamma^\sigma_{\alpha\mu},\\
R_{\mu\nu} &=& R_{\mu\rho\nu}^{~~~\rho} = (\partial_\rho \Gamma^\rho_{\mu\nu}) - (\partial_\mu \Gamma^\rho_{\rho\nu})
+ \Gamma^\alpha_{\mu\nu} \Gamma^\rho_{\alpha\rho} - \Gamma^\alpha_{\rho\nu} \Gamma^\rho_{\alpha\mu},\\
R &=& g^{\mu\nu} R_{\mu\nu} = R^\mu_\mu.
\end{eqnarray}
\section{Complex and K$\ddot{\mbox{a}}$hler manifolds}
Similarly, a complex $n$-dimensional manifold defined as a set of points that behaves locally like $\mathbb{C}^p$,
where $\mathbb{C}^p$ is the space of differential forms of degree $p$, and $n \in \mathbb{Z}$, i.e., they are parameterized by complex coordinates. So that the dimension of $n$ should be even.
Like the Riemannian manifold are manifested by a metric
$g_{\mu\nu} (x)$, complex manifolds are endowed by the so- called complex structure, a mixed matrix
$J^{~\nu}_\mu (x)$ defined on the manifold, where $\mu,\nu= 1,..,2m$, such that
\begin{equation}
J^{~\nu}_\mu ~ J^{~\rho}_\nu = - \delta^\rho_\mu, ~~~~ J^2=-\mathbb{1},
\end{equation}
this's called integrability condition for almost complex coordinates.
To say the manifold has a complex structure the so- called Nijenhuis tensor
\begin{equation}
N^\rho_{\mu\nu} = J^\alpha_\mu~
[ (\partial_\alpha J^\rho_\nu ) -
(\partial_\nu J^\rho_\alpha)] -
J^\alpha_\nu [ (\partial_\alpha J^\rho_\mu ) - (\partial_\mu J^\rho_\alpha)]
\end{equation}
should vanish everywhere. Note that this's not trivial for any manifold, for instance $S^6$ (6- dimensional sphere) has almost complex structure, while $S^4$ (4- dimensional sphere) has no any complex structure.
Now if the manifold has both Riemannian and complex structure, $g_{\mu\nu}$ and $J^{~\nu}_\mu$, respectively, the manifold is called
K$\ddot{\mbox{a}}$hler manifold. The compatibility conditions between Riemannian and complex structures are given by
\begin{enumerate}
\item $\nabla_\mu ~ J^\nu_{~\rho} =0$,
\item $g_{\mu\nu} J^\mu_{~\rho} J^\nu_{\sigma} = g_{\rho\sigma}$,
\end{enumerate}
the second condition leads to the metric $ ds^2= g_{\mu\nu} dx^\mu dx^\nu$ can be written as
\begin{equation}
ds^2= 2 g_{i\bar{j}} d\omega^i d\omega^{\bar{j}},
\end{equation}
the holomorphic coordinates $i,\bar{j}= 1,.., m$, where $2m$ is the dimension of the manifold. Introduce a 2- form ( anti-symmetric tensor)
\begin{equation}
\begin{split}
\kappa & = \frac{1}{2} g_{\mu\nu} J^\mu_{~\rho} ~ dx^\rho \wedge dx^\nu \\
& = i g_{i\bar{j}} d\omega^i \wedge d\omega^{\bar{j}},
\end{split}
\end{equation}
$\kappa$ is called the K$\ddot{\mbox{a}}$hler form. The first condition leads to :
\begin{equation}
d \kappa =0,
\label{eq1}
\end{equation}
means the metric $g_{i\bar{j}}$ satisfying ;
\begin{equation}
\begin{split}
\partial_i g_{j\bar{k}} &= \partial_j g_{i\bar{k}},\\
\partial_{\bar{j}} g_{i\bar{k}} &= \partial_{\bar{k}} g_{i\bar{j}},
\label{eq2}
\end{split}
\end{equation}
plus the complex conjugate equations. So “ locally “ on each coordinate patch on the manifold.
\begin{equation}
g_{i\bar{k}} = \partial_i f_{\bar{k}} ,
\end{equation}
$ f_{\bar{k}}$ can also written as a derivative of another function
\begin{equation}
g_{i\bar{k}} = \partial_i \partial_{\bar{j}} K,
\label{pot22}
\end{equation}
$K$ is called the K$\ddot{\mbox{a}}$hler potential , then
\begin{equation}
\kappa = i \partial_i \partial_{\bar{j}} K ~ d\omega^i \wedge d\omega^{\bar{j}}.
\end{equation}
Note that $K$ can not globally defined on the manifold, otherwise it can be defined locally for every coordinate patch, and if the coordinates had transformed, $K$ transforms in away makes all its relations work. Consequently, if $d\kappa =0$, and $ \kappa = d\Lambda$ , $\Lambda$ can only locally defined.
This’s obvious when we know that the volume of the
K$\ddot{\mbox{a}}$hler manifold is proportional to $\kappa^m$ , so if $\kappa$ is exact and assuming $M$ is a non-compact manifold, with no boundaries, then by Stoke’s theorem
\begin{equation}
\int_M d \Lambda = \int_{\partial M} \Lambda =0,
\end{equation}
which is a disaster because the volume of $M$ should be positive and definite.
Also the condition (\ref{eq1} ) or equivalently (\ref{eq2}) simplifies the properties of the manifold considerably, for example one finds that
\begin{equation}
\begin{aligned}
\Gamma^k_{ij} &= g^{k\bar{l}} (\partial_i g_{j\bar{l}}), \\
\Gamma^{\bar{k}}_{\bar{i}\bar{j}} &= g^{l\bar{k}} (\partial_{\bar{i}} g_{\bar{j}l}),
\end{aligned}
\end{equation}
are the only non-vanishing Christoffel symbols, indicating that parallel transport does not mix the holomorphic and the antiholomorphic components of a vector. Also the non-vanishing components of the Ricci tensor are found to be
\begin{equation}
R_{i\bar{j}} = \partial_i \partial_{\bar{j}} ~\mbox{log (det}~ g)
\end{equation}
\section{De Rham cohomology groups}
Consider a $p$- form $\omega$ if it closed form, i.e., $d \omega=0$, can we say $\omega$ is exact, i.e., $\omega=d\lambda$, where $\lambda$ is a $(p-1)$ form ? clearly if
$\omega$ is an exact form it is closed, but what about vise versa ? we will try to answer this question \cite{Nakahara2003}.
$\textbf{Poincar\'e's ~ Lemma}$:
On $n$- dimensional Euclidean space $\mathbb{R}^n$ a closed form is exact. This theorem can be generalized if we have a manifold $M$ that can be continuously deformed to a point. So if you have a closed form on any coordinate batch $U_i$ on $M$, this form is always exact, because $U_i$ can be contracted to a point in $\mathbb{R}^n$.
So now what kind of closed forms that are not exact ?
Define the space of closed forms:
\begin{equation}
Z^p \mathcal{M} = \{ \omega \in C^p \mathcal{M} ; d\omega =0 \},
\end{equation}
$Z^p$ is infinite dimension space because you can choose any $(p-1)$ form then take its exterior derivative. Also
\begin{equation}
B^p \mathcal{M} = \{ d\lambda : \lambda \in C^{p-1} \mathcal{M} \}.
\end{equation}
The cohomolgy is defined by
\begin{equation}
H^p \mathcal{M} = Z^p \mathcal{M}/ B^p \mathcal{M},
\end{equation}
the space of closed forms modulo by exact forms. In Mankowski flat space, cohomology is trivial by Poincar\'e's Lemma. The dimension of this space
\begin{equation}
b_k = \mbox{dim}~ H^k \mathcal{M},
\end{equation}
is called Betti number and it is a topological invariant, which means if you continuously deformed the manifold, this quantity keeps invariant. Also from the topological invariants, Euler characteristic which defined by
\begin{equation}
\chi = b_0-b_1 +b_2-b_3+ ...
\end{equation}
\section{Homology groups}
They are Abelian groups which partially count the number of holes in a topological space. In particular, homology groups form a measure of the hole structure of a space, but they are one particular measure and they do not always pick up everything.
Before going further, we need first introduce the following notation.
\subsection{Triangulation}
It means to approximate any topological space by a generalizaton of triangles.
When we talk about tringle, it will include its interior, its 3 vertices, 3 edges and one face. While a 3 dimentional version of a tringle is called tetrahedron. Generalize this to define $p$- simplex $\subset R^p$. They consists of ($p+1$) vertices, so the triangle of is 2- simplex, while the tetrahedron is 3- simplex.
$p$- simplex itself is a topological space becasue it is a subspace of the Euclidean space which's a topological space. Also they can be combined to form more complicated spaces. So we can have a continuous map from a topological space to another and we can say when two topological spaces are equivalent to each other. Consider two topological spaces $M_1$ and $M_2$, where
\begin{equation}
f ~~:~~ M_1 \to M_2, ~~~~~~~~~~~~~~~~ g ~~:~~ M_2 \to M_1
\end{equation}
If these mapping functions are inverse to each other
\begin{equation}
f \circ g = id, ~~~~~~~ g \circ f = id
\end{equation}
we say $M_1$ and $M_2$ are equivalent ($\sim$) to each other, consequently if $ k \sim M $, where $k$ is the simplicial complex, we say $k$ is a triangulation for $M$. In summary, given an arbitrary topological space $M$, to be traingulated , we cover it up by $k$ which is $\sim$ for $M$.
Of course traingulation is not unique, because $k$ can be more complicated, but this won't change the topology of space. One also can consider different $\sim$ relations between different triangulation.
Theorem: If you have smooth differential manifolds, you always have triangulation.
\subsection{Homology}
It is a dual space of cohomology space. Consider $k$ which contains $\sigma$ a $p$-simplex with $(p =0,..,n)$ and $n = dim M$. $\sigma$ has a rotation direction around its vertices , so $\sigma$ denoted by its vertices as
\begin{equation}
\sigma = \langle v_0,v_1,...,v_p \rangle
\end{equation}
Define $C_p$ a vector space generated by these simplexes. We can add or substract simplexes (as elements of vectors) and reverse their orintation
\begin{equation}
\langle v_0,v_1,...,v_p \rangle = - \langle v_1,v_0,...,v_p \rangle,
\end{equation}
Similar for the extrior operator in the differential forms, define
here $\partial$ a boundry operator
\begin{equation}
\partial \langle v_0 ....v_p \rangle = \sum^p_{i=0} (-1)^i~
\langle v_0,...,\hat{v_i}, ...,v_p \rangle,
\end{equation}
removing the $ith$ component and yields $(p-1)$ simplex and so that
here $\partial$ is called a boundry operator. Consider more complicated example of two triangles
\begin{equation}
\partial : C_p \to C_{p-1}
\end{equation}
and $\partial^2=0$. Define
\begin{equation}
Z_p(k)= \{ c \in C_p(k) : \partial c =0 \}
\end{equation}
i.e., elements of $C_p(k)$ which vanish by $\partial$. These means there are kind of simplical complex that have no boundries, because when they acted on by the boundry operator they vanish. So that the elements of $Z_p (k)$ are called cycles.
Now define
\begin{equation}
B_p(k)= \{ \partial c : c \in C_{p+1}(k) \}
\end{equation}
The boundry itself has no boundry. So $ B_p(k) \subset Z_p (k)$ . The homology defined as
\begin{equation}
H_p(k) = Z_p(k) / B_p(k),
\end{equation}
the cycles modulo by boundries. So if you want to compute the homology of any topoligcal space, what you do is to consider any trigulation of that space, then apply the above definition. Homology is a topological invariant (dose not depend on the way of triangulation).
Consider homology with integer or with real coefficients $H_p(M,Z)$ and $H_p(M,R)$ respectivly. The first detects more topolgy than the second. Consider for example SO(3) rotational group, this's a manifold because you can introduce coordintes on it ( so it has a topological structure ). Consider $S^3$ (3- dimentional sphere) with anti- polar points identified, where once you identify a point on the upper hemisphere, you identify a point on the lower hemisphere, so the whole sphere can be drawn just as a hemisphere, take
\begin{equation}
H_1 (SO(3), Z) = Z_2 = \{ 0,1 \}
\end{equation}
$Z_2$ is a space consists of 2 numbers and it is binary (1+1=0), which means you have one cycle when repeated you return to the same point. While $H_1 (SO(3), R)=0$ because $H_1 (SO(3), Z)$ is very discrete. $Z_2$ is called torsion and it is an element of homology which corresponds to some finite group. While $H_0 (SO(3), Z)= Z$ or $H_1 (SO(3), R)= R$ are trivial.
\subsection{Fundamental theorem of de Rham}
There are dual features between homology and cohomolgy. Consider
\begin{equation}
H^p \mathcal{M} \in \omega,
\end{equation}
where $\omega$ is a closed $p$-form, but not exact and
\begin{equation}
H_p \mathcal{M} \in c,
\end{equation}
where $c$ cycles in $p$- dimensions. Now integrate $p$-form over $p$- dimensional cycle
\begin{equation}
(c,w) = \int_c w.
\end{equation}
Due to Stoke's theorem this pair is independent of the change of $\omega$ by exact form or changing $c$ by a boundry
\begin{equation}
\int_\gamma d \lambda = \int_{\partial \gamma } \lambda ,
\end{equation}
where $\partial \gamma$ is the boundry of some orintable manifold $\gamma$. So
\begin{equation}
\int_c (\omega+d\lambda) = \int_c \omega + \int_{\partial c} \lambda = \int_c \omega,
\end{equation}
\begin{equation}
\int_{c+\partial \gamma} \omega = \int_c \omega + \int_{\partial \gamma} \omega = \int_c \omega,
\end{equation}
because the boundry of $c$ is absent. Also $(c,\omega)$ dose not affected by trivial shifts
\begin{equation}
(c+ \partial \gamma, \omega) = (c,\omega) ,
\end{equation}
\begin{equation}
(c, \omega + d\lambda) = (c,\omega),
\end{equation}
this manifest the pairing between $H_p\mathcal{M}$ (cycles or homology) and $H^p\mathcal{M}$ (closed forms or cohomology). dim $H_p = b_b$
(Batti number) is a topoligcal invariant and $H_p$ has basis of $\{c_1,...,c_{b_p}\}$. The two theorems of de Rham state that:
Theorem 1:
For any integers $v_i$, there are a closed $p$-form $\omega$, where is
\begin{equation}
\int_{c_i} \omega = v_i,
\end{equation}
$c_i$ is the period of integral that gives you this integer.
Theorem 2 :
If $\int_{c_i} \omega =0$ for all $c_i$, then $\omega=0$. This means $c_i$ basis are not degenerate. Note that
dim $H^p =$ dim $H_p = b_p $, where the basis of $H^p$ are $\{ \omega^1, \omega^2,...., \omega^{b_p}\}$, so that we can define a square matrix
\begin{equation}
(c^i, \omega^j) = \int_{c^i} \omega^j = \Delta_{ij},
\end{equation}
called a period matrix, and the de Rham theorems say: det $\Delta \neq 0$.
\section{Hodge- K$\ddot{\mbox{a}}$hler cohomology}
Rembeber a $k$-form on a manifold is defined by:
\begin{equation}
C^k\mathcal{M} \ni \omega = \frac{1}{k!} \omega_{\mu_1,.......\mu_k} ~~~~ dx^{\mu_1} \wedge ...... dx^{\mu_k},
\end{equation}
in de Rham cohomology we take the exterior derivative of this form and consider the form closed, then the space of the closed forms modulo by exact forms.
Now if we have complex coordinates $\omega^i$ and
$d\omega^{\bar{j}}$ that do not mix under holomorphic coodinate transfomation. So that the space of $k$-forms $C^k\mathcal{M}$ decomposing to
$C^{p.q}\mathcal{M}$ and $(p-q)$ form is defined instead by;
\begin{equation}
\omega= \frac{1}{p!~q!}~ \omega_{i_1,....., i_p, \bar{j}_1, ....., \bar{j}_q} ~~~ d\omega^{i_1} \wedge ..... d\omega^{i_p} ~~ \wedge d\omega^{\bar{j}_1} \wedge ...... \wedge d\omega^{\bar{j}_q},
\end{equation}
where $p$ is the number of $d\omega$’s and $q$ is the number of $d\bar{\omega}$’s. $H^{p.q}\mathcal{M}$ is called Hodge- de Rhame comology group, where the total degrees of freedom is given by $k=p+q$. Let’s consider now an arbitrary 2- form
\begin{equation}
C^{1.1} \mathcal{M} : \sigma_{i\bar{j}}~~ d\omega^i \wedge d\omega^{\bar{j}},
\end{equation}
if $d\sigma=0$ it will be a member of the de Rham class $H^{1.1}$. The K$\ddot{\mbox{a}}$hler form is a non trivial member of $H^{1.1}$, so that (locally) the dimension of $H^{1.1}$ is 1. The question now does $H^{1.1}$ contain any other known members? what about the Ricci tensor? it’s a 2-form ;
\begin{equation}
\mathcal{R}_{i\bar{j}}= \partial_i \partial_{\bar{j}}~ (\mbox{log}~ g),
\end{equation}
\begin{equation}
\mathcal{R} = \mathcal{R}_{i\bar{j}} d\omega^i \wedge d\omega^{\bar{j}},
\end{equation}
but it’s not clear if $d \mathcal{R}=0$ or not. May be a nice remark here is that Einstein field equations in the vaccum $\mathcal{R}_{\mu\nu}$ is trivial with $\mathcal{R}=d\Lambda=0$.
So in general is $R$ a trivial element of $H^{1.1}$? it has been found this question is so important and carries many interesting topology data about the manifold. $R$ is called the $1^{\mbox{st}}$ Chern- Class, and it was deduced that like when $\kappa$ is a member of $H^{1.1}$, the manifold contains a k$\ddot{\mbox{a}}$hler structure, similarly if $R$, can a metric $g_{i\bar{j}}$ found on the manifold with the same complex structure and with the same k$\ddot{\mbox{a}}$hler structure? that what was proved by Yau. K$\ddot{\mbox{a}}$hler manifolds satisfy the Ricci flatness condition are called Calabi- Yau manifolds.
\section{Calabi- Yau manifold}
In 1954 Calabi proposed the following conjecture: If $\mathcal{M}$ is a complex manifold with a K$\ddot{a}$hler metric and vanishing first Chern class, then there exists a unique Ricci flat metric for each K$\ddot{a}$hler class on $\mathcal{M}$. In 1976, Calabi's conjecture was proven by Yau \cite{CYM2-2}. Consider:
\begin{equation}
\partial_i \partial_{\bar{j}}~ log (\mbox{det}~ g)=0,
\end{equation}
then:
\begin{equation}
log (\mbox{det}~ g) = f(w) + \bar{f} (\bar{w}),
\end{equation}
so that $det~ g$ transforms as a determinant of ($i,\bar{j}$) matrix, where $i$ and $\bar{j}$ are the holomorphic and anti-holomorphic coordinates, respectively.
\begin{equation}
\mbox{det}~ g = \Omega(w) \bar{\Omega} (\bar{w}),
\label{volume}
\end{equation}
where $\Omega(w)$ is $(i,0)$ form and $\bar{\Omega}(\bar{w})$ is $(0,\bar{j})$ form. So that $\mathcal{R}_{i\bar{j}}=0 ~ \implies \exists ~ \Omega(i,0)$ form:
\begin{itemize}
\item $\bar{\partial} \Omega =0$, ~~~~~ (holomorphic)
\item $\Omega \neq 0.$ ~~~~~~~ any where on $\mathcal{M}$,
\end{itemize}
that what was proven by Yau. Examples of Calabi-Yau manifold in one complex dimension:
\begin{itemize}
\item A complex plane $\mathbb{C}$, with a flat metric
$dw~ d\bar{w}$, $w \in \mathbb{C}$.
\item $ \mathbb{C}^w = \mathbb{C} \backslash \{0\}$, which is a cylinder.
\item $T^2 = \mathbb{C} / \mathbb{Z} + \tau~ \mathbb{Z}$,
\end{itemize}
where $\tau$ is an integer $\in \mathbb{Z}$. $w=w^1+\tau~ w^2$, where $\tau$ is called the complex structure moduli, it parameterizes the shape of the torus. Since $T^2$ is a CY manifold, it should has a holomorphic form $\Omega$, $(i,0)$, $i=1$, because the complex dimension is 1, $\Omega= d~w$.
We can extract the complex strucure moduli by integrating $\Omega$ over a homology cycle of the CY manifold,
Fig. (\ref{to})
\begin{figure}[!t]
\centering
\includegraphics[scale=0.4]{./Fig/To.pdf}
\caption{The torus as Calabi- Yau manifold in one complex dimension.}
\label{to}
\end{figure}
\begin{equation}
\int_\alpha \Omega =1 , ~~~~~~~~~ \int_\beta \Omega = \tau,
\end{equation}
in this case the Hodge- de Rham cohomology will be given by:
\begin{equation}
H^{p,q} = \{ w \in \Omega^{p,q}, d w =0\}/ d\lambda.
\end{equation}
The dimension of $H^{p,q}$ is called the hodge number $h^{p,q}$. The hodge diamond of $T^2$ is given by Fig. (\ref{hod1}),
\begin{figure}[H]
\centering
\includegraphics[scale=0.3]{./Fig/hodge1.pdf}
\caption{The hodge diamond of $T^2$.}
\label{hod1}
\end{figure}
where:
\begin{equation}
\begin{split}
h^{1,0} &=1, ~~~~~~~~~~ \mbox{only}~ \Omega= dx ~ (1,0)~ \mbox{form}, \\
h^{1,1} &=1, ~~~~~~~~~~ dw~ d\bar{w} ~ \mbox{the volume.}
\end{split}
\end{equation}
While the hodge diamond of the 3- complex dimensions CY manifold is given by Fig. (\ref{hod3}),
\begin{figure}[t]
\centering
\includegraphics[scale=0.3]{./Fig/hodge3.pdf}
\caption{The hodge diamond of the 3-fold CY.}
\label{hod3}
\end{figure}
where $h^{11} = h^{22}$ and $h^{12}= h^{21}$. The Eular characterstic is given by $\chi= 2 (h^{11}- h^{12})$. $h^{2,1}$ determines the number of complex structure moduli on the complex strucutre space $M_C$ and $h^{1,1}$ determines the number of K$\ddot{\mbox{a}}$hler forms on $M_K$, where $M_K$ being a complexification of the parameters of the K$\ddot{\mbox{a}}$hler class. In addition, there exists a symmetry in the structures of $M_C$ and $M_K$ which lends support to the so-called mirror symmetry hypothesis of CY 3-folds \cite{CYM2-1}. In terms of homology groups, Calabi-Yau 3-folds admit a non-trivial $H^3$ that can be Hodge decomposed as follows:
\begin{equation}
H^3= H^{3,0} \oplus H^{2,1} \oplus H^{1,2} \oplus H^{0,3}.
\end{equation}
So that CY 3-folds have a single $(3,0)$ cohomology form;
$h_{3,0} =dim (H^{3,0})= 1$, which we will call $\Omega$ (holomorphic: $\bar{\partial} \Omega=0$ ).
\subsection{Deformation of complex structure}
$\mathbf {The~line~ sub- bundle}$:
in different complex structures, the decompositions of the tangent bundle into holomorphic and
anti-holomorphic parts are different \cite{Hori:2003ic}.
Therefore, what was a closed and holomorphic $(n, 0)$ form in one complex structure will no longer be of type $(n, 0)$, (nor holomorphic) in another complex structure. However, the form will still
be closed, as the exterior derivative is independent of complex structure.
In this description it is easy to see that, to linear order, a (3, 0) form
can only change into a linear combination of (3, 0) and (2, 1) forms. The
change can be measured by $H^{2,1}$.
The cohomology class in $H^3$ representing the holomorphic
$(n, 0)$ form must change over the moduli space of complex structures $M_C$ .
In fact, $H^3$ forms a bundle over the moduli space, and the Calabi-Yau form
is a complex line in this bundle. We can define a natural metric on this line bundle
\begin{equation}
h = ||\Omega||^2 = (\Omega,\Omega) = i \int \Omega \wedge \bar{\Omega},
\end{equation}
with
\begin{equation}
K= - log ||\Omega||^2 = - log \int \Omega \wedge \bar{\Omega},
\end{equation}
transforms as a K$\ddot{\mbox{a}}$hler potential,
$K \to K - f - f$. We therefore can define a metric on moduli space by
\begin{equation}
G_{m\bar{n}}= \partial_m \bar{\partial}_{\bar{n}} K,
\end{equation}
and this is well defined no matter the gauge choice of $\Omega$, since $f + \bar{f}$ is killed
by $\partial \bar{\partial}$.
We already decided that the tangent space to
moduli space is $H^{2,1}$ . To see the above metric in more detail,
let us write the variation of $\Omega$ with
respect to a coordinate direction $z^a$ as
\begin{equation}
\begin{aligned}
\partial_a \Omega &= (3,0) \mbox{piece} + (2,1) \mbox{piece}\\
&= k_a \Omega + \chi_a,
\label{q1}
\end{aligned}
\end{equation}
where there are no other terms since the variation of a holomorphic (1,0)
form has a (1, 0) and a (0, 1) piece. Then, using Eq. (\ref{q1}), one finds
\begin{equation}
\begin{aligned}
\partial_a \bar{\partial}_{\bar{b}} K &= \partial_a \big[ \frac{-1}{\int \Omega \wedge \bar{\Omega}} \int \Omega \wedge \bar{\partial}_{\bar{b}} \bar{\Omega} \big] \\
&= - \frac{1}{(\int \Omega \wedge \bar{\Omega})^2} \int \partial_a \Omega \wedge \Omega \int \Omega \wedge \bar{\partial}_{\bar{b}} \bar{\Omega} + \frac{1}{\int \Omega \wedge \bar{\Omega}} \int \partial_a \Omega \wedge \bar{\partial}_{\bar{b}} \bar{\Omega}\\
&= \frac{\int \chi_a \wedge \chi_{\bar{b}}}{(\int \Omega \wedge \bar{\Omega})}.
\end{aligned}
\end{equation}
While the variation of the Ricci-flat
metric corresponding to the ath direction
\begin{equation}
\delta_a g_{\bar{i}\bar{j}} = \frac{\partial g_{\bar{i}\bar{j}}}{\partial z^a} = - \frac{1}{||\Omega||^2} \bar{\Omega}^{~kl}_{\bar{i}} (\chi_{a~ k l \bar{j}}),
\end{equation}
Equivalent to
\begin{equation}
\delta g_{\bar{i}\bar{j}} = \frac{\partial g_{\bar{i}\bar{j}}}{\partial z^a} = - \frac{1}{||\Omega||^2} \bar{\Omega}^{~kl}_{\bar{i}} \chi_{a|~ k l \bar{j}}~~ z^a,
\end{equation}
$\chi_{a|~ kl \bar{i}} = - \frac{1}{2} \Omega_{kl}^{~~\bar{j}} ~\frac{\partial g_{\bar{i}\bar{j}}}{\partial z^a}.$ Also $z^a$ define the parameters, or moduli, of the complex structure ($z^a: a=1,...,h_{2,1}$) .
And each $\chi$ defines a $(2,1)$ cohomology class. The important observation here is that when the moduli treated as complex coordinates, it can define a special K$\ddot{\mbox{a}}$hler metric $G_{a\bar{b}} $ on $\mathcal{M}_C$ as follows \cite{Candelas:355}:
\begin{equation}
V_{CY}~ G_{a\bar{b}} (\delta z^a ) (\delta z^{\bar{b}} ) = \frac{1}{4} \int_{\mathcal{M}} g^{i\bar{j}} g^{k\bar{l}} (\delta g_{ik})~ (\delta g_{\bar{j}\bar{l}} ) , \label{41-ch2}
\end{equation}
where $V_{CY}$ is the volume of the Calabi-Yau.
From $(\ref{volume})$, the volume of the CY 3-fold is given by
\begin{equation}
Vol(\mathcal{M}) = i \int \Omega \wedge \bar{\Omega},
\end{equation}
which simply relates the K$\ddot{\mbox{a}}$hler potential of $\mathcal{M}_C$ to the volume of $\mathcal{M}$ by
\begin{equation}
Vol(\mathcal{M}) = e^{-K}.
\label{veve}
\end{equation}
\subsection{Periods and coordinates on moduli space}
Let ($\alpha_a, \beta^b$) be the dual cohomology basis of the homology 3-cycles $H^3$ on $\mathcal{M}_C$, where $ a, b, c = 0, . . , h_{2,1}$. Such that
\begin{equation}
\begin{aligned}
& \int_\mathcal{M} \alpha_a \wedge \beta^b = \int_{A^b} \alpha_a = \delta^b_a, \\
& \int_\mathcal{M} \beta^a \wedge \alpha_b = \int_{\beta_b} \beta^a = -\delta^a_b, \\
& \int_\mathcal{M} \alpha_a \wedge \alpha_b = \int_\mathcal{M} \beta^a \wedge \beta^b =0, \label{dual-coh}
\end{aligned}
\end{equation}
and let ($A^a , B_b$) be canonical homology basis, such that
\begin{equation}
\begin{aligned}
& A^a \cap B_b = \delta^a_b, \\
& B_a \cap A^b = - \delta^b_a, \\
& A^a \cap A^b = B_a \cap B_b=0.
\end{aligned}
\end{equation}
\begin{equation}
\mbox{dim}~ H^3 = h^{3,0} + h^{2,1} + h^{1,2} +h^{0,3} = 2 (1+ h^{2,1})
\end{equation}
So that we can expand the Calabi–Yau form $\Omega$ as \cite{CYM2-1}
\begin{equation}
\Omega = z^a \alpha_a - f_b \beta^b,
\end{equation}
for some $z^a$ , $f_b$ , $a, b = 1, . . , h^3, \mathcal{M}/2 = h^{2,1} (\mathcal{M}) + 1$.
The coordinates $z^a$ and $f_b$ will change as we move in Calabi–Yau moduli space,
since $\Omega$ will change. In fact, as we have mentioned, since the location of
$\Omega$ in $H^3$ $\mathcal{M}$ determines the complex structure, the $z^a$ and $f_b$ determine the
point in moduli space – even over determine it, as can be seen by counting
parameters (moduli space is $h^{2,1}$ ($\mathcal{M}$) dimensional).
It is immediate from the dual basis relations that $z^a$ and $f_b$ can be
expressed in terms of the “period integrals”
\begin{equation}
z^a = \int_{A^a} \Omega, ~~~f_b = \int_{B_b} \Omega.\label{ZZ}
\end{equation}
Therefore we can express the complex structure in terms of
periods $\int_C \Omega$ of the Calabi–Yau form.
In fact, it can be shown that the $z^a$ alone locally determine the complex
structure . We can therefore imagine solving for
the $f_b$ in terms of the $z^a$ . Then the $z^a$ are only redundant by one extra
variable, but there is also an overall scale of $\Omega$ that is arbitrary, and it is
often convenient to keep the $z^a$ as homogeneous coordinates on $\mathcal{M}$ .
Some of the ingredients of this structure require the knowledge of the Hodge duality relations
(with respect to $\mathcal{M}$) of the forms ($\alpha$, $\beta$) \cite{CYM2-2}:
\begin{equation}
\begin{aligned}
\star \alpha_a &= (\gamma_{ab}+ \gamma^{cd} \theta_{ac} \theta_{bd}) \beta^b - \gamma^{cb} \theta_{ca} \alpha_b \\
\star \beta^a &= \gamma^{ac} \theta_{cb} \beta^b - \gamma^{ab} \alpha_b, \label{hodge-dual}
\end{aligned}
\end{equation}
where $\theta_{ab}$ and $\gamma_{ab}$ are real matrices defined by
\begin{equation}
\begin{aligned}
\mathcal{N}_{ab} &= \bar{f}_{ab} + 2i \frac{N_{ac} z^c _{bd} z^d}{z^p N_{pq} z^q}\\
&= \theta_{ab}-i \gamma_{ab}
\end{aligned}
\end{equation}
where $f_{ab}=\partial_a f_b$ (the derivative is with respect to $z^a$ ), $\mathcal{N}_{ab}=Im(f_{ab})$, $\gamma^{ab} \gamma_{bc}=\delta^a_c$ and
$\mathcal{N}_{ab}$ is known as the {\textbf periods matrix}.
\chapter{Acknowledgments}
I write these words while I'm close to discussing my proposal for the doctorate, indeed I'm so overwhelmed by many emotions and aspirations and feels so much gratitude.
Praise to All-mighty God who created all that beauty and all that vast magnificent universe.
I'm grateful for all my professors who made the way for theoretical physics easy for me
and who made all that possible. I'm so grateful because I was so lucky to have a plentiful time before starting the Ph.D. to learn vital courses in theoretical physics. It was a precious chance to study, make research, and know-how the scientific research can be.
It's from my pleasure to meet Prof. Emam. At the beginning of the Ph.D., the field of gravitational physics was almost new for me, but with his consistent support and encouragement, the field of gravitational physics and supergravity became facile for me.
Your guidance and your continuous advice to take each research step in
a deeper viewpoint sharpen my experiences and make this work completed as it has been today. So I'm grateful for your concern, and patience.
I'm so thankful for Prof. Hala Hashem, for her valuable support, and assistance.
I would like to thank the head of the physics department at the faculty of science, Al-Azhar University. I would like to thank my colleagues and my professors at Al-Azhar University.
I'm so grateful for my father and my mother, love you so much, rest in peace.
Hope this dissertation is useful for anyone who reaches for it. Hope our work includes answers for bizarre questions about our universe and even raises many other inquiries.
\begin{abstract}
In this thesis we study BPS 3-brane embedded in five dimensional $\mathcal{N}=2$ supergravity. We have found that when modeling our universe as a brane world filled with matter, radiation and/or pure energy, the time evolution of the brane world is strongly correlated to the complex structure moduli of the underlying Calabi-Yau manifold.
Moreover most solutions exhibit an early time inflation and late time acceleration, which means our universe’s whole cosmological history may be explained by only the bulk effects, without introducing an inflation field or even cosmological constant term in some cases. The constructed solutions are numeric and have been manifested analytically as well. Finally, since our solutions to the field equations are constrained by how far we know about the Calabi-Yau manifold itself, we have studied the solely time dependence of the complex structure moduli and the metric of the complex structure space in case the dimension of the complex structure space of the Calabi-Yau manifold $h_{(2,1)}=1$ .
\end{abstract}
\setcounter{tocdepth}{2} %
\tableofcontents %
\listoffigures
\mainmatter %
\unnumberedchapter{Introduction}
The trials to extend general relativity (GR) emerged very early, indeed just after its discovery by Einstein in 1915 by a couple of years. The questions at those times rose about if GR is a unique theory of gravity or not. Later on, in the beginning of the sixties, extending GR became indispensable because it is not renormalizable and, therefore, can not be conventionally quantized.
On the other hand, supersymmery (SUSY) proved itself as a theory that can alleviate some of the UV divergences of quantum field theory, via cancellations between bosonic and fermionic loops, hence the UV divergences of quantum gravity become milder in supergravity.
Supergravity (SUGRA) is a supersymmetric extension of GR, or a theory of local supersymmetry. It involves the graviton, a spin 2 elementary particle described by Einstein gravity, and extra matter, in particular a fermionic partner of the graviton called the gravitino (in case of $\mathcal{N}=1$ SUGRA).
Supergravity manifested itself as a realistic theory that describes nature including the four types of interactions, and also it has been found that SUGRA in ten spacetime dimensions is the low energy limit of superstring theories. Moreover, there exist a unique $\mathcal{N}=1$ supergravity theory in eleven spacetime dimensions which is the low energy limit of a more general nonperturbative quantum field theory in $\mathcal{D}=11$ called M-theory, from which the known superstring theories in ten dimensions arise as perturbative limits.
In contemporary fundamental physics M- theory is known as the leading candidate for the theory of everything (TOE). Nowadays the search for the complete formulation of TOE or any empirical evidence for it, namely the compactification of its extra dimensions to construct candidate models of our four- dimensional world is one of the most vital quests of the new century, if not the millennium.
There are many works have been done before to compactify $\mathcal{N}=1~ \mathcal{D}=11$ SUGRA over Calabi- Yau (CY) 3-fold to $\mathcal{N}=2~ \mathcal{D}=5$ SUGRA. We use these results to find different BPS (Bogomol’nyi-Prasad-Sommerfield) 3-brane solutions for the reduced theory.
These solutions couple to new fields and moduli that depend on the particular topology of CY.
BPS solutions conserve supersymmetry by setting the SUSY transformations of the model fermions to zero,
so they are bosonic solutions. Worthy to mention here that SUSY is half broken in supergravity as we will see later.
In this framework, we show that when our universe is modeled as a 3-brane embedded in five dimensional spacetime,
one of the cosmological ambiguity of the universe can be interpreted, typically what happens in the inflationary epoch when
adding an extra spatial dimension in case of matter, radiation and/or pure energy dominant brane-universes ?
Also the solutions of the field equations does not interpret only the early time inflation of the brane-world, but when adding cosmological constants for the brane and the bulk, they show late time acceleration which means the recent accelerated time expansion of the universe can be explained as well. Further they introduce possible scenarios about the future of the universe time evolution.
We manifest on the problem of the initial conditions and the multiverses. That we solve the field equations for different initial conditions after the Big Bang leading to a variety of possible universes including several universes that follow the accepted history of our own.
Last but not least, these solutions depend on the geometery of the Calabi–Yau manifold because the time dependence of the brane-worlds is found to be strongly correlated to the hypermultiplet fields, particularly the complex structure moduli of the underlying Calabi-Yau submanifold. So that we study the moduli and the metric of the complex structure space of the Calabi-Yau manifold in case $h_{(2,1)} =1$.
So this work has two main aims, first the deeper understanding of the topology of the Calabi-Yau manifold,
and the dimensional reduction techniques, that gets us closer to the structure of superstring theories
and the peculiar M- theory, and second to find possible applications of our results to our universe’s cosmological evolution.
The dissertation is organized as follows:
\begin{description}
\item[Chapter One] We introduce a brief review about string theory and supergravity in eleven dimensions.
\item[Chapter Two] We give a review about the different types of manifold, their developments and properties,
then we will talk about Calabi-Yau manifold and its complex structure space.
\item[Chapter Three] We will introduce the dimensional reduction of $\mathcal{D} = 11$ SUGRA to $\mathcal{D} = 5$ SUGRA over Calabi-Yau manifold ,
ending up by the theory in its symplectic form.
\item[Chapter Four] We numerically find solutions for $\mathcal{N}=2~ \mathcal{D}= 5$ SUGRA by solving Einstein field equations in case of radiation only, and dust only filled branes. We have found that when the universe is modeled
as a 3-brane world, in both cases the time evolution of the Calabi–Yau complex structure moduli strongly correlated to the expansion of the brane-world, showing an early short period of rapid expansion, that can interpret the inflationary epoch.
\item[Chapter Five] We study a brane- world filled with all matter, radiation and energy (represented by a brane cosmological constant term) to interpret the time being era of the universe's evolution with time. We also added a bulk cosmological constant, to see the whole effect of the bulk and the CY moduli on the brane- universe scale factor.
Also we find out that the solutions can predict the future time evolution of our universe.
We study the moduli's variation with respect to the fifth extra dimension, where the cosmological constant problem can be resolved
analytically.
\item[Chapter Six] We use our solutions to the field equations in case of a world brane filled by dust, radiation and energy with the field equation of the moduli and its complex conjugate to get analytic expressions to the time dependence of the moduli and the metric of the CY complex structure space in case the dimension of the complex structure space $h_{(2,1)}=1$. We show the moduli and the metric component G in the complex plane. Then in this case we can find analytic solutions to the the potential of the complex structure space and the volume of the Calabi–Yau manifold which are both related to the metric. Finally we have solved
Einstein equations analytically to manifest all the numerical results introduced.
\item[Conclusion] We summerize the main results of our study and point out to possible streams of future research.
\end{description}
\input{./Chapter/ch1}
\input{./Chapter/ch2}
\input{./Chapter/ch3}
\input{./Chapter/ch4}
\input{./Chapter/ch5}
\input{./Chapter/ch6}
\input{./Chapter/conc}
|
\section{Related Work}
\subsection{Materials Property Prediction.}
In recent times, data-driven approaches have become quite popular to establish relationships between the atomic structure of crystalline materials and their properties with very high precision. Especially, graph neural network (GNN) based approaches have emerged as a powerful machine learning model tool to encode material’s complex topological structure along with node features in an enriched representation space. Models such as CGCNN~\cite{xie2018crystal} represent 3D crystal structure as a multi-graph and build a graph convolution neural network directly on the graph to update node features based on their local chemical and structural environment. GATGNN~\cite{louis2020graph} incorporates the idea of graph attention network to learn the importance of different inter atomic bonds whereas MEGNet~\cite{chen2019graph} introduces global state attributes for quantitative structure-state property relationship prediction in materials. While all these models implicitly represent many body interactions through multiple graph convolution layers, ALIGNN~\cite{choudhary2021atomistic} explicitly captures many body interactions by incorporating bond angles and local geometric distortions.
\subsection{Graph Pre-training Strategies}
There are attempts to pre-train GNNs to extract graph and node level representations. ~\cite{hu2020pretraining} develops an effective pre-training strategy for GNNs, where they perform both node level and graph level pre-training on GNNs to capture domain specific knowledge about nodes and edges, in addition to global graph-level knowledge. GPT-GNN~\cite{hu2020gpt} proposes a self-supervised attributed graph generation task to pre-train a GNN model, which captures structural and semantic properties of the graph. ~\cite{qiu2020gcc} presents GCC, which leverages the idea of contrastive learning to design the graph pre-training task as subgraph instance discrimination, to capture universal network topological properties across multiple networks. Another recent work is GraphCL~\cite{you2020graph}, which proposes different graph augmentation methods and maximises the agreement between two augmented views of the same graph via a contrastive loss.
In the field of crystal graphs, CrysXPP~\cite{das2022crysxpp} is the only model which comes close to a pre-trained model. In their work, an autoencoder is trained on a volume of un-tagged crystal graphs and the learned knowledge is (transferred to) used to initialize the encoder of CrysXPP, which is fine-tuned with property specific tagged data. \\
To the best of our knowledge, CrysGNN{} is the first attempt to develop a deep pre-trained model for the crystalline materials and for that, we curated a new large untagged crystal dataset with 800K crystal graphs.
\section{Crystal Representation}
We realize a crystal material as a multi-graph structure $\mathcal{G}_i =(\mathcal{V}_i, \mathcal{E}_i, \mathcal{X}_i, \mathcal{F}_i )$ as proposed in \cite{das2022crysxpp,xie2018crystal}. $\mathcal{G}_i$ is an undirected weighted multi-graph where $\mathcal{V}_i$ denotes the set of nodes or atoms present in a unit cell of the crystal structure. $\mathcal{E}_i=\{(u,v,k_{uv})\}$ denotes a multi-set of node pairs and $k_{uv}$ denotes number of edges between a node pair $(u,v)$. $\mathcal{X}_i=\{(x_{u} | u \in \mathcal{V}_i )\}$ denotes 92 dimensional node feature set proposed by CGCNN~\cite{xie2018crystal}. It includes different chemical properties like electronegativity, valance electron, covalent radius, etc. Details of these node features are given in Table \ref{tab:feature_desc}. Finally, $\mathcal{F}_i=\{\{s^k\}_{(u,v)} | (u,v) \in \mathcal{E}_i, k\in\{1..k_{uv}\}\}$ denotes the multi-set of edge weights where $s^k$ corresponds to the $k^{th}$ bond length between a node pair $(u,v)$, which signifies the inter-atomic bond distance between two atoms.
\begin{figure}[h]
\centering
\vspace*{-1mm}
\boxed{\includegraphics[width=\columnwidth, height=45mm]{Figure/crystal_graph_representation.pdf}}
\caption{Multi-graph Representation from Crystal 3D Structure.}
\label{fig:3d_crsytal_rep}
\end{figure}
\begin{table}[]
\centering
\small
\setlength{\tabcolsep}{4pt}
\scalebox{0.9}{
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Features} & \textbf{Unit} &\textbf{Range} & \textbf{Feature Dimension} \\
\hline
Group Number & - & 1,2, ..., 18 & 18 \\
\hline
Period Number & - & 1,2, ..., 9 & 9\\
\hline
Electronegativity & - & 0.5-4.0 & 10 \\
\hline
Covalent Radius & pm & 25-250 & 10 \\
\hline
Valence Electrons & - &1,2, ..., 12 & 12\\
\hline
First Ionization Energy & eV & 1.3–3.3 & 10\\
\hline
Electron Affinity & eV & -3–3.7 & 10 \\
\hline
Block & - & s, p, d, f & 4 \\
\hline
Atomic Volume & $cm^{3} /mol$ & 1.5–4.3 & 10 \\
\hline
\end{tabular}
}
\caption{Description of different atomic features $\mathcal{X}_i$ and their unit, range and dimensions. All of them together forms 92 dimensional node feature vector $\mathcal{X}_i$.}
\centering
\label{tab:feature_desc}
\end{table}
\section {Details of GNN Architecture in CrysGNN{}}
Here we describe GNN architectures used in CrysGNN for pre-training and property prediction model.
\subsubsection{Encoder.} We develop an graph convolution \cite{xie2018crystal} based encoding module, which takes crystal multi-graph structure $\mathcal{G} =(\mathcal{V}, \mathcal{E}, \mathcal{X}, \mathcal{F} )$ as input and encode structural semantics of the crystal graph into lower dimensional space. Considering $L$ graph convolution layers, neighbourhood aggregation of the $l$-th layer is represented as :
\begin{equation}
\label{eq:cgcnn}
z_u^{(l+1)} = z_u^{(l)} + \sum\limits_{v,k} \sigma ( {h^{(l)}_{(u,v)}}_k W_c^{(l)} + b_c^{(l)}) \odot g({h^{(l)}_{(u,v)}}_k W_s^{(l)}+b_s^{(l)})
\end{equation}
where $h_{(u,v)_k}^{(l)} = z_u^{(l)} \oplus z_v^{(l)} \oplus {s^{(k)}_{(u,v)}}$. Here $s^{(k)}_{(u,v)}$ represents inter-atomic bond length of the $k$-th bond between $(u,v)$ and $z_u^{(l)}$ denotes embedding of node $u$ after $l$-th layer, which is initialized to a transformed node feature vector $z_u^0 := x_u W_x$ where $W_x$ is the trainable parameter and $ x_u $ is the input node feature vector (Table. \ref{tab:feature_desc}). In Equation \ref{eq:cgcnn}, $W_c^{(l)},W_s^{(l)},b_c^{(l)},b_s^{(l)}$ are the convolution weight matrix, self weight matrix, convolution bias, and self bias of the $l$-th layer, respectively, $\oplus$ operator denotes concatenation, $\odot$ denotes element-wise multiplication, $\sigma$ is the sigmoid function indicating the edge importance and $g$ is a feed forward network. After $L$-layers of such propagation, $\mathcal{Z} = \{z_1,..., z_{|\mathcal{V}|}\}$ represents the set of final node embeddings, where $z_u :=z_u^L$ represents final embedding of node $u$.
\subsubsection{Decoder.} In the decoding module, we apply both node-level and graph-level self-supervised pre-training as discussed in \textbf{Node Level Decoding} and \textbf{Graph Level Decoding} sections in the main manuscript.
\subsubsection{Property Predictor.} For the property prediction model, we use four diverse sets of SOTA models, which have different encoder choices and on top of that they have a regressor module to predict the property value. We retrofit CrysGNN{} model with each of these SOTA models using the modified multitask loss (incorporating knowledge distillation loss with their property loss) as discussed in Equation \ref{eq:finetune_loss} (main manuscript).
\section{Crystal System and Space group Information}
By definition crystal is a periodic arrangement of repeating “motifs”( e.g. atoms, ions). The symmetry of a periodic pattern of repeated motifs is the total set of symmetry operations allowed by that pattern. The total set of such symmetry operations, applicable to the pattern is the pattern's symmetry, and is mathematically described by a so-called Space Group. The Space Group of a Crystal describes the symmetry of that crystal,
and as such it describes an important aspect of that crystal's internal structure.\\
In turn crystals across certain space groups show
similarities among each other, hence they are divided among seven Crystal Systems : Triclinic, Monoclinic, Orthorhombic, Tetragonal, Trigonal, Hexagonal, and Cubic or Isometric system ~\cite{kittel2018kittel}.\\
\textbf{Crystallographic Axes: } The crystallographic axes are imaginary lines that we can draw within the crystal lattice. These will define a coordinate system within the crystal. We refer to the axes as \textbf{a}, \textbf{b}, \textbf{c} and angle between them as $\alpha$, $\beta$ and $\gamma$. By using these crystallographic axes we can define six large groups or crystal systems.
\begin{itemize}
\item \textbf{Cubic or Isometric system :} The three crystallographic axes are all equal in length and intersect at right angles to each other.
\begin{equation}
\boxed{a=b=c\;; \;\; \alpha = \beta =\gamma = 90 \degree} \notag
\end{equation}
\item \textbf{Tetragonal system :} Three axes, all at right angles, two of which are equal in length (a and b)
and one (c) which is different in length (shorter or longer).
\begin{equation}
\boxed{a=b\neq c\;; \;\; \alpha = \beta =\gamma = 90 \degree} \notag
\end{equation}
\item \textbf{Orthorhombic system :} Three axes, all at right angles, all three have different length.
\begin{equation}
\boxed{a\neq b\neq c\;; \;\; \alpha = \beta =\gamma = 90 \degree} \notag
\end{equation}
\item \textbf{Monoclinic system :} Three axes, all unequal in length, two of which (a and c) intersect at an oblique angle (not 90 degrees), the third axis (b) is perpendicular to the other two axes.
\begin{equation}
\boxed{a\neq b\neq c\;; \;\; \alpha =\gamma = 90 \degree \neq \beta} \notag
\end{equation}
\item \textbf{Triclinic system :} The three axes are all unequal in length and intersect at three different angles (any angle but 90 degrees).
\begin{equation}
\boxed{a\neq b\neq c\;; \;\; \alpha \neq \beta \neq \gamma \neq 90 \degree} \notag
\end{equation}
\item \textbf{Hexagonal system :} Here we have four axes. Three of the axes fall in the same plane and intersect at the axial cross at 120°. These 3 axes, labeled $a_1$, $a_2$, and $a_3$, are the same length. The fourth axis, c, may be of {\em a} different length than the axes set. The c axis also passes through the intersection of the {\em a} axes
set at right angle to the plane formed by the {\em a} set.
\end{itemize}
\section{Details about datasets and different crystal properties.}
\textbf{OQMD:} The Open Quantum Materials Database (OQMD) is a high-throughput database based on DFT calculations of chemicals from the Inorganic Crystal Structure Database (ICSD) and embellishments of regularly occurring crystal structures \cite{OQMD}. Comprehensive computations for significant structure types in materials research, such as Heusler and perovskite compounds, are among the distinguishing aspects of this database. The database is public and a free resource.\\\\
\textbf{Materials project:} This is another free and public resource database that contains crystal structures and calculated materials properties~\cite{MP}. The dataset is made from the results obtained with density functional theory based calculations. The dataset is composed of electronic structure, thermodynamic, mechanical, and dielectric properties. It also provides a visual web-based search interface.\\\\
\textbf{JARVIS:} JARVIS (Joint Automated Repository for Various Integrated Simulations) is a data repository that incorporates not only DFT-based calculations, but also data from classical force-fields and machine learning approaches~\cite{JARVIS}. The databases are free and public as well.\\\\
We curated around {800K untagged crystal graph data from two popular materials databases, Materials Project (MP) and OQMD}, to pre-train CrysGNN{} model. Further to evaluate the performance of different SOTA models with distilled knowledge from CrysGNN{}, we select {MP 2018.6.1 version of Materials Project} and {2021.8.18 version of JARVIS-DFT}, another popular materials database, for property prediction as suggested by \cite{choudhary2021atomistic}. Please note, MP 2018.6.1 dataset is a subset of the dataset used for pre-training, whereas JARVIS-DFT is a separate dataset which is not seen during the pre-training.
MP 2018.6.1 consists of 69,239 materials with two properties namely bandgap and formation energy, whereas JARVIS-DFT consists of 55,722 materials with 19 properties which can be broadly classified into two categories : 1) properties like formation energy, bandgap, total energy, bulk modulus, etc. which depend greatly on crystal structures and atom features, and 2) properties like $\epsilon_x$, $\epsilon_y$, $\epsilon_z$, n-Seebeck, n-PF, etc. which depend on the precise description of the materials’ electronic structure. Details of these properties is provided in Table \ref{tbl-prop} \\
Moreover, all these properties in both Materials Project and JARVIS-DFT datasets are based on DFT calculations of chemicals. Therefore, to investigate how pre-trained knowledge helps to mitigate the DFT error, we also take a small dataset ~\cite[OQMD-EXP]{kirklin2015open} containing 1,500 available experimental data of formation energy. Details of each of these datasets are given in Table \ref{tbl-dataset}.
\begin{table}
\centering
\small
\setlength{\tabcolsep}{10 pt}
\begin{tabular}{ccc}
\toprule
Property & Unit & Data-size\\
\midrule
Formation\_Energy & $eV/(atom)$ & 69239 \\
Bandgap (OPT) & $eV$ & 69239 \\
\midrule
Formation\_Energy & $eV/(atom)$ & 55723 \\
Bandgap (OPT) & $eV$ & 55723 \\
Total\_Energy & $eV/(atom)$ & 55723\\
Ehull & $eV$ & 55371\\
Bandgap (MBJ) & $eV$ & 18172 \\
Bulk Modulus (Kv) & GPa & 19680\\
Shear Modulus (Gv) & GPa & 19680\\
SLME (\%) & No unit & 9068\\
Spillage & No unit & 11377 \\
$\epsilon_x$ (OPT) & No unit & 44491 \\
$\epsilon_y$ (OPT) & No unit & 44491\\
$\epsilon_z$ (OPT) & No unit & 44491 \\
$\epsilon_x$ (MBJ) & No unit & 16814\\
$\epsilon_y$ (MBJ) & No unit & 16814\\
$\epsilon_z$ (MBJ) & No unit & 16814 \\
n-Seebeck & $\mu VK^{-1}$ & 23210\\
n-PF & $\mu W(mK^2)^{-1}$ & 23210\\
p-Seebeck & $\mu VK^{-1}$ & 23210\\
p-PF & $\mu W(mK^2)^{-1}$ & 23210\\
\bottomrule
\end{tabular}
\caption{Summary of different crystal properties in Materials Project (Top) and JARVIS-DFT (Bottom) datasets.}
\label{tbl-prop}
\end{table}
\begin{table*}
\centering
\small
\setlength{\tabcolsep}{11 pt}
\scalebox{0.9}{
\begin{tabular}{c | c c | c c| c c | c c}
\toprule
Property & CGCNN & CGCNN & CrysXPP & CrysXPP & GATGNN & GATGNN & ALIGNN & ALIGNN\\
& & (Distilled) & & (Distilled) & & (Distilled) & & (Distilled)\\
\midrule
$\epsilon_x$ (OPT) & 24.93 & \textbf{23.86} & 24.47 & \textbf{23.33} & 26.79 & \textbf{25.23} & 21.42 & \textbf{21.37}\\
$\epsilon_y$ (OPT) & 25.06 & \textbf{24.08} & 24.84 & \textbf{23.98} & 26.77 & \textbf{25.02} & 21.66 & \textbf{21.25} \\
$\epsilon_z$ (OPT) & 24.99 & \textbf{23.85} & 24.13 & \textbf{23.76} & 26.09 & \textbf{24.81} & \textbf{20.51} & 21.03 \\
$\epsilon_x$ (MBJ) & 27.18 & \textbf{26.64} & 28.40 & \textbf{26.03} & \textbf{27.56} & 27.61 & 24.76 & \textbf{24.32}\\
$\epsilon_y$ (MBJ) & 27.14 & \textbf{25.80} & 26.92 & \textbf{25.34} & 27.21 & \textbf{26.89} & 23.99 & \textbf{23.83}\\
$\epsilon_z$ (MBJ) & 28.38 & \textbf{26.08} & 26.76 & \textbf{25.97} & 25.50 & \textbf{24.84} & \textbf{23.93} & 24.59\\
n-Seebeck & 50.32 & \textbf{46.66} & 48.79 & \textbf{46.53} & 52.03 & \textbf{49.46} & 42.85 & \textbf{42.13}\\
n-PF & 528.25 & \textbf{506.41} & 502.79 & \textbf{498.43} & 511.56 & \textbf{505.22} & 477.09 & \textbf{474.88} \\
p-Seebeck & 53.01 & \textbf{49.15} & 50.13 & \textbf{48.67} & 54.99 & \textbf{50.72} & 46.04 & \textbf{45.19}\\
p-PF & 513.57 & \textbf{501.11} & 493.35 & \textbf{489.54} & 516.84 & \textbf{494.92} & 460.11 & \textbf{447.81}\\
\bottomrule
\end{tabular}
}
\caption{Summary of the prediction performance (MAE) of different properties (belong to second class) in JARVIS-DFT dataset where performance improvement is modest. Model M is the vanilla variant of a SOTA model and M (Distilled) is the distilled variant using the pretrained CrysGNN{}. All the models are trained on $80\%$ data, validated on $10\%$ and evaluated on $10\%$ of the data. The best performance is highlighted in bold.}
\label{tbl-full-data-2}
\end{table*}
\section{Training Setup / Hyper-parameter Details / Computational Resources used.}
We use five convolution layers of the encoder module to train CrysGNN{} and train it for 200 epochs using Adam \cite{kingma2014adam} for optimization with a learning rate of 0.03. We keep the embedding dimension for each node as 64, batch size of data as 128, and equal weightage (0.25) for $\alpha$, $\beta$, $\gamma$, and $\lambda$ of Equation \ref{eq:pretrain_loss}. During property prediction, we used the default configuration for hyper-parameters for all four vanilla SOTA models. However, while predicting properties by the distilled version of the SOTA models we keep node embedding dimension for each SOTA as 64, and train each model for 500 epochs using Adam~\cite{kingma2014adam} optimization and keep $\delta$ as 0.5. We perform the experiments in shared servers having Intel E5-2620v4 processors which contain 16 cores/thread and four GTX 1080Ti 11GB GPUs each.
\section{Baseline Models.}
To evaluate the effectiveness of CrysGNN{}, we choose following four diverse state of the art algorithms for crystal property prediction.
\begin{enumerate}
\item \textbf{CGCNN}~\cite{xie2018crystal} : This work generates crystal graphs from inorganic crystal materials and builds a graph convolution based supervised model for predicting various properties of the crystals.
\item \textbf{GATGNN}~\cite{louis2020graph} : In this work, authors have incorporated a graph neural network with multiple graph-attention layers (GAT) and a global attention layer, which can learn efficiently the importance of different complex bonds shared among the atoms within each atom’s local neighborhood.
\item \textbf{ALIGNN}~\cite{choudhary2021atomistic} : In this work, authors adopted line graph neural networks to develop an alternative way to
include angular information into convolution layer which alternates between message passing on the bond graph and its bond-angle line graph.
\item \textbf{CrysXPP}~\cite{das2022crysxpp} : In this work, the authors train an autoencoder (CrysAE) on a volume of un-tagged crystal graphs and then the learned knowledge is (transferred to) used to initialize the encoder of CrysXPP, which is fine-tuned with property specific tagged data.Also they design a feature selector that helps to interpret the model’s prediction.
\end{enumerate}
\subsection{Baseline Implementations.} We used the available PyTorch implementations of all the baselines, viz, CGCNN\footnote{https://github.com/txie-93/cgcnn.git}, GATGNN\footnote{https://github.com/superlouis/GATGNN.git}, CrysXPP\footnote{https://github.com/kdmsit/crysxpp.git} and ALIGNN\footnote{https://github.com/usnistgov/alignn.git}. In order to retrofit CrysGNN{} to each SOTA model, we modified their supervised loss into the multi-task loss proposed in Equation \ref{eq:finetune_loss} (main manuscript) and train each model. To ensure a fair comparison between all the baselines and perform the knowledge distillation between node embeddings, we have ensured node embedding dimension for each SOTA and CrysGNN{} as 64.
\begin{table*}
\centering
\small
\setlength{\tabcolsep}{6 pt}
\scalebox{0.9}{
\begin{tabular}{cccccccccc}
\toprule
Property & Train-Val & CGCNN & CGCNN & CrysXPP & CrysXPP & GATGNN & GATGNN & ALIGNN & ALIGNN\\
& -Test(\%) & & (Distilled) & & (Distilled) & & (Distilled) & & (Distilled)\\
\midrule
\multirow{3}{*}{\shortstack{Bandgap \\ (MBJ)}}
& 20-10-70 & 0.588 & 0.453\textbf{*} (23.04) & 0.598 & 0.450\textbf{*} (24.82) & 0.541 & 0.521 (3.70) & 0.497 & 0.485 (2.53) \\
& 40-10-50 & 0.532 & 0.419\textbf{*} (21.41) & 0.496 & 0.405\textbf{*} (18.40) & 0.462 & 0.448\textbf{*} (2.81) & 0.404 & 0.395 (2.20)\\
& 60-10-30 & 0.449 & 0.364 (19.08) & 0.435 & 0.360 (17.36) & 0.449 & 0.439 (2.29) & 0.387 & 0.380 (1.98)\\
\midrule
\multirow{3}{*}{\shortstack{Bulk Modulus \\ (Kv)}}
& 20-10-70 & 16.91 & 16.26 (3.80) & 15.42 & 14.25\textbf{*} (7.59) & 14.80 & 14.19 (4.12) & 14.70 & 14.06 (4.35)\\
& 40-10-50 & 14.81 & 14.46 (2.36) & 15.13 & 14.02\textbf{*} (7.34) & 12.98 & 12.59 (3.00) & 12.47 & 12.11 (2.89)\\
& 60-10-30 & 14.23 & 14.05 (1.26) & 14.76 & 13.73 (6.98) & 12.01 & 11.75 (2.16) & 11.23 & 11.01 (1.96)\\
\midrule
\multirow{3}{*}{\shortstack{Shear Modulus \\(Gv)}}
& 20-10-70 & 13.89 & 12.50 (10.01) & 13.39 & 12.07\textbf{*} (9.86) & 12.83 & 12.42 (3.20) & 12.71 & 12.31 (3.15)\\
& 40-10-50 & 12.04 & 11.54\textbf{*} (4.15) & 12.16 & 11.01\textbf{*} (9.46) & 11.43 & 11.23 (1.75) & 10.98 & 10.67 (2.82)\\
& 60-10-30 & 11.75 & 11.31 (3.74) & 11.77 & 10.67 (9.35) & 10.65 & 10.47 (1.69) & 10.24 & 10.04 (1.95)\\
\midrule
\multirow{3}{*}{\shortstack{SLME (\%)}}
& 20-10-70 & 7.13 & 6.62 (7.17) & 7.05 & 5.90\textbf{*} (16.40) & 6.35 & 6.02 (5.20) & 6.36 & 6.27 (1.43)\\
& 40-10-50 & 6.14 & 5.78 (5.90) & 6.89 & 5.81 (15.67) & 5.78 & 5.63 (2.60) & 5.65 & 5.57 (1.42)\\
& 60-10-30 & 5.55 & 5.24 (5.68) & 5.41 & 4.84 (10.54) & 5.48 & 5.34 (2.55) & 4.88 & 4.82 (1.33)\\
\midrule
\multirow{3}{*}{\shortstack{Spillage}}
& 20-10-70 & 0.424 & 0.389\textbf{*} (8.43) & 0.422 & 0.403 (4.45) & 0.406 & 0.392\textbf{*} (3.5) & 0.402 & 0.392 (2.49)\\
& 40-10-50 & 0.408 & 0.391\textbf{*} (4.28) & 0.398 & 0.384 (3.57) & 0.396 & 0.388 (1.94) & 0.378 & 0.372 (1.72)\\
& 60-10-30 & 0.395 & 0.380 (3.80) & 0.379 & 0.368 (3.00) & 0.382 & 0.377 (1.36) & 0.362 & 0.357 (1.13)\\
\bottomrule
\end{tabular}
}
\caption{MAE values of five different properties in JARVIS-DFT dataset with the increase in training instances from 20 to 60\%. Model M is the vanilla variant of the SOTA model, M (Distilled) is distilled variant of it and relative improvement is mentioned in bracket. Distilled knowledge from CrysGNN{} improves the performance of all the baselines consistently. In few cases, we observe that MAE values of the distilled version of a model using even lesser training data is better than vanilla version of the same model using more training data and highlight it with *.}
\label{tbl-limited-data}
\end{table*}
\section{Additional Experimental Results}
\subsection{Evaluation on Electronic Properties.}
Properties of crystalline materials can be broadly classified into two categories : 1) properties like formation energy, bandgap, total energy, bulk modulus, etc. which depend greatly on crystal structures and atom features, and 2) properties like $\epsilon_x$, $\epsilon_y$, $\epsilon_z$, n-Seebeck, n-PF, etc. which depend on the precise description of the materials’ electronic structure. We discuss performance of SOTA models with distilled knowledge from CrysGNN{} for the first class of properties in {\bf ``Downstream Task Evaluation"} section and report the MAE in Table \ref{tbl-full-data-1}. We further evaluate the effectiveness of CrysGNN for the second class of properties for both the vanilla and distilled model and report the MAE in Table \ref{tbl-full-data-2}. We observe, for this class of properties all the SOTA models had a higher MAE and though pre-training enhances the performance, the improvement is modest. In specific, average improvement in CGCNN, CrysXPP, GATGNN and ALIGNN are 6\%, 4\%, 3.6\% and 0.8\%, respectively. A potential reason is, electronic dielectric constant, Seebeck coefficients, and power factors all depend greatly on the precise description of the materials' electronic structure, which is neither captured by the SOTA models nor by the pre-trained CrysGNN{} framework explicitly. Hence error is high for these properties by SOTA models and injecting distilled structural information from the pre-trained model is able to achieve only modest improvements.\\
\subsection{Effectiveness on sparse training dataset.}
A concise version of this result is presented in the main paper, we here elaborate some details.
To demonstrate the effectiveness of the pre-training in limited data settings, we conduct additional set of experiments under different training data split. In specific, we vary available training data from 20 to 60 \%, train different SOTA models and check their performance on test dataset. We report the MAE values of different baselines and their distilled version in Table \ref{tbl-limited-data} for five different properties, for which available data is very limited. We observe that the distilled version of any SOTA model consistently outperforms its vanilla version even in the limited training data setting, which illustrates the robustness of our pre-training framework.
Specifically, for CGCNN and GATGNN, the improvements are more for 20\% training data, because these deep models suffer from data scarcity issue and with distilled knowledge from the pre-trained model they are able to mitigate the issue. ALIGNN achieves moderate improvements as before but surprisingly, we found the least improvement (additional improvement for 20\% training data compared to other settings) for CrysXPP. CrysXPP itself pre-trains an autoencoder and transfers the learned information to the property predictor. Hence, the vanilla version performs well in the limited data settings. \\
Moreover, in some cases, we observe that the MAE values of the distilled version of a model using lesser training data is better than vanilla version of the model using more training data. For example, while predicting Bandgap (MBJ), CGCNN (Distilled) with 20\% and 40\% training data outperforms CGCNN with 40\% and 60\% training data, respectively. These results verify that even with lesser training data, the distilled information from the large pre-trained model gives performance comparable to using larger training data by the original model.
\section{Ablation Study : Analysis of Different Pre-training Loss}
A concise version of this result is presented in the main paper, we here elaborate some details.
We perform an ablation study to investigate the influence of different pre-training losses in enhancing the SOTA model performance. While pre-training CrysGNN{} (Eq. \ref{eq:pretrain_loss}), we capture both local chemical and global structural information via node and graph-level decoding, respectively.
Further, we are curious to know the influence of each of these decoding policies independently in the downstream property prediction task.
\tblak{In specific, we conduct the ablation experiments, where we pre-train CrysGNN{} with (a) only node-level decoding ($\mathcal{L}_{FR}$, $\mathcal{L}_{CR}$), (b) only graph-level decoding ($\mathcal{L}_{SG}$, $\mathcal{L}_{NTXent}$). Further, we perform ablations with individual graph-level losses, and pretrain with (c) removing $\mathcal{L}_{NTXent}$ (node-level with $\mathcal{L}_{SG}$ (space group)) and (d) removing $\mathcal{L}_{SG}$ (node-level with $\mathcal{L}_{NTXent}$(crystal system)).} We train two baseline models, CGCNN and ALIGNN, with distilled knowledge from all the aforementioned variants of the pre-trained model and evaluate the performance on four crystal properties. We report performance of both the experiments for four properties in Fig.\ref{fig:ablation_diff_loss} and for ten different properties in Fig.\ref{fig:ablation_diff_loss_appendix}.\\
We can observe clearly that all the variants offer significant performance gain in all four properties using the combined node and graph-level pre-training, compared to node-level or graph-level pre-training separately. Only exception is formation energy, where only node-level pre-training produces less error compared to other variants, in both the baseline. Formation energy of a crystal is defined as the difference between the energy of a unit cell comprised of $N$ chemical species and the sum of the chemical potentials of all the $N$ chemical species. Hence pre-training at the node-level (node features and connection) is adequate for enhancing performance of formation energy prediction and incorporating graph-level information works as a noisy information, which degrades the performance. \tblak{We also observe improvement in performance using both supervised and contrastive graph-level losses ($\mathcal{L}_{SG}$ and $\mathcal{L}_{NTXent}$), compared to using only one of them, which proves that the learned representation via supervised and contrastive learning is more expressive that using any one of them.}
Moreover, in ALIGNN, with either node or graph-level pre-training separately, performance degrades across different properties. ALIGNN explicitly captures the three body interactions which drive its performance, to replicate that inclusion of both node and graph information is necessary.
\subsection{Limitations and Future Work.}
Though our proposed pre-trained model is able to enhance the performance of all the state of the art baseline models, improvements for all types of SOTA models are not consistent and we found lesser improvement in case of complex models like GATGNN and ALIGNN. This provides scope for further investigation on designing a more complex and deeper pre-trained model. There are also scopes of exploring different graph representations, loss variations, distillation strategies, etc. Of course, the very fact that the simple model, CrysGNN{}, can provide substantial improvements, lays the foundation for such future exploration. Moreover, in this present work, we have focused on predicting crystal properties, which is a graph level regression task. However, the same framework can be used to explore the effect of pre-training on other graph level tasks (classification or clustering) or node level tasks which would be one of our future works.
\begin{figure*}[ht]
\centering
\boxed{\includegraphics[width=2\columnwidth]{Figure/Legend.pdf}}
\\
\vspace*{-1mm}
\subfloat[Shear Modulus (Gv)]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/shear_modulus.pdf}}}
\subfloat[SLME(\%)]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/slme.pdf}}}
\subfloat[Spillage]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/spillage.pdf}}}
\subfloat[$\epsilon_x$ (MBJ)]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/mepsx.pdf}}}\\
\subfloat[$\epsilon_y$ (MBJ) ]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/mepsy.pdf}}}
\subfloat[$\epsilon_z$ (MBJ) ]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/mepsz.pdf}}}
\subfloat[n-Seebeck]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/n-sb.pdf}}}
\subfloat[n-PF]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/n-pf.pdf}}}\\
\subfloat[p-Seebeck]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/p-sb.pdf}}}
\subfloat[p-PF]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/p-pf.pdf}}}
\caption{Summary of experiments of ablation study on importance of different pre-training loss components on CrysGNN{} training and eventually its effect on CGCNN and ALIGNN models on ten different properties from JARVIS DFT dataset(MAE for property prediction). (i) Vanilla: SOTA based model (without distillation) and all the other cases are SOTA models (distilled) from different pre-trained version of CrysGNN{}.
(ii) Node : only node-level pre-training, (iii) Graph : only graph-level pre-training, (iv) Node + L(SG) : node-level and $\mathcal{L}_{SG}$, (v) Node + L(NTXent) : node-level and $\mathcal{L}_{NTXent}$ and (vi) CrysGNN{}~: both node and graph-level pre-training.}
\label{fig:ablation_diff_loss_appendix}
\end{figure*}
\section{Conclusion}
\label{conclusion}
In this work, we present a novel but simple pre-trained GNN framework, CrysGNN{}, for crystalline materials, which captures both local chemical and global structural semantics of crystal graphs. To pre-train the model, we curate a huge dataset of 800k unlabelled crystal graphs.
Further, while predicting different crystal properties, we distill important knowledge from CrysGNN{} and inject it into different state of the art property predictors and enhance their performance. Extensive experiments on multiple popular datasets and diverse set of SOTA models show that with distilled knowledge from the pre-trained model, all the SOTA models outperform their vanilla versions.
Extensive experiments show its superiority over conventional fine-tune models and its inherent ability to remove DFT-induced bias.
The pretraining framework can be extended beyond structural graph information in a multi-modal setting to include other important (text and image) information about a crystal which would be our immediate future work.
\section{Acknowledgments}
This work was partially funded by Indo Korea Science and Technology Center, Bangalore, India, under the project name {\em Transfer learning and Weak Supervision for Accurate and Interpretable Prediction of Properties of Materials from their Crystal Graph Representation}. Also, it was (partially) funded by the Federal Ministry of Education and Research (BMBF), Germany under the project LeibnizKILabor with grant No.\,01DD20003. We thank the Ministry of Education, Govt of India, for supporting Kishalay with Prime Minister Research Fellowship during his Ph.D. tenure.
\section{Conclusion}
\label{conclusion}
In this work, we present a novel but simple pre-trained GNN framework, CrysGNN{}, for crystalline materials, which captures both local chemical and global structural semantics of crystal graphs. To pre-train the model, we curate a huge dataset of 800k unlabelled crystal graphs.
Further, while predicting different crystal properties, we distill important knowledge from CrysGNN{} and inject it into different state of the art property predictors and enhance their performance. Extensive experiments on multiple popular datasets and diverse set of SOTA models show that with distilled knowledge from the pre-trained model, all the SOTA models outperform their vanilla versions.
Extensive experiments show its superiority over conventional fine-tune models and its inherent ability to remove DFT-induced bias.
The pretraining framework can be extended beyond structural graph information in a multi-modal setting to include other important (text and image) information about a crystal which would be our immediate future work.
\section{Experimental Results}
\label{exp_results}
In this section, we evaluate how the distilled knowledge from CrysGNN{} enhances the performance of different state of the art property predictors on a diverse set of crystal properties from two popular benchmark materials datasets. We first briefly discuss the datasets used both in pre-training and downstream property prediction tasks. Then we report the results of different SOTA property predictors on the downstream property prediction tasks.
Next, we illustrate the effectiveness of our knowledge distillation method compared to the conventional fine-tuning approach.
We further conduct some ablation studies to show the influence of different pre-training losses in predicting different crystal properties and the performance of the system to sparse dataset. Finally, we demonstrate how distilled knowledge from the pre-trained model aids the SOTA models to remove DFT error bias, using very little experimental data.
\subsection{Datasets}
We curated 800K untagged crystal graph data from two popular materials databases, Materials Project (MP) and OQMD, to pre-train CrysGNN{} model. Further to evaluate the performance of different SOTA models with distilled knowledge from CrysGNN{}, we select MP 2018.6.1 version of Materials Project and 2021.8.18 version of JARVIS-DFT, another popular materials database, for property prediction as suggested by \cite{choudhary2021atomistic}. \tblak{Please note, MP 2018.6.1 dataset is a subset of the dataset used for pre-training, whereas JARVIS-DFT is a separate dataset which is not seen during the pre-training.}
MP 2018.6.1 consists of 69,239 materials with two properties bandgap and formation energy, whereas JARVIS-DFT consists of 55,722 materials with 19 properties \tblak{which can be broadly classified into two categories : 1) properties like formation energy, bandgap, total energy, bulk modulus, etc. which depend greatly on crystal structures and atom features, and 2) properties like $\epsilon_x$, $\epsilon_y$, $\epsilon_z$, n-Seebeck, n-PF, etc. which depend on the precise description of the materials’ electronic structure. In the following section, we will evaluate effectiveness of CrysGNN{} on the first class of properties. The impact of structural information is marginal on the second class of property hence all the SOTA perform poorly, there is however some modest improvement using CrysGNN{}; we have put the results and the discussion about it in the Appendix.}\\
Moreover, all these properties in both Materials Project and JARVIS-DFT datasets are based on DFT calculations of chemicals. Therefore, to investigate how pre-trained knowledge helps to mitigate the DFT error, we also take a small dataset OQMD-EXP~\cite{kirklin2015open}, containing 1,500 available experimental data of formation energy. Details of each of these datasets are given in Table \ref{tbl-dataset}. More detail about dataset, different crystal properties and experimental setup is in Appendix
\subsection{Downstream Task Evaluation}
\label{main_results}
To evaluate the effectiveness of CrysGNN{}, we choose four diverse state of the art algorithms for crystal property prediction, CGCNN~\cite{xie2018crystal}, GATGNN~\cite{louis2020graph}, CrysXPP~\cite{das2022crysxpp} and ALIGNN~\cite{choudhary2021atomistic}. To train these models for any specific property, we adopt the multi-task setting discussed in equation \ref{eq:finetune_loss} ,where we distill relevant knowledge from the pre-trained CrysGNN{} to each of these algorithms to predict different properties.
We report mean absolute error (MAE) of the predicted and actual value of a particular property to compare the performance of different participating methods. For each property, we
trained on $80\%$ data, validated on $10\%$ and evaluated on $10\%$ of the data.
We compare the results of distilled version of each SOTA model with its vanilla version (version reported in the respective papers), to show the effectiveness of the proposed framework.\\
\noindent\textbf{Results.}
\tblak{In Table \ref{tbl-full-data-1},
we report MAE of different crystal properties of Materials project and JARVIS-DFT datasets.
In the distilled version of the SOTA models, while training the model, we distill information from the pre-trained CrysGNN{} model. We observe that the distilled version of any state-of-the-art model outperforms the vanilla model across all the properties. In specific, average improvement in CGCNN, CrysXPP, GATGNN and ALIGNN are 16.20\%, 12.21\%, 8.02\% and 4.19\%, respectively.
These improvements are particularly significant as in most of the cases, the MAE is already low for SOTA models, still pretraining enables improvement over that. In fact, lower the MAE, higher the improvement.
We calculate Spearman's Rank Correlation between MAE for each property across different SOTA models and their improvement due to distilled knowledge and found it to be very high (0.72), which supports the aforementioned observations.
The average relative improvement across all properties for ALIGNN (4.19\%) and GATGNN (8.02\%) is lesser compared to CGCNN (16.20\%) and CrysXPP (12.21\%). A possible reason could be that
ALIGNN and GATGNN are more complex models (more number of parameters) than the pre-trained CrysGNN{} framework.
Hence designing a deeper pre-training model or additionally incorporating angle-based information (ALIGNN) or attention mechanism (GATGNN) as a part of pre-training framework may help to improve further. This requires further investigation and we keep it as a scope of future work.}\\
\noindent\textbf{Comparison with Existing Pre-trained Models.} \tblak{We further demonstrate the effectiveness of the knowledge distillation method vis-a-vis the conventional fine-tuning approaches.
Note that the encoding architecture is same for CrysGNN{}, CGCNN, and CrysXPP. CrysXPP is very similar to a {\em pretrained-finetuned} version of CGCNN.
Thus we compare distilled version of CGCNN with finetuned version of CrysGNN{} and CrysXPP.
Additionally, we consider Pretrain-GNN~\cite{hu2020pretraining} which is a popular pre-training algorithm for molecules.
We pre-train all the baseline models on our curated 800K untagged crystal data and fine-tune on seven properties in JARVIS dataset and report the MAE in Table \ref{tbl-distill-finetune}.
We feed multi-graph structure of the crystal material (as discussed in ``Crystal Graph Representation") in Pretrain-GNN and try different combinations of node-level pre-training strategy along with the graph-level supervised pre-training (as suggested in ~\cite{hu2020pretraining}) and report the minimum MAE for any specific property. For finetuned CrysGNN{}, we take the pre-trained encoder of CrysGNN{} and feed a multilayer perceptron to predict a specific property. We observe that distilled CGCNN outperforms finetuned version of CrysGNN{} and both the baselines with a significant margin for all the properties. Pretrain-GNN performs the worst and the potential reason is - it is designed considering simple two-dimensional structure of molecules with a minimal set of node and bond features, which is hard to generalize for crystal materials which have very complex structure with a rich set of node and edge features.}\\
\noindent\textbf{Effectiveness on sparse training dataset.}
Finally, to demonstrate the effectiveness of the pre-training in limited data settings, we conduct additional set of experiments under different training data split. In specific, we vary available training data from 20 to 60 \%, train different SOTA models and check their performance on test dataset. We observe that the distilled version of any SOTA model consistently outperforms its vanilla version even more in the limited training data setting, which illustrates the robustness of our pre-training framework. We report the MAE values of different baselines and their distilled version in Table \ref{tbl-limited-data-small}.
\subsection{Analysis of Different Pre-training Losses}
We perform an ablation study to investigate the influence of different pre-training losses in enhancing the SOTA model performance. While pre-training CrysGNN{} (Eq. \ref{eq:pretrain_loss}), we capture both local chemical and global structural information via node and graph-level decoding, respectively.
Further, we are curious to know the influence of each of these decoding policies independently in the downstream property prediction task.
\tblak{In specific, we conduct the ablation experiments, where we pre-train CrysGNN{} with (a) only node-level decoding ($\mathcal{L}_{FR}$, $\mathcal{L}_{CR}$), (b) only graph-level decoding ($\mathcal{L}_{SG}$, $\mathcal{L}_{NTXent}$). Further, we perform ablations with individual graph-level losses, and pretrain with (c) removing $\mathcal{L}_{NTXent}$ (node-level with $\mathcal{L}_{SG}$ (space group)) and (d) removing $\mathcal{L}_{SG}$ (node-level with $\mathcal{L}_{NTXent}$(crystal system)).} We train two baseline models, CGCNN and ALIGNN, with distilled knowledge from all the aforementioned variants of the pre-trained model and evaluate the performance on four crystal properties. \\
\begin{table}
\centering
\small
\setlength{\tabcolsep}{6 pt}
\scalebox{0.8}{
\begin{tabular}{c | c c c c}
\toprule
Property & CGCNN & CrysGNN{} & CrysXPP & Pretrain \\
& (Distilled) & (Finetuned) & & -GNN\\
\midrule
Formation Energy & \textbf{0.047} & 0.056 & 0.062 & 0.764 \\
Bandgap (OPT) & \textbf{0.160} & 0.183 & 0.190 & 0.688 \\
Total Energy & \textbf{0.053} & 0.069 & 0.072 & 1.451 \\
Ehull & \textbf{0.121} & 0.130 & 0.139 & 1.112 \\
Bandgap (MBJ) & \textbf{0.340} & 0.371 & 0.378 & 1.493 \\
Bulk Modulus (Kv) & \textbf{12.31} & 13.42 & 13.61 & 20.34 \\
Shear Modulus (Gv)& \textbf{10.87} & 11.07 & 11.20 & 16.51 \\
SLME (\%) & \textbf{4.791} & 5.452 & 5.110 & 9.853 \\
Spillage & \textbf{0.354} & 0.374 & 0.363 & 0.481 \\
\bottomrule
\end{tabular}
}
\caption{Comparison of the prediction performance (MAE) of seven properties in JARVIS-DFT between CrysGNN{} and existing pretrain-finetune models,
the best performance is highlighted in bold.}
\label{tbl-distill-finetune}
\end{table}
\begin{table}
\centering
\small
\setlength{\tabcolsep}{2.5 pt}
\scalebox{0.78}{
\begin{tabular}{c|c|cccc}
\toprule
Property & Train-Val & CGCNN & CrysXPP & GATGNN & ALIGNN\\
& -Test(\%) & (Distilled) & (Distilled) & (Distilled) &(Distilled)\\
\midrule
\multirow{3}{*}{\shortstack{Bandgap \\ (MBJ)}}
& 20-10-70 & 0.453 (23.04) & 0.450 (24.82) & 0.521 (3.70) & 0.485 (2.53) \\
& 40-10-50 & 0.419 (21.41) & 0.405 (18.40) & 0.448 (2.81) & 0.395 (2.20)\\
& 60-10-30 & 0.364 (19.08) & 0.360 (17.36) & 0.439 (2.29) & 0.380 (1.98)\\
\midrule
\multirow{3}{*}{\shortstack{Bulk Modulus \\ (Kv)}}
& 20-10-70 & 16.26 (3.80) & 14.25 (7.59) & 14.19 (4.12) & 14.06 (4.35)\\
& 40-10-50 & 14.46 (2.36) & 14.02 (7.34) & 12.59 (3.00) & 12.11 (2.89)\\
& 60-10-30 & 14.05 (1.26) & 13.73 (6.98) & 11.75 (2.16) & 11.01 (1.96)\\
\midrule
\multirow{3}{*}{\shortstack{Shear Modulus \\(Gv)}}
& 20-10-70 & 12.50 (10.01) & 12.07 (9.86) & 12.42 (3.20) & 12.31 (3.15)\\
& 40-10-50 & 11.54 (4.15) & 11.01 (9.46) & 11.23 (1.75) & 10.67 (2.82)\\
& 60-10-30 & 11.31 (3.74) & 10.67 (9.35) & 10.47 (1.69) & 10.04 (1.95)\\
\midrule
\multirow{3}{*}{\shortstack{SLME (\%)}}
& 20-10-70 & 6.62 (7.17) & 5.90 (16.40) & 6.02 (5.20) & 6.27 (1.43)\\
& 40-10-50 & 5.78 (5.90) & 5.81 (15.67) & 5.63 (2.60) & 5.57 (1.42)\\
& 60-10-30 & 5.24 (5.68) & 4.84 (10.54) & 5.34 (2.55) & 4.82 (1.33)\\
\midrule
\end{tabular}
}
\caption{MAE values of distilled version of all the SOTA models for four different properties in JARVIS-DFT dataset with the increase in training instances from 20 to 60\%. Relative improvement in the distilled model is mentioned in bracket.}
\label{tbl-limited-data-small}
\end{table}
\begin{table*}[!thb]
\centering
\small
\setlength{\tabcolsep}{6pt}
\scalebox{0.75}{
\begin{tabular}{l c c c c c c c c}
\toprule
\textbf{Experiment Settings} & CGCNN & CGCNN & CrysXPP & CrysXPP & GATGNN & GATGNN & ALIGNN & ALIGNN\\
& & (Distilled) & & (Distilled) & & (Distilled) & & (Distilled)\\
\midrule
\textbf{\vtop{\hbox{\strut Train on DFT }\hbox{\strut Test on Experimental}}}
& 0.265 & 0.244 (7.60) & 0.243 & 0.225 (7.40) & 0.274 & 0.232 (15.3) & 0.220 & 0.209 (5.05) \\
\midrule
\textbf{\vtop{\hbox{\strut Train on DFT and 20 \% Experimental }\hbox{\strut Test on 80 \% Experimental}}}
& 0.144 & 0.113 (21.7) & 0.138 & 0.118 (14.2) & 0.173 & 0.168 (2.70) & 0.099 & 0.094 (5.60) \\
\midrule
\textbf{\vtop{\hbox{\strut Train on DFT and 80 \% Experimental }\hbox{\strut Test on 20 \% Experimental}}}
& 0.094 & 0.073 (22.7) & 0.087 & 0.071 (18.4) & 0.113 & 0.109 (3.40) & 0.073 & 0.069 (5.90) \\
\bottomrule
\end{tabular}
}
\caption{MAE of predicting experimental values by different SOTA models and their distilled versions with full DFT data and different percentages of experimental data for formation energy in OQMD-EXP dataset. Relative improvement in the distilled model is mentioned in bracket.}
\label{tab:fe}
\end{table*}
\begin{figure}[!thb]
\centering
\boxed{\includegraphics[width=\columnwidth]{Figure/Legend.pdf}}
\\
\vspace*{-1mm}
\subfloat[Formation Energy]{
\boxed{\includegraphics[width=0.44\columnwidth, height=24mm]{Figure/formation_energy.pdf}}}
\subfloat[Bandgap (MBJ)]{
\boxed{\includegraphics[width=0.44\columnwidth,
height=24mm]{Figure/mbj_bg.pdf}}}\\
\subfloat[Total Energy]{
\boxed{\includegraphics[width=0.44\columnwidth, height=24mm]{Figure/total_energy.pdf}}}
\subfloat[Bulk Modulus (Kv)]{
\boxed{\includegraphics[width=0.44\columnwidth, height=24mm]{Figure/bulk_modulus.pdf}}}
\caption{\tblak{Summary of experiments of ablation study on importance of different pre-training loss components on CrysGNN{} training and eventually its effect on CGCNN and ALIGNN models on four different properties (MAE for property prediction). (i) Vanilla: SOTA based model (without distillation) and all the other cases are SOTA models (distilled) from different pre-trained version of CrysGNN{}.
(ii) Node : only node-level pre-training, (iii) Graph : only graph-level pre-training, (iv) Node + L(SG) : node-level and $\mathcal{L}_{SG}$, (v) Node + L(NTXent) : node-level and $\mathcal{L}_{NTXent}$ and (vi) CrysGNN{}~: both node and graph-level pre-training.}}
\label{fig:ablation_diff_loss}
\end{figure}
Experimental results are presented in Fig.~\ref{fig:ablation_diff_loss}.
We can observe clearly that all the variants offer significant performance gain in all four properties using the combined node and graph-level pre-training, compared to node-level or graph-level pre-training separately. Only exception is formation energy, where only node-level pre-training produces less error compared to other variants, in both the baseline. Formation energy of a crystal is defined as the difference between the energy of a unit cell comprised of $N$ chemical species and the sum of the chemical potentials of all the $N$ chemical species. Hence pre-training at the node-level (node features and connection) is adequate for enhancing performance of formation energy prediction and incorporating graph-level information works as a noisy information, which degrades the performance. \tblak{We also observe improvement in performance using both supervised and contrastive graph-level losses ($\mathcal{L}_{SG}$ and $\mathcal{L}_{NTXent}$), compared to using only one of them, which proves the learned representation via supervised and contrastive learning is more expressive that using any one of them.}
Moreover, in ALIGNN, with either node or graph-level pre-training separately, performance degrades across different properties. ALIGNN explicitly captures the three body interactions which drive its performance, to replicate that inclusion of both node and graph information is necessary.
\subsection{Removal of DFT error bias using experimental data}
One of the fundamental issues in material science is that experimental data for crystal properties are very rare~\cite{kubaschewski1993materials,bracht1995properties,turns1995understanding}. Hence existing SOTA models rely on DFT calculated data to train their parameters.
However, mathematical approximations in DFT calculation lead to erroneous predictions (error bias) compared to the actual experimental values of a particular property. Hence
DFT error bias is prevalent in all SOTA models.
\cite{das2022crysxpp} has shown that pre-training helps to remove DFT error bias when fine-tuned with experimental data. Hence,
we investigate whether SOTA models can remove the DFT error with distilled knowledge from pre-trained model, using a small amount of available experimental data. In specific, we consider OQMD-EXP dataset to conduct an experiment, where we train SOTA models and their distilled variants with available DFT data and different percentages of experimental data for formation energy. We report the MAE of different SOTA models and their distilled variant in Table \ref{tab:fe}. We observe, with more amount of experimental training data, all the SOTA models are minimizing the error consistently. Moreover, with distilled knowledge from pre-trained CrysGNN{}, all SOTA models are reducing MAE further and we observe consistently larger degree of improvement with more amount of experimental training data in almost all the models.
\section{Methodology}
\label{methodology}
\begin{figure*}
\centering
\vspace*{-1mm}
\subfloat[Node-level decoding]{
\boxed{\includegraphics[width=\columnwidth, height=40mm]{Figure/crysgnn_node.pdf}}}
\subfloat[Graph-level decoding]{
\boxed{\includegraphics[width=\columnwidth, height=40mm]{Figure/crysgnn_contrast.pdf}}}
\caption{Overview of both node and graph-level decoding methods for CrysGNN{}. (a) In node-level decoding, node feature attributes and connectivity between nodes are reconstructed in a self-supervised way. (b) In graph-level decoding, $G_2$ is the pivot graph and $G_1$ is from the same crystal system (Cubic), whereas $G_3, G_4, G_5$ are from different crystal systems. First we reconstruct space group information of $G_2$, then through contrastive loss, CrysGNN{} will maximize similarities between positive pair ($G_2,G_1$) and minimize similarities between negative pairs ($G_2, G_3$), ($G_2, G_4$) and ($G_2, G_5$) in embedding space.
\label{fig:CrysGNN}
\end{figure*}
Formally, we first curate a huge amount of property un-tagged crystal graphs $\mathcal{D}_{u} = \{\mathcal{G}_i\}$ from various materials datasets to pre-train a deep GNN model $f_{\theta}$, that learns intrinsic structural and chemical patterns of the crystal graphs.
Further, we use a training set of property tagged crystal graphs $\mathcal{D}_{t} =\{\mathcal{G}_i,y_i\}$ for property prediction, which is smaller in volume and \tblak{may or may not be disjoint from the original untagged set} $\mathcal{D}_{u}$. We train any supervised property predictor $\mathcal{P}_{\psi}$ using $\mathcal{D}_{t}$ to predict the property value given the crystal graph structure. While training the property predictor, we incorporate the idea of knowledge distillation to distill important structural and chemical information from the pre-trained model. This knowledge may prove to be useful to a property predictor which now need not learn from scratch, but be armed with distilled knowledge from the pre-trained model. Hence in this section, we first describe the CrysGNN{} pre-training strategy, followed by the knowledge distillation and property prediction process.
\subsection{CrysGNN{} Pre-training}
\label{pretraining}
We build a deep auto-encoder architecture CrysGNN{}, which comprises a graph convolution based encoder followed by an effective decoder. The autoencoder is (pre)trained end to end, using a large amount of property un-tagged crystal graphs $\mathcal{D}_{u} = \{\mathcal{G}_i\}$, where via node and graph-level self-supervised losses, the model can capture the structural and chemical information of the crystal graph data. First, we formalize the representation of a crystal 3D structure into a multi-graph structure, which will be an input to the encoder module.
\subsubsection{Crystal Graph Representation.}
\label{graph_rep}
We realize a crystal material as a multi-graph structure $\mathcal{G}_i =(\mathcal{V}_i, \mathcal{E}_i, \mathcal{X}_i, \mathcal{F}_i )$ as proposed in \cite{xie2018crystal}. $\mathcal{G}_i$ is an undirected weighted multi-graph where $\mathcal{V}_i$ denotes the set of nodes or atoms present in a unit cell of the crystal structure. $\mathcal{E}_i=\{(u,v,k_{uv})\}$ denotes a multi-set of node pairs and $k_{uv}$ denotes number of edges between a node pair $(u,v)$. $\mathcal{X}_i=\{(x_{u} | u \in \mathcal{V}_i )\}$ denotes the node feature set proposed by CGCNN~\cite{xie2018crystal}. It includes different chemical properties like electronegativity, valance electron, covalent radius, etc. Finally, $\mathcal{F}_i=\{\{s^k\}_{(u,v)} | (u,v) \in \mathcal{E}_i, k\in\{1..k_{uv}\}\}$ denotes the multi-set of edge weights where $s^k$ corresponds to the $k^{th}$ bond length between a node pair $(u,v)$, which signifies the inter-atomic bond distance between two atoms. Next, we formally define CrysGNN{} pre-training and knowledge distillation based property prediction strategy.
\begin{figure*}
\centering
\vspace*{-1mm}
\boxed{\includegraphics[width=1.3\columnwidth]{Figure/distillation_property_prediction.pdf}}
\caption{Overview of Property Prediction using Knowledge Distillation from CrysGNN{}.}
\label{fig:dist_prop}
\end{figure*}
\subsubsection{Self Supervision.}
\label{self_sup}
We first develop a graph convolution \cite{xie2018crystal} based encoding module, which takes crystal multi-graph structure $\mathcal{G} =(\mathcal{V}, \mathcal{E}, \mathcal{X}, \mathcal{F} )$ as input and encodes structural semantics of the crystal graph into lower dimensional space. Each layer of convolution, follows an iterative neighbourhood aggregation (or message passing) scheme to capture the structural information within node's (atom's) neighbourhood. After $L$-layers of such aggregation, the encoder returns the final set of node embeddings $\mathcal{Z} = \{z_1,..., z_{|\mathcal{V}|}\}$, where $z_u :=z_u^L$ represents the final embedding of node $u$. Details of the GNN architecture is in Appendix.
Next, we design an effective decoding module, which takes node embeddings $\mathcal{Z}$ as input and learns local chemical features and global structural information through node and graph-level decoding, respectively. Decoding node-level information will enable CrysGNN{} to learn local domain specific chemical features and connectivity information around an atom, while decoding graph-level features will help CrysGNN{} capture global structural knowledge. \\\\
\textbf{Node-Level Decoding.} For node-level decoding (Fig-\ref{fig:CrysGNN}(a)), we propose two self-supervised learning methods, where we reconstruct two important features that induce the local chemical environment of the crystal around a node (atom). For a given node $u$, we first reconstruct its node features $x_u$,
which represent different chemical properties of atom $u$. Given a node embedding $z_u$, which is encoded based on neighbouring structure around atom $u$, we apply a linear transformation on top of $z_u$ to reconstruct the node attributes. In crystalline graphs, node features correspond to different chemical properties associated with the constituent atoms
through reconstructing these features CrysGNN{} captures local chemical semantics around that atom.\\
Further, we reconstruct local connectivity around an atom, where given node embeddings of two nodes $u$ and $v$, we apply a bi-linear transformation module to generate combined transformed embedding of two nodes $z_{uv}$, which we pass through a feed forward network to predict the strength of association between two atoms. Through reconstructing this local connectivity around an atom, CrysGNN{} encodes the periodicity of the node i.e. the number of neighbours around it along with the relative position of its neighbours and their bond length.
\\\\
\textbf{Graph-level Decoding.}
We aim to capture periodic structure of a crystal material through graph-level decoding (Fig-\ref{fig:CrysGNN}(b)). We specifically leverage two concepts in doing so. (a). {\bf Space group} which is used to describe the symmetry of a unit cell of the crystal material. In materials science literature there are 230 unique space groups and each crystal (graph) has a unique space group number.
(b). {\bf Crystal system.} The space group level information can classify a crystal graph into 7 broad groups of crystal systems like Triclinic, Monoclinic, Orthorhombic, Tetragonal, Trigonal, Hexagonal, and Cubic. Several electronic and optical properties such as band gap, dielectric constant depend on the space-group and the crystal structure justifying its usage. More information about different space groups and crystal systems is in Appendix
\tblak{Given the set of node embeddings $\mathcal{Z} = \{z_1,..., z_{|\mathcal{V}|}\}$, we use a symmetric aggregation function to generate graph-level representation $\mathcal{Z}_{\mathcal{G}}$. First, we pass $\mathcal{Z}_{\mathcal{G}}$ through a feed-forward neural network to predict the space group number of graph $\mathcal{G}$.
Further, we develop a contrastive learning framework for pre-training of CrysGNN{}, where pre-training is performed by maximizing (minimizing) similarity between two crystal graphs belonging to the same (different) crystal system via contrastive loss in graph embedding space.}
A mini-batch of $N$ crystal graphs is randomly sampled and processed through contrastive learning to
align the positive pairs $\mathcal{Z}_{\mathcal{G}_i},\mathcal{Z}_{\mathcal{G}_j}$ of graph embeddings, which belong to the same crystal system and contrast the negative pairs which are from different crystal systems. Here we adopt the
normalized temperature-scaled cross-entropy loss (NT-Xent)\cite{sohn2016improved,van2018representation,wu2018unsupervised} and NT-Xent for the $i^{th}$ graph is defined:
\begin{equation}
\label{eq:contrastive}
\mathcal{L}_{i}=-log\frac{exp(sim(\mathcal{Z}_{\mathcal{G}_i},\mathcal{Z}_{\mathcal{G}_j})/\tau)}{\sum_{k=1}^{K} exp(sim(\mathcal{Z}_{\mathcal{G}_i},\mathcal{Z}_{\mathcal{G}_k})/\tau)}
\end{equation}
where $\tau$ denotes the temperature parameter and $sim(\mathcal{Z}_{\mathcal{G}_i},\mathcal{Z}_{\mathcal{G}_j})$ denotes cosine similarity function.
\tblak{The final loss $\mathcal{L}_{NTXent}$ is computed across all positive pairs in the minibatch.}
Overall we pre-train this deep auto-encoder architecture CrysGNN{} end to end to optimize the following loss :
\begin{equation}
\label{eq:pretrain_loss}
\mathcal{L}_{pretrain}=\alpha\mathcal{L}_{FR}+\beta\mathcal{L}_{CR}+\gamma\mathcal{L}_{SG}+\lambda\mathcal{L}_{NTXent}
\end{equation}
where $\mathcal{L}_{FR},\mathcal{L}_{CR}$ are the reconstruction losses for node feature, and local connectivity, $,\mathcal{L}_{SG}$ is the space group supervision loss, $\mathcal{L}_{NTXent}$ is the contrastive loss and $\alpha$, $\beta$, $\gamma$, $\lambda$ are the weighting coefficients of each loss. We denote the set of parameters in CrysGNN{} model as $\theta$ and the pre-trained CrysGNN{} as $f_{\theta}$.
\subsection{Distillation and Property Prediction}
We aim to retrofit the pre-trained CrysGNN{} model into any SOTA property predictor to enhance its learning process and improve performance (Fig-\ref{fig:dist_prop}). Hence we incorporate the idea of knowledge distillation to distill important structural and chemical information from the pre-trained model, which is useful for the downstream property prediction task, and feed it into the property prediction process.
Formally, given the pre-trained CrysGNN{} model $f_{\theta}$, any SOTA property predictor $\mathcal{P}_{\psi}$ and set of property tagged training data $\mathcal{D}_{t} =\{\mathcal{G}_i,y_i\}$, we aim to find optimal parameter values ${\psi^*}$ for $\mathcal{P}$. We train $\mathcal{P}_{\psi}$ using dataset $\mathcal{D}_{t}$ to optimize the following multitask loss:
\begin{equation}
\label{eq:finetune_loss}
\mathcal{L}_{prop}=\delta\mathcal{L}_{MSE}+(1-\delta)\mathcal{L}_{KD}
\end{equation}
where $\mathcal{L}_{MSE} = (\hat{y_i}-y_i)^2$ denotes the discrepancy between predicted and true property values by $\mathcal{P}_{\psi}$ (property prediction loss). We define knowledge distillation loss $\mathcal{L}_{KD}$ to match intermediate node feature representation between the pre-trained CrysGNN{} model and the SOTA property predictor $\mathcal{P}_{\psi}$ as follows:
\begin{equation}
\label{eq:kd_loss}
\mathcal{L}_{KD}= \lVert {\mathcal{Z}^{T}_{i}} - {\mathcal{Z}^{S}_{i}} \rVert^2
\end{equation}
where $\mathcal{Z}^{T}_{i}$ and $\mathcal{Z}^{S}_{i}$ denote intermediate node embeddings of the pre-trained CrysGNN{} and the property predictor $\mathcal{P}_{\psi}$ for crystal graph $\mathcal{G}_i$, respectively. \tblak{Note, both $\mathcal{Z}^{T}_{i}$ and $\mathcal{Z}^{S}_{i}$ are projected on the same latent space.} Finally, $\delta$ signifies relative weightage between two losses, which is a hyper-parameter to be tuned on validation data. During property prediction the pre-trained network is frozen and we backpropagate $\mathcal{L}_{prop}$ through the predictor $\mathcal{P}_{\psi}$ end to end.
\begin{table}
\centering
\small
\setlength{\tabcolsep}{2 pt}
\scalebox{0.68}{
\begin{tabular}{c | c | c | c | c | c }
\toprule
Task & Datasets & Graph Num. & Structural Info. & Properties Count & Data Type\\
\midrule
\multirow{2}{*}{\shortstack{Pre-training}}
& OQMD & 661K & \checkmark & x & DFT Calculated\\
& MP & 139K & \checkmark & x & DFT Calculated\\
\midrule
\multirow{3}{*}{\shortstack{Property \\ Prediction}} & MP 2018.6.1 & 69K & \checkmark & 2 & DFT Calculated\\
& JARVIS-DFT & 55K & \checkmark & 19 & DFT Calculated\\
& OQMD-EXP & 1.5K & \checkmark & 1 & Experimental\\
\bottomrule
\end{tabular}
}
\caption{Datasets Details}
\label{tbl-dataset}
\end{table}
\begin{table*}
\centering
\small
\setlength{\tabcolsep}{11 pt}
\scalebox{0.9}{
\begin{tabular}{c | c c | c c| c c | c c}
\toprule
Property & CGCNN & CGCNN & CrysXPP & CrysXPP & GATGNN & GATGNN & ALIGNN & ALIGNN \\
& & (Distilled) & & (Distilled) & & (Distilled) & & (Distilled)\\
\midrule
Formation Energy & 0.039 & \textbf{0.032} & 0.041 & \textbf{0.035} & 0.096 & \textbf{0.091} & 0.026 & \textbf{0.024} \\
Bandgap (OPT) & 0.388 & \textbf{0.293} & 0.347 & \textbf{0.287} & 0.427 & \textbf{0.403} & 0.271 & \textbf{0.253} \\
\midrule
Formation Energy & 0.063 & \textbf{0.047} & 0.062 & \textbf{0.048} & 0.132 & \textbf{0.117} & 0.036 & \textbf{0.035} \\
Bandgap (OPT) & 0.200 & \textbf{0.160} & 0.190 & \textbf{0.176} & 0.275 & \textbf{0.235} & 0.148 & \textbf{0.131} \\
Total Energy & 0.078 & \textbf{0.053} & 0.072 & \textbf{0.055} & 0.194 & \textbf{0.137} & 0.039 & \textbf{0.038} \\
Ehull & 0.170 & \textbf{0.121} & 0.139 & \textbf{0.114} & 0.241 & \textbf{0.203} & 0.091 & \textbf{0.083} \\
Bandgap (MBJ) & 0.410 & \textbf{0.340} & 0.378 & \textbf{0.350} & 0.395 & \textbf{0.386} & 0.331 & \textbf{0.325} \\
Spillage & 0.386 & \textbf{0.374} & 0.363 & \textbf{0.357} & 0.350 & \textbf{0.348} & 0.358 & \textbf{0.356} \\
SLME (\%) & 5.040 & \textbf{4.790} & 5.110 & \textbf{4.630} & 5.050 & \textbf{4.950} & 4.650 & \textbf{4.590} \\
Bulk Modulus (Kv) & 12.45 & \textbf{12.31} & 13.61 & \textbf{12.70} & 11.64 & \textbf{11.53} & 11.20 & \textbf{10.99} \\
Shear Modulus (Gv) & 11.24 & \textbf{10.87} & 11.20 & \textbf{10.56} & 10.41 & \textbf{10.35} & 9.860 &\textbf{9.800} \\
\bottomrule
\end{tabular}
}
\caption{Summary of the prediction performance (MAE) of different properties in Materials project (Top) and JARVIS-DFT (Bottom). Model M is the vanilla variant of a SOTA model and M (Distilled) is the distilled variant using the pretrained CrysGNN{}. The best performance is highlighted in bold.}
\label{tbl-full-data-1}
\end{table*}
\section{Related Work}
\label{related_work}
In recent times, data driven approaches~\cite{seko2015prediction,pilania2015structure,lee2016prediction,de2016statistical,seko2017representation,isayev2017universal,ward2017including,lu2018accelerated,im2019identifying} have become quite popular to establish relationship between the atomic structure of crystalline materials and their properties with very high precision. Especially, graph neural network (GNN) based approaches ~\cite{xie2018crystal,chen2019graph,louis2020graph,Wolverton2020,schmidt2021crystal,choudhary2021atomistic} have emerged as a powerful machine learning model tool to encode material’s complex topological structure along with node features in an enriched representation space.\\
There are attempts to pre-train GNNs to extract graph and node-level representations.~\cite{hu2020pretraining} develops an effective pre-training strategy for GNNs, where they perform both node-level and graph-level pre-training on GNNs to capture domain specific knowledge about nodes and edges, in addition to global graph-level knowledge. Followed by this work, there has been several other works on self-supervised graph pre-training ~\cite{hu2020gpt,qiu2020gcc,you2020graph}, which propose different graph augmentation methods and maximizes the agreement between two augmented views of the same graph via a contrastive loss. In the field of crystal graphs, CrysXPP~\cite{das2022crysxpp} is the only model which comes close to a pre-trained model. In their work, an autoencoder is trained on a volume of un-tagged crystal graphs and the learned knowledge is (transferred to) used to initialize the encoder of CrysXPP, which is fine-tuned with property specific tagged data. \\
\tblak{Although conceptually similar to the work done by ~\citet{hu2020pretraining}, our work differs in the following three key aspects: (1) pre-training strategy proposed by ~\citeauthor{hu2020pretraining} is very effective for molecular dataset, but it is difficult to extend directly to crystalline material because structural semantics are different between molecules and materials \cite{xie2021crystal}.
Molecules have non-periodic and finite structures, solid materials’ structures are infinite and periodic in nature.
(2) For graph-level pre-training, ~\citeauthor{hu2020pretraining} adapted supervised graph-level property prediction using a huge amount of labelled dataset from chemistry and biology domain, which makes it less effective in several other domains like material science where property labeled data is extremely scarce. Also, a crucial step in graph-level prediction is to find graph structural similarity between two sets of graphs, which they do not explore but mention as a future work. We do not make use of supervised pre-training which requires a large amount of property tagged material data. Instead, we leverage the idea of structural similarity of materials belonging to the similar space group, and via contrastive loss and space group classification loss, we try to capture this similarity. (3) Finally they follow conventional pre-train finetuning framework, whereas in CrysGNN{} we incorporate the idea of knowledge distillation \cite{romero2014fitnets,hinton2015distilling} to distill important information from the pre-trained model and inject it into the property prediction process. By design, this knowledge distillation based approach is more robust and independent of the underlying architecture of the property predictor, thus it can enhance the performance of a diverse set of SOTA models.}
\section{Introduction}
\label{intro}
Fast and accurate prediction of different material properties is a challenging and important task in material science. In recent times there has been an ample amount of data driven works ~\cite{seko2015prediction,pilania2015structure,lee2016prediction,de2016statistical,seko2017representation,isayev2017universal,ward2017including,lu2018accelerated,im2019identifying}
for predicting crystal properties which are as accurate as theoretical DFT [Density functional Theory] based approaches~\cite{orio2009density}, however, much faster than it. The architectural innovations of these approaches towards accurate property predictions come from incorporating specific domain knowledge into a deep encoding module. For example, in order to encode the neighbourhood structural information around a node (atom), GNN based approaches ~\cite{xie2018crystal,chen2019graph,louis2020graph,Wolverton2020,schmidt2021crystal} gained some popularity in this domain. Understanding the importance of many-body interactions, ALIGNN~\cite{choudhary2021atomistic} incorporates bond angular information into their encoder module and became SOTA for a large range of property predictions. However, as different properties expressed by a crystalline material are a complex function of different inherent structural and chemical properties of the constituent atoms, it is extremely difficult to explicitly incorporate them into the encoder architecture. Moreover, data sparsity across properties is a known issue~\cite{das2022crysxpp,jha2019enhancing}, which makes these models difficult to train for all the properties. To circumvent this problem we adopt the concept of self-supervised pre-training \cite{devlin2018bert,trinh2019selfie,chen2020simple,chen2020adversarial,he2020momentum,hu2020pretraining,hu2020gpt,qiu2020gcc,you2020graph} for crystalline materials which enables us to leverage a large amount of untagged material structures to learn the complex hidden features which otherwise are difficult to identify.\\
In this paper, we introduce a graph pre-training method which captures (a) connectivity of different atoms, (b) different atomic properties and (c) graph similarity from a large set of unlabeled data. To this effect, we curate a new large untagged crystal dataset with 800K crystal graphs and undertake a pre-training framework (named CrysGNN{}) with the dataset. CrysGNN{} learns the representation of a crystal graph by initiating self-supervised loss at both node (atom) and graph (crystal) level. At the node level, we pre-train the GNN model to reconstruct the node features and connectivity between nodes in a self-supervised way, whereas at the graph level, we adopt supervised and contrastive learning to learn structural similarities between graph structures using the space group and crystal system information of the crystal materials respectively.\\
We subsequently distill important structural and chemical information of a crystal from the pre-trained CrysGNN{} model and pass it to the property predictor.
The distillation process provides wider usage than the conventional pretrain-finetuning framework as transferring pre-trained knowledge to a property predictor and finetuning it requires a similar graph encoder architecture between the pre-trained model and the property predictor, which limits the knowledge transfer capability of the pre-trained model. On the other hand, using knowledge distillation \cite{romero2014fitnets,hinton2015distilling}, we can retrofit the pre-trained CrysGNN{} model into any existing state-of-the-art property predictor, irrespective of their architectural design, to improve their property prediction performance. Also experimental results (presented later) show that even in case of similar graph encoder, distillation performs better than finetuning.\\
With rigorous experimentation
across two popular benchmark materials datasets, we show that distilling necessary information from CrysGNN{}
\footnote{Source code, pre-trained model, and dataset of CrysGNN is made available at https://github.com/kdmsit/crysgnn}
to various property predictors results in substantial performance gains for GNN based architectures and complex ALIGNN model. The improvements range from 4.19\% to 16.20\% over several highly optimized SOTA models. We also perform ablation studies to investigate the influence of different pre-training losses in enhancing the SOTA model performance and observe significant performance gain employing the both node and graph-level pre-training, compared to node-level or graph-level pre-training separately. Also using both supervised and contrastive graph-level pre-training, we are able to learn more robust and expressive graph representation which enhances the property predictor performance.
This also helps to achieve even better improvements when the dataset is sparse. Moreover, the property-tagged dataset suffers from certain biases as it is theoretically (DFT) derived, hence the property predictor also suffers from such bias. We found that on being trained with small amount of experimental data, the DFT bias decreases substantially.
\section{Related Work}
\subsection{Materials Property Prediction.}
In recent times, data-driven approaches have become quite popular to establish relationships between the atomic structure of crystalline materials and their properties with very high precision. Especially, graph neural network (GNN) based approaches have emerged as a powerful machine learning model tool to encode material’s complex topological structure along with node features in an enriched representation space. Models such as CGCNN~\cite{xie2018crystal} represent 3D crystal structure as a multi-graph and build a graph convolution neural network directly on the graph to update node features based on their local chemical and structural environment. GATGNN~\cite{louis2020graph} incorporates the idea of graph attention network to learn the importance of different inter atomic bonds whereas MEGNet~\cite{chen2019graph} introduces global state attributes for quantitative structure-state property relationship prediction in materials. While all these models implicitly represent many body interactions through multiple graph convolution layers, ALIGNN~\cite{choudhary2021atomistic} explicitly captures many body interactions by incorporating bond angles and local geometric distortions.
\subsection{Graph Pre-training Strategies}
There are attempts to pre-train GNNs to extract graph and node level representations. ~\cite{hu2020pretraining} develops an effective pre-training strategy for GNNs, where they perform both node level and graph level pre-training on GNNs to capture domain specific knowledge about nodes and edges, in addition to global graph-level knowledge. GPT-GNN~\cite{hu2020gpt} proposes a self-supervised attributed graph generation task to pre-train a GNN model, which captures structural and semantic properties of the graph. ~\cite{qiu2020gcc} presents GCC, which leverages the idea of contrastive learning to design the graph pre-training task as subgraph instance discrimination, to capture universal network topological properties across multiple networks. Another recent work is GraphCL~\cite{you2020graph}, which proposes different graph augmentation methods and maximises the agreement between two augmented views of the same graph via a contrastive loss.
In the field of crystal graphs, CrysXPP~\cite{das2022crysxpp} is the only model which comes close to a pre-trained model. In their work, an autoencoder is trained on a volume of un-tagged crystal graphs and the learned knowledge is (transferred to) used to initialize the encoder of CrysXPP, which is fine-tuned with property specific tagged data. \\
To the best of our knowledge, CrysGNN{} is the first attempt to develop a deep pre-trained model for the crystalline materials and for that, we curated a new large untagged crystal dataset with 800K crystal graphs.
\section{Crystal Representation}
We realize a crystal material as a multi-graph structure $\mathcal{G}_i =(\mathcal{V}_i, \mathcal{E}_i, \mathcal{X}_i, \mathcal{F}_i )$ as proposed in \cite{das2022crysxpp,xie2018crystal}. $\mathcal{G}_i$ is an undirected weighted multi-graph where $\mathcal{V}_i$ denotes the set of nodes or atoms present in a unit cell of the crystal structure. $\mathcal{E}_i=\{(u,v,k_{uv})\}$ denotes a multi-set of node pairs and $k_{uv}$ denotes number of edges between a node pair $(u,v)$. $\mathcal{X}_i=\{(x_{u} | u \in \mathcal{V}_i )\}$ denotes 92 dimensional node feature set proposed by CGCNN~\cite{xie2018crystal}. It includes different chemical properties like electronegativity, valance electron, covalent radius, etc. Details of these node features are given in Table \ref{tab:feature_desc}. Finally, $\mathcal{F}_i=\{\{s^k\}_{(u,v)} | (u,v) \in \mathcal{E}_i, k\in\{1..k_{uv}\}\}$ denotes the multi-set of edge weights where $s^k$ corresponds to the $k^{th}$ bond length between a node pair $(u,v)$, which signifies the inter-atomic bond distance between two atoms.
\begin{figure}[h]
\centering
\vspace*{-1mm}
\boxed{\includegraphics[width=\columnwidth, height=45mm]{Figure/crystal_graph_representation.pdf}}
\caption{Multi-graph Representation from Crystal 3D Structure.}
\label{fig:3d_crsytal_rep}
\end{figure}
\begin{table}[]
\centering
\small
\setlength{\tabcolsep}{4pt}
\scalebox{0.9}{
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Features} & \textbf{Unit} &\textbf{Range} & \textbf{Feature Dimension} \\
\hline
Group Number & - & 1,2, ..., 18 & 18 \\
\hline
Period Number & - & 1,2, ..., 9 & 9\\
\hline
Electronegativity & - & 0.5-4.0 & 10 \\
\hline
Covalent Radius & pm & 25-250 & 10 \\
\hline
Valence Electrons & - &1,2, ..., 12 & 12\\
\hline
First Ionization Energy & eV & 1.3–3.3 & 10\\
\hline
Electron Affinity & eV & -3–3.7 & 10 \\
\hline
Block & - & s, p, d, f & 4 \\
\hline
Atomic Volume & $cm^{3} /mol$ & 1.5–4.3 & 10 \\
\hline
\end{tabular}
}
\caption{Description of different atomic features $\mathcal{X}_i$ and their unit, range and dimensions. All of them together forms 92 dimensional node feature vector $\mathcal{X}_i$.}
\centering
\label{tab:feature_desc}
\end{table}
\section {Details of GNN Architecture in CrysGNN{}}
Here we describe GNN architectures used in CrysGNN for pre-training and property prediction model.
\subsubsection{Encoder.} We develop an graph convolution \cite{xie2018crystal} based encoding module, which takes crystal multi-graph structure $\mathcal{G} =(\mathcal{V}, \mathcal{E}, \mathcal{X}, \mathcal{F} )$ as input and encode structural semantics of the crystal graph into lower dimensional space. Considering $L$ graph convolution layers, neighbourhood aggregation of the $l$-th layer is represented as :
\begin{equation}
\label{eq:cgcnn}
z_u^{(l+1)} = z_u^{(l)} + \sum\limits_{v,k} \sigma ( {h^{(l)}_{(u,v)}}_k W_c^{(l)} + b_c^{(l)}) \odot g({h^{(l)}_{(u,v)}}_k W_s^{(l)}+b_s^{(l)})
\end{equation}
where $h_{(u,v)_k}^{(l)} = z_u^{(l)} \oplus z_v^{(l)} \oplus {s^{(k)}_{(u,v)}}$. Here $s^{(k)}_{(u,v)}$ represents inter-atomic bond length of the $k$-th bond between $(u,v)$ and $z_u^{(l)}$ denotes embedding of node $u$ after $l$-th layer, which is initialized to a transformed node feature vector $z_u^0 := x_u W_x$ where $W_x$ is the trainable parameter and $ x_u $ is the input node feature vector (Table. \ref{tab:feature_desc}). In Equation \ref{eq:cgcnn}, $W_c^{(l)},W_s^{(l)},b_c^{(l)},b_s^{(l)}$ are the convolution weight matrix, self weight matrix, convolution bias, and self bias of the $l$-th layer, respectively, $\oplus$ operator denotes concatenation, $\odot$ denotes element-wise multiplication, $\sigma$ is the sigmoid function indicating the edge importance and $g$ is a feed forward network. After $L$-layers of such propagation, $\mathcal{Z} = \{z_1,..., z_{|\mathcal{V}|}\}$ represents the set of final node embeddings, where $z_u :=z_u^L$ represents final embedding of node $u$.
\subsubsection{Decoder.} In the decoding module, we apply both node-level and graph-level self-supervised pre-training as discussed in \textbf{Node Level Decoding} and \textbf{Graph Level Decoding} sections in the main manuscript.
\subsubsection{Property Predictor.} For the property prediction model, we use four diverse sets of SOTA models, which have different encoder choices and on top of that they have a regressor module to predict the property value. We retrofit CrysGNN{} model with each of these SOTA models using the modified multitask loss (incorporating knowledge distillation loss with their property loss) as discussed in Equation \ref{eq:finetune_loss} (main manuscript).
\section{Crystal System and Space group Information}
By definition crystal is a periodic arrangement of repeating “motifs”( e.g. atoms, ions). The symmetry of a periodic pattern of repeated motifs is the total set of symmetry operations allowed by that pattern. The total set of such symmetry operations, applicable to the pattern is the pattern's symmetry, and is mathematically described by a so-called Space Group. The Space Group of a Crystal describes the symmetry of that crystal,
and as such it describes an important aspect of that crystal's internal structure.\\
In turn crystals across certain space groups show
similarities among each other, hence they are divided among seven Crystal Systems : Triclinic, Monoclinic, Orthorhombic, Tetragonal, Trigonal, Hexagonal, and Cubic or Isometric system ~\cite{kittel2018kittel}.\\
\textbf{Crystallographic Axes: } The crystallographic axes are imaginary lines that we can draw within the crystal lattice. These will define a coordinate system within the crystal. We refer to the axes as \textbf{a}, \textbf{b}, \textbf{c} and angle between them as $\alpha$, $\beta$ and $\gamma$. By using these crystallographic axes we can define six large groups or crystal systems.
\begin{itemize}
\item \textbf{Cubic or Isometric system :} The three crystallographic axes are all equal in length and intersect at right angles to each other.
\begin{equation}
\boxed{a=b=c\;; \;\; \alpha = \beta =\gamma = 90 \degree} \notag
\end{equation}
\item \textbf{Tetragonal system :} Three axes, all at right angles, two of which are equal in length (a and b)
and one (c) which is different in length (shorter or longer).
\begin{equation}
\boxed{a=b\neq c\;; \;\; \alpha = \beta =\gamma = 90 \degree} \notag
\end{equation}
\item \textbf{Orthorhombic system :} Three axes, all at right angles, all three have different length.
\begin{equation}
\boxed{a\neq b\neq c\;; \;\; \alpha = \beta =\gamma = 90 \degree} \notag
\end{equation}
\item \textbf{Monoclinic system :} Three axes, all unequal in length, two of which (a and c) intersect at an oblique angle (not 90 degrees), the third axis (b) is perpendicular to the other two axes.
\begin{equation}
\boxed{a\neq b\neq c\;; \;\; \alpha =\gamma = 90 \degree \neq \beta} \notag
\end{equation}
\item \textbf{Triclinic system :} The three axes are all unequal in length and intersect at three different angles (any angle but 90 degrees).
\begin{equation}
\boxed{a\neq b\neq c\;; \;\; \alpha \neq \beta \neq \gamma \neq 90 \degree} \notag
\end{equation}
\item \textbf{Hexagonal system :} Here we have four axes. Three of the axes fall in the same plane and intersect at the axial cross at 120°. These 3 axes, labeled $a_1$, $a_2$, and $a_3$, are the same length. The fourth axis, c, may be of {\em a} different length than the axes set. The c axis also passes through the intersection of the {\em a} axes
set at right angle to the plane formed by the {\em a} set.
\end{itemize}
\section{Details about datasets and different crystal properties.}
\textbf{OQMD:} The Open Quantum Materials Database (OQMD) is a high-throughput database based on DFT calculations of chemicals from the Inorganic Crystal Structure Database (ICSD) and embellishments of regularly occurring crystal structures \cite{OQMD}. Comprehensive computations for significant structure types in materials research, such as Heusler and perovskite compounds, are among the distinguishing aspects of this database. The database is public and a free resource.\\\\
\textbf{Materials project:} This is another free and public resource database that contains crystal structures and calculated materials properties~\cite{MP}. The dataset is made from the results obtained with density functional theory based calculations. The dataset is composed of electronic structure, thermodynamic, mechanical, and dielectric properties. It also provides a visual web-based search interface.\\\\
\textbf{JARVIS:} JARVIS (Joint Automated Repository for Various Integrated Simulations) is a data repository that incorporates not only DFT-based calculations, but also data from classical force-fields and machine learning approaches~\cite{JARVIS}. The databases are free and public as well.\\\\
We curated around {800K untagged crystal graph data from two popular materials databases, Materials Project (MP) and OQMD}, to pre-train CrysGNN{} model. Further to evaluate the performance of different SOTA models with distilled knowledge from CrysGNN{}, we select {MP 2018.6.1 version of Materials Project} and {2021.8.18 version of JARVIS-DFT}, another popular materials database, for property prediction as suggested by \cite{choudhary2021atomistic}. Please note, MP 2018.6.1 dataset is a subset of the dataset used for pre-training, whereas JARVIS-DFT is a separate dataset which is not seen during the pre-training.
MP 2018.6.1 consists of 69,239 materials with two properties namely bandgap and formation energy, whereas JARVIS-DFT consists of 55,722 materials with 19 properties which can be broadly classified into two categories : 1) properties like formation energy, bandgap, total energy, bulk modulus, etc. which depend greatly on crystal structures and atom features, and 2) properties like $\epsilon_x$, $\epsilon_y$, $\epsilon_z$, n-Seebeck, n-PF, etc. which depend on the precise description of the materials’ electronic structure. Details of these properties is provided in Table \ref{tbl-prop} \\
Moreover, all these properties in both Materials Project and JARVIS-DFT datasets are based on DFT calculations of chemicals. Therefore, to investigate how pre-trained knowledge helps to mitigate the DFT error, we also take a small dataset ~\cite[OQMD-EXP]{kirklin2015open} containing 1,500 available experimental data of formation energy. Details of each of these datasets are given in Table \ref{tbl-dataset}.
\begin{table}
\centering
\small
\setlength{\tabcolsep}{10 pt}
\begin{tabular}{ccc}
\toprule
Property & Unit & Data-size\\
\midrule
Formation\_Energy & $eV/(atom)$ & 69239 \\
Bandgap (OPT) & $eV$ & 69239 \\
\midrule
Formation\_Energy & $eV/(atom)$ & 55723 \\
Bandgap (OPT) & $eV$ & 55723 \\
Total\_Energy & $eV/(atom)$ & 55723\\
Ehull & $eV$ & 55371\\
Bandgap (MBJ) & $eV$ & 18172 \\
Bulk Modulus (Kv) & GPa & 19680\\
Shear Modulus (Gv) & GPa & 19680\\
SLME (\%) & No unit & 9068\\
Spillage & No unit & 11377 \\
$\epsilon_x$ (OPT) & No unit & 44491 \\
$\epsilon_y$ (OPT) & No unit & 44491\\
$\epsilon_z$ (OPT) & No unit & 44491 \\
$\epsilon_x$ (MBJ) & No unit & 16814\\
$\epsilon_y$ (MBJ) & No unit & 16814\\
$\epsilon_z$ (MBJ) & No unit & 16814 \\
n-Seebeck & $\mu VK^{-1}$ & 23210\\
n-PF & $\mu W(mK^2)^{-1}$ & 23210\\
p-Seebeck & $\mu VK^{-1}$ & 23210\\
p-PF & $\mu W(mK^2)^{-1}$ & 23210\\
\bottomrule
\end{tabular}
\caption{Summary of different crystal properties in Materials Project (Top) and JARVIS-DFT (Bottom) datasets.}
\label{tbl-prop}
\end{table}
\begin{table*}
\centering
\small
\setlength{\tabcolsep}{11 pt}
\scalebox{0.9}{
\begin{tabular}{c | c c | c c| c c | c c}
\toprule
Property & CGCNN & CGCNN & CrysXPP & CrysXPP & GATGNN & GATGNN & ALIGNN & ALIGNN\\
& & (Distilled) & & (Distilled) & & (Distilled) & & (Distilled)\\
\midrule
$\epsilon_x$ (OPT) & 24.93 & \textbf{23.86} & 24.47 & \textbf{23.33} & 26.79 & \textbf{25.23} & 21.42 & \textbf{21.37}\\
$\epsilon_y$ (OPT) & 25.06 & \textbf{24.08} & 24.84 & \textbf{23.98} & 26.77 & \textbf{25.02} & 21.66 & \textbf{21.25} \\
$\epsilon_z$ (OPT) & 24.99 & \textbf{23.85} & 24.13 & \textbf{23.76} & 26.09 & \textbf{24.81} & \textbf{20.51} & 21.03 \\
$\epsilon_x$ (MBJ) & 27.18 & \textbf{26.64} & 28.40 & \textbf{26.03} & \textbf{27.56} & 27.61 & 24.76 & \textbf{24.32}\\
$\epsilon_y$ (MBJ) & 27.14 & \textbf{25.80} & 26.92 & \textbf{25.34} & 27.21 & \textbf{26.89} & 23.99 & \textbf{23.83}\\
$\epsilon_z$ (MBJ) & 28.38 & \textbf{26.08} & 26.76 & \textbf{25.97} & 25.50 & \textbf{24.84} & \textbf{23.93} & 24.59\\
n-Seebeck & 50.32 & \textbf{46.66} & 48.79 & \textbf{46.53} & 52.03 & \textbf{49.46} & 42.85 & \textbf{42.13}\\
n-PF & 528.25 & \textbf{506.41} & 502.79 & \textbf{498.43} & 511.56 & \textbf{505.22} & 477.09 & \textbf{474.88} \\
p-Seebeck & 53.01 & \textbf{49.15} & 50.13 & \textbf{48.67} & 54.99 & \textbf{50.72} & 46.04 & \textbf{45.19}\\
p-PF & 513.57 & \textbf{501.11} & 493.35 & \textbf{489.54} & 516.84 & \textbf{494.92} & 460.11 & \textbf{447.81}\\
\bottomrule
\end{tabular}
}
\caption{Summary of the prediction performance (MAE) of different properties (belong to second class) in JARVIS-DFT dataset where performance improvement is modest. Model M is the vanilla variant of a SOTA model and M (Distilled) is the distilled variant using the pretrained CrysGNN{}. All the models are trained on $80\%$ data, validated on $10\%$ and evaluated on $10\%$ of the data. The best performance is highlighted in bold.}
\label{tbl-full-data-2}
\end{table*}
\section{Training Setup / Hyper-parameter Details / Computational Resources used.}
We use five convolution layers of the encoder module to train CrysGNN{} and train it for 200 epochs using Adam \cite{kingma2014adam} for optimization with a learning rate of 0.03. We keep the embedding dimension for each node as 64, batch size of data as 128, and equal weightage (0.25) for $\alpha$, $\beta$, $\gamma$, and $\lambda$ of Equation \ref{eq:pretrain_loss}. During property prediction, we used the default configuration for hyper-parameters for all four vanilla SOTA models. However, while predicting properties by the distilled version of the SOTA models we keep node embedding dimension for each SOTA as 64, and train each model for 500 epochs using Adam~\cite{kingma2014adam} optimization and keep $\delta$ as 0.5. We perform the experiments in shared servers having Intel E5-2620v4 processors which contain 16 cores/thread and four GTX 1080Ti 11GB GPUs each.
\section{Baseline Models.}
To evaluate the effectiveness of CrysGNN{}, we choose following four diverse state of the art algorithms for crystal property prediction.
\begin{enumerate}
\item \textbf{CGCNN}~\cite{xie2018crystal} : This work generates crystal graphs from inorganic crystal materials and builds a graph convolution based supervised model for predicting various properties of the crystals.
\item \textbf{GATGNN}~\cite{louis2020graph} : In this work, authors have incorporated a graph neural network with multiple graph-attention layers (GAT) and a global attention layer, which can learn efficiently the importance of different complex bonds shared among the atoms within each atom’s local neighborhood.
\item \textbf{ALIGNN}~\cite{choudhary2021atomistic} : In this work, authors adopted line graph neural networks to develop an alternative way to
include angular information into convolution layer which alternates between message passing on the bond graph and its bond-angle line graph.
\item \textbf{CrysXPP}~\cite{das2022crysxpp} : In this work, the authors train an autoencoder (CrysAE) on a volume of un-tagged crystal graphs and then the learned knowledge is (transferred to) used to initialize the encoder of CrysXPP, which is fine-tuned with property specific tagged data.Also they design a feature selector that helps to interpret the model’s prediction.
\end{enumerate}
\subsection{Baseline Implementations.} We used the available PyTorch implementations of all the baselines, viz, CGCNN\footnote{https://github.com/txie-93/cgcnn.git}, GATGNN\footnote{https://github.com/superlouis/GATGNN.git}, CrysXPP\footnote{https://github.com/kdmsit/crysxpp.git} and ALIGNN\footnote{https://github.com/usnistgov/alignn.git}. In order to retrofit CrysGNN{} to each SOTA model, we modified their supervised loss into the multi-task loss proposed in Equation \ref{eq:finetune_loss} (main manuscript) and train each model. To ensure a fair comparison between all the baselines and perform the knowledge distillation between node embeddings, we have ensured node embedding dimension for each SOTA and CrysGNN{} as 64.
\begin{table*}
\centering
\small
\setlength{\tabcolsep}{6 pt}
\scalebox{0.9}{
\begin{tabular}{cccccccccc}
\toprule
Property & Train-Val & CGCNN & CGCNN & CrysXPP & CrysXPP & GATGNN & GATGNN & ALIGNN & ALIGNN\\
& -Test(\%) & & (Distilled) & & (Distilled) & & (Distilled) & & (Distilled)\\
\midrule
\multirow{3}{*}{\shortstack{Bandgap \\ (MBJ)}}
& 20-10-70 & 0.588 & 0.453\textbf{*} (23.04) & 0.598 & 0.450\textbf{*} (24.82) & 0.541 & 0.521 (3.70) & 0.497 & 0.485 (2.53) \\
& 40-10-50 & 0.532 & 0.419\textbf{*} (21.41) & 0.496 & 0.405\textbf{*} (18.40) & 0.462 & 0.448\textbf{*} (2.81) & 0.404 & 0.395 (2.20)\\
& 60-10-30 & 0.449 & 0.364 (19.08) & 0.435 & 0.360 (17.36) & 0.449 & 0.439 (2.29) & 0.387 & 0.380 (1.98)\\
\midrule
\multirow{3}{*}{\shortstack{Bulk Modulus \\ (Kv)}}
& 20-10-70 & 16.91 & 16.26 (3.80) & 15.42 & 14.25\textbf{*} (7.59) & 14.80 & 14.19 (4.12) & 14.70 & 14.06 (4.35)\\
& 40-10-50 & 14.81 & 14.46 (2.36) & 15.13 & 14.02\textbf{*} (7.34) & 12.98 & 12.59 (3.00) & 12.47 & 12.11 (2.89)\\
& 60-10-30 & 14.23 & 14.05 (1.26) & 14.76 & 13.73 (6.98) & 12.01 & 11.75 (2.16) & 11.23 & 11.01 (1.96)\\
\midrule
\multirow{3}{*}{\shortstack{Shear Modulus \\(Gv)}}
& 20-10-70 & 13.89 & 12.50 (10.01) & 13.39 & 12.07\textbf{*} (9.86) & 12.83 & 12.42 (3.20) & 12.71 & 12.31 (3.15)\\
& 40-10-50 & 12.04 & 11.54\textbf{*} (4.15) & 12.16 & 11.01\textbf{*} (9.46) & 11.43 & 11.23 (1.75) & 10.98 & 10.67 (2.82)\\
& 60-10-30 & 11.75 & 11.31 (3.74) & 11.77 & 10.67 (9.35) & 10.65 & 10.47 (1.69) & 10.24 & 10.04 (1.95)\\
\midrule
\multirow{3}{*}{\shortstack{SLME (\%)}}
& 20-10-70 & 7.13 & 6.62 (7.17) & 7.05 & 5.90\textbf{*} (16.40) & 6.35 & 6.02 (5.20) & 6.36 & 6.27 (1.43)\\
& 40-10-50 & 6.14 & 5.78 (5.90) & 6.89 & 5.81 (15.67) & 5.78 & 5.63 (2.60) & 5.65 & 5.57 (1.42)\\
& 60-10-30 & 5.55 & 5.24 (5.68) & 5.41 & 4.84 (10.54) & 5.48 & 5.34 (2.55) & 4.88 & 4.82 (1.33)\\
\midrule
\multirow{3}{*}{\shortstack{Spillage}}
& 20-10-70 & 0.424 & 0.389\textbf{*} (8.43) & 0.422 & 0.403 (4.45) & 0.406 & 0.392\textbf{*} (3.5) & 0.402 & 0.392 (2.49)\\
& 40-10-50 & 0.408 & 0.391\textbf{*} (4.28) & 0.398 & 0.384 (3.57) & 0.396 & 0.388 (1.94) & 0.378 & 0.372 (1.72)\\
& 60-10-30 & 0.395 & 0.380 (3.80) & 0.379 & 0.368 (3.00) & 0.382 & 0.377 (1.36) & 0.362 & 0.357 (1.13)\\
\bottomrule
\end{tabular}
}
\caption{MAE values of five different properties in JARVIS-DFT dataset with the increase in training instances from 20 to 60\%. Model M is the vanilla variant of the SOTA model, M (Distilled) is distilled variant of it and relative improvement is mentioned in bracket. Distilled knowledge from CrysGNN{} improves the performance of all the baselines consistently. In few cases, we observe that MAE values of the distilled version of a model using even lesser training data is better than vanilla version of the same model using more training data and highlight it with *.}
\label{tbl-limited-data}
\end{table*}
\section{Additional Experimental Results}
\subsection{Evaluation on Electronic Properties.}
Properties of crystalline materials can be broadly classified into two categories : 1) properties like formation energy, bandgap, total energy, bulk modulus, etc. which depend greatly on crystal structures and atom features, and 2) properties like $\epsilon_x$, $\epsilon_y$, $\epsilon_z$, n-Seebeck, n-PF, etc. which depend on the precise description of the materials’ electronic structure. We discuss performance of SOTA models with distilled knowledge from CrysGNN{} for the first class of properties in {\bf ``Downstream Task Evaluation"} section and report the MAE in Table \ref{tbl-full-data-1}. We further evaluate the effectiveness of CrysGNN for the second class of properties for both the vanilla and distilled model and report the MAE in Table \ref{tbl-full-data-2}. We observe, for this class of properties all the SOTA models had a higher MAE and though pre-training enhances the performance, the improvement is modest. In specific, average improvement in CGCNN, CrysXPP, GATGNN and ALIGNN are 6\%, 4\%, 3.6\% and 0.8\%, respectively. A potential reason is, electronic dielectric constant, Seebeck coefficients, and power factors all depend greatly on the precise description of the materials' electronic structure, which is neither captured by the SOTA models nor by the pre-trained CrysGNN{} framework explicitly. Hence error is high for these properties by SOTA models and injecting distilled structural information from the pre-trained model is able to achieve only modest improvements.\\
\subsection{Effectiveness on sparse training dataset.}
A concise version of this result is presented in the main paper, we here elaborate some details.
To demonstrate the effectiveness of the pre-training in limited data settings, we conduct additional set of experiments under different training data split. In specific, we vary available training data from 20 to 60 \%, train different SOTA models and check their performance on test dataset. We report the MAE values of different baselines and their distilled version in Table \ref{tbl-limited-data} for five different properties, for which available data is very limited. We observe that the distilled version of any SOTA model consistently outperforms its vanilla version even in the limited training data setting, which illustrates the robustness of our pre-training framework.
Specifically, for CGCNN and GATGNN, the improvements are more for 20\% training data, because these deep models suffer from data scarcity issue and with distilled knowledge from the pre-trained model they are able to mitigate the issue. ALIGNN achieves moderate improvements as before but surprisingly, we found the least improvement (additional improvement for 20\% training data compared to other settings) for CrysXPP. CrysXPP itself pre-trains an autoencoder and transfers the learned information to the property predictor. Hence, the vanilla version performs well in the limited data settings. \\
Moreover, in some cases, we observe that the MAE values of the distilled version of a model using lesser training data is better than vanilla version of the model using more training data. For example, while predicting Bandgap (MBJ), CGCNN (Distilled) with 20\% and 40\% training data outperforms CGCNN with 40\% and 60\% training data, respectively. These results verify that even with lesser training data, the distilled information from the large pre-trained model gives performance comparable to using larger training data by the original model.
\section{Ablation Study : Analysis of Different Pre-training Loss}
A concise version of this result is presented in the main paper, we here elaborate some details.
We perform an ablation study to investigate the influence of different pre-training losses in enhancing the SOTA model performance. While pre-training CrysGNN{} (Eq. \ref{eq:pretrain_loss}), we capture both local chemical and global structural information via node and graph-level decoding, respectively.
Further, we are curious to know the influence of each of these decoding policies independently in the downstream property prediction task.
\tblak{In specific, we conduct the ablation experiments, where we pre-train CrysGNN{} with (a) only node-level decoding ($\mathcal{L}_{FR}$, $\mathcal{L}_{CR}$), (b) only graph-level decoding ($\mathcal{L}_{SG}$, $\mathcal{L}_{NTXent}$). Further, we perform ablations with individual graph-level losses, and pretrain with (c) removing $\mathcal{L}_{NTXent}$ (node-level with $\mathcal{L}_{SG}$ (space group)) and (d) removing $\mathcal{L}_{SG}$ (node-level with $\mathcal{L}_{NTXent}$(crystal system)).} We train two baseline models, CGCNN and ALIGNN, with distilled knowledge from all the aforementioned variants of the pre-trained model and evaluate the performance on four crystal properties. We report performance of both the experiments for four properties in Fig.\ref{fig:ablation_diff_loss} and for ten different properties in Fig.\ref{fig:ablation_diff_loss_appendix}.\\
We can observe clearly that all the variants offer significant performance gain in all four properties using the combined node and graph-level pre-training, compared to node-level or graph-level pre-training separately. Only exception is formation energy, where only node-level pre-training produces less error compared to other variants, in both the baseline. Formation energy of a crystal is defined as the difference between the energy of a unit cell comprised of $N$ chemical species and the sum of the chemical potentials of all the $N$ chemical species. Hence pre-training at the node-level (node features and connection) is adequate for enhancing performance of formation energy prediction and incorporating graph-level information works as a noisy information, which degrades the performance. \tblak{We also observe improvement in performance using both supervised and contrastive graph-level losses ($\mathcal{L}_{SG}$ and $\mathcal{L}_{NTXent}$), compared to using only one of them, which proves that the learned representation via supervised and contrastive learning is more expressive that using any one of them.}
Moreover, in ALIGNN, with either node or graph-level pre-training separately, performance degrades across different properties. ALIGNN explicitly captures the three body interactions which drive its performance, to replicate that inclusion of both node and graph information is necessary.
\subsection{Limitations and Future Work.}
Though our proposed pre-trained model is able to enhance the performance of all the state of the art baseline models, improvements for all types of SOTA models are not consistent and we found lesser improvement in case of complex models like GATGNN and ALIGNN. This provides scope for further investigation on designing a more complex and deeper pre-trained model. There are also scopes of exploring different graph representations, loss variations, distillation strategies, etc. Of course, the very fact that the simple model, CrysGNN{}, can provide substantial improvements, lays the foundation for such future exploration. Moreover, in this present work, we have focused on predicting crystal properties, which is a graph level regression task. However, the same framework can be used to explore the effect of pre-training on other graph level tasks (classification or clustering) or node level tasks which would be one of our future works.
\begin{figure*}[ht]
\centering
\boxed{\includegraphics[width=2\columnwidth]{Figure/Legend.pdf}}
\\
\vspace*{-1mm}
\subfloat[Shear Modulus (Gv)]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/shear_modulus.pdf}}}
\subfloat[SLME(\%)]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/slme.pdf}}}
\subfloat[Spillage]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/spillage.pdf}}}
\subfloat[$\epsilon_x$ (MBJ)]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/mepsx.pdf}}}\\
\subfloat[$\epsilon_y$ (MBJ) ]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/mepsy.pdf}}}
\subfloat[$\epsilon_z$ (MBJ) ]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/mepsz.pdf}}}
\subfloat[n-Seebeck]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/n-sb.pdf}}}
\subfloat[n-PF]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/n-pf.pdf}}}\\
\subfloat[p-Seebeck]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/p-sb.pdf}}}
\subfloat[p-PF]{
\boxed{\includegraphics[width=0.50\columnwidth, height=30mm]{Figure/p-pf.pdf}}}
\caption{Summary of experiments of ablation study on importance of different pre-training loss components on CrysGNN{} training and eventually its effect on CGCNN and ALIGNN models on ten different properties from JARVIS DFT dataset(MAE for property prediction). (i) Vanilla: SOTA based model (without distillation) and all the other cases are SOTA models (distilled) from different pre-trained version of CrysGNN{}.
(ii) Node : only node-level pre-training, (iii) Graph : only graph-level pre-training, (iv) Node + L(SG) : node-level and $\mathcal{L}_{SG}$, (v) Node + L(NTXent) : node-level and $\mathcal{L}_{NTXent}$ and (vi) CrysGNN{}~: both node and graph-level pre-training.}
\label{fig:ablation_diff_loss_appendix}
\end{figure*}
\section{Experimental Results}
\label{exp_results}
In this section, we evaluate how the distilled knowledge from CrysGNN{} enhances the performance of different state of the art property predictors on a diverse set of crystal properties from two popular benchmark materials datasets. We first briefly discuss the datasets used both in pre-training and downstream property prediction tasks. Then we report the results of different SOTA property predictors on the downstream property prediction tasks.
Next, we illustrate the effectiveness of our knowledge distillation method compared to the conventional fine-tuning approach.
We further conduct some ablation studies to show the influence of different pre-training losses in predicting different crystal properties and the performance of the system to sparse dataset. Finally, we demonstrate how distilled knowledge from the pre-trained model aids the SOTA models to remove DFT error bias, using very little experimental data.
\subsection{Datasets}
We curated 800K untagged crystal graph data from two popular materials databases, Materials Project (MP) and OQMD, to pre-train CrysGNN{} model. Further to evaluate the performance of different SOTA models with distilled knowledge from CrysGNN{}, we select MP 2018.6.1 version of Materials Project and 2021.8.18 version of JARVIS-DFT, another popular materials database, for property prediction as suggested by \cite{choudhary2021atomistic}. \tblak{Please note, MP 2018.6.1 dataset is a subset of the dataset used for pre-training, whereas JARVIS-DFT is a separate dataset which is not seen during the pre-training.}
MP 2018.6.1 consists of 69,239 materials with two properties bandgap and formation energy, whereas JARVIS-DFT consists of 55,722 materials with 19 properties \tblak{which can be broadly classified into two categories : 1) properties like formation energy, bandgap, total energy, bulk modulus, etc. which depend greatly on crystal structures and atom features, and 2) properties like $\epsilon_x$, $\epsilon_y$, $\epsilon_z$, n-Seebeck, n-PF, etc. which depend on the precise description of the materials’ electronic structure. In the following section, we will evaluate effectiveness of CrysGNN{} on the first class of properties. The impact of structural information is marginal on the second class of property hence all the SOTA perform poorly, there is however some modest improvement using CrysGNN{}; we have put the results and the discussion about it in the Appendix.}\\
Moreover, all these properties in both Materials Project and JARVIS-DFT datasets are based on DFT calculations of chemicals. Therefore, to investigate how pre-trained knowledge helps to mitigate the DFT error, we also take a small dataset OQMD-EXP~\cite{kirklin2015open}, containing 1,500 available experimental data of formation energy. Details of each of these datasets are given in Table \ref{tbl-dataset}. More detail about dataset, different crystal properties and experimental setup is in Appendix
\subsection{Downstream Task Evaluation}
\label{main_results}
To evaluate the effectiveness of CrysGNN{}, we choose four diverse state of the art algorithms for crystal property prediction, CGCNN~\cite{xie2018crystal}, GATGNN~\cite{louis2020graph}, CrysXPP~\cite{das2022crysxpp} and ALIGNN~\cite{choudhary2021atomistic}. To train these models for any specific property, we adopt the multi-task setting discussed in equation \ref{eq:finetune_loss} ,where we distill relevant knowledge from the pre-trained CrysGNN{} to each of these algorithms to predict different properties.
We report mean absolute error (MAE) of the predicted and actual value of a particular property to compare the performance of different participating methods. For each property, we
trained on $80\%$ data, validated on $10\%$ and evaluated on $10\%$ of the data.
We compare the results of distilled version of each SOTA model with its vanilla version (version reported in the respective papers), to show the effectiveness of the proposed framework.\\
\noindent\textbf{Results.}
\tblak{In Table \ref{tbl-full-data-1},
we report MAE of different crystal properties of Materials project and JARVIS-DFT datasets.
In the distilled version of the SOTA models, while training the model, we distill information from the pre-trained CrysGNN{} model. We observe that the distilled version of any state-of-the-art model outperforms the vanilla model across all the properties. In specific, average improvement in CGCNN, CrysXPP, GATGNN and ALIGNN are 16.20\%, 12.21\%, 8.02\% and 4.19\%, respectively.
These improvements are particularly significant as in most of the cases, the MAE is already low for SOTA models, still pretraining enables improvement over that. In fact, lower the MAE, higher the improvement.
We calculate Spearman's Rank Correlation between MAE for each property across different SOTA models and their improvement due to distilled knowledge and found it to be very high (0.72), which supports the aforementioned observations.
The average relative improvement across all properties for ALIGNN (4.19\%) and GATGNN (8.02\%) is lesser compared to CGCNN (16.20\%) and CrysXPP (12.21\%). A possible reason could be that
ALIGNN and GATGNN are more complex models (more number of parameters) than the pre-trained CrysGNN{} framework.
Hence designing a deeper pre-training model or additionally incorporating angle-based information (ALIGNN) or attention mechanism (GATGNN) as a part of pre-training framework may help to improve further. This requires further investigation and we keep it as a scope of future work.}\\
\noindent\textbf{Comparison with Existing Pre-trained Models.} \tblak{We further demonstrate the effectiveness of the knowledge distillation method vis-a-vis the conventional fine-tuning approaches.
Note that the encoding architecture is same for CrysGNN{}, CGCNN, and CrysXPP. CrysXPP is very similar to a {\em pretrained-finetuned} version of CGCNN.
Thus we compare distilled version of CGCNN with finetuned version of CrysGNN{} and CrysXPP.
Additionally, we consider Pretrain-GNN~\cite{hu2020pretraining} which is a popular pre-training algorithm for molecules.
We pre-train all the baseline models on our curated 800K untagged crystal data and fine-tune on seven properties in JARVIS dataset and report the MAE in Table \ref{tbl-distill-finetune}.
We feed multi-graph structure of the crystal material (as discussed in ``Crystal Graph Representation") in Pretrain-GNN and try different combinations of node-level pre-training strategy along with the graph-level supervised pre-training (as suggested in ~\cite{hu2020pretraining}) and report the minimum MAE for any specific property. For finetuned CrysGNN{}, we take the pre-trained encoder of CrysGNN{} and feed a multilayer perceptron to predict a specific property. We observe that distilled CGCNN outperforms finetuned version of CrysGNN{} and both the baselines with a significant margin for all the properties. Pretrain-GNN performs the worst and the potential reason is - it is designed considering simple two-dimensional structure of molecules with a minimal set of node and bond features, which is hard to generalize for crystal materials which have very complex structure with a rich set of node and edge features.}\\
\noindent\textbf{Effectiveness on sparse training dataset.}
Finally, to demonstrate the effectiveness of the pre-training in limited data settings, we conduct additional set of experiments under different training data split. In specific, we vary available training data from 20 to 60 \%, train different SOTA models and check their performance on test dataset. We observe that the distilled version of any SOTA model consistently outperforms its vanilla version even more in the limited training data setting, which illustrates the robustness of our pre-training framework. We report the MAE values of different baselines and their distilled version in Table \ref{tbl-limited-data-small}.
\subsection{Analysis of Different Pre-training Losses}
We perform an ablation study to investigate the influence of different pre-training losses in enhancing the SOTA model performance. While pre-training CrysGNN{} (Eq. \ref{eq:pretrain_loss}), we capture both local chemical and global structural information via node and graph-level decoding, respectively.
Further, we are curious to know the influence of each of these decoding policies independently in the downstream property prediction task.
\tblak{In specific, we conduct the ablation experiments, where we pre-train CrysGNN{} with (a) only node-level decoding ($\mathcal{L}_{FR}$, $\mathcal{L}_{CR}$), (b) only graph-level decoding ($\mathcal{L}_{SG}$, $\mathcal{L}_{NTXent}$). Further, we perform ablations with individual graph-level losses, and pretrain with (c) removing $\mathcal{L}_{NTXent}$ (node-level with $\mathcal{L}_{SG}$ (space group)) and (d) removing $\mathcal{L}_{SG}$ (node-level with $\mathcal{L}_{NTXent}$(crystal system)).} We train two baseline models, CGCNN and ALIGNN, with distilled knowledge from all the aforementioned variants of the pre-trained model and evaluate the performance on four crystal properties. \\
\begin{table}
\centering
\small
\setlength{\tabcolsep}{6 pt}
\scalebox{0.8}{
\begin{tabular}{c | c c c c}
\toprule
Property & CGCNN & CrysGNN{} & CrysXPP & Pretrain \\
& (Distilled) & (Finetuned) & & -GNN\\
\midrule
Formation Energy & \textbf{0.047} & 0.056 & 0.062 & 0.764 \\
Bandgap (OPT) & \textbf{0.160} & 0.183 & 0.190 & 0.688 \\
Total Energy & \textbf{0.053} & 0.069 & 0.072 & 1.451 \\
Ehull & \textbf{0.121} & 0.130 & 0.139 & 1.112 \\
Bandgap (MBJ) & \textbf{0.340} & 0.371 & 0.378 & 1.493 \\
Bulk Modulus (Kv) & \textbf{12.31} & 13.42 & 13.61 & 20.34 \\
Shear Modulus (Gv)& \textbf{10.87} & 11.07 & 11.20 & 16.51 \\
SLME (\%) & \textbf{4.791} & 5.452 & 5.110 & 9.853 \\
Spillage & \textbf{0.354} & 0.374 & 0.363 & 0.481 \\
\bottomrule
\end{tabular}
}
\caption{Comparison of the prediction performance (MAE) of seven properties in JARVIS-DFT between CrysGNN{} and existing pretrain-finetune models,
the best performance is highlighted in bold.}
\label{tbl-distill-finetune}
\end{table}
\begin{table}
\centering
\small
\setlength{\tabcolsep}{2.5 pt}
\scalebox{0.78}{
\begin{tabular}{c|c|cccc}
\toprule
Property & Train-Val & CGCNN & CrysXPP & GATGNN & ALIGNN\\
& -Test(\%) & (Distilled) & (Distilled) & (Distilled) &(Distilled)\\
\midrule
\multirow{3}{*}{\shortstack{Bandgap \\ (MBJ)}}
& 20-10-70 & 0.453 (23.04) & 0.450 (24.82) & 0.521 (3.70) & 0.485 (2.53) \\
& 40-10-50 & 0.419 (21.41) & 0.405 (18.40) & 0.448 (2.81) & 0.395 (2.20)\\
& 60-10-30 & 0.364 (19.08) & 0.360 (17.36) & 0.439 (2.29) & 0.380 (1.98)\\
\midrule
\multirow{3}{*}{\shortstack{Bulk Modulus \\ (Kv)}}
& 20-10-70 & 16.26 (3.80) & 14.25 (7.59) & 14.19 (4.12) & 14.06 (4.35)\\
& 40-10-50 & 14.46 (2.36) & 14.02 (7.34) & 12.59 (3.00) & 12.11 (2.89)\\
& 60-10-30 & 14.05 (1.26) & 13.73 (6.98) & 11.75 (2.16) & 11.01 (1.96)\\
\midrule
\multirow{3}{*}{\shortstack{Shear Modulus \\(Gv)}}
& 20-10-70 & 12.50 (10.01) & 12.07 (9.86) & 12.42 (3.20) & 12.31 (3.15)\\
& 40-10-50 & 11.54 (4.15) & 11.01 (9.46) & 11.23 (1.75) & 10.67 (2.82)\\
& 60-10-30 & 11.31 (3.74) & 10.67 (9.35) & 10.47 (1.69) & 10.04 (1.95)\\
\midrule
\multirow{3}{*}{\shortstack{SLME (\%)}}
& 20-10-70 & 6.62 (7.17) & 5.90 (16.40) & 6.02 (5.20) & 6.27 (1.43)\\
& 40-10-50 & 5.78 (5.90) & 5.81 (15.67) & 5.63 (2.60) & 5.57 (1.42)\\
& 60-10-30 & 5.24 (5.68) & 4.84 (10.54) & 5.34 (2.55) & 4.82 (1.33)\\
\midrule
\end{tabular}
}
\caption{MAE values of distilled version of all the SOTA models for four different properties in JARVIS-DFT dataset with the increase in training instances from 20 to 60\%. Relative improvement in the distilled model is mentioned in bracket.}
\label{tbl-limited-data-small}
\end{table}
\begin{table*}[!thb]
\centering
\small
\setlength{\tabcolsep}{6pt}
\scalebox{0.75}{
\begin{tabular}{l c c c c c c c c}
\toprule
\textbf{Experiment Settings} & CGCNN & CGCNN & CrysXPP & CrysXPP & GATGNN & GATGNN & ALIGNN & ALIGNN\\
& & (Distilled) & & (Distilled) & & (Distilled) & & (Distilled)\\
\midrule
\textbf{\vtop{\hbox{\strut Train on DFT }\hbox{\strut Test on Experimental}}}
& 0.265 & 0.244 (7.60) & 0.243 & 0.225 (7.40) & 0.274 & 0.232 (15.3) & 0.220 & 0.209 (5.05) \\
\midrule
\textbf{\vtop{\hbox{\strut Train on DFT and 20 \% Experimental }\hbox{\strut Test on 80 \% Experimental}}}
& 0.144 & 0.113 (21.7) & 0.138 & 0.118 (14.2) & 0.173 & 0.168 (2.70) & 0.099 & 0.094 (5.60) \\
\midrule
\textbf{\vtop{\hbox{\strut Train on DFT and 80 \% Experimental }\hbox{\strut Test on 20 \% Experimental}}}
& 0.094 & 0.073 (22.7) & 0.087 & 0.071 (18.4) & 0.113 & 0.109 (3.40) & 0.073 & 0.069 (5.90) \\
\bottomrule
\end{tabular}
}
\caption{MAE of predicting experimental values by different SOTA models and their distilled versions with full DFT data and different percentages of experimental data for formation energy in OQMD-EXP dataset. Relative improvement in the distilled model is mentioned in bracket.}
\label{tab:fe}
\end{table*}
\begin{figure}[!thb]
\centering
\boxed{\includegraphics[width=\columnwidth]{Figure/Legend.pdf}}
\\
\vspace*{-1mm}
\subfloat[Formation Energy]{
\boxed{\includegraphics[width=0.44\columnwidth, height=24mm]{Figure/formation_energy.pdf}}}
\subfloat[Bandgap (MBJ)]{
\boxed{\includegraphics[width=0.44\columnwidth,
height=24mm]{Figure/mbj_bg.pdf}}}\\
\subfloat[Total Energy]{
\boxed{\includegraphics[width=0.44\columnwidth, height=24mm]{Figure/total_energy.pdf}}}
\subfloat[Bulk Modulus (Kv)]{
\boxed{\includegraphics[width=0.44\columnwidth, height=24mm]{Figure/bulk_modulus.pdf}}}
\caption{\tblak{Summary of experiments of ablation study on importance of different pre-training loss components on CrysGNN{} training and eventually its effect on CGCNN and ALIGNN models on four different properties (MAE for property prediction). (i) Vanilla: SOTA based model (without distillation) and all the other cases are SOTA models (distilled) from different pre-trained version of CrysGNN{}.
(ii) Node : only node-level pre-training, (iii) Graph : only graph-level pre-training, (iv) Node + L(SG) : node-level and $\mathcal{L}_{SG}$, (v) Node + L(NTXent) : node-level and $\mathcal{L}_{NTXent}$ and (vi) CrysGNN{}~: both node and graph-level pre-training.}}
\label{fig:ablation_diff_loss}
\end{figure}
Experimental results are presented in Fig.~\ref{fig:ablation_diff_loss}.
We can observe clearly that all the variants offer significant performance gain in all four properties using the combined node and graph-level pre-training, compared to node-level or graph-level pre-training separately. Only exception is formation energy, where only node-level pre-training produces less error compared to other variants, in both the baseline. Formation energy of a crystal is defined as the difference between the energy of a unit cell comprised of $N$ chemical species and the sum of the chemical potentials of all the $N$ chemical species. Hence pre-training at the node-level (node features and connection) is adequate for enhancing performance of formation energy prediction and incorporating graph-level information works as a noisy information, which degrades the performance. \tblak{We also observe improvement in performance using both supervised and contrastive graph-level losses ($\mathcal{L}_{SG}$ and $\mathcal{L}_{NTXent}$), compared to using only one of them, which proves the learned representation via supervised and contrastive learning is more expressive that using any one of them.}
Moreover, in ALIGNN, with either node or graph-level pre-training separately, performance degrades across different properties. ALIGNN explicitly captures the three body interactions which drive its performance, to replicate that inclusion of both node and graph information is necessary.
\subsection{Removal of DFT error bias using experimental data}
One of the fundamental issues in material science is that experimental data for crystal properties are very rare~\cite{kubaschewski1993materials,bracht1995properties,turns1995understanding}. Hence existing SOTA models rely on DFT calculated data to train their parameters.
However, mathematical approximations in DFT calculation lead to erroneous predictions (error bias) compared to the actual experimental values of a particular property. Hence
DFT error bias is prevalent in all SOTA models.
\cite{das2022crysxpp} has shown that pre-training helps to remove DFT error bias when fine-tuned with experimental data. Hence,
we investigate whether SOTA models can remove the DFT error with distilled knowledge from pre-trained model, using a small amount of available experimental data. In specific, we consider OQMD-EXP dataset to conduct an experiment, where we train SOTA models and their distilled variants with available DFT data and different percentages of experimental data for formation energy. We report the MAE of different SOTA models and their distilled variant in Table \ref{tab:fe}. We observe, with more amount of experimental training data, all the SOTA models are minimizing the error consistently. Moreover, with distilled knowledge from pre-trained CrysGNN{}, all SOTA models are reducing MAE further and we observe consistently larger degree of improvement with more amount of experimental training data in almost all the models.
\section{Introduction}
\label{intro}
Fast and accurate prediction of different material properties is a challenging and important task in material science. In recent times there has been an ample amount of data driven works ~\cite{seko2015prediction,pilania2015structure,lee2016prediction,de2016statistical,seko2017representation,isayev2017universal,ward2017including,lu2018accelerated,im2019identifying}
for predicting crystal properties which are as accurate as theoretical DFT [Density functional Theory] based approaches~\cite{orio2009density}, however, much faster than it. The architectural innovations of these approaches towards accurate property predictions come from incorporating specific domain knowledge into a deep encoding module. For example, in order to encode the neighbourhood structural information around a node (atom), GNN based approaches ~\cite{xie2018crystal,chen2019graph,louis2020graph,Wolverton2020,schmidt2021crystal} gained some popularity in this domain. Understanding the importance of many-body interactions, ALIGNN~\cite{choudhary2021atomistic} incorporates bond angular information into their encoder module and became SOTA for a large range of property predictions. However, as different properties expressed by a crystalline material are a complex function of different inherent structural and chemical properties of the constituent atoms, it is extremely difficult to explicitly incorporate them into the encoder architecture. Moreover, data sparsity across properties is a known issue~\cite{das2022crysxpp,jha2019enhancing}, which makes these models difficult to train for all the properties. To circumvent this problem we adopt the concept of self-supervised pre-training \cite{devlin2018bert,trinh2019selfie,chen2020simple,chen2020adversarial,he2020momentum,hu2020pretraining,hu2020gpt,qiu2020gcc,you2020graph} for crystalline materials which enables us to leverage a large amount of untagged material structures to learn the complex hidden features which otherwise are difficult to identify.\\
In this paper, we introduce a graph pre-training method which captures (a) connectivity of different atoms, (b) different atomic properties and (c) graph similarity from a large set of unlabeled data. To this effect, we curate a new large untagged crystal dataset with 800K crystal graphs and undertake a pre-training framework (named CrysGNN{}) with the dataset. CrysGNN{} learns the representation of a crystal graph by initiating self-supervised loss at both node (atom) and graph (crystal) level. At the node level, we pre-train the GNN model to reconstruct the node features and connectivity between nodes in a self-supervised way, whereas at the graph level, we adopt supervised and contrastive learning to learn structural similarities between graph structures using the space group and crystal system information of the crystal materials respectively.\\
We subsequently distill important structural and chemical information of a crystal from the pre-trained CrysGNN{} model and pass it to the property predictor.
The distillation process provides wider usage than the conventional pretrain-finetuning framework as transferring pre-trained knowledge to a property predictor and finetuning it requires a similar graph encoder architecture between the pre-trained model and the property predictor, which limits the knowledge transfer capability of the pre-trained model. On the other hand, using knowledge distillation \cite{romero2014fitnets,hinton2015distilling}, we can retrofit the pre-trained CrysGNN{} model into any existing state-of-the-art property predictor, irrespective of their architectural design, to improve their property prediction performance. Also experimental results (presented later) show that even in case of similar graph encoder, distillation performs better than finetuning.\\
With rigorous experimentation
across two popular benchmark materials datasets, we show that distilling necessary information from CrysGNN{}
\footnote{Source code, pre-trained model, and dataset of CrysGNN is made available at https://github.com/kdmsit/crysgnn}
to various property predictors results in substantial performance gains for GNN based architectures and complex ALIGNN model. The improvements range from 4.19\% to 16.20\% over several highly optimized SOTA models. We also perform ablation studies to investigate the influence of different pre-training losses in enhancing the SOTA model performance and observe significant performance gain employing the both node and graph-level pre-training, compared to node-level or graph-level pre-training separately. Also using both supervised and contrastive graph-level pre-training, we are able to learn more robust and expressive graph representation which enhances the property predictor performance.
This also helps to achieve even better improvements when the dataset is sparse. Moreover, the property-tagged dataset suffers from certain biases as it is theoretically (DFT) derived, hence the property predictor also suffers from such bias. We found that on being trained with small amount of experimental data, the DFT bias decreases substantially.
\section{Methodology}
\label{methodology}
\begin{figure*}
\centering
\vspace*{-1mm}
\subfloat[Node-level decoding]{
\boxed{\includegraphics[width=\columnwidth, height=40mm]{Figure/crysgnn_node.pdf}}}
\subfloat[Graph-level decoding]{
\boxed{\includegraphics[width=\columnwidth, height=40mm]{Figure/crysgnn_contrast.pdf}}}
\caption{Overview of both node and graph-level decoding methods for CrysGNN{}. (a) In node-level decoding, node feature attributes and connectivity between nodes are reconstructed in a self-supervised way. (b) In graph-level decoding, $G_2$ is the pivot graph and $G_1$ is from the same crystal system (Cubic), whereas $G_3, G_4, G_5$ are from different crystal systems. First we reconstruct space group information of $G_2$, then through contrastive loss, CrysGNN{} will maximize similarities between positive pair ($G_2,G_1$) and minimize similarities between negative pairs ($G_2, G_3$), ($G_2, G_4$) and ($G_2, G_5$) in embedding space.
\label{fig:CrysGNN}
\end{figure*}
Formally, we first curate a huge amount of property un-tagged crystal graphs $\mathcal{D}_{u} = \{\mathcal{G}_i\}$ from various materials datasets to pre-train a deep GNN model $f_{\theta}$, that learns intrinsic structural and chemical patterns of the crystal graphs.
Further, we use a training set of property tagged crystal graphs $\mathcal{D}_{t} =\{\mathcal{G}_i,y_i\}$ for property prediction, which is smaller in volume and \tblak{may or may not be disjoint from the original untagged set} $\mathcal{D}_{u}$. We train any supervised property predictor $\mathcal{P}_{\psi}$ using $\mathcal{D}_{t}$ to predict the property value given the crystal graph structure. While training the property predictor, we incorporate the idea of knowledge distillation to distill important structural and chemical information from the pre-trained model. This knowledge may prove to be useful to a property predictor which now need not learn from scratch, but be armed with distilled knowledge from the pre-trained model. Hence in this section, we first describe the CrysGNN{} pre-training strategy, followed by the knowledge distillation and property prediction process.
\subsection{CrysGNN{} Pre-training}
\label{pretraining}
We build a deep auto-encoder architecture CrysGNN{}, which comprises a graph convolution based encoder followed by an effective decoder. The autoencoder is (pre)trained end to end, using a large amount of property un-tagged crystal graphs $\mathcal{D}_{u} = \{\mathcal{G}_i\}$, where via node and graph-level self-supervised losses, the model can capture the structural and chemical information of the crystal graph data. First, we formalize the representation of a crystal 3D structure into a multi-graph structure, which will be an input to the encoder module.
\subsubsection{Crystal Graph Representation.}
\label{graph_rep}
We realize a crystal material as a multi-graph structure $\mathcal{G}_i =(\mathcal{V}_i, \mathcal{E}_i, \mathcal{X}_i, \mathcal{F}_i )$ as proposed in \cite{xie2018crystal}. $\mathcal{G}_i$ is an undirected weighted multi-graph where $\mathcal{V}_i$ denotes the set of nodes or atoms present in a unit cell of the crystal structure. $\mathcal{E}_i=\{(u,v,k_{uv})\}$ denotes a multi-set of node pairs and $k_{uv}$ denotes number of edges between a node pair $(u,v)$. $\mathcal{X}_i=\{(x_{u} | u \in \mathcal{V}_i )\}$ denotes the node feature set proposed by CGCNN~\cite{xie2018crystal}. It includes different chemical properties like electronegativity, valance electron, covalent radius, etc. Finally, $\mathcal{F}_i=\{\{s^k\}_{(u,v)} | (u,v) \in \mathcal{E}_i, k\in\{1..k_{uv}\}\}$ denotes the multi-set of edge weights where $s^k$ corresponds to the $k^{th}$ bond length between a node pair $(u,v)$, which signifies the inter-atomic bond distance between two atoms. Next, we formally define CrysGNN{} pre-training and knowledge distillation based property prediction strategy.
\begin{figure*}
\centering
\vspace*{-1mm}
\boxed{\includegraphics[width=1.3\columnwidth]{Figure/distillation_property_prediction.pdf}}
\caption{Overview of Property Prediction using Knowledge Distillation from CrysGNN{}.}
\label{fig:dist_prop}
\end{figure*}
\subsubsection{Self Supervision.}
\label{self_sup}
We first develop a graph convolution \cite{xie2018crystal} based encoding module, which takes crystal multi-graph structure $\mathcal{G} =(\mathcal{V}, \mathcal{E}, \mathcal{X}, \mathcal{F} )$ as input and encodes structural semantics of the crystal graph into lower dimensional space. Each layer of convolution, follows an iterative neighbourhood aggregation (or message passing) scheme to capture the structural information within node's (atom's) neighbourhood. After $L$-layers of such aggregation, the encoder returns the final set of node embeddings $\mathcal{Z} = \{z_1,..., z_{|\mathcal{V}|}\}$, where $z_u :=z_u^L$ represents the final embedding of node $u$. Details of the GNN architecture is in Appendix.
Next, we design an effective decoding module, which takes node embeddings $\mathcal{Z}$ as input and learns local chemical features and global structural information through node and graph-level decoding, respectively. Decoding node-level information will enable CrysGNN{} to learn local domain specific chemical features and connectivity information around an atom, while decoding graph-level features will help CrysGNN{} capture global structural knowledge. \\\\
\textbf{Node-Level Decoding.} For node-level decoding (Fig-\ref{fig:CrysGNN}(a)), we propose two self-supervised learning methods, where we reconstruct two important features that induce the local chemical environment of the crystal around a node (atom). For a given node $u$, we first reconstruct its node features $x_u$,
which represent different chemical properties of atom $u$. Given a node embedding $z_u$, which is encoded based on neighbouring structure around atom $u$, we apply a linear transformation on top of $z_u$ to reconstruct the node attributes. In crystalline graphs, node features correspond to different chemical properties associated with the constituent atoms
through reconstructing these features CrysGNN{} captures local chemical semantics around that atom.\\
Further, we reconstruct local connectivity around an atom, where given node embeddings of two nodes $u$ and $v$, we apply a bi-linear transformation module to generate combined transformed embedding of two nodes $z_{uv}$, which we pass through a feed forward network to predict the strength of association between two atoms. Through reconstructing this local connectivity around an atom, CrysGNN{} encodes the periodicity of the node i.e. the number of neighbours around it along with the relative position of its neighbours and their bond length.
\\\\
\textbf{Graph-level Decoding.}
We aim to capture periodic structure of a crystal material through graph-level decoding (Fig-\ref{fig:CrysGNN}(b)). We specifically leverage two concepts in doing so. (a). {\bf Space group} which is used to describe the symmetry of a unit cell of the crystal material. In materials science literature there are 230 unique space groups and each crystal (graph) has a unique space group number.
(b). {\bf Crystal system.} The space group level information can classify a crystal graph into 7 broad groups of crystal systems like Triclinic, Monoclinic, Orthorhombic, Tetragonal, Trigonal, Hexagonal, and Cubic. Several electronic and optical properties such as band gap, dielectric constant depend on the space-group and the crystal structure justifying its usage. More information about different space groups and crystal systems is in Appendix
\tblak{Given the set of node embeddings $\mathcal{Z} = \{z_1,..., z_{|\mathcal{V}|}\}$, we use a symmetric aggregation function to generate graph-level representation $\mathcal{Z}_{\mathcal{G}}$. First, we pass $\mathcal{Z}_{\mathcal{G}}$ through a feed-forward neural network to predict the space group number of graph $\mathcal{G}$.
Further, we develop a contrastive learning framework for pre-training of CrysGNN{}, where pre-training is performed by maximizing (minimizing) similarity between two crystal graphs belonging to the same (different) crystal system via contrastive loss in graph embedding space.}
A mini-batch of $N$ crystal graphs is randomly sampled and processed through contrastive learning to
align the positive pairs $\mathcal{Z}_{\mathcal{G}_i},\mathcal{Z}_{\mathcal{G}_j}$ of graph embeddings, which belong to the same crystal system and contrast the negative pairs which are from different crystal systems. Here we adopt the
normalized temperature-scaled cross-entropy loss (NT-Xent)\cite{sohn2016improved,van2018representation,wu2018unsupervised} and NT-Xent for the $i^{th}$ graph is defined:
\begin{equation}
\label{eq:contrastive}
\mathcal{L}_{i}=-log\frac{exp(sim(\mathcal{Z}_{\mathcal{G}_i},\mathcal{Z}_{\mathcal{G}_j})/\tau)}{\sum_{k=1}^{K} exp(sim(\mathcal{Z}_{\mathcal{G}_i},\mathcal{Z}_{\mathcal{G}_k})/\tau)}
\end{equation}
where $\tau$ denotes the temperature parameter and $sim(\mathcal{Z}_{\mathcal{G}_i},\mathcal{Z}_{\mathcal{G}_j})$ denotes cosine similarity function.
\tblak{The final loss $\mathcal{L}_{NTXent}$ is computed across all positive pairs in the minibatch.}
Overall we pre-train this deep auto-encoder architecture CrysGNN{} end to end to optimize the following loss :
\begin{equation}
\label{eq:pretrain_loss}
\mathcal{L}_{pretrain}=\alpha\mathcal{L}_{FR}+\beta\mathcal{L}_{CR}+\gamma\mathcal{L}_{SG}+\lambda\mathcal{L}_{NTXent}
\end{equation}
where $\mathcal{L}_{FR},\mathcal{L}_{CR}$ are the reconstruction losses for node feature, and local connectivity, $,\mathcal{L}_{SG}$ is the space group supervision loss, $\mathcal{L}_{NTXent}$ is the contrastive loss and $\alpha$, $\beta$, $\gamma$, $\lambda$ are the weighting coefficients of each loss. We denote the set of parameters in CrysGNN{} model as $\theta$ and the pre-trained CrysGNN{} as $f_{\theta}$.
\subsection{Distillation and Property Prediction}
We aim to retrofit the pre-trained CrysGNN{} model into any SOTA property predictor to enhance its learning process and improve performance (Fig-\ref{fig:dist_prop}). Hence we incorporate the idea of knowledge distillation to distill important structural and chemical information from the pre-trained model, which is useful for the downstream property prediction task, and feed it into the property prediction process.
Formally, given the pre-trained CrysGNN{} model $f_{\theta}$, any SOTA property predictor $\mathcal{P}_{\psi}$ and set of property tagged training data $\mathcal{D}_{t} =\{\mathcal{G}_i,y_i\}$, we aim to find optimal parameter values ${\psi^*}$ for $\mathcal{P}$. We train $\mathcal{P}_{\psi}$ using dataset $\mathcal{D}_{t}$ to optimize the following multitask loss:
\begin{equation}
\label{eq:finetune_loss}
\mathcal{L}_{prop}=\delta\mathcal{L}_{MSE}+(1-\delta)\mathcal{L}_{KD}
\end{equation}
where $\mathcal{L}_{MSE} = (\hat{y_i}-y_i)^2$ denotes the discrepancy between predicted and true property values by $\mathcal{P}_{\psi}$ (property prediction loss). We define knowledge distillation loss $\mathcal{L}_{KD}$ to match intermediate node feature representation between the pre-trained CrysGNN{} model and the SOTA property predictor $\mathcal{P}_{\psi}$ as follows:
\begin{equation}
\label{eq:kd_loss}
\mathcal{L}_{KD}= \lVert {\mathcal{Z}^{T}_{i}} - {\mathcal{Z}^{S}_{i}} \rVert^2
\end{equation}
where $\mathcal{Z}^{T}_{i}$ and $\mathcal{Z}^{S}_{i}$ denote intermediate node embeddings of the pre-trained CrysGNN{} and the property predictor $\mathcal{P}_{\psi}$ for crystal graph $\mathcal{G}_i$, respectively. \tblak{Note, both $\mathcal{Z}^{T}_{i}$ and $\mathcal{Z}^{S}_{i}$ are projected on the same latent space.} Finally, $\delta$ signifies relative weightage between two losses, which is a hyper-parameter to be tuned on validation data. During property prediction the pre-trained network is frozen and we backpropagate $\mathcal{L}_{prop}$ through the predictor $\mathcal{P}_{\psi}$ end to end.
\begin{table}
\centering
\small
\setlength{\tabcolsep}{2 pt}
\scalebox{0.68}{
\begin{tabular}{c | c | c | c | c | c }
\toprule
Task & Datasets & Graph Num. & Structural Info. & Properties Count & Data Type\\
\midrule
\multirow{2}{*}{\shortstack{Pre-training}}
& OQMD & 661K & \checkmark & x & DFT Calculated\\
& MP & 139K & \checkmark & x & DFT Calculated\\
\midrule
\multirow{3}{*}{\shortstack{Property \\ Prediction}} & MP 2018.6.1 & 69K & \checkmark & 2 & DFT Calculated\\
& JARVIS-DFT & 55K & \checkmark & 19 & DFT Calculated\\
& OQMD-EXP & 1.5K & \checkmark & 1 & Experimental\\
\bottomrule
\end{tabular}
}
\caption{Datasets Details}
\label{tbl-dataset}
\end{table}
\begin{table*}
\centering
\small
\setlength{\tabcolsep}{11 pt}
\scalebox{0.9}{
\begin{tabular}{c | c c | c c| c c | c c}
\toprule
Property & CGCNN & CGCNN & CrysXPP & CrysXPP & GATGNN & GATGNN & ALIGNN & ALIGNN \\
& & (Distilled) & & (Distilled) & & (Distilled) & & (Distilled)\\
\midrule
Formation Energy & 0.039 & \textbf{0.032} & 0.041 & \textbf{0.035} & 0.096 & \textbf{0.091} & 0.026 & \textbf{0.024} \\
Bandgap (OPT) & 0.388 & \textbf{0.293} & 0.347 & \textbf{0.287} & 0.427 & \textbf{0.403} & 0.271 & \textbf{0.253} \\
\midrule
Formation Energy & 0.063 & \textbf{0.047} & 0.062 & \textbf{0.048} & 0.132 & \textbf{0.117} & 0.036 & \textbf{0.035} \\
Bandgap (OPT) & 0.200 & \textbf{0.160} & 0.190 & \textbf{0.176} & 0.275 & \textbf{0.235} & 0.148 & \textbf{0.131} \\
Total Energy & 0.078 & \textbf{0.053} & 0.072 & \textbf{0.055} & 0.194 & \textbf{0.137} & 0.039 & \textbf{0.038} \\
Ehull & 0.170 & \textbf{0.121} & 0.139 & \textbf{0.114} & 0.241 & \textbf{0.203} & 0.091 & \textbf{0.083} \\
Bandgap (MBJ) & 0.410 & \textbf{0.340} & 0.378 & \textbf{0.350} & 0.395 & \textbf{0.386} & 0.331 & \textbf{0.325} \\
Spillage & 0.386 & \textbf{0.374} & 0.363 & \textbf{0.357} & 0.350 & \textbf{0.348} & 0.358 & \textbf{0.356} \\
SLME (\%) & 5.040 & \textbf{4.790} & 5.110 & \textbf{4.630} & 5.050 & \textbf{4.950} & 4.650 & \textbf{4.590} \\
Bulk Modulus (Kv) & 12.45 & \textbf{12.31} & 13.61 & \textbf{12.70} & 11.64 & \textbf{11.53} & 11.20 & \textbf{10.99} \\
Shear Modulus (Gv) & 11.24 & \textbf{10.87} & 11.20 & \textbf{10.56} & 10.41 & \textbf{10.35} & 9.860 &\textbf{9.800} \\
\bottomrule
\end{tabular}
}
\caption{Summary of the prediction performance (MAE) of different properties in Materials project (Top) and JARVIS-DFT (Bottom). Model M is the vanilla variant of a SOTA model and M (Distilled) is the distilled variant using the pretrained CrysGNN{}. The best performance is highlighted in bold.}
\label{tbl-full-data-1}
\end{table*}
\section{Related Work}
\label{related_work}
In recent times, data driven approaches~\cite{seko2015prediction,pilania2015structure,lee2016prediction,de2016statistical,seko2017representation,isayev2017universal,ward2017including,lu2018accelerated,im2019identifying} have become quite popular to establish relationship between the atomic structure of crystalline materials and their properties with very high precision. Especially, graph neural network (GNN) based approaches ~\cite{xie2018crystal,chen2019graph,louis2020graph,Wolverton2020,schmidt2021crystal,choudhary2021atomistic} have emerged as a powerful machine learning model tool to encode material’s complex topological structure along with node features in an enriched representation space.\\
There are attempts to pre-train GNNs to extract graph and node-level representations.~\cite{hu2020pretraining} develops an effective pre-training strategy for GNNs, where they perform both node-level and graph-level pre-training on GNNs to capture domain specific knowledge about nodes and edges, in addition to global graph-level knowledge. Followed by this work, there has been several other works on self-supervised graph pre-training ~\cite{hu2020gpt,qiu2020gcc,you2020graph}, which propose different graph augmentation methods and maximizes the agreement between two augmented views of the same graph via a contrastive loss. In the field of crystal graphs, CrysXPP~\cite{das2022crysxpp} is the only model which comes close to a pre-trained model. In their work, an autoencoder is trained on a volume of un-tagged crystal graphs and the learned knowledge is (transferred to) used to initialize the encoder of CrysXPP, which is fine-tuned with property specific tagged data. \\
\tblak{Although conceptually similar to the work done by ~\citet{hu2020pretraining}, our work differs in the following three key aspects: (1) pre-training strategy proposed by ~\citeauthor{hu2020pretraining} is very effective for molecular dataset, but it is difficult to extend directly to crystalline material because structural semantics are different between molecules and materials \cite{xie2021crystal}.
Molecules have non-periodic and finite structures, solid materials’ structures are infinite and periodic in nature.
(2) For graph-level pre-training, ~\citeauthor{hu2020pretraining} adapted supervised graph-level property prediction using a huge amount of labelled dataset from chemistry and biology domain, which makes it less effective in several other domains like material science where property labeled data is extremely scarce. Also, a crucial step in graph-level prediction is to find graph structural similarity between two sets of graphs, which they do not explore but mention as a future work. We do not make use of supervised pre-training which requires a large amount of property tagged material data. Instead, we leverage the idea of structural similarity of materials belonging to the similar space group, and via contrastive loss and space group classification loss, we try to capture this similarity. (3) Finally they follow conventional pre-train finetuning framework, whereas in CrysGNN{} we incorporate the idea of knowledge distillation \cite{romero2014fitnets,hinton2015distilling} to distill important information from the pre-trained model and inject it into the property prediction process. By design, this knowledge distillation based approach is more robust and independent of the underlying architecture of the property predictor, thus it can enhance the performance of a diverse set of SOTA models.}
\section{Acknowledgments}
This work was partially funded by Indo Korea Science and Technology Center, Bangalore, India, under the project name {\em Transfer learning and Weak Supervision for Accurate and Interpretable Prediction of Properties of Materials from their Crystal Graph Representation}. Also, it was (partially) funded by the Federal Ministry of Education and Research (BMBF), Germany under the project LeibnizKILabor with grant No.\,01DD20003. We thank the Ministry of Education, Govt of India, for supporting Kishalay with Prime Minister Research Fellowship during his Ph.D. tenure.
|
\section{Introduction}
Computational aesthetics evaluation \cite{galanter2012computational} enables computers to make qualitative or quantitative aesthetic judgments on works of art. These works of art usually include painting, music and design. It is meaningful for computers to realize beauty because this can guide AI generatation tasks.
Although the existing work of music generation is very mature, the quality of music score generated by AI is relatively poor compared with that created by human composers. This is probably because the essence of AI generation task is to predict the probability of the next music unit being played and the lack of prior music theory knowledge leads to the music generated by AI sounds unpleasant.
There are three main steps in the production of pop music: composition of music score, arrangement, and finally played by the performer. We hope that the quality of score can be evaluated from the stage of music score composition, so as to eliminate the interference of different performers' performance levels on the evaluation of music score quality.
Due to a lack of labeled aesthetic score on music scores like AVA \cite{murray2012ava} in the field of image aesthetic, we adopt the traditional aesthetic measure method to study the aesthetic model. Traditional aesthetic measure can analyze the beauty of objects from the perspective of information in objects.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.48\textwidth]{good_bad}
\caption{The quality of symbolic score can be easily evaluated through the Score Aesthetic Assessment Model (SAAM).}
\label{fig:good_bad}
\label{fig:good_bad}
\end{figure}
Our goal is to create a music score aesthetic assessment model that can objectively distinguish the good from the bad.
In this paper, Birkhoff's method \cite{birkhoff2013aesthetic} was selected to conduct a study of aesthetic quality assessment of music score from the perspective of information theory. Birkhoff formalizes the aesthetic measure of an object into the quotient between order and complexity:
\begin{equation}
M=\frac{O}{C} \label{1}
\end{equation}
Fig 1 briefly describes the content of our work. The main contributions of our work are as follows:
\begin{itemize}
\item We put forward a score aesthetic assessment model to objectively evaluate the quality of homophony music score as a baseline score aesthetic assessment model.
\item We put forward and update eight basic music features and four music aesthetic features in combination with information theory and music theory.
\end{itemize}
\begin{figure*}[htbp]
\centering
\includegraphics[width=\textwidth]{networknew}
\caption{First, we do not tokenize the symbolic score. We extract the attributes in the symbolic score, which requires preprocessing. After preprocessing, we get the pitch, rhythm, chord attributes of the score and the label of the score as the ground truss for classification. Then, we process the music attributes and extract 8 music features (small green boxes). Next, we train the samples through four logical regression models (LR stands for logical regression) and combined to extract four aesthetic features (light blue boxes). Finally, we input the four aesthetic attributes into our model, use sigmoid function to establish the loss of its error with the ground truth, and calculate the parameters of the aesthetic model.}
\label{fig:networknew}
\label{fig:networknew}
\end{figure*}
\section{Related Work}
There is only one work of music aesthetic measure using information theory, which is the aesthetic measure of audio. Audio Oracle \cite{dubnov2011audio} (AO) uses Information Rate (IR) as an aesthetic measure. However, it cannot clarify what kind of specific aesthetic intention the system has, because repetition or redundancy (proposed by IR) has essentially different meanings, interpretations and values in art, and it is questionable to use information rate as the aesthetic measure.
There are three levels of music generation: score generation, performance generation and audio generation. The work we discuss in this section is related to music score generation. In order to make full use of music theory and study complexity, we will not discuss monophonic music, but we will discuss homophony. Homophony is a kind of multipart music, which has a melodic part and an accompanying part. In the generation of homophony music field, there are tasks such as directly generating melody and chord \cite{lim2020style}, generating melody through chord \cite{inprovrnn}, and generating chord according to melody \cite{zhu2020pop}. Although the existing generation technology is rather mature, the quality of music score created by AI is still low, which sounds monotonous and lacks emotion.
In order to evaluate the quality of music, in the field of music generation, the evaluation part is often divided into objective evaluation and subjective evaluation. The objective evaluation is often to set some statistical metrics for the music generated by AI, and the result of objective evaluation is completely calculated by computer. Many music toolkits have objective evaluation metrics packaged for direct use, such as MV2H \cite{mcleod2018evaluating} and MusPy \cite{dong2020muspy}, etc. MV2H evaluated how many errors there are between the generated music and the ground truth. In Muspy, basic statistical metrics of symbolic score are provided. The subjective evaluation generally includes listening test and visual analysis, almost all AI generation tasks involve subjective evaluation experiments such as scoring and Turing test, which are considered necessary and essential.
\section{Score Aesthetic Assessment Model}
\subsection{Formalization of the Model}
Based on Birkhoff's theory, information theory and music theory, we propose four aesthetic features: harmony, symmetry, entropy and K-Complexity. We linearly combine the order measures of molecules and the complexity measures of denominators. Detailed measures explaination will be described in Sections 3.2 and 3.3. Fig 2 shows the process of our work. The music aesthetic measure formula is as follows:
\begin{equation}
Aesthetic\ Measure=\frac{\omega_1H + \omega_2S + \theta_1}{\omega_3E + \omega_4K + \theta_2} \label{6}
\end{equation}
Where $H$ is harmony, $S$ is symmetry, $E$ is entropy and $k$ is K complexity. $\omega$ is the weight and $\theta$ is the constant.
\subsection{Order Measures}
When objects have some characteristics of harmony, symmetry or order, they often have a certain sense of beauty. We quantify the order of music in two dimensions, harmony and symmetry. Harmony mainly calculates based on music theory knowledge, while symmetry mainly relies on some statistical information in music. Next, we use the linear combination of harmony and symmetry as the measure of order.
\subsubsection{Interval Harmony}
In music, the distance between two notes is called interval. In particular, in music, when an interval is 12 semitone, we call it an octave. Interval classification can be found in the supplementary material, interval are divided into five categories.
Mathematical and physical research shows that when two sound frequencies are a simple integer ratio, it is more pleasant to listen together. Therefore, we propose a calculation method of interval harmony, and the formula is as follows.
\begin{equation}
Interval\ Harmony=\sum_{i=1}^{12} \alpha_i * pir_i + \theta_{ih} \label{1}
\end{equation}
Among them, $\alpha_i$ is the weight of interval, $pir_i$ is the ratio of interval to total interval, $\theta_{ih}$ is the constant.
\subsubsection{Chord Progression Harmony}
According to Schoenberg's theory of harmony \cite{schoenberg1983theory}, the internal chord is divided into three functional harmonies: tonic triad (T), subdominant triad (S) and dominant triad (D). An example of chord functions and series are shown in supplementary material.
A complete harmony progression starts from the tonic triad, proceeds to the subordinate triad, proceeds to the dominant triad, and finally returns to the tonic triad to complete a complete cycle, which is called complete progression. In the usual sense, harmony progression is the connection of chords within a certain harmonic range in tonal music.
There are many ways to quantify harmony progression. In this paper, we refer to the method of María \cite{navarro2020computational}. In our work, we took the average value of the progression tension to obtain a quantitative chord progress harmony. It is calculated by referring to the following formula:
\begin{equation}
\begin{split}
Chord\ Progression\ Harmony = \lambda_1 d_1(T_i,T_{i-1})+ \\
\lambda_2 d_2(T_i, T_{key}) + \lambda_3 d_3(T_i-T_{key}, T_f) + \lambda_4 c(T_i) + \\
\lambda_5 m(T_i,P) + \lambda_6 h(T_i,P)
\end{split}
\end{equation}
Where $T_i$ is the i-th chord of progression $P$, $\lambda$ is the weight. For more information of parameters $c$, $m$, $h$, see \cite{navarro2020computational}.
\subsubsection{Self Similarity Fitness}
In the field of music generation, structure is often discussed as an important feature. Almost all music contains repetitive pieces. We will discuss the influence of repetitive structure on music aesthetics. We measure it with self similarity fitness.
Inspired by the aesthetics of images and art \cite{al2017symmetry}, the aesthetics beauty of music comes from the symmetry in musical compositions. Therefore, in our study, we refer to Müller's fitness method \cite{MuellerJG13_StructureAnaylsis_IEEE-TASLP} to measure the degree of repetition in a piece of music. The fitness formula is shown as follow:
\begin{equation}
Self\ Similarity\ Fitness = 2 \cdot \frac{\bar{\sigma}(\alpha) \cdot \bar{\gamma}(\alpha)}{\bar{\sigma}(\alpha)+\bar{\gamma}(\alpha)}
\end{equation}
Both $\bar{\sigma}(\alpha)$ and $\bar{\gamma}(\alpha)$ are related to a concept defined by Müller's method \cite{MuellerJG13_StructureAnaylsis_IEEE-TASLP}.
\subsubsection{Skewness}
Skewness is a concept proposed by jSymbolic \cite{mckay2018jsymbolic}. It proposes that both pitch and rhythm of music have the concept of skewness. The notes in music cannot lack pitch and rhythm. Skewness describes how asymmetrical the pitch / rhythm is to either the left or the right of the mean pitch / rhythm value.
The features are extracted based on jSymbolic, the calculation formula of pitch and rhythm skewness are not specifically described here. We combine pitch skewness and rhythm skewness linearly to get the formula of skewness:
\begin{equation}
Skewness = \beta_1 * PS + \beta_2 * RS + \theta_{sk}
\end{equation}
Where $PS$ is pitch skewness, $RS$ is rhythm skewness, $\beta$ represents their weight and $\theta_{sk}$ is a constant.
\subsection{Complexity Measures}
Bense \cite{bense1960programmierung} first uses Birkhoff's aesthetic measure formula to calculate aesthetics. They adapt statistical measure of information in aesthetic objects and believe that the objective measure of aesthetic objects is related to the complexity of objects. Their idea has to use information theory, and entropy is the core of it. Our aesthetic measure method considers two features: Shannon entropy and Kolmogorov complexity.
\subsubsection{Shannon Entropy}
Let $\Omega$ be a finite set, and $X$ be a random variable. The value $x$ in $\Omega$ has a distribution $p(x)=Pr[X=x]$. The Shannon entropy $H(X)$ of random variable $X$ is defined as follows:
\begin{equation}
H(X)=-\sum_{x\in\Omega}p(x)\log(x)
\end{equation}
The Shannon entropy $H(X)$ measures the average uncertainty of random variable $X$, which is widely used to evaluate the degree of chaos in the internal state of a system. In order to calculate the entropy of music, it is necessary to obtain the music attribute histogram.
Pitch and rhythm are the two basic elements of music. In our method, we consider pitch entropy and rhythm histogram entropy, and take their linear combination as the measure of entropy. This is used to describe the uncertainty in music. Our entropy formula is as follows:
\begin{equation}
Entropy=\eta_1 * PHE + \eta_2 * RHE + \theta_e
\end{equation}
Where $PHE$ and $RHS$ are pitch and rhythm histogram entropy, $\eta$ represents their weight and $\theta_e$ is a constant.
\subsubsection{Kolmogorov Complexity}
For a string $s$, Kolmogorov complexity $K(s)$ of the string $s$ refers to the shortest program to calculate the string $s$ on a computer. In essence, the Kolmogorov complexity of a string is the length of the final compressed version of the string. Then, we use the linear combination of entropy and Kolmogorov complexity as a measure of complexity.
Aesthetically speaking, redundancy makes people feel dull, resulting in negative emotions. According to the definition of Kolmogorov complexity, we also refer to the method of using Kolmogorov complexity in image aesthetics \cite{rigau2008informational}.
We believe that Kolmogorov complexity in music is also computable. It is actually the lossless compression ratio of music, which can be formalized as the following formula:
\begin{align}
Kolmogorov\ Complexity=\frac{NH_{m} - K}{NH_{m}} \label{1}
\end{align}
Where $NH_{m}$ is information content of a music, and $K$ is the simplest music information after compression.
\begin{figure*}[htbp]
\centering
\includegraphics[width=1\textwidth]{score}
\caption{As can be seen from the figure, the melody of composer score is in order, and it matches the chord progression very well. The melody of AI score seems very random and low-quality, which is specifically reflected in the irregular appearance of the rest and the melody do not follow the tension of the chord progression. \textbf{Note: The full score of normalized measure is 1.}}
\end{figure*}
\section{Implementation}
\subsection{Datasets}
There are many datasets of pop music, such as POP909 \cite{wang2020pop909} and Wikifonia\footnotemark{}. But we don't use them due to some reasons.
All music data formats in POP909 dataset are performance midi, which will cause different performers' aesthetic impact on the same music score. Although Wikifonia is a score dataset (single track), it can't be used by order measure.
\footnotetext{http://www.wikifonia.org/}
In order to eliminate the performance differences of different performers, objectively measure the aesthetic value of a piece of music, we finally download the scores of 100 pop songs on the Musescore\footnotemark{} website. Then we extract the chords of music scores in Musescore for generating 100 music scores. In datail, we take chord progression from the scores of Musescore, and determine scores' key signature (tonic) as the input parameter, then give them to Magenta's inprov\_rnn \cite{inprovrnn} enables it to generate music according to chord progression and key signature, which has the advantage of controlling the time length of two pairs of datasets to be approximately the same. The scores in Musescore are created by composers as positive samples, while the scores created by Magenta are generated by AI as negative samples.
\footnotetext{https://musescore.org/}
Since we use cross validation, we do not set validation sets. We split the dataset at a ratio of 7:3 for training testing.
\subsection{Preprocessing \& Computing Aesthetic Features}
We use music21 \cite{cuthbert2010music21} toolkit to load music scores. Music21 has note and chord attributes, which can easily obtain the information of music score and make calculations.
Firstly, we obtain the pitch histogram and rhythm histogram in the music score to calculate entropy. Then, we get all the intervals by calculating the note events that occur at the same time, and obtain a histogram. So, the histogram entropy of pitch and rhythm and interval harmony can be calculated.
Secondly, we try to get all the chords and the key signature in the score to calculate the chord progress harmony. The chord progress and the key signature will be saved in the json format and subsequently input to the pretrained model.
Thirdly, we use Musescore3 to batch process composer scores and AI scores into score midi. This is to facilitate the use of jSymbolic \cite{mckay2018jsymbolic} to extract features. The conversion of xml to score midi will not lose the score information. In this way, we get pitch and rhythm skewness.
Fourthly, we use Musescore3 to render the music score into audio, which controls that all music scores are played by Musescore3 to ensure that the music score information is lossless. Then we use the wav file format to calculate the self similarity fitness. Then, we refer to Monkey's Audio's\footnotemark{}\footnotetext{https://www.monkeysaudio.com/} lossless compression method to compress music into ape format, so as to calculate Kolmogorov complexity.
Fifthly, calculating the values of 8 basic music features (refer to Table 1), we will normalize them. Next, we use the method of logical regression to confirm the values of Harmony, Symmetry, Entropy, and Kolmogorov complexity.
Finally, we take the four normalized aesthetic features as inputs, use sigmoid function to map the aesthetic measure to 0 and 1 for classification, and use cross-entropy loss to make loss function, confirming the parameters of the aesthetic model in the way of gradient decline. We set the learning rate to 0.01. After 1000 iterations, the loss function converges.
\section{Experiments}
\subsection{Turing Test}
We need to verify whether people really think the music created by composers is more beautiful than the music generated by AI. So we did the Turing test to see if people can really distinguish between composer music and AI music.
We randomly sampled 10 pieces from the music created by the composer and the music generated by AI respectively, with each piece lasting about 15 seconds. Volunteers participating in the Turing test need to identify which music is created by the composer and which music is generated by AI in these 10 pairs of pieces. The volunteer also needs to choose the one in each pair that he thinks is more beautiful.
A total of 15 volunteers participate in the Turing test. Among 300 samples, their classification accuracy of music is 91.3\% (274). Assuming that the music created by the composer is more aesthetic, 87\% (261) of the samples are correctly classified. This proves that our assumption is correct.
\subsection{Results \& Discussion}
Since the music created by the composer has a higher aesthetic feeling than the music generated by AI, we let the machine learn the aesthetic score according to the label value. This is essentially a binary classification problem.
\begin{figure}[!htbp]
\centering
\includegraphics[width=0.352\textwidth]{distribution1}
\caption{The intersection of the distributions is quite small, showing that our model can distinguish between AI scores and composer scores very well.}
\label{fig:distribution}
\label{fig:distribution}
\end{figure}
After the weight is obtained by gradient descent, we bring the weight into the aesthetic model to view the distribution of the aesthetic model. The distribution is shown in Fig 4.
We use precision and F1-Measure as our metric to test our model. The precision of our model on the test set is 93.3\%, and the F1-Measure is 90.9\%. This proves that our model is valid. Fig 3 shows an example of score comparison.
\begin{table}[htbp]
\centering
\resizebox{1\columnwidth}{!}{
\begin{tabular}{c|c|c|c|c|c|c|c|c}
\hline
\textbf{Dataset} & \textbf{IH} & \textbf{CPH} & \textbf{SSF} & \textbf{PS} & \textbf{RS} & \textbf{PHE} & \textbf{RHE} & \textbf{KC} \\ \hline
AI & 0.97 & 1.98 & 0.06 & \textbf{0.49} & 2.16 & \textbf{1.34} & \textbf{1.60} & \textbf{0.69} \\ \hline
Composer & \textbf{1.40} & \textbf{2.04} & \textbf{0.19} & 0.53 & \textbf{0.99} & 2.16 & 1.79 & 0.73 \\ \hline
\end{tabular}
}
\caption{The similar CPH value shows the rationality of the AI dataset generated based on chords. \textbf{More detailed distribution comparison can be found in the supplementary material.} The results in bold have higher aesthetic scores.}
\end{table}
Table 1 shows the calculation results of 8 features without normalization. We can make the following analysis from it. Whether IH or CPH, the scores of composer is obviously better than that of AI. This is also in line with the music theory. Considering Symmetry, the SSF of composer is obviously higher than that of AI. This is because the music created by AI is too random and often has no repetitive pieces, which proves beauty is related to repetition. Although the PS of AI is smaller than that of the composer, the difference is not much. The RS of the composer is obviously smaller than that of AI. This is because composers tend to create more regular rhythm and pitch than AI. When it comes to entropy, the values of PE and HE are obviously higher in composer score than in AI score. Although this is contrary to Birkhoff's aesthetic measure, it is also reasonable. This is because composer tends to add some changes to pitch and rhythm, while music generated by AI does not change much around the tonic. As for K-complexity, the difference between composer's scores and AI's scores is not significant, but it can be seen from the table that the music created by AI is relatively low in compression and complexity. In conclusion, moderate order and complexity quotient can quantify aesthetic feeling to a certain extent.
\subsection{Ablation Study}
We conduct ablation experiments to remove harmony, symmetry, entropy and K-complexity respectively to train four different models. We compare them with the original model.
As shown in Fig 5, we observe the ROC curves of the original model and four models with one aesthetic feature removed respectively, and obtain their AUC values. The AUC value of our model is 0.93, which is obviously higher than that of the four models without aesthetic features.
If harmony is removed, the AUC value is only 0.77, which shows the importance of harmony and further proves that music theory plays a very important role in music aesthetics. If symmetry is removed, the AUC value of the model is 0.85, which indicates that symmetry may contribute slightly less to aesthetics than harmony. As for entropy, it is obviously important, but K-complexity seems not.
\begin{figure}[!htbp]
\centering
\includegraphics[width=0.38\textwidth]{Ablation}
\caption{The ROC curves are serrated due to lack of samples.}
\label{fig:ablation}
\label{fig:ablation}
\end{figure}
\section{Conclusion}
In summary, we propose a score aesthetic assessment model using Birkhoff's aesthetic measure to quantify a score's aesthetic. We also discover four categories of music aesthetic features, totaling eight basic aesthetic features. We have made some contributions to improve the quality of music scores. This might be helpful for music score quality assessment. However, our method still has shortcomings, for instance we have not taken the relationship between creativity and musical aesthetics into consideration. The aesthetic study of music audio quality assessment is worth exploring in the future.
\bibliographystyle{IEEEbib}
\section{Introduction}
Computational aesthetics evaluation \cite{galanter2012computational} enables computers to make qualitative or quantitative aesthetic judgments on works of art. These works of art usually include painting, music and design. It is meaningful for computers to realize beauty because this can guide AI generatation tasks.
Although the existing work of music generation is very mature, the quality of music score generated by AI is relatively poor compared with that created by human composers. This is probably because the essence of AI generation task is to predict the probability of the next music unit being played and the lack of prior music theory knowledge leads to the music generated by AI sounds unpleasant.
There are three main steps in the production of pop music: composition of music score, arrangement, and finally played by the performer. We hope that the quality of score can be evaluated from the stage of music score composition, so as to eliminate the interference of different performers' performance levels on the evaluation of music score quality.
Due to a lack of labeled aesthetic score on music scores like AVA \cite{murray2012ava} in the field of image aesthetic, we adopt the traditional aesthetic measure method to study the aesthetic model. Traditional aesthetic measure can analyze the beauty of objects from the perspective of information in objects.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.48\textwidth]{good_bad}
\caption{The quality of symbolic score can be easily evaluated through the Score Aesthetic Assessment Model (SAAM).}
\label{fig:good_bad}
\label{fig:good_bad}
\end{figure}
Our goal is to create a music score aesthetic assessment model that can objectively distinguish the good from the bad.
In this paper, Birkhoff's method \cite{birkhoff2013aesthetic} was selected to conduct a study of aesthetic quality assessment of music score from the perspective of information theory. Birkhoff formalizes the aesthetic measure of an object into the quotient between order and complexity:
\begin{equation}
M=\frac{O}{C} \label{1}
\end{equation}
Fig 1 briefly describes the content of our work. The main contributions of our work are as follows:
\begin{itemize}
\item We put forward a score aesthetic assessment model to objectively evaluate the quality of homophony music score as a baseline score aesthetic assessment model.
\item We put forward and update eight basic music features and four music aesthetic features in combination with information theory and music theory.
\end{itemize}
\begin{figure*}[htbp]
\centering
\includegraphics[width=\textwidth]{networknew}
\caption{First, we do not tokenize the symbolic score. We extract the attributes in the symbolic score, which requires preprocessing. After preprocessing, we get the pitch, rhythm, chord attributes of the score and the label of the score as the ground truss for classification. Then, we process the music attributes and extract 8 music features (small green boxes). Next, we train the samples through four logical regression models (LR stands for logical regression) and combined to extract four aesthetic features (light blue boxes). Finally, we input the four aesthetic attributes into our model, use sigmoid function to establish the loss of its error with the ground truth, and calculate the parameters of the aesthetic model.}
\label{fig:networknew}
\label{fig:networknew}
\end{figure*}
\section{Related Work}
There is only one work of music aesthetic measure using information theory, which is the aesthetic measure of audio. Audio Oracle \cite{dubnov2011audio} (AO) uses Information Rate (IR) as an aesthetic measure. However, it cannot clarify what kind of specific aesthetic intention the system has, because repetition or redundancy (proposed by IR) has essentially different meanings, interpretations and values in art, and it is questionable to use information rate as the aesthetic measure.
There are three levels of music generation: score generation, performance generation and audio generation. The work we discuss in this section is related to music score generation. In order to make full use of music theory and study complexity, we will not discuss monophonic music, but we will discuss homophony. Homophony is a kind of multipart music, which has a melodic part and an accompanying part. In the generation of homophony music field, there are tasks such as directly generating melody and chord \cite{lim2020style}, generating melody through chord \cite{inprovrnn}, and generating chord according to melody \cite{zhu2020pop}. Although the existing generation technology is rather mature, the quality of music score created by AI is still low, which sounds monotonous and lacks emotion.
In order to evaluate the quality of music, in the field of music generation, the evaluation part is often divided into objective evaluation and subjective evaluation. The objective evaluation is often to set some statistical metrics for the music generated by AI, and the result of objective evaluation is completely calculated by computer. Many music toolkits have objective evaluation metrics packaged for direct use, such as MV2H \cite{mcleod2018evaluating} and MusPy \cite{dong2020muspy}, etc. MV2H evaluated how many errors there are between the generated music and the ground truth. In Muspy, basic statistical metrics of symbolic score are provided. The subjective evaluation generally includes listening test and visual analysis, almost all AI generation tasks involve subjective evaluation experiments such as scoring and Turing test, which are considered necessary and essential.
\section{Score Aesthetic Assessment Model}
\subsection{Formalization of the Model}
Based on Birkhoff's theory, information theory and music theory, we propose four aesthetic features: harmony, symmetry, entropy and K-Complexity. We linearly combine the order measures of molecules and the complexity measures of denominators. Detailed measures explaination will be described in Sections 3.2 and 3.3. Fig 2 shows the process of our work. The music aesthetic measure formula is as follows:
\begin{equation}
Aesthetic\ Measure=\frac{\omega_1H + \omega_2S + \theta_1}{\omega_3E + \omega_4K + \theta_2} \label{6}
\end{equation}
Where $H$ is harmony, $S$ is symmetry, $E$ is entropy and $k$ is K complexity. $\omega$ is the weight and $\theta$ is the constant.
\subsection{Order Measures}
When objects have some characteristics of harmony, symmetry or order, they often have a certain sense of beauty. We quantify the order of music in two dimensions, harmony and symmetry. Harmony mainly calculates based on music theory knowledge, while symmetry mainly relies on some statistical information in music. Next, we use the linear combination of harmony and symmetry as the measure of order.
\subsubsection{Interval Harmony}
In music, the distance between two notes is called interval. In particular, in music, when an interval is 12 semitone, we call it an octave. Interval classification can be found in the supplementary material, interval are divided into five categories.
Mathematical and physical research shows that when two sound frequencies are a simple integer ratio, it is more pleasant to listen together. Therefore, we propose a calculation method of interval harmony, and the formula is as follows.
\begin{equation}
Interval\ Harmony=\sum_{i=1}^{12} \alpha_i * pir_i + \theta_{ih} \label{1}
\end{equation}
Among them, $\alpha_i$ is the weight of interval, $pir_i$ is the ratio of interval to total interval, $\theta_{ih}$ is the constant.
\subsubsection{Chord Progression Harmony}
According to Schoenberg's theory of harmony \cite{schoenberg1983theory}, the internal chord is divided into three functional harmonies: tonic triad (T), subdominant triad (S) and dominant triad (D). An example of chord functions and series are shown in supplementary material.
A complete harmony progression starts from the tonic triad, proceeds to the subordinate triad, proceeds to the dominant triad, and finally returns to the tonic triad to complete a complete cycle, which is called complete progression. In the usual sense, harmony progression is the connection of chords within a certain harmonic range in tonal music.
There are many ways to quantify harmony progression. In this paper, we refer to the method of María \cite{navarro2020computational}. In our work, we took the average value of the progression tension to obtain a quantitative chord progress harmony. It is calculated by referring to the following formula:
\begin{equation}
\begin{split}
Chord\ Progression\ Harmony = \lambda_1 d_1(T_i,T_{i-1})+ \\
\lambda_2 d_2(T_i, T_{key}) + \lambda_3 d_3(T_i-T_{key}, T_f) + \lambda_4 c(T_i) + \\
\lambda_5 m(T_i,P) + \lambda_6 h(T_i,P)
\end{split}
\end{equation}
Where $T_i$ is the i-th chord of progression $P$, $\lambda$ is the weight. For more information of parameters $c$, $m$, $h$, see \cite{navarro2020computational}.
\subsubsection{Self Similarity Fitness}
In the field of music generation, structure is often discussed as an important feature. Almost all music contains repetitive pieces. We will discuss the influence of repetitive structure on music aesthetics. We measure it with self similarity fitness.
Inspired by the aesthetics of images and art \cite{al2017symmetry}, the aesthetics beauty of music comes from the symmetry in musical compositions. Therefore, in our study, we refer to Müller's fitness method \cite{MuellerJG13_StructureAnaylsis_IEEE-TASLP} to measure the degree of repetition in a piece of music. The fitness formula is shown as follow:
\begin{equation}
Self\ Similarity\ Fitness = 2 \cdot \frac{\bar{\sigma}(\alpha) \cdot \bar{\gamma}(\alpha)}{\bar{\sigma}(\alpha)+\bar{\gamma}(\alpha)}
\end{equation}
Both $\bar{\sigma}(\alpha)$ and $\bar{\gamma}(\alpha)$ are related to a concept defined by Müller's method \cite{MuellerJG13_StructureAnaylsis_IEEE-TASLP}.
\subsubsection{Skewness}
Skewness is a concept proposed by jSymbolic \cite{mckay2018jsymbolic}. It proposes that both pitch and rhythm of music have the concept of skewness. The notes in music cannot lack pitch and rhythm. Skewness describes how asymmetrical the pitch / rhythm is to either the left or the right of the mean pitch / rhythm value.
The features are extracted based on jSymbolic, the calculation formula of pitch and rhythm skewness are not specifically described here. We combine pitch skewness and rhythm skewness linearly to get the formula of skewness:
\begin{equation}
Skewness = \beta_1 * PS + \beta_2 * RS + \theta_{sk}
\end{equation}
Where $PS$ is pitch skewness, $RS$ is rhythm skewness, $\beta$ represents their weight and $\theta_{sk}$ is a constant.
\subsection{Complexity Measures}
Bense \cite{bense1960programmierung} first uses Birkhoff's aesthetic measure formula to calculate aesthetics. They adapt statistical measure of information in aesthetic objects and believe that the objective measure of aesthetic objects is related to the complexity of objects. Their idea has to use information theory, and entropy is the core of it. Our aesthetic measure method considers two features: Shannon entropy and Kolmogorov complexity.
\subsubsection{Shannon Entropy}
Let $\Omega$ be a finite set, and $X$ be a random variable. The value $x$ in $\Omega$ has a distribution $p(x)=Pr[X=x]$. The Shannon entropy $H(X)$ of random variable $X$ is defined as follows:
\begin{equation}
H(X)=-\sum_{x\in\Omega}p(x)\log(x)
\end{equation}
The Shannon entropy $H(X)$ measures the average uncertainty of random variable $X$, which is widely used to evaluate the degree of chaos in the internal state of a system. In order to calculate the entropy of music, it is necessary to obtain the music attribute histogram.
Pitch and rhythm are the two basic elements of music. In our method, we consider pitch entropy and rhythm histogram entropy, and take their linear combination as the measure of entropy. This is used to describe the uncertainty in music. Our entropy formula is as follows:
\begin{equation}
Entropy=\eta_1 * PHE + \eta_2 * RHE + \theta_e
\end{equation}
Where $PHE$ and $RHS$ are pitch and rhythm histogram entropy, $\eta$ represents their weight and $\theta_e$ is a constant.
\subsubsection{Kolmogorov Complexity}
For a string $s$, Kolmogorov complexity $K(s)$ of the string $s$ refers to the shortest program to calculate the string $s$ on a computer. In essence, the Kolmogorov complexity of a string is the length of the final compressed version of the string. Then, we use the linear combination of entropy and Kolmogorov complexity as a measure of complexity.
Aesthetically speaking, redundancy makes people feel dull, resulting in negative emotions. According to the definition of Kolmogorov complexity, we also refer to the method of using Kolmogorov complexity in image aesthetics \cite{rigau2008informational}.
We believe that Kolmogorov complexity in music is also computable. It is actually the lossless compression ratio of music, which can be formalized as the following formula:
\begin{align}
Kolmogorov\ Complexity=\frac{NH_{m} - K}{NH_{m}} \label{1}
\end{align}
Where $NH_{m}$ is information content of a music, and $K$ is the simplest music information after compression.
\begin{figure*}[htbp]
\centering
\includegraphics[width=1\textwidth]{score}
\caption{As can be seen from the figure, the melody of composer score is in order, and it matches the chord progression very well. The melody of AI score seems very random and low-quality, which is specifically reflected in the irregular appearance of the rest and the melody do not follow the tension of the chord progression. \textbf{Note: The full score of normalized measure is 1.}}
\end{figure*}
\section{Implementation}
\subsection{Datasets}
There are many datasets of pop music, such as POP909 \cite{wang2020pop909} and Wikifonia\footnotemark{}. But we don't use them due to some reasons.
All music data formats in POP909 dataset are performance midi, which will cause different performers' aesthetic impact on the same music score. Although Wikifonia is a score dataset (single track), it can't be used by order measure.
\footnotetext{http://www.wikifonia.org/}
In order to eliminate the performance differences of different performers, objectively measure the aesthetic value of a piece of music, we finally download the scores of 100 pop songs on the Musescore\footnotemark{} website. Then we extract the chords of music scores in Musescore for generating 100 music scores. In datail, we take chord progression from the scores of Musescore, and determine scores' key signature (tonic) as the input parameter, then give them to Magenta's inprov\_rnn \cite{inprovrnn} enables it to generate music according to chord progression and key signature, which has the advantage of controlling the time length of two pairs of datasets to be approximately the same. The scores in Musescore are created by composers as positive samples, while the scores created by Magenta are generated by AI as negative samples.
\footnotetext{https://musescore.org/}
Since we use cross validation, we do not set validation sets. We split the dataset at a ratio of 7:3 for training testing.
\subsection{Preprocessing \& Computing Aesthetic Features}
We use music21 \cite{cuthbert2010music21} toolkit to load music scores. Music21 has note and chord attributes, which can easily obtain the information of music score and make calculations.
Firstly, we obtain the pitch histogram and rhythm histogram in the music score to calculate entropy. Then, we get all the intervals by calculating the note events that occur at the same time, and obtain a histogram. So, the histogram entropy of pitch and rhythm and interval harmony can be calculated.
Secondly, we try to get all the chords and the key signature in the score to calculate the chord progress harmony. The chord progress and the key signature will be saved in the json format and subsequently input to the pretrained model.
Thirdly, we use Musescore3 to batch process composer scores and AI scores into score midi. This is to facilitate the use of jSymbolic \cite{mckay2018jsymbolic} to extract features. The conversion of xml to score midi will not lose the score information. In this way, we get pitch and rhythm skewness.
Fourthly, we use Musescore3 to render the music score into audio, which controls that all music scores are played by Musescore3 to ensure that the music score information is lossless. Then we use the wav file format to calculate the self similarity fitness. Then, we refer to Monkey's Audio's\footnotemark{}\footnotetext{https://www.monkeysaudio.com/} lossless compression method to compress music into ape format, so as to calculate Kolmogorov complexity.
Fifthly, calculating the values of 8 basic music features (refer to Table 1), we will normalize them. Next, we use the method of logical regression to confirm the values of Harmony, Symmetry, Entropy, and Kolmogorov complexity.
Finally, we take the four normalized aesthetic features as inputs, use sigmoid function to map the aesthetic measure to 0 and 1 for classification, and use cross-entropy loss to make loss function, confirming the parameters of the aesthetic model in the way of gradient decline. We set the learning rate to 0.01. After 1000 iterations, the loss function converges.
\section{Experiments}
\subsection{Turing Test}
We need to verify whether people really think the music created by composers is more beautiful than the music generated by AI. So we did the Turing test to see if people can really distinguish between composer music and AI music.
We randomly sampled 10 pieces from the music created by the composer and the music generated by AI respectively, with each piece lasting about 15 seconds. Volunteers participating in the Turing test need to identify which music is created by the composer and which music is generated by AI in these 10 pairs of pieces. The volunteer also needs to choose the one in each pair that he thinks is more beautiful.
A total of 15 volunteers participate in the Turing test. Among 300 samples, their classification accuracy of music is 91.3\% (274). Assuming that the music created by the composer is more aesthetic, 87\% (261) of the samples are correctly classified. This proves that our assumption is correct.
\subsection{Results \& Discussion}
Since the music created by the composer has a higher aesthetic feeling than the music generated by AI, we let the machine learn the aesthetic score according to the label value. This is essentially a binary classification problem.
\begin{figure}[!htbp]
\centering
\includegraphics[width=0.352\textwidth]{distribution1}
\caption{The intersection of the distributions is quite small, showing that our model can distinguish between AI scores and composer scores very well.}
\label{fig:distribution}
\label{fig:distribution}
\end{figure}
After the weight is obtained by gradient descent, we bring the weight into the aesthetic model to view the distribution of the aesthetic model. The distribution is shown in Fig 4.
We use precision and F1-Measure as our metric to test our model. The precision of our model on the test set is 93.3\%, and the F1-Measure is 90.9\%. This proves that our model is valid. Fig 3 shows an example of score comparison.
\begin{table}[htbp]
\centering
\resizebox{1\columnwidth}{!}{
\begin{tabular}{c|c|c|c|c|c|c|c|c}
\hline
\textbf{Dataset} & \textbf{IH} & \textbf{CPH} & \textbf{SSF} & \textbf{PS} & \textbf{RS} & \textbf{PHE} & \textbf{RHE} & \textbf{KC} \\ \hline
AI & 0.97 & 1.98 & 0.06 & \textbf{0.49} & 2.16 & \textbf{1.34} & \textbf{1.60} & \textbf{0.69} \\ \hline
Composer & \textbf{1.40} & \textbf{2.04} & \textbf{0.19} & 0.53 & \textbf{0.99} & 2.16 & 1.79 & 0.73 \\ \hline
\end{tabular}
}
\caption{The similar CPH value shows the rationality of the AI dataset generated based on chords. \textbf{More detailed distribution comparison can be found in the supplementary material.} The results in bold have higher aesthetic scores.}
\end{table}
Table 1 shows the calculation results of 8 features without normalization. We can make the following analysis from it. Whether IH or CPH, the scores of composer is obviously better than that of AI. This is also in line with the music theory. Considering Symmetry, the SSF of composer is obviously higher than that of AI. This is because the music created by AI is too random and often has no repetitive pieces, which proves beauty is related to repetition. Although the PS of AI is smaller than that of the composer, the difference is not much. The RS of the composer is obviously smaller than that of AI. This is because composers tend to create more regular rhythm and pitch than AI. When it comes to entropy, the values of PE and HE are obviously higher in composer score than in AI score. Although this is contrary to Birkhoff's aesthetic measure, it is also reasonable. This is because composer tends to add some changes to pitch and rhythm, while music generated by AI does not change much around the tonic. As for K-complexity, the difference between composer's scores and AI's scores is not significant, but it can be seen from the table that the music created by AI is relatively low in compression and complexity. In conclusion, moderate order and complexity quotient can quantify aesthetic feeling to a certain extent.
\subsection{Ablation Study}
We conduct ablation experiments to remove harmony, symmetry, entropy and K-complexity respectively to train four different models. We compare them with the original model.
As shown in Fig 5, we observe the ROC curves of the original model and four models with one aesthetic feature removed respectively, and obtain their AUC values. The AUC value of our model is 0.93, which is obviously higher than that of the four models without aesthetic features.
If harmony is removed, the AUC value is only 0.77, which shows the importance of harmony and further proves that music theory plays a very important role in music aesthetics. If symmetry is removed, the AUC value of the model is 0.85, which indicates that symmetry may contribute slightly less to aesthetics than harmony. As for entropy, it is obviously important, but K-complexity seems not.
\begin{figure}[!htbp]
\centering
\includegraphics[width=0.38\textwidth]{Ablation}
\caption{The ROC curves are serrated due to lack of samples.}
\label{fig:ablation}
\label{fig:ablation}
\end{figure}
\section{Conclusion}
In summary, we propose a score aesthetic assessment model using Birkhoff's aesthetic measure to quantify a score's aesthetic. We also discover four categories of music aesthetic features, totaling eight basic aesthetic features. We have made some contributions to improve the quality of music scores. This might be helpful for music score quality assessment. However, our method still has shortcomings, for instance we have not taken the relationship between creativity and musical aesthetics into consideration. The aesthetic study of music audio quality assessment is worth exploring in the future.
\bibliographystyle{IEEEbib}
|
\section{Introduction}
\IEEEPARstart{P}{redicting} the depth from only a single image is particularly challenging, as this problem is inherently ambiguous~\cite{persson2021parameterization}.
Current data-driven methods predict depth by learning implicit structural and semantic information that provides enough monocular cues for depth estimation.
This requires the network to look at bigger parts of the scene~\cite{mertan2022single},
which has been addressed by many works~\cite{huynh2020guiding,bhat2021adabins} in the field of single-view depth estimation in recent years.
Previous work~\cite{huynh2020guiding} tackles the problem of extending receptive field by adding a non-local depth-attention module between encoder and decoder, but training this module requires additional supervision.
Adabins~\cite{bhat2021adabins} uses a post-processing module composed by vision transformer to refine the output of traditional encoder-decoder architecture, but the receptive field of ``traditional" parts of the network has not been enlarged.
Ranftl\emph{et al}\onedot~\cite{ranftl2021vision} introduce vision transformer as its feature extraction module,
which gives feature maps global receptive field at every stage, resulting in substantial improvements compared with fully-convolutional network~\cite{ranftl2021vision}.
However, the ability of encoding non-local geometric dependencies is still limited as it uses a traditional convolutional network as a decoder.
Panorama depth estimation method SliceNet~\cite{pintore2021slicenet} uses an RNN network to enhance its capability of retrieving global geometric information, but it requires additional constraints that panorama images are gravity aligned.
\begin{figure}
\centering
\includegraphics[width=0.99\linewidth]{figure/framework4.png}
\caption[Pipeline of the proposed method.]
{Pipeline of the proposed method. \textcolor{lightblue}{Light blue} parts work on planar ERP images,
while \textcolor{greyblue}{grey blue} parts work on unit spherical surface.
$\mathbf{f}_{{\Pi}\rightarrow{\mathbf{S^{2}}}}$ denotes mapping from image plane to unit sphere and $\mathbf{f}_{{\mathbf{S^{2}}}\rightarrow\Pi}$ denotes mapping from unit sphere to image plane.}
\label{fig:1}
\end{figure}
Besides challenges on extending receptive field, for panorama images, additional challenges arise from its unique \textbf{Equi-Rectangular Projection (ERP)} image representation.
This introduces severe distortion and networks directly working on ERP images are biased by these distortion towards the poles~\cite{albanis2021pano3d}, as the pixels from ERP images are unevenly distributed on the corresponding spherical surface.
Also, directly learning from ERP image representation cannot fulfill the implicit assumption of panorama images that the left side and the right side of the image are physically connected,
which results in spatial inconsistency on the depth map~\cite{sun2021indoor}.
Some previous methods~\cite{wang2020bifuse,jiang2021unifuse} employ an additional branch where cubemaps generated from panorama images are introduced to compensate for the distortion.
However, discontinuity happens on seams between different sides of cubemaps, thus the distortion on the ERP image cannot be fully compensated.
Other methods~\cite{zhuang2021acdnet,zioulis2018omnidepth,li2022omnifusion} take another way by using irregular convolution kernel or dilated convolution,
but their distortion removal is limited by the learning ability of the network.
In addition, all of the above methods need circular padding to ensure spatial consistency of the depth map, but the problem cannot be fully solved.
In a word, these problems require the panorama depth estimation method can represent ERP images in a distortion-reduced manner in which the pixels are evenly distributed, and possesses large receptive field that can better explore embedded global structural information.
Inspired by previous works on spherical CNNs~\cite{du2021spherical,cohen2018s2cnn},
we move the panorama depth estimation task onto spherical surface generated from ERP image using sampling method HEALPix~\cite{gorski2005healpix},
which guarantees that the sampled pixels distribute uniformly on sphere thus can largely reduces distortion.
Additional advantages is that the spherical representation naturally guarantees spatial consistency of the two sides of panorama images,
which reduces the need for remedial methods like circular padding.
We also introduce \textbf{Cross Attention Fusion (CAF)} module that utilizes learnable cross attention to fuse feature maps from skip connection and to enhance the capability of the decoder to retrieve global context information.
The pipeline of the proposed method can be found in Figure~\ref{fig:1}.
To the best of our knowledge, the proposed method is the first one that employs a vision transformer on the unit sphere in real-world vision tasks.
The proposed method outperforms previous methods significantly on nearly all public datasets,
especially on new Pano3D~\cite{albanis2021pano3d} datasets which focuses on method's generalization capablity:
the $RMSE$ error decreases $\mathbf{23.3}\%$ and $AbsRel$ error decreases $\mathbf{16.3}\%$.
Our method also obtains good results in spatial consistency and robustness to misalignment on gravity direction.
\textbf{Main contributions of this paper} are:
\begin{enumerate}
\item A novel end-to-end deep network that moves decoder network onto spherical surface
which eliminates the distortion of ERP images and maintains the spatial consistency of depth maps.
\item An spherical-transformer-based fusion module (CAF) that fuses features from skip connection in a learnable manner
and enhances the capability of retrieving global context information.
\end{enumerate}
The paper is organized as follows:
Section~\ref{sec:2} reviews and discusses related works;
Section~\ref{sec:3} contains the necessary details of the proposed method;
Section~\ref{sec:4} demonstrates experiments and ablation studies;
Section~\ref{sec:5} concludes the paper.
\section{Related Works}
\label{sec:2}
\paragraph{Depth estimation for perspective camera images}
new methods use ideas such as knowledge distillation~\cite{wang2021knowledge}, attention-based methods~\cite{bhat2021adabins,chen2021attention} and self-supervision~\cite{watson2021temporal}.
Extending the global context sensing capability of the network is also a focused research topic.
Methods~\cite{bhat2021adabins,huynh2020guiding} try to extend the global context of their method by adding vision transformer based processing module into the network, but the former work's~\cite{bhat2021adabins} traditional part of network's global context have not been enlarged, and the latter work~\cite{huynh2020guiding} requires additional supervising for the module.
Previous work~\cite{ranftl2021vision} adds attention-based modules as its feature extraction module, but its decoder remains to be the traditional structure.
\paragraph{Depth estimation for panorama images}
some methods~\cite{wang2020bifuse,jiang2021unifuse} use cubemaps as the image representation method,
and try to use information derived from cubemaps to compensate for ERP image's distortion,
which results may be affected by discontinuity of cubemaps on each side's edges.
There also exist methods~\cite{zhuang2021acdnet,zioulis2018omnidepth} use specialized convolution methods such as dilated convolution to tackle the distortion on ERP images and enlarge the receptive area of the network.
These methods do not use distortion-reduced methods to represent ERP images,
and hence results may be affected.
Some methods try to represent ERP images with 1D representations~\cite{pintore2021slicenet,sun2021hohonet};
however, distortion on latitude direction may not be fully removed by 1D representation.
Compared to these previous works, the proposed method uses distortion-reduced representation on the sphere, which is not affected by discontinuity from cubemap, and the receptive field of the proposed method can be further enlarged by vision methods.
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{figure/healpix.png}
\caption[HEALPixel.]
{We demonstrate the multi-scale sampling capability of HEALPix by two examples with different $n_{size}$: $n_{size}=4$ on the left and $n_{size}=8$ on the right. }
\label{fig:2}
\end{figure}
\paragraph{Spherical deep learning} most of these methods use dilated convolution or convolution kernel with special forms~\cite{Su2020KTN,fernandez2019CFL},
or compensate distortion using latitude and longitude~\cite{Wang2020360SDNET,Liu18CoordConv} in panorama image understanding tasks.
There are two works that are similar to the proposed method:
one work~\cite{cho2022spherical} utilizes the Icosahedron sampling method~\cite{Lee2022SpherePHD} and forms a novel transformer for representation learning of ERP images.
Another work~\cite{du2021spherical} utilizes HEALPix~\cite{gorski2005healpix} as the sampling method and proposes a CNN method for this unique pixel representation.
Compared to the above-mentioned methods, our method is the first with complete infrastructure for vision transform techniques used on the unit sphere, and the accuracy of panorama depth estimation using this method is significantly improved.
\section{Approach}
\label{sec:3}
We use the classic encoder-decoder deep network architecture with skip-connections.
Multi-scale feature maps extracted from the input ERP image by a vision transformer encoder network are transitioned from their planar representation to a uniform spherical pixel representation.
Then the multi-scale decoder network performs depth regression along the entire surface of the sphere to predict depth estimation results.
The multi-scale decoder network consisted by several fusion modules at different scales on the unit sphere is used to fuse the spherical feature maps from the skip connection which contains fine-grained local details learned in the encoder part.
\subsection{Panorama Image Representation}
\label{sec:3.1}
We use HEALPix (the Hierarchical Equal Area isoLatitude Pixelization)~\cite{gorski2005healpix} to generate uniformly distributed multi-scale pixels on the unit sphere;
for ${n}_{size}$ samples per-level,
the total number of sampled pixel would be ${n}_{pix}=12 * {{n}_{size}}^2$.
We use 4 coarse-to-fine spherical levels which pixel numbers are $[192, 768, 3072, 12288]$.
Exampled spherical grids can be found in Figure~\ref{fig:2}.
It is worth noticing that either using fewer (downsample) or more (upsample) pixels in HEALPix will not bring positive effects on final results. Tests can be found in Table~\ref{tab:4.0}.
We upsample and downsample the pixel number to $[768, 3072, 12288, 12288]$ and $[48, 192, 768, 3072]$ respectively, to study its effects on final accuracy results.
Due to limited GPU memory, we do not upsample the finest level.
The proposed pixel number achieves overall best accuracy results,
the reason is probably two-fold:
less pixel brings larger area of one particular pixel, which carries more global context information but less details;
more pixel helps the network retrieve details in depth map, but will lose global context information.
The proposed pixel number seems to be a good bargain between these two factors.
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{figure/circularpadding8.jpg}
\caption[Compare point cloud.]
{We compare the spatial consistency of UniFuse (\textbf{with circular padding})~\cite{jiang2021unifuse} and the proposed method. Methods like UniFuse cannot guarantee spatial consistency of the left and right end of the ERP images, while the proposed method solves this problem.}
\label{fig:4}
\end{figure}
\begin{table}
\centering
\scriptsize
\caption{Comparison of effects of different HEALPix pixel numbers on final depth map accuracy on M3D dataset~\cite{chang2017matterport3d}.}
\label{tab:4.0}
\begin{tabular}{c|c|c|c|c}
\hline\noalign{\smallskip}
\multirow[c]{2}{*}{Pixel number} & \multicolumn{3}{c|}{Error $\Downarrow$} & Accuracy (\%) $\Uparrow$ \\
& ${AbsRel}$ & ${RMSE}$ & ${RMSLE}$ & $\delta < 1.25$ \\
\noalign{\smallskip}\hline
\hline\noalign{\smallskip}
Downsample & 0.0902 & 0.4211 & 0.0582 & 92.07 \\
Original & \tablebestresult{0.0865} & \tablebestresult{0.4052} & \tablebestresult{0.0559} & \tablebestresult{92.63} \\
Upsample & 0.0866 & 0.4109 & 0.0565 & 92.54 \\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
HEALPix helps define two mappings between the surface and the image plane:
the mapping from image plane to unit sphere $\mathbf{f}_{{\Pi}\rightarrow{\mathbf{S^{2}}}}$ and mapping from unit sphere to image plane $\mathbf{f}_{{\mathbf{S^{2}}}\rightarrow\Pi}$,
which can be pre-computed using open-source software\footnote{\url{https://healpy.readthedocs.io/en/latest/}};
the pre-computed mapping can be used to formulate neighbourhood relation look-up table for each pixel,
which ensures easy implementation of self-attention calculation.
We use bilinear interpolation to calculate pixel color for each pixel on HEALPix sphere,
using other interpolation methods nearly makes no difference to final accuracy.
The encoder operates on the original ERP image,
and the generated feature maps with different channels are mapped onto the spherical surface with unified channel $D\in\{256, 320, 512\}$ using HEALPix.
Unit sphere-based image representation does not have boundaries,
thus operations on the unit sphere is naturally circular,
which makes the proposed method better at maintain spatial consistency on the left side and right side of ERP images, comparing with methods like UniFuse~\cite{jiang2021unifuse} with circular padding,
as both the receptive field and the padding’s width are limited. See Figure~\ref{fig:4} for one vivid example.
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{figure/CAF.png}
\caption[Compare point cloud.]
{Structure of CAF module that conducts self-attention calculation on sphere.
``LN", ``SPE", and ``ResCon" denote Layer Normalization, Spherical Positional Embedding (Section~\ref{sec:3.4}), and Residual Convolution, respectively. }
\label{fig:4.1}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=0.85\linewidth]{figure/depthmapresult5.jpg}
\caption[Result of depth map generated.]
{We collect a series of panorama images at different spots in a real-world scene,
and estimate a depth map for each image.
These depth maps are then registered to one global coordinate system by Structure from Motion.
We demonstrate the final mesh model generated from alignment results on the left side of the figure,
and depth maps for three exampled spots in the scene and
corresponding point clouds on the right side. }
\label{fig:3}
\end{figure*}
\subsection{Cross Attention Fusion}
\label{sec:3.3}
Multiple \textbf{Cross Attention Fusion (CAF)} modules that operates on multi-scale spherical features formulate the multi-scale decoder.
CAF is designed to enhance the network's global context acquisition capability and to fuse feature maps from skip connection to enhance the details of estimated depth map.
CAF adopts a ``global" cross attention calculation mechanism, which means that correlations are computed among pixels of feature maps at different positions globally.
In this way, CAF has the ability of long-dependencies and can learn better global context information from multi-path feature maps. Features from skip-connections contain ``local" fine-grained details as they come from pyramidal encoder network,
while the features from the decoder are relatively ``global" and possess larger receptive field.
Previous global-local fusion strategies mostly adopt trivial sum or concatenate,
which directly mixes up these feature maps with different contexts.
Our CAF module uses a learnable global attention based fusion strategy:
feature maps from skip connection with local context are treated as ``guidance" to enhance feature maps from decoder with global context, and vice versa.
In this way, CAF enhances the correlation between two feature maps by learning compensation values. The detailed structure of the CAF module can be found in Figure~\ref{fig:4.1}.
We use \textit{Base} to denote fusion module with direct addition, \textit{Fusion} to denote fusion module with our CAF. The \textit{Base} fusion module is only consisted of two simple residual convolution units from work~\cite{ranftl2021vision} using basic spherical convolution method in previous works~\cite{du2021spherical,cho2022spherical}.
\begin{figure*}
\centering
\includegraphics[width=0.85\linewidth]{figure/exhibit4.jpg}
\caption[Result of depth map generated.]
{Comparison between the proposed work and UniFuse~\cite{jiang2021unifuse} on several selected scenes from the M3D dataset. We use boxes to mark places worth noticing. Note that we use different colors to distinguish different noticeable places.}
\label{fig:3.1}
\end{figure*}
\begin{table}
\centering
\scriptsize
\caption{Comparison of effects of formulation of coordinates used in SPE on final depth map accuracy on M3D dataset~\cite{chang2017matterport3d}.
We use $[lat,lon]$ to represent SPE using point's longitude and latitude information.}
\label{tab:4.3}
\begin{tabular}{c|c|c|c|c}
\hline\noalign{\smallskip}
\multirow[c]{2}{*}{Coordinates} & \multicolumn{3}{c|}{Error $\Downarrow$} & Accuracy (\%) $\Uparrow$ \\
& ${AbsRel}$ & ${RMSE}$ & ${RMSLE}$ & $\delta < 1.25$ \\
\noalign{\smallskip}\hline
\hline\noalign{\smallskip}
$[x,y]$ & 0.0910 & 0.4238 & 0.0589 & 91.83 \\
$[y,z]$ & 0.0900 & 0.4202 & 0.0581 & 92.21 \\
$[x,z]$ & 0.0874 & 0.4162 & 0.0570 & 92.60 \\
$[lat,lon]$ & 0.0954 & 0.4416 & 0.0616 & 90.92 \\
$[x,y,z]$ & \tablebestresult{0.0865} & \tablebestresult{0.4052} & \tablebestresult{0.0559} & \tablebestresult{92.63} \\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
\begin{table}
\centering
\scriptsize
\caption{Comparison of effects of position embedding methods on final depth map accuracy on M3D dataset~\cite{chang2017matterport3d}}
\label{tab:6}
\begin{tabular}{c|c|c|c|c}
\hline\noalign{\smallskip}
\multirow[c]{2}{*}{Encoding method} & \multicolumn{3}{c|}{Error $\Downarrow$} & Accuracy (\%) $\Uparrow$ \\
& ${AbsRel}$ & ${RMSE}$ & ${RMSLE}$ & $\delta < 1.25$ \\
\noalign{\smallskip}\hline
\hline\noalign{\smallskip}
Learnable & 0.0881 & 0.4142 & 0.0570 & 92.369 \\
SPE & \tablebestresult{0.0865} & \tablebestresult{0.4052} & \tablebestresult{0.0559} & \tablebestresult{92.634} \\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
Given feature map $F_{i}$ from skip-connection and decoder network and positional embedding ${\zeta}_{i}$, the input of CAF can be defined as:
\begin{equation}
\begin{aligned}
X_{i} = LN(F_{i} + {\zeta}_{i}), i \in \{0, 1\}
\end{aligned}
\label{eq:1}
\end{equation}
CAF module tries to estimate two compensation terms ${Att}_{0}$ and ${Att}_{1}$ by cross-attention:
\begin{equation}
\begin{aligned}
{{Att}_{0}}(Q_{1}, K_{0}, V_{0}) &= {SoftMax}(Q_{1}K_{0} / \sqrt{d})V_{0} \\
{{Att}_{1}}(Q_{0}, K_{1}, V_{1}) &= {SoftMax}(Q_{0}K_{1} / \sqrt{d})V_{1}
\end{aligned}
\label{eq:2}
\end{equation}
the two values are added into inputs:
\begin{equation}
\begin{aligned}
X_{CA} &= {FFN}({LN}(X_{0} + {Att}_{0} + X_{1} + {Att}_{1})) \\
\end{aligned}
\label{eq:3}
\end{equation}
$FFN$ denotes feed-forward operation, and $LN$ denotes layer norm operation~\cite{cho2022spherical}, the two operations formulate a residual-like structure.
$Q_{i}$, $K_{i}$ and $V_{i}$ used in Equation~\ref{eq:2} are generated from input ${X}_{0}$ and ${X}_{1}$ using corresponding learnable parameters $W$:
\begin{equation}
\begin{aligned}
Q_{i} = W^{Q}X_{i}, K_{i} = W^{K}X_{i}, V_{i} = W^{V}X_{i}, i \in \{0, 1\}
\end{aligned}
\label{eq:4}
\end{equation}
As the global attention calculation requires large computing resources, we adopt the method used in previous work~\cite{chu2021twins} to perform global sub-sampled cross attention utilizing non-overlapped window partitions.
These partitions can be conveniently derived from the coarse-to-fine structure of HEALPix.
\begin{table}
\centering
\scriptsize
\caption{Comparison of effects of different loss function on final depth map accuracy on M3D datasets~\cite{chang2017matterport3d}.
We use ``RMSERel'' to represent Relative RMSE~\cite{Alvarez-Gila2017Adversarial}.}
\label{tab:4.1}
\begin{tabular}{c|c|c|c|c}
\hline\noalign{\smallskip}
\multirow[c]{2}{*}{Loss Function} & \multicolumn{3}{c|}{Error $\Downarrow$} & Accuracy (\%) $\Uparrow$ \\
& ${AbsRel}$ & ${RMSE}$ & ${RMSLE}$ & $\delta < 1.25$ \\
\noalign{\smallskip}\hline
\hline\noalign{\smallskip}
RMSERel & 0.0943 & 0.4200 & 0.0597 & 91.49 \\
RMSLE & \tablebestresult{0.0865} & \tablebestresult{0.4052} & \tablebestresult{0.0559} & \tablebestresult{92.63} \\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
\subsection{Spherical Positional Embedding}
\label{sec:3.4}
In order to make the CAF module better obtains pixel position information, we proposed a handcrafted positional embedding method, namely \textbf{Spherical Positional Embedding (SPE)}.
We concatenate the spherical coordinates along the channel axis, and then embed the three-channel input to the dimension of the feature maps:
\begin{equation}
\begin{aligned}
\zeta = C\times{[x_{sphere},y_{sphere},z_{sphere}]}^{T}
\end{aligned}
\label{eq:4}
\end{equation}
where $C\in\mathbb{R}^{M\times3}$ is an embedding matrix, $M$ is the channels of the feature map.
Reducing the dimensions of coordinate being used, or using latitude and longitude to form positional embedding,
will reduce the overall accuracy of the proposed method;
corresponding tests can be found in Table~\ref{tab:4.3}.
Large accuracy decrease when using latitude and longitude is probably because same area of pixel produce different changes in longitude when the pixel is placed on different latitude,
which may introduce implicit distortion into final results.
Compared with the learnable positional embedding mechanism, this embedding method utilizes the real space location on the spherical surface,
which helps the decoder network and CAF module take into account non-Euclidean spatial information embedded on the sphere, yielding better results.
Results in Table~\ref{tab:6} shows comparison between the two kinds of embedding methods, and the proposed SPE method produces better results, which $RMSE$ is decreased by $\mathbf{2.2}\%$.
\begin{table*}
\centering
\scriptsize
\caption{Comparison of our method with currently available methods on M3D datasets~\cite{chang2017matterport3d} and S2D3D datasets~\cite{armeni2017jointstanford2d3d}.
We use $^*$ to label results evaluated after performing scale alignment by medium or maximum depth.
${\triangle}$ denotes methods that does not contain optimization against distortion on panorama images.
Note that we use error and accuracy results of SliceNet~\cite{pintore2021slicenet} from ACDNet's paper~\cite{zhuang2021acdnet} in this table.}
\label{tab:1}
\begin{tabular}{c|c|c|c|c|c|c|c|c}
\hline\noalign{\smallskip}
\multirow[c]{2}{*}{Dataset} & \multirow[c]{2}{*}{Method} & \multicolumn{4}{c|}{Error Metric $\Downarrow$} & \multicolumn{3}{c}{Accuracy Metric (\%) $\Uparrow$} \\
& & ${MAE}$ & ${AbsRel}$ & ${RMSE}$ & ${RMSLE}$ & $\delta < 1.25$ & $\delta < {1.25}^{2}$ & $\delta < {1.25}^{3}$ \\
\noalign{\smallskip}\hline
\hline\noalign{\smallskip}
\multirow[c]{11}[3]{*}{M3D} & FCRN$^*$~\cite{laina2016fcrn} & 0.4008 & 0.2409 & 0.6704 & 0.1244 & 77.03 & 91.74 & 96.17 \\
& SliceNet$^*$~\cite{pintore2021slicenet} & 0.3296 & 0.1764 & 0.6133 & 0.1045 & 87.16 & 94.83 & 97.16 \\
& \textbf{Ours (base)}$^*$ & 0.2025 & 0.0776 & 0.3787 & 0.0513 & 93.58 & 98.31 & 99.44 \\
& \textbf{Ours (fusion)}$^*$ & \aligntablebestresult{0.1932} & \aligntablebestresult{0.0730} & \aligntablebestresult{0.3720} & \aligntablebestresult{0.0498} & \aligntablebestresult{93.96} & \aligntablebestresult{98.38} & \aligntablebestresult{99.46} \\
\cmidrule[1.5pt]{2-9}
& Bifuse~\cite{wang2020bifuse} & 0.3470 & 0.2048 & 0.6259 & 0.1134 & 84.52 & 93.19 & 96.32 \\
& HoHoNet~\cite{sun2021hohonet} & 0.2862 & 0.1488 & 0.5138 & 0.0871 & 87.86 & 95.19 & 97.71 \\
& UniFuse~\cite{jiang2021unifuse} & 0.2814 & 0.1063 & 0.4941 & 0.0701 & 88.97 & 96.23 & 98.31 \\
& ACDNet~\cite{zhuang2021acdnet} & 0.2670 & 0.1010 & 0.4629 & 0.0646 & 90.00 & 96.78 & 98.76 \\
& OmniFusion~\cite{li2022omnifusion} & $\times$ & 0.0900 & 0.4261 & 0.1483 & 91.89 & \tablebestresult{97.97} & \tablebestresult{99.31} \\
& \textbf{Ours (base)} & 0.2346 & 0.0906 & 0.4153 & 0.0577 & 91.92 & 97.55 & 99.12 \\
& \textbf{Ours (fusion)} & \tablebestresult{0.2263} & \tablebestresult{0.0865} & \tablebestresult{0.4052} & \tablebestresult{0.0559} & \tablebestresult{92.64} & 97.68 & 99.11 \\
\cmidrule[1.5pt]{2-9}
& ${\triangle}$AdaBins~\cite{bhat2021adabins} & 0.2560 & 0.0958 & 0.4484 & 0.1426 & 90.54 & 97.05 & 98.92 \\
\midrule[1.5pt]
\multirow[c]{11}[3]{*}{S2D3D} & FCRN$^*$~\cite{laina2016fcrn} & 0.3428 & 0.1837 & 0.5774 & 0.1100 & 72.30 & 92.07 & 97.31 \\
& SliceNet$^*$~\cite{pintore2021slicenet} & 0.1757 & 0.0995 & 0.3509 & 0.0801 & 90.29 & 96.26 & 98.44 \\
& \textbf{Ours (base)}$^*$ & 0.1745 & 0.0919 & \aligntablebestresult{0.3229} & 0.0593 & 92.12 & \aligntablebestresult{98.14} & \aligntablebestresult{99.31} \\
& \textbf{Ours (fusion)}$^*$ & \aligntablebestresult{0.1655} & \aligntablebestresult{0.0843} & 0.3232 & \aligntablebestresult{0.0575} & \aligntablebestresult{92.90} & 98.09 & 99.29 \\
\cmidrule[1.5pt]{2-9}
& Bifuse~\cite{wang2020bifuse} & 0.2343 & 0.1209 & 0.4142 & 0.0787 & 86.60 & 95.80 & 98.60 \\
& HoHoNet~\cite{sun2021hohonet} & 0.2027 & 0.1014 & 0.3834 & 0.0668 & 90.54 & 96.93 & 98.86 \\
& UniFuse~\cite{jiang2021unifuse} & 0.2082 & 0.1114 & 0.3691 & 0.0721 & 88.97 & 96.23 & 98.31 \\
& ACDNet~\cite{zhuang2021acdnet} & 0.1870 & 0.0984 & 0.3410 & 0.0664 & 88.72 & 97.04 & 98.95 \\
& OmniFusion~\cite{li2022omnifusion} & $\times$ & 0.0950 & 0.3474 & 0.1599 & 89.88 & 97.69 & \tablebestresult{99.24} \\
& \textbf{Ours (base)} & \tablebestresult{0.1760} & 0.0936 & \tablebestresult{0.3272} & \tablebestresult{0.0600} & 91.53 & \tablebestresult{98.30} & 99.18 \\
& \textbf{Ours (fusion)} & 0.1774 & \tablebestresult{0.0903} & 0.3383 & \tablebestresult{0.0600} & \tablebestresult{91.91} & 97.82 & 99.12 \\
\noalign{\smallskip}\hline
\end{tabular}
\end{table*}
\subsection{Training Loss}
\label{sec:3.5}
We use Root Mean Squared Logarithmic Error (RMSLE) as the loss function to supervise the training process, as shown in Equation~\ref{eq:7}.
$d_{i}$ and $\overline{d_{i}}$ refer to estimated and ground truth depth for pixel $i$ respectively.
\begin{equation}
\begin{aligned}
\mathcal{L}=\sqrt{\frac{1}{N}\sum_{i=1}^{N}(\log(d_{i})-\log(\overline{d_{i}}))^2}
\end{aligned}
\label{eq:7}
\end{equation}
We optimize for network parameters on all spherical pixels with valid depth value.
RMSLE can reduce the impact of large differences between prediction and ground truth on the overall loss and ensure that the loss is uniformly distributed in different depth ranges.
As the spherical pixels are already distributed uniformly on sphere,
the contribution of each pixel to the overall loss is spatially uniform.
We also notice that \textbf{Relative RMSE (RMSERel)}~\cite{Alvarez-Gila2017Adversarial} can also reduce the impact of large error;
we compare the accuracy results of the two kinds of losses in Table~\ref{tab:4.1}.
Results of RMSERel is worse than proposed RMSLE,
this is probably because RMSLE suppress losses on faraway objects more than RMSERel,
resulting in better spatial measurements on panorama depth estimation tasks,
as most of these tasks focus on indoor \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot near scenarios.
\section{Experiments}
\label{sec:4}
\begin{table*}
\centering
\scriptsize
\caption{Results on Pano3D~\cite{albanis2021pano3d} datasets.
``$w$'' means the error uniformly distribute by re-weighting using weighting methods provided by Pano3D.
Metrics without Pano3D official results are marked with "$\times$".}
\label{tab:3}
\begin{tabular}{c|c|c|c|c|c|c|c|c|c}
\hline\noalign{\smallskip}
\multirow[c]{2}[3]{*}{Dataset} &\multirow[c]{2}[3]{*}{Method} & \multicolumn{4}{c|}{Error Metric $\Downarrow$} & \multicolumn{4}{c}{Accuracy Metric (\%) $\Uparrow$} \\
& & ${\mathit{w}RMSE}$ & ${\mathit{w}RMSLE}$ & ${\mathit{w}AbsRel}$ & ${\mathit{w}SqRel}$ & ${\delta}_{1.1}^{{ico}^6}$ & ${\delta}_{1.25}^{{ico}^6}$ & ${\delta}_{{1.25}^{2}}^{{ico}^6}$ & ${\delta}_{{1.25}^{3}}^{{ico}^6}$ \\
\noalign{\smallskip}\hline
\hline\noalign{\smallskip}
\multirow[c]{4}{*}{\tabincell{c}{\textit{train on M3D} \\ test on GV2 tiny}} & ${\mathrm{UNet}}^{vnl}$ & 0.5794 & 0.1247 & 0.2151 & $\times$ & 31.98 & 62.05 & $\times$ & $\times$ \\
& ${\mathrm{ResNet}}^{comp}_{skip}$ & 0.4993 & 0.1273 & 0.1758 & $\times$ & 40.78 & 80.31 & $\times$ & $\times$ \\
& \textbf{Ours (base)} & \tablebestresult{0.3654} & 0.0800 & 0.1458 & 0.0675 & 45.44 & 80.93 & 96.03 & 99.13 \\
& \textbf{Ours (fusion)} & 0.3682 & \tablebestresult{0.0722} & \tablebestresult{0.1353} & \tablebestresult{0.0615} & \tablebestresult{47.82} & \tablebestresult{85.23} & \tablebestresult{97.45} & \tablebestresult{99.43} \\
\cmidrule[1.3pt]{1-10}
\multirow[c]{4}{*}{\tabincell{c}{\textit{train on M3D} \\ test on GV2 medium}} & ${\mathrm{UNet}}^{vnl}$ & 0.5901 & 0.1291 & 0.2269 & $\times$ & 31.21 & 61.02 & $\times$ & $\times$ \\
& ${\mathrm{ResNet}}^{comp}_{skip}$ & 0.4528 & 0.1618 & 0.1664 & $\times$ & 42.03 & 81.91 & $\times$ & $\times$ \\
& \textbf{Ours (base)} & 0.3502 & 0.0806 & 0.1500 & 0.0684 & 44.78 & 80.67 & 95.72 & 99.00 \\
& \textbf{Ours (fusion)} & \tablebestresult{0.3472} & \tablebestresult{0.0729} & \tablebestresult{0.1393} & \tablebestresult{0.0614} & \tablebestresult{46.95} & \tablebestresult{84.66} & \tablebestresult{97.16} & \tablebestresult{99.33} \\
\cmidrule[1.3pt]{1-10}
\multirow[c]{4}{*}{\tabincell{c}{\textit{train on M3D} \\ test on GV2 fullplus}} & ${\mathrm{UNet}}^{vnl}$ & 0.8772 & 0.1769 & 0.2730 & $\times$ & 22.46 & 46.09 & $\times$ & $\times$ \\
& ${\mathrm{ResNet}}^{comp}_{skip}$ & 0.6607 & 0.2308 & 0.1836 & $\times$ & \tablebestresult{41.18} & \tablebestresult{74.77} & $\times$ & $\times$ \\
& \textbf{Ours (base)} & 0.5560 & 0.1255 & 0.2000 & 0.1430 & 33.27 & 62.20 & 84.66 & 94.05 \\
& \textbf{Ours (fusion)} & \tablebestresult{0.5243} & \tablebestresult{0.1106} & \tablebestresult{0.1804} & \tablebestresult{0.1264} & 37.35 & 67.99 & \tablebestresult{88.47} & \tablebestresult{95.66} \\
\cmidrule[2.0pt]{1-10}
\multirow[c]{3}{*}{\tabincell{c}{\textit{train on GV2 medium} \\ test on GV2 fullplus}} & OmniFusion~\cite{li2022omnifusion} & 0.6159 & 0.1054 & 0.1618 & 0.1418 & 42.87 & 73.39 & 91.36 & 96.40 \\
& UniFuse~\cite{jiang2021unifuse} & 0.4498 & 0.0894 & 0.1434 & 0.1110 & 48.11 & 79.47 & 93.89 & 97.74 \\
& \textbf{Ours (fusion)} & \tablebestresult{0.4479} & \tablebestresult{0.0740} & \tablebestresult{0.1318} & \tablebestresult{0.0764} & \tablebestresult{51.25} & \tablebestresult{84.49} & \tablebestresult{97.00} & \tablebestresult{99.03} \\
\noalign{\smallskip}\hline
\end{tabular}
\end{table*}
\subsection{Datasets}
\label{sec:4.01}
\textbf{Matterport3D (M3D)}~\cite{chang2017matterport3d} and \textbf{Stanford2D3D (S2D3D)}~\cite{armeni2017jointstanford2d3d} are both real-world datasets collected using Matterport 3D Camera\footnote{\url{https://matterport.com/cameras}},
while \textbf{PanoSUNCG (PSunCG)}~\cite{wang2018self} and \textbf{3D60}~\cite{zioulis2018omnidepth} are both synthetic datasets.
Results evaluated on the above four datasets use metrics from previous works ~\cite{jiang2021unifuse,pintore2021slicenet,sun2021hohonet},
including mean absolute error ($MAE$), root mean square error ($RMSE$), logarithmic root mean square error ($RMSLE$), absolute relative error ($AbsRel$) and three threshold percentage~$\delta < {\alpha}^{n}~(\alpha=1.25, n={1, 2, 3})$.
\textbf{Pano3D}~\cite{albanis2021pano3d} is an evaluation benchmark for panorama depth estimation with the capability of evaluating the network's generalization ability.
We perform an evaluation on $1024 \times 512$ resolution compared with the baseline methods ${\mathrm{UNet}}^{vnl}$ and ${\mathrm{ResNet}}^{comp}_{skip}$ provided by Pano3D\footnote{\url{https://github.com/VCL3D/Pano3D}}.
\textbf{Note that metrics with prefix $w$ from evaluations on Pano3D datasets have considered error uniformity problems.}
In Figure~\ref{fig:3}, we demonstrate an aligned mesh generated by multiple depth maps estimated by the proposed method.
Comparison of depth map generated by UniFuse~\cite{jiang2021unifuse} and the proposed method on selected scenes from the M3D datasets are shown in Figure~\ref{fig:3.1},
which shows that the proposed method can generate depth maps with better details and is less fragile to image areas with special textures.
\subsection{Implementation Details}
\label{sec:4.02}
The Swin-B~\cite{liu2021swin} pertained on ImageNet-22k is used as the encoder network.
We use the same data augmentation techniques from UniFuse~\cite{jiang2021unifuse}.
All models are trained on 8 V100 GPU with batch size 32.
We use AdamW~\cite{loshchilov2017decoupled} optimizer with learning rate $\gamma=0.0003$.
A step decay learning rate scheduler with an interval epoch of 3 and a decay rate of 0.9 is used to accelerate the training process.
A five epochs warm-up with a starting learning rate of 0.00006 is adopted to get a more stable training convergence.
\begin{table}
\centering
\tiny
\caption{Comparison of different backbones on final accuracy using M3D datasets~\cite{chang2017matterport3d}.
We use EffNet B5 to represent EffcientNet B5 used in AdaBins~\cite{bhat2021adabins},
and ``D'' to represent the dimension of decoder.
Latency is tested on one NVIDIA TESLA P100 16GB.
Planar methods are marked with ${\triangle}$.}
\label{tab:4}
\begin{tabular}{c|c|c|c|c|c|c}
\hline
\noalign{\smallskip}
Method & Backbone & Params & Latency & AbsRel & RMSE & $\delta < 1.25$ \\
\noalign{\smallskip}\hline\noalign{\smallskip}
UniFuse~\cite{jiang2021unifuse} & ResNet18 & 30.3M & 25.5ms & 0.1063 & 0.4941 & 88.97 \\
Base & ResNet18 & \tablebestresult{19.9M} & \tablebestresult{17.3ms} & 0.1061 & 0.4761 & 88.67 \\
Fusion & ResNet18 & 33.9M & 25.2ms & \tablebestresult{0.0995} & \tablebestresult{0.4614} & \tablebestresult{90.31} \\
\noalign{\smallskip}\hline\noalign{\smallskip}
Base & ResNet34 & \tablebestresult{29.9M} & \tablebestresult{47.1ms} & 0.1040 & \tablebestresult{0.4680} & 89.33 \\
Fusion & ResNet34 & 43.0M & 81.3ms & \tablebestresult{0.0988} & 0.4720 & \tablebestresult{89.97} \\
\noalign{\smallskip}\hline\noalign{\smallskip}
ACDNet~\cite{zhuang2021acdnet} & ResNet50 & 87.0M & 72.4ms & 0.1010 & 0.4629 & 90.00 \\
Base D=256 & ResNet50 & \tablebestresult{32.9M} & \tablebestresult{58.6ms} & 0.1023 & 0.4638 & 89.43 \\
Base D=400 & ResNet50 & 45.6M & 92.9ms & 0.0999 & 0.4607 & 89.85 \\
Fusion D=256 & ResNet50 & 45.3M & 99.6ms & 0.0978 & 0.4552 & 90.44 \\
Fusion D=512 & ResNet50 & 88.7M & 144.3ms & \tablebestresult{0.0944} & \tablebestresult{0.4448} & \tablebestresult{91.07} \\
\noalign{\smallskip}\hline\noalign{\smallskip}
${\triangle}$Adabins~\cite{bhat2021adabins} & EffNet B5 & 78.0M & 199.9ms & 0.0958 & 0.4484 & 90.54 \\
Base & EffNet B5 & \tablebestresult{39.2M} & \tablebestresult{80.7ms} & 0.0947 & 0.4412 & 91.11 \\
Fusion & EffNet B5 & 53.2M & 111.2ms & \tablebestresult{0.0911} & \tablebestresult{0.4280} & \tablebestresult{91.90} \\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
\subsection{Overall Accuracy}
\label{sec:4.1}
On real-world datasets, the proposed method achieves overall best performance. As shown in Table~\ref{tab:1}, the decrease on $AbsRel$ of the proposed method is $\mathbf{5.0}\%$ (\textit{Fusion}) on S2D3D and $\mathbf{3.9}\%$ (\textit{Fusion}) on M3D; and the decrease on $RMSE$ is $\mathbf{4.2}\%$ (\textit{Base}) and $\mathbf{4.9}\%$ (\textit{Fusion}) respectively.
We also compare the proposed method with planar image method AdaBins~\cite{bhat2021adabins};
the results of the proposed method is better as it is better in dealing with distortion and retrieving global features.
It's worth noticing that AdaBins provide better results on M3D datasets comparing with ACDNet,
as self-attention module used in AdaBins largely enhances its capability of extracting features with global context.
\textit{Fusion} method does not outperform \textit{Base} on all evaluation metrics on S2D3D datasets,
as small number of training images (1040 only) in S2D3D makes the CAF module insufficiently trained.
\begin{table*}
\centering
\tiny
\caption{Comparison of our method with currently available methods on synthetic datasets.
We use $^*$ to label results evaluated using align methods.
Most methods trained using the PSunCG datasets can only estimate depth map at resolution $512 \times 256$,
but the proposed method can generate higher resolution ($1024 \times 512$),
where results are marked with ${\Diamond}$.}
\label{tab:2}
\begin{tabular}{c|c|c|c|c|c|c|c|c}
\hline\noalign{\smallskip}
\multirow[c]{2}[3]{*}{Dataset} & \multirow[c]{2}[3]{*}{Method} & \multicolumn{4}{c|}{Error Metric $\Downarrow$} & \multicolumn{3}{c}{Accuracy Metric $\Uparrow$} \\
& & ${MAE}$ & ${AbsRel}$ & ${RMSE}$ & ${RMSE}_{log}$ & $\delta < 1.25$ & $\delta < {1.25}^{2}$ & $\delta < {1.25}^{3}$ \\
\noalign{\smallskip}\hline
\hline\noalign{\smallskip}
\multirow[c]{11}[6]{*}{PSunCG} & OmniDepth$^*$~\cite{zioulis2018omnidepth} & 0.1624 & 0.1143 & 0.3710 & 0.0882 & 87.05 & 93.65 & 96.50 \\
& FCRN$^*$~\cite{laina2016fcrn} & 0.1346 & 0.0979 & 0.3973 & 0.0692 & 92.23 & 96.59 & 98.19 \\
& \textbf{Ours (base)}$^*$ & 0.0708 & 0.0355 & \aligntablebestresult{0.2639} & 0.0346 & 97.64 & 98.94 & 99.46 \\
& \textbf{Ours (fusion)}$^*$ & \aligntablebestresult{0.0696} & \aligntablebestresult{0.0340} & 0.2646 & \aligntablebestresult{0.0341} & \aligntablebestresult{97.81} & \aligntablebestresult{98.98} & \aligntablebestresult{99.48} \\
\cmidrule[1.3pt]{2-9}
& Bifuse~\cite{wang2020bifuse} & 0.0789 & 0.0592 & 0.2596 & 0.0443 & 95.90 & 98.38 & 99.07 \\
& UniFuse~\cite{jiang2021unifuse} & 0.0765 & 0.0485 & 0.2802 & 0.0416 & 96.55 & 98.46 & 99.10 \\
& \textbf{Ours (base)} & 0.0647 & 0.0347 & \tablebestresult{0.2628} & 0.0351 & 97.70 & 98.89 & 99.37 \\
& \textbf{Ours (fusion)} & \tablebestresult{0.0631} & \tablebestresult{0.0327} & 0.2631 & \tablebestresult{0.0346} & \tablebestresult{97.82} & \tablebestresult{98.92} & \tablebestresult{99.38} \\
\cmidrule[1.3pt]{2-9}
& ${\Diamond}$\textbf{Ours (base)} & 0.0572 & 0.0320 & 0.2220 & 0.0309 & 97.97 & 99.04 & 99.46 \\
& ${\Diamond}$\textbf{Ours (fusion)} & \tablebestresult{0.0553} & \tablebestresult{0.0298} & \tablebestresult{0.2217} & \tablebestresult{0.0303} & \tablebestresult{98.09} & \tablebestresult{99.07} & \tablebestresult{99.47} \\
\midrule[1.5pt]
\multirow[c]{10}{*}{3D60} & OmniDepth$^*$~\cite{zioulis2018omnidepth} & 0.1706 & 0.0931 & 0.3171 & 0.0725 & 90.92 & 97.02 & 98.51 \\
& FCRN$^*$~\cite{laina2016fcrn} & 0.1381 & 0.0699 & 0.2833 & 0.0473 & 95.32 & 99.05 & 99.66 \\
& \textbf{Ours (base)}$^*$ & 0.0766 & 0.0331 & 0.1802 & 0.0263 & 99.08 & \tablebestresult{99.71} & \tablebestresult{99.87} \\
& \textbf{Ours (fusion)}$^*$ & \tablebestresult{0.0733} & \tablebestresult{0.0316} & \tablebestresult{0.1770} & \tablebestresult{0.0257} & \tablebestresult{99.13} & \tablebestresult{99.71} & \tablebestresult{99.87} \\
\cmidrule[1.3pt]{2-9}
& Bifuse~\cite{wang2020bifuse} & 0.1143 & 0.0615 & 0.2440 & 0.0428 & 96.99 & 99.27 & 99.69 \\
& UniFuse~\cite{jiang2021unifuse} & 0.0996 & 0.0466 & 0.1968 & 0.0315 & 98.35 & 99.65 & 99.87 \\
& OmniFusion~\cite{li2022omnifusion} & $\times$ & 0.0430 & \tablebestresult{0.1808} & 0.0735 & 98.59 & 99.69 & \tablebestresult{99.89} \\
& \textbf{Ours (base)} & 0.0819 & 0.0367 & 0.1823 & 0.0271 & 99.05 & \tablebestresult{99.71} & 99.87 \\
& \textbf{Ours (fusion)} & \tablebestresult{0.0809} & \tablebestresult{0.0357} & 0.1818 & \tablebestresult{0.0270} & \tablebestresult{99.11} & \tablebestresult{99.71} & 99.87 \\
\noalign{\smallskip}\hline
\end{tabular}
\end{table*}
On Pano3D, based on results in Table~\ref{tab:3},
compared with best baseline results, our method reduces $wAbsRel$ by $\mathbf{23.0}\%$ (\textit{Fusion}) on tiny split and $\mathbf{16.3}\%$ (\textit{Fusion}) on the medium split.
The $wRMSE$ is reduced by $\mathbf{42.1}\%$ (\textit{Fusion}) and $\mathbf{43.5}\%$ (\textit{Fusion}) on these two splits.
We also train on GV2's medium and tested on fullplus to provide additional reference of the method's performance.
Results on Pano3D show that the proposed method has good generalization ability.
The proposed method is weaker on accuracy metrics on fullplus test splits;
this is probably because paying attention to the global context of the image makes the method fragile to large differences between training data and test data.
We believe that more training data can solve this problem.
On synthetic datasets, the proposed method is also overall the best, as shown in Table ~\ref{tab:2}:
comparing with previous work UniFuse~\cite{jiang2021unifuse},
on PSunCG datasets, the $RMSE$ decrease $\mathbf{6.5}\%$ (\textit{Fusion}),
and the $AbsRel$ decrease $\mathbf{32.6}\%$ (\textit{Fusion}).
However, on these two datasets, \textit{Fusion} does not outperform \textit{Base} too much.
This is because of the local correlation between the rendered depths from 3D models and the brightness of color images~\cite{jiang2021unifuse},
which derives the network to use ``local'' brightness information as the main cue for predicting depth,
thus the global context retrieving ability of \textit{Fusion} version does not contribute to final results.
\subsection{Ablation Study}
\label{sec:4.2}
\paragraph{\textbf{Different backbone networks}}
We adapt the proposed work with different backbones and show accuracy comparison in Table~\ref{tab:4}.
Changing the backbone of the proposed method from Swin to ResNet shows obvious performance downgrade,
but still better than methods like ACDNet and UniFuse with less number of parameters and better latency.
This proves that the improvement of the proposed work not only comes from the backbone we use but also the decoder on the spherical surface and our cross attention fusion module.
\begin{table}
\centering
\tiny
\caption{Accuracy test results using KITTI outdoor panorama datasets~\cite{Garanderie2018Eliminating}. Evaluation method is adapted from evaluation methods of M3D datasets~\cite{chang2017matterport3d}.}
\label{tab:4.2}
\begin{tabular}{c|c|c|c|c}
\hline\noalign{\smallskip}
\multirow[c]{2}{*}{Method} & \multicolumn{3}{c|}{Error $\Downarrow$} & Accuracy (\%) $\Uparrow$ \\
& ${AbsRel}$ & ${RMSE}$ & ${RMSLE}$ & $\delta < 1.25$ \\
\noalign{\smallskip}\hline
\hline\noalign{\smallskip}
OmniFusion~\cite{li2022omnifusion} & 0.0977 & 4.536 & 0.193 & 90.80 \\
UniFuse~\cite{jiang2021unifuse} & 0.0945 & 4.186 & 0.076 & 92.58 \\
Proposed & \tablebestresult{0.0657} & \tablebestresult{3.822} & \tablebestresult{0.060} & \tablebestresult{93.60} \\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
\paragraph{\textbf{Outdoor datasets}}
We evaluate the method's performance on KITTI outdoor datasets~\cite{Garanderie2018Eliminating} using our own train/test split,
and results are shown in Table~\ref{tab:4.2}.
Although the proposed method aims for solving depth estimation mostly for indoor scenes,
its capability of capturing global context helps the proposed method get overall best results.
\begin{table}
\centering
\tiny
\caption{Effects of error angles when the panorama images are not optimally aligned to gravity. Data with different tilt angles is retrieved by re-rendering from depth images from M3D datasets~\cite{chang2017matterport3d}. }
\label{tab:7}
\begin{tabular}{c|c|c|c|c|c}
\hline\noalign{\smallskip}
\multirow[c]{2}{*}{Methods} & \multirow[c]{2}{*}{\tabincell{c}{Tilt \\ angle}} & \multicolumn{3}{c|}{Error $\Downarrow$} & Accuracy (\%) $\Uparrow$ \\
& & ${AbsRel}$ & ${RMSE}$ & ${RMSLE}$ & $\delta < 1.25$\\
\noalign{\smallskip}\hline
\hline\noalign{\smallskip}
\multirow[c]{3}{*}{Bifuse~\cite{wang2020bifuse}} & ${0}^{\circ}$ & 0.2048 & 0.6259 & 0.1134 & 84.52 \\
& ${2}^{\circ}$ & 0.3888 & 0.9805 & 0.1852 & 61.44 \\
& ${5}^{\circ}$ & 0.4905 & 1.0225 & 0.2250 & 54.40 \\
\noalign{\smallskip}\hline\noalign{\smallskip}
\multirow[c]{3}{*}{SliceNet~\cite{pintore2021slicenet}} & ${0}^{\circ}$ & 0.1764 & 0.6133 & 0.1045 & 87.16 \\
& ${2}^{\circ}$ & 0.2645 & 0.7026 & 0.1334 & 72.56 \\
& ${5}^{\circ}$ & 0.3032 & 0.7720 & 0.1482 & 68.79 \\
\noalign{\smallskip}\hline\noalign{\smallskip}
\multirow[c]{3}{*}{\textbf{Ours}} & ${0}^{\circ}$ & 0.0865 & 0.4052 & 0.0559 & 92.64 \\
& ${2}^{\circ}$ & 0.0927 & 0.4282 & 0.0600 & 91.60 \\
& ${5}^{\circ}$ & 0.1091 & 0.4792 & 0.0711 & 88.14 \\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
\paragraph{\textbf{Imperfect alignment to gravity direction}} This may happen when the camera is not placed on a horizontal plane.
Results in Table~\ref{tab:7} demonstrate the performance downgrade of the proposed method under different tilt angles between the axis of the panorama image and gravity direction;
ground truth used for evaluation is retrieved by re-rendering from depth images from the M3D dataset. For tilt angles $2^{\circ}$ and $5^{\circ}$, the ${RMSE}$ of the proposed method increase $\mathbf{5.6}\%$ and $\mathbf{18.3}\%$ respectively, while performance downgrades for SliceNet~\cite{pintore2021slicenet} are $\mathbf{14.6}\%$ and $\mathbf{25.9}\%$ respectively.
This proves that the proposed method process better capability of resisting the jeopardization from imperfect gravity alignment.
\section{Conclusions}
\label{sec:5}
This paper proposes a novel method for panorama depth estimation.
By adapting uniform sampling on the unit sphere,
and by using self-attention calculation to enhance network's capability of retrieving global context,
the method achieved impressive accuracy on all publicly available datasets.
However, the proposed method can still be further improved:
by moving the encoder parts onto unit sphere, the method will possess better distortion-removal capability, which can improve the accuracy of the proposed method;
edge-preserving filters can be utilized in interpolation process of $\mathbf{f}_{{\Pi}\rightarrow{\mathbf{S^{2}}}}$ and $\mathbf{f}_{{\mathbf{S^{2}}}\rightarrow\Pi}$ to further improve depth estimation quality on edges.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:intro}
In the introduction to Shannon's seminal work on communication theory~\cite{shannon1949mathematical}, Warren Weaver defined a problem with three levels: classical information theory deals with Level A, or the \emph{technical problem}, which concerns itself with the accurate and efficient transmission of data. However, if communication is oriented toward affecting an agent's action, there are two further levels to the problem: Level B, or the \emph{semantic problem}, is to find the most effective way to convey the meaning of the message, even when irrelevant details are lost or misunderstood, while Level C, also called the \emph{effectiveness problem}, deals with the resulting behavior of the receiver: as long as the receiver takes the optimal decision, the effectiveness problem is solved, regardless of any differences in the received information.
While the Level B and C problems attracted limited attention for decades, the explosion of \gls{iiot} systems has drawn the research and industrial communities toward semantic and effective communication~\cite{popovski2020semantic}, optimizing remote control processes under severe communication constraints beyond Shannon's limits on Level A performance~\cite{gunduz2022beyond}. In particular, the effectiveness problem is highly relevant to robotic applications, in which independent mobile robots, such as drones or rovers, must operate based on information from remote sensors. In this case the sensors and cameras act as the transmitter in a communication problem, while the robot is the receiver: by solving the Level C problem, the sensors can transmit the information that best directs the robot's actions toward the optimal policy~\cite{stavrou2022rate}. We can also consider a case in which the robot is the transmitter, while the receiver is a remote controller, which must get the most relevant information to decide the control policy~\cite{wan2020cognitive}.
The rise of communication metrics that take the content of the message into account, such as the \gls{voi}~\cite{yates2020agesurvey}, represents an attempt to approach the problem in practical scenarios, and analytical studies have exploited information theory to define a semantic accuracy metric and minimize distortion~\cite{shao2022theory}. In particular, \emph{information bottleneck} theory~\cite{beck2022semantic} has been widely used to characterize Level B optimization~\cite{shao2021learning}. However, translating a practical system model into a semantic space is a non-trivial issue, and the semantic problem is a subject of active research~\cite{uysal2022semantic}.
The effectiveness problem is even more complex, as it implicitly depends on estimating the effect of communication distortion on the control policy and, consequently, on its performance~\cite{tung2021effective}. While the effect of simple scheduling policies is relatively easy to compute~\cite{kim2019learning}, and linear control systems can be optimized explicitly~\cite{zheng2020urgency}, realistic control tasks are highly complex, complicating an analytical approach to the Level C problem. Pure learning-based solutions that consider communication as an action in a multi-agent \gls{drl} problem, such as emergent communication, also have limitations~\cite{foerster2016learning}, as they can only deal with very simple scenarios due to significant convergence and training issues. In some cases, the information bottleneck approach can also~\cite{goyal2018transfer} be exploited to determine state importance, but the existing literature on optimizing Level C communication is very sparse, and limited to simpler scenarios~\cite{rcmab_pase}.
In this work, we consider a dual model which combines concepts from \gls{drl} and semantic source coding: we consider an ensemble \gls{vqvae} model~\cite{van2017neural}, each of which learns to represent observations using a different codebook. A \gls{drl} agent can then select the codebook to be used for each transmission, controlling the trade-off between accuracy and compression. Depending on the task of the receiver, the reward to the \gls{drl} agent can be tuned to solve the Level A, B, and C problems, optimizing the performance for each specific task.
In order to test the performance of the proposed framework, we consider the well-known CartPole problem, whose state can be easily converted into a semantic space, as its dynamics depend on a limited set of physical quantities. We show that dynamic codebook selection outperforms static strategies for all three levels, and that considering the Level C task can significantly improve the control performance with the same bitrate.
The rest of the paper is organized as follows: first, we present the general system model in Sec.~\ref{sec:system}. We then present the dynamic feature compression solution in Sec.~\ref{sec:solution}, which is evaluated by simulation in Sec.~\ref{sec:results}. Finally, Sec.~\ref{sec:conclusion} concludes the paper, along with remarks on possible extensions.
\section{System Model}
\label{sec:system}
We consider a general model in which two agents cooperate to solve a \emph{remote \gls{pomdp}} problem. In the following, we will denote random variables with capital letters, their possible values with lower-case letters, and sets with calligraphic or Greek capitals.
In the standard \gls{pomdp} formulation~\cite{pomdp}, one agent needs to optimally control a stochastic process defined by a tuple $\left< \Sigma, \mathcal{A}, \Omega, P, \xi, R, \gamma\right>$, where $\Sigma$ represents the set of system states, $\mathcal{A}$ is the set of feasible actions, and $\Omega$ is the observation set. The function $P : \Sigma \times \mathcal{A} \rightarrow \Phi(\Sigma)$, where $\Phi(\cdot)$ represents all possible probability distributions over a set, gives the state transition probability function, with $P_{ss'}^a$ denoting the conditional probability of entering state $s' \in \Sigma$ given the current state $s$ and the taken action $a \in \mathcal{A}$. On the other hand, $\xi : \Sigma \rightarrow \Phi(\Omega)$ is an observation function, which provides the conditional probabilities
$\xi_{os}$, i.e., the probability that the system is in state $s$, given the observation $o$. Finally, $R: \Sigma \times \mathcal{A} \rightarrow \mathbb{R}$ provides the average reward given to the agent when taking action $a$ in state $s$, denoted as $R_s^a$, and the scalar $\gamma \in [0, 1)$ is a discount factor.
Specifically, the \gls{pomdp} proceeds in discrete steps indexed by $t$, and at each step $t$ the agent can infer the system state $s_t$ only from the stochastic observation $o_t$, thus having partial information on it. Based on this observation, and on its policy $\pi : \Omega \rightarrow \Phi(\mathcal{A})$, which outputs a probability distribution over the action space for each possible observation, the agent interacts with the system by sampling an action ${a_t \sim \pi(a_t | o_t)}$. Then, given the real system state $s_t$ and the sampled action $a_t$, the agent receives a feedback from the system in the form of a reward ${r_t = R_{s_t}^{a_t}}$. The goal for the agent is to optimize its policy $\pi$ to maximize the expected cumulative discounted reward $G = \mathbb{E}\Big[ \sum_{t} \gamma^t R_t \Big]$. We can see that the \gls{pomdp} is indeed built on top of the \gls{mdp} formulation, with the addition of noisy observations.
\subsection{The Remote POMDP}
In this paper we consider a variant of the \gls{pomdp}, that we define \emph{remote \gls{pomdp}}, in which two agents are involved in the process. The first agent, i.e., the \emph{observer}, receives the observation $o_t \in \Omega$, and needs to convey such information to a second agent, i.e., the \emph{actor}, through a constrained communication channel, which limits the number of bits the observer can send. Consequently, the amount of information the observer can send to the actor is limited. The actor then chooses and takes an action. This system can formalize many control problems in future \gls{iiot} systems, as sensors and actuators can potentially be geographically distributed, and the amount of information they can exchange to accomplish a task is limited due to the shared wireless medium, which has to be allocated to the many devices installed in the factory, as well as by the energy limitations on the sensors.
In order to transmit the most relevant information to the actor, the observer chooses its transmission $m_t \in \mathcal{M}$, where $\mathcal{M}$ is the set of possible messages (which corresponds to the observer's action space). Then, based on the received message $m_t$, the actor optimizes its own policy $\pi^a(a_t | m_t)$ to maximize $G$, and to optimally solve the underlying control problem. As the objective of the observer is to minimize channel usage, i.e., communicate as few bits as possible, while maintaining the highest possible performance in the control task, we consider the problem of adaptive encoding.
In order to solve this problem, the observer needs to have a way to gauge the \emph{value} of information, as described in the introduction: information theory, and in particular rate-distortion theory, have provided the fundamental limits when optimizing for the technical problem, i.e., Level A, where the goal is to reconstruct the source signals with the highest fidelity~\cite{cover:IT}. The first and simplest way to solve the remote \gls{pomdp} problem is to blindly apply standard rate-distortion metrics to compress the sensor observations into messages to be sent to the agent. The observer's policy is then independent of the actor's task, and can be computed separately.
However, the observation may contain much more information than needed: for example, an image showing an object might be reduced to its coordinates, without losing any meaningful information for the controller. Lossily compressing the message to preserve the relevant information, removing redundant or irrelevant details, can beat the Shannon bound, easing communication requirements without losing performance. Compression is then a projection of the raw observations into a significantly smaller \emph{semantic space}.
As an example, in the CartPole problem analyzed in this work (see Section~\ref{sec:results}), one sensor observation is given by two consecutive 2D camera acquisitions, which only implicitly capture the needed physical system information, e.g., the angular position and velocity of the pole. By employing semantic source coding, the sensor can extract the underlying physical quantities from the raw observations, which can be useful for multiple applications at the receiver, considering not reconstruction accuracy but rather semantic fidelity, which measures the error on the underlying physical values rather than on the raw sensory observation, thus solving the Level B communication problem. What can be gained with the semantic formulation in terms of trade-off between size of the message and performance on the downstream task at the receiver, with respect to the Level A system, is the second question we try to answer with our analysis.
Going further, one may only want to convey the physical information related to a specific task, constructing a Level C system as envisioned in~\cite{shannon1949mathematical}. For this purpose, the sensor has access to feedback signals indicating the quality of the actions of the remote agent when decoding specific messages. This data can be used by the sensor to adjust the observations coding scheme to optimally trade off the size of the message and the performance of the remote agent on the specific control problem. How to design this scheme, and how this scheme compares with Level A and B designed systems, are the other questions we answer with our study.
\subsection{The Information Bottleneck}
\label{subsec:info_bottleneck}
We now consider the observer's encoding function. As the actor can use the received message $m_t$, along with its memory of past messages at previous steps, to choose an action $a_t$, the encoding function is $\Lambda: \Omega\times\mc{M}^{\tau} \rightarrow \mc{M}$, which considers a case with memory $\tau$. The function will generate a message $M_t = \Lambda(O_t,M_{t-\tau},\ldots,M_{t-1})$ to be sent to the actor at each step $t$. As mentioned above, the encoding function $\Lambda$ can be optimized to solve one of the three communication problems.
We assume that the underlying state of the physical system at time $t$ can be entirely described by the variable $S_t$, which is an element of the physical state space $\Sigma$. However, in our remote \gls{pomdp} formulation, the sensor has access to this state only through the observation $O_t$, which is in general a stochastic function of $S_t$ whose codomain is the observation space $\Omega$. In addition, we assume that the performance at time $t$ in a particular task $\mc{T}$
depends only on a sufficient statistic $i_{\mc{T}}(s)$ of any given state $s\in\Sigma$. Denoting the number of bits required to represent a realization of a random variable $X$~\cite{cover:IT} as $b(X)$, we assume that
\begin{align}
\label{eq:relation}
b(i_{\mc{T}}(S)) < b(S) < b(O) \quad \forall S \in \Sigma.
\end{align}
We can observe that ${i_{\mc{T}}(S) \rightarrow S \rightarrow O}$ is a hidden Markov model, as the information available in the current observation also summarizes the full history of past observations.
The random quantity $i_{\mc{T}}(S)$ represents the minimal description of the system with respect to the task $\mc{T}$, i.e., no additional data computed from $S$ adds meaningful information for the actor when solving $\mc{T}$. The state $S$ may also include task-irrelevant physical information on the system. However, both $S$ and $i_{\mc{T}}(S)$ are unknown quantities, as the observer only receives a noisy and high-dimensional representation of $S$ through $O$. This is a well-known issue in \gls{drl}: in the original paper presenting the \gls{dqn} architecture~\cite{mnih2015human}, the agent could only observe the screen while playing classic arcade videogames, and did not have access to the much more compact and precise internal state representation of the game. Introducing communication and dynamic encoding adds another layer of complexity.
\subsection{Encoding and Distortion}\label{subsec:coding}
In our remote \gls{pomdp}, the communication channel between the sensor and the actor is not ideal, thus the information the observer can convey to the actor at each step $t$ is limited to a maximum length of $L$ bits per message. Consequently, the problem introduces an information bottleneck between the observation $O_t$ and the final action $A_t$ the actor can take, which is based on the message $M_t$ conveyed through the channel. In principle, whenever $b(O)>L$, the constrained communication introduces a distortion $d(O, \hat{O})$, whose theoretical asymptotic limits are given by rate-distortion theory~\cite{cover:IT}. If we also consider \emph{memory}, i.e., the use of past messages in the estimation of $\hat{O}$, the mutual information between $O$ and the previous messages can be used to reduce the distortion, improving the quality of the estimate.
In the semantic problem, the aim is to extrapolate the real physical state of the system $S_t$ from the observation $O_t$, which can be a complex stochastic function. In general, the real state lies in a low-dimensional semantic space $\Sigma$. The term semantic is motivated by fact that, in this case, the observer is not just transmitting pure sensory data, but some meaningful physical information about the system. Consequently, the distortion to be considered in this case is some fidelity function $d(S_t, \hat{S}_t)$ measuring the reconstruction quality of the state $S_t$ in the semantic space, where the estimation $\hat{S}_t$ is performed by the actor based on its knowledge of the message, as well as its memory of past messages. To be even more efficient and specific with respect to the task $\mc{T}$, the observer may optimize the message $M_t$ to minimize the distortion
between the effective representation $i_{\mc{T}}(S_t)$, which contains only the task-specific information, and the knowledge available to the actor. Naturally, any message instance $m_t\in\mc{M}$ must be at most $L$ bits long, in order to respect the constraint. If we define a generic distortion
$d_{\alpha}$ between the real quantity and its reconstruction at the different levels, which are indexed by $\alpha\in\{A,B,C\}$, the observer's task
is then defined as follows:
\begin{equation}
\Lambda^*=\argmax_{\mathclap{\Lambda:\Omega\times\mc{M}^{\tau}\rightarrow\mc{M}}}\E{\sum_{t=0}^\infty \gamma^t\left[d_{\alpha} -\beta\ell(M_t)\right]\bigg| S_0,\Lambda},\label{eq:varcapproblem}
\end{equation}
where $\beta\geq0$ is a communication cost parameter and the function $\ell(\cdot)$ represents the length (in bits) of a message. The observation space of the observer is given by $\Omega\times\mc{M}^{\tau}$, while the observation space of the actor is $\mc{M}^{\tau+1}$, as it only receives the messages from the observer.
\section{Proposed Solution}
\label{sec:solution}
In this section, we introduce the architecture we used to represent $\Lambda$, the \gls{vqvae}, and discuss the remote \gls{pomdp} solution. As the \gls{vqvae} model is not adaptive, we consider an ensemble model with different quantization levels, limiting the choice of the observer to \emph{which} \gls{vqvae} model to use in the transmission. As the action space in~\eqref{eq:varcapproblem} is huge, techniques such as emergent communication that learn it explicitly are limited to scenarios with very simple tasks and immediate rewards. By restricting the problem to a smaller action space, we may find a slightly suboptimal solution, but we can deal with much more complex problems.
\subsection{Deep VQ-VAE Encoding}
\label{subsec:vae}
In order to represent the encoding function $\Lambda$, and to restrict the problem in~\eqref{eq:varcapproblem} to a more manageable action space, the observer exploits the \gls{vqvae} architecture introduced in~\cite{van2017neural}. The \gls{vqvae} is built on top of the more common \gls{vae} model, with the additional feature of finding an optimal discrete representation of the latent space. The \gls{vae} is used to reduce the dimensionality of an input vector $X \in \mathbb{R}^I$, by mapping it into a stochastic latent representation $Z \in \mathbb{R}^L \sim q_{\phi}(Z | X)$, where $L < I$. The stochastic encoding function $q_{\phi}(Z | X)$ is a parameterized probability distribution represented by a neural network with parameter vector $\phi$. To find optimal latent representations $Z$, the \gls{vae} jointly optimizes a decoding function $p_{\theta}(\hat{X}|Z)$ that aims to reconstruct $X$ from a sample $\hat{X} \sim p_{\theta}(\hat{X} | Z)$. This way, the parameter vectors $\phi$ and $\theta$ are usually jointly optimized to minimize the distortion $d(X, \hat{X})$ between the input and its reconstruction, given the constraint on $Z$, while reducing the distance between $q_{\phi}(Z|X)$, and some prior $q(Z)$~\cite{vae2013} used to impose some structure or complexity budget.
However, in practical scenarios, one needs to digitally encode the input $X$ into a discrete latent representation. To do this, the \gls{vqvae} quantizes the latent space by using $N$ $K$-dimensional codewords $z_1, \dots, z_N \in \mathbb{R}^K$, forming a dictionary with $N$ entries. Moreover, to better represent 3D inputs, the \gls{vqvae} quantizes the latent representation $Z$ using a set of $F$ blocks, each quantizing one feature $f(X)$ of the input, and chosen from a set of $N$ possible codewords. We denote the set containing all the $N^F$ possible concatenated blocks with $\mathcal{M}(N)$, as it represents the set of all possible messages the observer can use to convey to the actor the information on the observation $O$, by using $F$ discrete $N$-dimensional features. The peculiarity of the \gls{vqvae} architecture is that it jointly optimizes the codewords in $\mathcal{M}(N)$ together with the stochastic encoding and decoding functions $q_{\phi}$ and $p_{\theta}$, instead of simply applying fixed vector quantization on top of learned continuous latent variables $Z$. When the communication budget is fixed, i.e., the value of $L$ is constant, the protocol to solve the remote \gls{pomdp} is rather simple: first, the observer trains the \gls{vqvae} with $N = 2^{\frac{L}{F}}$ to minimize the technical, semantic, or effective distortion $d_{\alpha}$, depending on the problem; then, at each step $t$, the observer computes $\hat{m} \sim q_{\phi}(\cdot | o_t)$, and finds $ m_t = \argmin_{m \in \mathcal{M}(N)} \| m - \hat{m} \|_2$.
The message $m_t$ is sent to the actor, which can optimize its decision on the action to take accordingly.
\subsection{Dynamic Feature Compression}
\label{subsec:drl}
We can then consider the architecture shown in Fig.~\ref{fig:arch}, consisting of a set of \glspl{vqvae} $\mathcal{V} = \{ \Psi_1, \ldots, \Psi_V\}$, where each \gls{vqvae} has a different compression level $N_{\Psi}$: as we only consider the communication side of the problem, the actor is trained beforehand using the messages with the finest-grained quantization, which are compressed with the \gls{vqvae} $\Psi$ with larger value of $N$, i.e., larger codebook.
The actor can then perform three different tasks, corresponding to the three communication problems: it can decode the observation (Level A) with the highest possible accuracy, using the decoder part of the \gls{vqvae} architecture; it can estimate the hidden state (Level B) using a supervised learning solution; or it can perform a control action based on the received information and observe its effects (Level C).
\begin{figure}
\centering
\input{figures/architecture}
\caption{Dynamic feature compression architecture.}\vspace{-0.5cm}
\label{fig:arch}
\end{figure}
In all three cases, the dynamic compression is performed by the observer, based on the feedback from the actor. The observer side of the remote \gls{pomdp}, whose reward is given in~\eqref{eq:varcapproblem}, is restricted to the choice of $\Psi_v$, i.e., to selecting one of the possible codebooks learned by each \gls{vqvae} in the ensemble model. Naturally, the feedback depends on the communication problem that the observer is trying to solve: at Levels A and B, the observer aims at minimizing distortion in the observation and semantic space, respectively, while at Level C, the objective is to maximize the reward of the actor.
In all three cases, \emph{memory} is important: representing snapshots of the physical system in consecutive instants, subsequent observations have high correlations, and the actor can glean a significant amount of information from past messages. This is a significant advantage of dynamic compression, as it can adapt messages to the estimated knowledge of the receiver.
While the observer is adapting its transmissions to the actor's task, the actor's algorithms are fixed. They could themselves be adapted to the dynamic compression strategy, but this joint training is significantly more complex, and we consider it as a possible extension of this work.
\section{Numerical Evaluation}
\label{sec:results}
The underlying use case analyzed in this work is the well-known CartPole problem, as implemented in the \emph{OpenAI Gym} library.\footnote{\url{https://www.gymlibrary.dev/environments/classic_control/cart_pole/}} In this problem, a pole is installed in a cart, and the task is to control the cart position and velocity to keep the pole in equilibrium. The physical state of the system is fully described by the cart position $x_t$ and velocity $\dot{x}_t$, and the pole angle $\psi_t$ and angular velocity $\dot{\psi}_t$. Consequently, the true state of the system is ${s_t = ( x_t, \dot{x}_t, \psi_t, \dot{\psi}_t )}$, and the semantic state space is $\Sigma\subset\mathbb{R}^4$ (because of physical constraints, the range of each value does not actually span the whole real line).
At each step $t$, the observer senses the system by taking a color picture of the scene, which is in a space $\mc{P}=\left\{ 0, \dots, 255\right\}^{180 \times 360 \times 3}$. To take the temporal element into account, an observation $O_t$ includes two subsequent pictures, at times $t-1$ and $t$, so that the observation space is $\mc{O}=\mc{P}\times\mc{P}$. An example of the transmission process is given in Fig.~\ref{fig:cartpole}, which shows the original version sensed by the observer (above) and the reconstructed version at the receiver (below) when using a trained \gls{vqvae} with $N_{\Psi}=6$.
In the CartPole problem, the action space $\mathcal{A}$ contains just two actions $\mathtt{Left}$ and $\mathtt{Right}$, which push the cart to the left or to the right, respectively. At the end of each step, depending on the true state $s_t$, and on the taken action $a_t$, the environment will return a deterministic reward $ R_t = -x_{\text{max}}^{-1} |x_t|-\psi_{\text{max}}^{-1} |\psi_t|, \ \forall t$,
where $x_{\text{max}} = 4.8$ [m] and $\psi_{\text{max}} = 24^{\circ}$ are the maximal achievable values for the two quantities. The goal for the two agents is thus to maximize the cumulative discounted sum of the reward $R_t$, while limiting the communication cost.
\subsection{The Coding \& Decoding Functions}
\label{subsec:training_vqvae}
As mentioned in Sec.~\ref{subsec:coding}, the observer can optimize its coding function $\Lambda$ according to different criteria depending on the considered communication problem. However, as we explained in Sec.~\ref{sec:solution}, optimizing the choice of $\Lambda$ is usually not feasible due to the curse of dimensionality on the action space. Consequently, we rely on a pre-trained set $\mc{V}$ of \gls{vqvae} models, whose codebooks are optimized to solve the technical problem, i.e., minimizing the distortion on the observation measured using the \gls{mse}: $d_{A}(O, \hat{O}) = \text{MSE}(O, \hat{O})$. The training performance of the \gls{vqvae} with $N_{\Psi}=6$ is shown in Fig.~\ref{fig:vqvae_training}: the encoder converges to a good reconstruction performance, which can be measured by its \emph{perplexity}. The perplexity is simply $2^{H(p)}$, where $H(p)$ is the entropy of the codeword selection frequency, and a perplexity equal to the number of codewords is the theoretical limit, which is only reached if all codewords are selected with the same probability. The perplexity at convergence is 54.97, which is close to the theoretical limit for a real application.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\columnwidth]{figures/cartpole}
\includegraphics[width=0.8\columnwidth]{figures/cartpole_tx}
\caption{Example of the original and reconstructed observation.}\vspace{-0.6cm}
\label{fig:cartpole}
\end{figure}
\begin{figure}[t]
\centering
\subfloat[PSNR.\label{fig:vqvae}]{\input{figures/training_vqvae}}%
\subfloat[Perplexity.\label{fig:perplexity}]{\input{figures/perplexity_vqvae}}%
\caption{Training of the \gls{vqvae} model with $N=6$.}\vspace{-0.5cm}
\label{fig:vqvae_training}
\end{figure}
The observer then uses \gls{drl} to foresightedly optimize the quantization level $\Psi_t$ at each time step, maximizing~\eqref{eq:varcapproblem} for each communication problem. We train the observer to solve the level-specific coding problem by designing three different rewards, depending on the considered communication level:
\begin{enumerate}
\item \emph{Level A (technical problem):} The reward for the observer is $r_t = \text{PSNR}(o_t, \hat{o_t}) - \beta \ell(m_t)$, where the \gls{psnr} is an image fidelity measure proportional to the logarithm of the normalized \gls{mse} between the original and reconstructed image;
\item \emph{Level B (semantic problem):} The reward is $r_t = -\text{MSE}(s_t, \hat{s_t}) - \beta \ell(m_t)$, and the decoder needs to estimate the underlying physical state $s_t$ by minimizing the \gls{mse}, i.e., the distance between $s_t$ and $\hat{s}_t$ in the semantic space. In our case, the estimator used to obtain $\hat{s}_t$ is a pre-trained supervised \gls{lstm} neural network;
\item \emph{Level C (effective problem):} In this case the reward is $r_t = Q(o_t,\pi(a_t|m_t)) - \beta\ell(m_t)$, where $Q(o_t, a_t)$ is the Q-function of the actor, measuring its long-term performance in the CartPole task when sampling actions $a_t$ according to its policy $\pi$ informed by the message $m_t$. The policy $\pi$ is given by a \gls{dqn} agent implementing an \gls{lstm} architecture, pre-trained using data with the highest available message quality (6 bits per feature).
\end{enumerate}
We observe that, in this case, $i_{\mathcal{T}(s)} = s$, i.e., the task depends on all the semantic features contained in $s_t$. However, the $4$ values do not carry the same amount of information to the actor: depending on the system conditions, i.e., the state $s_t$, some pieces of information are more relevant than others.
\begin{figure*}[t]
\centering
\subfloat{\input{figures/prob_legend}}\vspace{-0.3cm}\\
\setcounter{subfigure}{0}
\subfloat[Technical problem.\label{fig:technical}]{\input{figures/tech_prob}}%
\subfloat[Semantic problem.\label{fig:semantic}]{\input{figures/sem_prob}}%
\subfloat[Effective problem.\label{fig:effective}]{\input{figures/eff_prob}}%
\caption{Performance of the communication schemes on the three levels of the remote \gls{pomdp}.}\vspace{-0.5cm}
\label{fig:performance}
\end{figure*}
\subsection{Results}
\label{subsec:results}
We assess the performance of the three different tasks in the CartPole scenario by simulation, measuring the results over 100 episodes after convergence. Fig.~\ref{fig:performance} shows the performance of the various schemes over the three problems, compared with a static \gls{vqvae} solution with a constant compression level. In the Level C evaluation, we also consider a static \gls{vqvae} solution in which the actor is not retrained for each $N_{\Psi}$, but is only trained for $N_{\Psi}=6$ (i.e., 48 bits per message). We trained the dynamic schemes with different levels of the communication cost $\beta$, so as to provide a full picture of the adaptation to the trade-off between performance and compression. We also introduce the notion of \emph{Pareto dominance}: an $n$-dimensional tuple $\eta=(\eta_1,\ldots,\eta_n)$, $\eta$ Pareto dominates $\eta'$ (which we denote as $\eta\succ \eta'$) if:
\begin{equation}
\eta\succ \eta' \iff \exists i:\eta_i>\eta_i' \wedge\eta_j\geq\eta_j'\,\forall j.
\end{equation}
We can extend this to different schemes, which have multiple performance points. The definition of Pareto dominance for schemes $x$ and $y$ is: $x\succ y \iff \exists \eta_x\succ\eta_y\,\forall \eta_y.$
We first consider the technical problem performance, shown in Fig.~\ref{fig:technical}: as expected, the Level A dynamic compression outperforms all other solutions, and its performance is Pareto dominant with respect to static compression. Interestingly, the Level B and Level C solutions perform worse than static compression: by concentrating on features in the semantic space or the task space, these solutions remove information that could be useful to reconstruct the full observation, but is meaningless for the specified task.
In the semantic problem, shown in Fig.~\ref{fig:semantic}, a lower \gls{mse} on the reconstructed state is better, and the Level B solution is Pareto dominant with respect to all others. The Level A solution also Pareto dominates static compression, while the Level C solution only outperforms it for higher compression levels, i.e., on the left side of the graph.
Finally, Fig.~\ref{fig:effective} shows the performance at the effectiveness level, summarized by the time that the cart-pole system is within the position and angle limits. The Level C solution significantly outperforms all others, but is not strictly Pareto dominant: when the communication constraint is very tight, setting a static compression and retraining the actor to deal with the specific \gls{vqvae} used may provide a slight performance advantage. However, this is also a possibility for a Level C dynamic compression: retraining the actor to deal with the dynamically compressed messages could match the performance of the retrained static scheme, and in general, almost perfect control can be achieved with less than half of the average bitrate of the static compressor. We also note that, in this case, the Level B solution performs worst: choosing the solution that minimizes the semantic distortion might not be matched to the task, as the state variables have equal weight, while a higher precision might be required when the quantization error might change the action the actor chooses.
\section{Conclusion}
\label{sec:conclusion}
In this work, we presented a dynamic feature compression scheme that can exploit an ensemble \gls{vqvae} to solve the semantic and effective communication problems. The dynamic scheme outperforms fixed quantization, and can be trained automatically with limited feedback, unlike emergent communication models that are unable to deal with complex tasks.
A natural extension of this model is to consider more complex tasks and wider communication channels, corresponding to realistic control scenarios. Another interesting direction for future work is to consider joint training of the actor and the observer, or cases with partial information available at both transmitter and receiver.
\bibliographystyle{IEEEtran}
\section{Introduction}
\label{sec:intro}
In the introduction to Shannon's seminal work on communication theory~\cite{shannon1949mathematical}, Warren Weaver defined a problem with three levels: classical information theory deals with Level A, or the technical problem, which concerns itself with the accurate and efficient transmission of data. However, if communication is oriented toward affecting an agent's action by communicating some information, there are two further levels to the problem: Level B, or the semantic problem, is to find the most effective way to convey the \emph{meaning} of the message, even when irrelevant details are lost or misunderstood, while Level C, also called the effectiveness problem, deals with the resulting behavior of the receiver: as long as the receiver takes the optimal decision, the effectiveness problem is solved, regardless of any differences in the received information.
While the Level B and C problems attracted limited attention for decades, the explosion of \gls{iiot} systems has drawn the research and industrial communities toward semantic communication~\cite{popovski2020semantic}, optimizing remote control processes under severe communication constraints beyond Shannon's limits on Level A performance~\cite{gunduz2022beyond}. In particular, the effectiveness problem is highly relevant to robotic applications, in which independent mobile robots, such as drones or rovers, must operate based on information from remote sensors. In this case the sensors and cameras act as the transmitter in a communication problem, while the robot is the receiver: by solving the Level C problem, the sensors can transmit the information that best directs the robot's actions toward the optimal policy~\cite{stavrou2022rate}. In the case of offloaded computation, the robot is the transmitter, while the remote controller is the receiver, which must get the most relevant information to decide the control policy~\cite{wan2020cognitive}.
The rise of metrics taking the content of the message into account such as \gls{voi}~\cite{yates2020agesurvey} in communications represents an attempt to approach the problem in practical scenarios, and analytical studies have exploited information theory to define a semantic accuracy metric and minimize distortion~\cite{shao2022theory}. In particular, \emph{information bottleneck} theory~\cite{beck2022semantic} has been widely used to characterize Level B optimization~\cite{shao2021learning}. However, translating a practical system model into a semantic space is a non-trivial issue, and the semantic problem is a subject of active research~\cite{uysal2022semantic}.
The effectiveness problem is even more complex, as it represents a dual problem: communication needs to be optimized not only toward minimizing semantic distortion, but toward distortion in directions that affect the control policy in meaningful ways~\cite{tung2021effective}. While the effect of simple scheduling policies is relatively easy to compute~\cite{kim2019learning}, and linear control systems can be optimized explicitly~\cite{zheng2020urgency}, realistic control tasks are highly complex, complicating an analytical approach to the Level C problem. Pure learning-based solutions that consider communication as an action in a multi-agent \gls{drl} problem, such as emergent communication, also have limitations~\cite{foerster2016learning}, as they can only deal with very simple scenarios due to significant convergence and training issues. In some cases, the information bottleneck approach can also~\cite{goyal2018transfer} be exploited to determine state importance, but the existing literature on optimizing Level C communication is very sparse, and limited to simpler scenarios~\cite{rcmab_pase}.
In this work, we consider a dual model which combines concepts from \gls{drl} and semantic source coding: we consider an ensemble \gls{vqvae} model~\cite{van2017neural}, each of which learns to represent observations using a different codebook. A \gls{drl} agent can then select the codebook to be used for each transmission, controlling the trade-off between accuracy and compression. Depending on the task of the receiver, the reward to the \gls{drl} agent can be tuned to solve the Level A, B, and C problems, optimizing the performance for each specific task.
In order to test the performance of the proposed framework, we consider the well-known CartPole problem, whose state can be easily converted into a semantic space, as its dynamics depend on a limited set of physical quantities. We show that dynamic codebook selection outperforms static strategies for all three levels, and that considering the Level C task can significantly improve control performance with the same bitrate.
The rest of the paper is organized as follows: first, we present the general system model in Sec.~\ref{sec:system}. We then present the dynamic feature compression solution in Sec.~\ref{sec:solution}, which is evaluated by simulation in Sec.~\ref{sec:results}. Finally, Sec.~\ref{sec:conclusion} concludes the paper, along with remarks on possible extensions.
\section{Numerical Evaluation}
\label{sec:results}
The underlying use case analyzed in this work is the well-known CartPole problem, as implemented in the \emph{OpenAI Gym} library\footnote{\url{https://www.gymlibrary.dev/environments/classic_control/cart_pole/}}. In this problem, a pole is installed in a cart, and the task is to control the cart position and velocity to keep the pole in equilibrium. The physical state of the system is fully described by the cart position $x_t$ and velocity $\dot{x}_t$, and the pole angle $\psi_t$ and angular velocity $\dot{\psi}_t$. Consequently, the true state of the system is ${s_t = ( x_t, \dot{x}_t, \psi_t, \dot{\psi}_t )}$, and the semantic state space is $\mathcal{S}\subset\mathbb{R}^4$ (because of physical constraints, the range of each value does not actually span the whole real line).
At each step $t$, the observer senses the system by taking a color picture of the scene, which is in a space $\mc{P}=\left\{ 0, \dots, 255\right\}^{180 \times 360 \times 3}$. To take the temporal element into account, an observation $O_t$ includes two subsequent pictures, at times $t-1$ and $t$, so that the observation space is $\mc{O}=\mc{P}\times\mc{P}$. An example of the transmission process is given in Fig.~\ref{fig:cartpole}, which shows the original version sensed by the observer (above) and the reconstructed version at the receiver (below).
In the CartPole problem, the action space $\mathcal{A}$ contains just two actions $\mathtt{Left}$ and $\mathtt{Right}$, which push the cart to the left or to the right, respectively. At the end of each step, depending on the true state $s_t$, and on the taken action $a_t$, the environment will return a deterministic reward:
\begin{align}
R_t = -x_{\text{max}}^{-1} |x_t|-\psi_{\text{max}}^{-1} |\psi_t|, \quad \forall t,
\end{align}
where $x_{\text{max}} = 4.8$ [m] and $\psi_{\text{max}} = 24^{\circ}$ [deg] are the maximal achievable values for the two quantities. The goal for the two agents is thus to maximize the cumulative discounted sum of the reward $R_t$, while limiting the communication cost.
\subsection{The Coding \& Decoding Functions}
\label{subsec:training_vqvae}
As mentioned in Section~\ref{subsec:coding}, the observer can optimize its coding function $\Lambda$ according to different criteria depending on the considered communication problem. However, as we explained in Section~\ref{sec:solution}, optimizing the choice of $\Lambda$ is usually not feasible due to the curse of dimensionality on the action space. Consequently, we rely on a pre-trained set $\mc{V}$ of \gls{vqvae} models, whose codebooks are optimized to solve the technical problem, i.e., minimizing the distortion on the observation measured using the \gls{mse}: $d_{A}(O, \hat{O}) = \text{MSE}(O, \hat{O})$. The training performance of the \gls{vqvae} with $N_{\Psi}=6$ is shown in Fig.~\ref{fig:vqvae_training}: the encoder converges to a good reconstruction performance, which can be measured by its \emph{perplexity}. The perplexity is simply $2^{H(p)}$, where $H(p)$ is the entropy of the codeword selection frequency, and a perplexity equal to the number of codewords is the theoretical limit, which is only reached if all codewords are selected with the same probability. The perplexity at convergence is 54.97, which is close to the theoretical limit for a real application.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\columnwidth]{figures/cartpole}
\includegraphics[width=0.8\columnwidth]{figures/cartpole_tx}
\caption{Example of the original and reconstructed observation.}\vspace{-0.4cm}
\label{fig:cartpole}
\end{figure}
\begin{figure}[t]
\centering
\subfloat[PSNR.\label{fig:vqvae}]{\input{figures/training_vqvae}}%
\subfloat[Perplexity.\label{fig:perplexity}]{\input{figures/perplexity_vqvae}}%
\caption{Training of the \gls{vqvae} model with $N=6$.}\vspace{-0.4cm}
\label{fig:vqvae_training}
\end{figure}
The observer then uses \gls{drl} to foresightedly optimize the quantization level $\Psi_t$ at each time step, maximizing~\eqref{eq:varcapproblem} for each communication problem. We train the observer to solve the level-specific coding problem by designing three different rewards, depending on the considered communication level:
\begin{enumerate}
\item \emph{Level A (technical problem):} The reward for the observer is $r_t = \text{PSNR}(o_t, \hat{o_t}) - \beta \ell(m_t)$, where the \gls{psnr} is an image fidelity measure proportional to the logarithm of the normalized \gls{mse};
\item \emph{Level B (semantic problem):} The reward is $r_t = -\text{MSE}(s_t, \hat{s_t}) - \beta \ell(m_t)$, and the decoder needs to estimate the underlying physical state $s_t$ by minimizing the \gls{mse}, i.e., the distance between $s_t$ and $\hat{s}_t$ in the semantic space. In our case, the estimator used to obtain $\hat{s}_t$ is a pre-trained supervised \gls{lstm} neural network;
\item \emph{Level C (effective problem):} In this case the reward is $r_t = Q(o_t,\pi(a_t|m_t)) - \beta\ell(m_t)$, where $Q(o_t, a_t)$ is the Q-function of the actor, measuring its long-term performance in the CartPole task when sampling actions $a_t$ according to its policy $\pi$ informed by the message $m_t$. The policy $\pi$ is given by a \gls{dqn} agent implementing an \gls{lstm} architecture, pre-trained using data with the highest available message quality (6 bits per feature).
\end{enumerate}
We observe that, in this case, $i_{\mathcal{T}(s)} = s$, i.e., the task depends on all the semantic features contained in $s_t$. However, the $4$ values do not carry the same amount of information to the actor: depending on the system conditions, i.e., the state $s_t$, some pieces of information are more relevant than others.
\begin{figure*}[t]
\centering
\subfloat{\input{figures/prob_legend}}\vspace{-0.2cm}\\
\setcounter{subfigure}{0}
\subfloat[Technical problem.\label{fig:technical}]{\input{figures/tech_prob}}%
\subfloat[Semantic problem.\label{fig:semantic}]{\input{figures/sem_prob}}%
\subfloat[Effective problem.\label{fig:effective}]{\input{figures/eff_prob}}%
\caption{Performance of the communication schemes on the three levels of the remote \gls{pomdp}.}\vspace{-0.4cm}
\label{fig:performance}
\end{figure*}
\subsection{Results}
\label{subsec:results}
We assess the performance of the three different tasks in the CartPole scenario by simulation, measuring the results over 100 episodes after convergence. Fig.~\ref{fig:performance} shows the performance of the various schemes over the three problems, compared with a static \gls{vqvae} solution with a constant compression level. We trained the dynamic schemes with different levels of the communication cost $\beta$, so as to provide a full picture of the adaptation to the trade-off between performance and compression. We also introduce the notion of \emph{Pareto dominance}: if performance is an $n$-dimensional tuple $\eta=(\eta_1,\ldots,\eta_n)$, $\eta$ Pareto dominates $\eta'$ (which we denote as $\eta\succ \eta'$) if:
\begin{equation}
\eta\succ \eta' \iff \exists i:\eta_i>\eta_i' \wedge\eta_j\geq\eta_j'\,\forall j.
\end{equation}
We can extend this to the performance of schemes, which have multiple performance points. The definition of Pareto dominance for schemes $x$ and $y$ is: $x\succ y \iff \exists \eta_x\succ\eta_y\,\forall \eta_y.$
We consider the technical problem performance, shown in Fig.~\ref{fig:technical}: as expected, the Level A dynamic compression outperforms all other solutions, and its performance is Pareto dominant with respect to static compression. Interestingly, the Level B and Level C solutions perform worse than static compression: by concentrating on features in the semantic space or the task space, these solutions remove information that could be useful to reconstruct the full observation, but is meaningless for the specified task.
In the semantic problem, shown in Fig.~\ref{fig:semantic}, a lower \gls{mse} on the reconstructed state is better, and the Level B solution is Pareto dominant with respect to all others. The Level A solution also Pareto dominates static compression, while the Level C solution only outperforms it for higher compression levels, i.e., on the left side of the graph.
Finally, we consider the performance at the effectiveness level, shown in Fig.~\ref{fig:effective}: we can note that the performance in the task, summarized by the time that the cart-pole system is within the position and angle limits, is much higher for the Level C solution. Interestingly, however, it is not strictly Pareto dominant: for extreme compression levels, setting a static compression and retraining the actor to deal with the specific \gls{vqvae} used can provide a slight performance advantage. However, this is also a possibility for a Level C dynamic compression: retraining the actor to deal with the dynamically compressed messages could match the performance of the retrained static scheme, and in general, almost perfect control can be achieved with less than half of the average bitrate as the static compressor. We also note that in this case, the Level B solution performs worst: choosing the solution that minimizes the semantic distortion might not be matched to the task, as the state variables have equal weight, while a higher precision might be required when the quantization error might change the action the actor chooses.
\section{Proposed Solution}
\label{sec:solution}
In this section, we introduce the architecture we used to represent $\Lambda$, the \gls{vqvae}, and discuss the remote \gls{pomdp} solution. As the \gls{vqvae} model is not adaptive, we consider an ensemble model with different quantization levels, limiting the choice of the observer to \emph{which} \gls{vqvae} model to use in the transmission. As the action space in~\eqref{eq:varcapproblem} is huge, techniques such as emergent communication that learn it explicitly are limited to very simple scenarios. By restricting the problem to a smaller action space, we may find a slightly suboptimal solution, but can deal with much more complex problems.
\subsection{Deep VQ-VAE Encoding}
\label{subsec:vae}
In order to represent the encoding function $\Lambda$, and to restrict the problem in~\eqref{eq:varcapproblem} to a more manageable action space, the observer exploits the \gls{vqvae} architecture introduced in~\cite{van2017neural}. The \gls{vqvae} is built on top of the more common \gls{vae} model, with the additional feature of finding an optimal discrete representation of the latent space. The \gls{vae} is used to reduce the dimensionality of an input vector $X \in \mathbb{R}^I$, by mapping it into a stochastic latent representation $Z \in \mathbb{R}^L \sim q_{\phi}(Z | X)$, where $L < I$. The stochastic encoding function $q_{\phi}(Z | X)$ is a parameterized probability distribution represented by a neural network with parameter vector $\phi$. To find optimal latent representations $Z$, the \gls{vae} jointly optimizes a decoding function $p_{\theta}(\hat{X}|Z)$ that aims to reconstruct $X$ from a sample $\hat{X} \sim p_{\theta}(\hat{X} | Z)$. This way, the parameter vectors $\phi$ and $\theta$ are usually jointly optimized to minimize the distortion $d(X, \hat{X})$ between the input and its reconstruction, given the constraint on $Z$, while reducing the distance between $q_{\phi}(Z|X)$, and some prior $q(Z)$~\cite{vae2013} used to impose some structure or complexity budget.
However, in practical scenarios, one needs to digitally encode the input $X$ into a discrete latent representation. To do this, the \gls{vqvae} quantizes the latent space by using $N$ $K$-dimensional codewords $z_1, \dots, z_N \in \mathbb{R}^K$, forming a dictionary with $N$ entries. Moreover, to better represent 3D inputs, the \gls{vqvae} quantizes the latent representation $Z$ using a set of $F$ codewords, each quantizing one feature $f(X)$ of the input, and chosen from a set of $N$ possible codewords. We denote the set containing the all $N^F$ codewords with $\mathcal{M}(N)$, as it represents the set of all possible messages the observer can use to convey to the actor the information on the observation $O$, by using $F$ discrete $N$-dimensional features. The peculiarity of the \gls{vqvae} architecture is that it jointly optimizes the codewords in $\mathcal{M}(N)$ together with the stochastic encoding and decoding functions $q_{\phi}$ and $p_{\theta}$, instead of simply applying fixed vector quantization on top of learned continuous latent variables $Z$. When the communication budget is fixed, i.e., constant value of $B$, the protocol to solve the remote \gls{pomdp} is rather simple: First, the observer trains the \gls{vqvae} with $N = 2^{\frac{B}{F}}$ to minimize the technical, semantic, or effective distortion $d_{\alpha}$, depending on the problem; Then, at each step $t$, the observer computes $\hat{m} \sim q_{\phi}(\cdot | o_t)$, and find $ m_t = \argmin_{m \in \mathcal{M}(N)} \| m - \hat{m} \|_2$.
The message $m_t$ is sent to the actor, which can optimize its decision on the action to take according to it.
\subsection{Dynamic Feature Compression}
\label{subsec:drl}
We can then consider the architecture shown in Fig.~\ref{fig:arch}, consisting of a set of \glspl{vqvae} $\mathcal{V} = \{ \Psi_1, \ldots, \Psi_V\}$, where each \gls{vqvae} has a different compression level $N_{\Psi}$: as we only consider the communication side of the problem, the actor is trained beforehand using the messages with the finest-grained quantization, which are compressed with the \gls{vqvae} $\Psi$ with larger value of $N$, i.e., larger codebook.
The actor can then perform three different tasks, corresponding to the three communication problem: it can decode the observation (Level A) with the highest possible accuracy, using the decoder part of the \gls{vqvae} architecture, it can estimate the hidden state (Level B) using a supervised learning solution, or it can control the environment.
\begin{figure}
\centering
\input{figures/architecture}
\caption{Dynamic feature compression architecture.}\vspace{-0.4cm}
\label{fig:arch}
\end{figure}
In all three cases, the dynamic compression is performed by the observer, based on the feedback from the actor. The observer side of the remote \gls{pomdp}, whose reward is given in~\eqref{eq:varcapproblem}, is restricted to the choice of $\Psi_v$, i.e., to selecting one of the possible codebooks learned by each \gls{vqvae} in the ensemble model. Naturally, the feedback depends on the communication problem that the observer is trying to solve: at Levels A and B, the observer aims at minimizing distortion in the observation and semantic space, respectively, while at Level C, the objective is to maximize the reward of the actor.
In all three cases, \emph{memory} is important: representing snapshots of the physical system in consecutive instants, subsequent observations have high correlations, and the actor can glean a significant amount of information from past messages. This is a significant advantage of dynamic compression, as it can adapt messages to the estimated knowledge of the receiver.
However, the performance could still be improved: the observer is adapting its transmissions to the actor's task, but the actor's algorithms are fixed, while they could themselves be adapted to the dynamic compression strategy. This joint training is significantly more complex, and we consider it as a possible extension of the work.
\section{System Model}
\label{sec:system}
We consider a general model in which two agents cooperate to solve a \emph{remote \gls{pomdp}} problem. In the standard \gls{pomdp} formulation~\cite{pomdp}, one agent needs to optimally control a stochastic process defined by a tuple $\left< \mathcal{S}, \mathcal{A}, \Omega, P, O, R, \gamma\right>$, where $\mathcal{S}$ represents a set of system states, $\mathcal{A}$ is the set of feasible actions, and $\Omega$ is the observation set. The function $P : \mathcal{S} \times \mathcal{A} \rightarrow \Phi(\mathcal{S})$, where $\Phi(\cdot)$ represents all possible probability distributions over a set, gives the state transition probability, whose element $P_{ss'}^a$ provides the probability distribution of the next state $s' \in \mathcal{S}$ given that the system is in state $s \in \mathcal{S}$, and the agent has taken action $a \in \mathcal{A}$. On the other hand, $O : \mathcal{S} \rightarrow \Phi(\Omega)$ is an observation function, which provides the conditional probabilities $O(s, o) = Pr \left[ S_t = s | O_t = o\right]$, i.e., the probability that the system is in state $s$, given the observation $o$. Finally, $R: \mathcal{S} \times \mathcal{A} \rightarrow \mathbb{R}$ provides the average reward given to the agent when taking action $a$ in state $s$, denoted as $R_s^a$, and the scalar $\gamma \in [0, 1)$ is a discount factor.
Specifically, the \gls{pomdp} proceeds in discrete steps indexed by $t$, and at each step $t$ the agent can infer the system state $s_t$ only from the stochastic observation $o_t$, thus having partial information on it. Based on this observation, and on its policy $\pi : \Omega \rightarrow \Phi(\mathcal{A})$, which outputs a probability distribution over the action space for each possible observation, the agent interacts with the system by sampling an action ${a_t \sim \pi(a_t | o_t)}$. Then, given the real system state $s_t$ and the sampled action $a_t$, the agent receives a feedback from the system in the form of a reward ${r_t = R_{s_t}^{a_t}}$. The goal for the agent is to optimize its policy $\pi$ to maximize the expected cumulative reward discounted reward $G = \mathbb{E}\Big[ \sum_{t} \gamma^t r_t \Big]$. We can see that the \gls{pomdp} is indeed built on top of the \gls{mdp} formulation, with the addition of noisy observations.
\subsection{The Remote POMDP}
In this paper we consider a variant of the \gls{pomdp}, that we define \emph{remote \gls{pomdp}}, in which two agents are involved in the process. The first agent, i.e., the \emph{observer}, receives the observation $o_t \in \Omega$, and needs to convey such information to a second agent, i.e., the \emph{actor}, through a constrained communication channel, which limits the number of bits the observer can send. The actor then chooses and takes an action. Consequently, the amount of information the observer can send to the actor is limited. This system can formalize many control problems in future \gls{iiot} systems, as sensors and actuators can potentially be geographically distributed, and the amount of information they can exchange to accomplish a task is limited due to the shared wireless medium, which has to be allocated to the large amount of devices installed in the factory, as well as by the energy limitations on the sensors.
In order to transmit the most relevant information to the actor, the observer chooses its transmission $m_t \in \mathcal{M}$, where $\mathcal{M}$ is the set of possible messages (which corresponds to the observer's action space). Then, based on the received message $m_t$, the actor optimizes its own policy $\pi^a(a_t | m_t)$ to maximize $G$, and to optimally solve the underlying control problem. As the objective of the observer is to minimize channel usage, i.e., communicate as few bits as possible, while maintaining the highest possible performance in the control task, we consider the problem of adaptive encoding.
In order to solve this problem, the observer needs to have a way to gauge the \emph{value} of information, as described in the introduction: information theory, and in particular rate-distortion theory, have provided the fundamental limits when optimizing for the technical problem, i.e., Level A, where the goal is to reconstruct the source signals with the highest fidelity~\cite{cover:IT}. The first and simplest way to solve the remote \gls{pomdp} problem is to blindly apply standard rate-distortion metrics to compress the sensor observations into messages to be sent to the agent. The observer's policy is then independent from the actor's task, and can be computed separately.
However, in robotics and remote control applications, which are the targets of our analysis, the receiver is only interested in the part of the message conveying relevant information on the underlying physical system, or on the task to be accomplished, and so further source redundancy may be exploited to reduce the size of the message. Consequently, it is natural to assume that the observations convey some sort of physical information on the underlying process to be controlled. A natural projection of the raw observations into a semantic space can be assumed, which captures the physical meaning contained in the observed signals, removing meaningless information due to noise and/or sensor additional data.
As an example, in the CartPole problem analyzed in this work (see Section~\ref{sec:results}), one sensor observation is given by two consecutive 2D camera acquisitions, which only implicitly capture the needed physical system information, e.g., the angular position and velocity of the pole. By employing semantic source coding, the sensor can extract the underlying physical quantities from the raw observations, which can be useful for multiple applications at the receiver, considering not reconstruction accuracy but rather semantic fidelity, which measures the error on the underlying physical values rather than on the raw sensory observation, thus solving the Level B communication problem. What can be gained with the semantic formulation in terms of trade-off between size of the message, and performance on the downstream task at the receiver, with respect to the Level A system is the second question we try to answer with our analysis.
Going further, one may only want to convey the physical information related to a specific task, constructing a Level C system as envisioned in~\cite{shannon1949mathematical}. For this purpose, the sensor has access to feedback signals indicating the quality of the actions of the remote agent when decoding specific messages. This data can be used by the sensor to adjust the observations coding scheme to optimally trade off size of the message, and performance of the remote agent on the specific control problem. How to design this scheme, and how this scheme compares with Level A and B designed systems, are the other questions we answer with our study.
\subsection{The Information Bottleneck}
\label{subsec:info_bottleneck}
We now consider the observer's encoding function. As the actor can use the received message $m_t$, along with its memory of past messages at previous steps, to choose an action $a_t$, the encoding function is $\Lambda: \Omega\times\mc{M}^{\tau} \rightarrow \mc{M}$, which considers a case with memory $\tau$. The function will generate a message $M_t = \Lambda(O_t,M_{t-\tau},\ldots,M_{t-1})$ to be sent to the actor at each step $t$. As mentioned above, the encoding function $\Lambda$ can be optimized to solve one of the three communication problems.
We assume that the underlying state of the physical system at time $t$ can be entirely described by the variable $s_t$, which is an element of the physical state space $\mc{S}$. However, in our remote \gls{pomdp} formulation, the sensor can access to this state only through the observation $o_t$, which is in general a stochastic function of $s_t$ whose codomain is the observation space $\Omega$. In addition to this, we assume that the performance at time $t$ in a particular task $\mc{T}$
depends only on a sufficient statistic $i_{\mc{T}}(s_t)$ of the state. Denoting the number of bits required to represent the random variable $X$~\cite{cover:IT} as $b(X)$, we assume that
\begin{align}
\label{eq:relation}
b(i_{\mc{T}}(S)) < b(S) < b(O) \quad \forall S \in \mc{S}.
\end{align}
We implicitly impose a Markov Chain structure ${i_{\mc{T}}(S) \rightarrow S \rightarrow O}$. The random quantity $i_{\mc{T}}(S)$ represents the minimal description of the system with respect to the task $\mc{T}$, i.e., no additional data computed from $s_t$ adds meaningful information for the actor when solving $\mc{T}$. The state $S$ may also include task-irrelevant physical information on the system. However, both $S$ and $i_{\mc{T}}(S)$ are unknown quantities, as the observer only receives a noisy and high-dimensional representation of $S$ through $O$. This is a well-known issue in \gls{drl}: in the original paper presenting the \gls{dqn} architecture~\cite{mnih2015human}, the agent could only observe the screen while playing classic arcade videogames, and did not have access to the much more compact and precise internal state representation of the game. Introducing communication and dynamic encoding adds another layer of complexity to the problem.
\subsection{Encoding and Distortion}\label{subsec:coding}
In our remote \gls{pomdp}, the communication channel between the sensor and the actor is not ideal, thus the information the observer can convey to the actor at each step $t$ is limited to a maximum number of bits $B$ per message. Consequently, the problem introduces an information bottleneck between the observation $o_t$ and the final action $a_t$ the actor can take, which is based on the message $M_t$ conveyed through the channel. In principle, whenever $b(O)>B$, the constrained communication introduces a distortion $d(O, \hat{O})$, whose theoretical asymptotic limits are given by rate-distortion theory~\cite{cover:IT}. If we also consider \emph{memory}, i.e., use past messages in the estimation of $\hat{O}$, the mutual information between $O$ and previous messages can be used to reduce the distortion, improving the quality of the estimate.
In the semantic problem, the aim is to extrapolate the real physical state of the system $s_t$ from the observation $o_t$, which can be a complex stochastic function. In general, the real state lays in a low-dimensional semantic space $\mc{S}$. The term semantic is motivated by fact that, in this case, the observer is not just transmitting pure sensory data, but some meaningful physical information about the system. Consequently, the distortion to be considered in the trade-off is in the semantic space $d(S_t, \hat{S}_t)$, where the estimation $\hat{S}_t$ is performed by the actor based on its knowledge of the message, as well as its memory of past messages. To be even more efficient and specific with respect to the task $\mc{T}$, the observer may optimize the message $M_t$ to minimize the distortion
between the effective representation $i_{\mc{T}}(s_t)$, which contains only the task-specific information, and the knowledge available to the actor. Naturally, any $M_t\in\mc{M}$ must be at most $B$ bits long, in order to respect the constraint. If we define a generic distortion
$d_{\alpha}$ between the real quantity and its reconstruction at the different levels, which are indexed by $\alpha\in\{A,B,C\}$, the observer's task
is then defined as follows:
\begin{equation}
\Lambda^*=\argmax_{\mathclap{\Lambda:\Omega\times\mc{M}^{\tau}\rightarrow\mc{M}}}\E{\sum_{t=0}^\infty \gamma^t\left[d_{\alpha} -\beta\ell(M_t)\right]\bigg| S_0,\Lambda},\label{eq:varcapproblem}
\end{equation}
where $\beta\geq0$ is a communication cost parameter and the function $\ell(\cdot)$ represents the length (in bits) of a message. The observation space of the observer is given by $\Omega\times\mc{M}^{\tau}$, while the observation space of the actor is $\mc{M}^{\tau+1}$, as it only receives the messages from the observer.
\section{Conclusion}
\label{sec:conclusion}
In this work we have presented a dynamic feature compression scheme that can exploit an ensemble \gls{vqvae} to solve the semantic and effective communication problems. The dynamic scheme significantly outperforms fixed quantization, and can be trained automatically with limited feedback, as well as overcoming the limitations of emergent communication models, which are unable to deal with complex tasks.
A natural extension to the model is to consider more complex tasks and wider communication channels, corresponding to realistic control scenarios. Another interesting direction for future work is to consider a joint training of the actor and observer, or cases with partial information is available at both the transmitter and the receiver.
|
\section{Example Appendix}
\end{document}
\section{Conclusion}
In this paper, we introduce the TikTalk, a Chinese multi-modal dialogue dataset. In TikTalk, we collect videos and video-based dialogue corpora from a video social application. These conversations contain visual, audio, and textual information, which are more similar to multi-modal chitchat dialogue in real-world scenarios. The analysis shows that generating high-quality responses to dialogues in TikTalk requires a deep understanding of various contexts and sophisticated reasoning capabilities combined with external knowledge. We evaluate some conversation baselines and make some improvements to existing models for the characteristics of TikTalk. The experimental results demonstrate that they still fall far short of human performance. There is a large room for improvement on TikTalk, and we hope this work will further contribute to the research of real-world multi-modal dialogue.
\section{TikTalk Dataset}
The TikTalk dataset is built on the Chinese version of TikTok, a.k.a Douyin, a popular video social application. With over 680 million monthly active users, Douyin has videos in more than 25 broad categories, such as education, food, games, travel, entertainment, etc. Each video on Douyin has a caption provided by its author and comments generated by users. Users would engage in further discussions under the videos and first-level comments, which are close to real-world face-to-face scenarios of multi-modal chitchats.
\subsection{Data Construction}
We collect videos posted on Douyin during 2021, along with titles, comments and their reply-to relations. To protect privacy, we did not crawl any user information. Due to the large base of videos, the quality of videos is mixed, and most of them have only first-level comments, which cannot constitute conversation corpora. Users can also give likes to videos and comments and sharing them with friends on Douyin. A video with a high number of likes proves that it has a high user rating, which can help us filter out low-quality data. Similarly, comments that receive more likes are marked as popular comments with higher ranks in the comments section and will get more replies. These mechanisms help us to filter out low-quality videos and comments easily when crawling data.
In the data cleaning process, we first obtain videos with high likes and multi-level comments from the crawled videos and split the tree-shaped comments into conversational corpora. Then, we use regular expressions to filter out useless content in the sentences, such as ``@[someone]'' (a way of sharing with friends) and repetitive words, as well as some unethical remarks. In addition, comments are often accompanied by emoticons that can indicate the user's emotions. We extract them from the conversation as additional information.
\subsection{Data Statistics}
Finally, among the 153,340 videos crawled, we obtained 38,703 videos and 367,670 dialogues to construct the TikTalk dataset. We divide the training, development, and test set containing 35,703, 1,000, and 2,000 videos, respectively. The detailed statistics for each split and in total are in Table \ref{tab:dataset_statistics}. With an average length of 34.03 seconds, our videos provide a wealth of video and audio information. The average number of turns per conversation is 2.25. With the shorter textual context of dialogues, the responses are more likely to be derived from videos or external knowledge. In addition, we build higher-quality development and test set for testing the performance of multi-modal dialogue models. We select such videos with 5 user responses as candidates in each of their dialogues for a given context.
\begin{table}[t]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{lcccccc}
\hline
\textbf{Split} & \textbf{videos} & \textbf{dialogues} & \textbf{turns} & \textbf{Avg-v} & \textbf{Avg-t} & \textbf{Avg-l}\\
\hline
\textbf{Train} & 35,703 & 363,589 & 816,170 & 34.01s & 2.24 & 16.42 \\
\textbf{Dev} & 1,000 & 1,353 & 3,495 & 36.93s & 2.58 & 16.06 \\
\textbf{Test} & 2,000 & 2,728 & 7,087 & 36.42s & 2.60 & 16.25 \\
\textbf{Total} & 38,703 & 367,670 & 826,752 & 34.03s & 2.25 & 16.41 \\
\hline
\hline
\end{tabular}}
\caption{Data division and statistics of TikTalk. ``Avg-v'', ``Avg-t'', and ``Avg-l'' denote the average duration of videos, the average number of turns in dialogues, and the average length of turns, respectively.}
\label{tab:dataset_statistics}
\end{table}
\subsection{Data Analysis}
\paragraph{Characteristics}
TikTalk is proposed to facilitate research on multi-modal dialogues, which have several main characteristics. Firstly, users initiate or participate in these conversations after watching videos on a daily basis, without any pre-determined intention or guidelines for creating the dataset. They are highly spontaneous and free, similar to the everyday chitchat people have in the real world. Next, the dialogue context contains multiple modalities, including associated images, audio, and text. It provides more diverse information sources for dialogues, but also places more restrictions on the chitchat scenarios and raises the standard for the plausibility of responses. Moreover, videos on the social platform are open-domain, and the discussion topics are diverse. Some external knowledge related to the video content or the textual context may be required when generating responses.
\label{sec:data comparison}
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{pics/info.pdf}
\caption{Comparison of context sources for responses in different dialogue datasets. (a) compares the proportion of all responses originating from various contexts in four dialogue datasets. (b) and (c) compare the proportions of different levels of conversations originating from various contexts in the IGC and TikTalk datasets.}
\label{fig:information_distribution}
\end{figure*}
\paragraph{Data Comparison}Observing the dialogues, we find that the responses often contain words that refer to messages in the context. For example, ``it'' in Figure \ref{fig:tiktalk_example}(a) refers to the seagull in the video. The conversation in Figure \ref{fig:tiktalk_example}(c) discusses the story described by the audio. The responses in Figure \ref{fig:tiktalk_example}(b) favor the viewpoint of the textual context, as well as analyze ``other shows'' beyond the video, which is associated with the user's personal experience and common sense. Based on the above analysis, we wonder about the ratio of responses in our dataset correlated with various information, including visual context or audio context in the video, textual context, and external knowledge as hidden context.
Therefore, we perform data annotation through crowdsourcing. For comparison, we select three other representative dialogue datasets and randomly pick 300 samples from each dataset. The detailed statistics are as follows: 1) LCCC\cite{wang2020large} is a Chinese textual dialogue dataset from social media. There are 550 responses in the 300 selected sessions; 2) OpenViDial\cite{meng2020openvidial} contains dialogues and aligned keyframes extracted from movies and TV series. We obtain 1,019 responses from the 300 keyframes; 3) IGC\cite{mostafazadeh2017image} encourage annotators to do more chitchats related to a given image and questions. There are 1,182 responses associated with the 300 selected images; 4) our TikTalk is a video-based chitchat dataset. There are 2,472 responses in the 300 selected videos.
We develop an annotation website that extracts nouns and pronouns from responses. We ask annotators to judge whether these words and whole-sentence responses refer to 1) vision context; 2) audio context; 3) text context; 4) hidden context as external knowledge, and 5) others. Note that the response can involve context information from multiple modalities. When the response is hard to understand or judge, they can skip labeling it.
The comparison results of the four data sets are shown in Figure \ref{fig:information_distribution}(a). Firstly, TikTalk has the highest proportion of information sources after removing textual context and external knowledge, indicating that the visuals and audio in the videos can provide more information. Secondly, TikTalk has the highest proportion of responses derived from external knowledge (about 33\%), which suggests that people introduce more new information correlated to the current conversation when the multi-modal context is richer. Third, Only 18\% of the responses in the OpenViDial dataset are derived from visual context, suggesting that OpenViDial is less suitable for investigating how visual context affects response generation in real-world dialogues. In contrast, this ratio is 48\% and 42\% in the IGC and TikTalk, which are more visually based. The IGC, on the other hand, has a higher ratio of visual relevance as a percentage of overall information sources, suggesting that it is more likely to be highly iamge dependent on images than TikTalk. Furthermore, except for OpenViDial, the ratio of mentions of textual contexts is close to 80\% for all three dialogue datasets. This data indicates that the dialogue in OpenViDial is less articulated than a daily dialogue and requires a greater understanding of the overall storyline.
To further explore the differences between IGC and TikTalk, we also investigate the dependence of their responses on different contexts when the number of turns of dialogue increases, as shown in Figure ~\ref{fig:information_distribution}(b) and (c). Note that there is a significant decrease in the proportion of vision with conversation in the IGC dataset, while there is no such trend in the TikTalk dataset. It may be because only one image is used in each conversation in IGC, and less information is available over time. However, since videos provide richer information, users' discussion on Douyin can focus on more varied contexts.
\section{Experiments}
\subsection{Task Definition}
We formulate the problem of multi-modal chitchat dialogues in TikTalk as follows. Suppose that we have a multi-modal dialogue dataset $\mathcal{D}$. Each dialogue is represented as $(C^{v}, C^{a}, C^{t}, R)$, where $C^{v}$ and $C^{a}$ represent the visual context and audio context in the video, $C^{t}$ is the textual context formed by the preceding utterances $\{u_{i}\}_{i=1}^{n}$, and $R$ is a textual response. The goal of multi-modal chitchat is to learn a model $P(R|C^{v},C^{a},C^{t})$ from $\mathcal{D}$ to generate an ideal response $R$ to the given contexts $C^{v}$, $C^{a}$, and $C^{t}$.
\subsection{Evaluate Metrics}
We carefully select high-quality development and test sets, with five user responses as candidates for a given context in each conversation under the videos. We adopt three automatic metrics to evaluate the performance of the models on TikTalk from multiple perspectives.
\paragraph{Relevance} BLEU-2 and BLEU-4~\cite{papineni2002bleu}, Meteor~\cite{denkowski2014meteor}, Rouge-L~\cite{lin2004rouge}, and CIDEr~\cite{vedantam2015cider}. For each dialogue, we calculate the metrics for the generated responses and the five real candidates separately and then take the average value as the final result. We adopt the public NLG evaluation code\footnote{https://github.com/Maluuba/nlg-eval} to calculate these metrics.
\paragraph{Ranking} 1) Recall@K: the proportion of correct responses among the first K responses in the sorted candidate set. 2) Mean Rank: the average rank of correct responses, with lower being better. We construct a candidate set with 100 samples for each conversation, including 5 ground truths and 95 randomly selected responses as negatives. During the inference phase, the model sorts the candidate set in descending order based on the log-likelihood score of generating each response. Since there are 5 positive responses for each conversation, we calculate the recall such that K equals 5 and 10.
\paragraph{Diversity} Dist-1 and Dist-2~\cite{li2015diversity}. The diversity of responses is crucial as multi-modal dialogue scenarios have more information sources and higher demands for response rationality, and generalized responses are of lower quality.
\subsection{Baselines}
We evaluate some state-of-the-art dialogue models from different tasks and settings. To adapt to the characteristics of TikTalk, we introduce audio and external knowledge as input to the existing image-based dialogue models and evaluate the performance of the models, respectively.
\paragraph{Livebot} is proposed for the task of live comment generation in videos~\cite{ma2019livebot}. It is a variant of Transformer~\cite{vaswani2017attention} and consists of three modules: a video encoder, a text encoder, and a comment decoder. It feeds video frames and textual contexts into two encoders, models the dependencies between the two modalities through them, and uses the decoder to generate responses.
\paragraph{DialoGPT}~\cite{zhang2019dialogpt} extends GPT-2 \cite{radford2019language}, an auto-regressive language model, and inherits the multi-layer Transformer decoder as model architecture. It is pre-trained on large-scale conversational corpora and reduces the probability of generating tedious and redundant responses through a reward and penalty mechanism.
\paragraph{Maria} is a visual conversational agent~\cite{liang2021maria} that uses the UniLM~\cite{dong2019unified} architecture for multi-modal modeling. It retrieves images from an image library for each dialogue to construct image-based dialogue data. After that, it feeds the textual context, the visual concepts retrieved from the images, and the concept labels to the multi-modal Transformer to generate responses. In TikTalk, since each conversation has a corresponding video, we extract consecutive video frames as visual input and adopt MCP (Masked Context Prediction) and MRP (Masked Response Prediction) as training objectives. We use the same visual inputs and modeling targets in the subsequent variants of Maria.
\paragraph{Maria-Audio} is the first variant of Maria. We use the Viggish Network~\cite{hershey2017cnn} to extract the global audio features in each video. Similar to video frames, we input consecutive audio frames as another segment.
\begin{table*}[t]
\centering
\resizebox{160mm}{!}{
\begin{tabular}{l|ccccc|ccc|cc}
\hline
Methods & BLEU-2 & BLEU-4& Meteor & Rouge-L& CIDEr & R@5 & R@10 & MeanR & Dist-1 & Dist-2\\
\hline
Livebot & 0.33 & 0.06 & 1.20 & 4.45 & 5.94 & 7.70 & 12.23 & 50.75 & 0.06 & 0.10\\
DialoGPT & 3.76 & 0.73 & 4.87 & 10.43 & 13.50 & \textbf{11.25} & \textbf{19.23} & 41.68 & 4.24 & 26.82\\
Maria& 4.28 & 1.46 & 5.05 & 11.58 & 26.08 & 9.77 & 16.82 & 42.50& 5.47 & 29.23\\
Maria-Audio &2.99 & 0.95 & 4.14 & 10.18 & 20.55 &9.30 & 15.82 & 43.45 & 5.33 & 24.58\\
Maria-C$^3$KG& \textbf{4.69} & \textbf{1.61} & \textbf{5.29}& \textbf{11.81} & \textbf{26.80} & 10.97 & 18.78 & \textbf{40.88} & \textbf{5.88} & \textbf{32.82}\\
\hline
Text-only & 3.33 & 1.16 & 4.55 & 11.15 & 24.82 & 9.15& 15.42 & 44.45 & 5.47 & 26.78 \\
Audio-text & 2.90 & 0.93 & 4.11 & 10.17 & 20.42 & 9.23 & 15.28 & 44.71 & 5.15 & 23.93\\
C$^3$KG-text & 4.31 & 1.54 & 5.07 & 11.63 & 26.64 &9.52 & 16.36 & 42.81 & 5.85 & 31.23\\
Maria-ATOMIC & 4.59 & 1.56 & 5.18 & \textbf{11.85} & \textbf{27.04} & 9.81 & 16.89 & 41.79 & 5.71 & 31.14\\
\hline
\end{tabular}}
\caption{Evaluation results of different baselines on TikTalk. We evaluate models in three types of automatic metrics, including relevance, ranking, and diversity metrics.}
\label{tab:baseline_performance}
\end{table*}
\paragraph{Maria-C$^3$KG} introduces common-sense convesational knowledge as input based on Maria. It consists of two modules, one retrieves conversational common sense from the knowledge base as hidden context, and the other takes information from multiple sources as input to generate responses.
As we observe in Figure~\ref{fig:information_distribution}, about 33\% of the responses in TikTalk are derived from external knowledge. Therefore, we use C$^3$KG~\cite{li2022c3kg}, a Chinese common-sense conversational knowledge graph, to serve as implicit contexts for the dialogue corpora. C$^3$KG translates ATOMIC~\cite{sap2019atomic} into Chinese and proposes some additional conversational knowledge. ATOMIC organizes common-sense knowledge in a triad of <head, relation, tail>. It collects people's emotions and reactions to daily events and builds knowledge by reasoning skills that can generate coherent responses. Based on it, C$^3$KG constructs three types of dialogue flows, including event, concept, and emotion-based empathy flow. Since dialogues are hierarchical, the first two could help transform events and topics. Emotion-based empathy flow helps fine-grained emotional understanding and reasoning. Note that we only introduce these dialogue flows added by C$^3$KG in this model.
For each conversation, we use Sentence-BERT ~\cite{reimers2019sentence} to encode textual information. By measuring the semantic similarity, we retrieve the heads of dialogue flow in C$^3$KG which are most relevant to the dialogue context. Next, we obtain the corresponding common-sense conversational knowledge and use them as knowledge context inputs. In addition, we added MKP (Masked Knowledge Prediction) as a training objective. We use a probability of 15\% to mask textual context and knowledge, which in response is 70\%. In our implementation, we adopt ResNet101~\cite{he2016deep} to extract the video frame features. The maximum length of the input sequence is 50 frames and 180 textual tokens. We initialize the parameters with a pre-trained model Bert-base-Chinese and train 20 epochs with a linear decay of learning rate from $3\times 10^{-5}$.
\subsection{Main Results}
The experimental results are shown in Table~\ref{tab:baseline_performance}. We can see that livebot yields a poor performance on all metrics, probably because it is designed for live comment generation, in which the comments do not originate from definite contexts and there are many repetitions and similar sentences. Compared with DialoGPT, Maria performs better in relevance and diversity metrics, suggesting that multi-modal contexts can provide richer information for response generation. While DialoGPT outperforms in the ranking metrics. This may be because it is pre-trained on a large number of textual conversation corpora, and could rank various correct responses better. In addition, Maria-Audio's performance decreases compared to the original model. It is because only 10\% of the responses in TikTalk are correlated with audio, and our simple changes may not be able to extract audio information effectively. Therefore, a better way of extracting and aligning multi-modal information should be explored. On the other hand, Maria-C$^3$KG improves over Maria in all metrics. It demonstrates that introducing common-sense conversational knowledge as hidden contexts is helpful in spontaneously generated chitchat dialogues.
Furthermore, the low values of these baseline metrics indicate that they are far from human performance in real-world multi-modal dialogues. This suggests that TikTalk differs significantly from previous tasks and datasets and requires more powerful multi-modal dialogue models. There is still a large room for improvement on TikTalk.
\subsection{Ablation Study}
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{pics/case4.PNG}
\caption{Data examples from the test set and responses generated by baselines(Translated from Chinese). (a)Five people are dancing in imitation of anime characters. (b)A singing show with roaming feathers as its stage effects. (c)A pair of siblings are hard at work farming. (d)A blogger performs a song.}
\label{fig:case_study}
\end{figure*}
The results of our ablation study are shown in the last four rows of Table~\ref{tab:baseline_performance}. We evaluate Maria, Maria-Audio, and Maria-C$^3$KG without visual information, and the Maria-ATOMIC model with ATOMIC introduced as external knowledge, respectively. We observe a decrease in the results of the first three models with visual context removed, consistent with the settings of the multi-modal dialogue scenario. Dynamic images provided by the video improve the quality of the dialogue response generation. In contrast, the model that incorporates audio still performs the worst, suggesting that simply fusing audio features leads to noise in the dialogues. How to introduce audio information needs to be carefully considered. In addition, Maria-C$^3$KG without video information, i.e., the C$^3$KG-text setting, shows less degradation in the results of relevance and diversity metrics than the original Maria without conversational common sense. Consecutive video frames may capture temporal visual information, but lack an understanding of fine-grained visual contents. Therefore, it leads to less relevance of visuals than external knowledge when generating responses, which is contrary to the data analysis in Section~\ref{sec:data comparison}. Moreover, Maria-ATOMIC performs lower than Maria-C$^3$KG on most metrics, which suggests that the dialogue flow contained in C$^3$KG is more effective in driving the conversation compared to the social common-sense knowledge in ATOMIC.
\subsection{Case Study}
We select a few data examples from the test set and responses generated by the above baselines and further investigate the characteristics of TikTalk and the performance of these models. The comparison results are shown in Figure~\ref{fig:case_study}. As can be seen, the human dialogues in these cases are highly personalized, while most of the responses generated by baselines are of poor quality. Specifically, in the multi-modal dialogue scenarios of Figure~\ref{fig:case_study}(a) and (b), the models produce universal responses, which fall far short of human performance. In Figure~\ref{fig:case_study}(c), Maria-C$^3$KG infers a reasonable response closer to one of the ground truths. The response from Maria-Audio in Figure~\ref{fig:case_study}(d) involves audio information. These two cases illustrate that the external knowledge and audio features we introduce may play a role. In summary, TikTalk requires complex understanding and reasoning capabilities due to the diversity of video scenarios and user personalities. While partial baselines can occasionally produce some reasonable responses, they stay far from meeting the expectations of real-world multi-modal dialogues.
\section{Introduction}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{pics/lizi2.PNG}
\caption{Examples of dialogues from our TikTalk dataset. (a) A seagull carrying food from a store. (b) A dance show. (c) A blogger talking about her strange experiences.}
\label{fig:tiktalk_example}
\end{figure}
The ability to perceive and interact with real-world multi-modal information is a significant manifestation of conversational robots' progress toward general AI. Recently, with the excellent performance of some large-scale pre-training language models\cite{zhang2019dialogpt, bao2019plato, adiwardana2020towards, roller2020recipes, zhou2021eva} on textual dialogue tasks, there has been widespread interest in introducing multi-modal information into conversations.
A variety of multi-modal dialogue datasets have been proposed to facilitate research in this area. Most previous works introduce images as an additional modality and collect conversational corpora from various sources, such as crowdsourcing, movies and TV series, social media platforms, textual dialogue datasets, etc. VisDial~\cite{das2017visual} and AVSD~\cite{alamri2019audio} focus on understanding visual content using multi-turn multi-modal question answering tasks. Livebot~\cite{ma2019livebot} and VideoIC~\cite{wang2020videoic} investigate the task of single-turn live comment generation in the video. IGC~\cite{mostafazadeh2017image}, Image-Chat~\cite{shuster2020image} and PhotoChat~\cite{zang2021photochat} ask crowd-workers to create engaged conversations based on the given images. OpenViDial~\cite{meng2020openvidial, wang2021openvidial} and $M^3$ED~\cite{zhao2022m3ed} extract dialogues of characters in movies and TVs and corresponding clips. MMChat~\cite{zheng2021mmchat} and MMDialog~\cite{feng2022mmdialog} collect images and conversations from social media platforms.
However, there are still some gaps between these previous datasets and the real-world daily chitchat dialogues. First of all, the conversational forms of the corpora of visual question answering are too homogeneous. The live comments in the video have no explicit context and are more commentary oriented, filled with repeated data used to liven up the atmosphere. In addition, the conversation contents extracted from movies and TV series or developed by the annotators from the given pictures are also not chitchat corpora naturally generated in real-world dialogue scenarios. The former is some characters' lines designed to drive the plots forward, and the latter may be highly image-dependent.
Secondly, real-world multi-modal dialogue scenarios may contain diverse contexts with temporal order, such as vision, audio, and text. In contrast, the static images in most datasets provide limited information, which restricts the diversity of conversation participants in terms of topics. Furthermore, there are distinctive expressions and popular cultures in different language contexts, which are difficult to translate or migrate to other languages. Therefore, it is necessary to construct multi-modal dialogue datasets for unique languages.
In this paper, we propose TikTalk, a Chinese multi-modal dialogue dataset containing video-based real-world human chitchat corpora, which are open-domain and spontaneously generated by users without intentionally creating a dataset. Our data is derived from videos and their corresponding multi-turn comments on a video social platform. Various multi-modal information of video, such as dynamic images, audio, etc., provide context of diverse topics and grounded conversations. Accordingly, TikTalk benchmark requires higher plausibility of responses within the restriction of rich context sources. It demands for a deep understanding of multi-modal information and advanced reasoning ability of conversation models.
In data analysis, we annotate which context source a dialogue response is grounded to and compare different datasets. The quantitative results show that in TikTalk, the ratio of responses related to images, audio, and requiring external knowledge is 42\%, 10\%, and 34\%, respectively. As shown in Figure~\ref{fig:tiktalk_example} (a), the chitchat originates from video content. The ``other shows'' and the analysis of the dances in Figure~\ref{fig:tiktalk_example} (b) require some external knowledge. While the conversation in Figure~\ref{fig:tiktalk_example} (c) builds upon audio content. The complex information sources in the real-world chitchat in TikTalk present a significant challenge for the multi-modal dialogue task, which provides an opportunity to test the capabilities of multi-modal grounded dialogue models.
We evaluate several baselines at TikTalk. Most of the current multi-modal conversation models are image-based\cite{zang2021photochat, liang2021maria, sun2021multimodal}. To address the above challenges, we obtain two new baselines by introducing audio and external knowledge as inputs for dialogue response generation to the existing image-based dialogue model, respectively. We adopt three types of automatic metrics: ranking, relevance, and diversity metrics, with case studies of the responses generated by models. Experimental results show that these models are still far from human performance, demonstrating considerable room for improvement on TikTalk.
\section{Related Work}
\begin{table*}[t]
\centering
\resizebox{\textwidth}{!}{
\begin{tabular}{lccccc}
\textbf{Dataset} & \textbf{Modalities} & \textbf{Dialogue Source} & \textbf{Utterance} & \textbf{Language} & \textbf{Origin} \\
\hline
\hline
VisDial~\cite{das2017visual} & Image-text & Crowdsourcing & 2.47M & English & MS COCO \\
\hline
AVSD~\cite{alamri2019audio} & Video-text & Crowdsourcing & 236K & English & Charades \\
\hline
IGC~\cite{mostafazadeh2017image} & Image-text & Crowdsourcing & 25.3K& English&VQG\\
\hline
Image-Chat~\cite{shuster2020image} & Image-text & Crowdsourcing & 401K & English & YFCC100M\\
\hline
PhotoChat~\cite{zang2021photochat} & Image-text & Crowdsourcing & 156K & English & OpenImage V4 \\
\hline
MMChat~\cite{zheng2021mmchat} & Image-text & Social Media &314K & Chinese & Weibo\\
\hline
MMDialog~\cite{feng2022mmdialog} & Image-text & Social Media & 4.92M & English & Social Media\\
\hline
OpenViDial~\cite{meng2020openvidial} & Image-text & Movies\&TVs & 1.1M & English & Movies\&TVs\\
\hline
OpenViDial 2.0 ~\cite{wang2021openvidial} & Image-text & Movies\&TVs & 5.6M & English & Movies\&TVs\\
\hline
$M^{3}$ED~\cite{zhao2022m3ed} & Video-text & TVs & 24.4K & Chinese & TVs\\
\hline
Livebot~\cite{ma2019livebot} & Video-text & Social Media & 896K & Chinese & Bilibili \\
\hline
VideoIC~\cite{wang2020videoic} & Video-text & Social Media & 5.33M & Chinese & Bilibili \\
\hline
\textbf{TikTalk(Ours)} & Video-text & Social Media & 827K & Chinese & Douyin\\
\hline
\end{tabular}}
\caption{A summary of main multi-modal dialogue datasets and their characteristics.}
\label{tab:dataset_comparison}
\end{table*}
Our work is related to two categories of previous works: multi-modal dialogue datasets and models.
\subsection{Multi-modal Dialogue Datasets}
Dialogue is a long-standing domain of research interest to NLP, speech and computer vision communities. Many textual dialogue datasets are available~\cite{sordoni2015neural, wu2016sequential, lison2016opensubtitles2016, zhang2018personalizing, dziri2018augmenting, wang2020large}. Also there have been efforts to introduce multi-modal information into dialogue datasets to facilitate the research on multi-modal dialogues. We summarize the main characteristics of the different multi-modal dialogue datasets in Table~\ref{tab:dataset_comparison}.
Several datasets ~\cite{das2017visual, de2017guesswhat} are for addressing image-based multi-turn question answering tasks. The Q\&As of AVSD~\cite{alamri2019audio} are derived from audio visual scenes. These datasets manually create question-and-answer pairs to measure models' performance on objective perception for given multi-modal information, rather than our dataset focused daily dialogues.
Some works use crowdsourcing or collect comment-reply pairs from social media to generate image-based dialogue corpora. IGC ~\cite{mostafazadeh2017image} and Image-Chat~\cite{shuster2020image} ask crowd-workers to talk about given images and related topics with explicit guidelines. In most cases, these utterances are highly image-dependent and expensive to build.
MMChat~\cite{zheng2021mmchat} and MMDialog~\cite{feng2022mmdialog} collect images and dialogues from social media platforms. MMDialog may contain images in both context and responses. Compared to these kinds of works, our dataset is grounded to a video containing temporal vision, audio, and text information, rather than static images.
Some video based dialogue datasets are from movies or TV shows. OpenViDial datasets ~\cite{meng2020openvidial, wang2021openvidial} extract dialogues from movies and TVs and use the corresponding screenshots as visual context. However, lines of characters are designed to drive plots and have some artistic nature. M$^3$ED~\cite{zhao2022m3ed} is a video-based emotional dialogue dataset from TVs. It selects videos with distinct facial expressions and voices for the emotion prediction in each utterance. In this paper, manual labeling results show that the dialogues from this kind of source are less grounded to vision information.
Some other works aim to generate comments upon a moment of a long video~\cite{ma2019livebot, wang2020videoic}, which focuses on video understanding rather than dialogues. Some works construct multi-modal dialogues by retrieving images with texts ~\cite{lee2021constructing, liang2021maria}; however, such dialogues are generated by users without considering multi-modal context and thus they are not appropriate to be regarded as typical real-world multi-modal dialogues. In contrast, our TikTalk dataset collect comment-reply pairs from short videos platform where users chat with each other after watching the videos, which is more natural real-world dialogues with multi-modal context.
\subsection{Multi-modal Dialogue Models}
With the rapid development of large-scale pre-training language models, such as BERT~\cite{devlin2018bert}, UniLM~\cite{dong2019unified}, and GPT-3~\cite{brown2020language}, the performance of textual dialogues has been significantly improved~\cite{zhang2019dialogpt, bao2019plato, bao2020plato, bao2021plato, adiwardana2020towards, roller2020recipes, zhou2021eva}. Due to the widespread interest in multi-modal dialogue, many models have emerged to address this problem.
Some previous works have made preliminary explorations on the retrieval model and the generation model~\cite{mostafazadeh2017image, shuster2020image, feng2022mmdialog}. \citet{huber2018emotional} present an image-based dialogue generation model that combines scene and sentiment understanding. ImgVAE~\cite{yang2021open} and Maria~\cite{liang2021maria} construct corresponding images for textual dialogues in advance and then generate responses based on multi-modal information. ImgVAE has an image reconstructor that generates images as input based on textual context; whereas, Maria retrieves a correlated image for each conversation by a cross-modal matching model. Divter~\cite{sun2021multimodal} could generate textual or visual responses for the given dialogue context. There are some other works incorporating vision and language for multi-modal dialogue tasks~\cite{zang2021photochat, zheng2021mmchat}. Since our dataset is derived from dialogues under videos, we propose baseline models that can take vision, audio, and knowledge graphs as input and output text responses.
|
\section{Introduction}
Real root isolation has been considered as a symbolic way to compute the real roots of a univariate polynomial or function. Real root isolation for polynomials is of fundamental importance in computational real algebraic geometry (\emph{e.g.}, it is a routing sub-algorithm for cylindrical algebraic decomposition \cite{collins1975quantifier}), and there are many efficient algorithms and popular tools in computer algebra systems such as Maple and Mathematica to isolate the real roots of polynomials.
There are also some research on real root isolation for non-polynomial functions: In \cite{mitchell1990robust}, interval arithmetic has been used to isolate real roots of a large class of analytic functions. The algorithm there mainly considers deciding the exclusion or the inclusion of a root in an interval. Recent years, Strzebo{\'n}ski presented a real root isolation procedure for univariate \emph{exp–log–arctan functions} \cite{strzebonski2008real,strzebonski2012real}.
The procedure is based on the technique of weak Fourier sequence and semi-Fourier sequence. Meanwhile, Strzebo{\'n}ski studied real root isolation for the so-called tame elementary functions using the concept of false derivative \cite{strzebonski2009real}. The completeness of these algorithms of Strzebo{\'n}ski relies on Schanuel's conjecture.
In \cite{achatz2008deciding,mccallum2012deciding}, for developing a decision procedure for a certain class of
sentences of first order logic involving integral polynomials and
a certain specific analytic transcendental function trans$(x)$, Achatz, McCallum and Weispfenning provided a recursive real root isolation algorithm for a certain kind of
generalised integral polynomial in trans$(x)$: $f(x,\text{trans}(x))$ (where $f(x, y)$ is a given polynomial in $y$ whose coefficients are elements of the ring of fractions of $\mathbb{Z}[x]$ with respect to powers of a specific
integral polynomial $d(x)$), using \emph{pseudo-differentiation} and
Rolle’s theorem in the spirit of \cite{collins1976polynomial}, and also a classical result of
Lindemann \cite{shidlovskii2011transcendental}. The list of
transcendental functions to which their decision method directly
applies includes $\exp(x),\ln(x)$ and $\arctan(x)$. The decision procedure when trans$(x)=\exp(x)$ is implemented and the corresponding root isolation algorithm is believed to be more efficient than the one
provided in \cite{maignan1998solving}. Inspired by \cite{achatz2008deciding} and \cite{strzebonski2011cylindrical}, Xu \emph{et al.} studied a class of quantified exponential polynomial formulas extending polynomial ones \cite{xu2015quantifier}.
The isolation algorithm for an exponential polynomial used therein is modified from the one in \cite{achatz2008deciding}.
In \cite{huang2018positive}, Huang \emph{et al.} developed some algorithms for isolating the positive roots of a class of functions called poly-powers, while in \cite{wang2021symbolic}, Wang and Xu focused on isolating the real roots of a certain class of radical expressions.
A {\em mixed trigonometric-polynomial} (MTP) refers to a function of the form $f(x,\sin x, \cos x)$ with $f$ being a trivariate polynomial having rational coefficients. In this paper, we present an algorithm ``isolating" all the real roots (there are probably infinitely many of them) of such an MTP. While all inverse trigonometrical functions are exp-log-arctan functions, this is not true for the MTPs. In fact, since an exp-log-arctan function has finitely many real roots (\cite[Claim 4]{strzebonski2012real}), any MTP with infinitely many real roots is not an exp-log-arctan function (\emph{e.g.}, $\sin x$, $\cos x$, $2x\sin x+1$, \emph{etc.}). On the other hand, an MTP is usually not a tame elementary function since the variable varies over the whole $\R$. Hence, our result is not covered by the ones of Strzebo{\'n}ski. However, in Section \ref{sec:IBR}, while isolating the real roots of an MTP in a bounded interval, we take advantages of the weak Fourier sequences of a function introduced in \cite{strzebonski2008real,strzebonski2012real} (it can be also practicable to use the false derivative technique in \cite{strzebonski2009real} since any MTP restricted to a bounded interval is tame). Nevertheless, we do not need Schanuel's conjecture since we only consider trigonometric functions in one variable. Note that both the methods used in \cite{strzebonski2012real} and \cite{mccallum2012deciding} are applicable for isolating the real roots of the function of the form $f(\arctan x,x)$, with $f$ being a bivariate polynomial having rational coefficients. This is equivalent to isolating the real roots of the function $f(\hat{x},\tan \hat{x})$ with $\hat{x}\in(-\frac{\pi}{2},\frac{\pi}{2})$ (see, \emph{e.g.}, \cite[Section 6]{mccallum2012deciding}). The goal of the present paper is to isolate all the real roots of an MTP,
which is a non-trivial generalisation.
The topic of automatically proving inequalities involving trigonometric functions, which is closely related to the topic of the present paper, has also drawn more and more attention these years due to their applications to engineering problems and different branches of science \cite{chen2020automated}. Usually, the term ``MTP inequalities" refers to inequalities of the form $g(x)>0$ or $g(x)\geq0$ with $g$ being an MTP. A so-called ``Natural Approach" introduced in \cite{mortici2011natural} by Mortici, approximates $\sin x$ and $\cos x$ by their Taylor polynomials and reduce the problem of proving an MTP inequality to the problem of proving some polynomial (or rational function) inequalities. That approach leads to a series of subsequent studies \cite{chen2012sharp,bercu2016pade,bercu2017natural,makragic2017method,chenliu2016}, in which a great deal of inequalities involving trigonometric functions are solved automatically. Later in \cite{chen2020automated}, Chen and Liu provided a complete algorithm proving MTP inequalities and discussed systematically, for the first time, the termination of the algorithm. However, these algorithms only prove MTP inequalities over bounded intervals (\emph{e.g.}, $(0,\frac{\pi}{2})$ or $(0,\frac{\pi}{4}]$, \emph{etc.}). Till recently, Chen and Ge \cite{chen2022automated} tried to develop a procedure automatically proving MTP inequalities on the unbounded interval $(0,\infty)$. Unfortunately, we found some mistakes therein. Therefore, the problems of proving an MTP inequality, and of isolating real roots in an unbounded interval remain open so far.
In this paper, an algorithm ``isolating" all the real roots of an MTP is provided (Algorithm \ref{alg:complete-algorithm}). It automatically divides the real roots into two parts: one consists of finitely many ``bounded" roots in an interval $[\mu_-,\mu_+]$ with $\mu_-<0<\mu_+$, while the other consists of probably countably many ``periodic" roots in $\mathbb{R}\backslash[\mu_-,\mu_+]$. For the bounded roots, the algorithm returns some isolating intervals and the multiplicity of each of them. For those periodic roots greater than $\mu_+$, it returns finitely many mutually disjoint small intervals $I_i\subset[-\pi,\pi]$, integers $c_i>0$ and multisets of root multiplicity $\{m_{j,i}\}_{j=1}^{c_i}$ such that any periodic root $t>\mu_+$ is in the set $(\sqcup_i\cup_{k\in\mathbb{N}}(I_i+2k\pi))$ and any interval $I_i+2k\pi\subset(\mu_+,\infty)$ contains exactly $c_i$ periodic roots with multiplicities $m_{1,i},...,m_{c_i,i},$ respectively. For those periodic roots less than $\mu_-$, it returns some similar results. The algorithm is implemented with {\tt Maple2021}. A large number of examples (either from the literature or randomly generated) have been tested to show its effectiveness and efficiency.
In particular, this is the first algorithm to decide whether or not a given MTP has infinitely many real roots: it does iff the algorithm returns at least one periodic root. And of course, the algorithm is also able to decide whether or not an MTP has a real root (when there are real roots, we can compute the multiplicity of all of them). This allows us to prove or disprove an MTP inequality over $\mathbb{R}$. Our results also indicate, for the first time, that the ``distributions" of the roots of an MTP in the ``periods" $(-\pi,\pi]+2k\pi$, with $k>0$ (\emph{resp.}, $k<0$) sufficiently large, share a same pattern, which seems to be a basic law about the real roots of MTPs.
Besides, the method used to isolate the roots in $[\mu_-,\mu_+]$ is applicable to any other bounded interval as well. It takes advantages of the weak Fourier sequence technique and deals with the intervals period-by-period without scaling the coordinate (see, \emph{e.g.}, \cite[Algorithm 4.1, Step 2]{chen2022automated}) so to keep the length of the sequence short. An interesting problem is to compare the efficiency of the false derivative technique with the weak Fourier technique for isolating real roots of an MTP in a bounded interval, which is beyond the scope of this paper.
As a direct application, we discuss briefly at the end of the paper how one can modified our new approaches to prove or disprove an MTP inequality on an interval of the form $(a,b)$, $[a,b)$, $(a,b]$, $[a,b]$, with $a,b\in\Q\cup\{\pm\infty\}$.
The rest of the paper is organized as follows.
The next section introduces the problem and some basic results that will be used from time to time. In Section \ref{sec:farroots}, we introduce the algorithm ``isolating" periodic roots while in Section \ref{sec:IBR} we provide the procedure computing the isolating intervals for bounded roots. The main algorithm ``isolating" all real roots of an MTP is presented in Section \ref{sec:wholealgorithm}. And in Section \ref{sec:experiments}, experimental results are provided to indicate the effectiveness and efficiency of the algorithm. Finally, Section \ref{sec:conclusion} summarizes the paper and provides some remarks relating to the topic of proving MTP inequalities.
\section{The Problem}\label{sec:Preprocessing}
\subsection{Some basic results}
Denote by $\mathbb{N}, \Z, \Q, \R$ and $\C$ the sets of natural, integer, rational, real and complex numbers, respectively. For any subset $A\subset\mathbb{R}$, set $A_{>0}=\{a\in A\,|\,a>0\}$. The meanings of the notation $A_{\geq0}$, $A_{\leq0}$ and $A_{<0}$ are similar. For any $\mathbb{A}\in\{\Z, \Q, \R,\C\}$, $\mathbb{A}[x_1,\ldots,x_m]$ denotes the corresponding ring of polynomials in the variables $x_1,\ldots,x_m$ with coefficients in $\mathbb{A}$. For any $f\in\mathbb{A}[x_1,\ldots,x_m]$, $\deg(f,x_i)$, $\lc(f,x_i)$ and $\discrim(f,x_i)$ denote the degree, the leading coefficient and the discriminant of $f$ with respect to $x_i$, respectively. Moreover, $\tdeg(f,[x_{i_1},\ldots,x_{i_{t}}])$ denotes the total degree of the subsequence of variables $[x_{i_1},\ldots,x_{i_{t}}]$.
{\bf The problem} considered in this paper is to solve the real roots of the following equation:
\begin{equation}\label{eq:goalequation}
f(x,\sin{x},\cos{x})=0,~{\rm where}~f\in \Q[x,y,z].
\end{equation}
We provide some basic observations in the following.
\begin{lemma}\label{lemma:NoLimit}
For any nonzero $f\in\Q[x,y,z]$, the real root set of $f(x,\sin{x},\cos{x})$ has no accumulation point in $\mathbb{R}$.
\end{lemma}
\begin{proof}
The complex-valued function $h(z)=f(z,\sin{z},\cos{z})$ is analytic over $\C$. By definition, $h|_{\mathbb{R}}=f(x,\sin{x},\cos{x})$. If the set of the real roots of $f$ has an accumulation point in $\mathbb{R}$, then the set of the complex roots of $h$ has an accumulation point in $\mathbb{C}$ as well, which indicates $h\equiv0$ over $\mathbb{C}$. But $h$ can not be identically $0$ over $\mathbb{C}$, since $f$ is not over $\mathbb{R}$.
\end{proof}
\begin{lemma}\label{lemma:piRealRoot}
For any $q(x)\in\Q[x]$, ``$r\pi$ is a real root of $q(x)$ for any $r\in\Q\setminus\{0\}$" iff ``$q(x)$ is the zero polynomial in $\Q[x]$".
\end{lemma}
\begin{proof}
The ``if" part is obvious. The ``only if" part follows from the fact that $r\pi$ is transcendental for any $r\in\Q\setminus\{0\}$.
\end{proof}
\begin{theorem}\label{thm:BothAlgeNum}\cite[Chapter 2.3]{shafarevich1998number}
If $\alpha\in\C$ and $\alpha\ne0$, then there is at least one transcendental number in each pair $(\alpha,\sin\alpha)$, $(\alpha,\cos\alpha)$, $(\alpha,\tan\alpha)$,
$(\alpha,\cot\alpha)$.\qed
\end{theorem}
\begin{definition}\label{definition:recip}
For a nonzero $g\in\Q[x,y]$, we define its \emph{reciprocal polynomial} $\mathcal{R}(g)$ as $g(x,\frac{1}{y})y^{\deg(g,y)}$. Additionally, we set $\mathcal{R}(0)=0$.
\end{definition}
From the definition it is obvious that
\;\;1) $\mathcal{R}(g_1g_2)=\mathcal{R}(g_1)\mathcal{R}(g_2)$ for any $g_1,g_2\in\Q[x,y]$,
\;\;2) if $g=\sum_{i=0}^{\deg(g,y)}a_i(x)y^i$ with $a_0(x)\neq0$ (\emph{i.e.}, $y\nmid g$), then $\mathcal{R}(\mathcal{R}(g))=g$, and
\;\;3) $y$ does not divide $\mathcal{R}(g)$ for any nonzero $g$.
\noindent{}We then have some key properties of $\mathcal{R}$ as follows.
\begin{proposition}\label{proposition:RkeepIrr}
Suppose $g\in\Q[x,y]$ and $g\neq cy$ for any $c\in\mathbb{Q}$. Then ``$g$ is irreducible in $\Q[x,y]$" implies ``$\mathcal{R}(g)$ is irreducible in $\Q[x,y]$".
\end{proposition}
\begin{proof}
The proof is inductive on the total degree of $g$: If $\tdeg(g,[x,y])=0$, the conclusion is trivially true since $g$ is not irreducible. Suppose that $\tdeg(g,[x,y])=n>0$ and the implication holds for any polynomial ($\neq cy$) of total degree less that $n$. Assume $\mathcal{R}(g)$ is reducible and its factorization is $c_0h_1^{p_1}\cdots h_t^{p_t}$ with $h_i$ irreducible, $\sum_{j=1}^{t}p_j\geq2$ and $c_0\in\mathbb{Q}\backslash\{0\}$. Moreover, each $h_i\neq cy$ for any nonzero $c\in\Q$ because $y\nmid\mathcal{R}(g)$. Since $g$ is irreducible and $g\neq cy$ for any $c\in\Q$, $y\nmid g$. Thus $g=\mathcal{R}(\mathcal{R}(g))=c_0\mathcal{R}(h_1)^{p_1}\cdots \mathcal{R}(h_t)^{p_t}$. By the inductive hypothesis, each $\mathcal{R}(h_i)$ is irreducible. Since $\sum_{j=1}^{t}p_j\geq2$, $g$ is reducible, which is a contradiction.
\end{proof}
For any polynomial $g\in\Q[x,y]$ of positive total degree, $g$ is said to be \emph{square-free}, if the factorization of $g$ is $cg_{1}\cdots g_{t}$ with $c\in\Q\setminus\{0\}$, $t\in\Z_{\ge1}$, each $g_i$ irreducible and $g_i$ and $g_j$ co-prime for any $1\le i\ne j\le t$. Any constant polynomial is not square-free in our definition.
\begin{proposition}\label{proposition:recipKeepSquarefree}
Suppose $g\in\Q[x,y]$ and $g\neq cy$ for any $c\in\mathbb{Q}$. Then ``$g$ is square-free" implies ``$\mathcal{R}(g)$ is square-free".
\end{proposition}
\begin{proof}
Factorize $g$ as $c_0y^\ell h_1\cdots h_t$, with $c_0\in\Q$ nonzero, $\ell=0\text{ or }1$ (corresponding to the cases $y\nmid g$ and $y\,|\,g$ respectively), $t\geq1$ (since $g\neq cy$ and $g\notin\Q$) and each $h_i$ irreducible.
First, $\mathcal{R}(g)=c_0\mathcal{R}(h_1)\cdots\mathcal{R}(h_t)$ since $\mathcal{R}(y^\ell)=1$.
Second, for each $h_i$, $h_i\neq cy$ for any $c\in\Q$ because $y\nmid h_i$.
So, by Proposition \ref{proposition:RkeepIrr}, $\mathcal{R}(h_i)$ is irreducible.
Finally, $\mathcal{R}(h_i)$ and $\mathcal{R}(h_j)$ are co-prime for $1\leq i\neq j\leq t$. Suppose they are not, then $\mathcal{R}(h_i)=c'\mathcal{R}(h_j)$ for a $c'\in\Q\backslash\{0\}$. Since $y\nmid h_i$ and $y\nmid h_j$, $h_i=\mathcal{R}(\mathcal{R}(h_i))=c'\mathcal{R}(\mathcal{R}(h_j))=c'h_j$, contradicting the fact that $h_i$ and $h_j$ are co-prime in the factorization of $g$.
\end{proof}
\subsection{Decomposing the problem into sub-problems}
Because
\[
\left\{
\begin{array}{rl}\vspace{2mm}
\sin x&= \frac{2\tan{\frac{x}{2}}}{1+\tan^2(\frac{x}{2})}\\
\cos x&= \frac{1-\tan^2{\frac{x}{2}}}{1+\tan^2({\frac{x}{2}})}
\end{array}\right.,\;\;x\neq 2k\pi+\pi~(k\in \Z),
\]
Eq. \eqref{eq:goalequation} is equivalent to
\begin{subnumcases}
{}
f(x,\sin{x},\cos{x})=0, ~~~~~~~~~~~~~~~~~~~~~x=2k\pi+\pi\label{meq:1},\\
f(x,\frac{2\tan{\frac{x}{2}}}{1+\tan^2{(\frac{x}{2})}},\frac{1-\tan^2{(\frac{x}{2})}}{1+\tan^2{(\frac{x}{2}})})=0,~ \hfill x\in (2k\pi-\pi,2k\pi+\pi), \label{meq:2_}
\end{subnumcases}
where Eq. \eqref{meq:1} is equivalent to $f(x,0,-1)=0,~x=(2k+1)\pi$.
By Lemma \ref{lemma:piRealRoot},
$(2k+1)\pi$ is a root of $f(x,0,-1)$ iff $f(x,0,-1)$ is the zero polynomial in $\Q[x]$.
Through cancelling the denominators, we define a polynomial as follows:
\begin{equation}\label{equation:gTurnedToTan}
g(x,t)=f(2x,\frac{2t}{1+t^2},\frac{1-t^2}{1+t^2})\cdot(1+t^2)^{\tdeg(f,[y,z])}\in\mathbb{Q}[x,t].
\end{equation}
Since $1+\tan^2{(\frac{x}{2})}>0$, solving the real roots of Eq. (\ref{meq:2_}) is then equivalent to solving the real roots of the equation
\begin{equation}
g(x,\tan{x})=0,~x\in(k\pi-\frac{\pi}{2},k\pi+\frac{\pi}{2}).\label{meq:2}
\end{equation}
\begin{example}\label{example:gtan}
Consider the following MTP
$$f(x,\sin{x},\cos{x})=x\cos^4{x}-2\sin{x}\cos^3{x}+\frac{4}{5}\sin^3{x}.$$
By Eq. \eqref{equation:gTurnedToTan}, we have
\begin{align*}
g(x,t)=10xt^8+20t^7-40xt^6-28t^5+60xt^4+92t^3-40xt^2-20t+10x.
\end{align*}
\end{example}
In Eq. \eqref{meq:2}, we have to deal with infinitely many intervals $(k\pi-\frac{\pi}{2},k\pi+\frac{\pi}{2})$ while $k$ varies in $\Z$. Fortunately, Proposition \ref{prop:firstbound} and Corollary \ref{cor:main_case} below, together with
Theorem \ref{thm:location-of-the-roots} in the next section will set us free from doing that.
\begin{proposition}\label{prop:firstbound}
Set nonzero polynomial $g\in\Q[x,y]$, $\delta>0$ and $U\subset\mathbb{R}$ to be a bounded set. If $|\lc(g,x)|>\delta$ for any $y\in U$, then there is an integer $N>0$ such that $x>N$ implies $\forall y\in U,\; g(x,y)\neq0$.
\end{proposition}
\begin{proof}
Suppose $\deg(g,x)=n$ and $a_i(y)\in\mathbb{Q}[y]\;(i=0,\ldots,n)$ are the coefficients of $g$ \emph{w.r.t.} $x$. Now that $U$ is bounded, there is $B_i>0$ such that $|a_i(y)|<B_i$ for any $y\in U$. For sufficiently large $x$, we have $\delta |x|^n>\sum_{i=0}^{n-1}B_i|x|^i$. Therefore, for any $y\in U$,
\[|a_n(y)x^n|>\delta|x|^n>\sum\limits_{i=0}^{n-1}B_i|x|^i\geq|\sum\limits_{i=0}^{n-1}a_i(y)x^i|.\]
This indicates that $g(x,y)\neq0$ for any $y\in U$.
\end{proof}
By Proposition \ref{prop:firstbound}, the value of the tangent function at every ``periodic'' real root of $g(x,\tan{x})$ appears near a real root of $\lc(g,x)$.
In order to locate these ``periodic'' real roots, we introduce the following definition.
\begin{definition}\label{def:potential-periodic-interval}
For any $g\in\Q[x,y]$ with $\deg(g,y)\ge 1$,
let $I$ be a finite list of open intervals $[(a_0,b_0),\ldots,(a_{s+1},b_{s+1})]$,
where $s\in\Z_{\ge0}$, $a_0=-\infty$, $b_{s+1}=+\infty$, $b_0<0$, $a_{s+1}>0$, $a_j\in \Q~(1\le j\le s+1)$ and $b_j\in \Q~(0\le j\le s)$, such that
\begin{enumerate}[label=$($\roman*$)$]
\item for each $j~(1\le j\le s)$, $a_j< b_j$ and $b_j \le a_{j+1}$,
\item for each real root of $\lc(g,x)$, there exists an interval $(a_j,b_j)~(1\le j\le s)$ containing it $($there is no root of $\lc(g,x)$ in $(a_0,b_0)$ and $(a_{s+1},b_{s+1}))$,
\item \label{item:potential-periodic-interval-3} each interval $(a_j,b_j)~(1\le j\le s)$ has exactly one real root of $\lc(g,x)$.
\end{enumerate}
We call $I$ a \emph{potential periodic interval set} of $g$.
Generally, if we replace ``exactly" in \ref{item:potential-periodic-interval-3} by ``at most", we call $I$ a \emph{general potential periodic interval set} of $g$.
\end{definition}
\begin{example}\label{example:potential-periodic-interval-set}
Consider the polynomial in Example \ref{example:gtan}:
\begin{align*
g(x,y)=10xy^8+20y^7-40xy^6-28y^5+60xy^4+92y^3-40xy^2-20y+10x.
\end{align*}
We have $\lc(g,x)=10y^8-40y^6+60y^4-40y^2+10$.
It has two real roots $-1$ and $1$ and the isolating intervals of them are $(-\frac{215}{128}, -\frac{19}{32})$ and $(\frac{19}{32}, \frac{215}{128})$, respectively.
Thus,
{\footnotesize
$$\left[\left(-\infty, -\frac{63}{16}\right), \left(-\frac{215}{128}, -\frac{19}{32}\right), \left(\frac{19}{32}, \frac{215}{128}\right), \left(\frac{63}{16}, +\infty\right)\right]$$}is a potential periodic interval set of $g$.
\end{example}
\begin{corollary}\label{cor:main_case}
Let $g\in\Q[x,y]$ be with $\deg(g,y)\ge 1$, $[(a_0,b_0),\ldots,(a_{s+1},b_{s+1})]$ be a general potential periodic interval set of $g$, and
\begin{align}
U:=\R\setminus\bigcup_{j=0}^{s+1}(a_i,b_i).
\end{align}
Then, there exists a positive integer $N$ such that
$$
\forall x\in \R ~\forall y\in U~(x>N \Rightarrow g(x,y)\neq 0).
$$
\end{corollary}
\begin{proof}
Note that $|\lc(g,x)|$ has a lower bound on $y\in U$ and $U$ is bounded.
Then, the conclusion follows from Proposition \ref{prop:firstbound}.
\end{proof}
For the polynomial $g\in\Q[x,y]$ in Eq. \eqref{meq:2},
let $U,a_j,b_j$ be the same as in Corollary \ref{cor:main_case}.
Take $N_+, N_-\in\Z_{\ge1}$ such that $\forall x\in \R ~\forall y\in U~(x>N_+ \Rightarrow g(x,y)\neq 0)$ and
$\forall x\in \R ~\forall y\in U~(x<-N_- \Rightarrow g(x,y)\neq 0)$.
There exist $k_+\in\Z_{\ge0}$ and $k_-\in\Z_{\le0}$ such that
$k_+\pi-\frac{\pi}{2}<N_+<k_+\pi+\frac{\pi}{2}$ and
$k_-\pi-\frac{\pi}{2}<-N_-<k_-\pi+\frac{\pi}{2}$, respectively.
Thus, Eq. \eqref{meq:2} can be transformed equivalently into the following three cases:
\begin{enumerate}[label=$($\ref{meq:2}\alph*$)$]
\item \label{item:MainThreeCase1}
\begin{align*}
g(x,\tan x)=0,~~~~&x\in(k\pi-\frac{\pi}{2},k\pi+\frac{\pi}{2})
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
\end{align*}
where $k_{-}\leq k \leq k_{+}\;(k\in\Z)$;
\item \label{item:MainThreeCase2}
\begin{align*}
g(x,\tan{x})=0,~~~~&x\in(k\pi-\frac{\pi}{2},k\pi+\arctan(b_0))\\
&x\in(k\pi+\arctan(a_i),k\pi+\arctan(b_i)),~{i=1,\ldots,s}\\
&x\in(k\pi+\arctan(a_{s+1}),k\pi+\frac{\pi}{2})
\end{align*}
where $k>k_{+}\;(k\in\Z)$;
\item \label{item:MainThreeCase3}
\begin{align*}
g(x,\tan{x})=0,~~~~&x\in(k\pi-\frac{\pi}{2},k\pi+\arctan(b_0))\\
&x\in(k\pi+\arctan(a_i),k\pi+\arctan(b_i)),~{i=1,\ldots,s}\\
&x\in(k\pi+\arctan(a_{s+1}),k\pi+\frac{\pi}{2})
\end{align*}
where $k<k_{-}\;(k\in\Z)$.
\end{enumerate}
A possible way to solve case \ref{item:MainThreeCase1} for $k_-\le k\le k_+$ is to go back to the original MTP $f(x,\sin x,\cos x)$ and consider the roots in $(2k_-\pi-\pi,2k_+\pi+\pi)$. Replace $x$ in $f(x,\sin x,\cos x)$ by $\ell x$ with $\ell\in\Z_{>0}$ large enough such that the roots in $(2k_-\pi-\pi,2k_+\pi+\pi)$ ``shrink" to the roots of $f(\ell x,\sin \ell x,\cos \ell x)$ in $(-\pi/2,\pi/2)$. Then unify the sine and cosine functions with the tangent function in $f(\ell x,\sin \ell x,\cos \ell x)$ to obtain some $\hat{g}(x,\tan x)$. Finally, deal with the roots of $\hat{g}(-x, -\tan x)$ and $\hat{g}(x, \tan x)$ in the interval $(0,\pi/4)$.
The reason why we do not use this method in Section \ref{sec:IBR} is that $\ell$ may be large and the ``degree" of $\tan x$ in $\hat{g}(x, \tan x)$ can be high. Instead, we will deal with case \ref{item:MainThreeCase1} period-by-period to avoid that problem.
The last case can be reduced to the second last one via replacing $g(x,\tan x)$ by $g(-x,-\tan x)$. In the next section, we focus on the asymptotic properties of the real roots corresponding to case \ref{item:MainThreeCase2}.
\section{Asymptotic Property of Roots}\label{sec:farroots}
In this section, we shall deal with case \ref{item:MainThreeCase2}, namely isolate the ``periodic'' roots of $g(x,\tan{x})$ where $g$ is a nonzero polynomial in $\Q[x,y]$.
\subsection{Detecting and locating periodic roots}
There are countably many ``periodic'' roots of $g(x,\tan{x})$, so they can not be located by the methods of traditional real root isolation.
However, these ``periodic'' roots can be divided into finitely many bunches via a general potential periodic interval set (recall Def. \ref{def:potential-periodic-interval}).
The roots in each bunch appear periodically, that is why we call them ``periodic'' roots.
In this subsection, we propose an algorithm to isolate these ``periodic'' roots by means of a general potential periodic interval set.
First, we introduce the definition of \emph{analytic delineable} and an important result about it from the work of McCallum \cite{mccallum1998improved}.
\begin{definition}[\cite{mccallum1998improved}]
Let $\vX$ denote the $(r-1)$-tuple $(x_1,\cdots,x_{r-1})$. An $r$-variate polynomial $f(\vX,x_r)$ over the reals is said to be \emph{analytic delineable} on a sub-manifold $S$ $($usually connected$)$ of $\R^{r-1}$, if
\begin{enumerate}[label=$($\arabic*$)$]
\item the portion of the real variety of $f$ that lies in the cylinder $S\times \R$ over $S$ consists of the union of the graphs of some $k\geq 0$ analytic functions $\theta_1<\cdots <\theta_k$ from $S$ into $\R$, and
\item there exist positive integers $m_1$, $\cdots$, $m_k$ such that for every $a\in S$, the multiplicity of the root $\theta_i(a)$ of $f(a,x_r)$ $($considered as a univariate polynomial in $x_r$$)$ is $m_i$.
\end{enumerate}
Also, an $r$-variate polynomial $f(\vX,x_r)$ over $\R$ is said to be \emph{degree-invariant} on a subset $S$ of $\R^{r-1}$ if the total degree of $f(p,x_r)$ $($as a univariate polynomial in $x_r$$)$ is the same for every point $p$ of S.
\end{definition}
\begin{proposition}[{\cite[Theorem 2]{mccallum1998improved}}]\label{prop:McCallum}
Let $r\geq 2$ be an integer and $f(\vX,x_r)$ be a polynomial in $\R[\vX,x_r]$ of positive total degree.
Suppose that $D(\vX):=\discrim(f,x_r)$ is a nonzero polynomial.
Let $S$ be a connected submanifold of $\R^{r-1}$ such that $f$ is degree-invariant and does not vanish identically on $S$, and $D(\vX)>0$ or $D(\vX)<0$ over $S$. Then, $f$ is analytic delineable on $S$.\qed
\end{proposition}
We restate a well-known result (which follows from Proposition \ref{prop:McCallum}) in Corollary \ref{corollary:bound-M} and redefine the concept of ``root function'' in Definition \ref{def:root-function-start-point}.
\begin{corollary}\label{corollary:bound-M}
Let $g$ be a square-free polynomial in $\Q[x,y]$ with $\deg(g,y)\ge 1$,
and $M$ be the maximal real root of the polynomial
$\lc(g,y)\cdot\discrim(g,y)$.
Then,
\begin{enumerate}
\item $g$ is analytic delineable on $(M, +\infty)$, and
\item there exist some analytic functions $\theta_1(x)<\cdots<\theta_r(x)\allowbreak\;(r\ge0)$ defined on $(M, +\infty)$ such that for any $x_0\in(M, +\infty)$, $g(x_0,\theta_i(x_0))=0$ and $\theta_1(x_0),\ldots,\theta_r(x_0)$ are all real roots of $g(x_0,y)$. \qed
\end{enumerate}
\end{corollary}
\begin{definition}\label{def:root-function-start-point}
Keep the notation in Corollary \ref{corollary:bound-M}.
We call those analytic functions $\theta_1(x),\ldots,\theta_r(x)$ the
\emph{terminal-root-functions} of $g$ and $M$ the \emph{start-point} of the terminal-root-functions of $g$.
\end{definition}
\begin{example}\label{example:terminal-root-function}
Consider the polynomial $g(x,y)$ in Example \ref{example:potential-periodic-interval-set}.
We have $\lc(g,y)=-10x$ and
$\discrim(g,y)=-703687441776640000\cdot(163840000x^8 + 261529600x^6 + 83912256x^4 - 13333725x^2 - 10857025)$.
After isolating all real roots of them, we see that the maximal real root $M$ of $\discrim(g,y)$, which is in $(0,1)$, is the start-point of the terminal-root-functions of $g$.
And there are two terminal-root-functions of $g$, which are defined on $(M,+\infty)$ $($see Fig. \ref{fig:graph-Terminal-Root-Function}$)$.
\input{fig/root_function_fig.tex}
\end{example}
In Theorem \ref{thm:location-of-the-roots} below, we compute a bound that guarantees well-defined terminal-root-functions $($as in Corollary \ref{corollary:bound-M}$)$, which are essential to locating the unbounded roots.
\begin{theorem}\label{thm:location-of-the-roots}
Let $g$ be a square-free polynomial in $\Q[x,y]$ with $\deg(g,y)\ge 1$,
$[(a_0,b_0),\ldots,(a_{s+1},b_{s+1})]$ be a general potential periodic interval set of $g$,
$k'$ be a natural number such that $M'=(k'+1/2)\pi$ is an upper bound of the start-point $M$ of the terminal-root-functions of $g$ and all real roots of the polynomial
$$\prod_{j=0}^{s}g(x,b_j)g(x,a_{j+1}),$$
and $\theta_1(x),\ldots,\theta_r(x)$ be the terminal-root-functions of $g$.
Then:
\begin{enumerate}[label=$($\arabic*$)$]
\item\label{item:thm:location-of-the-roots1} For each $i~(1\le i \le r)$, there exists
$j_i~(0\le j_i \le s+1)$ such that $\theta_i(x)\in (a_{j_i},b_{j_i})$ whenever $x\in (M',+\infty)$.
\item\label{item:thm:location-of-the-roots2}
For each real root $\xi$ of $g(x,\tan x)$ greater than $M'$, there exist $k>k'~(k\in \N)$ and $i$ $(1\le i \le r)$ such that
$$\xi\in\left(\arctan(a_{j_i})+k\pi, \arctan(b_{j_i})+k\pi\right).$$
\item\label{item:thm:location-of-the-roots3}
For each $i~(1\le i \le r)$ and any $k>k'~(k\in \N)$, there exists $$\xi\in\left(\arctan(a_{j_i})+\allowbreak k\pi, \arctan(b_{j_i})+k\pi\right)$$ such that $\tan {\xi}=\theta_i(\xi)$.
\end{enumerate}
\end{theorem}
\begin{example}\label{example:thm:location-of-the-roots}
Consider the polynomial $g(x,y)$ in Example \ref{example:potential-periodic-interval-set}.
In Example \ref{example:potential-periodic-interval-set}, we compute a potential periodic interval set of $g(x,y)$:
{\footnotesize
$$\left[\left(-\infty, -\frac{63}{16}\right), \left(-\frac{215}{128}, -\frac{19}{32}\right),\allowbreak \left(\frac{19}{32}, \frac{215}{128}\right), \left(\frac{63}{16}, +\infty\right)\right].$$}And the isolating intervals of the real roots of the polynomial
{\footnotesize
$$g(x,-\frac{63}{16})\cdot g(x,-\frac{215}{128})\cdot g(x,-\frac{19}{32})\cdot g(x,\frac{19}{32})\cdot g(x,\frac{215}{128})\cdot g(x,\frac{63}{16})$$}are $(-8, -7)$, $(-4, -3)$, $(-1, 0)$, $(0, 1)$, $(3, 4)$ and $(7, 8)$.
In Example \ref{example:terminal-root-function}, we see that the start-point $M$ of the terminal-root-functions of $g$ is less than $1$.
Thus, we take $k'=3$ such that $(k'+1/2)\pi$ is an upper bound of $M$ and the real roots computed above.
Note that the two terminal-root-functions of $g$ are bounded by $(-\frac{215}{128},-\frac{19}{32})$ when $x\in((k'+1/2)\pi,+\infty)$ $($see Fig. \ref{fig:thm:location-of-the-roots}$)$.
\input{fig/e37}
So, for every $k > 3~(k \in \Z)$, there are at least two real roots of $g(x,\tan{x})$ in $(k\pi-\arctan{\frac{215}{128}},k\pi-\arctan{\frac{19}{32}})$.
\end{example}
\begin{remark}\label{remark:decideexistence}
For each interval $(a_j,b_j)~(0\le j\le s+1)$, it may happen that none of the terminal-root-functions $\theta_i(x)$ is bounded by $(a_j,b_j)$ when $x>M'$. It is also possible that there can be one or more than one terminal-root-functions $\theta_{i_1}(x),\ldots,\theta_{i_t}(x)$, with $t\in\Z_{\ge1}$, that are bounded by $(a_j,b_j)$ when $x>M'$. By \ref{item:thm:location-of-the-roots2} and \ref{item:thm:location-of-the-roots3} in Theorem \ref{thm:location-of-the-roots}, there is no real root in the set $(M',\infty)\cap\tan^{-1}(a_j,b_j)$ in the former case, while in the latter case, there are at least $t$ real roots in each interval $(\arctan a_j+k\pi,\arctan b_j+k\pi)$ contained in $(M',\infty)$
$($in Line \ref{Line:alg1-detect-implicit-function} of Algorithm \ref{alg:AlgFarZero}, we distinguish these two cases$)$.
Denote these roots by $x^{(1)}_k<\cdots<x^{(t'_k)}_k$ for some $t'_k\geq t$. Since Theorem \ref{thm:location-of-the-roots} is valid for isolating intervals $(a_j,b_j)$ of arbitrarily small ``arc-tangent-length" $($\emph{i.e.}, $\arctan b_j-\arctan a_j)$ and any $M'$ large enough, we have
\[\lim_{k\rightarrow\infty}(x^{(t'_k)}_k-x^{(1)}_k)=0.\] This indicates that it is impossible to find real numbers $c_1,c_2\in(-\pi/2,\pi/2)$ such that they isolate, for instance, $x^{(1)}_k, x^{(2)}_k$ and $x^{(3)}_k$ $($if $t>2)$ uniformly. That is, there are no such $c_1,c_2$ satisfying \[x^{(1)}_k<c_1+k\pi<x^{(2)}_k<c_2+k\pi<x^{(3)}_k\] for all sufficiently large integer $k$. This is why we only ``isolate" them with the interval $(\arctan a_j,\arctan b_j)+k\pi$ in the sense that every $x^{(\iota)}_k$ is contained in it. Nevertheless, in the next subsection we will prove that $t'_k=t$ for sufficiently large $k$, which means we can compute the exact number $t$ of real roots in the interval $(\arctan a_j,\arctan b_j)+k\pi$ that is sufficiently far from zero.
\end{remark}
\noindent{}\emph{\bf Proof of Theorem \ref{thm:location-of-the-roots}.}
\ref{item:thm:location-of-the-roots1}
Note that $g(x,a_j)~(1\le j \le s+1)$ and $g(x,b_j)~(0\le j \le s)$ have no roots in $(M',+\infty)$.
Then, for every $i~(1\le i\le r)$ and for any $x\in(M',+\infty)$, $\theta_i(x)\not\in\{b_0,a_1,b_1,\ldots,b_s,a_{s+1}\}$.
By Proposition \ref{prop:McCallum}, $\theta_i(x)$ is continuous.
So, there exists $j_i~(0\le j_i\le s+1)$ such that $\theta_i((M',+\infty))\subseteq (a_{j_i},b_{j_i})$ or
there exists $j_i~(0\le j_i\le s)$ such that $\theta_i((M',+\infty))\subseteq(b_{j_i},a_{j_i+1})$.
By Corollary \ref{cor:main_case}, for sufficiently large $x$, $\theta_i(x)\in \cup_{j=0}^{s+1}(a_j,b_j)$. Therefore, $\theta_i((M',+\infty))\subseteq (a_{j_i},b_{j_i})$ for some $0\leq j_i\leq s+1$.
\ref{item:thm:location-of-the-roots2}
If $\xi$ is a real root of $g(x,\tan{x})$ such that $\xi>M'$, then there exists $i~(1\le i\le r)$ such that
$\theta_{i}(\xi)=\tan{\xi}$. By \ref{item:thm:location-of-the-roots1}, $\theta_i((M',+\infty))\subseteq (a_{j_i},b_{j_i})$. So, $\tan{\xi}\in (a_{j_i},b_{j_i})$. Thus, there exists $k~(k>k')$ such that $\xi\in (\arctan(a_{j_i})+k\pi,\arctan(b_{j_i})+k\pi)$.
\ref{item:thm:location-of-the-roots3}
By \ref{item:thm:location-of-the-roots1}, for each $i~(1\le i\le r)$, $\theta_i((M',+\infty))\subseteq (a_{j_i},b_{j_i})$. So, for any $k~(k>k')$, we have
\begin{align*}
\theta_i(\arctan(a_{j_i})+k\pi)>a_{j_i}=&\tan(\arctan(a_{j_i})+k\pi),\\
\theta_i(\arctan(b_{j_i})+k\pi)<b_{j_i}=&\tan(\arctan(b_{j_i})+k\pi).
\end{align*}
Hence there exists $\xi\in(\arctan(a_{j_i})+k\pi, \arctan(b_{j_i})+k\pi) $ such that $\tan \xi-\theta_i(\xi)=0$.
$\hfill{\square}$\par
\begin{comment}\textcolor{blue}{(The symbols ``$r$" and ``$s$" in Theorem \ref{thm:location-of-the-roots} and Algorithm \ref{alg:AlgFarZero} are ``swapped". And sometimes ``$i$" and ``$j$" are used casually, \emph{e.g.}, in the first two sentences of \S \ref{subsec:exactnumber}. It would be better if the meanings of these symbols were fixed.)}
\end{comment}
\vspace{3mm}
\begin{algorithm}[!ht]
\scriptsize
\DontPrintSemicolon
\LinesNumbered
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{
$g$, a square-free polynomial in $\Q[x,y]$ such that $\deg(g,y)\geq 1$;\\
$I=[(a_0,b_0),\ldots,(a_{s+1},b_{s+1})]$, a general potential periodic interval set of $g$.
}
\Output{
$k'$, a natural number;\\
$ret$, a finite set of pairs $((x_j^{-},x_j^{+}),c_j)$ such that for any $k>k'~(k\in\N)$, there are \emph{AT LEAST} $c_i$ roots for $g(x,\tan x)$ in the interval $(x_j^{-}+k\pi,x_j^{+}+k\pi)$.
}
\caption{\bf IsolatingFarZero}\label{alg:AlgFarZero}
\BlankLine
Find $k'\in\N$ such that
$(k'+\frac{1}{2})\pi$ is an upper bound for all real roots of $\lc(g,y)\cdot\discrim(g,y)\cdot\prod_{j=0}^{s}g(x,b_{j})g(x,a_{j+1})$. \label{Line:alg1-computation-of-M}\;
Choose an integer $x_0$ such that $x_0>(k'+\frac{1}{2})\pi$\label{Line:alg1-sample-point}.\;
$c_j\leftarrow0$ for $j=0,\ldots,s+1$\label{Line:alg1-ini-cj}\;
\For{each real root $r$ of $g(x_0,y)$}{\label{Line:alg1-for-1}
\For{$j$ from $0$ to $s+1$}
{
\If{$r\in (a_j,b_j)$}{
$c_{j}\leftarrow c_{j}+1$\label{Line:alg1-count-c}\;
\textbf{break}\label{Line:alg1-for-1-end}\;
}
}
}
$ret\leftarrow \emptyset$\label{Line:alg1-organizing-output}\;
\For{$j$ from $0$ to $s+1$\label{Line:alg1-detect-implicit-function-1}}{
\If{$c_j>0$\label{Line:alg1-detect-implicit-function}}{
$x_{j}^{-}\leftarrow\arctan a_j,~x_{j}^{+}\leftarrow\arctan b_j$\;
$ret \leftarrow ret \cup \{((x_{j}^{-},x_{j}^{+}),c_j)\}$\label{Line:alg1-output-end}\;
}
}
\Return $k'$, $ret$\;
\end{algorithm}
\begin{definition}
Let $g$, $k'$ be the same as in Theorem \ref{thm:location-of-the-roots}, and $\theta(x)$ be a terminal-root-function of $g(x,y)$.
We call the set
$\{x_0\in ((k'+1/2)\pi,+\infty) \mid \theta(x_0)=\tan{x_0} \}$
\emph{a bunch of periodic roots} of $g(x,\tan{x})$.
\end{definition}
Summarizing all above, we propose Algorithm \ref{alg:AlgFarZero} to isolate the ``periodic" roots of $g(x,\tan{x})$.
Given a square-free polynomial $g\in\Q[x,y]$ with $\deg(g,y)\geq1$, and a general potential periodic interval set $I$ of $g$, Algorithm \ref{alg:AlgFarZero} computes a natural number $k'$ and a set of pairs $((x_{i}^{-},x_{i}^{+}),c_i)$, such that for any $k>k'~(k\in\N)$, there are at least $c_i$ real roots of $g(x,\tan x)$ in the interval $(x_{i}^{-}+k\pi,x_{i}^{+}+k\pi)$.
\begin{theorem}\label{thm:AlgFarZeroisCorrect}
Algorithm \ref{alg:AlgFarZero} is correct.
\end{theorem}
\begin{proof}
The process of Algorithm \ref{alg:AlgFarZero} is based on Theorem \ref{thm:location-of-the-roots}.
And it is clear that Algorithm \ref{alg:AlgFarZero} is correct by the proof of Theorem \ref{thm:location-of-the-roots}.
\end{proof}
\subsection{Ensuring exact quantity of periodic roots in an ``isolating" interval}\label{subsec:exactnumber}
In the subsection, we would strengthen Theorem \ref{thm:location-of-the-roots}: for all sufficiently large $k$, $\theta_i(x)$ and $\tan x$ intersects exactly once in the interval $(\arctan a_{j_i}+k\pi,\arctan b_{j_i}+k\pi)$. And a lower bound for such $k$ can be efficiently determined.
Depending on whether the derivative $\theta_i'(x)\ (x>M')$ is strictly bounded by $1$ or not,
the discussion is divided into two cases.
For the bounded case, the key observation is: the derivative of $\tan x$ is $1/\cos^2 x$, which is always $\geq 1$. Therefore, since $\theta_i'(x)$ is smaller than $1$, $y=\tan x$ can not intersect with $\theta_i(x)$ twice in a period $(k\pi -\pi/2,k\pi+\pi/2)$, where $k\in\N$ and $(k+1/2)\pi>M'$.
For the unbounded case, we consider $1/\theta_i(x)$ instead of $\theta_i(x)$.
We first establish a lemma.
\begin{lemma}\label{lemma:existence-of-delta}
Let $g\in\Q[x,y]$ be square-free with $\deg(g,y)\geq1$.
There exists a finite set $S\subset\Q$ such that for any $\delta_0\in\Q\setminus S$,
$\res(\frac{\partial g}{\partial x}+\delta_0 \frac{\partial g}{\partial y},g,y)$
is not the zero polynomial in $\Q[x]$.
\end{lemma}
\begin{proof}
Regard $\delta$ as a variable. We claim that
$\res(\frac{\partial g}{\partial x}+\delta \frac{\partial g}{\partial y},g,y)$ is not the zero polynomial in $\Q[\delta,x]$.
Note that $\frac{\partial g}{\partial y}$ is not the zero polynomial, since $\deg(g,y)\ge1$.
Then, $q:=\gcd(\frac{\partial g}{\partial x}+\delta \frac{\partial g}{\partial y},~g)$ is a nonzero polynomial in $\Q[x,y]$.
We also note that $q\mid\left(\frac{\partial g}{\partial x}+\delta \frac{\partial g}{\partial y}\right)$ implies $q\mid\frac{\partial g}{\partial y}$. Now that $g$ is square-free, $q\mid g$ and $q\mid\frac{\partial g}{\partial y}$ imply $\deg(q,y)=0$. Thus, $\res(\frac{\partial g}{\partial x}+\delta \frac{\partial g}{\partial y},g,y)$ is not the zero polynomial in $\Q[\delta,x]$.
Hence we can let $S$ be the set of the common roots in $\Q$ of the coefficients of $\res(\frac{\partial g}{\partial x}+\delta \frac{\partial g}{\partial y},g,y)$ with respect to $x$. The proof is then completed.
\end{proof}
\begin{lemma}\label{thm:derivative-bound}
Let $g\in\Q[x,y]$ be a square-free polynomial with $\deg(g,y)\geq1$,
$\delta_0$ a rational number such that $\res(\frac{\partial g}{\partial x}+\delta_0 \frac{\partial g}{\partial y},g,y)$ is not the zero polynomial,
$\theta_1(x),\ldots,\theta_r(x)$ the terminal-root-functions of $g$, and $\hat{k}\in\N$ a natural number such that $\hat{M}=(\hat{k}+1/2)\pi$ is an upper bound of the start-point of the terminal-root-functions of $g$ and all real roots of $\res(\frac{\partial g}{\partial x}+\delta_0 \frac{\partial g}{\partial y},g,y)$.
Then, the derivative of each $\theta_i(x)$ satisfies either one of the following properties:
\begin{enumerate}[label=$($\arabic*$)$]
\item\label{item:thm:derivative-bound1} $\forall x>\hat{M}:$ $\theta_i'(x)<\delta_0$;
\item\label{item:thm:derivative-bound2} $\forall x>\hat{M}:$ $\theta_i'(x)>\delta_0$.
\end{enumerate}
We call $\hat{k}$ a \emph{$\delta_0$-bound-point} of $g$.
In particular, if $\delta_0<1$ and property \ref{item:thm:derivative-bound1} holds for some $\theta_i(x)$, then for any $k>\hat{k}~(k\in\N)$,
there is a unique $\xi\in (k\pi -\pi/2, k\pi+\pi/2)$ such that $\theta_i(\xi)=\tan{\xi}$.
\end{lemma}
\begin{proof}
For each $\theta_i'(x)$, we prove that one of the two cases holds.
Note that $\hat{M}$ is greater than the start-point of the terminal-root-functions of $g$ (recall Definition \ref{def:root-function-start-point}).
So, when $x\in(\hat{M},+\infty)$, $\lc(g,y)(x)\neq0$, $\discrim(g,y)(x)\ne0$ and $g(x,\theta_i(x))=0$.
Therefore, $\frac{\partial g}{\partial y}(x,\theta_i(x))\ne0$ and $\theta_i'(x)=-\frac{\partial g/\partial x}{\partial g/\partial y}(x,\theta_i(x))$ by the implicit function theorem.
By the definition of $\hat{M}$, we have
$\lc(g,y)(x)\neq0$ and $\res(\frac{\partial g}{\partial x}+\delta_0 \frac{\partial g}{\partial y},g,y)(x)\ne0$ for $x\in(\hat{M},+\infty)$.
Then, since $g(x,\theta_i(x))=0$, it is clearly that
$(\frac{\partial g}{\partial x}+\delta_0\frac{\partial g}{\partial y})(x,\theta_i(x))\ne 0$.
So, we have $\theta_i'(x)\ne\delta_0$.
Therefore, by the continuity of $\theta'_{i}(x)$, exactly one case holds.
As for the last assertion, the existence is by the intermediate value theorem applied to $\tan x-\theta_i(x)$ on the interval $(k\pi -\pi/2, k\pi+\pi/2)$. The uniqueness follows from the fact that $\tan x-\theta_i(x)$ is strictly increasing on that interval since $1/\cos^2x-\theta'_i(x) >1-\delta_0>0$.
\end{proof}
Keep the notation in Lemma \ref{thm:derivative-bound}.
By the ``In particular'' part of Lemma \ref{thm:derivative-bound},
when $\theta_i'(x)<\delta_0$ on the interval $(\hat{M},\infty)$ for some selected $\delta_0<1$,
we can guarantee that there exists exactly $1$ periodic root (induced by $\theta_i$) of $g(x,\tan{x})$ in $(k\pi-\pi/2,k\pi+\pi/2)$ with $k$ sufficiently large.
In order to deal with the case with $\theta_i'(x)>\delta_0$, we consider $1/\theta_i(x)$ instead of $\theta_i(x)$.
In fact, $1/\theta_i(x)$ is also a terminal-root-function of the reciprocal polynomial
$h:=\mathcal{R}(g)$ (see Def. \ref{definition:recip}). If $y\nmid g$, we have $\deg(h,y)=\deg(g,y)\geq1$ by the definition. Moreover, since $g$ is square-free and clearly $h\neq cy$ for any $c\in\Q$, $h$ is also square-free (recall Prop. \ref{proposition:recipKeepSquarefree}).
Finally, we obtain the following theorem, which strengthens Theorem \ref{thm:location-of-the-roots}.
\begin{theorem}\label{thm:Uniqueness}
Let $g\in\Q[x,y]$ be square-free such that $\deg(g,y)\geq1$ and $y\nmid g$. Set
$h=\mathcal{R}(g)$ and take
$\delta_1\in\Q ~(0<\delta_1<1)$, $\delta_2\in\Q ~(-1<\delta_2<0)$ such that neither $\res(\frac{\partial g}{\partial x}+\delta_1 \frac{\partial g}{\partial y},g,y)$ nor $\res(\frac{\partial h}{\partial x}+\delta_2 \frac{\partial h}{\partial y},h,y)$ is the zero polynomial.
Let $k'$ be the same as in Theorem \ref{thm:location-of-the-roots},
$k''$ a $\delta_1$-bound-point of $g$ and
$k'''$ a $\delta_2$-bound-point of $h$ guaranteed by Lemma \ref{lemma:existence-of-delta}.
Set $k^{*}:=\max\{k',k'',k'''\}$ and let $\theta_1(x),\ldots,\theta_r(x)$ be the terminal-root-functions of $g$.
Then, for any $k>k^{*}~(k\in\Z)$,
there is a unique $\xi\in (k\pi -\pi/2,k\pi +\pi/2)$ such that $\theta_i(\xi)=\tan \xi$.
\end{theorem}
\begin{proof}
Since $y\nmid g$, $\deg(g,y)\ge 1$ and $g$ is square-free, we see that $\deg(h,y)\geq1$ and $h$ is square-free.
Thus, $k'''$ is well-defined.
Set $\mu:=(k^{*}+1/2)\pi$.
By Lemma \ref{thm:derivative-bound}, suppose $\theta_1(x),\ldots,\theta_t(x)$ are in Case \ref{item:thm:derivative-bound1} (\emph{i.e.}, $\theta_i'(x)<\delta_1<1$ for $x\in (\mu,+\infty))$ and $\theta_{t+1}(x),\ldots,\theta_r(x)$ are in Case \ref{item:thm:derivative-bound2} (\emph{i.e.}, $\theta_i'(x)>\delta_1$ for $x\in (\mu,+\infty))$ with $0\le t\le r$.
Then, by the ``In particular'' part in Lemma \ref{thm:derivative-bound}, for each $\theta_i~(1\le i\le t)$ in Case \ref{item:thm:derivative-bound1}, we already have the conclusion.
In the following, we deal with those $\theta_i(x)$ with $t+1\leq i\leq r$.
First, we prove that for each $\theta_i(x)~(t+1\le i\le r)$, $\theta_i(x)>0$ on $(\mu,+\infty)$: Since $\theta_i'(x)$ is greater than a positive number $\delta_1$ by Lemma \ref{thm:derivative-bound}, $\theta_i(x)\rightarrow+\infty$ as $x\rightarrow+\infty$. By property \ref{item:thm:location-of-the-roots1} of Theorem \ref{thm:location-of-the-roots}, the image of $\theta_i$ on $(\mu,+\infty)$ is a subset of $(a_{s+1},+\infty)$ with $a_{s+1}>0$ therein (recall Definition \ref{def:potential-periodic-interval}).
Second, we prove the existence of $\xi$ for each $\theta_i(x)~(t+1\le i\le r)$.
Let $k$ be any integer greater than $k^{*}$.
Applying the intermediate value theorem to $\tan{x}-\theta_i(x)$ on $(k\pi-\pi/2,k\pi+\pi/2)$, we see that there exists $\xi\in(k\pi-\pi/2,k\pi+\pi/2)$ such that $\tan{\xi}=\theta_i(\xi)$.
Finally, we prove the uniqueness of $\xi$ for each $\theta_i(x)~(t+1\le i\le r)$: Suppose $\tan\xi=\theta_i(\xi)$ for some $\xi\in(k\pi-\pi/2,k\pi+\pi/2)$. Then, $\xi\in(k\pi,k\pi+\pi/2)$ since $\theta_i(x)>0$ for all $x\in(\mu,+\infty)$.
Note that $1/\theta_i(x)$ is well defined on $(\mu,+\infty)$, $h(x,1/\theta_i(x))=0$ whenever $x\in (\mu,+\infty)$, and $\mu$ is greater than the start-point of the terminal-root-functions of $h$.
So, $1/\theta_i(x)$ is one of the terminal-root-functions of $h$ when $x\in (\mu,+\infty)$.
By Lemma \ref{thm:derivative-bound}, we have
$(1/\theta_i(x))'<\delta_2$ or $(1/\theta_i(x))'>\delta_2$ for $x\in(\mu,+\infty)$ with some $-1<\delta_2<0$.
Since $\lim\limits_{x\rightarrow +\infty}\theta_i(x)=+\infty$ and $\lim\limits_{x\rightarrow +\infty}1/\theta_i(x)=0$, we have $(1/\theta_i(x))'>\delta_2$.
Because $-1/\sin^2 x-(1/\theta_i(x))'<-1-\delta_2<0$,
the function $\cot x-1/\theta_i(x)$ is strictly decreasing on the interval $(k\pi,k\pi+\pi/2)$.
Thus, $\xi$ is unique.
\end{proof}
\begin{example}
Consider the polynomial $g(x,y)$ in Example \ref{example:potential-periodic-interval-set}.
In Example \ref{example:thm:location-of-the-roots}, we have taken $k'=3$.
It remains to compute $k''$ and $k'''$ in Theorem \ref{thm:Uniqueness}.
We take
$\delta_1=\frac{999}{1000}$ and $k''=3$, which is a
$\delta_1$-bound-point of $g$.
Note that $h=\mathcal{R}(g)=10xy^8-20y^7-40xy^6+92y^5+60xy^4-28y^3-40xy^2+20y+10x.$
We take
$\delta_2=-\frac{999}{1000}$ and $k'''=0$, which is a $\delta_2$-bound-point of $h$.
Thus, we have $k^{*}=\max\{k',k'',k'''\}=3$.
By Example \ref{example:thm:location-of-the-roots} and Theorem \ref{thm:Uniqueness}, we see that for every $k > 3~(k \in \Z)$, there are exactly two real roots of $g(x,\tan{x})$ in the interval $(k\pi-\arctan{\frac{215}{128}},k\pi-\arctan{\frac{19}{32}})$.
\end{example}
We propose Algorithm \ref{alg:AlgDerBound} to compute
a better lower bound, which guarantees the uniqueness, for the ``periodic'' roots of $g(x,\tan{x})$.
Given a square-free polynomial $g\in\Q[x,y]$ such that $\deg(g,y)\geq1$ and $y\nmid g$, together with a general potential periodic interval set $I$,
Algorithm \ref{alg:AlgDerBound} computes a natural number $k^{*}$ and a set of pairs $((x_{i}^{-},x_{i}^{+}),c_i)$ such that for any $k>k^{*}~(k\in\N)$, there are exactly $c_i$ real roots of $g(x,\tan x)$ in the interval $(x_{i}^{-}+k\pi,x_{i}^{+}+k\pi)$.
The process of Algorithm \ref{alg:AlgDerBound} is as follows.
First, by Algorithm \ref{alg:AlgFarZero}, we compute $k'\in\N$
and a set of isolating intervals $ret$ such that for any $k>k'~(k\in\N)$ and any pair $((x_i^{-},x_{i}^{+}),c_i)$ in $ret$, there are at least $c_i$ real roots of $g(x,\tan x)$ in $(x_i^{-}+k\pi,x_i^{+}+k\pi)$.
Then, we compute $k''\in\N$ and $k'''\in\N$ by Lemma \ref{thm:derivative-bound} in Lines \ref{line:alg2-k''-st}--\ref{line:alg2-k''-ed} and Lines \ref{line:alg2-k'''-st}--\ref{line:alg2-k'''-ed}, respectively.
Finally, the algorithm returns the natural number $k^{*}:=\max\{k',k'',k'''\}$ and the set $ret$.
The correctness of Algorithm \ref{alg:AlgDerBound} is guaranteed by Theorem \ref{thm:Uniqueness} and the termination of it is obvious.
\begin{algorithm}[ht]\label{exactroot}
\scriptsize
\DontPrintSemicolon
\LinesNumbered
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{
$g$, a square-free polynomial in $\Q[x,y]$ such that $\deg(g,y)\geq 1$ and $y\nmid g$;\\
$I=[(a_0,b_0),\ldots,(a_{s+1},b_{s+1})]$, a general potential periodic interval set of $g$.
}
\Output{
$k^{*}$, a natural number;\\
$ret$, a finite set of pairs $((x_i^{-},x_i^{+}),c_i)$ such that
for any $k>k^{*}~(k\in\N)$, there are \emph{EXACTLY} $c_i$ roots for $g(x,\tan x)$ in the interval $(x_i^{-}+k\pi,x_i^{+}+k\pi)$.
}
\caption{\bf CompleteIsolatingFarZero}\label{alg:AlgDerBound}
\BlankLine
$k',\;ret\leftarrow{\bf IsolatingFarZero}(g,I)$\label{line:alg2-k'}\;
$d_1\leftarrow \mathtt{res}(\frac{\partial g}{\partial x}+\delta \frac{\partial g}{\partial y},g,y)$ \#\emph{In this step, $\delta$ is a variable and $d_1\in\Q[\delta,x]$.}\label{line:alg2-k''-st}\;
Choose $\delta_1 \in(0,1)\cap \mathbb{Q}$ such that $d_1(\delta_1,x)$ is not the zero polynomial in $\Q[x]$.\;
Find $k''\in\N$ such that
$(k''+\frac{1}{2})\pi$ is an upper bound for all real roots of the polynomial $\textcolor{white}{hhhhhhhhhhhh}d_1(\delta_1,x)\cdot\lc(g,y)\cdot\allowbreak\discrim(g,y)$.\label{line:alg2-k''-ed}\;
$h\leftarrow g(x,\frac{1}{y})\cdot y^{\deg(g,y)}$\label{line:alg2-k'''-st}\;
$d_2\leftarrow \mathtt{res}(\frac{\partial h}{\partial x}+\delta \frac{\partial h}{\partial y},h,y)$\;
Choose $\delta_2 \in(-1,0)\cap \mathbb{Q}$ such that $d_2(\delta_2,x)$ is not the zero polynomial in $\Q[x]$.\;
Find $k'''\in\N$ such that
$(k'''+\frac{1}{2})\pi$ is an upper bound for all real roots of the polynomial $\textcolor{white}{hhhhhhhhhhhh}d_2(\delta_2,x)\cdot\lc(h,y)\cdot\allowbreak\discrim(h,y)$.\label{line:alg2-k'''-ed}\;
$k^{*}\leftarrow \max\{k',k'',k'''\}$\;
\Return $k^{*}$, $ret$
\end{algorithm}
\input{chap/Boundedsolve}
\section{The Whole Isolation Algorithm: Taking Root Multiplicity into Account}\label{sec:wholealgorithm}
In this section, we shall combine the results in previous sections to give an algorithm to isolate all real roots of an MTP. Furthermore, the multiplicity of each root is also computed by the algorithm. To do this, we shall first introduce a proposition on root multiplicity.
\begin{proposition}\label{prop:multiplicity-one}
Let $g\in \mathbb{Q}[x,y]\backslash \mathbb{Q}$ be irreducible. Then except $x=0$, all real roots of $g(x,\tan x)$ in its domain are simple roots.
\end{proposition}
\begin{proof}
If $g\in\Q[x]\cup\Q[y]$, the conclusion is trivially true. When $g$ is bivariate the conclusion follows from \cite[Theorem 2.1]{chen2020automated}. Although the theorem only claims that any root of $g(x,\tan x)$ in $(0,\frac{\pi}{4}]$ is simple, its proof actually shows that any root in $(\mathbb{R}\backslash\{0\})\backslash(\mathbb{Z}+\frac{1}{2})\pi$ of $g(x,\tan x)$ is simple, which is just what we want.
\end{proof}
The following observation is similar to Lemma \ref{lemma:NoCommonRoot}. The proof given below is according to the proof of \cite[Theorem 7]{mccallum2012deciding}.
\begin{proposition}\label{proposition:nocommonroot}
Suppose that $p,q\in\mathbb{Q}[x,y]$ are co-prime, then there is no common root of $p(x,\tan x)$ and $q(x,\tan x)$ except $x=0$.
\end{proposition}
\begin{proof}
Replacing those ``$p$, $q$" in the proof of \cite[Theorem 7]{mccallum2012deciding} by $p$ and $q$ in this proposition, those ``$f^*(x)$, $g^*(x)$" therein by $p(x,\tan x)$ and $q(x,\tan x)$, and the function ``\text{trans}" therein by the tangent function, we see that if $\alpha\in(\mathbb{R}\backslash\{0\})\backslash(\mathbb{Z}+\frac{1}{2})\pi$ is a common root of $p(x,\tan x)$ and $q(x,\tan x)$, then both $\alpha$ and $\tan \alpha$ would be algebraic, which contradicts Theorem \ref{thm:BothAlgeNum}.
\end{proof}
Then the following theorem is straightforward.
\begin{theorem}\label{thm:counting-multiplicity}
Let $g\in\mathbb{Q}[x,y]\backslash\mathbb{Q}$, and the factorization of $g$ in $\mathbb{Q}[x,y]$ is
$g=c_0\prod_{i=1}^{n} g_i^{p_i}$,
where $g_i$ are distinct irreducible factors and $c_0\in\mathbb{Q}$. If $x_0\in\mathbb{R}\backslash\{0\}$ is a root of $g(x,\tan x)$, then there is a unique $g_i$ such that $g_i(x_0,\tan x_0)=0$ and the multiplicity of $x_0$ is $p_i$.\hfill\qed
\end{theorem}
By the discussions above, Algorithms \ref{alg:AlgDerBound} and \ref{alg:BoundRSI} can be extended to compute the multiplicity of each root: Factorize $g$ into $c_0\prod_{i=1}^{n} g_i^{p_i}$ and call Algorithms \ref{alg:AlgDerBound} and \ref{alg:BoundRSI} for each irreducible factor $g_i$, then the multiplicity is just the exponent of $g_i$ in $g$ except at $x=0$.
\begin{remark}\label{remark:thesameforcot}
Replacing the tangent function in Propositions \ref{prop:multiplicity-one}--\ref{proposition:nocommonroot} and Theorem \ref{thm:counting-multiplicity} by the cotangent function, the corresponding conclusions also hold. The proofs are all similar.
\end{remark}
The next lemma detects the possible roots of $f(x,\sin x,\cos x)$ at $k\pi~(k\in \mathbb{Z})$, of which the function $g(\frac{x}{2},\tan \frac{x}{2})$ does not take care.
\begin{lemma}\label{lemma:multiplicityofkpi}
It can be effectively decided whether $f(x,\sin x,\cos x)$ is identically zero for $f\in\mathbb{Q}[x,y,z]$. Moreover, the root multiplicity of $f(x,\sin x,\cos x)$ at $0$, $2k\pi$ $(k\in \mathbb{Z}\backslash\{0\})$, $(2k+1)\pi$ $(k\in \mathbb{Z})$ can be effectively computed.
\end{lemma}
\begin{proof}
Rewrite $f(x,\sin x,\cos x)$ as $g(\frac{x}{2},\tan\frac{x}{2})$ with $g\in\Q[x,t]$ in Eq. (\ref{equation:gTurnedToTan}), then $f(x,\sin x,\cos x)$ is identically zero iff $g(\frac{x}{2},\tan \frac{x}{2})$ is. Lemma 2.3 in \cite{chen2020automated} then shows $f(x,\sin x,\cos x)$ is identically zero iff $g$ is a zero polynomial.
If $f(x,\sin x,\cos x)$ is not identically zero, then the root multiplicity of $x=0$ is finite. Otherwise the analytic function $f(z,\sin z,\cos z)$ with $z$ varying in $\C$ would be identically zero, and so would be $f(x,\sin x,\cos x)$. By computing $\frac{\mathrm{d}^{n} }{\mathrm{d} x^{n}}f(x,\sin x,\cos x)|_{x=0}$ for $n=0,1,2,...,$ until $\frac{\mathrm{d}^{n_0} }{\mathrm{d} x^{n_0}}f(x,\sin x,\cos x)|_{x=0}\neq0$ for some $n_0$. Then $n_0$ is the multiplicity of $x=0$.
For $x=2k\pi$ with $k\in\Z\backslash\{0\}$, $f(x,\sin x,\cos x)|_{x=2k\pi}=0$ iff $g(x,\tan x)|_{x=k\pi}$\\$=0$ and the root multiplicity of $2k\pi$ \emph{w.r.t.} $f(x,\sin x,\cos x)$ equals the one of $k\pi$ \emph{w.r.t.} $g(x,\tan x)$. Moreover, we claim $g(k\pi,\tan k\pi)=0$ iff $t$ is a factor of $g$. The ``if" part is clear. Suppose $g(k\pi,\tan k\pi)=0$, then $0=h(k\pi,\tan k\pi)=h(k\pi,0)$ for some irreducible factor $h=\sum_{j=0}^{\deg(h,x)}a_j(t)x^j$ of $g$. Since $k\pi$ is transcendental, $a_j(0)=0$ for any $j$. Hence $t$ divides $h$ and $h=rt$ (with an $r\in\Q$), which proves the claim. Then, by Theorem \ref{thm:counting-multiplicity}, the multiplicity of $2k\pi$ is the exponent of $y$ in the factorization of $g$.
Similarly, we can rewrite $f(x,\sin x,\cos x)$ as $\hat{g}(\frac{x}{2},\cot \frac{x}{2})$, with
\[\hat{g}(x,t)=f(2x,\frac{2t}{1+t^2},\frac{t^2-1}{1+t^2})(1+t^2)^{\text{deg}(f,[y,z])}\in\Q[x,t].\]
Then, using Remark \ref{remark:thesameforcot} and the discussion similar to the last paragraph, one observes that the root multiplicity of $(2k+1)\pi$ equals the exponent of $t$ in the factorization of $\hat{g}$.\end{proof}
It is time to give the main algorithm (Algorithm \ref{alg:CompleteAlgorithm}) for isolating all real roots of $f(x,\sin x,\cos x)$:
\begin{enumerate}
\item Decide whether the MTP $f(x, \sin x,\cos x)$ is identically zero (Lines \ref{algline:whetherg0}--\ref{algline:error}).
\item Compute the root multiplicity of the points $x=k\pi$ with $k\in\Z$ using the method in the proof of Lemma \ref{lemma:multiplicityofkpi} and compute a polynomial $g\in\Q[x,y]$ according to the input $f$ such that $y\nmid$ $g$ (Lines \ref{algline:m1}--\ref{algline:m3}).
\item Compute the isolating intervals of $\lc(g,x)$ together with $(a_0,b_0)$, $(a_{s+1},b_{s+1})$, forming a potential periodic interval set $I$ and factorize $g$ (Lines \ref{algline:isolate}--\ref{algline:factorize}).
\item Call Algorithm \ref{alg:AlgDerBound} with each irreducible factor $g_i$ with $\deg(g_i,y)\geq1$ together with $I$ to compute some $k^*_1$, $k^*_2$ and isolate the real roots of $g_i(x,\tan x)$ outside $(-k^*_2\pi-\pi/2,k^*_1\pi+\pi/2)$ (Lines \ref{algline:deggy>0part}--\ref{algline:renewk-}).
\item Isolate nonzero roots of the factors $g_i$ in $\Q[x]$ (Lines \ref{algline:isolateAlge}--\ref{algline:remove00}). Note that these correspond to the only nonzero algebraic roots of $f(x,\sin x,\cos x)$. We call $[0,0]$ and the intervals obtained in the above steps the ``algebraic" ones and others the ``transcendental" ones: the former represent 0 or algebraic roots and the latter represent transcendental roots. An algebraic interval may intersect a transcendental one, but it won't cause any trouble since we know they represent different roots.
\item Unify those $k^*_{1\cdot 2}$ of each factor $g_i$ (with $\deg(g_i,y)\geq1$) by some $k_+$ and $k_-$, which characterize the bound of the ``bounded" roots (Lines \ref{algline:renewk+}, \ref{algline:renewk-} and \ref{algline:unifyk+}--\ref{algline:unifyk-}).
\item For each factor $g_i$ such that $\deg(g_i,y)\geq1$, isolate the real roots of $g_i(x,\tan x)$ in $(k_-\pi-\pi/2,k_+\pi+\pi/2)$ by Algorithm \ref{alg:BoundRSI} (Lines \ref{algline:k-tok+}--\ref{algline:refineboundedroots})
\item Return all the results.
\end{enumerate}
In addition, for any real numbers $\alpha,\beta$, we define the interval $[[\alpha,\beta]]$ as
\begin{equation*}
[[\alpha,\beta]]:=\begin{cases}
(\alpha,\beta), &\text{if}~\alpha<\beta,\\
[\alpha,\beta], &\text{if}~\alpha=\beta.
\end{cases}
\end{equation*}
The notation is used in the pseudocode of Algorithm \ref{alg:CompleteAlgorithm}.
\input{alg/alg7.tex}
\begin{comment}
\section{Effectively bounding the roots at ``infinity''}
In this paper, the ring of \emph{trigonometric-polynomials} refers to the following ring of functions on $\mathbb{R}$:
\[\qscx=\big\{\sum\limits_{\lambda\in\Lambda}c_\lambda x^{\lambda(3)}(\sin x)^{\lambda(1)}(\cos x)^{\lambda(2)}\,\big|\,c_\lambda\in\mathbb{Q}, \Lambda\subset\mathbb{N}^3\text{ is finite}\big\},\]
and each function in that ring is called \emph{a trigonometric-polynomial}. Usually, for a non-zero function $f\in\qscx$, we write it in the form
\[f=\sum\limits_{i=0}^na_ix^i\]
with each $a_i$ in the following ring of functions on $\mathbb{R}$:
\[\qsc=\big\{\sum\limits_{\lambda\in\Lambda}c_\lambda(\sin x)^{\lambda(1)}(\cos x)^{\lambda(2)}\,\big|\,c_\lambda\in\mathbb{Q}, \Lambda\subset\mathbb{N}^2\text{ is finite}\big\}\]
and $a_n$ not a zero-function on $\mathbb{R}$ (we call $a_n$ \emph{the leading coefficient} of $f$). Apparently, there are positive numbers $B_i$, $i=0,1,\ldots,n$, such that \begin{equation}\label{coebound}
|a_i|\leq B_i.
\end{equation}
The following lemma is also straightforward:
\begin{lemma}\label{lemma:nolimit}
For any $f\in\qscxe$ that is not identically $0,$ the set of real roots of $f$ $($if there are any$)$ has no limit point in $\mathbb{R}$.
\end{lemma}
\begin{proof}
Suppose that $g(y_1,y_2,x)\in\mathbb{Q}[y_1,y_2,x]$ is a polynomial such that $f(x)=g(\sin x,\cos x,x)$ for all $x\in\mathbb{R}$. Then the complex-valued function $h(z)=g(\sin z,\cos z,z)$ over $\mathbb{C}$ is entire. By definition, $h|_{\mathbb{R}}=f$. If the set of the real roots of $f$ has a limit point in $\mathbb{R}$, then the set of the complex roots of $h$ has a limit point in $\mathbb{C}$ as well, which means $h\equiv0$ over $\mathbb{C}$. But $h$ cannot be identically $0$ over $\mathbb{C}$ since $f$ is not over $\mathbb{R}$.
\end{proof}
\begin{corollary}\label{corollary:countableroots}
Any trigonometric-polynomial $f\in\qscxe$ that is not identically zero has at most countably many real roots $($counted without multiplicity$)$.
\end{corollary}
\begin{proof}
By Lemma \ref{lemma:nolimit}, for each real root $\xi$, there is a positive number $\epsilon_{\xi}$ so that the interval $(\xi-\epsilon_{\xi},\xi+\epsilon_{\xi})$ contains no other real roots of $f$ than $\xi$. Thus, for any two different real roots $\xi$ and $\xi'$ of $f$, we have $|\xi-\xi'|>\max\{\epsilon_{\xi},\epsilon_{\xi'}\}\geq\frac{\epsilon_{\xi}+\epsilon_{\xi'}}{2}$. This implies that $(\xi-\frac{\epsilon_{\xi}}{2},\xi+\frac{\epsilon_{\xi}}{2})\cap(\xi'-\frac{\epsilon_{\xi'}}{2},\xi'+\frac{\epsilon_{\xi'}}{2})=\emptyset$. Hence there is an injection from the set $\mathcal{O}=\{(\xi-\frac{\epsilon_{\xi}}{2},\xi+\frac{\epsilon_{\xi}}{2})\,|\,\xi\text{ is a real root of }f\}$ to $\mathbb{Q}$: for each interval in $\mathcal{O}$, one chooses a rational number in it as its image. So, $\mathcal{O}$ is either finite or countable, so is the set of the real roots of $f$.
\end{proof}
\begin{proposition}
For an $f\in\qscxe$ that is not identically zero$,$ any real root of it has finite multiplicity.
\end{proposition}
\begin{proof}
Set $g\in\mathbb{Q}[y_1,y_2,x]$ so that $g(\sin x,\cos x,x)=f(x)$ for all $x\in\mathbb{R}$ and $h(z)=g(\sin z,\cos z,z)$ to be an entire function over $\mathbb{C}$. Then $h|_\mathbb{R}=f$ and it is straightforward to prove by induction that $\frac{\mathrm{d}^kh}{\mathrm{d}z^k}|_\mathbb{R}=\frac{\mathrm{d}^kf}{\mathrm{d}x^k}$ for any nature number $k\in\mathbb{N}$. Assume that there is a real root $\xi\in\mathbb{R}$ of $f$ so that $\frac{\mathrm{d}^kf}{\mathrm{d}x^k}|_{x=\xi}=0$ for any $k\in\mathbb{N}$, then $\frac{\mathrm{d}^kh}{\mathrm{d}z^k}|_{z=\xi}=0$ for any $k\in\mathbb{N}$ as well. This means that, as an analytic function, $h\equiv0$ in an open disc centring at $\xi$ in the complex plane. Thus $h\equiv0$ in the whole complex plane and $f=h|_{\mathbb{R}}\equiv0$, which contradicts the assumption that $f$ is not identically zero.
\end{proof}
Before we talking about the roots at ``infinity'', we propose first a simple sufficient condition for a trigonometric-polynomial to have only finitely many real roots:
\begin{proposition}\label{proposition:finiteroot}
Let $f\in\qscxe$ be a trigonometric-polynomial with leading coefficient $a_n\in\qsce.$ If $a_n$ has no real root in $(-\pi,\pi]$, then $f$ has finitely many real roots in $\mathbb{R}$.
\end{proposition}
\begin{proof}
Since $a_n$ is a continues function, we see that it is strictly positive (or strictly negative) over the interval $(-\pi,\pi]$. Note that $a_n(-\pi)=a_n(\pi)$, the observation above is also true for the interval $[-\pi,\pi]$. The function $a_n$ is also continues restricted to the interval $[-\pi,\pi]$, which is a compact set, so there is a positive real number $B$ such that $a_n(x)\geq B$ (or, \emph{resp.}, $a_n(x)\leq -B$) for all $x\in[-\pi,\pi]$. Let $B_i$ be as in (\ref{coebound}) and $X_B$ a positive number depends on $B$ such that for any $x\in\mathbb{R}$ with $|x|>X_B$, the inequality
\begin{equation}\label{inq:Bxn}
B|x^n|>\sum\limits_{i=0}^{n-1}B_i|x^i|
\end{equation}
holds. For such an $x$, we have
\begin{equation}\label{inq:noroot}
\begin{array}{rcl}
|a_nx^n| &\geq& B|x^n| \\
&>&\sum\limits_{i=0}^{n-1}B_i|x^i|\\
&\geq&\sum\limits_{i=0}^{n-1}|a_i||x^i|\\
&\geq&|\sum\limits_{i=0}^{n-1}a_ix^i|.
\end{array}
\end{equation}
Thus for any $x\in\mathbb{R}$ with $|x|>X_B$, $f(x)\neq0$. So, all the real roots of $f$ (if there are any) belong to the interval $[-X_B,X_B]$. Now infinitely many roots of $f$ in $[-X_B,X_B]$ would result in a limit point of them, and hence, by Lemma \ref{lemma:nolimit}, one obtains $f\equiv0$ over $\mathbb{R}$. This contradicts with the fact that $f(x)\neq0$ for any $x\in\mathbb{R}$ with $|x|>X_B$.
\end{proof}
The following result deals with a more general case than the above proposition does:
\begin{proposition}\label{proposition:norootarea}
Suppose that $U\subset(-\pi,\pi]$ is a set such that $|a_n(x)|\geq B$ holds for some positive number $B$ for every $x\in(-\pi,\pi]\backslash U,$ and set
\[J^k_U=\{2k\pi+w\,|\,w\in(-\pi,\pi]\backslash U\}\]
for any integer $k,$ then there is a positive number $X_B$ $($which depends on $B)$ so that $f(x)\neq0$ for any $x\in\bigcup\limits_{k\in\mathbb{Z}}J_U^k$ with $|x|>X_B$.
\end{proposition}
One observes that Prop.\,\ref{proposition:finiteroot} takes care of the special case of the above proposition with $U=\emptyset$. The proof of Prop.\,\ref{proposition:norootarea} is almost the same to the one of Prop.\,\ref{proposition:finiteroot}: we choose the same $X_B$ so that (\ref{inq:Bxn}) holds and note that (\ref{inq:noroot}) holds now only for those $x\in\bigcup\limits_{k\in\mathbb{Z}}J_U^k$ satisfying $|x|>X_B$. We see that: for any sufficiently ``far'' interval $2k\pi+(-\pi,\pi]=(2k\pi-\pi,2k\pi+\pi]$, the roots of $f$ in it (if there are any) are necessarily contained in the subset $2k\pi+U=\{2k\pi+w\,|\,w\in U\}$. The following example illustrates the most frequently used form of the set $U$ throughout the paper when $a_n$ has roots, while the subsequent propositions tell us how to choose a ``good'' $U$ so that the roots of $f$ in every $2k\pi+U$ have ``good'' properties.
For convenience, the ``interval'' $[a,b]$ with $a>b$ is regarded as the empty set but as a usual closed interval when $a\leq b$.
\begin{example}\label{example:Uform}
As the leading coefficient of $f\in\qscx,$ $a_n$ is also a trigonometric-polynomial and $a_n\not\equiv0.$ Assume that $a_n$ has real roots$,$ then$,$ by the proof of Prop$.$\,\ref{proposition:finiteroot}$,$ there are only finitely many roots of $a_n$ in the compact set $[-\pi,\pi].$ Let $x_{0,1}<\cdots<x_{0,\ell}$ be all the roots of $a_n$ in $(-\pi,\pi]$ and $\{\epsilon_i\}_{i=1}^\ell$ be some positive numbers such that $-\pi<x_{0,1}-\epsilon_1$ and $(x_{0,i}-\epsilon_i,x_{0,i}+\epsilon_i)\cap(x_{0,j}-\epsilon_j,x_{0,j}+\epsilon_j)=\emptyset$ for every $1\leq i\neq j\leq\ell.$ Then$,$ we can choose
\begin{equation}\label{equation:U}
U=\big((-\pi,-\pi+\epsilon_0)\cup\bigcup\limits_{i=1}^{\ell}(x_{0,i}-\epsilon_i,x_{0,i}+\epsilon_i)\big)\cap(-\pi,\pi]
\end{equation}
with $0<\epsilon_0<x_{0,1}-\epsilon_1+\pi$ a positive number if $a_n(\pi)=0$ and $\epsilon_0=0$ otherwise$.$ In both cases$,$ we observe that $(-\pi,\pi]\backslash U$ is a subset of
\begin{equation}\label{PminusU}
[-\pi+\epsilon_0,x_{0,1}-\epsilon_1]\cup\bigcup\limits_{i=1}^{\ell-1}[x_{0,i}+\epsilon_i,x_{0,i+1}-\epsilon_{i+1}]\cup[x_{0,\ell}+\epsilon_{\ell},\pi],
\end{equation}
and that the closed intervals in the above formula pairwise disjoint$.$ For each non-empty closed interval $I$ occurring in \emph{(\ref{PminusU}),} the function $a_n$ is continuous and never vanishes in it$.$ Thus $|a_n|\geq B_I$ in $I$ for some positive number $B_I,$ and we can choose the number $B$ in Prop$.$\,{\ref{proposition:norootarea}} to be the minimum of these $B_I.$
\end{example}
In the following, we discuss further how to choose appropriate values of those $\epsilon_i$ in (\ref{equation:U}) so that the roots of $f$ in the sets $2k\pi+U$, which are sufficient ``far'', have ``good'' properties.
\begin{proposition}\label{proposition:sigleroot}
Let $f$ and $a_n$ be as in Prop. \ref{proposition:finiteroot}. If $a_n$ has a single root $x_0$ in $(-\pi,\pi],$ then there is an $\epsilon>0$ and a positive number $X_\epsilon$ $($which depends on $\epsilon)$ such that for any intervals $I^k_{x_0,\epsilon}=(x_0+2k\pi-\epsilon,x_0+2k\pi+\epsilon)$ $($with $k$ an integer$)$ contained in $[X_\epsilon,\infty)$ or $(-\infty,-X_\epsilon],$ $f$ has only a single root in it.
\end{proposition}
\begin{proof}
By assumption, we have $a_n(x_0)=0$ but $a_n'(x_0)\neq0$. Hence, there is a positive number $\epsilon>0$ such that $a_n'$ is strictly positive (or strictly negative) over the interval $[x_0-\epsilon,x_0+\epsilon]$. This implies that \[
a_n(x_0-\epsilon)a_n(x_0+\epsilon)<0.
\] Noting that there is a positive number $A$ such that $|a_n'|\geq A$ over the interval $[x_0-\epsilon,x_0+\epsilon]$, we define
\[
\eta=\min\{A,|a_n(x_0-\epsilon)|,|a_n(x_0+\epsilon)|\}.
\]
Let $B_i$, $i=0,1,\ldots,n-1$ be as in (\ref{coebound}) and set $f'=a_n'x^n+\sum\limits_{i=0}^{n-1}b_ix^i$ with $b_i\in\qsc$. Then there are positive numbers $C_i$, $i=0,\ldots,n-1$, such that $|b_i|\leq C_i$. Let $X_\epsilon$ be a positive number so that the following inequalities hold for any real number $x$ with $|x|\geq X_\epsilon$:
\[
\left\{
\begin{array}{rcl}
\eta|x^n|&>&\sum\limits_{i=0}^{n-1}B_i|x|^i,\\
\eta|x^n|&>&\sum\limits_{i=0}^{n-1}C_i|x|^i.
\end{array}
\right.
\]
Now, fix an integer $k$ such that $I_{x_0,\epsilon}^k\subset[X_\epsilon,\infty)$ or $I_{x_0,\epsilon}^k\subset(\infty,-X_\epsilon]$, then any $x\in I_{x_0,\epsilon}^k$ satisfies $|x|\geq X_\epsilon$. Thus,
\[
\begin{array}{rcl}
|a_n'x^n|&\geq&\eta|x^n| \\
&>& \sum\limits_{i=0}^{n-1}C_i|x|^i\\
&\geq&|\sum\limits_{i=0}^{n-1}b_ix^i|.
\end{array}
\]
This indicates that $f'$ does not vanish in $I_{x_0,\epsilon}^k$, \emph{i.e.}, $f'$ is strictly positive or strictly negative over that interval. On the other hand, if we set $x_*=x_0+2k\pi\pm\epsilon$, then $|x_*|\geq X_\epsilon$ and
\[
\begin{array}{rcl}
|a_n(x_*)x_*^n|&\geq&\eta|x_*^n| \\
&>& \sum\limits_{i=0}^{n-1}B_i|x_*|^i\\
&\geq&|\sum\limits_{i=0}^{n-1}a_ix_*^i|.
\end{array}
\]
This implies that $\text{sgn}(f(x_*))=\text{sgn}(a_n(x_*)x_*^n)$, thus
\[
\begin{array}{rcl}
&&\text{sgn}(f(x_0+2k\pi-\epsilon)f(x_0+2k\pi+\epsilon)) \\
&=& \text{sgn}(a_n(x_0+2k\pi-\epsilon)(x_0+2k\pi-\epsilon)^na_n(x_0+2k\pi+\epsilon)(x_0+2k\pi+\epsilon)^n)\\
&=&\text{sgn}(a_n(x_0-\epsilon)a_n(x_0+\epsilon))\\
&=&-1,
\end{array}
\]
which means
\begin{equation}\label{equation:differentsign}
f(x_0+2k\pi-\epsilon)f(x_0+2k\pi+\epsilon)<0.
\end{equation}Hence there is one and only one single root of $f$ in $I_{x_0,\epsilon}^k$.
\end{proof}
Prop. \ref{proposition:sigleroot} indicates that if $a_n$ has a single root then $f$ has infinitely many real roots. In fact, that would be true if $a_n$ has a root of an odd multiplicity:
\begin{proposition}
Let $f$ and $a_n$ be as usual. If $a_n$ has a root $x_0$ in $(-\pi,\pi]$ of an odd multiplicity$,$ then there is a positive number $\epsilon$ and an $X_\epsilon>0$ so that for any integer $k$ satisfying $I^k_{x_0,\epsilon}\subset(-\infty,-X_\epsilon]\cup[X_\epsilon,\infty),$ there is a root in the interval $I_{x_0,\epsilon}^k$.
\end{proposition}
\begin{proof}
Since $x_0$ is a root of an odd multiplicity, there is a positive number $\epsilon$ so that $a_n$ is strictly increasing or strictly decreasing in the interval $[x_0-\epsilon,x_0+\epsilon]$. This ensures that $a_n(x_0-\epsilon)a_n(x_0+\epsilon)<0$. Set \[\eta=\min\{|a_n(x_0-\epsilon)|,|a_n(x_0+\epsilon)|\}\] and chose a positive number $X_\epsilon$ so that $\eta|x^n|>\sum\limits_{i=0}^{n-1}B_i|x|^i$ holds for any real number $x$ with $|x|\geq X_\epsilon$. For any $k$ such that $I^k_{x_0,\epsilon}\subset(-\infty,-X_\epsilon]\cup[X_\epsilon,\infty)$, we can prove that $\text{sgn}(f(x_*))=\text{sgn}(a_n(x_*)x_*^n)$ with $x_*=x_0+2k\pi\pm\epsilon$ the endpoints of $I_{x_0,\epsilon}^k$, as we have done in the proof of Prop. \ref{proposition:sigleroot}. Thus, the inequality (\ref{equation:differentsign}) holds and there is a root in $I_{x_0,\epsilon}^k$.
\end{proof}
Extending the proof of Prop. \ref{proposition:sigleroot}, we have the following results:
\begin{proposition}\label{proposition:atleastkroots}
If $a_n$ has a root $x_0\in(-\pi,\pi]$ of multiplicity $m,$ then there is a positive number $\epsilon$ and an $X_\epsilon>0$ so that for any integer $k$ satisfying $I^k_{x_0,\epsilon}\subset(-\infty,-X_\epsilon]\cup[X_\epsilon,\infty),$ there are at most $m$ roots $($counted with multiplicity$)$ in the interval $I_{x_0,\epsilon}^k$.
\end{proposition}
\begin{proof}
The proof is inductive on the multiplicity $m$: The case where $m=1$ is clear by Prop. \ref{proposition:sigleroot}. In the following, we prove that the conclusion holds for the case where $m=\ell\geq2$, assuming that it does for the case where $m=\ell-1$.
Write $f'=a_n'x^n+\sum\limits_{i=0}^{n-1}b_ix^i$ with $b_i\in\qsc$. Then $x_0$ is a root of $a_n'$, the leading coefficient of $f'$, of multiplicity $\ell-1$. By assumption, there is a positive number $\epsilon$ and an $X_\epsilon>0$ so that for any integer $k$ satisfying $I^k_{x_0,\epsilon}\subset(-\infty,-X_\epsilon]\cup[X_\epsilon,\infty)$, there are at most $\ell-1$ real roots (counted with multiplicity) in the interval $I_{x_0,\epsilon}^k$. If $f$ has more than $\ell$ roots, say, $r_1<\cdots<r_t$ in the interval $I_{x_0,\epsilon}^k$, each with multiplicity $\ell_1\geq1,\dots,\ell_t\geq1$ and they satisfy $\ell_1+\cdots +\ell_t>\ell$, then every $r_j$ would be a root of $f'$ of multiplicity $\ell_j-1$ and there would be roots $w_1,\ldots,w_{t-1}$ of $f'$ satisfying $r_1<w_1<\cdots<r_{t-1}<w_{t-1}<r_t$. Thus, $f'$ has at least $\sum_{j=1}^t(\ell_j-1)+(t-1)=\sum_{j=1}^t\ell_t-1$ roots in $I_{x_0,\epsilon}^k$. But $\sum_{j=1}^t\ell_t-1>\ell-1$, this means $f'$ has more than $\ell-1$ roots in $I_{x_0,\epsilon}^k$, which contradicts the assumption.
\end{proof}
\end{comment}
\section{Experiments}\label{sec:experiments}
The main algorithm (Algorithm \ref{alg:complete-algorithm}) is implemented with {\tt Maple2021} as a tool named {\tt RootOfMTP}.
In order to indicate the effectiveness and efficiency of the tool, we prepare two types of testing examples, where one type is collected from the literature
and another is generated randomly.
On the one hand, we test the correctness of the tool on these examples through some numerical methods.
On the other hand, the experimental results of these examples show that the tool is powerful and is able to deal with some complicated examples.
\subsection{Commands}
The tool {\tt RootOfMTP} can be downloaded at
\begin{center}
{\bf \url{https://github.com/lihaokun/RootOfMTP}}.
\end{center}
All testing examples can also be downloaded there.
All tests were conducted on 16-Core Intel Core [email protected]
with 128GB of memory and Windows 11.
We illustrate how to use {\tt RootOfMTP} by a simple example.
Suppose we want to isolate all the real roots of the following MTP by {\tt RootOfMTP}:
\begin{align*}
x\sin x + \cos x - 1.
\end{align*}
We only need to run the following commands in {\tt Maple2021}:
{\scriptsize
\begin{verbatim}
read ".../RootOfMTP.mpl";
RootOfMTP(x*sin(x)+cos(x)-1,x,1);
\end{verbatim}}
Herein, the first command is to read the file, and the inputs of the second command are an MTP, the variable of the MTP and a rational number $\epsilon>0$ which specifies the maximal length of isolating intervals.
The output is
{\scriptsize
\begin{verbatim}
For every k <= -1 (k in Z), 2kPi+(0) is a real root with multiplicity 1.
For every k >= 1 (k in Z), 2kPi+(0) is a real root with multiplicity 1.
There is 1 real root with multiplicity 2 at 0.
For every k >= 2 (k in Z), there is 1 real root with multiplicity 1
in (2kPi+(2*arctan(63/16)), 2kPi+(Pi)).
For every k <= -2 (k in Z), there is 1 real root with multiplicity 1
in (2kPi+(-Pi), 2kPi+(-2*arctan(63/16))).
There is 1 real root with multiplicity 1 in each open interval of the list
[[-5/2*Pi-2*arctan(29666650363354128505/36893488147419103232),
-5/2*Pi-2*arctan(7242537696610193/9007199254740992)],
[-1/2*Pi-2*arctan(7030038563941/17592186044416),
-1/2*Pi-2*arctan(14741934773129570377/36893488147419103232)],
[1/2*Pi+2*arctan(14741934773129570377/36893488147419103232),
1/2*Pi+2*arctan(7030038563941/17592186044416)],
[5/2*Pi+2*arctan(7242537696610193/9007199254740992),
5/2*Pi+2*arctan(29666650363354128505/36893488147419103232)]].
\end{verbatim}}
\subsection{Examples from the literature}
We collect $21$ testing examples from the literature \cite{chenliu2016,chen2022automated}.
Since some examples are rational-function inequalities, we need to transform them to MTPs.
Take \cite[Ex.4]{chenliu2016} for example:
\begin{align*}
\frac{\sin^2 x}{x^2}+\frac{\tan x}{x}>\frac{x^2}{\sin^2 x}+\frac{x}{\tan x}.
\end{align*}
First, we move the terms in the right hand side to the left hand side.
Second, we replace $\tan x$ with $\frac{\sin x}{\cos x}$.
Third, we cancel the denominators and obtain an MTP: $-x^4\cos x-x^3\sin x\cos^2 x+x\sin^3 x+\sin^4 x\cos x$.
The experimental results are presented in Table \ref{tab:chen-result}.
Note that these examples are kind of simple.
The bounds $k_{-}$ and $k_{+}$ are close to $0$ and {\tt RootOfMTP} can deal with each of them within $2$ seconds.
\input{table/notation.tex}
\input{table/chen.tex}
\subsection{Random examples}
Now, we present the experimental results for randomly generated
polynomials. A randomly generated polynomial
\begin{align*}
{\bf rp}(d,p,c)\in\Q[x,y,z]
\end{align*}
is constructed as follows:
First, pick some monomials independently from the set
$\{x^{a_1}y^{a_2}z^{a_3}\mid a_i\in\N,~a_1+a_2+a_3\le d\}$ in a manner such that each monomial is picked with probability $p$ while discarded with probability $1-p$.
Then, randomly assign coefficients between $-c$ and $c$ to each picked monomial and sum them to obtain a polynomial.
We generate $160$ random polynomials $F_1,\ldots,F_{160}$ in this way from $16$ different classes, where
\begin{align*}
F_{1},\ldots,F_{10}&\in{\bf rp}(5,0.1,10), & F_{11},\ldots,F_{20}&\in{\bf rp}(10,0.015,10),\\
F_{21},\ldots,F_{30}&\in{\bf rp}(10,0.05,10), & F_{31},\ldots,F_{40}&\in{\bf rp}(10,0.1,10),\\
F_{41},\ldots,F_{50}&\in{\bf rp}(15,0.01,10), & F_{51},\ldots,F_{60}&\in{\bf rp}(15,0.05,10),\\
F_{61},\ldots,F_{70}&\in{\bf rp}(15,0.1,10), & F_{71},\ldots,F_{80}&\in{\bf rp}(20,0.0025,10),\\
F_{81},\ldots,F_{90}&\in{\bf rp}(5,0.1,1000), & F_{91},\ldots,F_{100}&\in{\bf rp}(10,0.015,1000),\\
F_{101},\ldots,F_{110}&\in{\bf rp}(10,0.05,1000), & F_{111},\ldots,F_{120}&\in{\bf rp}(10,0.1,1000),\\
F_{121},\ldots,F_{130}&\in{\bf rp}(15,0.01,1000), & F_{131},\ldots,F_{140}&\in{\bf rp}(15,0.05,1000),\\
F_{141},\ldots,F_{150}&\in{\bf rp}(15,0.1,1000), & F_{151},\ldots,F_{160}&\in{\bf rp}(20,0.0025,1000).
\end{align*}
Then, we replace $y$ and $z$ with $\sin x$ and $\cos x$ respectively in $F_{1},\ldots,F_{160}$ to get MTPs $M_{1},\ldots,M_{160}$.
The experimental results are presented in Table \ref{tab:random-result}.
The tool {\tt RootOfMTP} performs powerfully on these random examples: The tool can solve every example within $3000$ seconds, and $93\%$ of the examples were solved within 1 minute.
We notice that the computation of weak Fourier sequences is not very efficient.
For $59\%$ of the examples, the time for computing the weak Fourier sequences is more than one third of the total computing time.
Another surprising observation is that the bounds $k_{-}$ and $k_{+}$ are still small, even on these more difficult examples.
Finally, we present Figure \ref{fig:num-within-time} to show the performance of {\tt RootOfMTP} on all examples (either from the literature or randomly generated).
The horizontal axis represents time, while the vertical axis represents the number of solved examples within the corresponding time.
\input{table/random.tex}
\input{fig/fig_runtime.tex}
\section{Conclusions}\label{sec:conclusion}
This paper proposes, for the first time, a complete algorithm ``isolating" all the real roots of a given MTP. More precisely, the algorithm computes
\begin{itemize}
\item arbitrarily small intervals whose union covers all real roots of the given MTP (both the roots and the intervals can be countably many), and
\item the exact number of distinct roots in each interval together with the multiplicity of each of them.
\end{itemize}
As a result, this leads to the first complete algorithm (as far as we know) deciding whether there is any, or whether there is infinitely many real roots for a given MTP. The correctness of the algorithm indicates a phenomenon of mathematical interest: the real roots in every period sufficiently far away from zero share a same quantity, a same multiset of root multiplicity and even similar relative positions in each period.
The algorithm has been implemented and a large number of examples have been tested to illustrate its efficiency.
Instead of focusing only on real roots in a certain bounded interval (as some other approaches in the literature doing), our (sub-)algorithms detect and ``isolate" the roots in any bounded intervals of the form $[2k_-\pi-\pi,2k_+\pi+\pi]$, and unbounded intervals of the forms $(2k_+\pi+\pi,\infty)$ and $(-\infty,-2k_-\pi-\pi)$ with sufficiently large $k_\pm$. Based on this, our methods can easily be modified to ``isolate" and compute the multiplicity of all real roots of an MTP in any interval $I$ of the form $(a,b)$, $[a,b)$, $(a,b]$ or $[a,b]$, with $a,b\in\mathbb{Q}\cup\{\pm\infty\}$:
\begin{itemize}
\item Choose an interval $I^\text{e}\supset I$ so that the sub-algorithms ``isolate" the real roots in $I^\text{e}$ naturally.
\item If some of the ``transcendental" intervals cover some endpoints of $I$, replace them by some refined sub-intervals, each of which does not cover any endpoints but contains exactly one transcendental root.
\item After that, every ``isolating" interval covering some endpoints of $I$ is algebraic and contains exactly one root (which is $0$ or algebraic). If there are such intervals, then decide whether those covered endpoints are roots.
\item If the covered endpoints are not roots, refine these endpoint-covering intervals via bisection so that they do not cover the endpoints after refinement. Otherwise, refine them to $[a,a]$ or $[b,b]$ if they cover $a$ or $b$, respectively.
\end{itemize}
In that way we can compute the exact number of real roots in each ``isolating" interval, together with multiplicity of each of them. Based on that, one can further prove MTP inequalities over the interval $I$ if one uses the following simple claim:
\begin{claim}
For any MTP $f$ not identically zero and any interval $I$ $($bounded or not$)$ of positive length,
\begin{itemize}
\item $f>0$ on $I$ iff $f$ has no root in $I$ and $f(x_0)>0$ for some $x_0\in I$;
\item $f\geq0$ on $I$ iff every root of $f$ in the interior of $I$ $($if there are any$)$ has even multiplicity and $f(x_0)>0$ for some $x_0\in I$.
\end{itemize}
\end{claim}
\noindent{}In this way an algorithm proving MTP inequalities over any interval (bounded or not) with endpoints in $\mathbb{Q}\cup\{\pm\infty\}$ can then be derived. To our knowledge, this is the first algorithm proving MTP inequalities over unbounded intervals.
\section*{Acknowledgement}
This work was supported by the NSFC under grants No. 61732001 and No. 12071467, and by the National Key Research Project of China under grant No. 2018YFA0306702
\bibliographystyle{elsarticle-num}
\section{Isolating Bounded Real Roots}\label{sec:IBR}
In this section, we discuss case \ref{item:MainThreeCase1} in Section \ref{sec:Preprocessing}, i.e., given $k_+,k_-\in\Z$, isolating the real roots of
\begin{align}\label{eq:boudedeq}
g(x,\tan x)=0,~~~~x\in(k\pi-\frac{\pi}{2},k\pi+\frac{\pi}{2})
\end{align}
where $g\in\Q[x,y]\setminus\{0\}$, $k\in\Z$ and $k_-\le k\le k_+$.
We solve \eqref{eq:boudedeq} in four intervals $(k\pi-\frac{\pi}{2},k\pi-\frac{\pi}{4}),(k\pi-\frac{\pi}{4},k\pi),(k\pi,k\pi+\frac{\pi}{4})$ and $(k\pi+\frac{\pi}{4},k\pi+\frac{\pi}{2})$.
In each interval, the problem is equivalent to isolating the roots of some $G(r\pi+\arctan{y},y)$ on $(0,1)$ where $G\in\Q[x,y]$ and $r\in\Q$, and thus we can use the frame of \cite[Algorithm 47]{strzebonski2012real} to deal with it.
In Section \ref{subsec:WeakFourierSeq}, we use Algorithm \ref{alg:WeakFourierSeq} to compute a weak Fourier sequence for $G(r\pi+\arctan{y},y)$.
In fact, Algorithm \ref{alg:WeakFourierSeq} is a simplified version of \cite[Algorithm 40]{strzebonski2012real} for our specific needs.
In Section \ref{subsec:IsoRoot}, Algorithm \ref{alg:BoundRSI} is proposed to isolate the real roots of Eq. \eqref{eq:boudedeq}.
\subsection{Weak Fourier sequences}\label{subsec:WeakFourierSeq}
Let us first review the definition of weak Fourier sequences and an important result about it.
\begin{definition}\cite[Def. 10]{strzebonski2012real}\label{def:WeakFourierSeq}
Let $I\subseteq\R$ be an open interval, and $G_1,\ldots,G_k:I\rightarrow\R$ be a sequence of differentiable functions defined on $I$.
The sequence $G_1,\ldots,G_k$ is a \emph{weak Fourier sequence on $I$}, if for all $y\in I$,
\begin{enumerate}
\item for all $i\;(1\le i\le k-1)$, $\sign(G_{i+1}(y))=\sign(G_i'(y))$, and
\item $G_k(y)\ne0$.
\end{enumerate}
We say $G_1,\ldots,G_k$ is a \emph{weak Fourier sequence for $G$ on $I$}, if $G_1,\ldots,G_k$ is a weak Fourier sequence on $I$ and $G_1(y)=G(y)$ for all $y\in I$.
\end{definition}
\begin{notation}
Let $a,b\in\R$ and $G_1,\ldots,G_k$ be a weak Fourier sequence on the open interval $(a,b)$.
Then, for any $i,j\;(1\le i\le j\le k)$,
\begin{enumerate}
\item $\sign(G_{i}(x^{+})):=\sign(G_i(c))$, where $x\in[a,b)$, $c< b$ and $\forall y\in(x,c]\;G_i(y)\ne0$,
\item $\sign(G_{i}(x^{-})):=\sign(G_i(d))$, where $x\in(a,b]$, $a< d$ and $\forall y\in[d,x)\;G_i(y)\ne0$,
\item $\sgc_{i,j}(x)$, for $x\in(a,b)$, is the number of sign changes in the sequence $G_{i}(x),\ldots,G_{j}(x)$ with terms equal to zero removed,
\item $\sgc_{i,j}(x^{+})$, for $x\in[a,b)$, is the number of sign changes in the sequence $G_{i}(x^{+}),\ldots,G_{j}(x^{+})$,
\item $\sgc_{i,j}(x^{-})$, for $x\in(a,b]$, is the number of sign changes in the sequence $G_{i}(x^{-}),\ldots,G_{j}(x^{-})$.
\end{enumerate}
\end{notation}
\begin{theorem}\cite[Thm. 13]{strzebonski2012real}\label{thm:WeakFourierSeq}
Let $I\subseteq\R$ be an open interval and let
$G_1,\ldots,G_{k}:I\rightarrow\R$ be a weak Fourier sequence.
Then for any $x\in I$, $\sgc_{1,k}(x^{+})=\sgc_{1,k}(x)$ and $\sgc_{1,k}(x^{-})=\sgc_{1,k}(x)+r+2s$, where $r,s\in\N$ and $G_1(x)=\cdots=G_r(x)=0,G_{r+1}(x)\ne0$.
Moreover, $s=0$ unless there is a $t>r+1$ such that $G_{t}(x)=0$.\qed
\end{theorem}
Let $a,b$ be two real numbers such that $(a,b)\subseteq I$.
Remark that by the proof of \cite[Prop. 15]{strzebonski2012real}, if $\sgc_{1,k}(a^{+})=\sgc_{1,k}(b^{-})$, then $G_1$ has no real root in $(a,b)$, and if $\sgc_{1,k}(a^{+})=\sgc_{1,k}(b^{-})+1$, then $G_1$ has exactly one simple root in $(a,b)$.
To illustrate Algorithm \ref{alg:WeakFourierSeq}, we introduce the definition of the \emph{arctan-derivative}.
\begin{definition}\label{def:pseudo}
For any $G\in\Q(y)[x]$, let
\[
{\trueder G}:=\frac{\partial G}{\partial x}\cdot\frac{1}{1+y^2}+\frac{\partial G}{\partial y}\in\Q(y)[x].
\]
The \emph{arctan-derivative} of $G$ is defined as $G^{\pder}:={\trueder G}\cdot h(y)$, where $h(y)\in\Q[y]$ is the denominator of $\lc({\trueder G},x)\in\Q(y)$.
In particular, if $\lc({\trueder G},x)$ is a polynomial, then $h(y)=1$.
We call $h(y)$ the \emph{common-term} of $G^{\pder}$.
\end{definition}
\begin{remark}\label{remark:def:pseudo}
For any $G\in\Q(y)[x]$, $G^{\pder}\in\Q(y)[x]$ and $\lc(G^{\pder},x)$ is a polynomial in $\Q[y]$.
\end{remark}
Given $G(x,y)\in\Q[x,y]\setminus\{0\}$ and $r\in\Q$,
the process of Algorithm \ref{alg:WeakFourierSeq} is as follows:
Let $G_1$ be $G(x,y)$.
If $G_1$ is not a nonzero rational number, then compute the arctan-derivative $G_{1}^{\pder}$ of $G_1$, and set $G_2=G_{1}^{\pder}$.
If $G_2$ is not a nonzero rational number, then set $G_3=G_{2}^{\pder}$.
Repeat the process until some $G_k\;(k\ge1)$ is a nonzero rational number.
Then $G_1(r\pi+\arctan y,y),\allowbreak\ldots,G_k(r\pi+\arctan y,y)$ is a weak Fourier sequence for $G(r\pi+\arctan y,y)$ on $\R$.
\begin{algorithm}[!ht]
\scriptsize
\DontPrintSemicolon
\LinesNumbered
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{ $G(x,y)$, a polynomial in $\Q[x,y]\setminus\{0\}$; $r$, a rational number
}
\Output{$G_1(r\pi+\arctan{y},y),\ldots,G_k(r\pi+\arctan{y},y)$, a weak Fourier sequence for $G(r\pi+\arctan{y},y)$ on $\R$, where $G_i\in\Q(y)[x]$
}
\caption{{\bf WeakFourierSeq}}\label{alg:WeakFourierSeq}
\BlankLine
$G_1\leftarrow G(x,y)$\;
$i\leftarrow1$\;
\While{$G_i(x,y)$ is not a nonzero rational number}
{ $G_{i+1}(x,y)\leftarrow G_{i}^{\pder}(x,y)$\;
$i\leftarrow i+1$
}
\Return{$G_1(r\pi+\arctan{y},y),\ldots,G_k(r\pi+\arctan{y},y)$}
\end{algorithm}
Note that we can easily obtain a weak Fourier sequence from a \emph{semi-Fourier sequence} (defined in \cite[Def. 14]{strzebonski2012real}) by \cite[Prop. 15]{strzebonski2012real}.
So, Algorithm \ref{alg:WeakFourierSeq} and \cite[Algorithm 40]{strzebonski2012real} are essentially the same when computing a weak Fourier sequence for $G(r\pi+\arctan{y},y)$, and thus the termination and correctness of Algorithm \ref{alg:WeakFourierSeq} are clear.
However, in order to explain some properties that will be used later, we prepare some lemmas and re-prove the termination and correctness of Algorithm \ref{alg:WeakFourierSeq}.
\begin{lemma}\label{lemma:DerEquiv0}
If $G\in\Q(y)[x]\setminus\Q$ such that $G(x,y)$ is well-defined on the whole plane $\R^2$,
then $G^{\pder}\in\Q(y)[x]\setminus\{0\}$.
\end{lemma}
\begin{proof}
Assume that $G^{\pder}$ is the zero polynomial in $\Q(y)[x]$.
By Definition \ref{def:pseudo}, ${\trueder G}$ is also the zero polynomial.
Then, ${\trueder G}(\arctan{y},y)\equiv0$.
Note that ${\trueder G}(\arctan{y},y)=\frac{\der G(\arctan{y},y)}{\der y}$.
So, $G(\arctan{y},y)\in\Q$ for any $y\in\R$. Since the arctangent function is not algebraic, $G\in\Q$, a contradiction.
\end{proof}
\begin{lemma}\label{lemma:pseudoproperty}
If $G\in\Q(y)[x]\setminus\Q$ and $\lc(G,x)$ is a polynomial, then
\begin{align}\label{eq:lemma:pseudoproperty}
(\deg(G,x),\deg(\lc(G,x)))\succlex(\deg(G^{\pder},x),\deg(\lc(G^{\pder},x))),
\end{align}
where $\succlex$ denotes the lexicographic ordering.
\end{lemma}
\begin{proof}
Recall Remark \ref{remark:def:pseudo}. The degrees $\deg(G^{\pder},x)$ and $\deg(\lc(G^{\pder},x))$ are well-defined.
If $\deg(G,x)=0$, we have $G\in\Q(y)\setminus\Q$, and thus $\lc(G,x)=G$.
So, $G$ is a polynomial in $\Q[y]\setminus\Q$.
It is easy to check that $G^{\pder}=\frac{\partial G}{\partial y}\in\Q[y]$.
So, $\deg(\lc(G,x))=\deg(G)$, $\deg(G^{\pder},x)=0$ and $\deg(\lc(G^{\pder},x))=\deg(G)-1$.
Then inequality \eqref{eq:lemma:pseudoproperty} holds.
If $\deg(G,x)\ge1$, suppose that
\begin{align}\label{eq:G}
G=a_n(y)x^n+\sum_{j=0}^{n-1}a_j(y)x^
\end{align}
where $n=\deg(G,x)\ge1$, $a_n(y)\in\Q[y]$ and $a_j(y)\in\Q(y)$ for $j=0,\ldots,n-1$.
Then,
\begin{align}\label{eq:truederG}
{\trueder G}=a_n'(y)x^n+&\sum_{j=0}^{n-1}\big(a_j'(y)+\frac{(j+1)a_{j+1}(y)}{1+y^2}\big)x^{j}.
\end{align}
If $a_{n}'(y)$ is the zero polynomial, we have $\deg({\trueder G},x)<\deg(G,x)$.
By Definition \ref{def:pseudo}, $\deg(G^{\pder},x)=\deg({\trueder G},x)$.
So, $\deg(G^{\pder},x)<\deg(G,x)$ and inequality \eqref{eq:lemma:pseudoproperty} is true.
If $a_{n}'(y)$ is not the zero polynomial, then $\deg({\trueder G},x)=\deg(G,x)$.
Since $\lc({\trueder G},x)=a_{n}'(y)\in\Q[y]$, $G^{\pder}={\trueder G}$. Thus $\deg(G^{\pder},x)=\deg(G,x)$.
We also note that $\deg(\lc(G,x))=\deg(a_n(y))$ and $\deg(\lc(G^{\pder},x))=\deg(a_n(y))-1$.
The proof is completed.
\end{proof}
\begin{lemma}\label{lemma:DerForm}
If $G\in\Q(y)[x]$ has the following form
\begin{align}\label{eq:specialform}
G=a_n(y)x^{n}+\sum_{j=0}^{n-1}\frac{a_j(y)}{(1+y^2)^{m_j}}x^{j}
\end{align}
where $n=\deg(G,x)$, $a_i(y)\in\Q[y]\;(0\le i\le n)$ and $m_j\in\N~(0\le j\le n-1)$, then the common-term of $G^{\pder}$ is $(1+y^2)^m$ for some $m\in\N$.
Furthermore, $G^{\pder}$ also has the form in Eq. \eqref{eq:specialform}.
\end{lemma}
\begin{proof}
If $\deg(G,x)=0$, then $G=a_n(y)\in\Q[y]$, and thus $G^{\star}=G'(y)$, the common-term of $G^{\star}$ is $1$.
If $\deg(G,x)\ge1$, by Eq. \eqref{eq:G} and Eq. \eqref{eq:truederG} in the proof of Lemma \ref{lemma:pseudoproperty}, there exist $b_i(y)\in\Q[y]\;(0\le i\le n)$ and $d_j\in\N~(0\le j\le n-1)$ such that
\begin{align*}
{\trueder G}=b_n(y)x^n+\sum_{j=0}^{n-1}\frac{b_j(y)}{(1+y^2)^{d_j}}x^{j}
\end{align*}
If $b_n(y)$ is not the zero polynomial, the common-term of $G^{\pder}$ is $1$.
Otherwise, since $\deg(G,x)\ge1$ and $G$ is well-defined on the whole $\R^2$, by the proof of Lemma \ref{lemma:DerEquiv0}, ${\trueder G}$ is not the zero polynomial.
So, there exists an integer $0\le k\le n-1$ such that $b_k(y)$ is not the zero polynomial. Let $k_0$ be the maximal one of such $k$. Then, the common-term of $G^{\pder}$ is $(1+y^2)^{d_{k_0}}$.
Therefore, $G^{\pder}$ has the same form as in Eq. \eqref{eq:specialform}.
\end{proof}
\begin{theorem}\label{thm:alg:WeakFourierSeqTC}
Algorithm \ref{alg:WeakFourierSeq} terminates correctly.
\end{theorem}
\begin{proof}
(Termination)
For any $G(x,y)\in\Q[x,y]\setminus\{0\}$,
let $G_1:=G(x,y),G_2:=G_1^{\pder}(x,y),G_3:=G_2^{\pder}(x,y),\ldots$
We only need to prove that there exists $k\;(k\ge1)$ such that $G_k(x,y)$ is a nonzero rational number.
If $G_1\in\Q\setminus\{0\}$, then we take $k=1$.
If $G_1\in\Q[x,y]\setminus\Q$, then by Lemma \ref{lemma:pseudoproperty}, we have
$$(\deg(G_1,x),\deg(\lc(G_1,x)))\succlex(\deg(G_2,x),\deg(\lc(G_2,x))).$$
Recall Remark \ref{remark:def:pseudo}. For any $i~(i\ge2)$, $G_i\in\Q(y)[x]$ and $\lc(G_i,x)$ is a polynomial.
Therefore, assume that $G_2\notin\Q$, $G_3\notin\Q$,..., by Lemma \ref{lemma:pseudoproperty}, we have
\begin{align*}
(\deg(G_1,x),\deg(\lc(G_1,x)))&\succlex(\deg(G_2,x),\deg(\lc(G_2,x)))\\
&\succlex(\deg(G_3,x),\deg(\lc(G_3,x)))\succlex\cdots.
\end{align*}
So, there exists $m\;(m\ge2)$ such that $G_m\in\Q$.
Let $k\;(k\ge2)$ be the smallest integer such that $G_k\in\Q$.
Then, $G_{k-1}\notin\Q$.
Note that $G_{k-1}$ is well-defined on the whole $\R^2$ by Lemma \ref{lemma:DerForm}.
So, by Lemma \ref{lemma:DerEquiv0}, $G_k$ is a nonzero rational number.
(Correctness)
For $i=1,\ldots,k-1$, let $h_i(y)$ be the common-term of $G_i^{\pder}(x,y)$.
Recall Definition \ref{def:pseudo}.
We have $G_{i+1}(x,y)=h_i(y){\trueder G}_{i}(x,y)$.
Note that
$${\trueder G}_{i}(r\pi+\arctan{y},y)=G_i(r\pi+\arctan{y},y)'$$
and $G_{i+1}(x,y)\in\Q(y)[x]$.
Thus, we only need to prove that $\sign(h_i(y))>0$ on the whole $\R$ and $G_{i+1}(r\pi+\arctan{y},y)$ is defined on the whole $\R$.
If $\deg(G,x)=0$, then $G\in\Q[y]$.
So, $h_i(y)=1$ and $G_{i+1}(r\pi+\arctan{y},y)$ is defined on the whole $\R$.
If $\deg(G,x)\ge1$,
by the termination proof, there exists $m\;(1\le m\le k-1)$ such that $\deg(G_1,x)\ge1,\ldots,\deg(G_m,x)\ge1$ but $\deg(G_{m+1},x)=0$.
Applying Lemma \ref{lemma:DerForm} to $G_i\;(i=1,\dots,m)$, we have
$h_i(y)=(1+y^2)^{n_i}\;(n_i\in\N)$ and $G_{i+1}(r\pi+\arctan y,y)$ is defined on the whole $\R$.
Since $\deg(G_{m+1},x)=0$, $G_{m+1}\in\Q[y]$.
So, for any $m+1\le i\le k-1$, $h_i(x)=1$ and
$G_{i+1}\in\Q[y]$, which is also defined on the whole $\R$.
\end{proof}
\begin{remark}\label{remark:PolyWeakFourierSeq}
Let $G(x,y)\in\Q[x,y]\setminus\{0\}$, $r\in\Q$ and $G_1(r\pi+\arctan{y},y),\ldots,\allowbreak G_k(r\pi+\arctan{y},y)$ be a weak Fourier sequence for $G(r\pi+\arctan{y},y)$ on $\R$ computed by Algorithm \ref{alg:WeakFourierSeq}.
By the correctness proof of Theorem \ref{thm:alg:WeakFourierSeqTC}, $G_i(x,y)\in\Q[y]$ or
\begin{align*}
G_i(x,y)=a_{in_i}(y)x^{n_i}+\sum_{j=0}^{n_{i}-1}\frac{a_{ij}(y)}{(1+y^2)^{m_{ij}}}x^{j}
\end{align*}
where $n_i=\deg(G_i,x)\ge1$, $a_{in_i}(y)\in\Q[y]~(1\le i\le k)$ and $m_{ij}\in\N~(0\le j\le n_{i}-1)$.
For $i=1,\ldots,k$, let $m_i:=\max(m_{ij},0\le j\le n_i-1)$ and
\begin{align*}
F_i:=
\left\{
\begin{array}{cc}
G_i, & G_i\in\Q[y],\\
(1+y^2)^{m_{i}}G_i, & {\rm otherwise}.
\end{array}
\right.
\end{align*}
We have $F_i\in\Q[x,y]$, and $F_i(r\pi+\arctan{y},y)$ has the same signs and real roots as $G_i(r\pi+\arctan{y},y)$ on $\R$.
Thus, we consider $F_i$ instead of $G_i$ in the next subsection.
\end{remark}
\subsection{Isolating the roots}\label{subsec:IsoRoot}
Recall that the goal of Section \ref{sec:IBR} is to isolate the roots of the following equation:
\begin{align}\label{meq:boudedeq1_9}
g(x,\tan{x}),~x\in(k\pi-\frac{\pi}{2},k\pi+\frac{\pi}{2}),~{\rm where}~k\in\Z.
\end{align}
\subsubsection{The algorithm}
Given $g(x,y)\in\Q[x,y]\setminus\{0\}$ and $k\in\Z$,
we propose Algorithm \ref{alg:BoundRSI} to isolate the real roots of Eq. \eqref{meq:boudedeq1_9}.
The process of Algorithm \ref{alg:BoundRSI} is as follows.
We first decide whether $k\pi$, $-\frac{\pi}{4}+k\pi$ or $\frac{\pi}{4}+k\pi$ is a root of $g(x,\tan{x})$ by Lemma \ref{lemma:piRealRoot}.
Then, we compute all irreducible factors $g_1,\ldots,g_m$ of $g(x,y)$ in $\Q[x,y]$.
For every $g_i(x,\tan{x})$, we only need to isolate the real roots of it in four open intervals, i.e., $(k\pi-\frac{\pi}{2},k\pi-\frac{\pi}{4})$, $(k\pi-\frac{\pi}{4},k\pi)$, $(k\pi,k\pi+\frac{\pi}{4})$ and $(k\pi+\frac{\pi}{4},k\pi+\frac{\pi}{2})$.
If $g_i$ is a polynomial in $\Q[x]$, it is easy to deal with (there are many methods to isolate the real roots of univariate polynomials).
Otherwise, we have $\deg(g_i,y)\ge1$.
Let $g_{i}^{(1)}(x,y):=g_{i}(-x,-y)$.
It is easy to check that isolating the real roots of $g_i(x,\tan{x})$ in the first interval (\emph{resp.}, the second interval) is equivalent to isolating those of $g_{i}^{(1)}(x,\tan{x})$ in $(-k\pi+\frac{\pi}{4},-k\pi+\frac{\pi}{2})$ (\emph{resp.}, $(-k\pi,-k\pi+\frac{\pi}{4})$).
Note that these four cases can all be reduced to the problem of isolating the real roots of $G(x,\tan{x})$ on $(r\pi,r\pi+\frac{\pi}{4})$, where $G\in\{g_i,g_{i}^{(1)}\}$ is a rational bivariate irreducible polynomial with $\deg(G,y)\ge1$ and $r\in\{k_0,k_0+\frac{\pi}{4}\}~(k_0\in\Z)$.
And, Algorithm \ref{alg:SubIBR} can solve such problem, which we will explain in detail in the next subsection. Finally, in Algorithm \ref{alg:AlgDerBound} the isolating intervals obtained from each irreducible factor $g_i$ are refined so that they do not intersect. This is by using Remark \ref{remark:bisect} and the fact that the roots in them are all distinct simple roots (Prop. \ref{prop:multiplicity-one}--\ref{proposition:nocommonroot}).
\begin{algorithm}[!ht]
\scriptsize
\DontPrintSemicolon
\LinesNumbered
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{
$g(x,y)$, a polynomial in $\Q[x,y]\setminus\{0\}$\\
$k$, an integer
}
\Output{$ans$, a set of pairwise disjoint isolating intervals of real roots of $g(x,\tan{x})$ in $(k\pi-\frac{\pi}{2},k\pi+\frac{\pi}{2})$
}
\caption{{\bf IsolatingBoundedRoots}}\label{alg:BoundRSI}
\BlankLine
$ans\leftarrow\emptyset$\;
If $g(k\pi,0)=0$, then $ans\leftarrow ans\cup\{[k\pi,k\pi]\}$\;
If $g(-\frac{\pi}{4}+k\pi,-1)=0$, then $ans\leftarrow ans\cup\{[-\frac{\pi}{4}+k\pi,-\frac{\pi}{4}+k\pi]\}$\;
If $g(\frac{\pi}{4}+k\pi,1)=0$, then $ans\leftarrow ans\cup\{[\frac{\pi}{4}+k\pi,\frac{\pi}{4}+k\pi]\}$\;
$g_1,\ldots,g_m\leftarrow$ all irreducible factors of $g(x,y)$ in $\Q[x,y]$\;
\For{$i$ from $1$ to $m$}
{
\eIf{$g_i$ is a polynomial in $\Q[x]$}
{$S\leftarrow$ a set of isolating intervals of real roots of $g_i(x)$ in $(k\pi-\frac{\pi}{2},k\pi+\frac{\pi}{2})\setminus\{k\pi,-\frac{\pi}{4}+k\pi,\frac{\pi}{4}+k\pi\}$\;
$ans\leftarrow ans\cup S$\;
}
{
let $g_{i}^{(1)}\in\Q[x,y]$ such that $g_{i}^{(1)}(x,y)=g_{i}(-x,-y)$\;
$S_1\leftarrow{\bf SubIBR}(g_{i}^{(1)},-k+\frac{1}{4})$\;
$S_2\leftarrow{\bf SubIBR}(g_{i}^{(1)},-k)$\;
$S_3\leftarrow{\bf SubIBR}(g_{i},k)$\;
$S_4\leftarrow{\bf SubIBR}(g_{i},k+\frac{1}{4})$\;
$ans\leftarrow ans\cup\{(-b,-a)\mid (a,b)\in S_1\cup S_2\}\cup S_3\cup S_4$\;
}
}
Refine the intervals obtained in the loop above such that they do not intersect.\;
\Return{$ans$}\;
\end{algorithm}
\begin{comment}
Let $x_1:=x-k\pi$.
It is equivalent to isolating the real roots of $g_i(x_{1}+k\pi,\tan{x_1})$ in $(-\frac{\pi}{2},-\frac{\pi}{4})$, $(-\frac{\pi}{4},0)$, $(0,\frac{\pi}{4})$ and $(\frac{\pi}{4},\frac{\pi}{2})$.
For the first, the second and the fourth case, we can transform $g_i$ to some bivariate rational polynomial $\tilde{g_i}$ and we only need to consider the real roots of $\tilde{g_i}$ in $(0,\frac{\pi}{4})$.
Take the first case for example.
Let $x_2:=-x_{1}$ and $g_{i}^{(1)}:=g_{i}(-x,-y)$.
It is easy to check that solving $g_i(x_1+k\pi,\tan{x_1})$ where $x_1\in(-\frac{\pi}{2},-\frac{\pi}{4})$ is equivalent to solving $g_i^{(1)}(x_2-k\pi,\tan{x_2})$ where $x_2\in(\frac{\pi}{4},\frac{\pi}{2})$.
Let $x_3:=x_2-\frac{\pi}{4}$, then $x_3\in(0,\frac{\pi}{4})$. Note that $\tan(x_3+\frac{\pi}{4})=\frac{1+\tan{x_3}}{1-\tan{x_3}}$ and $1-\tan{x_3}>0$.
Then, let $g_{i}^{(2)}\in\Q[x,y]$ be the product of $g_{i}^{(1)}(x,\frac{1+y}{1-y})$ and the common denominator of $g_{i}^{(1)}(x,\frac{1+y}{1-y})$, and thus it becomes solving $g_{i}^{(2)}(x_3+(-k+\frac{1}{4})\pi,\tan{x_3})$ where $x_3\in(0,\frac{\pi}{4})$.
By Algorithm \ref{alg:SubIBR}, we can isolate the real roots of $g_{i}^{(2)}(x_3+(-k+\frac{1}{4})\pi,\tan{x_3})$ in $(0,\frac{\pi}{4})$.
Remark that $x_3=-x+k\pi-\frac{\pi}{4}$.
If $(a,b)$ is an isolating interval of $g_{i}^{(2)}(x_2+(-k+\frac{1}{4})\pi,\tan{x_2})$ in $(0,\frac{\pi}{4})$, it is obvious that $(-b+k\pi-\frac{\pi}{4},-a+k\pi-\frac{\pi}{4})$ is an isolating interval of $g_i(x,\tan{x})$ in $(k\pi-\frac{\pi}{2},k\pi-\frac{\pi}{4})$.
\end{comment}
\subsubsection{Sub-algorithms}
Let $g(x,y)$ be an irreducible polynomial in $\Q[x,y]$ with $\deg(g,y)\ge1$ and $r\in\{k,k+\frac{1}{4}\}$ where $k\in\Z$.
According to the frame of \cite[Algorithm 47]{strzebonski2012real}, Algorithm \ref{alg:SubIBR} is proposed to isolate the real roots of $g(x,\tan{x})$ in $(r\pi,r\pi+\frac{\pi}{4})$.
It is worth noting that \cite[Algorithm 47]{strzebonski2012real} and Algorithm \ref{alg:SubIBR} are not the same.
Some steps are added, and some steps are skipped or implemented in another way.
The algorithm \cite[Algorithm 47]{strzebonski2012real} can be used to isolate the real roots of an exp-log-arctan function $w(y)$ in an open interval $(a,b)$, where $a,b$ are rational numbers and $(a,b)$ are contained in the domain of $w(y)$.
Let us review the frame of the algorithm.
We need to compute a weak Fourier sequence $w_1,\ldots,w_k$ for $w(y)$ on $(a,b)$ first.
Let $ans$ be a set of computed isolating intervals of $w(y)$ (initialized as $\emptyset$), and $stack$ be a stack of triples $(c,d,m)$ such that $c<d$ are rational numbers and $w_m$ has a constant nonzero sign in $(c,d)$ (initialized as $(a,b,k)$).
In each loop step, we pick a triple $(c,d,m)$ from $stack$ and compute the difference value $\sgc_{1,m}(c^{+})-\sgc_{1,m}(d^{-})$ by \cite[Algorithm 44]{strzebonski2012real},
until $stack$ is empty.
\begin{enumerate}
\item If the value equals $0$, by Theorem \ref{thm:WeakFourierSeq}, $w(y)$ has no roots on $(c,d)$. So, we continue for the next loop step.
\item If the value equals $1$, by Theorem \ref{thm:WeakFourierSeq}, $w(y)$ has exactly one simple root on $(c,d)$. We add $(c,d)$ to $ans$.
\item Otherwise, let $\ell$ be the smallest element of $\{i\mid 2\le i\le m \;{\rm and}\; \sgc_{i,m}(c^{+})\allowbreak-\sgc_{i,m}(d^{-})=1\}$.
By Theorem \ref{thm:WeakFourierSeq}, $w_{\ell}(y)$ has exactly one simple root $sr$ on $(c,d)$.
\begin{enumerate}
\item If $sr$ is a root of $w(y)$, find two rational numbers $u,v$ such that $c<u<sr<v<d$ and $w(y)$ has only one root $sr$ in $[u,v]$
(recall that by Lemma \ref{lemma:NoLimit}, we can find such $u,v$).
Add $(u,v)$ to $ans$, and add $(c,u,\ell)$, $(v,d,\ell)$ to $stack$.
\item If $sr$ is not a root of $w(y)$. Find two rational numbers $u,v$ such that $c<u<sr<v<d$ and $w(y)$ has no roots in $[u,v]$.
Add $(c,u,\ell)$ and $(v,d,\ell)$ to $stack$.
\end{enumerate}
\end{enumerate}
\begin{comment}
Back to Algorithm \ref{alg:SubIBR}, we will explain some differences between the two algorithms and some implementation details in the following.
Instead of using \cite[Algorithm 40]{strzebonski2012real} to compute a semi-Fourier sequence, we directly compute a weak Fourier sequence $w_1=G_1(r\pi+\arctan{y},y),\ldots,w_k=G_k(r\pi+\arctan{y},y)$, where $G_i\in\Q(y)[x]$, for $w(y)$ on $(0,1)$ by Algorithm \ref{alg:WeakFourierSeq}.
Recall Remark \ref{remark:PolyWeakFourierSeq}.
We have $G_i\in\Q[x,y]$ in fact (see Algorithm \ref{alg:SubIBR}-Line \ref{algline:SubIBR-weakFseq}).
{\bf How to compute $\sgc_{i,m}(c^{+})$ or $\sgc_{i,m}(c^{-})$, where $c\in\Q$ and $1\le i\le m$?} We use Algorithm \cite[Algorithm 44]{strzebonski2012real}, i.e. Algorithm \ref{alg:SGC} in the paper, to compute it.
\end{comment}
\begin{comment}
\begin{definition}
A \emph{root constant} is an expression $\Root(G,a,b)$, where $G\in\R[x]$, $a<b$ are rational numbers such that $(a,b)$ is contained in the domain of $G$ and $G$ has the only root in $(a,b)$. $\Root(G,a,b)$ represents the only root of $G$ in $(a,b)$.
\end{definition}
\begin{lemma}
Let $G(x,y)$ be an irreducible polynomial in $\Q[x,y]$ and $r\in\Q$.
Then, $w(y)=G(r\pi+\arctan{y},y)$ has no multiple roots in $(0,1)$.
\end{lemma}
\end{comment}
\begin{comment}
\begin{lemma}
Let $w_1(y)=G_1(r\pi+\arctan{y},y),\ldots,w_k(y)=G_k(r\pi+\arctan{y},y)$ be a weak Fourier sequence for $w_1(y)$, where $G_i\in\Q[x,y]$ (see Remark \ref{remark:PolyWeakFourierSeq}) and $\Root(w_{\ell},c,d)~(l\ge2)$ be a simple root.
Then, there exists $u,v\in\Q$ such that $c<u<\Root(w_{\ell},c,d)<v<d$ and $\sgc_{1,l}$
\end{lemma}
\end{comment}
Before we explain Algorithm \ref{alg:SubIBR}, we first define a map $\varphi$ and introduce Lemma \ref{lemma:mappingirre} and Lemma \ref{lemma:NoCommonRoot}.
The map $\varphi:\;{\mathbb Q}[x,y]\to{\mathbb Q}[x,y_1]$ is defined as
\begin{align}\label{eq:def-map1}
\varphi(g(x,y))=g(x,\frac{1+y_1}{1-y_1})\cdot(1-y_1)^{\deg(g,y)}.
\end{align}
\begin{lemma}\label{lemma:mappingirre}
If $g$ is an irreducible polynomial in $\Q[x,y]\setminus\{c(y+1)\mid c\ne0\}$, then $\varphi(g)$ is an irreducible polynomial in $\Q[x,y_1]$.
\end{lemma}
\begin{proof}
Since $\deg(g(x,2y-1),y)=\deg(g(x,y),y)$, $\varphi(g)=\mathcal{R}(g(x,2y-1))|_{y=1-y_1}$ with $\mathcal{R}$ in Definition \ref{definition:recip}. Now that $g(x,y)\neq c(y+1)$ for any nonzero $c\in\Q$, $g(x,2y-1)\neq cy$ for any nonzero $c\in\Q$. Moreover, $g(x,2y-1)$ is irreducible since $g(x,y)$ is. By Proposition \ref{proposition:RkeepIrr}, $\mathcal{R}(g(x,2y-1))$ is irreducible. Thus so is $\varphi(g)=\mathcal{R}(g(x,2y-1))|_{y=1-y_1}$.
\end{proof}
The following lemma is similar to \cite[Lemma 3.2]{chen2022automated} and \cite[Theorem 7]{mccallum2012deciding}
\begin{lemma}\label{lemma:NoCommonRoot}
Let $F_1,F_2\in\Q[x,y]$, $k\in\Z$ and $r\in\{k,k+\frac{1}{4}\}$.
If ${\gcd}(F_1,F_2)=1$, then $F_1(r\pi+\arctan{y},y)$ and $F_2(r\pi+\arctan{y},y)$ have no common real roots in $(0,1)$.
\end{lemma}
\begin{proof}
Replacing those ``$p,q$" in the proof of \cite[Theorem 7]{mccallum2012deciding} by $F_1$ and $F_2$ in this lemma, those ``$f^*(x),g^*(x)$" therein by $F_1(r\pi+\arctan{y},y)$ and $F_2(r\pi+\arctan{y},y)$, and the function ``\text{trans}" therein by the function $r\pi+\arctan$, we see that: If $\beta\in(0,1)$ is a common root of $F_1(r\pi+\arctan{y},y)$ and $F_2(r\pi+\arctan{y},y)$, then both $\beta$ and $\alpha=r\pi+\arctan{\beta}$ would be algebraic, which contradicts Claim \ref{claim:atleastonetrans} below.
\end{proof}
\begin{claim}\label{claim:atleastonetrans}
For any $y\in(0,1)$, at least one of $r\pi+\arctan{y}$ and $y$ is transcendental.
\end{claim}
\begin{proof}
Note that \[
\tan(r\pi+\arctan y)=\left\{
\begin{array}{ccl}\vspace{2mm}
y,& \text{ if } &r=k\in\mathbb{Z},\\
\frac{1+y}{1-y},& \text{ if } &r=k+\frac{1}{4}\in\mathbb{Z}+\frac{1}{4}.
\end{array}
\right.
\]
Suppose $y$ is algebraic, then $\tan(r\pi+\arctan y)$ is also algebraic in either case. Since neither $y$ nor $(1+y)/(1-y)$ is zero, $r\pi+\arctan y\neq0$. Moreover, it is transcendental by Theorem \ref{thm:BothAlgeNum}.
\end{proof}
\begin{remark}\label{remark:bisect}
Suppose $h(x_1,x_2)=\sum_{i=0}^{\deg(h,x_1)}c_i(x_2)x_1^i\in\mathbb{Q}[x_1,x_2]$. For any $y\in(0,1)\cap\mathbb{Q}$, $r\pi+\arctan y$ is transcendental. Thus $h(r\pi+\arctan y,y)=0$ iff $c_i(y)=0$ for all $i$. We can therefore decide whether $h(r\pi+\arctan y,y)$ is zero for given $y\in(0,1)\cap\mathbb{Q}$. If it is nonzero, we can further compute its sign by approximating it more accurately. More generally, if $\phi$ is a real function $($\emph{e.g.}, $\phi=\tan)$ such that $\phi(y)$ is transcendental for any nonzero rational $y$ in its domain, then we can decide whether $h(\phi(y),y)$ is zero, and compute its sign whenever it is not zero.
\end{remark}
\begin{comment}
\begin{lemma}\label{lemma:NoCommonRootPre}
Let $k\in\Z$ and $r\in\{k,k+\frac{1}{4}\}$.
For any $y\in(0,1)$, at least one of $r\pi+\arctan{y}$ and $y$ is an transcendental number.
\end{lemma}
\begin{proof}
If $r=k$, assume that $k\pi+\arctan{y}$ and $y$ are both algebraic numbers.
Note that $\tan(k\pi+\arctan{y})=y$. Then, by Theorem \ref{thm:BothAlgeNum}, we have $y=0$, which makes a contradiction.
If $r=k+\frac{1}{4}$, assume that $k\pi+\frac{\pi}{4}+\arctan{y}$ and $y$ are both algebraic numbers.
Let $x:=k\pi+\frac{\pi}{4}+\arctan{y}$.
Then, $\tan{x}=\frac{1+y}{1-y}$.
Let $z:=\tan{x}$.
Then, we have $(y-1)z+y+1=0$.
Since $y$ is an algebraic number, $z$ is also an algebraic number.
Since $x$ and $z$ are both algebraic numbers, by Theorem \ref{thm:BothAlgeNum}, we have $z=0$.
Then, $y=-1$, which makes a contradiction.
\end{proof}
\end{comment}
Recall that our goal is to isolate the real roots of $g(x,\tan{x})$ in $(r\pi,r\pi+\frac{\pi}{4})$, where $g$ is an irreducible polynomial in $\Q[x,y]$ with $\deg(g,y)\ge1$ and $r\in\{k,k+\frac{1}{4}\}~(k\in\Z)$.
The process of Algorithm \ref{alg:SubIBR} is as follows.
If $g=y+1$, then $\tan{x}+1$ has no roots on $(r\pi,r\pi+\frac{\pi}{4})$, and thus return $\emptyset$.
Otherwise, let $F_1:=g$ when $r=k$, and $F_1:=\varphi(g)$ when $r=k+\frac{1}{4}$.
It is not difficult to check that our goal is equivalent to solving $F_1(x+r\pi,\tan{x})$ on $(0,\frac{\pi}{4})$.
Let $y:=\tan{x}$.
We only need to solve $F_1(r\pi+\arctan{y},y)$ on $(0,1)$.
Recall Remark \ref{remark:PolyWeakFourierSeq}.
We can compute a weak Fourier sequence
\begin{align}\label{eq:ComputeFourierSeq}
w_1=F_1(r\pi+\arctan{y},y),\ldots,w_k=F_k(r\pi+\arctan{y},y)~{\rm on}~(0,1),
\end{align}
where $F_i\in\Q[x,y]$.
Thus, the frame of \cite[Algorithm 47]{strzebonski2012real} can be used.
However, there are two main differences between \cite[Algorithm 47]{strzebonski2012real} and Algorithm \ref{alg:SubIBR}.
\begin{comment}
First, we use Algorithm \ref{alg:WeakFourierSeq} to compute a weak Fourier sequence
\begin{align}\label{eq:ComputeFourierSeq}
w_1=G_1(r\pi+\arctan{y},y),\ldots,w_k=G_k(r\pi+\arctan{y},y),
\end{align}
where $G_i\in\Q[x,y]$, for $G_1(r\pi+\arctan{y},y)$ on $(0,1)$ (see Line \ref{algline:SubIBR-weakFseq}), instead of \cite[Algorithm 40]{strzebonski2012real}.
Although \cite[Algorithm 40]{strzebonski2012real} is more general and can also be used to compute a weak Fourier sequence for $G_1(r\pi+\arctan{y},y)$, Algorithm \ref{alg:WeakFourierSeq} is better since it is proposed particularly for our purpose.
\end{comment}
First, when $w_{\ell}$ has exactly one simple root $sr$ on $(c,d)$ (see Line \ref{algline:SubIBR-sr}),
we propose a new method to decide whether $sr$ is a root of $w_1$.
If $F_{\ell}=QF_1~(Q\in\Q[y])$ holds and $Q$ has no roots on $(c,d)$, by \eqref{eq:ComputeFourierSeq}, $sr$ is the only root of $w_1$ on $(c,d)$.
If $F_{\ell}=QF_1$ holds and $Q$ has roots on $(c,d)$, then it is clear that $w_1$ has no roots on $(c,d)$.
If $F_{\ell}=QF_1$ does not hold, then $sr$ is not a root of $w_1$.
This method is correct: By Lemma \ref{lemma:mappingirre}, $F_1$ is an irreducible polynomial.
If $F_1\mid F_{\ell}$, we have $\deg(F_1,x)\le\deg(F_{\ell},x)$.
Then, by Lemma \ref{lemma:pseudoproperty}, $\deg(F_1,x)=\deg(F_{\ell},x)$.
So, $Q$ is a polynomial in $\Q[y]$.
Otherwise, ${\gcd}(F_1,F_{\ell})=1$. Then, by Lemma \ref{lemma:NoCommonRoot}, $sr$ is not a root of $w_1$.
Second, if $sr$ is not a root of $w_1$, we find two rational numbers $u,v$ such that $c\leq u<sr<v\leq d$ and $w_1$ has no roots in $[u,v]$ by Algorithm \ref{alg:RationalEndpoint} (see Line \ref{algline:SubIBR-RationalEP} of Algorithm \ref{alg:SubIBR}), instead of by \cite[Algorithm 45]{strzebonski2012real}. The process of Algorithm \ref{alg:RationalEndpoint} is as follows: Let $n:=1$.
While $n\ge1$, we compute an upper limit polynomial $W_{\max,n}\in\Q[y]$ and a lower limit polynomial $W_{\min,n}\in\Q[y]$ for $w_1(y)$ by two steps.
Since $w_1=F_1(r\pi+\arctan{y},y)$, we first estimate $\arctan{y}$ through Taylor series, and $T_{\max}(n,w_1),T_{\min}(n,w_1)\in\Q[\pi][y]$ are computed by \cite[Def. 3.1]{chen2020automated}.
Then, we use a bounded interval with rational endpoints to estimate $\pi$ in $T_{\max}(n,w_1)$ and $T_{\min}(n,w_1)$, and the polynomials $W_{\max,n},W_{\min,n}\in\Q[y]$ are computed by \cite[Def. 3.1]{xia2006real}.
Let $q:=\frac{c+d}{2}$.
If $W_{\max,n}$ has no roots on $[c,d]$ and $W_{\max,n}(q)<0$, or if $W_{\min,n}$ has no roots on $[c,d]$ and $W_{\min,n}(q)>0$, then $w_1$ has no roots on $[c,d]$.
Otherwise, increase $n$ by $1$.
Since $w_{\ell}$ has exactly one simple root $sr$ on $(c,d)$, we use a bisection method based on Remark \ref{remark:bisect} to find the half open interval where $sr$ is.
Update $(c,d)$ as the half interval and go for the next loop step.
Remark that if the interval $(c,d)$ is small enough and the upper (lower) limit polynomial is close enough, Algorithm \ref{alg:RationalEndpoint} terminates by Lemma \ref{lemma:NoLimit}.
\begin{algorithm}[!t]
\scriptsize
\DontPrintSemicolon
\LinesNumbered
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{
$g(x,y)$, an irreducible polynomial in $\Q[x,y]$ with $\deg(g,y)\ge1$\\
$r$, $r\in\{k,k+\frac{1}{4}\}$ where $k\in\Z$
}
\Output{$ans$, a set of isolating intervals of real roots of $g(x,\tan{x})$ in $(r\pi,r\pi+\frac{\pi}{4})$
}
\caption{{\bf SubIBR}
}\label{alg:SubIBR}
\BlankLine
{\bf if} $g=y+1$, {\bf then} {\bf return} $\emptyset$\;
\eIf{$r=k$}
{$F_1\leftarrow g$}
{$F_1\leftarrow g(x,\frac{1+y}{1-y})\cdot(1-y)^{\deg(g,y)}$\;}
let $w_1=F_1(r\pi+\arctan{y},y),\ldots,w_k=F_k(r\pi+\arctan{y},y)$, where $F_i\in\Q[x,y]$, be the sequence in Remark \ref{remark:PolyWeakFourierSeq}\label{algline:SubIBR-weakFseq}\;
$ans\leftarrow\emptyset$\;
$stack\leftarrow\{(0,1,k)\}$\;
\While{$stack\ne\emptyset$}
{
$(c,d,m)\leftarrow First(stack),\;stack\leftarrow Rest(stack)$\;
$sgc\leftarrow\sgc_{1,m}(c^{+})-\sgc_{1,m}(d^{-})$\;
{\bf if} $sgc=0$, {\bf then} continue\;
\If{$sgc=1$}{$ans\leftarrow ans\cup\{(r\pi+\arctan{c},r\pi+\arctan{d})\}$\; continue\;}
$\ell\leftarrow\min\;\{i\mid 2\le i\le m,~\sgc_{i,m}(c^{+})-\sgc_{i,m}(d^{-})=1\}$\;
$sr\leftarrow\Root(w_{\ell},c,d)$\label{algline:SubIBR-sr}\;
By Theorem \ref{thm:WeakFourierSeq}, $w_{\ell}$ has exactly one simple root $sr$ in $(c,d)$.
We use the notation $\Root(w_{\ell},c,d)$ to represent the only simple root of $w_{\ell}$ in $(c,d)$.\;
\eIf{$F_{\ell}=QF_1$ where $Q\in\Q[y]$\label{algline:SubIBR-wheHold}}
{
\eIf{$Q$ has no real roots on $(c,d)$}
{$ans\leftarrow ans\cup\{(r\pi+\arctan{c},r\pi+\arctan{d})\}$}
{continue}
}
{
$u,v\leftarrow{\bf RationalEndpoint}(w_1,w_{\ell},sr)$\label{algline:SubIBR-RationalEP}\;
add $(c,u,\ell)$ and $(v,d,\ell)$ to $stack$\;
}
}
\Return{$ans$}
\end{algorithm}
\begin{algorithm}[htbp]
\scriptsize
\DontPrintSemicolon
\LinesNumbered
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{
$w_1(y)=F_1(r\pi+\arctan{y},y)$, where $F_1(x,y)\in\Q[x,y]$, $r\in\Q$\\
$w_{\ell}(y)=F_{\ell}(r\pi+\arctan{y},y)$, where $F_{\ell}(x,y)\in\Q[x,y]$\\
$sr=\Root(w_{\ell},c,d)$, $w_{\ell}$ has exactly one simple root $sr$ in $(c,d)$ and $w_{1}(sr)\ne0$
}
\Output{two rational numbers $u$ and $v$ $(c\leq u<sr<v\leq d)$ such that $w_{1}$ has no roots on $[u,v]$
}
\caption{{\bf RationalEndpoint}}\label{alg:RationalEndpoint}
\BlankLine
$n\leftarrow1$\;
\While{$n\ge1$}{
Compute an upper limit polynomial $W_{\max,n}\in\Q[y]$ and a lower limit polynomial $W_{\min,n}\in\Q[y]$ for $w_1(y)$\;
$q\leftarrow\frac{c+d}{2}$\;
\eIf{$(W_{\max,n}$ has no roots on $[c,d]$ and $W_{\max,n}(q)<0)$
{\bf or} \\$\quad\,(W_{\min,n}$ has no roots on $[c,d]$ and $W_{\min,n}(q)>0)$
}
{\Return{$c,d$}}
{
$n\leftarrow n+1$\;
\uIf{$w_{\ell}(q)=0$}
{$c\leftarrow \frac{63q+c}{64},\;d\leftarrow\frac{63q+d}{64}$}
\uElseIf{$\sign(w_{\ell}(c)w_{\ell}(q))=-1$}
{$d\leftarrow q$}
\Else{$c\leftarrow q$}
}
}
\end{algorithm}
\begin{comment}
\begin{algorithm}[!ht]
\DontPrintSemicolon
\LinesNumbered
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{
\begin{itemize}
\item $c,d$, two rational numbers
\item $w_i,\ldots,w_m$, a weak Fourier sequence on $(c,d)$
\item $dir\in\{left,right\}$
\end{itemize}
}
\Output{
$sgc$, if $dir=left$, $sgc$ equals $\sgc_{i,m}(c^{+})$; if $dir=right$, $sgc$ equals $\sgc_{i,m}(d^{-})$
}
\caption{{\bf SGC} \cite[Algorithm 44]{strzebonski2012real}}\label{alg:SGC}
\BlankLine
$s_m\leftarrow\sign(w_m(\frac{c+d}{2}))$\;
{\footnotesize \# Recall Definition \ref{def:WeakFourierSeq}, $w_m(y)\ne0$ for any $y\in(c,d)$. So, $\sign(w_m(y))$ is constant on $(c,d)$.}\;
\eIf{$dir=left$}
{$l_i,\ldots,l_{m-1}\leftarrow\sign(w_i(c^{+})),\ldots,\sign(w_{m-1}(c^{+}))$\;
}
{$l_i,\ldots,l_{m-1}\leftarrow\sign(w_i(d^{-})),\ldots,\sign(w_{m-1}(d^{-}))$}
\For{$j$ {\bf from} $m-1$ {\bf to} $1$}
{If $l_j\ne0$, set $s_j\leftarrow l_j$\;
If $l_j=0$ and $dir=left$, if $s_{j+1}>0$ set $s_j\leftarrow 1$ and otherwise $s_j\leftarrow -1$\;
If $l_j=0$ and $dir=right$, if $s_{j+1}>0$ set $s_j\leftarrow -1$ and otherwise $s_j\leftarrow 1$\;
}
\Return{the number of sign changes in the sequence $s_1,\ldots,s_m$}\;
\end{algorithm}
\end{comment}
\begin{comment}
\begin{proposition}
If $f(x,y)\in\Q[x,y]\setminus\{0\}$ and $f_1(x,\arctan(x)),\ldots,f_k(x,\arctan(x))$ is a weak Fourier sequence for $f(x,\arctan(x))$ computed by Algorithm \ref{alg:WeakFourierSeq},
then for any nonzero rational number $r$, $f(x,y+r\pi)\in\Q[\pi][x,y]\setminus\{0\}$ and $f_1(x,\arctan(x)+r\pi),\ldots,f_k(x,\arctan(x)+r\pi)$ is a weak Fourier sequence for $f(x,\arctan(x)+r\pi)$.
\end{proposition}
\begin{proof}
It is obvious that $f(x,y+r\pi)\in\Q[\pi][x,y]\setminus\{0\}$.
Note that
$${\trueder f}(x,\arctan(x)+r\pi)=f'(x,\arctan(x)+r\pi).$$
So, by the correctness proof of Theorem \ref{thm:alg:WeakFourierSeqTC}, $f_1(x,\arctan(x)+r\pi),\ldots,f_k(x,\arctan(x)+r\pi)$ is a weak Fourier sequence for $f(x,\arctan(x)+r\pi)$.
\end{proof}
\end{comment}
\section{Isolating Bounded Real Roots}\label{sec:IBR}
In this section, we discuss case \ref{item:MainThreeCase1} in Section \ref{sec:Preprocessing}, i.e., given $k_+,k_-\in\Z$, isolating the real roots of
\begin{align}\label{eq:boudedeq}
g(x,\tan x)=0,~~~~x\in(k\pi-\frac{\pi}{2},k\pi+\frac{\pi}{2})
\end{align}
where $g\in\Q[x,y]\setminus\{0\}$, $k\in\Z$ and $k_-\le k\le k_+$.
We solve \eqref{eq:boudedeq} in four intervals $(k\pi-\frac{\pi}{2},k\pi-\frac{\pi}{4}),(k\pi-\frac{\pi}{4},k\pi),(k\pi,k\pi+\frac{\pi}{4})$ and $(k\pi+\frac{\pi}{4},k\pi+\frac{\pi}{2})$.
In each interval, the problem is equivalent to isolating the roots of some $G(r\pi+\arctan{y},y)$ on $(0,1)$ where $G\in\Q[x,y]$ and $r\in\Q$, and thus we can use the frame of \cite[Algorithm 47]{strzebonski2012real} to deal with it.
In Section \ref{subsec:WeakFourierSeq}, we use Algorithm \ref{alg:WeakFourierSeq} to compute a weak Fourier sequence for $G(r\pi+\arctan{y},y)$.
In fact, Algorithm \ref{alg:WeakFourierSeq} is a simplified version of \cite[Algorithm 40]{strzebonski2012real} for our specific needs.
In Section \ref{subsec:IsoRoot}, Algorithm \ref{alg:BoundRSI} is proposed to isolate the real roots of Eq. \eqref{eq:boudedeq}.
\subsection{Weak Fourier sequences}\label{subsec:WeakFourierSeq}
Let us first review the definition of weak Fourier sequences and an important result about it.
\begin{definition}\cite[Def. 10]{strzebonski2012real}\label{def:WeakFourierSeq}
Let $I\subseteq\R$ be an open interval, and $G_1,\ldots,G_k:I\rightarrow\R$ be a sequence of differentiable functions defined on $I$.
The sequence $G_1,\ldots,G_k$ is a \emph{weak Fourier sequence on $I$}, if for all $y\in I$,
\begin{enumerate}
\item for all $i\;(1\le i\le k-1)$, $\sign(G_{i+1}(y))=\sign(G_i'(y))$, and
\item $G_k(y)\ne0$.
\end{enumerate}
We say $G_1,\ldots,G_k$ is a \emph{weak Fourier sequence for $G$ on $I$}, if $G_1,\ldots,G_k$ is a weak Fourier sequence on $I$ and $G_1(y)=G(y)$ for all $y\in I$.
\end{definition}
\begin{notation}
Let $a,b\in\R$ and $G_1,\ldots,G_k$ be a weak Fourier sequence on the open interval $(a,b)$.
Then, for any $i,j\;(1\le i\le j\le k)$,
\begin{enumerate}
\item $\sign(G_{i}(x^{+})):=\sign(G_i(c))$, where $x\in[a,b)$, $c< b$ and $\forall y\in(x,c]\;G_i(y)\ne0$,
\item $\sign(G_{i}(x^{-})):=\sign(G_i(d))$, where $x\in(a,b]$, $a< d$ and $\forall y\in[d,x)\;G_i(y)\ne0$,
\item $\sgc_{i,j}(x)$, for $x\in(a,b)$, is the number of sign changes in the sequence $G_{i}(x),\ldots,G_{j}(x)$ with terms equal to zero removed,
\item $\sgc_{i,j}(x^{+})$, for $x\in[a,b)$, is the number of sign changes in the sequence $G_{i}(x^{+}),\ldots,G_{j}(x^{+})$,
\item $\sgc_{i,j}(x^{-})$, for $x\in(a,b]$, is the number of sign changes in the sequence $G_{i}(x^{-}),\ldots,G_{j}(x^{-})$.
\end{enumerate}
\end{notation}
\begin{theorem}\cite[Thm. 13]{strzebonski2012real}\label{thm:WeakFourierSeq}
Let $I\subseteq\R$ be an open interval and let
$G_1,\ldots,G_{k}:I\rightarrow\R$ be a weak Fourier sequence.
Then for any $x\in I$, $\sgc_{1,k}(x^{+})=\sgc_{1,k}(x)$ and $\sgc_{1,k}(x^{-})=\sgc_{1,k}(x)+r+2s$, where $r,s\in\N$ and $G_1(x)=\cdots=G_r(x)=0,G_{r+1}(x)\ne0$.
Moreover, $s=0$ unless there is a $t>r+1$ such that $G_{t}(x)=0$.\qed
\end{theorem}
Let $a,b$ be two real numbers such that $(a,b)\subseteq I$.
Remark that by the proof of \cite[Prop. 15]{strzebonski2012real}, if $\sgc_{1,k}(a^{+})=\sgc_{1,k}(b^{-})$, then $G_1$ has no real root in $(a,b)$, and if $\sgc_{1,k}(a^{+})=\sgc_{1,k}(b^{-})+1$, then $G_1$ has exactly one simple root in $(a,b)$.
To illustrate Algorithm \ref{alg:WeakFourierSeq}, we introduce the definition of the \emph{arctan-derivative}.
\begin{definition}\label{def:pseudo}
For any $G\in\Q(y)[x]$, let
\[
{\trueder G}:=\frac{\partial G}{\partial x}\cdot\frac{1}{1+y^2}+\frac{\partial G}{\partial y}\in\Q(y)[x].
\]
The \emph{arctan-derivative} of $G$ is defined as $G^{\pder}:={\trueder G}\cdot h(y)$, where $h(y)\in\Q[y]$ is the denominator of $\lc({\trueder G},x)\in\Q(y)$.
In particular, if $\lc({\trueder G},x)$ is a polynomial, then $h(y)=1$.
We call $h(y)$ the \emph{common-term} of $G^{\pder}$.
\end{definition}
\begin{remark}\label{remark:def:pseudo}
For any $G\in\Q(y)[x]$, $G^{\pder}\in\Q(y)[x]$ and $\lc(G^{\pder},x)$ is a polynomial in $\Q[y]$.
\end{remark}
Given $G(x,y)\in\Q[x,y]\setminus\{0\}$ and $r\in\Q$,
the process of Algorithm \ref{alg:WeakFourierSeq} is as follows:
Let $G_1$ be $G(x,y)$.
If $G_1$ is not a nonzero rational number, then compute the arctan-derivative $G_{1}^{\pder}$ of $G_1$, and set $G_2=G_{1}^{\pder}$.
If $G_2$ is not a nonzero rational number, then set $G_3=G_{2}^{\pder}$.
Repeat the process until some $G_k\;(k\ge1)$ is a nonzero rational number.
Then $G_1(r\pi+\arctan y,y),\allowbreak\ldots,G_k(r\pi+\arctan y,y)$ is a weak Fourier sequence for $G(r\pi+\arctan y,y)$ on $\R$.
\begin{algorithm}[!ht]
\scriptsize
\DontPrintSemicolon
\LinesNumbered
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{ $G(x,y)$, a polynomial in $\Q[x,y]\setminus\{0\}$; $r$, a rational number
}
\Output{$G_1(r\pi+\arctan{y},y),\ldots,G_k(r\pi+\arctan{y},y)$, a weak Fourier sequence for $G(r\pi+\arctan{y},y)$ on $\R$, where $G_i\in\Q(y)[x]$
}
\caption{{\bf WeakFourierSeq}}\label{alg:WeakFourierSeq}
\BlankLine
$G_1\leftarrow G(x,y)$\;
$i\leftarrow1$\;
\While{$G_i(x,y)$ is not a nonzero rational number}
{ $G_{i+1}(x,y)\leftarrow G_{i}^{\pder}(x,y)$\;
$i\leftarrow i+1$
}
\Return{$G_1(r\pi+\arctan{y},y),\ldots,G_k(r\pi+\arctan{y},y)$}
\end{algorithm}
Note that we can easily obtain a weak Fourier sequence from a \emph{semi-Fourier sequence} (defined in \cite[Def. 14]{strzebonski2012real}) by \cite[Prop. 15]{strzebonski2012real}.
So, Algorithm \ref{alg:WeakFourierSeq} and \cite[Algorithm 40]{strzebonski2012real} are essentially the same when computing a weak Fourier sequence for $G(r\pi+\arctan{y},y)$, and thus the termination and correctness of Algorithm \ref{alg:WeakFourierSeq} are clear.
However, in order to explain some properties that will be used later, we prepare some lemmas and re-prove the termination and correctness of Algorithm \ref{alg:WeakFourierSeq}.
\begin{lemma}\label{lemma:DerEquiv0}
If $G\in\Q(y)[x]\setminus\Q$ such that $G(x,y)$ is well-defined on the whole plane $\R^2$,
then $G^{\pder}\in\Q(y)[x]\setminus\{0\}$.
\end{lemma}
\begin{proof}
Assume that $G^{\pder}$ is the zero polynomial in $\Q(y)[x]$.
By Definition \ref{def:pseudo}, ${\trueder G}$ is also the zero polynomial.
Then, ${\trueder G}(\arctan{y},y)\equiv0$.
Note that ${\trueder G}(\arctan{y},y)=\frac{\der G(\arctan{y},y)}{\der y}$.
So, $G(\arctan{y},y)\in\Q$ for any $y\in\R$. Since the arctangent function is not algebraic, $G\in\Q$, a contradiction.
\end{proof}
\begin{lemma}\label{lemma:pseudoproperty}
If $G\in\Q(y)[x]\setminus\Q$ and $\lc(G,x)$ is a polynomial, then
\begin{align}\label{eq:lemma:pseudoproperty}
(\deg(G,x),\deg(\lc(G,x)))\succlex(\deg(G^{\pder},x),\deg(\lc(G^{\pder},x))),
\end{align}
where $\succlex$ denotes the lexicographic ordering.
\end{lemma}
\begin{proof}
Recall Remark \ref{remark:def:pseudo}. The degrees $\deg(G^{\pder},x)$ and $\deg(\lc(G^{\pder},x))$ are well-defined.
If $\deg(G,x)=0$, we have $G\in\Q(y)\setminus\Q$, and thus $\lc(G,x)=G$.
So, $G$ is a polynomial in $\Q[y]\setminus\Q$.
It is easy to check that $G^{\pder}=\frac{\partial G}{\partial y}\in\Q[y]$.
So, $\deg(\lc(G,x))=\deg(G)$, $\deg(G^{\pder},x)=0$ and $\deg(\lc(G^{\pder},x))=\deg(G)-1$.
Then inequality \eqref{eq:lemma:pseudoproperty} holds.
If $\deg(G,x)\ge1$, suppose that
\begin{align}\label{eq:G}
G=a_n(y)x^n+\sum_{j=0}^{n-1}a_j(y)x^
\end{align}
where $n=\deg(G,x)\ge1$, $a_n(y)\in\Q[y]$ and $a_j(y)\in\Q(y)$ for $j=0,\ldots,n-1$.
Then,
\begin{align}\label{eq:truederG}
{\trueder G}=a_n'(y)x^n+&\sum_{j=0}^{n-1}\big(a_j'(y)+\frac{(j+1)a_{j+1}(y)}{1+y^2}\big)x^{j}.
\end{align}
If $a_{n}'(y)$ is the zero polynomial, we have $\deg({\trueder G},x)<\deg(G,x)$.
By Definition \ref{def:pseudo}, $\deg(G^{\pder},x)=\deg({\trueder G},x)$.
So, $\deg(G^{\pder},x)<\deg(G,x)$ and inequality \eqref{eq:lemma:pseudoproperty} is true.
If $a_{n}'(y)$ is not the zero polynomial, then $\deg({\trueder G},x)=\deg(G,x)$.
Since $\lc({\trueder G},x)=a_{n}'(y)\in\Q[y]$, $G^{\pder}={\trueder G}$. Thus $\deg(G^{\pder},x)=\deg(G,x)$.
We also note that $\deg(\lc(G,x))=\deg(a_n(y))$ and $\deg(\lc(G^{\pder},x))=\deg(a_n(y))-1$.
The proof is completed.
\end{proof}
\begin{lemma}\label{lemma:DerForm}
If $G\in\Q(y)[x]$ has the following form
\begin{align}\label{eq:specialform}
G=a_n(y)x^{n}+\sum_{j=0}^{n-1}\frac{a_j(y)}{(1+y^2)^{m_j}}x^{j}
\end{align}
where $n=\deg(G,x)$, $a_i(y)\in\Q[y]\;(0\le i\le n)$ and $m_j\in\N~(0\le j\le n-1)$, then the common-term of $G^{\pder}$ is $(1+y^2)^m$ for some $m\in\N$.
Furthermore, $G^{\pder}$ also has the form in Eq. \eqref{eq:specialform}.
\end{lemma}
\begin{proof}
If $\deg(G,x)=0$, then $G=a_n(y)\in\Q[y]$, and thus $G^{\star}=G'(y)$, the common-term of $G^{\star}$ is $1$.
If $\deg(G,x)\ge1$, by Eq. \eqref{eq:G} and Eq. \eqref{eq:truederG} in the proof of Lemma \ref{lemma:pseudoproperty}, there exist $b_i(y)\in\Q[y]\;(0\le i\le n)$ and $d_j\in\N~(0\le j\le n-1)$ such that
\begin{align*}
{\trueder G}=b_n(y)x^n+\sum_{j=0}^{n-1}\frac{b_j(y)}{(1+y^2)^{d_j}}x^{j}
\end{align*}
If $b_n(y)$ is not the zero polynomial, the common-term of $G^{\pder}$ is $1$.
Otherwise, since $\deg(G,x)\ge1$ and $G$ is well-defined on the whole $\R^2$, by the proof of Lemma \ref{lemma:DerEquiv0}, ${\trueder G}$ is not the zero polynomial.
So, there exists an integer $0\le k\le n-1$ such that $b_k(y)$ is not the zero polynomial. Let $k_0$ be the maximal one of such $k$. Then, the common-term of $G^{\pder}$ is $(1+y^2)^{d_{k_0}}$.
Therefore, $G^{\pder}$ has the same form as in Eq. \eqref{eq:specialform}.
\end{proof}
\begin{theorem}\label{thm:alg:WeakFourierSeqTC}
Algorithm \ref{alg:WeakFourierSeq} terminates correctly.
\end{theorem}
\begin{proof}
(Termination)
For any $G(x,y)\in\Q[x,y]\setminus\{0\}$,
let $G_1:=G(x,y),G_2:=G_1^{\pder}(x,y),G_3:=G_2^{\pder}(x,y),\ldots$
We only need to prove that there exists $k\;(k\ge1)$ such that $G_k(x,y)$ is a nonzero rational number.
If $G_1\in\Q\setminus\{0\}$, then we take $k=1$.
If $G_1\in\Q[x,y]\setminus\Q$, then by Lemma \ref{lemma:pseudoproperty}, we have
$$(\deg(G_1,x),\deg(\lc(G_1,x)))\succlex(\deg(G_2,x),\deg(\lc(G_2,x))).$$
Recall Remark \ref{remark:def:pseudo}. For any $i~(i\ge2)$, $G_i\in\Q(y)[x]$ and $\lc(G_i,x)$ is a polynomial.
Therefore, assume that $G_2\notin\Q$, $G_3\notin\Q$,..., by Lemma \ref{lemma:pseudoproperty}, we have
\begin{align*}
(\deg(G_1,x),\deg(\lc(G_1,x)))&\succlex(\deg(G_2,x),\deg(\lc(G_2,x)))\\
&\succlex(\deg(G_3,x),\deg(\lc(G_3,x)))\succlex\cdots.
\end{align*}
So, there exists $m\;(m\ge2)$ such that $G_m\in\Q$.
Let $k\;(k\ge2)$ be the smallest integer such that $G_k\in\Q$.
Then, $G_{k-1}\notin\Q$.
Note that $G_{k-1}$ is well-defined on the whole $\R^2$ by Lemma \ref{lemma:DerForm}.
So, by Lemma \ref{lemma:DerEquiv0}, $G_k$ is a nonzero rational number.
(Correctness)
For $i=1,\ldots,k-1$, let $h_i(y)$ be the common-term of $G_i^{\pder}(x,y)$.
Recall Definition \ref{def:pseudo}.
We have $G_{i+1}(x,y)=h_i(y){\trueder G}_{i}(x,y)$.
Note that
$${\trueder G}_{i}(r\pi+\arctan{y},y)=G_i(r\pi+\arctan{y},y)'$$
and $G_{i+1}(x,y)\in\Q(y)[x]$.
Thus, we only need to prove that $\sign(h_i(y))>0$ on the whole $\R$ and $G_{i+1}(r\pi+\arctan{y},y)$ is defined on the whole $\R$.
If $\deg(G,x)=0$, then $G\in\Q[y]$.
So, $h_i(y)=1$ and $G_{i+1}(r\pi+\arctan{y},y)$ is defined on the whole $\R$.
If $\deg(G,x)\ge1$,
by the termination proof, there exists $m\;(1\le m\le k-1)$ such that $\deg(G_1,x)\ge1,\ldots,\deg(G_m,x)\ge1$ but $\deg(G_{m+1},x)=0$.
Applying Lemma \ref{lemma:DerForm} to $G_i\;(i=1,\dots,m)$, we have
$h_i(y)=(1+y^2)^{n_i}\;(n_i\in\N)$ and $G_{i+1}(r\pi+\arctan y,y)$ is defined on the whole $\R$.
Since $\deg(G_{m+1},x)=0$, $G_{m+1}\in\Q[y]$.
So, for any $m+1\le i\le k-1$, $h_i(x)=1$ and
$G_{i+1}\in\Q[y]$, which is also defined on the whole $\R$.
\end{proof}
\begin{remark}\label{remark:PolyWeakFourierSeq}
Let $G(x,y)\in\Q[x,y]\setminus\{0\}$, $r\in\Q$ and $G_1(r\pi+\arctan{y},y),\ldots,\allowbreak G_k(r\pi+\arctan{y},y)$ be a weak Fourier sequence for $G(r\pi+\arctan{y},y)$ on $\R$ computed by Algorithm \ref{alg:WeakFourierSeq}.
By the correctness proof of Theorem \ref{thm:alg:WeakFourierSeqTC}, $G_i(x,y)\in\Q[y]$ or
\begin{align*}
G_i(x,y)=a_{in_i}(y)x^{n_i}+\sum_{j=0}^{n_{i}-1}\frac{a_{ij}(y)}{(1+y^2)^{m_{ij}}}x^{j}
\end{align*}
where $n_i=\deg(G_i,x)\ge1$, $a_{in_i}(y)\in\Q[y]~(1\le i\le k)$ and $m_{ij}\in\N~(0\le j\le n_{i}-1)$.
For $i=1,\ldots,k$, let $m_i:=\max(m_{ij},0\le j\le n_i-1)$ and
\begin{align*}
F_i:=
\left\{
\begin{array}{cc}
G_i, & G_i\in\Q[y],\\
(1+y^2)^{m_{i}}G_i, & {\rm otherwise}.
\end{array}
\right.
\end{align*}
We have $F_i\in\Q[x,y]$, and $F_i(r\pi+\arctan{y},y)$ has the same signs and real roots as $G_i(r\pi+\arctan{y},y)$ on $\R$.
Thus, we consider $F_i$ instead of $G_i$ in the next subsection.
\end{remark}
\subsection{Isolating the roots}\label{subsec:IsoRoot}
Recall that the goal of Section \ref{sec:IBR} is to isolate the roots of the following equation:
\begin{align}\label{meq:boudedeq1_9}
g(x,\tan{x}),~x\in(k\pi-\frac{\pi}{2},k\pi+\frac{\pi}{2}),~{\rm where}~k\in\Z.
\end{align}
\subsubsection{The algorithm}
Given $g(x,y)\in\Q[x,y]\setminus\{0\}$ and $k\in\Z$,
we propose Algorithm \ref{alg:BoundRSI} to isolate the real roots of Eq. \eqref{meq:boudedeq1_9}.
The process of Algorithm \ref{alg:BoundRSI} is as follows.
We first decide whether $k\pi$, $-\frac{\pi}{4}+k\pi$ or $\frac{\pi}{4}+k\pi$ is a root of $g(x,\tan{x})$ by Lemma \ref{lemma:piRealRoot}.
Then, we compute all irreducible factors $g_1,\ldots,g_m$ of $g(x,y)$ in $\Q[x,y]$.
For every $g_i(x,\tan{x})$, we only need to isolate the real roots of it in four open intervals, i.e., $(k\pi-\frac{\pi}{2},k\pi-\frac{\pi}{4})$, $(k\pi-\frac{\pi}{4},k\pi)$, $(k\pi,k\pi+\frac{\pi}{4})$ and $(k\pi+\frac{\pi}{4},k\pi+\frac{\pi}{2})$.
If $g_i$ is a polynomial in $\Q[x]$, it is easy to deal with (there are many methods to isolate the real roots of univariate polynomials).
Otherwise, we have $\deg(g_i,y)\ge1$.
Let $g_{i}^{(1)}(x,y):=g_{i}(-x,-y)$.
It is easy to check that isolating the real roots of $g_i(x,\tan{x})$ in the first interval (\emph{resp.}, the second interval) is equivalent to isolating those of $g_{i}^{(1)}(x,\tan{x})$ in $(-k\pi+\frac{\pi}{4},-k\pi+\frac{\pi}{2})$ (\emph{resp.}, $(-k\pi,-k\pi+\frac{\pi}{4})$).
Note that these four cases can all be reduced to the problem of isolating the real roots of $G(x,\tan{x})$ on $(r\pi,r\pi+\frac{\pi}{4})$, where $G\in\{g_i,g_{i}^{(1)}\}$ is a rational bivariate irreducible polynomial with $\deg(G,y)\ge1$ and $r\in\{k_0,k_0+\frac{\pi}{4}\}~(k_0\in\Z)$.
And, Algorithm \ref{alg:SubIBR} can solve such problem, which we will explain in detail in the next subsection. Finally, in Algorithm \ref{alg:AlgDerBound} the isolating intervals obtained from each irreducible factor $g_i$ are refined so that they do not intersect. This is by using Remark \ref{remark:bisect} and the fact that the roots in them are all distinct simple roots (Prop. \ref{prop:multiplicity-one}--\ref{proposition:nocommonroot}).
\begin{algorithm}[!ht]
\scriptsize
\DontPrintSemicolon
\LinesNumbered
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{
$g(x,y)$, a polynomial in $\Q[x,y]\setminus\{0\}$\\
$k$, an integer
}
\Output{$ans$, a set of pairwise disjoint isolating intervals of real roots of $g(x,\tan{x})$ in $(k\pi-\frac{\pi}{2},k\pi+\frac{\pi}{2})$
}
\caption{{\bf IsolatingBoundedRoots}}\label{alg:BoundRSI}
\BlankLine
$ans\leftarrow\emptyset$\;
If $g(k\pi,0)=0$, then $ans\leftarrow ans\cup\{[k\pi,k\pi]\}$\;
If $g(-\frac{\pi}{4}+k\pi,-1)=0$, then $ans\leftarrow ans\cup\{[-\frac{\pi}{4}+k\pi,-\frac{\pi}{4}+k\pi]\}$\;
If $g(\frac{\pi}{4}+k\pi,1)=0$, then $ans\leftarrow ans\cup\{[\frac{\pi}{4}+k\pi,\frac{\pi}{4}+k\pi]\}$\;
$g_1,\ldots,g_m\leftarrow$ all irreducible factors of $g(x,y)$ in $\Q[x,y]$\;
\For{$i$ from $1$ to $m$}
{
\eIf{$g_i$ is a polynomial in $\Q[x]$}
{$S\leftarrow$ a set of isolating intervals of real roots of $g_i(x)$ in $(k\pi-\frac{\pi}{2},k\pi+\frac{\pi}{2})\setminus\{k\pi,-\frac{\pi}{4}+k\pi,\frac{\pi}{4}+k\pi\}$\;
$ans\leftarrow ans\cup S$\;
}
{
let $g_{i}^{(1)}\in\Q[x,y]$ such that $g_{i}^{(1)}(x,y)=g_{i}(-x,-y)$\;
$S_1\leftarrow{\bf SubIBR}(g_{i}^{(1)},-k+\frac{1}{4})$\;
$S_2\leftarrow{\bf SubIBR}(g_{i}^{(1)},-k)$\;
$S_3\leftarrow{\bf SubIBR}(g_{i},k)$\;
$S_4\leftarrow{\bf SubIBR}(g_{i},k+\frac{1}{4})$\;
$ans\leftarrow ans\cup\{(-b,-a)\mid (a,b)\in S_1\cup S_2\}\cup S_3\cup S_4$\;
}
}
Refine the intervals obtained in the loop above such that they do not intersect.\;
\Return{$ans$}\;
\end{algorithm}
\begin{comment}
Let $x_1:=x-k\pi$.
It is equivalent to isolating the real roots of $g_i(x_{1}+k\pi,\tan{x_1})$ in $(-\frac{\pi}{2},-\frac{\pi}{4})$, $(-\frac{\pi}{4},0)$, $(0,\frac{\pi}{4})$ and $(\frac{\pi}{4},\frac{\pi}{2})$.
For the first, the second and the fourth case, we can transform $g_i$ to some bivariate rational polynomial $\tilde{g_i}$ and we only need to consider the real roots of $\tilde{g_i}$ in $(0,\frac{\pi}{4})$.
Take the first case for example.
Let $x_2:=-x_{1}$ and $g_{i}^{(1)}:=g_{i}(-x,-y)$.
It is easy to check that solving $g_i(x_1+k\pi,\tan{x_1})$ where $x_1\in(-\frac{\pi}{2},-\frac{\pi}{4})$ is equivalent to solving $g_i^{(1)}(x_2-k\pi,\tan{x_2})$ where $x_2\in(\frac{\pi}{4},\frac{\pi}{2})$.
Let $x_3:=x_2-\frac{\pi}{4}$, then $x_3\in(0,\frac{\pi}{4})$. Note that $\tan(x_3+\frac{\pi}{4})=\frac{1+\tan{x_3}}{1-\tan{x_3}}$ and $1-\tan{x_3}>0$.
Then, let $g_{i}^{(2)}\in\Q[x,y]$ be the product of $g_{i}^{(1)}(x,\frac{1+y}{1-y})$ and the common denominator of $g_{i}^{(1)}(x,\frac{1+y}{1-y})$, and thus it becomes solving $g_{i}^{(2)}(x_3+(-k+\frac{1}{4})\pi,\tan{x_3})$ where $x_3\in(0,\frac{\pi}{4})$.
By Algorithm \ref{alg:SubIBR}, we can isolate the real roots of $g_{i}^{(2)}(x_3+(-k+\frac{1}{4})\pi,\tan{x_3})$ in $(0,\frac{\pi}{4})$.
Remark that $x_3=-x+k\pi-\frac{\pi}{4}$.
If $(a,b)$ is an isolating interval of $g_{i}^{(2)}(x_2+(-k+\frac{1}{4})\pi,\tan{x_2})$ in $(0,\frac{\pi}{4})$, it is obvious that $(-b+k\pi-\frac{\pi}{4},-a+k\pi-\frac{\pi}{4})$ is an isolating interval of $g_i(x,\tan{x})$ in $(k\pi-\frac{\pi}{2},k\pi-\frac{\pi}{4})$.
\end{comment}
\subsubsection{Sub-algorithms}
Let $g(x,y)$ be an irreducible polynomial in $\Q[x,y]$ with $\deg(g,y)\ge1$ and $r\in\{k,k+\frac{1}{4}\}$ where $k\in\Z$.
According to the frame of \cite[Algorithm 47]{strzebonski2012real}, Algorithm \ref{alg:SubIBR} is proposed to isolate the real roots of $g(x,\tan{x})$ in $(r\pi,r\pi+\frac{\pi}{4})$.
It is worth noting that \cite[Algorithm 47]{strzebonski2012real} and Algorithm \ref{alg:SubIBR} are not the same.
Some steps are added, and some steps are skipped or implemented in another way.
The algorithm \cite[Algorithm 47]{strzebonski2012real} can be used to isolate the real roots of an exp-log-arctan function $w(y)$ in an open interval $(a,b)$, where $a,b$ are rational numbers and $(a,b)$ are contained in the domain of $w(y)$.
Let us review the frame of the algorithm.
We need to compute a weak Fourier sequence $w_1,\ldots,w_k$ for $w(y)$ on $(a,b)$ first.
Let $ans$ be a set of computed isolating intervals of $w(y)$ (initialized as $\emptyset$), and $stack$ be a stack of triples $(c,d,m)$ such that $c<d$ are rational numbers and $w_m$ has a constant nonzero sign in $(c,d)$ (initialized as $(a,b,k)$).
In each loop step, we pick a triple $(c,d,m)$ from $stack$ and compute the difference value $\sgc_{1,m}(c^{+})-\sgc_{1,m}(d^{-})$ by \cite[Algorithm 44]{strzebonski2012real},
until $stack$ is empty.
\begin{enumerate}
\item If the value equals $0$, by Theorem \ref{thm:WeakFourierSeq}, $w(y)$ has no roots on $(c,d)$. So, we continue for the next loop step.
\item If the value equals $1$, by Theorem \ref{thm:WeakFourierSeq}, $w(y)$ has exactly one simple root on $(c,d)$. We add $(c,d)$ to $ans$.
\item Otherwise, let $\ell$ be the smallest element of $\{i\mid 2\le i\le m \;{\rm and}\; \sgc_{i,m}(c^{+})\allowbreak-\sgc_{i,m}(d^{-})=1\}$.
By Theorem \ref{thm:WeakFourierSeq}, $w_{\ell}(y)$ has exactly one simple root $sr$ on $(c,d)$.
\begin{enumerate}
\item If $sr$ is a root of $w(y)$, find two rational numbers $u,v$ such that $c<u<sr<v<d$ and $w(y)$ has only one root $sr$ in $[u,v]$
(recall that by Lemma \ref{lemma:NoLimit}, we can find such $u,v$).
Add $(u,v)$ to $ans$, and add $(c,u,\ell)$, $(v,d,\ell)$ to $stack$.
\item If $sr$ is not a root of $w(y)$. Find two rational numbers $u,v$ such that $c<u<sr<v<d$ and $w(y)$ has no roots in $[u,v]$.
Add $(c,u,\ell)$ and $(v,d,\ell)$ to $stack$.
\end{enumerate}
\end{enumerate}
\begin{comment}
Back to Algorithm \ref{alg:SubIBR}, we will explain some differences between the two algorithms and some implementation details in the following.
Instead of using \cite[Algorithm 40]{strzebonski2012real} to compute a semi-Fourier sequence, we directly compute a weak Fourier sequence $w_1=G_1(r\pi+\arctan{y},y),\ldots,w_k=G_k(r\pi+\arctan{y},y)$, where $G_i\in\Q(y)[x]$, for $w(y)$ on $(0,1)$ by Algorithm \ref{alg:WeakFourierSeq}.
Recall Remark \ref{remark:PolyWeakFourierSeq}.
We have $G_i\in\Q[x,y]$ in fact (see Algorithm \ref{alg:SubIBR}-Line \ref{algline:SubIBR-weakFseq}).
{\bf How to compute $\sgc_{i,m}(c^{+})$ or $\sgc_{i,m}(c^{-})$, where $c\in\Q$ and $1\le i\le m$?} We use Algorithm \cite[Algorithm 44]{strzebonski2012real}, i.e. Algorithm \ref{alg:SGC} in the paper, to compute it.
\end{comment}
\begin{comment}
\begin{definition}
A \emph{root constant} is an expression $\Root(G,a,b)$, where $G\in\R[x]$, $a<b$ are rational numbers such that $(a,b)$ is contained in the domain of $G$ and $G$ has the only root in $(a,b)$. $\Root(G,a,b)$ represents the only root of $G$ in $(a,b)$.
\end{definition}
\begin{lemma}
Let $G(x,y)$ be an irreducible polynomial in $\Q[x,y]$ and $r\in\Q$.
Then, $w(y)=G(r\pi+\arctan{y},y)$ has no multiple roots in $(0,1)$.
\end{lemma}
\end{comment}
\begin{comment}
\begin{lemma}
Let $w_1(y)=G_1(r\pi+\arctan{y},y),\ldots,w_k(y)=G_k(r\pi+\arctan{y},y)$ be a weak Fourier sequence for $w_1(y)$, where $G_i\in\Q[x,y]$ (see Remark \ref{remark:PolyWeakFourierSeq}) and $\Root(w_{\ell},c,d)~(l\ge2)$ be a simple root.
Then, there exists $u,v\in\Q$ such that $c<u<\Root(w_{\ell},c,d)<v<d$ and $\sgc_{1,l}$
\end{lemma}
\end{comment}
Before we explain Algorithm \ref{alg:SubIBR}, we first define a map $\varphi$ and introduce Lemma \ref{lemma:mappingirre} and Lemma \ref{lemma:NoCommonRoot}.
The map $\varphi:\;{\mathbb Q}[x,y]\to{\mathbb Q}[x,y_1]$ is defined as
\begin{align}\label{eq:def-map1}
\varphi(g(x,y))=g(x,\frac{1+y_1}{1-y_1})\cdot(1-y_1)^{\deg(g,y)}.
\end{align}
\begin{lemma}\label{lemma:mappingirre}
If $g$ is an irreducible polynomial in $\Q[x,y]\setminus\{c(y+1)\mid c\ne0\}$, then $\varphi(g)$ is an irreducible polynomial in $\Q[x,y_1]$.
\end{lemma}
\begin{proof}
Since $\deg(g(x,2y-1),y)=\deg(g(x,y),y)$, $\varphi(g)=\mathcal{R}(g(x,2y-1))|_{y=1-y_1}$ with $\mathcal{R}$ in Definition \ref{definition:recip}. Now that $g(x,y)\neq c(y+1)$ for any nonzero $c\in\Q$, $g(x,2y-1)\neq cy$ for any nonzero $c\in\Q$. Moreover, $g(x,2y-1)$ is irreducible since $g(x,y)$ is. By Proposition \ref{proposition:RkeepIrr}, $\mathcal{R}(g(x,2y-1))$ is irreducible. Thus so is $\varphi(g)=\mathcal{R}(g(x,2y-1))|_{y=1-y_1}$.
\end{proof}
The following lemma is similar to \cite[Lemma 3.2]{chen2022automated} and \cite[Theorem 7]{mccallum2012deciding}
\begin{lemma}\label{lemma:NoCommonRoot}
Let $F_1,F_2\in\Q[x,y]$, $k\in\Z$ and $r\in\{k,k+\frac{1}{4}\}$.
If ${\gcd}(F_1,F_2)=1$, then $F_1(r\pi+\arctan{y},y)$ and $F_2(r\pi+\arctan{y},y)$ have no common real roots in $(0,1)$.
\end{lemma}
\begin{proof}
Replacing those ``$p,q$" in the proof of \cite[Theorem 7]{mccallum2012deciding} by $F_1$ and $F_2$ in this lemma, those ``$f^*(x),g^*(x)$" therein by $F_1(r\pi+\arctan{y},y)$ and $F_2(r\pi+\arctan{y},y)$, and the function ``\text{trans}" therein by the function $r\pi+\arctan$, we see that: If $\beta\in(0,1)$ is a common root of $F_1(r\pi+\arctan{y},y)$ and $F_2(r\pi+\arctan{y},y)$, then both $\beta$ and $\alpha=r\pi+\arctan{\beta}$ would be algebraic, which contradicts Claim \ref{claim:atleastonetrans} below.
\end{proof}
\begin{claim}\label{claim:atleastonetrans}
For any $y\in(0,1)$, at least one of $r\pi+\arctan{y}$ and $y$ is transcendental.
\end{claim}
\begin{proof}
Note that \[
\tan(r\pi+\arctan y)=\left\{
\begin{array}{ccl}\vspace{2mm}
y,& \text{ if } &r=k\in\mathbb{Z},\\
\frac{1+y}{1-y},& \text{ if } &r=k+\frac{1}{4}\in\mathbb{Z}+\frac{1}{4}.
\end{array}
\right.
\]
Suppose $y$ is algebraic, then $\tan(r\pi+\arctan y)$ is also algebraic in either case. Since neither $y$ nor $(1+y)/(1-y)$ is zero, $r\pi+\arctan y\neq0$. Moreover, it is transcendental by Theorem \ref{thm:BothAlgeNum}.
\end{proof}
\begin{remark}\label{remark:bisect}
Suppose $h(x_1,x_2)=\sum_{i=0}^{\deg(h,x_1)}c_i(x_2)x_1^i\in\mathbb{Q}[x_1,x_2]$. For any $y\in(0,1)\cap\mathbb{Q}$, $r\pi+\arctan y$ is transcendental. Thus $h(r\pi+\arctan y,y)=0$ iff $c_i(y)=0$ for all $i$. We can therefore decide whether $h(r\pi+\arctan y,y)$ is zero for given $y\in(0,1)\cap\mathbb{Q}$. If it is nonzero, we can further compute its sign by approximating it more accurately. More generally, if $\phi$ is a real function $($\emph{e.g.}, $\phi=\tan)$ such that $\phi(y)$ is transcendental for any nonzero rational $y$ in its domain, then we can decide whether $h(\phi(y),y)$ is zero, and compute its sign whenever it is not zero.
\end{remark}
\begin{comment}
\begin{lemma}\label{lemma:NoCommonRootPre}
Let $k\in\Z$ and $r\in\{k,k+\frac{1}{4}\}$.
For any $y\in(0,1)$, at least one of $r\pi+\arctan{y}$ and $y$ is an transcendental number.
\end{lemma}
\begin{proof}
If $r=k$, assume that $k\pi+\arctan{y}$ and $y$ are both algebraic numbers.
Note that $\tan(k\pi+\arctan{y})=y$. Then, by Theorem \ref{thm:BothAlgeNum}, we have $y=0$, which makes a contradiction.
If $r=k+\frac{1}{4}$, assume that $k\pi+\frac{\pi}{4}+\arctan{y}$ and $y$ are both algebraic numbers.
Let $x:=k\pi+\frac{\pi}{4}+\arctan{y}$.
Then, $\tan{x}=\frac{1+y}{1-y}$.
Let $z:=\tan{x}$.
Then, we have $(y-1)z+y+1=0$.
Since $y$ is an algebraic number, $z$ is also an algebraic number.
Since $x$ and $z$ are both algebraic numbers, by Theorem \ref{thm:BothAlgeNum}, we have $z=0$.
Then, $y=-1$, which makes a contradiction.
\end{proof}
\end{comment}
Recall that our goal is to isolate the real roots of $g(x,\tan{x})$ in $(r\pi,r\pi+\frac{\pi}{4})$, where $g$ is an irreducible polynomial in $\Q[x,y]$ with $\deg(g,y)\ge1$ and $r\in\{k,k+\frac{1}{4}\}~(k\in\Z)$.
The process of Algorithm \ref{alg:SubIBR} is as follows.
If $g=y+1$, then $\tan{x}+1$ has no roots on $(r\pi,r\pi+\frac{\pi}{4})$, and thus return $\emptyset$.
Otherwise, let $F_1:=g$ when $r=k$, and $F_1:=\varphi(g)$ when $r=k+\frac{1}{4}$.
It is not difficult to check that our goal is equivalent to solving $F_1(x+r\pi,\tan{x})$ on $(0,\frac{\pi}{4})$.
Let $y:=\tan{x}$.
We only need to solve $F_1(r\pi+\arctan{y},y)$ on $(0,1)$.
Recall Remark \ref{remark:PolyWeakFourierSeq}.
We can compute a weak Fourier sequence
\begin{align}\label{eq:ComputeFourierSeq}
w_1=F_1(r\pi+\arctan{y},y),\ldots,w_k=F_k(r\pi+\arctan{y},y)~{\rm on}~(0,1),
\end{align}
where $F_i\in\Q[x,y]$.
Thus, the frame of \cite[Algorithm 47]{strzebonski2012real} can be used.
However, there are two main differences between \cite[Algorithm 47]{strzebonski2012real} and Algorithm \ref{alg:SubIBR}.
\begin{comment}
First, we use Algorithm \ref{alg:WeakFourierSeq} to compute a weak Fourier sequence
\begin{align}\label{eq:ComputeFourierSeq}
w_1=G_1(r\pi+\arctan{y},y),\ldots,w_k=G_k(r\pi+\arctan{y},y),
\end{align}
where $G_i\in\Q[x,y]$, for $G_1(r\pi+\arctan{y},y)$ on $(0,1)$ (see Line \ref{algline:SubIBR-weakFseq}), instead of \cite[Algorithm 40]{strzebonski2012real}.
Although \cite[Algorithm 40]{strzebonski2012real} is more general and can also be used to compute a weak Fourier sequence for $G_1(r\pi+\arctan{y},y)$, Algorithm \ref{alg:WeakFourierSeq} is better since it is proposed particularly for our purpose.
\end{comment}
First, when $w_{\ell}$ has exactly one simple root $sr$ on $(c,d)$ (see Line \ref{algline:SubIBR-sr}),
we propose a new method to decide whether $sr$ is a root of $w_1$.
If $F_{\ell}=QF_1~(Q\in\Q[y])$ holds and $Q$ has no roots on $(c,d)$, by \eqref{eq:ComputeFourierSeq}, $sr$ is the only root of $w_1$ on $(c,d)$.
If $F_{\ell}=QF_1$ holds and $Q$ has roots on $(c,d)$, then it is clear that $w_1$ has no roots on $(c,d)$.
If $F_{\ell}=QF_1$ does not hold, then $sr$ is not a root of $w_1$.
This method is correct: By Lemma \ref{lemma:mappingirre}, $F_1$ is an irreducible polynomial.
If $F_1\mid F_{\ell}$, we have $\deg(F_1,x)\le\deg(F_{\ell},x)$.
Then, by Lemma \ref{lemma:pseudoproperty}, $\deg(F_1,x)=\deg(F_{\ell},x)$.
So, $Q$ is a polynomial in $\Q[y]$.
Otherwise, ${\gcd}(F_1,F_{\ell})=1$. Then, by Lemma \ref{lemma:NoCommonRoot}, $sr$ is not a root of $w_1$.
Second, if $sr$ is not a root of $w_1$, we find two rational numbers $u,v$ such that $c\leq u<sr<v\leq d$ and $w_1$ has no roots in $[u,v]$ by Algorithm \ref{alg:RationalEndpoint} (see Line \ref{algline:SubIBR-RationalEP} of Algorithm \ref{alg:SubIBR}), instead of by \cite[Algorithm 45]{strzebonski2012real}. The process of Algorithm \ref{alg:RationalEndpoint} is as follows: Let $n:=1$.
While $n\ge1$, we compute an upper limit polynomial $W_{\max,n}\in\Q[y]$ and a lower limit polynomial $W_{\min,n}\in\Q[y]$ for $w_1(y)$ by two steps.
Since $w_1=F_1(r\pi+\arctan{y},y)$, we first estimate $\arctan{y}$ through Taylor series, and $T_{\max}(n,w_1),T_{\min}(n,w_1)\in\Q[\pi][y]$ are computed by \cite[Def. 3.1]{chen2020automated}.
Then, we use a bounded interval with rational endpoints to estimate $\pi$ in $T_{\max}(n,w_1)$ and $T_{\min}(n,w_1)$, and the polynomials $W_{\max,n},W_{\min,n}\in\Q[y]$ are computed by \cite[Def. 3.1]{xia2006real}.
Let $q:=\frac{c+d}{2}$.
If $W_{\max,n}$ has no roots on $[c,d]$ and $W_{\max,n}(q)<0$, or if $W_{\min,n}$ has no roots on $[c,d]$ and $W_{\min,n}(q)>0$, then $w_1$ has no roots on $[c,d]$.
Otherwise, increase $n$ by $1$.
Since $w_{\ell}$ has exactly one simple root $sr$ on $(c,d)$, we use a bisection method based on Remark \ref{remark:bisect} to find the half open interval where $sr$ is.
Update $(c,d)$ as the half interval and go for the next loop step.
Remark that if the interval $(c,d)$ is small enough and the upper (lower) limit polynomial is close enough, Algorithm \ref{alg:RationalEndpoint} terminates by Lemma \ref{lemma:NoLimit}.
\begin{algorithm}[!t]
\scriptsize
\DontPrintSemicolon
\LinesNumbered
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{
$g(x,y)$, an irreducible polynomial in $\Q[x,y]$ with $\deg(g,y)\ge1$\\
$r$, $r\in\{k,k+\frac{1}{4}\}$ where $k\in\Z$
}
\Output{$ans$, a set of isolating intervals of real roots of $g(x,\tan{x})$ in $(r\pi,r\pi+\frac{\pi}{4})$
}
\caption{{\bf SubIBR}
}\label{alg:SubIBR}
\BlankLine
{\bf if} $g=y+1$, {\bf then} {\bf return} $\emptyset$\;
\eIf{$r=k$}
{$F_1\leftarrow g$}
{$F_1\leftarrow g(x,\frac{1+y}{1-y})\cdot(1-y)^{\deg(g,y)}$\;}
let $w_1=F_1(r\pi+\arctan{y},y),\ldots,w_k=F_k(r\pi+\arctan{y},y)$, where $F_i\in\Q[x,y]$, be the sequence in Remark \ref{remark:PolyWeakFourierSeq}\label{algline:SubIBR-weakFseq}\;
$ans\leftarrow\emptyset$\;
$stack\leftarrow\{(0,1,k)\}$\;
\While{$stack\ne\emptyset$}
{
$(c,d,m)\leftarrow First(stack),\;stack\leftarrow Rest(stack)$\;
$sgc\leftarrow\sgc_{1,m}(c^{+})-\sgc_{1,m}(d^{-})$\;
{\bf if} $sgc=0$, {\bf then} continue\;
\If{$sgc=1$}{$ans\leftarrow ans\cup\{(r\pi+\arctan{c},r\pi+\arctan{d})\}$\; continue\;}
$\ell\leftarrow\min\;\{i\mid 2\le i\le m,~\sgc_{i,m}(c^{+})-\sgc_{i,m}(d^{-})=1\}$\;
$sr\leftarrow\Root(w_{\ell},c,d)$\label{algline:SubIBR-sr}\;
By Theorem \ref{thm:WeakFourierSeq}, $w_{\ell}$ has exactly one simple root $sr$ in $(c,d)$.
We use the notation $\Root(w_{\ell},c,d)$ to represent the only simple root of $w_{\ell}$ in $(c,d)$.\;
\eIf{$F_{\ell}=QF_1$ where $Q\in\Q[y]$\label{algline:SubIBR-wheHold}}
{
\eIf{$Q$ has no real roots on $(c,d)$}
{$ans\leftarrow ans\cup\{(r\pi+\arctan{c},r\pi+\arctan{d})\}$}
{continue}
}
{
$u,v\leftarrow{\bf RationalEndpoint}(w_1,w_{\ell},sr)$\label{algline:SubIBR-RationalEP}\;
add $(c,u,\ell)$ and $(v,d,\ell)$ to $stack$\;
}
}
\Return{$ans$}
\end{algorithm}
\begin{algorithm}[htbp]
\scriptsize
\DontPrintSemicolon
\LinesNumbered
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{
$w_1(y)=F_1(r\pi+\arctan{y},y)$, where $F_1(x,y)\in\Q[x,y]$, $r\in\Q$\\
$w_{\ell}(y)=F_{\ell}(r\pi+\arctan{y},y)$, where $F_{\ell}(x,y)\in\Q[x,y]$\\
$sr=\Root(w_{\ell},c,d)$, $w_{\ell}$ has exactly one simple root $sr$ in $(c,d)$ and $w_{1}(sr)\ne0$
}
\Output{two rational numbers $u$ and $v$ $(c\leq u<sr<v\leq d)$ such that $w_{1}$ has no roots on $[u,v]$
}
\caption{{\bf RationalEndpoint}}\label{alg:RationalEndpoint}
\BlankLine
$n\leftarrow1$\;
\While{$n\ge1$}{
Compute an upper limit polynomial $W_{\max,n}\in\Q[y]$ and a lower limit polynomial $W_{\min,n}\in\Q[y]$ for $w_1(y)$\;
$q\leftarrow\frac{c+d}{2}$\;
\eIf{$(W_{\max,n}$ has no roots on $[c,d]$ and $W_{\max,n}(q)<0)$
{\bf or} \\$\quad\,(W_{\min,n}$ has no roots on $[c,d]$ and $W_{\min,n}(q)>0)$
}
{\Return{$c,d$}}
{
$n\leftarrow n+1$\;
\uIf{$w_{\ell}(q)=0$}
{$c\leftarrow \frac{63q+c}{64},\;d\leftarrow\frac{63q+d}{64}$}
\uElseIf{$\sign(w_{\ell}(c)w_{\ell}(q))=-1$}
{$d\leftarrow q$}
\Else{$c\leftarrow q$}
}
}
\end{algorithm}
\begin{comment}
\begin{algorithm}[!ht]
\DontPrintSemicolon
\LinesNumbered
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{
\begin{itemize}
\item $c,d$, two rational numbers
\item $w_i,\ldots,w_m$, a weak Fourier sequence on $(c,d)$
\item $dir\in\{left,right\}$
\end{itemize}
}
\Output{
$sgc$, if $dir=left$, $sgc$ equals $\sgc_{i,m}(c^{+})$; if $dir=right$, $sgc$ equals $\sgc_{i,m}(d^{-})$
}
\caption{{\bf SGC} \cite[Algorithm 44]{strzebonski2012real}}\label{alg:SGC}
\BlankLine
$s_m\leftarrow\sign(w_m(\frac{c+d}{2}))$\;
{\footnotesize \# Recall Definition \ref{def:WeakFourierSeq}, $w_m(y)\ne0$ for any $y\in(c,d)$. So, $\sign(w_m(y))$ is constant on $(c,d)$.}\;
\eIf{$dir=left$}
{$l_i,\ldots,l_{m-1}\leftarrow\sign(w_i(c^{+})),\ldots,\sign(w_{m-1}(c^{+}))$\;
}
{$l_i,\ldots,l_{m-1}\leftarrow\sign(w_i(d^{-})),\ldots,\sign(w_{m-1}(d^{-}))$}
\For{$j$ {\bf from} $m-1$ {\bf to} $1$}
{If $l_j\ne0$, set $s_j\leftarrow l_j$\;
If $l_j=0$ and $dir=left$, if $s_{j+1}>0$ set $s_j\leftarrow 1$ and otherwise $s_j\leftarrow -1$\;
If $l_j=0$ and $dir=right$, if $s_{j+1}>0$ set $s_j\leftarrow -1$ and otherwise $s_j\leftarrow 1$\;
}
\Return{the number of sign changes in the sequence $s_1,\ldots,s_m$}\;
\end{algorithm}
\end{comment}
\begin{comment}
\begin{proposition}
If $f(x,y)\in\Q[x,y]\setminus\{0\}$ and $f_1(x,\arctan(x)),\ldots,f_k(x,\arctan(x))$ is a weak Fourier sequence for $f(x,\arctan(x))$ computed by Algorithm \ref{alg:WeakFourierSeq},
then for any nonzero rational number $r$, $f(x,y+r\pi)\in\Q[\pi][x,y]\setminus\{0\}$ and $f_1(x,\arctan(x)+r\pi),\ldots,f_k(x,\arctan(x)+r\pi)$ is a weak Fourier sequence for $f(x,\arctan(x)+r\pi)$.
\end{proposition}
\begin{proof}
It is obvious that $f(x,y+r\pi)\in\Q[\pi][x,y]\setminus\{0\}$.
Note that
$${\trueder f}(x,\arctan(x)+r\pi)=f'(x,\arctan(x)+r\pi).$$
So, by the correctness proof of Theorem \ref{thm:alg:WeakFourierSeqTC}, $f_1(x,\arctan(x)+r\pi),\ldots,f_k(x,\arctan(x)+r\pi)$ is a weak Fourier sequence for $f(x,\arctan(x)+r\pi)$.
\end{proof}
\end{comment}
\section{Introduction}
Real root isolation has been considered as a symbolic way to compute the real roots of a univariate polynomial or function. Real root isolation for polynomials is of fundamental importance in computational real algebraic geometry (\emph{e.g.}, it is a routing sub-algorithm for cylindrical algebraic decomposition \cite{collins1975quantifier}), and there are many efficient algorithms and popular tools in computer algebra systems such as Maple and Mathematica to isolate the real roots of polynomials.
There are also some research on real root isolation for non-polynomial functions: In \cite{mitchell1990robust}, interval arithmetic has been used to isolate real roots of a large class of analytic functions. The algorithm there mainly considers deciding the exclusion or the inclusion of a root in an interval. Recent years, Strzebo{\'n}ski presented a real root isolation procedure for univariate \emph{exp–log–arctan functions} \cite{strzebonski2008real,strzebonski2012real}.
The procedure is based on the technique of weak Fourier sequence and semi-Fourier sequence. Meanwhile, Strzebo{\'n}ski studied real root isolation for the so-called tame elementary functions using the concept of false derivative \cite{strzebonski2009real}. The completeness of these algorithms of Strzebo{\'n}ski relies on Schanuel's conjecture.
In \cite{achatz2008deciding,mccallum2012deciding}, for developing a decision procedure for a certain class of
sentences of first order logic involving integral polynomials and
a certain specific analytic transcendental function trans$(x)$, Achatz, McCallum and Weispfenning provided a recursive real root isolation algorithm for a certain kind of
generalised integral polynomial in trans$(x)$: $f(x,\text{trans}(x))$ (where $f(x, y)$ is a given polynomial in $y$ whose coefficients are elements of the ring of fractions of $\mathbb{Z}[x]$ with respect to powers of a specific
integral polynomial $d(x)$), using \emph{pseudo-differentiation} and
Rolle’s theorem in the spirit of \cite{collins1976polynomial}, and also a classical result of
Lindemann \cite{shidlovskii2011transcendental}. The list of
transcendental functions to which their decision method directly
applies includes $\exp(x),\ln(x)$ and $\arctan(x)$. The decision procedure when trans$(x)=\exp(x)$ is implemented and the corresponding root isolation algorithm is believed to be more efficient than the one
provided in \cite{maignan1998solving}. Inspired by \cite{achatz2008deciding} and \cite{strzebonski2011cylindrical}, Xu \emph{et al.} studied a class of quantified exponential polynomial formulas extending polynomial ones \cite{xu2015quantifier}.
The isolation algorithm for an exponential polynomial used therein is modified from the one in \cite{achatz2008deciding}.
In \cite{huang2018positive}, Huang \emph{et al.} developed some algorithms for isolating the positive roots of a class of functions called poly-powers, while in \cite{wang2021symbolic}, Wang and Xu focused on isolating the real roots of a certain class of radical expressions.
A {\em mixed trigonometric-polynomial} (MTP) refers to a function of the form $f(x,\sin x, \cos x)$ with $f$ being a trivariate polynomial having rational coefficients. In this paper, we present an algorithm ``isolating" all the real roots (there are probably infinitely many of them) of such an MTP. While all inverse trigonometrical functions are exp-log-arctan functions, this is not true for the MTPs. In fact, since an exp-log-arctan function has finitely many real roots (\cite[Claim 4]{strzebonski2012real}), any MTP with infinitely many real roots is not an exp-log-arctan function (\emph{e.g.}, $\sin x$, $\cos x$, $2x\sin x+1$, \emph{etc.}). On the other hand, an MTP is usually not a tame elementary function since the variable varies over the whole $\R$. Hence, our result is not covered by the ones of Strzebo{\'n}ski. However, in Section \ref{sec:IBR}, while isolating the real roots of an MTP in a bounded interval, we take advantages of the weak Fourier sequences of a function introduced in \cite{strzebonski2008real,strzebonski2012real} (it can be also practicable to use the false derivative technique in \cite{strzebonski2009real} since any MTP restricted to a bounded interval is tame). Nevertheless, we do not need Schanuel's conjecture since we only consider trigonometric functions in one variable. Note that both the methods used in \cite{strzebonski2012real} and \cite{mccallum2012deciding} are applicable for isolating the real roots of the function of the form $f(\arctan x,x)$, with $f$ being a bivariate polynomial having rational coefficients. This is equivalent to isolating the real roots of the function $f(\hat{x},\tan \hat{x})$ with $\hat{x}\in(-\frac{\pi}{2},\frac{\pi}{2})$ (see, \emph{e.g.}, \cite[Section 6]{mccallum2012deciding}). The goal of the present paper is to isolate all the real roots of an MTP,
which is a non-trivial generalisation.
The topic of automatically proving inequalities involving trigonometric functions, which is closely related to the topic of the present paper, has also drawn more and more attention these years due to their applications to engineering problems and different branches of science \cite{chen2020automated}. Usually, the term ``MTP inequalities" refers to inequalities of the form $g(x)>0$ or $g(x)\geq0$ with $g$ being an MTP. A so-called ``Natural Approach" introduced in \cite{mortici2011natural} by Mortici, approximates $\sin x$ and $\cos x$ by their Taylor polynomials and reduce the problem of proving an MTP inequality to the problem of proving some polynomial (or rational function) inequalities. That approach leads to a series of subsequent studies \cite{chen2012sharp,bercu2016pade,bercu2017natural,makragic2017method,chenliu2016}, in which a great deal of inequalities involving trigonometric functions are solved automatically. Later in \cite{chen2020automated}, Chen and Liu provided a complete algorithm proving MTP inequalities and discussed systematically, for the first time, the termination of the algorithm. However, these algorithms only prove MTP inequalities over bounded intervals (\emph{e.g.}, $(0,\frac{\pi}{2})$ or $(0,\frac{\pi}{4}]$, \emph{etc.}). Till recently, Chen and Ge \cite{chen2022automated} tried to develop a procedure automatically proving MTP inequalities on the unbounded interval $(0,\infty)$. Unfortunately, we found some mistakes therein. Therefore, the problems of proving an MTP inequality, and of isolating real roots in an unbounded interval remain open so far.
In this paper, an algorithm ``isolating" all the real roots of an MTP is provided (Algorithm \ref{alg:complete-algorithm}). It automatically divides the real roots into two parts: one consists of finitely many ``bounded" roots in an interval $[\mu_-,\mu_+]$ with $\mu_-<0<\mu_+$, while the other consists of probably countably many ``periodic" roots in $\mathbb{R}\backslash[\mu_-,\mu_+]$. For the bounded roots, the algorithm returns some isolating intervals and the multiplicity of each of them. For those periodic roots greater than $\mu_+$, it returns finitely many mutually disjoint small intervals $I_i\subset[-\pi,\pi]$, integers $c_i>0$ and multisets of root multiplicity $\{m_{j,i}\}_{j=1}^{c_i}$ such that any periodic root $t>\mu_+$ is in the set $(\sqcup_i\cup_{k\in\mathbb{N}}(I_i+2k\pi))$ and any interval $I_i+2k\pi\subset(\mu_+,\infty)$ contains exactly $c_i$ periodic roots with multiplicities $m_{1,i},...,m_{c_i,i},$ respectively. For those periodic roots less than $\mu_-$, it returns some similar results. The algorithm is implemented with {\tt Maple2021}. A large number of examples (either from the literature or randomly generated) have been tested to show its effectiveness and efficiency.
In particular, this is the first algorithm to decide whether or not a given MTP has infinitely many real roots: it does iff the algorithm returns at least one periodic root. And of course, the algorithm is also able to decide whether or not an MTP has a real root (when there are real roots, we can compute the multiplicity of all of them). This allows us to prove or disprove an MTP inequality over $\mathbb{R}$. Our results also indicate, for the first time, that the ``distributions" of the roots of an MTP in the ``periods" $(-\pi,\pi]+2k\pi$, with $k>0$ (\emph{resp.}, $k<0$) sufficiently large, share a same pattern, which seems to be a basic law about the real roots of MTPs.
Besides, the method used to isolate the roots in $[\mu_-,\mu_+]$ is applicable to any other bounded interval as well. It takes advantages of the weak Fourier sequence technique and deals with the intervals period-by-period without scaling the coordinate (see, \emph{e.g.}, \cite[Algorithm 4.1, Step 2]{chen2022automated}) so to keep the length of the sequence short. An interesting problem is to compare the efficiency of the false derivative technique with the weak Fourier technique for isolating real roots of an MTP in a bounded interval, which is beyond the scope of this paper.
As a direct application, we discuss briefly at the end of the paper how one can modified our new approaches to prove or disprove an MTP inequality on an interval of the form $(a,b)$, $[a,b)$, $(a,b]$, $[a,b]$, with $a,b\in\Q\cup\{\pm\infty\}$.
The rest of the paper is organized as follows.
The next section introduces the problem and some basic results that will be used from time to time. In Section \ref{sec:farroots}, we introduce the algorithm ``isolating" periodic roots while in Section \ref{sec:IBR} we provide the procedure computing the isolating intervals for bounded roots. The main algorithm ``isolating" all real roots of an MTP is presented in Section \ref{sec:wholealgorithm}. And in Section \ref{sec:experiments}, experimental results are provided to indicate the effectiveness and efficiency of the algorithm. Finally, Section \ref{sec:conclusion} summarizes the paper and provides some remarks relating to the topic of proving MTP inequalities.
\section{The Problem}\label{sec:Preprocessing}
\subsection{Some basic results}
Denote by $\mathbb{N}, \Z, \Q, \R$ and $\C$ the sets of natural, integer, rational, real and complex numbers, respectively. For any subset $A\subset\mathbb{R}$, set $A_{>0}=\{a\in A\,|\,a>0\}$. The meanings of the notation $A_{\geq0}$, $A_{\leq0}$ and $A_{<0}$ are similar. For any $\mathbb{A}\in\{\Z, \Q, \R,\C\}$, $\mathbb{A}[x_1,\ldots,x_m]$ denotes the corresponding ring of polynomials in the variables $x_1,\ldots,x_m$ with coefficients in $\mathbb{A}$. For any $f\in\mathbb{A}[x_1,\ldots,x_m]$, $\deg(f,x_i)$, $\lc(f,x_i)$ and $\discrim(f,x_i)$ denote the degree, the leading coefficient and the discriminant of $f$ with respect to $x_i$, respectively. Moreover, $\tdeg(f,[x_{i_1},\ldots,x_{i_{t}}])$ denotes the total degree of the subsequence of variables $[x_{i_1},\ldots,x_{i_{t}}]$.
{\bf The problem} considered in this paper is to solve the real roots of the following equation:
\begin{equation}\label{eq:goalequation}
f(x,\sin{x},\cos{x})=0,~{\rm where}~f\in \Q[x,y,z].
\end{equation}
We provide some basic observations in the following.
\begin{lemma}\label{lemma:NoLimit}
For any nonzero $f\in\Q[x,y,z]$, the real root set of $f(x,\sin{x},\cos{x})$ has no accumulation point in $\mathbb{R}$.
\end{lemma}
\begin{proof}
The complex-valued function $h(z)=f(z,\sin{z},\cos{z})$ is analytic over $\C$. By definition, $h|_{\mathbb{R}}=f(x,\sin{x},\cos{x})$. If the set of the real roots of $f$ has an accumulation point in $\mathbb{R}$, then the set of the complex roots of $h$ has an accumulation point in $\mathbb{C}$ as well, which indicates $h\equiv0$ over $\mathbb{C}$. But $h$ can not be identically $0$ over $\mathbb{C}$, since $f$ is not over $\mathbb{R}$.
\end{proof}
\begin{lemma}\label{lemma:piRealRoot}
For any $q(x)\in\Q[x]$, ``$r\pi$ is a real root of $q(x)$ for any $r\in\Q\setminus\{0\}$" iff ``$q(x)$ is the zero polynomial in $\Q[x]$".
\end{lemma}
\begin{proof}
The ``if" part is obvious. The ``only if" part follows from the fact that $r\pi$ is transcendental for any $r\in\Q\setminus\{0\}$.
\end{proof}
\begin{theorem}\label{thm:BothAlgeNum}\cite[Chapter 2.3]{shafarevich1998number}
If $\alpha\in\C$ and $\alpha\ne0$, then there is at least one transcendental number in each pair $(\alpha,\sin\alpha)$, $(\alpha,\cos\alpha)$, $(\alpha,\tan\alpha)$,
$(\alpha,\cot\alpha)$.\qed
\end{theorem}
\begin{definition}\label{definition:recip}
For a nonzero $g\in\Q[x,y]$, we define its \emph{reciprocal polynomial} $\mathcal{R}(g)$ as $g(x,\frac{1}{y})y^{\deg(g,y)}$. Additionally, we set $\mathcal{R}(0)=0$.
\end{definition}
From the definition it is obvious that
\;\;1) $\mathcal{R}(g_1g_2)=\mathcal{R}(g_1)\mathcal{R}(g_2)$ for any $g_1,g_2\in\Q[x,y]$,
\;\;2) if $g=\sum_{i=0}^{\deg(g,y)}a_i(x)y^i$ with $a_0(x)\neq0$ (\emph{i.e.}, $y\nmid g$), then $\mathcal{R}(\mathcal{R}(g))=g$, and
\;\;3) $y$ does not divide $\mathcal{R}(g)$ for any nonzero $g$.
\noindent{}We then have some key properties of $\mathcal{R}$ as follows.
\begin{proposition}\label{proposition:RkeepIrr}
Suppose $g\in\Q[x,y]$ and $g\neq cy$ for any $c\in\mathbb{Q}$. Then ``$g$ is irreducible in $\Q[x,y]$" implies ``$\mathcal{R}(g)$ is irreducible in $\Q[x,y]$".
\end{proposition}
\begin{proof}
The proof is inductive on the total degree of $g$: If $\tdeg(g,[x,y])=0$, the conclusion is trivially true since $g$ is not irreducible. Suppose that $\tdeg(g,[x,y])=n>0$ and the implication holds for any polynomial ($\neq cy$) of total degree less that $n$. Assume $\mathcal{R}(g)$ is reducible and its factorization is $c_0h_1^{p_1}\cdots h_t^{p_t}$ with $h_i$ irreducible, $\sum_{j=1}^{t}p_j\geq2$ and $c_0\in\mathbb{Q}\backslash\{0\}$. Moreover, each $h_i\neq cy$ for any nonzero $c\in\Q$ because $y\nmid\mathcal{R}(g)$. Since $g$ is irreducible and $g\neq cy$ for any $c\in\Q$, $y\nmid g$. Thus $g=\mathcal{R}(\mathcal{R}(g))=c_0\mathcal{R}(h_1)^{p_1}\cdots \mathcal{R}(h_t)^{p_t}$. By the inductive hypothesis, each $\mathcal{R}(h_i)$ is irreducible. Since $\sum_{j=1}^{t}p_j\geq2$, $g$ is reducible, which is a contradiction.
\end{proof}
For any polynomial $g\in\Q[x,y]$ of positive total degree, $g$ is said to be \emph{square-free}, if the factorization of $g$ is $cg_{1}\cdots g_{t}$ with $c\in\Q\setminus\{0\}$, $t\in\Z_{\ge1}$, each $g_i$ irreducible and $g_i$ and $g_j$ co-prime for any $1\le i\ne j\le t$. Any constant polynomial is not square-free in our definition.
\begin{proposition}\label{proposition:recipKeepSquarefree}
Suppose $g\in\Q[x,y]$ and $g\neq cy$ for any $c\in\mathbb{Q}$. Then ``$g$ is square-free" implies ``$\mathcal{R}(g)$ is square-free".
\end{proposition}
\begin{proof}
Factorize $g$ as $c_0y^\ell h_1\cdots h_t$, with $c_0\in\Q$ nonzero, $\ell=0\text{ or }1$ (corresponding to the cases $y\nmid g$ and $y\,|\,g$ respectively), $t\geq1$ (since $g\neq cy$ and $g\notin\Q$) and each $h_i$ irreducible.
First, $\mathcal{R}(g)=c_0\mathcal{R}(h_1)\cdots\mathcal{R}(h_t)$ since $\mathcal{R}(y^\ell)=1$.
Second, for each $h_i$, $h_i\neq cy$ for any $c\in\Q$ because $y\nmid h_i$.
So, by Proposition \ref{proposition:RkeepIrr}, $\mathcal{R}(h_i)$ is irreducible.
Finally, $\mathcal{R}(h_i)$ and $\mathcal{R}(h_j)$ are co-prime for $1\leq i\neq j\leq t$. Suppose they are not, then $\mathcal{R}(h_i)=c'\mathcal{R}(h_j)$ for a $c'\in\Q\backslash\{0\}$. Since $y\nmid h_i$ and $y\nmid h_j$, $h_i=\mathcal{R}(\mathcal{R}(h_i))=c'\mathcal{R}(\mathcal{R}(h_j))=c'h_j$, contradicting the fact that $h_i$ and $h_j$ are co-prime in the factorization of $g$.
\end{proof}
\subsection{Decomposing the problem into sub-problems}
Because
\[
\left\{
\begin{array}{rl}\vspace{2mm}
\sin x&= \frac{2\tan{\frac{x}{2}}}{1+\tan^2(\frac{x}{2})}\\
\cos x&= \frac{1-\tan^2{\frac{x}{2}}}{1+\tan^2({\frac{x}{2}})}
\end{array}\right.,\;\;x\neq 2k\pi+\pi~(k\in \Z),
\]
Eq. \eqref{eq:goalequation} is equivalent to
\begin{subnumcases}
{}
f(x,\sin{x},\cos{x})=0, ~~~~~~~~~~~~~~~~~~~~~x=2k\pi+\pi\label{meq:1},\\
f(x,\frac{2\tan{\frac{x}{2}}}{1+\tan^2{(\frac{x}{2})}},\frac{1-\tan^2{(\frac{x}{2})}}{1+\tan^2{(\frac{x}{2}})})=0,~ \hfill x\in (2k\pi-\pi,2k\pi+\pi), \label{meq:2_}
\end{subnumcases}
where Eq. \eqref{meq:1} is equivalent to $f(x,0,-1)=0,~x=(2k+1)\pi$.
By Lemma \ref{lemma:piRealRoot},
$(2k+1)\pi$ is a root of $f(x,0,-1)$ iff $f(x,0,-1)$ is the zero polynomial in $\Q[x]$.
Through cancelling the denominators, we define a polynomial as follows:
\begin{equation}\label{equation:gTurnedToTan}
g(x,t)=f(2x,\frac{2t}{1+t^2},\frac{1-t^2}{1+t^2})\cdot(1+t^2)^{\tdeg(f,[y,z])}\in\mathbb{Q}[x,t].
\end{equation}
Since $1+\tan^2{(\frac{x}{2})}>0$, solving the real roots of Eq. (\ref{meq:2_}) is then equivalent to solving the real roots of the equation
\begin{equation}
g(x,\tan{x})=0,~x\in(k\pi-\frac{\pi}{2},k\pi+\frac{\pi}{2}).\label{meq:2}
\end{equation}
\begin{example}\label{example:gtan}
Consider the following MTP
$$f(x,\sin{x},\cos{x})=x\cos^4{x}-2\sin{x}\cos^3{x}+\frac{4}{5}\sin^3{x}.$$
By Eq. \eqref{equation:gTurnedToTan}, we have
\begin{align*}
g(x,t)=10xt^8+20t^7-40xt^6-28t^5+60xt^4+92t^3-40xt^2-20t+10x.
\end{align*}
\end{example}
In Eq. \eqref{meq:2}, we have to deal with infinitely many intervals $(k\pi-\frac{\pi}{2},k\pi+\frac{\pi}{2})$ while $k$ varies in $\Z$. Fortunately, Proposition \ref{prop:firstbound} and Corollary \ref{cor:main_case} below, together with
Theorem \ref{thm:location-of-the-roots} in the next section will set us free from doing that.
\begin{proposition}\label{prop:firstbound}
Set nonzero polynomial $g\in\Q[x,y]$, $\delta>0$ and $U\subset\mathbb{R}$ to be a bounded set. If $|\lc(g,x)|>\delta$ for any $y\in U$, then there is an integer $N>0$ such that $x>N$ implies $\forall y\in U,\; g(x,y)\neq0$.
\end{proposition}
\begin{proof}
Suppose $\deg(g,x)=n$ and $a_i(y)\in\mathbb{Q}[y]\;(i=0,\ldots,n)$ are the coefficients of $g$ \emph{w.r.t.} $x$. Now that $U$ is bounded, there is $B_i>0$ such that $|a_i(y)|<B_i$ for any $y\in U$. For sufficiently large $x$, we have $\delta |x|^n>\sum_{i=0}^{n-1}B_i|x|^i$. Therefore, for any $y\in U$,
\[|a_n(y)x^n|>\delta|x|^n>\sum\limits_{i=0}^{n-1}B_i|x|^i\geq|\sum\limits_{i=0}^{n-1}a_i(y)x^i|.\]
This indicates that $g(x,y)\neq0$ for any $y\in U$.
\end{proof}
By Proposition \ref{prop:firstbound}, the value of the tangent function at every ``periodic'' real root of $g(x,\tan{x})$ appears near a real root of $\lc(g,x)$.
In order to locate these ``periodic'' real roots, we introduce the following definition.
\begin{definition}\label{def:potential-periodic-interval}
For any $g\in\Q[x,y]$ with $\deg(g,y)\ge 1$,
let $I$ be a finite list of open intervals $[(a_0,b_0),\ldots,(a_{s+1},b_{s+1})]$,
where $s\in\Z_{\ge0}$, $a_0=-\infty$, $b_{s+1}=+\infty$, $b_0<0$, $a_{s+1}>0$, $a_j\in \Q~(1\le j\le s+1)$ and $b_j\in \Q~(0\le j\le s)$, such that
\begin{enumerate}[label=$($\roman*$)$]
\item for each $j~(1\le j\le s)$, $a_j< b_j$ and $b_j \le a_{j+1}$,
\item for each real root of $\lc(g,x)$, there exists an interval $(a_j,b_j)~(1\le j\le s)$ containing it $($there is no root of $\lc(g,x)$ in $(a_0,b_0)$ and $(a_{s+1},b_{s+1}))$,
\item \label{item:potential-periodic-interval-3} each interval $(a_j,b_j)~(1\le j\le s)$ has exactly one real root of $\lc(g,x)$.
\end{enumerate}
We call $I$ a \emph{potential periodic interval set} of $g$.
Generally, if we replace ``exactly" in \ref{item:potential-periodic-interval-3} by ``at most", we call $I$ a \emph{general potential periodic interval set} of $g$.
\end{definition}
\begin{example}\label{example:potential-periodic-interval-set}
Consider the polynomial in Example \ref{example:gtan}:
\begin{align*
g(x,y)=10xy^8+20y^7-40xy^6-28y^5+60xy^4+92y^3-40xy^2-20y+10x.
\end{align*}
We have $\lc(g,x)=10y^8-40y^6+60y^4-40y^2+10$.
It has two real roots $-1$ and $1$ and the isolating intervals of them are $(-\frac{215}{128}, -\frac{19}{32})$ and $(\frac{19}{32}, \frac{215}{128})$, respectively.
Thus,
{\footnotesize
$$\left[\left(-\infty, -\frac{63}{16}\right), \left(-\frac{215}{128}, -\frac{19}{32}\right), \left(\frac{19}{32}, \frac{215}{128}\right), \left(\frac{63}{16}, +\infty\right)\right]$$}is a potential periodic interval set of $g$.
\end{example}
\begin{corollary}\label{cor:main_case}
Let $g\in\Q[x,y]$ be with $\deg(g,y)\ge 1$, $[(a_0,b_0),\ldots,(a_{s+1},b_{s+1})]$ be a general potential periodic interval set of $g$, and
\begin{align}
U:=\R\setminus\bigcup_{j=0}^{s+1}(a_i,b_i).
\end{align}
Then, there exists a positive integer $N$ such that
$$
\forall x\in \R ~\forall y\in U~(x>N \Rightarrow g(x,y)\neq 0).
$$
\end{corollary}
\begin{proof}
Note that $|\lc(g,x)|$ has a lower bound on $y\in U$ and $U$ is bounded.
Then, the conclusion follows from Proposition \ref{prop:firstbound}.
\end{proof}
For the polynomial $g\in\Q[x,y]$ in Eq. \eqref{meq:2},
let $U,a_j,b_j$ be the same as in Corollary \ref{cor:main_case}.
Take $N_+, N_-\in\Z_{\ge1}$ such that $\forall x\in \R ~\forall y\in U~(x>N_+ \Rightarrow g(x,y)\neq 0)$ and
$\forall x\in \R ~\forall y\in U~(x<-N_- \Rightarrow g(x,y)\neq 0)$.
There exist $k_+\in\Z_{\ge0}$ and $k_-\in\Z_{\le0}$ such that
$k_+\pi-\frac{\pi}{2}<N_+<k_+\pi+\frac{\pi}{2}$ and
$k_-\pi-\frac{\pi}{2}<-N_-<k_-\pi+\frac{\pi}{2}$, respectively.
Thus, Eq. \eqref{meq:2} can be transformed equivalently into the following three cases:
\begin{enumerate}[label=$($\ref{meq:2}\alph*$)$]
\item \label{item:MainThreeCase1}
\begin{align*}
g(x,\tan x)=0,~~~~&x\in(k\pi-\frac{\pi}{2},k\pi+\frac{\pi}{2})
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
\end{align*}
where $k_{-}\leq k \leq k_{+}\;(k\in\Z)$;
\item \label{item:MainThreeCase2}
\begin{align*}
g(x,\tan{x})=0,~~~~&x\in(k\pi-\frac{\pi}{2},k\pi+\arctan(b_0))\\
&x\in(k\pi+\arctan(a_i),k\pi+\arctan(b_i)),~{i=1,\ldots,s}\\
&x\in(k\pi+\arctan(a_{s+1}),k\pi+\frac{\pi}{2})
\end{align*}
where $k>k_{+}\;(k\in\Z)$;
\item \label{item:MainThreeCase3}
\begin{align*}
g(x,\tan{x})=0,~~~~&x\in(k\pi-\frac{\pi}{2},k\pi+\arctan(b_0))\\
&x\in(k\pi+\arctan(a_i),k\pi+\arctan(b_i)),~{i=1,\ldots,s}\\
&x\in(k\pi+\arctan(a_{s+1}),k\pi+\frac{\pi}{2})
\end{align*}
where $k<k_{-}\;(k\in\Z)$.
\end{enumerate}
A possible way to solve case \ref{item:MainThreeCase1} for $k_-\le k\le k_+$ is to go back to the original MTP $f(x,\sin x,\cos x)$ and consider the roots in $(2k_-\pi-\pi,2k_+\pi+\pi)$. Replace $x$ in $f(x,\sin x,\cos x)$ by $\ell x$ with $\ell\in\Z_{>0}$ large enough such that the roots in $(2k_-\pi-\pi,2k_+\pi+\pi)$ ``shrink" to the roots of $f(\ell x,\sin \ell x,\cos \ell x)$ in $(-\pi/2,\pi/2)$. Then unify the sine and cosine functions with the tangent function in $f(\ell x,\sin \ell x,\cos \ell x)$ to obtain some $\hat{g}(x,\tan x)$. Finally, deal with the roots of $\hat{g}(-x, -\tan x)$ and $\hat{g}(x, \tan x)$ in the interval $(0,\pi/4)$.
The reason why we do not use this method in Section \ref{sec:IBR} is that $\ell$ may be large and the ``degree" of $\tan x$ in $\hat{g}(x, \tan x)$ can be high. Instead, we will deal with case \ref{item:MainThreeCase1} period-by-period to avoid that problem.
The last case can be reduced to the second last one via replacing $g(x,\tan x)$ by $g(-x,-\tan x)$. In the next section, we focus on the asymptotic properties of the real roots corresponding to case \ref{item:MainThreeCase2}.
\section{Asymptotic Property of Roots}\label{sec:farroots}
In this section, we shall deal with case \ref{item:MainThreeCase2}, namely isolate the ``periodic'' roots of $g(x,\tan{x})$ where $g$ is a nonzero polynomial in $\Q[x,y]$.
\subsection{Detecting and locating periodic roots}
There are countably many ``periodic'' roots of $g(x,\tan{x})$, so they can not be located by the methods of traditional real root isolation.
However, these ``periodic'' roots can be divided into finitely many bunches via a general potential periodic interval set (recall Def. \ref{def:potential-periodic-interval}).
The roots in each bunch appear periodically, that is why we call them ``periodic'' roots.
In this subsection, we propose an algorithm to isolate these ``periodic'' roots by means of a general potential periodic interval set.
First, we introduce the definition of \emph{analytic delineable} and an important result about it from the work of McCallum \cite{mccallum1998improved}.
\begin{definition}[\cite{mccallum1998improved}]
Let $\vX$ denote the $(r-1)$-tuple $(x_1,\cdots,x_{r-1})$. An $r$-variate polynomial $f(\vX,x_r)$ over the reals is said to be \emph{analytic delineable} on a sub-manifold $S$ $($usually connected$)$ of $\R^{r-1}$, if
\begin{enumerate}[label=$($\arabic*$)$]
\item the portion of the real variety of $f$ that lies in the cylinder $S\times \R$ over $S$ consists of the union of the graphs of some $k\geq 0$ analytic functions $\theta_1<\cdots <\theta_k$ from $S$ into $\R$, and
\item there exist positive integers $m_1$, $\cdots$, $m_k$ such that for every $a\in S$, the multiplicity of the root $\theta_i(a)$ of $f(a,x_r)$ $($considered as a univariate polynomial in $x_r$$)$ is $m_i$.
\end{enumerate}
Also, an $r$-variate polynomial $f(\vX,x_r)$ over $\R$ is said to be \emph{degree-invariant} on a subset $S$ of $\R^{r-1}$ if the total degree of $f(p,x_r)$ $($as a univariate polynomial in $x_r$$)$ is the same for every point $p$ of S.
\end{definition}
\begin{proposition}[{\cite[Theorem 2]{mccallum1998improved}}]\label{prop:McCallum}
Let $r\geq 2$ be an integer and $f(\vX,x_r)$ be a polynomial in $\R[\vX,x_r]$ of positive total degree.
Suppose that $D(\vX):=\discrim(f,x_r)$ is a nonzero polynomial.
Let $S$ be a connected submanifold of $\R^{r-1}$ such that $f$ is degree-invariant and does not vanish identically on $S$, and $D(\vX)>0$ or $D(\vX)<0$ over $S$. Then, $f$ is analytic delineable on $S$.\qed
\end{proposition}
We restate a well-known result (which follows from Proposition \ref{prop:McCallum}) in Corollary \ref{corollary:bound-M} and redefine the concept of ``root function'' in Definition \ref{def:root-function-start-point}.
\begin{corollary}\label{corollary:bound-M}
Let $g$ be a square-free polynomial in $\Q[x,y]$ with $\deg(g,y)\ge 1$,
and $M$ be the maximal real root of the polynomial
$\lc(g,y)\cdot\discrim(g,y)$.
Then,
\begin{enumerate}
\item $g$ is analytic delineable on $(M, +\infty)$, and
\item there exist some analytic functions $\theta_1(x)<\cdots<\theta_r(x)\allowbreak\;(r\ge0)$ defined on $(M, +\infty)$ such that for any $x_0\in(M, +\infty)$, $g(x_0,\theta_i(x_0))=0$ and $\theta_1(x_0),\ldots,\theta_r(x_0)$ are all real roots of $g(x_0,y)$. \qed
\end{enumerate}
\end{corollary}
\begin{definition}\label{def:root-function-start-point}
Keep the notation in Corollary \ref{corollary:bound-M}.
We call those analytic functions $\theta_1(x),\ldots,\theta_r(x)$ the
\emph{terminal-root-functions} of $g$ and $M$ the \emph{start-point} of the terminal-root-functions of $g$.
\end{definition}
\begin{example}\label{example:terminal-root-function}
Consider the polynomial $g(x,y)$ in Example \ref{example:potential-periodic-interval-set}.
We have $\lc(g,y)=-10x$ and
$\discrim(g,y)=-703687441776640000\cdot(163840000x^8 + 261529600x^6 + 83912256x^4 - 13333725x^2 - 10857025)$.
After isolating all real roots of them, we see that the maximal real root $M$ of $\discrim(g,y)$, which is in $(0,1)$, is the start-point of the terminal-root-functions of $g$.
And there are two terminal-root-functions of $g$, which are defined on $(M,+\infty)$ $($see Fig. \ref{fig:graph-Terminal-Root-Function}$)$.
\input{fig/root_function_fig.tex}
\end{example}
In Theorem \ref{thm:location-of-the-roots} below, we compute a bound that guarantees well-defined terminal-root-functions $($as in Corollary \ref{corollary:bound-M}$)$, which are essential to locating the unbounded roots.
\begin{theorem}\label{thm:location-of-the-roots}
Let $g$ be a square-free polynomial in $\Q[x,y]$ with $\deg(g,y)\ge 1$,
$[(a_0,b_0),\ldots,(a_{s+1},b_{s+1})]$ be a general potential periodic interval set of $g$,
$k'$ be a natural number such that $M'=(k'+1/2)\pi$ is an upper bound of the start-point $M$ of the terminal-root-functions of $g$ and all real roots of the polynomial
$$\prod_{j=0}^{s}g(x,b_j)g(x,a_{j+1}),$$
and $\theta_1(x),\ldots,\theta_r(x)$ be the terminal-root-functions of $g$.
Then:
\begin{enumerate}[label=$($\arabic*$)$]
\item\label{item:thm:location-of-the-roots1} For each $i~(1\le i \le r)$, there exists
$j_i~(0\le j_i \le s+1)$ such that $\theta_i(x)\in (a_{j_i},b_{j_i})$ whenever $x\in (M',+\infty)$.
\item\label{item:thm:location-of-the-roots2}
For each real root $\xi$ of $g(x,\tan x)$ greater than $M'$, there exist $k>k'~(k\in \N)$ and $i$ $(1\le i \le r)$ such that
$$\xi\in\left(\arctan(a_{j_i})+k\pi, \arctan(b_{j_i})+k\pi\right).$$
\item\label{item:thm:location-of-the-roots3}
For each $i~(1\le i \le r)$ and any $k>k'~(k\in \N)$, there exists $$\xi\in\left(\arctan(a_{j_i})+\allowbreak k\pi, \arctan(b_{j_i})+k\pi\right)$$ such that $\tan {\xi}=\theta_i(\xi)$.
\end{enumerate}
\end{theorem}
\begin{example}\label{example:thm:location-of-the-roots}
Consider the polynomial $g(x,y)$ in Example \ref{example:potential-periodic-interval-set}.
In Example \ref{example:potential-periodic-interval-set}, we compute a potential periodic interval set of $g(x,y)$:
{\footnotesize
$$\left[\left(-\infty, -\frac{63}{16}\right), \left(-\frac{215}{128}, -\frac{19}{32}\right),\allowbreak \left(\frac{19}{32}, \frac{215}{128}\right), \left(\frac{63}{16}, +\infty\right)\right].$$}And the isolating intervals of the real roots of the polynomial
{\footnotesize
$$g(x,-\frac{63}{16})\cdot g(x,-\frac{215}{128})\cdot g(x,-\frac{19}{32})\cdot g(x,\frac{19}{32})\cdot g(x,\frac{215}{128})\cdot g(x,\frac{63}{16})$$}are $(-8, -7)$, $(-4, -3)$, $(-1, 0)$, $(0, 1)$, $(3, 4)$ and $(7, 8)$.
In Example \ref{example:terminal-root-function}, we see that the start-point $M$ of the terminal-root-functions of $g$ is less than $1$.
Thus, we take $k'=3$ such that $(k'+1/2)\pi$ is an upper bound of $M$ and the real roots computed above.
Note that the two terminal-root-functions of $g$ are bounded by $(-\frac{215}{128},-\frac{19}{32})$ when $x\in((k'+1/2)\pi,+\infty)$ $($see Fig. \ref{fig:thm:location-of-the-roots}$)$.
\input{fig/e37}
So, for every $k > 3~(k \in \Z)$, there are at least two real roots of $g(x,\tan{x})$ in $(k\pi-\arctan{\frac{215}{128}},k\pi-\arctan{\frac{19}{32}})$.
\end{example}
\begin{remark}\label{remark:decideexistence}
For each interval $(a_j,b_j)~(0\le j\le s+1)$, it may happen that none of the terminal-root-functions $\theta_i(x)$ is bounded by $(a_j,b_j)$ when $x>M'$. It is also possible that there can be one or more than one terminal-root-functions $\theta_{i_1}(x),\ldots,\theta_{i_t}(x)$, with $t\in\Z_{\ge1}$, that are bounded by $(a_j,b_j)$ when $x>M'$. By \ref{item:thm:location-of-the-roots2} and \ref{item:thm:location-of-the-roots3} in Theorem \ref{thm:location-of-the-roots}, there is no real root in the set $(M',\infty)\cap\tan^{-1}(a_j,b_j)$ in the former case, while in the latter case, there are at least $t$ real roots in each interval $(\arctan a_j+k\pi,\arctan b_j+k\pi)$ contained in $(M',\infty)$
$($in Line \ref{Line:alg1-detect-implicit-function} of Algorithm \ref{alg:AlgFarZero}, we distinguish these two cases$)$.
Denote these roots by $x^{(1)}_k<\cdots<x^{(t'_k)}_k$ for some $t'_k\geq t$. Since Theorem \ref{thm:location-of-the-roots} is valid for isolating intervals $(a_j,b_j)$ of arbitrarily small ``arc-tangent-length" $($\emph{i.e.}, $\arctan b_j-\arctan a_j)$ and any $M'$ large enough, we have
\[\lim_{k\rightarrow\infty}(x^{(t'_k)}_k-x^{(1)}_k)=0.\] This indicates that it is impossible to find real numbers $c_1,c_2\in(-\pi/2,\pi/2)$ such that they isolate, for instance, $x^{(1)}_k, x^{(2)}_k$ and $x^{(3)}_k$ $($if $t>2)$ uniformly. That is, there are no such $c_1,c_2$ satisfying \[x^{(1)}_k<c_1+k\pi<x^{(2)}_k<c_2+k\pi<x^{(3)}_k\] for all sufficiently large integer $k$. This is why we only ``isolate" them with the interval $(\arctan a_j,\arctan b_j)+k\pi$ in the sense that every $x^{(\iota)}_k$ is contained in it. Nevertheless, in the next subsection we will prove that $t'_k=t$ for sufficiently large $k$, which means we can compute the exact number $t$ of real roots in the interval $(\arctan a_j,\arctan b_j)+k\pi$ that is sufficiently far from zero.
\end{remark}
\noindent{}\emph{\bf Proof of Theorem \ref{thm:location-of-the-roots}.}
\ref{item:thm:location-of-the-roots1}
Note that $g(x,a_j)~(1\le j \le s+1)$ and $g(x,b_j)~(0\le j \le s)$ have no roots in $(M',+\infty)$.
Then, for every $i~(1\le i\le r)$ and for any $x\in(M',+\infty)$, $\theta_i(x)\not\in\{b_0,a_1,b_1,\ldots,b_s,a_{s+1}\}$.
By Proposition \ref{prop:McCallum}, $\theta_i(x)$ is continuous.
So, there exists $j_i~(0\le j_i\le s+1)$ such that $\theta_i((M',+\infty))\subseteq (a_{j_i},b_{j_i})$ or
there exists $j_i~(0\le j_i\le s)$ such that $\theta_i((M',+\infty))\subseteq(b_{j_i},a_{j_i+1})$.
By Corollary \ref{cor:main_case}, for sufficiently large $x$, $\theta_i(x)\in \cup_{j=0}^{s+1}(a_j,b_j)$. Therefore, $\theta_i((M',+\infty))\subseteq (a_{j_i},b_{j_i})$ for some $0\leq j_i\leq s+1$.
\ref{item:thm:location-of-the-roots2}
If $\xi$ is a real root of $g(x,\tan{x})$ such that $\xi>M'$, then there exists $i~(1\le i\le r)$ such that
$\theta_{i}(\xi)=\tan{\xi}$. By \ref{item:thm:location-of-the-roots1}, $\theta_i((M',+\infty))\subseteq (a_{j_i},b_{j_i})$. So, $\tan{\xi}\in (a_{j_i},b_{j_i})$. Thus, there exists $k~(k>k')$ such that $\xi\in (\arctan(a_{j_i})+k\pi,\arctan(b_{j_i})+k\pi)$.
\ref{item:thm:location-of-the-roots3}
By \ref{item:thm:location-of-the-roots1}, for each $i~(1\le i\le r)$, $\theta_i((M',+\infty))\subseteq (a_{j_i},b_{j_i})$. So, for any $k~(k>k')$, we have
\begin{align*}
\theta_i(\arctan(a_{j_i})+k\pi)>a_{j_i}=&\tan(\arctan(a_{j_i})+k\pi),\\
\theta_i(\arctan(b_{j_i})+k\pi)<b_{j_i}=&\tan(\arctan(b_{j_i})+k\pi).
\end{align*}
Hence there exists $\xi\in(\arctan(a_{j_i})+k\pi, \arctan(b_{j_i})+k\pi) $ such that $\tan \xi-\theta_i(\xi)=0$.
$\hfill{\square}$\par
\begin{comment}\textcolor{blue}{(The symbols ``$r$" and ``$s$" in Theorem \ref{thm:location-of-the-roots} and Algorithm \ref{alg:AlgFarZero} are ``swapped". And sometimes ``$i$" and ``$j$" are used casually, \emph{e.g.}, in the first two sentences of \S \ref{subsec:exactnumber}. It would be better if the meanings of these symbols were fixed.)}
\end{comment}
\vspace{3mm}
\begin{algorithm}[!ht]
\scriptsize
\DontPrintSemicolon
\LinesNumbered
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{
$g$, a square-free polynomial in $\Q[x,y]$ such that $\deg(g,y)\geq 1$;\\
$I=[(a_0,b_0),\ldots,(a_{s+1},b_{s+1})]$, a general potential periodic interval set of $g$.
}
\Output{
$k'$, a natural number;\\
$ret$, a finite set of pairs $((x_j^{-},x_j^{+}),c_j)$ such that for any $k>k'~(k\in\N)$, there are \emph{AT LEAST} $c_i$ roots for $g(x,\tan x)$ in the interval $(x_j^{-}+k\pi,x_j^{+}+k\pi)$.
}
\caption{\bf IsolatingFarZero}\label{alg:AlgFarZero}
\BlankLine
Find $k'\in\N$ such that
$(k'+\frac{1}{2})\pi$ is an upper bound for all real roots of $\lc(g,y)\cdot\discrim(g,y)\cdot\prod_{j=0}^{s}g(x,b_{j})g(x,a_{j+1})$. \label{Line:alg1-computation-of-M}\;
Choose an integer $x_0$ such that $x_0>(k'+\frac{1}{2})\pi$\label{Line:alg1-sample-point}.\;
$c_j\leftarrow0$ for $j=0,\ldots,s+1$\label{Line:alg1-ini-cj}\;
\For{each real root $r$ of $g(x_0,y)$}{\label{Line:alg1-for-1}
\For{$j$ from $0$ to $s+1$}
{
\If{$r\in (a_j,b_j)$}{
$c_{j}\leftarrow c_{j}+1$\label{Line:alg1-count-c}\;
\textbf{break}\label{Line:alg1-for-1-end}\;
}
}
}
$ret\leftarrow \emptyset$\label{Line:alg1-organizing-output}\;
\For{$j$ from $0$ to $s+1$\label{Line:alg1-detect-implicit-function-1}}{
\If{$c_j>0$\label{Line:alg1-detect-implicit-function}}{
$x_{j}^{-}\leftarrow\arctan a_j,~x_{j}^{+}\leftarrow\arctan b_j$\;
$ret \leftarrow ret \cup \{((x_{j}^{-},x_{j}^{+}),c_j)\}$\label{Line:alg1-output-end}\;
}
}
\Return $k'$, $ret$\;
\end{algorithm}
\begin{definition}
Let $g$, $k'$ be the same as in Theorem \ref{thm:location-of-the-roots}, and $\theta(x)$ be a terminal-root-function of $g(x,y)$.
We call the set
$\{x_0\in ((k'+1/2)\pi,+\infty) \mid \theta(x_0)=\tan{x_0} \}$
\emph{a bunch of periodic roots} of $g(x,\tan{x})$.
\end{definition}
Summarizing all above, we propose Algorithm \ref{alg:AlgFarZero} to isolate the ``periodic" roots of $g(x,\tan{x})$.
Given a square-free polynomial $g\in\Q[x,y]$ with $\deg(g,y)\geq1$, and a general potential periodic interval set $I$ of $g$, Algorithm \ref{alg:AlgFarZero} computes a natural number $k'$ and a set of pairs $((x_{i}^{-},x_{i}^{+}),c_i)$, such that for any $k>k'~(k\in\N)$, there are at least $c_i$ real roots of $g(x,\tan x)$ in the interval $(x_{i}^{-}+k\pi,x_{i}^{+}+k\pi)$.
\begin{theorem}\label{thm:AlgFarZeroisCorrect}
Algorithm \ref{alg:AlgFarZero} is correct.
\end{theorem}
\begin{proof}
The process of Algorithm \ref{alg:AlgFarZero} is based on Theorem \ref{thm:location-of-the-roots}.
And it is clear that Algorithm \ref{alg:AlgFarZero} is correct by the proof of Theorem \ref{thm:location-of-the-roots}.
\end{proof}
\subsection{Ensuring exact quantity of periodic roots in an ``isolating" interval}\label{subsec:exactnumber}
In the subsection, we would strengthen Theorem \ref{thm:location-of-the-roots}: for all sufficiently large $k$, $\theta_i(x)$ and $\tan x$ intersects exactly once in the interval $(\arctan a_{j_i}+k\pi,\arctan b_{j_i}+k\pi)$. And a lower bound for such $k$ can be efficiently determined.
Depending on whether the derivative $\theta_i'(x)\ (x>M')$ is strictly bounded by $1$ or not,
the discussion is divided into two cases.
For the bounded case, the key observation is: the derivative of $\tan x$ is $1/\cos^2 x$, which is always $\geq 1$. Therefore, since $\theta_i'(x)$ is smaller than $1$, $y=\tan x$ can not intersect with $\theta_i(x)$ twice in a period $(k\pi -\pi/2,k\pi+\pi/2)$, where $k\in\N$ and $(k+1/2)\pi>M'$.
For the unbounded case, we consider $1/\theta_i(x)$ instead of $\theta_i(x)$.
We first establish a lemma.
\begin{lemma}\label{lemma:existence-of-delta}
Let $g\in\Q[x,y]$ be square-free with $\deg(g,y)\geq1$.
There exists a finite set $S\subset\Q$ such that for any $\delta_0\in\Q\setminus S$,
$\res(\frac{\partial g}{\partial x}+\delta_0 \frac{\partial g}{\partial y},g,y)$
is not the zero polynomial in $\Q[x]$.
\end{lemma}
\begin{proof}
Regard $\delta$ as a variable. We claim that
$\res(\frac{\partial g}{\partial x}+\delta \frac{\partial g}{\partial y},g,y)$ is not the zero polynomial in $\Q[\delta,x]$.
Note that $\frac{\partial g}{\partial y}$ is not the zero polynomial, since $\deg(g,y)\ge1$.
Then, $q:=\gcd(\frac{\partial g}{\partial x}+\delta \frac{\partial g}{\partial y},~g)$ is a nonzero polynomial in $\Q[x,y]$.
We also note that $q\mid\left(\frac{\partial g}{\partial x}+\delta \frac{\partial g}{\partial y}\right)$ implies $q\mid\frac{\partial g}{\partial y}$. Now that $g$ is square-free, $q\mid g$ and $q\mid\frac{\partial g}{\partial y}$ imply $\deg(q,y)=0$. Thus, $\res(\frac{\partial g}{\partial x}+\delta \frac{\partial g}{\partial y},g,y)$ is not the zero polynomial in $\Q[\delta,x]$.
Hence we can let $S$ be the set of the common roots in $\Q$ of the coefficients of $\res(\frac{\partial g}{\partial x}+\delta \frac{\partial g}{\partial y},g,y)$ with respect to $x$. The proof is then completed.
\end{proof}
\begin{lemma}\label{thm:derivative-bound}
Let $g\in\Q[x,y]$ be a square-free polynomial with $\deg(g,y)\geq1$,
$\delta_0$ a rational number such that $\res(\frac{\partial g}{\partial x}+\delta_0 \frac{\partial g}{\partial y},g,y)$ is not the zero polynomial,
$\theta_1(x),\ldots,\theta_r(x)$ the terminal-root-functions of $g$, and $\hat{k}\in\N$ a natural number such that $\hat{M}=(\hat{k}+1/2)\pi$ is an upper bound of the start-point of the terminal-root-functions of $g$ and all real roots of $\res(\frac{\partial g}{\partial x}+\delta_0 \frac{\partial g}{\partial y},g,y)$.
Then, the derivative of each $\theta_i(x)$ satisfies either one of the following properties:
\begin{enumerate}[label=$($\arabic*$)$]
\item\label{item:thm:derivative-bound1} $\forall x>\hat{M}:$ $\theta_i'(x)<\delta_0$;
\item\label{item:thm:derivative-bound2} $\forall x>\hat{M}:$ $\theta_i'(x)>\delta_0$.
\end{enumerate}
We call $\hat{k}$ a \emph{$\delta_0$-bound-point} of $g$.
In particular, if $\delta_0<1$ and property \ref{item:thm:derivative-bound1} holds for some $\theta_i(x)$, then for any $k>\hat{k}~(k\in\N)$,
there is a unique $\xi\in (k\pi -\pi/2, k\pi+\pi/2)$ such that $\theta_i(\xi)=\tan{\xi}$.
\end{lemma}
\begin{proof}
For each $\theta_i'(x)$, we prove that one of the two cases holds.
Note that $\hat{M}$ is greater than the start-point of the terminal-root-functions of $g$ (recall Definition \ref{def:root-function-start-point}).
So, when $x\in(\hat{M},+\infty)$, $\lc(g,y)(x)\neq0$, $\discrim(g,y)(x)\ne0$ and $g(x,\theta_i(x))=0$.
Therefore, $\frac{\partial g}{\partial y}(x,\theta_i(x))\ne0$ and $\theta_i'(x)=-\frac{\partial g/\partial x}{\partial g/\partial y}(x,\theta_i(x))$ by the implicit function theorem.
By the definition of $\hat{M}$, we have
$\lc(g,y)(x)\neq0$ and $\res(\frac{\partial g}{\partial x}+\delta_0 \frac{\partial g}{\partial y},g,y)(x)\ne0$ for $x\in(\hat{M},+\infty)$.
Then, since $g(x,\theta_i(x))=0$, it is clearly that
$(\frac{\partial g}{\partial x}+\delta_0\frac{\partial g}{\partial y})(x,\theta_i(x))\ne 0$.
So, we have $\theta_i'(x)\ne\delta_0$.
Therefore, by the continuity of $\theta'_{i}(x)$, exactly one case holds.
As for the last assertion, the existence is by the intermediate value theorem applied to $\tan x-\theta_i(x)$ on the interval $(k\pi -\pi/2, k\pi+\pi/2)$. The uniqueness follows from the fact that $\tan x-\theta_i(x)$ is strictly increasing on that interval since $1/\cos^2x-\theta'_i(x) >1-\delta_0>0$.
\end{proof}
Keep the notation in Lemma \ref{thm:derivative-bound}.
By the ``In particular'' part of Lemma \ref{thm:derivative-bound},
when $\theta_i'(x)<\delta_0$ on the interval $(\hat{M},\infty)$ for some selected $\delta_0<1$,
we can guarantee that there exists exactly $1$ periodic root (induced by $\theta_i$) of $g(x,\tan{x})$ in $(k\pi-\pi/2,k\pi+\pi/2)$ with $k$ sufficiently large.
In order to deal with the case with $\theta_i'(x)>\delta_0$, we consider $1/\theta_i(x)$ instead of $\theta_i(x)$.
In fact, $1/\theta_i(x)$ is also a terminal-root-function of the reciprocal polynomial
$h:=\mathcal{R}(g)$ (see Def. \ref{definition:recip}). If $y\nmid g$, we have $\deg(h,y)=\deg(g,y)\geq1$ by the definition. Moreover, since $g$ is square-free and clearly $h\neq cy$ for any $c\in\Q$, $h$ is also square-free (recall Prop. \ref{proposition:recipKeepSquarefree}).
Finally, we obtain the following theorem, which strengthens Theorem \ref{thm:location-of-the-roots}.
\begin{theorem}\label{thm:Uniqueness}
Let $g\in\Q[x,y]$ be square-free such that $\deg(g,y)\geq1$ and $y\nmid g$. Set
$h=\mathcal{R}(g)$ and take
$\delta_1\in\Q ~(0<\delta_1<1)$, $\delta_2\in\Q ~(-1<\delta_2<0)$ such that neither $\res(\frac{\partial g}{\partial x}+\delta_1 \frac{\partial g}{\partial y},g,y)$ nor $\res(\frac{\partial h}{\partial x}+\delta_2 \frac{\partial h}{\partial y},h,y)$ is the zero polynomial.
Let $k'$ be the same as in Theorem \ref{thm:location-of-the-roots},
$k''$ a $\delta_1$-bound-point of $g$ and
$k'''$ a $\delta_2$-bound-point of $h$ guaranteed by Lemma \ref{lemma:existence-of-delta}.
Set $k^{*}:=\max\{k',k'',k'''\}$ and let $\theta_1(x),\ldots,\theta_r(x)$ be the terminal-root-functions of $g$.
Then, for any $k>k^{*}~(k\in\Z)$,
there is a unique $\xi\in (k\pi -\pi/2,k\pi +\pi/2)$ such that $\theta_i(\xi)=\tan \xi$.
\end{theorem}
\begin{proof}
Since $y\nmid g$, $\deg(g,y)\ge 1$ and $g$ is square-free, we see that $\deg(h,y)\geq1$ and $h$ is square-free.
Thus, $k'''$ is well-defined.
Set $\mu:=(k^{*}+1/2)\pi$.
By Lemma \ref{thm:derivative-bound}, suppose $\theta_1(x),\ldots,\theta_t(x)$ are in Case \ref{item:thm:derivative-bound1} (\emph{i.e.}, $\theta_i'(x)<\delta_1<1$ for $x\in (\mu,+\infty))$ and $\theta_{t+1}(x),\ldots,\theta_r(x)$ are in Case \ref{item:thm:derivative-bound2} (\emph{i.e.}, $\theta_i'(x)>\delta_1$ for $x\in (\mu,+\infty))$ with $0\le t\le r$.
Then, by the ``In particular'' part in Lemma \ref{thm:derivative-bound}, for each $\theta_i~(1\le i\le t)$ in Case \ref{item:thm:derivative-bound1}, we already have the conclusion.
In the following, we deal with those $\theta_i(x)$ with $t+1\leq i\leq r$.
First, we prove that for each $\theta_i(x)~(t+1\le i\le r)$, $\theta_i(x)>0$ on $(\mu,+\infty)$: Since $\theta_i'(x)$ is greater than a positive number $\delta_1$ by Lemma \ref{thm:derivative-bound}, $\theta_i(x)\rightarrow+\infty$ as $x\rightarrow+\infty$. By property \ref{item:thm:location-of-the-roots1} of Theorem \ref{thm:location-of-the-roots}, the image of $\theta_i$ on $(\mu,+\infty)$ is a subset of $(a_{s+1},+\infty)$ with $a_{s+1}>0$ therein (recall Definition \ref{def:potential-periodic-interval}).
Second, we prove the existence of $\xi$ for each $\theta_i(x)~(t+1\le i\le r)$.
Let $k$ be any integer greater than $k^{*}$.
Applying the intermediate value theorem to $\tan{x}-\theta_i(x)$ on $(k\pi-\pi/2,k\pi+\pi/2)$, we see that there exists $\xi\in(k\pi-\pi/2,k\pi+\pi/2)$ such that $\tan{\xi}=\theta_i(\xi)$.
Finally, we prove the uniqueness of $\xi$ for each $\theta_i(x)~(t+1\le i\le r)$: Suppose $\tan\xi=\theta_i(\xi)$ for some $\xi\in(k\pi-\pi/2,k\pi+\pi/2)$. Then, $\xi\in(k\pi,k\pi+\pi/2)$ since $\theta_i(x)>0$ for all $x\in(\mu,+\infty)$.
Note that $1/\theta_i(x)$ is well defined on $(\mu,+\infty)$, $h(x,1/\theta_i(x))=0$ whenever $x\in (\mu,+\infty)$, and $\mu$ is greater than the start-point of the terminal-root-functions of $h$.
So, $1/\theta_i(x)$ is one of the terminal-root-functions of $h$ when $x\in (\mu,+\infty)$.
By Lemma \ref{thm:derivative-bound}, we have
$(1/\theta_i(x))'<\delta_2$ or $(1/\theta_i(x))'>\delta_2$ for $x\in(\mu,+\infty)$ with some $-1<\delta_2<0$.
Since $\lim\limits_{x\rightarrow +\infty}\theta_i(x)=+\infty$ and $\lim\limits_{x\rightarrow +\infty}1/\theta_i(x)=0$, we have $(1/\theta_i(x))'>\delta_2$.
Because $-1/\sin^2 x-(1/\theta_i(x))'<-1-\delta_2<0$,
the function $\cot x-1/\theta_i(x)$ is strictly decreasing on the interval $(k\pi,k\pi+\pi/2)$.
Thus, $\xi$ is unique.
\end{proof}
\begin{example}
Consider the polynomial $g(x,y)$ in Example \ref{example:potential-periodic-interval-set}.
In Example \ref{example:thm:location-of-the-roots}, we have taken $k'=3$.
It remains to compute $k''$ and $k'''$ in Theorem \ref{thm:Uniqueness}.
We take
$\delta_1=\frac{999}{1000}$ and $k''=3$, which is a
$\delta_1$-bound-point of $g$.
Note that $h=\mathcal{R}(g)=10xy^8-20y^7-40xy^6+92y^5+60xy^4-28y^3-40xy^2+20y+10x.$
We take
$\delta_2=-\frac{999}{1000}$ and $k'''=0$, which is a $\delta_2$-bound-point of $h$.
Thus, we have $k^{*}=\max\{k',k'',k'''\}=3$.
By Example \ref{example:thm:location-of-the-roots} and Theorem \ref{thm:Uniqueness}, we see that for every $k > 3~(k \in \Z)$, there are exactly two real roots of $g(x,\tan{x})$ in the interval $(k\pi-\arctan{\frac{215}{128}},k\pi-\arctan{\frac{19}{32}})$.
\end{example}
We propose Algorithm \ref{alg:AlgDerBound} to compute
a better lower bound, which guarantees the uniqueness, for the ``periodic'' roots of $g(x,\tan{x})$.
Given a square-free polynomial $g\in\Q[x,y]$ such that $\deg(g,y)\geq1$ and $y\nmid g$, together with a general potential periodic interval set $I$,
Algorithm \ref{alg:AlgDerBound} computes a natural number $k^{*}$ and a set of pairs $((x_{i}^{-},x_{i}^{+}),c_i)$ such that for any $k>k^{*}~(k\in\N)$, there are exactly $c_i$ real roots of $g(x,\tan x)$ in the interval $(x_{i}^{-}+k\pi,x_{i}^{+}+k\pi)$.
The process of Algorithm \ref{alg:AlgDerBound} is as follows.
First, by Algorithm \ref{alg:AlgFarZero}, we compute $k'\in\N$
and a set of isolating intervals $ret$ such that for any $k>k'~(k\in\N)$ and any pair $((x_i^{-},x_{i}^{+}),c_i)$ in $ret$, there are at least $c_i$ real roots of $g(x,\tan x)$ in $(x_i^{-}+k\pi,x_i^{+}+k\pi)$.
Then, we compute $k''\in\N$ and $k'''\in\N$ by Lemma \ref{thm:derivative-bound} in Lines \ref{line:alg2-k''-st}--\ref{line:alg2-k''-ed} and Lines \ref{line:alg2-k'''-st}--\ref{line:alg2-k'''-ed}, respectively.
Finally, the algorithm returns the natural number $k^{*}:=\max\{k',k'',k'''\}$ and the set $ret$.
The correctness of Algorithm \ref{alg:AlgDerBound} is guaranteed by Theorem \ref{thm:Uniqueness} and the termination of it is obvious.
\begin{algorithm}[ht]\label{exactroot}
\scriptsize
\DontPrintSemicolon
\LinesNumbered
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{
$g$, a square-free polynomial in $\Q[x,y]$ such that $\deg(g,y)\geq 1$ and $y\nmid g$;\\
$I=[(a_0,b_0),\ldots,(a_{s+1},b_{s+1})]$, a general potential periodic interval set of $g$.
}
\Output{
$k^{*}$, a natural number;\\
$ret$, a finite set of pairs $((x_i^{-},x_i^{+}),c_i)$ such that
for any $k>k^{*}~(k\in\N)$, there are \emph{EXACTLY} $c_i$ roots for $g(x,\tan x)$ in the interval $(x_i^{-}+k\pi,x_i^{+}+k\pi)$.
}
\caption{\bf CompleteIsolatingFarZero}\label{alg:AlgDerBound}
\BlankLine
$k',\;ret\leftarrow{\bf IsolatingFarZero}(g,I)$\label{line:alg2-k'}\;
$d_1\leftarrow \mathtt{res}(\frac{\partial g}{\partial x}+\delta \frac{\partial g}{\partial y},g,y)$ \#\emph{In this step, $\delta$ is a variable and $d_1\in\Q[\delta,x]$.}\label{line:alg2-k''-st}\;
Choose $\delta_1 \in(0,1)\cap \mathbb{Q}$ such that $d_1(\delta_1,x)$ is not the zero polynomial in $\Q[x]$.\;
Find $k''\in\N$ such that
$(k''+\frac{1}{2})\pi$ is an upper bound for all real roots of the polynomial $\textcolor{white}{hhhhhhhhhhhh}d_1(\delta_1,x)\cdot\lc(g,y)\cdot\allowbreak\discrim(g,y)$.\label{line:alg2-k''-ed}\;
$h\leftarrow g(x,\frac{1}{y})\cdot y^{\deg(g,y)}$\label{line:alg2-k'''-st}\;
$d_2\leftarrow \mathtt{res}(\frac{\partial h}{\partial x}+\delta \frac{\partial h}{\partial y},h,y)$\;
Choose $\delta_2 \in(-1,0)\cap \mathbb{Q}$ such that $d_2(\delta_2,x)$ is not the zero polynomial in $\Q[x]$.\;
Find $k'''\in\N$ such that
$(k'''+\frac{1}{2})\pi$ is an upper bound for all real roots of the polynomial $\textcolor{white}{hhhhhhhhhhhh}d_2(\delta_2,x)\cdot\lc(h,y)\cdot\allowbreak\discrim(h,y)$.\label{line:alg2-k'''-ed}\;
$k^{*}\leftarrow \max\{k',k'',k'''\}$\;
\Return $k^{*}$, $ret$
\end{algorithm}
\input{chap/Boundedsolve}
\section{The Whole Isolation Algorithm: Taking Root Multiplicity into Account}\label{sec:wholealgorithm}
In this section, we shall combine the results in previous sections to give an algorithm to isolate all real roots of an MTP. Furthermore, the multiplicity of each root is also computed by the algorithm. To do this, we shall first introduce a proposition on root multiplicity.
\begin{proposition}\label{prop:multiplicity-one}
Let $g\in \mathbb{Q}[x,y]\backslash \mathbb{Q}$ be irreducible. Then except $x=0$, all real roots of $g(x,\tan x)$ in its domain are simple roots.
\end{proposition}
\begin{proof}
If $g\in\Q[x]\cup\Q[y]$, the conclusion is trivially true. When $g$ is bivariate the conclusion follows from \cite[Theorem 2.1]{chen2020automated}. Although the theorem only claims that any root of $g(x,\tan x)$ in $(0,\frac{\pi}{4}]$ is simple, its proof actually shows that any root in $(\mathbb{R}\backslash\{0\})\backslash(\mathbb{Z}+\frac{1}{2})\pi$ of $g(x,\tan x)$ is simple, which is just what we want.
\end{proof}
The following observation is similar to Lemma \ref{lemma:NoCommonRoot}. The proof given below is according to the proof of \cite[Theorem 7]{mccallum2012deciding}.
\begin{proposition}\label{proposition:nocommonroot}
Suppose that $p,q\in\mathbb{Q}[x,y]$ are co-prime, then there is no common root of $p(x,\tan x)$ and $q(x,\tan x)$ except $x=0$.
\end{proposition}
\begin{proof}
Replacing those ``$p$, $q$" in the proof of \cite[Theorem 7]{mccallum2012deciding} by $p$ and $q$ in this proposition, those ``$f^*(x)$, $g^*(x)$" therein by $p(x,\tan x)$ and $q(x,\tan x)$, and the function ``\text{trans}" therein by the tangent function, we see that if $\alpha\in(\mathbb{R}\backslash\{0\})\backslash(\mathbb{Z}+\frac{1}{2})\pi$ is a common root of $p(x,\tan x)$ and $q(x,\tan x)$, then both $\alpha$ and $\tan \alpha$ would be algebraic, which contradicts Theorem \ref{thm:BothAlgeNum}.
\end{proof}
Then the following theorem is straightforward.
\begin{theorem}\label{thm:counting-multiplicity}
Let $g\in\mathbb{Q}[x,y]\backslash\mathbb{Q}$, and the factorization of $g$ in $\mathbb{Q}[x,y]$ is
$g=c_0\prod_{i=1}^{n} g_i^{p_i}$,
where $g_i$ are distinct irreducible factors and $c_0\in\mathbb{Q}$. If $x_0\in\mathbb{R}\backslash\{0\}$ is a root of $g(x,\tan x)$, then there is a unique $g_i$ such that $g_i(x_0,\tan x_0)=0$ and the multiplicity of $x_0$ is $p_i$.\hfill\qed
\end{theorem}
By the discussions above, Algorithms \ref{alg:AlgDerBound} and \ref{alg:BoundRSI} can be extended to compute the multiplicity of each root: Factorize $g$ into $c_0\prod_{i=1}^{n} g_i^{p_i}$ and call Algorithms \ref{alg:AlgDerBound} and \ref{alg:BoundRSI} for each irreducible factor $g_i$, then the multiplicity is just the exponent of $g_i$ in $g$ except at $x=0$.
\begin{remark}\label{remark:thesameforcot}
Replacing the tangent function in Propositions \ref{prop:multiplicity-one}--\ref{proposition:nocommonroot} and Theorem \ref{thm:counting-multiplicity} by the cotangent function, the corresponding conclusions also hold. The proofs are all similar.
\end{remark}
The next lemma detects the possible roots of $f(x,\sin x,\cos x)$ at $k\pi~(k\in \mathbb{Z})$, of which the function $g(\frac{x}{2},\tan \frac{x}{2})$ does not take care.
\begin{lemma}\label{lemma:multiplicityofkpi}
It can be effectively decided whether $f(x,\sin x,\cos x)$ is identically zero for $f\in\mathbb{Q}[x,y,z]$. Moreover, the root multiplicity of $f(x,\sin x,\cos x)$ at $0$, $2k\pi$ $(k\in \mathbb{Z}\backslash\{0\})$, $(2k+1)\pi$ $(k\in \mathbb{Z})$ can be effectively computed.
\end{lemma}
\begin{proof}
Rewrite $f(x,\sin x,\cos x)$ as $g(\frac{x}{2},\tan\frac{x}{2})$ with $g\in\Q[x,t]$ in Eq. (\ref{equation:gTurnedToTan}), then $f(x,\sin x,\cos x)$ is identically zero iff $g(\frac{x}{2},\tan \frac{x}{2})$ is. Lemma 2.3 in \cite{chen2020automated} then shows $f(x,\sin x,\cos x)$ is identically zero iff $g$ is a zero polynomial.
If $f(x,\sin x,\cos x)$ is not identically zero, then the root multiplicity of $x=0$ is finite. Otherwise the analytic function $f(z,\sin z,\cos z)$ with $z$ varying in $\C$ would be identically zero, and so would be $f(x,\sin x,\cos x)$. By computing $\frac{\mathrm{d}^{n} }{\mathrm{d} x^{n}}f(x,\sin x,\cos x)|_{x=0}$ for $n=0,1,2,...,$ until $\frac{\mathrm{d}^{n_0} }{\mathrm{d} x^{n_0}}f(x,\sin x,\cos x)|_{x=0}\neq0$ for some $n_0$. Then $n_0$ is the multiplicity of $x=0$.
For $x=2k\pi$ with $k\in\Z\backslash\{0\}$, $f(x,\sin x,\cos x)|_{x=2k\pi}=0$ iff $g(x,\tan x)|_{x=k\pi}$\\$=0$ and the root multiplicity of $2k\pi$ \emph{w.r.t.} $f(x,\sin x,\cos x)$ equals the one of $k\pi$ \emph{w.r.t.} $g(x,\tan x)$. Moreover, we claim $g(k\pi,\tan k\pi)=0$ iff $t$ is a factor of $g$. The ``if" part is clear. Suppose $g(k\pi,\tan k\pi)=0$, then $0=h(k\pi,\tan k\pi)=h(k\pi,0)$ for some irreducible factor $h=\sum_{j=0}^{\deg(h,x)}a_j(t)x^j$ of $g$. Since $k\pi$ is transcendental, $a_j(0)=0$ for any $j$. Hence $t$ divides $h$ and $h=rt$ (with an $r\in\Q$), which proves the claim. Then, by Theorem \ref{thm:counting-multiplicity}, the multiplicity of $2k\pi$ is the exponent of $y$ in the factorization of $g$.
Similarly, we can rewrite $f(x,\sin x,\cos x)$ as $\hat{g}(\frac{x}{2},\cot \frac{x}{2})$, with
\[\hat{g}(x,t)=f(2x,\frac{2t}{1+t^2},\frac{t^2-1}{1+t^2})(1+t^2)^{\text{deg}(f,[y,z])}\in\Q[x,t].\]
Then, using Remark \ref{remark:thesameforcot} and the discussion similar to the last paragraph, one observes that the root multiplicity of $(2k+1)\pi$ equals the exponent of $t$ in the factorization of $\hat{g}$.\end{proof}
It is time to give the main algorithm (Algorithm \ref{alg:CompleteAlgorithm}) for isolating all real roots of $f(x,\sin x,\cos x)$:
\begin{enumerate}
\item Decide whether the MTP $f(x, \sin x,\cos x)$ is identically zero (Lines \ref{algline:whetherg0}--\ref{algline:error}).
\item Compute the root multiplicity of the points $x=k\pi$ with $k\in\Z$ using the method in the proof of Lemma \ref{lemma:multiplicityofkpi} and compute a polynomial $g\in\Q[x,y]$ according to the input $f$ such that $y\nmid$ $g$ (Lines \ref{algline:m1}--\ref{algline:m3}).
\item Compute the isolating intervals of $\lc(g,x)$ together with $(a_0,b_0)$, $(a_{s+1},b_{s+1})$, forming a potential periodic interval set $I$ and factorize $g$ (Lines \ref{algline:isolate}--\ref{algline:factorize}).
\item Call Algorithm \ref{alg:AlgDerBound} with each irreducible factor $g_i$ with $\deg(g_i,y)\geq1$ together with $I$ to compute some $k^*_1$, $k^*_2$ and isolate the real roots of $g_i(x,\tan x)$ outside $(-k^*_2\pi-\pi/2,k^*_1\pi+\pi/2)$ (Lines \ref{algline:deggy>0part}--\ref{algline:renewk-}).
\item Isolate nonzero roots of the factors $g_i$ in $\Q[x]$ (Lines \ref{algline:isolateAlge}--\ref{algline:remove00}). Note that these correspond to the only nonzero algebraic roots of $f(x,\sin x,\cos x)$. We call $[0,0]$ and the intervals obtained in the above steps the ``algebraic" ones and others the ``transcendental" ones: the former represent 0 or algebraic roots and the latter represent transcendental roots. An algebraic interval may intersect a transcendental one, but it won't cause any trouble since we know they represent different roots.
\item Unify those $k^*_{1\cdot 2}$ of each factor $g_i$ (with $\deg(g_i,y)\geq1$) by some $k_+$ and $k_-$, which characterize the bound of the ``bounded" roots (Lines \ref{algline:renewk+}, \ref{algline:renewk-} and \ref{algline:unifyk+}--\ref{algline:unifyk-}).
\item For each factor $g_i$ such that $\deg(g_i,y)\geq1$, isolate the real roots of $g_i(x,\tan x)$ in $(k_-\pi-\pi/2,k_+\pi+\pi/2)$ by Algorithm \ref{alg:BoundRSI} (Lines \ref{algline:k-tok+}--\ref{algline:refineboundedroots})
\item Return all the results.
\end{enumerate}
In addition, for any real numbers $\alpha,\beta$, we define the interval $[[\alpha,\beta]]$ as
\begin{equation*}
[[\alpha,\beta]]:=\begin{cases}
(\alpha,\beta), &\text{if}~\alpha<\beta,\\
[\alpha,\beta], &\text{if}~\alpha=\beta.
\end{cases}
\end{equation*}
The notation is used in the pseudocode of Algorithm \ref{alg:CompleteAlgorithm}.
\input{alg/alg7.tex}
\begin{comment}
\section{Effectively bounding the roots at ``infinity''}
In this paper, the ring of \emph{trigonometric-polynomials} refers to the following ring of functions on $\mathbb{R}$:
\[\qscx=\big\{\sum\limits_{\lambda\in\Lambda}c_\lambda x^{\lambda(3)}(\sin x)^{\lambda(1)}(\cos x)^{\lambda(2)}\,\big|\,c_\lambda\in\mathbb{Q}, \Lambda\subset\mathbb{N}^3\text{ is finite}\big\},\]
and each function in that ring is called \emph{a trigonometric-polynomial}. Usually, for a non-zero function $f\in\qscx$, we write it in the form
\[f=\sum\limits_{i=0}^na_ix^i\]
with each $a_i$ in the following ring of functions on $\mathbb{R}$:
\[\qsc=\big\{\sum\limits_{\lambda\in\Lambda}c_\lambda(\sin x)^{\lambda(1)}(\cos x)^{\lambda(2)}\,\big|\,c_\lambda\in\mathbb{Q}, \Lambda\subset\mathbb{N}^2\text{ is finite}\big\}\]
and $a_n$ not a zero-function on $\mathbb{R}$ (we call $a_n$ \emph{the leading coefficient} of $f$). Apparently, there are positive numbers $B_i$, $i=0,1,\ldots,n$, such that \begin{equation}\label{coebound}
|a_i|\leq B_i.
\end{equation}
The following lemma is also straightforward:
\begin{lemma}\label{lemma:nolimit}
For any $f\in\qscxe$ that is not identically $0,$ the set of real roots of $f$ $($if there are any$)$ has no limit point in $\mathbb{R}$.
\end{lemma}
\begin{proof}
Suppose that $g(y_1,y_2,x)\in\mathbb{Q}[y_1,y_2,x]$ is a polynomial such that $f(x)=g(\sin x,\cos x,x)$ for all $x\in\mathbb{R}$. Then the complex-valued function $h(z)=g(\sin z,\cos z,z)$ over $\mathbb{C}$ is entire. By definition, $h|_{\mathbb{R}}=f$. If the set of the real roots of $f$ has a limit point in $\mathbb{R}$, then the set of the complex roots of $h$ has a limit point in $\mathbb{C}$ as well, which means $h\equiv0$ over $\mathbb{C}$. But $h$ cannot be identically $0$ over $\mathbb{C}$ since $f$ is not over $\mathbb{R}$.
\end{proof}
\begin{corollary}\label{corollary:countableroots}
Any trigonometric-polynomial $f\in\qscxe$ that is not identically zero has at most countably many real roots $($counted without multiplicity$)$.
\end{corollary}
\begin{proof}
By Lemma \ref{lemma:nolimit}, for each real root $\xi$, there is a positive number $\epsilon_{\xi}$ so that the interval $(\xi-\epsilon_{\xi},\xi+\epsilon_{\xi})$ contains no other real roots of $f$ than $\xi$. Thus, for any two different real roots $\xi$ and $\xi'$ of $f$, we have $|\xi-\xi'|>\max\{\epsilon_{\xi},\epsilon_{\xi'}\}\geq\frac{\epsilon_{\xi}+\epsilon_{\xi'}}{2}$. This implies that $(\xi-\frac{\epsilon_{\xi}}{2},\xi+\frac{\epsilon_{\xi}}{2})\cap(\xi'-\frac{\epsilon_{\xi'}}{2},\xi'+\frac{\epsilon_{\xi'}}{2})=\emptyset$. Hence there is an injection from the set $\mathcal{O}=\{(\xi-\frac{\epsilon_{\xi}}{2},\xi+\frac{\epsilon_{\xi}}{2})\,|\,\xi\text{ is a real root of }f\}$ to $\mathbb{Q}$: for each interval in $\mathcal{O}$, one chooses a rational number in it as its image. So, $\mathcal{O}$ is either finite or countable, so is the set of the real roots of $f$.
\end{proof}
\begin{proposition}
For an $f\in\qscxe$ that is not identically zero$,$ any real root of it has finite multiplicity.
\end{proposition}
\begin{proof}
Set $g\in\mathbb{Q}[y_1,y_2,x]$ so that $g(\sin x,\cos x,x)=f(x)$ for all $x\in\mathbb{R}$ and $h(z)=g(\sin z,\cos z,z)$ to be an entire function over $\mathbb{C}$. Then $h|_\mathbb{R}=f$ and it is straightforward to prove by induction that $\frac{\mathrm{d}^kh}{\mathrm{d}z^k}|_\mathbb{R}=\frac{\mathrm{d}^kf}{\mathrm{d}x^k}$ for any nature number $k\in\mathbb{N}$. Assume that there is a real root $\xi\in\mathbb{R}$ of $f$ so that $\frac{\mathrm{d}^kf}{\mathrm{d}x^k}|_{x=\xi}=0$ for any $k\in\mathbb{N}$, then $\frac{\mathrm{d}^kh}{\mathrm{d}z^k}|_{z=\xi}=0$ for any $k\in\mathbb{N}$ as well. This means that, as an analytic function, $h\equiv0$ in an open disc centring at $\xi$ in the complex plane. Thus $h\equiv0$ in the whole complex plane and $f=h|_{\mathbb{R}}\equiv0$, which contradicts the assumption that $f$ is not identically zero.
\end{proof}
Before we talking about the roots at ``infinity'', we propose first a simple sufficient condition for a trigonometric-polynomial to have only finitely many real roots:
\begin{proposition}\label{proposition:finiteroot}
Let $f\in\qscxe$ be a trigonometric-polynomial with leading coefficient $a_n\in\qsce.$ If $a_n$ has no real root in $(-\pi,\pi]$, then $f$ has finitely many real roots in $\mathbb{R}$.
\end{proposition}
\begin{proof}
Since $a_n$ is a continues function, we see that it is strictly positive (or strictly negative) over the interval $(-\pi,\pi]$. Note that $a_n(-\pi)=a_n(\pi)$, the observation above is also true for the interval $[-\pi,\pi]$. The function $a_n$ is also continues restricted to the interval $[-\pi,\pi]$, which is a compact set, so there is a positive real number $B$ such that $a_n(x)\geq B$ (or, \emph{resp.}, $a_n(x)\leq -B$) for all $x\in[-\pi,\pi]$. Let $B_i$ be as in (\ref{coebound}) and $X_B$ a positive number depends on $B$ such that for any $x\in\mathbb{R}$ with $|x|>X_B$, the inequality
\begin{equation}\label{inq:Bxn}
B|x^n|>\sum\limits_{i=0}^{n-1}B_i|x^i|
\end{equation}
holds. For such an $x$, we have
\begin{equation}\label{inq:noroot}
\begin{array}{rcl}
|a_nx^n| &\geq& B|x^n| \\
&>&\sum\limits_{i=0}^{n-1}B_i|x^i|\\
&\geq&\sum\limits_{i=0}^{n-1}|a_i||x^i|\\
&\geq&|\sum\limits_{i=0}^{n-1}a_ix^i|.
\end{array}
\end{equation}
Thus for any $x\in\mathbb{R}$ with $|x|>X_B$, $f(x)\neq0$. So, all the real roots of $f$ (if there are any) belong to the interval $[-X_B,X_B]$. Now infinitely many roots of $f$ in $[-X_B,X_B]$ would result in a limit point of them, and hence, by Lemma \ref{lemma:nolimit}, one obtains $f\equiv0$ over $\mathbb{R}$. This contradicts with the fact that $f(x)\neq0$ for any $x\in\mathbb{R}$ with $|x|>X_B$.
\end{proof}
The following result deals with a more general case than the above proposition does:
\begin{proposition}\label{proposition:norootarea}
Suppose that $U\subset(-\pi,\pi]$ is a set such that $|a_n(x)|\geq B$ holds for some positive number $B$ for every $x\in(-\pi,\pi]\backslash U,$ and set
\[J^k_U=\{2k\pi+w\,|\,w\in(-\pi,\pi]\backslash U\}\]
for any integer $k,$ then there is a positive number $X_B$ $($which depends on $B)$ so that $f(x)\neq0$ for any $x\in\bigcup\limits_{k\in\mathbb{Z}}J_U^k$ with $|x|>X_B$.
\end{proposition}
One observes that Prop.\,\ref{proposition:finiteroot} takes care of the special case of the above proposition with $U=\emptyset$. The proof of Prop.\,\ref{proposition:norootarea} is almost the same to the one of Prop.\,\ref{proposition:finiteroot}: we choose the same $X_B$ so that (\ref{inq:Bxn}) holds and note that (\ref{inq:noroot}) holds now only for those $x\in\bigcup\limits_{k\in\mathbb{Z}}J_U^k$ satisfying $|x|>X_B$. We see that: for any sufficiently ``far'' interval $2k\pi+(-\pi,\pi]=(2k\pi-\pi,2k\pi+\pi]$, the roots of $f$ in it (if there are any) are necessarily contained in the subset $2k\pi+U=\{2k\pi+w\,|\,w\in U\}$. The following example illustrates the most frequently used form of the set $U$ throughout the paper when $a_n$ has roots, while the subsequent propositions tell us how to choose a ``good'' $U$ so that the roots of $f$ in every $2k\pi+U$ have ``good'' properties.
For convenience, the ``interval'' $[a,b]$ with $a>b$ is regarded as the empty set but as a usual closed interval when $a\leq b$.
\begin{example}\label{example:Uform}
As the leading coefficient of $f\in\qscx,$ $a_n$ is also a trigonometric-polynomial and $a_n\not\equiv0.$ Assume that $a_n$ has real roots$,$ then$,$ by the proof of Prop$.$\,\ref{proposition:finiteroot}$,$ there are only finitely many roots of $a_n$ in the compact set $[-\pi,\pi].$ Let $x_{0,1}<\cdots<x_{0,\ell}$ be all the roots of $a_n$ in $(-\pi,\pi]$ and $\{\epsilon_i\}_{i=1}^\ell$ be some positive numbers such that $-\pi<x_{0,1}-\epsilon_1$ and $(x_{0,i}-\epsilon_i,x_{0,i}+\epsilon_i)\cap(x_{0,j}-\epsilon_j,x_{0,j}+\epsilon_j)=\emptyset$ for every $1\leq i\neq j\leq\ell.$ Then$,$ we can choose
\begin{equation}\label{equation:U}
U=\big((-\pi,-\pi+\epsilon_0)\cup\bigcup\limits_{i=1}^{\ell}(x_{0,i}-\epsilon_i,x_{0,i}+\epsilon_i)\big)\cap(-\pi,\pi]
\end{equation}
with $0<\epsilon_0<x_{0,1}-\epsilon_1+\pi$ a positive number if $a_n(\pi)=0$ and $\epsilon_0=0$ otherwise$.$ In both cases$,$ we observe that $(-\pi,\pi]\backslash U$ is a subset of
\begin{equation}\label{PminusU}
[-\pi+\epsilon_0,x_{0,1}-\epsilon_1]\cup\bigcup\limits_{i=1}^{\ell-1}[x_{0,i}+\epsilon_i,x_{0,i+1}-\epsilon_{i+1}]\cup[x_{0,\ell}+\epsilon_{\ell},\pi],
\end{equation}
and that the closed intervals in the above formula pairwise disjoint$.$ For each non-empty closed interval $I$ occurring in \emph{(\ref{PminusU}),} the function $a_n$ is continuous and never vanishes in it$.$ Thus $|a_n|\geq B_I$ in $I$ for some positive number $B_I,$ and we can choose the number $B$ in Prop$.$\,{\ref{proposition:norootarea}} to be the minimum of these $B_I.$
\end{example}
In the following, we discuss further how to choose appropriate values of those $\epsilon_i$ in (\ref{equation:U}) so that the roots of $f$ in the sets $2k\pi+U$, which are sufficient ``far'', have ``good'' properties.
\begin{proposition}\label{proposition:sigleroot}
Let $f$ and $a_n$ be as in Prop. \ref{proposition:finiteroot}. If $a_n$ has a single root $x_0$ in $(-\pi,\pi],$ then there is an $\epsilon>0$ and a positive number $X_\epsilon$ $($which depends on $\epsilon)$ such that for any intervals $I^k_{x_0,\epsilon}=(x_0+2k\pi-\epsilon,x_0+2k\pi+\epsilon)$ $($with $k$ an integer$)$ contained in $[X_\epsilon,\infty)$ or $(-\infty,-X_\epsilon],$ $f$ has only a single root in it.
\end{proposition}
\begin{proof}
By assumption, we have $a_n(x_0)=0$ but $a_n'(x_0)\neq0$. Hence, there is a positive number $\epsilon>0$ such that $a_n'$ is strictly positive (or strictly negative) over the interval $[x_0-\epsilon,x_0+\epsilon]$. This implies that \[
a_n(x_0-\epsilon)a_n(x_0+\epsilon)<0.
\] Noting that there is a positive number $A$ such that $|a_n'|\geq A$ over the interval $[x_0-\epsilon,x_0+\epsilon]$, we define
\[
\eta=\min\{A,|a_n(x_0-\epsilon)|,|a_n(x_0+\epsilon)|\}.
\]
Let $B_i$, $i=0,1,\ldots,n-1$ be as in (\ref{coebound}) and set $f'=a_n'x^n+\sum\limits_{i=0}^{n-1}b_ix^i$ with $b_i\in\qsc$. Then there are positive numbers $C_i$, $i=0,\ldots,n-1$, such that $|b_i|\leq C_i$. Let $X_\epsilon$ be a positive number so that the following inequalities hold for any real number $x$ with $|x|\geq X_\epsilon$:
\[
\left\{
\begin{array}{rcl}
\eta|x^n|&>&\sum\limits_{i=0}^{n-1}B_i|x|^i,\\
\eta|x^n|&>&\sum\limits_{i=0}^{n-1}C_i|x|^i.
\end{array}
\right.
\]
Now, fix an integer $k$ such that $I_{x_0,\epsilon}^k\subset[X_\epsilon,\infty)$ or $I_{x_0,\epsilon}^k\subset(\infty,-X_\epsilon]$, then any $x\in I_{x_0,\epsilon}^k$ satisfies $|x|\geq X_\epsilon$. Thus,
\[
\begin{array}{rcl}
|a_n'x^n|&\geq&\eta|x^n| \\
&>& \sum\limits_{i=0}^{n-1}C_i|x|^i\\
&\geq&|\sum\limits_{i=0}^{n-1}b_ix^i|.
\end{array}
\]
This indicates that $f'$ does not vanish in $I_{x_0,\epsilon}^k$, \emph{i.e.}, $f'$ is strictly positive or strictly negative over that interval. On the other hand, if we set $x_*=x_0+2k\pi\pm\epsilon$, then $|x_*|\geq X_\epsilon$ and
\[
\begin{array}{rcl}
|a_n(x_*)x_*^n|&\geq&\eta|x_*^n| \\
&>& \sum\limits_{i=0}^{n-1}B_i|x_*|^i\\
&\geq&|\sum\limits_{i=0}^{n-1}a_ix_*^i|.
\end{array}
\]
This implies that $\text{sgn}(f(x_*))=\text{sgn}(a_n(x_*)x_*^n)$, thus
\[
\begin{array}{rcl}
&&\text{sgn}(f(x_0+2k\pi-\epsilon)f(x_0+2k\pi+\epsilon)) \\
&=& \text{sgn}(a_n(x_0+2k\pi-\epsilon)(x_0+2k\pi-\epsilon)^na_n(x_0+2k\pi+\epsilon)(x_0+2k\pi+\epsilon)^n)\\
&=&\text{sgn}(a_n(x_0-\epsilon)a_n(x_0+\epsilon))\\
&=&-1,
\end{array}
\]
which means
\begin{equation}\label{equation:differentsign}
f(x_0+2k\pi-\epsilon)f(x_0+2k\pi+\epsilon)<0.
\end{equation}Hence there is one and only one single root of $f$ in $I_{x_0,\epsilon}^k$.
\end{proof}
Prop. \ref{proposition:sigleroot} indicates that if $a_n$ has a single root then $f$ has infinitely many real roots. In fact, that would be true if $a_n$ has a root of an odd multiplicity:
\begin{proposition}
Let $f$ and $a_n$ be as usual. If $a_n$ has a root $x_0$ in $(-\pi,\pi]$ of an odd multiplicity$,$ then there is a positive number $\epsilon$ and an $X_\epsilon>0$ so that for any integer $k$ satisfying $I^k_{x_0,\epsilon}\subset(-\infty,-X_\epsilon]\cup[X_\epsilon,\infty),$ there is a root in the interval $I_{x_0,\epsilon}^k$.
\end{proposition}
\begin{proof}
Since $x_0$ is a root of an odd multiplicity, there is a positive number $\epsilon$ so that $a_n$ is strictly increasing or strictly decreasing in the interval $[x_0-\epsilon,x_0+\epsilon]$. This ensures that $a_n(x_0-\epsilon)a_n(x_0+\epsilon)<0$. Set \[\eta=\min\{|a_n(x_0-\epsilon)|,|a_n(x_0+\epsilon)|\}\] and chose a positive number $X_\epsilon$ so that $\eta|x^n|>\sum\limits_{i=0}^{n-1}B_i|x|^i$ holds for any real number $x$ with $|x|\geq X_\epsilon$. For any $k$ such that $I^k_{x_0,\epsilon}\subset(-\infty,-X_\epsilon]\cup[X_\epsilon,\infty)$, we can prove that $\text{sgn}(f(x_*))=\text{sgn}(a_n(x_*)x_*^n)$ with $x_*=x_0+2k\pi\pm\epsilon$ the endpoints of $I_{x_0,\epsilon}^k$, as we have done in the proof of Prop. \ref{proposition:sigleroot}. Thus, the inequality (\ref{equation:differentsign}) holds and there is a root in $I_{x_0,\epsilon}^k$.
\end{proof}
Extending the proof of Prop. \ref{proposition:sigleroot}, we have the following results:
\begin{proposition}\label{proposition:atleastkroots}
If $a_n$ has a root $x_0\in(-\pi,\pi]$ of multiplicity $m,$ then there is a positive number $\epsilon$ and an $X_\epsilon>0$ so that for any integer $k$ satisfying $I^k_{x_0,\epsilon}\subset(-\infty,-X_\epsilon]\cup[X_\epsilon,\infty),$ there are at most $m$ roots $($counted with multiplicity$)$ in the interval $I_{x_0,\epsilon}^k$.
\end{proposition}
\begin{proof}
The proof is inductive on the multiplicity $m$: The case where $m=1$ is clear by Prop. \ref{proposition:sigleroot}. In the following, we prove that the conclusion holds for the case where $m=\ell\geq2$, assuming that it does for the case where $m=\ell-1$.
Write $f'=a_n'x^n+\sum\limits_{i=0}^{n-1}b_ix^i$ with $b_i\in\qsc$. Then $x_0$ is a root of $a_n'$, the leading coefficient of $f'$, of multiplicity $\ell-1$. By assumption, there is a positive number $\epsilon$ and an $X_\epsilon>0$ so that for any integer $k$ satisfying $I^k_{x_0,\epsilon}\subset(-\infty,-X_\epsilon]\cup[X_\epsilon,\infty)$, there are at most $\ell-1$ real roots (counted with multiplicity) in the interval $I_{x_0,\epsilon}^k$. If $f$ has more than $\ell$ roots, say, $r_1<\cdots<r_t$ in the interval $I_{x_0,\epsilon}^k$, each with multiplicity $\ell_1\geq1,\dots,\ell_t\geq1$ and they satisfy $\ell_1+\cdots +\ell_t>\ell$, then every $r_j$ would be a root of $f'$ of multiplicity $\ell_j-1$ and there would be roots $w_1,\ldots,w_{t-1}$ of $f'$ satisfying $r_1<w_1<\cdots<r_{t-1}<w_{t-1}<r_t$. Thus, $f'$ has at least $\sum_{j=1}^t(\ell_j-1)+(t-1)=\sum_{j=1}^t\ell_t-1$ roots in $I_{x_0,\epsilon}^k$. But $\sum_{j=1}^t\ell_t-1>\ell-1$, this means $f'$ has more than $\ell-1$ roots in $I_{x_0,\epsilon}^k$, which contradicts the assumption.
\end{proof}
\end{comment}
\section{Experiments}\label{sec:experiments}
The main algorithm (Algorithm \ref{alg:complete-algorithm}) is implemented with {\tt Maple2021} as a tool named {\tt RootOfMTP}.
In order to indicate the effectiveness and efficiency of the tool, we prepare two types of testing examples, where one type is collected from the literature
and another is generated randomly.
On the one hand, we test the correctness of the tool on these examples through some numerical methods.
On the other hand, the experimental results of these examples show that the tool is powerful and is able to deal with some complicated examples.
\subsection{Commands}
The tool {\tt RootOfMTP} can be downloaded at
\begin{center}
{\bf \url{https://github.com/lihaokun/RootOfMTP}}.
\end{center}
All testing examples can also be downloaded there.
All tests were conducted on 16-Core Intel Core [email protected]
with 128GB of memory and Windows 11.
We illustrate how to use {\tt RootOfMTP} by a simple example.
Suppose we want to isolate all the real roots of the following MTP by {\tt RootOfMTP}:
\begin{align*}
x\sin x + \cos x - 1.
\end{align*}
We only need to run the following commands in {\tt Maple2021}:
{\scriptsize
\begin{verbatim}
read ".../RootOfMTP.mpl";
RootOfMTP(x*sin(x)+cos(x)-1,x,1);
\end{verbatim}}
Herein, the first command is to read the file, and the inputs of the second command are an MTP, the variable of the MTP and a rational number $\epsilon>0$ which specifies the maximal length of isolating intervals.
The output is
{\scriptsize
\begin{verbatim}
For every k <= -1 (k in Z), 2kPi+(0) is a real root with multiplicity 1.
For every k >= 1 (k in Z), 2kPi+(0) is a real root with multiplicity 1.
There is 1 real root with multiplicity 2 at 0.
For every k >= 2 (k in Z), there is 1 real root with multiplicity 1
in (2kPi+(2*arctan(63/16)), 2kPi+(Pi)).
For every k <= -2 (k in Z), there is 1 real root with multiplicity 1
in (2kPi+(-Pi), 2kPi+(-2*arctan(63/16))).
There is 1 real root with multiplicity 1 in each open interval of the list
[[-5/2*Pi-2*arctan(29666650363354128505/36893488147419103232),
-5/2*Pi-2*arctan(7242537696610193/9007199254740992)],
[-1/2*Pi-2*arctan(7030038563941/17592186044416),
-1/2*Pi-2*arctan(14741934773129570377/36893488147419103232)],
[1/2*Pi+2*arctan(14741934773129570377/36893488147419103232),
1/2*Pi+2*arctan(7030038563941/17592186044416)],
[5/2*Pi+2*arctan(7242537696610193/9007199254740992),
5/2*Pi+2*arctan(29666650363354128505/36893488147419103232)]].
\end{verbatim}}
\subsection{Examples from the literature}
We collect $21$ testing examples from the literature \cite{chenliu2016,chen2022automated}.
Since some examples are rational-function inequalities, we need to transform them to MTPs.
Take \cite[Ex.4]{chenliu2016} for example:
\begin{align*}
\frac{\sin^2 x}{x^2}+\frac{\tan x}{x}>\frac{x^2}{\sin^2 x}+\frac{x}{\tan x}.
\end{align*}
First, we move the terms in the right hand side to the left hand side.
Second, we replace $\tan x$ with $\frac{\sin x}{\cos x}$.
Third, we cancel the denominators and obtain an MTP: $-x^4\cos x-x^3\sin x\cos^2 x+x\sin^3 x+\sin^4 x\cos x$.
The experimental results are presented in Table \ref{tab:chen-result}.
Note that these examples are kind of simple.
The bounds $k_{-}$ and $k_{+}$ are close to $0$ and {\tt RootOfMTP} can deal with each of them within $2$ seconds.
\input{table/notation.tex}
\input{table/chen.tex}
\subsection{Random examples}
Now, we present the experimental results for randomly generated
polynomials. A randomly generated polynomial
\begin{align*}
{\bf rp}(d,p,c)\in\Q[x,y,z]
\end{align*}
is constructed as follows:
First, pick some monomials independently from the set
$\{x^{a_1}y^{a_2}z^{a_3}\mid a_i\in\N,~a_1+a_2+a_3\le d\}$ in a manner such that each monomial is picked with probability $p$ while discarded with probability $1-p$.
Then, randomly assign coefficients between $-c$ and $c$ to each picked monomial and sum them to obtain a polynomial.
We generate $160$ random polynomials $F_1,\ldots,F_{160}$ in this way from $16$ different classes, where
\begin{align*}
F_{1},\ldots,F_{10}&\in{\bf rp}(5,0.1,10), & F_{11},\ldots,F_{20}&\in{\bf rp}(10,0.015,10),\\
F_{21},\ldots,F_{30}&\in{\bf rp}(10,0.05,10), & F_{31},\ldots,F_{40}&\in{\bf rp}(10,0.1,10),\\
F_{41},\ldots,F_{50}&\in{\bf rp}(15,0.01,10), & F_{51},\ldots,F_{60}&\in{\bf rp}(15,0.05,10),\\
F_{61},\ldots,F_{70}&\in{\bf rp}(15,0.1,10), & F_{71},\ldots,F_{80}&\in{\bf rp}(20,0.0025,10),\\
F_{81},\ldots,F_{90}&\in{\bf rp}(5,0.1,1000), & F_{91},\ldots,F_{100}&\in{\bf rp}(10,0.015,1000),\\
F_{101},\ldots,F_{110}&\in{\bf rp}(10,0.05,1000), & F_{111},\ldots,F_{120}&\in{\bf rp}(10,0.1,1000),\\
F_{121},\ldots,F_{130}&\in{\bf rp}(15,0.01,1000), & F_{131},\ldots,F_{140}&\in{\bf rp}(15,0.05,1000),\\
F_{141},\ldots,F_{150}&\in{\bf rp}(15,0.1,1000), & F_{151},\ldots,F_{160}&\in{\bf rp}(20,0.0025,1000).
\end{align*}
Then, we replace $y$ and $z$ with $\sin x$ and $\cos x$ respectively in $F_{1},\ldots,F_{160}$ to get MTPs $M_{1},\ldots,M_{160}$.
The experimental results are presented in Table \ref{tab:random-result}.
The tool {\tt RootOfMTP} performs powerfully on these random examples: The tool can solve every example within $3000$ seconds, and $93\%$ of the examples were solved within 1 minute.
We notice that the computation of weak Fourier sequences is not very efficient.
For $59\%$ of the examples, the time for computing the weak Fourier sequences is more than one third of the total computing time.
Another surprising observation is that the bounds $k_{-}$ and $k_{+}$ are still small, even on these more difficult examples.
Finally, we present Figure \ref{fig:num-within-time} to show the performance of {\tt RootOfMTP} on all examples (either from the literature or randomly generated).
The horizontal axis represents time, while the vertical axis represents the number of solved examples within the corresponding time.
\input{table/random.tex}
\input{fig/fig_runtime.tex}
\section{Conclusions}\label{sec:conclusion}
This paper proposes, for the first time, a complete algorithm ``isolating" all the real roots of a given MTP. More precisely, the algorithm computes
\begin{itemize}
\item arbitrarily small intervals whose union covers all real roots of the given MTP (both the roots and the intervals can be countably many), and
\item the exact number of distinct roots in each interval together with the multiplicity of each of them.
\end{itemize}
As a result, this leads to the first complete algorithm (as far as we know) deciding whether there is any, or whether there is infinitely many real roots for a given MTP. The correctness of the algorithm indicates a phenomenon of mathematical interest: the real roots in every period sufficiently far away from zero share a same quantity, a same multiset of root multiplicity and even similar relative positions in each period.
The algorithm has been implemented and a large number of examples have been tested to illustrate its efficiency.
Instead of focusing only on real roots in a certain bounded interval (as some other approaches in the literature doing), our (sub-)algorithms detect and ``isolate" the roots in any bounded intervals of the form $[2k_-\pi-\pi,2k_+\pi+\pi]$, and unbounded intervals of the forms $(2k_+\pi+\pi,\infty)$ and $(-\infty,-2k_-\pi-\pi)$ with sufficiently large $k_\pm$. Based on this, our methods can easily be modified to ``isolate" and compute the multiplicity of all real roots of an MTP in any interval $I$ of the form $(a,b)$, $[a,b)$, $(a,b]$ or $[a,b]$, with $a,b\in\mathbb{Q}\cup\{\pm\infty\}$:
\begin{itemize}
\item Choose an interval $I^\text{e}\supset I$ so that the sub-algorithms ``isolate" the real roots in $I^\text{e}$ naturally.
\item If some of the ``transcendental" intervals cover some endpoints of $I$, replace them by some refined sub-intervals, each of which does not cover any endpoints but contains exactly one transcendental root.
\item After that, every ``isolating" interval covering some endpoints of $I$ is algebraic and contains exactly one root (which is $0$ or algebraic). If there are such intervals, then decide whether those covered endpoints are roots.
\item If the covered endpoints are not roots, refine these endpoint-covering intervals via bisection so that they do not cover the endpoints after refinement. Otherwise, refine them to $[a,a]$ or $[b,b]$ if they cover $a$ or $b$, respectively.
\end{itemize}
In that way we can compute the exact number of real roots in each ``isolating" interval, together with multiplicity of each of them. Based on that, one can further prove MTP inequalities over the interval $I$ if one uses the following simple claim:
\begin{claim}
For any MTP $f$ not identically zero and any interval $I$ $($bounded or not$)$ of positive length,
\begin{itemize}
\item $f>0$ on $I$ iff $f$ has no root in $I$ and $f(x_0)>0$ for some $x_0\in I$;
\item $f\geq0$ on $I$ iff every root of $f$ in the interior of $I$ $($if there are any$)$ has even multiplicity and $f(x_0)>0$ for some $x_0\in I$.
\end{itemize}
\end{claim}
\noindent{}In this way an algorithm proving MTP inequalities over any interval (bounded or not) with endpoints in $\mathbb{Q}\cup\{\pm\infty\}$ can then be derived. To our knowledge, this is the first algorithm proving MTP inequalities over unbounded intervals.
\section*{Acknowledgement}
This work was supported by the NSFC under grants No. 61732001 and No. 12071467, and by the National Key Research Project of China under grant No. 2018YFA0306702
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
\ \ \ \ \ In this work we attempt to construct a correspondence between models in regularity structures \cite{reg} and geometric Rough Paths from classical Rough Path theory \cite{Lyons98,Gubinelli2004}. Results of this kind have already been obtained in the case of Branched Rough Paths which are another type of rough paths defined on trees (see \cite{Gub06}) instead of words. An approach given in \cite{BC} constructs a bijection between the two spaces $\bf{BRP}^{\gamma}$ of branched rough paths and $\bf{ARP}^{\gamma}$ of anisotropic rough paths. The main idea is to use an algebraic result from \cite{Foi02,cha10} that directly relates the underlying Hopf algebras. Inspired by this approach, we endeavour to show that certain Hopf algebras appearing in the context of regularity structures also relate to Hopf algebras with simpler presentation such as quotients of the tensor Hopf algebra $(T(\mathcal{B}), \otimes, \Delta_{\shuffle})$ that appears in the context of geometric Rough Paths.
Trying to find other combinatorial structures than decorated trees introduced in \cite{reg,BHZ} for SPDEs has been developed in the recent multi-index formalism in \cite{LOT,OSSW}. The main difference with our approach, is that in their context one can hope at the best for a post-Lie morphism between decorated trees and multi-indices. Whereas, we obtain in this work an isomorphism. This duality between trees and words for coding expansions has been considered in numerical analysis (see \cite{Murua2017}). We also expect this work to have an impact in the context of Low regularity integrators in \cite{BS} for dispersive PDEs where similar decorated trees are used.
Our approach further relies on an indispensable algebraic tool, which is the notion of a post-Lie algebra. In \cite{BM22}, it has been shown that certain Hopf algebras appearing in the context of regularity structures can be built directly from certain pre-Lie algebraic structures -a special case of post-Lie algebras- that are simpler to describe. This is accomplished via means of a recursive construction of the product by Guin and Oudom, first appearing in \cite{Guin1,Guin2}. This fact can reveal important information about the Hopf algebras involved. Given a pre-Lie algebra $(E, \curvearrowright)$ the Guin-Oudom procedure constructs a product on the symmetric space over the undrerlying vector space $E$. It also endows the space with the shuffle coproduct $\Delta_{\shuffle}$ thus turning it into a Hopf algebra, which turns out to be isomorphic to the universal enveloping algebra of the commutator Lie algebra $E_{Lie}$ associated to $E$.
It was already known that the graded dual of the Butcher-Connes-Kreimer Hopf algebra $\mathcal{H}_{\text{\tiny{BCK}}}$ \cite{Butcher,CK1,CK2}, which is the Grossman-Larson Hopf algebra $\mathcal{H}_{\text{\tiny{GL}}}$ \cite{GL}, can be generated in this manner by the free pre-Lie algebra over a set of generators which can be described as the linear span of trees endowed with the grafting product. In the work of \cite{BM22} it is proven that the graded dual of a deformed version of the Butcher-Connes-Kreimer Hopf algebra is also generated in this manner by a deformed version of the grafting product. This deformed version of the grafting product is then shown to be isomorphic to the original grafting product in the category of pre-Lie algebras via an isomorphism $\Theta$. This is illustrated below via the following diagram:
\begin{equs}\label{diag_1}
\begin{aligned}
\xymatrix{\curvearrowright
\ar[rr]^{\scriptsize{\text{Guin-Oudom}}}\ar[d]_{\Theta} && \star \ar[d]^{\Phi} && \ar[ll]^{\scriptsize{\text{Dual}}} \Delta_{\text{\tiny{BCK}}} \\ \wh \curvearrowright
\ar[rr]^{\scriptsize{\text{Guin-Oudom}}} && \tilde{\star} && \ar[ll]^{\scriptsize{\text{Dual}}} \Delta_{\text{\tiny{DBCK}}}
}
\end{aligned}
\end{equs}
where $ \wh \curvearrowright $ is the deformed grafting obtained from $ \curvearrowright $ by $ \Theta $. The coproducts $\Delta_{\text{\tiny{BCK}}} $ and $\Delta_{\text{\tiny{DBCK}}} $ are respectly the Butcher-Connes-Kreimer and the deformed Butcher-Connes-Kreimer coproducts. The products $ \star $ and $ \tilde{\star} $ are the Grossman-Larson and deformed Grossman-Larson products.
The isomorphism $ \Phi $ between these two products is obtained by applying the Guin-Oudom functor to $ \Theta $ (see Theorem~\ref{theorem_functor}).
Furthermore, the work of \cite{BK} completes this programme, in the sense that post-Lie algebras that generate the graded duals of the Hopf algebras $ \mathcal{H}_2 $ used for the recentering in singular SPDEs (see \cite{reg,BHZ,BM22}). Again, for each Hopf algebra one has an original and deformed version and these are correspondingly proven to be generated by a post-Lie product or a deformed version thereof. This could be summarise in the following diagram:
\begin{equs}\label{diag_2}
\begin{aligned}
\xymatrix{\curvearrowright^{\text{\tiny{Lie}}}
\ar[rr]^{\scriptsize{\text{Guin-Oudom}}}\ar[d]_{\Theta^{\text{\tiny{Lie}}}} && \star^{\text{\tiny{Lie}}} \ar[d]^{\Phi^{\text{\tiny{Lie}}}} && \ar[ll]^{\scriptsize{\text{Dual}}} \Delta_{\text{\tiny{BCK}}}^{\text{\tiny{Lie}}} \\ \wh \curvearrowright^{\text{\tiny{Lie}}}
\ar[rr]^{\scriptsize{\text{Guin-Oudom}}} && \star_2 && \ar[ll]^{\scriptsize{\text{Dual}}} \Delta_{2}
}
\end{aligned}
\end{equs}
We have added the notation $ \text{\tiny{Lie}} $ to stress the fact that one starts with a Lie algebra and therefore the previous operations are extended to new objects. The Guin-Oudom procedure used is the one for post-Lie algebras developed in \cite{ELM}. The map $ \Delta_2 $ is the coproduct for $ \mathcal{H}_2 $.
The map $ \Psi $ allows to say that the deformed Butcher-Connes-Kreimer Hopf algebra is isomorphic to the tensor Hopf algebra (see Theorem~\ref{Main}). Indeed, the basis $ B $ given by the Chapoton-Foissy isomorphism $ \Psi_{\text{\tiny{CF}}} $ is transported via $ \Phi $ in the sense that one has:
\begin{equs}
\Psi_{\text{\tiny{CF}}} : \tau_1 \star \ldots \star \tau_r \mapsto
\tau_1 \otimes \ldots \otimes \tau_r, \quad \tau_i \in B
\end{equs}
Then, the new isomorphism $ \Psi_{\Phi} $ is given by
\begin{equs}
\Psi_{\Phi} : \Phi(\tau_1) \ \tilde{\star} \ldots \tilde{\star} \ \Phi(\tau_r) \mapsto
\Phi(\tau_1) \otimes \ldots \otimes \Phi(\tau_r)
\end{equs}
where $ \Phi(\tau_1) \, \otimes \ldots \otimes \, \Phi(\tau_n) \in T(\Phi(\mathcal{B})) $. This gives a clear description of the basis that can be used in the context of the deformed Butcher-Connes-Kreimer Hopf algebra. We also know that elements of $ \mathcal{B} $ are linear combination of planted trees that are primitives elements for the Butcher-Connes-Kreimer coproduct of the form $ \mathcal{I}_{a}(\tau) $. Here, in the notation $ \tau $ is a linear combination of decorated trees and $ \mathcal{I}_{a} $ correspond of the grafting of these trees onto a new root via an edge decorated by $ a $.
This result does cover the Hopf algebras used in \cite{BS} but not the one at play in the context of Regularity Structures. Indeed, not only are planted trees used for describing solutions of singular SPDEs but so are classical monomials $ X_i $. In the expansion, these objects are associated to some operators that do not commute, motivating the introduction of a natural Lie bracket.
The grafting product has to be compatible with this underlying Lie bracket and that is encapsulated in the form of a post-Lie product recently introduced for SPDEs in \cite{BK}. Therefore, the Lie- algebraic structure has to be taken into account when one extends the isomorphism introduced by Chapoton and Foissy. In our main result (see Theorem~\ref{main_result_paper}) the alphabet $ A $ is given by the $ X_i $ and the $ \Phi(\mathcal{I}_a(\tau)) \in \Phi(\mathcal{B}) $ and we denote by $ W $ words on this alphabet. The space $ \tilde{W} $ is given as the quotient of $W$ by the Hopf ideal $ \mathcal{J} $ generated by the elements
\begin{equs}
\{ X_i \otimes \Phi( \mathcal{I}_{a}(\tau) ) - \Phi( \mathcal{I}_{a}(\tau) ) \otimes X_i - \uparrow^{i} \Phi( \mathcal{I}_{a}(\tau)) - \Phi(\mathcal{I}_{a - e_i}(\tau)) \}
\end{equs}
where $ \uparrow^{i} $ corresponds to changing a node decoration by adding $ e_i $ to it. The $ e_i $ are the canonical basis of $ \mathbb{N}^{d+1} $. Then, there exists a Hopf algebra isomorphism $ \Psi $ between decorated trees and $ \tilde{W} $ equipped with $ \star_2 $.
The map $ \Psi $ is given by
\begin{equs}
\Psi : \prod_{i=1}^n \CI_{a_i}(\tau_i) X^k \rightarrow \Psi_{\Phi}(\prod_{i=1}^n \CI_{a_i}(\tau_i)) \otimes_{j=0}^{d} \otimes_{i=1}^{k_j} X_i.
\end{equs}
where $ k = (k_0,...,k_d) \in \mathbb{N}^{d+1} $, $ X^k = \prod_{i=0}^{d} X_i^{k_i} $ and $ \prod_{i=1}^n \CI_{a_i}(\tau_i) X^k $ corresponds to a certain order as two planted trees commute but not $ X_i $ and a planted tree which is encoded in the Lie bracket. The product $ \prod_{i=1}^n $ is commutative.
The paper will be structured as follows: In Section~\ref{section::2} we give an outline of the results in \cite{BC}, where the authors construct a correspondence between geometric Rough Paths and branched Rough Paths. This is accomplished using a result of Foissy and Chapoton that shows that the Grossman-Larson Hopf algebra is in fact isomorphic to the tensor Hopf algebra.
As rough paths may be seen as parametrized families of characters of these Hopf algebras one can use composition with the Hopf algebra isomorphism to directly transverse across structures see Theorem~\ref{chiso}. In this section, we also present the generalised version of the classical Butcher-Connes-Kreimer Hopf algebra $ \mathcal{H}_{\text{\tiny{DBCK}}} $ that can accomodate decorations on the edges and vertices of the forests. We also present the generalized version of the Grossman-Larson Hopf algebra $ \mathcal{H}_{\text{\tiny{GL}}} $ which again comes with decorations on the edges and vertices and is dual to $ \mathcal{H}_{\text{\tiny{DBCK}}} $. We also present the Chapoton-Foissy isomorphism) in this context (see Theorem~\ref{Foissy_Chapoton_basis}) that we will use in the sequel.
In Section~\ref{section::3}, we present the grafting and deformed grafting pre-Lie products and explain how they generate the Grossman-Larson product (see Theorem~\ref{Free pre-Lie -> Grossman - Larson}) and a deformed version thereof (see Theorem~\ref{Deformed Grafting0}), via a contruction given by Guin and Oudom . We then present the isomorphism $\Theta$ appearing in \cite{BM22} between the grafting and deformed grafting pre-Lie algebras. We use this to prove that this deformed structure is isomorphic as a Hopf Algebra to the original Grossman-Larson Hopf algebra. Hence, by virtue of the result of Foissy and Chapoton, it is also isomorphic to the tensor Hopf algebra (see Theorem~\ref{Main}).
In Section~\ref{section::4} we present the post-Lie algebraic formalism and the generalisation of the Guin-Oudom construction in that context. We recall the results in \cite{BK} and make essential use of them to prove our main result. The post-Lie algebraic formalism allows for a precise encoding of the action of the deformed grafting product alongside its interaction with the $\uparrow^i$ operators. Uploading this data onto the universal enveloping algebra via the construction in \cite{ELM} allows for a finer analysis of the $\mathcal{H}_{2}$ Hopf algebra and the relation it bears to the simpler deformed Grossman-Larson Hopf algebra. This information, together with the isomorphism $\Psi_{\Phi}$ ultimately leads to the main result Theorem~\ref{main_result_paper} which is an isomorphism between the $\mathcal{H}_{2}$ Hopf algebra and an appropriate quotient of the shuffle Hopf algebra.
Finally in Section~\ref{section::5}, we explore some applications in the context of regularity structures. We show how, using this isomorphism one may move from one encoding to another. This is done by composing the elements of the structure group with the isomorphism and obtaining a new structure group acting on the relevant quotient of the tensor Hopf algbra. This is similar in spirit to the approach in \cite{BC}. The main algebraic result of the section is Theorem~\ref{T_+ isomorphism} which identifies $ \mathcal{T}_+ $ the vector space used for the structure group as isomorphic to a Hopf subalgebra of the quotient of the shuffle Hopf algebra described before. Then, we propose an attempt to rewrite Theorem~\ref{chiso} int the context of Regularity Structures (see Theorem~\ref{chiso_bis}).
\section{From non-geometric to geometric rough paths}
\label{section::2}
In this section, we present the state of the art in the context of moving from geometric to branched rough paths. We formulate the results using the notion of anisotropic and branched $\gamma$ rough paths, with the corresponding spaces denoted by $\mathbf{ARP}^{\gamma}$ and $\mathbf{BRP}^{\gamma}$. We shall give the definitions of the relevant Hopf algebras and then proceed to outline the results in \cite{HK15}, \cite{BC} and \cite{TZ}. Our approach in the next sections is a generalization of the approach presented in \cite{BC}.
\ \ \ \ Before moving further, we introduce some notation. We shall denote by $T_{E}^{V}$ the set of all rooted trees with vertices decorated by $V$ and edges decorated by $E$ and by $F_{E}^{V}$ the set of forests which consists of monomials over $T_{E}^{V}$. We then denote by $\mathcal{T}_{E}^{V}$ the formal linear span of $T_{E}^{V}$. We also denote by $\CF_{E}^{V}$ the forest algebra, which consists of all polynomials over $T_{E}^{V}$. It is the free commutative algebra over the vector space $\mathcal{T}_{E}^{V}$.
Furthermore, we denote by $ P_E^V $ the set of planted trees and by $ \mathcal{P}_E^V $ their linear span. \label{planted_p} A planted tree is of the form $ \mathcal{I}_{a}(\tau) $ where $ \tau \in \mathcal{T}_E^V $ and $ \mathcal{I}_{a}(\tau) $ denotes the grafting of the tree $ \tau $ onto a new root via an edge decorated by $a \in E$. We also use $N_{\tau}$, $E_{\tau}$ and $L_{\tau}$ for the set of vertices, edges and leaves of a tree $\tau \in T_{E}^{V}$.
When there is no decorations on the edge we remove the subscript $ E $ and write for instance $ T^V $. We may equip these structures with different products and at these times we will essentially use the notation to refer to the underlying vector spaces.
We will use $\catname{Vec}$ for the category of vector spaces and $\catname{Alg}$ and $\catname{CAlg}$ for the category of algebras and commutative algebras respectively. We shall use $S: \catname{Vec} \rightarrow \catname{CAlg}$ for the symmetric algebra functor taking a vector space $V$ to the free commutative algebra over $V$. Similarly, we use $T: \catname{Vec} \rightarrow \catname{Alg}$ for the tensor algebra functor taking a vector space $V$ to the free associative algebra over $V$.
We also define an admissible cut of a tree to be any selection of edges that are pairwise incomparable with respect to the partial ordering induced by the tree. If $h \in \CF_{E}^{V}$ then we use $\text{Adm}(h)$ to denote the set of admissible cuts of $h$.
\begin{definition}
We define the Butcher-Connes-Kreimer coproduct $\Delta_{\tiny{\text{BCK}}}$ on the symmetric algebra $S(\mathcal{P}_{E}^{V})$ by setting, for $h \in \mathcal{P}_{E}^{V}$:
\begin{equs} \label{Connes_Kreimer_1}
\Delta_{\tiny{\text{BCK}}} (h) := \sum_{C \in \scriptsize{\text{Adm}}(h) } R^C(h) \otimes \tilde{P}^C(h)
\end{equs}
Here, we have used $\tilde{P}^C(h)$ to denote the pruned forest that is formed by collecting all the edges at or above the cut, including the ones upon which the cut was performed, so that the edges that were attached to the same node in $h$ are part of the same tree. The term $R^C(h)$ corresponds to the "trunk", that is the subforest formed by the edges not lying above the ones upon which the cut was performed. In the case of decorated trees with no decorations on the edges, we consider the classical Butcher-Connes-Kreimer coproduct given by:
\begin{equs} \label{Connes_Kreimer_2}
\hat{\Delta}_{\tiny{\text{BCK}}} (h) := \sum_{C \in \scriptsize{\text{Adm}}(h) } R^C(h) \otimes P^C(h)
\end{equs}
where this time, we do not keep the edges in the cut $ C $ with $ P^{C}(h) $.
\end{definition}
The Butcher-Connes-Kreimer Hopf Algebra $\mathcal{H}_{\tiny{\text{BCK}}}$ is the graded bialgebra with underlying algebraic structure given by the natural symmetric product on $ S(\mathcal{P}^V_E) $ and coalgebra structure given by $ \Delta_{\tiny{\text{BCK}}} $. The grading is defined to be the number of edges. As a graded connected bialgebra, it is also a Hopf algebra. We denote the usual Butcher-Connes-Kreimer Hopf algebra by $\hat{\mathcal{H}}_{\tiny{\text{BCK}}}$ given by $ S(\mathcal{T}^V) $ equipped with the forest product and coproduct $ \hat{\Delta}_{\tiny{\text{BCK}}} $. For this Hopf algebra, the grading is the number of nodes.
\begin{remark} In the context of rough paths, the coproduct $ \hat{\Delta}_{\tiny{\text{BCK}}} $ has been mostly used. In the context of SPDEs with several equations, one has to keep track of the various operators needed for a given iterated integral. This is perfomed by decorations on the edges. Let us mention, that a variant of the Butcher-Connes-Kreimer coproduct has been used in the context of Volterra-type rough paths in \cite{BK1} where the edges cut are also kept for keeping track the fact that they were attached to the same node. This is crucial for proving a generalized Chen's relation involving a convolution-type product.
\end{remark}
\begin{remark}
We shall frequently use Sweedler's notation and will write $\Delta_{_{\tiny{\text{BCK}}}}(h) = \sum_{(h)} h^{(1)} \otimes h^{(2)}$ to denote the sum ranging over the expansion of the coproduct $\Delta_{\tiny{\text{BCK}}}$. We will also frequently use Sweedler's notation for the other coproducts that will be introduced.
\end{remark}
One can provide a recursive formula for the Butcher-Connes-Kreimer coproduct $ \Delta_{\text{\tiny{BCK}}}$:
\begin{equs} \label{recur_def_deltaCK}
\begin{aligned}
\Delta_{\text{\tiny{BCK}}} \mathbf{1} & = \mathbf{1} \otimes \mathbf{1} \\
\Delta_{\text{\tiny{BCK}}}
\CI_a(X^k \tau) & =
\left( \mathrm{id} \otimes \CI_a(X^k \cdot) \right) \Delta_{\text{\tiny{BCK}}} \tau + \CI_a(X^k \tau) \otimes \mathbf{1},
\end{aligned}
\end{equs}
and it extends multiplicatively to the product of $ S(\mathcal{P}^V_E) $. Here, $ \tau $ belongs to $ S(\mathcal{P}^V_E) $. The map $ \CI_a(X^k \cdot) $ is an operator from $ S(\mathcal{P}^V_E) $ into $ S(\mathcal{P}^V_E) $ that grafts the decorated tree $ X^k \tau $ onto a new root via an edge decorated by $ a $.
Associated to this coproduct, one has an associative product denoted as the Grossman-Larson product $ \star $ defined as:
\begin{equs}
\sigma \star \tau := \left( \sigma \otimes \tau \right) \Delta_{\text{\tiny{BCK}}}
\end{equs}
where we use the following identification by viewing $ \tau \in S(\mathcal{P}^V_E) $ as a linear functional in $ S(\mathcal{P}^V_E) $ such that : $ \langle\tau,\bar \tau \rangle = S(\tau) $ if $ \tau = \bar \tau $ and zero elsewhere. The coefficient $ S(\tau) $ corresponds to the internal symmetry factor of the forest $ \tau $. It is given by $ \prod_{i} |\text{Aut}(\tau_i)| $ where the $ \tau_i $ are the trees of $ \tau $ and $ |\text{Aut}(\tau_i)| $ is the number of automorphisms preserving $ \tau_i $.
We define a second coproduct on $ S(\mathcal{P}^V_E) $ as
\begin{equs}
\Delta \tau = \tau \otimes \mathbf{1} + \mathbf{1} \otimes \tau
\end{equs}
and then extends multiplicatively for the symmetric product of $ S(\mathcal{P}^V_E) $.
For the rest of the section, we set $V = \{0,...,d\}$. We denote by $ T^V_n $ decorated trees with $ n $ nodes.
We also set $ F^V_N $ to be of forests with at most $ N $ nodes.
We denote by $ \mathcal{G} $ (resp. $ \mathcal{G}_N $) the set of characters from the Hopf algebra $\mathcal{H} $ (resp. $\mathcal{H}_N = \langle \CF_N \rangle $) into $\R$. These are linear algebra morphisms forming a group with respect to the convolution product $ \star_0 $
\begin{equs}
\label{convolution}
X \star_0 Y := (X \otimes Y ) \hat{\Delta}_{\text{\tiny{BCK}}}
\end{equs}
The unit for the convolution product is the co-unit $\mathbf{1}^*$ which is non-zero only on the empty tree.
\begin{definition
Let $\gamma\in\,]0,1[$, we define a branched $\gamma$-rough path as a map $X:[0,1]^2\to\cal G$ such that $X_{tt}=\mathbf{1}^{*}$ and such that Chen's relation is satisfied:
\begin{equs} \label{chen}
X_{su} \star_0 X_{ut}=X_{st}, \qquad s,u,t\in[0,1],
\end{equs}
and the analytical condition
\begin{equs} \label{analytical}
|\langle X_{st},\tau\rangle|\lesssim|t-s|^{\gamma|\tau|},
\end{equs}
for every $ \tau $ that does not contain the zero decoration on the nodes. The term $ |\tau| $ corresponds to the number of nodes. Otherwise, we have
\begin{equs} \label{analytical_0}
\sup_{0 \leq s,t \leq 1} \frac{\langle X_{st}, \tau \rangle}{|t-s|^{(1-\gamma)|\tau|_0 + \gamma |\tau|}} < \infty,
\end{equs}
where $ |\tau|_0 $ counts the number of times the decoration $ 0 $ appears in $ \tau $. In the sequel, we will consider the biggest $ N \in \N $ such that $ \gamma N \leq 1 $. The branched $ \gamma $-rough paths are taking values in $ \mathcal{G}_{N} $.
We denote this space by $\bf{BRP}^\gamma$.
\end{definition}
We also introduce the shuffle Hopf algebra. Given an alphabet $ A $ and we consider the linear span of the words on this alphabet denoted by $ T(A) $. We set $ \varepsilon $ as the empty word. The product on $ T(A)$ is the shuffle product defined by
\begin{equs}
\varepsilon \shuffle v=v\shuffle \varepsilon =v, \quad (au\shuffle bv) = a(u\shuffle bv) + b(au\shuffle v)
\end{equs}
for all $u,v\in T(A)$ and $a,b\in A$.
We first define the shuffle coproduct dual to the deshuffle coproduct defined for every $ a \in A $ as
\begin{equs}
\Delta_{\shuffle} a = a \otimes \varepsilon + \varepsilon \otimes a
\end{equs}
and then extends multiplicatively for the tensor product $ \otimes $. The coproduct $\bar\Delta:T(A)\to T(A)\otimes T(A)$ is the deconcatenation of words:
\[ \bar\Delta(a_1\dotsm a_n) = a_1\dotsm a_n\otimes \varepsilon + \varepsilon \otimes a_1\dotsm a_n + \sum_{k=1}^{n-1}a_1\dotsm a_k\otimes a_{k+1}\dotsm a_n.\]
Equipped with this product and coproduct $ T(A) $ is a Hopf algebra. The grading of $ T(A) $
is given by the length of words $\ell(a_1\dotsm a_n) = n$. We denote by $ \mathcal{G}_A $ the group of characters associated to $ T(A) $ and by $ * $ the convolution product.
\begin{definition}
An anisotropic $\gamma$-rough path, with $\gamma=(\gamma_a, \, a\in A)$, $ 0<\gamma_a<1$, is a map $X:[0,1]^2\to \mathcal{G}_{A}$ such that $ X_{tt} = \varepsilon^{*} $ where $ \varepsilon^{*}$ is the counit. It satisfies
\begin{equs}
X_{su} * X_{ut}=X_{st}, \qquad |\langle X_{st},v\rangle|\lesssim|t-s|^{\hat{\gamma}\omega(v)}
\end{equs}
for all $(s,u,t)\in[0,1]^3$ and words $v$. Here, $\hat{\gamma}=\min_{a\in A}\gamma_a$ and for a word $v=a_1\dotsm a_k$ of length $k$ we define
\begin{equation}
\label{eq:weight}
\omega(v)=\frac{\gamma_{a_1}+\dotsc+\gamma_{a_k}}{\hat{\gamma}}=\frac{1}{\hat{\gamma}}\sum_{a\in A} n_a(v)\gamma_a
\end{equation}
where $n_a(v)$ is the number of times the letter $ a $ appears in $ v $.
The space of anisotropic $\gamma$-rough paths is denoted by $ \bf{ARP}^{\gamma} $. When the $ \gamma_a $ are all equal to a fixed $\gamma $, one recovers the classical geometric rough paths.
\end{definition}
As for the branched rough paths, we perform a truncation and consider paths taking values in $ \mathcal{G}_{\mathcal{T}_N,N}$. Elements of $ \mathcal{G}_{\mathcal{T}_N,N}$ are characters over $ T_N(\mathcal{T}_N) $ which are words $ v = \tau_1 \otimes \ldots \otimes \tau_n $ built on the alphabet $ \mathcal{T}_N $ such that $ \sum_{i=1}^n |\tau_i| \leq N $.
The first approach to moving from trees into words is given by the Hairer-Kelly map $ \Psi_{\text{\tiny{HK}}} $ in the context of geometric rough paths in \cite{HK15}.
This map first introduced in \cite{HK15} is given in \cite[Def. 4, Sec. 6]{Br173} as the the unique Hopf algebra morphism from $\mathcal{H}$ to the shuffle Hopf algebra $(T(\mathcal{T}), \shuffle)$ obeying:
\[ \Psi_{\text{\tiny{HK}}} = (\Psi_{\text{\tiny{HK}}} \otimes P_{\mathbf{1}}) \hat{\Delta}_{\text{\tiny{BCK}}}
\]
where $P_{\mathbf{1}}:=\mathrm{id}-\mathbf{1}^{*}$ is the augmentation projector.
The following theorem given in \cite{TZ} established a correspondence between anisotropic rough paths and branched rough paths:
\begin{theorem}\label{Hairer-Kelly}
Let $X$ be a branched $\gamma$-rough path. There exists an anisotropic geometric rough path $\bar{X}$ indexed by words on the alphabet $\mathcal{T}_N$, $ N = \lfloor 1/\gamma \rfloor $, with exponents $(\gamma_\tau, \tau\in\mathcal{T}_N)$, and such that $\langle X,\tau\rangle = \langle\bar{X},\Psi_{\text{\tiny{HK}}}(\tau)\rangle$.
\end{theorem}
\begin{remark}
The previous theorem relies on the Lyons-Victoir extension given in \cite{LV07} theorem which is not canonical.
The authors in \cite{TZ} identified a transitive free action of the additive group $\mathcal{C}^\gamma$ on $ \bf{BRP}^\gamma $. The abelian group $\mathcal{C}^\gamma$ is given by
\[
\cal C^\gamma:=\{(g^\tau)_{\tau\in \mathcal{T}_N}: \, g^\tau_0=0, \, g^\tau\in C^{\gamma_\tau}([0,1]), \, \forall\, \tau\in\mathcal{T}, |\tau|\leq N\}.
\]
Explicit expressions for $ g $ have been given in \cite{Br20} for the BPHZ renormalisation at the level of rough paths introduced in \cite{BCFP}.
Parametrisation in the context of Regularity Structures has been considered in \cite{BB21b}.
\end{remark}
Lastly, the approach most relevant to this work, given in \cite{BC}, constructs a bijection between the two spaces
$ \bf{BRP}^{\gamma} $ and $ \bf{ARP}^{\gamma} $. The main idea is to use an algebraic result from \cite{Foi02,cha10}. We denote by $ \hat{\mathcal{H}}_{\text{\tiny{GL}}} $ the Grossmann-Larson Hopf algebras (resp. $ \mathcal{H}_{\text{\tiny{GL}}} $) defined on $ S(\mathcal{T}^V_E) $ (resp. $ S(\mathcal{P}^V_E) $ ) equipped with the product $ \star_0 $ (resp. $ \star $) and the coproduct $ \Delta $. We recall this result of Foissy and Chapoton:
\begin{theorem} \label{Foissy_Chapoton_basis}
There exists a subspace $ \hat{\mathcal{B}} = \langle\tau_1, \tau_2,...\rangle $ of $ \mathcal{T} $ (resp. $ \mathcal{B} $ and $ \mathcal{P}^V_E $) such that $ \hat{\mathcal{H}}_{\text{\tiny{GL}}} $ (resp. $ \mathcal{H}_{\text{\tiny{GL}}} $) is isomorphic as a Hopf algebra to the tensor Hopf algebra $ (T(\hat{\mathcal{B}}), \otimes, \Delta_{\shuffle})$ (resp. $ (T(\mathcal{B}), \otimes, \Delta_{\shuffle})$) which consists of the linear span of the set of words from the alphabet $\hat{\mathcal{B}}$ (resp. $ \mathcal{B} $), endowed with the tensor product and the shuffle coproduct.
\end{theorem}
We provide an outline here of the construction. First, one proves the existence of a set $B = \{ \tau_{1}, \tau_{2}, ... \}$ that consists of primitive elements of the Hopf algebra $\mathcal{H}_{\text{\tiny{GL}}}$ belonging to $ \mathcal{P}^V_E $ such that every $\tau \in \mathcal{H}_{\text{\tiny{GL}}}$ has a unique representation of the form:
\begin{equs} \label{uniquedec}
\tau = \sum_{R} \lambda_R \tau_{r_1} \star \ldots \star \tau_{r_n}
\end{equs}
where the sum is performed over finitely many multi-indices $R = (r_1,\ldots,r_n)$.
Then, one can exhibit an isomorphism $ \Psi_{\text{\tiny{CF}}} $ between the two Hopf algebras $ \mathcal{H}_{\text{\tiny{GL}}} $ and $ T(\mathcal{B}) $ where $ \mathcal{B}$ is the linear span of $B$ as follows:
\begin{equs}
\Psi_{\text{\tiny{CF}}} : \tau_1 \star \ldots \star \tau_r \mapsto
\tau_1 \otimes \ldots \otimes \tau_r
\end{equs}
where $ \tau_1 \otimes \ldots \otimes \tau_r \in T(\mathcal{B}) $. This will be the isomorphism that we will use in the next section. We will obtain an isomorphism of the deformed Grossman-Larson Hopf algebra $\mathcal{H}_{\text{\tiny{DGL}}}$ with the tensor Hopf algebra $ (T(\mathcal{B}), \otimes, \Delta_{\shuffle})$.
In the context of rough paths, one uses the isomorphism $ \hat{\Psi}_{\text{\tiny{CF}}} $ between the two spaces $ \mathcal{H}_{N}^{*} $ and $ T_N(\hat{\mathcal{B}}_{N}) $ based on the basis $ \hat{\mathcal{B}}_N $ (see \cite[Lemma 4.2]{BC}):
\begin{equs}
\hat{\Psi}_{\text{\tiny{CF}}} : \tau_1 \star_0 \ldots \star_0 \tau_r \mapsto
\tau_1 \otimes \ldots \otimes \tau_r
\end{equs}
where $ \tau_1 \otimes \ldots \otimes \tau_n \in T_N(\hat{\mathcal{B}}_{N}) $. One has from \cite{BC}
\begin{theorem} \label{chiso} Let $ X \in \bf{BRP}^{\gamma} $, then $ \hat{X} := \hat{\Psi}_{\text{\tiny{CF}}}(X) \in \bf{ARP}^{\gamma}$.
\end{theorem}
In \cite{Br20}, the action of the renormalisation on this construction has been described. The family of renomalisation maps considered are BPHZ renormalisation map $ M $ (inspired from the BPHZ renormalisation of Feynman diagrams \cite{WZ69,Hepp,BP} which was used in the context of Regularity Structures \cite{BHZ,CH16}) whose adjoints $ M^{*} $ are morphisms for the product $ \star_0 $:
\begin{equs}
M^{*} \left( \tau \star_0 \sigma \right) = M^{*} \tau \star_0 M^{*} \sigma
\end{equs}
Then, one is able to define a renormalisation map $ \hat{M}^{*} $ on $ T_N(\hat{\mathcal{B}}_{N}) $ that commutes with the isomorphism $ \hat{\Psi}_{\text{\tiny{CF}}} $ (see \cite[Theorem 4.7]{Br20}):
\begin{equs}
\hat{M}^{*} \hat{X} = \hat{M}^{*} \hat{\Psi}_{\text{\tiny{CF}}}(X) = \hat{\Psi}_{\text{\tiny{CF}}}( M^{*} X).
\end{equs}
BPHZ renormalisation maps in the context of rough paths have been first considered in \cite{BCFP} with some examples provided in \cite{BCF}.
\section{An isomorphism for the deformed Grossman-Larson Hopf algebra}
\label{section::3}
In this section, we introduced pre-Lie and multi-pre-Lie algebras with the main example being the grafting product for decorated trees and its deformations given in \cite{BCCH,BM22}. Then, we apply the Guin-Oudom procedure \cite{Guin1, Guin2} to the grafting product for deriving the Grossman-Larson Hopf algebra $\mathcal{H}_{\text{\tiny{GL}}}$, the graded dual of the Butcher-Connes-Kreimer Hopf algebra $\mathcal{H}_{\text{\tiny{BCK}}}$. Using the functor given by Guin-Oudom, we are able to lift the isomorphism $ \Theta $ introduced in \cite{BM22} (see \eqref{recursive_theta}) at the level of the grafting products to an isomorphism $ \Phi $ for the deformed Grossman-Larson Hopf algebras (see Theorem~\ref{theorem_functor}).
This allows us to state our main result which is Theorem~\ref{Main} that translates the Chapoton-Foissy isomorphsm in the context of the deformed Grossman-Larson Hopf algebra: One just applies the isomorphism $ \Phi $ to the basis previously obtained in Theorem~\ref{Foissy_Chapoton_basis}.
\ We begin this section by giving the definition of a pre-Lie algebra.
\begin{definition}
A pre-Lie algebra is an algebra $(\mathcal{P}, \curvearrowright)$ over a field $\mathbf{k}$ of characteristic $0$, whose product satisfies the following relation for every $ x,y,z \in \mathcal{P} $
$$
x \curvearrowright (y \curvearrowright z) - (x \curvearrowright y) \curvearrowright z = y \curvearrowright (x \curvearrowright z) - (y \curvearrowright x) \curvearrowright z.
$$
\end{definition}
\begin{remark}
Note that every associative algebra is a pre-Lie algebra as in this case the associator vanishes and the left- and right-hand sides above are both equal to zero.
\end{remark}
A pre-Lie algebra gives rise to a Lie algebra:
\begin{proposition}
If $(E, \curvearrowright)$ is a pre-Lie algebra, then the commutatator $[x, y] = x \curvearrowright y - y \curvearrowright x$ is a Lie bracket.
\end{proposition}
\begin{remark}
Here is an equivalent definition: An algebra $(E, \curvearrowright)$ over a field $\mathbf{k}$ of characteristic $0$, whose commutator is a Lie bracket and left multiplication by $\curvearrowright$ gives a representation of the commutator Lie algebra.
\end{remark}
\begin{remark}
Not every Lie algebra comes from an associative algebra (with the commutator as it's Lie bracket). For example, free Lie algebras do not arise from any associative algebra. They do, however, arise from free pre-Lie algebras, see \cite{cha10}. It is interesting to study the implications of a Lie algebra $L$ arising from a pre-Lie structure. An explicit recursive procedure, given by Guin and Oudom (see Theorem~\ref{Guin-Oudom}), for constructing an associative product on the symmetric space over $L$ is one implication. The free cocommutative coalgebra endowed with this associative product turns out to be a Hopf algebra that is isomorphic to the universal enveloping algebra of $L$. This can be seen as exploiting the pre-Lie structure to obtain extra information about the universal envelope of $L$.
\end{remark}
We also give the definition of a multi-pre-Lie algebra first introduced in \cite{BCCH}. Although a seemingly richer structure, all the information can be condensed into a single pre-Lie algebra. It is nonetheless a useful notion when describing certain families of products.
\begin{definition}
A multi-pre-Lie algebra indexed by a set $E$ is a vector space $\mathcal{P}$ over a field $\mathbf{k}$ of characteristic $0$, endowed with a family $(\curvearrowright^{\alpha})_{\alpha \in E}$ of bilinear products such that for every $ x, y,z \in \mathcal{P} $
$$
x \curvearrowright^{a} (y \curvearrowright^{b} z) - (x \curvearrowright^{a} y) \curvearrowright^{b} z = y \curvearrowright^{b} (x \curvearrowright^{a} z) - (y \curvearrowright^{b} x) \curvearrowright^{a} z.
$$
As is shown below (see \cite{Foi21}), can summarise all the data of a multi-pre-Lie algebra into a single pre-Lie algebra.
\end{definition}
\begin{lemma}
If $\mathcal{P}$ is a multi-pre-Lie algebra over a field $\mathbf{k}$ of characteristic $0$ and indexed by a set $E$, then $\mathcal{P} \otimes \mathbf{k}E$ is a pre-Lie algebra when endowed with the product
\begin{equs}
(x \otimes a) \curvearrowright (y \otimes b) = (x \curvearrowright^{a} y) \otimes b
\end{equs}
for any $a, b \in E$ and for any $x,y,z \in \mathcal{P}$.
\end{lemma}
\begin{example}\rm
\label{rooted_r} A family of pre-Lie products on $ \mathcal{T}^V_E $ is given by grafting by means of decorated edges, namely:
\begin{equation}
\sigma \curvearrowright^a \tau:=\sum_{v\in N_{\tau} } \sigma \curvearrowright^a_v \tau,
\end{equation}
\label{grafting_a}
where $\sigma $ and $\tau$ are two decorated rooted trees, $ N_\tau $ is the set of vertices of $ \tau $ and where $\sigma \curvearrowright^a_v \tau$ is obtained by grafting the tree $\sigma$ on the tree $\tau$ at vertex $v$ by means of a new edge decorated by $a\in E$.
\end{example}
Another example is a deformed version of the above family of grafting products.
\begin{example}
We suppose here that the vertices are decorated by elements of a monoid $\Omega$ and that $\Omega = \N^{d+1}$ here, endowed with componentwise addition. The operators $\uparrow_{v}^{\omega}$ act by adding $\omega$ to the decoration $\mathfrak{n}_{v}$ of a given tree. A grading is given by
\begin{equs}
|\mathbf n|_{\mathfrak{s}}:=s_1n_1+\cdots +s_{d+1}n_{d+1}
\end{equs}
where $\mathfrak s:=(s_1,\ldots, s_{d+1})\in \N_{>0}^{d+1}$ is fixed. We suppose that $V = S \times \N^{d+1}$ and $E = S' \times \N^{d+1}$ where $ S $ and $ S' $ are two finite sets. Then $\Omega$ acts freely on both $E$ and $V$ in a graded way. We denote by $+$ the addition in $\Omega$ as well as both actions of $\Omega$ on $E$ and $V$. A family of deformed grafting products on $\mathcal T_E^V$ is defined as follows:
\begin{equation} \label{deformation_preLie}
\sigma \widehat{\curvearrowright}^a \tau:=\sum_{v\in N_{\tau}}\sum_{\ell\in\N^{d+1}}{\mathfrak{n}_v \choose \ell} \sigma \curvearrowright_v^{a-\ell}(\uparrow_v^{-\ell} \tau).
\end{equation}
\label{deformed_grafting_a}
Here $ \mathfrak{n}_v \in\N^{d+1}$ denotes the second component of the decoration at the vertex $ v $. The generic term is self-explanatory if there exists a (unique) pair $(b,\alpha)\in E\times V$ such that $a=\ell+b$ and $\mathfrak{n}_v =\ell+\alpha$. It vanishes by convention if this condition is not satisfied. We define the \textsl{grading} of a tree in $\mathcal T_E^V$ by the sum of the gradings of its edges given by $ |\cdot|_{\text{grad}} $:
\begin{equation} \label{grading}
|\tau|_{\text{grad}}:=\sum_{e\in E_{\tau}}\big|\mathfrak{e}(e) \big|_{\mathfrak{s}}.
\end{equation}
where $ E_{\tau} $ are the edges of $ \tau $ and $ \mathfrak{e}(e) $ is the decoration of the edge $ e $.
Then, $ \widehat{\curvearrowright}^a $ is a deformation of $ \curvearrowright^a $ in the sense that:
\begin{equs}
\sigma \widehat{\curvearrowright}^a \tau = \sigma \curvearrowright^a \tau +
\hbox{ lower grading terms}.
\end{equs}
\end{example}
We now proceed to give the definitions of pre-Lie products that will be of interest to us. We first define the grafting product, a pre-Lie product on planted trees with edge and vertex decorations that gives rise to the free multi-pre-Lie algebra equipped with a family of pre-Lie products over a prescribed set of generators. We will then define a deformed version of this grafting product that is of greater interest to us, which turns out to be isomorphic to the original grafting product.
\begin{definition}[Grafting product]
The grafting product on the space $\mathcal{P}_{E}^{V}$ is defined as follows for planted trees and is then extended by linearity:
\begin{equs} \label{def_pre_a}
\mathcal{I}_{a}(\sigma) \curvearrowright \mathcal{I}_{b}( \tau) = \mathcal{I}_{b}( \sigma \curvearrowright^{a} \tau),
\end{equs}
for any $a,b \in E$ and any $\tau, \sigma \in \mathcal{T}_{E}^{V}$.
\end{definition}
\begin{definition}[Deformed grafting product]
The deformed grafting product on the space $\mathcal{P}_{E}^{V}$ is defined as follows for planted trees and is then extended by linearity:
\begin{equs}
\quad \mathcal{I}_{a}(\sigma) \, \wh\curvearrowright \, \mathcal{I}_{b}( \tau) = \mathcal{I}_{b}( \sigma \wh\curvearrowright^{a} \tau)
\end{equs}
for any $a,b \in E$ and any $\tau, \sigma \in \mathcal{T}_{E}^{V}$.
\end{definition}
\begin{remark}
The grafting and deformed grafting products are clearly pre-Lie products under the identification of the space $\mathcal{P}_{E}^{V}$ with $\mathcal{T}_{E}^{V} \otimes \mathbf{k}E$ by identifying an element $\CI_{a}(\tau)$ with $\tau \otimes a$.
\end{remark}
We will now introduce a deformed version of the original Butcher-Connes-Kreimer coproduct, which we call the deformed Butcher-Connes-Kreimer (DBCK) coproduct and denote by $\Delta_{\text{\tiny{DBCK}}}$.
\begin{definition}
We suppose that the set $V$ of vertex decorations coincides with the commutative monoid $\Omega$. Then, the deformed Butcher-Connes-Kreimer coproduct is defined by the maps $ \Delta_{\text{\tiny{DBCK}}} : S(\mathcal{P}_E^V ) \rightarrow S(\mathcal{P}_E^V ) \otimes S(\mathcal{P}_E^V )$ and $ \bar{\Delta}_{\text{\tiny{DBCK}}} : \mathcal{T}_E^V \rightarrow S(\mathcal{P}_E^V ) \otimes \mathcal{T}_{E}^V $ \label{delta_dck} \label{delta_dck_bar} defined recursively by:
\begin{equs} \label{recur_def_deltaDCK}
\begin{aligned}
\Delta_{\text{\tiny{DBCK}}} \CI_a(\tau) & =
\left( \mathrm{id} \otimes \CI_a \right) \bar{\Delta}_{\text{\tiny{DBCK}}} \tau + \CI_a(\tau) \otimes \mathbf{1},
\quad \bar{\Delta}_{\text{\tiny{DBCK}}} X^k = \mathbf{1} \otimes X^{k}
\\ \bar{\Delta}_{\text{\tiny{DBCK}}} \CI_a(\tau) & = \left( \mathrm{id} \otimes \CI_a \right) \bar{\Delta}_{\text{\tiny{DBCK}}} \tau + \sum_{\ell \in \N^{d+1}} \frac{1}{\ell !} \CI_{a + \ell}(\tau) \otimes X^{\ell}.
\end{aligned}
\end{equs}
The map $ \Delta_{\text{\tiny{DBCK}}} $ is extended using the product of $ S(\mathcal{T}_E^V) $. We use the tree product for extending the map $ \bar{\Delta}_{\text{\tiny{DCK}}} $.
The infinite sum over $ \ell $ makes sense via a bigrading introduced in \cite[Section 2.3]{BHZ}.
\end{definition}
\begin{definition}
The Deformed Butcher-Connes-Kreimer (DBCK) Hopf algebra $\mathcal{H}_{\text{\tiny{DBCK}}}$ is the graded bialgebra on $\CF_{E}^{V} = S(\mathcal{P}_E^V )$ equipped with the forest product (i.e. the product of the symmetric algebra) and the $\Delta_{\text{\tiny{DBCK}}}$ coproduct. As a graded, connected bialgebra, it is also a Hopf algebra.
\end{definition}
For a Hopf algebra $\mathcal{H}$ with coproduct $\Delta$ and a given element $p \in \mathcal{H}$, recall that $p$ is primitive if $\Delta(p) = \mathbf{1} \otimes p + p \otimes \mathbf{1}$. We shall denote the space of primitive elements of $\mathcal{H}$ by $Prim(\mathcal{H})$. Note that $Prim(\mathcal{H})$ is a linear subspace of $\mathcal{H}$. Equipping $Prim(\mathcal{H})$ with the commutator Lie bracket $[h_{1}, h_{2}] = h_{1}h_{2} - h_{2}h_{1}$ it is also a Lie algebra.
It is well-known that the primitive elements of a Hopf Algebra carry the structure of a Lie algebra. When $\mathcal{H}$ is a cocommutative graded connected Hopf algebra with finite-dimensional graded components, the Milnor-Moore theorem tells us that $\mathcal{H} \cong U(Prim(\mathcal{H}))$, i.e. that $\mathcal{H}$ is isomorphic as a Hopf algebra to the universal enveloping algebra over it's primitives. When $\mathcal{H}$ is cofree-cocommutatve and right-sided, the primitive elements of $\mathcal{H}$ admit a finer structure, that of a pre-Lie algebra. An explicit description of any Hopf algebra obeying these conditions via means of it's underlying pre-Lie algebra is given by the Guin-Oudom procedure \cite{Guin1,Guin2}, which gives a recursive construction of the algebra's associative product on the symmetric algebra $S(Prim(\mathcal{H}))$ over the primitives:
\begin{theorem}[Guin-Oudom]
\label{Guin-Oudom}
Let $(\mathcal{P}, \curvearrowright )$ be a pre-Lie algebra and let $S(\mathcal{P})$ denote the symmetric space over the underlying vector space.
For every $u,v, w \in S(\mathcal{P})$, $x, y \in \mathcal{P}$ we start by defining a product $\bullet$ on $S(\mathcal{P})$ as follows:
\begin{equs}\label{e::symmetric product}
\begin{aligned}
\mathbf{1} \bullet w&= w, \quad
u\bullet \mathbf{1} =\mathbf{1}^{\star}(u),\\
w \bullet uv &=\sum_{(w)} ( w^{(1)} \bullet u)(w^{(2)} \bullet v),\\
x v \bullet y &= x \, \curvearrowright \, ( v\bullet y) - (x \, \curvearrowright \, v) \bullet y
\end{aligned}
\end{equs}
where $\mathbf{1}^\star$ stands for the counit, the summation over $(w)$ is shorthand for summing over the terms of the expansion for the shuffle coproduct $\Delta_{\shuffle}$ and where $ \curvearrowright $ is extended to $ \mathcal{P} \otimes S(\mathcal{P}) $ in the following way:
\begin{align*}
x \, \curvearrowright \, x_1\ldots x_k
&=\sum_{i=1}^k x_1\ldots (x \, \curvearrowright \, x_i)\ldots x_k,
\end{align*}
with $ x_i \in \mathcal{P} $. \label{star_0}
We now define the associative product $ \star $ as follows:
\begin{equs} \label{def_*}
w \star v = \sum_{(w)}\left( (w^{(1)} \bullet v) w^{(2)} \right)
\end{equs}
Then, the associative product $\star$ on $S(\mathcal{P})$ is such that the Hopf algebra $(S(\mathcal{P}), \star, \Delta_{\shuffle})$ where $\Delta_{\shuffle}$ denotes the shuffle coproduct, is isomorphic to the universal enveloping algebra $U(\mathcal{P})$ of the Lie Algebra associated to $\mathcal{P}$, equipped with it's standard Hopf-algebraic structure. Furthermore, the induced mapping from the category $\catname{PreLie}$ to the category $\catname{Hopf}$ of Hopf algebras is a functor. A morphism $\phi$ in $\catname{PreLie}$ is mapped to $S(\phi)$ where $S$ is the symmetric space functor.
\end{theorem}
Given a pre-Lie algebra $\mathcal{P}$ the Guin-Oudom procedure describes a way to impose a Hopf Algebra structure on $S(\mathcal{P})$ by using the pre-Lie product to obtain an associative product on $S(\mathcal{P})$. This turns out to be isomorphic to $U(\mathcal{P})$. In fact, one obtains a functor from the category $\catname{PreLie}$ to the category $\catname{Hopf}$ of Hopf algebras, see the proof of Proposition 3.1 in \cite{Guin2}. Furthermore, Loday and Ronco, in \cite{LR} prove that this mapping is an equivalence of categories from $\catname{PreLie}$ to a certain category $\catname{CHA}$ of cofree-cocommutative right-sided combinatorial Hopf Algebras. For example, under this correspondence, the free pre-Lie algebra on one generator gives rise to the Grossman-Larson Hopf algebra.
\begin{theorem}\label{Free pre-Lie -> Grossman - Larson}
The grafting pre-Lie algebra with product $\curvearrowright$ and edge decorations from the set $E$ is the free-pre Lie algebra over $E$. Furthermore the product obtained by the Guin-Oudom construction above is the Grossman-Larson product and therefore $\mathcal{H}_{\text{\tiny{GL}}} = (S(\mathcal{P}_{E}^{V}), \star, \Delta_{\shuffle})$ is isomorphic to the universal enveloping algebra $U(\mathfrak{h})$, where $\mathfrak{h}$ is the Lie algebra induced by the grafting product $\curvearrowright$.
\end{theorem}
We also have analogous results for the deformed grafting product:
\begin{theorem}\label{Deformed Grafting0}
The Hopf algebra $\mathcal{H}_{\text{\tiny{DGL}}}$ = $(S(\mathcal{P}_E^V), \tilde{\star}, \Delta_{\shuffle})$, where $\tilde{\star}$ is obtained from $\widehat{\curvearrowright}$ via the Guin-Oudom construction given above, is isomorphic to the universal enveloping algebra $U(\mathfrak{g})$, where $\mathfrak{g}$ is the commutator Lie algebra induced by $\widehat{\curvearrowright}$.
\end{theorem}
We shall denote by $\mathcal{H}_{\text{\tiny{DGL}}}$ the deformed Grossman-Larson Hopf algebra. This is due to the following theorem \cite[Theorem 3.4]{BM22}:
\begin{theorem
The product $\tilde{\star}$ is dual to the deformed Butcher-Connes-Kreimer coproduct $\Delta_{\text{\tiny{DBCK}}}$.
\end{theorem}
In \cite[Theorem 2.7]{BM22}, the authors prove that there exists an isomorphism $\Theta$ between the pre-Lie algebra $E_{\text{\tiny{DGL}}}$ associated with the deformed Grossman-Larson Hopf algebra and the pre-Lie algebra $E_{\text{\tiny{GL}}}$ associated with the original Grossman-Larson Hopf algebra. One can describe recursively the isomorphism $\Theta$ by
\begin{equs} \label{recursive_theta}
\begin{aligned}
\Theta\left( \mathcal{I}_a(X^k) \right) & = \mathcal{I}_a(X^k) \\
\Theta \left( \mathcal{I}_a(X^k \prod_{i=1}^n \mathcal{I}_{a_i}(\tau_i)) \right)
& = \prod_{i=1}^n \Theta \left( \mathcal{I}_{a_i}(\tau_i)) \right) \, \widehat{\curvearrowright} \, \mathcal{I}_a(X^k).
\end{aligned}
\end{equs}
Using the isomorphism $\Theta$, together with Theorem~\ref{Guin-Oudom}, we can now prove:
\begin{theorem} \label{theorem_functor}
The deformed Grossman-Larson Hopf algebra $\mathcal{H}_{\text{\tiny{DGL}}}$ is isomorphic as a Hopf algebra to the original Grossman-Larson Hopf algebra $\mathcal{H}_{\text{\tiny{GL}}}$.
\end{theorem}
\begin{proof}
Let $\Theta$ denote the isomorphism between the pre-Lie Algebras associated to $\mathcal{H}_{\text{\tiny{GL}}}$ and $\mathcal{H}_{\text{\tiny{DGL}}}$. We let $G: \catname{PreLie} \rightarrow \catname{Hopf}$ denote the Guin-Oudom functor. Then, $\Phi := G(\Theta)$ is an isomorphism between $\mathcal{H}_{\text{\tiny{GL}}}$ and $\mathcal{H}_{\text{\tiny{DGL}}}$ in the category of Hopf algebras.
\end{proof}
Then, our previously defined isomorphism $\Phi: \mathcal{H}_{\text{\tiny{GL}}} \rightarrow \mathcal{H}_{\text{\tiny{DGL}}}$ induces a linear isomorphism of the space of primitive elements for $\mathcal{H}_{\text{\tiny{CK}}}$, seen as a subspace of $\mathcal{H}_{\text{\tiny{GL}}}$, onto it's image. This allows us to prove the following theorem:
\begin{proposition} \label{prop_iso_tilde}
Let $\sigma \in \mathcal{H}_{\text{\tiny{DGL}}}$. Then, there exists a set of elements $\mathcal{B} = \{\sigma_{1}, \sigma_{2}, ... \} \subseteq \mathcal{H}_{\text{\tiny{DGL}}}$ such that
\begin{equs}
\sigma = \sum_{R} \lambda_{R} \sigma_{r_{1}} \tilde{\star} ... \tilde{\star} \sigma_{r_{n}}
\end{equs}
with a unique such decomposition.
\end{proposition}
\begin{proof}
We pick the unique $\tau$ such that $\sigma = \Phi(\tau)$. We know that $\tau = \sum_{R} \lambda_{R}\tau_{r_{1}} \star ... \star \tau_{r_{n}}$ for some primitive elements $\tau_{r_{i}}$. Hence,
\begin{equs}
\sigma = \Phi(\tau) = \sum_{R} \lambda_{R}\Phi(\tau_{r_{1}}) \tilde{\star} ... \tilde{\star} \Phi(\tau_{r_{n}})
\end{equs}
We then pick $\sigma_{r_{i}} = \Phi(\tau_{r_{i}})$. This completes the proof.
\end{proof}
From the previous proposition, given the basis $ \mathcal{B} $, the basis one can use for $ \tilde{\star} $ is $ \Phi(\mathcal{B}) = \langle\Phi(\tau_1), \Phi(\tau_2),...\rangle $.
Then, one can exhibit an isomorphism $ \Psi_{\Phi} $ between the two spaces $ S(\mathcal{P}_{E}^{V}) $ and $ T(\Phi(\mathcal{B})) $ based on the basis $ \Phi(\mathcal{B})$:
\begin{equs}
\Psi_{\Phi} : \Phi(\tau_1) \ \tilde{\star} \ldots \tilde{\star} \ \Phi(\tau_r) \mapsto
\Phi(\tau_1) \otimes \ldots \otimes \Phi(\tau_r)
\end{equs}
where $ \Phi(\tau_1) \, \otimes \ldots \otimes \, \Phi(\tau_n) \in T(\Phi(\mathcal{B})) $. As a corollary, by composition of isomorphisms, we obtain:
\begin{theorem}\label{Main}
There exists a subspace $\mathcal{B} = \langle\tau_1, \tau_2,...\rangle$ of $\mathcal{P}_{E}^{V} $ such that $\mathcal{H}_{\text{\tiny{DGL}}}$ is isomorphic as a Hopf algebra to the tensor Hopf algebra $(T(\mathcal{B}), \otimes, \Delta_{\shuffle})$ endowed with the tensor product and the shuffle coproduct.
\end{theorem}
\section{Extension of the Chapoton-Foissy isomorphism}
\label{section::4}
\ \ \ \ \ \ In this section, we prove our main result, Theorem~\ref{main_result_paper}, which asserts that the $\mathcal{H}_{2}$ Hopf algebra, used in the context of regularity structures for recentering the ensuing Taylor-type expansions around different points, is actually isomorphic to a simple quotient of the tensor Hopf algebra. This quotient comes from the Lie bracket between planted trees and extra elements $ X_i$ which are parts of $ \mathcal{H}_2 $ but were absent in the previous section. The main difficulty is to check that the basis given by Theorem~\ref{Main} is stable under this quotient. This is proved in Proposition~\ref{stable_ideal} and relies on properties of the two derivations $\uparrow^i$ and $\mathcal{D}^i$. They commute with the isomorphism $\Psi$, as shown in Proposition~\ref{conjecture_psi} and leave invariant the primitives of the Butcher-Connes-Kreimer Hopf algebra, see Corollary~\ref{primitive_BCK}. The formulation of the main result and its proof rely strongly on the post-Lie algebras introduced in \cite{BK} for describing $ \mathcal{H}_2 $. Finally, we give a non-trivial extension of the Chapoton-Foissy isomorphism. This is a consequence of having better understood the two main algebraic components at play in the context of the Regularity Structures currently used, which are the deformation and the post-Lie structure.
We shall begin by introducing the concept of a post-Lie algebra, which generalizes that of a pre-Lie algebra. We also describe the recursive construction of an associative product on the universal envelope of a post-Lie algebra that directly generalizes the construction of Guin and Oudom. It was first introduced in \cite{ELM}.
\begin{definition} \label{definition_post_lie}
A post-Lie algebra is a Lie algebra $ (\mathfrak{g}, [.,.]) $ equipped with a bilinear product $ \triangleright $ satisfying the following identities:
\begin{equs} \label{ident1}
\begin{aligned}
x \triangleright [y,z] & = [x \triangleright y,z] + [y, x \triangleright z] \\
[x,y] \triangleright z & = a_{\triangleright}(x,y,z) - a_{\triangleright}(y,x,z)
\end{aligned}
\end{equs}
with $ x,y,x \in \mathfrak{g} $ and the commutator $ a_{\triangleright}(x,y,z) $ is given by:
\begin{equs}
a_{\triangleright}(x,y,z) = x \triangleright ( y \triangleright z ) - (x \triangleright y) \triangleright z.
\end{equs}
\end{definition}
When $ (\mathfrak{g}, [.,.]) $ is the abelian Lie algebra, we obtain the notion of a pre-Lie algebra. One can define a new Lie bracket $[[ .,.]]$ given by:
\begin{equation} \label{bracket}
[[x,y]] = [x,y] + x \triangleright y - y \triangleright x.
\end{equation}
The post-Lie product $ \triangleright $ can be extended to a product on the universal enveloping algebra $ U(\mathfrak{g}) $ by first defining it on $ \mathfrak{g} \otimes U(\mathfrak{g})$:
\begin{equs}
x \triangleright \mathbf{1} = 0, \quad x \triangleright y_1 ... y_n = \sum_{i=1}^n y_1 ... (x \triangleright y_i) ... y_n.
\end{equs}
and then extending it to $ U(\mathfrak{g}) \otimes U(\mathfrak{g}) $ by defining:
\begin{equs}
\mathbf{1} \triangleright A & = A, \quad
x A \triangleright y = x \triangleright (A \triangleright y) -(x \triangleright A) \triangleright y, \\
A \triangleright B C & = \sum_{(A)} (A^{(1)} \triangleright B)(A^{(2)} \triangleright C).
\end{equs}
where $ A, B, C \in U(\mathfrak{g}) $ and $ x, y \in \mathfrak{g} $. Here, $ (A) $ correspond to the shuffle coproduct. One defines an associative product $*$ on $ U(\mathfrak{g})$, the universal enveloping algebra of $\mathfrak{g}$:
\begin{equs} \label{product_1}
A * B = \sum_{(A)} A^{(1)} (A^{(2)} \triangleright B).
\end{equs}
Then, a result that generalizes that of Guin and Oudom, allows us to exploit the underlying post-Lie structure on $\mathfrak{g}$ in order to gain additional insight on the structure of $U(\mathfrak{g})$. This is formalised in the following theorem:
\begin{theorem} \label{main_theorem_section_2}
The Hopf algebra $ (U(\mathfrak{g}),*,\Delta) $ is isomorphic to the enveloping
algebra $ U(\bar{\mathfrak{g}}) $ where $ \bar{\mathfrak{g}} $ is the Lie algebra equipped with the Lie bracket $ [[.,.]] $.
\end{theorem}
This result has been used in \cite{BK}, in the context of Regularity Structures, in order to show that the $\star_{2}$ product, dual to the $\Delta_{2}$ coproduct appearing in \cite{BM22} and introduced in \cite{reg,BHZ}, comes directly from a post-Lie product by applying the above procedure. Below, we briefly recall this.
We define the following spaces:
\begin{equs}
\mathcal{V} & = \Big \langle \{ \CI_a(\tau), \, a \in \mathbb{N}^{d+1}, \, \tau \in \mathcal{T}^V_E \} \cup \{ X_{i} \}_{i = 0,..., d} \Big \rangle_{\mathbb{R}}, \\
\tilde{\mathcal{V}} & = \Big \langle \CI_a(\tau), \, a \in \mathbb{N}^{d+1}, \, \tau \in \mathcal{T}^V_E \Big \rangle_{\mathbb{R}}.
\end{equs}
We denote by $\uparrow^{k}_{v}$ the operator acting on decorated trees by adding $k$ to the decoration of the node $v$.
We then define, for a tree $\tau \in \mathcal{T}_{E}^{V}$ the operator $ \uparrow^{i} $ as follows:
\begin{equs}
\uparrow^{i} \tau = \sum_{v \in N_{\tau}} \uparrow^{e_i}_v \tau.
\end{equs}
This operator acts as a derivation on the multi-pre-Lie algebra of grafting products in the sense that:
\begin{equs} \label{derivation_node}
\uparrow^{i} \left( \sigma \curvearrowright^a \tau \right) = (\uparrow^{i} \sigma) \curvearrowright^a \tau + \sigma \curvearrowright^a \, ( \uparrow^{i} \tau)
\end{equs}
The derivation property \eqref{derivation_node} is not preserved under the deformation. One has the following identity similar to \cite[Proposition 4.4]{BK}.
\begin{equs} \label{non_commutation}
\uparrow^{i} \left( \sigma \widehat{\curvearrowright}^a \tau \right) = ( \uparrow^{i} \sigma ) \, \widehat{\curvearrowright}^a \, \tau + \sigma \, \widehat{\curvearrowright}^a \, ( \uparrow^{i} \tau) - \sigma \, \widehat{\curvearrowright}^{a-e_i} \tau,
\end{equs}
for all decorated trees $ \sigma, \tau $ and $ a \in \mathbb{N}^{d+1} $, $ i \in \lbrace 0,...,d \rbrace $. Looking at the above formula, one observes that the pair of operators $\tau \mapsto \sigma \widehat{\curvearrowright}^a \tau$ and $\tau \mapsto \uparrow^{i} \tau$ does not satisfy the commutativity relation satisfied by the operators $\tau \mapsto \sigma \curvearrowright^{a} \tau$ and $\tau \mapsto \uparrow^{i} \tau$. The non-commutative relation \ref{non_commutation} motivates the introduction of a Lie bracket together with a product that is a derivation for that bracket, that encode these relations in the form of a post-Lie algebra.
We begin by introducing a product $ \widehat{\triangleright} $ on $\mathcal{V}$:
\begin{equs} \label{prodcut_post_Lie}
X_i \, \widehat{\triangleright} \, \CI_{a}(\tau) & = \CI_{a}( \uparrow^i \tau), \quad \CI_{a}(\tau) \, \widehat{\triangleright} \, X_{i} = X_i \, \widehat{\triangleright} \, X_{j} = 0
\\
\CI_{a}(\sigma) \, \widehat{\triangleright} \, \CI_{b}(\tau) & = \CI_{a}(\sigma) \, \widehat{\curvearrowright} \, \CI_{b}(\sigma).
\end{equs}
We now proceed to define the appropriate Lie bracket, motivated by \ref{non_commutation}:
\begin{definition} \label{def_lie_trees}
We define the Lie bracket on $\mathcal{V}$ as $[x, y]_{0} = 0$ for $x, y \in \tilde{\mathcal{V}}$, $[x, y]_{0} = 0$ for $x, y \in \langle \ X_{i} \ \rangle_{\mathbb{R}}$ and as
\begin{equation}
[\CI_a(\tau),X_i]_0 = \CI_{a-e_i}(\tau)
\end{equation}
\end{definition}
With these definitions at hand, we have the following theorem (see \cite[Theorem 4.4]{BK}):
\begin{theorem}
The triple $( \mathcal{V}, [.,.]_{0}, \widehat{\triangleright})$ is a post-Lie algebra.
\end{theorem}
The bracket induced by the post-Lie algebra encodes all the (non-)commutativity relations between operators acting on decorated trees. However, most of these actually commute with one another, forming a pre-Lie algebra that lives inside the Lie algebra $\tilde{V}$. The extra post-Lie structure allows one to, roughly speaking, split the bracket into a commutative and non-commutative part. Hence the non-commutativity relations are actually encoded more succinctly by the $[,]_{0}$ bracket.
We denoted by $ U(\mathcal{V}_0) $ the enveloping algebra with the Lie bracket $ [.,.]_0 $ and by $ U(\mathcal{V}) $ the enveloping algebra with the Lie bracket $ [[.,.]] $. We also set $ * $ to be the product obtained by the generalization of the Guin-Oudom procedure given in \eqref{product_1}. As a mere application of Theorem~\ref{main_theorem_section_2}, one gets
\begin{theorem} \label{main_result_trees}
The Hopf algebra $U(\mathcal{V})$ is isomorphic to the Hopf algebra $(U(\mathcal{V}_0), *, \Delta)$.
\end{theorem}
Then, the main result of \cite{BK} is
\begin{theorem}
The Hopf algebra $(U(\mathcal{V}_0), *, \Delta)$ is isomorphic to the Hopf algebra $\mathcal{H}_{2} = (\mathcal{T}_{E}^{V}, \star_{2}, \Delta)$ as presented in \cite{BM22}.
\end{theorem}
\begin{remark}
An explicit formula for the $\star_{2}$ product for $\sigma = X^{k} \prod_{i \in I} \CI_{a_{i}}(\sigma_{i})$ and $\tau \in \mathcal{T}_{E}^{V}$ is given by
\begin{equs}
\CI_{b}(\sigma \star_{2} \tau) := \tilde{\uparrow}_{N_{\tau}}^{k} \( \prod_{i \in I} \CI_{a_{i}}(\sigma_{i}) \, \widehat{\curvearrowright} \, \CI_{b}(\tau) \),
\quad
\tilde{\uparrow}_{N_{\tau}}^{k} = \sum_{k = \sum_{v \in N_{\tau}}k_{v} } \uparrow_{v}^{k_{v}}
\end{equs}
\end{remark}
We shall now decompose trees of the form $ X^k \prod_i \mathcal{I}_{a_i}(\tau_i) $ with decoration $ k $ at the root. So far, we have been successful in doing this for trees with no root decoration. For these terms, we will need to utilize the underlying post-Lie structure and the fact that $X_{i}$ does not comute with any term of the form $ \mathcal{I}_{a}(\tau) $. Instead one has:
\begin{equs}
X_i \star_2 \mathcal{I}_{a}(\tau) - \mathcal{I}_{a}(\tau) \star_2 X_i = \uparrow_{e_i} \mathcal{I}_{a}(\tau) - \mathcal{I}_{a - e_i}(\tau)
\end{equs}
where $ \star_2 $ is the product constructed from the post-Lie product. The restriction of this product on the space spanned by planted trees coincides with $ \tilde{\star} $. What we obtain will then be an isomorphism with a space of words quotiented by the following relation:
\begin{equs}
X_i \otimes \mathcal{I}_{a}(\tau) - \mathcal{I}_{a}(\tau) \otimes X_i = \uparrow_{e_i} \mathcal{I}_{a}(\tau) - \mathcal{I}_{a - e_i}(\tau)
\end{equs}
where now the trees with a single node are treated as letters. Let us explain how this works for a decorated of the form $ X^k \prod_{i=1}^n I_{a_i}(\tau_i) $ when one wants to decompose the following terms:
\begin{equs}
X^k \prod_{i=1}^n \CI_{a_i}(\tau_i) \star_2 \tau
\end{equs}
We begin by making the following remarks that shall prove useful in what follows:
\begin{itemize}
\item By choosing a different ordering in the Poincare-Birkhoff-Witt theorem, we clearly see that the set of elements of the form
\begin{equs}
\prod_{i=1}^n \CI_{a_i}(\tau_i)X^{k}
\end{equs}
where $\textbf{F} = \CI_{a_{1}}(\tau_{1}) \cdot \cdot \cdot \CI_{a_{n}}(\tau_n) $ ranges over all forests of planted trees and $\textbf{m} \in \mathbb{N}^{d+1}$ is a basis for $U(\mathcal{V}_0)$.
\vspace{10pt}
\item The operator $\tau \mapsto
\prod_{i=1}^n \CI_{a_i}(\tau_i)X^{k} \star_{2} \tau$
is equal to the operator $\tau \mapsto
\prod_{i=1}^n \CI_{a_i}(\tau_i) \star_{2} X^{k} \star_{2} \tau$
\end{itemize}
We introduce a second derivation $ \mathcal{D}^{i} $ defined by
\begin{equs}
\mathcal{D}^{i} \tau = \sum_{e \in E_\tau}
\mathcal{D}^{i}_e \tau
\end{equs}
where $ \mathcal{D}^{i}_e $ adds $ -e_i$ to the decoration of the edge $ e $ if possible. Otherwise, it is equal to zero.
\begin{proposition} \label{conjecture_psi}
For every $ \tau \in \mathcal{P}^V_E $, one has:
\begin{equs}
\uparrow^{i} \Phi(\tau) = \Phi( \uparrow^{i} \tau ) - \Phi( \mathcal{D}^{i} \tau)
\end{equs}
\end{proposition}
\begin{proof}
We first consider a decorated tree $ \tau $ of the form
\begin{equs}
\tau = \tau_{1} \curvearrowright^{b} \mathcal{I}_a(\tau_{2})
\end{equs}
then one has
\begin{equs}
\uparrow^{i} \Phi(\tau) & =
\uparrow^{i} \left( \Phi(\tau_1) \, \widehat{\curvearrowright}^{b} \, \Phi(\mathcal{I}_a(\tau_{2})) \right)
\\ & = ( \uparrow^{i} \Phi(\tau_1) ) \, \widehat{\curvearrowright}^{b} \, \mathcal{I}_a( \Phi(\tau_{2})) + \Phi(\tau_1) \, \widehat{\curvearrowright}^{b} \, (\uparrow^{i} \mathcal{I}_a( \Phi(\tau_{2}))) \\ & - \Phi(\tau_1) \, \widehat{\curvearrowright}^{b-e_i} \, \mathcal{I}_a( \Phi(\tau_{2}))
\end{equs}
where we have used \eqref{non_commutation}. Then, one can apply an induction hypothesis on $ \mathcal{I}_b(\tau_1) $ and $ \mathcal{I}_a(\tau_2) $ and one gets
\begin{equs}
\uparrow^{i} \Phi(\mathcal{I}_b(\tau_1)) & = \Phi(\mathcal{I}_b( \uparrow^{i} \tau_1 )) - \Phi( \mathcal{I}_b(\mathcal{D}^{i} \tau_1)) \\
\uparrow^{i} \Phi(\mathcal{I}_a(\tau_2)) & = \Phi(\mathcal{I}_a( \uparrow^{i} \tau_2 )) - \Phi( \mathcal{I}_a(\mathcal{D}^{i} \tau_2))
\end{equs}
Then, one observes that
\begin{equs}
\Phi( \mathcal{I}_a(\mathcal{D}^{i} \tau)) & = \Phi(\tau_1 \curvearrowright^{b-e_i} \mathcal{I}_a(\tau_{2}) + \Phi( \mathcal{D}^i \tau_1 \curvearrowright^{b} \mathcal{I}_a(\tau_{2}) + \Phi( \tau_1 \curvearrowright^{b} \mathcal{I}_a(\mathcal{D}^i \tau_{2})
\\ \Phi(\mathcal{I}_a( \uparrow^{i} \tau )) & = \Phi( \uparrow^{i} \tau_1 \curvearrowright^{b} \mathcal{I}_a(\tau_{2})) + \Phi( \tau_1 \curvearrowright^{b} \mathcal{I}_a( \uparrow^{i} \tau_{2}))
\end{equs}
We conclude by using again the morphism property of $ \Phi $ that gives us for example:
\begin{equs}
\Phi( \uparrow^{i} \tau_1 \curvearrowright^{b} \mathcal{I}_a(\tau_{2})) = \Phi( \uparrow^{i} \tau_1) \, \widehat{\curvearrowright} \, \Psi(\mathcal{I}_a(\tau_{2}))
\end{equs}
and the fact that $ \tau $ is generated by the family $ (\widehat{\curvearrowright}^{b})_{b} $.
\end{proof}
\begin{proposition} \label{commutattion_derivations}
One has the following commutation identities:
\begin{equs}
\Delta_{\text{\tiny{BCK}}}\uparrow^i & = \left( \uparrow^i \otimes \mathbf{1} \right) \Delta_{\text{\tiny{BCK}}} + \left( \mathbf{1} \otimes \uparrow^i \right) \Delta_{\text{\tiny{BCK}}}
\\ \Delta_{\text{\tiny{BCK}}}\mathcal{D}^i & = \left( \mathcal{D}^i \otimes \mathbf{1} \right) \Delta_{\text{\tiny{BCK}}} + \left( \mathbf{1} \otimes \mathcal{D}^i \right) \Delta_{\text{\tiny{BCK}}}
\end{equs}
with the convention that $ \mathcal{D}^i \mathbf{1} = \uparrow^i \mathbf{1} = 0 $.
\end{proposition}
\begin{proof}
This is just a consequence of the fact that $ \uparrow^i $ and $ \mathcal{D}^i $ are derivations for $ \curvearrowright^a $ and therefore for the Grossman-Larson product $ \star $. By going to the dual, one gets the desired identities.
\end{proof}
\begin{corollary} \label{primitive_BCK}
The set of primitives elements for $\mathcal{H}_{\text{\tiny{BCK}}}$ is stable under the action of the derivations $\uparrow^{e_{i}}$ as well as the derivations $\mathcal{D^i}$.
\end{corollary}
\begin{proof}
Let $ \tau $ a primitive elements, one has
\begin{equs}
\Delta_{\text{\tiny{BCK}}} \uparrow^i \tau = \left( \uparrow^i \otimes \mathbf{1} \right) \Delta_{\text{\tiny{BCK}}} \tau + \left( \mathbf{1} \otimes \uparrow^i \right) \Delta_{\text{\tiny{BCK}}} \tau
\end{equs}
where we have used Proposition~\ref{commutattion_derivations} . Then, from the primitiveness of $ \tau $
\begin{equs}
\Delta_{\text{\tiny{BCK}}} \tau
= \tau \otimes \mathbf{1} + \mathbf{1} \otimes \tau
\end{equs}
which allows us to get using the fact that $ \uparrow^i \mathbf{1} = 0 $:
\begin{equs}
\Delta_{\text{\tiny{BCK}}} \uparrow^i \tau = \uparrow^i \tau \otimes \mathbf{1} +
\mathbf{1} \otimes \uparrow^i \tau
\end{equs}
The proof works as the same for $ \mathcal{D}^i $.
\end{proof}
\begin{proposition} \label{stable_ideal}
If $\sigma = \Psi(\tau)$ for some primitive element $\tau$ with respect to the $\Delta_{\text{\tiny{BCK}}}$ coproduct, then $\uparrow^{e_i}\sigma$ and $\mathcal{D}^i \sigma$ are also in the image of $Prim(\mathcal{H}_{\text{\tiny{BCK}}})$.
\end{proposition}
\begin{proof}
This is a consequence of Proposition~\ref{conjecture_psi} and Corollary~\ref{primitive_BCK}.
\end{proof}
We can now state and prove our main result:
\begin{theorem} \label{main_result_paper} Let $\mathcal{T}$ be the linear span of decorated trees equipped with two products: $ \tilde{\star} $ is the product dual to the deformed Butcher-Connes-Kreimer coproduct and $ \star_2 $ is the product of $\mathcal{H}_{2}$.
We let $W$ be the linear space of the words whose letters are from the alphabet $ A $ whose letters are the $X_i$ and $ \Phi(\CI_{a}(\tau))$ where $ \tau $ is a primitive element for $ \Delta_{\text{\tiny{BCK}}} $.
We define $ \tilde{W} $ as the quotient of $W$ by the Hopf ideal $ \mathcal{J} $ generated by the elements
\begin{equs}
\{ X_i \otimes \Phi( \mathcal{I}_{a}(\tau) ) - \Phi( \mathcal{I}_{a}(\tau) ) \otimes X_i - \uparrow^{i} \Phi( \mathcal{I}_{a}(\tau)) - \Phi(\mathcal{I}_{a - e_i}(\tau)) \}
\end{equs}
Then, there exists a Hopf algebra isomorphism $ \Psi $ between $ \mathcal{T} $ and $ \tilde{W} $ equipped with $ \star_2 $.
The map $ \Psi $ is given by
\begin{equs}
\Psi : \prod_{i=1}^n \CI_{a_i}(\tau_i) X^k \rightarrow \Psi(\prod_{i=1}^n \CI_{a_i}(\tau_i)) \otimes_{j=0}^{d} \otimes_{i=1}^{k_j} X_i.
\end{equs}
\end{theorem}
\begin{proof}
We first apply the isomorphism described in Proposition~\ref{prop_iso_tilde} on $\sigma = \prod_{i=1}^n \CI_{a_i}(\tau_i)$ by writing
\begin{equs}
\prod_{i=1}^n \CI_{a_i}(\tau_i) = \sum_{R} \lambda_{R} \, \sigma_{r_{1}} \star_{2} ... \star_{2} \sigma_{r_{n}}
\end{equs}
with $ \sigma_{r_i} \in A $. We then map $ \prod_{i=1}^n \CI_{a_i}(\tau_i) X^k $ as follows:
\begin{equs}
X^{k}\prod_{i=1}^{n} \CI_{a_{i}}(\tau_i) \mapsto \sum_{R} \lambda_{R} \sigma_{r_{1}} \otimes ... \otimes \sigma_{r_{n}} \otimes X_{1}^{\otimes k_{1}} \otimes ... \otimes X_{m}^{\otimes k_{m}}
\end{equs}
By virtue of Proposition~\ref{stable_ideal}, this clearly gives an isomorphism onto the Hopf algebra $\tilde{W}$. Indeed, given a letter $ \Phi(\mathcal{I}_a(\tau)) $, one has that $ \uparrow^i \Phi(\mathcal{I}_a(\tau)) $ and $ \Phi(\mathcal{I}_{a - e_i}(\tau)) $ are linear combination of letters of $ W $.
\end{proof}
\section{Applications in Regularity structures}
\label{section::5}
In this section we restrict ourselves to the setting that is more specific to the theory of regularity structures, specifically the structures first appearing in the works \cite{reg,BHZ}. For an introduction to the theory see \cite{FrizHai,EMS,BaiHos}. This involves considering a Hopf subalgebra of the $\mathcal{H}_{2}$ Hopf algebra, that consists of trees with branches of positive degree. We shall use the theorem proved in the previous section to embed this into the tensor Hopf algebra. This allows for an encoding of the iterated integrals appearing when solving the equations, in the form of words. We begin by defining the space:
\begin{equs}
T_{+} := \{ X^{k}\prod_{i=1}^{n} \CI_{a_{i}}(\tau_{i}) \ | \ \alpha( \CI_{a_{i}}(\tau_{i})) > 0, \, \tau \in T^V_E \}
\end{equs}
We also define $\mathcal{T}_{+}$ to be the linear span of $T_{+}$. Here, $ \alpha $ is a degree map computing a number associated to a tree. This corresponds of some kind of regularity of the stochastic integral associated to the tree. It takes into account the decoration on the edges that could both encode distributional noises or convolution with kernel that provides a smoothing effect via Schauder estimates. We refrain to give a precise definition that could be found in many works \cite{reg,BHZ}.
For each subcritical singular SPDE one constructs a Hopf subalgebra $\mathcal{T}^+_{R}$ of $\mathcal{T}_{+}$ by attaching a generating rule $R$ to the nonlinearity $F$ of the equation. The rule induces a recursive procedure that generates the entire Hopf subalgebra $\mathcal{T}_{R}^+$. This procedure may be thought of as formal Picard iteration. The resulting Hopf subalgebra is then used to describe the regularity structure for the given equation. In the next theorem, we denote by $ \cdot $ the product on $ \mathcal{T}_+ $.
\begin{theorem
\label{T_+ isomorphism}
The Hopf algebra $(\mathcal{T}_{+}, \star_{2}, \Delta_{\shuffle})$, which is the graded dual of $(\mathcal{T}_{+}, \cdot, \Delta_{2})$, is isomorphic to a Hopf subalgebra of $T(A)/\mathcal{J}$.
\end{theorem}
\begin{proof}
By Theorem~\ref{main_result_paper}, we have an isomorphism $\Phi: \mathcal{H}_{2} \rightarrow T(A)/\mathcal{J}$. By simply restricting $\Phi$ to $\mathcal{T}_{+}$ we obtain a Hopf algebra ismorphism of $\mathcal{T}_{+}$ onto its image.
\end{proof}
Let explain how this algebraic result allows to interpret Regularity Structures as some kind of geometric Rough Paths.
Let us be more specific about the main ideas of this paper. Solutions $ u $ of local subcritical singular stochastic partial equations (SPDEs) are locally described by
\begin{equs}
u(y) - u(x) = \sum_{\tau \in \mathcal{T}_R} u_{\tau}(x)(\Pi_x \tau)(y), \quad (\Pi_x \tau)(y) \lesssim |y-x|^{\alpha(\tau)}
\end{equs}
where $ x,y \in \mathbb{R}^{d+1} $, $ (\Pi_x \tau)(y) $ are stochastic iterated integrals recentred around the point $ x $ such that one has a behaviour close to $ x $ according to the degree of the given decorated tree. The $ u_{\tau}(x) $ are some kind of derivatives. Then, the theory of Regularity Structures provides a reexpansion map $ \Gamma_{xy} $ that allows us to move the recentring:
\begin{equs}
\Pi_y = \Pi_x \Gamma_{xy}
\end{equs}
The collection of these two maps $ (\Pi_x,\Gamma_{xy}) $ is what is referred to as a model \cite[Def. 3.1]{reg}. One important algebraic construction is to represent $ \Gamma_{xy} $ via a character $ \gamma_{xy} : \mathcal{T}_+ \rightarrow \mathbb{R}$ a multiplicative for the tree product. This description is given via a co-action $ \Delta : \mathcal{T} \rightarrow \mathcal{T} \otimes \mathcal{T}_+ $
\begin{equs} \label{chen_RS}
\Gamma_{xy} = \left( \mathrm{id} \otimes \gamma_{xy} \right) \Delta, \quad | \gamma_{xy}(\tau) | \lesssim | y-x |^{\alpha(\tau)}
\end{equs}
The character $ \gamma_{xy} $ can be viewed as extension of branched Rough Paths to the multidimentional case as $ x, y \in \mathbb{R}^{d+1} $. Moreover, it satisfies some Chen's relation:
\begin{equs}
\gamma_{xy} \star_2 \gamma_{yz} = \gamma_{xz}
\end{equs}
We denote the space of such maps as $ \bf{TM}^{\alpha} $ called $ \alpha$-Tree-indexed Models. Maps $ \gamma_{xy} $ defined as character on $ \Psi(\mathcal{T}_+) $ are $ \alpha$-Geometric Models denoted by $ \bf{GM}^{\alpha} $. They satisfies the following properties:
\begin{equs} \label{def_GM}
\gamma_{xy} \otimes \gamma_{yz} = \gamma_{xz},
\quad | \gamma_{xy}(\Psi(\tau)) | \lesssim | y-x |^{\alpha(\tau)}
\end{equs}
We could have used the terminology of Anisotropic Rough Paths but the characters are defined on a quotient of a tensor Hopf algebra and not the tensor Hopf algebra itself. One can rephrase our main algebraic result as:
\begin{theorem} \label{chiso_bis} Let $ X \in \bf{TM}^{\alpha} $, then $ \hat{X} := \Psi(X) \in \bf{GM}^{\alpha}$.
\end{theorem}
\begin{proof}
The analytical bounds are easily satisfies by realising that:
\begin{equs}
\langle \Psi(X)_{xy}, \Psi(\tau) \rangle
= \langle X_{xy}, \tau \rangle.
\end{equs}
The algebraic identities are such as Chen's relation are preserved by the map $ Psi $.
\end{proof}
\begin{remark} As in \cite{Br20}, one can investigate the action of the renormalisation on this construction by looking at maps $ M $ that are morphisms for the product $ \star_2 $ which are BPHZ renormalisation maps. One of the main issue is that $ \mathcal{T}_+ $ may not be stable under $ M $ due to the constraint imposed on the degree being positive. Extended decorations on trees have been introduced in \cite{BHZ} in order to guarantee that $ M $ is degree preserving in the sense that for any decorated tree $ \tau $. One can easily check that $ M $ commutes with $ \Phi $ and then it is possible to find a map $ \tilde{M} $ defined on $ T(A)/\mathcal{J} $ such that it will commute with $ \Psi $:
\begin{equs}
\tilde{M} \Psi = \Psi M.
\end{equs}
This will be an equivalent of Theorem 4.7 in \cite{Br20}.
\end{remark}
|
\section{Introduction}
The crystal bases introduced in
\cite{K1, L} are powerful
tools to study the representation theory of quantized enveloping algebras $U_q(\mathfrak{g})$
for symmetrizable Kac-Moody Lie algebras $\mathfrak{g}$.
Realizing the crystal bases as combinatorial objects like as Young tableaux, LS paths, Nakajima monomials \cite{K, KN, Lit2, Nj},
one can reveal skeleton structures of representations or negative part $U_q^-(\mathfrak{g})\subset U_q(\mathfrak{g})$.
We are interested in the crystal bases $B(\lambda)$ for irreducible integrable highest weight modules $V(\lambda)$ with a dominant integral weight $\lambda$
and the crystal bases $B(\infty)$ for the negative part $U_q^-(\mathfrak{g})$.
In \cite{N99, NZ}, {\it polyhedral realizations} of $B(\lambda)$ and $B(\infty)$ are introduced,
which realize $B(\lambda)$ and $B(\infty)$ as sets of integer points
in polytopes and polyhedral cones, respectively.
There are embeddings of crystals $\Psi_{\iota}:B(\infty)\hookrightarrow \mathbb{Z}^{\infty}_{\iota}$
and $\Psi_{\iota}^{(\lambda)}:B(\lambda)\hookrightarrow \mathbb{Z}^{\infty}_{\iota}\otimes R_{\lambda}$
associated with an infinite sequence $\iota$ from the index set $I=\{1,2,\cdots,n\}$ of $\mathfrak{g}$,
where $R_{\lambda}$ is a crystal, which has a single element (Example \ref{r-ex}) and
\[
\mathbb{Z}^{\infty}_{\iota}
=\{(\cd,a_r,\cd,a_2,a_1)| a_r\in\ZZ
\,\,{\rm and}\,\,a_r=0\,\,{\rm for}\,\,r\gg 0\}
\]
has a crystal structure associated with $\iota$.
Then procedures to compute
linear inequalities defining
${\rm Im}(\Psi_{\iota}) (\cong B(\infty))$
and ${\rm Im}(\Psi^{(\lambda)}_{\iota}) (\cong B(\lambda))$ are provided.
It is most fundamental problem in the theory of polyhedral realizations
to find explicit forms of
${\rm Im}(\Psi_{\iota})$ and ${\rm Im}(\Psi^{(\lambda)}_{\iota})$.
The polyhedral realizations for $B(\infty)$ are deeply related to the {\it string cones} in \cite{BZ, Lit}. In fact,
in case of
$\mathfrak{g}$ is a finite dimensional simple Lie algebra and $\iota=(\cdots,i_{N+1},i_N,\cdots,i_2,i_1)$ is a sequence
such that
$(i_N,\cdots,i_2,i_1)$ is a reduced word of the longest element in the Weyl group $W$,
the image ${\rm Im}(\Psi_{\iota})$ coincides with the set of integer points in the string cone associated to the reduced word
$(i_1,i_2\cdots,i_N)$ as a set, which is a polyhedral convex cone.
When $\mathfrak{g}$ is of type ${\rm A}_n$, several combinatorial expressions of string cones
are given in
\cite{GKS16, GP}.
If $\mathfrak{g}$ is a finite dimensional simple Lie algebra and $\iota$ is a specific one then
explicit forms of ${\rm Im}(\Psi_{\iota})$ and ${\rm Im}(\Psi^{(\lambda)}_{\iota})$ are
given in \cite{H1, KS, Lit, N99,NZ}.
In case of
$\mathfrak{g}$ is a classical affine Lie algebra and $\iota$ is specific one,
explicit forms of them are
given in \cite{H2, HN, N99, NZ}.
When $\mathfrak{g}$ is a finite dimensional classical Lie algebra
and $\iota$ is adapted (Definition \ref{adapt}), we explicitly describe inequalities defining polyhedral realizations of $B(\infty)$ and $B(\lambda)$
in terms of rectangular Young tableaux \cite{KaN, KaN2}.
In the previous paper \cite{Ka}, we consider the case
$\mathfrak{g}$ is a classical affine Lie algebra of type
${\rm A}^{(1)}_{n-1}$, ${\rm C}^{(1)}_{n-1}$, ${\rm A}^{(2)}_{2n-2}$ or ${\rm D}^{(2)}_{n}$
and $\iota$ is adapted.
Then we give an explicit form of linear inequalities defining polyhedral realizations of $B(\infty)$
in terms of extended Young diagrams and proper Young walls.
Extended Young diagrams are used
to realize irreducible integrable highest weight modules $V(\Lambda)$ of $U_q(\mathfrak{g})$
as Fock space representations
for an affine Lie algebra $\mathfrak{g}$ of type ${\rm A}^{(1)}_n$, ${\rm C}^{(1)}_n$, ${\rm A}^{(2)}_{2n}$ or ${\rm D}^{(2)}_{n+1}$
and almost all fundamental weights $\Lambda$ \cite{Ha, JMMO, KMM}.
The proper Young walls are introduced in \cite{Kang}
to combinatorially realize the crystal base $B(\lambda)$ of level $1$ representations $V(\lambda)$ of $U_q(\mathfrak{g})$
for several affine Lie algebras $\mathfrak{g}$.
The detail of result in \cite{Ka} is as follows:
Defining ${\rm EYD}_k$ as the set of
extended Young diagrams with $y_{\infty}=k\in I$ (see Definition \ref{EYD-sub}),
one can assign linear functions $L^{{\rm A}^{(1)}}_{s,k,\iota}(T)$
and $L^{{\rm C}^{(1)}}_{s,k,\iota}(T)$
for each $T\in {\rm EYD}_k$ and $s\in\mathbb{Z}$ (see (\ref{ovlL2})).
We also define ${\rm REYD}^{{\rm A}^{(2)}}_{k}$ and ${\rm REYD}^{{\rm D}^{(2)}}_{k}$
as sets of revised extended Young diagrams (Definition \ref{ad-rem-pt2}, \ref{ad-rem-pt3})
and ${\rm YW}^{{\rm A}^{(2)}}_{k}$ and ${\rm YW}^{{\rm D}^{(2)}}_{k}$
as the sets of proper Young walls (Definition \ref{def-YW}, \ref{def-YW2}).
One can assign linear functions $L^{{\rm A}^{(2)}}_{s,k,\iota}(T)$ and
$L^{{\rm D}^{(2)}}_{s,k,\iota}(T)$ for each revised extended Young diagram or proper Young wall $T$ ((\ref{L1kdef}), (\ref{L11-def}))
and $s\in\mathbb{Z}$.
Setting
${\rm Comb}_{\iota}^{X}[\infty]$ as in (\ref{comb-inf-1}) and (\ref{comb-inf-2}),
we obtain
\[
{\rm Im}(\Psi_{\iota})=
\{
\mathbf{a}\in\mathbb{Z}^{\infty}_{\iota}|
\varphi(\mathbf{a})\geq0\text{ for any }\varphi\in {\rm Comb}_{\iota}^{X^L}[\infty]
\},
\]
where, $X$ is the type of $\mathfrak{g}$ and $X^L={\rm A}^{(1)}, {\rm D}^{(1)}, {\rm A}^{(2)}$ or ${\rm C}^{(2)}$
when $X={\rm A}^{(1)}_{n-1}$, ${\rm C}^{(1)}_{n-1}$, ${\rm A}^{(2)}_{2n-2}$ or ${\rm D}^{(2)}_{n}$, respectively.
That is, the inequalities defining ${\rm Im}(\Psi_{\iota})$
can be expressed in terms of combinatorial objects that deeply related
to fundamental representations of $U_q( ^L\mathfrak{g})$, where
$^L\mathfrak{g}$ is the classical affine Lie algebra whose generalized Cartan matrix is
the transposed matrix of that of $\mathfrak{g}$.
In this paper,
assuming that
$\mathfrak{g}$ is a classical affine Lie algebra of type
${\rm A}^{(1)}_{n-1}$, ${\rm C}^{(1)}_{n-1}$, ${\rm A}^{(2)}_{2n-2}$ or ${\rm D}^{(2)}_{n}$
and $\iota$ is adapted,
we give an explicit form of inequalities defining
${\rm Im}(\Psi_{\iota}^{(\lambda)})$ in terms of
combinatorial objects like as extended Young diagrams and Young walls.
Depending on $\iota$ and $k\in I$,
we will define
${\rm Comb}^X_{k,\iota}[\lambda]$
as a set
of
(i) a single function,
(ii) functions parametrized by boxes with integers $\mathbb{Z}_{\leq k}$,
(iii) functions parametrized by boxes with integers $\mathbb{Z}_{\geq k+1}$,
or
(iv) functions parametrized by (revised) extended Young diagrams or Young walls
((\ref{comb-lam-1}), (\ref{comb-lam-2}), (\ref{comb-lam-3})).
The functions in (ii), (iii) are related to crystal bases for extremal weight modules of $U_q( ^L\mathfrak{g})$ (see Remark \ref{rem-1}).
Then our main results Theorem \ref{thmA1} and \ref{thmA2} say
\[
{\rm Im}(\Psi_{\iota}^{(\lambda)})
=\left\{ \mathbf{a}\in\mathbb{Z}^{\infty}
\left|
\varphi(\mathbf{a})\geq0\\
\text{ for any }
\varphi\in
{\rm Comb}^{X^L}_{\iota}[\infty]\cup
\bigcup_{k\in I}
{\rm Comb}^{X^L}_{k,\iota}[\lambda]
\right.
\right\}.
\]
In this way, inequalities defining ${\rm Im}(\Psi_{\iota}^{(\lambda)})$
can be described by
combinatorial objects related
to representations of $U_q( ^L\mathfrak{g})$,
As an application, we will give a combinatorial
description of $\varepsilon_k^*$-functions on $B(\infty)$ (Theorem \ref{thm3}).
The organization
of this paper is as follows:
In Sect.2, we will review the notion of crystals and polyhedral realizations.
A procedure to compute explicit forms of ${\rm Im}(\Psi_{\iota}^{(\lambda)})$ is
given in this section.
In Sect.3, we recall several combinatorial objects like as (revised) extended Young diagrams
and Young walls.
Sect.4 is devoted to present
our main results, which describe explicit forms of ${\rm Im}(\Psi_{\iota}^{(\lambda)})$
in terms of combinatorial objects seen in Sect.3.
In Sect.5 and Sect.6, we give proofs of main theorems.
\vspace{2mm}
\nd
{\bf Acknowledgements}
This work was supported by JSPS KAKENHI Grant Number JP20J00186.
\section{Polyhedral realizations and a procedure}
In this section, we review crystals and polyhedral realizations.
We also give a modified procedure computing explicit forms of polyhedral realizations for $B(\lambda)$.
\subsection{Notations}
Let $\mathfrak{g}$ denote a symmetrizable Kac-Moody algebra over $\mathbb{Q}$ with the index set $I=\{1,2,\cdots,n\}$ and
a generalized Cartan matrix $A=(a_{i,j})_{i,j\in I}$. Let $\mathfrak{h}$ be a Cartan subalgebra,
$\langle \cdot,\cdot \rangle : \mathfrak{h} \times \mathfrak{h}^*\rightarrow \mathbb{Q}$
be the canonical pairing,
$P\subset \mathfrak{h}^*$ a weight lattice,
$\{\alpha_i\}_{i\in I}$ a set of simple roots and $\{h_i\}_{i\in I}$ a set of simple coroots.
It holds $\langle h_{i},\alpha_j \rangle=a_{i,j}$.
We also define
$P^*:=\{h\in\mathfrak{h} | \langle h,P \rangle\subset\mathbb{Z}\}$ and
$P^+:=\{\lambda\in P | \langle h_i,\lambda \rangle \in\mathbb{Z}_{\geq0} \text{ for all }i\in I \}$.
For each $i\in I$, the $i$-th fundamental weight $\Lambda_i\in P^+$ is defined as
$\langle h_j,\Lambda_i \rangle=\delta_{i,j}$.
Let $U_q(\mathfrak{g})$ be the quantized universal enveloping algebra
with generators $e_i$, $f_i$ ($i\in I$) and $q^h$ ($h\in P^*$) and $U_q^-(\mathfrak{g})$ be
the subalgebra of $U_q(\mathfrak{g})$ generated by $f_i$ ($i\in I$).
It is known that the algebra $U_q^-(\mathfrak{g})$ has
a crystal base $(L(\infty),B(\infty))$
and the irreducible integrable highest weight module $V(\lambda)$ of $U_q(\mathfrak{g})$
has a crystal base $(L(\lambda),B(\lambda))$ for $\lambda\in P^+$ (see \cite{K0,K1}).
\begin{rem}\label{rem1}
In this paper, we mainly treat Lie algebras of type
${\rm A}^{(1)}_{n-1}$, ${\rm C}^{(1)}_{n-1}$, ${\rm A}^{(2)}_{2n-2}$ and ${\rm D}^{(2)}_{n}$.
For the sake of simplicity, we often write them by
${\rm A}^{(1)}$, ${\rm C}^{(1)}$, ${\rm A}^{(2)}$ and ${\rm D}^{(2)}$.
\end{rem}
\nd
We will use the following
numbering of vertices in affine Dynkin diagrams:
\[
\begin{xy}
(-8,0) *{{\rm A}_{1}^{(1)} : }="A1",
(0,0) *{\bullet}="1",
(0,-3) *{1}="1a",
(10,0)*{\bullet}="2",
(10,-3) *{2}="2a",
\ar@{<=>} "1";"2"^{}
\end{xy}
\]
\[
\begin{xy}
(-15,5) *{{\rm A}_{n-1}^{(1)} \ (n\geq 3) : }="A1",
(20,10) *{\bullet}="n",
(20,13) *{n}="na",
(0,0) *{\bullet}="1",
(0,-3) *{1}="1a",
(10,0)*{\bullet}="2",
(10,-3)*{2}="2a",
(20,0)*{\ \cdots\ }="3",
(30,0)*{\bullet}="4",
(30,-3)*{n-2}="4a",
(40,0)*{\bullet}="5",
(40,-3)*{n-1}="5a",
(65,5) *{{\rm C}_{n-1}^{(1)} \ (n\geq 3) : }="C1",
(80,5) *{\bullet}="11",
(80,2) *{1}="11a",
(90,5)*{\bullet}="22",
(90,2)*{2}="22a",
(100,5)*{\ \cdots\ }="33",
(110,5)*{\bullet}="44",
(110,2)*{n-1}="44a",
(120,5)*{\bullet}="55",
(120,2)*{n}="55a",
\ar@{-} "1";"2"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\ar@{-} "4";"5"^{}
\ar@{-} "1";"n"^{}
\ar@{-} "n";"5"^{}
\ar@{=>} "11";"22"^{}
\ar@{-} "22";"33"^{}
\ar@{-} "33";"44"^{}
\ar@{<=} "44";"55"^{}
\end{xy}
\]
\[
\begin{xy}
(-15,5) *{{\rm A}_{2n-2}^{(2)}\ (n\geq 3) : }="A2",
(0,5) *{\bullet}="1",
(0,2) *{1}="1a",
(10,5)*{\bullet}="2",
(10,2)*{2}="2a",
(20,5)*{\ \cdots\ }="3",
(30,5)*{\bullet}="4",
(30,2)*{n-1}="4a",
(40,5)*{\bullet}="5",
(40,2)*{n}="5a",
(65,5) *{{\rm D}_{n}^{(2)} \ (n\geq 3) : }="D2",
(80,5) *{\bullet}="11",
(80,2) *{1}="11a",
(90,5)*{\bullet}="22",
(90,2)*{2}="22a",
(100,5)*{\ \cdots\ }="33",
(110,5)*{\bullet}="44",
(110,2)*{n-1}="44a",
(120,5)*{\bullet}="55",
(120,2)*{n}="55a",
\ar@{=>} "1";"2"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\ar@{=>} "4";"5"^{}
\ar@{<=} "11";"22"^{}
\ar@{-} "22";"33"^{}
\ar@{-} "33";"44"^{}
\ar@{=>} "44";"55"^{}
\end{xy}
\]
Replacing our numbering $1,2,\cdots,n-1,n,n+1$ of ${\rm A}_{n}^{(1)}$, ${\rm C}_{n}^{(1)}$ and ${\rm D}_{n+1}^{(2)}$
with $0,1,2,\cdots,n$, one obtains the numbering in \cite{JMMO, Kang, KMM}.
Replacing our numbering $1,2,\cdots,n-1,n,n+1$ of ${\rm A}_{2n}^{(2)}$ with $n,n-1,\cdots,1,0$, we get the numbering in \cite{Kang, KMM}.
\subsection{Crystals}
We review the notion of {\it crystals} following \cite{K3}:
\begin{defn}
A set $\mathcal{B}$ together with the maps
${\rm wt}:\mathcal{B}\rightarrow P$,
$\varepsilon_k,\varphi_k:\mathcal{B}\rightarrow \mathbb{Z}\sqcup \{-\infty\}$
and $\tilde{e}_k$,$\tilde{f}_k:\mathcal{B}\rightarrow \mathcal{B}\sqcup\{0\}$
($k\in I$)
is said to be a {\it crystal} if
the following relation holds: For $b,b'\in\mathcal{B}$ and $k\in I$,
\begin{itemize}
\item $\varphi_k(b)=\varepsilon_k(b)+\langle h_k,{\rm wt}(b)\rangle$,
\item ${\rm wt}(\tilde{e}_kb)={\rm wt}(b)+\alpha_k$ if $\tilde{e}_k(b)\in\mathcal{B}$,
\quad ${\rm wt}(\tilde{f}_kb)={\rm wt}(b)-\alpha_k$ if $\tilde{f}_k(b)\in\mathcal{B}$,
\item $\varepsilon_k(\tilde{e}_k(b))=\varepsilon_k(b)-1,\ \
\varphi_k(\tilde{e}_k(b))=\varphi_k(b)+1$\ if $\tilde{e}_k(b)\in\mathcal{B}$,
\item $\varepsilon_k(\tilde{f}_k(b))=\varepsilon_k(b)+1,\ \
\varphi_k(\tilde{f}_k(b))=\varphi_k(b)-1$\ if $\tilde{f}_k(b)\in\mathcal{B}$,
\item $\tilde{f}_k(b)=b'$ if and only if $b=\tilde{e}_k(b')$,
\item if $\varphi_k(b)=-\infty$ then $\tilde{e}_k(b)=\tilde{f}_k(b)=0$.
\end{itemize}
Here, $0$ and $-\infty$ are additional elements which do not belong to $\mathcal{B}$ and $\mathbb{Z}$, respectively.
The maps $\tilde{e}_k$,$\tilde{f}_k$ are called {\it Kashiwara operators}.
\end{defn}
The sets $B(\lambda)$ $(\lambda\in P^+)$, $B(\infty)$
have crystal structures. Let us review another crystal $R_{\lambda}$:
\begin{ex}\label{r-ex}
Let $R_{\lambda}:=\{r_{\lambda}\}$ ($\lambda\in P$) be the set consisting of a single element $r_{\lambda}$ and
defining
\[
{\rm wt}(r_{\lambda})=\lambda,\ \ \varepsilon_k(r_{\lambda})= -\lan h_k,\lambda\ran,\ \ \varphi_k(r_{\lambda})=0,\ \
\tilde{e}_k(r_{\lambda})=\tilde{f}_k(r_{\lambda})=0.
\]
Then $R_{\lambda}$ is a crystal.
\end{ex}
\begin{ex}\label{star-ex}
Let $\{\tilde{e}_i\}_{i\in I},\{\tilde{f}_i\}_{i\in I},\{\varepsilon_i\}_{i\in I},\{\varphi_i\}_{i\in I}, {\rm wt}$ be the crystal
structure on $B(\infty)$ and
$*:U_q(\frak{g})\rightarrow U_q(\frak{g})$ the antiautomorphism
such that $e_i^*=e_i$, $f_i^*=f_i$ and $(q^h)^*=q^{-h}$ in \cite{K1}.
It is known that the antiautomorphism $*$ induces a bijection $*:B(\infty)\rightarrow B(\infty)$ satisfying $*\circ *=id$.
One can define a crystal $B(\infty)^{*}$ as follows : The underlying set $B(\infty)^{*}$ equals $B(\infty)$,
the maps on $B(\infty)^{*}$ are defined as $\tilde{e}_i^*:=*\circ \tilde{e}_i\circ *$,
$\tilde{f}_i^*:=*\circ \tilde{f}_i\circ *$,
$\varepsilon_i^*:=\varepsilon_i\circ *$,
$\varphi_i^*:=\varphi_i\circ *$ and ${\rm wt}^*:={\rm wt}$.
\end{ex}
\begin{defn}
Let $\mathcal{B}_1$, $\mathcal{B}_2$ be two crystals.
\begin{enumerate}
\item[$(1)$]
A map $\psi : \mathcal{B}_1\sqcup\{0\}\rightarrow \mathcal{B}_2\sqcup\{0\}$
is said to be a strict morphism
and denoted by $\psi : \mathcal{B}_1\rightarrow \mathcal{B}_2$
if $\psi(0)=0$ and the following holds:
\begin{itemize}
\item For $k\in I$, if $b\in\mathcal{B}_1$ and $\psi(b)\in \mathcal{B}_2$ then
\[
{\rm wt}(\psi(b))={\rm wt}(b),\quad
\varepsilon_k(\psi(b))=\varepsilon_k(b),\quad
\varphi_k(\psi(b))=\varphi_k(b),
\]
\item $\tilde{e}_k(\psi(b))=\psi(\tilde{e}_k(b))$ and $\tilde{f}_k(\psi(b))=\psi(\tilde{f}_k(b))$ for $k\in I$ and $b\in\mathcal{B}_1$, where
we understand $\tilde{e}_k(0)=\tilde{f}_k(0)=0$.
\end{itemize}
\item[$(2)$] If a strict morphism
$\psi : \mathcal{B}_1\sqcup\{0\}\rightarrow \mathcal{B}_2\sqcup\{0\}$ is injective then
$\psi$ is said to be a strict embedding and denoted by $\psi:\mathcal{B}_1 \hookrightarrow \mathcal{B}_2$.
\item[$(3)$] If a strict morphism
$\psi : \mathcal{B}_1\sqcup\{0\}\rightarrow \mathcal{B}_2\sqcup\{0\}$ is bijective then
$\psi$ is said to be an isomorphism.
\end{enumerate}
\end{defn}
\begin{defn}
For two crystals $\mathcal{B}_1$, $\mathcal{B}_2$,
the tensor product $\mathcal{B}_1\otimes\mathcal{B}_2$
is defined to be the set
$\mathcal{B}_1\times\mathcal{B}_2$ whose crystal structure is defined as follows: For $k\in I$, $b_1\in\mathcal{B}_1$,
and $b_2\in\mathcal{B}_2$,
\begin{enumerate}
\item[$(1)$] ${\rm wt}(b_1\otimes b_2)={\rm wt}(b_1)+{\rm wt}(b_2)$,
\item[$(2)$] $\varepsilon_k(b_1\otimes b_2)={\rm max}\{\varepsilon_k(b_1), \varepsilon_k(b_2)-\langle h_k, {\rm wt}(b_1)\rangle\}$,
\item[$(3)$] $\varphi_k(b_1\otimes b_2)={\rm max}\{\varphi_k(b_2), \varphi_k(b_1)+\langle h_k, {\rm wt}(b_2)\rangle\}$,
\item[$(4)$]$\tilde{e}_k (b_1\otimes b_2)=
\begin{cases}
\tilde{e}_k b_1\otimes b_2 & {\rm if}\ \varphi_k(b_1)\geq \varepsilon_k(b_2),\\
b_1\otimes \tilde{e}_kb_2 & {\rm if}\ \varphi_k(b_1)< \varepsilon_k(b_2),
\end{cases}$
\item[$(5)$]$\tilde{f}_k (b_1\otimes b_2)=
\begin{cases}
\tilde{f}_k b_1\otimes b_2 & {\rm if}\ \varphi_k(b_1)> \varepsilon_k(b_2),\\
b_1\otimes \tilde{f}_kb_2 & {\rm if}\ \varphi_k(b_1)\leq \varepsilon_k(b_2).
\end{cases}$
\end{enumerate}
Here, we understand $0\otimes b_2=b_1\otimes0=0$.
\end{defn}
\subsection{Embeddings of crystal bases}\label{def-z}
One defines
\[
\ZZ^{\ify}
:=\{(\cd,a_r,\cd,a_2,a_1)| a_r\in\ZZ\text{ for }r\in\mathbb{Z}_{\geq1}
\text{ and it holds }a_r=0\,\,{\rm for}\ r\gg 0\}
\]
and takes an infinite sequence
$\io=(\cd,i_r,\cd,i_2,i_1)$ of indices from $I$ such that
\begin{equation}
{\hbox{
$i_r\ne i_{r+1}$ for $r\in\mathbb{Z}_{\geq1}$ and $\sharp\{r\in\mathbb{Z}_{\geq1}| i_r=k\}=\ify$ for all $k\in I$.}}
\label{seq-con}
\end{equation}
The set $\ZZ^{\ify}$ has a crystal structure associated with $\iota$ (see subsection 2.4 of \cite{NZ}).
Let $\ZZ^{\ify}_{\iota}$ denote its crystal.
\begin{thm}\cite{K3,NZ}
\label{emb}
There uniquely exists a strict embedding of crystals
\begin{equation}
\Psi_{\io}:B(\ify)\hookrightarrow \ZZ^{\ify}_{\io}
\label{psi}
\end{equation}
such that
$\Psi_{\io} (u_{\ify}) = \textbf{0}$, where
$u_{\ify}\in B(\ify)$ is
the highest weight vector, $\textbf{0}:=(\cd,0,\cd,0,0)$.
\end{thm}
Recall the crystal $R_{\lambda}$ defined in Example \ref{r-ex}.
We can explicitly write the crystal structure on $\mathbb{Z}^{\infty}[\lambda]:=\mathbb{Z}^{\infty}_{\iota}\otimes R_{\lambda}$ as follows:
Since $R_{\lambda}$ has only one element, one can identify $\mathbb{Z}^{\infty}[\lambda]$ with
$\mathbb{Z}^{\infty}$ as a set.
We consider the infinite dimensional vector space
\[
\QQ^{\ify}:=\{\textbf{a}=
(\cd,a_r,\cd,a_2,a_1)| a_r \in \QQ\text{ for }r\in\mathbb{Z}_{\geq1}
\text{ and it holds }
a_r = 0\,\,{\rm for\ } r \gg 0\},
\]
and its dual space $(\QQ^{\ify})^*:={\rm Hom}(\QQ^{\ify},\QQ)$.
Let $x_r\in (\QQ^{\ify})^*$ denote the linear function defined as $x_r((\cd,a_r,\cd,a_2,a_1)):=a_r$
for $r\in\mathbb{Z}_{\geq1}$ and
\[
\sigma_r:=x_r+\sum_{j>r}\langle h_{i_r},\alpha_{i_j}\rangle x_j\in(\QQ^{\ify})^*\quad (r\in\mathbb{Z}_{\geq1}),\qquad
\sigma_0^{(k)}:=-\langle h_k,\lambda \rangle+ \sum_{j\geq1} \langle h_k,\alpha_{i_j}\rangle x_j\quad (k\in I),
\]
where
$\sigma_0^{(k)}$ is a map from $\QQ^{\ify}$ to $\mathbb{Q}$ defined as
$\sigma_0^{(k)}(\mathbf{a})=-\langle h_k,\lambda \rangle+ \sum_{j\geq1} \langle h_k,\alpha_{i_j}\rangle a_j$
for $\mathbf{a}=(\cd,a_j,\cd,a_2,a_1)\in\QQ^{\ify}$.
Since $(\cd,a_j,\cd,a_2,a_1)\in\QQ^{\ify}$ satisfies $a_j=0$ for $j\gg0$,
these $\sigma_r$ and $\sigma_0^{(k)}$ are well-defined.
For $\mathbf{a}=(\cd,a_j,\cd,a_2,a_1)\in\QQ^{\ify}$ and $k\in I$,
let $\sigma^{(k)}(\mathbf{a}):={\rm max} \{\sigma_r(\mathbf{a}) | r\in\mathbb{Z}_{\geq1},\ i_r=k\}$
and
\[
M^{(k)}(\mathbf{a}):=\{r\in\mathbb{Z}_{\geq1} | i_r=k,\ \sigma_r(\mathbf{a})=\sigma^{(k)}(\mathbf{a})\}.
\]
We see that $\sigma^{(k)}(\mathbf{a})\geq0$ and the set
$M^{(k)}(\mathbf{a})$ is finite if and only if
$\sigma^{(k)}(\mathbf{a})>0$. Then ${\rm wt} : \mathbb{Z}^{\infty}[\lambda]\rightarrow P$,
$\varepsilon_k, \varphi_k : \mathbb{Z}^{\infty}[\lambda]\rightarrow \mathbb{Z}$ $(k\in I)$ are defined by
\[
{\rm wt}(\mathbf{a}):=\lambda-\sum_{j=1}^{\infty} a_j\alpha_{i_j},\quad
\varepsilon_k(\mathbf{a}):={\rm max}\{\sigma^{(k)}(\mathbf{a}),\sigma_0^{(k)}(\mathbf{a})\},\quad
\varphi_k(\mathbf{a}):=\langle h_k,{\rm wt}(\mathbf{a})\rangle + \varepsilon_k(\mathbf{a})
\]
for $\mathbf{a}=(\cd,a_j,\cd,a_2,a_1)\in\mathbb{Z}^{\infty}[\lambda]$. Next, let us consider
Kashiwara operators $\tilde{f}_k, \tilde{e}_k: \mathbb{Z}^{\infty}[\lambda]\sqcup\{0\}\rightarrow
\mathbb{Z}^{\infty}[\lambda]\sqcup\{0\}$. They are defined by $\tilde{f}_k(0)=\tilde{e}_k(0)=0$ and
\begin{equation}\label{Kashi-def1}
(\tilde{f}_k(\mathbf{a}))_r : =
a_r +\delta_{r,{\rm min} M^{(k)}(\mathbf{a})} \text{ if }\sigma^{(k)}(\mathbf{a})>\sigma_0^{(k)}(\mathbf{a}) ; \text{ otherwise }\tilde{f}_k(\mathbf{a})=0,
\end{equation}
\begin{equation}\label{Kashi-def2}
(\tilde{e}_k(\mathbf{a}))_r : =
a_r -\delta_{r,{\rm max} M^{(k)}(\mathbf{a})} \text{ if }\sigma^{(k)}(\mathbf{a})>0\text{ and }\sigma^{(k)}(\mathbf{a})\geq\sigma_0^{(k)}(\mathbf{a})
; \text{ otherwise }\tilde{e}_k(\mathbf{a})=0
\end{equation}
for $\mathbf{a}=(\cd,a_j,\cd,a_2,a_1)\in\mathbb{Z}^{\infty}[\lambda]$.
Then above $\{\tilde{e}_k\}_{k\in I}$, $\{\tilde{f}_k\}_{k\in I}$,
$\{\varepsilon_k\}_{k\in I}$, $\{\varphi_k\}_{k\in I}$ and ${\rm wt}$ give
the crystal structure of the set $\mathbb{Z}^{\infty}[\lambda]$ \cite{N99}.
Let $\mathbb{Z}^{\infty}_{\iota}[\lambda]$ denote this crystal.
\vspace{2mm}
For $\lambda\in P^+$,
setting
\begin{equation}\label{Btilde}
\tilde{B}(\lambda)
:=\{
b\otimes r_{\lambda}\in B(\infty)\otimes R_{\lambda} |
\varepsilon_i^*(b)\leq \lan h_i,\lambda\ran \text{ for all }i\in I
\},
\end{equation}
we obtain $B(\lambda)\cong \tilde{B}(\lambda)$ \cite{N99},
where $\varepsilon_i^*$ is the function in Example \ref{star-ex}.
Hence, there exists
a strict embedding of crystals $\Omega : B(\lambda)\hookrightarrow B(\infty)\otimes R_{\lambda}$.
Using the embedding $\Psi_{\iota}$ in Theorem \ref{emb},
the following theorem holds:
\begin{thm}\cite{N99}\label{emb2}
The map
\[
\Psi^{(\lambda)}_{\io}:B(\lambda)\overset{\Omega}{\hookrightarrow} B(\infty)\otimes R_{\lambda}
\overset{\Psi_{\iota}\otimes \text{id}}{\hookrightarrow} \mathbb{Z}^{\infty}_{\iota}\otimes R_{\lambda}
=\ZZ^{\ify}_{\iota}[\lambda]
\]
is the unique strict embedding of crystals
such that $\Psi^{(\lambda)}_{\io}(u_{\lambda})=(\cdots,0,0,0)$. Here,
$u_{\lambda}$ is the highest weight vector in $B(\lambda)$.
\end{thm}
\subsection{Modified Nakashima-Zelevinsky's procedure}\label{poly-uqm}
Let us review a procedure to compute an explicit form of ${\rm Im}(\Psi_{\iota})\subset\mathbb{Z}^{\infty}_{\iota}$ gave in
the previous paper \cite{Ka}.
As in the previous subsection,
we fix an infinite sequence
\[
\io=(\cd,i_r,\cd,i_2,i_1)
\]
of indices
satisfying (\ref{seq-con})
and for $r\in \mathbb{Z}_{\geq1}$, define $x_r\in(\QQ^{\ify})^*$ as
$x_r(\cd,a_r,\cd,a_2,a_1)=a_r$. For $r\in \mathbb{Z}_{<1}$, we set $x_r:=0$.
For $r\in \mathbb{Z}_{\geq1}$, let
\[
r^{(+)}:={\rm min}\{\ell\in\mathbb{Z}_{\geq1}\ |\ \ell>r\,\,{\rm and }\,\,i_r=i_{\ell}\},\ \
r^{(-)}:={\rm max} \{\ell\in\mathbb{Z}_{\geq1}\ |\ \ell<r\,\,{\rm and }\,\,i_r=i_{\ell}\}\cup\{0\},
\]
and for $r\in\mathbb{Z}_{\geq1}$, let
\begin{equation}
\beta_r:=
\sigma_r-\sigma_{r^{(+)}}
=x_r+\sum_{r<j<r^{(+)}}\lan h_{i_r},\al_{i_j}\ran x_j+x_{r^{(+)}}\in (\QQ^{\ify})^*
\label{betak}
\end{equation}
and $\beta_0:=0$. For $r\in\mathbb{Z}_{\geq1}$,
one defines the map
$S_r'=S_{r,\io}':(\QQ^{\ify})^*\rightarrow (\QQ^{\ify})^*$ as follows:
For $\vp=\sum_{\ell\in\mathbb{Z}_{\geq1}}c_{\ell}x_{\ell}\in(\QQ^{\ify})^*$,
\begin{equation}
S_r'(\vp):=
\begin{cases}
\vp-\beta_r & {\mbox{ if }}\ c_r>0,\\
\vp+\beta_{r^{(-)}} & {\mbox{ if }}\ c_r< 0,\\
\vp & {\mbox{ if }}\ c_r= 0.
\end{cases}
\label{Sk}
\end{equation}
We often write $S_r'(\vp)$ as $S_r'\vp$. Let us define
\begin{eqnarray}
\Xi_{\io}' &:= &\{S_{j_{\ell}}'\cd S_{j_2}'S_{j_1}'x_{j_0}\,|\,
\ell\in\mathbb{Z}_{\geq0},j_0,j_1,\cd,j_{\ell}\in\mathbb{Z}_{\geq1}\}, \label{xiprime}\\
\Sigma_{\io}' & := &
\{\textbf{a}\in \ZZ^{\ify}\subset \QQ^{\ify}\,|\,\vp(\textbf{a})\geq0\,\,{\rm for}\,\,
{\rm any}\,\,\vp\in \Xi_{\io}'\}.\nonumber
\end{eqnarray}
We say the sequence $\io$ satisfies the $\Xi'$-{\it positivity condition}
when it holds
\begin{equation}
{\hbox{for any
$\vp=\sum_{\ell\in\mathbb{Z}_{\geq1}} c_{\ell}x_{\ell}\in \Xi_{\io}'$, if $\ell^{(-)}=0$ then $c_{\ell}\geq0$}}.
\label{posi}
\end{equation}
\begin{thm}\label{inf-thm}\cite{Ka}
Let $\io$ be a sequence of indices satisfying $(\ref{seq-con})$
and $(\ref{posi})$. Then we have
${\rm Im}(\Psi_{\io})=\Sigma_{\io}'$.
\end{thm}
\subsection{Modified Nakashima's procedure}\label{poly-uql}
In this subsection, we give a procedure to compute an explicit form of the image
${\rm Im}(\Psi^{(\lambda)}_{\io})$ ($\cong B(\lambda)$) of
$\Psi^{(\lambda)}_{\io}$ defined in Theorem \ref{emb2}.
Recall that we defined $\beta_r$ in $(\ref{betak})$.
Let $\beta^{(\pm)}_r$ be the functions given by $\beta^{(+)}_r=\beta_{r}$ and
\begin{equation*}
\beta^{(-)}_r =
\begin{cases}
x_{r^{(-)}}+\sum_{r^{(-)}<j<r} \lan h_{i_r},\alpha_{i_j}\ran x_j +x_r=\sigma_{r^{(-)}}-\sigma_r & {\rm if}\ r^{(-)}>0,\\
- \lan h_{i_r},\lambda\ran + \sum_{1\leq j<r} \lan h_{i_r},\alpha_{i_j}\ran x_j + x_r
=\sigma_0^{(i_r)}-\sigma_r
& {\rm if}\ r^{(-)}=0.
\end{cases}
\end{equation*}
Note that $\beta^{(-)}_r= \beta_{r^{(-)}}$ if $r^{(-)}>0$.
Using this notation, for each $r\in\mathbb{Z}_{\geq1}$
and $\varphi=c+\sum_{r\geq1} \varphi_rx_r$ with $c$, $\varphi_r\in\mathbb{Q}$,
we define a new function $\what{S}_r'(\varphi)$ as follows:
\begin{equation}\label{Shat}
\what{S}_r'(\varphi):=
\begin{cases}
\varphi - \beta^{(+)}_r & {\rm if}\ \varphi_r>0, \\
\varphi + \beta^{(-)}_r & {\rm if}\ \varphi_r<0, \\
\varphi & {\rm if}\ \varphi_r=0.
\end{cases}
\end{equation}
Note that except for the case $\varphi_r<0$ and $r^{(-)}=0$,
it holds
\begin{equation}\label{twoS}
\what{S}_r'(\varphi)=S_r'(\varphi-c)+c,
\end{equation}
where $S_r'$ is defined in (\ref{Sk}).
For $k\in I$, one defines
\begin{equation}\label{lmi-def}
\iota^{(k)}:={\rm min}\{r\in\mathbb{Z}_{\geq1} | i_r=k\},\qquad
\lambda^{(k)} :=\lan h_k, \lambda \ran - \sum_{1\leq j<\iota^{(k)}} \lan h_k, \alpha_{i_j} \ran x_j - x_{\iota^{(k)}}.
\end{equation}
For $\iota$ and $\lambda\in P^+$, let $\Xi'_{\iota}[\lambda]$ be the set of all
functions generated by applying $\what{S}'_r$ on the functions $x_j$ ($j \geq 1$) and $\lambda^{(k)}$
($k\in I$), that is,
\begin{equation}\label{xilamdef}
\begin{array}{l}
\Xi'_{\iota}[\lambda]: =\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1}x_{j_0} | \ell\in\mathbb{Z}_{\geq0},\ j_0,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \} \\
\qquad\qquad \cup \{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ k\in I,\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}.
\end{array}
\end{equation}
By (\ref{twoS}), if $\iota$ satisfies the $\Xi'$-positivity condition (\ref{posi}) then
we see that
\begin{equation}\label{twoS2}
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1}x_{j_0} | \ell\in\mathbb{Z}_{\geq0}, j_0,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
=\{ S'_{j_{\ell}}\cdots S'_{j_1}x_{j_0} | \ell\in\mathbb{Z}_{\geq0}, j_0,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}.
\end{equation}
Now we set
\begin{equation}\label{siglamdef}
\Sigma_{\iota}'[\lambda]:=
\{
\mathbf{a}\in\mathbb{Z}^{\infty}_{\iota}[\lambda] | \varphi(\mathbf{a}) \geq0\ \ {\rm for\ any\ }\varphi\in \Xi'_{\iota}[\lambda]
\}.
\end{equation}
\begin{defn}\label{ample1}
The pair $(\iota,\lambda)$ is said to be $\Xi'$-{\it ample}
if $\textbf{0}:=(\cdots,0,0,0)\in \Sigma'_{\iota}[\lambda]$.
\end{defn}
\begin{thm}\label{Nthm1}
For $\lambda\in P^+$,
we suppose that $(\iota,\lambda)$ is $\Xi'$-ample.
Let $\Psi^{(\lambda)}_{\iota}:B(\lambda)\hookrightarrow \mathbb{Z}^{\infty}_{\iota}[\lambda]$
be the embedding of crystals in Theorem \ref{emb2}. Then
the image ${\rm Im}(\Psi^{(\lambda)}_{\iota}) (\cong B(\lambda))$ coincides with
$\Sigma'_{\iota}[\lambda]$.
\end{thm}
\nd
{\it Proof.} The proof is a slight modification of that of Theorem 4.1 in \cite{N99}.
Since ${\rm Im}(\Psi^{(\lambda)}_{\iota})$ is isomorphic to $B(\lambda)$ and $\Psi^{(\lambda)}_{\iota}$
is a strict embedding of crystals, it holds
\[
{\rm Im}(\Psi^{(\lambda)}_{\iota})=
\{
\tilde{f}_{j_{\ell}}\cdots\tilde{f}_{j_1}\Psi_{\iota}^{(\lambda)}(u_{\lambda}) | j_1,\cdots,j_{\ell}\in I,\ \ell\in\mathbb{Z}_{\geq0} \}\setminus
\{0\}
\]
and $\Psi_{\iota}^{(\lambda)}(u_{\lambda})=\mathbf{0}=(\cdots,0,0,0)$.
By the definition of $\tilde{f}_j$ $(j\in I)$ on $\mathbb{Z}^{\infty}_{\iota}[\lambda]$ in subsection \ref{def-z},
one can easily check
\[
{\rm Im}(\Psi_{\iota}^{(\lambda)})\subset
\mathbb{Z}^{\infty}_{\geq0,\iota}[\lambda]:=\{ (\cdots,a_3,a_2,a_1)\in\mathbb{Z}^{\infty}_{\iota}[\lambda] | a_r\geq0\ (r\in\mathbb{Z}_{\geq1}) \}.
\]
First, let us show
\[
{\rm Im}(\Psi_{\iota}^{(\lambda)})\subset \Sigma_{\iota}'[\lambda].
\]
For any $k\in I$, let us show
\begin{equation}\label{Nthm-pr-1}
\tilde{f}_k \Sigma'_{\iota}[\lambda] \subset \Sigma'_{\iota}[\lambda]\cup\{0\}.
\end{equation}
For $\textbf{a}=(\cdots,a_3,a_2,a_1)\in \Sigma'_{\iota}[\lambda]$, we assume
\[
\tilde{f}_k\textbf{a}=(\cdots,a_{r+1},a_r+1,a_{r-1},\cdots,a_2,a_1)
\]
with some $r\in\mathbb{Z}_{\geq1}$. Note that $i_r=k$.
For any $\varphi=c+\sum_{j\in\mathbb{Z}_{\geq1}}\varphi_jx_j\in \Xi'_{\iota}[\lambda]$ ($c,\varphi_j\in\mathbb{Z}$),
we need to show
\begin{equation}\label{Nthm-pr-2}
\varphi(\tilde{f}_k \mathbf{a})\geq0.
\end{equation}
Since
$\varphi(\tilde{f}_k \mathbf{a})=\varphi(\mathbf{a})+\varphi_r \geq \varphi_r$,
we may assume that $\varphi_r<0$. It follows from the definition of $\tilde{f}_k$ that
\[
\sigma_r(\mathbf{a})>\sigma_{r^{(-)}}(\mathbf{a})\ \text{if }r^{(-)}>0,\qquad
\sigma_r(\mathbf{a})>\sigma_{0}^{(k)}(\mathbf{a})\ \text{if }r^{(-)}=0
\]
so that $\beta^{(-)}_{r}(\mathbf{a})\leq -1$.
Hence,
\[
\varphi(\tilde{f}_k(\mathbf{a}))=
\varphi(\mathbf{a})+\varphi_r
\geq
\varphi(\mathbf{a})-\varphi_r\beta^{(-)}_{r}(\mathbf{a})
=(\what{S}_r')^{-\varphi_r} (\varphi)(\mathbf{a})\geq0.
\]
Thus, we get (\ref{Nthm-pr-2}) and (\ref{Nthm-pr-1}).
By the assumption $\mathbf{0}\in \Sigma_{\iota}'[\lambda]$ and
(\ref{Nthm-pr-1}) imply ${\rm Im}(\Psi_{\iota}^{(\lambda)})\subset \Sigma_{\iota}'[\lambda]$.
Next, to prove
\[
{\rm Im}(\Psi_{\iota}^{(\lambda)})\supset \Sigma_{\iota}'[\lambda],
\]
let us show for any $k\in I$,
\begin{equation}\label{Nthm-pr-3}
\tilde{e}_k \Sigma'_{\iota}[\lambda] \subset \Sigma'_{\iota}[\lambda]\cup\{0\}.
\end{equation}
For $\textbf{a}=(\cdots,a_3,a_2,a_1)\in \Sigma'_{\iota}[\lambda]$, we assume
\[
\tilde{e}_k\textbf{a}=(\cdots,a_{r+1},a_r-1,a_{r-1},\cdots,a_2,a_1)
\]
with some $r\in\mathbb{Z}_{\geq1}$. Note that $i_r=k$.
For any $\varphi=c+\sum_{j\in\mathbb{Z}_{\geq1}}\varphi_jx_j\in \Xi'_{\iota}[\lambda]$ ($c,\varphi_j\in\mathbb{Z}$),
we need to show
\begin{equation}\label{Nthm-pr-4}
\varphi(\tilde{e}_k \mathbf{a})\geq0.
\end{equation}
Since $\varphi(\tilde{e}_k\mathbf{a})=\varphi(\mathbf{a})-\varphi_r\geq -\varphi_r$,
we may assume $\varphi_r>0$. By the definition of $\tilde{e}_k$, it follows $\sigma_r(\mathbf{a})-\sigma_{r^{(+)}}(\mathbf{a})>0$
so that $\beta^{(+)}_r(\mathbf{a})\geq1$.
Thus,
\[
\varphi(\tilde{e}_k(\mathbf{a}))=
\varphi(\mathbf{a})-\varphi_r
\geq
\varphi(\mathbf{a})-\varphi_r\beta^{(+)}_{r}(\mathbf{a})
=(\what{S}_r')^{\varphi_r} (\varphi)(\mathbf{a})\geq0.
\]
Hence, one obtains (\ref{Nthm-pr-4}) and (\ref{Nthm-pr-3}).
Considering
$\Sigma'_{\iota}[\lambda]\subset
\mathbb{Z}^{\infty}_{\geq0,\iota}[\lambda]$ and
(\ref{Nthm-pr-3}), for any $\textbf{a}=(\cdots,a_3,a_2,a_1)\in \Sigma'_{\iota}[\lambda]$,
taking $\ell\in\mathbb{Z}_{>0}$ sufficiently larger than $0$, it holds
\[
\tilde{e}_{j_1}\cdots \tilde{e}_{j_{\ell}}(\mathbf{a})=0
\]
for any $j_1,\cdots,j_{\ell}\in I$ (indeed, one can take $\ell=\sum_{j\geq1}a_j+1$).
Hence,
all we need to show is that if $\mathbf{a}\in \Sigma_{\iota}'[\lambda]$ satisfies
$\tilde{e}_k(\mathbf{a})=0$ for all $k\in I$ then $\mathbf{a}=\mathbf{0}$. In fact,
it means for any $\mathbf{y}\in\Sigma_{\iota}'[\lambda]$, there exist $j_1,\cdots,j_{\ell}\in I$
such that $\tilde{e}_{j_1}\cdots \tilde{e}_{j_{\ell}}(\mathbf{y})=\mathbf{0}$, which implies
\[
\mathbf{y}=\tilde{f}_{j_{\ell}}\cdots \tilde{f}_{j_{1}}(\mathbf{0})\in {\rm Im}(\Psi_{\iota}^{(\lambda)}).
\]
We assume $\mathbf{a}\in \Sigma_{\iota}'[\lambda]$, $\mathbf{a}\neq\mathbf{0}$ and $\tilde{e}_k(\mathbf{a})=0$ for all $k\in I$
and deduce a contradiction from this assumption.
Considering the definition of $\tilde{e}_k$, it holds either
\begin{equation}\label{Nthm-pr-5}
\sigma^{(k)}(\mathbf{a})\leq 0\quad \text{or}\quad \sigma^{(k)}(\mathbf{a})< \sigma^{(k)}_0(\mathbf{a}).
\end{equation}
The assumption
$\mathbf{a}=(\cdots,a_3,a_2,a_1)\neq\mathbf{0}$ means
there exists $j\in\mathbb{Z}_{\geq1}$ such that $a_j>0$ and $a_r=0$ when $r>j$. Thus,
$\sigma_j(\mathbf{a})=a_j>0$ and $\sigma^{(i_j)}(\mathbf{a})\geq a_j>0$.
The relation $\sigma^{(i_j)}(\mathbf{a})<\sigma^{(i_j)}_0(\mathbf{a})$ is an easy consequence of
(\ref{Nthm-pr-5}). Putting $k:=i_j$,
\[
0<\sigma^{(k)}_0(\mathbf{a})-\sigma^{(k)}(\mathbf{a})
\leq \sigma^{(k)}_0(\mathbf{a})-\sigma_{\iota^{(k)}}(\mathbf{a})=\beta^{(-)}_{\iota^{(k)}}(\mathbf{a}).
\]
Hence,
\[
\lambda^{(k)}(\mathbf{a})=-\beta^{(-)}_{\iota^{(k)}}(\mathbf{a})<0,
\]
which contradicts $\mathbf{a}\in \Sigma_{\iota}'[\lambda]$. Therefore, one obtains $\mathbf{a}=\mathbf{0}$.\qed
\subsection{A description of $\varepsilon_k^*$}
In general, it seems to be difficult to find a value of $\varepsilon_k^*$ in Example \ref{star-ex}
by a direct calculation.
As an application of polyhedral realizations, we give a combinatorial description of
$\varepsilon_k^*$, which is an analog of subsection 4.3 of \cite{N99}.
For $k\in I$, let
\[
\xi^{(k)}:=- \sum_{1\leq j<\iota^{(k)}} \lan h_k, \alpha_{i_j} \ran x_j - x_{\iota^{(k)}}\in (\mathbb{Q}^{\infty})^*.
\]
Note that for any $\lambda\in P^+$, one can write
$\xi^{(k)}=\lambda^{(k)}-\lan h_k, \lambda \ran$ by using $\lambda^{(k)}$ in
(\ref{lmi-def}). Using the maps in (\ref{Sk}), we set
\[
\Xi'^{(k)}_{\iota}:=\{S'_{j_{\ell}}\cdots S'_{j_1}\xi^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in\mathbb{Z}_{\geq1}\}.
\]
We say the sequence $\iota$ satisfies the $\Xi'$-{\it strict positivity condition} when it holds the following:
\begin{equation}\label{strict-cond}
\text{for any }
\vp=\sum_{\ell\geq1} \varphi_{\ell}x_{\ell}\in \left(\bigcup_{k\in I}\Xi'^{(k)}_{\iota}\setminus\{\xi^{(k)}\}\right)\cup\Xi'_{\iota},
\text{ if }\ell^{(-)}=0\text{ then }\varphi_{\ell}\geq0,
\end{equation}
where $\Xi'_{\iota}$ is defined in (\ref{xiprime}).
\begin{thm}\label{thm2}
If $\iota$ satisfies the $\Xi'$-strict positivity condition then for $k\in I$ and $x\in {\rm Im}(\Psi_{\iota})$, we have
\[
\varepsilon^*_k(x)={\rm max}\{-\varphi(x)| \varphi\in \Xi'^{(k)}_{\iota}\}.
\]
\end{thm}
\nd
{\it Proof.} The proof is a slight modification of Theorem 4.2 of \cite{N99}. We take an arbitrary $\lambda\in P^+$.
Since $\iota$ satisfies the $\Xi'$-strict positivity condition, it also satisfies
$\Xi'$-positivity condition in (\ref{posi}) and it holds
\begin{equation}\label{thm2-pr0}
{\rm Im}(\Psi_{\iota})=
\{
\mathbf{a}\in\mathbb{Z}_{\iota}^{\infty} | \varphi(\mathbf{a})\geq0\ \text{for any }\varphi\in
\{S'_{j_{\ell}}\cdots S'_{j_1}x_{j_0} | \ell\in\mathbb{Z}_{\geq0}, j_0,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
\}
\end{equation}
by Theorem \ref{inf-thm}. As checked in (\ref{twoS2}), we have
\begin{equation}\label{thm2-pr1}
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1}x_{j_0} | \ell\in\mathbb{Z}_{\geq0}, j_0,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
=\{ S'_{j_{\ell}}\cdots S'_{j_1}x_{j_0} | \ell\in\mathbb{Z}_{\geq0}, j_0,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}.
\end{equation}
Next, for $\ell\in\mathbb{Z}_{\geq0}$ and $j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1}$,
let us show if $\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} \neq0$ then
\begin{equation}\label{thm2-pr2}
\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} = \lan h_k,\lambda\ran + S'_{j_{\ell}}\cdots S'_{j_1} \xi^{(k)}
\end{equation}
by induction on $\ell$. When $\ell=0$ the claim is clear so that we assume $\ell>0$
and (\ref{thm2-pr2}) holds. We put $\varphi:=\lan h_k,\lambda\ran+\sum_{r\in\mathbb{Z}_{\geq1}}
\varphi_rx_r=\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} = \lan h_k,\lambda\ran + S'_{j_{\ell}}\cdots S'_{j_1} \xi^{(k)}$
with some $\varphi_r\in\mathbb{Z}$.
For any $r\in\mathbb{Z}_{\geq1}$, except for the case $\varphi_r<0$ and $r^{(-)}=0$, it holds
\[
\what{S}'_r\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} = \lan h_k,\lambda\ran + S'_rS'_{j_{\ell}}\cdots S'_{j_1} \xi^{(k)}
\]
by (\ref{twoS}). Thus, we may assume $\varphi_r<0$ and $r^{(-)}=0$. Then
$\Xi'$-strict positivity condition implies $S'_{j_{\ell}}\cdots S'_{j_1} \xi^{(k)}=\xi^{(k)}$ so that
$\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} =\lambda^{(k)}$.
It follows from $\varphi_r<0$ and the definition of $\xi^{(k)}$ that $r=\iota^{(k)}$
and $\what{S}'_r\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)}=0$, this is not the case of (\ref{thm2-pr2}).
In this way, we have proved (\ref{thm2-pr2}).
Since $S'_{j_{\ell}}\cdots S'_{j_1} \xi^{(k)}$ and
$S'_{j_{\ell}}\cdots S'_{j_1}x_{j_0}$ are linear maps and $\lan h_k,\lambda\ran\geq0$,
we see that $(\iota,\lambda)$ is $\Xi'$-ample.
Combining Theorem \ref{Nthm1} with (\ref{thm2-pr0}), (\ref{thm2-pr1}) and (\ref{thm2-pr2}),
it holds
\begin{eqnarray}
{\rm Im}(\Psi_{\iota}^{(\lambda)})
&=&
\{
\mathbf{a}\otimes r_{\lambda}\in
{\rm Im}(\Psi_{\iota})\otimes R_{\lambda}
| \varphi(\mathbf{a})\geq0\ \text{for any }\varphi\in
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1}\lambda^{(k)} | k\in I,\ \ell\in\mathbb{Z}_{\geq0}, j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
\}\nonumber\\
&=&
\{
\mathbf{a}\otimes r_{\lambda}\in
{\rm Im}(\Psi_{\iota})\otimes R_{\lambda}
| -\varphi(\mathbf{a})\leq\lan h_k,\lambda\ran\ \text{for any }k\in I\text{ and }\varphi\in
\Xi_{\iota}'^{(k)}\}
\}\nonumber\\
&=&
\{
\mathbf{a}\otimes r_{\lambda}\in
{\rm Im}(\Psi_{\iota})\otimes R_{\lambda}
| {\rm max}\{-\varphi(\mathbf{a}) |\varphi\in
\Xi_{\iota}'^{(k)}\} \leq\lan h_k,\lambda\ran\ \text{for any }k\in I\}
\}.
\label{Btilde2}
\end{eqnarray}
Note that
${\rm max}\{-\varphi(\mathbf{a}) |\varphi\in
\Xi_{\iota}'^{(k)}\}\geq0$. In fact,
taking $r\in\mathbb{Z}_{\geq1}$
such that $a_s=0$ for $s\geq r$
and
applying maps $S'_{j_{\ell}},\cdots, S'_{j_1}$ to $\xi^{(k)}$ properly,
$S'_{j_{\ell}},\cdots, S'_{j_1}\xi^{(k)}\in
\Xi_{\iota}'^{(k)}$ becomes a sum of
$-x_t$ ($t\geq1$) and $x_{s}$ ($s\geq r$) by the definition of $S'$
so that $-S'_{j_{\ell}},\cdots, S'_{j_1}\xi^{(k)}(\mathbf{a})\geq0$.
Comparing (\ref{Btilde2}) with (\ref{Btilde}), one obtains
$\varepsilon^*_k(x)={\rm max}\{-\varphi(x)| \varphi\in \Xi'^{(k)}_{\iota}\}$ on $x\in{\rm Im}(\Psi_{\iota})$.
In fact, for a fixed $x\in{\rm Im}(\Psi_{\iota})$,
because $\lambda\in P^+$ is arbitrary, one can take $\lambda$ as
$\varepsilon_k^*(x)=\lan h_k,\lambda\ran$ for all $k\in I$.
Using (\ref{Btilde}), it follows $x\otimes r_{\lambda}\in {\rm Im}(\Psi_{\iota}^{(\lambda)})$
so that $\varepsilon_k^*(x)=\lan h_k,\lambda\ran\geq {\rm max}\{-\varphi(x) |\varphi\in
\Xi_{\iota}'^{(k)}\}$ holds by (\ref{Btilde2}). Similarly, taking $\lambda$ as
$ {\rm max}\{-\varphi(x) |\varphi\in
\Xi_{\iota}'^{(k)}\}=\lan h_k,\lambda\ran$ for all $k\in I$, the expression (\ref{Btilde2})
yields $x\otimes r_{\lambda}\in {\rm Im}(\Psi_{\iota}^{(\lambda)})$
and
$ {\rm max}\{-\varphi(x) |\varphi\in
\Xi_{\iota}'^{(k)}\}=\lan h_k,\lambda\ran\geq \varepsilon_k^*(x)$ holds by (\ref{Btilde}). \qed
\section{Combinatorial objects}
In this section, we recall several combinatorial objects ((revised) extended Young diagrams, Young walls), which express explicit forms of ${\rm Im}(\Psi_{\iota}^{(\lambda)})$.
\subsection{Extended Young diagrams}\label{EYD-sub}
\begin{defn}\cite{Ha,JMMO}
Let $y_{\infty}$ be a fixed integer. A sequence $(y_r)_{r\in\mathbb{Z}_{\geq0}}$
is called an {\it extended Young diagram} of charge $y_{\infty}$ if the following conditions hold:
\begin{itemize}
\item $y_r\in\mathbb{Z}$, $y_r\leq y_{r+1}$ for all $r\in\mathbb{Z}_{\geq0}$,
\item $y_r=y_{\infty}$ for $r\gg0$.
\end{itemize}
\end{defn}
Each extended Young diagram is described as a Young diagram
drawn in $\mathbb{R}_{\geq0}\times \mathbb{R}_{\leq y_{\infty}}$ as follows:
For $(y_r)_{r\in\mathbb{Z}_{\geq0}}$, we draw a line between the points $(r,y_r)$ and $(r+1,y_r)$ and when
$y_r<y_{r+1}$ draw
a line
between $(r+1,y_r)$ and $(r+1,y_{r+1})$
for each $r\in\mathbb{Z}_{\geq0}$.
\begin{ex}\label{ex-1}
Let $T=(y_r)_{r\in\mathbb{Z}_{\geq0}}$ be
the extended Young diagram of charge
$y_{\infty}=1$ defined by
$y_0=-3$, $y_1=-2$, $y_2=y_3=-1$, $y_4=0$, $y_5=1, y_6=1,\cdots$.
Then $T$ is drawn in $\mathbb{R}_{\geq0}\times \mathbb{R}_{\leq 1}$ as
\[
\begin{xy}
(0,0) *{}="1",
(45,0)*{}="2",
(0,-35)*{}="3",
(-5,0)*{(0,1)}="4",
(6,2) *{1}="10",
(6,-1) *{}="1010",
(12,2) *{2}="11",
(12,-1) *{}="1111",
(18,2) *{3}="12",
(18,-1) *{}="1212",
(24,2) *{4}="13",
(24,-1) *{}="1313",
(30,2) *{5}="14",
(30,-1) *{}="1414",
(-3,-6)*{0\ }="5",
(1,-6)*{}="55",
(-4,-12)*{-1\ }="6",
(1,-12)*{}="66",
(-4,-18)*{-2\ }="7",
(1,-18)*{}="77",
(-4,-24)*{-3\ }="8",
(1,-24)*{}="88",
(6,-24)*{}="810",
(6,-18)*{}="8107",
(12,-18)*{}="810711",
(12,-12)*{}="8107116",
(24,-12)*{}="810711613",
(24,-6)*{}="8107116130",
(30,-6)*{}="81071161300",
(30,0)*{}="810711613000",
(-4,-30)*{-4\ }="9",
(1,-30)*{}="99",
\ar@{-} "1";"2"^{}
\ar@{-} "1";"3"^{}
\ar@{-} "5";"55"^{}
\ar@{-} "6";"66"^{}
\ar@{-} "7";"77"^{}
\ar@{-} "8";"810"^{}
\ar@{-} "810";"8107"^{}
\ar@{-} "8107";"810711"^{}
\ar@{-} "810711";"8107116"^{}
\ar@{-} "8107116";"810711613"^{}
\ar@{-} "810711613";"8107116130"^{}
\ar@{-} "8107116130";"81071161300"^{}
\ar@{-} "81071161300";"810711613000"^{}
\ar@{-} "9";"99"^{}
\ar@{-} "10";"1010"^{}
\ar@{-} "11";"1111"^{}
\ar@{-} "12";"1212"^{}
\ar@{-} "13";"1313"^{}
\ar@{-} "14";"1414"^{}
\end{xy}
\]
\end{ex}
\nd
Note that drawing an extended Young diagram
$(y_r)_{r\in\mathbb{Z}_{\geq0}}$ in
$\mathbb{R}_{\geq0}\times \mathbb{R}_{\leq y_{\infty}}$,
if $y_r< y_{r+1}$ then
the points $(r+1,y_r)$ and $(r+1,y_{r+1})$ are corners.
\begin{defn}\cite{JMMO}
For an extended Young diagram
$(y_r)_{r\in\mathbb{Z}_{\geq0}}$, if $y_r< y_{r+1}$ then
the point $(r+1,y_r)$ is said to be a {\it convex corner} and $(r+1,y_{r+1})$
is said to be a {\it concave corner}. The point $(0,y_0)$ is also called a concave corner.
\end{defn}
\begin{ex}
Let $T$ be the
extended Young diagram as in Example \ref{ex-1}.
The points $(1,-3)$, $(2,-2)$, $(4,-1)$ and $(5,0)$ are convex corners and
$(0,-3)$, $(1,-2)$, $(2,-1)$, $(4,0)$ and $(5,1)$ are concave corners.
\end{ex}
Next, we consider a coloring of each corner in extended Young diagrams
by $I$.
\begin{defn}\cite{JMMO,KMM}\label{pi1-def}
\begin{enumerate}
\item
We define a map $\pi_{{\rm A}^{(1)}}:\mathbb{Z}\rightarrow \{1,2,\cdots,n\}=I$ as
\[
\pi_{{\rm A}^{(1)}}(\ell+rn)=\ell
\]
for any $r\in\mathbb{Z}$ and $\ell\in\{1,2,\cdots,n\}$.
\item
Defining a map
$\{1,2,\cdots,2n-2\}\rightarrow \{1,2,\cdots,n\}$
as
\[
\ell\mapsto \ell,\ 2n-\ell\mapsto \ell \qquad (2\leq \ell\leq n-1),
\]
\[
1\mapsto1,\ n\mapsto n
\]
and extend it to a map $\pi_{{\rm C}^{(1)}}:\mathbb{Z}\rightarrow\{1,2,\cdots,n\}=I$ by periodicity $2n-2$.
\item Defining a map $\{1,2,3,\cdots,2n-1\}\rightarrow \{1,2,\cdots,n\}$ as
\[
\ell\mapsto \ell,\ \ 2n-\ell\mapsto \ell \quad (1\leq \ell\leq n-1),\ \ n\mapsto n
\]
and extend it to the map $\pi_{{\rm A}^{(2)}}:\mathbb{Z}\rightarrow \{1,2,\cdots,n\}=I$ with periodicity $2n-1$.
\item Defining a map $\{1,2,3,\cdots,2n\}\rightarrow \{1,2,\cdots,n\}$ as
\[
\ell\mapsto \ell,\ \ 2n+1-\ell\mapsto \ell \quad (1\leq \ell\leq n),
\]
and extend it to the map $\pi_{{\rm D}^{(2)}}:\mathbb{Z}\rightarrow \{1,2,\cdots,n\}=I$ with periodicity $2n$.
\end{enumerate}
\end{defn}
Originally,
the maps $\pi_{{\rm A}^{(1)}}$, $\pi_{{\rm C}^{(1)}}$, $\pi_{{\rm A}^{(2)}}$ and $\pi_{{\rm D}^{(2)}}$ in this definition were
introduced to define actions of Chevalley generators (or Kashiwara operators) of type ${\rm A}^{(1)}_n$, ${\rm C}^{(1)}_n$, ${\rm A}^{(2)}_{2n}$
and ${\rm D}^{(2)}_{n+1}$ on extended Young diagrams. Each corner $(i,j)$ is colored by
$\pi_{{\rm A}^{(1)}}(i+j)$, $\pi_{{\rm C}^{(1)}}(i+j)$, $\pi_{{\rm A}^{(2)}}(i+j)$ and $\pi_{{\rm D}^{(2)}}(i+j)\in I$, respectively
and roughly speaking, a concave corner colored by $k\in I$
becomes a convex corner by an action of $f_k$ or $\tilde{f}_k$ \cite{JMMO, KMM}.
\subsection{Revised extended Young diagrams}
Next, we reviewed the notion of revised extended Young diagrams (REYD) introduced in \cite{Ka}, which are
used to express inequalities of ${\rm A}^{(2)}_{2n-2}$ and ${\rm C}^{(1)}_{n-1}$.
\begin{defn}\label{AEYD}\cite{Ka}
For $k\in I\setminus\{1\}$, we define ${\rm REYD}^{{\rm A}^{(2)}}_{k}$ as the set of sequences $(y_t)_{t\in\mathbb{Z}}$ such that
\begin{itemize}
\item[$(1)$] $y_t\in\mathbb{Z}$ for all $t\in\mathbb{Z}$,
\item[$(2)$] $y_{t}=k$ for $t\gg0$ and $y_t=k+t$ for $t\ll0$,
\item[$(3)$] for $t\in\mathbb{Z}$ such that $k+t\not\equiv 0$ (mod $2n-1$), it holds either $y_{t+1}=y_t$ or $y_{t+1}=y_t+1$,
\item[$(4)$] for $t\in\mathbb{Z}_{>0}$ such that $k+t\equiv 0$ (mod $2n-1$), we have $y_{t+1}\geq y_t$,
\item[$(5)$] for $t\in\mathbb{Z}_{<0}$ such that $k+t\equiv 0$ (mod $2n-1$), we have $y_{t+1}\leq y_t+1$.
\end{itemize}
\end{defn}
\begin{defn}\label{DEYD}\cite{Ka}
For $k\in I\setminus\{1,n\}$, we define ${\rm REYD}^{{\rm D}^{(2)}}_{k}$ as the set of sequences $(y_t)_{t\in\mathbb{Z}}$ such that
\begin{itemize}
\item[$(1)$] $y_t\in\mathbb{Z}$ for all $t\in\mathbb{Z}$,
\item[$(2)$] $y_{t}=k$ for $t\gg0$ and $y_t=k+t$ for $t\ll0$,
\item[$(3)$] for $t\in\mathbb{Z}$ such that $k+t\not\equiv 0, n$ (mod $2n$), it holds either $y_{t+1}=y_t$ or $y_{t+1}=y_t+1$,
\item[$(4)$] for $t\in\mathbb{Z}_{>0}$ such that $k+t\equiv 0$ or $n$ (mod $2n$), we have $y_{t+1}\geq y_t$,
\item[$(5)$] for $t\in\mathbb{Z}_{<0}$ such that $k+t\equiv 0$ or $n$ (mod $2n$), we have $y_{t+1}\leq y_t+1$.
\end{itemize}
\end{defn}
Just as in ordinary extended Young diagrams,
each element in ${\rm REYD}^{{\rm A}^{(2)}}_{k}$ and
${\rm REYD}^{{\rm D}^{(2)}}_{k}$
is described as an infinite diagram
drawn on $\mathbb{R}\times \mathbb{R}_{\leq k}$.
For example,
let $n=3$, $k=2$ and $T=(y_t)_{t\in\mathbb{Z}}$ be
the element in ${\rm REYD}^{{\rm A}^{(2)}}_{2}$
defined as
\[
y_{\ell}=\ell+2\ (\ell\leq -3),\ y_{-2}=0,\ y_{-1}=y_0=y_1=-2,\ y_2=y_3=-1,\ y_4=y_5=y_6=y_7=1,\ y_t=2 (t\geq 8).
\]
Then $T$ is described as
\begin{equation}\label{reydA-ex1}
\begin{xy}
(-42,-18) *{T=}="YY",
(-33,0) *{}="-6",
(-6,2) *{-1}="-1",
(-12,2) *{-2}="-2",
(-18,2) *{-3}="-3",
(-24,2) *{-4}="-4",
(-30,2) *{-5}="-5",
(-6,-1) *{}="-1a",
(-12,-1) *{}="-2a",
(-18,-1) *{}="-3a",
(-24,-1) *{}="-4a",
(-30,-1) *{}="-5a",
(0,0) *{}="1",
(56,0)*{}="2",
(0,-40)*{}="3",
(0,2)*{(0,2)}="4",
(6,2) *{1}="10",
(6,-1) *{}="1010",
(12,2) *{2}="11",
(12,-1) *{}="1111",
(18,2) *{3}="12",
(18,-1) *{}="1212",
(24,2) *{4}="13",
(24,-1) *{}="1313",
(30,2) *{5}="14",
(36,2) *{6}="6newnew",
(42,2) *{7}="7newnew",
(36,-1) *{}="6newnew-1",
(42,-1) *{}="7newnew-1",
(30,-1) *{}="1414",
(-3,-6)*{1\ }="5",
(1,-6)*{}="55",
(-3,-12)*{0\ }="6",
(-6,-12)*{}="6a",
(-12,-12)*{}="6aa",
(-12,-18)*{}="6aaa",
(-18,-18)*{}="6aaaa",
(-18,-24)*{}="st1",
(-24,-24)*{}="st2",
(-24,-30)*{}="st3",
(-30,-30)*{}="st4",
(-33,-33)*{\cdots}="stdot",
(1,-12)*{}="66",
(-6,-18)*{}="7",
(-6,-24)*{}="7new",
(-2,-16)*{-1\ }="7a",
(-2,-18)*{}="m1l",
(1,-18)*{}="m1r",
(-2,-23)*{-2\ }="8",
(1,-24)*{}="88",
(6,-18)*{}="8107",
(6,-24)*{}="8107new",
(12,-18)*{}="810711",
(12,-24)*{}="810711new",
(12,-12)*{}="8107116",
(24,-12)*{}="810711613",
(24,-18)*{}="810711613new",
(24,-6)*{}="8107116130",
(48,-6)*{}="8107116130new",
(48,2)*{8}="8107116130new2",
(30,-6)*{}="81071161300",
(30,0)*{}="810711613000",
(-4,-30)*{-3\ }="9",
(1,-30)*{}="99",
\ar@{-} "1";"-6"^{}
\ar@{-} "1";"2"^{}
\ar@{-} "1";"3"^{}
\ar@{-} "-1";"-1a"^{}
\ar@{-} "-2";"-2a"^{}
\ar@{-} "-3";"-3a"^{}
\ar@{-} "-4";"-4a"^{}
\ar@{-} "-5";"-5a"^{}
\ar@{-} "5";"55"^{}
\ar@{-} "6";"66"^{}
\ar@{-} "6newnew";"6newnew-1"^{}
\ar@{-} "7newnew";"7newnew-1"^{}
\ar@{-} "m1r";"m1l"^{}
\ar@{-} "810711";"810711613new"^{}
\ar@{-} "810711613";"810711613new"^{}
\ar@{-} "810711613";"8107116130"^{}
\ar@{-} "8107116130";"8107116130new"^{}
\ar@{-} "8107116130new2";"8107116130new"^{}
\ar@{-} "7new";"6a"^{}
\ar@{-} "7new";"810711new"^{}
\ar@{-} "810711";"810711new"^{}
\ar@{-} "6aa";"6a"^{}
\ar@{-} "6aaa";"6aa"^{}
\ar@{-} "6aaaa";"6aaa"^{}
\ar@{-} "st1";"6aaaa"^{}
\ar@{-} "st1";"st2"^{}
\ar@{-} "st2";"st3"^{}
\ar@{-} "st3";"st4"^{}
\ar@{-} "8";"88"^{}
\ar@{-} "9";"99"^{}
\ar@{-} "10";"1010"^{}
\ar@{-} "11";"1111"^{}
\ar@{-} "12";"1212"^{}
\ar@{-} "13";"1313"^{}
\ar@{-} "14";"1414"^{}
\end{xy}
\end{equation}
As an analogy of concave corners and convex corners of extended Young diagrams,
one can define admissible points and removable points of REYD:
\begin{defn}\label{ad-rem-pt}\cite{Ka}
We set ${\rm REYD}_{k}={\rm REYD}^{{\rm A}^{(2)}}_{k}$ or ${\rm REYD}_{k}={\rm REYD}^{{\rm D}^{(2)}}_{k}$ in Definition \ref{AEYD}, \ref{DEYD} .
Let $T=(y_t)_{t\in\mathbb{Z}}\in{\rm REYD}_{k}$ and $i\in\mathbb{Z}$.
\begin{enumerate}
\item Let $T'=(y_t')_{t\in\mathbb{Z}}$ be the sequence of integers such that
$y_i'=y_i-1$ and $y_t'=y_t$ $(t\neq i)$.
The point $(i,y_i)$ is said to be an admissible point of $T$ if $T'\in{\rm REYD}_{k}$.
\item Let $T''=(y_t'')_{t\in\mathbb{Z}}$ be the sequence of integers such that
$y_{i-1}''=y_{i-1}+1$ and $y_t''=y_t$ $(t\neq i-1)$.
The point $(i,y_{i-1})$ is said to be
a removable point of $T$
if $T''\in{\rm REYD}_{k}$.
\end{enumerate}
\end{defn}
\begin{defn}\label{ad-rem-pt2}\cite{Ka}
Let $T=(y_t)_{t\in\mathbb{Z}}\in{\rm REYD}^{{\rm A}^{(2)}}_{k}$ and $i\in\mathbb{Z}$.
\begin{enumerate}
\item
The point $(i,y_i)$ is said to be a double $1$-admissible point of $T$
if $y_{i-1}<y_i=y_{i+1}$ and it holds either
\begin{itemize}
\item $i+k\equiv 1$ (mod $2n-1$) and $i<0$ or
\item $i+k\equiv 0$ (mod $2n-1$) and $i>0$.
\end{itemize}
\item
The point $(i,y_{i-1})$ is said to be a double $1$-removable point of $T$
if $y_{i-2}=y_{i-1}<y_{i}$ and it holds
either
\begin{itemize}
\item
$i+k-1\equiv 1$ (mod $2n-1$) and $i>1$ or
\item
$i+k-1\equiv 0$ (mod $2n-1$) and $i<1$.
\end{itemize}
\item Other admissible (resp. removable) points $(i,y_i)$ (resp. $(i,y_{i-1})$) than (i) (resp. (ii)) are said to be single $\pi_{{\rm A}^{(2)}}(i+k)$-admissible (resp. $\pi_{{\rm A}^{(2)}}(i+k-1)$-removable) points.
\end{enumerate}
\end{defn}
\begin{defn}\label{ad-rem-pt3}\cite{Ka}
Let $T=(y_t)_{t\in\mathbb{Z}}\in{\rm REYD}^{{\rm D}^{(2)}}_{k}$, $i\in\mathbb{Z}$
and $\ell\in\{0,n\}$.
\begin{enumerate}
\item The point $(i,y_i)$ is said to be a double $\pi_{{\rm D}^{(2)}}(\ell)$-admissible point of $T$
if $y_{i-1}<y_i=y_{i+1}$ and it holds either
\begin{itemize}
\item
$i+k\equiv \ell+1$ (mod $2n$) and $i<0$ or
\item
$i+k\equiv \ell$ (mod $2n$) and $i>0$.
\end{itemize}
\item The point $(i,y_{i-1})$ is said to be a double $\pi_{{\rm D}^{(2)}}(\ell)$-removable point of $T$
if $y_{i-2}=y_{i-1}<y_{i}$ and it holds
either
\begin{itemize}
\item
$i+k-1\equiv \ell+1$ (mod $2n$) and $i>1$ or
\item
$i+k-1\equiv \ell$ (mod $2n$) and $i<1$.
\end{itemize}
\item Other admissible (resp. removable) points $(i,y_i)$ (resp. $(i,y_{i-1})$) than (i) (resp. (ii)) are said to be single $\pi_{{\rm D}^{(2)}}(i+k)$-admissible (resp. $\pi_{{\rm D}^{(2)}}(i+k-1)$-removable) points.
\end{enumerate}
\end{defn}
Note that in Definition \ref{ad-rem-pt2} and \ref{ad-rem-pt3}, a double admissible (resp. removable) point
$(i,y_i)$ (resp. $(i,y_{i-1})$) is an admissible (resp. a removable) point in Definition \ref{ad-rem-pt} by $y_{i-1}<y_i=y_{i+1}$
(resp. $y_{i-2}=y_{i-1}<y_{i}$).
\begin{ex}
Let us consider the revised extended Young diagram $T$ in (\ref{reydA-ex1}).
The points $(-2,0)$, $(-1,-2)$ and $(4,1)$ are single $1$-admissible points in $T$.
The point $(2,-1)$ is a single $2$-admissible point and
the point $(8,2)$ is a double $1$-admissible point in $T$.
We also see that the point $(2,-2)$ is a single $3$-removable point in $T$.
The point $(4,-1)$ is a single $1$-removable point and
the point $(8,1)$ is a single $2$-removable point in $T$.
\end{ex}
\subsection{Young walls}
Following \cite{Kang}, let us review the notion of Young walls of type ${\rm A}^{(2)}_{2n-2}$ and ${\rm D}^{(2)}_n$.
We consider
$I$-colored blocks of two different
shapes:
\begin{enumerate}
\item[(1)] block with unit width, unit height and unit thickness:
\[
\begin{xy}
(3,3) *{j}="0",
(0,0) *{}="1",
(6,0)*{}="2",
(6,6)*{}="3",
(0,6)*{}="4",
(3,9)*{}="5",
(9,9)*{}="6",
(9,3)*{}="7",
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\ar@{-} "5";"4"^{}
\ar@{-} "5";"6"^{}
\ar@{-} "3";"6"^{}
\ar@{-} "2";"7"^{}
\ar@{-} "6";"7"^{}
\end{xy}
\]
\item[(2)] block with unit width, half-unit height and unit thickness:
\[
\begin{xy}
(3,1.5) *{j}="0",
(0,0) *{}="1",
(6,0)*{}="2",
(6,3)*{}="3",
(0,3)*{}="4",
(3,6)*{}="5",
(9,6)*{}="6",
(9,3)*{}="7",
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\ar@{-} "5";"4"^{}
\ar@{-} "5";"6"^{}
\ar@{-} "3";"6"^{}
\ar@{-} "2";"7"^{}
\ar@{-} "6";"7"^{}
\end{xy}
\]
\end{enumerate}
The block (1) with color $j\in I$ is simply expressed as
\begin{equation}\label{smpl1}
\begin{xy}
(3,3) *{j}="0",
(0,0) *{}="1",
(6,0)*{}="2",
(6,6)*{}="3",
(0,6)*{}="4",
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\end{xy}
\end{equation}
and (2) with color $j\in I$ is expressed as
\begin{equation}\label{smpl2}
\begin{xy}
(1.5,1.5) *{\ \ j}="0",
(0,0) *{}="1",
(6,0)*{}="2",
(6,3.5)*{}="3",
(0,3.5)*{}="4",
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\end{xy}
\end{equation}
We call
the blocks (\ref{smpl1}) and (\ref{smpl2}) $j$-blocks.
For example, if colored blocks are stacked as follows
\[
\begin{xy}
(-9.5,1.5) *{\ 1}="000",
(-3.5,1.5) *{\ 1}="00",
(1.5,1.5) *{\ \ 1}="0",
(1.5,6.5) *{\ \ 2}="02",
(-4,6.5) *{\ \ 2}="002",
(1.5,12.5) *{\ \ 3}="03",
(1.5,18.5) *{\ \ 4}="04",
(9,3)*{}="2-a",
(9,6.5)*{}="3-a",
(9,12.5)*{}="3-1-a",
(9,18.5)*{}="3-2-a",
(9,24.5)*{}="3-3-a",
(3,24.5)*{}="4-3-a",
(-3,12.5)*{}="5-1-a",
(0,12.5)*{}="5-1-ab",
(-9,6.5)*{}="7-a",
(-6,6.5)*{}="7-ab",
(0,0) *{}="1",
(6,0)*{}="2",
(6,3.5)*{}="3",
(0,3.5)*{}="4",
(-6,3.5)*{}="5",
(-6,0)*{}="6",
(-12,3.5)*{}="7",
(-12,0)*{}="8",
(6,9.5)*{}="3-1",
(6,15.5)*{}="3-2",
(6,21.5)*{}="3-3",
(0,9.5)*{}="4-1",
(0,15.5)*{}="4-2",
(0,21.5)*{}="4-3",
(-6,9.5)*{}="5-1",
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\ar@{-} "5";"6"^{}
\ar@{-} "5";"4"^{}
\ar@{-} "1";"6"^{}
\ar@{-} "7";"8"^{}
\ar@{-} "7";"5"^{}
\ar@{-} "6";"8"^{}
\ar@{-} "3";"3-1"^{}
\ar@{-} "4-1";"3-1"^{}
\ar@{-} "4-1";"4"^{}
\ar@{-} "5-1";"5"^{}
\ar@{-} "4-1";"5-1"^{}
\ar@{-} "4-1";"4-2"^{}
\ar@{-} "3-2";"3-1"^{}
\ar@{-} "3-2";"4-2"^{}
\ar@{-} "3-2";"3-3"^{}
\ar@{-} "4-3";"4-2"^{}
\ar@{-} "4-3";"3-3"^{}
\ar@{-} "2";"2-a"^{}
\ar@{-} "3";"3-a"^{}
\ar@{-} "3-1";"3-1-a"^{}
\ar@{-} "3-2";"3-2-a"^{}
\ar@{-} "3-3";"3-3-a"^{}
\ar@{-} "2-a";"3-3-a"^{}
\ar@{-} "4-3";"4-3-a"^{}
\ar@{-} "3-3-a";"4-3-a"^{}
\ar@{-} "5-1";"5-1-a"^{}
\ar@{-} "5-1-ab";"5-1-a"^{}
\ar@{-} "7";"7-a"^{}
\ar@{-} "7-ab";"7-a"^{}
\end{xy}
\]
then
it is simply expressed as
\begin{equation}\label{smpl3}
\begin{xy}
(-9.5,1.5) *{\ 1}="000",
(-3.5,1.5) *{\ 1}="00",
(1.5,1.5) *{\ \ 1}="0",
(1.5,6.5) *{\ \ 2}="02",
(-4,6.5) *{\ \ 2}="002",
(1.5,12.5) *{\ \ 3}="03",
(1.5,18.5) *{\ \ 4}="04",
(0,0) *{}="1",
(6,0)*{}="2",
(6,3.5)*{}="3",
(0,3.5)*{}="4",
(-6,3.5)*{}="5",
(-6,0)*{}="6",
(-12,3.5)*{}="7",
(-12,0)*{}="8",
(6,9.5)*{}="3-1",
(6,15.5)*{}="3-2",
(6,21.5)*{}="3-3",
(0,9.5)*{}="4-1",
(0,15.5)*{}="4-2",
(0,21.5)*{}="4-3",
(-6,9.5)*{}="5-1",
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\ar@{-} "5";"6"^{}
\ar@{-} "5";"4"^{}
\ar@{-} "1";"6"^{}
\ar@{-} "7";"8"^{}
\ar@{-} "7";"5"^{}
\ar@{-} "6";"8"^{}
\ar@{-} "3";"3-1"^{}
\ar@{-} "4-1";"3-1"^{}
\ar@{-} "4-1";"4"^{}
\ar@{-} "5-1";"5"^{}
\ar@{-} "4-1";"5-1"^{}
\ar@{-} "4-1";"4-2"^{}
\ar@{-} "3-2";"3-1"^{}
\ar@{-} "3-2";"4-2"^{}
\ar@{-} "3-2";"3-3"^{}
\ar@{-} "4-3";"4-2"^{}
\ar@{-} "4-3";"3-3"^{}
\end{xy}
\end{equation}
A diagram formed by stacking blocks is called a wall.
For $X={\rm A}^{(2)}_{2n-2}$, ${\rm D}^{(2)}_{n}$
and $\lambda\in P^+$ of level $1$,
we define a wall $Y_{\lambda}$ called a {\it ground state wall}:
\begin{itemize}
\item For $X={\rm A}^{(2)}_{2n-2}$ and $\lambda=\Lambda_1$, one defines
\[
Y_{\Lambda_1}=
\begin{xy}
(-15.5,1.5) *{\ \cdots}="0000",
(-9.5,1.5) *{\ 1}="000",
(-3.5,1.5) *{\ 1}="00",
(1.5,1.5) *{\ \ 1}="0",
(0,0) *{}="1",
(6,0)*{}="2",
(6,3.5)*{}="3",
(0,3.5)*{}="4",
(-6,3.5)*{}="5",
(-6,0)*{}="6",
(-12,3.5)*{}="7",
(-12,0)*{}="8",
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\ar@{-} "5";"6"^{}
\ar@{-} "5";"4"^{}
\ar@{-} "1";"6"^{}
\ar@{-} "7";"8"^{}
\ar@{-} "7";"5"^{}
\ar@{-} "6";"8"^{}
\end{xy}
\]
Here, the wall $Y_{\Lambda_1}$ has infinitely many $1$-blocks of half-unit height
and
extends infinitely to the left.
\item
For $X={\rm D}^{(2)}_{n}$ and $\lambda=\Lambda_1$, $\Lambda_n$, one defines
\[
Y_{\Lambda_1}=
\begin{xy}
(-15.5,1.5) *{\ \cdots}="0000",
(-9.5,1.5) *{\ 1}="000",
(-3.5,1.5) *{\ 1}="00",
(1.5,1.5) *{\ \ 1}="0",
(0,0) *{}="1",
(6,0)*{}="2",
(6,3.5)*{}="3",
(0,3.5)*{}="4",
(-6,3.5)*{}="5",
(-6,0)*{}="6",
(-12,3.5)*{}="7",
(-12,0)*{}="8",
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\ar@{-} "5";"6"^{}
\ar@{-} "5";"4"^{}
\ar@{-} "1";"6"^{}
\ar@{-} "7";"8"^{}
\ar@{-} "7";"5"^{}
\ar@{-} "6";"8"^{}
\end{xy}
\]
and
\[
Y_{\Lambda_n}=
\begin{xy}
(-15.5,1.5) *{\ \cdots}="0000",
(-9.5,1.5) *{\ n}="000",
(-3.5,1.5) *{\ n}="00",
(1.5,1.5) *{\ \ n}="0",
(0,0) *{}="1",
(6,0)*{}="2",
(6,3.5)*{}="3",
(0,3.5)*{}="4",
(-6,3.5)*{}="5",
(-6,0)*{}="6",
(-12,3.5)*{}="7",
(-12,0)*{}="8",
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\ar@{-} "5";"6"^{}
\ar@{-} "5";"4"^{}
\ar@{-} "1";"6"^{}
\ar@{-} "7";"8"^{}
\ar@{-} "7";"5"^{}
\ar@{-} "6";"8"^{}
\end{xy}
\]
\end{itemize}
\begin{defn}\cite{Kang}\label{def-YW}
For $X={\rm A}^{(2)}_{2n-2}$, ${\rm D}^{(2)}_{n}$
and $\lambda\in P^+$ of level $1$,
a wall $Y$ is called a {\it Young wall} of ground state $\lambda$
of type $X$ if it satisfies the following:
\begin{enumerate}
\item The wall $Y$ is obtained
by stacking finitely many colored blocks
on the ground state wall $Y_{\lambda}$.
\item The colored blocks are stacked following the patterns we give below for each type and $\lambda$.
\item Let $h_j$ be the height of $j$-th column of the wall $Y$ from the right. Then we get $h_j\geq h_{j+1}$
for all $j\in\mathbb{Z}_{\geq1}$.
\end{enumerate}
The patterns mentioned in (ii) are as follows:
In the case ${\rm A}^{(2)}_{2n-2}$ and $\lambda=\Lambda_1$:
\[
\begin{xy}
(-15.5,-2) *{\ 1}="000-1",
(-9.5,-2) *{\ 1}="00-1",
(-3.5,-2) *{\ 1}="0-1",
(1.5,-2) *{\ \ 1}="0-1",
(-15.5,1.5) *{\ 1}="0000",
(-9.5,1.5) *{\ 1}="000",
(-3.5,1.5) *{\ 1}="00",
(1.5,1.5) *{\ \ 1}="0",
(1.5,6.5) *{\ \ 2}="02",
(-4,6.5) *{\ \ 2}="002",
(-10,6.5) *{\ \ 2}="0002",
(-16,6.5) *{\ \ 2}="00002",
(1.5,12.5) *{\ \ 3}="03",
(-4,12.5) *{\ \ 3}="003",
(-10,12.5) *{\ \ 3}="0003",
(-16,12.5) *{\ \ 3}="00003",
(1.5,20.5) *{\ \ \vdots}="04",
(-4,20.5) *{\ \ \vdots}="004",
(-10,20.5) *{\ \ \vdots}="0004",
(-16,20.5) *{\ \ \vdots}="00004",
(1.5,27.5) *{\ \ _{n-1}}="05",
(-4,27.5) *{\ \ _{n-1}}="005",
(-10,27.5) *{\ \ _{n-1}}="0005",
(-16,27.5) *{\ \ _{n-1}}="00005",
(1.5,33.5) *{\ \ n}="06",
(-4,33.5) *{\ \ n}="006",
(-25,34) *{\cdots}="dots",
(-10,33.5) *{\ \ n}="0006",
(-16,33.5) *{\ \ n}="00006",
(1.5,39.5) *{\ \ _{n-1}}="07",
(-4,39.5) *{\ \ _{n-1}}="007",
(-10,39.5) *{\ \ _{n-1}}="0007",
(-16,39.5) *{\ \ _{n-1}}="00007",
(1.5,48) *{\ \ \vdots}="08",
(-4,48) *{\ \ \vdots}="008",
(-10,48) *{\ \ \vdots}="0008",
(-16,48) *{\ \ \vdots}="00008",
(1.5,55) *{\ \ 2}="09",
(-4,55) *{\ \ 2}="009",
(-10,55) *{\ \ 2}="0009",
(-16,55) *{\ \ 2}="00009",
(1.5,60) *{\ \ 1}="010",
(-4,60) *{\ \ 1}="0010",
(-10,60) *{\ \ 1}="00010",
(-16,60) *{\ \ 1}="000010",
(1.5,64) *{\ \ 1}="011",
(-4,64) *{\ \ 1}="0011",
(-10,64) *{\ \ 1}="00011",
(-16,64) *{\ \ 1}="000011",
(1.5,69) *{\ \ 2}="012",
(-4,69) *{\ \ 2}="0012",
(-10,69) *{\ \ 2}="00012",
(-16,69) *{\ \ 2}="000012",
(-18,-3.5) *{}="0.5-l",
(-20,-3.5) *{}="0-l",
(-20,0) *{}="1-l",
(-20,3.5) *{}="2-l",
(-20,9.5) *{}="3-l",
(-20,15.5) *{}="4-l",
(-20,24.5) *{}="5-l",
(-20,30.5) *{}="6-l",
(-20,36.5) *{}="7-l",
(-20,42.5) *{}="8-l",
(-20,52) *{}="9-l",
(-20,58) *{}="10-l",
(-20,62) *{}="11-l",
(-20,66) *{}="12-l",
(-20,72) *{}="13-l",
(-18,74) *{}="13.5-l",
(6,-3.5)*{}="r--1",
(6,0)*{}="r",
(6,3.5)*{}="r-0",
(0,-3.5) *{}="b1",
(-6,-3.5)*{}="b2",
(-12,-3.5)*{}="b3",
(0,74) *{}="t1",
(-6,74)*{}="t2",
(-12,74)*{}="t3",
(6,9.5)*{}="r-1",
(6,15.5)*{}="r-2",
(6,24.5)*{}="r-3",
(6,30.5)*{}="r-4",
(6,36.5)*{}="r-5",
(6,42.5)*{}="r-6",
(6,52)*{}="r-7",
(6,58)*{}="r-8",
(6,62)*{}="r-9",
(6,66)*{}="r-10",
(6,72)*{}="r-11",
(6,74)*{}="r-11.5",
\ar@{-} "b1";"t1"^{}
\ar@{-} "b2";"t2"^{}
\ar@{-} "b3";"t3"^{}
\ar@{-} "0.5-l";"13.5-l"^{}
\ar@{-} "r-11.5";"r--1"^{}
\ar@{-} "0-l";"r--1"^{}
\ar@{-} "1-l";"r"^{}
\ar@{-} "2-l";"r-0"^{}
\ar@{-} "3-l";"r-1"^{}
\ar@{-} "4-l";"r-2"^{}
\ar@{-} "5-l";"r-3"^{}
\ar@{-} "6-l";"r-4"^{}
\ar@{-} "7-l";"r-5"^{}
\ar@{-} "8-l";"r-6"^{}
\ar@{-} "9-l";"r-7"^{}
\ar@{-} "10-l";"r-8"^{}
\ar@{-} "11-l";"r-9"^{}
\ar@{-} "12-l";"r-10"^{}
\ar@{-} "13-l";"r-11"^{}
\end{xy}
\]
In the case $X={\rm D}^{(2)}_{n}$,
\[
\begin{xy}
(-40,35) *{\lambda=\Lambda_1:}="weight",
(-15.5,-2) *{\ 1}="000-1",
(-9.5,-2) *{\ 1}="00-1",
(-3.5,-2) *{\ 1}="0-1",
(1.5,-2) *{\ \ 1}="0-1",
(-15.5,1.5) *{\ 1}="0000",
(-9.5,1.5) *{\ 1}="000",
(-3.5,1.5) *{\ 1}="00",
(1.5,1.5) *{\ \ 1}="0",
(1.5,6.5) *{\ \ 2}="02",
(-4,6.5) *{\ \ 2}="002",
(-10,6.5) *{\ \ 2}="0002",
(-16,6.5) *{\ \ 2}="00002",
(1.5,12.5) *{\ \ 3}="03",
(-4,12.5) *{\ \ 3}="003",
(-10,12.5) *{\ \ 3}="0003",
(-16,12.5) *{\ \ 3}="00003",
(1.5,20.5) *{\ \ \vdots}="04",
(-4,20.5) *{\ \ \vdots}="004",
(-10,20.5) *{\ \ \vdots}="0004",
(-16,20.5) *{\ \ \vdots}="00004",
(1.5,27.5) *{\ \ _{n-1}}="05",
(-4,27.5) *{\ \ _{n-1}}="005",
(-10,27.5) *{\ \ _{n-1}}="0005",
(-16,27.5) *{\ \ _{n-1}}="00005",
(1.5,32) *{\ \ _n}="06",
(-4,32) *{\ \ _n}="006",
(-25,34) *{\cdots}="dots",
(-10,32) *{\ \ _n}="0006",
(-16,32) *{\ \ _n}="00006",
(1.5,35) *{\ \ _n}="06.5",
(-4,35) *{\ \ _n}="006.5",
(-10,35) *{\ \ _n}="0006.5",
(-16,35) *{\ \ _n}="00006.5",
(1.5,39.5) *{\ \ _{n-1}}="07",
(-4,39.5) *{\ \ _{n-1}}="007",
(-10,39.5) *{\ \ _{n-1}}="0007",
(-16,39.5) *{\ \ _{n-1}}="00007",
(1.5,48) *{\ \ \vdots}="08",
(-4,48) *{\ \ \vdots}="008",
(-10,48) *{\ \ \vdots}="0008",
(-16,48) *{\ \ \vdots}="00008",
(1.5,55) *{\ \ 2}="09",
(-4,55) *{\ \ 2}="009",
(-10,55) *{\ \ 2}="0009",
(-16,55) *{\ \ 2}="00009",
(1.5,60) *{\ \ 1}="010",
(-4,60) *{\ \ 1}="0010",
(-10,60) *{\ \ 1}="00010",
(-16,60) *{\ \ 1}="000010",
(1.5,64) *{\ \ 1}="011",
(-4,64) *{\ \ 1}="0011",
(-10,64) *{\ \ 1}="00011",
(-16,64) *{\ \ 1}="000011",
(1.5,69) *{\ \ 2}="012",
(-4,69) *{\ \ 2}="0012",
(-10,69) *{\ \ 2}="00012",
(-16,69) *{\ \ 2}="000012",
(-18,-3.5) *{}="0.5-l",
(-20,-3.5) *{}="0-l",
(-20,0) *{}="1-l",
(-20,3.5) *{}="2-l",
(-20,9.5) *{}="3-l",
(-20,15.5) *{}="4-l",
(-20,24.5) *{}="5-l",
(-20,30.5) *{}="6-l",
(-20,33.5) *{}="65-l",
(-20,36.5) *{}="7-l",
(-20,42.5) *{}="8-l",
(-20,52) *{}="9-l",
(-20,58) *{}="10-l",
(-20,62) *{}="11-l",
(-20,66) *{}="12-l",
(-20,72) *{}="13-l",
(-18,74) *{}="13.5-l",
(6,-3.5)*{}="r--1",
(6,0)*{}="r",
(6,3.5)*{}="r-0",
(0,-3.5) *{}="b1",
(-6,-3.5)*{}="b2",
(-12,-3.5)*{}="b3",
(0,74) *{}="t1",
(-6,74)*{}="t2",
(-12,74)*{}="t3",
(6,9.5)*{}="r-1",
(6,15.5)*{}="r-2",
(6,24.5)*{}="r-3",
(6,30.5)*{}="r-4",
(6,33.5)*{}="r-45",
(6,36.5)*{}="r-5",
(6,42.5)*{}="r-6",
(6,52)*{}="r-7",
(6,58)*{}="r-8",
(6,62)*{}="r-9",
(6,66)*{}="r-10",
(6,72)*{}="r-11",
(6,74)*{}="r-11.5",
\ar@{-} "b1";"t1"^{}
\ar@{-} "b2";"t2"^{}
\ar@{-} "b3";"t3"^{}
\ar@{-} "0.5-l";"13.5-l"^{}
\ar@{-} "r-11.5";"r--1"^{}
\ar@{-} "0-l";"r--1"^{}
\ar@{-} "1-l";"r"^{}
\ar@{-} "2-l";"r-0"^{}
\ar@{-} "3-l";"r-1"^{}
\ar@{-} "4-l";"r-2"^{}
\ar@{-} "5-l";"r-3"^{}
\ar@{-} "6-l";"r-4"^{}
\ar@{-} "65-l";"r-45"^{}
\ar@{-} "7-l";"r-5"^{}
\ar@{-} "8-l";"r-6"^{}
\ar@{-} "9-l";"r-7"^{}
\ar@{-} "10-l";"r-8"^{}
\ar@{-} "11-l";"r-9"^{}
\ar@{-} "12-l";"r-10"^{}
\ar@{-} "13-l";"r-11"^{}
\end{xy}\qquad \qquad
\begin{xy}
(-40,35) *{\lambda=\Lambda_n:}="weight",
(-15.5,-2) *{\ _n}="000-1",
(-9.5,-2) *{\ _n}="00-1",
(-3.5,-2) *{\ _n}="0-1",
(1.5,-2) *{\ \ _n}="0-1",
(-15.5,1.5) *{\ _n}="0000",
(-9.5,1.5) *{\ _n}="000",
(-3.5,1.5) *{\ _n}="00",
(1.5,1.5) *{\ \ _n}="0",
(1.5,6.5) *{\ \ _{n-1}}="02",
(-4,6.5) *{\ \ _{n-1}}="002",
(-10,6.5) *{\ \ _{n-1}}="0002",
(-16,6.5) *{\ \ _{n-1}}="00002",
(1.5,12.5) *{\ \ _{n-2}}="03",
(-4,12.5) *{\ \ _{n-2}}="003",
(-10,12.5) *{\ \ _{n-2}}="0003",
(-16,12.5) *{\ \ _{n-2}}="00003",
(1.5,20.5) *{\ \ \vdots}="04",
(-4,20.5) *{\ \ \vdots}="004",
(-10,20.5) *{\ \ \vdots}="0004",
(-16,20.5) *{\ \ \vdots}="00004",
(1.5,27.5) *{\ \ 2}="05",
(-4,27.5) *{\ \ 2}="005",
(-10,27.5) *{\ \ 2}="0005",
(-16,27.5) *{\ \ 2}="00005",
(1.5,32) *{\ \ 1}="06",
(-4,32) *{\ \ 1}="006",
(-25,34) *{\cdots}="dots",
(-10,32) *{\ \ 1}="0006",
(-16,32) *{\ \ 1}="00006",
(1.5,35) *{\ \ 1}="06.5",
(-4,35) *{\ \ 1}="006.5",
(-10,35) *{\ \ 1}="0006.5",
(-16,35) *{\ \ 1}="00006.5",
(1.5,39.5) *{\ \ 2}="07",
(-4,39.5) *{\ \ 2}="007",
(-10,39.5) *{\ \ 2}="0007",
(-16,39.5) *{\ \ 2}="00007",
(1.5,48) *{\ \ \vdots}="08",
(-4,48) *{\ \ \vdots}="008",
(-10,48) *{\ \ \vdots}="0008",
(-16,48) *{\ \ \vdots}="00008",
(1.5,55) *{\ \ _{n-1}}="09",
(-4,55) *{\ \ _{n-1}}="009",
(-10,55) *{\ \ _{n-1}}="0009",
(-16,55) *{\ \ _{n-1}}="00009",
(1.5,60) *{\ \ _{n}}="010",
(-4,60) *{\ \ _{n}}="0010",
(-10,60) *{\ \ _{n}}="00010",
(-16,60) *{\ \ _{n}}="000010",
(1.5,64) *{\ \ _{n}}="011",
(-4,64) *{\ \ _{n}}="0011",
(-10,64) *{\ \ _{n}}="00011",
(-16,64) *{\ \ _{n}}="000011",
(1.5,69) *{\ \ _{n-1}}="012",
(-4,69) *{\ \ _{n-1}}="0012",
(-10,69) *{\ \ _{n-1}}="00012",
(-16,69) *{\ \ _{n-1}}="000012",
(-18,-3.5) *{}="0.5-l",
(-20,-3.5) *{}="0-l",
(-20,0) *{}="1-l",
(-20,3.5) *{}="2-l",
(-20,9.5) *{}="3-l",
(-20,15.5) *{}="4-l",
(-20,24.5) *{}="5-l",
(-20,30.5) *{}="6-l",
(-20,33.5) *{}="65-l",
(-20,36.5) *{}="7-l",
(-20,42.5) *{}="8-l",
(-20,52) *{}="9-l",
(-20,58) *{}="10-l",
(-20,62) *{}="11-l",
(-20,66) *{}="12-l",
(-20,72) *{}="13-l",
(-18,74) *{}="13.5-l",
(6,-3.5)*{}="r--1",
(6,0)*{}="r",
(6,3.5)*{}="r-0",
(0,-3.5) *{}="b1",
(-6,-3.5)*{}="b2",
(-12,-3.5)*{}="b3",
(0,74) *{}="t1",
(-6,74)*{}="t2",
(-12,74)*{}="t3",
(6,9.5)*{}="r-1",
(6,15.5)*{}="r-2",
(6,24.5)*{}="r-3",
(6,30.5)*{}="r-4",
(6,33.5)*{}="r-45",
(6,36.5)*{}="r-5",
(6,42.5)*{}="r-6",
(6,52)*{}="r-7",
(6,58)*{}="r-8",
(6,62)*{}="r-9",
(6,66)*{}="r-10",
(6,72)*{}="r-11",
(6,74)*{}="r-11.5",
\ar@{-} "b1";"t1"^{}
\ar@{-} "b2";"t2"^{}
\ar@{-} "b3";"t3"^{}
\ar@{-} "0.5-l";"13.5-l"^{}
\ar@{-} "r-11.5";"r--1"^{}
\ar@{-} "0-l";"r--1"^{}
\ar@{-} "1-l";"r"^{}
\ar@{-} "2-l";"r-0"^{}
\ar@{-} "3-l";"r-1"^{}
\ar@{-} "4-l";"r-2"^{}
\ar@{-} "5-l";"r-3"^{}
\ar@{-} "6-l";"r-4"^{}
\ar@{-} "65-l";"r-45"^{}
\ar@{-} "7-l";"r-5"^{}
\ar@{-} "8-l";"r-6"^{}
\ar@{-} "9-l";"r-7"^{}
\ar@{-} "10-l";"r-8"^{}
\ar@{-} "11-l";"r-9"^{}
\ar@{-} "12-l";"r-10"^{}
\ar@{-} "13-l";"r-11"^{}
\end{xy}
\]
Note that the first row of each pattern from the bottom is
the ground state wall.
\end{defn}
\begin{ex}\label{ex-YW}
The following is a Young wall of ground state $\Lambda_1$ of type ${\rm A}^{(2)}_{4}$:
\begin{equation*}
\begin{xy}
(-15.5,-2) *{\ 1}="000-1",
(-9.5,-2) *{\ 1}="00-1",
(-3.5,-2) *{\ 1}="0-1",
(1.5,-2) *{\ \ 1}="0-1",
(-21.5,-2) *{\dots}="00000",
(-9.5,1.5) *{\ 1}="000",
(-3.5,1.5) *{\ 1}="00",
(1.5,1.5) *{\ \ 1}="0",
(1.5,6.5) *{\ \ 2}="02",
(-4,6.5) *{\ \ 2}="002",
(1.5,12.5) *{\ \ 3}="03",
(1.5,18.5) *{\ \ 2}="04",
(0,0) *{}="1",
(0,-3.5) *{}="1-u",
(6,0)*{}="2",
(6,-3.5)*{}="2-u",
(6,3.5)*{}="3",
(0,3.5)*{}="4",
(-6,3.5)*{}="5",
(-6,0)*{}="6",
(-6,-3.5)*{}="6-u",
(-12,3.5)*{}="7",
(-12,0)*{}="8",
(-12,-3.6)*{}="8-u",
(-18,3.5)*{}="9",
(-18,0)*{}="10-a",
(-18,0)*{}="10",
(-18,-3.5)*{}="10-u",
(6,9.5)*{}="3-1",
(6,15.5)*{}="3-2",
(6,21.5)*{}="3-3",
(0,9.5)*{}="4-1",
(0,15.5)*{}="4-2",
(0,21.5)*{}="4-3",
(-6,9.5)*{}="5-1",
\ar@{-} "8";"10-a"^{}
\ar@{-} "10-u";"10-a"^{}
\ar@{-} "10-u";"2-u"^{}
\ar@{-} "8";"8-u"^{}
\ar@{-} "6";"6-u"^{}
\ar@{-} "2";"2-u"^{}
\ar@{-} "1";"1-u"^{}
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\ar@{-} "5";"6"^{}
\ar@{-} "5";"4"^{}
\ar@{-} "1";"6"^{}
\ar@{-} "7";"8"^{}
\ar@{-} "7";"5"^{}
\ar@{-} "6";"8"^{}
\ar@{-} "3";"3-1"^{}
\ar@{-} "4-1";"3-1"^{}
\ar@{-} "4-1";"4"^{}
\ar@{-} "5-1";"5"^{}
\ar@{-} "4-1";"5-1"^{}
\ar@{-} "4-1";"4-2"^{}
\ar@{-} "3-2";"3-1"^{}
\ar@{-} "3-2";"4-2"^{}
\ar@{-} "3-2";"3-3"^{}
\ar@{-} "4-3";"4-2"^{}
\ar@{-} "4-3";"3-3"^{}
\end{xy}
\end{equation*}
\end{ex}
\begin{defn}\cite{Kang}\label{def-YW2}
Let $X={\rm A}^{(2)}_{2n-2}$ or ${\rm D}^{(2)}_{n}$ and
$Y$ be a Young wall of ground state $\lambda$.
\begin{enumerate}
\item A column in $Y$ is called a {\it full column} if its height is a multiple of the unit length.
\item The wall $Y$ is said to be {\it proper} if none of two full columns in $Y$ have the same height.
\end{enumerate}
\end{defn}
\begin{defn}\cite{Kang}\label{def-YW2a}
Let $X={\rm A}^{(2)}_{2n-2}$ or ${\rm D}^{(2)}_{n}$ and
$Y$ be a proper Young wall of ground state $\lambda$.
\begin{enumerate}
\item
Let $i\in I$ and $Y'$ be a wall obtained from $Y$ by
removing an $i$-block. This $i$-block is said to be a {\it removable} $i$-{\it block} of $Y$
if $Y'$ is a proper Young wall.
\item Let $i\in I$. If we obtain a proper Young wall by adding
an $i$-block to a place of $Y$ then the place is said to be an
$i$-{\it admissible} {\it slot} of $Y$.
\end{enumerate}
\end{defn}
\begin{defn}\cite{Ka}
Let $X={\rm A}^{(2)}_{2n-2}$ or ${\rm D}^{(2)}_{n}$ and
$Y$ be a proper Young wall of ground state $\lambda$. We take $t\in\{1,n\}$.
\begin{enumerate}
\item[(i)]
Let $Y'$ be a wall
obtained by adding two $t$-blocks of shape (\ref{smpl2}) to the top of a column in $Y$:
\[
Y=
\begin{xy}
(14,10) *{\leftarrow A}="A",
(-3,3) *{\cdots}="dot1",
(12,3) *{\cdots}="dot2",
(0,-4) *{}="-1",
(8,-4)*{}="-2",
(4,-6)*{\vdots}="dot3",
(0,0) *{}="1",
(8,0)*{}="2",
(8,8)*{}="3",
(0,8)*{}="4",
(8,12)*{}="5",
(0,12)*{}="6",
(8,16)*{}="7",
(0,16)*{}="8",
\ar@{-} "1";"-1"^{}
\ar@{-} "-2";"2"^{}
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\ar@{--} "3";"5"^{}
\ar@{--} "4";"6"^{}
\ar@{--} "6";"5"^{}
\ar@{--} "7";"5"^{}
\ar@{--} "6";"8"^{}
\ar@{--} "7";"8"^{}
\end{xy}\qquad
Y'=
\begin{xy}
(14,14) *{}="B",
(4,10)*{t}="t1",
(4,14)*{t}="t2",
(-3,3) *{\cdots}="dot1",
(12,3) *{\cdots}="dot2",
(0,-4) *{}="-1",
(8,-4)*{}="-2",
(4,-6)*{\vdots}="dot3",
(0,0) *{}="1",
(8,0)*{}="2",
(8,8)*{}="3",
(0,8)*{}="4",
(8,12)*{}="5",
(0,12)*{}="6",
(8,16)*{}="7",
(0,16)*{}="8",
\ar@{-} "1";"-1"^{}
\ar@{-} "-2";"2"^{}
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\ar@{-} "3";"5"^{}
\ar@{-} "4";"6"^{}
\ar@{-} "6";"5"^{}
\ar@{-} "7";"5"^{}
\ar@{-} "6";"8"^{}
\ar@{-} "7";"8"^{}
\end{xy}
\]
In $Y$, a slot is named $A$ as above.
The slot $A$ in $Y$ is said to be {\it double} $t$-{\it admissible} if $Y'$ is also a proper Young wall.
\item[(ii)]
Let $Y''$ be a wall
obtained by removing two $t$-blocks of shape (\ref{smpl2}) from the top of a column in $Y$:
\[
Y=
\begin{xy}
(14,14) *{\leftarrow B}="B",
(4,10)*{t}="t1",
(4,14)*{t}="t2",
(-3,3) *{\cdots}="dot1",
(12,3) *{\cdots}="dot2",
(0,-4) *{}="-1",
(8,-4)*{}="-2",
(4,-6)*{\vdots}="dot3",
(0,0) *{}="1",
(8,0)*{}="2",
(8,8)*{}="3",
(0,8)*{}="4",
(8,12)*{}="5",
(0,12)*{}="6",
(8,16)*{}="7",
(0,16)*{}="8",
\ar@{-} "1";"-1"^{}
\ar@{-} "-2";"2"^{}
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\ar@{-} "3";"5"^{}
\ar@{-} "4";"6"^{}
\ar@{-} "6";"5"^{}
\ar@{-} "7";"5"^{}
\ar@{-} "6";"8"^{}
\ar@{-} "7";"8"^{}
\end{xy}\quad
Y''=
\begin{xy}
(14,10) *{}="A",
(-3,3) *{\cdots}="dot1",
(12,3) *{\cdots}="dot2",
(0,-4) *{}="-1",
(8,-4)*{}="-2",
(4,-6)*{\vdots}="dot3",
(0,0) *{}="1",
(8,0)*{}="2",
(8,8)*{}="3",
(0,8)*{}="4",
(8,12)*{}="5",
(0,12)*{}="6",
(8,16)*{}="7",
(0,16)*{}="8",
\ar@{-} "1";"-1"^{}
\ar@{-} "-2";"2"^{}
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\ar@{--} "3";"5"^{}
\ar@{--} "4";"6"^{}
\ar@{--} "6";"5"^{}
\ar@{--} "7";"5"^{}
\ar@{--} "6";"8"^{}
\ar@{--} "7";"8"^{}
\end{xy}
\]
A block in $Y$ is named $B$ as above.
The block $B$ in $Y$ is said to be {\it double} $t$-{\it removable}
if $Y''$ is also a proper Young wall.
\item[(iii)] Other $i$-admissible slot (resp. removable $i$-block)
is said to be single admissible (resp. single removable) for $i\in I$.
\end{enumerate}
\end{defn}
\begin{ex}\label{ex-A2-1}
We consider the following Young wall of type ${\rm A}^{(2)}_{4}$ in Example \ref{ex-YW}.
\begin{equation*}
\begin{xy}
(-15.5,-2) *{\ 1}="000-1",
(-9.5,-2) *{\ 1}="00-1",
(-3.5,-2) *{\ 1}="0-1",
(1.5,-2) *{\ \ 1}="0-1",
(-21.5,-2) *{\dots}="00000",
(-9.5,1.5) *{\ 1}="000",
(-3.5,1.5) *{\ 1}="00",
(1.5,1.5) *{\ \ 1}="0",
(1.5,6.5) *{\ \ 2}="02",
(-4,6.5) *{\ \ 2}="002",
(1.5,12.5) *{\ \ 3}="03",
(1.5,18.5) *{\ \ 2}="04",
(22.5,18.5) *{\ \ \leftarrow\text{removable }2\text{-block}}="arrow1",
(27,23) *{\ \ \leftarrow\text{double }1\text{-admissible slot}}="arrow2",
(-21.5,12.5) *{3\text{-admissible slot}\rightarrow}="arrow3",
(-28.5,2.5) *{\text{removable }1\text{-block}\rightarrow}="arrow4",
(0,0) *{}="1",
(0,-3.5) *{}="1-u",
(6,0)*{}="2",
(6,-3.5)*{}="2-u",
(6,3.5)*{}="3",
(0,3.5)*{}="4",
(-6,3.5)*{}="5",
(-6,0)*{}="6",
(-6,-3.5)*{}="6-u",
(-12,3.5)*{}="7",
(-12,0)*{}="8",
(-12,-3.6)*{}="8-u",
(-18,3.5)*{}="9",
(-18,0)*{}="10-a",
(-18,0)*{}="10",
(-18,-3.5)*{}="10-u",
(6,9.5)*{}="3-1",
(6,15.5)*{}="3-2",
(6,21.5)*{}="3-3",
(0,9.5)*{}="4-1",
(0,15.5)*{}="4-2",
(0,21.5)*{}="4-3",
(-6,9.5)*{}="5-1",
\ar@{-} "8";"10-a"^{}
\ar@{-} "10-u";"10-a"^{}
\ar@{-} "10-u";"2-u"^{}
\ar@{-} "8";"8-u"^{}
\ar@{-} "6";"6-u"^{}
\ar@{-} "2";"2-u"^{}
\ar@{-} "1";"1-u"^{}
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\ar@{-} "5";"6"^{}
\ar@{-} "5";"4"^{}
\ar@{-} "1";"6"^{}
\ar@{-} "7";"8"^{}
\ar@{-} "7";"5"^{}
\ar@{-} "6";"8"^{}
\ar@{-} "3";"3-1"^{}
\ar@{-} "4-1";"3-1"^{}
\ar@{-} "4-1";"4"^{}
\ar@{-} "5-1";"5"^{}
\ar@{-} "4-1";"5-1"^{}
\ar@{-} "4-1";"4-2"^{}
\ar@{-} "3-2";"3-1"^{}
\ar@{-} "3-2";"4-2"^{}
\ar@{-} "3-2";"3-3"^{}
\ar@{-} "4-3";"4-2"^{}
\ar@{-} "4-3";"3-3"^{}
\end{xy}
\end{equation*}
Since it holds $h_1=4$, $h_2=2$, $h_3=1$ and $h_{\ell}=\frac{1}{2}$ ($\ell\geq4$),
the first, second and third columns are full columns, but the fourth column is not since its height is half-unit.
We see that the Young wall is proper.
It has only one double $1$-admissible slot and other admissible slots and removable blocks are single.
\end{ex}
\section{Combinatorial descriptions of polyhedral realizations for $B(\lambda)$}
In this section, we present our main results.
\subsection{Adaptedness and several notation}\label{seno}
\begin{defn}\label{adapt}\cite{KaN}
Let $A=(a_{i,j})_{i,j\in I}$ be the symmetrizable generalized Cartan matrix of $\mathfrak{g}$.
We assume that a sequence $\io=(\cdots,i_3,i_2,i_1)$
satisfies $(\ref{seq-con})$.
We say $\iota$ is {\it adapted} to $A$ if the following condition holds :
For $i,j\in I$ such that $a_{i,j}<0$, the subsequence of $\iota$ consisting of all $i$, $j$ is
\[
(\cdots,i,j,i,j,i,j,i,j)\quad {\rm or}\quad (\cdots,j,i,j,i,j,i,j,i).
\]
In case of the Cartan matrix is fixed, the sequence $\iota$ is shortly said to be {\it adapted}.
\end{defn}
\begin{ex}
In case that
$\mathfrak{g}$ is of type ${\rm A}^{(1)}_2$, $\iota=(\cdots,2,1,3,2,1,3,2,1,3)$,
it holds $a_{1,2}=a_{2,3}=a_{1,3}=-1$.
\begin{itemize}
\item
The subsequence consisting of $1$ and $2$ is $(\cdots,2,1,2,1,2,1)$.
\item
The subsequence consisting of $2$ and $3$ is $(\cdots,2,3,2,3,2,3)$.
\item
The subsequence consisting of $1$ and $3$ is $(\cdots,1,3,1,3,1,3)$.
\end{itemize}
Thus, $\iota$ is an adapted sequence.
\end{ex}
In the rest of article, let us fix a sequence $\iota=(\cdots,i_3,i_2,i_1)$ adapted to $A$.
One defines a set of integers $(p_{i,j})_{i,j\in I;a_{i,j}<0}$ by
\begin{equation}\label{pij}
p_{i,j}=\begin{cases}
1 & {\rm if}\ {\rm the\ subsequence\ of\ }\iota{\rm\ consisting\ of}\ i,j\ {\rm is}\ (\cdots,j,i,j,i,j,i),\\
0 & {\rm if}\ {\rm the\ subsequence\ of\ }\iota{\rm\ consisting\ of}\ i,j\ {\rm is}\ (\cdots,i,j,i,j,i,j).
\end{cases}
\end{equation}
Note that if $a_{i,j}<0$ then we have
\begin{equation}\label{pij2}
p_{i,j}+p_{j,i}=1.
\end{equation}
We identify each single index $r\in\mathbb{Z}_{\geq1}$ with a pair $(s,k)\in \mathbb{Z}_{\geq1}\times I$
if $i_r=k$ holds and $k$ appears $s$ times in $i_r$, $i_{r-1}$, $\cdots,i_1$.
For example, when $\iota=(\cdots,2,1,3,2,1,3,2,1,3)$, single indices
$\cdots,6,5,4,3,2,1$ are identified with
\[
\cdots,(2,2),(2,1),(2,3),(1,2),(1,1),(1,3).
\]
The notation $x_{r}$, $\beta_{r}$, $\beta^{\pm}_{r}$ and $\what{S}'_{r}$ in the subsections \ref{poly-uqm}, \ref{poly-uql} are also rewritten as
\[
x_r=x_{s,k},\quad \beta_r=\beta_{s,k},\quad \beta^{\pm}_r=\beta^{\pm}_{s,k},\quad \what{S}'_r=\what{S}'_{s,k}.
\]
If $(s,k)\notin \mathbb{Z}_{\geq1}\times I$ then we understand $x_{s,k}=0$.
By this identification and the ordinary order on $\mathbb{Z}_{\geq1}$, that is, $1<2<3<4<5<6<\cdots$,
an order on $\mathbb{Z}_{\geq1}\times I$ is defined. In case of $\iota=(\cdots,2,1,3,2,1,3,2,1,3)$, the order is
\[
\cdots>(2,2)>(2,1)>(2,3)>(1,2)>(1,1)>(1,3).
\]
Using the notation in (\ref{pij}), by the adaptedness of $\iota$, one can write $\beta_{s,k}$ as
\begin{equation}\label{pij3}
\beta_{s,k}=x_{s,k}+x_{s+1,k}+\sum_{j\in I; a_{k,j}<0} a_{k,j}x_{s+p_{j,k},j}.
\end{equation}
The above
$\beta_{s,k}$ can be regarded as an analog of
simple root $\alpha_{k}$ of $^L\mathfrak{g}$
since $\alpha_{k}$ is expressed by
\[
\alpha_{k}=\Lambda_{k}+\Lambda_{k}+\sum_{j\in I; a_{k,j}<0} a_{k,j}\Lambda_{j}
\]
on $\bigoplus_{i\in I}\mathbb{Z}h_i$.
This is a reason why the
inequalities defining ${\rm Im}(\Psi^{(\lambda)}_{\iota})$
is expressed by combinatorial objects related to $^L\mathfrak{g}$ rather than $\mathfrak{g}$
in Theorem \ref{thmA1}, \ref{thmA2}.
\subsection{Non-negative integers $P^X_k$}
Let $X={\rm A}^{(1)},\ {\rm C}^{(1)},\ {\rm A}^{(2)}$ or ${\rm D}^{(2)}$ and
\[
\begin{cases}
k\in I & \text{if }X={\rm A}^{(1)},\ \text{or }{\rm C}^{(1)},\\
k\in I\setminus\{1\} & \text{if }X={\rm A}^{(2)},\\
k\in I\setminus\{1,n\} & \text{if }X={\rm D}^{(2)}.
\end{cases}
\]
For $t\in\mathbb{Z}$, let $P^X_k(t)\in\mathbb{Z}_{\geq0}$ denote the non-negative integer defined as follows:
Setting $P^X_k(k):=0$, one recursively defines as
\begin{equation}\label{pk1}
P^X_k(t):=P^X_k(t-1)+p_{\pi_X(t),\pi_X(t-1)}\ (\text{for } t>k),
\end{equation}
\begin{equation}\label{pk2}
P^X_k(t):=P^X_k(t+1)+p_{\pi_X(t),\pi_X(t+1)}\ (\text{for } t<k),
\end{equation}
where the notation $\pi_X$ is defined in Definition \ref{pi1-def} and we understand $p_{j,j}:=0$ for all $j\in I$.
Next,
considering the map
$\{1,2,\cdots,2n-2\}\rightarrow \{1,2,\cdots,n\}$
defined as
\[
\ell\mapsto \ell,\ 2n-\ell\mapsto \ell \qquad (2\leq \ell\leq n-1),
\]
\[
1\mapsto1,\ n\mapsto n,
\]
we extend it to a map
\begin{equation}\label{piprime}
\pi':\mathbb{Z}_{\geq1}\rightarrow\{1,2,\cdots,n\}
\end{equation}
by periodicity $2n-2$.
For
$X={\rm A}^{(2)}$ or ${\rm D}^{(2)}$
and
\[
k=
\begin{cases}
1 & \text{if }X={\rm A}^{(2)},\\
1 \text{ or }n & \text{if }X={\rm D}^{(2)},
\end{cases}
\]
we recursively define integers $P^X_k(\ell)$ ($\ell\in\mathbb{Z}_{\geq k}$) as
\[
P^X_k(k):=0,\ \ P^X_k(\ell)=P^X_k(\ell-1)+p_{\pi'(\ell),\pi'(\ell-1)}\ (\ell>k).
\]
\subsection{Type ${\rm A}_{n-1}^{(1)}$ and ${\rm D}^{(2)}_n$-case}
For any pair $(i,j)\in\mathbb{Z}\times \mathbb{Z}$, $s\in\mathbb{Z}$, $k\in I$ and $X={\rm A}^{(1)}$ or ${\rm C}^{(1)}$, we set
\begin{equation}\label{ovlL1}
L^X_{s,k,\iota}(i,j):=x_{s+P^X_k(i+j)+{\rm min}\{k-j,i\},\pi_X(i+j)}\in (\mathbb{Q}^{\infty})^*.
\end{equation}
For an extended Young diagram $T$ with $y_{\infty}=k$, we define
\begin{equation}\label{ovlL2}
L^X_{s,k,\iota}(T)
:=\sum_{P:{\rm concave\ corner\ of\ }T} L^X_{s,k,\iota}(P) - \sum_{P:{\rm convex\ corner\ of\ }T} L^X_{s,k,\iota}(P)
\in (\mathbb{Q}^{\infty})^*.
\end{equation}
For $k\in I$,
let ${\rm EYD}_{k}$ denote the set of extended Young diagrams with $y_{\infty}=k$.
One defines
\begin{equation}\label{r-tilde-A0}
\fbox{$r$}^{X}_k:=x_{P^X_k(r),\pi_X(r)}
-x_{1+P^X_k(r-1),\pi_X(r-1)} \qquad (r\in\mathbb{Z}_{\geq k+1}),
\end{equation}
\begin{equation}\label{r-tilde-A1}
\fbox{$\tilde{r}$}_k^{X}:=x_{P^X_k(r-1),\pi_X(r-1)}
-x_{1+P^X_k(r),\pi_X(r)} \qquad (r\in\mathbb{Z}_{\leq k}).
\end{equation}
We also set
\begin{equation}\label{comb-lam-1}
{\rm Comb}_{k,\iota}^{X}[\lambda]:=
\begin{cases}
\{-x_{1,k}+\langle h_k,\lambda \rangle\} & \text{if }(1,k)<(1,\pi_X(k+1)),\ (1,k)<(1,\pi_X(k-1)),\\
\{\fbox{$\tilde{r}$}_k^{X}+\langle h_k,\lambda \rangle\ | r\in\mathbb{Z}_{\leq k}\} & \text{if }(1,k)<(1,\pi_X(k+1)),\ (1,k)>(1,\pi_X(k-1)),\\
\{\fbox{$r$}^{X}_k+\langle h_k,\lambda \rangle\ | r\in\mathbb{Z}_{\geq k+1}\} & \text{if }(1,k)>(1,\pi_X(k+1)),\ (1,k)<(1,\pi_X(k-1)),\\
\{
L^X_{0,k,\iota}(T)+\langle h_k,\lambda \rangle | T\in{\rm EYD}_{k}\setminus\{\phi^k\}
\} & \text{if }(1,k)>(1,\pi_X(k+1)),\ (1,k)>(1,\pi_X(k-1))\\
\end{cases}
\end{equation}
and
\begin{equation}\label{comb-inf-1}
{\rm Comb}_{\iota}^{X}[\infty]:=
\{ L^X_{s,k,\iota}(T) | s\in\mathbb{Z}_{\geq1},\ k\in I,\ T\in {\rm EYD}_k \}.
\end{equation}
Here, $\phi^k\in{\rm EYD}_k$ is the extended Young diagram without boxes, that is, $\phi^k=(y_{\ell})_{\ell\in\mathbb{Z}_{\geq0}}$,
$y_{\ell}=k$ for any $\ell\in\mathbb{Z}_{\geq0}$.
\begin{thm}\label{thmA1}
We suppose that $\mathfrak{g}$ is of type $X={\rm A}_{n-1}^{(1)}$ $(n\geq2)$ or $X={\rm D}_{n}^{(2)}$ $(n\geq3)$ and $\iota$ is adapted.
Then for any $\lambda\in P^+$, the pair $(\iota,\lambda)$ is $\Xi'$-ample and
\[
{\rm Im}(\Psi_{\iota}^{(\lambda)})
=\left\{ \mathbf{a}\in\mathbb{Z}^{\infty}
\left|
\varphi(\mathbf{a})\geq0\\
\text{ for any }
\varphi\in
{\rm Comb}^{X^L}_{\iota}[\infty]\cup
\bigcup_{k\in I}
{\rm Comb}^{X^L}_{k,\iota}[\lambda]
\right.
\right\}.
\]
Here, $({\rm A}^{(1)}_{n-1})^L={\rm A}^{(1)}_{n-1}$ and $({\rm D}^{(2)}_{n})^L={\rm C}^{(1)}_{n-1}$.
\end{thm}
\begin{rem}\label{rem-1}
Let us put
\[
[r]^X_k:=
\frac{X_{P^X_k(r),\pi_X(r)}}
{X_{1+P^X_k(r-1),\pi_X(r-1)}},\qquad
[\tilde{r}]^X_k:=
\frac{X_{P^X_k(r-1),\pi_X(r-1)}}
{X_{1+P^X_k(r),\pi_X(r)}},
\]
where tropicalizing these Laurent monomials, we get linear functions
in (\ref{r-tilde-A0}), (\ref{r-tilde-A1}).
Then
the set
$\{[r]^X_k | r\in\mathbb{Z}\}$ (resp. $\{[\tilde{r}]^X_k | r\in\mathbb{Z}\}$) coincides with
the connected component of $\mathcal{M}_c$ in \cite{K}
with $c_{m,j}=p_{m,j}$ for $m,j\in I$ such that $a_{m,j}<0$
containing
$\frac{X_{P^X_k(k+1),\pi_X(k+1)}}
{X_{1,k}}$
(resp. $\frac{X_{P^X_k(k-1),\pi_X(k-1)}}
{X_{1,k}}$),
which is isomorphic to the crystal base $B(\Lambda_{\pi_X(k+1)}-\Lambda_{k})$ (resp. $B(\Lambda_{\pi_X(k-1)}-\Lambda_{k})$)
of the extremal weight module $V(\Lambda_{\pi_X(k+1)}-\Lambda_{k})$ (resp. $V(\Lambda_{\pi_X(k-1)}-\Lambda_{k})$) of $U_q(\mathfrak{g})$.
In this way, functions in ${\rm Comb}^{X^L}_{\iota}[\infty]$ and
${\rm Comb}^{X^L}_{k,\iota}[\lambda]$ are deeply related to representation
theory of $U_q(^L\mathfrak{g})$.
In Theorem 3.2 of \cite{HN}, monomial realizations of level 0 extremal
fundamental weight crystals are given.
\end{rem}
\begin{ex}\label{ex-2}
Let us consider the case
$\mathfrak{g}$ is of type ${\rm A}_2^{(1)}$ and $\iota=(\cdots,3,1,2,3,1,2)$, which is the same setting
as in Example 4.4 of \cite{Ka}.
In the paper \cite{Ka},
we computed several inequalities defining
${\rm Im}(\Psi_{\iota})=\{\mathbf{a}\in \mathbb{Z}^{\infty} |
\varphi(\mathbf{a})\geq0\ \text{ for any }\varphi\in {\rm Comb}^{{\rm A}^{(1)}}_{\iota}[\infty]\}$:
\[
{\rm Im}(\Psi_{\iota})
=\left\{
\textbf{a}\in\mathbb{Z}^{\infty} \left|
\begin{array}{l}
s\in\mathbb{Z}_{\geq1},\ a_{s,1}\geq0,
a_{s+1,2}+a_{s,3}-a_{s+1,1}\geq0,
a_{s+1,3}+a_{s,3}-a_{s+2,2}\geq0,\\
2a_{s+1,2}-a_{s+1,3}\geq0,
a_{s+1,2}+a_{s+1,1}-a_{s+2,2}\geq0,
a_{s+1,2}+a_{s+1,3}-a_{s+2,1}\geq0,\cdots\\
a_{s,2}\geq0,\
a_{s,1}+a_{s,3}-a_{s+1,2}\geq0,
a_{s,1}+a_{s+1,1}-a_{s+1,3}\geq0,\\
2a_{s,3}-a_{s+1,1}\geq0,
a_{s,3}+a_{s+1,2}-a_{s+1,3}\geq0,
a_{s,3}+a_{s+1,1}-a_{s+2,2}\geq0,\cdots\\
a_{s,3}\geq0,
a_{s+1,1}+a_{s+1,2}-a_{s+1,3}\geq0,
a_{s+2,2}+a_{s+1,2}-a_{s+2,1}\geq0,\\
2a_{s+1,1}-a_{s+2,2}\geq0,
a_{s+1,1}+a_{s+1,3}-a_{s+2,1}\geq0,
a_{s+1,1}+a_{s+2,2}-a_{s+2,3}\geq0,\cdots
\end{array}
\right.
\right\}.
\]
Let us compute several inequalities coming from ${\rm Comb}_{k,\iota}^{{\rm A}^{(1)}}[\lambda]$ for $k=1,2,3$.
One obtains
\[
\cdots,\ P_1^{{\rm A}^{(1)}}(-1)=1,\ P_1^{{\rm A}^{(1)}}(0)=0,\ P_1^{{\rm A}^{(1)}}(1)=0,\ P_1^{{\rm A}^{(1)}}(2)=1,\ P_1^{{\rm A}^{(1)}}(3)=1,\cdots,
\]
\[
\cdots,\ P_2^{{\rm A}^{(1)}}(-1)=1,\ P_2^{{\rm A}^{(1)}}(0)=0,\ P_2^{{\rm A}^{(1)}}(1)=0,\ P_2^{{\rm A}^{(1)}}(2)=0,\ P_2^{{\rm A}^{(1)}}(3)=0,\ P_2^{{\rm A}^{(1)}}(4)=1,\cdots,
\]
\[
\cdots,\ P_3^{{\rm A}^{(1)}}(1)=1,\ P_3^{{\rm A}^{(1)}}(2)=1,\ P_3^{{\rm A}^{(1)}}(3)=0,\ P_3^{{\rm A}^{(1)}}(4)=1,\ P_3^{{\rm A}^{(1)}}(5)=2,\cdots .
\]
By the definition (\ref{comb-lam-1}), we have ${\rm Comb}_{2,\iota}^{{\rm A}^{(1)}}[\lambda]=\{-x_{1,2}+\langle h_2,\lambda \rangle\}$
and
\[
{\rm Comb}_{1,\iota}^{{\rm A}^{(1)}}[\lambda]=
\{\fbox{$r$}^{{\rm A}^{(1)}}_1+\langle h_1,\lambda \rangle\ | r\in\mathbb{Z}_{\geq 2}\}.
\]
One can compute
\[
\fbox{$2$}^{{\rm A}^{(1)}}_1=x_{1,2}-x_{1,1}, \
\fbox{$3$}^{{\rm A}^{(1)}}_1=x_{1,3}-x_{2,2},\
\fbox{$4$}^{{\rm A}^{(1)}}_1=x_{2,1}-x_{2,3},\
\]
\[
\fbox{$5$}^{{\rm A}^{(1)}}_1=x_{3,2}-x_{3,1}, \
\fbox{$6$}^{{\rm A}^{(1)}}_1=x_{3,3}-x_{4,2}.
\]
Next, let us turn to ${\rm Comb}_{3,\iota}^{{\rm A}^{(1)}}[\lambda]$.
The following diagrams are elements in ${\rm EYD}_{3}\setminus\{\phi^3\}$:
\[
\begin{xy}
(40,-9) *{T_3^1:=}="Y1",
(55,0) *{}="1a",
(85,0)*{}="2a",
(55,-30)*{}="3a",
(50,0)*{(0,3)}="4a",
(61,2) *{1}="10a",
(61,-1) *{}="1010a",
(61,-6) *{}="1-0a",
(67,2) *{2}="11a",
(67,-1) *{}="1111a",
(73,2) *{3}="12a",
(73,-1) *{}="1212a",
(79,2) *{4}="13a",
(79,-1) *{}="1313a",
(50,-6)*{\ 2\ }="5a",
(56,-6)*{}="55a",
(50,-12)*{\ 1\ }="6a",
(56,-12)*{}="66a",
(50,-18)*{\ 0\ }="7a",
(56,-18)*{}="77a",
(50,-24)*{\ -1}="8a",
(56,-24)*{}="88a",
(95,-9) *{T_3^2:=}="Y2",
(110,0) *{}="1b",
(140,0)*{}="2b",
(110,-30)*{}="3b",
(105,0)*{(0,3)}="4b",
(116,2) *{1}="10b",
(116,-1) *{}="1010b",
(122,-6) *{}="1-0b",
(122,2) *{2}="11b",
(122,-1) *{}="1111b",
(128,2) *{3}="12b",
(128,-1) *{}="1212b",
(134,2) *{4}="13b",
(134,-1) *{}="1313b",
(105,-6)*{\ 2\ }="5b",
(111,-6)*{}="55b",
(105,-12)*{\ 1\ }="6b",
(111,-12)*{}="66b",
(105,-18)*{\ 0\ }="7b",
(111,-18)*{}="77b",
(105,-24)*{\ -1}="8b",
(111,-24)*{}="88b",
\ar@{-} "1a";"2a"^{}
\ar@{-} "1a";"3a"^{}
\ar@{-} "5a";"55a"^{}
\ar@{-} "6a";"66a"^{}
\ar@{-} "7a";"77a"^{}
\ar@{-} "8a";"88a"^{}
\ar@{-} "10a";"1010a"^{}
\ar@{-} "11a";"1111a"^{}
\ar@{-} "12a";"1212a"^{}
\ar@{-} "13a";"1313a"^{}
\ar@{-} "55a";"1-0a"^{}
\ar@{-} "1010a";"1-0a"^{}
\ar@{-} "1b";"2b"^{}
\ar@{-} "1b";"3b"^{}
\ar@{-} "5b";"55b"^{}
\ar@{-} "6b";"66b"^{}
\ar@{-} "7b";"77b"^{}
\ar@{-} "8b";"88b"^{}
\ar@{-} "10b";"1010b"^{}
\ar@{-} "11b";"1111b"^{}
\ar@{-} "12b";"1212b"^{}
\ar@{-} "13b";"1313b"^{}
\ar@{-} "55b";"1-0b"^{}
\ar@{-} "11b";"1-0b"^{}
\end{xy}
\]
\[
\begin{xy}
(-15,-9) *{T_3^3:=}="Y3",
(0,0) *{}="1",
(30,0)*{}="2",
(0,-30)*{}="3",
(-5,0)*{(0,3)}="4",
(6,2) *{1}="10",
(6,-1) *{}="1010",
(12,2) *{2}="11",
(12,-1) *{}="1111",
(18,2) *{3}="12",
(18,-1) *{}="1212",
(24,2) *{4}="13",
(24,-1) *{}="1313",
(-5,-6)*{\ 2\ }="5",
(1,-6)*{}="55",
(6,-12)*{}="6-12",
(-5,-12)*{\ 1\ }="6",
(1,-12)*{}="66",
(-5,-18)*{\ 0\ }="7",
(1,-18)*{}="77",
(-5,-24)*{\ -1}="8",
(1,-24)*{}="88",
(40,-9) *{T_3^4:=}="Y4",
(55,0) *{}="1a",
(85,0)*{}="2a",
(55,-30)*{}="3a",
(50,0)*{(0,3)}="4a",
(61,2) *{1}="10a",
(61,-1) *{}="1010a",
(61,-6) *{}="6-6a",
(61,-12)*{}="6-12a",
(67,-6)*{}="12-6a",
(67,2) *{2}="11a",
(67,-1) *{}="1111a",
(73,2) *{3}="12a",
(73,-1) *{}="1212a",
(79,2) *{4}="13a",
(79,-1) *{}="1313a",
(50,-6)*{\ 2\ }="5a",
(56,-6)*{}="55a",
(50,-12)*{\ 1\ }="6a",
(56,-12)*{}="66a",
(50,-18)*{\ 0\ }="7a",
(56,-18)*{}="77a",
(50,-24)*{\ -1}="8a",
(56,-24)*{}="88a",
(95,-9) *{T_3^5:=}="Y5",
(110,0) *{}="1b",
(140,0)*{}="2b",
(110,-30)*{}="3b",
(105,0)*{(0,3)}="4b",
(116,2) *{1}="10b",
(116,-1) *{}="1010b",
(122,-12) *{}="1-0b",
(122,2) *{2}="11b",
(122,-1) *{}="1111b",
(128,2) *{3}="12b",
(128,-1) *{}="1212b",
(134,2) *{4}="13b",
(134,-1) *{}="1313b",
(105,-6)*{\ 2\ }="5b",
(111,-6)*{}="55b",
(105,-12)*{\ 1\ }="6b",
(111,-12)*{}="66b",
(105,-18)*{\ 0\ }="7b",
(111,-18)*{}="77b",
(105,-24)*{\ -1}="8b",
(111,-24)*{}="88b",
\ar@{-} "1";"2"^{}
\ar@{-} "1";"3"^{}
\ar@{-} "5";"55"^{}
\ar@{-} "6";"66"^{}
\ar@{-} "6";"6-12"^{}
\ar@{-} "1010";"6-12"^{}
\ar@{-} "7";"77"^{}
\ar@{-} "8";"88"^{}
\ar@{-} "10";"1010"^{}
\ar@{-} "11";"1111"^{}
\ar@{-} "12";"1212"^{}
\ar@{-} "13";"1313"^{}
\ar@{-} "1a";"2a"^{}
\ar@{-} "1a";"3a"^{}
\ar@{-} "5a";"55a"^{}
\ar@{-} "6a";"66a"^{}
\ar@{-} "7a";"77a"^{}
\ar@{-} "8a";"88a"^{}
\ar@{-} "10a";"1010a"^{}
\ar@{-} "11a";"1111a"^{}
\ar@{-} "12a";"1212a"^{}
\ar@{-} "13a";"1313a"^{}
\ar@{-} "6-12a";"66a"^{}
\ar@{-} "6-12a";"6-6a"^{}
\ar@{-} "12-6a";"6-6a"^{}
\ar@{-} "12-6a";"1111a"^{}
\ar@{-} "1b";"2b"^{}
\ar@{-} "1b";"3b"^{}
\ar@{-} "5b";"55b"^{}
\ar@{-} "6b";"66b"^{}
\ar@{-} "7b";"77b"^{}
\ar@{-} "8b";"88b"^{}
\ar@{-} "10b";"1010b"^{}
\ar@{-} "11b";"1111b"^{}
\ar@{-} "12b";"1212b"^{}
\ar@{-} "13b";"1313b"^{}
\ar@{-} "6b";"1-0b"^{}
\ar@{-} "11b";"1-0b"^{}
\end{xy}
\]
The diagram $T_3^1$ has concave corners $(1,3)$ and $(0,2)$ and a convex corner $(1,2)$. Thus,
\[
L^{{\rm A}^{(1)}}_{0,3,\iota}(T^1_3)=x_{P_3^{{\rm A}^{(1)}}(4),\pi_{{\rm A}^{(1)}}(4)}+x_{P_3^{{\rm A}^{(1)}}(2),\pi_{{\rm A}^{(1)}}(2)}
-x_{P_3^{{\rm A}^{(1)}}(2),\pi_{{\rm A}^{(1)}}(3)}=x_{1,1}+x_{1,2}-x_{1,3}.
\]
Similarly, it follows
\[
L^{{\rm A}^{(1)}}_{0,3,\iota}(T^2_3)=x_{P_3^{{\rm A}^{(1)}}(5),\pi_{{\rm A}^{(1)}}(5)}+x_{P_3^{{\rm A}^{(1)}}(2),\pi_{{\rm A}^{(1)}}(2)}-x_{P_3^{{\rm A}^{(1)}}(4)+1,\pi_{{\rm A}^{(1)}}(4)}
=x_{2,2}+x_{1,2}-x_{2,1},
\]
\[
L^{{\rm A}^{(1)}}_{0,3,\iota}(T^3_3)=x_{P_3^{{\rm A}^{(1)}}(4),\pi_{{\rm A}^{(1)}}(4)}+x_{P_3^{{\rm A}^{(1)}}(1),\pi_{{\rm A}^{(1)}}(1)}-x_{P_3^{{\rm A}^{(1)}}(2)+1,\pi_{{\rm A}^{(1)}}(2)}
=2x_{1,1}-x_{2,2},
\]
\begin{eqnarray*}
L^{{\rm A}^{(1)}}_{0,3,\iota}(T^4_3)&=&x_{P_3^{{\rm A}^{(1)}}(1),\pi_{{\rm A}^{(1)}}(1)}+
x_{P_3^{{\rm A}^{(1)}}(3)+1,\pi_{{\rm A}^{(1)}}(3)}
+x_{P_3^{{\rm A}^{(1)}}(5),\pi_{{\rm A}^{(1)}}(5)}\\
& &-x_{P_3^{{\rm A}^{(1)}}(2)+1,\pi_{{\rm A}^{(1)}}(2)}-x_{P_3^{{\rm A}^{(1)}}(4)+1,\pi_{{\rm A}^{(1)}}(4)}
=
x_{1,1}+x_{1,3}-x_{2,1},
\end{eqnarray*}
\[
L^{{\rm A}^{(1)}}_{0,3,\iota}(T^5_3)=x_{P_3^{{\rm A}^{(1)}}(1),\pi_{{\rm A}^{(1)}}(1)}+x_{P_3^{{\rm A}^{(1)}}(5),\pi_{{\rm A}^{(1)}}(5)}
-x_{P_3^{{\rm A}^{(1)}}(3)+2,\pi_{{\rm A}^{(1)}}(3)}
=x_{1,1}+x_{2,2}-x_{2,3}.
\]
In this way, we obtain several inequalities defining ${\rm Im}(\Psi_{\iota}^{(\lambda)})$:
\[
{\rm Im}(\Psi_{\iota}^{(\lambda)})
=\left\{
\textbf{a}\in\mathbb{Z}^{\infty} \left|
\begin{array}{l}
s\in\mathbb{Z}_{\geq1},\ a_{s,1}\geq0,
a_{s+1,2}+a_{s,3}-a_{s+1,1}\geq0,
a_{s+1,3}+a_{s,3}-a_{s+2,2}\geq0,\\
2a_{s+1,2}-a_{s+1,3}\geq0,
a_{s+1,2}+a_{s+1,1}-a_{s+2,2}\geq0,
a_{s+1,2}+a_{s+1,3}-a_{s+2,1}\geq0,\cdots\\
a_{s,2}\geq0,\
a_{s,1}+a_{s,3}-a_{s+1,2}\geq0,
a_{s,1}+a_{s+1,1}-a_{s+1,3}\geq0,\\
2a_{s,3}-a_{s+1,1}\geq0,
a_{s,3}+a_{s+1,2}-a_{s+1,3}\geq0,
a_{s,3}+a_{s+1,1}-a_{s+2,2}\geq0,\cdots\\
a_{s,3}\geq0,
a_{s+1,1}+a_{s+1,2}-a_{s+1,3}\geq0,
a_{s+2,2}+a_{s+1,2}-a_{s+2,1}\geq0,\\
2a_{s+1,1}-a_{s+2,2}\geq0,
a_{s+1,1}+a_{s+1,3}-a_{s+2,1}\geq0,
a_{s+1,1}+a_{s+2,2}-a_{s+2,3}\geq0,\cdots\\
a_{1,2}\leq \langle h_2,\lambda\rangle,
a_{1,2}-a_{1,1}+\langle h_1,\lambda\rangle\geq0,
a_{1,3}-a_{2,2}+\langle h_1,\lambda\rangle\geq0,\\
a_{2,1}-a_{2,3}+\langle h_1,\lambda\rangle\geq0,
a_{3,2}-a_{3,1}+\langle h_1,\lambda\rangle\geq0,
a_{3,3}-a_{4,2}+\langle h_1,\lambda\rangle\geq0,\cdots\\
a_{1,1}+a_{1,2}-a_{1,3}+\langle h_3,\lambda\rangle\geq0,
a_{2,2}+a_{1,2}-a_{2,1}+\langle h_3,\lambda\rangle\geq0,\\
2a_{1,1}-a_{2,2}+\langle h_3,\lambda\rangle\geq0,
a_{1,1}+a_{1,3}-a_{2,1}+\langle h_3,\lambda\rangle\geq0,\\
a_{1,1}+a_{2,2}-a_{2,3}+\langle h_3,\lambda\rangle\geq0, \cdots
\end{array}
\right.
\right\}.
\]
\end{ex}
\subsection{Type ${\rm A}_{2n-2}^{(2)}$ and ${\rm C}_{n-1}^{(1)}$-case}
In this subsection, let
$X={\rm A}^{(2)}$ or $X={\rm D}^{(2)}$.
We set
\[
I_X:=
\begin{cases}
I\setminus \{1\} & \text{if }X={\rm A}_{2n-2}^{(2)},\\
I\setminus\{1,n\} & \text{if }X={\rm D}_{n}^{(2)}.
\end{cases}
\]
First, we take $k$ as $k\in I_X$.
For $(i,j)\in \mathbb{Z}\times\mathbb{Z}$, $s\in\mathbb{Z}$,
one
defines
\[
L^X_{s,k,{\rm ad}}(i,j)=x_{s+P^X_k(i+k)+[i]_-+k-j,\pi_X(i+k)},\quad
L^X_{s,k,{\rm re}}(i,j)=x_{s+P^X_k(i+k-1)+[i-1]_-+k-j,\pi_X(i+k-1)},
\]
where $[i]_-={\rm min}\{i,0\}$.
For each $T\in{\rm REYD}^{X}_{k}$, we set
\begin{eqnarray}
L^X_{s,k,\iota}(T)&:=&
\sum_{t\in I} \left(
\sum_{P:\text{single }\text{-admissible point of }T} L^X_{s,k,{\rm ad}}(P)
-\sum_{P:\text{single }\text{-removable point of }T} L^X_{s,k,{\rm re}}(P)\right)\label{L1kdef}\\
& & +
\sum_{P:\text{double }\text{-admissible point of }T} 2L^X_{s,k,{\rm ad}}(P)
-\sum_{P:\text{double }\text{-removable point of }T} 2L^X_{s,k,{\rm re}}(P)\in (\mathbb{Q}^{\infty})^*.\nonumber
\end{eqnarray}
One defines for $r\in\mathbb{Z}_{\geq k+1}$,
\begin{equation}\label{r-tilde-D0}
\fbox{$r$}^{X}_k:=
\begin{cases}
2x_{P^X_k(r),\pi_X(r)}
-x_{1+P^X_k(r-1),\pi_X(r-1)}, & \text{if }\pi_X(r)\in I\setminus I_X\text{ and }\pi_X(r)\neq \pi_X(r-1),\\
x_{P^X_k(r),\pi_X(r)}
-2x_{1+P^X_k(r-1),\pi_X(r-1)}, & \text{if }\pi_X(r-1)\in I\setminus I_X\text{ and }\pi_X(r)\neq \pi_X(r-1),\\
x_{P^X_k(r),\pi_X(r)}
-x_{1+P^X_k(r-1),\pi_X(r-1)}, & \text{otherwise},
\end{cases}
\end{equation}
for $r\in\mathbb{Z}_{\leq k}$,
\begin{equation}\label{r-tilde-D1}
\fbox{$\tilde{r}$}^{X}_k:=
\begin{cases}
2x_{P^X_k(r-1),\pi_X(r-1)}
-x_{1+P^X_k(r),\pi_X(r)} & \text{if }\pi_X(r-1)\in I\setminus I_X\text{ and }\pi_X(r)\neq \pi_X(r-1),\\
x_{P^X_k(r-1),\pi_X(r-1)}
-2x_{1+P^X_k(r),\pi_X(r)} & \text{if }\pi_X(r)\in I\setminus I_X\text{ and }\pi_X(r)\neq \pi_X(r-1),\\
x_{P^X_k(r-1),\pi_X(r-1)}
-x_{1+P^X_k(r),\pi_X(r)} & \text{otherwise}.
\end{cases}
\end{equation}
We also define
\begin{multline}
{\rm Comb}_{k,\iota}^{X}[\lambda]:=\label{comb-lam-2}\\
\begin{cases}
\{-x_{1,k}+\langle h_k,\lambda \rangle\} & \text{if }(1,k)<(1,\pi_X(k+1)),\ (1,k)<(1,\pi_X(k-1)),\\
\{\fbox{$\tilde{r}$}^{X}_k+\langle h_k,\lambda \rangle\ | r\in\mathbb{Z}_{\leq k}\}
& \text{if }(1,k)<(1,\pi_X(k+1)),\ (1,k)>(1,\pi_X(k-1)),\\
\left\{\fbox{$r$}^{X}_k+\langle h_k,\lambda \rangle\ | r\in\mathbb{Z}_{\geq k+1}\right\}
& \text{if }(1,k)>(1,\pi_X(k+1)),\ (1,k)<(1,\pi_X(k-1)),\\
\{
L^X_{0,k,\iota}(T)+\langle h_k,\lambda \rangle | T\in{\rm REYD}^{X}_{k}\setminus\{\phi^k\}
\} & \text{if }(1,k)>(1,\pi_X(k+1)),\ (1,k)>(1,\pi_X(k-1)).\\
\end{cases}
\end{multline}
Here, $\phi^k\in{\rm REYD}^{X}_{k}$ is the revised extended Young diagram without boxes, that is, $\phi^k=(y_{\ell})_{\ell\in\mathbb{Z}}$,
$y_{\ell}=k$ for any $\ell\in\mathbb{Z}$.
Next, we take $k$ as $k\in I\setminus I_X$.
We draw Young walls on $\mathbb{R}_{\leq0}\times \mathbb{R}_{\geq k}$.
For example, the Young wall in Example \ref{ex-A2-1} is drawn as follow:
\begin{equation*}
\begin{xy}
(-15.5,-2) *{\ 1}="000-1",
(-9.5,-2) *{\ 1}="00-1",
(-3.5,-2) *{\ 1}="0-1",
(1.5,-2) *{\ \ 1}="0-1",
(-21.5,-2) *{\dots}="00000",
(-9.5,1.5) *{\ 1}="000",
(-3.5,1.5) *{\ 1}="00",
(1.5,1.5) *{\ \ 1}="0",
(1.5,6.5) *{\ \ 2}="02",
(-4,6.5) *{\ \ 2}="002",
(1.5,12.5) *{\ \ 3}="03",
(1.5,18.5) *{\ \ 2}="04",
(12,-6) *{(0,1)}="origin",
(0,0) *{}="1",
(0,-3.5) *{}="1-u",
(6,0)*{}="2",
(6,-3.5)*{}="2-u",
(6,-7.5)*{}="-1-u",
(6,3.5)*{}="3",
(0,3.5)*{}="4",
(-6,3.5)*{}="5",
(-6,0)*{}="6",
(-6,-3.5)*{}="6-u",
(-12,3.5)*{}="7",
(-12,0)*{}="8",
(-12,-3.6)*{}="8-u",
(-18,3.5)*{}="9",
(-18,0)*{}="10-a",
(-18,0)*{}="10",
(-18,-3.5)*{}="10-u",
(-24,-3.5)*{}="11-u",
(20,-3.5)*{}="0-u",
(6,9.5)*{}="3-1",
(6,15.5)*{}="3-2",
(6,21.5)*{}="3-3",
(6,25)*{}="3-4",
(0,9.5)*{}="4-1",
(0,15.5)*{}="4-2",
(0,21.5)*{}="4-3",
(-6,9.5)*{}="5-1",
\ar@{-} "-1-u";"2-u"^{}
\ar@{->} "2-u";"0-u"^{}
\ar@{-} "10-u";"11-u"^{}
\ar@{-} "8";"10-a"^{}
\ar@{-} "10-u";"10-a"^{}
\ar@{-} "10-u";"2-u"^{}
\ar@{-} "8";"8-u"^{}
\ar@{-} "6";"6-u"^{}
\ar@{-} "2";"2-u"^{}
\ar@{-} "1";"1-u"^{}
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\ar@{-} "5";"6"^{}
\ar@{-} "5";"4"^{}
\ar@{-} "1";"6"^{}
\ar@{-} "7";"8"^{}
\ar@{-} "7";"5"^{}
\ar@{-} "6";"8"^{}
\ar@{-} "3";"3-1"^{}
\ar@{-} "4-1";"3-1"^{}
\ar@{-} "4-1";"4"^{}
\ar@{-} "5-1";"5"^{}
\ar@{-} "4-1";"5-1"^{}
\ar@{-} "4-1";"4-2"^{}
\ar@{-} "3-2";"3-1"^{}
\ar@{-} "3-2";"4-2"^{}
\ar@{-} "3-2";"3-3"^{}
\ar@{-} "4-3";"4-2"^{}
\ar@{-} "4-3";"3-3"^{}
\ar@{->} "3-3";"3-4"^{}
\end{xy}
\end{equation*}
Let $i\in\mathbb{Z}_{\geq0}$, $\ell\in\mathbb{Z}_{\geq k}$ and $S$ be a slot or block
\[
S=
\begin{xy}
(-8,15) *{(-i-1,\ell+1)}="0000",
(17,15) *{(-i,\ell+1)}="000",
(15,-3) *{(-i,\ell)}="00",
(-7,-3) *{(-i-1,\ell)}="0",
(0,0) *{}="1",
(12,0)*{}="2",
(12,12)*{}="3",
(0,12)*{}="4",
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\end{xy}
\]
in $\mathbb{R}_{\leq0}\times \mathbb{R}_{\geq k}$.
If $S$ is colored by $t\in I$ in the pattern of Definition \ref{def-YW} (ii)
then we define
\begin{equation}\label{l1def1}
L^X_{s,k,{\rm ad}}(S):=x_{s+P^X_k(\ell)+i,t},\quad L^X_{s,k,{\rm re}}(S):=x_{s+P^X_k(\ell)+i+1,t}.
\end{equation}
Let $i\in\mathbb{Z}_{\geq0}$, $\ell\in\mathbb{Z}_{\geq k}$ and $S'$ be
a slot or block colored by $t=1$ or $n$ in $\mathbb{R}_{\leq 0}\times \mathbb{R}_{\geq k}$
such that the place is one of the following two:
\[
S'=
\begin{xy}
(-8,10) *{(-i-1,\ell+\frac{1}{2})}="0000",
(17,10) *{(-i,\ell+\frac{1}{2})}="000",
(15,-3) *{(-i,\ell)}="00",
(-7,-3) *{(-i-1,\ell)}="0",
(0,0) *{}="1",
(12,0)*{}="2",
(12,6)*{}="3",
(0,6)*{}="4",
(32,3) *{{\rm or}}="or",
(52,10) *{(-i-1,\ell+1)}="a0000",
(77,10) *{(-i,\ell+1)}="a000",
(75,-3) *{(-i,\ell+\frac{1}{2})}="a00",
(54,-3) *{(-i-1,\ell+\frac{1}{2})}="a0",
(60,0) *{}="a1",
(72,0)*{}="a2",
(72,6)*{}="a3",
(60,6)*{}="a4",
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\ar@{-} "a1";"a2"^{}
\ar@{-} "a1";"a4"^{}
\ar@{-} "a2";"a3"^{}
\ar@{-} "a3";"a4"^{}
\end{xy}
\]
Then we set
\begin{equation}\label{l1def2}
L^X_{s,k,{\rm ad}}(S'):=x_{s+P^X_k(\ell)+i,t},\quad L^X_{s,k,{\rm re}}(S'):=x_{s+P^X_k(\ell)+i+1,t}.
\end{equation}
If there exists $S$ or $S'$ colored by $t\in I$ as above,
it holds $\pi'(\ell)=t$, where $\pi'$ is defined as (\ref{piprime}).
For a proper Young wall $Y$ of type $X$ of ground state $\Lambda_k$, one defines
\begin{eqnarray}
L^X_{s,k,\iota}(Y)&:=&
\sum_{t\in I} \left(
\sum_{P:\text{single admissible slot}} L^X_{s,k,{\rm ad}}(P)
-\sum_{P:\text{single removable block}} L^X_{s,k,{\rm re}}(P)\right)\nonumber\\
& & +
\sum_{P:\text{double admissible slot}} 2L^X_{s,k,{\rm ad}}(P)
-\sum_{P:\text{double removable block}} 2L^X_{s,k,{\rm re}}(P). \label{L11-def}
\end{eqnarray}
Let ${\rm YW}^{X}_{k}$ denote the set of all proper Young walls of type $X$ of ground state $\Lambda_k$.
We define
\begin{equation}\label{comb-lam-3}
{\rm Comb}_{k,\iota}^{X}[\lambda]:=
\begin{cases}
\{-x_{1,k}+\langle h_k,\lambda \rangle\} & \text{if }(1,k)<(1,\pi'(k+1)),\\
\{
L^X_{0,k,\iota}(Y)+\langle h_k,\lambda \rangle | Y\in{\rm YW}^{X}_{k}\setminus\{\phi^k\}
\} & \text{if }(1,k)>(1,\pi'(k+1)).
\end{cases}
\end{equation}
Here, $\phi^k=Y_{\Lambda_k}\in {\rm YW}^{X}_{k}$ is the ground state wall.
One also defines
\begin{equation}\label{comb-inf-2}
{\rm Comb}^{X}_{\iota}[\infty]
:=\{
L^{X}_{s,k,\iota}(T)|
s\in\mathbb{Z}_{\geq1},\ k\in I_{X}\
\text{and}\ T\in{\rm REYD}^{X}_{k}
\}\cup
\{
L^{X}_{s,k,\iota}(Y)|
s\in\mathbb{Z}_{\geq1},\ k\in I\setminus I_{X}\text{ and }Y\in{\rm YW}^{X}_{k}
\}.
\end{equation}
\begin{thm}\label{thmA2}
We suppose that $\mathfrak{g}$ is of type $X={\rm A}_{2n-2}^{(2)}$ or $X={\rm C}_{n-1}^{(1)}$ $(n\geq3)$ and $\iota$ is adapted.
Then
for any $\lambda\in P^+$, the pair
$(\iota,\lambda)$ is $\Xi'$-ample and
\[
{\rm Im}(\Psi^{(\lambda)}_{\iota})
=
\left\{
\mathbf{a}\in\mathbb{Z}^{\infty} \left|
\varphi(\mathbf{a})\geq0\text{ for any }
\varphi\in
{\rm Comb}^{X^L}_{\iota}[\infty]\cup
\bigcup_{k\in I}
{\rm Comb}^{X^L}_{k,\iota}[\lambda]
\right.
\right\}.
\]
Here, $({\rm A}^{(2)}_{2n-2})^L={\rm A}^{(2)}_{2n-2}$, $({\rm C}^{(1)}_{n-1})^L={\rm D}^{(2)}_{n}$.
\end{thm}
Similarly to Remark \ref{rem-1}, we see that
functions in ${\rm Comb}^{X^L}_{\iota}[\infty]$ and ${\rm Comb}^{X^L}_{k,\iota}[\lambda]$
are related to representation theory of $U_q(^L\mathfrak{g})$.
\begin{ex}
We consider the case $\mathfrak{g}$ is of type ${\rm A}_4^{(2)}$ and $\iota=(\cdots,3,1,2,3,1,2)$, which
is the same setting as in Example 4.9 of \cite{Ka}.
Let us
compute several inequalities defining ${\rm Im}(\Psi_{\iota}^{(\lambda)})$.
In the same example of \cite{Ka},
we already found several inequalities defining
${\rm Im}(\Psi_{\iota})=\{\mathbf{a}\in \mathbb{Z}^{\infty} |
\varphi(\mathbf{a})\geq0\ \text{ for any }\varphi\in {\rm Comb}^{{\rm A}^{(2)}}_{\iota}[\infty]\}$:
\[
{\rm Im}(\Psi_{\iota})
=\left\{
\textbf{a}\in\mathbb{Z}^{\infty} \left|
\begin{array}{l}
s\in\mathbb{Z}_{\geq1},\ a_{s,2}\geq0,
2a_{s,1}+a_{s,3}-a_{s+1,2}\geq0,
a_{s,1}+a_{s,3}-a_{s+1,1}\geq0,\\
a_{s,1}+2a_{s+1,2}-a_{s+1,3}-a_{s+1,1}\geq0,\\
a_{s,1}+a_{s+1,2}+a_{s+1,1}-a_{s+2,2}\geq0,
a_{s,1}+a_{s+1,2}-a_{s+2,1}\geq0,\cdots\\
a_{s,3}\geq0,\
2a_{s+1,2}-a_{s+1,3}\geq0,
2a_{s+1,1}+a_{s+1,2}-a_{s+2,2}\geq0,\cdots\\
a_{s,1}\geq0,
a_{s+1,2}-a_{s+1,1}\geq0,
a_{s+1,3}+a_{s+1,1}-a_{s+2,2}\geq0,\\
a_{s+1,3}-a_{s+2,1}\geq0,
a_{s+2,2}+a_{s+1,1}-a_{s+2,3}\geq0,\cdots
\end{array} \right.
\right\}.
\]
Let us compute several inequalities coming from ${\rm Comb}_{k,\iota}^{{\rm A}^{(2)}}[\lambda]$ for $k=1,2,3$.
We obtain
\[
P_1^{{\rm A}^{(2)}}(1)=0,\ P_1^{{\rm A}^{(2)}}(2)=1,\ P_1^{{\rm A}^{(2)}}(3)=1,\ P_1^{{\rm A}^{(2)}}(4)=2,\cdots,
\]
\[
\cdots,\ P_2^{{\rm A}^{(2)}}(-1)=1,\ P_2^{{\rm A}^{(2)}}(0)=0,\ P_2^{{\rm A}^{(2)}}(1)=0,\ P_2^{{\rm A}^{(2)}}(2)=0,\ P_2^{{\rm A}^{(2)}}(3)=0,\ P_2^{{\rm A}^{(2)}}(4)=1,\cdots,
\]
\[
\cdots,\ P_3^{{\rm A}^{(2)}}(1)=1,\ P_3^{{\rm A}^{(2)}}(2)=1,\ P_3^{{\rm A}^{(2)}}(3)=0,\ P_3^{{\rm A}^{(2)}}(4)=1,\ P_3^{{\rm A}^{(2)}}(5)=1,\cdots.
\]
It follows from definitions that
${\rm Comb}_{2,\iota}^{{\rm A}^{(2)}}[\lambda]=\{-x_{1,2}+\langle h_2,\lambda\rangle\}$.
Next, we consider ${\rm Comb}_{1,\iota}^{{\rm A}^{(2)}}[\lambda]$.
The following Young walls are elements in ${\rm YW}^{{\rm A}^{(2)}}_{1}\setminus\{\phi^1\}$:
\[
\begin{xy}
(30.5,6) *{Y_1:=}="0000-1a",
(52.5,1.5) *{\ 1}="000-1-1a",
(34.5,-2) *{\ 1}="000-1a",
(40.5,-2) *{\ 1}="00-1a",
(46.5,-2) *{\ 1}="0-1a",
(52.5,-2) *{\ 1}="0-1sa",
(29.5,-2) *{\dots}="00000a",
(62,-6) *{(0,1)}="origina",
(50,0) *{}="1a",
(56,18) *{}="ya",
(56,3.5)*{}="6+35a",
(50,3.5)*{}="0+35a",
(50,0)*{}="0+0a",
(50,-3.5) *{}="1-ua",
(56,0)*{}="2a",
(56,-3.5)*{}="2-ua",
(56,-7.5)*{}="-1-ua",
(44,0)*{}="6a",
(44,-3.5)*{}="6-ua",
(38,0)*{}="8a",
(38,-3.6)*{}="8-ua",
(32,3.5)*{}="9a",
(32,0)*{}="10-ta",
(32,0)*{}="10ta",
(32,-3.5)*{}="10-uta",
(27,-3.5)*{}="11-uta",
(70,-3.5)*{}="0-ua",
(80.5,6) *{Y_2:=}="0000-1b",
(102.5,1.5) *{\ 1}="000-1-1b",
(102.5,6.5) *{\ 2}="000-1-1-2b",
(84.5,-2) *{\ 1}="000-1b",
(90.5,-2) *{\ 1}="00-1b",
(96.5,-2) *{\ 1}="0-1b",
(102.5,-2) *{\ 1}="0-1sb",
(79.5,-2) *{\dots}="00000b",
(112,-6) *{(0,1)}="originb",
(100,0) *{}="1b",
(106,18) *{}="yb",
(106,9.5)*{}="6+95b",
(100,9.5)*{}="0+95b",
(106,3.5)*{}="6+35b",
(100,3.5)*{}="0+35b",
(100,0)*{}="0+0b",
(100,-3.5) *{}="1-ub",
(106,0)*{}="2b",
(106,-3.5)*{}="2-ub",
(106,-7.5)*{}="-1-ub",
(94,0)*{}="6b",
(94,-3.5)*{}="6-ub",
(88,0)*{}="8b",
(88,-3.6)*{}="8-ub",
(82,3.5)*{}="9b",
(82,0)*{}="10-tb",
(82,0)*{}="10tb",
(82,-3.5)*{}="10-utb",
(77,-3.5)*{}="11-utb",
(120,-3.5)*{}="0-ub",
\ar@{-} "-1-ua";"2-ua"^{}
\ar@{->} "2-ua";"0-ua"^{}
\ar@{->} "2a";"ya"^{}
\ar@{-} "10-uta";"11-uta"^{}
\ar@{-} "8a";"10-ta"^{}
\ar@{-} "10-uta";"10-ta"^{}
\ar@{-} "10-uta";"2-ua"^{}
\ar@{-} "8a";"8-ua"^{}
\ar@{-} "6a";"6-ua"^{}
\ar@{-} "2a";"2-ua"^{}
\ar@{-} "1a";"1-ua"^{}
\ar@{-} "1a";"2a"^{}
\ar@{-} "1a";"6a"^{}
\ar@{-} "6a";"8a"^{}
\ar@{-} "6+35a";"0+35a"^{}
\ar@{-} "0+0a";"0+35a"^{}
\ar@{-} "-1-ub";"2-ub"^{}
\ar@{->} "2-ub";"0-ub"^{}
\ar@{->} "2b";"yb"^{}
\ar@{-} "10-utb";"11-utb"^{}
\ar@{-} "8b";"10-tb"^{}
\ar@{-} "10-utb";"10-tb"^{}
\ar@{-} "10-utb";"2-ub"^{}
\ar@{-} "8b";"8-ub"^{}
\ar@{-} "6b";"6-ub"^{}
\ar@{-} "2b";"2-ub"^{}
\ar@{-} "1b";"1-ub"^{}
\ar@{-} "1b";"2b"^{}
\ar@{-} "1b";"6b"^{}
\ar@{-} "6b";"8b"^{}
\ar@{-} "6+35b";"0+35b"^{}
\ar@{-} "0+0b";"0+35b"^{}
\ar@{-} "0+95b";"6+95b"^{}
\ar@{-} "0+35b";"0+95b"^{}
\end{xy}
\]
\[
\begin{xy}
(30.5,6) *{Y_3:=}="0000-1a",
(52.5,1.5) *{\ 1}="000-1-1a",
(52.5,6.5) *{\ 2}="000-1-1-2a",
(34.5,-2) *{\ 1}="000-1a",
(40.5,-2) *{\ 1}="00-1a",
(46.5,-2) *{\ 1}="0-1a",
(46.5,1.5) *{\ 1}="0-1-1a",
(52.5,-2) *{\ 1}="0-1sa",
(29.5,-2) *{\dots}="00000a",
(62,-6) *{(0,1)}="origina",
(50,0) *{}="1a",
(56,18) *{}="ya",
(56,9.5)*{}="6+95a",
(50,9.5)*{}="0+95a",
(56,3.5)*{}="6+35a",
(50,3.5)*{}="0+35a",
(50,0)*{}="0+0a",
(50,-3.5) *{}="1-ua",
(56,0)*{}="2a",
(56,-3.5)*{}="2-ua",
(56,-7.5)*{}="-1-ua",
(44,3.5)*{}="-6+35a",
(44,0)*{}="6a",
(44,-3.5)*{}="6-ua",
(38,0)*{}="8a",
(38,-3.6)*{}="8-ua",
(32,3.5)*{}="9a",
(32,0)*{}="10-ta",
(32,0)*{}="10ta",
(32,-3.5)*{}="10-uta",
(27,-3.5)*{}="11-uta",
(70,-3.5)*{}="0-ua",
(80.5,6) *{Y_4:=}="0000-1b",
(102.5,1.5) *{\ 1}="000-1-1b",
(102.5,6.5) *{\ 2}="000-1-1-2b",
(102.5,12.5) *{\ 3}="000-1-1-2-3b",
(84.5,-2) *{\ 1}="000-1b",
(90.5,-2) *{\ 1}="00-1b",
(96.5,-2) *{\ 1}="0-1b",
(102.5,-2) *{\ 1}="0-1sb",
(79.5,-2) *{\dots}="00000b",
(112,-6) *{(0,1)}="originb",
(100,0) *{}="1b",
(106,18) *{}="yb",
(106,15.5)*{}="6+155b",
(100,15.5)*{}="0+155b",
(106,9.5)*{}="6+95b",
(100,9.5)*{}="0+95b",
(106,3.5)*{}="6+35b",
(100,3.5)*{}="0+35b",
(100,0)*{}="0+0b",
(100,-3.5) *{}="1-ub",
(106,0)*{}="2b",
(106,-3.5)*{}="2-ub",
(106,-7.5)*{}="-1-ub",
(94,0)*{}="6b",
(94,-3.5)*{}="6-ub",
(88,0)*{}="8b",
(88,-3.6)*{}="8-ub",
(82,3.5)*{}="9b",
(82,0)*{}="10-tb",
(82,0)*{}="10tb",
(82,-3.5)*{}="10-utb",
(77,-3.5)*{}="11-utb",
(120,-3.5)*{}="0-ub",
\ar@{-} "-1-ua";"2-ua"^{}
\ar@{->} "2-ua";"0-ua"^{}
\ar@{->} "2a";"ya"^{}
\ar@{-} "10-uta";"11-uta"^{}
\ar@{-} "8a";"10-ta"^{}
\ar@{-} "10-uta";"10-ta"^{}
\ar@{-} "10-uta";"2-ua"^{}
\ar@{-} "8a";"8-ua"^{}
\ar@{-} "6a";"6-ua"^{}
\ar@{-} "2a";"2-ua"^{}
\ar@{-} "1a";"1-ua"^{}
\ar@{-} "1a";"2a"^{}
\ar@{-} "1a";"6a"^{}
\ar@{-} "6a";"8a"^{}
\ar@{-} "-6+35a";"0+35a"^{}
\ar@{-} "-6+35a";"6a"^{}
\ar@{-} "6+35a";"0+35a"^{}
\ar@{-} "0+0a";"0+35a"^{}
\ar@{-} "0+95a";"6+95a"^{}
\ar@{-} "0+35a";"0+95a"^{}
\ar@{-} "-1-ub";"2-ub"^{}
\ar@{->} "2-ub";"0-ub"^{}
\ar@{->} "2b";"yb"^{}
\ar@{-} "10-utb";"11-utb"^{}
\ar@{-} "8b";"10-tb"^{}
\ar@{-} "10-utb";"10-tb"^{}
\ar@{-} "10-utb";"2-ub"^{}
\ar@{-} "8b";"8-ub"^{}
\ar@{-} "6b";"6-ub"^{}
\ar@{-} "2b";"2-ub"^{}
\ar@{-} "1b";"1-ub"^{}
\ar@{-} "1b";"2b"^{}
\ar@{-} "1b";"6b"^{}
\ar@{-} "6b";"8b"^{}
\ar@{-} "6+35b";"0+35b"^{}
\ar@{-} "0+0b";"0+35b"^{}
\ar@{-} "0+95b";"6+95b"^{}
\ar@{-} "0+35b";"0+95b"^{}
\ar@{-} "0+155b";"6+155b"^{}
\ar@{-} "0+95b";"0+155b"^{}
\end{xy}
\]
We see that $Y_1$ has a single $2$-admissible slot
and a single removable $1$-block so that $L^{{\rm A}^{(2)}}_{0,1,\iota}(Y_1)=x_{P_1^{{\rm A}^{(2)}}(2),2}-
x_{P_1^{{\rm A}^{(2)}}(1)+1,1}
=x_{1,2}-x_{1,1}$. Similarly,
\[
L^{{\rm A}^{(2)}}_{0,1,\iota}(Y_2)=x_{P_1^{{\rm A}^{(2)}}(3),3}+x_{P_1^{{\rm A}^{(2)}}(1)+1,1}-x_{P_1^{{\rm A}^{(2)}}(2)+1,2}=x_{1,3}+x_{1,1}-x_{2,2}, \]
\[
L^{{\rm A}^{(2)}}_{0,1,\iota}(Y_3)=x_{P_1^{{\rm A}^{(2)}}(3),3}-x_{P_1^{{\rm A}^{(2)}}(1)+2,1}=x_{1,3}-x_{2,1},
\]
\[
L^{{\rm A}^{(2)}}_{0,1,\iota}(Y_4)=x_{P_1^{{\rm A}^{(2)}}(4),2}+x_{P_1^{{\rm A}^{(2)}}(1)+1,1}-x_{P_1^{{\rm A}^{(2)}}(3)+1,3}=x_{2,2}+x_{1,1}-x_{2,3}.
\]
Next, we consider ${\rm Comb}_{3,\iota}^{{\rm A}^{(2)}}[\lambda]$.
The following diagrams are elements in ${\rm REYD}^{{\rm A}^{(2)}}_{3}\setminus\{\phi^3\}$:
\[
\begin{xy}
(40,-9) *{T_3^1:=}="Y1",
(55,0) *{}="1a",
(85,0)*{}="2a",
(50,0)*{(0,3)}="4a",
(61,2) *{1}="10a",
(61,-1) *{}="1010a",
(61,-6) *{}="1010-1a",
(67,2) *{2}="11a",
(67,-1) *{}="1111a",
(73,2) *{3}="12a",
(73,-1) *{}="1212a",
(79,2) *{4}="13a",
(79,-1) *{}="1313a",
(55,-6)*{}="5a",
(49,-6)*{}="55a",
(49,-12)*{}="6a",
(43,-12)*{}="66a",
(43,-18)*{}="7a",
(37,-18)*{}="77a",
(37,-24)*{}="8a",
(31,-24)*{}="88a",
(95,-9) *{T_3^2:=}="Y2",
(110,0) *{}="1b",
(110,-12) *{}="1b-12",
(140,0)*{}="2b",
(105,0)*{(0,3)}="4b",
(116,2) *{1}="10b",
(116,-1) *{}="1010b",
(116,-6) *{}="1010-1b",
(122,2) *{2}="11b",
(122,-1) *{}="1111b",
(128,2) *{3}="12b",
(128,-1) *{}="1212b",
(134,2) *{4}="13b",
(134,-1) *{}="1313b",
(110,-6)*{}="5b",
(104,-6)*{}="55b",
(104,-12)*{}="6b",
(98,-12)*{}="66b",
(98,-18)*{}="7b",
(92,-18)*{}="77b",
(92,-24)*{}="8b",
(86,-24)*{}="88b",
\ar@{-} "1a";"2a"^{}
\ar@{-} "1a";"5a"^{}
\ar@{-} "5a";"55a"^{}
\ar@{-} "55a";"6a"^{}
\ar@{-} "6a";"66a"^{}
\ar@{-} "66a";"7a"^{}
\ar@{-} "7a";"77a"^{}
\ar@{-} "77a";"8a"^{}
\ar@{-} "8a";"88a"^{}
\ar@{-} "10a";"1010a"^{}
\ar@{-} "1010a";"1010-1a"^{}
\ar@{-} "5a";"1010-1a"^{}
\ar@{-} "11a";"1111a"^{}
\ar@{-} "12a";"1212a"^{}
\ar@{-} "13a";"1313a"^{}
\ar@{-} "1b";"2b"^{}
\ar@{-} "1b";"1b-12"^{}
\ar@{-} "66b";"1b-12"^{}
\ar@{-} "1b";"5b"^{}
\ar@{-} "5b";"55b"^{}
\ar@{-} "55b";"6b"^{}
\ar@{-} "6b";"66b"^{}
\ar@{-} "66b";"7b"^{}
\ar@{-} "7b";"77b"^{}
\ar@{-} "77b";"8b"^{}
\ar@{-} "8b";"88b"^{}
\ar@{-} "10b";"1010b"^{}
\ar@{-} "1010b";"1010-1b"^{}
\ar@{-} "5b";"1010-1b"^{}
\ar@{-} "11b";"1111b"^{}
\ar@{-} "12b";"1212b"^{}
\ar@{-} "13b";"1313b"^{}
\end{xy}
\]
\[
\begin{xy}
(-15,-9) *{T_3^3:=}="Y3",
(0,0) *{}="1",
(30,0)*{}="2",
(-5,0)*{(0,3)}="4",
(6,2) *{1}="10",
(6,-1) *{}="1010",
(6,-6) *{}="1010-1",
(12,2) *{2}="11",
(12,-6) *{}="12-6",
(12,-1) *{}="1111",
(18,2) *{3}="12",
(18,-1) *{}="1212",
(24,2) *{4}="13",
(24,-1) *{}="1313",
(0,-6)*{}="5",
(0,-12) *{}="0-12",
(-6,-6)*{}="55",
(-6,-12)*{}="6",
(-6,-18)*{}="6-18",
(0,-18)*{}="0-18",
(-12,-12)*{}="66",
(-12,-18)*{}="7",
(-18,-18)*{}="77",
(-18,-24)*{}="8",
(-24,-24)*{}="88",
(40,-9) *{T_3^4:=}="Y4",
(55,0) *{}="1a",
(85,0)*{}="2a",
(50,0)*{(0,3)}="4a",
(61,2) *{1}="10a",
(61,-1) *{}="1010a",
(61,-6) *{}="6-6a",
(61,-12) *{}="6-12a",
(67,-6) *{}="1010-1a",
(67,2) *{2}="11a",
(67,-1) *{}="1111a",
(73,2) *{3}="12a",
(73,-1) *{}="1212a",
(79,2) *{4}="13a",
(79,-1) *{}="1313a",
(55,-6)*{}="5a",
(49,-6)*{}="55a",
(49,-12)*{}="6a",
(43,-12)*{}="66a",
(43,-18)*{}="7a",
(37,-18)*{}="77a",
(37,-24)*{}="8a",
(31,-24)*{}="88a",
\ar@{-} "1";"2"^{}
\ar@{-} "1";"5"^{}
\ar@{-} "5";"55"^{}
\ar@{-} "55";"6"^{}
\ar@{-} "6";"66"^{}
\ar@{-} "5";"0-12"^{}
\ar@{-} "6";"0-12"^{}
\ar@{-} "66";"7"^{}
\ar@{-} "7";"77"^{}
\ar@{-} "77";"8"^{}
\ar@{-} "8";"88"^{}
\ar@{-} "10";"1010"^{}
\ar@{-} "11";"1111"^{}
\ar@{-} "11";"12-6"^{}
\ar@{-} "12";"1212"^{}
\ar@{-} "13";"1313"^{}
\ar@{-} "12-6";"1010-1"^{}
\ar@{-} "5";"1010-1"^{}
\ar@{-} "1a";"2a"^{}
\ar@{-} "1a";"5a"^{}
\ar@{-} "5a";"55a"^{}
\ar@{-} "55a";"6a"^{}
\ar@{-} "6a";"66a"^{}
\ar@{-} "66a";"7a"^{}
\ar@{-} "7a";"77a"^{}
\ar@{-} "77a";"8a"^{}
\ar@{-} "8a";"88a"^{}
\ar@{-} "10a";"1010a"^{}
\ar@{-} "1111a";"1010-1a"^{}
\ar@{-} "6-12a";"6-6a"^{}
\ar@{-} "6-12a";"6a"^{}
\ar@{-} "1010-1a";"6-6a"^{}
\ar@{-} "11a";"1111a"^{}
\ar@{-} "12a";"1212a"^{}
\ar@{-} "13a";"1313a"^{}
\end{xy}
\]
One can compute
\[
L^{{\rm A}^{(2)}}_{0,3,\iota}(T_3^1)=x_{P_3^{{\rm A}^{(2)}}(4),2}+x_{P_3^{{\rm A}^{(2)}}(2),2}-
x_{P_3^{{\rm A}^{(2)}}(3)+1,3}
=2x_{1,2}-x_{1,3},\]
\[
L^{{\rm A}^{(2)}}_{0,3,\iota}(T_3^2)=2x_{P_3^{{\rm A}^{(2)}}(1),1}+x_{P_3^{{\rm A}^{(2)}}(4),2}-x_{P_3^{{\rm A}^{(2)}}(2)+1,2}
=2x_{1,1}+x_{1,2}-x_{2,2},
\]
\[
L^{{\rm A}^{(2)}}_{0,3,\iota}(T_3^3)=2x_{P_3^{{\rm A}^{(2)}}(1),1}+
x_{P_3^{{\rm A}^{(2)}}(3)+1,3}+
2x_{P_3^{{\rm A}^{(2)}}(5),1}
-x_{P_3^{{\rm A}^{(2)}}(4)+1,2}
-x_{P_3^{{\rm A}^{(2)}}(2)+1,2}
=4x_{1,1}+x_{1,3}
-2x_{2,2},
\]
\[
L^{{\rm A}^{(2)}}_{0,3,\iota}(T_3^4)=2x_{P_3^{{\rm A}^{(2)}}(1),1}+
2x_{P_3^{{\rm A}^{(2)}}(5),1}
-x_{P_3^{{\rm A}^{(2)}}(3)+2,3}
=4x_{1,1}-x_{2,3}.
\]
Note that in $T_3^2$, $T_3^3$ and $T_3^4$, the point $(-2,1)$ is double $1$-admissible
and
in $T_3^3$ and $T_3^4$, the point $(2,3)$ is double $1$-admissible.
Hence, we get several inequalities defining ${\rm Im}(\Psi_{\iota}^{(\lambda)})$:
\[
{\rm Im}(\Psi_{\iota}^{(\lambda)})
=\left\{
\textbf{a}\in\mathbb{Z}^{\infty} \left|
\begin{array}{l}
s\in\mathbb{Z}_{\geq1},\ a_{s,2}\geq0,
2a_{s,1}+a_{s,3}-a_{s+1,2}\geq0,
a_{s,1}+a_{s,3}-a_{s+1,1}\geq0,\\
a_{s,1}+2a_{s+1,2}-a_{s+1,3}-a_{s+1,1}\geq0,\\
a_{s,1}+a_{s+1,2}+a_{s+1,1}-a_{s+2,2}\geq0,
a_{s,1}+a_{s+1,2}-a_{s+2,1}\geq0,\cdots\\
a_{s,3}\geq0,\
2a_{s+1,2}-a_{s+1,3}\geq0,
2a_{s+1,1}+a_{s+1,2}-a_{s+2,2}\geq0,\cdots\\
a_{s,1}\geq0,
a_{s+1,2}-a_{s+1,1}\geq0,
a_{s+1,3}+a_{s+1,1}-a_{s+2,2}\geq0,\\
a_{s+1,3}-a_{s+2,1}\geq0,
a_{s+2,2}+a_{s+1,1}-a_{s+2,3}\geq0,\cdots\\
a_{1,2}\leq \langle h_2,\lambda \rangle,
a_{1,2}-a_{1,1}+\langle h_1,\lambda \rangle\geq0,
a_{1,3}+a_{1,1}-a_{2,2}+\langle h_1,\lambda \rangle\geq0,\\
a_{1,3}-a_{2,1}+\langle h_1,\lambda \rangle\geq0,
a_{2,2}+a_{1,1}-a_{2,3}+\langle h_1,\lambda \rangle\geq0,\\
2a_{1,2}-a_{1,3}+\langle h_3,\lambda \rangle\geq0,
2a_{1,1}+a_{1,2}-a_{2,2}+\langle h_3,\lambda \rangle\geq0,\\
4a_{1,1}+a_{1,3}-2a_{2,2}+\langle h_3,\lambda \rangle\geq0,
4a_{1,1}-a_{2,3}+\langle h_3,\lambda \rangle\geq0,\cdots
\end{array} \right.
\right\}.
\]
\end{ex}
\subsection{Combinatorial description of $\varepsilon_k^*$}
\begin{thm}\label{thm3}
We suppose that $\mathfrak{g}$ is of type $X={\rm A}_{n-1}^{(1)}$ $(n\geq2)$, ${\rm D}_{n}^{(2)}$,
${\rm A}_{2n-2}^{(2)}$ or ${\rm C}_{n-1}^{(1)}$ $(n\geq3)$ and $\iota$ is adapted.
Then for $k\in I$ and $x\in{\rm Im}(\Psi_{\iota})$, we have
\[
\varepsilon_k^*(x)
={\rm max}\{-\varphi(x) | \varphi\in{\rm Comb}_{k,\iota}^{X^L}[0]\}.
\]
\end{thm}
\begin{ex}
Let
$\mathfrak{g}$ be of type ${\rm A}_2^{(1)}$ and $\iota=(\cdots,3,1,2,3,1,2)$, which is the same setting
as in Example \ref{ex-2}. In Example 4.4 of \cite{Ka},
we shown that
for any $a_{2,3},a_{2,1},a_{2,2},a_{1,3},a_{1,1},a_{1,2}\in\mathbb{Z}_{\geq0}$ such that
$a_{1,1}-a_{2,2}\geq0$, $a_{1,3}-a_{2,1}\geq0$, $a_{2,2}-a_{2,3}\geq0$ and $a_{2,1}-a_{2,3}\geq0$,
it follows
$(\cdots,0,0,0,a_{2,3},a_{2,1},a_{2,2},a_{1,3},a_{1,1},a_{1,2})\in{\rm Im}(\Psi_{\iota})$.
Taking $x\in {\rm Im}(\Psi_{\iota})$ as
\begin{equation}\label{x-def}
x=(\cdots,0,0,1,2,3,2,3,3),
\end{equation}
let us confirm Theorem \ref{thm3}. By Section 2 of \cite{NZ},
we have
\[
x^*=
\tilde{f}_2^3 \tilde{f}_1^3 \tilde{f}_3^2 \tilde{f}_2^3 \tilde{f}_1^2 \tilde{f}_3
(\cdots,0,0,0,0,0,0)
=(\cdots,0,0,1,0,1,1,1,1,1,1,2,2,1,1,1)
\]
and $\varepsilon_1^*(x)=1$, $\varepsilon_2^*(x)=3$
and $\varepsilon_3^*(x)=0$.
As seen in Example \ref{ex-2}, it holds
${\rm Comb}_{2,\iota}^{{\rm A}^{(1)}}[0]=\{-x_{1,2}\}$. Thus,
${\rm max}\{-\varphi(x) | \varphi\in{\rm Comb}_{2,\iota}^{{\rm A}^{(1)}}[0]\}={\rm max}\{3,0\}=3$.
We also know that
${\rm Comb}_{1,\iota}^{{\rm A}^{(1)}}[0]=
\{\fbox{$r$}^{{\rm A}^{(1)}}_1| r\in\mathbb{Z}_{\geq 2}\}$.
By $P_1^{{\rm A}^{(1)}}(4)=2$, $P_1^{{\rm A}^{(1)}}(m)\geq3$ for $m\geq5$ and
(\ref{r-tilde-A0}), it follows
$\fbox{$r$}^{{\rm A}^{(1)}}_1(x)=0$ for $r\in \mathbb{Z}_{\geq5}$.
Hence
\begin{eqnarray*}
{\rm max}\{-\varphi(x) | \varphi\in{\rm Comb}_{1,\iota}^{{\rm A}^{(1)}}[0]\}&=&
{\rm max}
\{
-\fbox{$2$}^{{\rm A}^{(1)}}_1(x),\ -\fbox{$3$}^{{\rm A}^{(1)}}_1(x),\ -\fbox{$4$}^{{\rm A}^{(1)}}_1(x)
\}\\
&=&
{\rm max}\{0,1,-1\}=1.
\end{eqnarray*}
In this way, we can confirm
$\varepsilon_k^*(x)
={\rm max}\{-\varphi(x) | \varphi\in{\rm Comb}_{k,\iota}^{{\rm A}^{(1)}}[0]\}$ for $k=1,2$.
Next, we consider the case $k=3$. The definition (\ref{comb-lam-1}) says
${\rm Comb}_{3,\iota}^{{\rm A}^{(1)}}[0]=\{L^{{\rm A}^{(1)}}_{0,3,\iota}(T) | T\in {\rm EYD}_3\setminus\{\phi^3\} \}$.
Note that
for $\varphi = \sum_{m\in \mathbb{Z}_{\geq1},j\in I} c_{m,j}x_{m,j}\in {\rm Comb}_{3,\iota}^{{\rm A}^{(1)}}[0]$,
if $c_{m,j}\geq0$ for all $m\in\{1,2\}$, $j\in I$ then
$-\varphi(x)\leq 0$ by (\ref{x-def}).
By $P^{{\rm A}^{(1)}}_3(3)=0$, $P^{{\rm A}^{(1)}}_3(4)=1$, $P^{{\rm A}^{(1)}}_3(t)\geq2$ for $t\in\mathbb{Z}_{\geq5}$
and $P^{{\rm A}^{(1)}}_3(2)=P^{{\rm A}^{(1)}}_3(1)=P^{{\rm A}^{(1)}}_3(0)=1$, $P^{{\rm A}^{(1)}}_3(t)\geq2$
for $t\in\mathbb{Z}_{\leq-1}$, if $(i,j)$ is a convex corner of $T\in {\rm EYD}_3$ and
$P^{{\rm A}^{(1)}}_3(i+j)+{\rm min}\{3-j,i\}\leq2$ then
\[
(i,j)=(1,2),\ (1,1),\ (1,0),\ (1,-1),\ (2,2)\ \text{or } (2,1).
\]
If $T$ has a convex corner $(i,j)=(1,2)$ then $T=T_{3}^1$ in Example \ref{ex-2} and
$L^{{\rm A}^{(1)}}_{0,3,\iota}(T^1_3)=x_{1,1}+x_{1,2}-x_{1,3}$ and
$-L^{{\rm A}^{(1)}}_{0,3,\iota}(T^1_3)(x)= 2-3-3 \leq0$.
If $T$ has a convex corner $(i,j)=(1,1)$ (resp. $(1,0)$, $(1,-1)$, $(2,2)$) then
the point $(0,1)$ (resp. $(0,0)$, $(0,-1)$, $(2,3)$) is a concave corner in $T$.
Note that $L^{{\rm A}^{(1)}}_{0,3,\iota}(1,1)=x_{2,2}$,
$L^{{\rm A}^{(1)}}_{0,3,\iota}(1,0)=x_{2,1}$,
$L^{{\rm A}^{(1)}}_{0,3,\iota}(1,-1)=x_{2,3}$,
$L^{{\rm A}^{(1)}}_{0,3,\iota}(2,2)=x_{2,1}$
and
$L^{{\rm A}^{(1)}}_{0,3,\iota}(0,1)=x_{1,1}$,
$L^{{\rm A}^{(1)}}_{0,3,\iota}(0,0)=x_{1,3}$,
$L^{{\rm A}^{(1)}}_{0,3,\iota}(0,-1)=x_{2,2}$,
$L^{{\rm A}^{(1)}}_{0,3,\iota}(2,3)=x_{2,2}$.
If $T$ has a convex corner $(i,j)=(2,1)$ then either $(2,2)$
or $(2,3)$ is a concave corner in $T$ and
$L^{{\rm A}^{(1)}}_{0,3,\iota}(2,1)=x_{2,3}$.
By the above argument and (\ref{x-def}), we see that
${\rm max}\{-\varphi(x) | \varphi\in{\rm Comb}_{3,\iota}^{{\rm A}^{(1)}}[0]\}=0=\varepsilon_3^*(x)$.
\end{ex}
\section{Action of $\what{S}'$}
For the proof of theorems in the previous section,
we recall action of $\what{S}'$ on (revised) extended Young diagrams and Young walls.
\subsection{On extended Young diagrams}
In the next proposition,
for two extended Young diagrams $T=(y_r)_{r\in\mathbb{Z}_{\geq0}},\ T'=(y_r')_{r\in\mathbb{Z}_{\geq0}} \in {\rm EYD}_{k}$,
we consider the setting a point $(i,j)$ is a concave corner in $T$ and
$T'$ is obtained by
the following replacement of the concave corner
with a convex corner
and calculate how the values $L^X_{s,k,\iota}(T)$ (defined in (\ref{ovlL2})) are changed.
\begin{equation}\label{change}
\begin{xy}
(-6,13) *{(i,j)}="0000",
(4,-3) *{(i,j-1)}="000",
(10,15) *{(i+1,j)}="00",
(72,0) *{(i+1,j-1)}="a0000",
(54,-3) *{(i,j-1)}="a000",
(60,15) *{(i+1,j)}="a00",
(30,8) *{\rightarrow}="a00",
(0,0) *{\bullet}="1",
(12,12)*{\bullet}="3",
(0,12)*{\bullet}="4",
(50,0) *{\bullet}="a1",
(62,12)*{\bullet}="a3",
(62,0)*{\bullet}="a4",
\ar@{-} "a1";"a4"^{}
\ar@{-} "a3";"a4"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "3";"4"^{}
\end{xy}
\end{equation}
\begin{prop}{\rm \cite{Ka}}\label{prop-closednessAD}
We suppose that $\mathfrak{g}$ is of type $X={\rm A}_{n-1}^{(1)}$ or $X={\rm D}_{n}^{(2)}$.
Let $T=(y_r)_{r\in\mathbb{Z}_{\geq0}},\ T'=(y_r')_{r\in\mathbb{Z}_{\geq0}} \in {\rm EYD}_{k}$
and put $j:=y_i$ with some $i\in\mathbb{Z}_{\geq0}$.
We suppose that
\[
y_i'=y_i-1(=j-1),\quad y_r'=y_r\ (r\neq i)
\]
and take $s\in\mathbb{Z}_{\geq0}$. If $s+P_k^{X^L}(i+j)+{\rm min}\{i,k-j\}\geq1$
then
\[
L^{X^L}_{s,k,\iota}(T')=L^{X^L}_{s,k,\iota}(T)-\beta_{s+P_k^{X^L}(i+j)+{\rm min}\{i,k-j\},\pi_{X^L}(i+j)}.
\]
\end{prop}
\subsection{On revised extended Young diagrams}
In this subsection, we assume $k\in I_X$.
\begin{prop}\label{A2closed}\cite{Ka}
We assume $\mathfrak{g}$ is of type ${\rm A}^{(2)}_{2n-2}$.
Let $T=(y_t)_{t\in\mathbb{Z}}$ be a sequence in ${\rm REYD}^{{\rm A}^{(2)}}_{k}$ (Definition \ref{AEYD}), $s\in\mathbb{Z}_{\geq0}$
and $i\in\mathbb{Z}$.
\begin{enumerate}
\item We assume that the point $(i,y_i)$ is single or double admissible and a sequence
$T'=(y_t')_{t\in\mathbb{Z}}\in{\rm REYD}^{{\rm A}^{(2)}}_{k}$
satisfies $y_{i}'=y_i-1$ and $y_t'=y_t$ $(t\neq i)$. Putting $j:=y_i$, if $s+P_k^{{\rm A}^{(2)}}(i+k)+[i]_-+k-j\geq1$ then
\[
L^{{\rm A}^{(2)}}_{s,k,\iota}(T')-L^{{\rm A}^{(2)}}_{s,k,\iota}(T)=-\beta_{s+P_k^{{\rm A}^{(2)}}(i+k)+[i]_-+k-j,\pi_{{\rm A}^{(2)}}(i+k)}.
\]
\item We assume that the point $(i,y_{i-1})$ is single or double removable and a sequence $T''=(y_t'')_{t\in\mathbb{Z}}\in {\rm REYD}^{{\rm A}^{(2)}}_{k}$ satisfies $y_{i-1}''=y_{i-1}+1$ and $y_t''=y_t$ $(t\neq i-1)$. Putting $j:=y_{i-1}''=y_{i-1}+1$,
if $s+P_k^{{\rm A}^{(2)}}(i+k-1)+[i-1]_-+k-j\geq1$
then
\[
L^{{\rm A}^{(2)}}_{s,k,\iota}(T'')-L^{{\rm A}^{(2)}}_{s,k,\iota}(T)=\beta_{s+P_k^{{\rm A}^{(2)}}(i+k-1)+[i-1]_-+k-j,\pi_{{\rm A}^{(2)}}(i+k-1)}.
\]
\end{enumerate}
\end{prop}
\begin{prop}\label{D2closed}\cite{Ka}
We assume $\mathfrak{g}$ is of type ${\rm C}^{(1)}_{n-1}$.
Let $T=(y_t)_{t\in\mathbb{Z}}$ be a sequence in ${\rm REYD}^{{\rm D}^{(2)}}_{k}$ (Definition \ref{DEYD}), $s\in\mathbb{Z}_{\geq0}$ and $i\in\mathbb{Z}$.
\begin{enumerate}
\item We assume that the point $(i,y_i)$ is single or double admissible and a sequence
$T'=(y_t')_{t\in\mathbb{Z}}\in{\rm REYD}^{{\rm D}^{(2)}}_{k}$ satisfies
$y_{i}'=y_i-1$ and $y_t'=y_t$ $(t\neq i)$. Putting $j:=y_i$, if $s+P_k^{{\rm D}^{(2)}}(i+k)+[i]_-+k-j\geq1$ then
\[
L^{{\rm D}^{(2)}}_{s,k,\iota}(T')-L^{{\rm D}^{(2)}}_{s,k,\iota}(T)=-\beta_{s+P_k^{{\rm D}^{(2)}}(i+k)+[i]_-+k-j,\pi_{{\rm D}^{(2)}}(i+k)}.
\]
\item We assume that the point $(i,y_{i-1})$ is single or double removable and $T''=(y_t'')_{t\in\mathbb{Z}}\in{\rm REYD}^{{\rm D}^{(2)}}_{k}$
satisfies $y_{i-1}''=y_{i-1}+1$ and $y_t''=y_t$ $(t\neq i-1)$. Putting $j:=y_{i-1}''=y_{i-1}+1$,
if $s+P_k^{{\rm D}^{(2)}}(i+k-1)+[i-1]_-+k-j\geq1$ then
\[
L^{{\rm D}^{(2)}}_{s,k,\iota}(T'')-L^{{\rm D}^{(2)}}_{s,k,\iota}(T)=\beta_{s+P_k^{{\rm D}^{(2)}}(i+k-1)+[i-1]_-+k-j,\pi_{{\rm D}^{(2)}}(i+k-1)}.
\]
\end{enumerate}
\end{prop}
\subsection{On proper Young walls}
\begin{prop}\label{prop-closednessAw-YW}\cite{Ka}
We assume $\mathfrak{g}$ is of type ${\rm A}^{(2)}_{2n-2}$.
\begin{enumerate}
\item
For $t\in I\setminus\{1\}$,
we assume that a wall $Y\in {\rm YW}^{{\rm A}^{(2)}}_{1}$ has a $t$-admissible slot
\[
\begin{xy}
(-8,15) *{(-i-1,\ell+1)}="0000",
(17,15) *{(-i,\ell+1)}="000",
(15,-3) *{(-i,\ell)}="00",
(-7,-3) *{(-i-1,\ell)}="0",
(0,0) *{}="1",
(12,0)*{}="2",
(12,12)*{}="3",
(0,12)*{}="4",
\ar@{--} "1";"2"^{}
\ar@{--} "1";"4"^{}
\ar@{--} "2";"3"^{}
\ar@{--} "3";"4"^{}
\end{xy}
\]
Let $Y'$ be a wall obtained from $Y$
by adding the $t$-block to the slot.
If $Y'\in{\rm YW}^{{\rm A}^{(2)}}_{1}$
then for $s\in\mathbb{Z}_{\geq0}$
such that $s+P_1^{{\rm A}^{(2)}}(\ell)+i\geq1$,
it follows
\[
L^{{\rm A}^{(2)}}_{s,1,\iota}(Y')=L^{{\rm A}^{(2)}}_{s,1,\iota}(Y)-\beta_{s+P_1^{{\rm A}^{(2)}}(\ell)+i,t}.
\]
\item We assume that a wall $Y\in {\rm YW}^{{\rm A}^{(2)}}_{1}$ has a $1$-admissible slot
\[
\begin{xy}
(-8,9) *{(-i-1,\ell+\frac{1}{2})}="0000",
(17,9) *{(-i,\ell+\frac{1}{2})}="000",
(15,-3) *{(-i,\ell)}="00",
(-7,-3) *{(-i-1,\ell)}="0",
(0,0) *{}="1",
(12,0)*{}="2",
(12,6)*{}="3",
(0,6)*{}="4",
(38,3)*{{\rm or}}="or",
(60,9) *{(-i-1,\ell+1)}="a0000",
(85,9) *{(-i,\ell+1)}="a000",
(83,-3) *{(-i,\ell+\frac{1}{2})}="a00",
(61,-3) *{(-i-1,\ell+\frac{1}{2})}="a0",
(68,0) *{}="a1",
(80,0)*{}="a2",
(80,6)*{}="a3",
(68,6)*{}="a4",
\ar@{--} "1";"2"^{}
\ar@{--} "1";"4"^{}
\ar@{--} "2";"3"^{}
\ar@{--} "3";"4"^{}
\ar@{--} "a1";"a2"^{}
\ar@{--} "a1";"a4"^{}
\ar@{--} "a2";"a3"^{}
\ar@{--} "a3";"a4"^{}
\end{xy}
\]
Let $Y'$ be a wall obtained from $Y$
by adding the $1$-block to the slot.
If $Y'\in{\rm YW}^{{\rm A}^{(2)}}_{1}$
then for $s\in\mathbb{Z}_{\geq0}$
such that $s+P_1^{{\rm A}^{(2)}}(\ell)+i\geq1$,
it follows
\[
L^{{\rm A}^{(2)}}_{s,1,\iota}(Y')=L^{{\rm A}^{(2)}}_{s,1,\iota}(Y)-\beta_{s+P_1^{{\rm A}^{(2)}}(\ell)+i,1}.
\]
\end{enumerate}
\end{prop}
\begin{prop}\label{prop-closednessCw-YW}\cite{Ka}
We assume $\mathfrak{g}$ is of type ${\rm C}^{(1)}_{n-1}$ and $k\in\{1,n\}$.
\begin{enumerate}
\item
Let $t\in I\setminus\{1,n\}$ and
we assume that a wall $Y\in {\rm YW}^{{\rm D}^{(2)}}_{k}$ has a $t$-admissible slot
\[
\begin{xy}
(-8,15) *{(-i-1,\ell+1)}="0000",
(17,15) *{(-i,\ell+1)}="000",
(15,-3) *{(-i,\ell)}="00",
(-7,-3) *{(-i-1,\ell)}="0",
(0,0) *{}="1",
(12,0)*{}="2",
(12,12)*{}="3",
(0,12)*{}="4",
\ar@{--} "1";"2"^{}
\ar@{--} "1";"4"^{}
\ar@{--} "2";"3"^{}
\ar@{--} "3";"4"^{}
\end{xy}
\]
Let $Y'$ be a wall obtained from $Y$
by adding the $t$-block to the slot.
If $Y'\in{\rm YW}^{{\rm D}^{(2)}}_{k}$
then for $s\in\mathbb{Z}_{\geq0}$
such that $s+P_k^{{\rm D}^{(2)}}(\ell)+i\geq1$,
it follows
\[
L^{{\rm D}^{(2)}}_{s,k,\iota}(Y')=L^{{\rm D}^{(2)}}_{s,k,\iota}(Y)-\beta_{s+P_k^{{\rm D}^{(2)}}(\ell)+i,t}.
\]
\item Let $t$ be $t=1$ or $t=n$. We assume that a wall $Y\in {\rm YW}^{{\rm D}^{(2)}}_{k}$ has a $t$-admissible slot
\[
\begin{xy}
(-8,9) *{(-i-1,\ell+\frac{1}{2})}="0000",
(17,9) *{(-i,\ell+\frac{1}{2})}="000",
(15,-3) *{(-i,\ell)}="00",
(-7,-3) *{(-i-1,\ell)}="0",
(0,0) *{}="1",
(12,0)*{}="2",
(12,6)*{}="3",
(0,6)*{}="4",
(38,3)*{{\rm or}}="or",
(60,9) *{(-i-1,\ell+1)}="a0000",
(85,9) *{(-i,\ell+1)}="a000",
(83,-3) *{(-i,\ell+\frac{1}{2})}="a00",
(61,-3) *{(-i-1,\ell+\frac{1}{2})}="a0",
(68,0) *{}="a1",
(80,0)*{}="a2",
(80,6)*{}="a3",
(68,6)*{}="a4",
\ar@{--} "1";"2"^{}
\ar@{--} "1";"4"^{}
\ar@{--} "2";"3"^{}
\ar@{--} "3";"4"^{}
\ar@{--} "a1";"a2"^{}
\ar@{--} "a1";"a4"^{}
\ar@{--} "a2";"a3"^{}
\ar@{--} "a3";"a4"^{}
\end{xy}
\]
Let $Y'$ be a wall obtained from $Y$
by adding the $t$-block to the slot.
If $Y'\in{\rm YW}^{{\rm D}^{(2)}}_{k}$
then for $s\in\mathbb{Z}_{\geq0}$ such that $s+P_k^{{\rm D}^{(2)}}(\ell)+i\geq1$,
it follows
\[
L^{{\rm D}^{(2)}}_{s,k,\iota}(Y')=L^{{\rm D}^{(2)}}_{s,k,\iota}(Y)-\beta_{s+P_k^{{\rm D}^{(2)}}(\ell)+i,t}.
\]
\end{enumerate}
\end{prop}
\section{Proofs}
In this section, we complete proofs of theorems in Section 4.
\subsection{Type ${\rm A}_{n-1}^{(1)}$-case and ${\rm D}_{n}^{(2)}$-case}
In this subsection, let $\mathfrak{g}$ be of type $X={\rm A}_{n-1}^{(1)}$ or
$X={\rm D}_{n}^{(2)}$ and prove Theorem \ref{thmA1}.
First, we consider the first set
$\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1}x_{j_0} | \ell\in\mathbb{Z}_{\geq0}, j_0,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}$
of right hand side of (\ref{xilamdef}).
By Theorem 4.3, 4.5 of \cite{Ka}, the sequence $\iota$ satisfies the $\Xi'$-positivity condition
and it follows from (\ref{twoS2}) and the proofs of Theorem 4.3, 4.5 of \cite{Ka} that
\begin{eqnarray}
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1}x_{j_0} | \ell\in\mathbb{Z}_{\geq0}, j_0,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
&=&\{ S'_{j_{\ell}}\cdots S'_{j_1}x_{j_0} | \ell\in\mathbb{Z}_{\geq0}, j_0,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}\nonumber\\
&=&\{ S'_{j_{\ell}}\cdots S'_{j_1}x_{s,k} | \ell\in\mathbb{Z}_{\geq0}, j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1},\ k\in I,\ s\in\mathbb{Z}_{\geq1} \}
\nonumber\\
&=& \{ L^{X^L}_{s,k,\iota}(T)| k\in I,\ s\in\mathbb{Z}_{\geq1}, T\in {\rm EYD}_k \}\label{half1}\\
&=&{\rm Comb}_{\iota}^{X^L}[\infty].\nonumber
\end{eqnarray}
Next, let us consider the second set
$\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ k\in I,\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}$
of (\ref{xilamdef}). We fix an index $k\in I$. In this proof, we simply write $P_k^{X^L}$ and $\pi_{X^L}$ as
$P_k$ and $\pi$.
\vspace{2mm}
\nd
\underline{Case 1 : it holds $(1,k)<(1,\pi(k-1)), (1,\pi(k+1))$}
\vspace{2mm}
Since $\lambda^{(k)}=-x_{1,k}+\langle h_k,\lambda \rangle$, it is clear that
$\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
=\{0,\lambda^{(k)}\}$.
\vspace{2mm}
\nd
\underline{Case 2 : it holds $(1,k)>(1,\pi(k-1))$ and $(1,k)<(1,\pi(k+1))$}
\vspace{2mm}
In this case, the definition of $\lambda^{(k)}$ implies
$\lambda^{(k)}=-x_{1,k}+x_{1,\pi(k-1)}+\langle h_k,\lambda \rangle$.
Using the notation in (\ref{r-tilde-A1}), one can express $\lambda^{(k)}$ as $\lambda^{(k)}=\fbox{$\tilde{k}$}^{{\rm X}^{L}}_k
+\langle h_k,\lambda \rangle$.
Following the definition of $\what{S}'$,
we get the following diagram of actions of $\what{S}'$:
\[
\begin{xy}
(-8,0) *{-\langle h_k,\lambda \rangle}="00",
(0,0) *{\ }="0",
(10,0) *{\fbox{$\tilde{k}$}^{X^L}}="r-1r-1",
(40,0)*{\fbox{$\overset{\sim}{k-1}$}^{X^L}}="r-2r-2",
(70,0)*{\fbox{$\overset{\sim}{k-2}$}^{X^L}}="r-3r-3",
(100,0)*{\fbox{$\overset{\sim}{k-3}$}^{X^L}}="r-4r-4",
(130,0)*{\cdots}="dot",
\ar@{->} "r-1r-1";"00"^{\what{S}'_{1,k}}
\ar@/^/ @{->} "r-1r-1";"r-2r-2"^{\what{S}'_{P_k(k-1),\pi(k-1)}}
\ar@/^/ @{->} "r-2r-2";"r-3r-3"^{\what{S}'_{P_k(k-2),\pi(k-2)}}
\ar@/^/ @{->} "r-3r-3";"r-4r-4"^{\what{S}'_{P_k(k-3),\pi(k-3)}}
\ar@/^/ @{->} "r-4r-4";"dot"^{\what{S}'_{P_k(k-4),\pi(k-4)}}
\ar@/_/ @{<-} "r-1r-1";"r-2r-2"_{\what{S}'_{1+P_k(k-1),\pi(k-1)}}
\ar@/_/ @{<-} "r-2r-2";"r-3r-3"_{\what{S}'_{1+P_k(k-2),\pi(k-2)}}
\ar@/_/ @{<-} "r-3r-3";"r-4r-4"_{\what{S}'_{1+P_k(k-3),\pi(k-3)}}
\ar@/_/ @{<-} "r-4r-4";"dot"_{\what{S}'_{1+P_k(k-4),\pi(k-4)}}
\end{xy}
\]
Here, we omitted index $k$ of boxes.
Other actions of $\what{S}'$ are trivial.
Note that by $p_{\pi(k-1),\pi(k)}=1$, if $r<k$ then
\begin{equation}\label{stpspr1}
1+P_k(r)\geq2,
\end{equation}
which is the left index of $-x_{1+P_k(r),\pi(r)}$ in (\ref{r-tilde-A1}).
Thus,
\[
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
=\{0,\fbox{$\tilde{r}$}^{X^L}_k+\langle h_k,\lambda \rangle\ | r\in\mathbb{Z}_{\leq k}\}.
\]
\vspace{2mm}
\nd
\underline{Case 3 : it holds $(1,k)<(1,\pi(k-1))$ and $(1,k)>(1,\pi(k+1))$}
\vspace{2mm}
In this case, we have
$\lambda^{(k)}=-x_{1,k}+x_{1,\pi(k+1)}+\langle h_k,\lambda \rangle$.
Using the notation in (\ref{r-tilde-A0}),
it holds
$\lambda^{(k)}=\fbox{$k+1$}^{X^L}_k+\langle h_k,\lambda \rangle$.
One obtains the following diagram of actions of $\what{S}'$:
\[
\begin{xy}
(-8,0) *{-\langle h_k,\lambda \rangle}="00",
(0,0) *{\ }="0",
(10,0) *{\fbox{$k+1$}^{X^L}}="r-1r-1",
(40,0)*{\fbox{$k+2$}^{X^L}}="r-2r-2",
(70,0)*{\fbox{$k+3$}^{X^L}}="r-3r-3",
(100,0)*{\fbox{$k+4$}^{X^L}}="r-4r-4",
(130,0)*{\cdots}="dot",
\ar@{->} "r-1r-1";"00"^{\what{S}'_{1,k}}
\ar@/^/ @{->} "r-1r-1";"r-2r-2"^{\what{S}'_{P_k(k+1),\pi(k+1)}}
\ar@/^/ @{->} "r-2r-2";"r-3r-3"^{\what{S}'_{P_k(k+2),\pi(k+2)}}
\ar@/^/ @{->} "r-3r-3";"r-4r-4"^{\what{S}'_{P_k(k+3),\pi(k+3)}}
\ar@/^/ @{->} "r-4r-4";"dot"^{\what{S}'_{P_k(k+4),\pi(k+4)}}
\ar@/_/ @{<-} "r-1r-1";"r-2r-2"_{\what{S}'_{1+P_k(k+1),\pi(k+1)}}
\ar@/_/ @{<-} "r-2r-2";"r-3r-3"_{\what{S}'_{1+P_k(k+2),\pi(k+2)}}
\ar@/_/ @{<-} "r-3r-3";"r-4r-4"_{\what{S}'_{1+P_k(k+3),\pi(k+3)}}
\ar@/_/ @{<-} "r-4r-4";"dot"_{\what{S}'_{1+P_k(k+4),\pi(k+4)}}
\end{xy}
\]
Here, we omitted index $k$.
Other actions of $\what{S}'$ are trivial. Note that for $r>k+1$, it holds
\begin{equation}\label{stpspr2}
1+P_k(r-1)\geq2,
\end{equation}
which is the left index of $-x_{1+P_k(r-1),\pi(r-1)}$ in (\ref{r-tilde-A0}).
Thus,
\[
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
=\{0,\fbox{$r$}^{X^L}_k+\langle h_k,\lambda \rangle\ | r\in\mathbb{Z}_{\geq k+1}\}.
\]
\vspace{2mm}
\nd
\underline{Case 4 : it holds $(1,k)>(1,\pi(k-1))$ and $(1,k)>(1,\pi(k+1))$}
\vspace{2mm}
In this case, it holds $\lambda^{(k)}=-x_{1,k}+x_{1,\pi(k-1)}+x_{1,\pi(k+1)}+
\lan h_k,\lambda \ran$. Note that if $(i,j)\in\mathbb{Z}_{\geq0}\times\mathbb{Z}_{\leq k}\setminus\{(0,k)\}$
then
\begin{equation}\label{geq-1}
P_k(i+j)+{\rm min}\{k-j,i\}\geq1.
\end{equation}
First, let us prove
\begin{equation}\label{A-pr-1}
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
\subset
\{L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle | T\in{\rm EYD}_{k}\setminus\{\phi^k\}\}\cup\{0\}.
\end{equation}
We put the right hand side of (\ref{A-pr-1}) as $Z$:
$Z:=\{L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle | T\in{\rm EYD}_{k}\setminus\{\phi^k\}\}\cup\{0\}$.
We take $T^1_k\in{\rm EYD}_{k}\setminus\{\phi^k\}$ as follows:
\[
\begin{xy}
(40,-9) *{T^1_k:=}="Y1",
(55,0) *{}="1a",
(85,0)*{}="2a",
(55,-30)*{}="3a",
(50,0)*{(0,k)}="4a",
(61,2) *{1}="10a",
(61,-1) *{}="1010a",
(61,-6) *{}="1-0a",
(67,2) *{2}="11a",
(67,-1) *{}="1111a",
(73,2) *{3}="12a",
(73,-1) *{}="1212a",
(79,2) *{4}="13a",
(79,-1) *{}="1313a",
(50,-6)*{k-1}="5a",
(56,-6)*{}="55a",
(50,-12)*{k-2}="6a",
(56,-12)*{}="66a",
(50,-18)*{k-3}="7a",
(56,-18)*{}="77a",
(50,-24)*{k-4}="8a",
(56,-24)*{}="88a",
\ar@{-} "1a";"2a"^{}
\ar@{-} "1a";"3a"^{}
\ar@{-} "5a";"55a"^{}
\ar@{-} "6a";"66a"^{}
\ar@{-} "7a";"77a"^{}
\ar@{-} "8a";"88a"^{}
\ar@{-} "10a";"1010a"^{}
\ar@{-} "11a";"1111a"^{}
\ar@{-} "12a";"1212a"^{}
\ar@{-} "13a";"1313a"^{}
\ar@{-} "55a";"1-0a"^{}
\ar@{-} "1010a";"1-0a"^{}
\end{xy}
\]
The points $(1,k)$ and $(0,k-1)$ are concave corners and
$(1,k-1)$ is the unique convex corners in $T^1_k$.
Considering $P_k(k+1)=p_{\pi(k+1),\pi(k)}=1$, $P_k(k-1)=p_{\pi(k-1),\pi(k)}=1$,
it follows
\[
L^{X^L}_{0,k,\iota}(T^1_k)+\lan h_k,\lambda \ran=
-x_{1,k}+x_{1,\pi(k-1)}+x_{1,\pi(k+1)}+
\lan h_k,\lambda \ran=\lambda^{(k)}.
\]
Hence, $\lambda^{(k)}\in \{L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle | T\in{\rm EYD}_{k}\setminus\{\phi^k\}\}$.
Thus, to prove (\ref{A-pr-1}), we need to show
$Z$
is closed under the action of $\what{S}'_r$ for all $r\in\mathbb{Z}_{\geq1}$.
For any $T=(y_r)_{r\in\mathbb{Z}_{\geq1}}\in {\rm EYD}_{k}\setminus\{\phi^k\}$, if the coefficient of $x_r$ is positive in
$L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle$ then
$T$ has a concave corner $(i,j)$ such that $j=y_i$ and
\[
r=(P_k(i+j)+{\rm min}\{k-j,i\},\pi(i+j))
\]
by the definition of $L^{X^L}_{0,k,\iota}$ in (\ref{ovlL1}), (\ref{ovlL2}). Note that
we identify $\mathbb{Z}_{\geq1}$ with $\mathbb{Z}_{\geq1}\times I$ by the method of subsection \ref{seno}.
Following the definition of $\what{S}'_r$,
one obtains
\[
\what{S}'_r(L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle)
=L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle - \beta_{P_k(i+j)+{\rm min}\{k-j,i\},\pi(i+j)}.
\]
Taking $T'=(y_r')_{r\in\mathbb{Z}_{\geq1}}\in {\rm EYD}_{k}\setminus\{\phi^k\}$ such that
$y_i'=y_i-1$, $y_r'=y_r$ ($r\neq i$), it follows by Proposition \ref{prop-closednessAD}
that
\begin{eqnarray*}
\what{S}'_r(L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle)
&=&L^{X^L}_{0,k,\iota}(T)- \beta_{P_k(i+j)+{\rm min}\{k-j,i\},\pi(i+j)}+\langle h_k,\lambda \rangle\\
&=&
L^{X^L}_{0,k,\iota}(T')+\langle h_k,\lambda \rangle\in
Z.
\end{eqnarray*}
If the coefficient of $x_r$ is $0$ then
it is clear $\what{S}'_r(L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle)=L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle
\in
Z$.
We assume
the coefficient of $x_r$ is negative in
$L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle$.
Then
$T$ has a convex corner $(i,j)$ such that $j=y_{i-1}$ and
\[
r=(P_k(i+j)+{\rm min}\{k-j,i\},\pi(i+j)).
\]
Note that if $(i,j)=(1,k-1)$ then $T=T^1_k$ and $\what{S}'_r(L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle)=0$.
Thus, we may assume $(i,j)\neq(1,k-1)$.
Since $(i,j)$ is convex, we see that ${\rm min}\{k-j,i\}\geq1$.
In case of $i+j\neq k$, one can verify $P_k(i+j)\geq1$ by
$p_{\pi(k+1),\pi(k)}=p_{\pi(k-1),\pi(k)}=1$ so that $P_k(i+j)+{\rm min}\{k-j,i\}\geq2$.
In case of $i+j=k$,
since we assumed $(i,j)\neq(1,k-1)$,
it follows $i=k-j\geq2$ and $P_k(i+j)+{\rm min}\{k-j,i\}={\rm min}\{k-j,i\}\geq2$.
Hence,
it holds
\begin{equation}\label{A-pr-1a}
P_k(i+j)+{\rm min}\{k-j,i\}\geq2
\end{equation}
and
taking $T''=(y_r'')_{r\in\mathbb{Z}_{\geq1}}$ as $y_{i-1}''=y_{i-1}+1$, $y_r''=y_r$ ($r\neq i-1$),
the point $(i-1,y_{i-1}'')=(i-1,j+1)$ is a concave corner in
$T''$ and
it follows
from Proposition \ref{prop-closednessAD} that
\begin{eqnarray}
\what{S}'_r(L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle)
&=&
L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle+
\beta_{P_k(i+j)+{\rm min}\{k-j,i\}-1,\pi(i+j)}\nonumber\\
&=&
L^{X^L}_{0,k,\iota}(T)+
\beta_{P_k((i-1)+(j+1))+{\rm min}\{k-(j+1),i-1\},\pi((i-1)+(j+1))}+\langle h_k,\lambda \rangle \nonumber\\
&=&
L^{X^L}_{0,k,\iota}(T'')+\langle h_k,\lambda \rangle
\in
Z. \label{A-pr1-b}
\end{eqnarray}
Thus, we obtain (\ref{A-pr-1}).
Next, let us show
\begin{equation}\label{A-pr-2}
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
\supset Z.
\end{equation}
Since we know $0\in \{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}$, let us prove
for any $T\in{\rm EYD}_{k}\setminus\{\phi^k\}$,
it holds
\begin{equation}\label{A-pr-3}
L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle\in
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}.
\end{equation}
We identify each $T\in {\rm EYD}_k\setminus\{\phi^k\}$ as a Young diagram consisting of several boxes, where each box is the square
whose length of sides are $1$. Let us show (\ref{A-pr-3}) by induction on the number of boxes in $T$.
If $T$ has only one box then $T=T^1_k$ and
$L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle=-x_{1,k}+x_{1,\pi(k-1)}+x_{1,\pi(k+1)}+
\lan h_k,\lambda \ran=\lambda^{(k)}\in
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}$.
Thus,
we may assume that $T$ has at least two boxes.
By Proposition \ref{prop-closednessAD}, (\ref{geq-1}) and
$L^{X^L}_{0,k,\iota}(T^k_1)+\langle h_k,\lambda \rangle=\lambda^{(k)}$, we see that
\[
L^{X^L}_{0,k,\iota}(T)+\lan h_k,\lambda \ran
=
\lambda^{(k)} - \sum_{(t,d)\in\mathbb{Z}_{\geq1}\times I} c_{t,d}\beta_{t,d}
\]
with some non-negative integers $\{c_{t,d}\}$ such that $c_{t,d}=0$ except for finitely many $(t,d)$.
Since $T\neq T^1_k$,
there exists $(t',d')\in\mathbb{Z}_{\geq1}\times I$ such that
\[
(t',d')={\rm max}\{(t,d)\in\mathbb{Z}_{\geq1}\times I | c_{t,d}\neq0 \}.
\]
The definition (\ref{betak}) of $\beta_{t,d}$ and Definition \ref{adapt} mean
the coefficient of $x_{t'+1,d'}$ is negative in $L^{X^L}_{0,k,\iota}(T)$. We define $(t'',d'')$ as
\[
(t'',d'')={\rm min}\{(t,d)\in\mathbb{Z}_{\geq1}\times I | \text{the coefficient of }x_{t,d}\ \text{in } L^{X^L}_{0,k,\iota}(T)\text{ is negative} \}.
\]
By the definition of $L^{X^L}_{0,k,\iota}(T)$,
$T$ has a convex corner $(i+1,j-1)$ such that $i\in\mathbb{Z}_{\geq0}$, $j\in\mathbb{Z}_{\leq k}$ and $L^{X^L}_{0,k,\iota}(i+1,j-1)=x_{t'',d''}$.
Just as in (\ref{A-pr-1a}), one can prove
$t''-1\geq1$.
Let $T'\in {\rm EYD}_k\setminus\{\phi^k\}$ be
the extended Young diagram obtained from $T$ by replacing the
convex corner $(i+1,j-1)$ by a concave corner.
Just as in (\ref{A-pr1-b}), we obtain
\[
L^{X^L}_{0,k,\iota}(T)+\beta_{t''-1,d''}=L^{X^L}_{0,k,\iota}(T').
\]
The minimality of $(t'',d'')$ means the coefficient of $x_{t''-1,d''}$ in
$L^{X^L}_{0,k,\iota}(T')$ is positive and
\begin{equation}\label{A-pr-4}
\what{S}'_{t''-1,d''}(L^{X^L}_{0,k,\iota}(T')+\lan h_k,\lambda \ran)=L^{X^L}_{0,k,\iota}(T')+\lan h_k,\lambda \ran-\beta_{t''-1,d''}=L^{X^L}_{0,k,\iota}(T)+\lan h_k,\lambda \ran.
\end{equation}
Note that the number of boxes in $T'$ is smaller than $T$. Hence,
the induction assumption implies
\[
L^{X^L}_{0,k,\iota}(T')+\lan h_k,\lambda \ran=
\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)}
\]
with some $j_1,\cdots,j_{\ell}\in\mathbb{Z}_{\geq1}$. Combining with
(\ref{A-pr-4}), one obtains
$L^{X^L}_{0,k,\iota}(T)+\lan h_k,\lambda \ran\in \{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}$, which implies (\ref{A-pr-3}) so that $(\ref{A-pr-2})$.
By the above argument, it follows
\begin{equation}\label{stpspr3}
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
={\rm Comb}^{X^L}_k[\lambda]\cup\{0\}.
\end{equation}
Combining with (\ref{half1}), one can confirm the $\Xi'$-ample condition $\textbf{0}=(\cdots,0,0,0)\in\Sigma'_{\iota}[\lambda]$
(Definition \ref{ample1}). Using Theorem \ref{Nthm1}, we get our claim. \qed
\subsection{Type ${\rm A}_{2n-2}^{(2)}$-case and ${\rm C}_{n-1}^{(1)}$-case}
In this subsection, let $\mathfrak{g}$ be of type $X={\rm A}_{2n-2}^{(2)}$ or
$X={\rm C}_{n-1}^{(1)}$ and prove Theorem \ref{thmA2}.
First, we consider the first set
$\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1}x_{j_0} | \ell\in\mathbb{Z}_{\geq0}, j_0,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}$
of right hand side of (\ref{xilamdef}).
By Theorem 4.8, 4.11 of \cite{Ka}, the sequence $\iota$ satisfies the $\Xi'$-positivity condition
and it follows from (\ref{twoS2}) and the proofs of Theorem 4.8, 4.11 of \cite{Ka} that
\begin{eqnarray}
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1}x_{j_0} | \ell\in\mathbb{Z}_{\geq0}, j_0,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
&=&\{ S'_{j_{\ell}}\cdots S'_{j_1}x_{j_0} | \ell\in\mathbb{Z}_{\geq0}, j_0,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}\nonumber\\
&=&\{ S'_{j_{\ell}}\cdots S'_{j_1}x_{s,k} | \ell\in\mathbb{Z}_{\geq0}, j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1},\ k\in I,\ s\in\mathbb{Z}_{\geq1} \}
\nonumber\\
&=& \{ L^{X^L}_{s,k,\iota}(T)| k\in I_{X^L},\ s\in\mathbb{Z}_{\geq1}, T\in {\rm REYD}^{X^L}_{k} \}\nonumber\\
& &\cup
\{ L^{X^L}_{s,k,\iota}(T)| k\in I\setminus I_{X^L},\ s\in\mathbb{Z}_{\geq1}, T\in {\rm YW}^{X^L}_{k} \}\label{half2}\\
&=&{\rm Comb}_{\iota}^{X^L}[\infty].\nonumber
\end{eqnarray}
Next, we consider the second set
$\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ k\in I,\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}$
of (\ref{xilamdef}). We fix an index $k\in I$. In this proof, we simply write $P_k^{X^L}$ and $\pi_{X^L}$ as
$P_k$ and $\pi$. We take $k$ as $k\in I_{X^L}$ and
for $j=k-1$, $k+1$,
\[
c_{j}:=
\begin{cases}
2 & \text{if }\pi(j)\in I\setminus I_X,\\
1 & \text{otherwise}.
\end{cases}
\]
\vspace{2mm}
\nd
\underline{Case 1 : it holds $(1,k)<(1,\pi(k-1)), (1,\pi(k+1))$}
\vspace{2mm}
By $\lambda^{(k)}=-x_{1,k}+\langle h_k,\lambda \rangle$, it is easy to see
$\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
=\{0,\lambda^{(k)}\}$.
\vspace{2mm}
\nd
\underline{Case 2 : it holds $(1,k)>(1,\pi(k-1))$ and $(1,k)<(1,\pi(k+1))$}
\vspace{2mm}
The definition of $\lambda^{(k)}$ implies
$\lambda^{(k)}=-x_{1,k}+c_{k-1}x_{1,\pi(k-1)}+\langle h_k,\lambda \rangle$.
Using the notation in (\ref{r-tilde-D1}), it holds $\lambda^{(k)}=\fbox{$\tilde{k}$}^{{\rm X}^{L}}_k+\langle h_k,\lambda \rangle$.
By the definition of $\what{S}'$,
the actions of $\what{S}'$ are as follows:
\[
\begin{xy}
(-8,0) *{-\langle h_k,\lambda \rangle}="00",
(0,0) *{\ }="0",
(10,0) *{\fbox{$\tilde{k}$}^{X^L}}="r-1r-1",
(40,0)*{\fbox{$\overset{\sim}{k-1}$}^{X^L}}="r-2r-2",
(70,0)*{\fbox{$\overset{\sim}{k-2}$}^{X^L}}="r-3r-3",
(100,0)*{\fbox{$\overset{\sim}{k-3}$}^{X^L}}="r-4r-4",
(130,0)*{\cdots}="dot",
\ar@{->} "r-1r-1";"00"^{\what{S}'_{1,k}}
\ar@/^/ @{->} "r-1r-1";"r-2r-2"^{\what{S}'_{P_k(k-1),\pi(k-1)}}
\ar@/^/ @{->} "r-2r-2";"r-3r-3"^{\what{S}'_{P_k(k-2),\pi(k-2)}}
\ar@/^/ @{->} "r-3r-3";"r-4r-4"^{\what{S}'_{P_k(k-3),\pi(k-3)}}
\ar@/^/ @{->} "r-4r-4";"dot"^{\what{S}'_{P_k(k-4),\pi(k-4)}}
\ar@/_/ @{<-} "r-1r-1";"r-2r-2"_{\what{S}'_{1+P_k(k-1),\pi(k-1)}}
\ar@/_/ @{<-} "r-2r-2";"r-3r-3"_{\what{S}'_{1+P_k(k-2),\pi(k-2)}}
\ar@/_/ @{<-} "r-3r-3";"r-4r-4"_{\what{S}'_{1+P_k(k-3),\pi(k-3)}}
\ar@/_/ @{<-} "r-4r-4";"dot"_{\what{S}'_{1+P_k(k-4),\pi(k-4)}}
\end{xy}
\]
\[
\begin{xy}
(-5,0) *{}="00",
(0,0) *{\ }="0",
(10,0) *{\cdots}="r-1r-1",
(40,0)*{\fbox{$\tilde{2}$}^{X^L}}="r-2r-2",
(70,0)*{\fbox{$\tilde{1}$}^{X^L}}="r-3r-3",
(100,0)*{\fbox{$\tilde{0}$}^{X^L}}="r-4r-4",
(130,0)*{\cdots}="dot",
\ar@/^/ @{->} "r-1r-1";"r-2r-2"^{\what{S}'_{P_k(2),\pi(2)}}
\ar@/^/ @{->} "r-2r-2";"r-3r-3"^{\what{S}'_{P_k(1),\pi(1)}}
\ar@/^/ @{->} "r-3r-3";"r-4r-4"^{\what{S}'_{P_k(0),\pi(0)}}
\ar@/^/ @{->} "r-4r-4";"dot"^{\what{S}'_{P_k(-1),\pi(-1)}}
\ar@/_/ @{<-} "r-1r-1";"r-2r-2"_{\what{S}'_{1+P_k(2),\pi(2)}}
\ar@/_/ @{<-} "r-2r-2";"r-3r-3"_{\what{S}'_{1+P_k(1),\pi(1)}}
\ar@/_/ @{<-} "r-3r-3";"r-4r-4"_{\what{S}'_{1+P_k(0),\pi(0)}}
\ar@/_/ @{<-} "r-4r-4";"dot"_{\what{S}'_{1+P_k(-1),\pi(-1)}}
\end{xy}
\]
Here, we omitted index $k$.
Other actions of $\what{S}'$ are trivial.
Note that by $p_{\pi(k-1),\pi(k)}=1$, if $r<k$ then
\begin{equation}\label{stpspr4}
1+P_k(r)\geq2,
\end{equation}
which is the left index of $-x_{1+P_k(r),\pi(r)}$ in (\ref{r-tilde-D1}).
Thus,
\[
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
=\{0,\fbox{$\tilde{r}$}^{X^L}_k+\langle h_k,\lambda \rangle\ | r\in\mathbb{Z}_{\leq k}\}.
\]
\vspace{2mm}
\nd
\underline{Case 3 : it holds $(1,k)>(1,\pi(k+1))$ and $(1,k)<(1,\pi(k-1))$}
\vspace{2mm}
It is easy to see
$\lambda^{(k)}=-x_{1,k}+c_{k+1}x_{1,\pi(k+1)}+\langle h_k,\lambda \rangle$ so that
$\lambda^{(k)}=\fbox{$k+1$}^{{\rm X}^{L}}_k+\langle h_k,\lambda \rangle$ by using the notation in (\ref{r-tilde-D0}).
By the definition of $\what{S}'$,
the actions of $\what{S}'$ are as follows:
\[
\begin{xy}
(-8,0) *{-\langle h_k,\lambda \rangle}="00",
(0,0) *{\ }="0",
(10,0) *{\fbox{$k+1$}^{X^L}}="r-1r-1",
(40,0)*{\fbox{$k+2$}^{X^L}}="r-2r-2",
(70,0)*{\fbox{$k+3$}^{X^L}}="r-3r-3",
(100,0)*{\fbox{$k+4$}^{X^L}}="r-4r-4",
(130,0)*{\cdots}="dot",
\ar@{->} "r-1r-1";"00"^{\what{S}'_{1,k}}
\ar@/^/ @{->} "r-1r-1";"r-2r-2"^{\what{S}'_{P_k(k+1),\pi(k+1)}}
\ar@/^/ @{->} "r-2r-2";"r-3r-3"^{\what{S}'_{P_k(k+2),\pi(k+2)}}
\ar@/^/ @{->} "r-3r-3";"r-4r-4"^{\what{S}'_{P_k(k+3),\pi(k+3)}}
\ar@/^/ @{->} "r-4r-4";"dot"^{\what{S}'_{P_k(k+4),\pi(k+4)}}
\ar@/_/ @{<-} "r-1r-1";"r-2r-2"_{\what{S}'_{1+P_k(k+1),\pi(k+1)}}
\ar@/_/ @{<-} "r-2r-2";"r-3r-3"_{\what{S}'_{1+P_k(k+2),\pi(k+2)}}
\ar@/_/ @{<-} "r-3r-3";"r-4r-4"_{\what{S}'_{1+P_k(k+3),\pi(k+3)}}
\ar@/_/ @{<-} "r-4r-4";"dot"_{\what{S}'_{1+P_k(k+4),\pi(k+4)}}
\end{xy}
\]
\[
\begin{xy}
(-5,0) *{}="00",
(0,0) *{\ }="0",
(10,0) *{\cdots}="r-1r-1",
(40,0)*{\fbox{$n-1$}^{X^L}}="r-2r-2",
(70,0)*{\fbox{$n$}^{X^L}}="r-3r-3",
(100,0)*{\fbox{$n+1$}^{X^L}}="r-4r-4",
(130,0)*{\cdots}="dot",
\ar@/^/ @{->} "r-1r-1";"r-2r-2"^{\what{S}'_{P_k(n-2),\pi(n-2)}}
\ar@/^/ @{->} "r-2r-2";"r-3r-3"^{\what{S}'_{P_k(n-1),\pi(n-1)}}
\ar@/^/ @{->} "r-3r-3";"r-4r-4"^{\what{S}'_{P_k(n),\pi(n)}}
\ar@/^/ @{->} "r-4r-4";"dot"^{\what{S}'_{P_k(n+1),\pi(n+1)}}
\ar@/_/ @{<-} "r-1r-1";"r-2r-2"_{\what{S}'_{1+P_k(n-2),\pi(n-2)}}
\ar@/_/ @{<-} "r-2r-2";"r-3r-3"_{\what{S}'_{1+P_k(n-1),\pi(n-1)}}
\ar@/_/ @{<-} "r-3r-3";"r-4r-4"_{\what{S}'_{1+P_k(n),\pi(n)}}
\ar@/_/ @{<-} "r-4r-4";"dot"_{\what{S}'_{1+P_k(n+1),\pi(n+1)}}
\end{xy}
\]
Here, we omitted index $k$.
Other actions of $\what{S}'$ are trivial.
Note that for $r>k+1$, it holds
\begin{equation}\label{stpspr5}
1+P_k(r-1)\geq2,
\end{equation}
which is the left index of $-x_{1+P_k(r-1),\pi(r-1)}$ in (\ref{r-tilde-D0}).
Thus,
\[
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
=\{0,\fbox{$r$}^{X^L}_k+\langle h_k,\lambda \rangle\ | r\in\mathbb{Z}_{\geq k+1}\}.
\]
\vspace{2mm}
\nd
\underline{Case 4 : it holds $(1,k)>(1,\pi(k+1))$, $(1,\pi(k-1))$}
\vspace{2mm}
In this case, one obtains $\lambda^{(k)}=-x_{1,k}+c_{k-1}x_{1,\pi(k-1)}+c_{k+1}x_{1,\pi(k+1)}+
\lan h_k,\lambda \ran$.
First, let us prove
\begin{equation}\label{D-pr-1}
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
\subset
\{L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle | T\in{\rm REYD}^{X^L}_{k}\setminus\{\phi^k\}\}\cup\{0\}.
\end{equation}
We put the right hand side of (\ref{D-pr-1}) as $Z$:
$Z:=\{L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle | T\in{\rm REYD}^{X^L}_{k}\setminus\{\phi^k\}\}\cup\{0\}$.
Let
$T^1_k:=(y_{\ell})_{\ell\in\mathbb{Z}}\in{\rm REYD}^{X}_{k}\setminus\{\phi^k\}$ be the following:
\[
\begin{xy}
(-33,0) *{}="-6",
(-6,2) *{-1}="-1",
(-12,2) *{-2}="-2",
(-18,2) *{-3}="-3",
(-24,2) *{-4}="-4",
(-30,2) *{-5}="-5",
(-6,-1) *{}="-1a",
(-12,-1) *{}="-2a",
(-18,-1) *{}="-3a",
(-24,-1) *{}="-4a",
(-30,-1) *{}="-5a",
(0,0) *{}="1",
(50,0)*{}="2",
(0,-40)*{}="3",
(0,2)*{(0,k)}="4",
(6,2) *{1}="10",
(6,-1) *{}="1010",
(12,2) *{2}="11",
(12,-1) *{}="1111",
(18,2) *{3}="12",
(18,-1) *{}="1212",
(24,2) *{4}="13",
(24,-1) *{}="1313",
(30,2) *{5}="14",
(30,-1) *{}="1414",
(6,-8)*{k-1}="k-1",
(6,-12)*{k-2}="k-2",
(6,-18)*{k-3}="k-3",
(6,-24)*{k-4}="k-4",
(6,-30)*{k-5}="k-5",
(-1,-6)*{}="5",
(1,-6)*{}="55",
(6,-6)*{}="555",
(-1,-12)*{}="6",
(-6,-12)*{}="6a",
(-12,-12)*{}="6aa",
(-12,-18)*{}="6aaa",
(-18,-18)*{}="6aaaa",
(-18,-24)*{}="st1",
(-24,-24)*{}="st2",
(-24,-30)*{}="st3",
(-30,-30)*{}="st4",
(-33,-33)*{\cdots}="stdot",
(1,-12)*{}="66",
(-1,-18)*{}="t",
(1,-18)*{}="tt",
(-6,-6)*{}="7",
(1,-18)*{}="77",
(-1,-24)*{}="8",
(1,-24)*{}="88",
(-1,-30)*{}="9",
(1,-30)*{}="99",
\ar@{-} "1";"-6"^{}
\ar@{-} "1";"2"^{}
\ar@{-} "1";"3"^{}
\ar@{-} "-1";"-1a"^{}
\ar@{-} "-2";"-2a"^{}
\ar@{-} "-3";"-3a"^{}
\ar@{-} "-4";"-4a"^{}
\ar@{-} "-5";"-5a"^{}
\ar@{-} "5";"55"^{}
\ar@{-} "55";"555"^{}
\ar@{-} "1010";"555"^{}
\ar@{-} "6";"66"^{}
\ar@{-} "t";"tt"^{}
\ar@{-} "8";"88"^{}
\ar@{-} "7";"5"^{}
\ar@{-} "7";"6a"^{}
\ar@{-} "6aa";"6a"^{}
\ar@{-} "6aaa";"6aa"^{}
\ar@{-} "6aaaa";"6aaa"^{}
\ar@{-} "st1";"6aaaa"^{}
\ar@{-} "st1";"st2"^{}
\ar@{-} "st2";"st3"^{}
\ar@{-} "st3";"st4"^{}
\ar@{-} "9";"99"^{}
\ar@{-} "10";"1010"^{}
\ar@{-} "11";"1111"^{}
\ar@{-} "12";"1212"^{}
\ar@{-} "13";"1313"^{}
\ar@{-} "14";"1414"^{}
\end{xy}
\]
The point $(1,k-1)$ is a single $k$-removable point in $T^1_k$.
We see that the point $(1,k)$ is a double (resp. single) $\pi(k+1)$-admissible point
if $c_{k+1}=2$ (resp. $c_{k+1}=1$). Similarly,
the point $(-1,k-1)$ is a double (resp. single) $\pi(k-1)$-admissible point
if $c_{k-1}=2$ (resp. $c_{k-1}=1$).
Hence, it holds $L_{0,k,\iota}^{X^L}(T^1_k)+\lan h_k,\lambda \ran = \lambda^{(k)}$ so that
$\lambda^{(k)}\in Z$.
To show (\ref{D-pr-1}), let us prove $Z$
is closed under the action of $\what{S}'_r$ for all $r\in\mathbb{Z}_{\geq1}$.
For any $T=(y_r)_{r\in\mathbb{Z}}\in {\rm REYD}^{X}_{k}\setminus\{\phi^k\}$, if the coefficient of $x_r$ is positive in
$L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle$ then
$T$ has an admissible point $(i,j)$ such that $j=y_i$ and
\[
r=(P_k(i+k)+[i]_-+k-j,\pi(i+k))
\]
by the definition of $L^{X^L}_{0,k,\iota}$ in (\ref{L1kdef}).
Let
$T'=(y_t')_{t\in\mathbb{Z}}\in{\rm REYD}^{X}_{k}$ be the sequence defined as
$y_{i}'=y_i-1$ and $y_t'=y_t$ $(t\neq i)$.
It follows from Proposition \ref{A2closed}, \ref{D2closed} that
\[
\what{S}'_{r}(L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle)
=L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle-\beta_{r}=
L^{X^L}_{0,k,\iota}(T')+\langle h_k,\lambda \rangle\in Z.
\]
If the coefficient of $x_r$ is negative in
$L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle$ then
$T$ has a removable point $(i,j)$ such that $j=y_{i-1}$ and
\[
r=(P_k(i+k-1)+[i-1]_-+k-j,\pi(i+k-1)).
\]
Note that $k-j\geq1$ since $(i,j)$ is removable.
If it holds $T=T^1_k$ then
$L^{X^L}_{0,k,\iota}(T)=-x_{1,k}+c_{k-1}x_{1,\pi(k-1)}+c_{k+1}x_{1,\pi(k+1)}$ and
$r=(1,k)$. Thus,
\[
\what{S}'_r(L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle)=0.
\]
Hence, we may assume that $T\neq T^1_k$.
In case of $i-1>0$, since $p_{\pi(k+1),\pi(k)}=1$, it holds $P_k(i+k-1)\geq1$ so that
$P_k(i+k-1)+[i-1]_-+k-j=P_k(i+k-1)+k-j\geq2$.
In case of $i-1=0$, if $k-j=1$ then $T=T^1_k$, which is absurd. If $k-j\geq2$ then
$P_k(i+k-1)+[i-1]_-+k-j=P_k(i+k-1)+k-j\geq2$.
In case of $i-1<0$, it holds $k-j\geq 2-i$
and $P_k(i+k-1)\geq p_{\pi(k-1),\pi(k)}=1$
so that
$P_k(i+k-1)+[i-1]_-+k-j=P_k(i+k-1)+i-1+k-j\geq2$.
Hence,
\begin{equation}\label{D-pr-1a}
P_k(i+k-1)+[i-1]_-+k-j\geq2.
\end{equation}
Let
$T''=(y_t'')_{t\in\mathbb{Z}}\in{\rm REYD}^{X^L}_{k}$
be the sequence such that $y_{i-1}''=y_{i-1}+1$ and $y_t''=y_t$ $(t\neq i-1)$.
It follows by Proposition \ref{A2closed}, \ref{D2closed} that
\begin{multline*}
\what{S}'_{r}(L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle)
=L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle+\beta_{r^-}\\
=
L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle+\beta_{P_k(i+k-1)+[i-1]_-+k-j-1,\pi(i+k-1)}=
L^{X^L}_{0,k,\iota}(T'')+\langle h_k,\lambda \rangle\in Z.
\end{multline*}
Therefore, one obtains $(\ref{D-pr-1})$.
Next, we prove
\begin{equation}\label{D-pr-2}
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
\supset Z.
\end{equation}
Since $0\in\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}$
is clear, we show that for any
$T=(y_{\ell})_{\ell\in\mathbb{Z}}\in {\rm REYD}^{X^L}_{k}\setminus\{\phi^k\}$, it holds
\begin{equation}\label{D-pr-3}
L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle \in
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1}\}.
\end{equation}
Note that each element in ${\rm REYD}^{X^L}_{k}$ can be obtained from $\phi^k$
by putting several boxes whose length of sides are $1$.
The proof of (\ref{D-pr-3}) proceeds by induction on the number of boxes of $T$.
When the number of boxes is $1$, it holds $T=T^1_k$ and
$L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle
=\lambda^{(k)}\in\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1}\}$.
Hence, we may assume
that the number of boxes of $T$ is at least $2$.
Putting $m:={\rm min}\{\ell\in\mathbb{Z}|y_{\ell}<k+\ell\}$, one obtains $m\leq 1$ by the definition of ${\rm REYD}^{X^L}_{k}$.
The equality $m=1$ means $y_{\ell}=k+\ell$ for all $\ell\in\mathbb{Z}_{\leq0}$, in particular, $y_0=k$,
which implies $y_r=k$ for any $r\in\mathbb{Z}_{\geq0}$ by $y_0\leq y_r\leq k$. Hence $T=\phi^k$, which contradicts
$T\neq \phi^k$ so that $m\leq0$ and $y_m<k$.
Putting
$m_1:={\rm min}\{y_{\ell} | m\leq \ell\}$,
we obtain $m_1<k$ and there exists $i\in\mathbb{Z}_{\geq m}$ such that $y_i=m_1$ and $y_{i+1}>y_i$.
Note that the point $(i+1,y_i)$ is removable.
If $i=0$ and $y_0=k-1$ then $T=T^1_k$, which is absurd.
Thus, if $i=0$ then $k-1>y_0$.
In the case $i>0$, it follows from $p_{\pi(k+1),\pi(k)}=1$ that
$P_k(i+k)+[i]_-+k-y_i-1 \geq p_{\pi(k+1),\pi(k)}=1$.
In the case $i=0$, it holds $k-1>y_0$ and
$P_k(i+k)+[i]_-+k-y_i-1 =k-y_0-1\geq1$.
In the case $i<0$, using $p_{\pi(k-1),\pi(k)}=1$ and $y_i\leq k+i-1$, we obtain
$P_k(i+k)+[i]_-+k-y_i-1 \geq1$. Thus,
\begin{equation}\label{D-pr-4}
P_k(i+k)+[i]_- + k-y_i-1 \geq1.
\end{equation}
We define
$T''=(y_{\ell}'')_{\ell\in\mathbb{Z}}\in{\rm REYD}^{X^L}_{k}$
as $y_{i}''=y_{i}+1$ and $y_{\ell}''=y_{\ell}$ $(\ell\neq i)$.
Considering Proposition \ref{A2closed}, \ref{D2closed}, $L^{X^L}_{0,k,\iota}(T)$ is in the form
\[
L^{X^L}_{0,k,\iota}(T)
=L^{X^L}_{0,k,\iota}(T'')-\beta_{t_1,d_1}
\]
with $(t_1,d_1)=(P_k(i+k)+[i]_- + k-y_i-1,\pi(i+k))\in \mathbb{Z}_{\geq1}\times I$.
Repeating this argument,
\[
L^{X^L}_{0,k,\iota}(T)+\lan h_k, \lambda \ran
=L^{X^L}_{0,k,\iota}(T_k^1)-\sum_{(t,d)\in\mathbb{Z}_{\geq1}\times I} c_{t,d}\beta_{t,d}+\lan h_k, \lambda \ran
=\lambda^{(k)}-\sum_{(t,d)\in\mathbb{Z}_{\geq1}\times I} c_{t,d}\beta_{t,d}
\]
with non-negative integers $\{c_{t,d}\}$ such that $c_{t,d}=0$
except for finitely many $(t,d)\in\mathbb{Z}_{\geq1}\times I$.
Taking the assumption $T\neq T^1_k$ into account,
there exists $(t',d')\in\mathbb{Z}_{\geq1}\times I$ such that
\[
(t',d')
={\rm max} \{(t,d)\in\mathbb{Z}_{\geq1}\times I | c_{t,d}>0\}.
\]
By the definition of $\beta_{t,d}$,
the coefficient of $x_{t'+1,d'}$ is negative in
$L^{X^L}_{0,k,\iota}(T)$. We set
\[
(t'',d''):=
{\rm min}
\{(t,d)\in\mathbb{Z}_{\geq1}\times I | \text{the coefficient of }x_{t,d}\text{ in }L^{X^L}_{0,k,\iota}(T)\text{ is negative}\}.
\]
The definition of $L^{X^L}_{0,k,\iota}$ says
there exists a removable point $(\xi,y_{\xi-1})$ in $T$
such that $L^{X^L}_{0,k,\text{re}}(\xi,y_{\xi-1})= x_{t'',d''}$
and $t''=P_k(\xi+k-1)+[\xi-1]_-+k-y_{\xi-1}$.
Just as in (\ref{D-pr-1a}), it holds
$t''\geq2$.
Let
$T''=(y_t'')_{t\in\mathbb{Z}}\in {\rm REYD}^{X^L}_{k}$
be the sequence such that $y_{\xi-1}''=y_{\xi-1}+1$ and $y_t''=y_t$ $(t\neq \xi-1)$.
Using Proposition \ref{A2closed} and \ref{D2closed}, we have
\[
L^{X^L}_{0,k,\iota}(T)
=L^{X^L}_{0,k,\iota}(T'')-\beta_{t''-1,d''}.
\]
Considering the minimality of $(t'',d'')$ and above equality,
we see that the coefficient of $x_{t''-1,d''}$
in $L^{X^L}_{0,k,\iota}(T'')$ is positive. Thus,
\begin{equation}\label{D-pr-5}
L^{X^L}_{0,k,\iota}(T)
=\what{S}'_{t''-1,d''}L^{X^L}_{0,k,\iota}(T'').
\end{equation}
Since the number of boxes in $T''$ is less than that of $T$,
the induction assumption yields
\[
L^{X^L}_{0,k,\iota}(T'')+\langle h_k,\lambda \rangle
\in
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1}\}.
\]
In conjunction with (\ref{D-pr-5}), we get
$L^{X^L}_{0,k,\iota}(T)+\langle h_k,\lambda \rangle
\in
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1}\}$.
Finally, we take $k\in I\setminus I_{X^L}$.
\vspace{2mm}
\nd
\underline{Case 1 : it holds $(1,k)<(1,\pi'(k+1))$}
\vspace{2mm}
It holds $\lambda^{(k)}=-x_{1,k}+\langle h_k,\lambda \rangle$ and one can easily check
$\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
=\{0,\lambda^{(k)}\}$.
\vspace{2mm}
\nd
\underline{Case 2 : it holds $(1,k)>(1,\pi'(k+1))$}
\vspace{2mm}
We have $\lambda^{(k)}=-x_{1,k}+x_{1,\pi'(k+1)}+\langle h_k,\lambda \rangle$.
First, let us prove
\begin{equation}\label{D-pr-6}
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
\subset
\{L^{X^L}_{0,k,\iota}(Y)+\langle h_k,\lambda \rangle | Y\in{\rm YW}^{X^L}_{k}\setminus\{Y_{\Lambda_k}\}\}\cup\{0\}.
\end{equation}
We put the right hand side of (\ref{D-pr-6}) as $Z$:
$Z:=\{L^{X^L}_{0,k,\iota}(Y)+\langle h_k,\lambda \rangle | Y\in{\rm YW}^{X^L}_{k}\setminus\{Y_{\Lambda_k}\}\}\cup\{0\}$.
Let $Y^1_k\in {\rm YW}^{X^L}_{k}\setminus\{Y_{\Lambda_k}\}$ be the following:
\begin{equation*}
\begin{xy}
(-15.5,-2) *{\ k}="000-1",
(-9.5,-2) *{\ k}="00-1",
(-3.5,-2) *{\ k}="0-1",
(1.5,-2) *{\ \ k}="0-1",
(-21.5,-2) *{\dots}="00000",
(1.5,1.5) *{\ \ k}="0",
(12,-6) *{(0,k)}="origin",
(0,0) *{}="1",
(0,-3.5) *{}="1-u",
(6,0)*{}="2",
(6,-3.5)*{}="2-u",
(6,-7.5)*{}="-1-u",
(6,3.5)*{}="3",
(0,3.5)*{}="4",
(-6,3.5)*{}="5",
(-6,0)*{}="6",
(-6,-3.5)*{}="6-u",
(-12,3.5)*{}="7",
(-12,0)*{}="8",
(-12,-3.6)*{}="8-u",
(-18,3.5)*{}="9",
(-18,0)*{}="10-a",
(-18,0)*{}="10",
(-18,-3.5)*{}="10-u",
(-24,-3.5)*{}="11-u",
(20,-3.5)*{}="0-u",
(6,9.5)*{}="3-1",
(6,15.5)*{}="3-2",
(6,21.5)*{}="3-3",
(6,25)*{}="3-4",
(0,9.5)*{}="4-1",
(0,15.5)*{}="4-2",
(0,21.5)*{}="4-3",
(-6,9.5)*{}="5-1",
\ar@{-} "-1-u";"2-u"^{}
\ar@{->} "2-u";"0-u"^{}
\ar@{-} "10-u";"11-u"^{}
\ar@{-} "8";"10-a"^{}
\ar@{-} "10-u";"10-a"^{}
\ar@{-} "10-u";"2-u"^{}
\ar@{-} "8";"8-u"^{}
\ar@{-} "6";"6-u"^{}
\ar@{-} "2";"2-u"^{}
\ar@{-} "1";"1-u"^{}
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\ar@{-} "1";"6"^{}
\ar@{-} "6";"8"^{}
\ar@{-} "3";"3-1"^{}
\ar@{-} "3-2";"3-1"^{}
\ar@{-} "3-2";"3-3"^{}
\ar@{->} "3-3";"3-4"^{}
\end{xy}
\end{equation*}
Since $Y^1_k$ has a single removable $1$-block and
a single $1$-admissible slot, we have
$Z\ni L^{X^L}_{0,k,\iota}(Y^1_k)+\langle h_k,\lambda \rangle=\lambda^{(k)}$.
We need to show $Z$ is closed under the action of $\what{S}'_r$ for all $r\in\mathbb{Z}_{\geq1}$.
Let $Y\in{\rm YW}^{X^L}_{k}\setminus\{Y_{\Lambda_k}\}$.
If the coefficient of $x_r$ is positive in $L^{X^L}_{0,k,\iota}(Y)$
then
the definition of $L^{X^L}_{0,k,\iota}$ in (\ref{L11-def}) says
there exists a $t$-admissible slot $S$ in $Y$ in the form
\[
\begin{xy}
(-8,15) *{(-i-1,\ell+1)}="0000",
(17,15) *{(-i,\ell+1)}="000",
(15,-3) *{(-i,\ell)}="00",
(-7,-3) *{(-i-1,\ell)}="0",
(0,0) *{}="1",
(12,0)*{}="2",
(12,12)*{}="3",
(0,12)*{}="4",
\ar@{--} "1";"2"^{}
\ar@{--} "1";"4"^{}
\ar@{--} "2";"3"^{}
\ar@{--} "3";"4"^{}
\end{xy}
\]
or
\[
\begin{xy}
(-8,9) *{(-i-1,\ell+\frac{1}{2})}="0000",
(17,9) *{(-i,\ell+\frac{1}{2})}="000",
(15,-3) *{(-i,\ell)}="00",
(-7,-3) *{(-i-1,\ell)}="0",
(0,0) *{}="1",
(12,0)*{}="2",
(12,6)*{}="3",
(0,6)*{}="4",
(38,3)*{{\rm or}}="or",
(60,9) *{(-i-1,\ell+1)}="a0000",
(85,9) *{(-i,\ell+1)}="a000",
(83,-3) *{(-i,\ell+\frac{1}{2})}="a00",
(61,-3) *{(-i-1,\ell+\frac{1}{2})}="a0",
(68,0) *{}="a1",
(80,0)*{}="a2",
(80,6)*{}="a3",
(68,6)*{}="a4",
\ar@{--} "1";"2"^{}
\ar@{--} "1";"4"^{}
\ar@{--} "2";"3"^{}
\ar@{--} "3";"4"^{}
\ar@{--} "a1";"a2"^{}
\ar@{--} "a1";"a4"^{}
\ar@{--} "a2";"a3"^{}
\ar@{--} "a3";"a4"^{}
\end{xy}
\]
such that $L^{X^L}_{0,k,{\rm ad}}(S)=x_{r}$. Hence,
$r=(P_k(\ell)+i,t)$.
Since $Y\neq Y_{\Lambda_k}$, at least one of $\ell\geq k+1$ and $i\geq1$ holds, which
implies $P_k(\ell)+i\geq1$ by $p_{\pi'(k+1),k}=1$. Let $Y'\in {\rm YW}^{X^L}_{k}\setminus\{Y_{\Lambda_k}\}$
be the proper Young wall obtained from $Y$ by adding a $t$-block to the slot $S$.
It follows from Proposition \ref{prop-closednessAw-YW} and \ref{prop-closednessCw-YW}
that
\[
\what{S}'_r(L^{X^L}_{0,k,\iota}(Y)+\langle h_k,\lambda \rangle)=
L^{X^L}_{0,k,\iota}(Y)+\langle h_k,\lambda \rangle-\beta_{r}
=
L^{X^L}_{0,k,\iota}(Y)+\langle h_k,\lambda \rangle-\beta_{P_k(\ell)+i,t}
=
L^{X^L}_{0,k,\iota}(Y')+\langle h_k,\lambda \rangle\in Z.
\]
If the coefficient of $x_r$ is negative in $L^{X^L}_{0,k,\iota}(Y)$
then
there exists a removable $t$-block $B$ in $Y$ in the form
\[
\begin{xy}
(-8,15) *{(-i-1,\ell+1)}="0000",
(17,15) *{(-i,\ell+1)}="000",
(15,-3) *{(-i,\ell)}="00",
(-7,-3) *{(-i-1,\ell)}="0",
(0,0) *{}="1",
(12,0)*{}="2",
(12,12)*{}="3",
(0,12)*{}="4",
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\end{xy}
\]
or
\[
\begin{xy}
(-8,9) *{(-i-1,\ell+\frac{1}{2})}="0000",
(17,9) *{(-i,\ell+\frac{1}{2})}="000",
(15,-3) *{(-i,\ell)}="00",
(-7,-3) *{(-i-1,\ell)}="0",
(0,0) *{}="1",
(12,0)*{}="2",
(12,6)*{}="3",
(0,6)*{}="4",
(38,3)*{{\rm or}}="or",
(60,9) *{(-i-1,\ell+1)}="a0000",
(85,9) *{(-i,\ell+1)}="a000",
(83,-3) *{(-i,\ell+\frac{1}{2})}="a00",
(61,-3) *{(-i-1,\ell+\frac{1}{2})}="a0",
(68,0) *{}="a1",
(80,0)*{}="a2",
(80,6)*{}="a3",
(68,6)*{}="a4",
\ar@{-} "1";"2"^{}
\ar@{-} "1";"4"^{}
\ar@{-} "2";"3"^{}
\ar@{-} "3";"4"^{}
\ar@{-} "a1";"a2"^{}
\ar@{-} "a1";"a4"^{}
\ar@{-} "a2";"a3"^{}
\ar@{-} "a3";"a4"^{}
\end{xy}
\]
such that $L^{X^L}_{0,k,{\rm re}}(S)=x_{r}$ and
$r=(P_k(\ell)+i+1,t)$.
In the case $P_k(\ell)+i+1=1$, it holds $P_k(\ell)=i=0$ so that $\ell=k$,
which imply $t=k$ and $B$ equals
\[
\begin{xy}
(60,9) *{(-1,1)}="a0000",
(85,9) *{(0,1)}="a000",
(83,-3) *{(0,\frac{1}{2})}="a00",
(61,-3) *{(-1,\frac{1}{2})}="a0",
(68,0) *{}="a1",
(80,0)*{}="a2",
(80,6)*{}="a3",
(68,6)*{}="a4",
\ar@{-} "a1";"a2"^{}
\ar@{-} "a1";"a4"^{}
\ar@{-} "a2";"a3"^{}
\ar@{-} "a3";"a4"^{}
\end{xy}
\]
and $Y=Y^1_k$. Thus, $\what{S}'_r(L^{X^L}_{0,k,\iota}(Y)+\langle h_k,\lambda \rangle)=\what{S}'_{1,k}(\lambda^{(k)})=0$.
Hence, we may assume that
\begin{equation}\label{D-pr-a}
P_k(\ell)+i+1>1.
\end{equation}
Let $Y''\in {\rm YW}^{X^L}_{k}\setminus\{Y_{\Lambda_k}\}$
be the proper Young wall obtained from $Y$ by removing the $t$-block $B$.
It follows from Proposition \ref{prop-closednessAw-YW} and \ref{prop-closednessCw-YW}
that
\[
\what{S}'_r(L^{X^L}_{0,k,\iota}(Y)+\langle h_k,\lambda \rangle)=
L^{X^L}_{0,k,\iota}(Y)+\langle h_k,\lambda \rangle+\beta_{r^-}
=
L^{X^L}_{0,k,\iota}(Y)+\langle h_k,\lambda \rangle+\beta_{P_k(\ell)+i,t}
=
L^{X^L}_{0,k,\iota}(Y'')+\langle h_k,\lambda \rangle\in Z.
\]
If the coefficient of $x_r$ is zero then
$\what{S}'_r(L^{X^L}_{0,k,\iota}(Y)+\langle h_k,\lambda \rangle)=
L^{X^L}_{0,k,\iota}(Y)+\langle h_k,\lambda \rangle\in Z$.
Therefore, $Z$ is closed under the action of $\what{S}'_r$ and
one gets
(\ref{D-pr-6}).
Next, let us show
\begin{equation}\label{D-pr-7}
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
\supset
Z=\{L^{X^L}_{0,k,\iota}(Y)+\langle h_k,\lambda \rangle | Y\in{\rm YW}^{X^L}_{k}\setminus\{Y_{\Lambda_k}\}\}\cup\{0\}.
\end{equation}
For $Y\in{\rm YW}^{X^L}_{k}$,
we say the number of blocks in $Y$ is $m$ if $Y$ is obtained from
$Y_{\Lambda_k}$ by adding $m$ blocks.
For $Y\in{\rm YW}^{X^L}_{k}\setminus\{Y_{\Lambda_k}\}$,
let us prove
$L^{X^L}_{0,k,\iota}(Y)+\langle h_k,\lambda \rangle\in \{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}$
by induction on the number of blocks in $Y$.
When the number is $1$, that is, $Y=Y^1_k$, it holds
$L^{X^L}_{0,k,\iota}(Y)+\langle h_k,\lambda \rangle=\lambda^{(k)}
\in \{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}$.
Hence, we may assume $Y\neq Y^1_k$.
Using Proposition \ref{prop-closednessAw-YW}, \ref{prop-closednessCw-YW},
one can describe $L^{X^L}_{0,k,\iota}(Y)$ as
\[
L^{X^L}_{0,k,\iota}(Y)+\langle h_k,\lambda \rangle=\lambda^{(k)}-\sum_{(t,d)\in\mathbb{Z}_{\geq 1}\times I} c_{t,d}\beta_{t,d}
\]
with non-negative integers $\{c_{t,d}\}$. Except for finitely many $(t,d)$, we have $c_{t,d}=0$.
As assumed $Y\neq Y^1_k$, one can take $(t',d')\in\mathbb{Z}_{\geq 1}\times I$ as
$(t',d')
={\rm max}\{(t,d)\in\mathbb{Z}_{\geq 1}\times I |c_{t,d}>0 \}$.
Since we supposed $\iota$ is adapted,
the coefficient of $x_{t'+1,d'}$ is negative in
$L^{X^L}_{0,k,\iota}(Y)$.
Let us take $(t'',d'')$ as
\[
(t'',d'')={\rm min}\{(t,d)\in\mathbb{Z}_{\geq 1}\times I | \text{the coefficient of }x_{t,d}\text{ in }L^{X^L}_{0,k,\iota}(Y)\text{ is negative}\}.
\]
There is a removable block $B$ in $Y$ such that
$L^{X^L}_{0,k,{\rm re}}(B)=x_{t'',d''}$.
Let $Y'\in{\rm YW}^{X^L}_{k}$ be the proper Young wall
obtained from $Y$ by removing the block $B$. From $Y\neq Y^1_k$, one can check $t''\geq2$ by a
similar argument to (\ref{D-pr-a}).
By Proposition \ref{prop-closednessAw-YW} and \ref{prop-closednessCw-YW}, we have
\[
L^{X^L}_{0,k,\iota}(Y)=
L^{X^L}_{0,k,\iota}(Y')-\beta_{t''-1.d''}.
\]
Note that
the minimality of $(t'',d'')$ means the coefficient of $x_{t''-1,d''}$ is positive in
$L^{X^L}_{0,k,\iota}(Y')$ so that
\begin{equation}\label{D-pr-8}
\what{S}'_{t''-1,d''}(L^{X^L}_{0,k,\iota}(Y')+\lan h_k,\lambda\ran)
=L^{X^L}_{0,k,\iota}(Y')+\lan h_k,\lambda\ran-\beta_{t''-1,d''}=
L^{X^L}_{0,k,\iota}(Y)+\lan h_k,\lambda\ran.
\end{equation}
By induction assumption, one can write
$L^{X^L}_{0,k,\iota}(Y')+\lan h_k,\lambda\ran
=
\what{S}'_{j_{\xi}}\cdots \what{S}'_{j_1} \lambda^{(k)}
$
with some
$\xi\in\mathbb{Z}_{\geq0}$ and $j_1,\cdots,j_{\xi}\in \mathbb{Z}_{\geq1} $.
In conjunction with (\ref{D-pr-8}),
one gets $L^{X^L}_{0,k,\iota}(Y)+\lan h_k,\lambda\ran \in
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}$
and $(\ref{D-pr-7})$.
In this way, we get
\begin{equation}\label{stpspr6}
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} \lambda^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}
={\rm Comb}^{X^L}_k[\lambda]\cup\{0\}.
\end{equation}
Combining with (\ref{half2}), we see that the $\Xi'$-ample condition $\textbf{0}=(\cdots,0,0,0)\in\Sigma'_{\iota}[\lambda]$
(Definition \ref{ample1}) holds. Using Theorem \ref{Nthm1}, we get our claim. \qed
\subsection{Proof of Theorem \ref{thm3}}
In this subsection, we prove Theorem \ref{thm3}. First, let us show $\iota$ satisfies
$\Xi'$-strict positivity condition (\ref{strict-cond}).
Since we assumed $\iota$ is adapted, it follows from Theorem 4.3, 4.5, 4.8 and 4.11 in \cite{Ka} that
$\iota$ satisfies the $\Xi'$-positivity condition.
Therefore, it holds
\[
\text{if }\ell^{(-)}=0\text{ then }\varphi_{\ell}\geq0\text{ for any }
\vp=\sum_{\ell\geq1} \varphi_{\ell}x_{\ell}\in \Xi'_{\iota}.
\]
To show (\ref{strict-cond}), we need to prove
for any $k\in I$ and $\vp=\sum_{\ell\geq1} \varphi_{\ell}x_{\ell}\in \Xi'^{(k)}_{\iota}\setminus\{\xi^{(k)}\}$,
if $\ell^{(-)}=0$ then $\varphi_{\ell}\geq0$.
One can write $\vp=S'_{j_{r}}\cdots S'_{j_1}\xi^{(k)}$
with some
$r\in\mathbb{Z}_{\geq1}$ and $j_1,\cdots,j_{r}\in\mathbb{Z}_{\geq1}$.
We may assume that for any $t=1,2,\cdots,r$, it holds
$S'_{j_t}(S'_{j_{t-1}}\cdots S'_{j_1}\xi^{(k)})\neq S'_{j_{t-1}}\cdots S'_{j_1}\xi^{(k)}$.
By (\ref{twoS}), this assumption means
$\vp=S'_{j_{r}}\cdots S'_{j_1}\xi^{(k)}=\what{S}'_{j_r}\cdots \what{S}'_{j_1}\xi^{(k)}$.
Note that when $\lambda=0$, it holds $\lambda^{(k)}=\xi^{(k)}$.
Thus, the relation (\ref{stpspr3}) and (\ref{stpspr6}) yields $\vp\in {\rm Comb}_{k,\iota}^{X^L}[0]$.
Therefore, by the same argument as in (\ref{stpspr1}), (\ref{stpspr2}), (\ref{A-pr-1a}), (\ref{stpspr4}), (\ref{stpspr5}), (\ref{D-pr-1a}) and (\ref{D-pr-a}), one obtains if $\ell^{(-)}=0$ then $\varphi_{\ell}\geq0$. Hence,
$\Xi'$-strict positivity condition holds.
Therefore, as a consequence of (\ref{thm2-pr2}), Theorem \ref{thm2}, (\ref{stpspr3}) and (\ref{stpspr6}),
we see that
\begin{eqnarray*}
\varepsilon_k^*(x)
&=&{\rm max}\{-\varphi(x)|\varphi\in \Xi'^{(k)}_{\iota}\}\\
&=&
{\rm max}\{-\varphi(x)|\varphi\in
\{\what{S}'_{j_{\ell}}\cdots \what{S}'_{j_1} 0^{(k)} | \ell\in\mathbb{Z}_{\geq0},\ j_1,\cdots,j_{\ell}\in \mathbb{Z}_{\geq1} \}\setminus\{0\}\}\\
&=&
{\rm max}\{-\varphi(x)|\varphi\in {\rm Comb}_{k,\iota}^{X^L}[0]\}.
\end{eqnarray*}
\qed
|
\section{Introduction \label{sec:intro}}
Quantum nondemolition (QND) measurements~\cite{braginsky1980,grangier1998} are one of the established tools for measuring and engineering quantum mechanical systems.
Typically, a quantum measurement disturbs the quantum state of a system, due to the backaction of the measurement on the system. Furthermore, due to the free evolution of the Hamiltonian, repeated measurements do not leave the state invariant, as often considered in a textbook example of projective quantum measurements. In order to qualify as a QND measurement, the measurement observable and the free Hamiltonian must commute. In a sequence of measurements, the observables in the sequence must also commute. In this situation, QND measurements can be considered a type of measurement that least disturbs the system, and were investigated in the context of
precision measurement applications such as gravitation wave detection \cite{eberle2010,pitkin2011}, squeezed state preparation \cite{scully1997,loudon2000,byrnes2021} in optical \cite{brune1990,holland1991,ueda1992}, and mechanical oscillators \cite{lecocq2015}.
In atomic systems, QND measurements are well-known for realizing squeezed states~\cite{takahashi1999,kuzmich2000,higbie2005,meppelink2010,schleier-smith2010,vasilakis2015,moller2017,behbood2014generation,takano2009spin,bao2020spin}. The general principle is shown in Fig. \ref{fig1}. A superposition of clockwise and anti-clockwise circularly polarized light illuminates an atomic cloud. The dispersive coupling of the light with the atomic ground states produces a spin-dependent phase that is different for the two types of circularly polarized light. A waveplate interferes the two light modes such that the spin dependent phase becomes an observable, and a polarizing beam splitter and photodetectors measures the light to detect this. Due to the spin-dependence of the phase, this amounts to an indirect measurement of the spins of the atoms. Typically, the interaction between the atoms and the light is weak, such that there is a limited collapse of the atoms' wave function. Such a procedure produces a squeezed state in the spin variables due to the partial collapse of the wavefunction due to measurement. Besides squeezed state preparation, works focusing on non-Gaussian correlated quantum states preparation, such as supersinglets~\cite{cabello2003Review}, Schrodinger cat states, and N00N states, and preparation of correlated states between spatially separated atomic ensembles~\cite{julsgaard2001} have been proposed. These states are typically highly entangled states that play a central role in various quantum protocols such as entanglement purification~\cite{bennett1996b,bennett1996c}, quantum teleportation~\cite{bennett1996,horodecki1999,pyrkov2014}, remote state preparation~\cite{chaudhary2021}, clock synchronization~\cite{jozsa2000,ilo-okeke2018}, quantum computing~\cite{byrnes2012,abdelrahman2014}.
In this paper, we develop a theory of QND measurements in terms of a positive operator-value measure (POVM). POVMs are generalized measurement operators that do not necessarily follow the form of a projection operator. In our case, depending upon the strength of the interactions, the POVM can result in a weak or strong measurement. Our theory is valid for a wide range of regimes, and can be applied to any atom-light interaction time, any initial coherent state amplitude and any photonic measurement outcome. By writing the QND measurement as a POVM, it allows one to better grasp what is feasible with the technique, in a succinct mathematical form. We look at particular limits of the theory, such as the short interaction time limit where the effect of the measurement produces a Gaussian modulation of the amplitude of the initial state, written in terms of total spin $ J_z $ eigenstates. Another limit we examine is the projective limit where the state collapses to a single $J_z $ eigenstate. We also show that for long atom-light interaction times the measurements have some highly non-Gaussian effects, where rather exotic measurements can be performed, capable of realizing Schrodinger cat states.
This paper is organized as follows. In Sec.~\ref{sec:Dispersiveimaging} we introduce the formalism to define our measurement operator for QND measurements. In Sec. \ref{sec:ampfunc} we discuss further the amplitude function which plays a central role to modify the nature of the atomic wavefunction. In Sec.~\ref{sec:ProjectionMecurement} we show how a projective measurement can be obtained in a particular limit of the QND measurement. We show several examples of the measurement operator to produce various states in Sec. \ref{sec:examples}. Finally, we summarize and conclude in Sec. \ref{sec:Discussion}.
\begin{figure}[t]
\includegraphics[width=\columnwidth]{Fig1}
\caption{Setup for quantum nondemolition measurement of atomic gases considered in this paper. Atoms in the cloud interact with an incident laser beam consisting of two circularly polarized light modes detuned from atomic resonant transition. The spin dependent atom-light interactions cause parts of the laser beam to accumulate a phase relative to the other parts.
The atoms in the atomic cloud consist of ground states of spin-$j$, which produce a state that can be written in terms of the collective spin $ | J, m_z \rangle $ of the cloud. By passing the light beams through a wave plate (WP), the two light polarizations are interfered.
The polarizing beam splitter (PBS) and photodetectors detects the light in the linear polarization basis, producing photocounts $ n_c $ and $ n_d $. }
\label{fig1}
\end{figure}
\section{POVM for QND measurements}\label{sec:Dispersiveimaging}
We consider the experimental configuration as shown in Fig.~\ref{fig1}. Coherent light consisting of left and right circularly polarized light, which we denote by bosonic annihilation operators as $ a,b $, illuminate the atomic ensemble. The light is chosen at a frequency which is detuned from atomic resonance transition, producing a second order dispersive interaction. Within the rotating wave approximation and applying the adiabatic elimination, the effective Hamiltonian governing the evolution is described as~\cite{ilo-okeke2014}
\begin{align}
\hat{H} = \hbar g J_z S_z ,
\label{hamint}
\end{align}
where $g$ is the atom light coupling frequency, $J_z$ is the \emph{z}-component of the total spin operator of the atomic ensemble. The atomic degrees of freedom consist of \emph{N} particles of spin-\emph{j}, $J_\mathrm{max} = Nj$ is the maximum spin angular momentum, and $J$ takes on values $J = 0, \cdots, J_\mathrm{max}$. The eigenstates are denoted $ J_z | J, m_z \rangle = m_z | J, m_z \rangle$, where $ m_z = -J, -J+1, ... , J $. The spins typically refer to the hyperfine ground states of the atoms. The light modes are written using the Stokes operator $ S_z = (a^\dagger a - b^\dagger b)/2$. The form of the interaction (\ref{hamint}) originates from the fact that the two types of circularly polarized light interact differently to the spins of the atoms.
The initial state of the light that illuminates the atomic ensemble is taken to be in a coherent state of the form
\begin{align}
\lvert \gamma \rangle \lvert \chi\rangle = e^{-|\gamma|^2/2 } e^{-|\chi|^2/2 } e^{\gamma a^\dagger} e^{\chi b^\dagger} \lvert 0\rangle,
\end{align}
where $ \chi,\gamma $ are complex amplitude. The optimal case involves choosing $ \chi = \gamma $, which corresponds to linearly polarized light, but we will keep $ \chi,\gamma $ arbitrary for the sake of generality. The light then interacts with the atoms according to the interaction (\ref{hamint}), which entangles the atoms and the light. The state at this point becomes
\begin{align}
& e^{-i g t J_z S_z} | \gamma \rangle | \chi \rangle | \psi_0 \rangle =
| e^{-i g t J_z } \gamma \rangle | e^{i g t J_z} \chi \rangle | \psi_0 \rangle \\
& = \sum_{J=0}^{J_{\max}} \sum_{m_z = - J}^{J} \langle J, m_z | \psi_0 \rangle
| e^{-i g t m_z} \gamma \rangle | e^{i g t m_z} \chi \rangle | J, m_z \rangle ,
\label{halfwaystate}
\end{align}
where we used the fact that $ e^{i \theta a^\dagger a } | \chi \rangle = | e^{i \theta } \chi \rangle $ for a coherent state. In the second line, we used the completeness identity
\begin{align}
\sum_{J}\sum_{m_z=-J}^{J} \lvert J,m_z\rangle\langle J,m_z\rvert = \mathbbm{1} .
\label{completeness}
\end{align}
We note that the terms inside the sum of (\ref{halfwaystate}) have no dependence on $ J $. Hence the measurement operator acts on all spin sectors and affects the superposition in $ m_z $ only.
After the light interacts with the atoms, the light fields carry the information about the atoms in their phase. The information is accessed by interfering the modes $ a, b$ with each other using a waveplate according to the following transformation
\begin{equation}
\label{eq:BeamSplitter}
\hat{a}^\dagger = \frac{\hat{c}^\dagger + i\hat{d}^\dagger}{\sqrt{2}},\qquad \hat{b}^\dagger = \frac{+i\hat{c}^\dagger + \hat{d}^\dagger}{\sqrt{2}}.
\end{equation}
The state after the transformation becomes
\begin{align}
e^{-( \lvert \chi\rvert^2 + \lvert\gamma\rvert^2 )/2}
\exp \left[ \left( \frac{ e^{-i g t J_z/2 } \gamma + i e^{i g t J_z/2} \chi}{\sqrt{2}} \right) c^\dagger \right] \nonumber \\
\times \exp \left[ \left( \frac{ i e^{-i g t J_z/2 } \gamma + e^{i g t J_z/2} \chi}{\sqrt{2}} \right) d^\dagger \right] | 0\rangle | \psi_0 \rangle
\end{align}
At the final step, the photon numbers in the modes $ c, d $ are detected, yielding $n_c$ and $n_d$ photons respectively. The unnormalized state after the measurement is
\begin{align}
\label{eq:MeasuredState}
\lvert \psi_{n_c.n_d}\rangle & =\frac{
e^{-( \lvert \chi\rvert^2 + \lvert\gamma\rvert^2 )/2} }{\sqrt{n_c!n_d!}}
\left(\frac{\gamma e^{-igt J_z/2 } + i\chi e^{igt J_z/2 } }{\sqrt{2}}\right)^{n_c}\nonumber\\
&\times \left(\frac{i\gamma e^{-igt J_z/2 } + \chi e^{igt J_z/2 } }{\sqrt{2}}\right)^{n_d}
\lvert\psi_{0}\rangle.
\end{align}
On the right hand side of (\ref{eq:MeasuredState}), we see that the photon operators have disappeared, leaving only the atomic operators.
Clearly the state $\lvert\psi_{n_c,n_d}\rangle$ can be obtained by the action of the operator $\hat{M}_{n_c,n_d}$ acting on the initial state $\lvert\psi_0\rangle$, $\lvert\psi_{n_c,n_d}\rangle = \hat{M}_{n_c,n_d}\lvert\psi_0\rangle$, where the operator $\hat{M}_{n_c,n_d}$ is defined as~\cite{ilo-okeke2016}
\begin{align}
\label{eq:measurementoperator}
\hat{M}_{n_c,n_d} = & \frac{
e^{-( \lvert \chi\rvert^2 + \lvert\gamma\rvert^2 )/2} }{\sqrt{n_c!n_d!}}
\left(\frac{\gamma e^{-igt J_z/2 } + i\chi e^{igt J_z/2 } }{\sqrt{2}}\right)^{n_c}\nonumber\\
&\times \left(\frac{i\gamma e^{-igt J_z/2 } + \chi e^{igt J_z/2 } }{\sqrt{2}}\right)^{n_d}
\end{align}
and $n_c$ and $n_d$ are the number of photons detected at detectors $c$ and $d$, respectively.
The above measurement operator satisfies the decomposition of unity
\begin{equation}
\label{eq:decomposeunity}
\sum_{n_c,n_d} \hat{M}^\dagger_{n_c,n_d}\hat{M}_{n_c,n_d} = \mathbbm{1},
\end{equation}
as a consequence of the photon probability being normalized.
We now use the fact that a function of a operator $ \hat{A} $ can be decomposed into function of its eigenstates $ \lvert a_n\rangle $ with eigenvalue $ a_n $ %
\begin{align}
\label{eq:OperatorEqaution}
f(\hat{A})= \sum_{n} f(a_n) \lvert a_n\rangle\langle a_n \rvert .
\end{align}
We may hence write the measurement operator (\ref{eq:measurementoperator}) as
\begin{equation}
\label{eq:measurementoperator_spectral}
\begin{split}
&\hat{M}_{n_c,n_d} = e^{-( \lvert \chi\rvert^2 + \lvert\gamma\rvert^2 )/2} \left(\frac{\lvert \gamma\rvert^2 + \lvert\chi\rvert^2}{2} \right)^{\frac{n_c + n_d}{2}} \\
& \times \sum_{J=0}^{J_\mathrm{max}}\sum_{m_z=-J}^{J} e^{i(n_c \phi_{c}(m_z) + n_d \phi_{d}(m_z) )} \\
&\times A_{n_c n_d} (m_z) \lvert J,m_z\rangle\langle J, m_z\rvert,
\end{split}
\end{equation}
where we defined the amplitude
\begin{align}
A_{n_c n_d} (m_z) & = \frac{\left( 1 + \cos2\eta\cos2\phi(m_z) \right)^{\frac{n_c}{2}}}{\sqrt{n_c!}}\nonumber \\
& \times \frac{\left(1 - \cos2\eta \cos2\phi(m_z) \right)^{\frac{n_d}{2}}}{\sqrt{n_d!}}
\label{afuncdef}
\end{align}
phases
\begin{equation}
\label{eq:phases}
\begin{split}
\phi(m_z) & = \frac{gt m_z}{2} + \frac{\phi_{\chi\gamma}}{2} + \frac{\pi}{4} ,\\
\phi_{c} (m_z) & = \arctan\left(\tan\eta \tan\phi(m_z) \right),\\
\phi_{d} (m_z) & = \arctan\left( \cot \eta \tan\phi (m_z) \right) - \frac{\pi}{2}, \\
\phi_{\chi\gamma} & = \text{arg} (\chi) - \text{arg} (\gamma)
\end{split}
\end{equation}
and we have removed irrelevant global phases. Here we defined $\eta$ as the ratio
%
\begin{equation}
\label{eq:lightamplitude}
\tan\eta = \frac{\lvert\chi\rvert - \lvert\gamma\rvert}{\lvert\chi\rvert + \lvert\gamma\rvert}.
\end{equation}
In (\ref{eq:measurementoperator_spectral}) we used the completeness relation.
We note that in the case that $ | \chi | = | \gamma |$, $ \eta = 0 $ and the $m_z $ dependence on $ \phi_{c} (m_z), \phi_{d} (m_z)$ disappears. The amplitude function (\ref{afuncdef}) is purely real and has the effect of modulating the wavefunction.
The probability of a particular photonic measurement outcome is given by
\begin{align}
p_{n_c n_d} = & \langle \psi_0 | \hat{M}_{n_c,n_d}^\dagger \hat{M}_{n_c,n_d} | \psi_0 \rangle \nonumber \\
= & e^{-( \lvert \chi\rvert^2 + \lvert\gamma\rvert^2 )} \left(\frac{ \lvert \gamma\rvert^2 + \lvert\chi\rvert^2}{2} \right)^{n_c + n_d } \nonumber \\
& \times \sum_{J=0}^{J_\mathrm{max}}\sum_{m_z=-J}^{J} | A_{n_c n_d} (m_z) \langle J, m_z\rvert \psi_0 \rangle |^2 .
\label{probformula}
\end{align}
\section{The amplitude function}
\label{sec:ampfunc}
The final form of the POVM for QND measurements (\ref{eq:measurementoperator_spectral}), while exact, is not yet in a form that is easily grasped intuitively. In this section we give a more intuitive description of the effect of the POVM. Specifically we study the form of the amplitude function $ A_{n_c n_d} (m_z) $.
\subsection{Parameter dependence}
We first study the amplitude function $ A_{n_c n_d} (m_z) $ numerically, and capture its basic behavior as a function of various parameters. Figure \ref{fig2} shows the functional dependence of the amplitude function with $ m_z $, while varying various parameters.
In Fig. \ref{fig2}(a), we show the effect of different ratios of the measurement outcome
\begin{align}
r_{n_c n_d} = \frac{n_d - n_c}{n_c + n_d},
\label{normratio}
\end{align}
keeping $ n_c + n_d $ fixed. We see that the amplitude function generally has the form of a Gaussian, but is centered at different values of $ m_z $, depending on the value of $ r $. By solving $ d A/dm_z = 0 $ we can find location of the Gaussian to be
\begin{align}
\cos 2 \eta \sin ( gt m_z + \phi_{\chi \gamma}) = r_{n_c n_d} = \frac{n_d - n_c}{n_c + n_d} .
\label{peakposition}
\end{align}
For a choice of time $ gt = \pi/N $, the factor $ -\pi/2 \le gt m_z \le \pi/2 $ and there is a one-to-one relation between the value of $ r_{n_c n_d} $ and the spin $ m_z $. The factor of $ \cos 2 \eta $ acts to reduce the amplitude of the sine function such that a smaller range of the $ r_{n_c n_d} $ outcomes occur.
In Fig. \ref{fig2}(b), we keep $ r_{n_c n_d} = 0 $ by setting $ n_c = n_d $ but vary the total photon count $ n_c + n_d $. We see that larger total photon numbers tends to narrow the distribution. A similar narrowing effect is also seen in Fig. \ref{fig2}(c), where the photon counts $ n_c, n_d $ are kept constant, but the interaction time $ g t $ is varied. We see that longer interaction times tend to produce sharper peaks, looking around $ m_z = 0 $. However, longer interactions tends to also produce additional peaks, due to the additional solutions of (\ref{peakposition}). Such additional peaks occur when $ gt > \pi/N $ and there is no longer a one-to-one relation between $ m_z $ and $ r_{n_c n_d} $.
Finally, Fig. \ref{fig2}(d) we show the effect of varying the ensemble size $ N $. In this case, we rescale the horizontal axis so that $ -1 \le m/J \le 1 $, where $ J = N/2$. The interaction time is set to $ gt = \pi/N $. We see that the amplitude function under these rescalings are visually identical. Thus we consider $ gt = \pi/N $ to be the equivalent time regardless of the ensemble size since it affects the spins in an comparable way.
\begin{figure}[t]
\includegraphics[width=\columnwidth]{Fig2}
\caption{The amplitude function $ A_{n_c n_d} (m_z) $ as defined in (\ref{afuncdef}). Plotted are the amplitude functions as a function of $ m_z $ normalized to their maximum values. (a) Varying the ratio (\ref{normratio}) for interaction time $ gt = \pi/ N$. (b) Varying the total photon number while keeping $ r $ fixed for $ gt = \pi/ N$. (c) Varying the interaction time $ g t $ while for $ n_c = n_d = 25 $. (d) Varying $ N $ for interaction times $ gt = \pi/ N$ keeping $ n_c = n_d = 25 $ fixed. Common parameters are $ \gamma = 5.1, \chi = 5 $ with the exception of (b) where $ \gamma = 2.55, \chi = 2.5 $ was chosen for the $ n_c = n_d = 6 $ case and $ \gamma = 10.2, \chi = 10 $ was chosen for the $ n_c = n_d = 102 $ case. }
\label{fig2}
\end{figure}
\subsection{Analytical approximation}
The plots of Fig. \ref{fig2} reveal that the amplitude function takes the form of a Gaussian as long as the interaction time is in the range $0< gt \le \pi/N$. We now make an analytical approximation of the amplitude function in this regime, assuming $n_c,n_d\gg1$ such that $ | r_{n_c n_d} | \ll \cos 2 \eta $. The amplitude function can be approximated in this regime as
\begin{equation}
\begin{split}
A_{n_c n_d} (m_z) & \approx \frac{ \left(\frac{2e }{n_c + n_d}\right)^{\frac{n_c+n_d }{2}} }{\left(4\pi^2 n_c n_d \right)^{\frac{1}{4}}} e^{-\frac{1}{2\sigma^2}\left(m_z - m_0\right)^2},
\end{split}
\label{eq:amp_measurementoperator}
\end{equation}
where we have used the Stirling's approximation to rewrite the factorials. This is a Gaussian with variance
\begin{equation}
\label{eq:width}
\sigma^2 = \left( \frac{g^2t^2}{8} \frac{n_c + n_d}{n_cn_d}\left[(n_c + n_d)^2\cos^22\eta - (n_c-n_d)^2\right]\right)^{-1},
\end{equation}
and maximum position at
\begin{equation}
\label{eq:peakposition}
m_0 = \frac{1}{gt}\left[\arcsin\left(\frac{1}{\cos2\eta} \frac{n_d - n_c}{n_d + n_c}\right) - \phi_{\chi\gamma} \right].
\end{equation}
This is the same equation as (\ref{peakposition}) except that only the principal value of the arcsine has been taken.
Let us examine the form of the formula for the variance and compare it to the results of Fig. \ref{fig2}. For the case $ | \chi | = | \gamma | $, the variance simplies to
\begin{align}
\sigma^2 = \frac{2}{g^2 t^2 (n_c + n_d)} \hspace{1cm} (\eta = 0 ) .
\label{sigmasimple}
\end{align}
For this case we can easily see that the variance reduces for longer interaction times and larger photon number. We also see that choosing a time $ gt = \pi/N $ makes the width $ \sigma \propto N $. Since the spin variable $ J_z \propto N $, the $ N $-dependence cancels out as was seen in Fig. \ref{fig2}(d).
Substituting (\ref{eq:amp_measurementoperator}) in (\ref{eq:measurementoperator_spectral}) gives the approximated form of the POVM
\begin{equation}
\label{eq:measurementoperator_approx}
\begin{split}
& \hat{M}_{n_c,n_d} = \frac{e^{\frac{n_c + n_d - (\lvert\gamma\rvert^2 + \lvert \chi\rvert^2 ) }{2}}}{\left(4\pi^2 n_c n_d \right)^{\frac{1}{4}}} \left({ \frac{\lvert\gamma\rvert^2 + \lvert\chi\rvert^2}{n_c + n_d}} \right)^{\frac{n_c + n_d}{2}} \\
&\times \sum_{J=0}^{J_\mathrm{max}}\sum_{m_z=-J}^{J}
e^{i(n_c \phi_{c} + n_d\phi_{d})} e^{-\frac{1}{2\sigma^2}\left(m_z - m_0\right)^2}\lvert J,m_z\rangle\langle J, m_z\rvert.
\end{split}
\end{equation}
Due to the conservation of photon number during the sequence, the most likely total photon number $n_c + n_d$ to be detected is equal to $\lvert\gamma\rvert^2 + \lvert\chi\rvert^2$. This may be seen by examining the factors outside the sums in (\ref{eq:measurementoperator_approx}).
%
%
The width $\sigma$ offers a way to characterize the measurement operator $\hat{M}$. For $\sigma$ being large in comparison to the spacing $ \Delta m_z = 1 $ between the eigenvalues of $J_z$, several eigenvalues $| J, m_z \rangle $ would be contained in within its width. Hence a measurement with such width $\sigma > 1 $ would modify the amplitude of the initial state, by applying an Gaussian envelope function centered around $m_0$. Such measurement is often called weak measurement since it gives some information of the eigenvalue spectrum of $J_z$, but not all. According to the trade-off relation between the strength of the measurement and the backaction \cite{ilo-okeke2016}, this also means that the effect on the state is small. As a result, weak measurement would preserve the linear combination of states in the states $ | J, m_z \rangle $ within the width of the measurement operator and does not lead to a total collapse of the quantum state.
In the opposite limit where the width is small $\sigma < 1 $ in comparison to the spacing $\Delta m_z = 1 $, only a handful of eigenvalues (and in the very extreme case only one eigenvalue) is contained in the width of the measurement operator $\hat{M}$, and would be centred around $m_0$. As a result, any measurement performed with such width $\sigma \ll 1 $ would produce only one eigenvalue $|J_z,m_z \rangle $ near the peak value $m_0$.
This is a destructive measurement and leads to the collapse of the quantum state of the system.
\section{Projection Measurement Limit}
\label{sec:ProjectionMecurement}
In the previous section we saw that by adjusting the width of the amplitude function one could tune the measurement operator from a weak measurement to a projective measurement, where there is a total collapse of the quantum state. In this section, we explicitly derive the form of the projection operator, in the limit of a very narrow width of the amplitude function.
Our starting point to derive the projective limit is (\ref{eq:measurementoperator_approx}). Making the change of variables
\begin{equation}
\label{eq:relativebasis}
\begin{split}
u & = \frac{n_d + n_c}{2},\\
v & = \frac{n_d-n_c}{2}.
\end{split}
\end{equation}
and recognizing that the dominant contribution to the amplitude of the POVM comes from points around $u = (\lvert \gamma\rvert^2 + \lvert\chi\rvert^2)/2$, we write
\begin{equation}
\label{eq:OperatorRelativeBasis}
\begin{split}
\hat{M}(u,v)& = \frac{e^{iv\left(\phi_{d,0} - \phi_{c,0} \right)}}{\left[4\pi^2(u^2 -v^2)\right]^{\frac{1}{4}}} e^{-\frac{\left(u - \frac{\lvert \chi\rvert^2 + \lvert\gamma\rvert^2}{2}\right)^2}{\lvert \chi\rvert^2 + \lvert\gamma\rvert^2}}\\
&\times e^{iu\left( \phi_{c,0} + \phi_{d,0} \right)} \sum_{J=0}^{J_\mathrm{max}}\sum_{m_z=-J}^{J}
e^{-\frac{1}{2\sigma^2 }\left(m_z - m_0 \right)^2 } \\
& \times e^{-igt\left(m_z - m_0 \right) \left( u \xi_+ -v \xi_- \right)} \lvert J,m_z\rangle\langle J, m_z\rvert,
\end{split}
\end{equation}
where $\phi_{c,0} = \phi_{c}(m_0 )$, $\phi_{d,0} = \phi_{d}(m_0)$, and
\begin{equation}
\label{eq:PhaseFirstDerivative}
\begin{split}
\xi_+ & = 1- \xi_d - \xi_c \\
\xi_- & = \xi_d ( m_0) - \xi_c (m_0 ) \\
\xi_c & = \frac{1}{2}\frac{\tan\eta}{\cos^2(\phi(m_0))} \frac{1}{1 +\tan^2\eta\tan^2(\phi(m_0)) },\\
\xi_d & = \frac{1}{2}\frac{\tan\eta}{\cos^2(\phi(m_0))} \frac{1}{\tan^2(\phi(m_0)) + \tan^2\eta}.
\end{split}
\end{equation}
In writing (\ref{eq:OperatorRelativeBasis}), we have expanded the phases $\phi_{c}$ and $\phi_{d}$ around $m_0$ assuming they vary linearly in the neighborhood of $m_0$.
The operator $\hat{M}(u,v)$ satisfies the decomposition of unity, $\sum_{n_c,n_d} \hat{M}^\dagger_{nc,n_d}\hat{M}_{n_c,n_d}= \mathbbm{1} = \int du dv \hat{M}^\dagger (u,v) \hat{M} (u,v)$,
\begin{align}
\mathbbm{1} = & \sum_{J,m_z} \lvert J,m_z\rangle\langle J,m_z\rvert \int 2 du\,dv
\frac{e^{-2\frac{\left(u - \frac{\lvert \chi\rvert^2 + \lvert\gamma\rvert^2}{2}\right)^2}{\lvert \chi\rvert^2 + \lvert\gamma\rvert^2}}}{\sqrt{4\pi^2(u^2 -v^2)}} \nonumber\\
&\times e^{-\frac{(m_z - m_0)^2}{\sigma^2}},\nonumber\\
=& \sum_{J,m_z} \lvert J,m_z\rangle\langle J,m_z\rvert \int du \frac{e^{-2\frac{\left(u - \frac{\lvert \chi\rvert^2 + \lvert\gamma\rvert^2}{2}\right)^2}{\lvert \chi\rvert^2 + \lvert\gamma\rvert^2}}}{\sqrt{\pi u}},\nonumber\\
=&\sum_{J,m_z} \lvert J,m_z\rangle\langle J,m_z\rvert.
\end{align}
where the factor of $2$ in the numerator is the Jacobian of transformation from $n_c$, $n_d$ to $u$, $v$, and we have used (\ref{eq:peakposition}) in the evaluation of the integral over $v$, $dv = gt\, u\cos2\eta \cos(gt m_0) dm_0 $ in the first line. In the second line, the $u$ in the denominator is evaluated at $(\lvert\chi\rvert^2 + \lvert\gamma\rvert^2)/2$, while the identity (the last line) simply follows by the definition of completeness relation. Hence we define a projection operator for the operator $\hat{M}(u,v)$ in this limit, $\sigma\ll1$, as
\begin{equation}
\label{eq:OperatorRelativeBasis2}
\begin{split}
\hat{M} (u,v) =& \frac{e^{-\frac{\left(u - \frac{\lvert \chi\rvert^2 + \lvert\gamma\rvert^2}{2}\right)^2}{\lvert \chi\rvert^2 + \lvert\gamma\rvert^2}}} {\left[\pi g^2t^2 u (u^2\cos^2{2\eta} -v^2)\right]^{\frac{1}{4}}} \\
&\times\sum_{J=\text{int} (\lvert m_0\rvert ) }^{J_\mathrm{max}}
\lvert J, \text{int} ( m_0 ) \rangle \langle J, \text{int} ( m_0 ) \rvert,
\end{split}
\end{equation}
where $m_0$ is related to $v$ via (\ref{eq:peakposition}), and $ \text{int} ( x )$ rounds $ x $ to the nearest integer. A factor of $(\pi\sigma^2)^{1/4}$ was introduced to ensure the normalization of the projector $\hat{M} (u,v)$, and we have dropped all irrelevant global phases.
Finally, the form of the (\ref{eq:OperatorRelativeBasis2}) suggests that the projection operator can be written directly in terms of $u$ and $m_0$, if one multiplies it with the square root of the Jacobian $ gt\, u\cos2\eta \cos(gt m_0) $ of transformation from $v$ to $m_0$, as
%
\begin{equation}
\label{eq:ProjetionOperatorDirect}
\begin{split}
\hat{M} (u,m_0) & = \frac{e^{-\frac{\left(u - \frac{\lvert \chi\rvert^2 + \lvert\gamma\rvert^2}{2}\right)^2}{\lvert \chi\rvert^2 + \lvert\gamma\rvert^2}} }{\sqrt[4]{\pi u}} \\
&\times\sum_{J= \mathrm{int}(\lvert m_0\rvert)}^{J_\mathrm{max}}\lvert J, \mathrm{int}( m_0)\rangle\langle J,\mathrm{int}( m_0)\rvert.
\end{split}
\end{equation}
%
The operator $\hat{M} (u,m_0) $ consists of product of two parts, the classical part (the measurement current amplitude) that is read of on the measuring device and the projector that performs quantum operation on a quantum state of a system. However, the quantum operator part --- the projection operator --- depends on the classical properties of the measurement device to define it and its action on the quantum state via (\ref{eq:peakposition}).
\section{Examples}
\label{sec:examples}
Here we show some examples of the formalism introduced in the previous sections where we apply the QND measurements to produce various types of states. These will serve to illustrate several properties of the QND measurements. Here we will consider the case where an atomic ensemble consists of $ N $ atoms occupying two hyperfine ground states such that the effective spin of each atom can be considered to be $ j = 1/2 $. A typical choice for the two levels may be $ F = 1, m_F = -1 $ and $ F= 2, m_F = 1 $ for $^{87} \text{Rb} $.
\subsection{Spin squeezed states}
Let us assume the initial state of the atoms are prepared in a spin coherent state and given by \cite{byrnes2021}
\begin{align}
|\psi_0 \rangle & = | \theta \rangle \rangle = \left( \cos \frac{\theta}{2} | \uparrow \rangle +
\sin \frac{\theta}{2} | \downarrow \rangle \right)^{\otimes N} \nonumber \\
& = \sum^{J}_{m_z =-J} \sqrt{\binom{N}{ N/2 + m_z}} \nonumber \\
& \times \cos^{ N/2 +m_z } \frac{\theta}{2} \sin^{N/2-m_z} \frac{\theta}{2} | J, m_z \rangle ,
\label{spincoherent}
\end{align}
where we denoted the two states of each atom by $ | \uparrow \rangle$, $| \downarrow \rangle $.
This state has average values of the total spin %
\begin{align}
\langle J_x \rangle & = \frac{N}{2} \sin \theta \nonumber \\
\langle J_z \rangle & = \frac{N}{2} \cos \theta ,
\end{align}
and normalized variance
\begin{align}
\sigma^2_x \equiv \frac{\langle J_z^2 \rangle - \langle J_z \rangle^2}{N^2} = \frac{\sin \theta }{4N} .
\end{align}
The $ \sigma_x \propto 1/\sqrt{N} $ shows that a spin coherent state is in the standard quantum limit. For an angle choice $ \theta = \pi/2$, the spin coherent state is an eigenstate of $ J_x $
\begin{align}
J_x | \theta = \pi/2 \rangle \rangle = \frac{N}{2} | \theta = \pi/2 \rangle \rangle .
\end{align}
Applying the measurement operator to the spin coherent state (\ref{spincoherent}), we obtain the unnormalized state
\begin{align}
& \hat{M}_{n_c,n_d} | \psi_0 \rangle = e^{-( \lvert \chi\rvert^2 + \lvert\gamma\rvert^2 )/2} \left(\frac{\lvert \gamma\rvert^2 + \lvert\chi\rvert^2}{2} \right)^{\frac{n_c + n_d}{2}} \\
& \times \sum_{m_z=-J}^{J} e^{i(n_c \phi_{c}(m_z) + n_d \phi_{d}(m_z) )} A_{n_c n_d} (m_z) \nonumber \\
& \sqrt{\binom{N}{ N/2 + m_z}} \cos^{ N/2 +m_z } \frac{\theta}{2} \sin^{N/2-m_z} \frac{\theta}{2} \lvert J,m_z\rangle .
\end{align}
The probability of obtaining a particular photon outcome is
\begin{align}
p_{n_c n_d} & = e^{- \lvert \chi\rvert^2 - \lvert\gamma\rvert^2 } \left(\frac{\lvert \gamma\rvert^2 + \lvert\chi\rvert^2}{2} \right)^{n_c + n_d} \nonumber \\
& \times \sum_{m_z=-J}^{J} A^2 (n_c,n_d,m_z) \binom{N}{ N/2 + m_z} \nonumber \\
& \times \cos^{ N +2 m_z } \frac{\theta}{2} \sin^{N-2 m_z} \frac{\theta}{2} .
\end{align}
In Fig. \ref{fig4}(a)(b) we plot the effect of applying the QND operator to the states $ | \theta = \pi/2 \rangle \rangle$. First examining the photon outcomes, we see that the most likely outcome occurs when $ n_c = n_d = (|\chi|^2 + | \gamma|^2)/2 $. Examining different interaction times as shown in Fig. \ref{fig4}(b) we see that the effect of the QND measurement is to create a squeezing effect with respect to $ J_z $, where the original Gaussian distribution of the spin coherent state becomes a narrower distribution. In Fig. \ref{fig4}(b) the squeezing is centered around $ m_z =0 $ due to the choice of $ n_c = n_d $, for different choice, the squeezing will be centered around a different $ m_z $ value. For the longest interaction time $ gt = 4 \pi/N $, additional peaks appear due to the multiple solutions of (\ref{peakposition}). For current experimental realizations of QND, typically the interaction is in the short time regime, where only a single peak occurs.
Figure \ref{fig4}(c)(d) shows the effect of applying the QND operator on the state $ | \theta = \pi/4 \rangle \rangle$. This corresponds to a state that is polarized along the spin direction $ (J_z + J_x)/\sqrt{2} $, i.e. 45 degrees between the $ z $ and $ x $-axes on the Bloch sphere. We see that in this case the most likely outcome is $ n_d = |\chi|^2 + | \gamma|^2 $ and $ n_c = 0 $. Taking this as the typical case, we examine the effect on the wavefunction as shown in Fig. \ref{fig4}(d). We again see a squeezing effect with increased interaction times. The location of the peak depends upon the interaction time this time due to the relation (\ref{peakposition}).
\begin{figure}[t]
\includegraphics[width=\columnwidth]{Fig3}
\caption{Performing a QND measurement on a spin coherent state. (a)(c) The photon probability $ p_{n_c n_d} $ for and $gt = \pi/N$. (b)(d) Resulting wavefunction amplitude $ \langle J, m_z | \hat{M}_{n_c,n_d} | \psi_0 \rangle $ for the interaction times as marked. (a)(b) are results for $ | \psi_0 \rangle = | \theta = \pi/2 \rangle \rangle $; (c)(d) $ | \psi_0 \rangle = | \theta = \pi/4 \rangle \rangle $. The parameters of the figures are $\chi= 5$, $\gamma = 5.1$, $ N = 100$, $ J = N/2 = 10$. \label{fig4} }
\end{figure}
\subsection{Dicke states}
Next let us examine the case that the initial state is prepared in an eigenstate of $ J_z $, also known as a Dicke state
\begin{align}
| \psi_0 \rangle = |J, m_z \rangle .
\end{align}
Such a state may be produced in the limit that the QND measurement is a projective measurement, as described in Sec. \ref{sec:ProjectionMecurement}. While such states are non-trivial to prepare from an experimental perspective, they will illustrate conceptually the notion of QND measurements and show the relationship between the states and photonic outcomes.
After a QND measurement, the resulting unnormalized state becomes
\begin{align}
\hat{M}_{n_c,n_d} | \psi_0 \rangle & = e^{-( \lvert \chi\rvert^2 + \lvert\gamma\rvert^2 )/2} \left(\frac{\lvert \gamma\rvert^2 + \lvert\chi\rvert^2}{2} \right)^{\frac{n_c + n_d}{2}} \\
& \times A_{n_c n_d} (m_z) \lvert J,m_z\rangle ,
\end{align}
up to a global phase. Due to the measurement operator (\ref{eq:measurementoperator_spectral}) being diagonal with respect to the spin eigenstates $ | J, m_z \rangle $, any measurement on any of the spin eigenstates leaves the state unchanged
\begin{align}
\hat{M}_{n_c,n_d} | J, m_z \rangle \propto | J, m_z \rangle .
\end{align}
We can see the essential effect of the QND measurement where repeated measurements on the spin eigenstates remain unchanged. More generally, any mixed state consisting of Dicke state will be unchanged
\begin{align}
\hat{M}_{n_c,n_d} \rho_D \hat{M}_{n_c,n_d}^\dagger \propto \rho_D
\end{align}
where
\begin{align}
\rho_D = \sum_{m_z=-J}^J p(m_z) |J, m_z \rangle
\langle J, m_z | .
\end{align}
The probability of this outcome is
\begin{align}
p_{n_c n_d} & = e^{- \lvert \chi\rvert^2 - \lvert\gamma\rvert^2 } \left(\frac{\lvert \gamma\rvert^2 + \lvert\chi\rvert^2}{2} \right)^{n_c + n_d} A^2 (n_c,n_d,m_z) .
\end{align}
We plot the photonic probability for different values of $ m_z $ in the Dicke state in Fig. \ref{fig3}(a)(b)(c). We see that depending on the value of $ m_z$, the photonic probabilities emerge with different values of $ n_c, n_d$. First, the photons emerge along the line
\begin{align}
n_c + n_d = |\chi |^2 + | \gamma|^2 ,
\end{align}
which is guaranteed by photon number conservation --- the number of photons at the input should be equal to the number on the output. How the photons are distributed along this line, i.e. the value of $ n_c - n_d $, depends upon the value of $ m_z $ according to (\ref{eq:peakposition}).
The peak value of the photonic distribution in Figs. \ref{fig3}(a)(b)(c) recovers the value of $ m_z $ using the map in Fig. \ref{fig3}(d), which maps the relation (\ref{eq:peakposition}). We note that due to the finite width of the photonic distribution, at a single shot level, one is not guaranteed to recover the exact $ m_z $ of the initial Dicke state. This can be attributed to shot noise of the photons which gives fluctuations around the mean. Hence even in the limit of a completely squeezed Dicke state, the QND measurement gives photonic outcomes which involve noise fluctuations. We note that for $ m_z = \pm J $, the photonic distributions lie along the $n_c = 0 $ or $ n_d = 0 $ axes respectively, greatly reducing the effect of photonic shot noise (Fig. \ref{fig3}(a)). In these cases, the variance of $ r_{n_c n_d} $ is zero and the value of $ m_z $ can be read even at the single shot level.
\begin{figure}[t]
\includegraphics[width=\columnwidth]{Fig4}
\caption{The probability density of getting $n_c$ and $n_d$ photons in a measurement of a Dicke state $ | J, m_z \rangle$ with (a) $m_z=-J$; (b) $m_z = 0 $; (c) $ m_z = J/2$. Map of the relation (\ref{eq:peakposition}) giving the peak value of $ m_0 $ as a function $ n_c, n_d $. The parameters of the figures are $gt = \pi/N$, $\chi= 5$, $\gamma = 5.1$, $ N = 100$, $ J = N/2 = 50$.
\label{fig3}
}
\end{figure}
\subsection{Cat state generation}
We finally consider the case of long interaction times, where the amplitude function develops non-Gaussian characteristics, as already seen in Fig. \ref{fig2}(c). Specifically, let us consider interaction times $ gt = \pi/ 2 $, and $ \phi_{\chi \gamma} = 0 $, $ \eta = 0 $, and even $ N $. Figure \ref{fig5} shows an example of an application of the QND measurement for the time $ gt = \pi/2 $ to the spin coherent state (\ref{spincoherent}). In Fig. \ref{fig5}(a) the photon probability is shown, which shows that there are three main outcomes, corresponding to (i) $ n_c =0, n_d \approx |\chi|^2 + | \gamma|^2 $;(ii) $ n_c = n_d \approx (|\chi|^2 + | \gamma|^2)/2 $; and (iii) $ n_d =0, n_c \approx |\chi|^2 + | \gamma|^2 $. In these three cases the amplitude function simplifies to the following expressions. For $ n_c>0, n_d > 0 $
\begin{align}
A_{n_c n_d} (m_z) = \left\{
\begin{array}{ll}
\frac{2^{(n_c + n_d)/2 }}{\sqrt{n_c! n_d!}}
& \hspace{1cm} m_z \in \text{even} \\
0 & \hspace{1cm} \text{otherwise}
\end{array}
\right. ,
\label{afunccase1}
\end{align}
for $ n_c = 0 $, $ n_d \gg 1 $,
\begin{align}
A_{n_c n_d} (m_z) \approx \left\{
\begin{array}{ll}
\frac{2^{n_d/2 }}{\sqrt{ n_d!}}
& \hspace{1cm} m_z (\text{mod } 4) = 1 \\
0 & \hspace{1cm} \text{otherwise}
\end{array}
\right. ,
\label{afunccase2}
\end{align}
and for $ n_c \gg 1 $, $ n_d =0 $,
\begin{align}
A_{n_c n_d} (m_z) \approx \left\{
\begin{array}{ll}
\frac{2^{n_c/2 }}{\sqrt{ n_c!}}
& \hspace{1cm} m_z (\text{mod } 4) = 3 \\
0 & \hspace{1cm} \text{otherwise}
\end{array}
\right. .
\label{afunccase3}
\end{align}
Figure \ref{fig5}(b) shows the amplitude functions (\ref{afunccase1})-(\ref{afunccase3}). We see that for these parameters the effect of the QND measurement is to suppress the amplitudes according to the the parity of $ m_z $.
Applying the QND POVM to a spin coherent state for the $ n_c>0, n_d > 0 $ case, we obtain
\begin{align}
M_{n_c, n_d} | \theta = \pi/2 \rangle \rangle & =
\frac{1}{\sqrt{2^{N/2}}} \sum_{m_z \in \text{even} } \sqrt{\binom{N}{ N/2 + m_z}} | J, m_z \rangle \nonumber \\
& =\frac{1}{2} ( | \theta = \pi/2 \rangle \rangle + | \theta = - \pi/2 \rangle \rangle )
\end{align}
which is a Schrodinger cat state. Figure \ref{fig5}(c) shows the amplitude of the wavefunction before and after applying the QND measurement. We see that only the even values of $ m_z $ have a non-zero amplitude, as expected from (\ref{afunccase1}). Figure \ref{fig5}(d) shows the spin Wigner function defined as \cite{dowling1994wigner,byrnes2021}
\begin{align}
W(\theta, \phi) = \sum_{L=0}^{2J} \sum_{M=-L}^{L} \rho_{LM} Y_{LM} (\theta, \phi), \label{wigner}
\end{align}
where $Y_{LM} (\theta, \phi) $ are the spherical harmonic functions. Here, $\rho_{LM}$ is defined as
\begin{align}
\rho_{LM}= & \sum_{m_z=-J}^{J} \sum_{m_z'=-J}^{J} (-1)^{j-m_z-M}
\langle J m_z; J -m_z' | L M \rangle \nonumber \\
& \times \langle J m_z | \rho | J m_z' \rangle,
\end{align}
where $ \langle J m_z; J m_z' | L M\rangle $ is a Clebsch-Gordan coefficient\index{Clebsch-Gordan coefficient} for combining two angular momentum eigenstates $ |J m_z \rangle $ and $ |J m_z' \rangle $ to $ | L M \rangle $, and $ \rho $ is the density matrix. The Wigner function shows the characteristic interference fringes between the two dominant probability peaks around $ \theta = \pm \pi/2 $. This indicates the highly non-classical nature of the quantum state that is formed using QND measurements.
\begin{figure}[t]
\includegraphics[width=\columnwidth]{Fig5}
\caption{Generation of a Schrodinger cat state using QND measurements. (a) The photonic probability distribution; (b) normalized amplitude function $A_{n_c n_d} (m_z) $; (c) the state before $ \psi (m_z ) = \langle J, m_z | \psi_0 \rangle = \langle J, m_z | \theta = \pi/2 \rangle \rangle $ (dashed line) and after application of the QND measurement $ \psi (m_z ) = \langle J, m_z |M_{n_c n_d} | \theta = \pi/2 \rangle \rangle $ (solid line); (d) the Wigner function of the resulting state; The parameters of the figures are $gt = \pi/2$, $\chi= 5$, $\gamma = 5.1$, $ N = 10$, $ J = N/2 = 5$.
\label{fig5}
}
\end{figure}
\section{Summary and Conclusions \label{sec:Discussion}}
We have developed a theory of QND measurements for a single atomic ensemble. The main result of this paper is the POVM (\ref{eq:measurementoperator_spectral}) which summarizes the effect of performing a QND measurement. The basic effect of the measurement is described by the amplitude function (\ref{afuncdef}), which modulates the initial atomic wavefunction. For small interaction times, the amplitude function takes the form of a Gaussian, with a width that decreases with both photon number and interaction time. The position of the Gaussian depends upon the normalized difference of the photon counts according to relation (\ref{peakposition}). For longer interaction times, the amplitude function develops non-Gaussian characteristics, which can be taken advantage of to create more exotic non-classical states such as Schrodinger cat states. We also showed how in the large photon number limit the QND measurement reduces to a projective measurement onto the state $ | J, m_z \rangle $.
Current experiments typically work in the regime of short interaction times, hence the QND measurement should be well-approximated by the Gaussian form
(\ref{eq:measurementoperator_approx}). As seen in Fig. \ref{fig3}, even after completely collapsing the state to $ | J, m_z \rangle $, photonic shot noise prevents one from precisely determining the true value of $ m_z $ for a single shot measurement. Thus to determine $ m_z $, one would need to use a large photon number to reduce shot noise, make a large number of repeated measurements, or both. One way to evade this would be to tune the phase $ \phi_{\chi \gamma}$ such that the photon probability emerges entirely at $ n_c = 0 $ (or equivalently $ n_d = 0 $). In this way, effectively the photon number distribution is squeezed and one obtains a more accurate measurement of $ m_z $ from (\ref{peakposition}). The precise understanding of QND measurements may allow for novel applications beyond squeezing and precision measurements~\cite{ilo-okeke2022,mao2022measurement,kondappan2022imaginary}. One example is in realizing imaginary time evolution, where it was shown that combining QND measurements with a adaptive unitary can deterministically prepare eigenstates of various Hamiltonians, including cluster states \cite{mao2022measurement,kondappan2022imaginary}.
\begin{acknowledgments}
This work is supported by the National Natural Science Foundation of China (62071301); NYU-ECNU Institute of Physics at NYU Shanghai; the Joint Physics Research Institute Challenge Grant; the Science and Technology Commission of Shanghai Municipality (19XD1423000,22ZR1444600); the NYU Shanghai Boost Fund; the China Foreign Experts Program (G2021013002L); the NYU Shanghai Major-Grants Seed Fund; Tamkeen under the NYU Abu Dhabi Research Institute grant CG008. E.O.I.O. and B. C. A. acknowledge the support of the Tertiary Education Trust fund (TETFUND), Nigeria.
\end{acknowledgments}
\section*{AUTHOR DECLARATIONS}
\subsection*{Conflict of Interest}
The authors declare no conflict of interest.
\section*{Data availability}
The data that support the findings of this study are available from the corresponding author upon reasonable request.
%
%
|
\section{Introduction}
\begin{figure}[!ht]
\centering
\includegraphics[width=7cm]{MilieuPerio.png}
\qquad\begin{tikzpicture}[scale=0.7]
\pgfmathsetmacro{\cubexa}{8}
\pgfmathsetmacro{\cubeya}{1}
\pgfmathsetmacro{\cubeza}{1}
\draw[black,dotted] (-0.5,-\cubexa/2,-0.5) -- ++(0,-0.8,0);
\draw[black,dotted] (-\cubexa/2,-0.5,-0.5) -- ++(-0.8,0,0);
\draw[black,fill=gray!30] (0.5,0.5,-0.5) -- ++(0,0,-\cubexa/2+0.5) -- ++(-\cubeya,0,0) -- ++(0,0,\cubexa/2-0.5) -- cycle;
\draw[black,fill=gray!30] (0.5,0.5,-0.5) -- ++(0,-\cubeza,0) -- ++(-\cubeya,0,0) -- ++(0,\cubeza,0) -- cycle;
\draw[black,fill=gray!30] (0.5,0.5,-0.5) -- ++(0,0,-\cubexa/2+0.5) -- ++(0,-\cubeza,0) -- ++(0,0,\cubexa/2-0.5) -- cycle;
\draw[black,fill=gray!30] (0.5,\cubexa/2,0.5) -- ++(0,-\cubexa,0) -- ++(0,0,-\cubeya) -- ++(0,\cubexa,0) -- cycle;
\draw[black,fill=gray!30] (0.5,\cubexa/2,0.5) -- ++(-\cubeza,0,0) -- ++(0,0,-\cubeya) -- ++(\cubeza,0,0) -- cycle;
\draw[black,fill=gray!30] (0.5,\cubexa/2,0.5) -- ++(0,-\cubexa,0) -- ++(-\cubeza,0,0) -- ++(0,\cubexa,0) -- cycle;
\draw[black,fill=gray!30] (\cubexa/2,0.5,0.5) -- ++(-\cubexa,0,0) -- ++(0,-\cubeya,0) -- ++(\cubexa,0,0) -- cycle;
\draw[black,fill=gray!30] (\cubexa/2,0.5,0.5) -- ++(0,0,-\cubeza) -- ++(0,-\cubeya,0) -- ++(0,0,\cubeza) -- cycle;
\draw[black,fill=gray!30] (\cubexa/2,0.5,0.5) -- ++(-\cubexa,0,0) -- ++(0,0,-\cubeza) -- ++(\cubexa,0,0) -- cycle;
\draw[black,fill=gray!30] (0.5,\cubexa/2,0.5) -- ++(0,-\cubexa/2+0.5,0) -- ++(0,0,-\cubeya) -- ++(0,\cubexa/2-0.5,0) -- cycle;
\draw[black,fill=gray!30] (0.5,\cubexa/2,0.5) -- ++(-\cubeza,0,0) -- ++(0,0,-\cubeya) -- ++(\cubeza,0,0) -- cycle;
\draw[black,fill=gray!30] (0.5,\cubexa/2,0.5) -- ++(0,-\cubexa,0) -- ++(-\cubeza,0,0) -- ++(0,\cubexa,0) -- cycle;
\draw[black,fill=gray!30] (0.5,0.5,\cubexa/2) -- ++(0,0,-\cubexa/2+0.5) -- ++(-\cubeya,0,0) -- ++(0,0,\cubexa/2-0.5) -- cycle;
\draw[black,fill=gray!30] (0.5,0.5,\cubexa/2) -- ++(0,-\cubeza,0) -- ++(-\cubeya,0,0) -- ++(0,\cubeza,0) -- cycle;
\draw[black,fill=gray!30] (0.5,0.5,\cubexa/2) -- ++(0,0,-\cubexa/2+0.5) -- ++(0,-\cubeza,0) -- ++(0,0,\cubexa/2-0.5) -- cycle;
\draw[black,dotted] (\cubexa/2,0.5,0.5) -- ++(0.8,0,0);
\draw[black,dotted] (\cubexa/2,-0.5,0.5) -- ++(0.8,0,0);
\draw[black,dotted] (\cubexa/2,0.5,-0.5) -- ++(0.8,0,0);
\draw[black,dotted] (\cubexa/2,-0.5,-0.5) -- ++(0.8,0,0);
\draw[black,dotted] (-\cubexa/2,0.5,0.5) -- ++(-0.8,0,0);
\draw[black,dotted] (-\cubexa/2,-0.5,0.5) -- ++(-0.8,0,0);
\draw[black,dotted] (-\cubexa/2,0.5,-0.5) -- ++(-0.8,0,0);
\draw[black,dotted] (0.5,\cubexa/2,0.5) -- ++(0,0.8,0);
\draw[black,dotted] (0.5,\cubexa/2,-0.5) -- ++(0,0.8,0);
\draw[black,dotted] (-0.5,\cubexa/2,0.5) -- ++(0,0.8,0);
\draw[black,dotted] (-0.5,\cubexa/2,-0.5) -- ++(0,0.8,0);
\draw[black,dotted] (0.5,-\cubexa/2,0.5) -- ++(0,-0.8,0);
\draw[black,dotted] (0.5,-\cubexa/2,-0.5) -- ++(0,-0.8,0);
\draw[black,dotted] (-0.5,-\cubexa/2,0.5) -- ++(0,-0.8,0);
\draw[black,dotted] (0.5,0.5,\cubexa/2) -- ++(0,0,0.8);
\draw[black,dotted] (-0.5,0.5,\cubexa/2) -- ++(0,0,0.8);
\draw[black,dotted] (0.5,-0.5,\cubexa/2) -- ++(0,0,0.8);
\draw[black,dotted] (-0.5,-0.5,\cubexa/2) -- ++(0,0,0.8);
\draw[black,dotted] (0.5,0.5,-\cubexa/2) -- ++(0,0,-0.8);
\draw[black,dotted] (-0.5,0.5,-\cubexa/2) -- ++(0,0,-0.8);
\draw[black,dotted] (0.5,-0.5,-\cubexa/2) -- ++(0,0,-0.8);
\draw[black,dotted] (-0.5,-0.5,-\cubexa/2) -- ++(0,0,-0.8);
\end{tikzpicture}
\caption{Left: unbounded periodic cubic lattice made of thin bars of width $\varepsilon$ with square cross-sections. Right: geometry $\Omega$ of the near field, aka boundary layer, problem at the junction regions. \label{PeriodicLattice}}
\end{figure}
With the profusion of works related to graphene in physics, an important effort has been made in the mathematical community to understand the asymptotic behaviour of the spectrum of the Dirichlet operator in quantum waveguides made of thin ligaments, of characteristic width $\varepsilon\ll1$, forming unbounded periodic lattices. Various geometries have been considered and we refer the reader to \cite{Kuch02} for a review article.
To address such problems, the general approach can be summarized as follows. Using the Floquet-Bloch-Gelfand theory \cite{Gelf50,Kuch82,Skri87,Kuch93}, one shows that the spectrum of the operator in the periodic domain has a band-gap structure, the bands being generated by the eigenvalues of a spectral problem set on the periodicity cell with quasi-periodic boundary conditions involving the Floquet-Bloch parameter. The first step consists in applying techniques of dimension reduction to derive a 1D model for this spectral problem on the periodicity cell. Then one studies precisely this 1D model depending on the Floquet-Bloch parameter to get information on the spectral bands.\\
\newline
Let us mention that in the past, slapdash and casual conclusions have been made concerning the 1D model problem. This model consists of ordinary differential equations on the ligaments obtained when taking $\varepsilon\to0$ supplemented by transmission conditions at the nodes of the graph. Certain authors have inappropriately applied L. Pauling's model \cite{Paul36} and imposed Kirchoff transmission conditions at the nodes. These Kirchoff conditions boil down to impose continuity of the field and zero outgoing flux (the sum of the derivatives of the field along the outgoing directions at the node vanishes). This is correct for the Laplacian with Neumann boundary conditions (BC) and has been rigorously justified in \cite{KuZe03,ExPo05,Post12}. However it has been shown by D. Grieser in \cite{Grie08} (see also \cite{MoVa07}) that for the Dirichlet problem, in general the right conditions to impose at the nodes are Dirichlet ones. More precisely, it has been proved in \cite{Grie08} that the transmission conditions to impose depend on the existence or absence of so-called threshold resonances for the near field operator defined as the Laplacian in the geometry obtained when zooming at the junction regions (denoted $\Omega$ in the sequel, see Figure \ref{PeriodicLattice} right). We say that there is a threshold resonance if there is a non zero bounded function which solves the homogeneous problem at the frequency coinciding with the bottom of the essential spectrum (the threshold) of the Laplace operator. For the Neumann problem, the threshold is $\Lambda_\dagger=0$ and there is a threshold resonance because the constants solve $-\Delta u=0$ in $\Omega$ + Neumann BC. Due to this property, one must impose Kirchoff conditions at the nodes. For the Dirichlet problem, the continuous spectrum starts at a positive threshold $\Lambda_\dagger>0$ and in general the only solution to the problem
\begin{equation}\label{PbThreshold}
\begin{array}{|rcll}
-\Delta u&=&\Lambda_\dagger u &\mbox{ in } \Omega\\[3pt]
u&=&0&\mbox{ on } \partial\Omega
\end{array}
\end{equation}
which remains bounded at infinity is zero, i.e. there is no threshold resonance. Because of this feature, one should impose Dirichlet condition at the nodes of the 1D model (see (\ref{farfield_BC}) for the precise moment where this pops up in the analysis below).
\begin{figure}[!ht]
\centering
\begin{tikzpicture}[scale=1]
\draw[fill=gray!30,draw=none](-2,-1/2) rectangle (2,1/2);
\draw[fill=gray!30,draw=none](-1/2,-1/2) rectangle (1/2,2);
\draw[black] (-2,-1/2)--(2,-1/2);
\draw[black] (-2,1/2)--(-1/2,1/2)--(-1/2,2);
\draw[black] (2,1/2)--(1/2,1/2)--(1/2,2);
\draw[black,dotted] (-2.5,-1/2)--(2.5,-1/2);
\draw[black,dotted] (-2.5,1/2)--(-2,1/2);
\draw[black,dotted] (2.5,1/2)--(2,1/2);
\draw[black,dotted] (-1/2,2.5)--(-1/2,2);
\draw[black,dotted] (1/2,2.5)--(1/2,2);
\end{tikzpicture}\qquad\begin{tikzpicture}[scale=1]
\begin{scope}[rotate=20]
\draw[fill=gray!30,draw=none](-1/2,0) rectangle (1/2,2);
\draw[black] (-1/2,1/2)--(-1/2,2);
\draw[black] (1/2,0)--(1/2,2);
\draw[black,dotted] (-1/2,2.5)--(-1/2,2);
\draw[black,dotted] (1/2,2.5)--(1/2,2);
\end{scope}
\draw[fill=gray!30,draw=none](-2,-1/2) rectangle (2,1/2);
\begin{scope}[rotate=20]
\draw[black,dashed] (0,0)--(0,2.5);
\end{scope}
\draw[black,dashed] (-2.5,0)--(2.5,0);
\draw[black] (-2,-1/2)--(2,-1/2);
\draw[black] (-2,1/2)--(-0.7,1/2);
\draw[black] (2,1/2)--(0.35,1/2);
\draw[black,dotted] (-2.5,-1/2)--(2.5,-1/2);
\draw[black,dotted] (-2.5,1/2)--(-2,1/2);
\draw[black,dotted] (2.5,1/2)--(2,1/2);
\pgfmathparse{.5*cos(110)}\let\x\pgfmathresult
\pgfmathparse{.5*sin(110)}\let\y\pgfmathresult
\draw[blue,->] (\x,\y) arc (110:180:.5cm);
\node[left,blue] at (145:.5) {\small $\alpha$};
\end{tikzpicture}\qquad\raisebox{-0.6cm}{\begin{tikzpicture}[scale=0.75]
\draw[fill=gray!30,draw=none](-2,-1/2) rectangle (2,1/2);
\draw[fill=gray!30,draw=none](-1/2,-2) rectangle (1/2,2);
\draw[black] (-2,-1/2)--(-1/2,-1/2)--(-1/2,-2);
\draw[black] (2,-1/2)--(1/2,-1/2)--(1/2,-2);
\draw[black] (-2,1/2)--(-1/2,1/2)--(-1/2,2);
\draw[black] (2,1/2)--(1/2,1/2)--(1/2,2);
\draw[black,dotted] (2.5,-1/2)--(2,-1/2);
\draw[black,dotted] (-2.5,-1/2)--(-2,-1/2);
\draw[black,dotted] (-2.5,1/2)--(-2,1/2);
\draw[black,dotted] (2.5,1/2)--(2,1/2);
\draw[black,dotted] (-1/2,2.5)--(-1/2,2);
\draw[black,dotted] (1/2,2.5)--(1/2,2);
\draw[black,dotted] (-1/2,-2.5)--(-1/2,-2);
\draw[black,dotted] (1/2,-2.5)--(1/2,-2);
\begin{scope}[rotate=45]
\draw[fill=gray!60,draw=none](-0.707,-0.707) rectangle (0.707,0.707);
\end{scope}
\end{tikzpicture}}
\caption{Examples of T- and X-shaped 2D geometries.\label{2Dexamples}}
\end{figure}
\noindent From there, authors have worked to establish rigorous results showing the absence of non zero bounded solution to (\ref{PbThreshold}). First, different planar
quantum waveguides made of T-, X- and Y-shaped junctions of thin ligaments have been considered in \cite{Naza10a,Naza14,Naza14a,NaRU14,NaRU15,Naza17,Pank17}. In these articles, additionally it has been proved that the near field operator in $\Omega$, depending on the considered geometry, may have discrete spectrum (one or several eigenvalue below the continuous spectrum). When the latter exists, the low-frequency range of the spectrum in the periodic domain is not described by the above mentioned 1D model with Dirichlet conditions at the nodes. Instead, the first spectral bands, their number being equal to the multiplicity of the discrete spectrum, are generated by functions which are localized at the junctions regions. Let us mention that for certain exceptional geometries, for example for a sequence of angles $\alpha$ in the central domain of Figure \ref{2Dexamples}, one may have non zero solutions to (\ref{PbThreshold}) which remain bounded at infinity, i.e. existence of threshold resonance. In these situations, at least if the dimension of the space of bounded solutions to (\ref{PbThreshold}) is one, the good 1D model describing the spectral bands which are not associated with the discrete spectrum in $\Omega$, has certain Kirchoff transmission conditions at the nodes which depend on the geometry. We emphasize that this leads to very different spectra for the operator in the periodic medium. More precisely, when there is no threshold resonance, the bands of the spectrum in the periodic material become very small as $\varepsilon\to0$ and the spectral gaps enlarge. In other words, the spectrum becomes rather sparse, for most of the spectral parameters waves cannot propagate and the limit 1D ligaments are somehow disconnected. This is what we will obtain below in our configuration. On the other hand, when Problem (\ref{PbThreshold}) admits a space of dimension one of bounded solutions, the spectral bands in the periodic material are much larger.\\
\newline
Afterwards, 3D geometries were considered in \cite{BaMaNa17} (see also the corresponding note \cite{BaMaNa15}). In these articles, the authors consider quantum waveguides for which the near field domain $\Omega$ is a cruciform junction of two cylinders whose cross section coincides with the unit disc. The passage from the planar case to the spatial case requires the non obvious adaptation of the methods. In particular the characterization of the discrete spectrum of the near field operator and the proof of absence of threshold resonances are much more involved.
In the present work, we study an even more intricate 3D geometry for which the near field geometry is the union of three waveguides. Rather precise Friedrichs estimates are required to prove that the discrete spectrum of the near field operator contains exactly one eigenvalue and to show that at the threshold, zero is the only bounded solution. This work also complements the study of \cite{BaMaNa17} thanks to the numerical experiments. Let us mention that the cruciform junction of two 3D cylinders with square cross section reduces to a 2D problem in a X-shaped geometry (see again Figure \ref{2Dexamples} right) and due to factoring out, is of no interest.\\
\newline
The outline is as follows. First we describe the problem, introduce the notation and present the main results. Then we study the discrete spectrum of the near field operator. In section \ref{sectionAbsenceofBoundedSol}, we demonstrate the absence of threshold resonance for the near field operator. Section \ref{SectionModels} is dedicated to the analysis of the main theorem of the article (Theorem \ref{MainThmPerio}) with the derivation of asymptotic models for the spectral bands in the original periodic domain. Finally we show some numerics to complement the results and conclude with some appendix containing the proof of two lemmas needed in sections \ref{Section_DiscreteSpectrum}, \ref{sectionAbsenceofBoundedSol}.
\section{Notation and main results}
For $j=1,2,3$, introduce the cylinder with square cross section
\begin{equation}\label{defBars}
L_j := \{x:=(x_1,x_2,x_2)\in\mathbb{R}^3\,|\,|x_k|<1/2\mbox{ for }k\in\{1,2,3\}\setminus\{j\}\}
\end{equation}
and for $\varepsilon>0$ small, $m,n\in\mathbb{Z}:=\{0,\pm1,\pm2,\dots\}$, set
\[
\begin{array}{rcl}
L^{mn\varepsilon}_1 & := & \{x\in\mathbb{R}^3\,|\,|x_2-m|<\varepsilon/2,\,|x_3-n|<\varepsilon/2\}\\[2pt]
L^{mn\varepsilon}_2 & := & \{x\in\mathbb{R}^3\,|\,|x_1-m|<\varepsilon/2,\,|x_3-n|<\varepsilon/2\}\\[2pt]
L^{mn\varepsilon}_3 & := & \{x\in\mathbb{R}^3\,|\,|x_1-m|<\varepsilon/2,\,|x_2-n|<\varepsilon/2\}.
\end{array}
\]
Finally define the unbounded periodic domain
\[
\Pi^\varepsilon:=\bigcup_{m,n\in\mathbb{Z}}L^{mn\varepsilon}_1\cup L^{mn\varepsilon}_2\cup L^{mn\varepsilon}_3
\]
(see Figure \ref{PeriodicLattice} left). Consider the Dirichlet spectral problem for the Laplace operator
\begin{equation}\label{PbSpectral}
\begin{array}{|rcll}
-\Delta u^\varepsilon&=&\lambda^\varepsilon\,u^\varepsilon &\mbox{ in }\Pi^\varepsilon\\[3pt]
u^\varepsilon&=&0&\mbox{ on } \partial\Pi^\varepsilon.
\end{array}
\end{equation}
The variational form associated with this problem writes
\begin{equation}\label{variationPb}
\int_{\Pi^\varepsilon}\nabla u^\varepsilon\cdot\nabla v^\varepsilon\,dx=\lambda^\varepsilon\int_{\Pi^\varepsilon} u^\varepsilon v^\varepsilon\,dx,\qquad\forall v^\varepsilon\in\mrm{H}^1_0(\Pi^\varepsilon).
\end{equation}
Here $\mrm{H}^1_0(\Pi^\varepsilon)$ stands for the Sobolev space of functions of $\mrm{H}^1(\Pi^\varepsilon)$ which vanish on the boundary $\partial\Pi^\varepsilon$. Classically (see e.g \cite[\S10.1]{BiSo87}), the variational problem (\ref{variationPb}) gives rise to an unbounded, positive definite, selfadjoint operator $A^\varepsilon$ in the Hilbert space $\mrm{L}^2(\Pi^\varepsilon)$, with domain $\mathcal{D}(A^\varepsilon)\subset\mrm{H}^1_0(\Pi^\varepsilon)$. Note that this operator is sometimes called the quantum graph Laplacian \cite{BeKu13,Post12}. Since $\Pi^\varepsilon$ is unbounded, the embedding $\mrm{H}^1_0(\Pi^\varepsilon)\subset \mrm{L}^2(\Pi^\varepsilon)$ is not compact and $A^\varepsilon$ has a non empty essential component $\sigma_e(A^\varepsilon)$ (\cite[Thm. 10.1.5]{BiSo87}). Actually, due to the periodicity, we have $\sigma_e(A^\varepsilon)=\sigma(A^\varepsilon)$. The Gelfand transform (see the surveys \cite{Kuch82,Naza99a} and books \cite{Skri87,Kuch93})
\[
u^{\varepsilon}(x)\mapsto U^\varepsilon(x,\eta)=\cfrac{1}{(2\pi)^{3/2}}\sum_{\iota\in\mathbb{Z}^3} e^{i\eta\cdot \iota}u^{\varepsilon}(x+\iota),\quad\eta=(\eta_1,\eta_2,\eta_3),
\]
changes Problem (\ref{PbSpectral}) into the following spectral problem with quasi-periodic at the faces located at $x_1=\pm1/2$, $x_2=\pm1/2$, $x_3=\pm1/2$,
\begin{equation}\label{PbSpectralCell}
\begin{array}{|rcll}
-\Delta U^\varepsilon(x,\eta)&=&\Lambda^\varepsilon\,U^\varepsilon(x,\eta) & x\in\omega^\varepsilon\\[3pt]
U^\varepsilon(x,\eta)&=&0&x\in\partial\omega^\varepsilon\cap\partial\Pi^\varepsilon\\[3pt]
U^\varepsilon(-1/2,x_2,x_3,\eta) & = & e^{i\eta_1}U^\varepsilon(+1/2,x_2,x_3,\eta) & (x_2,x_3)\in(-\varepsilon/2;\varepsilon/2)^2\\[3pt]
\partial_{x_1}U^\varepsilon(-1/2,x_2,x_3,\eta) & = & e^{i\eta_1}\partial_{x_1}U^\varepsilon(+1/2,x_2,x_3,\eta) & (x_2,x_3)\in(-\varepsilon/2;\varepsilon/2)^2\\[3pt]
U^\varepsilon(x_1,-1/2,x_3,\eta) & = & e^{i\eta_2}U^\varepsilon(x_1,+1/2,x_3,\eta) & (x_1,x_3)\in(-\varepsilon/2;\varepsilon/2)^2\\[3pt]
\partial_{x_2}U^\varepsilon(x_1,-1/2,x_3,\eta) & = & e^{i\eta_2}\partial_{x_2}U^\varepsilon(x_1,+1/2,x_3,\eta) & (x_1,x_3)\in(-\varepsilon/2;\varepsilon/2)^2\\[3pt]
U^\varepsilon(x_1,x_2,-1/2,\eta) & = & e^{i\eta_3}U^\varepsilon(x_1,x_2,+1/2,\eta) & (x_1,x_2)\in(-\varepsilon/2;\varepsilon/2)^2\\[3pt]
\partial_{x_3}U^\varepsilon(x_1,x_2,-1/2,\eta) & = & e^{i\eta_3}\partial_{x_3}U^\varepsilon(x_1,x_2,+1/2,\eta) & (x_1,x_2)\in(-\varepsilon/2;\varepsilon/2)^2,
\end{array}
\end{equation}
set in the periodicity cell
\[
\omega^\varepsilon:=\omega^\varepsilon_1\cup\omega^\varepsilon_2\cup\omega^\varepsilon_3\quad\mbox{ with }\quad\begin{array}{|l}
\omega^\varepsilon_1:=(-1/2;1/2)\times(-\varepsilon/2;\varepsilon/2)^2\\[2pt]
\omega^\varepsilon_2:=(-\varepsilon/2;\varepsilon/2)\times(-1/2;1/2)\times(-\varepsilon/2;\varepsilon/2)\\[2pt]
\omega^\varepsilon_3:=(-\varepsilon/2;\varepsilon/2)^2\times(-1/2;1/2).
\end{array}
\]
Problem (\ref{PbSpectralCell}) is formally selfadjoint for any value of the dual variable $\eta\in\mathbb{R}^3$. Additionally it is $2\pi$-periodic with respect to each of the $\eta_j$ because the transformation $\eta_j\mapsto\eta_j+2\pi$ leaves invariant the quasiperiodicity conditions. For any $\eta\in[0;2\pi)^3$, the spectrum of (\ref{PbSpectralCell}) is discrete, made of a monotone increasing positive sequence of eigenvalues
\[
0<\Lambda_1^\varepsilon(\eta)\le \Lambda_2^\varepsilon(\eta)\le \dots \le \Lambda_p^\varepsilon(\eta)\le \dots
\]
where the $\Lambda_p^\varepsilon(\eta)$ are counted according to their multiplicity. The functions
\[
\eta\mapsto \Lambda_p^\varepsilon(\eta)
\]
are continuous (\cite[Chap. 9]{Kato95}) so that the sets
\begin{equation}\label{SpectralBands}
\Upsilon^\varepsilon_p=\{\Lambda_p^\varepsilon(\eta),\,\eta\in[0;2\pi)^3\}
\end{equation}
are connected compact segments. Finally, according to the theory (see again \cite{Gelf50,Kuch82,Skri87,Naza99a,Kuch93}), the spectrum of the operator $A^{\varepsilon}$ has the form
\[
\sigma(A^\varepsilon)=\bigcup_{p\in\mathbb{N}^\ast}\Upsilon^\varepsilon_p
\]
where $\mathbb{N}^\ast:=\{1,2,\dots\}$. At this stage, we see that to clarify the behaviour of the spectrum of $A^\varepsilon$ with respect to $\varepsilon\to0^+$, we need to study the dependence of the $\Upsilon^\varepsilon_p$ with respect to $\varepsilon$.\\
\newline
As already mentioned in the introduction, the analysis developed for example in \cite{Grie08,Naza05,Naza14} shows that the asymptotic behaviour of the $\Upsilon^\varepsilon_p$ with respect to $\varepsilon$ depends on the features of the Dirichlet Laplacian in the geometry obtained when zooming at the junction region of the periodicity cell $\omega^\varepsilon$. More precisely, introduce the unbounded domain
\begin{equation}\label{DefOmega}
\Omega:=L_1\cup L_2\cup L_3
\end{equation}
(see Figure \ref{PeriodicLattice} right) where the $L_j$ are the cylinders with unit square cross section appearing in (\ref{defBars}). In $\Omega$, consider the Dirichlet spectral problem for the Laplace operator
\begin{equation}\label{PbSpectralZoom}
\begin{array}{|rcll}
-\Delta v&=&\mu\,v &\mbox{ in }\Omega\\[3pt]
v&=&0&\mbox{ on } \partial\Omega
\end{array}
\end{equation}
which is now independent of $\varepsilon$. We denote by $A^\Omega$ the unbounded, positive definite, selfadjoint operator naturally associated with this problem defined in $\mrm{L}^2(\Omega)$ and of domain $\mathcal{D}(A^\Omega)\subset\mrm{H}^1_0(\Omega)$. Its continuous spectrum $\sigma_c(A^\Omega)$ occupies the ray $[2\pi^2;+\infty)$ and the threshold point $\Lambda_\dagger:=2\pi^2$ is the first eigenvalue of the Dirichlet problem in the cross sections of the branches of $\Omega$ (which are unit squares). The main goal of this article is to show the following results.
\begin{theorem}\label{ThmDiscreteSpectrum}
The discrete spectrum of the operator $A^{\Omega}$ contains exactly one eigenvalue $\mu_1\in(0;2\pi^2)$.
\end{theorem}
\begin{remark}
Numerically, in Section \ref{SectionNumerics}, we find $\mu_1\approx 12.9\approx1.3\pi^2$. Note that the eigenfunctions associated with $\mu_1$ decay at infinity as $O(-\sqrt{2\pi^2-\mu_1}\,|x_j|)$ in $L_j$, $j=1,2,3$.
\end{remark}
\noindent As classical in literature, we shall say that there is a threshold resonance for Problem (\ref{PbSpectralZoom}) if there is non trivial function which solves (\ref{PbSpectralZoom}) with the threshold value $\mu=2\pi^2$ of the spectral parameter.
\begin{theorem}\label{ThmNoBoundedSol}
There is no threshold resonance for Problem (\ref{PbSpectralZoom}).
\end{theorem}
\noindent From Theorems \ref{ThmDiscreteSpectrum} and \ref{ThmNoBoundedSol}, we will derive the final result for the spectrum of $A^\varepsilon$ in the initial unbounded periodic lattice:
\begin{theorem}\label{MainThmPerio}
There are positive values $\varepsilon_p>0$ and $c_p>0$ such that for the spectral bands introduced in (\ref{SpectralBands}), we have the estimates
\[
\begin{array}{ll}
\Upsilon^\varepsilon_1\subset \varepsilon^{-2}\mu_1+\varepsilon^{-2}e^{-\sqrt{2\pi^2-\mu_1}/\varepsilon}(-c_1;c_1),&\ \varepsilon\in(0;\varepsilon_1]\\[8pt]
\Upsilon^\varepsilon_{1+q+3(p-1)}\subset \varepsilon^{-2}2\pi^2+p^2\pi^2+\varepsilon\,(-c_p;c_p),&\ \varepsilon\in(0;\varepsilon_p],\ q=1,2,3,\ p\in\mathbb{N}^\ast.\\
\end{array}
\]
\end{theorem}
\noindent Let us comment this result. First, as already mentioned in the introduction, the spectrum of the operator $A^\varepsilon$ goes to $+\infty$ as $\varepsilon\to0$. Additionally this spectrum becomes very sparse. Indeed Theorem \ref{MainThmPerio} implies the following results. The length of the spectral bands are infinitesimal as $\varepsilon\to0$. Moreover, between the bands $\Upsilon^\varepsilon_1$ and $\Upsilon^\varepsilon_2$, there is a gap, that is a segment of spectral parameters $\Lambda$ such that waves cannot propagate, of size $O(\varepsilon^{-2}(2\pi^2-\mu_1))$ while between $\Upsilon^\varepsilon_{1+q+3(p-1)}$ and $\Upsilon^\varepsilon_{1+q+3p}$, the gap is of width $O(\pi^2(2p-1))$. Therefore the main message here is that in the thin lattice $\Pi^\varepsilon$ the propagation of waves is hampered and occurs for very narrow intervals of frequencies.
\section{Properties of the discrete spectrum of $A^{\Omega}$}\label{Section_DiscreteSpectrum}
The goal of this section is to prove Theorem \ref{ThmDiscreteSpectrum}. We start by showing that the discrete spectrum of $A^{\Omega}$ is non empty (for related multidimensional problems, see \cite{Naza12a}).
\begin{proposition}\label{PropoAtLeastOne}
The discrete spectrum of $A^{\Omega}$ has at least one eigenvalue.
\end{proposition}
\begin{proof}
Define the 2D X-shaped geometry $\Omega^{2D}:=\Omega^{2D}_1\cup\Omega^{2D}_2$ with $\Omega^{2D}_1:=\mathbb{R}\times(-1/2;1/2)$ and $\Omega^{2D}_2:=(-1/2;1/2)\times\mathbb{R}$ (see Figure \ref{2Dexamples} right). According to \cite{ScRW89,ABGM91} (see also \cite[Thm. 2.1]{Naza14b} as well as the discussion at the end of this proof), we know that the Dirichlet Laplacian in $\Omega^{2D}$ admits exactly one eigenvalue $\mu^{2D}$ below the continuous spectrum which coincides with $[\pi^2;+\infty)$. Let $\varphi\in\mrm{H}^1_0(\Omega^{2D})$ be a corresponding eigenfunction. In the domain $\Omega\subset\mathbb{R}^3$ introduced in (\ref{DefOmega}), consider the function $v$ such that
\[
\begin{array}{ll}
v(x_1,x_2,x_3) = \varphi(x_1,x_2)\cos(\pi x_3)\ \mbox{ in }L_1\cup L_2,\qquad\qquad &v(x_1,x_2,x_3)=0\ \mbox{ in }\Omega\setminus\overline{L_1\cup L_2}.
\end{array}
\]
We have
\[
\int_{\Omega}|\nabla v|^2\,dx=\int_{L_1\cup L_2}|\nabla v|^2\,dx=-\int_{L_1\cup L_2}\Delta v v\,dx=(\mu^{2D}+\pi^2)\int_{\Omega} v^2\,dx.
\]
But according to the max-min principle (cf. \cite[Thm. 10.2.2]{BiSo87}), we know that
\begin{equation}\label{CaracInfSup}
\inf \sigma(A^{\Omega})=\inf_{w\in\mrm{H}^1_0(\Omega)\setminus\{0\}}\cfrac{\displaystyle\int_{\Omega}|\nabla w|^2\,dx}{\displaystyle\int_{\Omega}w^2\,dx}\,.
\end{equation}
Inserting the above $v$ in the right hand side of (\ref{CaracInfSup}), we deduce that the discrete spectrum of $A^{\Omega}$ contains an eigenvalue below $\mu^{2D}+\pi^2$.\\
\newline
We end this proof by describing the elegant trick proposed in \cite{ScRW89,ABGM91} to show the existence of an eigenvalue in the discrete spectrum of the Dirichlet Laplacian in $\Omega^{2D}$. Consider the square coloured in dark gray of Figure \ref{2Dexamples} right. It has side $\sqrt{2}$. Therefore the first eigenvalue of the Dirichlet Laplacian in this geometry is equal to $\pi^2$. Hence, in any domain containing strictly this square and included in $\Omega^{2D}$, the first eigenvalue of the Dirichlet Laplacian is strictly less than $\pi^2$. Extending a corresponding eigenfunction by zero to $\Omega^{2D}$ and using the max-min principle, we infer that the discrete spectrum of the Dirichlet Laplacian in $\Omega^{2D}$ is not empty.
\end{proof}
\noindent It remains to show that the discrete spectrum of $A^{\Omega}$ has at most one eigenvalue. To proceed, first we establish a result of symmetry.
\begin{lemma}\label{LemmaSym}
Let $v\in\mrm{H}^1_0(\Omega)$ be an eigenfunction of the operator $A^{\Omega}$ associated with an eigenvalue $\mu<2\pi^2$. Then $v$ is symmetric with respect to the three planes $(Ox_2x_3)$, $(Ox_3x_1)$, $(Ox_1x_2)$.
\end{lemma}
\begin{proof}
We present the proof of symmetry with respect to the plane $(Ox_2x_3)$. The two other symmetries can be established similarly. Introduce the function $\varphi$ such that
\[
\varphi(x_1,x_2,x_3)=v(x_1,x_2,x_3)-v(-x_1,x_2,x_3).
\]
Our goal is to prove that $\varphi\equiv0$. Set $\Omega^+_1:=\{(x_1,x_2,x_3)\in\Omega\,|\,x_1>0\}$. The function $\varphi$ satisfies $-\Delta\varphi=\mu\varphi$ in $\Omega^+_1$ and $\varphi=0$ on $\partial\Omega^+_1$. Therefore we have
\begin{equation}\label{IdentityEigen}
\int_{\Omega^+_1}|\nabla\varphi|^2\,dx=\mu\int_{\Omega^+_1}\varphi^2\,dx.
\end{equation}
\begin{figure}[!ht]
\centering\begin{tikzpicture}[scale=0.75]
\pgfmathsetmacro{\cubexa}{8}
\pgfmathsetmacro{\cubeya}{1}
\pgfmathsetmacro{\cubeza}{1}
\begin{scope}[rotate=90]
\draw[black,dotted] (0,-\cubexa/2,-0.5) -- ++(0,-0.8,0);
\draw[black,fill=gray!30] (0.5,0.5,-1) -- ++(0,0,-\cubexa/2) -- ++(-\cubeya/2,0,0) -- ++(0,0,\cubexa/2) -- cycle;
\draw[black,fill=gray!30] (0.5,0.5,-1) -- ++(0,-\cubeza,0) -- ++(-\cubeya/2,0,0) -- ++(0,\cubeza,0) -- cycle;
\draw[black,fill=gray!30] (0.5,0.5,-1) -- ++(0,0,-\cubexa/2) -- ++(0,-\cubeza,0) -- ++(0,0,\cubexa/2) -- cycle;
\draw[black,fill=gray!30] (0.5,\cubexa/2+1,0.5) -- ++(0,-\cubexa/2,0) -- ++(0,0,-\cubeya) -- ++(0,\cubexa/2,0) -- cycle;
\draw[black,fill=gray!30] (0.5,\cubexa/2+1,0.5) -- ++(-\cubeza/2,0,0) -- ++(0,0,-\cubeya) -- ++(\cubeza/2,0,0) -- cycle;
\draw[black,fill=gray!30] (0.5,\cubexa/2+1,0.5) -- ++(0,-\cubexa/2,0) -- ++(-\cubeza/2,0,0) -- ++(0,\cubexa/2,0) -- cycle;
\draw[black,fill=gray!30] (\cubexa/2+1,0.5,0.5) -- ++(-\cubexa/2,0,0) -- ++(0,-\cubeya,0) -- ++(\cubexa/2,0,0) -- cycle;
\draw[black,fill=gray!30] (\cubexa/2+1,0.5,0.5) -- ++(0,0,-\cubeza) -- ++(0,-\cubeya,0) -- ++(0,0,\cubeza) -- cycle;
\draw[black,fill=gray!30] (\cubexa/2+1,0.5,0.5) -- ++(-\cubexa/2,0,0) -- ++(0,0,-\cubeza) -- ++(\cubexa/2,0,0) -- cycle;
\draw[black,fill=gray!30] (0.5,0.5,0.5) -- ++(-0.5,0,0) -- ++(0,-\cubeya,0) -- ++(0.5,0,0) -- cycle;
\draw[black,fill=gray!30] (0.5,0.5,0.5) -- ++(0,0,-\cubeza) -- ++(0,-\cubeya,0) -- ++(0,0,\cubeza) -- cycle;
\draw[black,fill=gray!30] (0.5,0.5,0.5) -- ++(-0.5,0,0) -- ++(0,0,-\cubeza) -- ++(0.5,0,0) -- cycle;
\draw[black,fill=gray!30] (0.5,-1,0.5) -- ++(0,-\cubexa/2,0) -- ++(0,0,-\cubeya) -- ++(0,\cubexa/2,0) -- cycle;
\draw[black,fill=gray!30] (0.5,-1,0.5) -- ++(-\cubeza/2,0,0) -- ++(0,0,-\cubeya) -- ++(\cubeza/2,0,0) -- cycle;
\draw[black,fill=gray!30] (0.5,-1,0.5) -- ++(0,-\cubexa/2,0) -- ++(-\cubeza/2,0,0) -- ++(0,\cubexa/2,0) -- cycle;
\draw[black,fill=gray!30] (0.5,0.5,\cubexa/2+1) -- ++(0,0,-\cubexa/2) -- ++(-\cubeya/2,0,0) -- ++(0,0,\cubexa/2) -- cycle;
\draw[black,fill=gray!30] (0.5,0.5,\cubexa/2+1) -- ++(0,-\cubeza,0) -- ++(-\cubeya/2,0,0) -- ++(0,\cubeza,0) -- cycle;
\draw[black,fill=gray!30] (0.5,0.5,\cubexa/2+1) -- ++(0,0,-\cubexa/2) -- ++(0,-\cubeza,0) -- ++(0,0,\cubexa/2) -- cycle;
\draw[black,dotted] (\cubexa/2+1,0.5,0.5) -- ++(0.8,0,0);
\draw[black,dotted] (\cubexa/2+1,-0.5,0.5) -- ++(0.8,0,0);
\draw[black,dotted] (\cubexa/2+1,0.5,-0.5) -- ++(0.8,0,0);
\draw[black,dotted] (\cubexa/2+1,-0.5,-0.5) -- ++(0.8,0,0);
\draw[black,dotted] (0.5,\cubexa/2+1,0.5) -- ++(0,0.8,0);
\draw[black,dotted] (0.5,\cubexa/2+1,-0.5) -- ++(0,0.8,0);
\draw[black,dotted] (0,\cubexa/2+1,0.5) -- ++(0,0.8,0);
\draw[black,dotted] (0,\cubexa/2+1,-0.5) -- ++(0,0.8,0);
\draw[black,dotted] (0.5,-\cubexa/2-1,0.5) -- ++(0,-0.8,0);
\draw[black,dotted] (0.5,-\cubexa/2-1,-0.5) -- ++(0,-0.8,0);
\draw[black,dotted] (0,-\cubexa/2-1,0.5) -- ++(0,-0.8,0);
\draw[black,dotted] (0.5,0.5,\cubexa/2+1) -- ++(0,0,0.8);
\draw[black,dotted] (0,0.5,\cubexa/2+1) -- ++(0,0,0.8);
\draw[black,dotted] (0.5,-0.5,\cubexa/2+1) -- ++(0,0,0.8);
\draw[black,dotted] (0,-0.5,\cubexa/2+1) -- ++(0,0,0.8);
\draw[black,dotted] (0.5,0.5,-\cubexa/2-1) -- ++(0,0,-0.8);
\draw[black,dotted] (0,0.5,-\cubexa/2-1) -- ++(0,0,-0.8);
\draw[black,dotted] (0.5,-0.5,-\cubexa/2-1) -- ++(0,0,-0.8);
\node at (3,0,0.5){$L_1^+$};
\node at (0,2.6,-1){$S_2^+$};
\node at (0,-3.6,-1){$S_2^-$};
\node at (0,-0.5,2){$S_3^+$};
\node at (0,-0.5,-4.5){$S_3^-$};
\draw[black,->] (1.6,-1.6,0) -- ++(-1,1,0);
\node at (1.8,-2,0){$Q^+_1$};
\end{scope}
\end{tikzpicture}\qquad \begin{tikzpicture}[scale=0.75]
\pgfmathsetmacro{\cubexa}{8}
\pgfmathsetmacro{\cubeya}{1}
\pgfmathsetmacro{\cubeza}{1}
\begin{scope}[rotate=90]
\draw[black,fill=gray!30] (0.5,\cubexa/2+1,0.5) -- ++(0,-\cubexa/2,0) -- ++(0,0,-\cubeya/2) -- ++(0,\cubexa/2,0) -- cycle;
\draw[black,fill=gray!30] (0.5,\cubexa/2+1,0.5) -- ++(-\cubeza/2,0,0) -- ++(0,0,-\cubeya/2) -- ++(\cubeza/2,0,0) -- cycle;
\draw[black,fill=gray!30] (0.5,\cubexa/2+1,0.5) -- ++(0,-\cubexa/2,0) -- ++(-\cubeza/2,0,0) -- ++(0,\cubexa/2,0) -- cycle;
\draw[black,fill=gray!30] (\cubexa/2+1,0.5,0.5) -- ++(-\cubexa/2,0,0) -- ++(0,-\cubeya/2,0) -- ++(\cubexa/2,0,0) -- cycle;
\draw[black,fill=gray!30] (\cubexa/2+1,0.5,0.5) -- ++(0,0,-\cubeza/2) -- ++(0,-\cubeya/2,0) -- ++(0,0,\cubeza/2) -- cycle;
\draw[black,fill=gray!30] (\cubexa/2+1,0.5,0.5) -- ++(-\cubexa/2,0,0) -- ++(0,0,-\cubeza/2) -- ++(\cubexa/2,0,0) -- cycle;
\draw[black,fill=gray!30] (0.5,0.5,0.5) -- ++(-0.5,0,0) -- ++(0,-\cubeya/2,0) -- ++(0.5,0,0) -- cycle;
\draw[black,fill=gray!30] (0.5,0.5,0.5) -- ++(0,0,-\cubeza/2) -- ++(0,-\cubeya/2,0) -- ++(0,0,\cubeza/2) -- cycle;
\draw[black,fill=gray!30] (0.5,0.5,0.5) -- ++(-0.5,0,0) -- ++(0,0,-\cubeza/2) -- ++(0.5,0,0) -- cycle;
\draw[black,fill=gray!30] (0.5,0.5,\cubexa/2+1) -- ++(0,0,-\cubexa/2) -- ++(-\cubeya/2,0,0) -- ++(0,0,\cubexa/2) -- cycle;
\draw[black,fill=gray!30] (0.5,0.5,\cubexa/2+1) -- ++(0,-\cubeza/2,0) -- ++(-\cubeya/2,0,0) -- ++(0,\cubeza/2,0) -- cycle;
\draw[black,fill=gray!30] (0.5,0.5,\cubexa/2+1) -- ++(0,0,-\cubexa/2) -- ++(0,-\cubeza/2,0) -- ++(0,0,\cubexa/2) -- cycle;
\draw[black,dotted] (\cubexa/2+1,0.5,0.5) -- ++(0.8,0,0);
\draw[black,dotted] (\cubexa/2+1,0,0.5) -- ++(0.8,0,0);
\draw[black,dotted] (\cubexa/2+1,0.5,0) -- ++(0.8,0,0);
\draw[black,dotted] (\cubexa/2+1,0,0) -- ++(0.8,0,0);
\draw[black,dotted] (0.5,\cubexa/2+1,0.5) -- ++(0,0.8,0);
\draw[black,dotted] (0.5,\cubexa/2+1,0) -- ++(0,0.8,0);
\draw[black,dotted] (0,\cubexa/2+1,0.5) -- ++(0,0.8,0);
\draw[black,dotted] (0,\cubexa/2+1,0) -- ++(0,0.8,0);
\draw[black,dotted] (0.5,0.5,\cubexa/2+1) -- ++(0,0,0.8);
\draw[black,dotted] (0,0.5,\cubexa/2+1) -- ++(0,0,0.8);
\draw[black,dotted] (0.5,0,\cubexa/2+1) -- ++(0,0,0.8);
\draw[black,dotted] (0,0,\cubexa/2+1) -- ++(0,0,0.8);
\draw[black,->] (1.6,1.6,0) -- ++(-1,-1,0);
\node at (1.8,1.8,0){$Q^+$};
\end{scope}
\end{tikzpicture}
\caption{Left: exploded decomposition of $\Omega^+_1$. Right: exploded decomposition of $\Omega^+$. \label{ExplodedDomain}}
\end{figure}
\noindent Define the domains
\[
\begin{array}{rcl}
L^+_1 & := & \{x\in L_1\,|\,x_1>1/2\}\\[2pt]
S_2 & := & \{x\in L_2\,|\,x_1>0\}=\Omega^+_1\cap L_2\\[2pt]
S_3 & := & \{x\in L_3\,|\,x_1>0\}=\Omega^+_1\cap L_3\\[2pt]
S^{\pm}_2 & := & \{x\in S_2\,|\,\pm x_2>1/2\}\\[2pt]
S^{\pm}_3 & := & \{x\in S_3\,|\,\pm x_3>1/2\}\\[2pt]
Q^+_1& := & \{x\in\mathbb{R}^3\,|\,0<x_1<1/2,\,|x_2|<1/2,\,|x_3|<1/2\}=S_2\cap S_3
\end{array}
\]
(see the exploded view of Figure \ref{ExplodedDomain} left). Using the Poincar\'e inequality in a section of $L^+_1$, which is a square of unit side, and then integrating with respect to the $x_1$ variable, first we find
\begin{equation}\label{PoincareTranche0}
2\pi^2 \int_{L^+_1}\varphi^2\,dx \le \int_{L^+_1}|\nabla\varphi|^2\,dx.
\end{equation}
In $S^\pm_2$, $S^\pm_3$, since the transverse section is smaller than in $L^+_1$, we have better estimates, namely
\begin{equation}\label{PoincareTranche}
5\pi^2\int_{S_2^\pm}\varphi^2\,dx \le \int_{S_2^\pm}(\partial_{x_1}\varphi)^2+(\partial_{x_3}\varphi)^2\,dx,\qquad 5\pi^2\int_{S_3^\pm}\varphi^2\,dx \le \int_{S_3^\pm}(\partial_{x_1}\varphi)^2+(\partial_{x_2}\varphi)^2\,dx.
\end{equation}
The difficulty to complete the proof however is that in $Q^+_1$, using a 1D Poincar\'e inequality with respect to the $x_1$ variable, we can simply get
\begin{equation}\label{PoincareTranche0bis}
\pi^2 \int_{Q^+_1}\varphi^2\,dx \le \int_{Q^+_1}(\partial_{x_1}\varphi)^2\,dx.
\end{equation}
This is enough to conclude directly that $\varphi\equiv0$ when $\mu\in(0;\pi^2)$ but not in the general situation $\mu\in(0;2\pi^2)$ that we wish to deal with. Therefore we have to refine the strategy. What we will do is to exploit the ``extra'' terms coming from the efficiency of estimates (\ref{PoincareTranche}) to control the $\mrm{L}^2$-norm of $\varphi$ in $Q^+_1$. More precisely, the Lemma 5.1 of \cite{BaMaNa17} that we recall in Appendix (Lemma \ref{LemmaPoincareFriedrichs}) guarantees that for a given $a>0$, we have the Friedrichs inequality
\begin{equation}\label{PoincareModified}
\kappa(a)\int_{0}^{1/2}\phi^2\,dt \le \int_{0}^{+\infty}(\partial_t\phi)^2\,dt+a^2\int_{1/2}^{+\infty}\phi^2\,dt,\qquad \forall \phi\in\mrm{H}^1(0;+\infty),
\end{equation}
where $\kappa(a)$ is the smallest positive root of the transcendental equation
\begin{equation}\label{TransEqn}
\sqrt{\kappa}\tan\bigg(\cfrac{\sqrt{\kappa}}{2}\bigg)=a.
\end{equation}
In particular, solving (\ref{TransEqn}) with $a=\pi\sqrt{5/2}$, we find $\kappa(a)>\pi^2/2$. Therefore, using (\ref{PoincareModified}), we can write
\begin{equation}\label{PoincareTranche01}
\cfrac{\pi^2}{2}\int_{Q^+_1}\varphi^2\,dx \le \int_{S_2}(\partial_{x_2}\varphi)^2\,dx+\cfrac{5\pi^2}{2}\int_{S_2^+\cup S_2^-}\varphi^2\,dx
\end{equation}
as well as
\begin{equation}\label{PoincareTranche02}
\cfrac{\pi^2}{2}\int_{Q^+_1}\varphi^2\,dx \le \int_{S_3}(\partial_{x_3}\varphi)^2\,dx+\cfrac{5\pi^2}{2}\int_{S_3^+\cup S_3^-}\varphi^2\,dx.
\end{equation}
Then inserting (\ref{PoincareTranche}) in (\ref{PoincareTranche01}), (\ref{PoincareTranche02}) and summing up the resulting estimates, we obtain
\begin{equation}\label{PoincareTranche1}
\begin{array}{rcl}
\pi^2\displaystyle\int_{Q^+_1}\varphi^2\,dx &\le&\displaystyle \int_{S_2}(\partial_{x_2}\varphi)^2\,dx+\cfrac{1}{2}\int_{S_2^+\cup S_2^-}(\partial_{x_3}\varphi)^2\,dx+\displaystyle \int_{S_3}(\partial_{x_3}\varphi)^2\,dx+\cfrac{1}{2}\int_{S_3^+\cup S_3^-}(\partial_{x_2}\varphi)^2\,dx\\[10pt]
& & +\displaystyle\cfrac{1}{2}\int_{S_2^+\cup S_2^-\cup S_3^+\cup S_3^-}(\partial_{x_1}\varphi)^2\,dx.
\end{array}
\end{equation}
On the other hand, (\ref{PoincareTranche}) also yields
\begin{equation}\label{PoincareTranche2}
2\pi^2\int_{S^+_2\cup S^-_2\cup S^+_3\cup S^-_3}\varphi^2\,dx \le \cfrac{1}{2} \int_{S^+_2\cup S^-_2\cup S^+_3\cup S^-_3}(\partial_{x_1}\varphi)^2\,dx+\cfrac{1}{2}\int_{S_2^+\cup S_2^-}(\partial_{x_3}\varphi)^2\,dx+\cfrac{1}{2}\int_{S_3^+\cup S_3^-}(\partial_{x_2}\varphi)^2\,dx.
\end{equation}
Finally, summing up (\ref{PoincareTranche0}), (\ref{PoincareTranche0bis}), (\ref{PoincareTranche1}) and (\ref{PoincareTranche2}), we get
\[
2\pi^2 \int_{\Omega^+_1}\varphi^2\,dx \le \int_{\Omega^+_1}|\nabla\varphi|^2\,dx.
\]
This estimate together with the identity (\ref{IdentityEigen}) imply $\varphi\equiv0$.
\end{proof}
\noindent We can now establish the main result of this section.\\[10pt]
\textit{Proof of Theorem \ref{ThmDiscreteSpectrum}.} From Proposition \ref{PropoAtLeastOne}, we know that there is at least one eigenvalue $\mu_1$ of $A^{\Omega}$ below $2\pi^2$. Assume that $A^{\Omega}$ has a second eigenvalue $\mu_2$ such that $\mu_2<2\pi^2$. Set $\Omega^+:=\{x\in\Omega\,|\,x_1>0,\,x_2>0,\,x_3>0\}$ (see Figure \ref{ExplodedDomain} right). Then according to the result of symmetry of Lemma \ref{LemmaSym}, the problem
\[
\begin{array}{|rcll}
-\Delta v &=&\mu v &\mbox{in }\Omega^+\\[3pt]
v&=&0 &\mbox{on }\Sigma_0:=\partial\Omega\cap \partial \Omega^+\\[3pt]
\partial_nv&=&0 &\mbox{on }\partial \Omega^+\setminus\Sigma_0
\end{array}
\]
admits the two eigenvalues $\mu_1$, $\mu_2$. Besides, from the max-min principle (\cite[Thm. 10.2.2]{BiSo87}), we have
\[
\mu_2=\max_{E\subset\mathscr{E}_1}\inf_{w\in E\setminus\{0\}}\cfrac{\displaystyle\int_{\Omega^+}|\nabla w|^2\,dx}{\displaystyle\int_{\Omega^+} w^2\,dx}\,
\]
where $\mathscr{E}_1$ denotes the set of subspaces of $\mrm{H}^1_{0}(\Omega^+;\Sigma_0):=\{\varphi\in\mrm{H}^1(\Omega^+)\,|\,\varphi=0\mbox{ on }\Sigma_0\}$ of codimension one. In particular, we have
\begin{equation}\label{Contra1}
\mu_2 \ge \inf_{w\in E\setminus\{0\}}\cfrac{\displaystyle\int_{\Omega^+}|\nabla w|^2\,dx}{\displaystyle\int_{\Omega^+} w^2\,dx}
\end{equation}
with $E=\{\varphi\in\mrm{H}^1_{0}(\Omega^+;\Sigma_0)\,|\,\int_{Q^+}\varphi\,dx=0\}$, $Q^+:=(0;1/2)^3$. However from the Poincar\'e inequality, we can write for $w\in E\setminus\{0\}$,
\begin{equation}\label{IdentityEigenD}
2\pi^2\int_{\Omega^+\setminus\overline{Q^+}}w^2\,dx \le \int_{\Omega^+\setminus\overline{Q^+}} |\nabla w|^2\,dx
\end{equation}
and there holds, according to the max-min principle,
\begin{equation}\label{Contra2}
\inf_{w\in E\setminus\{0\}}\cfrac{\displaystyle\int_{Q^+}|\nabla w|^2\,dx}{\displaystyle\int_{Q^+} w^2\,dx}=4\pi^2
\end{equation}
because the first positive eigenvalue of the Neumann Laplacian in $Q^+$ is equal to $4\pi^2$. Using (\ref{IdentityEigenD}) and (\ref{Contra2}) in (\ref{Contra1}) leads to $\mu_2\ge2\pi$ which contradicts the initial assumption. Therefore $A^{\Omega}$ cannot have two eigenvalues below the continuous spectrum.\hfill $\square$
\section{Absence of threshold resonance}\label{sectionAbsenceofBoundedSol}
In this section, we establish Theorem \ref{ThmNoBoundedSol}. To proceed, we apply the tools of \cite{Pank17,BaNa21} that we recall now. For $R\ge1/2$ and $j=1,2,3$, define the truncated cylinder
\[
L^R_j := \{x\in L_j\,|\,|x_j|<R\}
\]
and set $\Omega^R:=L^R_1\cup L^R_2 \cup L^R_3$. In $\Omega^R$, consider the spectral problem with mixed boundary conditions
\begin{equation}\label{SpectralRMixed}
\begin{array}{|rcll}
-\Delta v & = & \mu v &\mbox { in }\Omega^R\\[2pt]
v & = & 0 &\mbox{ on }\partial\Omega^R\cap\partial\Omega\\[2pt]
\partial_n v & = & 0 &\mbox{ on }\partial\Omega^R\setminus\partial\Omega
\end{array}
\end{equation}
where $n$ stand for the unit normal vector to $\partial\Omega^R$ directed to the exterior of $\Omega^R$. Set $\mrm{H}^1_{0}(\Omega^R;\partial\Omega^R\cap\partial\Omega):=\{\varphi\in\mrm{H}^1(\Omega^R)\,|\,\varphi=0\mbox{ on }\partial\Omega^R\cap\partial\Omega\}$ and finally denote by $B^R$ the unbounded operator associated with (\ref{SpectralRMixed}) of domain $\mathcal{D}(B^R)\subset\mrm{H}^1_{0}(\Omega^R;\partial\Omega^R\cap\partial\Omega)$.\\
\newline
The Theorem 3 of \cite{Pank17} (see also the necessary and sufficient condition of \cite[Thm. 1]{BaNa21}) guarantees that there is no threshold resonant for Problem (\ref{PbSpectralZoom}) if there exists for some $R\ge0$ such that the second eigenvalue $\mu^R_2$ of $B_R$ satisfies $\mu^R_2>2\pi^2$.\\
\newline
Therefore from now our goal is to show that the second eigenvalue of $B_R$ is larger than $2\pi^2$ (Proposition \ref{PropoBoundedSecond} below). We start with a result of symmetry similar to Lemma \ref{LemmaSym}.
\begin{lemma}\label{LemmaSymR}
For $R$ large enough, if $v\in\mrm{H}^1_{0}(\Omega^R;\partial\Omega^R\cap\partial\Omega)$ is an eigenfunction of the operator $B^R$ associated with an eigenvalue $\mu\le2\pi^2$, then $v$ is symmetric with respect to the planes $(Ox_2x_3)$, $(Ox_3x_1)$, $(Ox_1x_2)$.
\end{lemma}
\begin{proof}
The demonstration follows the lines of the one of Lemma \ref{LemmaSym}. However we write the details for the sake of clarity. We focus our attention on the symmetry with respect to the plane $(Ox_2x_3)$, the two other ones being similar. Introduce the function $\varphi$ such that
\[
\varphi(x_1,x_2,x_3)=v(x_1,x_2,x_3)-v(-x_1,x_2,x_3).
\]
We wish to show that $\varphi\equiv0$. Set $\Omega^{R+}_1:=\{x\in\Omega^R\,|\,x_1>0\}$. We have
\begin{equation}\label{IdentityEigenR}
\int_{\Omega^{R+}_1}|\nabla\varphi|^2\,dx=\mu\int_{\Omega^{R+}_1}\varphi^2\,dx.
\end{equation}
Define the domains
\[
\begin{array}{rcl}
L^{R+}_1 & := & \{x\in L^R_1\,|\,x_1>1/2\}\\[2pt]
S^R_2 & := & \{x\in L^R_2\,|\,x_1>0\}=\Omega^{R+}_1\cap L^R_2\\[2pt]
S^R_3 & := & \{x\in L^R_3\,|\,x_1>0\}=\Omega^{R+}_1\cap L^R_3\\[2pt]
S^{R\pm}_2 & := & \{x\in S^R_2\,|\,\pm x_2>1/2\}\\[2pt]
S^{R\pm}_3 & := & \{x\in S^R_3\,|\,\pm x_3>1/2\}.
\end{array}
\]
In $L^{R+}_1$ the Poincar\'e inequality gives
\begin{equation}\label{PoincareTranche0R}
2\pi^2 \int_{L^{R+}_1}\varphi^2\,dx \le \int_{L^{R+}_1}|\nabla\varphi|^2\,dx.
\end{equation}
On the other hand, in $Q^+_1=(0;1/2)\times(-1/2;1/2)^2$ there holds
\begin{equation}\label{PoincareTranche0bisR}
\pi^2 \int_{Q^+_1}\varphi^2\,dx \le \int_{Q^+_1}(\partial_{x_1}\varphi)^2\,dx.
\end{equation}
The Lemma \ref{LemmaPoincareFriedrichsR} in Appendix (see also Lemma 5.1 of \cite{BaMaNa17}) guarantees that for $R>1/2$ and $a>0$, we have
\begin{equation}\label{PoincareModifiedR}
\kappa(a,R)\int_{0}^{1/2}\phi^2\,dt \le \int_{0}^{R}(\partial_t\phi)^2\,dt+a^2\int_{1/2}^{R}\phi^2\,dt,\qquad \forall\phi\in\mrm{H}^1(0;R),
\end{equation}
where $\kappa(a,R)>0$ converges to the constant $\kappa(a)$ appearing in (\ref{TransEqn}) as $R\to+\infty$. For $a=\pi\sqrt{5/2}$, as already said, one finds $\kappa(a)>\pi^2/2$. Introduce $\delta>0$ such that $\kappa(\pi\sqrt{5/2})>\pi^2/2+\delta$. We know that there is $R_0$ large enough such that we have $\kappa(\pi\sqrt{5/2},R)>\pi^2/2+\delta/2$ for all $R\ge R_0$. We infer that we have
\begin{equation}\label{PoincareTranche01R}
\cfrac{\pi^2+\delta}{2}\int_{Q^+_1}\varphi^2\,dx \le \int_{S^R_2}(\partial_{x_2}\varphi)^2\,dx+\cfrac{5\pi^2}{2}\int_{S_2^{R+}\cup S_2^{R-}}\varphi^2\,dx
\end{equation}
and
\begin{equation}\label{PoincareTranche02R}
\cfrac{\pi^2+\delta}{2}\int_{Q^+_1}\varphi^2\,dx \le \int_{S^R_3}(\partial_{x_3}\varphi)^2\,dx+\cfrac{5\pi^2}{2}\int_{S_3^{R+}\cup S_3^{R-}}\varphi^2\,dx.
\end{equation}
But from the Poincar\'e inequality in the transverse section of $S_2^{R\pm}$, $S_3^{R\pm}$, we know that
\begin{equation}\label{PoincareTrancheR}
5\pi^2\int_{S_2^{R\pm}}\varphi^2\,dx \le \int_{S_2^{R\pm}}(\partial_{x_1}\varphi)^2+(\partial_{x_3}\varphi)^2\,dx,\qquad 5\pi^2\int_{S_3^{R\pm}}\varphi^2\,dx \le \int_{S_3^{R\pm}}(\partial_{x_1}\varphi)^2+(\partial_{x_2}\varphi)^2\,dx.
\end{equation}
Inserting (\ref{PoincareTrancheR}) in (\ref{PoincareTranche01R}), (\ref{PoincareTranche02R}) and summing up the resulting estimates, we obtain
\begin{equation}\label{PoincareTranche1R}
\begin{array}{rcl}
(\pi^2+\delta)\displaystyle\int_{Q^+_1}\varphi^2\,dx &\le&\displaystyle \int_{S^R_2}(\partial_{x_2}\varphi)^2\,dx+\displaystyle \int_{S_3^R}(\partial_{x_3}\varphi)^2\,dx\\[10pt]
& &+\displaystyle\cfrac{1}{2}\int_{S_2^{R+}\cup S_2^{R-}}(\partial_{x_3}\varphi)^2\,dx+\cfrac{1}{2}\int_{S_3^{R+}\cup S_3^{R-}}(\partial_{x_2}\varphi)^2\,dx\\[10pt]
& & +\displaystyle\cfrac{1}{2}\int_{S_2^{R+}\cup S_2^{R-}\cup S_3^{R+}\cup S_3^{R-}}(\partial_{x_1}\varphi)^2\,dx.
\end{array}
\end{equation}
On the other hand, (\ref{PoincareTrancheR}) also yields
\begin{equation}\label{PoincareTranche2R}
\begin{array}{l}
\dsp2\pi^2\int_{S^{R+}_2\cup S^{R-}_2\cup S^{R+}_3\cup S^{R-}_3}\varphi^2\,dx \\[10pt]
\displaystyle\le \cfrac{1}{2} \int_{S^{R+}_2\cup S^{R-}_2\cup S^{R+}_3\cup S^{R-}_3}(\partial_{x_1}\varphi)^2\,dx+\cfrac{1}{2}\int_{S_2^{R+}\cup S_2^{R-}}(\partial_{x_3}\varphi)^2\,dx+\cfrac{1}{2}\int_{S_3^{R+}\cup S_3^{R-}}(\partial_{x_2}\varphi)^2\,dx.
\end{array}
\end{equation}
Finally, summing up (\ref{PoincareTranche0R}), (\ref{PoincareTranche0bisR}), (\ref{PoincareTranche1R}) and (\ref{PoincareTranche2R}), we get
\[
\delta\displaystyle\int_{Q^+_1}\varphi^2\,dx+2\pi^2 \int_{\Omega^{R+}_1}\varphi^2\,dx \le \int_{\Omega^{R+}_1}|\nabla\varphi|^2\,dx.
\]
This estimate together with the identity (\ref{IdentityEigenR}) imply $\int_{Q^+_1}\varphi^2\,dx=0$ and so $\varphi\equiv0$ in $Q^+_1$. From the unique continuation principle, this gives $\varphi\equiv0$ in $\Omega^{R+}_1$.
\end{proof}
\begin{proposition}\label{PropoBoundedSecond}
For $R$ large enough, the second eigenvalue $\mu_2^R$ of $B^R$ satisfies $\mu_2^R>2\pi^2$.
\end{proposition}
\begin{proof}
One applies Lemma \ref{LemmaSymR} to reduce the analysis to $\Omega^{R+}:=\{x\in\Omega^R\,|\,x_1>0,\,x_2>0,\,x_3>0\}$. Here in particular we use the assumption that $R$ is large enough. The rest of the proof is completely similar to the one of Theorem \ref{ThmDiscreteSpectrum}.
\end{proof}
\section{Model problems for the spectral bands}\label{SectionModels}
In this section, we establish Theorem \ref{MainThmPerio} and obtain models for the spectral bands $\Upsilon^\varepsilon_p$ appearing in (\ref{SpectralBands}). We recall that the spectrum of $A^\varepsilon$ is such that $\sigma(A^\varepsilon)=\bigcup_{p\in\mathbb{N}^\ast}\Upsilon^\varepsilon_p$.
\subsection{Study of $\Upsilon^\varepsilon_1$}
By definition, we have
\begin{equation}\label{Def_FirstBande}
\Upsilon^\varepsilon_1=\{\Lambda_1^\varepsilon(\eta),\,\eta\in[0;2\pi)^3\}
\end{equation}
where $\Lambda_1^\varepsilon(\eta)$ is the first eigenvalue of (\ref{PbSpectralCell}). Therefore our goal is to obtain an asymptotic expansion of $\Lambda_1^\varepsilon(\eta)$ with respect to $\varepsilon$ as $\varepsilon\to0^+$.\\
\newline
Pick $\eta\in[0;2\pi)^3$. Let $u^\varepsilon(\cdot,\eta)$ be an eigenfunction associated with $\Lambda_1^\varepsilon(\eta)$. As a first approximation when $\varepsilon\to0^+$, it is natural to consider the expansion
\begin{equation}\label{ExpansionTR0}
\Lambda_1^\varepsilon(\eta)=\varepsilon^{-2}\mu_1+\dots,\qquad u^\varepsilon(x,\eta)=v_1(x_1/\varepsilon)+\dots
\end{equation}
where $\mu_1\in(0;2\pi^2)$ stands for the eigenvalue of the discrete spectrum of the operator $A^{\Omega}$ introduced in Theorem \ref{ThmDiscreteSpectrum} and $v_1\in\mrm{H}^1_0(\Omega)$ is a corresponding eigenfunction that we choose such that $\|v_1\|_{\mrm{L}^2(\Omega)}=1$. Indeed, inserting $(\varepsilon^{-2}\mu_1,v_1(\cdot/\varepsilon))$ in Problem (\ref{PbSpectralCell}) only leaves a small discrepancy on the square faces of $\partial\omega^\varepsilon$ because $v_1$ is exponentially decaying at infinity. Let us write more precisely the decomposition of $v_1$ at infinity because this will be useful in the sequel. To proceed and keep short notation, we shall work with the coordinates $(z_j,y_j)$, $j=1,\dots,6$, such that
\begin{equation}\label{Def_coor_local}
\begin{array}{|l}
z_1=x_1,\quad y_1=(x_2,x_3);\\[3pt]
z_2=x_2,\quad y_2=(x_3,x_1);\\[3pt]
z_3=x_3,\quad y_3=(x_1,x_2);
\end{array}\qquad\qquad \begin{array}{|l}
z_4=-x_1,\quad y_4=(x_2,-x_3);\\[3pt]
z_5=-x_2,\quad y_5=(x_3,-x_1);\\[3pt]
z_6=-x_3,\quad y_6=(x_1,-x_2).
\end{array}
\end{equation}
We also define for $j=1,\dots,6$, the branch
\[
\mathcal{L}_j:=\{x\in\Omega\,|\,z_j>1/2\}.
\]
Then Fourier decomposition together with the result of symmetry of Lemma \ref{LemmaSym} and the fact that $\mu_1$ is a simple eigenvalue\footnote{This is needed to show that $K$ is the same in the branches $\mathcal{L}_j$, $j=1,\dots,6$.} guarantee that for $j=1,\dots,6$, we have
\[
v_1(x)=K\,e^{-\beta_1 z_j}\,U_{\dagger}(y_j)+O(e^{-\beta_2 z_j})\quad\mbox{ in }\mathcal{L}_j\quad\mbox{ as } z_j\to+\infty.
\]
Here $K\in\mathbb{R}^*$ is independent of $j$, $\beta_1:=\sqrt{2\pi^2-\mu_1}$, $\beta_2:=\sqrt{5\pi^2-\mu_1}$ and
\begin{equation}\label{defUdagger}
U_{\dagger}(s_1,s_2)=2\cos(\pi s_1)\cos(\pi s_2).
\end{equation}
The first model (\ref{ExpansionTR0}) is interesting but does not comprise the dependence with respect $\eta$. To improve it, consider the more refined ans\"atze
\begin{equation}\label{ExpansionTR1}
\Lambda_1^\varepsilon(\eta)=\varepsilon^{-2}\mu_1+\varepsilon^{-2}e^{-\beta_1/\varepsilon}M(\eta)+\dots,\qquad u^\varepsilon(x,\eta)=v_1(x_1/\varepsilon)+e^{-\beta_1/\varepsilon}V(x_1/\varepsilon,\eta)+\dots
\end{equation}
where the quantities $M(\eta)$, $V(\cdot,\eta)$ are to determine. Inserting (\ref{ExpansionTR1}) into (\ref{PbSpectralCell}), first we obtain that $V(\cdot,\eta)$ must satisfy
\begin{equation}\label{Pb_corrector_TR}
\begin{array}{|rccl}
-\Delta V(\cdot,\eta)-\mu_1V(\cdot,\eta)&=&M(\eta)v_1 & \mbox{ in }\Omega\\[3pt]
V(\cdot,\eta) &=&0 &\mbox{ on }\partial\Omega.
\end{array}
\end{equation}
In order to have a non zero solution to (\ref{Pb_corrector_TR}), we must look for a $V(\cdot,\eta)$ which is growing at infinity. Due to (\ref{Pb_corrector_TR}), the simplest growth that we can allow is
\[
V(x,\eta)=B_j\,e^{\beta_1 z_j}\,U_{\dagger}(y_j)+O(e^{-\beta_1 z_j})\quad\mbox{ in }\mathcal{L}_j\quad\mbox{ as } z_j\to+\infty
\]
where the $B_j$ are some constants. Then the quasi-periodic conditions satisfied by $u^\varepsilon(\cdot,\eta)$ at the faces located at $x_1=\pm1/2$, $x_2=\pm1/2$, $x_3=\pm1/2$ (see (\ref{PbSpectralCell})) lead us to choose the $B_j$ such that
\[
\begin{array}{|l}
K+B_4=e^{i\eta_1}(K+B_1)\\[2pt]
K-B_4=e^{i\eta_1}(-K+B_1)
\end{array}\qquad \begin{array}{|l}
K+B_5=e^{i\eta_2}(K+B_2)\\[2pt]
K-B_5=e^{i\eta_2}(-K+B_2)
\end{array}\qquad\begin{array}{|l}
K+B_6=e^{i\eta_3}(K+B_3)\\[2pt]
K-B_6=e^{i\eta_3}(-K+B_3).
\end{array}
\]
Solving these systems, we obtain $B_j=K\,e^{-i\eta_j}$, $B_{j+3}=K\,e^{+i\eta_j}$ for $j=1,2,3$. Now since $\mu_1$ is a simple eigenvalue of $A^\Omega$, multiplying (\ref{Pb_corrector_TR}) by $v_1$, integrating by part in $\Omega^R$ and taking the limit $R\to+\infty$, we find that there is a solution if and only if the following compatibility condition
\[
M(\eta)\|v_1\|_{\mrm{L}^2(\Omega)}=-2\beta_1K^2\sum_{j=1}^3e^{i\eta_j}+e^{-i\eta_j}\qquad\Leftrightarrow\qquad M(\eta)=-4\beta_1K^2\sum_{j=1}^3\cos(\eta_j)
\]
is satisfied. This defines the value of $M(\eta)$ in the expansion (\ref{ExpansionTR1}). From (\ref{Def_FirstBande}), this gives the inclusion $\Upsilon^\varepsilon_1\subset \varepsilon^{-2}\mu_1+\varepsilon^{-2}e^{-\sqrt{2\pi^2-\mu_1}/\varepsilon}(-c_1;c_1)$ for $\varepsilon$ small enough of Theorem \ref{MainThmPerio}. For $c_1$, we can take any constant larger than $12\beta_1K^2$. Note that we decided to focus on a rather formal presentation above for the sake of conciseness. We emphasize that all these results can be completely justified by proving rigorous error. This has been realized in detail in \cite{Naza17,Naza18} for similar problems and can be repeated with obvious modifications.
\subsection{Study $\Upsilon^\varepsilon_k$, $k\ge 2$}
We turn our attention to the asymptotic of the spectral bands of higher frequency
\begin{equation}\label{Def_FirstBande_2}
\Upsilon^\varepsilon_k=\{\Lambda_k^\varepsilon(\eta),\,\eta\in[0;2\pi)^3\},\qquad k\ge2,
\end{equation}
as $\varepsilon\to0^+$. Pick $\eta\in[0;2\pi)^3$ and introduce $u^\varepsilon(\cdot,\eta)$ an eigenfunction associated with $\Lambda_k^\varepsilon(\eta)$. In the sequel, to simplify, we remove the subscript ${}_k$ and do not indicate the dependence on $\eta$. As a first approximation when $\varepsilon\to0^+$, we consider the expansion
\begin{equation}\label{ExpansionTR0_2}
\Lambda^\varepsilon=\varepsilon^{-2}2\pi^2+\nu+\dots,\qquad u^\varepsilon(x)=v^\varepsilon(x)+\dots
\end{equation}
with $v^\varepsilon$ of the form
\[
v^\varepsilon(x)=\begin{array}{|rl}
\gamma^\pm_{1}(x_1)\,U_{\dagger}(x_2/\varepsilon,x_3/\varepsilon)&\mbox{ in }\omega^{\varepsilon\pm}_1:=\{x\in\omega^{\varepsilon}_1\,|\,\pm x_1>\varepsilon/2\},\\[3pt]
\gamma^\pm_{2}(x_2)\,U_{\dagger}(x_3/\varepsilon,x_1/\varepsilon)&\mbox{ in }\omega^{\varepsilon\pm}_2:=\{x\in\omega^{\varepsilon}_2\,|\,\pm x_2>\varepsilon/2\},\\[3pt]
\gamma^\pm_{3}(x_3)\,U_{\dagger}(x_1/\varepsilon,x_2/\varepsilon)&\mbox{ in }\omega^{\varepsilon\pm}_3:=\{x\in\omega^{\varepsilon}_3\,|\,\pm x_3>\varepsilon/2\},\\[3pt]
\end{array}
\]
where the functions $\gamma^\pm_{j}$, $j=1,2,3$, are to determine ($U_{\dagger}$ is defined in (\ref{defUdagger})). Inserting (\ref{ExpansionTR0_2}) into Problem (\ref{PbSpectralCell}), we obtain for $j=1,2,3$,
\begin{equation}\label{farfield}
\begin{array}{|rcl}
\partial^2_s\gamma^+_{j}+\nu\gamma^+_{j}&=&0\quad\mbox{ in }(0;1/2)\\[3pt]
\partial^2_s\gamma^-_{j}+\nu\gamma^-_{j}&=&0\quad\mbox{ in }(-1/2;0)\\[3pt]
\gamma^-_{j}(-1/2)&=&e^{i\eta_j}\gamma^+_{j}(+1/2)\\[3pt]
\partial_s\gamma^-_{j}(-1/2)&=&e^{i\eta_j}\partial_s\gamma^+_{j}(+1/2).
\end{array}
\end{equation}
To uniquely define the $\gamma^\pm_{j}$, we need to complement (\ref{farfield}) with some conditions at the origin. To proceed, we match the behaviour of the $\gamma^\pm_{j}$ with the one of some inner field expansion of $u^\varepsilon$. More precisely, in a neighbourhood of the origin we look for an expansion of $u^\varepsilon$ of the form
\begin{equation}\label{NearField}
u^\varepsilon(x)=W(x_1/\varepsilon)+\dots
\end{equation}
with $W$ to determine. Inserting (\ref{NearField}) and (\ref{ExpansionTR0_2}) in (\ref{PbSpectralCell}), we find that $W$ must satisfy
\begin{equation}\label{NearFieldPb}
\begin{array}{|rcll}
\Delta W+2\pi^2W&=&0&\mbox{ in }\Omega\\
W&=&0 &\mbox{ on }\partial\Omega.
\end{array}
\end{equation}
Now we come to the point where Theorem \ref{ThmNoBoundedSol} appears in the analysis. Indeed it guarantees that the only solution of (\ref{NearFieldPb}) which is bounded at infinity is the null function. Therefore we take $W\equiv0$ and impose
\begin{equation}\label{farfield_BC}
\gamma^\pm_{j}(0)=0.
\end{equation}
Then solving the spectral problem (\ref{farfield}), (\ref{farfield_BC}), we obtain
\begin{equation}\label{DefModel1D}
\nu=p^2\pi^2 \mbox{ for }p\in\mathbb{N}^\ast,\qquad \begin{array}{|l}
\gamma^+_{j}(s)=\sin(p\pi s)\\[2pt]
\gamma^-_{j}(s)=-e^{i\eta_j}\sin(p\pi s).
\end{array}
\end{equation}
Note that $\nu$ is a triple eigenvalue (geometric multiplicity equal to three) because the problems (\ref{farfield}), (\ref{farfield_BC}) for $j=1,2,3$ are uncoupled. Additionally $\nu$ is independent of $\eta\in[0;2\pi)^3$. This latter fact is not completely satisfactory and in the sequel we wish to improve the model obtained above. Let us refine the expansion proposed in (\ref{ExpansionTR0_2}) and work with
\begin{equation}\label{ExpansionTR0_3}
\Lambda^\varepsilon=\cfrac{2\pi^2}{\varepsilon^2}+p^2\pi^2+\varepsilon\tilde{\nu}+\dots,\quad u^\varepsilon(x)=\begin{array}{|rl}
(a_1\gamma^\pm_{1}(x_1)+\varepsilon\tilde{\gamma}^{\pm}_{1}(x_1))\,U_{\dagger}(x_2/\varepsilon,x_3/\varepsilon)&\mbox{ in }\omega^{\varepsilon\pm}_1\\[3pt]
(a_2\gamma^\pm_{2}(x_2)+\varepsilon\tilde{\gamma}^{\pm}_{2}(x_2))\,U_{\dagger}(x_3/\varepsilon,x_1/\varepsilon)&\mbox{ in }\omega^{\varepsilon\pm}_2\\[3pt]
(a_3\gamma^\pm_{3}(x_3)+\varepsilon\tilde{\gamma}^{\pm}_{3}(x_3))\,U_{\dagger}(x_1/\varepsilon,x_2/\varepsilon)&\mbox{ in }\omega^{\varepsilon\pm}_3\\[3pt]
\end{array}+\dots .
\end{equation}
Here $a:=(a_1,a_2,a_3)\in\mathbb{R}^3$, $\tilde{\nu}$ as well as the $\tilde{\gamma}^{\pm}_{j}$, $j=1,2,3$, are to determine. Since we are working with eigenfunctions, we can impose the normalization condition $a_1^2+a_2^2+a_3^2=1$. Inserting (\ref{ExpansionTR0_3}) into Problem (\ref{PbSpectralCell}) and extracting the terms in $\varepsilon$, we get for $j=1,2,3$,
\begin{equation}\label{farfield_3}
\begin{array}{|rcl}
\partial^2_s\tilde{\gamma}^+_{j}+p^2\pi^2\tilde{\gamma}^+_{j}&=&-\tilde{\nu}a_j\gamma^+_j\quad\mbox{ in }(0;1/2)\\[3pt]
\partial^2_s\tilde{\gamma}^-_{j}+p^2\pi^2\tilde{\gamma}^-_{j}&=&-\tilde{\nu}a_j\gamma^-_j\quad\mbox{ in }(-1/2;0)\\[3pt]
\tilde{\gamma}^-_{j}(-1/2)&=&e^{i\eta_j}\tilde{\gamma}^+_{j}(+1/2)\\[3pt]
\partial_s\tilde{\gamma}^-_{j}(-1/2)&=&e^{i\eta_j}\partial_s\tilde{\gamma}^+_{j}(+1/2).
\end{array}
\end{equation}
To define properly the $\tilde{\gamma}^\pm_{j}$, we need to add to (\ref{farfield_3}) conditions at the origin. To identify them, again we match with the behaviour of some inner field representation of $u^\varepsilon$. In a neighbourhood of the origin we look for an expansion of $u^\varepsilon$ of the form
\begin{equation}\label{NearField_3}
u^\varepsilon(x)=\varepsilon \tilde{W}(x_1/\varepsilon)+\dots.
\end{equation}
Inserting (\ref{NearField_3}) and (\ref{ExpansionTR0_3}) in (\ref{PbSpectralCell}), we find that $\tilde{W}$ must satisfy
\begin{equation}\label{NearFieldPb_3}
\begin{array}{|rcll}
\Delta \tilde{W}+2\pi^2\tilde{W}&=&0&\mbox{ in }\Omega\\
\tilde{W}&=&0 &\mbox{ on }\partial\Omega.
\end{array}
\end{equation}
This problem admits solutions $\tilde{W}_{j}$, $j=1,\dots,6$, with the expansions
\begin{equation}\label{DefPolMat1}
\tilde{W}_j(x)=
\begin{array}{|rl}
(z_j+M_{jj})\,U_{\dagger}(y_j)+\dots&\mbox{ in }\mathcal{L}_j\mbox{ as }z_j\to+\infty \\[2pt]
M_{jk}\,U_{\dagger}(y_k)+\dots&\mbox{ in }\mathcal{L}_k\mbox{ as }z_k\to+\infty,\quad k\ne j.
\end{array}
\end{equation}
Here we use the coordinates $(z_j,y_j)$ introduced in (\ref{Def_coor_local}). Note that at infinity $\tilde{W}_j$ is growing only in the branch $\mathcal{L}_j$. Let us explain how to show the existence of these functions (see \cite[Chap. 4, Prop. 4.13]{NaPl94} for more details). Introduce some $\psi\in\mathscr{C}^{\infty}(\mathbb{R})$ such that $\psi(s)=1$ for $s>1$ and $\psi(s)=0$ for $s\le 1/2$. Then for $j=1,\dots,6$, define $f_j$ such that $f_j(x)=(\Delta+2\pi^2)(\psi(z_j) \,z_jU_{\dagger}(y_j))$. Observe that $f_j$ is compactly supported. Then the theory of \cite[Chap. 5]{NaPl94} together with Theorem \ref{ThmNoBoundedSol} above ensure that there is a solution to the problem
\[
\begin{array}{|rcll}
\Delta \hat{W}_j+2\pi^2\hat{W}_j&=&-f_j&\mbox{ in }\Omega\\
\hat{W}_j&=&0 &\mbox{ on }\partial\Omega
\end{array}
\]
which is bounded at infinity. Finally we take $\tilde{W}_j$ such that $\tilde{W}_j(x)=\psi(z_j) \,z_jU_{\dagger}(y_j)+\hat{W}_j(x)$. The coefficients $M_{jk}$ form the so-called polarization matrix
\begin{equation}\label{def_Pola_matrix}
\mathbb{M}:=\left(M_{jk}\right)_{1\le j,k\le 6}
\end{equation}
which is real and symmetric even in non symmetric geometries (see \cite[Chap. 5, Prop. 4.13]{NaPl94}).\\
\newline
For the functions $\gamma^{\pm}_{j}$ in (\ref{DefModel1D}), we have the Taylor expansion, as $s\to0$,
\begin{equation}\label{ExpanFarField}
\gamma^+_{j}(s)=0+p\pi s+\dots=\varepsilon p\pi\,\cfrac{s}{\varepsilon}+\dots,\qquad \gamma^-_{j}(s)=0-e^{i\eta_j}p\pi s+\dots=-\varepsilon e^{i\eta_j} p\pi \,\cfrac{s}{\varepsilon}+\dots.
\end{equation}
Comparing (\ref{ExpanFarField}) with (\ref{DefPolMat1}) leads us to choose $\tilde{W}$ in the expansion $u^\varepsilon(x)=\varepsilon \tilde{W}(x_1/\varepsilon)+\dots$ (see (\ref{NearField_3})) such that
\[
\tilde{W}=p\pi \sum_{j=1}^3a_j\,(\tilde{W}_j+e^{i\eta_j}\tilde{W}_{3+j}).
\]
This sets the constant behaviour of $\tilde{W}$ at infinity and we now match the later with the behaviour of the $\tilde{\gamma}^\pm_{j}$ at the origin to close system (\ref{farfield_3}). This step leads us to impose
\begin{equation}\label{ConditionAtZero}
\begin{array}{ll}
\displaystyle\tilde{\gamma}^+_1(0)=p\pi \sum_{j=1}^3 a_j\,(M_{j1}+e^{i\eta_j}M_{3+j1}); &\quad\displaystyle \tilde{\gamma}^-_1(0)=p\pi \sum_{j=1}^3 a_j\,M_{j4}+e^{i\eta_j}M_{3+j4});\\[14pt]
\displaystyle\tilde{\gamma}^+_2(0)=p\pi \sum_{j=1}^3 a_j\,(M_{j2}+e^{i\eta_j}M_{3+j2}); &\quad\displaystyle \tilde{\gamma}^-_2(0)=p\pi \sum_{j=1}^3 a_j\,(M_{j5}+e^{i\eta_j}M_{3+j5});\\[14pt]
\displaystyle\tilde{\gamma}^+_3(0)=p\pi \sum_{j=1}^3 a_j\,(M_{j3}+e^{i\eta_j}M_{3+j3}); &\quad\displaystyle \tilde{\gamma}^-_3(0)=p\pi \sum_{j=1}^3 a_j\,(M_{j6}+e^{i\eta_j}M_{3+j6}).
\end{array}
\end{equation}
Equations (\ref{farfield_3}), (\ref{ConditionAtZero}) form a boundary value problem for a system of ordinary differential equations. For this problem, there is a kernel and co-cokernel. In order to have a solution, the following compatibility conditions, obtained by multiplying (\ref{farfield_3}) by $\overline{\gamma^\pm_j}$ and integrating by parts,
\[
\overline{\partial_s\gamma^+_j(0)}\,\tilde{\gamma}^+_j(0)-\overline{\partial_s\gamma^-_j(0)}\,\tilde{\gamma}^-_j(0)=-\tilde{\nu}a_j\left(\int_{-1/2}^{0}|\gamma^-_j|^2\,ds+\int_{0}^{1/2}|\gamma^+_j|^2\,ds\right)
\]
must be satisfied for $j=1,2,3$ (note that we used that $\gamma^{\pm}_j(0)=0$ according to (\ref{farfield_BC})). Since $\overline{\partial_s\gamma^+_j(0)}=p\pi$ and $\overline{\partial_s\gamma^-_j(0)}=-p\pi e^{-i\eta_j}$, this gives
\[
2(p\pi)^2\bigg(\sum_{k=1}^3 a_k\,(M_{kj}+e^{i\eta_k}M_{3+kj})+e^{-i\eta_j}\sum_{k=1}^3 a_k\,(M_{k3+j}+e^{i\eta_k}M_{3+k3+j})\bigg)=-\tilde{\nu}a_j.
\]
In a more compact form and by rewriting the dependence with respect to $\eta$, we obtain
\begin{equation}\label{CompactForm}
\mathbb{A}(\eta) a^{\top}=-\tilde{\nu}(\eta)a^{\top},
\end{equation}
with
\[
\mathbb{A}(\eta):=\Theta(\eta)\mathbb{M}\Theta^\ast(\eta)\in\mathbb{C}^{3\times 3},\quad
\Theta(\eta):=\left(\begin{array}{cccccc}
1 & 0 & 0 & e^{-i\eta_1} & 0 & 0\\[2pt]
0& 1 & 0 & 0 & e^{-i\eta_2} & 0\\[2pt]
0 & 0 & 1 & 0 & 0 & e^{-i\eta_3}
\end{array}\right),\quad\Theta^{\ast}(\eta)=\overline{\Theta(\eta)}^{\top}.
\]
Above we used that the polarization matrix $\mathbb{M}$ defined in (\ref{def_Pola_matrix}) is symmetric. By solving the spectral problem (\ref{CompactForm}), we get the values for $\tilde{\nu}(\eta)$ and $a$. Once $\tilde{\nu}(\eta)$ and $a$ are known, one can compute the solution to the system (\ref{farfield_3}), (\ref{ConditionAtZero}) to obtain the expressions of the $\tilde{\gamma}_j(\eta)$. This ends the definition of the terms appearing in the expansions (\ref{ExpansionTR0_3}). Let us exploit and comment these results.\\
\newline
$\star$ First, observe that the $\tilde{\nu}(\eta)$ are real. Indeed, since $\mathbb{M}$ is real and symmetric, we infer that $\mathbb{A}(\eta)$ is hermitian. \\
\newline
$\star$ We have obtained
\begin{equation}\label{ExpanConclu}
\Lambda^\varepsilon(\eta)=\varepsilon^{-2}2\pi^2+p^2\pi^2+\varepsilon\tilde{\nu}(\eta)+\dots\,.
\end{equation}
Note that in this expansion, the third term, contrary to the first two ones, depends on $\eta$. For $\eta\in[0;2\pi)^3$, denote by $\tilde{\nu}_1(\eta)$, $\tilde{\nu}_2(\eta)$, $\tilde{\nu}_3(\eta)$ the three eigenvalues of
(\ref{CompactForm}) numbered such that
\[
\tilde{\nu}_1(\eta) \le \tilde{\nu}_2(\eta) \le \tilde{\nu}_3(\eta).
\]
Define the quantity
\begin{equation}\label{defAleph}
\aleph:=\bigcup_{j=1}^3\{\tilde{\nu}_j(\eta),\,\eta\in[0;2\pi)^3\}.
\end{equation}
Since $\Upsilon^\varepsilon_{1+q+3(p-1)}=\{\Lambda_{1+q+3(p-1)}^\varepsilon(\eta),\,\eta\in[0;2\pi)^3\}$, this analysis shows that we have the inclusion
\begin{equation}\label{MainInclusion}
\Upsilon^\varepsilon_{1+q+3(p-1)}\subset \varepsilon^{-2}2\pi^2+p^2\pi^2+\varepsilon\,(-c_p;c_p)
\end{equation}
for $\varepsilon$ small enough as stated in Theorem \ref{MainThmPerio}. For $c_p$, we can take any value such that $\overline{\aleph}\subset (-c_p;c_p)$. \\
\newline
$\star$ Due to the symmetries of $\Omega$, $\mathbb{M}$ is of the form
\begin{equation}\label{MatPolSimple}
\left(\begin{array}{cccccc}
r_m & t^{\perp}_m & t^{\perp}_m & t_m & t^{\perp}_m & t^{\perp}_m \\[2pt]
t^{\perp}_m & r_m & t^{\perp}_m & t^{\perp}_m & t_m & t^{\perp}_m \\[2pt]
t^{\perp}_m & t^{\perp}_m & r_m & t^{\perp}_m & t^{\perp}_m & t_m \\[2pt]
t_m & t^{\perp}_m & t^{\perp}_m & r_m & t^{\perp}_m & t^{\perp}_m \\[2pt]
t^{\perp}_m & t_m & t^{\perp}_m & t^{\perp}_m & r_m & t^{\perp}_m \\[2pt]
t^{\perp}_m & t^{\perp}_m & t_m & t^{\perp}_m & t^{\perp}_m & r_m
\end{array}\right)
\end{equation}
where $r_m$, $t_m$, $t_m^\perp$ are real coefficients. For $\eta=(0,0,0)$, we find that the eigenvalues of (\ref{CompactForm}) are
\[
2r_m+2t_m+8t^{\perp}_m,\qquad 2r_m+2t_m-4t^{\perp}_m,\qquad 2r_m+2t_m-4t^{\perp}_m.
\]
For $\eta=(\pi,\pi,\pi)$, we obtain
\[
2r_m-2t_m,\qquad 2r_m-2t_m,\qquad 2r_m-2t_m.
\]
In the numerics of \S\ref{Section_PolMat}, we compute $\mathbb{M}$. From the values of $r_m$, $t_m$, $t^{\perp}_m$ obtained in (\ref{ValuePolMat}), we find for example
\[
2r_m+2t_m+8t^{\perp}_m \ne 2r_m-2t_m.
\]
This shows that the set $\aleph$ appearing in (\ref{defAleph}) has a non empty interior. Additionally, since $\nu_1(\pi)=\nu_2(\pi)=\nu_3(\pi)$ and the $\nu_j$ depend continuously on $\eta$, we deduce that $\aleph$ is a connected segment. This is obtained in the numerics of \S\ref{Section_PolMat} (see (\ref{NumAleph})) and due to the symmetries of $\omega^\varepsilon$, this was somehow expected.\\
\newline
Finally, let us mention again that all the formal presentation above can be justified rigorously by a direct adaptation of the proofs of error estimates presented in \cite{Naza17,Naza18}.
\section{Numerics}\label{SectionNumerics}
\subsection{Spectrum of $A^{\Omega}$}
We start the numerics by computing the spectrum of the operator $A^{\Omega}$ defined after (\ref{PbSpectralZoom}). More precisely, in order to approximate also the eigenvalues which are embedded in the continuous spectrum of $A^{\Omega}$ and to reveal complex resonances which would be located close to the real axis, we work with Perfectly Matched Layers \cite{Bera94,KiPa10,BoCP18} (see also the techniques of analytic dilatations \cite{aguilar1971class,balslev1971spectral,moiseyev1998quantum}). For $\theta\in(0;\pi/2)$ and $L>1/2$, define the complex valued parameters
\[
\alpha^\theta_1=\begin {array}{|ll}
1 & \mbox{for }|x_1|\le L\\[2pt]
e^{-i\theta} & \mbox{for }|x_1|>L
\end{array},\qquad \alpha^\theta_2=\begin {array}{|ll}
1 & \mbox{for }|x_2|\le L\\[2pt]
e^{-i\theta} & \mbox{for }|x_2|>L
\end{array},\qquad \alpha^\theta_3=\begin {array}{|ll}
1 & \mbox{for }|x_3|\le L\\[2pt]
e^{-i\theta} & \mbox{for }|x_3|>L.
\end{array}
\]
Here the coefficient $\theta$ will drive the rotation of the continuous spectrum while $L$ marks the beginning of the PML region. Then consider the spectral problem
\begin{equation}\label{ProblemPMLs}
\begin{array}{|rcll}
-\alpha^\theta_1\cfrac{\partial}{\partial x_1}\Big(\alpha^\theta_1\cfrac{\partial u}{\partial x_1}\,\Big)+\alpha^\theta_2\cfrac{\partial}{\partial x_2}\Big(\alpha^\theta_2\cfrac{\partial u}{\partial x_2}\,\Big)+\alpha^\theta_3\cfrac{\partial}{\partial x_3}\Big(\alpha^\theta_3\cfrac{\partial u}{\partial x_3}\,\Big)&=&\Lambda u &\mbox{ in } \Omega\\[3pt]
u&=&0&\mbox{ on } \partial\Omega.
\end{array}
\end{equation}
Denote by $A^\Omega_\theta$ the unbounded operator associated with (\ref{ProblemPMLs}). Observe that $A^\Omega_\theta$ is not selfadjoint due to the complex parameters $\alpha^\theta_j$. However the theory guarantees that the real eigenvalues of $A^\Omega_\theta$ coincide exactly with the real eigenvalues of $A^{\Omega}$. What is interesting is that one can show that the essential spectrum of $A^\Omega_\theta$, that is the values of $\Lambda$ such that $A^\Omega_\theta-\Lambda\,\mrm{Id}:\mathcal{D}(A^\Omega_\theta)\to\mrm{L}^2(\Omega)$ is not Fredholm, corresponds to the set
\[
\bigcup_{m,n\in\mathbb{N}^\ast}\{\pi^2(m^2+n^2)+t\,e^{-2i\theta},\,t\ge0\},
\]
so that the real eigenvalues of $A^\Omega_\theta$ are isolated in the spectrum. As a consequence, we can compute them by truncating the branches of $\Omega$ at a certain distance without producing spectral pollution. Then we approximate the spectrum in this bounded geometry by using a classical P1 finite element method. We construct the matrices with the library \texttt{Freefem++} \cite{Hech12} and compute the spectrum with \texttt{Matlab}\footnote{\textit{Matlab}, \url{http://www.mathworks.com/}.}.\\
\newline
In Figure \ref{FigSpectrum}, we display in the complex plane the approximation of the spectrum of $A^\Omega_\theta$ obtained with this approach. We observe that the branches of essential spectrum of $A^\Omega_\theta$ are somehow discretized. This is due to the fact that the approximated problem is set in finite dimension. Moreover, we note that $A^\Omega_\theta$, and so $A^\Omega$, have eigenvalues on the real line. In accordance with Theorem \ref{ThmDiscreteSpectrum}, we find exactly one eigenvalue $\mu_1\approx 12.9\approx1.3\pi^2$ on the segment $(0;\pi^2)$. We also note the presence of eigenvalues embedded in the continuous spectrum for the operator $A^{\Omega}$. For the first one, we get $\mu_2\approx 46.7\approx4.7\pi^2$.
\begin{figure}[!ht]
\centering
\includegraphics[width=10cm]{Spectrum.pdf}
\caption{Approximation of the spectrum of $A^\Omega_\theta$ for $\theta=\pi/4$ ($L=1/2$).}
\label{FigSpectrum}
\end{figure}
\noindent In Figure \ref{FigEigenfunctions}, we represent an eigenfunction (trapped mode) associated with the eigenvalue $\mu_1$ of the discrete spectrum of $A^{\Omega}$. As guaranteed by Lemma \ref{LemmaSym}, we observe that it is indeed symmetric with respect to the planes $(Ox_2x_3)$, $(Ox_3x_1)$, $(Ox_1x_2)$.
\begin{figure}[!ht]
\centering
\includegraphics[width=5.8cm]{TMx0}\qquad\quad\includegraphics[width=5.8cm]{TMy0}
\caption{Eigenfunction associated with the eigenvalue $\mu_1$: cuts $x_1=0$ (left) and $x_2=0$ (right).}
\label{FigEigenfunctions}
\end{figure}
\subsection{Polarization matrix and threshold scattering matrix for the operator $A^\Omega$}\label{Section_PolMat}
In this section, we explain how to compute the polarization matrix $\mathbb{M}$ introduced after (\ref{CompactForm}) and whose properties allow one to assess the second corrector term in the expansion $\Lambda^\varepsilon=\varepsilon^{-2}2\pi^2+p^2\pi^2+\varepsilon\tilde{\nu}+\dots$ of the eigenvalues generating the spectral bands $\Upsilon^\varepsilon_k$, $k\ge 2$. We work in the geometry $\Omega$ defined in (\ref{DefOmega}) and study the problem (\ref{PbSpectralZoom}) at the threshold, namely
\begin{equation}\label{PbSpiderThreshold}
\begin{array}{|rcll}
\Delta v +2\pi^2v&=& 0 &\mbox{on }\Omega\\
v&=&0 &\mbox{on }\partial\Omega.
\end{array}
\end{equation}
To obtain $\mathbb{M}$, we will first compute the so-called threshold scattering matrix that we define now. In $\mathcal{L}_j$, $j=1,\dots,6$, set
\[
w_{j}^{\pm}(x)=\cfrac{z_j \mp i}{\sqrt{2}}\,U_{\dagger}(y_j).
\]
Let us work again with the function $\psi\in\mathscr{C}^{\infty}(\mathbb{R})$ introduced after (\ref{DefPolMat1}) such that $\psi(s)=1$ for $s>1$ and $\psi(s)=0$ for $s\le 1/2$. For $j=1,\dots,6$, define $\psi_j$ such that $\psi_j(x)=\psi(z_j,y_j)$ (observe that $\psi_j$ is non zero only in the branch $\mathcal{L}_j$). For $j=1,\dots,6$, the theory of \cite[Chap. 5]{NaPl94} guarantees that problem (\ref{PbSpiderThreshold}) admits a solution with the decomposition
\begin{equation}\label{decompo_vj}
v_j=\psi_j\,w_j^{-}+\sum_{k=1}^6\psi_k\,s_{jk}\,w_k^++\tilde{u}_j,
\end{equation}
where the $s_{jk}$ are complex numbers and the $\tilde{u}_j$ decay exponentially at infinity. The matrix
\[
\mathbb{S}:=\left(s_{jk}\right)_{1\le j,k\le 6}
\]
is called the threshold scattering matrix. It is symmetric ($\mathbb{S}=\mathbb{S}^{\top}$) but not necessarily hermitian and unitary ($\mathbb{S}\,\overline{\mathbb{S}}^{\top}=\mrm{Id}$). It is known (see relation (7.9) in \cite{Naza16}) that $\mathbb{M}$ coincides with the Cayley transform of $\mathbb{S}$, i.e. we have
\begin{equation}\label{DefMatPol}
\mathbb{M}=i(\mrm{Id}+\mathbb{S})^{-1}(\mrm{Id}-\mathbb{S}).
\end{equation}
Note that one can show that we have $\dim\,\ker\,(\mrm{Id}+\mathbb{S})=\dim\,(\mathscr{B}/\mathscr{B}_{\mrm{tr}})$ (the quotient space) where $\mathscr{B}$ denotes the space of bounded solutions of (\ref{PbSpiderThreshold}) and $\mathscr{B}_{\mrm{tr}}$ the space of trapped modes of (\ref{PbSpiderThreshold}). For the proof, we refer the reader for example to Theorem 1 in \cite{Naza14}. Since Theorem \ref{ThmNoBoundedSol} ensures that $\mathscr{B}$ reduces to the null function, we infer that $\mrm{Id}+\mathbb{S}$ is invertible which guarantees that $\mathbb{M}$ is well defined via formula (\ref{DefMatPol}). Additionally, due to the symmetries of $\Omega$, $\mathbb{S}$ is of the form
\begin{equation}\label{DefMatScaSimplifiee}
\left(\begin{array}{cccccc}
r & t^{\perp} & t^{\perp} & t & t^{\perp} & t^{\perp} \\[2pt]
t^{\perp} & r & t^{\perp} & t^{\perp} & t & t^{\perp} \\[2pt]
t^{\perp} & t^{\perp} & r & t^{\perp} & t^{\perp} & t \\[2pt]
t & t^{\perp} & t^{\perp} & r & t^{\perp} & t^{\perp} \\[2pt]
t^{\perp} & t & t^{\perp} & t^{\perp} & r & t^{\perp} \\[2pt]
t^{\perp} & t^{\perp} & t & t^{\perp} & t^{\perp} & r
\end{array}\right)
\end{equation}
where $r$, $t$, $t^{\perp}$ are complex reflection and transmission coefficients. Therefore it is sufficient to compute $v_1$. To proceed, we shall work in the bounded domain $\Omega^R$ (see before (\ref{SpectralRMixed})) and impose approximated radiation conditions on the artificial cuts. Denote by $n$ the unit normal to $\partial\Omega^R$ directed to the exterior of $\Omega^R$, set
\[
\begin{array}{l}
\Gamma_1:=\{R\}\times(-1/2;1/2)^2,\quad \Gamma_2:=(-1/2;1/2)\times\{R\}\times(-1/2;1/2),\quad \Gamma_3:=(-1/2;1/2)^2\times\{R\}\\[4pt]
\Gamma_4:=\{-R\}\times(-1/2;1/2)^2,\ \Gamma_5:=(-1/2;1/2)\times\{-R\}\times(-1/2;1/2),\ \Gamma_6:=(-1/2;1/2)^2\times\{-R\}
\end{array}
\]
and $\Gamma:=\cup_{j=1}^6\Gamma_j$. On $\Gamma_1$, according to (\ref{decompo_vj}), we have
\[
\partial_n(v_1-\psi_1w_{1}^{-})=\partial_{z_1}(v_1-w_{1}^{-})=2^{-1/2}r\,U_{\dagger}+\dots
\]
where the dots stand for terms which are small for large values of $R$. On the other hand on $\Gamma_1$, there holds
\[
v_1-w_{1}^{-}=2^{-1/2}r\,(R-i)\,U_{\dagger}+\dots .
\]
Therefore, this gives, still on $\Gamma_1$,
\begin{equation}\label{Approx_RC1}
\partial_n v_1=\cfrac{v_1-w_{1}^{-}}{R-i}+\partial_{z_1} w_{1}^{-}+\dots=\cfrac{v_1}{R-i}+w_{1}^{-}\left(\cfrac{1}{R+i}-\cfrac{1}{R-i}\right)+\dots=\cfrac{v_1}{R-i}-\cfrac{2i}{R^2+1}\,w_{1}^{-}+\dots.
\end{equation}
On $\Gamma_j$, $j\ne1$, the situation is simpler because $v_1$ is outgoing in the corresponding branches and we have
\begin{equation}\label{Approx_RC2}
\partial_n v_1=\cfrac{v_1}{R-i}+\dots .
\end{equation}
Finally, using the Robin conditions (\ref{Approx_RC1}), (\ref{Approx_RC2}) as approximated radiation conditions, we consider the variational formulation
\begin{equation}\label{Pb_Approximation}
\begin{array}{|l}
\mbox{Find }\hat{v}_1\in\mrm{H}^1_{0}(\Omega^R;\partial\Omega^R\cap\partial\Omega)\mbox{ such that for all }v\in\mrm{H}^1_{0}(\Omega^R;\partial\Omega^R\cap\partial\Omega) \\[6pt]
\displaystyle\int_{\Omega^R}\nabla \hat{v}_1\cdot\nabla v\,dx-\cfrac{1}{R-i}\int_{\Gamma} \hat{v}_1 v\,d\sigma-(2\pi)^2\displaystyle\int_{\Omega^R}\hat{v}_1 v\,dx=-\cfrac{2i}{R^2+1}\int_{\Gamma_1} w_{1}^{-} v\,d\sigma.
\end{array}
\end{equation}
One can prove that $\hat{v}_1$ yields a good approximation of $v_1$ with an error which is exponentially decaying with $R$. In practice, we solve the problem (\ref{Pb_Approximation}) with a P1 finite element method thanks to \texttt{Freefem++}. Then replacing $v_1$ by $\hat{v}_1$ in the exact formulas
\begin{equation}\label{FormulaApproxCoef}
r=\cfrac{2}{R^2+1}\,\int_{\Gamma_1} (v_1-w_{1}^{-})\,w_{1}^{-}\,d\sigma,\quad t=\cfrac{2}{R^2+1}\,\int_{\Gamma_4} v_1\,w_4^{-}\,d\sigma,\quad t^\perp=\cfrac{2}{R^2+1}\,\int_{\Gamma_2} v_1\,w_2^{-}\,d\sigma,
\end{equation}
we get an approximation of the threshold scattering matrix $\mathbb{S}$ given by (\ref{DefMatScaSimplifiee}). Finally with (\ref{DefMatPol}), we obtain an approximation of the polarization matrix $\mathbb{M}$ which appears in the $3\times3$ spectral problems (\ref{CompactForm}).\\
\newline
Our computations give
\[
r\approx 0.66+0.11i,\qquad t\approx 0.08-0.70i,\qquad t^{\perp}\approx -0.08-0.08i.
\]
The eigenvalues of $\mathbb{S}\in\mathbb{C}^{6\times6}$ are approximately equal to
\[
0.44 - 0.9i\mbox{ (simple)},\quad 0.9 - 0.44i\mbox{ (double)}\quad\mbox{ and }\quad 0.58 + 0.81i\mbox{ (triple).}
\]
They have modulus one which is consistent with the fact that $\mathbb{S}$ is unitary. Moreover, we indeed observe that they are different from $-1$ which is coherent with the discussion following (\ref{DefMatPol}) (absence of threshold resonance). On the other hand, for the coefficients of $\mathbb{M}$ (see (\ref{MatPolSimple})), we get
\begin{equation}\label{ValuePolMat}
r_m\approx 0.08,\qquad t_m\approx -0.44,\qquad t^{\perp}_m\approx -0.06.
\end{equation}
With these values, solving the $3\times3$ eigenvalue problem (\ref{CompactForm}) for $\eta\in[0;2\pi)^3$, we find that the segment $\aleph$ defined in (\ref{defAleph}) satisfies
\begin{equation}\label{NumAleph}
\aleph\approx (-1.24;1.04).
\end{equation}
\noindent Let us mention that the Robin conditions (\ref{Approx_RC1}), (\ref{Approx_RC2}) are rather crude approximations of the exact radiation conditions. To get good errors estimates, we should take rather large values of $R$. However in practice, large $R$ are no so simple to handle and can create important numerical errors. Therefore a compromise must be found and we take $R=2.5$. Admittedly, this point should more investigated.
\section*{Apprendix}
\subsection{Friedrichs inequality}
We reproduce here the Lemma 5.1 of \cite{BaMaNa17}.
\begin{lemma}\label{LemmaPoincareFriedrichs}
Assume that $a>0$. Then we have the Friedrichs inequality
\begin{equation}\label{Fried_ineq}
\kappa(a)\int_{0}^{1/2}\phi^2\,dt \le \int_{0}^{+\infty}(\partial_t\phi)^2\,dt+a^2\int_{1/2}^{+\infty}\phi^2\,dt,\qquad \forall \phi\in\mrm{H}^1(0;+\infty),
\end{equation}
where $\kappa(a)$ is the smallest positive root of the transcendental equation
\begin{equation}\label{TransEqnLemma}
\sqrt{\kappa}\tan\bigg(\cfrac{\sqrt{\kappa}}{2}\bigg)=a.
\end{equation}
\end{lemma}
\begin{proof}
For $a>0$, consider the spectral problem
\begin{equation}\label{Eigen1D}
\begin{array}{|rcll}
-\partial^2_t \phi+a^2\mathbbm{1}_{(1/2;+\infty)} \phi &=& \lambda(a)\,\mathbbm{1}_{(0;1/2)}\phi&\mbox{ in }(0;+\infty)\\[4pt]
\partial_t\phi(0)&=&0
\end{array}
\end{equation}
where $\mathbbm{1}_{(1/2;+\infty)}$, $\mathbbm{1}_{(0;1/2)}$ stand for the indicator functions of the sets $(1/2;+\infty)$, $(0;1/2)$ respectively. Let us equip $\mrm{H}^1(0;+\infty)$ with the inner product
\[
(\phi,\phi')_a=\int_{0}^{+\infty}\partial_t\phi\,\partial_t\phi'\,dt+a^2\phi\,\phi'\,dt.
\]
With the Riesz representation theorem, define the linear and continuous operator $T:\mrm{H}^1(0;+\infty)\to \mrm{H}^1(0;+\infty)$ such that
\[
(T\phi,\phi')_a=\int_{0}^{1/2}\phi\,\phi'\,dt.
\]
With this definition, we find that $(\lambda(a),\phi)$ is an eigenpair of (\ref{Eigen1D}) if and only if we have
\[
T\phi=(\lambda(a)+a^2)^{-1}\phi.
\]
Since $T$ is bounded and symmetric, it is self-adjoint. Additionally the Rellich theorem ensures that $T$ is compact. This guarantees that the spectrum of (\ref{Eigen1D}) coincides with a sequence of positive eigenvalues whose only accumulation point is $+\infty$. Let us denote by $\kappa(a)$ the smallest eigenvalue of (\ref{Eigen1D}). From classical results concerning compact self-adjoint (see e.g. \cite[Thm. 2.7.2]{BiSo87}), we know that
\begin{equation}\label{Def_quotient}
(\kappa(a)+a^2)^{-1}=\underset{\phi\in\mrm{H}^1(0;+\infty),\,(\phi,\phi)_a=1}{\sup} (T\phi,\phi)_a.
\end{equation}
Rearranging the terms, we find that (\ref{Def_quotient}) provides the desired estimates (\ref{Fried_ineq}). Now we compute $\kappa(a)$. Solving the ordinary differential equation (\ref{Eigen1D}) with $\lambda(a)=\kappa(a)$, we obtain, up to a multiplicative constant,
\[
\phi(t)=\begin{array}{|ll}
\cos(\sqrt{\kappa(a)}t) & \mbox{ for }t\in(0;1/2)\\[3pt]
c\,e^{-at} & \mbox{ for }t\ge 1/2
\end{array}
\]
where $c$ is a constant to determine. Writing the transmission conditions at $t=1/2$, we find that a non zero solution exists if and only if $\kappa(a)>0$ satisfies the relation (\ref{TransEqnLemma}).
\end{proof}
\begin{lemma}\label{LemmaPoincareFriedrichsR}
Assume that $a>0$ and $R>1/2$. Then we have the Friedrichs inequality
\begin{equation}\label{Fried_ineq_R}
\kappa(a,R)\int_{0}^{1/2}\phi^2\,dt \le \int_{0}^{R}(\partial_t\phi)^2\,dt+a^2\int_{1/2}^{R}\phi^2\,dt,\qquad \forall \phi\in\mrm{H}^1(0;R),
\end{equation}
where $\kappa(a,R)$ is the smallest positive root of the transcendental equation
\begin{equation}\label{TransEqnLemmaR}
\sqrt{\kappa}\tan\bigg(\cfrac{\sqrt{\kappa}}{2}\bigg)=a\,\tanh(a(R-1/2)).
\end{equation}
Therefore, we have $\lim_{R\to+\infty}\kappa(a,R)=\kappa(a)$ where $\kappa(a)$ is the constant appearing in Lemma \ref{LemmaPoincareFriedrichs}.
\end{lemma}
\begin{proof}
The demonstration is completely similar to the one of Lemma \ref{LemmaPoincareFriedrichs} above. We find that the largest constant $ \kappa(a,R)$ such that (\ref{Fried_ineq_R}) holds coincides with the smallest eigenvalue of the problem
\[
\begin{array}{|rcll}
-\partial^2_t \phi+a^2\mathbbm{1}_{(1/2;R)} \phi &=& \kappa(a,R)\mathbbm{1}_{(0;1/2)}\phi&\mbox{ in }(0;R)\\[4pt]
\partial_t\phi(0)=\partial_t\phi(R)&=&0.
\end{array}
\]
Solving it, we find that if $\phi$ is a corresponding eigenfunction, up to a multiplicative constant, we have
\[
\phi(t)=\begin{array}{|ll}
\cos(\sqrt{\kappa(a,R)}t) & \mbox{ for }t\in(0;1/2)\\[3pt]
c\,\cosh(a(t-R)) & \mbox{ for }t\in(1/2;R)
\end{array}
\]
for some constant $c$. This times, writing the transmission conditions at $t=1/2$, we find that a non zero solution exists when $\kappa(a,R)$ satisfies (\ref{TransEqnLemmaR}). Finally we obtain that $\kappa(a,R)\to\kappa(a)$ because $\tanh(a(R-1/2))$ in (\ref{TransEqnLemmaR}) tends to $1$ when $R\to+\infty$.
\end{proof}
\section{Acknowledgements}
The work of the second author was supported by the Russian Science Foundation, project 22-11-00046.
|
\section{Introduction}
\hspace{.2 in} Asymmetric, positive data occur widely in real-world applications. Examples documented in the fields of health, economics, and agriculture include the body mass index (\cite{pfeffermann2007small}), poverty-related measurements (\cite{molina2010small}), and sheet and rill erosion (\cite{berg2014small}). These types of data are often used to gain a deeper understanding of the characteristics of sub-populations (sub-domains) defined by geographic regions or socio-demographic groups. Such subdivisions are usually more granular than planned estimation domains and therefore have sample sizes that are small or even zero. This motivates the definition of a small area (domain) as any sub-population where the area-specific data are insufficient to assure direct domain estimates of acceptable precision. Estimation procedures for small domains commonly employ indirect estimators based on small area models that incorporate between-area variation and auxiliary variables \citep{rao2015small, jiang2006mixed, pfeffermann2013new, morales2020small}. A fundamental small area model is the unit-level linear mixed model of \cite{battese1988error}. This model assumes normal distributions and is not immediately suitable for positive, skewed data.
\hspace{.2 in } A common approach for skewed data is to apply the unit-level linear mixed model of \cite{battese1988error}, after an appropriate transformation. In the framework of a unit-level lognormal model, \cite{berg2014small} develop closed-form expressions for an empirical Bayes predictor of a small area mean. \cite{lyu2020empirical} and \cite{zimmermann2018small} extend the lognormal model to zero-inflated data and informative sampling, respectively. \cite{berg2014small}, \cite{lyu2020empirical}, and \cite{zimmermann2018small} focus on prediction of means, but many small area parameters are more complex functions of the model response variable. \cite{molina2010small} obtain a Monte Carlo approximation for the best predictor of a general parameter, assuming that transformed study variables follow the nested error linear regression model. \cite{guadarrama2018small} develops predictors of general parameters under an informative sample design. \cite{rojas2020data} extend \cite{molina2010small} to data-driven transformations that are more general than the log transformation. The use of a transformation (such as a log transformation) is often undesirable. The transformation of the data to a normal distribution can be artificial. In addition, the interpretation of results based on transformed data is difficult.
\hspace{.2 in} An alternative to a transformation is to model study variable directly. The gamma distribution enables the analyst to model skewed data in the original scale, without need for a transformation. \cite{hobza2020small} compares several small area predictors, developed under a generalized linear mixed model (GLMM) with a gamma response distribution. A challenge with the gamma GLMM is that the likelihood involves intractable integrals. \cite{hobza2020small} estimate the model parameters by maximizing the Laplace approximation to the log-likelihood, but this procedure may perform poorly if the gamma distribution differs from a normal distribution. \cite{dreassi2014small} use Bayesian inference procedures to construct small area estimates under the assumptions of a gamma GLMM. \cite{berg2016small} compare predictors based on lognormal and gamma distributions through simulation, and find that the predictors based on the gamma distribution are more robust to model misspecification. \cite{Graf2019} develop empirical best small area predictors of both means and more general parameters under the assumptions of a generalized gamma inverse-gamma distribution. Unlike the gamma GLMM, the model of \cite{Graf2019} leads to tractable integrals and predictors with closed-form expressions. The works of \cite{hobza2020small} and \cite{Graf2019} provide the impetus for the research in this paper.
\hspace{.2 in} We study small area predictors based on gamma distributions. We focus on a unit-level model with a gamma response distribution and gamma distributed random effects, which we call the ``gamma-gamma'' model. The gamma-gamma model is a special case of the more general model of \cite{Graf2019}, with a slightly different parametrization. We compare predictors based on the gamma-gamma model to the predictors based on the gamma GLMM of \cite{hobza2020small}. We develop predictors for the gamma-gamma model in the context of an informative sample design. Our approach to informative sampling transfers the fundamental concepts of \cite{pfeffermann2007small} to the gamma-gamma framework.
\hspace{.2 in } Although the models in this paper are not new, our work has several important contributions. First, the extension of the gamma-gamma model to informative sampling is the most substantive contribution because \cite{Graf2019} only consider noninformative designs. Second, we formally compare predictors based on the gamma-gamma model to predictors based on the gamma GLMM through simulation. \cite{Graf2019} only compare their model to a lognormal model, and \cite{hobza2020small} exclusively consider the gamma GLMM. Our third contribution is in the area of MSE estimation. We evaluate the properties of several MSE estimators for the gamma-gamma model through simulation. We also propose an MSE estimator that has not yet been used in combination with the gamma-gamma model. \cite{Graf2019} propose an MSE estimator but do not evaluate its properties through simulation. Our final two contributions are relatively minor but are valuable nonetheless. We develop predictors for the gamma-gamma model using a hierarchical formulation that is computationally easier to implement than the formulation based on marginal distributions in \cite{Graf2019}. Finally, we generalize the procedures of \cite{hobza2020small} slightly to prediction of parameters that are more general than the class of additive parameters. \cite{hobza2020small} define predictors for additive parameters of the form $N_{i}^{-1}\sum_{j = 1}^{N_{i}} \tilde{h}(y_{ij})$, where $y_{i1},\ldots, y_{i N_{i}}$ denote the variables of interest for the $N_{i}$ elements of the population for area $i$. We define predictors for more general parameters of the form $h(y_{i1},\ldots, y_{i N_{i}})$. An important type of non-additive parameter that we consider in our study is the population quantile.
The procedures discussed in this paper are relevant to studies of sheet and rill erosion, or soil loss due to the flow of water. Small area estimates of sheet and rill erosion are valuable for assessing the efficacy of conservation programs. Sheet and rill erosion is positive, and past studies have documented the distribution of sheet and rill erosion to be skewed right. We apply the methods detailed in this paper to construct small area estimates of sheet and rill erosion using data from a national survey of cropland in the United States.
\hspace{.2 in} Our study of small area prediction based on gamma distributions is organized as follows. In Section 2, we define the gamma-gamma model and the gamma GLMM. In Section 3, we propose two MSE estimators for the gamma-gamma model. The first MSE estimator differs from the MSE estimator of \cite{Graf2019}. The second applies a bootstrap bias correction to the MSE estimator of \cite{Graf2019}. In Section 4, we extend the gamma-gamma model to an informative sample design. In Section 5, we present three simulation studies that (1) compare the gamma-gamma model to the gamma GLMM, (2) evaluate the alternative MSE estimators, and (3) study the properties of the predictors under varying degrees of informativeness. The data analysis is presented in Section 6. We summarize the main conclusions in Section 7.
\section{Small Area Estimation based on Gamma Distributions}
\hspace{.2 in} We establish a common notation that we will use for both the gamma-gamma model and the gamma GLMM. Let $i = 1,\ldots, D$ denote the areas, and let $j= 1,\ldots, N_{i}$ index the elements in the population for area $i$. Let $y_{ij}$ denote the response variable for unit $j$ in area $i$, where the support of $y_{ij}$ is $(0,\infty)$. Let $\bm{x}_{ij}$ denote the covariate associated to element $(i,j)$. Assume that $y_{ij}$ is observed for a sample of $n_{i}$ elements in area $i$. Without loss of generality, let $j = 1,\ldots, n_{i}$ index the sampled elements, and let $j = n_{i} + 1,\ldots, N_{i}$ index the non-sampled elements. We let $\bm{y}_{i} = (\bm{y}_{is}',\bm{y}_{ir}')'$, where $\bm{y}_{is} = (y_{i1},\ldots, y_{i n_{i}})'$, and $\bm{y}_{ir} = (y_{i n_{i}+1},\ldots, y_{i N_{i}})'$. Assume the covariate $\bm{x}_{ij}$ is known for all elements of the population. We consider prediction of a general parameter defined as
\begin{align}\label{def_parameter}
\theta_{i} = h(y_{i1}, \ldots, y_{i N_{i}}),
\end{align}
where $h(\cdot)$ is a specified function.
Suppose a population model is specified for $y_{ij}$. Under the model, the minimum mean square error predictor of $\theta_{i}$ is defined as
\begin{align*}
\tilde{\theta}_{i}^{BP} = E_{p}(\theta_{i} \mid \bm{y}_{is}),
\end{align*}
which is called the best predictor (BP) of the area parameter. In this section, we develop best predictors of $\theta_{i}$ under two models that assume a gamma distribution for the response variables. Section 2.1 and Section 2.2 describe the gamma-gamma model and the gamma, GLMM respectively.
\subsection{Unit-Level Gamma-Gamma Small Area Model}
Assume that the population is generated under a gamma-gamma small area model defined as
\begin{align}\label{Gam-Gam}
y_{ij}\mid u_{i} &\stackrel{ind}{\sim}\text{Gamma}(\alpha,\eta_{ij}), \text{\hspace{0.35cm} $j=1,\hdots,N_{i}$, $i=1,\hdots,D$,},
\end{align}
where $\eta_{ij} = \exp(\bm{x}_{ij}^{T}\bm{\gamma})u_{i},$ $u_{i} \stackrel{iid}{\sim} \text{Gamma}(\delta,\delta),$ and $\bm{\gamma}_{1} = (\gamma_{0},\hdots,\gamma_{p+1})^{T}$. We use the notation $\text{Gamma}(a,b)$ to denote a gamma distribution with shape parameter $a$ and rate parameter $b$. The model (\ref{Gam-Gam}) can be viewed as a special case of \cite{Graf2019}, with a slightly different parametrization. We use a gamma distribution for $u_{i}$, while \cite{Graf2019} use an inverse-gamma distribution for the area random effect in a transformed scale.
We first consider small area prediction of the mean defined as
\begin{align*}
\bar{y}_{N_i} = \frac{1}{N_i}\sum_{j=1}^{N_i}y_{ij}, \text{\hspace{0.1in}$i=1,\hdots,D$.}
\end{align*}
{\it Theorem 1} gives the best predictor of $\bar{y}_{N_{i}}$ under the model (\ref{Gam-Gam}). Although {\it Theorem 1} can be cast as a special case of results in \cite{Graf2019}, we state {\it Theorem 1} and its proof here for completeness. The formulas in {\it Theorem 1} are slightly different than the formulas in \cite{Graf2019} because we parameterize the distribution of the random effect differently.
{\it Theorem 1:} Under the gamma-gamma model, the best predictor of the small area mean as
\begin{align}\label{Pred_mean}
\tilde{\bar{y}}_{N_i}^{BP}(\bm{\gamma}, \alpha, \delta, \bm{y}_{is})&= \frac{1}{N_i}\Bigg[\sum_{j=1}^{n_i}y_{ij}+\sum_{j=n_{i}+1}^{N_i}\alpha \exp(-\bm{x}_{ij}^{T}\bm{\gamma})\frac{\big\{\sum_{j=1}^{n_{i}}y_{ij}\exp(\bm{x}_{ij}^{T}\bm{\gamma})\big\}+\delta}{n_{i}\alpha+\delta-1}\Bigg].
\end{align}
A proof of {\it Theorem 1} is given in Appendix A. We use the notation ${\bar{y}}_{N_i}^{BP}(\bm{\gamma}, \alpha, \delta, \bm{y}_{is})$ to emphasize dependence of the best predictor on the unknown model parameters and the observed data.
\hspace{.2 in } We next consider prediction of more general parameters of the form (\ref{def_parameter}). Depending on the complexity of the real-valued function $h(\cdot)$ in equation (\ref{def_parameter}), an analytic expression of the best predictor, such as the mean predictor in the equation (\ref{Pred_mean}), may not exist. We adopt the approach of \cite{molina2010small} and use a Monte Carlo approximation for the best predictor. By the proof of Theorem 1, the conditional distribution of $u_{i}$ given the data is given by
\begin{align*}
u_{i} \mid \bm{y}_{is} \sim \mbox{Gamma}(\alpha + \delta, \sum_{j =1 }^{n_{i}} y_{ij}\mbox{exp}(\bm{x}_{ij}'\bm{\gamma}) + \tau).
\end{align*}
This convenient form for the distribution of $u_{i}$ enables us to develop a simple algorithm for approximating the best predictor of $\theta_{i}$. For $\ell = 1,\ldots, L$, repeat the following steps:
\begin{enumerate}
\item Generate $u_{i}^{(\ell)} \sim {\rm Gamma}(n_{i}\alpha+\delta,\sum_{j=1}^{n_{i}}y_{ij}\exp(\bm{x}_{ij}^{T}\bm{\gamma})+\delta)$, $i=1,\hdots,D$.
\item Generate $y_{ij}^{*(r)}\sim {f}(y_{ij}|u_i^{(\ell)};\bm{x}_{ij})$, $j=n_{i}+1,\hdots N_i$.
\item Define
\begin{align*}
{\theta}_{i}^{(\ell)} = h(\bm{y}_{is}',y_{in_{i+1}}^{*(\ell)}, \ldots, y_{iN_{i}}^{*(\ell)}).
\end{align*}
\end{enumerate}
An approximation for the best predictor of the area parameter is defined as
\begin{align}\label{BP}
\tilde{\theta}_{i}^{BP}(\bm{\gamma}, \alpha, \delta) = \frac{1}{L}\sum_{
\ell=1}^{L}{\theta}_{i}^{(\ell)}.
\end{align}
The notation $\tilde{\theta}_{i}^{BP}(\bm{\gamma}, \alpha, \delta)$ emphasizes dependence of the best predictor on the unknown $\bm{\gamma}$, $\alpha$, and $\delta$.
The algorithm above is slightly simpler than the algorithm of \cite{Graf2019}. We exploit the convenient form of the conditional distribution of $u_{i}$ to generate $y_{ij}$ for nonsampled elements through a hierarchical process that involves first generating $u_{i}$ and then generating $y_{ij}$ given $u_{i}$. In contrast, \cite{Graf2019} generate $y_{ij}$ from the marginal distribution of $y_{ij} \mid \bm{y}_{is}$. Simulating from the conditional distributions, as in our algorithm, is easier than simulating from the marginal distribution, as in \cite{Graf2019}.
The best predictor is a function of the unknown model parameters, denoted as $\bm{\psi} = (\alpha, \delta,\bm{\gamma}^{T})^{T}$. Calculation of a predictor requires an estimator of the model parameters. We propose to use maximum likelihood estimation. {\it Theorem 2} gives the closed-form expression for the likelihood.
{\it Theorem 2}: The likelihood for the model parameters, $\bm{\psi} = (\alpha, \delta,\bm{\gamma}^{T})^{T}$, under the model (\ref{Gam-Gam}) is of the form
\begin{align}\label{likelihood}
L(\bm{\psi}; \bm{y}_{is}) = \prod_{i=1}^{D} f(\bm{y}_{is}\mid\bm{\psi}),
\end{align}
where
\begin{align*}
f(\bm{y}_{is}\mid \bm{\psi}) &=\frac{\delta^{\delta}}{\{\Gamma(\alpha)\}^{n_{i}}\Gamma(\delta)}\prod_{j=1}^{n_i}y_{ij}^{\alpha-1}\exp\bigg(\alpha (\sum_{j=1}^{n_i}\bm{x}_{ij})^{T}\bm{\gamma}\bigg)\frac{\Gamma(n_{i}\alpha+\delta)}{\bigg(\sum_{j=1}^{n_{i}}\{y_{ij}\exp(\bm{x}_{ij}^{T}\bm{\gamma})\}+\delta\bigg)^{n_{i}\alpha+\delta}}.
\end{align*}
A proof of {\it Theorem 2} is given in Appendix B. Let $\hat{\bm{\psi}} = (\hat{\alpha}, \hat{\delta}, \hat{\bm{\gamma}}')'$ denote the maximum likelihood estimator defined as
\begin{align*}
\hat{\bm{\psi}} = argmax_{\bm{\psi}} L(\bm{\psi}; \bm{y}_{is}).
\end{align*}
Given the maximum likelihood estimator, we define an empirical Bayes predictor by substitution of $\hat{\bm{\psi}}$ with $\bm{\psi}$. The empirical best predictor of the mean is defined as
\begin{align}\label{ebhatclsd}
\hat{\bar{y}}_{N_{i}} = \tilde{\bar{y}}_{N_{i}}(\hat{\alpha}, \hat{\beta}, \hat{\bm{\gamma}}, \bm{y}_{is}).
\end{align}
The predictor (\ref{ebhatclsd}) is obtained by evaluating the closed-form expression for the best predictor of the mean in (\ref{Pred_mean}) at the maximum likelihood estimators. To define an empirical best predictor of a general parameter, we repeat steps 1-3 above with the maximum likelihood estimators in place of the true parameters. This enables us to define the empirical best predictor of a general parameter by
\begin{align}\label{ebhatgamgam}
\hat{\theta}_{i} = \tilde{\theta}_{i}(\hat{\alpha}, \hat{\beta}, \hat{\bm{\gamma}}, \bm{y}_{is}) .
\end{align}
We refer to the predictor (\ref{ebhatgamgam}) as the EB predictor. When we use the EB predictor (\ref{ebhatgamgam}) to predict the mean, we obtain a MC approximation for the closed-form predictor (\ref{ebhatclsd}).
\subsection{Gamma GLMM}
Define a unit-level gamma GLMM by
\begin{align}\label{Hobza}
y_{ij}\mid v_{i} &\stackrel{ind}{\sim} \text{Gamma}(\nu, \frac{\nu}{\mu_{ij}}), \text{ $i = 1,\hdots, D,$ $j = 1, \hdots, n_{i}$},
\end{align}
where $g(\mu_{ij}) =\bm{x}_{ij}^{T}\bm{\beta} + v_{i},$ and $v_{i}\stackrel{iid}{\sim} N(0, \phi^{2})$. We use the log link function for the mean parameter $\mu_{ij}$, such that $g(\mu_{ij}) = \mbox{log}(\mu_{ij})$. This model is comparable to the gamma-gamma model in that it has a constant same shape parameter and the area random effects are modeled with only one parameter. Let $\bm{\psi}^{\rm GLMM} = (\bm{\beta}', \phi, \nu)$ denote the parameters of the gamma GLMM. As mentioned in \cite{hobza2020small}, one can fit the model (\ref{Hobza}) using the {\tt R} function {\tt glmer} from the {\tt lme4} package. Let $\hat{\bm{\psi}}^{\rm GLMM} = (\hat{\bm{\beta}}', \hat{\phi}, \hat{\nu})'$ denote the resulting estimates.
\cite{hobza2020small} proposed three types of predictors for unit-level GLMMs, with emphasis on the gamma GLMM (\ref{Hobza}). \cite{hobza2020small} restricts attention to additive parameters of the form $N_{i}^{-1}\sum_{j = 1}^{N_{i}} \tilde{h}(y_{ij})$. We slightly modify their procedures for the purpose of constructing predictors of more general parameters with the form (\ref{def_parameter}) that are not necessarily additive. The first predictor is called an EBP predictor. Note that, for $i=1,\hdots,D$,
\begin{align}\label{Hobza EBP}
E(\theta_{i}\mid\bm{y}_{is})
&= \int h(\bm{y}_{is}, \bm{y}_{ir})f(\bm{y}_{ir}\mid \bm{y}_{is})d\bm{y}_{ir} \nonumber\\
&=\frac{\int h(\bm{y}_{is}, \bm{y}_{ir})\int f(\bm{y}_{ir}\mid {v}_{i})f(\bm{y}_{is}\mid v_{i})f(v_{i})d{v}_{i}d\bm{y}_{ir}}{\int f(\bm{y}_{is}\mid v_{i})f(v_{i})dv_{i}} \nonumber \\
&=\frac{\int \bigg\{\int h(\bm{y}_{is},\bm{y}_{ir})f(\bm{y}_{ir}\mid v_{i})d\bm{y}_{ir}\bigg\}f(\bm{y}_{is}\mid v_{i})f(v_{i})dv_{i}}{\int h(\bm{y}_{is},\bm{y}_{ir})f(\bm{y}_{ir}\mid v_{i})d\bm{y}_{ir}}.
\end{align}
Using equation (\ref{Hobza EBP}), we define the EBP under the model (\ref{Hobza}) using procedures similar to those suggested by Hobza et al. (2020). The iterative Monte Carlo algorithm is as follows:
\begin{enumerate}
\item For $\ell_{1}=1,\hdots,L_{1},$ generate $v_{i}^{(\ell_{1})}\sim N(0, \hat{\phi}^{2})$.
\begin{enumerate}
\item for $\ell_{2}=1,\hdots,L_{2},$ generate $\bm{y}_{ir}^{(\ell_{1},\ell_{2})}\sim \hat{f}(\bm{y}_{ir}\mid v_{i}^{(\ell_{1})};\hat{\bm{\psi}}^{\rm GLMM}),$ where
\begin{align*}
\hat{f}(\bm{y}_{ir}\mid v_{i}^{(\ell_{1})};\hat{\bm{\psi}}^{\rm GLMM}) = \prod_{j = n_{i}+1}^{N_{i}} g_{ij}(y_{ij} \mid v_{i}^{(\ell_{1})}, \hat{\bm{\beta}}, \hat{\nu}),
\end{align*}
and $ g_{ij}(y_{ij} \mid v_{i}^{(\ell_{1})}, \hat{\bm{\beta}}, \hat{\nu})$ is the density of a gamma distribution with shape parameter $\hat{\nu}$ and rate parameter $\hat{\nu}/\mbox{exp}(\bm{x}_{ij}'\hat{\bm{\beta}} + v_{i}^{(\ell_{1})})$.
\item Calculate
\begin{align*}
\hat{A}_{hi}^{(\ell_1)}&=\hat{f}(\bm{y}_{is}\mid v_{i}^{(\ell_{1})};\hat{\bm{\psi}}^{\rm GLMM})\times \frac{1}{L_{2}}\sum_{\ell_{2}=1}^{L_{2}}h(\bm{y}_{is},\bm{y}_{ir}^{(\ell_{1},\ell_{2})})
\end{align*}
and $\hat{B}_{hi}^{(\ell_1)} = \hat{f}(\bm{y}_{is}\mid v_{i}^{(\ell_{1})};\hat{\bm{\psi}}^{\rm GLMM})$, where $\hat{f}(\bm{y}_{is}\mid v_{i}^{(\ell_{1})};\hat{\bm{\psi}}^{\rm GLMM}) = \prod_{j = 1}^{n_{i}} g_{ij}(y_{ij} \mid v_{i}^{(\ell_{1})}, \hat{\bm{\beta}}, \hat{\nu})$.
\item Approximate the EBP of $\theta_{i}$ as
\begin{align}\label{HZ_EBP}
\hat{\theta}_{i}^{\rm EB\_HZ} = \frac{\sum_{\ell_{1}}^{L_{1}}\hat{A}_{hi}^{(\ell_1)}}{\sum_{\ell_{1}}^{L_{1}}\hat{B}_{hi}^{(\ell_1)}}.
\end{align}
\end{enumerate}
\end{enumerate}
The second predictor, called the plug-in predictor, is defined as
\begin{align}\label{HZ_PI}
\hat{\theta}_{i}^{\rm PI} &= h(\bm{y}_{is}, \tilde{\bm{\mu}}_{ir}),
\end{align}
where $\tilde{\bm{\mu}}_{ir} = (\tilde{{\mu}}_{in_{i}+1},\hdots,\tilde{{\mu}}_{iN_{i}})^{T}$, $\tilde{{\mu}}_{ij}= \exp(\bm{x}_{ij}^{T}\hat{\bm{\beta}}+\hat{v}_{i})$, and $\hat{v}_{i}$ is the predicted random effect obtained using \texttt{ranef}.
The last predictor, called the marginal predictor, is defined as
\begin{align}\label{HZ_M}
\hat{\theta}_{i}^{\rm M} &= \frac{1}{L} \sum_{\ell =1}^{L}h(\bm{y}_{is},\bm{y}_{ir}^{M(\ell)}),
\end{align}
where $\bm{y}_{ir}^{M(\ell)}\sim f(\bm{y}_{ir}\mid \hat{v}_{i};\hat{\bm{\psi}}^{\rm GLMM})$ for $\ell = 1,\hdots,L.$
\section{MSE Estimation}
In this section, we define two estimators of the MSE of $\hat{\theta}_{i}$. The MSE estimator of Section 3.1 is an adaptation of the general procedure of \cite{https://doi.org/10.48550/arxiv.2210.12221} to the small area context. In Section 3.2, we explain an existing parametric bootstrap MSE estimator in the context of the gamma-gamma model. We later compare the two MSE estimators through simulation in Section 5.2. We do not investigate MSE estimation for the gamma GLMM because we find, through the simulations of Section 5.1, that the gamma-gamma model is generally preferable to the gamma GLMM.
\subsection{Proposed MSE Estimator}
Suppose we use $L=\infty$ in the prediction procedure of Section 2.1 in order to ignore the variability from the MC approximation used to construct the EB predictor (\ref{ebhatgamgam}). Then, note that the MSE of the predictor $\hat{\theta}_{i}$ can be decomposed as
\begin{equation}
\begin{aligned}\label{msedecomp1}
{\rm MSE}(\hat{\theta}_{i} ) = M_{i1}+M_{i2},
\end{aligned}
\end{equation}
where $M_{i1} =E[V(\theta_i\mid\bm{y}_{is}; \bm{\psi})] $, $M_{i2} = E[(\hat{\theta}_{i,\infty} - \tilde{\theta}_{i,\infty}(\alpha, \delta, \bm{\gamma}) )^{2}]$, and $(\hat{\theta}_{i,\infty}, \tilde{\theta}_{i,\infty}(\alpha, \delta, \bm{\gamma})) = \mbox{lim}_{\ell \rightarrow\infty}(\hat{\theta}_{i}, \tilde{\theta}_{i}(\alpha, \delta, \bm{\gamma}, \bm{y}_{is})$. \cite{https://doi.org/10.48550/arxiv.2210.12221} provide a more rigorous development of the decomposition (\ref{msedecomp1}). Also, see \cite{rao2015small} and \cite{reluga2021simultaneous} for a similar decomposition of the MSE.
The first term $M_{i1}$, called the leading term, is the MSE of the best predictor, and its unbiased estimator is $V(\theta_{i}\mid \bm{y}_{is};\bm{\psi})$. In practice, due to the unknown model parameters $\bm{\psi}$, we use $V(\theta_{i}\mid \bm{y}_{is};\hat{\bm{\psi}})$ as the leading term estimator, and approximate it as
\begin{equation}
\begin{aligned}
\label{leadingtermest}
V(\theta_{i}\mid \bm{y}_{is};\hat{\bm{\psi}}) &\approx \frac{1}{L-1}\sum_{\ell= 1}^{L}(\hat{\theta}_{i}^{(\ell)} - \hat{\theta}_{i}^{EB})^{2}=:\hat{M}_{1i},
\end{aligned}
\end{equation}
where $\hat{\theta}_{i}^{(\ell)}$ are obtained through the EBP procedure in Section 2.1.
The extra variation induced by replacing ${\bm{\psi}}$ with $\hat{\bm{\psi}}$ in the best predictor is accounted for by the second component $M_{i2}$. The analytical form of $M_{i2}$ is difficult to obtain, so we use the parametric bootstrap to approximate it. For $b = 1,\ldots, B$, repeat the following steps:
\begin{enumerate}
\item Generate the bootstrap sample $\bm{y}_{is}^{*(b)}= ({y}_{i1}^{*(b)},\hdots, {y}_{in_{i}}^{(b)})^{T} $, $i=1,\hdots,D$ from model (\ref{Gam-Gam}) as
$ y_{ij}^{*(b)}\stackrel{ind}{\sim}{\rm Gamma}(\hat{\alpha}, \exp(\bm{x}_{ij}^{T}\hat{\bm{\gamma}})u_{i}^{*(b)}) \text{, $j=1,\hdots, n_{i}$, $i=1,\hdots,D$}$, where $u_{i}^{*(b)}\stackrel{{iid}}{\sim} {\rm Gamma}(\hat{\delta},\hat{\delta})\text{, $i=1,\hdots,D$}.$
\item Estimate the bootstrap version of the model parameter estimates, $\hat{\bm{\psi}}^{*(b)}$, by maximizing the likelihood with the bootstrap data generated in step 1. Specifically, $\hat{\bm{\psi}}^{*(b)} = argmax_{\bm{\psi}}L(\bm{\psi}; \bm{y}_{is}^{*(b)})$.
\item Calculate the bootstrap predictor, $\hat{\theta}_{i}^{*(b)} = \tilde{\theta}_{i}(\hat{\bm{\psi}}^{*(b)}; \bm{y}_{is}^{*(b)})$. Note that the bootstrap predictor is obtained by applying the algorithm defined in Section 2.1 with the bootstrap parameter estimator and the original data. Implementation of this algorithm results in simulated samples $\theta_{i}^{(\ell, b)}$. Calculate the bootstrap MC approximation for $V\{\theta_{i}\mid\bm{y}_{is};\hat{\bm{\psi}}^{*(b)}\}$, denoted as $\hat{M}_{1i}^{*(b)}$, as $\hat{M}
_{1i}^{*(b)} = (L-1)^{-1}\sum_{\ell = 1}^{L} (\theta_{i}^{(\ell, b)} - \hat{\theta}
_{i}^{*(b)})^{2}$.
\end{enumerate}
Then, define the estimator of $M_{2i}$ as:
\begin{align}\label{M2}
\hat{M}_{2i} = \frac{1}{B}\sum_{b=1}^{B}(\hat{\theta}_{i}^{EB*(b)} - \hat{\theta}_{i}^{EB})^{2}.
\end{align}
A preliminary estimator of the MSE of $\theta_{i}$ is be defined as
\begin{align}\label{noBC}
{\rm mse}_{i}^{\rm noBC} = \hat{M}_{1i}+\hat{M}_{2i}.
\end{align}
The label ``noBC'' is used to indicate that the MSE estimator (\ref{noBC}) does not incorporate a correction for the bias of the estimator of the leading term.
However, the estimator of leading term $\hat{M}_{1i}$ is not an unbiased estimator for $M_{1i}$ due to the replacement ${\bm{\psi}}$ with $\hat{\bm{\psi}}$. To adjust this bias, we may estimate it by utilizing $\hat{M}_{1i}^{*(b)}$, $b=1,\hdots,B$, which is the byproduct of the bootstrap procedure. We can define an additive bias correction as $\hat{M}_{1i}^{\rm Add} = \hat{M}_{1i}- ({\bar{M}}_{1i}^{*B} - \hat{M}_{1i})$, or a multiplicative correction as, $\hat{M}_{1i}^{\rm Mult} = \hat{M}_{1i}^{2} [\bar{M}_{1i}^{*B}]^{-1}$, where $\bar{M}_{1i}^{*B} = \frac{1}{B}\sum_{b=1}^{B}V\{\theta_{i}\mid\bm{y}_{is};\hat{\bm{\psi}}^{*(b)}\}$.
These classic additive and multiplicative bias-correction approaches are straightforward, but \cite{hall2006parametric} mention several issues with those approaches. The additive and multiplicative bias-correction could produce a negative leading term estimator when ${\bar{M}}_{1i}^{*B} > \hat{M}_{1i}$ and unreliable estimators, respectively. Thus, \cite{hall2006parametric} suggest a different bias-correction defined as
\begin{align*
\hat{M}_{1i}^{\rm HM} &= \begin{cases}
\hat{M}_{1i}^{\rm Add}, & \hat{M}_{1i}\geq {\bar{M}}_{1i}^{*B},\\
\hat{M}_{1i}\exp\big[-\{\bar{M}_{1i}^{*B}- \hat{M}_{1i})\big\}/\bar{M}_{1i}^{*B}\big], & \text{if }\hat{M}_{1i}< {\bar{M}}_{1i}^{*B}.
\end{cases}
\end{align*}
As a special case of the general bias corrections given in \cite{hall2006parametric}, we further define a compromise between $\hat{M}_{1i}^{\rm Add}$ and $\hat{M}_{1i}^{\rm Mult}$ as
\begin{align*}
\hat{M}_{1i}^{\rm Comp} &= \begin{cases}
\hat{M}_{1i}^{\rm Add}, & \hat{M}_{1i}\geq {\bar{M}}_{1i}^{*B},\\
\hat{M}_{1i}^{\rm Mult}, & \text{if }\hat{M}_{1i}< {\bar{M}}_{1i}^{*B}.
\end{cases}
\end{align*}
In summary, the bias-corrected MSE estimators are constructed by
\begin{align}\label{HM}
{\rm mse}_{i}^{\rm HM} = \hat{M}_{1i}^{\rm HM}+\hat{M}_{2i},
\end{align}
and
\begin{align}\label{Comp}
{\rm mse}_{i}^{\rm Comp} = \hat{M}_{1i}^{\rm Comp}+\hat{M}_{2i}.
\end{align}
\subsection{Existing Parametric Bootstrap MSE Estimators}
Instead of estimating $M_{i1}$ and $M_{i2}$ separately, one can use the parametric bootstrap to estimate ${\rm MSE}(\hat{\theta}_{i})$ directly. \cite{molina2007small}, \cite{Graf2019}, and \cite{hobza2020small} use the single-stage bootstrap method to estimate the MSE. One may implement the double-bootstrap algorithm suggested by \cite{hall2006parametric} to correct the single-stage estimator. However, the double-bootstrap is computationally expensive and may not be feasible for a large population. Thus, we consider a simpler double-bootstrap motivated by \cite{erciulescu2014parametric} and \cite{reluga2021simultaneous}, where we generate only one bootstrap replicate in the second-stage bootstrap. The following algorithm describes how to obtain those estimators.
\begin{enumerate}
\item Obtain the estimate of the model parameter $\hat{\bm{\psi}}$.
\item For $b_{1} = 1, \hdots, B_{1}$, independently generate the bootstrap population $\bm{y}_{i}^{*(b_{1})}= ({y}_{i1}^{*(b_{1})},\hdots, {y}_{iN_{i}}^{(b_{1})})^{T} $, $i=1,\hdots,D$ from model (\ref{Gam-Gam}) as $ y_{ij}^{*(b_{1})}\stackrel{ind}{\sim}{\rm Gamma}(\hat{\alpha}, \exp(\bm{x}_{ij}^{T}\hat{\bm{\gamma}})u_{i}^{*(b_{1})})$, $j=1,\hdots, N_{i}$, $i=1,\hdots,D$ where $u_{i}^{*(b_{1})}\stackrel{{iid}}{\sim} {\rm Gamma}(\hat{\delta},\hat{\delta})$ for $ i=1,\hdots,D$.
\item Calculate the bootstrap version of a small area parameter $\theta_{i}^{*(b_{1})}$ with the bootstrap population $\bm{y}_{i}^{*(b_{1})}$, the EBP $\hat{\theta}_{i}^{*(b_{1})}$ with the bootstrap sample $\bm{y}_{is}^{*(b_{1})}$, and $D^{*(b_{1})} = \{\hat{\theta}_{i}^{*(b_{1})}- \theta_{i}^{*(b_{1})}\}^2$.
\item Estimate the bootstrap model parameter estimates $\hat{\bm{\psi}}^{*(b_{1})} = (\hat{\alpha}^{*(b_{1})},\hat{\delta}^{*(b_{1})},\hat{\bm{\gamma}}^{(b_{1})T})^{T}$ with the bootstrap sample $\bm{y}_{is}^{*(b_{1})}$.
\begin{enumerate}
\item For $b_{2} = 1, \hdots, B_{2}$, independently generate the bootstrap population $\bm{y}_{i}^{**(b_{2})}= ({y}_{i1}^{**(b_{2})},\hdots, {y}_{iN_{i}}^{**(b_{2})})^{T} $, $i=1,\hdots,D$ from model (\ref{Gam-Gam}):
\begin{align*}
y_{ij}^{**(b_{2})}&\stackrel{ind}{\sim}{\rm Gamma}(\hat{\alpha}^{*(b_{2})}, \exp(\bm{x}_{ij}^{T}\hat{\bm{\gamma}}^{*(b_{1})})u_{i}^{**(b_{1})}) \text{, $j=1,\hdots, N_{i}$, $i=1,\hdots,D$}\\
u_{i}^{**(b_{2})}&\stackrel{{iid}}{\sim} {\rm Gamma}(\hat{\delta}^{*(b_{1})},\hat{\delta}^{*(b_{1})})\text{, $i=1,\hdots,D$}.
\end{align*}
(We set $B_{2}=1$ to employ the simpler double-bootstrap.)
\item Calculate the bootstrap version of a small area parameter $\theta_{i}^{**(b_{2})}$ with the bootstrap population $\bm{y}_{i}^{**(b_{2})}$, the EBP $\hat{\theta}_{i}^{**(b_{2})}$ with the bootstrap sample $\bm{y}_{is}^{**(b_{2})}$, and $D^{**(b_{2})} =\{ \hat{\theta}_{i}^{**(b_{2})}- \theta_{i}^{**(b_{2})}\}^2$.
\item Set ${\rm mse}_{i}^{**(b_1)} = \frac{1}{B_{2}}\sum_{b_{2}=1}^{B_{2}}D^{**(b_{2})}.$
\end{enumerate}
\item Finally, define the single-stage and the double-bootstrap MSE estimators as
\begin{align}
{\rm mse}_{i}^{S} = \frac{1}{B_{1}}\sum_{b_{1}=1}^{B_{1}} D^{*(b_{1})},
\end{align}
and
\begin{align}
{\rm mse}_{i}^{D} = 2{\rm mse}_{i}^{S}- \frac{1}{B_{1}}\sum_{b_{1}=1}^{B_{1}}{\rm mse}_{i}^{**(b_1)}.
\end{align}
\end{enumerate}
Note that the single-stage MSE estimator is comparable to (\ref{noBC}) and the (simpler) double-stage MSE estimator to the proposed bias-corrected MSE estimators.
\section{Extension of SAE Gamma-Gamma model Under Informative Sampling}
We extend the gamma-gamma model to an informative sampling design. We utilize well-known relationships among the population, sample, and sample-complement distributions of $y_{ij}$ established in \cite{pfeffermann2007small}. We assume the same model for the first moment of the sampling weight in \cite{pfeffermann2007small}.
For completeness, we restate key relationships defined in \cite{pfeffermann2007small} with respect to the second-stage unit $y_{ij}$ and the corresponding sampling weight $w_{ij}$. These are given by
\begin{align}\label{relationship1}
f(y_{ij}\mid \bm{x}_{ij},u_{i},I_{i}=1,I_{ij}=1)&=\frac{E(w_{ij}\mid \bm{x}_{ij}, u_{i},I_{i}=1, I_{ij}=1)}{E(w_{ij}\mid \bm{x}_{ij}, u_{i},y_{ij},I_{i}=1, I_{ij}=1)}f(y_{ij}\mid \bm{x}_{ij},u_{i},I_{i}=1)
\end{align}
and
\begin{align}\label{relationship2}
f(y_{ij}\mid \bm{x}_{ij},u_{i},I_{i}=1,I_{ij}=0)&=\frac{E[(w_{ij}-1)\mid \bm{x}_{ij}, u_{i},y_{ij},I_{i}=1, I_{ij}=1]}{E[(w_{ij}-1)\mid \bm{x}_{ij}, u_{i},I_{i}=1, I_{ij}=1]}f(y_{ij}\mid \bm{x}_{ij},u_{i},I_{i}=1,I_{ij}=1),
\end{align}
where $I_{i}$ and $I_{ij}$ are the sample indicators for an area $i$ and unit $j$ in the area $i$, respectively, $w_{ij}=1/P(j \in s_{i})$, and $s_{i}=\{j: I_{ij}=1\}$. These relationships imply that we can deduce adequate information about other distributions from observed units and their weights. For simplicity, we suppose all areas are selected such that $I_{i} = 1$ for $i= 1,\ldots, D$.
For the complex design, we suppose the sample distribution is given by
\begin{align}\label{sampledist}
y_{ij} \mid \bm{x}_{ij},u_{i},I_{i}=1, I_{ij}=1 &\stackrel{ind}{\sim} \text{Gamma}(\alpha_{s},\eta_{s,ij})\text{\hspace{0.2cm} $j=1,\hdots,n_{i}$, $i=1,\hdots,D$},
\end{align}
where $ \eta_{s,ij} = \exp(\bm{x}_{ij}^{T}\gamma_{s})u_{i}\nonumber$ and $u_{i}\stackrel{iid}{\sim} \text{Gamma}(\delta_{s}, \delta_{s})$. Here, the model parameters in (\ref{sampledist}) are differentiated from those of the population distribution (\ref{Gam-Gam}) using the subscript $s$. We further assume that the expected values of the sampling weight satisfies
\begin{align}\label{sp_wgt}
E(w_{ij}\mid \bm{x}_{ij}, u_{i},y_{ij},I_{i}=1, I_{ij}=1)&=E(w_{ij}\mid \bm{x}_{ij},y_{ij},I_{i}=1,I_{i}=1)\\
&=\kappa_{i}\exp(\bm{x}_{ij}^{T}\bm{a}-by_{ij}),\nonumber
\end{align}
for $b>0$ where $\kappa_{i} = N_{i}^{-1}\sum_{j=1}^{N_{i}}\exp(-\bm{x}_{ij}^{T}\bm{a}+by_{ij})$. Denote the collection of fixed model parameters by ${\bm{\psi}}^{\rm Info} = (\alpha_{s},\delta_{s},\bm{\gamma}_{s}^{T},\bm{a}^{T},b,\bm{\kappa}^{T})^{T}$, where $\bm{\kappa} = (\kappa_{1}, \hdots, \kappa_{D})^{T}$. Then, using the relationship (\ref{relationship2}), the following sample-complement distributions under the informative sample scheme can be derived as
\begin{align}\label{fciyij}
f(y_{ij}\mid \bm{x}_{ij}, u_{i}, I_{i}=1, I_{ij}=0; \bm{\psi}^{Info})
&= \frac{[E(w_{ij}\mid \bm{x}_{ij},y_{ij},I_{i}=1,I_{ij}=1)-1] f(y_{ij} \mid \bm{x}_{ij},u_{i},I_{i}=1, I_{ij}=1)}{E(w_{ij}\mid \bm{x}_{ij},I_{i}=1,I_{ij}=1)-1} \nonumber \\
&=\frac{\lambda_{ij}}{\lambda_{ij}-1}f(y_{ij}\mid\bm{x}_{ij},u_{i},I_{i}=1)-\frac{1}{\lambda_{ij}-1}f(y_{ij}\mid\bm{x}_{ij},u_{i},I_{i}=1,I_{ij}=1),
\end{align}
where
\begin{align*}
\lambda_{ij} &= E(w_{ij}\mid\bm{x}_{ij}, I_{i=1}, I_{ij}=1)\\
&= \kappa_{i}\exp(\bm{a}^{T}\bm{x}_{ij})E[\exp(-by_{ij})\mid\bm{x}_{ij}, u_{i},I_{i=1}, I_{ij}=1]\\
&= \kappa_{i}\exp(\bm{a}^{T}\bm{x}_{ij})\bigg(1+\frac{b}{\eta_{s,ij}}\bigg)^{-\alpha},
\end{align*}
and the population distribution $f(y_{ij}\mid \bm{x}_{ij},u_{i},I_{i}=1)$ is given by
\begin{align}\label{fp}
f(y_{ij}\mid \bm{x}_{ij},u_{i},I_{i}=1)&\propto E(w_{ij}\mid \bm{x}_{ij}, u_{i},y_{ij},I_{i}=1, I_{ij}=1)f(y_{ij}\mid \bm{x}_{ij},u_{i},I_{i}=1, I_{ij}=1) \text{\hspace{0.2cm} (\ref{relationship1})}\nonumber\\
&= \frac{(\eta_{s,ij}+b)^{\alpha_{s}}}{\Gamma(\alpha_{s})} y_{ij}^{\alpha_{s}-1}\exp(-y_{ij}(\eta_{ij}+b))
\end{align}
When the observed values are not related to sampling probability (that is, $b=0$), the population and sample-complement distribution, $f(y_{ij}\mid \bm{x}_{ij},u_{i},I_{i}=1)$ and $f(y_{ij}\mid \bm{x}_{ij},u_{i},I_{i}=1, I_{ij}=0)$, are the same as the sample distribution $f(y_{ij}\mid \bm{x}_{ij},u_{i},I_{i}=1, I_{ij}=1)$. In this case, the algorithm for the (empirical) best predictor is identical to that in Section 2.1.
However, under the informative design, that is $b\neq0$, we need to reflect the informative sampling scheme by using the sample-complement distribution (\ref{fciyij}). The procedure requires an estimator of $\bm{\psi}^{Info}$. We define $(\hat{\alpha}_{s}, \hat{\delta}_{s}, \hat{\bm{\gamma}}_{s}')'$ to be the maximum likelihood estimator under the sample model. The estimator of $(\bm{a}^{T},b,\bm{\kappa}_{i})^{T}$ is obtained by minimizing
\begin{align*}
{\rm SSE}(\bm{a}^{T},b,\bm{\kappa}^{T}) = \sum_{i=1}^{D}\sum_{j=1}^{n_{i}}\big\{w_{ij} - \kappa_{i}\exp(\bm{x}_{ij}^{T}\bm{a}-by_{ij})\big\},
\end{align*}
as in \cite{pfeffermann2007small}. The procedure for the empirical best predictor under the informative design is then implemented as follows. For $\ell = 1, \ldots, L$, repeat the following steps:
\begin{enumerate}
\item Generate $u_{i}^{(\ell)} \sim \text{Gamma}(n_{i}\hat{\alpha}_{s}+\hat{\delta}_{s},\sum_{j=1}^{n_{i}}y_{ij}\exp(\bm{x}_{ij}^{T}\hat{\bm{\gamma}}_{s})+\hat{\delta}_{s})$, $i=1,\hdots,D$.
\item Generate $y_{ij}^{*(\ell)}\sim \hat{f}(y_{ij}\mid \bm{x}_{ij}, u_i^{(\ell)},I_{i}=1,I_{ij}=0; \hat{\bm{\psi}}^{Info})$, $j=n_{i}+1,\hdots N_i$.
\item Define
\begin{align*}
\hat{\theta}_{i}^{(\ell)} = h(\bm{y}_{is},y_{in_{i+1}}^{*(\ell)}, \ldots, y_{iN_{i}}^{*(\ell)}).
\end{align*}
\end{enumerate}
Then, the empirical best predictor of the area parameter is defined as
\begin{align}\label{EBinfor}
\hat{\theta}_{i}^{\rm EB\_INFO} &= \frac{1}{L}\sum_{
\ell=1}^{L}\hat{\theta}_{i}^{(\ell)}.
\end{align}
{\it Remark 1:} We use inversion sampling to generate $y_{ij}^{*(r)}$ from the sample-complement distribution $f(y_{ij}\mid\bm{x}_{ij},u_{i},I_{i}=1,I_{ij}=0; \hat{\bm{\psi}}^{Info})$ in Step 3 of the procedure. Specifically, we decompose Step 3 into two steps as follows:
\begin{enumerate}
\item[3.1] Generate $q_{ij}\stackrel{iid}{\sim} \text{Unif}(0,1)$, $j = n_{i+1}.\hdots,N_{i}$.
\item [3.2] Set $y_{ij}^{*(\ell)} = \hat{F}^{-1}(q_{ij}\mid \bm{x}_{ij}, u_{i}^{(\ell)}, I_{i}=1, I_{ij}=0; \hat{\bm{\psi}}^{Info})$, $j = n_{i+1}.\hdots,N_{i}$, where
\begin{align*}
F(y_{ij}\mid \bm{x}_{ij}, u_{i}, I_{i}=1, I_{ij}=0; \hat{\bm{\psi}}^{Info}) &= \int_{0}^{y_{ij}}f(y_{ij}\mid\bm{x}_{ij}, u_{i}, I_{i}=1, I_{ij}=0)dy_{ij}\\
&=\frac{\lambda_{ij}}{\lambda_{ij}-1}\frac{\gamma\{\alpha_{s},(\eta_{s,ij}+b)y_{ij}\big\}}{\Gamma(\alpha_{s})}-\frac{1}{\lambda_{ij}-1}\frac{\gamma\{\alpha_{s},\eta_{s,ij}y_{ij}\big\}}{\Gamma(\alpha_{s})},
\end{align*}
and $\gamma(a,x) = \int_{0}^{x}t^{s-1}\mbox{exp}(-t)dt$ is the incomplete gamma function.
\end{enumerate}
{\it Remark 2:} For sufficiently large $E(w_{ij}\mid\bm{x}_{ij}, I_{i=1}, I_{ij}=1)$, one may use the population distribution, $f(y_{ij}\mid \bm{x}_{ij}, u_{i}, I_{i}=1)$, as an approximation for the sample-complement distribution, $f(y_{ij}\mid \bm{x}_{ij}, u_{i}, I_{i}=1, I_{ij}=0)$. Simulating from the population distribution is easier than simulating from the sample complement distribution because the population distribution is a gamma distribution. We use the exact complement distribution because we found that the difference between the population distribution and the complement distribution can be important when the degree of informativeness is large.
\section{Simulation Study}
We carry out three simulation experiments to evaluate the procedures defined in Sections 2-4. In all simulation studies, we consider $D=100$ areas, each with population size $N_{i} = 100$. We stratify the areas into two strata, where Stratum $H_{1}$ and Stratum $H_{2}$ are composed of areas $1\le i \le 50$ and $51\le i \le D$, respectively. Assign the area sample size of $n_{i}=10$ if $i\in H_{1}$ and $n_{i}=20$ if $i\in H_{2}$. Samples of size $n_{i}$ are selected independently across the areas using simple random sampling without replacement. Then, for the population units, we simulate the study variables $y_{ij}$ in conjunction with the values of the auxiliary variable $x_{ij} \sim U(0,2)$ for $j=1,\hdots,N_{i}$, where $x_{ij}$ are held constant throughout MC simulations. The true model for simulating a population is defined in each experiment.
\hspace{.2 in} In addition to the small area mean, we take into account three non-additive parameters. The first two are the $25$th and $75$th sample quantiles, denoted as $Q_{0.25}$ and $Q_{0.75}$, respectively.
These are calculated through the function \texttt{quantile} in \texttt{R} with the default method.
The second is the Gini Coefficient (abbreviated Gini) defined as
$${\rm Gini}_i = \frac{\sum_{k=1}^{N_i}\sum_{\ell=1}^{N_i}\mid y_{ik} - y_{i\ell}\mid}{2N_i^{2}\bar{Y}_{i}},$$
where the value is obtained by the function \texttt{gini} of \texttt{R} package \texttt{reldist}.
\subsection{Simulation 1}
The objectives of Simulation 1 are to evaluate the performance of the gamma-gamma predictors in comparison to the three predictors suggested by \cite{hobza2020small} and to examine the robustness of the gamma-gamma predictors against model misspecification. To attain these goals, the values of target variable $y$ are generated from either the gamma-gamma model (\ref{Gam-Gam}) or the GLMM (\ref{Hobza}).
\subsubsection{The Gamma-Gamma model}
We first generate data from the model (\ref{Gam-Gam}). We set $\bm{\gamma}=(1,0.5)^{T}$, $\delta = 4$, and consider the three values of $\alpha \in \{1,2.5,5\}$. Note that the skewness of the distribution descreases as $\alpha$ increases.
In each simulation iteration, we compute the considered predictors, $\hat{\theta}_{i}^{(m),\text{ } {\rm pred}}$, ${\rm pred} \in\{\rm EB, EB\_HZ, EB\_M, EB\_PI\}$, where ${\rm EB}$, ${\rm EB\_HZ}$, ${\rm EB\_M}$ and ${\rm EB\_PI}$ are defined in (\ref{BP}), (\ref{HZ_EBP}), (\ref{HZ_PI}), and (\ref{HZ_M}), respectively. For $\bar{Y}$, the closed-form expression (\ref{Pred_mean}) is also used, and is denoted as EB\_clsd. Further, calculate the direct estimator of each considered parameter, denoted as Dir, which can be obtained by passing the area-specific sample obtained in step 3 as the argument of the area parameter function.
\hspace{0.2 in} We compare the predictors using the relative bias (RB) and relative root MSE (RRMSE). The RB and RRMSE for predictor $pred$ are defined as
\begin{align*}
{\rm RB}_{i} = \frac{M^{-1}\sum_{m=1}^{M}(\hat{\theta}_{i}^{(m), \rm pred}-\theta_{i}^{(m)})}{M^{-1}\sum_{m=1}^{M}\theta_{i}^{(m)}}, \hspace{0.2cm} \mbox{ and } {\rm RRMSE}_{i} = \frac{\sqrt{M^{-1}\sum_{m=1}^{M}(\hat{\theta}_{i}^{(m), \rm pred}-\theta_{i}^{(m)})^{2}}}{M^{-1}\sum_{m=1}^{M}\theta_{i}^{(m)}},
\end{align*}
where $\theta_{i}^{(m)}$ denotes the area population parameter obtained in MC simulation $m$. The averages of RB and RRMSE (in \%) for areas within the same stratum (same sample size) are shown in Table \ref{sim1.1_EB_RB_RE}.
The $\rm EB$ and $\rm EB\_clsd$ predictors are superior to the alternatives for this configuration. EB has RB closest to zero for most cases, and the RB of EB is uniformly below 1\% in absolute value. The RRMSE of $\rm EB$ is consistently smaller than the RRMSE of the $\rm EB\_HZ$, $M$, PI, or Dir predictors. This is expected because $\rm EB$ is an estimator of the optimal (minimum MSE) predictor for this simulation model. For the mean, $\rm EB\_clsd$ is more efficient than $\rm EB$, which is expected because $\rm EB$ is an MC approximation for $\rm EB\_clsd$. The loss of efficiency from use of the MC approximation through $\rm EB$, relative to $\rm EB\_clsd$, is slight.
The properties of the $\rm EB\_HZ$, $M$, and $\rm PI$ predictors reflect the patterns described in \cite{hobza2020small}. The RB of the PI predictor makes an important contribution to the RRMSE for all parameters, except for the mean. This occurs because the PI predictor replaces a non-sampled unit with its estimated conditional mean. Therefore, the $25$th quantile is predicted to be greater than the actual value, while the $75$th quantile is predicted to be lower. A similar phenomenon occurs with the PI predictor of Gini. In terms of RRMSE, $M$ compares favorably to $\rm EB\_HZ$, which is consistent with the simulation results of \cite{hobza2020small}.
As expected, the efficiency of the direct estimator depends heavily on the sample size. The direct estimator is inefficient for $Q_{0.25}$ and for $\rm Gini$, as a result of the small sample size and the nonlinearity of these parameters. For the mean, the direct estimator is more efficient than $\rm EB\_HZ$.
\begin{table}[H]
\centering
\scalebox{0.75}{
\begin{tabular}{ccc|ccccc|ccccc}
\toprule
\toprule
\multirow{2}{*}{\large{\bfseries Parameter }} & \multirow{2}{*}{\large{\bfseries $\alpha$}} & \multirow{2}{*}{\large{\bfseries $n_{i}\,$}} &\multicolumn{5}{c}{\large{\bfseries RB (\%) }}&\multicolumn{5}{c}{\large{\bfseries RRMSE (\%) }}\\
\cmidrule(r){4-13}
& & &{\large{\bfseries EB}} & {\large{\bfseries $EB\_HZ$}} & {\large{\bfseries $M$}} & {\large{\bfseries PI}} & {\large{\bfseries Dir}} & {\large{\bfseries EB}} & {\large{\bfseries $EB\_HZ$}} & {\large{\bfseries $M$}} & {\large{\bfseries PI}} & {\large{\bfseries Dir}} \\
\midrule
h1 & 1.00 & 10 & 0.05 (0.05) & -2.02 & -6.59 & -6.59 & 0.02 & 33.48 (33.40) & 38.36 & 34.99 & 34.98 & 40.21 \\
& & 20 & -0.02 (-0.03) & -1.50 & -3.67 & -3.67 & -0.05 & 23.49 (23.45)& 28.74 & 24.29 & 24.28 & 26.39 \\
& 2.50 & 10 & 0.05 (0.05) & -1.98 & -6.57 & -6.57 & -0.01 & 33.14 (33.07) & 38.26 & 34.93 & 34.92 & 39.60 \\
& & 20 & 0.04 (0.04) & -1.47 & -3.61 & -3.61 & 0.04 & 23.59 (23.52)& 29.70 & 24.40 & 24.39 & 26.54 \\
& 5.00 & 10 & 0.04 (0.04)& -4.12 & -5.08 & -5.08 & 0.03 & 22.45 (22.39)& 36.39 & 26.30 & 26.30 & 26.39 \\
& & 20 & 0.02 (0.02) & -3.23 & -2.74 & -2.74 & -0.00 & 15.32 (15.29) & 28.91 & 16.91 & 16.91 & 17.65 \\
\midrule
h2 & 1.00 & 10 & 0.09 (0.09) & 2.17 & -2.04 & 142.96 & 26.74 & 39.03 & 43.92 & 39.64 & 168.48 & 88.62 \\
& & 20 & 0.02 & 2.20 & 0.59 & 144.00 & 11.65 & 30.24 & 35.53 & 30.48 & 171.84 & 54.62 \\
& 2.50 & 10 & 0.10 & 2.19 & -2.04 & 142.92 & 26.72 & 38.68 & 43.65 & 39.45 & 167.81 & 88.29 \\
& & 20 & 0.10 & 2.23 & 0.65 & 144.12 & 11.76 & 30.45 & 36.62 & 30.63 & 172.54 & 55.09 \\
& 5.00 & 10 & 0.04 & -5.34 & -5.89 & 39.17 & 11.16 & 24.88 & 40.10 & 29.07 & 48.72 & 44.31 \\
& & 20 & 0.04 & -4.45 & -3.40 & 40.89 & 4.96 & 17.98 & 34.12 & 19.76 & 49.48 & 28.40 \\
\midrule
h3 & 1.00 & 10 & 0.04 & -2.05 & -6.40 & -15.06 & -1.91 & 34.72 & 39.97 & 36.16 & 40.59 & 47.07 \\
& & 20 & -0.01 & -1.53 & -3.47 & -12.51 & -0.95 & 25.04 & 30.97 & 25.81 & 30.84 & 32.11 \\
& 2.50 & 10 & 0.04 & -2.00 & -6.38 & -15.02 & -1.89 & 34.37 & 39.75 & 36.06 & 40.51 & 46.16 \\
& & 20 & 0.02 & -1.52 & -3.44 & -12.46 & -0.87 & 25.21 & 32.22 & 26.04 & 31.08 & 32.28 \\
& 5.00 & 10 & 0.03 & -4.08 & -4.95 & -11.26 & -2.34 & 23.40 & 38.14 & 27.16 & 31.26 & 31.41 \\
& & 20 & 0.01 & -3.22 & -2.60 & -9.05 & -1.09 & 16.51 & 30.94 & 17.94 & 22.26 & 21.84 \\
\midrule
h4 & 1.00 & 10 & -0.02 & -1.57 & -1.87 & -59.85 & -9.60 & 5.77 & 6.05 & 6.14 & 60.19 & 19.30 \\
& & 20 & -0.02 & -1.39 & -1.70 & -51.87 & -4.32 & 5.44 & 5.68 & 5.76 & 52.22 & 12.36 \\
& 2.50 & 10 & -0.01 & -1.54 & -1.84 & -59.81 & -9.60 & 5.77 & 6.04 & 6.13 & 60.15 & 19.29 \\
& & 20 & -0.02 & -1.38 & -1.70 & -51.84 & -4.29 & 5.45 & 5.68 & 5.77 & 52.19 & 12.34 \\
& 5.00 & 10 & -0.02 & 1.04 & 0.68 & -49.16 & -9.58 & 6.40 & 6.57 & 6.47 & 49.62 & 21.13 \\
& & 20 & -0.01 & 0.97 & 0.57 & -43.25 & -4.29 & 6.04 & 6.26 & 6.09 & 43.71 & 13.65 \\
\bottomrule \bottomrule
\end{tabular}
}
\caption{RB ($\%$) and RRMSE ($\%$) for EB, EB\_{HZ}, M, PI, and Dir of considered parameters by sample size under the gamma-gamma model. In parentheses are the RB (\%) and RRMSE (\%) values for EB\_clsd.}
\label{sim1.1_EB_RB_RE}
\end{table}
\subsubsection{GLMM}
In this subsection, we generate data from the model (\ref{Hobza}) with $\nu = \{1,2.5,5\}$, $\bm{\beta} = (0.5,0.05)^{T}$, and $\phi^{2} =0.1^{2}$. To gain a deeper understanding of the predictors proposed by Hobza et al. (2020), the best predictors, denoted as BP\_HZ and BP\_M, were additionally calculated using the value of the true model parameter.
What is most interesting about this configuration is that the empirical best predictors for the gamma-gamma model remain competitive, even though the data are generated from the gamma GLMM. For $\nu \in \{1.00, 2.50\}$, the RRMSEs of the EB and EB\_clsd predictors are below the RRMSEs of the alternative predictors. For Gini, the EB predictor has uniformly smallest RRMSE. For the other parameters and $\nu = 5$, the increase in RRMSE from EB, relative to EB\_HZ is slight. As illustrated in Figure~\ref{RE_GMM}, the problems with the EB\_HZ and $M$ predictors occur due to the effect of the variance of parameter estimators. The HZ and $M$ predictors calculated with the true parameters (abbreviated BP\_HZ and BP\_M in Figure~\ref{RE_GMM}) are most efficient in terms of RRMSE for the GLMM configurations. This result has two implications. The first is that estimating the model parameters with the Laplace approximation for the likelihood may not work well when the skewness of the distribution is large. The second is that the empirical best predictors for the gamma-gamma model appear robust to this form of model misspecification. Specifically, the empirical best predictors for the gamma-gamma model maintain reasonable efficiency, even when the data are generated from the gamma GLMM.
The results for the PI and Dir predictors for the GLMM are similar to the results for these predictors for the gamma-gamma model. The use of the estimated mean as the predictor for a nonsampled element for the PI predictor causes the PI predictor to have an important bias for nonlinear parameters. The direct estimator is also inefficient for nonlinear parameters as a result of the small area sample size.
\begin{table}[H]
\centering
\scalebox{0.7}{
\begin{tabular}{ccc|ccccc|ccccc}
\toprule
{\large{\bfseries Parameter }} & {\large{\bfseries $\nu$}} & {\large{\bfseries $n_{i}$}} & {\large{\bfseries EB}} & {\large{\bfseries EB\_HZ}} & {\large{\bfseries $M$}} & {\large{\bfseries PI}} & {\large{\bfseries Dir}} & {\large{\bfseries EB}} & {\large{\bfseries EB\_HZ}} & {\large{\bfseries $M$}} & {\large{\bfseries PI}} & {\large{\bfseries Dir}} \\
\midrule
h1 & 1.00 & 10 & 0.13 (0.13) & -0.26 & -0.83 & -0.83 & -0.07 & 13.20 (13.17) & 13.28 & 13.31 & 13.29 & 30.23 \\
& & 20 & 0.11 (0.12) & -0.19 & -0.67 & -0.67 & 0.18 & 11.88 (11.86)& 11.98 & 11.99 & 11.98 & 20.31 \\
& 2.50 & 10 & 0.10 (0.10)& -0.38 & -1.02 & -1.02 & 0.12 & 10.24 (10.22) & 10.38 & 10.41 & 10.40 & 19.16 \\
& & 20 & 0.02 (0.02)& -0.30 & -0.79 & -0.79 & -0.03 & 8.81 (8.79)& 8.93 & 8.94 & 8.93 & 12.75 \\
& 5.00 & 10 & 0.15 (0.14)& -0.34 & -0.79 & -0.79 & 0.03 & 8.71 (8.68)& 8.64 & 8.66 & 8.65 & 13.49 \\
& & 20 & 0.08 (0.08)& -0.23 & -0.55 & -0.55 & -0.00 & 7.07 (7.05) & 7.02 & 7.03 & 7.02 & 8.98 \\
\midrule
h2 & 1.00 & 10 & 0.44 & 1.44 & 0.95 & 224.52 & 25.76 & 21.13 & 21.54 & 21.51 & 226.07 & 70.23 \\
& & 20 & 0.24 & 1.20 & 0.85 & 222.61 & 11.63 & 19.73 & 20.03 & 20.01 & 224.24 & 44.09 \\
& 2.50 & 10 & 0.17 & 0.25 & -0.28 & 77.71 & 10.02 & 13.02 & 13.14 & 13.14 & 79.11 & 33.13 \\
& & 20 & 0.02 & 0.19 & -0.16 & 77.31 & 4.29 & 11.78 & 11.89 & 11.89 & 78.59 & 21.82 \\
& 5.00 & 10 & 0.08 & -0.21 & -0.57 & 42.28 & 5.61 & 10.10 & 10.04 & 10.05 & 43.58 & 21.16 \\
& & 20 & -0.01 & -0.15 & -0.36 & 42.08 & 2.41 & 8.58 & 8.54 & 8.54 & 43.10 & 14.14 \\
\midrule
h3 & 1.00 & 10 & 0.10 & -0.31 & -0.84 & -26.12 & -2.56 & 15.01 & 15.06 & 15.10 & 30.32 & 35.41 \\
& & 20 & 0.13 & -0.20 & -0.63 & -26.05 & -0.99 & 13.71 & 13.78 & 13.79 & 29.85 & 24.84 \\
& 2.50 & 10 & 0.03 & -0.54 & -1.17 & -23.10 & -2.70 & 11.26 & 11.40 & 11.44 & 25.73 & 22.60 \\
& & 20 & -0.00 & -0.40 & -0.88 & -22.89 & -1.33 & 9.83 & 9.93 & 9.95 & 25.03 & 15.58 \\
& 5.00 & 10 & 0.16 & -0.38 & -0.84 & -18.61 & -2.29 & 9.35 & 9.29 & 9.31 & 20.87 & 16.09 \\
& & 20 & 0.11 & -0.26 & -0.59 & -18.35 & -1.09 & 7.77 & 7.72 & 7.74 & 20.06 & 11.20 \\
\midrule
h4 & 1.00 & 10 & -0.07 & -0.61 & -0.66 & -82.81 & -9.07 & 5.62 & 5.82 & 5.84 & 83.06 & 19.03 \\
& & 20 & -0.07 & -0.56 & -0.63 & -70.24 & -4.13 & 5.30 & 5.47 & 5.49 & 70.51 & 12.21 \\
& 2.50 & 10 & -0.07 & -0.66 & -0.79 & -82.18 & -9.13 & 6.40 & 6.49 & 6.52 & 82.47 & 21.13 \\
& & 20 & -0.05 & -0.58 & -0.75 & -70.05 & -3.99 & 6.05 & 6.13 & 6.16 & 70.34 & 13.60 \\
& 5.00 & 10 & 0.12 & -0.26 & -0.42 & -80.77 & -8.95 & 6.73 & 6.76 & 6.78 & 81.08 & 21.85 \\
& & 20 & 0.12 & -0.21 & -0.42 & -69.06 & -3.99 & 6.32 & 6.34 & 6.36 & 69.38 & 14.16 \\
\bottomrule
\end{tabular}
}
\caption{RB($\%$) and RRMSE($\%$) for EB, EB\_clsd, EB\_HZ, $M$,PI,Dir of considred parametrs by sample size under the gamma mixed model}
\label{RB_RE_GMM}
\end{table}
\begin{figure}[H]
\captionsetup{font=footnotesize}
\centering
\includegraphics[scale=0.7]{EB_RE_HZ.pdf}
\caption{Box plots of RRMSE(\%) for considered predictors under the GLMM. $BP$ indicates the proposed best predictor with true model parameters.}
\label{RE_GMM}
\end{figure}
\subsection{Simulation 2}
In Simulation 1, we observed that the empirical best predictors for the gamma-gamma model perform well even when the model is misspecified. Thus, we concentrate on the performance of the MSE estimators under the framework of the gamma-gamma model. We simulate data from the gamma-gamma model. The predictors and MSE estimators are cunstructed under the assumptions of the gamma-gamma model.
We evaluate the MSE estimators described in Section 3.2 on the basis of two criteria. The first is a measure of the relative bias of the MSE estimator, as a measure of the unconditional MSE of the predictor. This is defined as
\begin{align}\label{RB_MSE}
RB_{A}^{\rm uncond} = \frac{(10000D)^{-1}\sum_{i=1}^{D}\sum_{m=1}^{10000}{\rm mse}_{i}^{(m, A)}-{\rm MSE^{UCond}}}{{\rm MSE^{UCond}}},
\end{align}
where ${\rm mse}_{i}^{(m, A)}$ is the type $A$ MSE estimator obtained in MC simulation $m$, \newline $A\in \{\rm noBC, Add, Mult, HM, Comp, S, D\}$ and ${\rm MSE}^{\rm UCond} = {(10000D)}^{-1}\sum_{i=1}^{D}\sum_{m=1}^{10000}(\hat{\theta}_{i}^{\rm EB(m)}-{\theta}_{i}^{(m)})^{2}$. Note that the proposed estimators can also be regarded as the estimators for the conditional MSE defined as $E\{(\hat{\theta}_{i}^{EB}-{\theta}_{i})^{2}\mid\bm{y}_{is}\}$. Thus, we define the conditional RB as
\begin{align}\label{RB_MSE_Cond}
RB_{A}^{\rm cond} = \frac{(10000D)^{-1}\sum_{i=1}^{D}\sum_{m=1}^{10000}{\rm mse}_{i}^{(m, A)}-{\rm MSE^{\rm cond}}}{{\rm MSE^{\rm cond}}},
\end{align}
where
\begin{align}
{\rm MSE}^{\rm Cond} &=D^{-1}\sum_{i=1}^{D}( \bar{M}_{1i}+\bar{M}_{2i}),
\end{align}
$\bar{M}_{1i} =(10000)^{-1}\sum_{m=1}^{M}(\hat{\theta}_{i}^{{\rm B}(m)}-\hat{\theta}_{i}^{(m)})^{2}$ and $\bar{M}_{2i} =(10000)^{-1}\sum_{m=1}^{M}(\hat{\theta}_{i}^{{\rm EB}(m)}-\hat{\theta}_{i}^{{\rm B}(m)})^{2}$. \cite{lohr2009jackknife} evaluate the conditional relative bias of the MSE estimators in their simulations, and \cite{booth1998standard} discuss the value of the conditional MSE estimators in prediction problems.
Several of the MSE estimators defined in Section 3 incorporate corrections for the estimators of the bias of the estimator of the leading term. We therefore check whether there exists a bias for the leading term estimators. For this, define a test-statistic as
\begin{align}\label{leadtest}
T^{\rm Bias} = \frac{\bar{\omega}}{{\rm sd}_{\omega}/\sqrt{10000}},
\end{align}
where $\omega^{(m)}=D^{-1}\sum_{d=1}^{D}(\hat{M}_{1i}^{(m)}-{M}_{1i}^{(m)})$, $\bar{\omega}=10000^{-1}\sum_{m=1}^{10000}\omega^{(m)}$ and\\ ${\rm sd}_{\omega} =\sqrt{(10000-1)^{-1}\sum_{m=1}^{10000}(\omega^{(m)}-\bar{\omega})^2}$.
Figure~\ref{Sim2} displays the relative biases of the alternative MSE estimators. The single-bootstrap MSE estimator ($S$) has a positive bias for $\bar{Y}$, $Q_{0.25}$, and $Q_{0.75}$. The double-bootstrap procedure ($D$) can over-correct this bias, producing important negative biases when $\alpha = 1$. For Gini, both $S$ and $D$ have negative biases. The MSE estimator noBC consistently has relative bias close to zero. The bias corrections Comp and HM lead to slight increases in the estimated MSE. The t-statistics in Table~\ref{BiasTeststat} shed insight into the relative biases of the noBC, Comp, and HM MSE estimators. The estimator of the leading term does not have a significant bias for all parameters, except for the Gini coefficient. Therefore, the noBC MSE estimator has RB close to zero for $\bar{Y}$, $Q_{0.25}$, and $Q_{0.75}$. For the Gini coefficient, the bias of the estimator of the leading term is important. As illustrated by the conditional relative bias of the MSE estimators, the HM and Comp bias corrections effectively correct the bias of the estimator of the leading term for the Gini coefficient.
\begin{figure}[H]
\centering
\includegraphics[width=15cm]{Sim2.pdf}
\caption{RB (\%) of MSE Estimators}
\label{Sim2}
\end{figure}
\begin{table}[H]
\centering
\begin{tabular}{c|ccc}
\toprule\toprule
\multirow{2}{*}{Parameter}&\multicolumn{3}{c}{$\alpha$}\\
\cline{2-4}
& 1 & 2.5 & 5 \\ \hline
h1 & -0.13 & 0.35 & -0.96 \\
h2 & -0.08 & -0.15 & -0.03 \\
h3 & -0.20 & 0.29 & -0.45 \\
h4 & -4.62 & -5.60 & -4.85 \\
\bottomrule\bottomrule
\end{tabular}
\caption{ $T^{\rm Bias}$ by the area parameter for each scenario}
\label{BiasTeststat}
\end{table}
\subsection{Simulation 3}
In this simulation study, the setup described in Section 5.1.1 is slightly modified to take into account an informative sample design. Instead of using simple random sampling when selecting $n_{i}$ units within an area $i$, we use systematic sampling. The inclusion probability for element $j$ within area $i$ is defined as
$$\pi_{ij} =
\frac{n_{i}\exp(ax_{ij}+by_{ij}+\tau_{ij}/20)}{\sum_{k=1}^{N_{i}}\exp(ax_{ik}+by_{ik}+\tau_{ik}/20)},$$
where $\tau_{ij}\stackrel{iid}{\sim}\text{Gamma}(\delta,\delta)$. We hold $a$ constant at $0.05$. We vary $b$ from $0.05$ to $0.2$ with an increment of 0.05 to alter the degree of the informativeness in the sample design. Then, the performance of the alternative predictors is examined through the RB and RRMSE defined in Section 5.1.1.
Table~\ref{sim3_RB_RE} contains the average RBs and RRMSEs of the alternative predictors, where the average is across areas of the same sample size. EB\_INFO always attains smallest RBs and RRMSEs for all scenarios. The RBs of the EB, EB\_HZ, and $M$, that ignore the unequal sample probabilities, significantly increase as $b$ increases. Among these predictors, $M$ performs best in terms of both RB and RRMSE. EB was robust to the model misspecification, but seems to be more sensitive to the informative sampling strategy, as shown by its bigger RBs. Fortunately, the EB-Info procedure offers a theoretically defensible solution for the informative sample design.
\begin{sidewaystable}[ht]
\centering
\scalebox{0.8}{
\begin{tabular}{ccc|cccccc|cccccc}
\toprule \toprule
\multirow{2}{*}{\large{\bfseries Parameter }} &\multirow{2}{*} {\large{\bfseries $\alpha$}} & \multirow{2}{*}{\large{\bfseries $b$}} & \multicolumn{6}{c}{\large{\bfseries{RB (\%)}}}& \multicolumn{6}{c}{\large{\bfseries{RRMSE (\%)}}}\\\cmidrule{4-15}
& & &{\large{\bfseries EB\_INFO}} & {\large{\bfseries EB}} & {\large{\bfseries EB\_HZ}} & {\large{\bfseries M}} & {\large{\bfseries PI}} & {\large{\bfseries Dir}} & {\large{\bfseries EB\_INFO}} & {\large{\bfseries EB}} & {\large{\bfseries EB\_HZ}} & {\large{\bfseries M}} & {\large{\bfseries PI}} & {\large{\bfseries Dir}} \\
\midrule
h1 & 1.00 & 0.05 & -0.01 & 2.58 & 0.60 & -2.77 & -2.77 & 1.97 & 27.29 & 28.50 & 31.41 & 27.54 & 27.53 & 35.84 \\
& & 0.10 & -0.13 & 5.21 & 2.98 & -0.36 & -0.36 & 4.90 & 26.41 & 30.75 & 30.59 & 26.90 & 26.89 & 40.89 \\
& & 0.20 & -0.42 & 10.84 & 8.12 & 4.73 & 4.73 & 11.34 & 25.85 & 36.58 & 31.51 & 28.52 & 28.50 & 55.51 \\
& 5.00 & 0.05 & -0.03 & 2.55 & -2.82 & -0.75 & -0.75 & 2.81 & 13.03 & 15.13 & 31.11 & 13.82 & 13.82 & 21.30 \\
& & 0.10 & -0.13 & 5.17 & -0.75 & 1.66 & 1.66 & 6.67 & 13.05 & 18.61 & 30.33 & 13.42 & 13.42 & 28.43 \\
& & 0.20 & -0.64 & 9.97 & 3.54 & 6.07 & 6.07 & 14.08 & 13.51 & 23.67 & 27.19 & 15.29 & 15.28 & 38.65 \\
\midrule
h2 & 1.00 & 0.05 & 0.13 & 2.38 & 4.36 & 1.49 & 147.83 & 21.20 & 34.04 & 35.03 & 38.59 & 34.30 & 178.81 & 76.84 \\
& & 0.10 & 0.12 & 4.78 & 6.66 & 3.86 & 152.29 & 24.42 & 33.50 & 37.06 & 38.77 & 34.63 & 188.27 & 88.72 \\
& & 0.20 & 0.04 & 9.92 & 11.60 & 8.83 & 161.54 & 32.29 & 33.43 & 42.68 & 40.29 & 37.83 & 206.04 & 141.07 \\
& 5.00 & 0.05 & 0.12 & 2.19 & -5.32 & -2.70 & 18.32 & 7.69 & 14.86 & 16.58 & 36.88 & 16.20 & 25.50 & 31.35 \\
& & 0.10 & 0.14 & 4.44 & -3.76 & -0.73 & 20.44 & 11.51 & 14.96 & 19.69 & 36.39 & 15.47 & 28.48 & 43.40 \\
& & 0.20 & -0.03 & 8.64 & -0.08 & 3.03 & 24.40 & 19.73 & 15.22 & 24.23 & 31.67 & 15.99 & 33.40 & 62.77 \\
\midrule
h3 & 1.00 & 0.05 & -0.00 & 2.48 & 0.47 & -2.67 & -11.30 & 0.38 & 28.80 & 29.98 & 33.56 & 29.06 & 32.41 & 41.65 \\
& & 0.10 & -0.13 & 4.99 & 2.71 & -0.38 & -8.93 & 3.23 & 28.17 & 32.38 & 33.03 & 28.70 & 30.32 & 47.63 \\
& & 0.20 & -0.34 & 10.49 & 7.69 & 4.60 & -3.80 & 9.72 & 27.77 & 38.32 & 33.60 & 30.51 & 28.47 & 62.14 \\
& 5.00 & 0.05 & -0.04 & 2.61 & -2.59 & -0.36 & -3.90 & 1.18 & 14.03 & 16.09 & 33.60 & 14.72 & 16.33 & 25.02 \\
& & 0.10 & -0.16 & 5.29 & -0.50 & 2.13 & -1.41 & 5.15 & 14.15 & 19.61 & 33.03 & 14.57 & 14.88 & 31.19 \\
& & 0.20 & -0.69 & 10.28 & 3.97 & 6.74 & 3.19 & 12.93 & 14.67 & 24.83 & 29.21 & 16.78 & 15.06 & 40.09 \\
\midrule
h4 & 1.00 & 0.05 & -0.00 & 0.11 & -1.33 & -1.64 & -55.40 & -6.83 & 5.58 & 5.59 & 5.81 & 5.89 & 55.74 & 15.78 \\
& & 0.10 & -0.01 & 0.23 & -1.24 & -1.56 & -54.96 & -6.76 & 5.57 & 5.58 & 5.77 & 5.85 & 55.31 & 15.71 \\
& & 0.20 & -0.03 & 0.46 & -1.05 & -1.37 & -54.05 & -6.59 & 5.54 & 5.57 & 5.70 & 5.77 & 54.40 & 15.60 \\
& 5.00 & 0.05 & -0.01 & 0.45 & 3.39 & 2.82 & -35.65 & -6.42 & 6.27 & 6.29 & 7.65 & 6.89 & 36.23 & 17.62 \\
& & 0.10 & -0.05 & 0.89 & 3.91 & 3.29 & -34.81 & -6.09 & 6.24 & 6.30 & 7.95 & 7.06 & 35.40 & 17.50 \\
& & 0.20 & -0.25 & 1.61 & 4.56 & 3.94 & -33.34 & -5.68 & 6.26 & 6.43 & 8.09 & 7.36 & 33.94 & 17.42 \\
\bottomrule
\end{tabular}
}
\caption{RB ($\%$) and RRMSE ($\%$) for EB, EB\_{HZ}, M, PI, Dir of considered parameters by sample size under the informative sample design}
\label{sim3_RB_RE}
\end{sidewaystable}
\section{Application to Ohio Soil Erosion Data}
This analysis is based on a survey of cropland conducted as part of the Conservation Effects Assessment Project (CEAP), which is a nationwide evaluation to quantify the effectiveness of conservation efforts on croplands. The sample for the CEAP survey contains a subset of crop fields in a massive panel survey called the National Resources Inventory (NRI). The NRI measures numerous variables related to land cover/use and soil characteristics on non-federal US lands. One variable of interest is sheet and rill erosion, soil loss due to rainfall or water runoff.
There are two measurements of the soil erosion: the Universal Soil Loss Equation (USLE) and Revised Universal Soil Loss Equation version 2 (RUSLE2). The former is a classic soil erosion measurement (See Wischmeier and Smith, 1965) and the latter is the enhaced version of the former in that it considers additional variables and daily-based factors. USLE is obtained at every NRI sample point. The measure of RUSLE2 that we use is only computed for the CEAP subsample. Our objective is to estimate functions of RUSLE2 for Ohio counties using the USLE as a covariate.
Let $y_{ij}$ and $x_{ij}$ denote the RUSLE2 and USLE for the $j$th sample point of the $i$th county in Ohio. Note that 73 counties among 88 counties are sampled, and nonsampled counties are excluded in this analysis. Further, we regress $log(1/\pi_{ij})$ on $x_{ij}$ and $y_{ij}$ with areas fixed effects. The coefficients for both $a$ and $b$ were not significant, so we do not consider the predictor under an informative sample design. We fit the gamma-gamma model (\ref{Gam-Gam}) and the GLMM (\ref{Hobza}) to the sample.
The fitted models are assessed by using the generalized residuals. The gneralized residual for the gamma-gamma model is defined as
$r_{ij} = G^{-1}(y_{ij}\mid \bm{x}_{ij}, \hat{u}_{i};\hat{\bm{\psi}})$, where $G$ is the CDF of a gamma distribution with shape parameter $\hat{\alpha}$
and rate parameter $\mbox{exp}(\bm{x}_{ij}'\hat{\bm{\gamma}})\hat{u}_{i}$ with
$$\hat{u}_{i}= \frac{ \hat{\alpha} + \hat{\delta}}{ \sum_{j = 1}^{n_{i}}y_{ij}\mbox{exp}(\bm{x}_{ij}'\hat{\bm{\gamma}}) + \hat{\tau} }.$$ For the GLMM, the generalized residual is defined as $r_{ij} = G^{-1}(y_{ij}\mid \bm{x}_{ij}, \hat{u}_{i};\hat{\bm{\psi}}^{GLMM})$, where $G$ is the CDF of a gamma distribution with shape parameter $\hat{\nu}$ and rate parameter $\hat{\nu}^{-1}\mbox{exp}(-\bm{x}_{ij}'\hat{\bm{\beta}} + \hat{v}_{i})$ with $\hat{v}_{i}$ being the predicted random effect from {\tt ranef} in {\tt R}. The set of the generalized residuals $\{r_{ij}:j=1,\hdots,n_{i}, i=1,\hdots,73\}$ behaves like a sample from the uniform distribution under the correct model. The justification for the fitted model can be made through comparing the distribution of $\{\Phi^{-1}(r_{ij}):j=1,\hdots,n_{i}, i=1,\hdots,73\}$ with the normal distribution. As shown in Figure \ref{gen_res}, the residuals from both models almost fall along the 45 degree reference line. Both the GLMM and gamma-gamma models appear to fit the data adequately.
We next consider the following county-level parameters: the mean ($\bar{Y}_{i}$), the $25$th ($Q_{0.25})$ \& the $75$th ($Q_{0.75}$) quantiles, and the proportion greater than $0.22$ ($P_{m}$) of a county $i$, $i=1,\hdots,73$. The value used for $P_{m}$ is the sample median of RUSLE2 in Ohio. Based on the simulation results in Section 5.1.1, we take into account EB, EB\_HZ, and M in Section 3. As shown in Figure \ref{data_pred}, the predicted values produced by each predictor are distributed similarly for a given parameter. We construct normal theory 95\% confidence intervals for the EB county predictors using the noBC, HM, and S MSE estimators. The double-bootstrap produced negative MSE estimators for some counties, so this procedure cannot be used to construct CIs. CIs constructed with S tend to be longer than CIs constructed by noBC and HM. This reflects the result of the simulation study where the $S$ MSE estimator tended to have a positive bias.
\begin{figure}[H]
\centering
\includegraphics[width = 16cm]{generalized_errors.pdf}
\caption{Generalized residuals of the gamma-gamma model (Left) and the GLMM (Right)}
\label{gen_res}
\end{figure}
\begin{table}[H]
\centering
\begin{tabular}{cc|cc}
\toprule
$\hat{\bm{\psi}}^{\rm Gam-Gam}$ & Estimates & $\hat{\bm{\psi}}^{\rm GLMM}$ & Estimates \\
\hline
$\hat{\alpha}$ & 1.659 (1.617, 1.700) & $\hat{\nu}$& 1.381 (1.224, 1.539) \\
$\hat{\delta}$ & 4.922 (3.931, 5.912) & $\hat{\phi}^{2}$ &0.458 (0.370, 0.546) \\
$\hat{\gamma}_{0}$ & 2.183 (0.365, 4.003) & $\hat{\beta}_{0}$&-1.602 (-1.734, -1.471)\\
$\hat{\gamma}_{1}$ & -0.156 (-0.19, -0.126) & $\hat{\beta}_{1}$& 0.153 (0.139, 0.168) \\
\bottomrule
\end{tabular}
\caption{Estimates of the model parameters with their confidence intervals}
\label{model_par_est}
\end{table}
\begin{figure}[H]
\centering
\includegraphics[width=15cm]{Prediction.pdf}
\caption{Predicted values of the considered parameters by EB, EB\_HZ, M}
\label{data_pred}
\end{figure}
\begin{figure}[H]
\centering%
\includegraphics[width = 18cm]{county_plots.pdf}
\caption{Prediction of the considered small area parameters with EB predictor. Counties colored by grey are nonsampled.}
\label{county_plot}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width = 16cm]{CI.pdf}
\caption{95 \% normal-theory Confidence Intervals. Counties are sorted by their sample size.}
\label{data_CI}
\end{figure}
\section{Conclusion}
\hspace{.2 in } In this work, we demonstrate that the gamma distribution is a useful model for constructing small area predictors with a skewed response variable. We focus heavily on the EB predictor for the gamma-gamma model proposed in \cite{Graf2019}. We demonstrate that this predictor has good properties, even when the true model is the gamma-GLMM. We also demonstrate that the general MSE estimator of \cite{https://doi.org/10.48550/arxiv.2210.12221} has good properties for in the context of the gamma-gamma model. Finally, we extend the gamma-gamma model to an informative sample design. Our results generally provide more support for the gamma-gamma model than for the gamma-GLMM.
\bibliographystyle{agsm}
|
\chapter{Chapter Heading}
\label{introA}
Use the template \emph{appendix.tex} together with the Springer document class SVMono (monograph-type books) or SVMult (edited books) to style appendix of your book.
\section{Section Heading}
\label{sec:A1}
Instead of simply listing headings of different levels we recommend to let every heading be followed by at least a short passage of text. Furtheron please use the \LaTeX\ automatism for all your cross-references and citations.
\subsection{Subsection Heading}
\label{sec:A2}
Instead of simply listing headings of different levels we recommend to let every heading be followed by at least a short passage of text. Furtheron please use the \LaTeX\ automatism for all your cross-references and citations as has already been described in Sect.~\ref{sec:A1}.
For multiline equations we recommend to use the \verb|eqnarray| environment.
\begin{eqnarray}
\vec{a}\times\vec{b}=\vec{c} \nonumber\\
\vec{a}\times\vec{b}=\vec{c}
\label{eq:A01}
\end{eqnarray}
\subsubsection{Subsubsection Heading}
Instead of simply listing headings of different levels we recommend to let every heading be followed by at least a short passage of text. Furtheron please use the \LaTeX\ automatism for all your cross-references and citations as has already been described in Sect.~\ref{sec:A2}.
Please note that the first line of text that follows a heading is not indented, whereas the first lines of all subsequent paragraphs are.
\begin{figure}[t]
\sidecaption[t]
\includegraphics[scale=.65]{figure}
\caption{Please write your figure caption here}
\label{fig:A1}
\end{figure}
\begin{table}
\caption{Please write your table caption here}
\label{tab:A1}
\begin{tabular}{p{2cm}p{2.4cm}p{2cm}p{4.9cm}}
\hline\noalign{\smallskip}
Classes & Subclass & Length & Action Mechanism \\
\noalign{\smallskip}\hline\noalign{\smallskip}
Translation & mRNA$^a$ & 22 (19--25) & Translation repression, mRNA cleavage\\
Translation & mRNA cleavage & 21 & mRNA cleavage\\
Translation & mRNA & 21--22 & mRNA cleavage\\
Translation & mRNA & 24--26 & Histone and DNA Modification\\
\noalign{\smallskip}\hline\noalign{\smallskip}
\end{tabular}
$^a$ Table foot note (with superscript)
\end{table}
\part{Part Title}
\noindent Use the template \emph{part.tex} together with the document class SVMono (monograph-type books) or SVMult (edited books) to style your part title page and, if desired, a short introductory text (maximum one page) on its verso page.
\end{partbacktext}
\section*{Problems of Chapter~\ref{intro}}
\begin{sol}{prob1}
The solution\index{problems}\index{solutions} is revealed here.
\end{sol}
\begin{sol}{prob2}
\textbf{Problem Heading}\\
(a) The solution of first part is revealed here.\\
(b) The solution of second part is revealed here.
\end{sol}
\chapter{
\huge{An Introduction of System-Scientific Approaches to Cognitive Security}
\\
\quad \\
\Large{Linan Huang and Quanyan Zhu}}
\label{chap:intro}
\chaptermark{An Introduction of System-Scientific Approaches to Cognitive Security}
\abstract{
Human cognitive capacities and the needs of human-centric solutions for ``\textit{Industry 5.0}'' make humans an indispensable component in \acrfullpl{cps}, referred to as \acrfullpl{hcps}, where \acrshort{ai}-powered technologies are incorporated to assist and augment humans.
The close integration between humans and technologies in Section \ref{sec: AI-Powered Human-Cyber-Physical Systems} and cognitive attacks in Section \ref{subsec:Cognitive Attack} poses emerging security challenges, where attacks can exploit vulnerabilities of human cognitive processes, affect their behaviors, and ultimately damage the \acrshort{hcps}.
\newline\indent
Defending \acrshortpl{hcps} against cognitive attacks requires a new security paradigm, which we refer to as ``\textit{cognitive security}'' in Section \ref{subsec:Cognitive Security}.
The vulnerabilities of human cognitive systems and the associated methods of exploitation distinguish cognitive security from ``\textit{cognitive reliability}'' and give rise to a distinctive \textit{CIA triad}, as shown in Sections \ref{subsubsec:Cognitive Reliability vs. Cognitive Security} and \ref{subsubsec:CIAtriad}, respectively. Section \ref{subsubsec:Cognitive and Technical Defenses} introduces cognitive and technical defense methods that deter the kill chain of cognitive attacks and achieve cognitive security.
System scientific perspectives in Section \ref{sec:System-Scientific Perspectives for Cognitive Security} offer a promising direction to address the new challenges of cognitive security by developing \textit{quantitative, modular, multi-scale, and transferable} solutions.
Fig. \ref{fig:chap1structure} illustrates the structure of Chapter \ref{chap:intro}.
}
\keywords{Cognitive Security, Cognitive Reliability, Human-Cyber-Physical Systems, Human-Centered AI, Cognitive Attacks, Cognitive Vulnerability, Cyber Security, Data Science, System Science, CIA Triad}
\begin{figure}[tbh]
\sidecaption[t]
\includegraphics[width=1 \textwidth]{fig/chap1structure1.pdf}
\caption{
The structure of Chapter \ref{chap:intro} concerning \acrshort{ai}-powered HCPS in Section \ref{sec: AI-Powered Human-Cyber-Physical Systems}, cognitive security in Section \ref{sec:Cognitive Security in HCPS}, and system science in Section \ref{sec:System-Scientific Perspectives for Cognitive Security} as the object, goal, and method of this book, respectively.
}
\label{fig:chap1structure}
\end{figure}
\section{AI-Powered Human-Cyber-Physical Systems}
\label{sec: AI-Powered Human-Cyber-Physical Systems}
\acrfullpl{cps} are ``smart systems that include engineered interacting networks of physical and computational components'', as defined by the National Institute of Standards and Technology (NIST) in $2017$ \cite{griffor2017framework}.
Despite the increasing automation and intelligence in \acrshortpl{cps}, humans play indispensable roles in accomplishing \acrshort{cps} tasks, as illustrated by the mission stack in Fig. \ref{fig:TwoTrends}.
The rapid development of \acrfullpl{ai} and big data has facilitated a close integration of \acrshort{ai}-powered technologies along with the mission completion, as illustrated by the \acrshort{ai} stack in Fig. \ref{fig:TwoTrends}.
We elaborate on the human-involved mission stack and the \acrshort{ai} stack in Sections \ref{subsec:human roles in CPS} and \ref{subsec:mission stack and AI stack}, respectively.
\begin{figure}[tbh]
\sidecaption[t]
\includegraphics[width=1 \textwidth]{fig/TwoTrends3.pdf}
\caption{
The overview diagram of human roles and \acrshort{ai}-powered technologies in \acrfullpl{cps}.
The mission stack at the top illustrates the task-driven flow chart that consists of human, cyber, and physical processes in green, blue, and yellow, respectively.
The \acrshort{ai} stack at the bottom illustrates the \acrshort{ai}-powered technologies that interact with the human-cyber-physical process to assist with mission completion.
The \acrshort{ai}-powered technologies are designed to be scalable and transferable. In addition, they need to be customized, explainable, ethical, respectful of privacy, and trustworthy to achieve human-centric objectives.
}
\label{fig:TwoTrends}
\end{figure}
\subsection{Human Roles in Mission Stack}
\label{subsec:human roles in CPS}
Humans \textit{cannot} and \textit{should not} be superseded in \acrshortpl{cps}.
On the one hand, due to their distinguished cognitive power and analytical capabilities (e.g., logical reasoning, symbolic abstraction, knowledge transfer, understanding others' intent and emotions, etc.), humans have been playing irreplaceable roles, including determining demands, designing mechanisms, and responding to incidents.
On the other hand, the inevitable progress of ``Industry 5.0'' \cite{industry5} has reemphasized human-centric solutions to create personalized products, services, and experiences.
The increased level of automation aims to \textit{support} rather than \textit{supersede} humans.
The mission stack in Fig. \ref{fig:TwoTrends} summarizes the system-level integration of human touch with cyber and physical processes, referred to as \acrfullpl{hcps}, for mission completion.
After a human participant receives a stimulus, he processes it and outputs responsive behaviors that influence the cyber and physical processes.
The cognitive process can be affected by externalities related to individual factors (e.g., personality, awareness, and expertise), environmental factors (e.g., workload and stress), and social factors (e.g., peer pressure and culture).
Human participants have different roles (e.g., users, operators, security analysts, and administrators) in mission completion, and the associated cognitive process can take different forms. Some examples are provided as follows.
\begin{description}[Example 1]
\item[Example 1] {End users, including employees and contractors, use the computing facilities of the providers to maintain the corporation's normal operations.
For example, the stimulus for employees could be working emails. Their cognitive processes affect the accuracy and timeliness of the phishing\footnote{Phishing is coined as a combination of the words ``password'' and ``fishing'' to describe the practice of tricking Internet users into revealing sensitive data, including passwords.} recognition, which results in either secure behaviors or falling victim to phishing. }
\item[Example 2] {Security analysts investigate alerts in real time for alert triage and response.
The alerts, the triage process, and the response are the stimulus, the cognitive process, and the behaviors, respectively, in this alert management scenario.}
\end{description}
\subsection{Incorporating AI Stack into Mission Stack}
\label{subsec:mission stack and AI stack}
The advances in \acrshort{ai} have accelerated the automation and smartification process in cyber and physical layers, as shown by the blue and yellow double-headed arrows in Fig. \ref{fig:TwoTrends}.
\acrshort{ai} has been widely applied to sensing, control, communication, and data processing in a variety of applications such as biomedical monitoring, robotics systems, and digital twins \cite{Song2022,salau2022recent,groshev2021toward,doghri2022cyber}.
In these \acrshort{cps} applications, \acrshort{ai} not only serves as a technology for reasoning, planning, learning, and processing, but also enables the manipulation of physical objects.
For example, autonomous driving cars adopt \acrshort{ai} to sense the environment (e.g., road condition, weather, and the movement of pedestrians and other cars) and determine the optimal driving setting (e.g., speed, brake, and steer).
\acrshort{ai}-powered technologies in these \acrshort{cps} applications should be scalable \cite{Barmer2021} and transferable \cite{pan2009survey}.
Compared to \textit{technology-based} \acrshort{ai} design that enables humans to adapt to the technical system, \textit{human-centered} \acrshort{ai} aims to design systems that are aware of human cognitive processes to augment human perception, cognition, and decision-making capabilities in the rapidly evolving, complex, and uncertain \acrshort{cps} environment.
Besides the desirable features of scalability and transferability for \acrshort{ai}-powered technologies, we further require human-centered \acrshort{ai} to be customized, explainable, ethical, respectful of privacy, and trustworthy.
Such requirements are based on a thorough understanding of the human cognitive process and fulfilled by designing proper human-assistive technologies (e.g., \acrfullpl{hmi}), as shown by the down-side and up-side green arrows in Fig. \ref{fig:TwoTrends}, respectively.
The design of human-centered \acrshort{ai} should also be adaptive to \acrshort{cps} applications of different functions, features, requirements, and constraints.
For example, it is essential for a driving-assistive system to be explainable to facilitate trust and minimize a driver's decision time to take proper action \cite{chaczko2020exploration}.
Meanwhile, an \acrshort{ai}-enabled \acrshort{hmi} in the control room of a power grid should feature the following functions \cite{marottowards}:
\begin{itemize}
\item Supporting efficient invocation and dismissal because the number of actions an operator can do in a time window is limited, especially in a time-constrained environment.
\item Taking into account and learn from user behavior and feedback because grid operators are well-trained experts, capable of evaluating the assistant’s answers and providing feedback.
\item Conveying the consequences of user actions because operators may not be able to estimate the risk correctly and timely.
\end{itemize}
\section{Cognitive Security in HCPS}
\label{sec:Cognitive Security in HCPS}
\acrshortpl{cps} have been under threat of various attacks since their emergence.
Established on the mission stack and the \acrshort{ai} stack illustrated in Fig. \ref{fig:TwoTrends}, we incorporate the attack stack and the defense stack in Section \ref{subsec:Attack and Defense Stack} to form the four-stack dissection of \acrshort{hcps} security, as illustrated in Fig. \ref{fig:4stack}.
After the above panorama of \acrshort{hcps} security, we zoom into the focus of his book, i.e., cognitive process, cognitive vulnerability, cognitive attack, and cognitive security, in Sections \ref{subsec:Cognitive Process}, \ref{subsec:Cognitive Vulnerability}, \ref{subsec:Cognitive Attack}, and \ref{subsec:Cognitive Security}, respectively.
\subsection{Attack and Defense Stack}
\label{subsec:Attack and Defense Stack}
Fig. \ref{fig:4stack} illustrates the vertical dissection of \acrshort{hcps} security, including the attack stack and the defense stack in orange and gray, respectively.
In the past decade, attacks have evolved to be targeted, intelligent, multi-staged, and multi-phased.
The typical life cycle of these attacks consists of the following four stages.
In the reconnaissance stage, attackers gather intelligence to determine the attack goal and identify vulnerabilities in the target network. In the planning stage, attackers tailor their strategies to the selected vulnerabilities and choose attack tools. In the execution stage, the attackers deploy the malware to gain an initial foothold, expand access, gain credentials, and escalate privilege. Finally, the attack in the exploitation stage exfiltrates confidential data, interrupts cyber services, or inflicts physical damage.
The four gray boxes in the defense stack illustrate four defensive courses of action. The prevention stage includes the precautionary and proactive defense methods used in advance of attacks. Intrusion prevention techniques, including firewalls and demilitarized zones (DMZ), may be ineffective, especially for advanced attacks such as \acrfullpl{apt}. Therefore, intrusion detection and response are necessary to protect against them. The attribution stage includes post-event analysis, threat intelligence acquisition, and an accountability system.
\begin{figure}[tbh]
\sidecaption[t]
\includegraphics[width=1 \textwidth]{fig/4stack4.pdf}
\caption{
The vertical dissection of \acrshort{hcps} security consists of attack and defense stacks over the mission and \acrshort{ai} stacks.
The four orange boxes in the attack stack illustrate the typical life cycle of an attack.
First, attackers tailor their attack strategies and select attack tools in the planning stage based on the reconnaissance results (e.g., exploitable human and technical vulnerabilities).
Then, an attack is launched and inflicts damage in the execution and exploitation stages, respectively.
As illustrated by the four gray boxes in the defense stack, a typical defense consists of four stages: prevention, detection, response, and attribution.
}
\label{fig:4stack}
\end{figure}
\subsection{Feedforward and Feedback Cognitive Processes}
\label{subsec:Cognitive Process}
Fig. \ref{fig:CogProZoomin} provides a zoomed-in version of the dynamic cognitive process that consists of feedforward and feedback processes, represented by solid and dashed arrows, respectively.
We first illustrate the feedforward cognitive process, where the stimulus is the input and the behavior is the output.
First, perception \cite{bruce2003visual} decodes and gathers the sensory information from external stimuli.
Due to the limited cognitive capacity, attention is used to filter information and store essential or urgent items in working memory.
Then, humans process the information for mental operations, including decision-making, reasoning, and learning, which lead to behaviors.
Based on the tasks and scenarios, human mental operations may need to retrieve past experiences that are stored in the \acrfull{ltm}
The cognitive process also includes several feedback loops listed below:
\begin{itemize}
\item Behaviors have an impact on the \acrshort{cps} and change the subsequent stimulus.
\item Mental operations learn new knowledge and store experience in the \acrshort{ltm}.
\item Besides passively filtering the collected information in the feedforward process, attention also actively affects perception by directing our awareness to relevant stimuli while suppressing distracting information. Such \textit{selective attention} \cite{johnston1986selective} has been demonstrated in phenomena such as the \textit{cocktail party effect} \cite{arons1992review}.
\item \textit{Selective attention} is a result of the mental operations and is usually goal-driven and endogenous (referred to as the \textit{top-down attention}), compared to the \textit{bottom-up attention} that is endogenously driven by the stimuli.
\end{itemize}
\begin{figure}[tbh]
\sidecaption[t]
\includegraphics[width=1 \textwidth]{fig/CogProZoomin3.pdf}
\caption{
A zoomed-in version of the cognitive process that includes sensation, attention, memory, and mental operations.
The feedforward path in solid arrows illustrates the information flow of gathering, filtering, storing, retrieving, and processing.
In the feedback loops, behaviors alter future stimuli, mental operations affect \acrshort{ltm} and attention, and selective attention determines perception, respectively, as illustrated by the dashed arrows.
}
\label{fig:CogProZoomin}
\end{figure}
\subsection{Exploitable Cognitive Vulnerability}
\label{subsec:Cognitive Vulnerability}
Each component of the cognitive process in Fig. \ref{fig:CogProZoomin} possesses distinct vulnerabilities exploitable by attackers.
\cite{cybenko2002cognitive}
We discuss vulnerabilities concerning perception, attention, memory, and mental operations in Sections \ref{subsubsec:v_perception}, \ref{subsubsec:v_attention}, \ref{subsubsec:v_memory}, \ref{subsubsec:v_mental}, respectively.
\subsubsection{Vulnerability of Perception}
\label{subsubsec:v_perception}
Perception is a complex process that involves visual, auditory, somatosensory, olfactory, gustatory, and vestibular systems.
These systems follow patterns that can be exploited by attackers.
For example, knowing the perception limits (e.g., it takes approximately $0.3$ and $1$ seconds to see and hear a signal, respectively) and contributing factors (e.g., how light, color, and noise affect human perception), attackers can create environments where humans are prone to perception errors.
Human perception systems can sometimes undergo distortions of the senses (e.g., visual and auditory illusions) and produce false sensory information.
By understanding the contributing factors and causes of these sensory illusions, attackers can craft phishing websites and emails with fewer identifiable phishing indicators.
Besides exploitable patterns and illusions, human perception is also susceptible to manipulation.
Attacks have adopted many psychological techniques, including priming \cite{molden2014understanding}, to manipulate perception in \acrshortpl{hcps}.
Priming is a well-known phenomenon in psychology wherein the presence of a stimulus impacts how later stimuli are processed or interpreted (e.g., humans recognize the word ``food'' more quickly after the presentation of the word ``kitchen'' than the word ``office'') \cite{lindsay2020attention}.
The majority of priming is positive and increases the sensitivity of the related stimulus, while negative priming \cite{tipper1985negative} slows down the processing speed.
Both \textit{positive} and \textit{negative} priming can be weaponized for perception hacking. For example, attacks may use positive priming to emphasize certain ideas in phishing or use negative priming to deemphasize the phishing indicators.
Based on whether the stimulus is consciously perceptible or not, priming is classified as \textit{supraliminal} and \textit{subliminal}, respectively \cite{elgendi2018subliminal}.
Due to its stealthiness, subliminal priming can be a primary candidate for perception hacking.
As shown in the experiment results \cite{huang2020detecting}, by subtly presenting words or images that are physically or semantically similar to the judgments preferred by the attackers, attackers can influence the accuracy or false alarm rate of the inspectors.
\subsubsection{Vulnerability of Attention}
\label{subsubsec:v_attention}
Attention can be described as an overall level of alertness or ability to engage with surroundings \cite{lindsay2020attention}.
Humans rely on their attention mechanisms to control their limited computational resources in the brain for the maximum information acquisition.
Despite the remarkable success of human attention, it suffers from reduced performance under multitasking, long duration, stress, fatigue, and heavy cognitive load.
Moreover, as the result of the selectiveness of our attention to prevent us from getting lost in irrelevant information, we can go through failures of selection in space and time.
On the one hand, failures of selection in space has been demonstrated in experiments of \textit{change blindness} \cite{simons2005change} and \textit{change deafness} \cite{vitevitch2003change}, where observer does not notice a change in a visual and auditory stimulus, respectively.
On the other hand, experiments of \textit{attentional blink} \cite{shapiro1997attentional} and \textit{repetition blindness} \cite{kanwisher1987repetition} have shown that failures can occur along with time; i.e., when new information (even of small amount) continues to arrive, processing it leads to the miss of other information.
Attacks can exploit these \textit{spatial} and \textit{temporal} attentional vulnerabilities either reactively or proactively.
\textit{Reactive} attention attacks exploit inattention to evade detection and do not attempt to change human attention patterns. For example, many \acrfull{se} and phishing attacks result from a lack of attention. We provide a defense framework against reactive attention attacks in \cite{huang2022advert}.
In contrast, \textit{proactive} attention attacks aim to strategically influence human attention patterns.
For example, an attacker can generate a large volume of feints and hide real attacks among them to overload human operators, delay their responses, and reduce the accuracy of their judgements.
We refer to this new form of attacks as the \acrfull{idos} attacks and present a formal description of \acrshort{idos} attacks, their impacts, and associated defense methods in \cite{huang2022radams}.
\subsubsection{Vulnerability of Memory}
\label{subsubsec:v_memory}
Relying on networks of neurons in the brain, human memory suffers from restricted capacity, limited speed of information storage and retrieval, forgetting, and memory errors.
While digital storage devices share the first two memory vulnerabilities, the latter two are unique to human memory.
According to \citet{schacter2002seven}, the latter two belong to the \textit{sins of omissio} and \textit{commission}, respectively.
Forgetting is the spontaneous or gradual loss of information already stored in an individual's short- or long-term memory.
Unlike a digital storage device, humans cannot `forget on demand'; i.e., items will linger in memory even then they are no longer needed \cite{sasse2001transforming}.
Memory errors refer to the wrong recall of information. This can include remembering things that have not happened, giving the wrong source for a memory, or making up things that did not happen.
Memory errors are caused in part by the structure of neuron networks as well as a feature of human memorization.
As shown in the Deese–Roediger–McDermott paradigm \cite{deese1959prediction}, humans incorrectly recall an absent word as it is related to a list of words that belong to a common theme.
Many factors (e.g., the degree of attention, motivation, emotional state, and environment where memorization takes place) can affect human memory.
For example, the \textit{emotional enhancement of memory} \cite{hamann2001cognitive} has demonstrated that emotional stimuli are more easily remembered than neutral stimuli.
Human memory vulnerabilities directly lead to security risks. For example, humans use simple and meaningful passwords, reuse the same password over different sites, and even write down the passwords to remember them, which makes the passwords insecure.
As will be introduced in Definition \ref{def:Cognitive Reliability}, cognitive reliability methods, including Single Sign-On (SSO)\footnote{Single SSO allows a user to log in several related systems with a single ID and password, it reduces the total number of passwords to remember.} \cite{radha2012survey}, cognitive passwords \cite{zviran1990cognitive}, and graphical passwords \cite{biddle2012graphical}, are introduced to mitigate memory vulnerability concerning passwords.
Attackers can also actively exploit those memory vulnerabilities and manipulate the above factors to create attack vectors.
For example, attackers can reduce the attack frequency to exploit the forgetting vulnerability.
As demonstrated in \cite{sawyer2018hacking} and \cite{kaivanto2014effect}, lower frequency and likelihood of phishing events increase victim’s susceptibility to phishing cyberattacks.
Due to the \textit{suggestibility} of human memory (i.e., humans are inclined to accept and act on the suggestions of others), attackers can design phishing emails to trigger memory errors and inject false memories by designing misleading hints.
Moreover, they can use emotional language to enhance the operator's false memory and facilitate trust.
\subsubsection{Vulnerability of Mental Operations}
\label{subsubsec:v_mental}
Mental operation vulnerabilities primarily refer to a variety of cognitive biases and exploitable traits.
In the history of human development, we have developed cognitive shortcuts and biases for rapid, although less accurate or reasonable, responses to survive in highly dynamic and uncertain environments.
However, those cognitive biases expand the \textit{attack surface} and make humans susceptible to \acrshort{se} attacks.
We list some of the cognitive biases and the potential adversarial exploitation \cite{Rohleder_2019} as follows.
\begin{itemize}
\item \textbf{Anchoring}: An individual's decisions are influenced by a particular reference point (i.e., an ``anchor''). Pretexting \acrshort{se} can create a situation and initial context to increase the attack’s apparent legitimacy and the likelihood of success. Anchoring bias keeps people from questioning the initial impression and accepting the scam.
\item \textbf{Framing}: An individual can draw different conclusions from the same information, depending on how that information is presented (e.g., as a loss or a gain). Attackers can utilize the framing effect to craft the content of phishing emails and distort human risk perception.
\item \textbf{Optimism bias}: People tend to have unrealistic optimism; e.g., overestimating the likelihood of positive events and underestimating the likelihood of negative events.
Users frequently believe others are more susceptible than they are \cite{cox2020stuck}.
Since they think they are immune to attacks, they tend to resist preventive defense measures, such as patching, virus scanning, clearing cache, checking for secure sites before entering credit card information, or paying attention to spear phishing.
\item \textbf{Ingroup bias}: People are social animals and give preferential treatment to in-group members over out-group ones.
An attacker can pretend to be affiliated with the group to gain trust and influence the decisions of group members.
\end{itemize}
In marketing and persuasion, \citet{cialdini2007influence} deduced six principles of influence from experimental and field studies based on exploitable personal traits.
Attackers can also take advantage of the following traits to complete the compromise.
\begin{itemize}
\item \textbf{Reciprocity}:
\acrshort{se} attackers frequently offer victims something to set the stage for reciprocity. For example, people are less likely to refuse an inappropriate request from someone who has provided them with a gift in advance.
\item \textbf{Social proof}:
Individuals are easily influenced by the decisions of a large group of people. An attacker can impersonate the involvement of a victim's friends in order to compel victims to act.
\item \textbf{Authority}:
People tend to conform to authority, and attackers can exploit that by pretending to be a system administrator or a security expert.
\item \textbf{Liking}:
Since it is much easier to influence a person who likes you, attackers can attempt to be likeable by being polite and using concepts, languages, and appearances familiar to the target.
\item \textbf{Scarcity}:
Something in short supply can increase its sense of value. Social engineers may use scarcity (e.g., a limited-time offer) to create a feeling of urgency and spur the victim's action.
\item \textbf{Commitments and consistency}:
Once people make a choice, they receive pressure from others and themselves to adhere to it.
\acrshort{se} attacks can induce the victim to make a seemingly insignificant commitment and then escalate the requests. People tend to accept the escalation of commitment as long as the subsequent requests are consistent with the prior commitment.
\end{itemize}
\subsection{Cognitive Attack}
\label{subsec:Cognitive Attack}
In Section \ref{subsec:Cognitive Vulnerability}, we discuss four types of exploitable vulnerabilities in the cognitive process.
Cognitive attacks refer to the adversarial processes that involve the exploitation of these human vulnerabilities, as defined in Definition \ref{def:Cognitive Attacks}.
\begin{definition}[Cognitive Attacks]
\label{def:Cognitive Attacks}
Cognitive attacks are a class of cyber-physical-human processes that manipulate the behaviors of human actors for malicious purposes, including theft or damage of confidential data and disruption or misdirection of the \acrshort{hcps} services, by exploiting their cognitive vulnerabilities.
\end{definition}
Analogous to cyber kill chains\footnote{
An example cyber kill chains developed by Lockheed Martin in $2011$ can be found at \url{https://www.lockheedmartin.com/en-us/capabilities/cyber/cyber-kill-chain.html}}, we present the kill chain of cognitive attacks in Section \ref{subsubsec:Kill Chain}.
Then, in Section \ref{subsubsec:Examples Paths}, we zoom into the execution phase of the kill chain and use \acrshort{se} and \acrshort{idos} attacks as two examples to illustrate the cross-layer attack paths of cognitive attacks.
\subsubsection{Kill Chain of Cognitive Attacks}
\label{subsubsec:Kill Chain}
Fig. \ref{fig:killchain} illustrates the conceptual kill chain of cognitive attacks that consists of six stages in blue arrows.
We map the six kill-chain stages into the four attack phases of the attack stack in Fig. \ref{fig:4stack} at the bottom of Fig. \ref{fig:killchain} in orange.
In the reconnaissance phase, the attacker analyzes the information collected from the \acrshort{hcps}.
For example, an attacker can formulate the behavioral baseline of a user or an operator during the reconnaissance. Such a baseline can help the attack identify the human vulnerabilities of the target users and exploit them in the execution phase.
In the planning phase, the attackers identify the valuable assets and tailor their \acrfull{ttp} accordingly.
We divide the execution phase into three stages. First, attackers exploit human vulnerabilities directly or indirectly, where we show several examples of exploitation paths in Section \ref{subsubsec:Examples Paths}.
Second, attackers monitor the human target's responses (e.g., emotions and behaviors) and adapt their \acrshort{ttp}.
For example, attackers can choose to increase or decrease the attack frequency for cautious and careless users, respectively.
After the attackers obtain the optimal attack setting, they continue to reinforce the cognitive exploitation and propagate the compromise to other victims related to the initial human target.
For example, after gaining trust, attackers can ask the initial victim to forward emails to his colleague, who will become less doubtful about the legitimacy of the phishing emails.
In the exploitation phase, cognitive exploitation begins to take effect in human behaviors and subsequently the cyber and physical layers.
\begin{figure}[tbh]
\sidecaption[t]
\includegraphics[width=1 \textwidth]{fig/killchain.pdf}
\caption{
The illustration of the kill chain of cognitive attacks concerning the four-stage attack phases of reconnaissance, planning, execution, and exploitation.
In the execution stage, cognitive attacks exploit human vulnerabilities, monitor their responses, and adapt the exploitation \acrshort{ttp} accordingly.
Then, they choose the best attack setting to reinforce the compromise effect and spread it to a group of victims from the initial victim.
}
\label{fig:killchain}
\end{figure}
\subsubsection{Examples Paths of Cognitive Attacks}
\label{subsubsec:Examples Paths}
Exploiting human vulnerabilities is a critical step in the execution attack stage. It can take different forms and involve one or many of the cognitive vulnerabilities described in Section \ref{subsec:Cognitive Vulnerability}.
In Fig. \ref{fig:Cognitiveattack}, we use \acrshort{se} and \acrshort{idos} attacks as two examples to illustrate the various forms and procedures of cognitive attacks.
\acrshort{se} attacks directly manipulate the stimulus (e.g., phishing content) and external factors (e.g., peer pressure) to compromise human users.
In this type of cognitive attack, cognitive compromise is used as the stepping stone to enter the \acrshort{cps} and perpetrate the technical compromise.
However, in other cognitive attacks, the technical compromise may also serve as a precondition to exploit human cognitive vulnerabilities, as illustrated by the attack path of \acrshort{idos} attacks.
\acrshort{idos} attacks first generate a lot of feint attacks to trigger cyber alerts that are displayed through a \acrshort{hmi}.
Human operators investigating these alerts in real-time can suffer from \textit{cognitive overload}, which leads to reduced accuracy and speed in processing the alerts.
Since human operators may be unable to respond to alerts associated with real and significant attacks, these attacks have the potential to disrupt both cyber and physical processes.
Examples of real-world \acrshort{idos} attacks widely exist but are usually implicit in many attack incidents.
The following three incidents in Examples \ref{example:Sony 2011}, \ref{example:BIPS 2013}, and \ref{example:tesla 2021} use \acrfull{ddos} attacks as a ``smoke shell'' to attract security analysts' attention while simultaneously launching other stealthy attacks.
\begin{svgraybox}
\begin{Example}[Sony PSN Data Breach 2011]
\label{example:Sony 2011}
In April 2011, the \acrfull{soc} of Sony was occupied dealing with a \acrshort{ddos} attack, and they overlooked the extraction of confidential information, including the names, addresses, dates of birth, and passwords, of 77 million PlayStation Network (PSN) customers.
Authorities in the U.K. fined the Sony company the equivalent of almost \$400,000 over the failures in the PSN data breach \cite{SonyPSNDDoS}.
\qed{}
\end{Example}
\begin{Example}[BIPS BTC Stolen 2013]
\label{example:BIPS 2013}
Europe’s primary bitcoin payment processor for merchants and free online wallet service, BIPS, became the target of a massive \acrshort{ddos} attack on November 15th, 2013.
While the \acrshort{soc} was busy to get the system back online, the attacker launched a subsequent attack to gain access and compromise several wallets, resulting in the theft of 1,295 BTC (approximately $1$ million dollars) \cite{BIPSddos}.
\qed{}
\end{Example}
\begin{Example}[Tesla Ransomware 2021]
\label{example:tesla 2021}
A recent thwarted cybersecurity attack against Tesla in 2021 planned to use \acrshort{ddos} attack to divert security analysts' attention from malware that extracted confidential data \cite{idosexam}.
\qed{}
\end{Example}
\end{svgraybox}
\begin{figure}[tbh]
\sidecaption[t]
\includegraphics[width=1 \textwidth]{fig/Cognitiveattack2.pdf}
\caption{
Example paths of cognitive attacks are depicted in the dashed arrow.
\acrshort{se} attacks directly exploit human cognitive vulnerabilities by changing stimuli and influencing external factors.
\acrshort{idos} attacks first compromise the cyber process and use technical compromise as a stepping stone to affect the \acrshort{hmi} (e.g., by creating feints to increase the operator's cognitive load), which indirectly manipulates the cognitive process.
Both types of cognitive attacks have an impact on the cyber and physical processes.
}
\label{fig:Cognitiveattack}
\end{figure}
\subsection{Cognitive Security}
\label{subsec:Cognitive Security}
The development of the \acrshort{ai}-powered \acrshort{cps} begins with reliability assurance to guarantee that the system and \acrshort{ai} technologies can remain trustworthy and efficient under uncertainty, disturbances, and failures.
The presence of attacks requires us to incorporate defenses, which leads to the concept of security.
Analogous to reliability and security in \acrshort{cps}, we define and distinguish cognitive reliability and cognitive security in Definitions \ref{def:Cognitive Reliability} and \ref{def:Cognitive Security}, respectively, to form the four quadrants in Fig. \ref{fig:Quad}.
\begin{figure}[tbh]
\sidecaption[t]
\includegraphics[width=1 \textwidth]{fig/Quad3.pdf}
\caption{
Four quadrants of security and reliability in \acrshortpl{cps} and \acrshortpl{hcps}.
The horizontal arrows distinguish security and reliability based on whether attacks are present or absent, respectively.
The vertical arrows distinguish security (or reliability) in \acrshortpl{hcps} or \acrshortpl{cps} based on whether cognitive vulnerabilities are involved or not.
Cognitive security generalizes \acrshort{se} and its mitigation through the additional investigation of the cross-layer impact on \acrshortpl{hcps} and the development of human-technical defense technologies to deter the kill chain of cognitive attacks.
}
\label{fig:Quad}
\end{figure}
\begin{definition}[Cognitive Reliability]
\label{def:Cognitive Reliability}
Cognitive reliability is the capacity of a \acrshort{hcps} to maintain the continuity of operations, fulfil a designated mission, and provide services at the desired level under stated conditions, including challenging environments of uncertainty, disturbance, and error.
\end{definition}
\begin{definition}[Cognitive Security]
\label{def:Cognitive Security}
Cognitive security is the practice of deploying people, policies, processes, and technologies to withstand cognitive attacks in Definition \ref{def:Cognitive Attacks} and defend essential \acrshort{hcps} components, including humans, critical system structures, services, and sensitive information.
\end{definition}
\subsubsection{Cognitive Reliability vs. Cognitive Security}
\label{subsubsec:Cognitive Reliability vs. Cognitive Security}
From Definition \ref{def:Cognitive Reliability}, cognitive reliability focuses on \acrfull{hra} and human-centered system design.
On the one hand, \acrshort{hra} identifies potential human error events, evaluates the contributing factors, estimates the probability of those errors, and analyzes their impact on the \acrshort{cps}.
A variety of methods exist for \acrshort{hra}, including techniques based on probabilistic risk analysis and cognition control, where \citet{di2013overview} refers to these two classes of techniques as the first and second generations of \acrshort{hra}, respectively.
These \acrshort{hra} techniques have been widely used in life-critical industrial control systems (e.g., nuclear power plants) to minimize the adverse consequences of human behaviors on the technical systems.
On the other hand, human-centered system design includes ergonomics and behavioral science to understand human cognition processes, adapt the system to human factors to enhance usability, and reduce erroneous behaviors.
For example, based on the behavioral science findings that recognition is significantly easier than recall, text-based challenge-response mechanisms have been applied in user-to-computer authentication as an improvement over unaided password recall \cite{zviran1990cognitive}.
Other cognitive science results have been used in user authentication and email encryption for usable security \cite{payne2008brief}; e.g., a graphical password emerges because people memorize pictures better than texts.
Cognitive security in Definition \ref{def:Cognitive Security} is a concept associated with cognitive attacks in Section \ref{subsec:Cognitive Attack}.
Unlike human-induced failures in cognitive reliability, cognitive attacks can directly take advantage of cognitive weaknesses and use them as stepping stones to maximize the attack gain.
The term ``at the desired level under stated conditions'' in Definition \ref{def:Cognitive Reliability} indicates that cognitive reliability has a specific goal of system efficiency and usability for a set of defined conditions.
Performance under these conditions enables cognitive functions to ``withstand attacks" or errors up to a certain capacity.
For attacks that are beyond such capacity, defense methods are needed to provide further protection, as suggested in Definition \ref{def:Cognitive Security}.
Protection methods come at the cost of reduced efficiency (e.g., false alarms of an \acrfull{ids} disrupting normal operation) and usability (e.g., additional effort needed to comply with security procedures).
Security designs that aim to augment the capacity needs to takes into account the tradeoff among security, efficiency, and usability in a holistic manner. Such a rationale leads to the system-scientific perspectives, which will be discussed in Section \ref{sec:System-Scientific Perspectives for Cognitive Security}.
\subsubsection{CIA Triad of Cognitive Security}
\label{subsubsec:CIAtriad}
In the context of cognitive security, we discuss confidentiality, integrity, and availability, i.e., the CIA triad, to provide a guide for theoretical foundations.
Confidentiality prevents unauthorized users from accessing the sequential information flow of the cognitive process shown in Fig. \ref{fig:CogProZoomin}.
Phishing attacks, for example, violate confidentiality by masquerading as legitimate entities and collecting confidential information.
The attacker can use the information to acquire access credentials, steal data, and deploy malware (e.g., ransomware).
The integrity of cognitive security assesses whether the cognitive process depicted in Fig. \ref{fig:CogProZoomin} has been manipulated to induce biased actions.
Examples of cognitive attacks that compromise integrity include misinformation propagation and belief manipulation techniques (e.g., gaslighting).
The availability of cognitive security guarantees the normal functioning of the cognitive process in Fig. \ref{fig:CogProZoomin}.
\acrshort{idos} attacks compromise availability by depleting the limited cognitive resources, including attention, memory, and decision capacity.
\subsubsection{Cognitive and Technical Defenses for Cognitive Security}
\label{subsubsec:Cognitive and Technical Defenses}
Cognitive security synthesizes cognitive and technical defenses to break the kill chain of cognitive attacks, as shown in Fig. \ref{fig:cognitivesecurity}.
Since cognitive attacks include technical exploitation across the cyber and physical layers, as shown in the IDoS attack path in Section \ref{subsubsec:Examples Paths}, cyber and physical defenses are indispensable components.
\begin{figure}[tbh]
\sidecaption[t]
\includegraphics[width=1 \textwidth]{fig/cognitivesecurity4.pdf}
\caption{
The defense methods in blue dashed boxes to break the kill chain of cognitive attacks and achieve cognitive security.
The cyber and physical defenses take effect on the cyber and physical processes, respectively.
User training and human-centered mechanism design affect the cognitive process through externalities. Incorporating the \acrshort{ai} stack, cognitive security includes the \acrshort{ai}-powered enhancer to protect \acrshortpl{hcps} from the adversarial exploitation of \acrshort{ai} technologies.
}
\label{fig:cognitivesecurity}
\end{figure}
Besides the cyber and physical defenses, cognitive security also focuses on defending humans and \acrshort{ai} technologies.
On the one hand, user training can help users gain security knowledge, raise security awareness, and cultivate critical thinking, which consolidates the cognitive process.
On the other hand, we can design mechanisms to avoid selfish behaviors and incentivize users to behave securely.
\acrshort{ai}-technologies in \acrshort{ai} stack can also be susceptible to cognitive attacks.
Keeping the security requirement in mind, we need to design an \acrshort{ai}-powered enhancer to guarantee the security and resilience of the \acrshort{ai}-powered technologies.
In particular, human-centered \acrshort{ai} technologies in \acrshort{ai} stack need to not only tilt the cognitive biases when augmenting human capacities, but also be aware of and defend against the cognitive attacks.
For example, \acrshort{ai}-enabled security assistive technologies can argument \acrshort{hmi} to defend against the threat from \acrshort{idos} attacks.
To achieve the goal, we fuse various biosensor data, including optical, motion, and acoustic signals and electrical biosignals, to reflect the status of the current cognitive process in real-time.
These signals can be fed back to the \acrshort{ai}-enhancer for adaptive control.
We can analogize the human process augmented by \acrshort{ai}-powered technologies in Fig. \ref{fig:cognitivesecurity} as children under the supervision of parents.
While user training and mechanism design teach and motivate children to behave securely, these methods are subject to social and cognitive limitations.
To break the limitation, the \acrshort{ai}-powered enhancer alternatively lets parents be aware of and protect against the potential attacks on their children.
\subsubsection{Scope of Cognitive Security}
\label{subsubsec:Scope of Cognitive Security}
To better illustrate the scope of cognitive security, we distinguish the following related concepts from cognitive security in Definition \ref{def:Cognitive Security} through two running examples.
First, the exploitation and defense of human cognition are essential components of cognitive security, which is significantly different from the realm of CPS security and creates a new battlefield of \acrshort{hcps} security.
Second, as illustrated in Fig. \ref{fig:Quad}, cognitive security generalizes the concept of \acrshort{se} and cognitive hacking \cite{cybenko2002cognitive} that usually rely on weaponized information and exploit vulnerabilities of mental operation.
On top of the cognitive manipulation in the human-human interaction, we further investigate its impacts and dependency on the \acrshort{cps} during the human-machine interaction.
As illustrated in the \acrshort{idos} example, technical exploitation can exacerbate the attentional vulnerability. Such human-technical attacks that directly exploit human cognition limitations have not been considered in \acrshort{se} and cognitive hacking.
Third, cognitive security in this book does not aim to enhance humans at the neuroscience or biological level through the means of medications or medical therapies.
Instead, we focus on designing mechanisms and assistive technologies to non-intrusively influence and externally affect human cognitive processes, change their behaviors, and ultimately enhance the security of the entire \acrshort{hcps}.
Fourth, cognitive security does not aim to answer human-engineering questions such as how to craft convincing misinformation or how to design an anti-phishing interface.
Instead, we leverage theories and empirical results from behavioral science and psychology at the system-level.
Fifth, technical solutions for cognitive security are aware of and adapt to adverse exploitation of cognitive vulnerabilities.
For example, under the threat of \acrshort{idos} attacks, it may be insufficient to adopt the traditional method of alert triage to indirectly reduce cognitive load.
Therefore, we further de-emphasize alerts strategically to directly manage attention and make the alert inspection process compatible with each operator's cognitive load.
Sixth, cognitive defense is an essential but not the only way to achieve cognitive security.
We may not need to defend against all types of cognitive attacks. It is sufficient to break the kill chain of cognitive attacks.
Moreover, since some types of cognitive attacks are not defensible or too costly to defend, we can focus on compensation and correction mechanisms to reduce the risk and impact of the attacks.
Finally, cognitive security in this book does not mean developing security measures to have human cognitive capacities (e.g., cognitive computing \acrshort{soc} \cite{demertzis2018next}, phishing detector \cite{garces2019detection}, and incident management systems \cite{andrade2019cognitive}), nor does it mean the exploitation of human cognitive strengths (e.g., creativity and flexibility) to augment technical defense systems.
Moreover, cognitive security does not focus on increasing the usability of the security measures \cite{greenstadt2008cognitive,payne2008brief}, despite its contribution to cognitive security.
\section{System-Scientific Perspectives for Cognitive Security}
\label{sec:System-Scientific Perspectives for Cognitive Security}
Following Sections \ref{sec: AI-Powered Human-Cyber-Physical Systems} and \ref{sec:Cognitive Security in HCPS}, humans are an essential component in accomplishing \acrshort{cps} missions, and cognitive security needs to incorporate rather than eschew human cognition.
Human cognition, however, is of high complexity, uncertainty, and individual difference, which leads to the following research questions.
How to incorporate established theories of human factors into \acrshortpl{hcps} for defensive purposes?
How to customize the defense and make it adaptive to various security applications?
What are useful metrics and measures to quantify the impact of human factors on a \acrshort{cps} and the effectiveness of the defense methods?
This book adopts system-scientific perspectives, including decision theory, optimization, game theory, \acrshort{ai}, physiological metrics, and psychology theories to address these questions for the following reasons.
\begin{itemize}
\item First, as illustrated in Section \ref{subsec:Cognitive Attack}, cognitive attacks occur at the \textit{cognition level} (e.g., \acrshort{se}) and the system level (e.g., \acrshort{idos} attacks) rather than the \textit{neuronal level} through medicine. Thus, the attack and defense interactions need to be investigated at the \textit{system level} to understand the entire attack phase, anticipate the interaction of attacks with the \acrshort{hcps}, and find defensible points among these phases.
\item Second, cognitive attacks (e.g., \acrshort{idos} attacks) exploit distinct vulnerabilities in human, cyber, and physical processes.
A single tool is not sufficient to deter, detect, and prevent these cross-layer, cross-disciplinary attacks.
To safeguard diverse vulnerabilities, we need interdisciplinary tools to synthesize defense methods from multiple areas.
\item Third, since cognitive attacks have an impact on human, cyber, and physical processes, it is important to quantify the risk to each component, the prorogation of the risk, and the risk to the entire \acrshort{hcps} \cite{huang2017large,huang2018distributed,chen2019game}.
Moreover, we need to design optimal controllables to defend \acrshortpl{hcps} cost-effectively.
Therefore, we need scientific and quantitative approaches to quantify security metrics and bounds, assess tradeoffs, characterize fundamental limits, and design optimal control \cite{wiener2019cybernetics,meadows2008thinking,xu2016cross}.
\item Fourth, since both cognitive attackers and defenders are intelligent players, they take actions based on the predictions of the others. Game theory becomes a natural tool to capture the interactions in adversarial environment for quantitative analysis and design \cite{manshaei2013game,zhu2015game,huang2020dynamic,zhu2018game}.
\item Fifth, models may not accurately describe \acrshort{hcps} components (e.g., human cognition) that are highly dynamic and sometimes unobservable. Data collected by biosensors (e.g., electroencephalogram (EEG) and eye-tracking devices) and \acrshort{ai} can be incorporated to provide system-level adaptive solutions in response to the measurable observations.
\item Finally, theories from psychology and social science can be adopted to understand perception, attention, memory, and decisions. As a multi-disciplinary methods, system-scientific approach incorporate those findings and results to create holistic hybrid data-driven and model-based system frameworks of \acrshortpl{hcps}.
\end{itemize}
\subsection{Advantages of System-Scientific Approaches}
\label{sec:Benefits of System-Scientific Approaches}
System-scientific approaches provide a new paradigm for cognitive security and bring the following advantages.
\begin{itemize}
\item \textbf{Emergence}:
Emergence widely appears in philosophy, psychology, and art, where ``the whole is something else than the sum of its parts'' \cite{koffka2013principles}\footnote{often misquoted as `the whole is greater than the sum of its parts'.}.
As shown in Fig. \ref{fig:TwoTrends}, the close integration of \acrshort{ai} and humans in \acrshortpl{cps} leads to the new concept of \acrshortpl{hcps}. The interactions among human, cyber, physical, and \acrshort{ai} components create new attack surfaces and attack paths, which potentially amplifies the human cognitive vulnerabilities and promotes the need for research in the emerging field of cognitive security.
\item \textbf{Black Box and Function Simulation}: A system perspective focuses on the input and output (or the transfer characteristics) of the system, where the system itself is treated as a black box without any knowledge of its internal workings.
By treating human cognition systems as black boxes, we focus on the \textit{behavior-level impact} rather than the \textit{cognitive- or neuro-level mechanisms} that can be complicated and not well-understood.
The input-output view also enables us to simulate a system's function without establishing the internal model of the system.
\item \textbf{Modular and Multi-Scale Design}:
A system can be divided into different levels of subsystems that compose a multi-scale system model.
For example, the \acrshort{hcps} system consists of human, cyber, and physical subsystems, where the human subsystem further contains cognition systems consisting of perception, attention, memory, and mental operations.
We can zoom into the proper level of subsystems based on our goal.
Since these subsystems interact through their inputs and outputs, each subsystem can be replaced and revised to achieve a modular design.
\item \textbf{Quantitative Strategies and Optimization}:
The system-scientific models enable a quantitative description of the situation and thus the formulation of optimal design problems that lead to cost-effective security mechanisms. The integration of multiple system-scientific approaches, including deterministic and stochastic methods, data-driven and model-based tools, and static and dynamic frameworks, strengthens their strong points and mitigates their weak points.
\end{itemize}
\bibliographystyle{author/spbasic}
|
\section{Introduction}
Sagittarius A$^*$ (Sgr A$^{*}$), the super-massive black hole located at our Galactic center (SMBH; with mass $M \sim 4.1\times10^{6} \ {M_\mathrm{\odot}}$ and distance $D\sim 8.1 \ \mathrm{kpc}$), provides an excellent laboratory to investigate extreme physics and accretion process due to its proximity (e.g., \citealt{Boehle2016, Gillessen2017, GRAVITY2018, GRAVITY2019, Ghez2008}; see reviews by \citealt{Melia&Falcke2001,2010RvMP...82.3121G,YuanNarayan2014}). It is one of the most under-luminous SMBH
with a bolometric luminosity in its quiescent state as low as $L_\mathrm{bol} \sim 5 \times 10^{35} \ \mathrm{erg}\ \mathrm{s}^{-1} \sim 10^{-9} \ L_\mathrm{Edd} $, where ${L_\mathrm{Edd}}$ is the Eddington luminosity. The accretion rate is as low as $(5.2-9.5)\times 10^{-9} \ {M}_{\odot} \mathrm{yr}^{-1}$ \citep{2000ApJ...545..842Q,Marrone2006,EHT2022a}, so it has been an ideal target for studying hot accretion flows around black holes (\citealt{1995Natur.374..623N,1998ApJ...492..554N,Yuan2003,Yuan2004,2020ApJ...896L...6R,EHT2022a,EHT2022e}; see \citealt{YuanNarayan2014} for a review).
While Sgr A* spends most of its time in the quiescent state, it also exhibits multi-wavelength flares. These flares are most significant in the near infrared (NIR) and X-ray wavebands, which generally occur several times a day with a typical timescale of half an hour to an hour \citep{Dodds-Eden2011,Witzel2012,Neilsen2013,Li2015,Ponti2015,Yuan2016}. During the NIR flares, the flux density can be enhanced by a factor of 10 within $\sim 10$ minutes \citep{Genzel2003, Ghez2004, Trippe2007, Witzel2021}. X-ray flares are always associated with NIR flares \citep{2004A&A...427....1E,2006A&A...450..535E}, the luminosity of which is $L_{\mathrm{{3-79 \ kev}}} \sim 2 \times10^{35} \ \mathrm{erg s^{-1}}$, two orders higher than that of the quiescent state \citep{Baganoff2001,2003A&A...407L..17P,2008A&A...488..549P,Dodds-Eden2011,Eckart2013,Neilsen2013,Neilsen2015,Zhang2017, Boyce2019, Do2019,Haggard2019, GRAVITY2021}. Flares have also been observed at lower frequencies such as submm and radio wavebands, with some time delay with respect to the high frequency flares \citep{2003ApJ...586L..29Z,2006A&A...450..535E,Zadeh2006}. The high linear polarization degree ($\sim 20\%-40\%$) of the flares at NIR wavelengths suggests a synchrotron origin, but the emission mechanism for the X-ray flares is not fully conclusive. Many models have been proposed, the possible radiative processes invoked in the models include inverse Compton (IC) scattering of submillimeter photons, synchrotron-self-Compton (SSC) scattering of IR/
NIR photons, or pure synchrotron radiation from non-thermal electron \citep{2001A&A...379L..13M,Yuan2004,Zadeh2006,Broderick2006,Eckart2008,Dodds-Eden2009,Yuan2009,Dodds-Eden2010, Zadeh2012,Ponti2017, Subroweit2020}. It is argued in some works that, although IC and SSC models can self-consistently explain the simultaneous IR/X-ray multiwavelength observations and the time delay between flares at high and low frequencies, synchrotron radiation model has the advantage that it can be achieved with physically plausible
magnetic field strength and electron density compared with IC and SCC models \citep{Dodds-Eden2009,Dodds-Eden2010}.
A breakthrough for understanding Sgr A* flares was made by \citet{GRAVITY2018} (hereafter \citetalias{GRAVITY2018}), in which they reported the first detection of the continuous positional changes of Sgr A* during three prominent bright NIR flares with the near-infrared GRAVITY-Very Large Telescope Interferometer (VLTI) beam-combining instrument. The position centroids exhibit a nearly circular trajectory around the central black hole with a velocity of $\sim 0.3c$ and an orbital period of tens of minutes. During these flares, polarization angle also exhibits continuous rotation, presenting a roughly same period as that of the centroid motion, which is also consistent with the timescale of the flares. The centroid motion and polarization angle rotation can be well fitted by a radiating ``hot spot'' rotating around the black hole with its distance to the black hole being several $r_g$ in the projected plane.
It is not trivial to quantitatively explain the GRAVITY observational results, as emphasized by, e.g., \citet{Matsumoto2020}. The centroids trace 3/4 of a full loop in 30 min and this time corresponds to a Keplerian orbit radius of $\sim 7 r_g$ \citepalias{GRAVITY2018}. But this radius is smaller than the detected distance between centroids and their medians most of the time, which is $\sim$(10-20)$r_g$ (refer to Figure 1 in \citet{GRAVITY2018} and Figure 1 in \citet{Matsumoto2020}). In another word, the hot spot has to rotate with a super-Keplerian velocity to match both orbit radius and period. Therefore a "hot spot" confined on the equatorial plane of an accretion flow is hard to explain the results, because the theory of hot accretion flow which predicts that the flow should have a sub-Keplerian azimuthal velocity \citep{YuanNarayan2014}. Furthermore, the centroid trajectory of the July 22 flare shows a gradually increasing orbital radius (refer to Figure 1 in \citet{GRAVITY2018} and Figure 1 in \citet{Matsumoto2020}), which also indicates the "hot spot" should not stay in the accretion flow otherwise the radius would decrease as it is advected to the black hole with the accretion flow.
Recently many models have been proposed to investigate the nature of flares based on the (GR)MHD simulation of black hole accretion flows, and these models often invoke magnetic reconnection \citep{Guo2014,2016ApJ...818L...9G,Sironi2014, Werner2016, Werner2018} occurred in the accretion flow as the mechanism of electron acceleration \citep[e.g.,][]{2020MNRAS.495.1549N,Nathanail2022,Petersen2020,Zhao2020,2020MNRAS.497.4999D,Chatterjee2021,Porth2021,Scepi2022,2022ApJ...926..136W,2022ApJ...937L..34K,Ripperda2022}. The radiation of these energetic electrons accelerated in the magnetic reconnection events is then responsible for the observed ``hot spot''.
However, with a few exceptions as we will discuss in Section \ref{comparison}, none of these works has presented detailed calculations of radiation and comparisons with observations especially with the GRAVITY observations.
An additional important clue to the nature of flares comes from the radio observations carried out by VLA at 43 and 22 GHz \citep{Zadeh2006}. A radio flare was detected at these two frequencies, with the peak flare emission at 43 GHz leading the 22 GHz flare by $\sim 20-40$ minutes. Such phenomenon is a strong evidence for the ejection and expansion of a radio-emitting plasma blob from the accretion flow during the flare activity. In fact, such kind of physical association between flares and episodic ejection has been observed widely in other black hole sources, such as radio galaxy 3C120 \citep{Marscher2002, Olmstead2008,Chatterjee2009,Casadio2015}, blazar PKS 1510–089 \citep{Park2019}, black hole X-ray binary GRS 1915+105 \citep{Mirabel1994, Fender1999, Jones2005}, low-luminosity AGN M81 \citep{King2016} and M87 \citep{Hada2014}. Although it is possible that not all flares are associated with ejection of blobs, these observations suggest that a physical model of flares should at least be able to provide a framework of simultaneously explaining the associated ejection.
Stimulated by this important observational result, \citet{Yuan2009} (hereafter \citetalias{Yuan2009}) have proposed a model to interpret the flares and the associated ejections of blobs. The model is by analogy with the standard model of coronal mass ejection (CME) on the Sun \citep{Lin2000}, since CME is often associated with solar flares, which is very similar to the black hole sources as we have mentioned above. The schematic figure of the system is shown in Figure 2 of that paper (see also Figure \ref{fig:schematic} in the present work). There exist closed magnetic field lines in the coronal region of the accretion flow, which could be emerged out of the main body of the accretion flow due to Parker instability. The turbulent and differential rotation of the accretion flow twist the field lines and result in reconnection and the formation of flux ropes in the coronal region, similar to the formation of prominence in the solar corona. The reconnection enhances the magnetic pressure and reduces the magnetic tension force. Consequently, the flux ropes are ejected out by the magnetic pressure force.
This scenario has been confirmed recently by the detailed analysis to the three dimensional GRMHD simulation data of black hole accretion flow \citep{Miki2022}. The accelerated electrons in the current sheet will flow into the flux rope and the flare loop regions through the reconnection outflow, as shown by Figure \ref{fig:schematic}. The radiation of these energetic electrons explains the observed flares, while the ejected flux rope is responsible for the observed ejected blobs. \citet{Li2017} (hereafter \citetalias{Li2017}) have applied the \citetalias{Yuan2009} model to Sgr A* and successfully interpreted the NIR and X-ray light curves of Sgr A* flares by calculating the dynamics of the ejected blobs and the evolution of the energy distribution of energetic electrons accelerated by the magnetic reconnection process.
In the present work, we investigate whether we can explain the new observational results from \citetalias{GRAVITY2018} based on the framework of \citetalias{Yuan2009} and \citetalias{Li2017}.
The structure of the paper is as follow. In Section \ref{sec:2} we calculate the dynamic properties of hot spots based on \citetalias{Yuan2009} and \citetalias{Li2017}. We then evolve the non-thermal electrons for the radiation calculation injected from the reconnection region in Section \ref{sec:3}. The ray-traced calculation is described in Section \ref{sec:4}. We show the results and compare with observations in Section \ref{sec:5}. We discuss and summarize our results in Section \ref{sec:discussion} and Section \ref{sec:6}, respectively.
\begin{figure*}
\includegraphics[width=1.\textwidth]{modelnew.png}
\caption{A schematic illustration of our model (left panel) and a zoomed-in plot of the flux rope in its 2D cross section (right panel). In the left panel, the main radiation region is from the center of the flux rope where it is represented by a yellow hot spot. The helical trajectory of the hot spot is depicted by the red curve and the hot spots in different times are plotted. Note that the hot spot is expanding as it is ejected outward. In the right panel, the flux rope is ejected outward in a fixed angel $\theta_B$ inclined with respect to the normal direction of the equatorial plane (i.e., $z$-direction). The loop and the flux rope are connected by the current sheet (marked by the red line).}
\label{fig:schematic}
\end{figure*}
\section{Dynamics of the flux rope}
\label{sec:2}
\subsection{The basic scenario of the Model}
The original \citetalias{Yuan2009} model is two-dimensional. To calculate the radiation, we need to convert this two-dimensional model to three-dimension, which is shown schematically by the left plot of Figure \ref{fig:schematic}.
We start our calculations from a flux rope newly formed due to magnetic reconnection. The flare loop stays on the equatorial plane, keeping a constant distance $R_{\mathrm{lp}}$ away from the black hole and following a sub-Keplerian rotation velocity of the underlying RIAFs \citep{YuanNarayan2014}:
\begin{equation}
\Omega_{\mathrm{lp}}= k\Omega_{\mathrm{0}},
\end{equation}
where $0 \textless k \textless 1$ is the proportionality factor, and $\Omega_{\mathrm{0}}$ is the Keplerian angular velocity at $R_{\mathrm{lp}}$ \citep{Bardeen1972}:
\begin{equation}
\Omega_{\mathrm{0}} = \frac{M^{1 / 2}}{R_{\mathrm{lp}}^{3 / 2} + a M^{1 / 2}},
\end{equation}
where $M$ and $a$ are the mass and the dimensionless angular momentum of the black hole.
The ejected flux rope follows a spiral motion since it has angular momentum, its trajectory is shown in the left panel of Figure~\ref{fig:schematic} by the yellow sphere and red line. The motion of the flux rope can be decomposed into two components, i.e. the azimuthal (or toroidal) and poloidal motions. The inclusion of the toroidal motion of the ejected flux rope is one of the main improvements we make in this work for the dynamics of the flux rope compared to \citetalias{Yuan2009} and \citetalias{Li2017}, where the toroidal motion is not considered for simplicity. The poloidal motion of the flux will be calculated in Section \ref{dynamics}. Note that the poloidal motion of the ejected flux rope has an inclination angle $\theta_B$ with respect to the normal direction of the accretion flow, as shown in the figure. This is because the poloidal motion of the ejected flux rope has to follow the direction of the large-scale poloidal magnetic field, which is inclined with respect to the accretion flow in this way, as indicated by our GRMHD numerical simulations \citep{Miki2022}. As we will see later, this will explain the increasing distance of the hot spots observed by GRAVITY \citepalias{GRAVITY2018}.
In our model, the rotation of the ejected flux rope is super-Keplerian. The reason is as follows. From the GRMHD simulations by \citet{Miki2022}, we find that the flux rope has the same angular velocity as its footpoint, i.e., the flare loop, which is $\Omega_{\mathrm{rp}}=\Omega_{\mathrm{lp}}$, within a distance of $\sim 60 r_g$ from the black hole. This distance marks the Alfvén surface where Alfvén speed $V_A (\equiv B/\sqrt{4\pi \rho})$ is equal to the wind speed.
Physically, this is because the strength of the field is so strong that it can resist bending caused by differential rotation when the field line is close to the central black hole; the field line thus acts as a rigid rod that rotates around the black hole with the same angular velocity. Plasmoids enveloped in the flux rope co-rotate with the field line due to the flux frozen-in.
Due to the constant angular velocity of the ejected flux rope but increasing distance to the rotation axis, the plamoids will thus possess a larger angular velocity compared to the the local Keplerian velocity as it is ejected outward, i.e., its motion is super-Keplerian. We note that this is in good consistency with the GRAVITY observations \citepalias{GRAVITY2018} and emphasized by \citet{Matsumoto2020}.
Beyond this Alfvén surface, the field strength decreases to a certain extent that magnetic tension cannot effectively suppress differential rotation. Therefore the magnetic field lines start to be more twisted and eventually become toroidal-component dominated.
In this region (e.g., $r\gtrsim60r_g$), the angular velocity of the flux rope is calculated instead according to the conservation of angular momentum in Newtonian frame, i.e.,
\begin{equation}
\Omega_{\mathrm{rp}}=\Omega_{\mathrm{lp}}r_0^2/r(t)^2,
\end{equation}
where $r_0=60r_g$.
\subsection{The poloidal motion of the ejected flux rope}
\label{dynamics}
The detailed calculation of the poloidal motion of the flux rope can be referred to \citetalias{Yuan2009} and \citetalias{Li2017}. Here we only present a brief description.
The dynamics of the ejected flux rope is controlled by three forces, namely the magnetic pressure, magnetic tension, and gravity. Due to the reconnection, the magnetic pressure force will become significantly larger than the sum of magnetic tension and the gravity.
As a result, the flux rope will be ejected out by the magnetic pressure force. Magnetic dissipation in the current sheet will convert the magnetic energy to heat and accelerate electrons in the current sheet, and these electrons will be transported to the flux rope and the flare loop region located above and below the current sheet, respectively, through the reconnection outflow.
The two foot points of the flux rope embedded in the magnetic arches are anchored at the surface of the accretion flow. We approximately treat $\pi L_0$ as the total length of the magnetic arch, which should increase with the eruption of the flux rope. In this work, we take it as a constant and $L_0 = 50\ r_{\rm g}$ as in \citetalias{Li2017} for simplicity. The effective length of the magnetic arch should be determined by the length of the arch when the emission from the flux rope is strongest. During the flares, the flux rope can be ejected outward ($h\gtrsim50\ r_{\rm g}$) as we show below (i.e., Figure~\ref{fig:1}), which would give a large effective $L_{0}$ of $\sim50\ r_{\rm g}$. Note that there is some degeneracy between the value of $L_0$ and other model parameters so we don’t treat it as a free parameter. The central part of the flux rope has a tilted height $h$ (see the right panel of Figure \ref{fig:schematic}).
The explicit illustration of the system is shown in the right panel of Figure \ref{fig:schematic}. The dynamical evolution of the flux rope is depicted by five quantities, namely the ``height'' of bottom and top tips $p$ and $q$ of the current sheet, the ``height''\footnote{Here we use quotes mark because $\theta_B\neq 0$.} of the flux rope center $h$, the poloidal velocity of the flux rope $\dot{h}$ and the total mass enveloped in the flux rope $m$. These five quantities can be deduced by solving the equation of motion
\begin{equation}
m \gamma_{\mathrm{b}}^{3} \frac{d^{2} h}{d t^{2}}=\frac{1}{c}\left|\boldsymbol{I} \times \boldsymbol{B}_{\mathrm{ext}}\right|-F_{\mathrm{g}},
\end{equation}
where $\gamma_{\mathrm{b}}=1 / \sqrt{1-\dot{h}^{2} / c^{2}}$ is the Lorentz
factor of the flux rope, $\boldsymbol{I}$ is the integrated current intensity inside
the flux rope, $\boldsymbol{B}_{\mathrm{ext}}$ is the total external magnetic field (apart from $\boldsymbol{B}_{\mathrm{I}}$ generated by $\boldsymbol{I}$ itself) measured at the center of the flux rope, and $F_{\rm g}$ is the gravity from the central SMBH. The first term on the right-hand side depicts magnetic force and the second term depicts the gravitational force.
Following \citetalias{Yuan2009} and \citetalias{Li2017}, in order to obtain an analytical solution of the motion of the flux rope, and by analogy with the magnetic field in the solar surface, the spatial distribution of the magnetic field is assumed to be described by
\begin{equation}
B(\zeta)=\frac{2 i A_{0} \lambda\left(h^{2}+\lambda^{2}\right) \sqrt{\left(\zeta^{2}+p^{2}\right)\left(\zeta^{2}+q^{2}\right)}}{\pi\left(\zeta^{2}-\lambda^{2}\right)\left(\zeta^{2}+h^{2}\right) \sqrt{\left(\lambda^{2}+p^{2}\right)\left(\lambda^{2}+q^{2}\right)}},
\label{eq:2}
\end{equation}
where $\zeta=r+i z, A_{0}=B_{0} \pi \lambda_{0}$ is the source field strength and $\lambda=5r_g$ is the
half-distance between the two field line foot points of the flare loop. Note that the magnetic field is time-dependent due to the evolution of $p,q$ and $h$. We emphasize that this distribution of magnetic field is a weakness of our model. We should in principle adopt the realistic magnetic field, which is, however, very uncertain. We expect that the main effect will be on the prediction of the polarization, as we will discuss in Section~\ref{polarization}.
There are seven free parameters in this model: (i) The distance of the flare loop center from the central black hole $R_{\mathrm{lp}}$. We choose it to be $6.5 r_g$, where $r_{g}$ is the gravitational radius of the SMBH\footnote{\citet{Miki2022} find that only flux ropes formed beyond $10-15 r_g$ can be ejected out while those formed inside this radius usually stay within the accretion flow. In our fiducial model, the flux rope is formed at $h\sim 10 r_g$ although its ``footpoints'' are at $6.5 r_g$ (refer to Figure \ref{fig:1}). This corresponds to its spherical radius of $\sim 12 r_g$, thus the flux rope can be ejected out. }. This is in the favored range given by \citet{GRAVITY2020a}. We note that our model is not sensitive to the exact value of $R_{\mathrm{lp}}$. (ii) The proportionality factor $k$, which is set as $k=0.93$. That means the rotation velocity is sub-Keplerian as predicted by the theory of RIAFs \citep{YuanNarayan2014}. (iii) The tilted angle between the direction of the poloidal motion of the flux rope and spin axis. We set this angle $\theta_B = 173^{\circ}$ (or $\theta_B = 7^{\circ}$).
This small tilted angle indicates the ejected blobs are highly collimated, similar with jets. (iv) The background magnetic field $B_0$ in the vicinity of the reconnection region. We choose $B_0=120 \ \mathrm{G}$, which is a factor of a few larger compared to the average magnitude of 10-50 Gs (e.g.,\citealt{Yuan2003,Dodds-Eden2010,Boyce2019,EHT2022e}). This is reasonable given that the magnetic field has strong spatial fluctuation and we focus on strong flares in highly magnetized regions where the magnetic field should be greater than the average value. (v) The electron number density in the inner accretion flow region $n_{\mathrm{e}} = 1.75\times 10^7\ \mathrm{cm^{-3}}$ and (vi) the density ratio $\xi=10$ for the flux rope with respect to the background density $n_{\mathrm{e}}$. (vii) The Alfv\'{e}n Mach number $M_{\mathrm{A}}$ defined as the reconnection inflow speed divided by the local Alfv\'{e}n speed near the reconnection regions, determining the efficiency of the reconnection. In this work we set $M_{\mathrm{A}}=0.5$, which is slightly higher than the predicated values from the magnetic reconnection theory \citep{Liu2017,Guo2020}. Note that the adopted parameters above should not be regarded as a unique solution for modeling the NIR flares as we did not fully explore the parameter dependence in the current work due to the complexity of the radiative transfer calculations as described below. We note that the values of these main parameters adopted in the present paper are basically the same as those in
\citetalias{Li2017}. In this way, we hope to investigate whether our original
\citetalias{Li2017} model, with extension of including more consistent general relativity radiative transfer calculations, can also explain the new \citetalias{GRAVITY2018} results. These parameters together with which will be introduced in non-thermal electron evolution and ray-tracing calculation are summarized in Table \ref{tab:1}.
\begin{figure}
\includegraphics[width=\columnwidth]{pqh.png}
\caption{Dynamical evolution of the flux rope. The flux rope can be ejected out to a height of $h \sim 110 r_g$ with velocity of 0.5 speed of light within 150 minutes after the reconnection occurs.}
\label{fig:1}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{bmean.png}
\caption{Evolution of the magnetic field in the flare loop and ejected flux rope. }
\label{fig:2}
\end{figure}
As the system evolves, energy stored in the magnetic field releases and transforms to other forms of energy, e.g., the kinetic energy of the bulk motion of the flux rope, and energy of the heated and accelerated electrons. Figure \ref{fig:1} shows the evolution of $p,q,h$ and $\dot{h}$. During the whole evolution stage, the size of flare loop, which is determined by the lower tip of current sheet $p$, increases slowly ($\lesssim 10\ r_{\rm g}$). This indicates that the expansion of the flare loop region is insignificant and it is justified that the loop always stays on the equatorial plane. By contrast, the evolution of $h$, which measures the height of the flux rope, is dramatic, indicating the eruptive process of episodic ejection. In our fiducial model, the flux rope is formed at $h \sim 10r_{\rm g}$ or $r \sim 12r_{\rm g}$ in spherical coordinates. We found that if the initial electron density $n_\mathrm{e} \cdot \xi$ in the flux rope is too large, or the magnetic field magnitude in the reconnection region $B_0$ is too weak or the newly formed flux rope is too close to the black hole, the catastrophic eruption would fail and the flux rope would no longer be ejected out. This phenomenon is ubiquitous in solar observations. It is interesting to mention that similar scenario is also found in GRMHD simulation \citep{Miki2022}. They found that at small radii, inside of $\sim 10-15 r_{\rm g}$, few flux ropes can be ejected out, likely due to the strong gravitational force of the black hole which is similar to the absence of wind at small radii. As shown by Figure \ref{fig:1}, the flux rope is ejected out over 100 $r_g$ and reaches a velocity of $\sim 0.5 c$ within 150 min after the eruption occurs. This value is similar to the ejection speed obtained in GRMHD simulations \citep{Miki2022}. The mean size of the flux rope can be roughly expressed by ($h-q$), which is $\sim 100 r_g$ after 150 min. But we will see that such a large flux rope in the later stage leave no observational signatures due to its strong cooling. The length of the current sheet ($\sim q-p$) is of the order of several $r_{\rm g}$, again consistent with that found in the GRMHD simulations \citep{Miki2022}.
The evolution of the spatially-averaged magnetic field within the flux rope and flare loop during the reconnection process is shown in Figure \ref{fig:2}. It can be seen that the magnetic field in the flux rope drops rapidly, which means the reconnection proceeds efficiently owing to the high $M_{\rm A}$.
\begin{figure}
\includegraphics[width=\columnwidth]{Edot.png}
\caption{Total energy release rate during the magnetic reconnection in the model. The observed timescale of a typical flare event is about one hour. }
\label{fig:3}
\end{figure}
The electromagnetic energy release rate during the reconnection is given by
\begin{equation}
\dot{E}(t)=\pi L_{0} S(t),
\end{equation}
where $S(t)\propto E(t) \times B(t) $ is the Poynting flux in the current sheet. The reconnection event shows a peak power $\dot{E}_{\mathrm{{max}}} \sim 2.5 \times 10^{37}$ erg $\mathrm{s^{-1}}$ with a typical timescale of about one hour, as shown by Figure \ref{fig:3}, which is controlled by Alfv\'{e}n and reconnection time scales. This is comparable to the typical timescale of flares in Sgr A$^{*}$. As a consequence, the total energy released in this process is $ \textgreater 10^{40}$ erg,
much larger than the estimation of $\ \gtrsim 10^{38}$ erg limited by two observed simultaneous near-infrared and X-ray flares \citep{Ponti2017}. This shows that there is enough energy available to reconnection to account for the observed energy released during the flares.
\section{Evolution of non-thermal electrons accelerated by reconnection}
\label{sec:3}
Magnetic reconnection is well recognized as a high efficient way to energize electrons from thermal pool to high energy non-thermal electrons. The acceleration process is found to be related to the ambient plasma-$\beta$ and magnetization $\sigma$ \citep[e.g.,][]{Ball2018}. In the present paper, we assume $10\%$ of the released energy is delivered to non-thermal electrons, which is quite plausible in the low-$\beta$ magnetically-dominated region of hot accretion flow surface based on the previous theoretical and simulation works \citep{Ball2018,Chatterjee2021}.
Particle-in-cell (PIC) simulations of magnetic reconnection indicate that the accelerated electrons follow a power-law distribution \citep{Guo2014,Sironi2014,Werner2017,Werner2018}, which can be written as
\begin{equation}
\dot{N}_\mathrm{pl}(\gamma)=\dot{n}_{\mathrm{pl}} \gamma^{-p_{\mathrm{e}}}, \quad \gamma_{\min } \leq \gamma \leq \gamma_{\max },
\end{equation}
where $\dot{n}_{\mathrm{pl}}$ is the non-thermal electron injection rate, $p_{\mathrm{e}}$ is the power law index, $\gamma_{\mathrm{min}}$ and $\gamma_{\mathrm{max}}$ are the low- and high-energy cutoffs of the distribution. The explicit value of $p$ is still not well determined, which could range from 1.0-2.5 depending on many factors like plasma-$\beta$ and magnetization $\sigma$ \citep[see, e.g., ][]{Guo2020,Li2021}. Observations show that the infrared spectral index of Sgr A* is $\alpha=-0.6 \pm 0.2\left(F_{\nu} \propto \nu^{\alpha}\right)$ \citep{Hornstein2007,Witzel2018}, which is expected for pure synchrotron radiation emitted from a power law energy distribution electrons with $p_{\mathrm{e}}=1-2 \alpha \sim 2.2$. The X-ray spectral index, on the contrary, has a large uncertainty. The value ranges from $-1.5 \pm 0.3$ to $0.5_{-1.3}^{+0.9}$ (\citealt{Porquet2003,Baganoff2001,Baganoff2003}). Then the inferred power-law index could be as steep as $p_{\rm e}\sim4.0$ if the X-ray comes from a population of single power-law electrons. However, the electron cooling timescale due to synchrotron losses is $ t_{\text {cool }}=1.29 \times 10^{12} \nu^{-1 / 2} B^{-3 / 2} \mathrm{~s}$, which is less than 10 mins for $B\sim 50G$ and $\gamma \sim 10^3$. The synchrotron cooling thus should be important in shaping the single power-law function into the broken power law. This is also suggested by the steep spectral
index between IR and X-ray wavelengths of $\alpha_{\rm K-X}\sim-1.1$ to $<-1.5$.
Based on these, we set the power law index $p_\mathrm{e}=2.5$ for the initial injected electrons, which is also in agreement with the analytical studies of the first order Fermi process in current sheets which usually predicts $p_{\mathrm{e}} \sim 2.5$ \citep{Pino2005}. We choose the maximum Lorentz factor of electrons $\gamma_{\mathrm{max}}=10^6$. Note that, as long as $\gamma_{\mathrm{max}}$ is high enough (e.g., $\gtrsim 10^4$), the value of $\gamma_{\mathrm{max}}$ is not so important when we focus on the NIR flares with $p_{\mathrm{e}}>2$ as considered here. The values of $\gamma_{\mathrm{min}}$ and $\dot{n}_{\mathrm{pl}}$ can be determined jointly by assuming the thermal and non-thermal electrons are smoothly connected at $\gamma_{\mathrm{min}}$ with an injected fraction $10\%$ of thermal energy from reconnection event, as shown in Figure~\ref{fig:4} (see \citetalias{Li2017} for details).
\begin{figure}
\includegraphics[width=\columnwidth]{breakpl.png}
\caption{Electron energy distribution for the flare loop at $t = 120$ min after the reconnection starts. The electron temperature $T_{\mathrm{e}}=3.4\times 10^9 \ \mathrm{K}$, and mean magnetic field $B=45.2\ \mathrm{G}$. The distribution can be divided into three parts, i.e., the thermal part denoted by the dashed blue line, and two power-law parts with power-law indexes being $p=2.5$ and $p=3.5$, transited at $\gamma \sim 10^3$. }
\label{fig:4}
\end{figure}
These accelerated non-thermal electrons flow into the flare loop and the flux rope. However, how the electrons distribute in each region are unclear and non-trivial. In this work, we simply assume that half of the electrons flow into each region. Given that cooling will be important if high-energy electrons contribute significantly to the NIR flares, we take into account cooling due to the expansion of flaring region and due to synchrotron radiation \citep{Dodds-Eden2010, Li2017, Petersen2020,Chatterjee2021,Scepi2022}. We evolve the electron energy spectrum by considering injection as well as cooling according to the following continuity equation \citep{Blumenthal1970}:
\begin{equation}
\frac{\partial N_{\mathrm{e}}(\gamma, t)}{\partial t}=Q_{\mathrm{inj}}(\gamma, t)-\frac{\partial[\dot{\gamma} N_{\mathrm{e}}(\gamma,t)]}{\partial \gamma}.
\end{equation}
The escaping of non-thermal electrons from flaring regions are neglected. The detailed discussion of this equation can be referred to \citetalias{Li2017} and \citet{Dodds-Eden2009}. Figure \ref{fig:4} shows the energy spectrum of the flare loop at $t = 120$ min. There is clearly a break in the power law distribution around $\gamma \sim 10^3$; the power-law index beyond this $\gamma$ transits from $p=2.5$ to $p^\prime =3.5$. Because the radiation at NIR is likely from the electrons around this break, it is noted that a careful treatment of the electron distribution, i.e. a broken power law instead of a single power law distribution, should be taken.
\section{Calculation of radiative transfer}
\label{sec:4}
To calculate the radiation of the system, we employ the publicly available numerical code GRTRANS \citep{Dexter2009,Dexter2016}, which provides a self-consistent fully relativistic ray tracing radiative transfer calculation. As in most works, a ``fast light'' treatment is adopted to neglect the light travel time. Therefore, the direct emission and lensed emission will reach the camera at the same time. We note that this simplification may impose some modifications on the image and flux at each snapshot, and thus on the light curves, centroid motion and polarization as well \citep{Younsi2015, Ball2021}.
In Section \ref{sec:3}, we have deduced the total non-thermal electron number $N_{\mathrm{e}} (\gamma,t)$ at each time in the flare loop and flux rope respectively. We further assume that these nonthermal electrons within the flux rope and flare loop regions satisfy a Gaussian distribution \citep{Broderick2006}
\begin{equation}
n_{\mathrm{S}(r,t)}=n_{\mathrm{S}}^{0} \exp \left[-\frac{\lvert \Delta r\rvert^2}{2 R_{\mathrm{0}}^{2}}\right],
\label{eq:density}
\end{equation}
where $\Delta r \equiv r-r_{\rm S}$ is the displacement from the region (hereafter ``spot'') center, and $R_{\mathrm{0}}$ controls the size of the spot. As the acceleration and spatial distribution of non-thermal electrons may depend on the microphysics of reconnection \citep[e.g.,][]{Guo2020}, i.e., the magnetization parameters, which is also hard to determine from the first principle for our semi-analytical modeling, we thus for simplicity fix $R_{\mathrm{0}}=2 \ r_g$ in both regions to be consistent with $R_{\mathrm{0}} \lesssim2.5 \ r_g$ favored by \citet{GRAVITY2020a}. The electrons are confined in $R_s = p \sim$ several $r_g$ for the flare loop and $h-q \sim$ tens of $r_g$ for the flux rope. We have to notice that the initial shape of the flux rope should not be spherical but arched (see left plot of Figure \ref{fig:schematic}). However, as the eruption proceeds, the complete structure of the flux rope cannot be well maintained. The flux rope will soon be entangled because its foot points are anchored in the different radii of the accretion flow where the angular velocities are different. This entanglement would exacerbate kink stability which could further disintegrate the end parts of the flux rope and finally only the main central part would be preserved. Therefore it is reasonable to adopt a spherical hot spot model for simplicity. The number density at the spot center $n_{\mathrm{S}}^{0}$ can be determined by integrating the distribution function, together with Eq. \ref{eq:density}
\begin{equation}
\int^{R_S}_{0}4\pi r^2 n_{\mathrm{S}}dr = N_e.
\end{equation}
The configuration and magnitude of the magnetic field is described by Eq. \ref{eq:2}. It only has a poloidal component $\left(B_{r} \text { and } B_{\theta}\right)$. As suggested by \citetalias{GRAVITY2018}, a poloidal field is preferred to explain the rotation of polarization angle. In realistic situations, field configuration can be much more complicated. On the one hand, accretion and rotation can twist the field lines. On the other hand, the turbulence driven by magneto-rotational instability (MRI) can cause the field disordered. This effect is out the scope of our current work and we leave it for our future work based on 3-D GRMHD simulation of \citet{Miki2022} where magnetic field is self-consistently evolved. But as discussed before, a pure poloidal magnetic field is still an acceptable simplification
given that the magnetic field is dominated by the poloidal component inside the Alfvén surface.
The radiative transfer is calculated in the Kerr metric and the physical quantities used to fulfil the calculation have been transformed into Boyer-Lindquist coordinates. More details about how to transform the velocity and magnetic field into four-vector forms can be referred from Appendix \ref{Appendix}. We choose the same coordinate systems to describe the motion of the hot spots as was adopted by \citet{Matsumoto2020} (refer to Figure 2 in their paper).
Dimensionless spin of the black hole $a=0.9$ and our results are not sensitive to its exact value. We choose this value because it agrees with the parameter survey conducted by \citet{GRAVITY2020a} and \citet{EHT2022e}. The fiducial inclination angle, defined as the angle between the line of sight and the normal direction of the accretion flow, is set to be $i=173^{\circ}$ to explain the apparent circular centroid motion and to ensure the clockwise motion of the spot. The almost face-on configuration is also favored by both GRAVITY \citep{GRAVITY2020a} and EHT \citep{EHT2022a,EHT2022e} results, which give $180^{\circ}\ga i \ga 140^{\circ}$. The longitude of the ascending node $\Omega$ is chosen to match both centroid trajectory and light curves in the flare epoch. Two types of ejected flux rope are considered, with one being ejected towards the observer (forward model, $\theta_B=173^\circ$) while another being ejected away from the observer (backward model, $\theta_B=7^\circ$). In both cases, the brightness centroid exhibits apparent clockwise motion in the projected sky since their rotation direction follows the accretion flow. The field-of-view of the camera is [200,200] $\mu as$ in both x and y directions with a numerical resolution of [1,1] $\mu as$ in each direction. For other parameters, the mass of the black hole $M = 4.1\times10^{6} \ {M_\mathrm{\odot}}$ and the distance from us is $8.1 \ \mathrm{kpc}$.
With these setups, the polarized synchrotron radiation and images at NIR K-band at different time epochs from accelerated non-thermal electrons obtained from Section\ \ref{sec:3} are calculated using the GRTRANS code. The results are then compared with observations from \citetalias{GRAVITY2018}. Because of the large parameter space and non-triviality of radiative transfer calculations, we do not resort to any rigorous statistical method like Markov Chain Monte Carlo to fit the observational data. Instead, we manually adjust some additional geometrical model parameters combined with the model parameters adopted in the dynamical evolution stage until our model can match salient features of \citetalias{GRAVITY2018} observations. The effect of some model parameters will be explored to diagnose their respective roles on different aspects of the results and some of them have been discussed in \citetalias{Li2017}.
\begin{table*}
\caption{Main parameters of fiducial model. 'f' and 'b' represent forward and backward moving flux rope respectively. And the ray-tracing parameters for the flare loop are as same as the forward flux rope. }
\label{tab:1}
\begin{tabular}{cccccccc}
\hline \hline Dynamics & $R_{\mathrm{lp}}(r_g)$ & $\theta_B$ (f/b) & $k$ & $B_0 (G)$ & $n_{\mathrm{e}} (\mathrm{cm^{-3}})$ & $\xi$ & $M_A$ \\
\hline \ & 6.5 &$173^\circ$/ $7^\circ$ & 0.93 & 120G & $ 1.75\times 10^7 $ & $10 $ & $0.5$ \\
\hline \hline Electron evolution & $p_{\mathrm{e}}$ & $\gamma_{\mathrm{max}}$ & $L_0 (r_g)$\ & $R_0 (r_g)$ & & & \\
\hline \ & 2.5 & $10^6$ & 50 & 2 & & &\\
\hline \hline Ray-tracing & $a$ & $i$ & $\phi_0$ (f/b) & $\Omega$ (f/b) &$M ( {M_\mathrm{\odot}})$ & $D$ (Kpc) & \\
\hline \ & 0.9 & $173^{\circ}$ & $256.4^{\circ}$/$285^{\circ}$ & $270^{\circ}$/$110^{\circ}$ & $4.1\times10^{6}$ & $8.1$ & \\
\hline \hline
\end{tabular}
\end{table*}
\section{Results}
\label{sec:5}
\subsection{Light curves}
Figure \ref{fig:lightcurve} shows the light curves produced by different regions and their comparison with observations. The flare loop, and the forward and backward moving flux rope of our fiducial model are represented by different lines, and the observational results of July 22 flare from \citetalias{GRAVITY2018} are shown as blue dots. The emission from the forward flux rope can well explain the main features of the observed light curve, including the timescale and luminosity. We can see from the figure that the emission of the flux rope region dominates over the flare loop by more than a factor of ten and shows much stronger variability than the flare loop.
There are several reasons for this enhancement. Firstly, the energetic electrons in the flux rope get more concentrated in the high magnetic field region. Secondly, both the Doopler beaming effect due to the high ejection/rotation speed (to the forward flux rope) and the general relativistic lensing effect (to the backward flux rope) can amplify the emission of the flux rope. As we will discuss later in Section~\ref{sec:discussion}, the former effect tends to be stronger.
It should be noted that this is based on the assumption that we re-distribute energy and number of the non-thermal electrons to two flare regions equivalently. Results could be different if we inject a larger fraction of non-thermal electrons into the flare loop. By contrast, \citetalias{Li2017} finds the emission from the loop and flux ropes are comparable, which is because the effects mentioned above are not considered there.
\begin{figure}
\includegraphics[width=\columnwidth]{flux1.png}
\caption{Light curves produced by the forward
and backward moving flux ropes and the flare loop.
The blue dots represent the observation data of the July 22nd flare from \citetalias{GRAVITY2018}. The flux is in units of the flux of S2 ($K_S = 14$, or 15 mJy). The starting time of our model is shifted to match the observational data. Similar treatment is made to the following figures.}
\label{fig:lightcurve}
\end{figure}
In our following analyses, we attribute the observed hot spot observed by \citetalias{GRAVITY2018} to the forward ejected flux rope rather than the flare loop and backward flux rope since the radiation of the forward flux rope can fit the light curve better. The flux rope moving toward us (forward flux rope) shows about 4-5 times stronger emission than the rope away from us (backward flux rope). This is mainly due to the beaming effect which can be in analogy with jets and counter jets. Although the emission from the current backward flux rope does not fit the observational data, such a result may be parameter dependent. we suspect that under some model parameters, the general relativistic lensing effect may be able to more strongly amplify the radiation of the backward flux rope and result in a stronger flare than the forward flux rope.
In addition, we will see in the next section that the backward flux rope can also reproduce the hot spot trajectory. Hence we cannot fully rule out this possibility.
\subsection{Centroid motion of the hot spot}
One of the most important results from GRAVITY observations is the centroid motion of the hot spot during the Sgr A$^{*}$\ flares. The July 22 flare exhibits an apparent clockwise circular motion. The centroid position for the other two flares analysed in \citetalias{GRAVITY2018} does not form an obvious loop pattern due to the large uncertainties but still presents counter-clockwise rotating trend. Hence we focus on the orbit motion of the July 22 flare. The continuous positional changes exhibit $\sim 4/3$ of a full loop over 30 mins, corresponding to an orbit period of $\sim 40$ mins. The medians of the centroid are coincident with the location of the black hole determined from the mass center of the S2 orbit.
Figure~\ref{fig:12} shows the ray-tracing images of the forward and backward moving flux rope. One compact hot spot appears in the projected sky for the forward flux rope while two bright spots appear on each side of the black hole for the backward model. This is simply due to the strong gravitational lensing effect.
\begin{figure}
\includegraphics[width=\columnwidth]{motion.png}
\caption{Ray-tracing images of the forward (left panel) and backward (right panel) moving blobs. The time interval between two adjacent panels is 10 mins. For the backward blob, two bright spots appear on the camera screen due to strong gravitational lensing, which has noticeable influence on the brightness centroid position. When the blob comes closer to the back side of the black hole, the lensing effect becomes stronger and the bright spot begins to be present as bright arcs as shown in the bottom left panel. Note that we do not consider any observational beam effect.}
\label{fig:12}
\end{figure}
Figure \ref{fig:11} shows the centroid motion of the two flux ropes and the flare loop obtained by our ray-tracing calculations\footnote{We reasonably assume that usually only one flux rope exist during a flare, but the flare loop of course always exists. In this case, the observed hot spot should be the radiation flux-weighted combination of the flux rope and the flare loop, which is dominated by the flux rope.}, together with their comparisons with observations. The exact locations of each component are determined by the flux-weighted calculation. The helical motion of the ejected flux ropes in our model is represented as orbital motion in the projected plane. On the projected plane, the trajectories of the two flux ropes are helices (magenta and cyan lines in Figure \ref{fig:11}), while that of the flare loop is nearly a circle (purple line in Figure \ref{fig:11}). The 30-min trajectory of the forward flux rope during the flare time is highlighted by the solid magenta line.
Both the forward and backward flux ropes can satisfactorily explain the centroid motion of the NIR flares \citepalias{GRAVITY2018}. A remarkable feature is the increasing projected distance over time that fits with observations. More importantly, the ejected flux rope scenario with spiral motion makes the blob maintain its high angular velocity attained at small radius as it voyages to larger radius, which can then explain the observed super-Keplerian motion of the hot spot.
\begin{figure}
\includegraphics[width=\columnwidth]{centroidnew.png}
\caption{Centroid motion of the hot spot and its comparison with the July 22nd flare result. The 80 mins trajectory (in accordance with Figure\ \ref{fig:lightcurve}) of the flux rope is plotted and 30 min trajectory of the forward flux rope during the flare is thickened by the solid magenta line. The arrow in the bottom right indicates that these spots exhibit clockwise looped motion on the camera plane.}
\label{fig:11}
\end{figure}
From the Figure \ref{fig:11} we can rule out the flare loop model because its orbital radius is too small to be compatible with the observation. The orbit lies almost entirely inside the observation data points and the distance (Figure \ref{fig:11}) to the black hole is $\sim$ 3 times smaller than the last observed centroid. While we can enlarge the orbital radius by changing model parameters, it is difficult for the flare loop to show any helical motion as suggested by the observations.
We find that the orbit of the flare loop in our model does not complete a closed orbit either. The reason is that our magnetic field and non-thermal electron distribution are evolving with time
and the emission region is extended in space rather than a point source. Therefore, the emission-weighted centroid may shift from the original location after one orbit.
\subsection{Polarization}
\label{polarization}
\begin{figure*}
\includegraphics[width=0.8\textwidth]{polarization.png}
\caption{Polarized emission from the forward flux rope (up panels), backward flux rope (middle panels) and flare loop (lower panels). By doing so, the Stokes parameters $I, Q, U$ are integrated over the entire source region. The polarization period of all model is $\sim$ 20 min, approximately a half of the orbit period. }
\label{fig:15}
\end{figure*}
The polarization results are very sensitive to the configuration of the magnetic field. As we have emphasized before, the spatial distribution adopted in our model is not that realistic.
But it is still worthwhile to investigate the polarization signatures of our model and compare them with observations.
The evolution of polarization angle and fraction from our model is shown in Figure \ref{fig:15}. We plot the polarization $Q/I$ and $U/I$ in one orbital period, i.e., 40 mins.
The July 28 flare shows a rotation of polarization angle with a period of $P_{\mathrm{pol}} \approx 46 \pm 6 \min$ \citepalias{GRAVITY2018}(see also \citealt{GRAVITY2020c}). The normalized Stokes components $U/I$ and $Q/I$ trace out a loop in time, which indicates the existence of an ordered magnetic field configuration.
It is also worth noting that the observed July 22 flare exhibits a different polarization period than the other two flares. The period of its polarization swing is $p_{\mathrm{pol}} \approx 57 \pm 8$ min, which is almost twice of the orbital period.
The highly ordered polar magnetic field in our model can indeed result in a polarization angle swing, but the oscillation frequency of stokes $Q$ and $U$ components are substantially higher, with a period of $\sim$ 20 mins for all three models, just a half of the orbital period.
The overall evolution of $Q/I$ and $U/I$ for the two flux rope models are very similar; they trace nearly two full loops with same amplitude in the $Q-U$ diagram.
The evolution pattern of polarized emission for the flare loop is slightly different. They exhibit an ``outer'' and an ``inner'' loop in the $Q-U$ diagram respectively. The $Q-U$ signature will trace the outer loop when the ``hot spot'' is approaching to the observer and follow the inner loop when receding from the observer (see \citealt{Vos2022} for an explicit discussion). The forward flux loop is always moving toward the observer and backward rope always backing away from the observer hence lacking this "inner outer double-looped" feature.
Overall, we found that it is very hard to reproduce the oscillation period provided by \citetalias{GRAVITY2018} using our simple model. This is likely due to our simplified magnetic field setup taken from the solar corona mass ejection model \citep{Lin2000}. The magnetic field in the black hole accretion system, on the contrary, should be more complicated due to the strong differential rotation and turbulence. The key caveat of our magnetic field configuration is the neglect of the toroidal component of the magnetic field for simplicity. While we believe that this simplification should not affect the dynamics of the ejected flux rope, it does significantly affect the polarization result. In our next work, we will use our GRMHD numerical simulation data of \citet{Miki2022} and adopt the magnetic field configuration obtained in the simulation to revisit this problem.
Another quantity is the fraction of polarization. \citetalias{GRAVITY2018} showed a polarization degree of $20\%-40\%$. In our model, it can reach $60\%-70\%$ for the flux rope and $40\%-50\%$ for the flare loop, which is also related to our highly order magnetic field configuration adopted. This high polarization fraction is almost close to the limit set by the emission of power-law electrons in uniform magnetic field lines. For a population of power-law distributed electron with uniform magnetic field lines, the degree of polarization from synchrotron radiation is $\Pi=(p+1)/(p+7/3)$, which should be $\sim 75\%$ if $p=2.5$ or $3.5$ as in the case of our model \citep{Rybicki1986}.
It is naturally expected that the involving of turbulence in more realistic situations can change the local magnetic field configuration to some extent, and result in the field reversal in small scale, which will decrease the polarization fraction.
In addition, external Faraday rotation and conversion can lead to depolarization when the light travels through magnetized plasma, which is not taken into consideration either.
\section{Discussion}\label{sec:discussion}
\subsection{Model Parameter Exploration}
In our fiducial model, the backward and forward flux ropes have similar dynamical characteristics and non-thermal electrons, and thus should show similar intrinsic emissions. However, the backward blob is apparently dimmer on the camera screen than the forward one. As we have stated above, we suspect this discrepancy mainly results from Doppler beaming and gravitational lensing.
To diagnose their respective roles, we have performed a series of ray-tracing runs by varying the viewing angle, position and velocity of the flux rope while keeping other parameters fixed. For the inclination angle $i>90$ chosen in this work, the observer is thus located on the lower side of the equatorial plane (refer to the left plot of Figure \ref{fig:schematic}).
To test the beaming effect driven by radial velocity, the radial velocity is manually assigned while the height of the blob $h$ and its angular velocity $\Omega_{\mathrm{rp}}$ is calculated as before. The sign of radial velocity is defined to be positive if the blob is moving toward us and negative if away from us.
The effect of different radial velocities is shown in Figure \ref{fig:7} with different lines (dotted lines for the backward rope and solid lines for the forward rope). By comparing with the case with no radial velocity ($\dot{h}=0$), we see that Doppler effects can boost or de-boost the intrinsic flux by a factor of 2 for our fiducial model ($\dot{h}\sim0.3-0.4c$).
By comparing the light curves with $\dot{h}=0$ but located at two sides of the disk plane, we find that the gravitational lensing effect makes the blob behind the black hole have a greater flux. We will see this is also true for the blob with different velocities if we compare the solid and dotted lines with the same colors, even though they are now mixed with Doppler effect. The influence of lensing effect in our model is in general not so significant compared to the beaming effect. We suspect that this is because (i) the outward radial velocity is close to be relativistic and (ii) the projected position of the blob on the sky plane is relatively far away from the black hole. The former enhances the beaming effect while the latter weakens the lensing effect.
\begin{figure}
\includegraphics[width=\columnwidth]{fluxtest.png}
\caption{Light curves produced by the flux ropes with different locations and velocities. $u (l)$ denotes that the flux rope is on the upper (lower) side of the equatorial plane, $\dot{h}(-\dot{h})$ denotes that the flux rope is moving toward (away) from the observer. The initial $\phi_0=285^{\circ}$ and $\Omega=270$ for both cases. Note that for our choice of inclination angle, the camera is on the lower side of the equatorial plane. Only combinations ($u,-\dot{h}$) (backward blob) and ($l,\dot{h}$) (forward blob) are physically plausible, while others are to explore the role of Doppler boosting and gravitational lensing effect.}
\label{fig:7}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{forwardinc.png}
\includegraphics[width=\columnwidth]{backinc.png}
\caption{Light curves produced by the flux ropes with different inclination angles. The upper (lower) panel shows the forward (backward) flux rope. When the viewing angle becomes larger ($180^{\circ}-i$ is larger), the flux enhancement becomes more significant, and an apparent "double peak" structure due to beaming effect appears.}
\label{fig:lc_incl}
\end{figure}
To diagnose the influence of the inclination angle $i$, we perform a series of ray-tracing runs for the flux rope.
Figure \ref{fig:lc_incl} shows the light curves of the forward (upper panel) and backward (lower panel) flux ropes respectively with various values of inclination angle $i$. For both cases, increasing the value of $180^{\circ}-i$ will lead to the increase of the peak flux. Furthermore, depending on the model parameters, a second peak could appear in the light curve. The appearance of the first peak is due to the evolution of electron acceleration during the magnetic reconnection, while the appearance of the second peak is due to the Doppler effect. These two effects are not necessarily synchronous. This remarkable double-peaked feature in the light curve is detected in the July 28th flare. Our results are comparable to the observations in the sense of the ratio of the two peak fluxes and their time interval. We note that \citet{Ball2021} also presented a model to explain the presence of the two peaks, although their physical mechanism is different from ours.
\begin{figure}
\includegraphics[width=\columnwidth]{forwardphi.png}
\includegraphics[width=\columnwidth]{backphi.png}
\caption{Light curves produced by the forward (upper panel) and backward (lower panel) flux ropes with different $\phi_0$. The two thick lines denote the fiducial model. An interesting feature is the appearance of a second peak in the light curve for a certain range of $\phi_{0}$.}
\label{fig:9}
\end{figure}
Another interesting dependence of light curve pattern is on the initial azimuthal angle $\phi_0$. We show in Figure \ref{fig:9} the light curves of forward (upper panel) and backward (lower panel) flux ropes with different $\phi_0$. The overall variation of the light curves for the two models is similar. The time when the most significant enhancement through beaming occurs varies with $\phi_0$. Therefore, the overall profile will be sharp if the time when the external enhancement is strongest is coincident with the time when the intrinsic radiation reaches its peak, otherwise it will be flat if the two epochs are asynchronous.
\begin{figure}
\includegraphics[width=\columnwidth]{for_cen_test.png}
\includegraphics[width=\columnwidth]{back_cen_test.png}
\caption{The centroid motion of the forward (upper panel) and backward (lower panel) moving flux ropes with different inclination angles. We only present the trajectory within 40 min during the flaring epochs.}
\label{fig:13}
\end{figure}
Lastly, we present different centroid trajectories by changing the inclination angle $i$ for both flux rope models. The results are shown in Figure \ref{fig:13}.
As expected, the projected trajectory varies from close to a circular pattern to a helix as the inclination angle $i$ decreases from $180^\circ$ to $90^\circ$. This is almost the case for the forward moving flux rope where the light bending effect is less important (upper panel of Figure \ref{fig:13}). The situation for the backward moving rope is more complicated. Strong gravitational lensing makes the images of the hot spot on the camera appear as two bright spots on each side of the black hole rather than a single one (see Figure \ref{fig:12}). The flux-weighted centroid thus shows a slightly smaller centroid projected distance as shown in the lower panel of Figure \ref{fig:13}.
\subsection{Comparison with Previous Works}
\label{comparison}
By contrast to the previous works (\citetalias{Yuan2009} and \citetalias{Li2017}), the flux rope now is ejected
along a fixed conical angle $\theta_{\mathrm{B}}$ rather than along a cylinder above the equatorial plane as we have described above (see right panel of Figure \ref{fig:schematic}). The new scenario agrees well with the flux rope trajectory found in the 3D GRMHD simulation of \citet{Miki2022}\footnote{\citet{Miki2022} find that collimation effect exists for the ejected flux rope, i.e, the corresponding value of $\theta$ in spherical coordinate of the flux rope gradually decreases with its outward propagation. Our definition of $\theta$ is different from the standard definition in a spherical coordinate because the flare loop is not located at the position of the black hole (refer to Figure \ref{fig:schematic}). Because of this reason, our constant-$\theta_B$ model effectively has taken into account the collimation effect. }. A similar ejected flux rope scenario has been studied by \citet{Younsi2015} and \citet{Ball2021}. By analogy with the same CME model, \citet{Younsi2015} considered the case where the flux rope is ejected out but confined on a cylinder surface. Therefore, we expect that their model is difficult to be compatible with the super-Keplerian motion of the observed hot spot.
Our flux rope trajectory is also different from work of \citet{Ball2021}. They considered a phenomenological hot spot model and the spot is moving along a fixed $\theta_0$ in spherical coordinates with the generatrix of the cone points towards to the central black hole.
There are also some GRMHD simulations aimed at interpreting the flares from Sgr A$^{*}$. In \citet{Porth2021}, the magnetically dominant plasmoids are embedded within the accretion flow, and the motion of these plasmoids is thus substantially sub-Keplerian, which is in tension with the GRAVITY observations. \citet{Petersen2020} also focus on the accretion flow rather than ejection. Their model can reproduce the flux and spectrum of the IR flares, but fails to explain the centroid motion of the hot spot observed by GRAVITY. In some GRMHD simulation works, it has been found that the plasmoids can be be ejected out from the accretion flow \citep{Nathanail2022,Ripperda2022}, but a quantitative calculation of the radiative transfer and comparison with GRAVITY observations in terms of the centroid motion and flare light curves usually lacked in these works.
\subsection{Implications to the observed quasi-periodic variability of the NIR flares from Sgr A*}
While the present paper focuses on the interpretation of the GRAVITY observational results, it is interesting to discuss other infrared observations to the flares of Sgr A$^{*}$. Among these observations, one prominent result we think presents an important constraint on the nature of Sgr A* flare is the periodicity of the flares detected by the high-resolution infrared observations of \citet{Genzel2003}. They have performed the power spectrum analysis to two light curves lasting $\sim 130$ minutes and found a significant peak with a time period of $\sim 17$ minutes. The periodicity is confirmed later by other infrared observations \citep{2006A&A...455....1E,2007MNRAS.375..764T,2010RvMP...82.3121G}\footnote{ We note that some observations do not find such a periodic component from the light curves and the variability is fully consistent with the red-noise \citep[e.g.,][]{2009ApJ...691.1021D}. As pointed out by \citet{2010RvMP...82.3121G}, the apparent discrepancy seems to be the
question whether it is reasonable for the observer to select certain parts of the light curve to analyze the significance of the periodicity, which may be washed out in the noise-like variable state of longer
time series.}, although the period changes somewhat from case to cases. This period is consistent with the orbital period at 3–5$r_g$, assuming the accretion flow in Sgr A* is described by a magnetically arrested disk (MAD).
Such a quasi-periodic variability is hard to be explained by the orbital motion of a single hot spot. Our detailed modeling shown in the present paper indicates that the flare due to a single hot spot usually lasts for less than one orbital timescale. The physical reasons are: (i) As we have discussed above, synchrotron cooling timescale for these electrons is very short in the high magnetic field environment. So the duration of the flare is determined by the magnetic reconnection, which cannot last for too long either. (ii) The ejected blob can rush out to a very large distance in a short time, which causes the magnetic field embedded in the blob drops rapidly. Therefore, the magnitude of the radiation would decline and even becomes weaker than the threshold of the detection limit.
Instead, we suggest that the periodic formation of the flux ropes found in the numerical simulations of \citet{Miki2022} may be responsible for the quasi-periodic variability. It is found that flux ropes are formed periodically, with the period being the local orbital timescale. Specifically, they found that the flux ropes formed within $10-15\ r_g$ are hard to be ejected out but stay within the accretion flow and finally advected into the black hole horizon. So it is predicted that, the flares with short period of variability as found in \citet{Genzel2003} should not be accompanied by blob ejections, thus their trajectories projected on the sky should not become larger with time.
\section{Summary}
\label{sec:6}
High-resolution GRAVITY observations have resolved the infrared flares of Sgr A* and a compact “hot
spot” was detected orbiting around the black hole with a high speed. Among other features, two prominent results obtained in the observations are that: 1) the trajectory of the hot spot gradually increases with time; and 2) the radius of the circular trajectory combined with the orbital time indicates that the hot spot is rotating with super-Keplerian speed \citep{GRAVITY2018}. These results present important constraint on the nature of flares in Sgr A$^{*}$.
\citet{Yuan2009} has proposed a ``CME'' model to explain the flares by analogy with the coronal mass ejection in the Sun. The basic scenario is that, magnetic reconnection at the surface of the accretion flow results in the formation of flux ropes, which are then ejected out with high velocity. The scenario has been confirmed by our recent GRMHD simulations \citep{Miki2022}. Electrons accelerated in the reconnection current sheet escape to the regions of the flare loop and of the ejected flux rope, whose radiation are responsible for the flares. The first application of the model to the interpretation of Sgr A* was presented in \citet{Li2017}. In that work, the dynamics of the ejected blobs, the energy distribution of accelerated nonthermal electrons and their radiation were calculated and compared with the light curve and spectrum of infrared and X-ray flares. It was found that both the radiation from the ejected blob and flare loop provided the comparable contribution to the observed flares.
In the present paper, we further develop the \citet{Li2017} model by also taking into account the toroidal motion of the ejected blobs in addition to its poloidal motion, which was neglected in \citet{Li2017} for simplicity (refer to Figure \ref{fig:schematic} for the schematic figure).
In our model, the ejected flux rope follows a spiral motion roughly confined in the surface of a cone. Therefore the distance of the hot spot from the black hole in the projected plane gradually becomes larger with time. Moreover, as suggested by the numerical simulations of \citet{Miki2022}, the angular velocity of the blob roughly remains the same value as its launching location as long as it is not too far away from this launching location. Thus, with the increase of radius in the spiral motion, its rotation becomes super-Keplerian, fully consistent with \citet{GRAVITY2018}. We consider the electron acceleration by reconnection and the evolution of the energy distribution of accelerated electrons. The radiation calculation is elaborated by carrying out fully general relativistic ray-tracing radiative transfer calculations, which is another main improvement compared to \citet{Li2017}.
We have considered three components as the candidate model of Sgr A* flares, namely the forward and backward motion flux ropes and the flare loop. We find that the radiation from the ejected blobs will dominate over that from the flare loop. The forward flux rope can well reproduce both the light curve and centroid motion for the July 22 flare reported by \citet{GRAVITY2018}. Based on the trajectory of the hot spot, we can rule out the flare loop model due to a much smaller, and nearly circular pattern of the centroid motion. But we cannot fully rule out the backward flux rope since this needs a thorough parameter survey, which is out the scope of the current work. In addition, the flux rope models can generate double-peak structure in the light curve by adjusting the inclination angle or the initial azimuthal angle where the rope is launched. This could be responsible for another flare observed by GRAVITY on July 28 \citep{GRAVITY2018}.
One main caveat of our current model is that the magnetic field configuration is adopted from solar coronal mass ejection model, which is in order to obtain an analytical solution of the dynamics of the ejected flux rope. Such a highly ordered magnetic field configuration however overestimates the degree of polarization and produces a shorter polarization swing period. We expect that the future modeling by making use of the GRMHD simulation data with a more reasonable field configuration would alleviate this discrepancy.
\section*{Acknowledgements}
This work is supported in part by the Natural Science Foundation of China (grants 12133008, 12192220, and 12192223). We thank the beneficial discussions with Hai Yang, Haocheng Zhang, Antonios Nathanail, and useful comments from the referee. The calculations have made use of the High Performance Computing
Resource in the Core Facility for Advanced Research Computing
at Shanghai Astronomical Observatory.
\section*{Data availability}
The data underlying this article will be shared on reasonable request to the corresponding author.
\bibliographystyle{mnras}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.